text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'The non-linearities of the dynamics of Earth artificial satellites are encapsulated by two formal integrals that are customarily computed by perturbation methods. Standard procedures begin with a Hamiltonian simplification that removes non-essential short-period terms from the Geopotential, and follow with the removal of both short- and long-period terms by means of two different canonical transformations that can be carried out in either order. We depart from the tradition and proceed by standard normalization to show that the Hamiltonian simplification part is dispensable. Decoupling first the motion of the orbital plane from the in-plane motion reveals as a feasible strategy to reach higher orders of the perturbation solution, which, besides, permits an efficient evaluation of the long series that comprise the analytical solution.'
author:
- Martin Lara
date: ' draft of April 16, 2020'
title: Solution to the main problem of the artificial satellite by reverse normalization
---
Introduction
============
The dynamics of close Earth satellites under gravitational effects are mostly driven by perturbations of the Keplerian motion induced by the Earth oblateness. For this reason, the approximation obtained when truncating the Legendre polynomials expansion of the Geopotential to the only contribution of the zonal harmonic of the second degree, whose coefficient is usually denoted $J_2$, is traditionally known as the *main problem* of artificial satellite theory. The $J_2$-truncation of the gravitational potential is known to give rise to nonintegrable dynamics [@Danby1968; @IrigoyenSimo1993; @Broucke1994; @CellettiNegrini1995] that comprise short- and long-period effects, as well as secular terms [@Kozai1959; @Kaula1961]. However, due to the smallness of the $J_2$ coefficient of the Earth, the full system can be replaced by a separable approximation, which is customarily obtained by removing the periodic effects by means of perturbation methods [@Nayfeh2004].
When written in the action-angle variables of the Kepler problem, also called Delaunay variables, the main problem Hamiltonian immediately shows that the right ascension of the ascending node is a cyclic variable. In consequence, its conjugate momentum, the projection of the angular momentum vector along the Earth’s rotation axis, is an integral of the main problem, which, therefore, is just of two degrees of freedom. Then, following Brouwer [@Brouwer1959], the main problem Hamiltonian is normalized in two steps. First, the short period effects are removed by means of a canonical transformation that, after truncation to some order of the perturbation approach, turns the conjugate momentum to the mean anomaly (the Delaunay action) into a formal integral. Next, a new canonical transformation converts the total angular momentum into another formal integral. The main problem Hamiltonian is in this way completely reduced to a function of only the momenta in the new variables, whose Hamilton equations are trivially integrable.
In spite of the normalization procedure is standard from the point of view of perturbation theory, it happens that not all the action-angle variables of the Kepler problem appear explicitly in the Geopotential, as is usually advisable in the construction of perturbation solutions [@Kinoshita1972; @LaraFerrer2013; @LaraFerrer2015; @Lara2018]. In particular, the mean anomaly remains implicit in the gravitational potential through its dependence on the polar angle. Unavoidably, Kepler’s equation must be solved to show the mean anomaly explicit, a fact that entails expanding the elliptic motion in powers of the eccentricity [@Hansen1855; @BrouwerClemence1961]. This standard procedure is quite successful when dealing with orbits of low eccentricities, like is typical in a variety of astronomy problems [@Delaunay1860; @Tisserand1889]. However, it involves handling very long Fourier series in the case of orbits with moderate eccentricities [@DepritRom1970; @Kinoshita1977SAOSR], and hence the application of this method to different problems of interest in astrodynamics is de facto prevented.
On the contrary, the integrals appearing in the solution of the artificial satellite problem can be solved in closed form of the eccentricity [@Brouwer1959; @Kozai1959]. It only requires the help of the standard relation between the differentials of the true and mean anomalies that is derived from the preservation of the total angular momentum of the Keplerian motion. Regrettably, the closed form approach soon finds difficulties in achieving higher orders of the short-period elimination, which stem from the impossibility of obtaining the antiderivative of the equation of the center (the difference between the true and mean anomalies) in closed form of the eccentricity in the realm of trigonometric functions [@Jefferys1971; @OsacarPalacian1994]. Nonetheless, the difficulties are overcome by the artifact of grouping the equation of the center with other functions appearing in the procedure previously to approaching their integration [@Kozai1962; @Aksnes1971; @Deprit1982]. Alternatively, the application of a preliminary Hamiltonian simplification, the *elimination of the parallax* [@Deprit1981; @LaraSanJuanLopezOchoa2013b], eases the consequent removal of short-period effects to some extent [@CoffeyDeprit1982; @Healy2000].
From the point of view of the perturbation approach, removing the short-period effects in the first place seems the more natural in view of the degeneracy of the Kepler problem. Indeed, the Kepler Hamiltonian in action-angle variables, on which the perturbation approach hinges on, only depends on the Delaunay action [@GoldsteinPooleSafko2001; @LaraTossa2016]. However, the order in which the formal integrals are sequentially introduced in the perturbation solution is not relevant in a total normalization procedure, whose result is unique [@Arnold1989]. In fact, it happens that relegating the transformation of the Delaunay action into a formal integral to the last step of the perturbation approach provides clear simplifications in dealing with the equation of the center [@AlfriendCoffey1984]. In this way the task of extending the solution of the main problem to higher orders is notably simplified.
Converting the total angular momentum into a formal integral of the main problem requires making cyclic the argument of the perigee, up to some truncation order of the perturbation approach, in the transformed Hamiltonian. However, as it appeared in the literature, the transformation called by their authors the *elimination of the perigee* [@AlfriendCoffey1984] is not the typical normalization procedure, although it operates analogous results. Indeed, on the one hand, the elimination of the perigee is only applied to a Hamiltonian obtained after the elimination of the parallax, to which simplification it could seem to be unavoidably attached. On the other hand, when extended to higher orders, it removes more terms than those needed in a partial normalization. Last, the fact that the technique was originally devised in the canonical set of polar variables, to which the argument of the perigee does not pertain, neither helps in grasping the essence of the transformation. Reimplementation of the procedure in the usual set of action-angle variables makes the process of converting the argument of the perigee into a cyclic variable much more evident [@LaraSanJuanLopezOchoa2013c], but it still bears the same differences with respect to a classical normalization procedure.
We disregard the claimed benefits of Hamiltonian simplification procedures and compute the solution to the main problem of the artificial satellite theory by standard normalization. It is called *reverse* normalization because we exchange the order in which the formal integrals are traditionally introduced when solving the artificial satellite problem. More precisely, the total angular momentum is transformed into a formal integral in the first place, in this way decoupling the motion of the orbital plane from the satellite’s motion on that plane.[^1] Then, a second canonical transformation converts the mean anomaly into a cyclic variable, in this way achieving the total reduction of the main problem Hamiltonian.
The procedure for making the argument of the perigee cyclic in the first place follows an analogous strategy to the one devised in the classical elimination of the perigee transformation [@AlfriendCoffey1984]. However, in our approach it is applied directly to the original main problem Hamiltonian, and differs from the original technique, as well as from an analogous procedure carried out in [@SanJuanetal2013], in which the parallactic terms (inverse powers of the radius with exponents higher than 2) are not removed from the new, partially normalized Hamiltonian. In spite of that, we did not find trouble in dealing with the equation of the center in closed form in the subsequent Delaunay normalization [@Deprit1982], a convenience that might had been anticipated from the discussions in [@Lara2019CMDA].
The Hamiltonian reduction has been approached in Delaunay variables. Unfortunately, these variables share the deficiencies of their partner Keplerian elements, which are singular for circular orbits and for equatorial orbits. Because of that, the secular terms are reformulated in a set of non-singular variables that replaces the mean anomaly, the argument of the perigee, and the total angular momentum, by the mean argument of the latitude and the projections of the eccentricity vector in the nodal frame, which are sometimes denoted *semi-equinoctial* variables [@Konopliv1990]. For the periodic corrections, we find convenience in using polar variables, which in the particular case of the main problem are free from small divisors of any kind except for those related to the critical inclination resonance, and are known to produce compact expressions that yield faster evaluation [@Izsak1963; @LaraVilhenaSanchezPrado2015].
We extended the complete normalization to the order six of the perturbation approach, which, to our knowledge, is the maximum order that has been reported in the literature (yet limited to partial normalization cases) [@Healy2000; @SanJuanetal2013]. The aim of computing such a high order is not to enter a competition. On the contrary, we did it simply because the particular value of the Earth’s $J_2$ coefficient is $\sim10^{-3}$, and hence the computed solution should be exact to the numerical precision of standard floating-point arithmetic [@Kahan1997] already at the fifth order. We checked that it is exactly the case, and extending the computations to the sixth order only served us to verify that we don’t find observable improvements in our tests. In order to compute this high-order solution we resorted to the practicalities of standard commercial software, in which Deprit’s perturbation algorithm based on the Lie transforms method [@Deprit1969] is easily implemented. On account of the current computational power, it should be quite feasible to extend the perturbation solution, if desired, to even higher orders, although we didn’t try that.
On the other hand, making cyclic the argument of the perigee with the standard normalization in action-angle variables seems to be a particularly efficient procedure from the computational point of view, resulting in a generating function whose size is astonishingly smaller than alternative proposals in the literature. This fact, combined with the reduction in the number of transformations required by the perturbation solution to just two, as opposite to the three transformations needed when the elimination of the parallax is carried out in the first place, might make this latter simplification dispensable, as well as the discussions in [@DepritMiller1989] questionable, at least in what respects to the computation of explicit analytical solutions as opposite to partial normalizations to be integrated semianalytically.
Tests carried out on typical Earth orbits of different types confirm that the computed solution bears exactly the expected characteristics of a perturbation solution of perturbed Keplerian motion. In particular, we verified the degree of convergence of successive approximations. We confirmed too that, as expected, the quality of the analytical solution degrades in the vicinity of the critical inclination resonance —a case that, of course, would require a specific treatment [@CoffeyDepritMiller1986; @Jupp1988; @Lara2015IR]. In addition, we checked that the computation of the initialization constants of the analytical solution from corresponding initial conditions gets a clear benefit of using a higher order of the periodic corrections than the order needed for ephemeris evaluation, yet this additional accuracy is not needed in all the variables and can be limited to the computation of the formal integral given by the Delaunay action [@BreakwellVagners1970; @DepritRom1970; @HautesserresLaraCeMDA2017].
Construction of the analytical solution
=======================================
The solution of Laplace’s equation in spherical coordinates provides the gravitational potential in the form of an expansion in harmonic functions. In the case of the Earth, the mass distribution is almost symmetric with respect to the rotation axis. For this reason, the Geopotential is simplified in different applications to just the contribution of the zonal harmonics (see [@Danby1992], for instance) $$\label{zonalpot}
\mathcal{V}=\frac{\mu}{r}\sum_{m\ge{0}}\frac{R_\oplus^m}{r^m}J_mP_{m}(\sin\varphi),$$ where $r$ is distance from the origin, $\varphi$ is latitude, $\mu$ is the Earth’s gravitational parameter, $R_\oplus$ is the Earth’s mean equatorial radius, $J_m$ are zonal harmonic coefficients, and $P_{m}$ denote Legendre polynomials of degrees $m$.
The flow stemming from the potential (\[zonalpot\]) admits Hamiltonian formulation. Besides, because $J_2\sim10^{-3}$ whereas $J_m$ ($m>2$) are $\mathcal{O}(J_2^2)$ or smaller, the Hamiltonian $$\label{Hammain}
\mathcal{H}=-\frac{\mu}{2a}-\frac{\mu}{r}\frac{R_\oplus^2}{r^2}J_2\frac{1}{2}\left(1-3\sin^2\varphi\right),$$ is representative of the main characteristics of the dynamics of close Earth satellites, and is commonly known as the *main problem* of artificial satellite theory. The symbol $a$ in Eq. (\[Hammain\]) stands for the orbit semi-major axis, and, from standard trigonometric relations, $\sin\varphi=\sin{I}\sin\theta$, where $I$ and $\theta$ are used to denote the inclination and the argument of the latitude, respectively.
When using the classic set of Keplerian elements, the argument of the latitude is $\theta=f+\omega$, where $f$ and $\omega$ denote the true anomaly and the argument of the periapsis, respectively. The radius is computed from the conic equation $$\label{conic}
r=\frac{p}{1+e\cos{f}},$$ where $p=a(1-e^2)$ is the parameter of the conic and $e$ is the eccentricity of the orbit. Note, however, that, because we are using Hamiltonian formulation, the symbols appearing in Eq. (\[Hammain\]) need to be expressed as functions of some set of canonical variables. While the Keplerian variables are not canonical, they are conveniently expressed in the set of Delaunay canonical variables $(\ell,g,h,L,G,H)$, in which $\ell$ is the mean anomaly, $g=\omega$, $h$ is the argument of the ascending node, $L=\sqrt{\mu{a}}$ is known as the Delaunay action, $G=L\sqrt{1-e^2}$ is the total angular momentum, and $H=G\cos{I}$ is the projection of the angular momentum along the Earth’s rotation axis. That the latter is an integral of the main problem becomes evident after checking that $h$ is an ignorable variable in Hamiltonian (\[Hammain\]), which, in consequence, is of just two degrees of freedom.
Besides, the main problem Hamiltonian is conservative, and, therefore, Eq. (\[Hammain\]) remains constant (the total energy) for a given set of initial conditions. On the other hand, the existence of a third integral has not been proved. Then, exact solutions of the main problem are not known. Alternatively, it is approached with the usual tools of non-linear dynamics, such as Poincaré surfaces of section or the computation of families of periodic orbits [@Danby1968; @Broucke1994; @SADSaM1999; @Lara1999]. Still, in some regions of phase space, and for particular energy values, the third integral can be constructed formally with the help of perturbation methods, obtaining in this way useful analytical approximations to the main problem dynamics in these particular regions.
Perturbation approach
---------------------
We rewrite the main problem in the form of a perturbation Hamiltonian $$\label{mainP}
\mathcal{H}=\mathcal{H}(\ell,g,-,L,G,H)\equiv\sum_{m\ge0}\frac{\varepsilon^m}{m!}K_{m,0},$$ in which $\varepsilon$ is a formal small parameter ($\varepsilon=1$) used to manifest the strength of each summand of Eq. (\[mainP\]), and $$\begin{aligned}
\label{mainPK0}
K_{0,0} &=& -\frac{\mu}{2a}, \\ \label{mainPK1}
K_{1,0} &=& -\frac{\mu}{r}\frac{R_\oplus^2}{r^2}\frac{1}{4}J_2\left[2-3s^2+3s^2\cos(2f+2\omega)\right], \\ \label{mainPKm}
K_{m,0} &=& 0, \quad m\ge2.\end{aligned}$$ The symbol $s$ in Eq. (\[mainPK1\]) stands for the usual abbreviation of the sine of the inclination.
The aim is to find a transformation of variables $$(\ell,g,h,L,G,H;\varepsilon)\mapsto(\ell',g',h',L',G',H'),$$ given also by an expansion in powers of the small parameter, such that, up to some truncation order $m=k$, the Hamiltonian in the new variables is transformed into a separable Hamiltonian. Namely, $$\label{mainPp}
\mathcal{H}'=\sum_{m=0}^{k}\frac{\varepsilon^m}{m!}K_{0,m}(-,-,-,L',G',H)+\mathcal{O}(\varepsilon^{k+1}).$$
The desired transformation is derived from a scalar generating function $$\label{Lie:Gen}
\mathcal{W}=\sum_{m\ge0}\frac{\varepsilon^m}{m!}W_{m+1}.$$ which, in our case, is computed using Deprit’s perturbation algorithm based on Lie transforms [@Deprit1969]. The procedure is summarized in finding a particular solution of the *homological equation* $$\label{homoeq}
\mathcal{L}_0(W_m)\equiv\{W_m;K_{0,0}\}=\widetilde{K}_{0,m}-K_{0,m},$$ in which $\{W_m;K_{0,0}\}$ stands for the computation of the Poisson bracket of $W_m$ and the zeroth order term of the Hamiltonian. Terms $\widetilde{K}_{0,m}$ are known from the original Hamiltonian as well as from previous computations to the step $m$, which are carried out using Deprit’s fundamental recursion $$\label{Lie:triangle}
{F}_{n,q}={F}_{n+1,q-1}+
\sum_{m=0}^n{n\choose{m}}\{{F}_{n-m,q-1};W_{m+1}\}.$$ Equation (\[Lie:triangle\]) can be used to formulate any function $\mathcal{F}$ of the canonical set of original variables like a function of the new variables, the Hamiltonian (\[mainP\]) being just one among the different possibilities. Finally, the term $K_{0,m}$ is chosen at our will, with the only condition of making the homological equation solvable in $W_m$.
When using Delaunay variables the homological equation of the main problem Hamiltonian is solved by indefinite integration. Indeed, Eq. (\[mainPK0\]) turns into $K_{0,0}=-\frac{1}{2}\mu^2/L^2$, and hence the *Lie derivative* operator $\mathcal{L}_0$ takes the extremely simple form $$\mathcal{L}_0(W_m)=n\frac{\partial{W}_m}{\partial\ell},$$ in which $n=\sqrt{\mu/a^3}=\mu^2/L^3$ is the mean motion of the unperturbed problem. Then, from Eq. (\[homoeq\]), $$\label{homosoll}
W_m=\frac{1}{n}\int(\widetilde{K}_{0,m}-K_{0,m})\,\mathrm{d}\ell+C_m,$$ where the functions $C_m\equiv{C}_m(-,g,L,G;H)$ play the role of integration “constants” that verify $\mathrm{d}C_m/\mathrm{d}\ell=0$. Equation (\[homosoll\]) is solved in closed form of the eccentricity with the help of the differential relation between the true and mean anomalies provided by the preservation of the angular momentum of the Kepler problem. That is, $G=r^2\mathrm{d}f/\mathrm{d}t$, from which $$\label{dl2df}
a^2\eta\,\mathrm{d}\ell=r^2\mathrm{d}f,$$ where $\eta=\sqrt{1-e^2}=G/L$ is usually known as the *eccentricity function*. Hence,$$\label{homosol}
W_m=\frac{1}{n}\int(\widetilde{K}_{0,m}-K_{0,m})\frac{r^2}{a^2\eta}\,\mathrm{d}f+C_m.$$
The transformation that makes the main problem Hamiltonian separable, up to the truncation order, is split into two different canonical transformations. With the first one $$(\ell,g,h,L,G,H;\varepsilon)\mapsto(\ell',g',h',L',G',H'),$$ we make the argument of the perigee cyclic, thus converting the total angular momentum into a formal integral. That is, the motion of the satellite in the orbital plane, whose inclination remains constant in the new variables, is decoupled from the motion of the orbital plane. Therefore, the reduced system representing the motion on the orbital plane could be integrated separately. Rather, we carry out a second canonical transformation $$(\ell',g',h',L',G',H';\varepsilon)\mapsto(\ell'',g'',h'',L'',G'',H''),$$ that makes ignorable the mean anomaly in the transformed Hamiltonian in double-prime variables. In this way, the Delaunay action is converted into a formal integral too, and the complete Hamiltonian reduction is achieved up to the truncation order of the perturbation solution, whose Hamilton equations are thus trivially integrable.
The computation of the secular terms from the normalized Hamiltonian is only part of the perturbation solution. To be complete, it requires the correct initialization of the constants of the perturbation solution from corresponding initial conditions using the *inverse* transformation (from original to double prime variables), on the one hand, and the recovery of periodic effects using the *direct* transformation (from double prime to original variables) to obtain the ephemeris corresponding to the secular terms propagation. Both the direct and inverse transformations are obtained from standard application of the Lie transforms method. Because there are no specific artifices related to the computation of the current solution in what respects to that part, we do not provide details on their computation and an interested reader is referred to the literature [@Deprit1969].
Normalization of the total angular momentum
-------------------------------------------
At the first order, we check from Eq. (\[Lie:triangle\]) that the known terms involved in the homological equation consist only of $\widetilde{K}_{0,1}=K_{1,0}$. Therefore, we chose the new Hamiltonian term $$\label{K01g}
K_{0,1}=-\frac{\mu}{r}\frac{R_\oplus^2}{r^2}\frac{1}{4}J_2\left(2-3s^2\right),$$ which is the part of Eq. (\[mainPK1\]) that is free from the argument of the perigee, as desired. Then, Eq. (\[homosol\]) turns into $$W_1=-\frac{3}{4}GJ_2\frac{R_\oplus^2}{p^2}s^2\int\frac{p}{r}\cos(2f+2\omega)\,\mathrm{d}f+C_1,$$ which is solved by standard integration after replacing the radius with the conic equation (\[conic\]). We obtain $$\label{W1g}
W_1=-\epsilon G \frac{1}{2}s^2\sum_{i=1}^3 3^{\left\lfloor2-\frac{1}{2}i\right\rfloor } e^{\left| i-2\right| } \sin (if+2\omega)+C_1,$$ where $C_1$ is hold arbitrary by now, the symbol $\lfloor\;\rfloor$ stands for integer part, and we abbreviated $$\label{mieps}
\epsilon=\frac{1}{4}J_2\frac{R_\oplus^2}{p^2},$$ in which $p=G^2/\mu$, and, therefore $\epsilon$ is a function of the total angular momentum. Recall that the symbols $p$, $s$, $e$, $\omega$, and $f$ in Eqs. (\[K01g\]), (\[W1g\]), and (\[mieps\]) are functions of the Delaunay variables.
At the second order we compute $K_{0,2}$ from Eq. (\[Lie:triangle\]), to obtain $$\label{triangle2}
K_{0,2}=\{K_{0,1},W_1\}+K_{1,1},$$ in which $K_{1,1}$ is computed using again Eq. (\[Lie:triangle\]). Namely, $$\label{K11}
K_{1,1}=\{K_{0,0},W_2\}+\{K_{1,0},W_1\}+K_{2,0}.$$ On account of $K_{2,0}=0$ from Eq. (\[mainPKm\]), the known terms hitherto of the homological equation (\[homoeq\]) are $$\widetilde{K}_{0,2}=\{K_{0,1},W_1\}+\{K_{1,0},W_1\},$$ whose computation only involves partial differentiation. Before approaching the solution of Eq. (\[homosol\]), we first check that the term $\widetilde{K}_{0,2}$ is made of trigonometric coefficients $T_i$ whose arguments always involve the true anomaly as argument, except for the terms $$\begin{aligned}
T_0 &=& \epsilon^2\frac{3}{4}\frac{\mu}{r}\frac{p}{r}s^2\left[e^2(23s^2-16\right)+8 \left(4 s^2-3)\right],
\\
T_1 &=&-\frac{3}{2}\epsilon^2G^2\frac{1}{r^2}(15s^2-14)s^2e^2\cos2\omega,
\\
T_2 &=& 6\epsilon{G}\frac{1}{r^2}(5s^2-4)\frac{\partial{C}_1}{\partial{g}}.\end{aligned}$$ Integration of these 3 terms in Eq. (\[homosol\]) would yield corresponding terms that grow unbounded with $f$. The term $T_0$ is free from the argument of the periapsis, and hence it can be cancelled by choosing the new Hamiltonian $K_{0,2}$ having a corresponding term $T_0$. On the contrary, the terms $T_1$ and $T_2$ depend on the argument of the perigee, a fact that prevents their appearance in the new Hamiltonian under our requirement of making $g$ a cyclic variable. Nevertheless, since we had left $C_1$ arbitrary, both terms will cancel each other if $C_1$ is determined from the partial differential equation $T_1+T_2=0$. We readily check that the particular solution $$\label{C1g}
C_1=\epsilon{G}\frac{15s^2-14}{8(5s^2-4)}s^2e^2\sin2\omega,$$ matches this purpose. Remark that the appearance of the divisor $5s^2-4$ prevents application of the solution to the resonant cases that happen at the so-called *critical inclinations* in which $\sin^2I=4/5$. That is, the supplementary inclinations $I=63.435^\circ$ and $I=116.565^\circ$.
After computing the partial derivatives of Eq. (\[C1g\]) with respect to $L$ and $G$, which also appear among the coefficients of the terms remaining in $\widetilde{K}_{0,2}$, the known terms of the homological equation become fully determined. Then, we can safely choose $K_{0,2}$ so that it comprises those terms of $\widetilde{K}_{0,2}$ that are free from the argument of the periapsis. We obtain, $$\label{K02g}
K_{0,2}=\epsilon^2\frac{\mu}{r}\frac{p^2}{r^2}\frac{3s^2}{8(5s^2-4)^2}\sum_{j=0}^2\frac{p^j}{r^j}\sum_{k=0}^{\lfloor1-\frac{1}{2}j\rfloor}e^{2k}\gamma_{2,j,k},$$ in which the inclination polynomials $\gamma_{2,j,k}\equiv\gamma_{2,j,k}(s)$ are provided in Table \[t:K2gpoly\].
[@ll@]{} ${}_{0,0}$ & $-8 \left(200 s^6-455 s^4+345 s^2-88\right)$\
${}_{0,1}$ & $375 s^6-930 s^4+780s^2-224$\
${}_{1,0}$ & $5 \left(805 s^6-1878 s^4+1464 s^2-384\right)$\
${}_{2,0}$ & $-825 s^6+1990s^4-1616 s^2+448$\
The second order term of the generating function is then readily computed from Eq. (\[homosol\]), yielding $$\begin{aligned}
\nonumber
W_2 &=& \frac{\epsilon^2G}{32(5s^2-4)^2}\sum_{l=1}^2\sum_{\substack{k=l-2\\k\ne0}}^{2l+2}\sum_{j=0}^1\Gamma_{2,j,k,l}e^{2j+k^*} \\ \label{W2g}
&& \times s^{2l}\sin(kf+2l\omega)+C_2,\end{aligned}$$ with $k^*\equiv{k}\bmod2$, and the inclination polynomials $\Gamma_{2,j,k,l}$ are given in Table \[t:W2gpoly\].
[@ll@]{} ${}_{1,-1,1}$ & $-12(5 s^2-4) (7 s^2-6) (15s^2-14)$\
${}_{0,1,1}$ & $-48(5 s^2-4) (195 s^4-340s^2+148)$\
${}_{1,1,1}$ & $24(5 s^2-4)^2 (15 s^2-14)$\
${}_{1,1,2}$ & $-3 (225 s^4-430 s^2+208)$\
${}_{0,2,1}$ & $-96 (5s^2-4)^2 (9 s^2-8)$\
${}_{1,2,1}$ & $-24 (5 s^2-4) (65s^4-116 s^2+52)$\
${}_{1,2,2}$ & $-60 (50 s^4-87 s^2+38)$\
${}_{0,3,1}$ & $-64 (5 s^2-4)^2 (8 s^2-7)$\
${}_{1,3,1}$ & $4 (3s^2-2) (5 s^2-4) (15 s^2-14)$\
${}_{0,3,2}$ & $-4 (5s^2-4) (135 s^2-122)$\
${}_{1,3,2}$ & $-8 (75 s^4-135s^2+61)$\
${}_{1,4,1}$ & $-12 (5 s^2-4)^2 (7 s^2-6)$\
${}_{0,4,2}$ & $24 (5 s^2-4)^2$\
${}_{1,4,2}$ & $-12 (5 s^2-4) (25s^2-23)$\
${}_{0,5,2}$ & $24 (5 s^2-4)^2$\
${}_{1,5,2}$ & $-3 (5s^2-4) (15 s^2-14)$\
${}_{1,6,2}$ & $6 (5 s^2-4)^2$\
${}_{0,0,2}$ & $(15 s^2-14)^2(15s^2-13)$\
${}_{0,0,1}$ & $8(5s^2-4)^2(1215s^4-1997s^2+824)$\
${}_{1,0,1}$ & $-2(5s^2-4)(15s^2-14)(45s^4+36s^2-56)$\
Analogously as we did with $C_1$, the arbitrary function $C_2$ in which Eq. (\[W2g\]) depends upon is determined at the next step of the perturbation approach in such a way that the appearance of unbounded terms in the solution of the homological equation of the third order are avoided. Thus, we compute $\widetilde{K}_{0,3}$ by successive evaluations of the fundamental recursion (\[Lie:triangle\]). After identifying the problematic terms of $\widetilde{K}_{0,3}$, they are cancelled by computing $$\label{C2g}
C_2=\frac{\epsilon^2G}{64(5s^2-4)^3}\sum_{l=1}^2\sum_{j=0}^{2-l}\Gamma_{2,j,0,l}e^{2(j+l)}s^{2l}\sin2l\omega,$$ which contributes three additional inclination functions that are also listed in Table \[t:W2gpoly\]. In this way, the second order term of the generating function given by Eq. (\[W2g\]) becomes fully determined.
The needed partial derivatives of $C_2$ appearing in $\widetilde{K}_{0,3}$ are then computed, and the third order term of the new Hamiltonian is chosen, as before, to comprise those terms of $\widetilde{K}_{0,3}$ that are free from the argument of the periapsis. We obtain $$\label{K03g}
K_{0,3}=\epsilon^3\frac{\mu}{r}\frac{p^2}{r^2}\frac{3s^2}{32(5s^2-4)^3}\sum_{j=0}^4\frac{p^j}{r^j}\sum_{k=0}^{\lfloor2-\frac{1}{2}j\rfloor}e^{2k}\gamma_{3,j,k},$$ where the inclination polynomials $\gamma_{3,j,k}$ are displayed in Table \[t:K3g\].
[@ll@]{} ${}_{0,0}$ & $-8(5s^2-4)(313525 s^8-899030 s^6+933656 s^4$\
& $-409296 s^2+61824)$\
${}_{0,1}$ & $4(1551625 s^{10}-5675700 s^8+8148960 s^6-5706408 s^4$\
& $+1930272 s^2-248064)$\
${}_{0,2}$ & $-2 (40500 s^{10}-99525 s^8+64840 s^6+18788 s^4$\
& $-33936 s^2+9408)$\
${}_{1,0}$ & $2 (5 s^2-4) (2631475 s^8-7558270 s^6+7872692 s^4$\
& $-3470616 s^2+530304)$\
${}_{1,1}$ & $-3457125 s^{10}+12282750 s^8-17085020 s^6$\
& $+11554040 s^4-3756000 s^2+459648$\
${}_{2,0}$ & $-2 (5 s^2-4) (1584375 s^8-4536150 s^6+4716436 s^4$\
& $-2082712 s^2+321408)$\
${}_{2,1}$ & $138375 s^{10}-128250 s^8-351900 s^6+612440 s^4$\
& $-326368 s^2+56448$\
${}_{3,0}$ & $8 (5 s^2-4) (93300 s^8-259915 s^6+264982 s^4-116928 s^2$\
& $+18816)$\
${}_{4,0}$ & $-20 s^2 (5 s^2-4) (15 s^2-14) (45 s^4+36 s^2-56)$\
Then, after solving Eq. (\[homosol\]), we find that the third order term of the generating function can be arranged in the form $$\begin{aligned}
\nonumber
W_3 &=& \frac{\epsilon^3G}{8960 (5 s^2-4)^4}\sum_{l=1}^3\sum_{\substack{k=l-4 \\ k\ne0}}^{2l+4}\sum_{j=0}^2\Gamma_{3,j,k,l}e^{2j+k^*} \\ \label{W3g}
&& \times s^{2l}\sin(kf+2lg)+C_3\end{aligned}$$ in which the inclination polynomials $\Gamma_{3,j,k,l}$ that do not vanish are listed in Table \[t:W3g\]. The constant $C_3$ is determined at the next order of the perturbation approach in the same way as we did previously. We obtain $$C_3=\frac{\epsilon^3G}{1536(5s^2-4)^5}\sum_{l=1}^3\sum_{j=0}^{3-l}\Gamma_{3,j,0,l}e^{2(j+l)}s^{2l}\sin2l\omega,$$ which contributes six additional inclination polynomials $\Gamma_{3,j,0,l}$, that are also listed in Table \[t:W3g\].
[@ll@]{} ${}_{2,-3,1}$ & $35 \left(15 s^2-14\right) \left(87375 s^{10}-335550 s^8+505080
s^6-371184 s^4+132096 s^2-17920\right)$\
${}_{2,-2,1}$ & $105 \left(15 s^2-14\right)
\left(399375 s^{10}-1863400 s^8+3389440 s^6-3023632 s^4+1328128
s^2-230400\right)$\
${}_{1,-1,1}$ & $-840 \left(5 s^2-4\right) \left(228125 s^{10}-549325
s^8+255940 s^6+324664 s^4-352992 s^2+93824\right)$\
${}_{2,-1,1}$ & $210 \left(15
s^2-14\right) \left(100875 s^{10}-275600 s^8+228220 s^6-6408 s^4-70272
s^2+23296\right)$\
${}_{2,-1,2}$ & $-105 \left(5 s^2-4\right) \left(15 s^2-14\right)
\left(13725 s^6-37680 s^4+34228 s^2-10304\right)$\
${}_{0,1,1}$ & $-1680 \left(5
s^2-4\right)^2 \left(486525 s^8-1594290 s^6+1955772 s^4-1064576
s^2+216960\right)$\
${}_{1,1,1}$ & $840 \left(5 s^2-4\right) \left(1531125 s^{10}-6503075
s^8+10982780 s^6-9224760 s^4+3855648 s^2-641920\right)$\
${}_{2,1,1}$ & $-420 \left(15
s^2-14\right) \left(61875 s^{10}-138825 s^8+51640 s^6+92200 s^4-89088
s^2+22400\right)$\
${}_{1,1,2}$ & $1680 \left(5 s^2-4\right) \left(240750 s^8-775475
s^6+932445 s^4-495822 s^2+98320\right)$\
${}_{2,1,2}$ & $105 \left(5 s^2-4\right)
\left(226125 s^8-787950 s^6+1015020 s^4-572056 s^2+118944\right)$\
${}_{2,1,3}$ & $-840
\left(15 s^2-14\right) \left(1125 s^6-3300 s^4+3235 s^2-1058\right)$\
${}_{0,2,1}$ & $-1680
\left(5 s^2-4\right)^3 \left(41615 s^6-97838 s^4+76016 s^2-19488\right)$\
${}_{1,2,1}$ & $
-1680 \left(5 s^2-4\right) \left(666875 s^{10}-2586600 s^8+4014940 s^6-3117320
s^4+1210368 s^2-187904\right)$\
${}_{2,2,1}$ & $210 \left(5398125 s^{12}-27480750
s^{10}+57999400 s^8-64973520 s^6+40757888 s^4-13579264 s^2+1878016\right)$\
${}_{1,2,2}$ & $
-3360 \left(5 s^2-4\right)^2 \left(42850 s^6-108830 s^4+92099
s^2-25958\right)$\
${}_{2,2,2}$ & $840 \left(5 s^2-4\right) \left(123750 s^8-359475
s^6+378010 s^4-167724 s^2+25624\right)$\
${}_{2,2,3}$ & $-105 \left(639375 s^8-2259750
s^6+2991200 s^4-1757840 s^2+387104\right)$\
${}_{0,3,1}$ & $-1120 \left(5 s^2-4\right)^2
\left(270650 s^8-828285 s^6+945816 s^4-477232 s^2+89664\right)$\
${}_{1,3,1}$ & $280
\left(5 s^2-4\right) \left(634500 s^{10}-2623725 s^8+4300340 s^6-3496152 s^4+1412352
s^2-227584\right)$\
${}_{2,3,1}$ & $-70 \left(15 s^2-14\right) \left(76875 s^{10}-202950
s^8+167700 s^6-16544 s^4-37376 s^2+12544\right)$\
${}_{0,3,2}$ & $-560 \left(5
s^2-4\right)^2 \left(605775 s^6-1524950 s^4+1277728 s^2-356256\right)$\
${}_{1,3,2}$ & $560
\left(5 s^2-4\right)^2 \left(9150 s^6-6435 s^4-9741 s^2+7154\right)$\
${}_{2,3,2}$ & $35
\left(5 s^2-4\right) \left(104625 s^8-68850 s^6-286620 s^4+378936
s^2-127232\right)$\
${}_{1,3,3}$ & $-280 \left(5 s^2-4\right) \left(52875 s^6-129225
s^4+102900 s^2-26456\right)$\
${}_{2,3,3}$ & $-140 \left(15 s^2-14\right) \left(7875
s^6-21225 s^4+19120 s^2-5756\right)$\
${}_{1,4,1}$ & $-840 \left(5 s^2-4\right)
\left(516875 s^{10}-1956050 s^8+2950940 s^6-2217472 s^4+829440
s^2-123392\right)$\
${}_{2,4,1}$ & $420 \left(5 s^2-4\right) \left(173625 s^{10}-668250
s^8+1023720 s^6-780120 s^4+295872 s^2-44800\right)$\
${}_{0,4,2}$ & $-6720 \left(5
s^2-4\right)^3 \left(730 s^4-1153 s^2+444\right)$\
${}_{1,4,2}$ & $-3360 \left(5
s^2-4\right)^2 \left(66050 s^6-166215 s^4+139230 s^2-38808\right)$\
${}_{2,4,2}$ & $420
\left(5 s^2-4\right)^2 \left(20925 s^6-38700 s^4+19984 s^2-1976\right)$\
${}_{1,4,3}$ & $
840 \left(5 s^2-4\right) \left(10125 s^6-32150 s^4+33500 s^2-11456\right)$\
${}_{2,4,3}$ & $
-2100 \left(5 s^2-4\right) \left(3525 s^6-9240 s^4+7996 s^2-2280\right)$\
${}_{1,5,1}$ & $
-168 \left(5 s^2-4\right) \left(115000 s^{10}-434875 s^8+663700 s^6-512080 s^4+199936
s^2-31488\right)$\
${}_{2,5,1}$ & $-105 s^2 \left(5 s^2-4\right) \left(15 s^2-14\right)
\left(825 s^6-1990 s^4+1616 s^2-448\right)$\
${}_{0,5,2}$ & $-336 \left(5 s^2-4\right)^3
\left(15425 s^4-24050 s^2+9112\right)$\
${}_{1,5,2}$ & $-84 \left(5 s^2-4\right)^2
\left(551625 s^6-1390850 s^4+1167040 s^2-325728\right)$\
${}_{2,5,2}$ & $105 \left(5
s^2-4\right)^2 \left(15 s^2-14\right) \left(225 s^4+288 s^2-364\right)$\
${}_{0,5,3}$ & $
3360 \left(5 s^2-4\right)^2 \left(1575 s^4-2795 s^2+1256\right)$\
${}_{1,5,3}$ & $840
\left(5 s^2-4\right) \left(8250 s^6-24975 s^4+25080 s^2-8336\right)$\
${}_{2,5,3}$ & $-420
\left(5 s^2-4\right) \left(15 s^2-14\right)^2 \left(15 s^2-13\right)$\
${}_{2,6,1}$ & $35
\left(5 s^2-4\right) \left(171375 s^{10}-616950 s^8+871680 s^6-600128 s^4+199168
s^2-25088\right)$\
${}_{1,6,2}$ & $-280 \left(5 s^2-4\right)^3 \left(8265 s^4-12874
s^2+4872\right)$\
${}_{2,6,2}$ & $-280 \left(5 s^2-4\right)^2 \left(11475 s^6-29280
s^4+24828 s^2-6992\right)$\
${}_{0,6,3}$ & $560 \left(5 s^2-4\right)^3 \left(1335
s^2-1166\right)$\
${}_{1,6,3}$ & $560 \left(5 s^2-4\right)^2 \left(8325 s^4-14910
s^2+6764\right)$\
${}_{2,6,3}$ & $70 \left(5 s^2-4\right) \left(21375 s^6-61950 s^4+59960
s^2-19328\right)$\
${}_{1,7,2}$ & $-60 \left(5 s^2-4\right)^3 \left(8385 s^4-13226
s^2+5080\right)$\
${}_{0,7,3}$ & $10080 \left(5 s^2-4\right)^3 \left(90
s^2-79\right)$\
${}_{1,7,3}$ & $12600 \left(5 s^2-4\right)^2 \left(105 s^4-191
s^2+88\right)$\
${}_{2,8,2}$ & $-840 \left(3 s^2-2\right) \left(5 s^2-4\right)^3 \left(15
s^2-14\right)$\
${}_{1,8,3}$ & $4200 \left(5 s^2-4\right)^3 \left(96
s^2-85\right)$\
${}_{2,8,3}$ & $210 \left(5 s^2-4\right)^2 \left(525 s^4-990
s^2+472\right)$\
${}_{1,9,3}$ & $7560 \left(5 s^2-4\right)^3 \left(10
s^2-9\right)$\
${}_{2,10,3}$ & $315 \left(5 s^2-4\right)^3 \left(15 s^2-14\right)$\
${}_{0,0,3}$ & $2 \left(15 s^2-14\right)^3 \left(825 s^4-1445s^2+634\right)$\
${}_{0,0,2}$ & $-6 \left(5 s^2-4\right)^2 \left(2171250 s^8-7719525s^6+10225470 s^4-5983260 s^2+1305248\right)$\
${}_{1,0,2}$ & $-3 \left(5 s^2-4\right)
\left(15 s^2-14\right)^2 \left(1800 s^6+2655 s^4-8208 s^2+3928\right)$\
${}_{0,0,1}$ & $48
\left(5 s^2-4\right)^2 \left(9060750 s^{10}-34431275 s^8+51858720 s^6-38675200
s^4+14258176 s^2-2072064\right)$\
${}_{1,0,1}$ & $-12 \left(5 s^2-4\right) \left(93223125
s^{12}-421210500 s^{10}+784654200 s^8-771469840 s^6+422629664 s^4-122600960
s^2+14780416\right)$\
${}_{2,0,1}$ & $6 \left(15 s^2-14\right) \left(2328750 s^{12}-8703375
s^{10}+13317150 s^8-10848180 s^6+5157560 s^4-1450624 s^2+200704\right)$\
The computation of additional orders finds similar structures. Thus, for instance, the fourth order term of the new Hamiltonian takes the form $$\label{K04g}
K_{0,4}=\epsilon^4\frac{\mu}{r}\frac{p^2}{r^2}\frac{9s^2}{1280(5s^2-4)^6}\sum_{j=0}^5\frac{p^j}{r^j}\sum_{k=0}^{\lfloor3-\frac{1}{2}j\rfloor}e^{2k}\gamma_{4,j,k},$$ and the fourth order term of the generating function takes the form $$\begin{aligned}
\nonumber
W_4 &=& \frac{\epsilon^4G}{30105600(5s^2-4)^6}\sum_{l=1}^4\sum_{\substack{k=l-6 \\ k\ne0}}^{2l+6}\sum_{j=0}^6\Gamma_{4,j,k,l}e^{2j+k^*} \\ \label{W4g}
&& \times s^{2l}\sin(kf+2l\omega)+C_4.\end{aligned}$$ While the former involves 15 inclination polynomials $\gamma_{4,j,k}$, the later comprises up to 124 non-vanishing inclination polynomials $\Gamma_{4,j,k,l}$, of degree 9 in $s^2$. Therefore, these and other polynomials resulting from following orders are not listed due to their length.
The normalization of $G$ has been extended up to the order six without major trouble, except for the increase of the computational burden of successive orders due to the notable growth of the length of the series to be handled, on the one hand, and the increasing size of the rational coefficients resulting from the integer arithmetic used, on the other. Thus, the new Hamiltonian terms $K_{0,5}$ and $K_{0,6}$ take analogous forms to the previous orders, with 24 inclination polynomials $\gamma_{5,j,k}$, and 35 $\gamma_{6,j,k}$. Similarly, the generating function terms $W_{5}$ and $W_{6}$ have been arranged in the same form as previous orders of the perturbation approach, with 254 non-vanishing coefficients $\Gamma_{5,j,k,l}$ and 429 $\Gamma_{6,j,k,l}$, 15 of which correspond to the integration constant $C_5$ and 21 to $C_6$.
At the end, since the generating function is known, the transformation equations from the original Delaunay variables to the new ones, and vice versa, are readily computed by standard application of the fundamental recursion (\[Lie:triangle\]) (see [@Deprit1969] for details). The procedure ends by replacing the original variables by the new ones in the computed terms $K_{0,m}$.
We remark that the procedure described here is not equivalent to the combination of the elimination of the parallax and the elimination of the perigee into a single transformation, in spite of the fact that the total angular momentum is converted into a formal integral in both cases. On the contrary, the perigee has been removed here keeping as much short-period terms as possible in the new Hamiltonian. With this strategy, the size of the generating function of the first partial normalization is astonishingly smaller than the one that would be obtained with other alternatives in the literature. To check that, we fully expanded both the Hamiltonian and the generating function and reckoned the number of terms. This procedure is the one that has been traditionally used in the literature to assess the complexity of a perturbation solution [@DepritRom1970; @CoffeyDeprit1982]. Thus, for instance, we find 5, 56, 367, 1152, 2627, and 4897 expanded terms for the 1st, 2nd …, 6th order terms of the generating function, respectively, whereas much longer expressions have been reported in the literature. For instance, the sixth order of the generating function reported in [@SanJuanetal2013] entails 39630 terms, which is almost one order of magnitude larger than the one obtained with the current approach. It is worth mentioning that further simplifications could be obtained in particular cases, like when constraining the application of the analytical solution to the case of low eccentricity orbits, in which case many of the terms involved can be neglected [@Lara2008; @GaiasColomboLara2020].
The comparisons are, nevertheless, inconclusive due to the diversity of approaches used in the computation of the variety of solutions reported in the literature, which involve representation in different variables, on the one hand, and yield distinct one-degree-of-freedom Hamiltonians, on the other. The partially normalized Hamiltonian obtained here is definitely longer than the one obtained after the classical elimination of the perigee. However, this is not of concern in the complete, as opposite to partial, normalization of the main problem. Indeed, after the following elimination of short-period terms either procedure should arrive to the same Hamiltonian. We didn’t find reported data for the second normalization after the elimination of the perigee, but one would expect that the figures should be balanced to some extent —the generating function of the short-period elimination probably being heavier in our case than in other prospective approaches. While the needed data for the thorough comparison is not available, our claims must restrict to the checked fact that our approach makes both transformations of manageable size —as we will show in the next section, where the short-period effects are removed in a Delaunay normalization. This feature makes the evaluation of higher orders of our perturbation solution certainly practicable.
Delaunay normalization
----------------------
The partially normalized Hamiltonian is just of one degree of freedom, yet it is not separable. To get an explicit analytical solution we remove the short-period terms that are associated to the radius by means of a Delaunay normalization [@Deprit1982]. The partially reduced Hamiltonian from which we start takes again the form of Eq. (\[mainP\]), but it is now written in prime Delaunay variables $(\ell',g',h',L',G',H')$. In this Hamiltonian the variables $g'$ and $h'$ are cyclic, and hence $H'=H$ and $G'$ remain constant for given initial conditions. The zeroth order term is the same as in Eq. (\[mainPK0\]), the term $K_{1,0}$ is given by Eq. (\[K01g\]), whereas terms $K_{m,0}$ with $m\ge2$ are no longer void, and, on the contrary are given by Eqs. (\[K02g\]), (\[K03g\]), (\[K04g\]), for $m=2,3,4$, respectively, and analogous equations for higher orders that, due to its length, are not printed in the paper. Moreover, the definition in Eq. (\[mieps\]) turns $\epsilon$ into a physical constant, rather than a function, when witten in the prime variables. In consequence $\epsilon$ can replace now the formal small parameter $\varepsilon$ used before in the Lie transforms procedure, assumed, of course, that the corresponding scaling of the Hamiltonian terms is properly made.
The homological equation to be solved at each step of the new Lie transformation is the same as before, either in the form given by Eq. (\[homosoll\]) or Eq. (\[homosol\]), except for choosing a non-zero integration constant does not provide any advantage in the current case. In fact, since the new Hamiltonian must be free from terms depending on the mean anomaly, which are obtained by averaging, the homological equation can be further particularized. That is, $$\label{averaging}
K_{0,m}=\langle\widetilde{K}_{0,m}\rangle_\ell=\frac{1}{2\pi}\int_0^{2\pi}\widetilde{K}_{0,m}\frac{r^2}{a^2\eta}\,\mathrm{d}f,$$ and hence $$\begin{aligned}
\nonumber
W_m &=& -\frac{\ell}{n}K_{0,m}+\frac{1}{n}\int\widetilde{K}_{0,m}\frac{r^2}{a^2\eta}\,\mathrm{d}f \\
\label{homolo}
&=& K_{0,m}\frac{\phi}{n}+\frac{1}{n}\int\left(\widetilde{K}_{0,m}\frac{r^2}{a^2\eta}-K_{0,m}\right)\mathrm{d}f,\end{aligned}$$ in which $\phi=f-\ell$ is the equation of the center, and the terms under the integral sign in the final form of the equation are purely periodic in $f$.
Thus, at the first order of the Lie transforms procedure we find $$\label{K01l}
K_{0,1}=\langle{K}_{1,0}\rangle_\ell\equiv \frac{\mu}{p}\eta^3 \left(3 s^2-2\right),$$ from which $$\label{W1l}
W_1=G'\left(3s^2-2\right)(e\sin{f}+\phi ).$$ Recall that the symbols $p$, $\eta$, $e$, etc. are now functions of the Delaunay prime variables.
At the second order, the known terms are given again by Eqs. (\[triangle2\]) and (\[K11\]), from which, using Eq. (\[averaging\]), $$\label{K02l}
K_{0,2}=-\frac{\mu}{p}\frac{3}{4}\eta^3\sum_{j=0}^2\lambda_{2,j}\eta^j,$$ with $\lambda_{2,0}=5(7s^4-16s^2+8)$, $\lambda_{2,1}=4(3s^2-2)^2$, and $\lambda_{2,2}=5s^4+8s^2-8$. The second order term of the generating function is then trivially integrated from Eq. (\[homolo\]), to yield $$\begin{aligned}
\nonumber
W_2 &=&
-\frac{G'\beta}{32(5s^2-4)^2}\sum_{j=1}^3\sum_{k=0}^{3-\lfloor\frac{1}{2}j\rfloor}\Lambda_{2,j,k}\eta^ke^j\sin{jf} \\ \label{W2l}
&& -G'\frac{3}{4}\phi\sum_{j=0}^1\Phi_{2,j}e^{2j},\end{aligned}$$ where $\beta=1/(1+\eta)$, $\Phi_{2,0}=8(s^2-1)(5s^2-4)$, $\Phi_{2,1}=8-8s^2-5s^4$, and the inclination polynomials $\Lambda_{2,j,k}$ are provided in Table \[t:Lambda2\].
[@ll@]{} ${}_{1,0}$ & $15 (3 s^2-2) (805 s^6-2448 s^4+2400 s^2-768)$\
${}_{1,1}$ & $3 (3s^2-2) (2225 s^6-8160 s^4+8928 s^2-3072)$\
${}_{1,2}$ & $3 (825 s^8-3030s^6+4064 s^4-2368 s^2+512)$\
${}_{1,3}$ & $-3 s^2 (975 s^6-2250 s^4+1728 s^2-448)$\
${}_{2,0}$ & $6(1925 s^8-6210 s^6+7452 s^4-3936 s^2+768)$\
${}_{2,1}$ & $6 (125 s^8-930 s^6+1660s^4-1120 s^2+256)$\
${}_{3,0}$ & $2625 s^8-7270 s^6+7408 s^4-3264 s^2+512$\
${}_{3,1}$ & $s^2 (825 s^6-1990s^4+1616 s^2-448)$\
Integrals cease to be trivial at the third order. Indeed, the formal computation of $K_{0,3}$ using the fundamental recursion (\[Lie:triangle\]) yields two different types of terms.
The first type consists of terms depending on the equation of the center, which can be reduced to the form $$\frac{\mu}{r}\frac{p}{r}P(e)Q(s)\phi\sin{mf},$$ with $P$ and $Q$ denoting arbitrary eccentricity and inclination polynomials. Definite integration of these kinds of terms is carried out from expressions in [@Metris1991], whereas the indefinite integration is achieved by parts, to get $$ma^2\eta\int \frac{\sin{mf}}{r^2}\phi\,\mathrm{d}\ell=\frac{\sin{mf}}{m}-\phi\cos{mf}-\int\cos{mf}\mathrm{d}\ell,$$ in which the antiderivatives of cosines of multiples of the true anomaly are carried out after expressing them in terms of $r$ and $R=\mathrm{d}r/\mathrm{d}t$, rather than in trigonometric functions, as discussed in [@Jefferys1971].
The second type consists of terms free from the equation of the center. In these terms, the trigonometric functions of the true anomaly can be replaced by inverse powers of the radius, without involving the radial velocity. We found that the exponents of the inverse of $r$ range from 0 to 8 missing the exponent 1. Therefore, both definite and indefinite integrals of terms of the second type are trivially solved.
Proceeding in this way, we compute the third order Hamiltonian term $$\label{K03l}
K_{0,3}=\frac{\mu}{p}\frac{9\eta^3}{16(5s^2-4)^2}\sum_{j=0}^4\lambda_{3,j}\eta^j,$$ with the inclination polynomials $\lambda_{3,j}$ given in Table \[t:lambda3\].
[@ll@]{} ${}_{0}$ & $5(28700 s^{10}-107205 s^8+158960 s^6-118492 s^4$\
& $+45152 s^2-7168)$\
${}_{1}$ & $60(3 s^2-2)(5 s^2-4)^2(7 s^4-16 s^2+8)$\
${}_{2}$ & $-2(28675s^{10}-98005 s^8+130852 s^6-87164 s^4$\
& $+30176 s^2-4608)$\
${}_{3}$ & $20(3 s^2-2)(5 s^2-4)^2(5 s^4+8 s^2-8)$\
${}_{4}$ & $-s^2(15 s^2-14)(450s^6-925 s^4+590 s^2-112)$\
The corresponding term of the generating function is $$\begin{aligned}
\nonumber
W_3 &=& \frac{G'\beta^2}{128(5s^2-4)^3}\sum_{j=1}^6\sum_{k=0}^{7-2\lfloor\frac{1}{2}j\rfloor}\Lambda_{3,j,k}\eta^{k-1}e^j\sin{jf} \\ \label{W03l}
&& +\frac{3G'}{16(5s^2-4)^2}\phi\sum_{j=0}^3\sum_{k=0}^{4}\Phi_{3,j,k}\eta^ke^j\cos{jf},\end{aligned}$$ in which the inclination polynomials $\Phi_{3,j,k}$ are given in Table \[t:W3lc\], and those $\Lambda_{3,j,k}$ in Table \[t:W3ls\].
[@ll@]{} ${}_{0,0}$ & $5 (89100 s^{10}-323615 s^8+466320 s^6-337684 s^4$\
& $+125216 s^2-19456)$\
${}_{0,2}$ & $-2(112125 s^{10}-374775 s^8+488460 s^6-314932 s^4$\
& $+103840 s^2-14848)$\
${}_{0,3}$ & $8 (3 s^2-2) (5 s^2-4)^2 (5 s^4+8 s^2-8)$\
${}_{0,4}$ & $-3 s^2 (15s^2-14) (450 s^6-925 s^4+590 s^2-112)$\
[@ll@]{} ${}_{1,0}$ & $-864 (3 s^2-2)^3 (5 s^2-4)^3$\
${}_{1,1}$ & $3 (22218875 s^{12}-104346550 s^{10}+202703740 s^8-209869352s^6+123038240s^4-39033472 s^2+5275648)$\
${}_{1,2}$ & $12(10925500s^{12}-50711075s^{10}+97386820s^8-99715748s^6+57863024s^4-18199872s^2+2445312)$\
${}_{1,3}$ & $3(27560125s^{12}-119080550s^{10}+212650740s^8-202245448s^6+109190304s^4-32208768s^2+4128768)$\
${}_{1,4}$ & $12(3155125s^{12}-10820800s^{10}+13899620s^8-7620256s^6+944256s^4+613248s^2-167936)$\
${}_{1,5}$ & $3(5410625s^{12}-17331450s^{10}+19448180s^8-6842968s^6-2742560s^4+2556544s^2-491520)$\
${}_{1,6}$ & $-12(59625s^{12}-415275s^{10}+942920s^8-994980s^6+529776s^4-134592s^2+12288)$\
${}_{1,7}$ & $-3 s^2 (77625 s^{10}-568950 s^8+1256420s^6-1222216 s^4+550816 s^2-94080)$\
${}_{2,0}$ & $-1044 (3 s^2-2)^3(5 s^2-4)^3$\
${}_{2,1}$ & $24 (131000s^{12}-1121875s^{10}+3061340s^8-3989664s^6+2758768s^4-983648s^2+143360)$\
${}_{2,2}$ & $96(51625s^{12}-437800s^{10}+1183290s^8-1528682s^6+1049344s^4-372240s^2+54144)$\
${}_{2,3}$ & $-12(5s^2-4)(16375s^{10}+64070s^8-257508 s^6+297320 s^4-145792 s^2+26624)$\
${}_{2,4}$ & $-12(263625s^{12}-1000750s^{10}+1526820s^8-1206712s^6+539616s^4-142592s^2+19968)$\
${}_{2,5}$ & $-12s^2 (162375 s^{10}-576100 s^8+787020s^6-506248 s^4+145984 s^2-12992)$\
${}_{3,0}$ & $-656 (3 s^2-2)^3(5 s^2-4)^3$\
${}_{3,1}$ & $-934875s^{12}+605000s^{10}+4973120s^8-10412952s^6+8554272s^4-3281280s^2+491520$\
${}_{3,2}$ & $-2080125s^{12}+3562000s^{10}+2881300s^8-11103360s^6+10155456s^4-4038912s^2+614400$\
${}_{3,3}$ & $-(5 s^2-4) (254925 s^{10}-526480 s^8+318084 s^6-10672s^4-40064 s^2+8192)$\
${}_{3,4}$ & $3(28875s^{12}-147800s^{10}+254260s^8-160992s^6-11968s^4+54016s^2-16384)$\
${}_{3,5}$ & $-12 s^2 (4500s^{10}-4125 s^8-16075 s^6+31670 s^4-20664 s^2+4704)$\
${}_{4,0}$ & $-240 (3s^2-2)^3 (5 s^2-4)^3$\
${}_{4,1}$ & $6 (5 s^2-4)(50325 s^{10}-157660 s^8+180520 s^6-90312 s^4+18112 s^2-1024)$\
${}_{4,2}$ & $12 (5 s^2-4) (47475 s^{10}-147000 s^8+164852 s^6-79056s^4+14208 s^2-512)$\
${}_{4,3}$ & $6 s^2 (5 s^2-4) (44625s^8-136340 s^6+149184 s^4-67800 s^2+10304)$\
${}_{5,0}$ & $-48 (3s^2-2)^3 (5 s^2-4)^3$\
${}_{5,1}$ & $6 s^2 (5 s^2-4)^2(180 s^6-609 s^4+530 s^2-112)$\
${}_{5,2}$ & $3 s^2 (5 s^2-4)(1125 s^8-7440 s^6+11516 s^4-6144 s^2+896)$\
${}_{5,3}$ & $-3 s^4 (5s^2-4) (15 s^2-14) (45 s^4+36 s^2-56)$\
${}_{6,0}$ & $-4(3 s^2-2)^3 (5 s^2-4)^3$\
In the process of carrying out the normalization to higher orders we need to deal with trigonometric series of notably increasing length. However, we only found terms of the same two types as before in the solution of the homological equation, which, in consequence, is analogously integrated. In this way, the second normalization has been extended up to the order 6 in the small parameter, in agreement with the order to which the perigee was previously eliminated. Corresponding Hamiltonian and generating function terms are analogously arranged in the form of Eqs. (\[K03l\]) and (\[W03l\]), respectively, yet the inclination polynomials comprise much more longer listings, as expected. Thus, for instance, at the fourth order the normalized Hamiltonian term $$\label{K04l}
K_{0,4}=\frac{\mu}{p}\frac{9\eta^3}{64(5s^2-4)^3}\sum_{j=0}^6\lambda_{4,j}\eta^j,$$ contributes 7 new inclination polynomials, which are polynomials of degree 7 in the square of the sine of the inclination, whereas the generating function term $$\begin{aligned}
\nonumber
W_4 &=&
\frac{G'\beta^3}{20480(5s^2-4)^6}\sum_{j=1}^6\sum_{k=0}^{7-2\lfloor\frac{1}{2}j\rfloor}\Lambda_{4,j,k}\eta^{k-3}e^j\sin{jf} \\ \label{W04l}
&& +\frac{3G'}{256(5s^2-4)^3}\phi\sum_{j=0}^5\sum_{k=0}^{6}\Phi_{4,j,k}\eta^ke^j\cos{jf},\end{aligned}$$ is made of 20 non-zero coefficients $\Phi_{4,j,k}$, of degree 7 in $s^2$, and 72 non-zero coefficients $\Lambda_{4,j,k}$, which are of degree 10 in $s^2$. Note that, because of the denominators factoring the summations, the maximum degree amounts to 4 in any case, in agreement with the order of the perturbation. At the 5th order we find 9 coefficients of the type $\lambda$ and 41 nonvanishing coefficients of the type $\Phi$, which are of degree 11 in $s^2$ although they are divided by $(5s^2-4)^6$, as well as 134 nonvanishing trigonometric polynomials of the type $\Lambda$, which are of degree 12 in $s^2$ yet they must be divided by $(5s^2-4)^7$. Finally, the figures of the 6th order are 11 and 70 of degree 13 in $s^2$ for $\lambda$ and $\Phi$, respectively, and 218 for $\Lambda$ of degree 16.
Once reached the desired order of the second normalization, the procedure ends by changing prime variables by double-prime variables in the new Hamiltonian terms $K_{0,m}$.
In order to provide comparative figures of the computational burden of this normalization with other approaches that might be carried out, we expanded the series that comprise the solution and reckon the number of separate terms, as we already did in the normalization of the total angular momentum. We found that the 1st, 2nd, …6th-order terms of the generating function of the Delaunay normalization comprise 4, 48, 257, 931, 2266, and 4826 terms respectively. Note that different arrangements from the one chosen by us —Eqs. (\[W03l\]), (\[W04l\]), etc.— may provide different figures than those reported, yet should be of analogous magnitude. Following the same procedure with the completely reduced Hamiltonian, we reckon up to 2, 9, 29, 55, 106, and 152 coefficients, for the 1st, 2nd, …6th-order term, respectively.
Secular terms
-------------
After neglecting higher order effects of $J_2$, the normalized Hamiltonian is $$\label{Hn}
\mathcal{K}''=\mathcal{K}''(-,-,-,L'',G'',H'')\equiv\sum_{m\ge0}^{\tilde{m}}\frac{\epsilon^m}{m!}K_{m,0},$$ in which $\tilde{m}\le6$ for the different approximations provided by the computed perturbation solution, and terms $K_{m,0}$ are obtained by replacing prime by double-prime variables in corresponding terms given by Eqs. (\[K01l\]), (\[K02l\]), (\[K03l\]), as well as higher order Hamiltonian terms $K_{0,m}$ that have not been displayed. That is, the symbols $p$, $\eta$, and $s$ in these equations are assumed to be functions of the double-prime Delaunay momenta. Recall that $\epsilon$ is obtained from Eq. (\[mieps\]) by making $p=G''^2/\mu$.
The corresponding solution to the flow stemming from Eq. (\[Hn\]) is $$\begin{aligned}
\ell'' &=& \ell''_0+n_\ell{t}, \\
g''&=& g''_0+{n}_\omega{t} \\
h'' &=& h''_0+n_ht,\end{aligned}$$ in which, from Hamilton equations, $$n_\ell=\frac{\partial\mathcal{K}''}{\partial{L}''}, \qquad
n_g=\frac{\partial\mathcal{K}''}{\partial{G}''}, \qquad
n_h=\frac{\partial\mathcal{K}''}{\partial{H}''},$$ $L''=L''_0$, $G''=G'_0$, and $H''=H_0$, are the initialization constants of the analytical solution, and $\ell''_0$, $g''_0$, $h''_0$, $L''_0$, and $G'_0$, are computed applying consecutively the inverse transformation of the normalization of the angular momentum and the Delaunay normalization to corresponding initial conditions of the original problem.
On the other hand, Delaunay variables are singular for equatorial orbits, in which the argument of the node is not defined, as well as for circular orbits, in which the argument of the periapsis is not defined. In particular, the singularity for circular orbits reveals immediately by the appearance of the eccentricity in denominators of the transformation equations of both the mean anomaly and the argument of the perigee (see Eqs. (20) and (21) of [@Brouwer1959], for instance). This fact not only makes singular the transformation of these elements for exactly circular orbits, but prevents convergence of the respective perturbation series for small values of the eccentricity.
Different sets of non-singular variables can be used to avoid these issues [@Lyddane1963]. In particular, troubles related with low eccentricities are commonly avoided by replacing the mean anomaly, the argument of the periapsis, and the total angular momentum with the non-canonical variables given by the mean distance to the node $F=\ell+g$, also called mean argument of the latitude, and the components of the eccentricity vector in the nodal frame $C=e\cos\omega$, $S=e\sin\omega$, also called semi-equinoctial elements [@Konopliv1990; @Cook1992]. In these variables, the secular terms of the main problem are given by [@DepritRom1970] $$\begin{aligned}
\label{Fsecular}
F &=& F_0+n_Ft, \\ \label{Csecular}
C &=& e\cos(g_0+n_g{t})=C_0\cos{n}_g{t}-S_0\sin{n}_g{t}, \\ \label{Ssecular}
S &=& e\sin(g_0+n_g{t})=S_0\cos{n}_g{t}+C_0\sin{n}_g{t}, \\ \label{Lsecular}
L &=& L_0, \\ \label{hsecular}
h &=& h_0+n_ht, \\ \label{Hsecular}
H &=& H_0,\end{aligned}$$ in which $n_F=n_\ell+n_g$, $C_0=e\cos{g}_0$, $S_0=e\sin{g}_0$, and $e=(1-G_0^2/L_0^2)^{1/2}$. The double-prime notation has been omitted for simplicity.
After standard partial differentiation, we obtain $$\begin{aligned}
\label{nFsecular}
n_F &=& n+n\sum_{m=1}^{\tilde{m}}
\frac{\epsilon^m}{(5s^2-4)^m}\,\sum_{i=0}^{2m-1}\Psi_{m,i}(s)\eta^i, \\ \label{ngsecular}
n_g &=& n\sum_{m=1}^{\tilde{m}}
\frac{\epsilon^m}{(5s^2-4)^m}\sum_{i=0}^{2m-2}\omega_{m,i}(s)\eta^i, \\ \label{nhsecular}
n_h &=& n{c}\sum_{m=1}^{\tilde{m}}
\frac{\epsilon^m}{(5s^2-4)^m}\sum_{i=0}^{2m-2}\Omega_{m,i}(s)\eta^i,\end{aligned}$$ where the inclination polynomials $\Psi_{m,i}$, $\omega_{m,i}$, and $\Omega_{m,i}$ are provided in Tables \[t:Psi\], \[t:omega\], and \[t:Omega\], respectively, up to the third order of the perturbation approach. In these tables we can check that, in spite of the general arrangement of the secular frequencies in Eqs. (\[nFsecular\])–(\[nhsecular\]), the critical inclination divisors $5s^2-4$ start to appear only at the third order truncation.
[@llll@]{} ${}_{1,0}$ & $-3 (5 s^2-4)^2$\
${}_{1,1}$ & $-3 (3 s^2-2) (5s^2-4)$\
${}_{2,0}$ & $\frac{15}{8} (5 s^2-4)^2 (77 s^4-172s^2+88)$\
${}_{2,1}$ & $\frac{9}{8} (5 s^2-4)^2 (155 s^4-256s^2+104)$\
${}_{2,2}$ & $\frac{3}{8} (5 s^2-4)^2 (189 s^4-156s^2+8)$\
${}_{2,3}$ & $\frac{15}{8} (5 s^2-4)^2 (5 s^4+8s^2-8)$\
${}_{3,0}$ &$-\frac{15}{32}(2439500s^{12}-11312175s^{10}+21772080s^8$\
& $-22346500s^6+12956400s^4-4043136 s^2+533248)$\
${}_{3,1}$ & $-\frac{45}{32}(5 s^2-4) (62300 s^{10}-260365 s^8+431504 s^6$\
& $-356508s^4+147552s^2-24576)$\
${}_{3,2}$ & $\frac{3}{16} (1835625 s^{12}-7723875 s^{10}+13291500s^8$\
& $-12015300 s^6+6064176 s^4-1644928 s^2+192256)$\
${}_{3,3}$ & $\frac{15}{16}(5s^2-4)(18175s^{10}-85105 s^8+153172 s^6-136540 s^4$\
& $+61408s^2-11264)$\
${}_{3,4}$ & $\frac{3}{32} (213750 s^{12}-1441125 s^{10}+3537000s^8-4313100 s^6$\
& $+2835280 s^4-967808 s^2+135424)$\
${}_{3,5}$ &\
[@ll@]{} ${}_{1,0}$ & $-3 (5 s^2-4)^2$\
${}_{2,0}$ & $\frac{15}{8} (5s^2-4)^2 (77 s^4-172 s^2+88)$\
${}_{2,1}$ & $9 (3 s^2-2)(5 s^2-4)^3$\
${}_{2,2}$ & $\frac{3}{8} (5 s^2-4)^2 (45 s^4+36s^2-56)$\
${}_{3,0}$ & $-\frac{15}{32}(2439500s^{12}-11312175s^{10}+21772080s^8$\
& $-22346500s^6+12956400s^4-4043136 s^2+533248)$\
${}_{3,1}$ & $-\frac{45}{4} (5 s^2-4)^3 (168 s^6-497 s^4+460s^2-136)$\
${}_{3,2}$ & $\frac{3}{16} (2150625 s^{12}-9409875 s^{10}+16968300s^8$\
& $-16218180 s^6+8729136 s^4-2535808 s^2+315136)$\
${}_{3,3}$ & $-\frac{15}{4}(5 s^2-4)^3 (105 s^6+39 s^4-228 s^2+104)$\
${}_{3,4}$ & $\frac{3}{32} (438750 s^{12}-1771125 s^{10}+2865000 s^8-2345100 s^6$\
& $+999760s^4-199808 s^2+12544)$\
[@ll@]{} ${}_{1,0}$ & $-6 (5 s^2-4)$\
${}_{2,0}$ & $\frac{15}{2} (5s^2-4)^2 (7 s^2-8)$\
${}_{2,1}$ & $18 (3 s^2-2) (5s^2-4)^2$\
${}_{2,2}$ & $\frac{3}{2} (5 s^2-4)^2 (5s^2+4)$\
${}_{3,0}$ & $-\frac{15}{8}(215250s^{10}-823025 s^8+1255040s^6-953760s^4$\
& $+361088s^2-54464)$\
${}_{3,1}$ & $-\frac{45}{4}(5s^2-4)^3 (63 s^4-124 s^2+56)$\
${}_{3,2}$ & $\frac{3}{8}(430125s^{10}-1553550s^8+2222340s^6-1570224s^4$\
& $+546432s^2-74624)$\
${}_{3,3}$ & $-\frac{15}{4} (5 s^2-4)^3 (45 s^4+28 s^2-40)$\
${}_{3,4}$ & $\frac{3}{8}(50625s^{10}-168375s^8+215900s^6-130800s^4$\
& $+35840s^2-3136)$\
Computing the periodic corrections of the semi-equinoctial variables would require to carry out expansions of the true anomaly. Therefore, we rather formulate them in polar variables. In this way we avoid the trouble of small denominators in the $J_2$-problem, yet, of course, the nodes of exactly equatorial orbits remain undefined. For simplicity we do not deal with this latter case, which, if desired, could be approached using different sets of non-singular variables.
Performance of the solution
===========================
The performance of the analytical solution is illustrated in three different cases. The first one is a low-altitude, almost-circular orbit with the orbital parameters of the PRISMA mission [@PerssonJacobssonGill2005]. The second is a high-eccentricity orbit with the typical characteristics of a geostationary transfer orbit (GTO), which we borrowed from [@Konopliv1991]. The third test has been specifically carried out to check the behavior of the analytical solution when approaching the critical inclination resonance. For this last case we selected an orbit with orbital parameters similar to the TOPEX orbit, which departs only $\sim3^\circ$ from the inclination resonance condition [@Frauenholzetal1998]. Orbital elements corresponding to these three cases are presented in Table \[t:testcases\]. The reference, true orbits have been propagated numerically in extended precision to assure that all the computed points are accurate with 15 significant digits in the decimal representation. We requested that precision because this is the maximum number of digits with which exact decimal operations are guaranteed in standard double-precision [@Kahan1997]. Because the analytical solution is evaluated in double-precision the reference numeric orbit is considered exact.
[@lllllrr@]{} & & & & & &\
PRISMA & $6878.137$ & $0.001$ & $97.42$ & $168.162$ & $20$ & $30$\
TOPEX & $7707.270$ & $0.0001$ & $66.04$ & $180.001$ & $270$ & $180$\
GTO & $24460.00$ & $0.73$ & $30.$ & $170.1$ & $280$ & $0$\
Two different kind of errors are associated to the truncation order of the analytical solution. On the one hand, the truncation of the secular terms introduces an error in the computation of the frequencies of the analytical solution, Eqs. (\[nFsecular\])–(\[nhsecular\]), which will make the perturbation solution to degrade with time. On the other hand, the truncation affects also the generating function from which the periodic corrections are derived. The latter fundamentally affects the amplitude of the periodic errors, and, therefore, the quality of the ephemeris provided by the analytical solution is not expected to deteriorate significantly with time by effect of this error. However, this is only true for the *direct* transformation, a case in which we will see that it is acceptable to use a lower order truncation than the truncation used for the secular terms. On the contrary, the truncation order of the *inverse* corrections has a direct effect in the precision with which the initialization constants are computed from a given set of initial conditions. Therefore, not computing the inverse transformation with the same accuracy as that of the secular terms will also contribute to the deterioration of the latter. In practice, the highest accuracy of the inverse transformation can be limited to the computation of the periodic corrections of the initial semimajor axis (or its partner canonical variable, the Delaunay action) because it affects directly the secular mean motion $n_\ell$, while the other elements only affect the frequencies of the secular node and perigee, which are $\mathcal{O}(J_2)$ when compared to $n_\ell$.
Accuracy of the periodic corrections
------------------------------------
First of all, we check the accuracy of the periodic corrections for increasing orders of the perturbation solution. If the periodic corrections were exact, transforming different states of the same orbit provided by the reference solution will result in the same, constant secular values of the momenta, and in an exactly linear growing of the secular angles. On the contrary, the secular values obtained from the different states of the true orbit oscillate periodically due to the truncation order of the solution. This is illustrated in Fig. \[f:PRISMAdar\] for the semimajor axis of the PRISMA test orbit, where the relative errors with respect to a constant reference value, which has been obtained as the arithmetic mean of the computed secular values, are shown for different truncations of the perturbation solution. We recall that the expected errors of a perturbation solution are of the order of the neglected terms, as follows from Eq. (\[mainPp\]). Therefore, for a truncation to the order $m$, we would expect errors of the order $\varepsilon^{m+1}$.
![Relative errors of the semimajor axis of the PRISMA-type orbit for, from top to bottom, the 1st, 2nd, 3rd, and 4th order truncations of the perturbation solution. The latter shows the relative errors of the 5th order truncation superimposed. Abscissas are hours.[]{data-label="f:PRISMAdar"}](Prismadar1 "fig:") ![Relative errors of the semimajor axis of the PRISMA-type orbit for, from top to bottom, the 1st, 2nd, 3rd, and 4th order truncations of the perturbation solution. The latter shows the relative errors of the 5th order truncation superimposed. Abscissas are hours.[]{data-label="f:PRISMAdar"}](Prismadar2 "fig:") ![Relative errors of the semimajor axis of the PRISMA-type orbit for, from top to bottom, the 1st, 2nd, 3rd, and 4th order truncations of the perturbation solution. The latter shows the relative errors of the 5th order truncation superimposed. Abscissas are hours.[]{data-label="f:PRISMAdar"}](Prismadar3 "fig:") ![Relative errors of the semimajor axis of the PRISMA-type orbit for, from top to bottom, the 1st, 2nd, 3rd, and 4th order truncations of the perturbation solution. The latter shows the relative errors of the 5th order truncation superimposed. Abscissas are hours.[]{data-label="f:PRISMAdar"}](Prismadar5 "fig:")
We found that the relative errors of the secular semimajor axis of the first order truncation (top plot of Fig. \[f:PRISMAdar\]) are of the order of $J_2^2$, as it should be the case, which amounts to 3 meters in absolute value. The relative errors of the second order truncation (second to top plot) are of the order of $J_2^3$, or less than 3 millimeter in absolute value. The third and fourth order truncations of the solution yield relative errors of the order of $J_2^4$ and $J_2^5$, respectively (second from bottom and bottom plots, respectively), or absolute errors of micrometers and hundredths of micrometers, respectively. The latter are very close to the 15 exact digits reachable working in double precision. Still, additional improvements are obtained when the truncation of the periodic corrections is extended to the fifth order of $J_2$, now effectively reaching the numerical precision, as shown in the bottom plot of Fig. \[f:PRISMAdar\], in which the relative errors of the fifth order truncation (black dots) are superimposed to the previous case.
The behavior is analogous in the case of the GTO orbit, yet now the errors notable peak at perigee passages. This is illustrated in Fig. \[f:GTOdir\], where, from top to bottom, we present now the relative errors of the inclination $I=\arccos(H/G')$, rather than the total angular momentum $G'$, for the 1st, 2nd, …, 4th order truncations of the periodic corrections, respectively. The relative errors of the 4th order truncation in the bottom plot of Fig. \[f:GTOdir\] (black dots) are superimposed to the third order ones (gray line) for reference. Corresponding absolute errors are of the order of tens of milliarc seconds for the 1st order, hundredths of mas for the second, hundredths of microarc seconds for the third, and below the level of 1 thousandth of $\mu\mathrm{as}$ for the fourth order. Due to the larger semimajor axis, the GTO orbit is less perturbed in general, and we found that the numerical precision is achieved at the fourth order of the perturbation approach.
![Relative errors of the inclination of the GTO orbit for, from top to bottom, the 1st, 2nd, 3rd, and 4th order truncations of the perturbation solution. The latter is superimposed to the relative errors of the 3rd order truncation. Abscissas are hours.[]{data-label="f:GTOdir"}](GTOdir1 "fig:") ![Relative errors of the inclination of the GTO orbit for, from top to bottom, the 1st, 2nd, 3rd, and 4th order truncations of the perturbation solution. The latter is superimposed to the relative errors of the 3rd order truncation. Abscissas are hours.[]{data-label="f:GTOdir"}](GTOdir2 "fig:") ![Relative errors of the inclination of the GTO orbit for, from top to bottom, the 1st, 2nd, 3rd, and 4th order truncations of the perturbation solution. The latter is superimposed to the relative errors of the 3rd order truncation. Abscissas are hours.[]{data-label="f:GTOdir"}](GTOdir3 "fig:") ![Relative errors of the inclination of the GTO orbit for, from top to bottom, the 1st, 2nd, 3rd, and 4th order truncations of the perturbation solution. The latter is superimposed to the relative errors of the 3rd order truncation. Abscissas are hours.[]{data-label="f:GTOdir"}](GTOdir4 "fig:")
In the case of the TOPEX orbit the inclination of $\sim66^\circ$ makes the coefficient $5s^2-4$ to become less than 2 tenths. However, this small divisor is not harmful at all in a first order truncation because, as follows from Eq. (\[C1g\]), it is multiplied by the square of the eccentricity, which is really small for the TOPEX orbit (about one thousandth, on average). Hence, the relative errors in the semimajor axis introduced by the periodic corrections are slightly better than those previously found for PRISMA, as shown in the top plot of Fig. \[f:TOPEXdar\]. The improvements should be a consequence of the larger semimajor axis of the TOPEX orbit when compared with that of PRISMA.
![Relative errors of the semimajor axis of the TOPEX-type orbit for different truncations of the perturbation solution. Abscissas are hours.[]{data-label="f:TOPEXdar"}](Topexdar1 "fig:") ![Relative errors of the semimajor axis of the TOPEX-type orbit for different truncations of the perturbation solution. Abscissas are hours.[]{data-label="f:TOPEXdar"}](Topexdar2 "fig:") ![Relative errors of the semimajor axis of the TOPEX-type orbit for different truncations of the perturbation solution. Abscissas are hours.[]{data-label="f:TOPEXdar"}](Topexdar3 "fig:") ![Relative errors of the semimajor axis of the TOPEX-type orbit for different truncations of the perturbation solution. Abscissas are hours.[]{data-label="f:TOPEXdar"}](Topexdar5 "fig:")
The effects of the offending divisors are more apparent when using higher order truncations of the periodic corrections. Indeed, the coefficients of $\sin(f+2\omega)$ and $\sin(3f+4\omega)$ in Eq. (\[W2g\]), as well as the coefficient of $\sin{f}$ in Eq. (\[W2l\]), are multiplied by the factor $e/(5s^2-4)^2$, which, while being bigger than before, it is neither of worry for TOPEX yet, as demonstrated in the second to the top plot in Fig. \[f:TOPEXdar\], where it can be checked that the relative errors in the semimajor axis remain of $\mathcal{O}(J_2^3)$. On the contrary, at the third order we find the small divisor alone in coefficients of $\sin2m(f+\omega)$, $m=1,2,3$, in Eq. (\[W3g\]), in this way lessening to some extent the corrector effect of the terms factored by $J_2^3$. This slight deterioration of convergence is noted in the second-from-bottom plot of Fig. \[f:TOPEXdar\], where we see that the relative errors in the semimajor axis are now about 10 times bigger than corresponding errors of PRISMA. Bigger offending coefficients are found in higher orders of the periodic corrections, which make that the relative errors of the fourth order truncation are clearly larger than corresponding ones of PRISMA. These relative errors are depicted in the bottom plot of Fig. \[f:TOPEXdar\] (gray line) superimposed with the errors of the fifth order truncation (black dots), showing that, at the end, the numerical precision is reachable also in this case with the fifth order truncation of the solution. The absolute error of the different truncations are now of the order of one meter, a few tenths of cm, sevral hundredths of mm, tenths of $\mu\mathrm{m}$, and thousandths of $\mu\mathrm{m}$, respectively.
Accuracy of the analytical solution
-----------------------------------
The reference orbits of the three test cases are now compared with the ones provided by the analytical solution. That is, starting from initial conditions in Table \[t:testcases\], we first apply the inverse periodic corrections to initialize the constants of the secular solution. Then, the secular terms are evaluated in Eqs. (\[Fsecular\])–(\[Hsecular\]) at the same times $t_i$ as those in which the true solution has been obtained from the numerical integration. It follows the application of the direct periodic corrections to each secular term to get the ephemeris predicted by the analytical solution at the times $t_i$. The accuracy of the different truncations of the analytical solution is assessed by computing the root-sum-square (RSS) of the difference between the position and velocity predicted by the analytical solution and those of the reference orbit and the times $t_i$.
The results of the PRISMA test case are shown in Fig. \[f:PRISMARSS\], in which ordinates are displayed in a logarithmic scale to ease comparisons. The letter $S$ in the labels ($S$:$P$) stand for the truncation order of both the inverse corrections and the secular terms, whereas $P$ indicates the truncation order of the direct periodic corrections. As we already pointed out, the latter do not need to be computed to the same accuracy as the inverse transformation. We observe in Fig. \[f:PRISMARSS\] that a simple first order truncation of the analytical solutions —curve labeled (1:1)— starts with a RSS error of approximately 1 meter in position, but, due to the errors introduced by the first order truncation of the secular terms, the RSS errors reach more than 10 km after one month. The solution is notably improved when taking the second order terms of the inverse corrections and the secular terms into account. This is shown with the curve labeled (2:1), that only reaches about 30 m at the end of the one-month propagation. The improvements are obtained with only a slight increase of the computational burden, because the inverse corrections and the secular frequencies are evaluated only once. The (3:2) propagation starts from a much smaller RSS error, of less than 1 cm, that only grows to about 10 cm by the end of day 30. Errors fall clearly below the mm level in the case of the (4:3) truncation, and to just a few $\mu\mathrm{m}$ in the (5:4) case. No further improvement of the RSS errors is observed if the direct periodic corrections are taken also up to the fifth order, yet a slight improvement is achieved in that case in the preservation of the energy integral, that reaches in this last case the numerical precision.
![RSS errors of different (S:P) truncations of the secular (S) and periodic terms (P) of the analytical solution in the PRISMA test case. Abscissas are days.[]{data-label="f:PRISMARSS"}](PrismaRSSr)
The behavior of the analytical solution is analogous in the case of the high-eccentricity GTO orbit in what respects to the secular terms, yet the errors of the periodic corrections are now of larger amplitude, as clearly observed in Fig. \[f:GTORSS\]. Now, the periodic oscillations of the (1:1) truncation may allow the RSS errors to grow to many tens of km. A secular trend in the RSS errors of about half meter per day of the (2:1) solution remains almost hidden along the 30 days propagation under periodic oscillations of about 30 m. A similar behavior is observed in the (3:2) propagation, where the amplitude of the oscillations is now reduced to the cm order. This amplitude is further reduced to hundredths of mm with the (4:3) truncation. Finally the (5:4) solution improves slightly the propagation, and the RSS errors remain of just a few $\mu\mathrm{m}$ along the whole propagation interval, thus showing the same quality as in the previous test case.
![RSS errors of different (S:P) truncations of the analytical solution in the GTO test case. Abscissas are days.[]{data-label="f:GTORSS"}](GtoRSSr)
Analogous tests carried out for TOPEX are presented in Fig. \[f:TOPEXRSS\]. The results are similar to those previously presented in Fig. \[f:PRISMARSS\] for PRISMA, and the apparent discrepancies when using the (2:1) solution are only due to the logarithmic scale, which encompasses a different range in each figure. In fact, the periodic oscillations of the RSS errors are of the same amplitude in both cases ($\sim2$ m), yet the secular trend grows at a low rate of less than 2 cm per day in the case of TOPEX while it does almost two orders of magnitude faster for PRISMA ($\sim1$ m/day). This fact should be attributed to the less perturbed orbit of TOPEX due to its higher altitude. The (3:2) solution provides quite similar RSS errors in both cases, TOPEX and PRISMA, whereas the (4:3) solution performs worse for TOPEX, whose RSS errors grow now at a secular rate about 5 times faster than in the case of PRISMA. These behavior is in agreement with the slight deterioration of the solution previously observed when testing the periodic corrections due to the proximity to the critical inclination. Things are balanced with the (5:4) order solution —also in agreement with the previously observed behavior due to the increased precision of the analytical solution— for which the RSS errors reach only the micrometer level at the end of the 30 day propagation.
![RSS errors of different (S:P) truncations of the analytical solution in the TOPEX test case. Abscissas are days.[]{data-label="f:TOPEXRSS"}](TopexRSSr)
As it is confirmed by the examples carried out, at the end of a long-term propagation the secular trend of the errors prevails over their periodic oscillations. Therefore, in practice, perturbation solutions in which the periodic terms are recovered to a lower order than the order at which the secular frequencies are truncated make full sense. The computational burden of these kinds of solutions is notably alleviated, and hence are definitely much more practicable for ephemeris computation. This is further illustrated for TOPEX in Fig. \[f:TOPEXRSSb\], where it is shown that the (5:3) truncation of the analytical solution might replace the more accurate (5:4) truncation in a long-term propagation with substantial reductions in computing time and minimum degradation of accuracy, which, besides, remains much more uniform along the whole propagation interval. Note that the scale of the ordinates axis has been changed from meters in Fig. \[f:TOPEXRSS\] to mm in Fig. \[f:TOPEXRSSb\].
![TOPEX: RSS errors of different (5:P) truncations. Abscissas are days.[]{data-label="f:TOPEXRSSb"}](TopexRSSrb)
Conclusions
===========
The main problem of artificial satellite theory is a two degrees of freedom model that captures the bulk of the gravitational effects undergone by low Earth orbits. Still, it lacks of enough integrals to obtain a closed form solution. On the other hand, perturbation methods disclose as specially successful in providing analytical solutions that can replace the true dynamics within a high degree of approximation. Using them, it is shown that there are wide regions of phase space in which the main problem behaves as if it were separable.
In particular, the main problem Hamiltonian has been completely reduced by reverse normalization. That is, the motion of the orbital plane is first decoupled from the motion on that plane, and then the remaining short-period terms are removed by averaging. The normalization is achieved without need of resorting to Hamiltonian simplification procedures, in this way radically departing from a well stablished tradition in artificial satellite theory. The computed analytical solution is exact to all practical purposes in the sense that, working in double-precision floating point aritmethic, there are no substantial differences with respect to exact numerical integrations (computed with extended precision) for reasonably long time intervals. In agreement with the particular value of the Earth’s oblateness coefficient, this precision is achieved only when the perturbation approach is extended up to, at least, the fifth order of the small parameter.
Support by the Spanish State Research Agency and the European Regional Development Fund under Projects ESP2016 -76585-R and ESP2017-87271-P (MINECO/ AEI/ERDF, EU) is recognized.
#### **Conflict of Interest** {#conflict-of-interest .unnumbered}
: The author declares that he has no conflict of interest.
[10]{} K. [Aksnes]{}. . , 4(1):119–121, September 1971.
K. T. [Alfriend]{} and S. L. [Coffey]{}. . , 32(2):163–172, February 1984.
V. I. Arnol’d. , volume 60 of [*Graduate Texts in Mathematics*]{}. Springer-Verlag, New York, 2nd edition, 1989.
J. V. [Breakwell]{} and J. [Vagners]{}. . , 2:253–264, June 1970.
R. A. [Broucke]{}. . , 58(2):99–123, February 1994.
D. [Brouwer]{}. . , 64:378–397, November 1959.
D. Brouwer and G. M. Clemence. . Academic Press, New York and London, 1961.
A. [Celletti]{} and P. [Negrini]{}. . , 61(3):253–260, March 1995.
S. Coffey and A. Deprit. . , 5(4):366–371, 1982.
S. L. [Coffey]{}, A. [Deprit]{}, and B. R. [Miller]{}. . , 39(4):365–406, December 1986.
R. A. [Cook]{}. . In B. Kaufman, K. T. Alfriend, R. L. Roehrich, and R. R. Dasenbrock, editors, [*Astrodynamics 1991*]{}, volume 76 of [ *Advances in the Astronautical Sciences*]{}, pages 2205–2221, P.O. Box 28130, San Diego, California 92198, USA, 1992. American Astronautical Society, Univelt, Inc.
J. M. A. [Danby]{}. . , 73(10):1031–1038, December 1968.
J. M. A. [Danby]{}. . Willmann-Bell, Richmond VA, 2nd edition, 1992.
C. E. [Delaunay]{}. , volume 28 of [*Mémoires de l’Academie des Sciences de l’Institut Impérial de France.*]{} , Paris, 1860.
A. [Deprit]{}. Canonical transformations depending on a small parameter. , 1(1):12–30, 1969.
A. Deprit. The elimination of the parallax in satellite theory. , 24(2):111–153, 1981.
A. [Deprit]{}. . , 26:9–21, January 1982.
A. [Deprit]{} and B. [Miller]{}. . , 45:189–200, 1989.
A. Deprit and A. Rom. . , 2(2):166–206, June 1970.
R. B. [Frauenholz]{}, R. S. [Bhat]{}, B. E. [Shapiro]{}, and R. K. [Leavitt]{}. . , 35:212–224, March 1998.
G. [Gaias]{}, C. [Colombo]{}, and M. [Lara]{}. . , [on line]{}:1–13, March 2020.
H. [Goldstein]{}, C. P. [Poole]{}, and J. L. [Safko]{}. . Addison-Wesley, 3rd edition, 2001.
P. A. [Hansen]{}. . , 2(3):183–281, 1855. .
D. [Hautesserres]{} and M. [Lara]{}. . , 127:505–526, April 2017.
L. M. [Healy]{}. . , 76(2):79–120, 2000.
M. [Irigoyen]{} and C. [Simó]{}. . , 55(3):281–287, March 1993.
I. G. [Izsak]{}. . , 6:81–107, 1963.
W. H. [Jefferys]{}. , 3:390–394, September 1971.
A. H. [Jupp]{}. . , 43(1-4):127–138, 1988.
W. [Kahan]{}. . Technical report, Electrical Engineering and Computer Science, University of California, Berkeley, CA, October 1997.
W. M. [Kaula]{}. . , 5:104–133, July 1961.
H. [Kinoshita]{}. . , 24:423–457, 1972.
H. [Kinoshita]{}. . , 379, July 1977.
A. [Konopliv]{}. . , 47:305, 1990.
A. S. [Konopliv]{}. . Interoffice Memorandum IOM 314.3 - 970, Jet Propulsion Laboratory, 4800 Oak Grove Dr, Pasadena, CA 91109, USA, March 1991.
Y. [Kozai]{}. . , 64:367–377, November 1959.
Y. [Kozai]{}. . , 67:446–461, September 1962.
M. Lara. . Technical Report DTS/MPI/MS/MN/99-053, Centre National d’Études Spatiales, 18, avenue Edouard Belin - 31401 Toulouse Cedex 9, France, May 1999.
M. Lara. . , 47(3-4):177–188, 1999.
M. [Lara]{}. . , 31(1):172–181, January 2008.
M. [Lara]{}. . In G. [G[ó]{}mez]{} and J. J. [Masdemont]{}, editors, [ *Astrodynamics Network AstroNet-II: The Final Conference*]{}, volume 44 of [ *Astrophysics and Space Science Proceedings*]{}, pages 151–166. Springer, Cham, 2016.
M. [Lara]{}. . , 129:137–151, September 2017.
M. [Lara]{} and S. [Ferrer]{}. . , 51(4):289–303, July 2013.
M. [Lara]{} and S. [Ferrer]{}. . , 36(5):055040, September 2015. .
M. [Lara]{}, J. F. [San-Juan]{}, and L. M. [L[ó]{}pez-Ochoa]{}. . , 120(1):39–56, September 2014.
M. [Lara]{}, J. F. [San-Juan]{}, and L. M. [L[ó]{}pez-Ochoa]{}. . In S. B. Broschart, J. D. Turner, K. C. Howell, and F. R. Hoots, editors, [*Astrodynamics 2013*]{}, volume 150 of [ *[Advances in the Astronautical Sciences]{}*]{}, pages 315–331, P.O. Box 28130, San Diego, California 92198, USA, January 2014. American Astronautical Society, Univelt, Inc.
M. [Lara]{}, R. [Vilhena de Moraes]{}, D. M. [Sanchez]{}, and A. F. B. A. [Prado]{}. . In R. Furfaro, S. Casotto, A. Trask, and S. Zimmer, editors, [*AAS/AIAA Spaceflight Mechanics Meeting 2015*]{}, volume 155 of [*[Advances in the Astronautical Sciences]{}*]{}, pages 437–455, P.O. Box 28130, San Diego, California 92198, USA, 2015. American Astronautical Society, Univelt, Inc.
M. [Lara]{}. . , 110:239–246, May 2015. .
M. [Lara]{}. . , 93(4):2019–2038, apr 2018.
M. [Lara]{}. . , 131(9):20, September 2019.
R. H. [Lyddane]{}. . , 68(8):555–558, October 1963.
G. [Metris]{}. . , 52:79–84, March 1991.
A. H. Nayfeh. . Wiley-VCH Verlag GmbH [&]{} Co. KGaA, Weinheim, Germany, 2004.
C. [Osácar]{} and J. F. [Palacián]{}. . , 60(2):207–223, October 1994.
S. [Persson]{}, B. [Jacobsson]{}, and E. [Gill]{}. . In [*Proceedings of the 56th International Astronautical Congress (IAC), October 17 - 21 2005, Fukuoka, Japan*]{}, pages 1–10. International Astronautical Federation (IAF), International Astronautical Federation (IAF), October 2005.
J. F. [San-Juan]{}, D. [Ortigosa]{}, L. M. [L[ó]{}pez-Ochoa]{}, and R. [L[ó]{}pez]{}. . , 60:137–148, June 2013.
F. [Tisserand]{}. . Gauthier-Villars et fils, Quai des Grands-Augustins, 55, Paris, 1889.
[^1]: The advantages of decoupling the motion of the instantaneous orbital plane from the in-plane motion are well known, and are commonly pursued in the search for efficient numerical integration methods, q.v. [@Lara2017if] and references therein.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
A. Ilderton\
Centre for Particle Theory, University of Durham,\
Durham DH1 3LE, UK.\
\
[email protected]
title: 'Radial evolution in anti-de Sitter spacetime'
---
General results {#general-results .unnumbered}
---------------
Consider $D+1$ dimensional spacetime with metric $g_{\mu\nu}$ and let $\nabla^2$ be the Laplacian, $$\nabla^2 = -\frac{1}{\sqrt{g}}{\partial}_a\, \sqrt{g}g^{ab}{\partial}_b.$$ Suppose we have a region $V$ of spacetime bounded by a surface ${\partial}V$. Let $\phi$ be a solution to Poisson’s equation in $V$, $$\nabla^2 \phi(x) = \rho(x) \quad\forall x\in V$$ for some function $\rho$. Let $G$ be a Green’s function for Poisson’s equation obeying $$\nabla^2 G(q,z) = \delta(q-z)/\sqrt{g(z)},$$ which holds for $z$, $q$ both within and in some neighbourhood of $V$ and the boundary, then we have $$\phi(x)\nabla^2G(x,z) - G(x,z)\nabla^2\phi(x) = \phi(x)\delta(x-z)/\sqrt{g(z)} - G(x,z)\rho(x).$$ Integrating both sides over $V$ and applying Green’s theorem gives $$\begin{split}
\int\limits_{{\partial}V}\!{\mathrm{d}}y\,\, G(y,z)&\frac{{\partial}}{{\partial}n}\phi(y)-\phi(y)\frac{{\partial}}{{\partial}n} G(y,z) \\
&= \int\limits_V\!{\mathrm{d}}x\sqrt{g(x)}\,\,\phi(x)\delta(x-z) - \int\limits_V\!{\mathrm{d}}x\sqrt{g(x)}\,\,G(x,z)\rho(x)
\end{split}$$ where ${\mathrm{d}}y$ is a measure on ${\partial}V$ and ${\partial}/{\partial}n$ is the normal derivative on the boundary, $$\frac{{\partial}}{{\partial}n} = n^\mu\frac{{\partial}}{{\partial}x^\mu}\bigg|_{{\partial}V}.$$ Now choose $z\in V$ and let $\phi(x) = G(q,x)$ for some $q\not\in V$, so that $\rho(x)=\delta(x-q)$ which vanishes in our integration range. Substituting in the above we find $$\label{factor}
\int\limits_{{\partial}V}\!{\mathrm{d}}y\,\, G(q,y)\overleftrightarrow{\frac{{\partial}}{{\partial}n}}G(y,z)= G(q,z).$$ where $\leftrightarrow = \leftarrow - \rightarrow$. A simple extension shows that the derivation above also holds when $\phi$ and $G$ are a solution and Green’s function of the Helmholtz equation $\nabla^2 + m^2 =\rho$, rather than Poisson’s equation. As applied to a scalar field theory this implies that the field propagator may be factorised. This result, in certain situations, also applies to the propagator of bosonic string theory, and has been applied to the construction of the vacuum wave functional [@me2] and investigating T-duality [@us1] [@us2].
It is interesting to note that this property can be derived from considering first quantised particles. The off-shell free space propagator from the point $x_i$ to the point $x_f$ may be written as a sum over all paths from $x_i$ to $x_f$ equipped with all possible worldline metrics $e$ [@Brink] [@PolyakovBook], $$\label{prop-path}
G(x_f; x_i) = \int{\!\mathscr{D}}(x,\,e)\,\, e^{-\int_0^1 {\mathrm{d}}\xi\sqrt{e}\,\, ({\dot x\cdot\dot x}/(2e)+m^2/2)}\bigg|_{x(0)=x_i}^{x(1)=x_f} \\$$ The path is parameterised by $\xi$ and $x(0)$ is the point $x_i$, $x(1)$ the point $x_f$. The intrinsic metric $e$ transforms as a (0,2) tensor, $$e(\xi) = \bigg(\frac{{\mathrm{d}}\xi'}{{\mathrm{d}}\xi}\bigg)^2e(\xi'),$$ so the propagator is manifestly reparametrisation invariant[^1]. This is the particle analogue of the Polyakov integral in string theory.
Now, with reference to the preceding second quantised discussion, consider a particle path from $z\in V$ to $q\not\in V$. The derivation of the property (\[factor\]) begins with the simple observation that at some point the path must cross ${\partial}V$ at some (intrinsic, not physical) time. This implies the sum over all paths defining the propagator can factorised so that formally $$\sum_{\textrm{paths }z\rightarrow q} e^{-\textrm{length }z\rightarrow q}= \sum\limits_y \bigg( \sum_{\textrm{paths }z\rightarrow y}
e^{-\textrm{length }z\rightarrow y)}\bigg) \bigg( \sum_{\textrm{paths }y\rightarrow q}
e^{-\textrm{length }y\rightarrow q)}\bigg)$$ where $y\in{\partial}V$. To make this factorisation explicit, insert into (\[prop-path\]) a resolution of the identity, $$\label{prop-trick}
1=\int\!{\mathrm{d}}\xi'\sqrt{e(\xi')}\,\, J(x)\,\delta\big(F(x(\xi'))\big)$$ where $F(x)=0$ defines the surface ${\partial}V$. This delta function has support for any path $x^\mu(\xi)$ as we have described. $e$ is here an arbitrary worldline metric which will soon drop out. Taylor expanding $F(x)$ the Jacobian $J$ which makes the insertion unity is easily found to be $$J=\frac{\nabla_\mu S{\partial}_\xi x^\mu(\xi)}{\sqrt{e(\xi)}}\bigg|_{S(x(\xi))=0}.$$ Note that this Jacobian is reparametrisation invariant. Taking the integral over $\xi'$ outside and distinguishing between worldline times earlier and later than $\xi'$, the $x^\mu$ integrals in the propagator can be written $$\begin{split}
\int\!{\mathrm{d}}\xi'\sqrt{e(\xi')}\,\left[\prod_{\xi<\xi'}\int\!{\mathrm{d}}x^\mu(\xi)\right]\int\!{\mathrm{d}}x^\mu(\xi')\,&J\,\delta\big(F(x)\big)\left[ \prod_{\xi>\xi'} \int\!{\mathrm{d}}x^\mu(\xi)\right] \\
&\exp\left(-\sum\limits_{\xi<\xi'} S[x(\xi)] - \sum\limits_{\xi>\xi'} S[x(\xi)]\right).
\end{split}$$ Integrating over the delta function gives $$\int\limits_{{\partial}V}\!{\mathrm{d}}y \frac{J}{|\nabla F|} = \int\limits_{{\partial}V}{\mathrm{d}}y\,\, n_\mu\frac{\partial_\xi x^\mu}{\sqrt{e}}.$$ We can write $\partial_\xi x(\xi)$ as a two sided derivative $${\partial}_\xi x^\mu(\xi')= \lim_{h\rightarrow 0} \frac{x^\mu(\xi'+h)- x^\mu(\xi'-h)}{2h}$$ which splits the path integration into a pair of terms of two products, one of which has an insertion. Everything is invariant under reparametrisations of the worldline and so have no explicit $\xi'$ dependence, the integral over which gives a finite volume, leaving $$\begin{split}
G_0=\frac{1}{2}\int\limits_{{\partial}V}\!{\mathrm{d}}y\int{\!\mathscr{D}}(x,e)\,\, &\frac{n.\dot{x}(\xi_\text{final})}{\sqrt{e(\xi_\text{final})}}e^{-S[x]} \int{\!\mathscr{D}}(x,e)\,\,e^{-S[x]} \\
+ \frac{1}{2}\int\limits_{{\partial}V}\!{\mathrm{d}}y\int{\!\mathscr{D}}(x,e)\,\, e^{-S[x]}& \int{\!\mathscr{D}}(x,e)\,\,\frac{n.\dot{x}(\xi_\text{initial})}{\sqrt{e(\xi_\text{initial})}}e^{-S[x]}
\end{split}$$ It can be seen from an integration by parts in the action that the insertions of the Jacobian can be taken outside the integrals as derivatives with respect to boundary data, $$\frac{{\partial}}{{\partial}x^\mu(\xi_\text{final})} = -\frac{1}{2}\frac{\dot{x}^\mu(\xi_\text{final})}{\sqrt{e(\xi_\text{final})}}$$ or with an overall plus sign for the initial time, recovering the result (\[factor\]).
Example - AdS space {#example---ads-space .unnumbered}
-------------------
Let us give an explicit example which may be of interest. Since its proposal [@Maldec] there has been huge interest in studying the correspondence between string theory in anti-de Sitter space and its holographic dual conformal field theory on the boundary ([@review] for a comprehensive review). A variety of new approaches to studying the correspondence have recently been employed (see for example [@Berkovits] and [@Bena]...[@Hatsu]). Berkovits’ approach [@Berkovits2] gives a quantisable action for the superstring in AdS space times a compact manifold, with Ramond Ramond flux, but the complexity of the action makes hard work of calculating scattering amplitudes [@Berkovits3]...[@Triv].
Given the difficulties of string calculations in anti-de Sitter space, it is worthwhile to pursue unconventional approaches. We believe the factorisation property discussed plays a crucial role in defining the radial dependence of scalar fields in AdS, much as we found for time evolution in flat space [@us1] [@us2]. Translation invariance simplified the flat space calculations so we do not expect our Feynman diagram arguments to be so simple in AdS, but we can derive the essential property.
We will represent Euclideanised anti-de Sitter spacetime as the upper half space $x^0>0$ with metric $$\begin{split}
{\mathrm{d}}s^2 &= \frac{1}{{x^0}^2}\big({\mathrm{d}}x^0{\mathrm{d}}x^0+{\mathrm{d}}{{\bf x}}^i{\mathrm{d}}{{\bf x}}^i\big)=: A_{ab}{\mathrm{d}}x^a {\mathrm{d}}x^b, \\
\quad A(x^0)&:={\text{Det}\,}A_{ab} = \bigg(\frac{1}{x^0}\bigg)^{D+1}
\end{split}$$ where $i=1\ldots D$ and $x^0$ is the ‘radial’ direction. For a scalar field of mass $m$ define $\nu=\sqrt{ D^2/4 +m^2}$, then the scalar field propagator which vanishes on the boundary of the spacetime (compactified $\mathbb{R}^D$) is, labelling the final and initial values of $x^0$ as $r_f$ and $r_i$ respectively [@Burgess] [@Muck], $$\label{AdS-prop}
G_\text{AdS}(r_f,{{\bf x}}_f;r_i,{{\bf x}}_i) = -\int\!\frac{{\mathrm{d}}^D{{\bf k}}}{(2\pi)^D}\, (r_f r_i)^{D/2}e^{-i{{\bf k}}.({{\bf x}}_f-{{\bf x}}_i)}K_\nu(|{{\bf k}}|r_f)I_\nu(|{{\bf k}}|r_i)$$ when $r_f>r_i$ and the Bessel functions $I_\nu$ and $K_\nu$ [@Grad] are exchanged otherwise. Now let $V$ be the space $x^0<r$, let $z=(r_i,{{\bf x}}_i)$ and let $q(r_f,{{\bf x}}_f)$ with $r_f >r > r_i$. Our general result implies that the propagator factorises as $$\label{AdS-fact}\begin{split}
G_\text{AdS}(r_f,{{\bf x}}_f;r_i,{{\bf x}}_i)= \int\!{\mathrm{d}}^D{{\bf y}}A(r)\,\,&\bigg(r^2\frac{\partial}{\partial r}G_\text{AdS}(r_f,{{\bf x}}_f;r,{{\bf y}})\bigg)G_\text{AdS}(r,{{\bf y}};r_i,{{\bf x}}_i) \\
&- G_\text{AdS}(r_f,{{\bf x}}_f;r,{{\bf y}})\bigg(r^2\frac{\partial}{\partial r}G_\text{AdS}(r,{{\bf y}};r_i,{{\bf x}}_i)\bigg).
\end{split}$$ To prove this directly is a simple matter. Inserting the explicit representation of the propagator (\[AdS-prop\]) into the right hand side of (\[AdS-fact\]) we find $$\begin{split}
-\int\!\frac{{\mathrm{d}}^D({{\bf k}},{{\bf p}},{{\bf y}})}{(2\pi)^{2D}} &(r_f r_i)^{D/2}\,\,e^{-i{{\bf k}}.({{\bf x}}_f-{{\bf y}})}e^{-i{{\bf p}}.({{\bf y}}-{{\bf x}}_i)}K_\nu(|{{\bf k}}|r_f)I_\nu(|{{\bf p}}|r_i) \\
&\times r^{1-D/2}\bigg\{ K_\nu(|{{\bf p}}|r)\frac{\partial}{\partial r} r^{D/2}I_\nu(|{{\bf k}}|r) - I_\nu(|{{\bf k}}|r)\frac{\partial}{\partial r} r^{D/2}K_\nu(|{{\bf p}}|r)\bigg\}.
\end{split}$$ The integral over ${{\bf y}}$ gives a momentum conserving delta function which allows us to do the integral over ${{\bf p}}$, say. For brevity write $s\equiv|{{\bf k}}|r$ and the result of these integrations is $$\begin{split}\label{big}
-\int\!\frac{{\mathrm{d}}^D{{\bf k}}}{(2\pi)^D}&(r_f r_i)^{D/2}\,\,e^{-i{{\bf k}}.({{\bf x}}_f-{{\bf x}}_i)} K_\nu(|{{\bf k}}|r_f)I_\nu(|{{\bf k}}|r_i) \\
&\times s^{1-D/2}\bigg\{ K_\nu(s)\frac{\partial}{\partial s} s^{D/2}I_\nu(s) - I_\nu(s)\frac{\partial}{\partial s} s^{D/2}K_\nu(s)\bigg\} \\
= -\int\!\frac{{\mathrm{d}}^D{{\bf k}}}{(2\pi)^D}&(r_f r_i)^{D/2}\,\,e^{-i{{\bf k}}.({{\bf x}}_f-{{\bf x}}_i)} K_\nu(|{{\bf k}}|r_f)I_\nu(|{{\bf k}}|r_i) \\
&\times s^{1-\nu}\bigg\{ K_\nu(s)\frac{\partial}{\partial s} s^\nu I_\nu(s) - I_\nu(s)\frac{\partial}{\partial s} s^\nu K_\nu(s)\bigg\}
\end{split}$$ plus two terms coming from the derivatives of $s^{D/2}$ which cancel. We require the final line of the above to be unity in order to recover $G_\text{AdS}$. Applying the Bessel function properties [@Grad] $$\frac{\partial}{\partial s}s^\nu I_{\nu}(s) = s^\nu I_{\nu-1}(s),\qquad \frac{\partial}{\partial s}s^\nu K_{\nu}(s) = -s^\nu K_{\nu-1}(s),$$ the final line of (\[big\]) becomes $$sI_{\nu-1}(s)K_\nu(s)+sI_\nu(s)K_{\nu-1}(s) = 1,$$ a standard Bessel function identity for unity, leaving $$-\int\!\frac{{\mathrm{d}}^D{{\bf k}}}{(2\pi)^D}(r_f r_i)^{D/2}\,\,e^{-i{{\bf k}}.({{\bf x}}_f-{{\bf x}}_i)} K_\nu(|{{\bf k}}|r_f)I_\nu(|{{\bf k}}|r_i) \equiv G_\text{AdS}(r_f,{{\bf x}}_f;r_i,{{\bf x}}_i).$$ We have proven (\[AdS-fact\]). For $r_f<r<r_i$ the right hand side of (\[AdS-fact\]) picks up a minus sign. These results also hold in the limit in which $r_f=r$ or $r=r_i$.
In the AdS/CFT correspondence translations in the bulk radial direction correspond to conformal transformations in the boundary field theory. Since the correspondence holds between scalar field theory in $AdS_{D+1}$ and conformal field theory on $\mathbb{R}^D$ [@WittenAdS], there is an opportunity to study how our results relate to a holographic description in terms of scaling of operators in the boundary conformal field theory. Toward this end we note that our results hold for the propagator $$\begin{split}
G_\epsilon(r_f,{{\bf x}}_f;r_i,{{\bf x}}_i) &:= G_\text{AdS}(r_f,{{\bf x}}_f;r_i,{{\bf x}}_i) \\
&\hspace{5pt}+ \int\!\frac{{\mathrm{d}}^D{{\bf k}}}{(2\pi)^D}(r_f r_i)^{D/2}e^{-i{{\bf k}}.({{\bf x}}_f-{{\bf x}}_i)}K_\nu(|{{\bf k}}|r_f)K_\nu(|{{\bf k}}|r_i)\frac{I_\nu(|{{\bf k}}|\epsilon)}{K_\nu(|{{\bf k}}|\epsilon)}
\end{split}$$ considered in [@Muck] which vanishes not on the spacetime boundary but on the near-boundary surface $x^0=\epsilon$ and is used to regulate boundary divergences when investigating the conformal field theory.
[99]{} A. Ilderton, “The vacuum state functional of interacting string field theory", hep-th/0506173
A. Ilderton and P. Mansfield, “Time evolution in string field theory and T-duality" Phys.Lett.[**B607**]{}(2005) p294 \[hep-th/0410267\]
A. Ilderton and P. Mansfield, “Timelike T-duality in the string field Schrödinger functional" \[arxiv:hep-th/0411166\]
L. Brink, P. Di Vecchia, P. Howe “A locally supersymmetric and reparametrisation invariant action for the spinning string",
A.M. Polyakov, “Gauge fields and strings”, Harwood Academic Publishers (1987) ISBN 3-7186-0492-6
J.M. Maldacena, “The large $N$ limit of superconformal field theories and supergravity”, hep-th/9711200
O. Aharony et al, “Large N field theories, string theory and gravity”, hep-th/ , hep-th/9905111
N.Berkovits, “ICTP lectures on covariant quantization of the superstring”, published in ‘Trieste 2002, Superstrings and related matters’, hep-th/0209059
I. Bena, J. Polchinski, R. Roiban, “Hidden symmetries of the AdS$_5\times\mathbb{S}^5$ superstring”, , hep-th/0305116
L. Dolan, C. R. Nappi, E. Witten, “Yangian symmetry in D = 4 superconformal Yang-Mills theory”, hep-th/0401243.
M. Hatsuda and K. Yoshida, “Classical integrability and super Yangian of superstring on AdS$_5\times\mathbb{S}^5$”, hep-th/0407044.
N. Berkovits “Super-Poincaré covariant quantization of the superstring”, , hep-th/0001035
N. Berkovits, “Superstring vertex operators in an AdS$_5\times\mathbb{S}^5$ background” , hep-th/0009168;
L. Dolan, E. Witten, “Vertex operators for AdS3 background with Ramond Ramond flux”, , hep-th/9910205
K. Bobkov, L. Dolan “Three-graviton amplitude in Berkovits-Vafa-Witten variables”, , hep-th/0201027
G. Trivedi, “Correlation functions in Berkovits’ pure spinor formulation” , hep-th/0205217
C.P.Burgess and C.A.Lútken, “Propagators and effective potentials in anti-de Sitter space”,
W. Múck and K.S.Viswanathan, “Conformal field theory correlators from classical scalar field theory on anti-de Sitter space”,
I.S.Gradshteyn and I.M.Ryzhik, “Table of integrals, series and products” ISBN 0-12-294760-6
E. Witten, “Anti-de Sitter space and holography”, Adv. Theor. Math. Phys. [**2**]{}, 253 (1998) \[arXiv:hep-th/9802150\]
[^1]: The over counting of equivalent paths is removed by dividing by the volume of the space of reparametrisations, we will not repeat the explicit evaluation of the integrals as this is well documented, see for example [@PolyakovBook].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Orthogonal Time Frequency Space (OTFS) modulation has been recently proposed to be robust to channel induced Doppler shift in high mobility wireless communication systems. However, to the best of our knowledge, none of the prior works on OTFS have derived it from first principles. In this paper, using the ZAK representation of time-domain (TD) signals, we rigorously derive an orthonormal basis of approximately time and bandwidth limited signals which are also localized in the delay-Doppler (DD) domain. We then consider DD domain modulation based on this orthonormal basis, and derive OTFS modulation. To the best of our knowledge, this is the first paper to rigorously derive OTFS modulation from first principles. We show that irrespective of the amount of Doppler shift, the received DD domain basis signals are localized in a small interval of size roughly equal to the inverse time duration along the Doppler domain and of size roughly equal to the inverse bandwidth along the delay domain (time duration refers to the length of the time-interval where the TD transmit signal has been limited). With sufficiently large time duration and bandwidth, there is little interference between information symbols modulated on different basis signals, which allows for joint DD domain equalization of all information symbols. This explains the inherent robustness of DD domain modulation to channel induced Doppler shift when compared with Orthogonal Frequency Division Multiplexing (OFDM). The degree of localization of the DD domain basis signals is inversely related to the time duration of the transmit signal, which explains the trade-off between robustness to Doppler shift and latency.'
author:
-
title: Derivation of OTFS Modulation from First Principles
---
Orthonormal Basis, Delay-Doppler, ZAK Representation, OTFS, Doppler Shift.
Introduction
============
Next generation wireless communication systems are expected to support reliable and high data rate communication even at very high mobile speed [@IMT2020]. However, the modulation waveform used in Fifth Generation (5G) communication systems is based on Orthogonal Frequency Division Multiplexing (OFDM) for which communication reliability and data rate is known to degrade in high mobility scenarios [@NR5G]. Recently, Orthogonal Time Frequency Space (OTFS) modulation has been proposed to be robust to channel induced Doppler shift when compared to OFDM [@HadaniOTFS1; @HadaniOTFS2; @HadaniOTFS3]. In OTFS modulation, information is embedded in the delay-Doppler (DD) domain. The information bearing DD signal is then converted to a time-domain (TD) transmit signal. At the receiver, the received TD signal is converted to a DD domain signal from which the information symbols are decoded, i.e., modulation, demodulation and channel estimation are all performed in the DD domain [@channel; @mimootfs; @MCMC; @Estimation].
To the best of our knowledge, none of the prior works have rigorously derived OTFS modulation from first principles. In the absence of a rigorous mathematical derivation of OTFS modulation and its basis waveforms, it is difficult for communication engineers to fully understand the robustness of OTFS modulation to Doppler shift. A deeper understanding of DD domain modulation and basis waveforms is required to design modulation and demodulation methods which are robust to channel induced Doppler shift in very high mobility scenarios (e.g., high speed train, air-to-ground communication). Therefore, in this paper, using the ZAK representation/transform[^1] of TD signals, we rigorously derive OTFS modulation from first principles. The novel contributions of this paper are:
- In Section \[seczak\], using the ZAK representation of TD signals, we derive an expression for TD signals which are neither time-limited nor bandwidth limited, but which are perfectly localized (i.e., Dirac-delta impulse) in the DD domain. In Lemma \[plem1st\] we show that these TD signals are an impulse train which is time-shifted and multiplied by a complex exponential. In Theorem \[thmbasis\] we show that these signals form a basis for all TD signals.
- The TD basis signals derived in Section \[seczak\] are neither time-limited nor bandwidth limited. Therefore, in Section \[secbw\], by approximately limiting the TD basis signals of Section \[seczak\] along time and frequency domains, we obtain the expression for TD signals which are approximately time and bandwidth limited. We then derive the ZAK representation of these TD signals in Theorem \[thm21\]. We show that due to time and bandwidth limitation, the corresponding DD domain signal is no more perfectly localized at a point but is instead spread over an interval whose size along the delay and Doppler domains is roughly equal to the inverse bandwidth and inverse time duration respectively.
- Further, in Theorem \[lem18\] in Section \[secbw\], for a given $(T , \Delta f)$, $\Delta f = 1/T$, and positive integers $M, N$, we derive a basis of orthonormal signals which are approximately time-limited to $NT$ seconds and bandwidth limited to $M \Delta f$ Hz, and are also localized in an interval of size inverse-bandwidth and inverse-time duration along the delay and Doppler domain respectively. The dimensionality of this basis is equal to the time-bandwidth product $NT \times M \Delta f = MN$. Our derivation therefore shows the important result that the additional constraint of DD domain localization does not reduce the dimensionality of approximately time and bandwidth limited signals.
- Using the orthonormal basis derived in Section \[secbw\], in Section \[propmodulation\] we consider DD domain modulation, where DD domain information symbols linearly modulate the orthonormal DD domain basis signals derived in Section \[secbw\].
- In Theorem \[thmotfsmod\] in Section \[deriveOTFS\], we derive OTFS modulation from the DD domain modulation derived in Section \[propmodulation\]. To the best of our knowledge, this paper is the first to rigorously derive OTFS modulation from first principles.
- In Section \[Zakrecv\] we derive an expression for the spectral efficiency (SE) achieved by the DD domain modulation derived in Section \[propmodulation\].
- In Section \[secbetter\], we study the localization of the received basis signals in the DD domain. We show that the energy transmitted on a particular DD domain basis signal interferes with only a small fraction of the other $(MN -1)$ basis signals when compared to the fraction of interfered sub-carriers in OFDM. With increasing Doppler shift, the variation in the fraction of interfered DD domain basis signals is much smaller than the variation in the fraction of interfered sub-carriers in OFDM. This explains the inherent robustness of DD domain modulation to channel induced Doppler shift.
- It is also observed that for a given $M$ (i.e., given bandwidth $M \Delta f$), the fraction of interfered DD domain basis signals decreases with increasing $N$ which makes it easier to perform joint equalization of all $MN$ information symbols in the DD domain. However, with increasing $N$ the time duration $NT$ increases, which increases latency.
[*Notations:*]{} The continuous-time Dirac-delta signal with impulse at $t=0$ is denoted by $\delta(t)$. The discrete-time impulse signal is denoted by $\delta[ k], k \in {\mathbb Z}$, where $\delta[k] = 1$ for $k=0$ and is zero otherwise. For any matrix ${\bf A}$, $\left\vert {\bf A} \right\vert$ denotes the determinant of ${\bf A}$. Also, $A[p,q]$ denotes the element in the $p$-th row and $q$-th column of matrix ${\bf A}$. The zero mean circular symmetric complex Gaussian distribution with variance $\sigma^2$ is denoted by ${\mathcal C}{\mathcal N}(0, \sigma^2)$. The conjugate of a complex number $z \in {\mathbb C}$ is denoted by $z^*$. The real part of a complex number $z$ is denoted by $Re(z)$. For any real number $x$, $\lfloor x \rfloor$ denotes the greatest integer smaller than or equal to $x$. For any integer $M$ and real number $x$, $[ x ]_{_M}$ denotes the smallest unique non-negative real number such that $\left(x - [ x ]_{_M} \right)$ is an integer multiple of $M$. The symbol $\%$ denote percent, e.g. $12.5 \%$ is $0.125$. For any set ${\mathcal A}$, $\left\vert {\mathcal A} \right\vert$ denotes its cardinality. The abbreviation R.H.S. stands for “right hand side" and w.r.t. stands for “with respect to". For any real $x$, $\mbox{\small{sinc}}(x) {\stackrel {\Delta} {=} }\frac{\sin(\pi x)}{\pi x}$. For any two sets ${\mathcal A}$ and ${\mathcal B}$, ${\mathcal A} \subseteq {\mathcal B}$ means that ${\mathcal A}$ is a subset of ${\mathcal B}$.
The ZAK Representation of Time-Domain (TD) Signals {#seczak}
==================================================
Let $x(t)$ be a complex time-continuous signal. For any $T > 0$ we define the ZAK representation of $x(t)$ by the two-dimensional signal [@Janssen] $$\begin{aligned}
\label{zdef1}
{\mathcal Z}_x(\tau, \nu) & {\stackrel {\Delta} {=} }& \sqrt{T} \, \sum\limits_{n=-\infty}^{\infty} \, x(\tau + nT) \, e^{-j 2 \pi n \nu T } \,,\, \nonumber \\
& & -\infty < \tau < \infty \,,\, -\infty < \nu < \infty. \end{aligned}$$ The following result from [@Saif2020] shows that a channel induced time shift (due to path delay) and frequency shift (due to Doppler) to a TD signal $x(t)$ corresponds to simple shifts along the $\tau-$ and $\nu-$ domains in its ZAK representation. Therefore, subsequently we refer to the $\tau-$ and $\nu-$ domains as the “delay" and “Doppler" domain respectively, i.e., jointly we refer to them as the delay-Doppler (DD) domain.
\[zlem52\] Let there be only one channel path with a delay of $\tau_0$ and a Doppler shift of $\nu_0$. With $x(t)$ as the transmit signal, the ZAK representation of the noise-free received signal $r(t) = x(t - \tau_0) e^{j 2 \pi \nu_0 (t - \tau_0)}$ is given by $$\begin{aligned}
\label{zakmod2}
{\mathcal Z}_r(\tau, \nu) & = & e^{j 2 \pi \nu_0 (\tau - \tau_0)} {\mathcal Z}_x(\tau - \tau_0 , \nu - \nu_0)\end{aligned}$$ i.e., delay and Doppler shift in TD results in a shift of $\tau_0$ and $\nu_0$ along the $\tau-$ and $\nu-$ domains respectively.
See proof of Theorem $1$ in [@Saif2020].
In the following we present important results on ZAK representation which will be useful later. These results are available in [@Janssen] for normalized $T = \Delta f = 1$. Here we present these results for a general $T$ and $\Delta f = 1/T$, and for the general audience we also provide much simpler and detailed step by step proof of these results in the appendix. These results from [@Janssen] have been mentioned as “Result", whereas our original/novel results stated and proved in this current paper have been referred to as “Lemma" or “Theorem".
The following result states that the ZAK representation of a TD signal is [*quasi-periodic*]{} along the delay and Doppler domain.
\[zlem1\] For any $x(t)$, the corresponding ZAK representation ${\mathcal Z}_x(\tau,\nu)$ is periodic along the Doppler domain with a period of $\Delta f = 1/T$ and is quasi-periodic along the delay domain with a period $T$, i.e., $$\begin{aligned}
\label{zeqn2}
{\mathcal Z}_x(\tau + T, \nu) & = & e^{j 2 \pi \nu T} \, {\mathcal Z}_x(\tau,\nu) \,,\, \nonumber \\
{\mathcal Z}_x \left(\tau,\nu + \Delta f \right) & = & {\mathcal Z}_x(\tau,\nu).\end{aligned}$$
See Appendix \[prf\_zlem1\].
From (\[zeqn2\]) it follows that for any integer $n$, ${\mathcal Z}_x(\tau + nT, \nu) = e^{j 2 \pi \nu T} \, {\mathcal Z}_x(\tau + (n-1)T,\nu)$, repeated use of which gives $$\begin{aligned}
\label{zquasieqn1}
{\mathcal Z}_x(\tau + nT, \nu) = e^{j 2 \pi \nu n T} \, {\mathcal Z}_x(\tau ,\nu), \,\,\, n \in {\mathbb Z}.\end{aligned}$$ Similarly, from (\[zeqn2\]) it also follows that $$\begin{aligned}
\label{zquasieqn2}
{\mathcal Z}_x(\tau, \nu + m \Delta f) = {\mathcal Z}_x(\tau ,\nu), \,\,\, m \in {\mathbb Z}.\end{aligned}$$
Conversely, it is also true that if a DD domain signal ${\mathcal Z}_x(\tau,\nu)$ satisfies the quasi-periodicity conditions in (\[zeqn2\]), then there exists a unique TD signal $x(t)$ whose ZAK representation is ${\mathcal Z}_x(\tau,\nu)$ (see $(2.35)$ and $(2.36)$ in [@Janssen]).
The TD signal $x(t)$ and its Fourier transform ${\mathcal F}_x(f) = \int\limits_{-\infty}^{\infty} x(t) e^{-j 2 \pi f t} \, dt$ can be obtained from the ZAK representation ${\mathcal Z}_x(\tau,\nu)$ as stated in the following result.
\[zlem3\] The TD signal $x(t)$ can be recovered from its ZAK representation by $$\begin{aligned}
\label{zeqn6}
x(t) & = & \sqrt{T} \int\limits_{0}^{\Delta f} \, {\mathcal Z}_x(t, \nu) \, d\nu\end{aligned}$$ and the Fourier transform of $x(t)$ is given by $$\begin{aligned}
\label{zeqn7}
{\mathcal F}_x(f) & = & \frac{1}{\sqrt{T}} \int\limits_{0}^T {\mathcal Z}_x(\tau, f) \, e^{-j 2 \pi f \tau} \, d\tau.\end{aligned}$$
See Appendix \[prf\_zlem3\].
We know that it is not possible to simultaneously localize a signal in the time as well as in the frequency domain, i.e., there exists no TD signal $x(t)$ which is zero outside some interval $[T_1 \,,\, T_2]$ and whose Fourier transform ${\mathcal F}_x(f)$ is also zero outside some interval $[F_1 \,,\, F_2]$, where $T_1, T_2, F_1, F_2$ are all finite. However, there exists TD signals which are simultaneously localized in the delay as well as the Doppler domain. A DD domain signal localized at $\tau = \tau_0$ along the delay domain and at $\nu = \nu_0$ along the Doppler domain ($0 \leq \tau_0 < T$, $0 \leq \nu_0 < \Delta f$) is given by $$\begin{aligned}
\label{lceqn1}
{\mathcal Z}_{(p, \tau_0, \nu_0)}(\tau, \nu) & \hspace{-2mm} {\stackrel {\Delta} {=} }& \hspace{-2mm}
\sum\limits_{m=-\infty}^{\infty} \sum\limits_{n = -\infty}^{\infty} {\Big (} e^{j 2 \pi \nu_0 n T} \delta(\tau - \tau_0 - nT) \nonumber \\
& & \hspace{24mm} \delta(\nu - \nu_0 - m \Delta f) {\Big )}\end{aligned}$$ where $\delta(\tau)$ and $\delta(\nu)$ are the Dirac-delta impulse signal along the delay and Doppler domains respectively. In Fig. \[fig1\] we illustrate the signal ${\mathcal Z}_{(p, \tau_0, \nu_0)}(\tau, \nu)$ in the DD domain. In Fig. \[fig1\], the location of the impulses (see R.H.S. in (\[lceqn1\])) are denoted by dark dots. The complex value and co-ordinate of each DD domain impulse is mentioned next to it. Only a portion of the DD domain is illustrated as the delay and Doppler domains extend infinitely in both directions.
\[fig1\]
\[fig2\]
This DD domain signal ${\mathcal Z}_{(p, \tau_0, \nu_0)}(\tau, \nu)$ satisfies the quasi-periodicity conditions in (\[zeqn2\]) and therefore the corresponding TD signal $p_{(\tau_0,\nu_0)}(t)$ is given by the following Lemma.
\[plem1st\] The TD signal $p_{(\tau_0, \nu_0)}(t)$ having ZAK representation ${\mathcal Z}_{(p, \tau_0, \nu_0)}(\tau, \nu)$ in (\[lceqn1\]) is given by $$\begin{aligned}
\label{plemeqn2}
p_{(\tau_0, \nu_0)}(t) & = & \sqrt{T} \sum\limits_{n=-\infty}^{\infty} e^{j 2 \pi \nu_0 n T } \, \delta(t - \tau_0 - nT).\end{aligned}$$
See Appendix \[prf\_plem1st\].
Therefore, for a given $(\tau_0, \nu_0)$, the TD signal $p_{(\tau_0, \nu_0)}(t)$ is essentially an impulse train multiplied by the complex exponential $e^{j 2 \pi \nu_0 (t - \tau_0) }$. The impulses are spaced $T$ seconds apart and the impulse in the time-interval $[0 \,,\, T)$ is located at $t = \tau_0$ (see Fig. \[fig2\], where we have plotted $\frac{1}{\sqrt{T}} \left\vert p_{(\tau_0, \nu_0)}(t) \right\vert$ vs. $t/T$ for $\tau_0 = 0.6 T, \nu_0 = 0.2 \Delta f$). Further, we note that for any $(\tau_0, \nu_0), 0 \leq \tau_0 < T, 0 \leq \nu_0 < \Delta f$, $p_{(\tau_0, \nu_0)}(t)$ can be obtained from the TD signal $p_{(0, 0)}(t)$ by firstly multiplying $p_{(0, 0)}(t)$ by $e^{j 2 \pi \nu_0 t}$ (equivalent to a shift by $\nu_0$ along the Doppler domain) and then delaying this product signal by $\tau_0$ (equivalent to a shift by $\tau_0$ along the delay domain).
The next theorem states that the TD signals $p_{(\tau_0, \nu_0)}(t) \,,\, 0 \leq \tau_0 < T \,,\, 0 \leq \nu_0 < \Delta f$, form a basis for the space of TD signals.
\[thmbasis\] Any TD signal $x(t)$ can be expressed in terms of the basis signals $p_{(\tau_0, \nu_0)}(t) \,,\, 0 \leq \tau_0 < T \,,\, 0 \leq \nu_0 < \Delta f$, i.e. $$\begin{aligned}
\label{thmbasiseqn1}
x(t) & = & \int_{0}^T \int_{0}^{\Delta f} \hspace{-2mm} c_x(\tau_0, \nu_0) \, p_{(\tau_0, \nu_0)}(t) \, d\tau_0 \, d\nu_0, \nonumber \\
c_x(\tau_0, \nu_0) & {\stackrel {\Delta} {=} }& \int_{-\infty}^{\infty} p^*_{(\tau_0, \nu_0)}(t) \, x(t) \, dt\end{aligned}$$ where the coefficient $c_x(\tau_0, \nu_0)$ corresponding to the basis signal $p_{(\tau_0, \nu_0)}(t)$ is the value of the ZAK representation of $x(t)$ at $\tau = \tau_0$ and $\nu = \nu_0$, i.e. $$\begin{aligned}
\label{thmbasiseqn2}
c_x(\tau_0, \nu_0) & = & {\mathcal Z}_x(\tau_0, \nu_0).\end{aligned}$$
See Appendix \[prf\_thmbasis\].
An Orthonormal Basis for Time and Bandwidth Limited Signals Which are Localized in DD Domain {#secbw}
============================================================================================
In this section we consider TD signals which are approximately time-limited to the interval $[0 \,,\, NT)$ and band-limited to the interval $[0 \,,\, M \Delta f)$ where $M$ and $N$ are positive integers. We are specifically interested in those signals whose ZAK representation is localized in the DD domain, since such signals can be used to modulate and demodulate information symbols in the DD domain with little inter-symbol interference. These signals are also expected to be robust to channel induced Doppler shift, since from Result \[zlem52\] we know that the effect of Doppler shift is to only shift the signal along the Doppler domain.
Although the TD signals $p_{(\tau_0, \nu_0)}(t), 0 \leq \tau_0 < T, 0 \leq \nu_0 < \Delta f$ in Lemma \[plem1st\] are localized in the DD domain (see (\[lceqn1\])), they are neither time-limited nor band-limited. Therefore, we obtain another basis of approximately time and bandwidth limited signals, by approximately limiting the signals $p_{(\tau_0, \nu_0)}(t), 0 \leq \tau_0 < T, 0 \leq \nu_0 < \Delta f$ in the time and frequency domain. For this, we firstly multiply each basis signal $p_{(\tau_0, \nu_0)}(t)$ by a signal $q(t)$ which is approximately limited to the time-interval $[0 \,,\, NT)$, followed by convolution of the product signal $q(t) p_{(\tau_0, \nu_0)}(t)$ with another TD signal $s(t)$ which is approximately band-limited to the frequency domain interval $[ 0 \,,\, M \Delta f)$. These time- and band-limited signals are given by
[ $$\begin{aligned}
\label{basiseqn1}
\psi^{(q,s)}_{(\tau_0, \nu_0)}(t) & \hspace{-3mm} {\stackrel {\Delta} {=} }& \hspace{-3mm} \left( p_{(\tau_0, \nu_0)}(t) \, q(t) \right) \, \star s(t),\, 0 \leq \tau_0 < T , \, 0 \leq \nu_0 < \Delta f, \nonumber \\
q(t) & \approx & 0 \,,\, t \notin [0 \,,\, NT), \nonumber \\
\left\vert {\mathcal F}_s(f) \right\vert & \hspace{-3mm} = & \hspace{-3mm} \left\vert \int_{-\infty}^{\infty} \hspace{-2mm} s(t) e^{-j 2 \pi f t} \, dt \right\vert \, \approx \, 0 \,,\, f \notin [0 \,,\, M \Delta f),\end{aligned}$$ ]{} where $\star$ denotes the TD convolution operator. Although the signals $\psi^{(q,s)}_{(\tau_0, \nu_0)}(t), 0 \leq \tau_0 < T\,,\, 0 \leq \nu_0 < \Delta f,$ are approximately time and bandwidth limited, the degree of localization of these signals in the DD domain is not immediately obvious. Therefore, next we derive the ZAK representation of these signals. Subsequently, in this paper we consider the ideal time-limited waveform $$\begin{aligned}
\label{qteqn}
q(t) & {\stackrel {\Delta} {=} }\begin{cases}
1 &, \,\, 0 \leq t < NT \\
0 &, \,\, \mbox{\small{otherwise}}
\end{cases}\end{aligned}$$ and the ideal band-limited waveform $$\begin{aligned}
\label{steqn}
s(t) & \hspace{-3mm} = & \hspace{-3mm} \int_{0}^{M \Delta f} \hspace{-3mm} e^{j 2 \pi f t} df \, = \, e^{j \pi M \Delta f t} \, M \Delta f \, \mbox{\small{sinc}}(M \Delta f t), \nonumber \\
\mbox{\small{sinc}}(x) & {\stackrel {\Delta} {=} }& \frac{\sin(\pi x)}{\pi x}.\end{aligned}$$ Using (\[plemeqn2\]), (\[qteqn\]) and (\[steqn\]) in (\[basiseqn1\]), we get $$\begin{aligned}
\label{basisqeqn}
\psi^{(q,s)}_{(\tau_0, \nu_0)}(t) & = & \sqrt{T} \sum\limits_{n=0}^{N-1} e^{j 2 \pi \nu_0 n T} \, s(t - \tau_0 - nT)\end{aligned}$$ where $s(t)$ is given by (\[steqn\]). From this expression, it is clear that due to bandwidth limitation, the train of impulses in $p_{(\tau_0, \nu_0)}(t)$ appears as train of $\mbox{\small{sinc}}(\cdot)$ pulses in $\psi^{(q,s)}_{(\tau_0, \nu_0)}(t)$, each $\mbox{\small{sinc}}(\cdot)$ pulse having width roughly twice the inverse bandwidth. Further, due to time-limitation of $p_{(\tau_0, \nu_0)}(t)$, this train of $\mbox{\small{sinc}}(\cdot)$ pulses is restricted to the interval $[ 0 \,,\, NT)$. In Fig. \[fig3\], an illustration has been provided for $\frac{\sqrt{T}}{M} \left\vert \psi^{(q,s)}_{(\tau_0, \nu_0)}(t) \right\vert$ when $M = 12, N=14$, $\tau_0 = 0.6 T, \nu_0 = 0.2 \Delta f$.
The signal $p_{(\tau_0, \nu_0)}(t)$ is ideally localized in the DD domain. However, we expect that time and bandwidth limitation of $p_{(\tau_0, \nu_0)}(t)$, will affect its degree of localization in the DD domain. In order to understand this, in the following theorem we derive the expression for the ZAK representation of $\psi^{(q,s)}_{(\tau_0, \nu_0)}(t)$.
\[thm21\] The ZAK representation of $\psi^{(q,s)}_{(\tau_0, \nu_0)}(t)$ is given by $$\begin{aligned}
\label{eqn43}
{\mathcal Z}_{\psi, \tau_0, \nu_0}(\tau, \nu) & = & {\mathcal Z}_q\left( \tau_0 , \nu - \nu_0 \right) \, {\mathcal Z}_s\left( \tau - \tau_0, \nu \right)\end{aligned}$$ where $ {\mathcal Z}_q(\tau,\nu)$ and ${\mathcal Z}_s(\tau,\nu)$ are the ZAK representations of $q(t)$ and $s(t)$. These representations are given by
\[fig3\]
[ $$\begin{aligned}
\label{expreqn1}
\hspace{-2mm} {\mathcal Z}_q(\tau, \nu) & \hspace{-3mm} = & \hspace{-3mm} \sqrt{T} e^{j 2 \pi \nu \left\lfloor \frac{\tau}{T} \right\rfloor T } e^{-j \pi \nu (N-1) T} \, \frac{\sin\left( \pi \nu N T \right)}{\sin\left( \pi \nu T \right)}, \nonumber \\
\hspace{-2mm} {\mathcal Z}_s(\tau, \nu) & \hspace{-3mm} = & \hspace{-3mm} \frac{1}{\sqrt{T}} e^{j 2 \pi \nu \tau} e^{-j 2 \pi \left\lfloor \frac{\nu}{\Delta f} \right\rfloor \Delta f \tau} \, e^{j \pi (M-1) \Delta f \tau} \frac{\sin( \pi M \Delta f \tau)}{\sin(\pi \Delta f \tau)}. \nonumber \\\end{aligned}$$ ]{}
See Appendix \[prf\_thm21\].
Using the expressions for ${\mathcal Z}_q(\tau, \nu)$ and ${\mathcal Z}_s(\tau, \nu)$ from (\[expreqn1\]) into the R.H.S. of (\[eqn43\]), we get
[ $$\begin{aligned}
\hspace{-4mm} \left\vert {\mathcal Z}_{\psi, \tau_0, \nu_0}(\tau, \nu) \right\vert^2 & \hspace{-3mm} = & \hspace{-3mm} \frac{\sin^2\left( \pi (\nu - \nu_0) N T \right)}{\sin^2\left( \pi (\nu - \nu_0) T \right)} \, \frac{\sin^2( \pi M \Delta f (\tau - \tau_0))}{\sin^2(\pi \Delta f (\tau - \tau_0))}. \nonumber \\\end{aligned}$$ ]{}
\[fig4\]
From this expression it is clear that the ZAK representation of $\psi^{(q,s)}_{(\tau_0, \nu_0)}(t)$ has most of its energy localized around the point $(\tau_0, \nu_0)$ in the DD domain, in an interval of width $1/(M \Delta f)$ and $\Delta f/N$ respectively along the delay and Doppler domains. This is illustrated through Fig. \[fig4\] where we have plotted $\left\vert \frac{1}{MN} {\mathcal Z}_{\psi, \tau_0, \nu_0}(\tau, \nu) \right\vert^2$ in the DD domain for $M = 12, N=14$, $\tau_0 = 0.6T, \nu_0 = 0.2 \Delta f$. This implies that two DD domain signals ${\mathcal Z}_{\psi, \tau_1, \nu_1}(\tau, \nu)$ and ${\mathcal Z}_{\psi, \tau_2, \nu_2}(\tau, \nu)$ (localized at $(\tau_1, \nu_1)$ and $(\tau_2, \nu_2)$ respectively), will not interfere significantly, if the points $(\tau_1, \nu_1)$ and $(\tau_2, \nu_2)$ are [*separated*]{} by roughly $1/(M \Delta f)$ along the delay domain and by roughly $\Delta f/N$ along the Doppler domain (i.e., $\vert \tau_2 - \tau_1 \vert \approx 1/(M \Delta f)$ and $\vert \nu_2 - \nu_1 \vert \approx \Delta f/N$).
Hence, for transmission of information, information symbols can linearly modulate the DD domain signals ${\mathcal Z}_{\psi, \tau_0, \nu_0}(\tau, \nu)$. With sufficient separation in the DD domain, these DD domain signals will not interfere significantly and therefore the information symbols can be recovered back from the modulated signal. Since $0 \leq \tau_0 < T, 0 \leq \nu_0 < \Delta f$ and the required separation along the delay and Doppler domains is at least $1/(M \Delta f)$ and $\Delta f/N$ respectively, we consider DD domain signals for $\tau_0 = l/(M \Delta f) = lT/M, l=0,1,\dots, M-1$ and $\nu_0 = k \Delta f/N = k/(NT), k=0,1,\cdots, N-1$. From (\[steqn\]) and (\[basisqeqn\]), the $(k,l)$-th time-domain (TD) basis signal is then given by
[ $$\begin{aligned}
\label{eqn27}
\alpha_{(k,l)}(t) & {\stackrel {\Delta} {=} }& \frac{1}{\sqrt{MN}} \psi^{(q,s)}_{\left( \tau_0 = \frac{lT}{M}, \nu_0 = \frac{k}{NT} \right)}(t) \nonumber \\
& \hspace{-28mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-16mm} \sqrt{\frac{T}{MN}} \sum\limits_{n=0}^{N-1}e^{j 2 \pi \frac{ n k}{N} } \, \int_{0}^{M \Delta f} \hspace{-3mm} e^{j 2 \pi f \left( t - \frac{lT}{M} - nT \right)} df , \nonumber \\
& & k = 0,1,\cdots, N-1, \, l=0,1,\cdots, M-1,\end{aligned}$$ ]{} where step (a) follows from (\[basisqeqn\]) and the expression of $s(t)$ in (\[steqn\]). The factor $1/\sqrt{MN}$ ensures that these basis signals have unit energy. Subsequently we denote this basis by $\boldsymbol {\alpha} = \left\{ \alpha_{(k,l)}(t) \right\}_{k=0,\cdots, N-1, l=0,1,\cdots,M-1}$. The following theorem shows that $\boldsymbol {\alpha}$ is an $MN$-dimensional orthonormal basis.
\[lem18\] The basis $\boldsymbol {\alpha}$ is an $MN$-dimensional orthonormal basis, i.e. $$\begin{aligned}
\label{lem18eqn}
\int_{-\infty}^{\infty} \hspace{-2mm} \alpha_{(k_1,l_1)}(t) \, \alpha^*_{(k_2,l_2)}(t) \, dt & = & \delta[k_1 - k_2] \, \delta[l_1 - l_2], \nonumber \\
& & \hspace{-45mm} k_1,k_2 = 0,1,\cdots, N-1 \,,\, l_1,l_2 = 0,1,\cdots, M-1.\end{aligned}$$
See Appendix \[prf\_lem18\].
For a given $(M, N, T)$, we have therefore derived an $MN$-dimensional orthonormal basis $\boldsymbol {\alpha}$ of TD signals which are localized in the DD domain, and which are approximately time-limited to the interval $[0 \,,\, NT)$ and band-limited to the interval $[0 \,,\, M \Delta f)$. The dimensionality of the space of signals which are time-limited to $NT$ seconds and band-limited to $M \Delta f$ Hz but not necessarily localized in the DD domain, is known to be the time-bandwidth product $M \Delta f \times N T = MN$. Therefore, the additional constraint of localization in the DD domain does not reduce the dimensionality of the space of time and bandwidth limited signals.
Delay-Doppler (DD) Domain Modulation {#propmodulation}
====================================
From the discussion in the previous section, it is natural to consider DD domain modulation where the complex information symbols $x[k,l]$ linearly modulate the basis signals $\alpha_{(k,l)}(t), k=0,1,\cdots, N-1, l=0,1,\cdots, M-1$. The TD transmit signal is therefore given by
[ $$\begin{aligned}
\label{eqn44}
x(t) & \hspace{-3mm} = & \hspace{-3mm} \sum\limits_{k=0}^{N-1}\sum\limits_{l=0}^{M-1} x[k,l] \, \alpha_{(k,l)}(t) \, {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}\, \sqrt{\frac{T}{MN}} \sum\limits_{n=0}^{N-1} x_n(t -nT) , \nonumber \\
x_n(t) & \hspace{-3mm} {\stackrel {\Delta} {=} }& \hspace{-3mm} \sum\limits_{k=0}^{N-1} \sum\limits_{l=0}^{M-1} x[k,l] \, e^{j 2 \pi \frac{n k}{N}} \int_{0}^{M \Delta f} \hspace{-3mm} e^{j 2 \pi f \left( t - \frac{lT}{M} \right)} \, df \nonumber \\
& \hspace{-3mm} = & \hspace{-3mm} \int_{-\infty}^{\infty} \hspace{-2mm} X_n(f) \, e^{j 2 \pi f t} \, df \,\,,\,\, n=0,1,\cdots, N-1, \nonumber \\
X_n(f) & \hspace{-3mm} {\stackrel {\Delta} {=} }& \begin{cases}
\sum\limits_{k=0}^{N-1} \sum\limits_{l=0}^{M-1} x[k,l] \, e^{j 2 \pi \left( \frac{n k}{N} - \frac{f l T}{M} \right)} , \, 0 \leq f < M \Delta f \\
0 \, \hspace{38mm} , \, \mbox{\small{otherwise}}
\end{cases}, \nonumber \\
& & n=0,1,\cdots, N-1\end{aligned}$$ ]{} where in step (a) we have used the expression of $\alpha_{(k,l)}(t)$ from (\[eqn27\]). The expression of $x(t)$ in (\[eqn44\]) suggests a two stage modulation method, where in the first stage the DD domain information symbols $x[k,l], k=0,1,\cdots, N-1, l=0,1,\cdots, M-1$ are transformed to the signals $X_n(f), n=0,1,\cdots, N-1$. In the second stage, for each $n=0,1,\cdots, N-1,$ $X_n(f)$ is then transformed to the TD signal $x_n(t)$ which is time-shifted by $nT$. These time-shifted signals are then added to get $x(t)$.
Derivation of OTFS Modulation {#deriveOTFS}
=============================
Using the DD domain modulation derived in Section \[propmodulation\], we derive the OTFS modulation equations in this section. To the best of our knowledge, all prior work on OTFS modulation/demodulation have only reported the equations for transforming the DD domain information symbols to the TD transmit signal, and have not derived the modulation equations from first principles. An important contribution of our current paper lies in the novel derivation of OTFS modulation equations from first principles.
The following observation is useful in the derivation of OTFS modulation from (\[eqn44\]). In (\[eqn44\]) we note that for any $(k,l), k=0,1,\cdots, N-1, l=0,1,\cdots, M-1$, the $(k,l)$-th term in the expression for $x_n(t)$ i.e., $x_{n,k,l}(t) {\stackrel {\Delta} {=} }x[k,l] e^{j 2 \pi \frac{n k}{N}} \int_{0}^{M \Delta f} \hspace{-2mm} e^{j 2 \pi f \left( t - \frac{lT}{M} \right)} \, df = M \Delta f \, x[k,l] \, e^{j 2 \pi \frac{n k}{N}} e^{j \pi M \Delta f \left( t - \frac{lT}{M} \right)} \mbox{\small{sinc}}\left(M \Delta f \left( t - \frac{lT}{M} \right) \right)$, has most of its energy in a time-interval of the type $\left[ \frac{(l- \zeta)T}{M} \,,\, \frac{(l+\zeta)T}{M} \right]$, $\zeta \geq 1$. This is because $\left\vert x_{n,k,l}(t) \right\vert^2 = (M \Delta f)^2 \vert x[k,l] \vert^2 \mbox{\small{sinc}}^2\left(M \Delta f \left( t - \frac{lT}{M} \right) \right)$ and
[ $$\begin{aligned}
\label{otfseqn87}
\hspace{-1mm} \frac{\int_{\frac{(l- \zeta)T}{M}}^{\frac{(l+ \zeta)T}{M}} \left\vert x_{n,k,l}(t) \right\vert^2 dt}{\int_{-\infty}^{\infty} \left\vert x_{n,k,l}(t) \right\vert^2 dt} & \hspace{-3mm} = & \hspace{-3mm}
\frac{\int_{\frac{(l- \zeta)T}{M}}^{\frac{(l+ \zeta)T}{M}} (M \Delta f )^2 \mbox{\small{sinc}}^2\hspace{-1mm}\left(M \Delta f \left( t - \frac{lT}{M} \right) \right) dt}{ {\int_{-\infty}^{\infty} (M \Delta f )^2 \mbox{\small{sinc}}^2\hspace{-1mm}\left(M \Delta f \left( t - \frac{lT}{M} \right) \right) dt}} \nonumber \\
& \hspace{-3mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-3mm} \int_{-\zeta}^{\zeta} \hspace{-2mm} \mbox{\small{sinc}}^2(t') dt' \end{aligned}$$ ]{} where step (a) follows from the change of integration variable to $t' {\stackrel {\Delta} {=} }M \Delta f \left(t - \frac{lT}{M} \right)$. Assuming integer $\zeta \geq 1$ and $M \geq 2 \zeta$, each term $x_{n,k,l}(t)$ corresponding to $l=
\zeta,\zeta + 1,\cdots, M- \zeta,$ has at least $\int_{-\zeta}^{\zeta} \mbox{\small{sinc}}^2(t') dt' $ fraction of its energy in the time-interval $[ 0 \,,\, T)$. This follows from (\[otfseqn87\]) and the fact that for each $l = \zeta,\zeta + 1,\cdots, M- \zeta,$ $\left[ \frac{(l - \zeta) T}{M} \,,\, \frac{(l+ \zeta)T}{M} \right) \subset [0 \,,\, T)$. Next, for each $i=1,2,\cdots, \zeta -1$, the terms $x_{n,k,l}(t)$ corresponding to $l=i, M-i,$ have at least $\int_{-i}^{i} \mbox{\small{sinc}}^2(t') dt'$ fraction of their energy inside $[ 0 \,,\, T)$ (since for $l=i, M-i$, $\left[ \frac{(l-i)T}{M} \,,\, \frac{(l+i)T}{M} \right) \, \subset \, [0 \,,\, T)$). However, the terms $x_{n,k,l}(t)$ corresponding to $l=0$ have at least half of their energy outside the interval $[ 0 \,,\, T)$. Since $x_n(t) = \sum\limits_{k=0}^{N-1} \sum\limits_{l=0}^{M-1} x_{n,k,l}(t)$, out of all the terms $x_{n,k,l}(t)$ corresponding to $l=0,1,\cdots, M-1$, a fraction $(M- 2 \zeta + 1)/M$ correspond to $l=\zeta,\zeta + 1,\cdots, M- \zeta,$ and a fraction $2/M$ corresponds to $l=i, M-i$ for each $i=1,2,\cdots, \zeta -1$. Therefore, for $M \geq 2 \zeta$, at least a fraction $\gamma(\zeta, M) {\stackrel {\Delta} {=} }\left( \frac{M + 1 - 2\zeta}{M} \int_{-\zeta}^{\zeta} \mbox{\small{sinc}}^2(t') dt' \, + \, \frac{2}{M} \sum\limits_{i=1}^{\zeta - 1} \int_{-i}^{i} \mbox{\small{sinc}}^2(t') dt' \right)$ of the total energy of $x_n(t)$ lies in the interval $[0 \,,\, T)$. From the expression of this fraction $\gamma(\zeta, M)$ it is clear that for a fixed $\zeta \geq 1$, $\lim_{M \rightarrow \infty} \gamma(\zeta, M) = \int_{-\zeta}^{\zeta} \mbox{\small{sinc}}^2(t') dt' $. Since $\lim_{\zeta \rightarrow \infty} \int_{-\zeta}^{\zeta} \mbox{\small{sinc}}^2(t') dt' = \int_{-\infty}^{\infty} \mbox{\small{sinc}}^2(t') dt' = 1$, it follows that with a sufficiently large $M$, $x_n(t)$ has almost all of its energy in the time-interval $[0 \,,\, T)$.
The following theorem uses this observation to derive OTFS modulation equations from the DD domain modulation equations in (\[eqn44\]). Also, since $x_n(t), n=0,1,\cdots, N-1,$ are approximately time-limited to $[0 \,,\, T)$, it follows that the $n$-th term in step (a) of (\[eqn44\]) i.e., $x_n(t - nT)$ is approximately time-limited to the interval $[ nT \,,\, (n+1)T)$. As $X_n(f)$ is the Fourier transform of $x_n(t)$, we subsequently refer to $X_n(f)$ as the $n$-th time-frequency (TF) signal.
(Derivation of OTFS modulation) \[thmotfsmod\] For sufficiently large $M$, the DD domain modulated signal $x(t)$ in (\[eqn44\]) is given by
[ $$\begin{aligned}
\label{otfsthmeqn}
\hspace{-8mm} x(t) & \hspace{-4mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{\approx}}}& \hspace{-4mm} \frac{1}{\sqrt{MN}} \sum\limits_{n=0}^{N-1} \sum\limits_{m=0}^{M-1} \hspace{-1mm} g(t - nT) \, X_{\mbox{\tiny{TF}}}[n,m] \, e^{j 2 \pi m \Delta f (t - nT)}, \nonumber \\
X_{\mbox{\tiny{TF}}}[n,m] & \hspace{-2mm} {\stackrel {\Delta} {=} }& \hspace{-2mm} \sum\limits_{k=0}^{N-1}\sum\limits_{l=0}^{M-1} x[k,l] \, e^{j 2 \pi \left( \frac{n k}{N} - \frac{m l}{M} \right)}, \nonumber \\
& & \hspace{-1mm} n=0,1,\cdots, N-1, m=0,1,\cdots, M-1, \nonumber \\
g(t) & \hspace{-2mm} {\stackrel {\Delta} {=} }& \hspace{-2mm} \begin{cases}
\frac{1}{\sqrt{T}} \,,\, t \in [0 \,,\, T) \\
0 \,\,\,,\, \mbox{\small{otherwise}}
\end{cases}.\end{aligned}$$ ]{} The R.H.S. in step (a) is [*exactly the equation for OTFS modulation*]{} in [@HadaniOTFS1; @HadaniOTFS2; @HadaniOTFS3], with a rectangular transmit pulse $g(t)$.
Let us consider the product of $X_n(f)$ with the frequency domain pulse train $\Delta f \hspace{-2mm} \sum\limits_{m=-\infty}^{\infty} \hspace{-2mm} \delta(f - m \Delta f)$, i.e. $$\begin{aligned}
\label{otfsderiveqn7}
{\Tilde X_n}(f) & {\stackrel {\Delta} {=} }& X_n(f) \, \left[ \Delta f \hspace{-2mm} \sum\limits_{m=-\infty}^{\infty} \hspace{-2mm} \delta(f - m \Delta f) \right].\end{aligned}$$ Since the inverse Fourier transform of $\Delta f \sum\limits_{m=-\infty}^{\infty} \delta(f - m \Delta f)$ is $ \sum\limits_{k=-\infty}^{\infty} \delta(t - k T)$, the inverse Fourier transform of ${\Tilde X_n}(f)$ (which we denote by ${\Tilde x_n}(t)$) is given by $$\begin{aligned}
\label{otfsderiveqn5}
{\Tilde x_n}(t) & \hspace{-3mm} = & \hspace{-3mm} x_n(t) \star \sum\limits_{k=-\infty}^{\infty} \hspace{-2mm} \delta(t - k T) = \sum\limits_{k=-\infty}^{\infty} \hspace{-2mm} x_n(t - kT) \end{aligned}$$ where $\star$ denotes TD convolution. Note that ${\Tilde x_n}(t)$ is a periodic TD signal with period $T$. As we have already observed, for a sufficiently large $M$, $x_n(t)$ is approximately time-limited to the interval $[0 \,,\, T)$. Therefore, from (\[otfsderiveqn5\]) it follows that for $0 \leq t < T$ we have $$\begin{aligned}
{\Tilde x_n}(t) & \approx & x_n(t) \,\,,\,\, 0 \leq t < T.\end{aligned}$$ In other words $$\begin{aligned}
\label{otfsderiveqn11}
x_n(t) & \approx & \sqrt{T} \, g(t) \, {\Tilde x_n}(t), \end{aligned}$$ where $g(t)$ is the rectangular TD signal defined in (\[otfsthmeqn\]). Taking the inverse Fourier transform of ${\Tilde X_n}(f)$, from (\[otfsderiveqn7\]) we get $$\begin{aligned}
\label{otfsderiveqn8}
{\Tilde x_n}(t) & \hspace{-3mm} = & \hspace{-3mm} \int_{-\infty}^{\infty} \hspace{-2mm} X_n(f) \, \Delta f \hspace{-2mm} \sum\limits_{m=-\infty}^{\infty} \hspace{-2mm} \delta(f - m \Delta f) \, e^{j 2 \pi ft} \, df \nonumber \\
& \hspace{-3mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-3mm} \int_{0}^{M \Delta f} \hspace{-2mm} X_n(f) \, \Delta f \hspace{-2mm} \sum\limits_{m=-\infty}^{\infty} \hspace{-2mm} \delta(f - m \Delta f) \, e^{j 2 \pi ft} \, df \nonumber \\
& = & \Delta f \sum\limits_{m=0}^{M-1} X_n(m \Delta f) \, e^{j 2 \pi m \Delta f t}\end{aligned}$$ where step (a) follows from the fact that $X_n(f)$ is limited to the frequency domain interval $[0 \,,\, M \Delta f)$ (see (\[eqn44\])). From the expression of $X_n(f)$ in (\[eqn44\]) (in terms of the information symbols $x[k,l]$), for $m=0,1,\cdots, M-1$ we get $$\begin{aligned}
\label{xnfmdfeqn}
X_n(m \Delta f) & = & \sum\limits_{k=0}^{N-1} \sum\limits_{l=0}^{M-1} x[k,l] \, e^{j 2 \pi \left( \frac{n k}{N} - \frac{(m \Delta f) l T}{M} \right)} \nonumber \\
& {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \sum\limits_{k=0}^{N-1} \sum\limits_{l=0}^{M-1} x[k,l] \, e^{j 2 \pi \left( \frac{n k}{N} - \frac{m l}{M} \right)} \nonumber \\
& {\mathrel{\overset{\makebox[0pt]{{\tiny(b)}}}{=}}}& X_{\mbox{\tiny{TF}}}[n,m],\end{aligned}$$ where step (a) follows from the fact that $T \Delta f = 1$ and step (b) follows from the definition of $X_{\mbox{\tiny{TF}}}[n,m]$ in (\[otfsthmeqn\]). Using (\[xnfmdfeqn\]) in (\[otfsderiveqn8\]) we get $$\begin{aligned}
\label{otfs11eqn2}
{\Tilde x_n}(t) = \Delta f \sum\limits_{m=0}^{M-1} X_{\mbox{\tiny{TF}}}[n,m] \, e^{j 2 \pi m \Delta f t}.\end{aligned}$$ Using (\[otfs11eqn2\]) in (\[otfsderiveqn11\]) we get $$\begin{aligned}
\label{apintgeqn}
x_n(t) & \approx & \frac{1}{\sqrt{T}} \, g(t) \, \sum\limits_{m=0}^{M-1} X_{\mbox{\tiny{TF}}}[n,m] \, e^{j 2 \pi m \Delta f t}.\end{aligned}$$ Using (\[apintgeqn\]) in (\[eqn44\]) we get $$\begin{aligned}
x(t) & \hspace{-3mm} = & \hspace{-3mm} \sqrt{\frac{T}{MN}} \sum\limits_{n=0}^{N-1} x_n(t -nT) \nonumber \\
& & \hspace{-10mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{\approx}}}\frac{1}{\sqrt{MN}} \sum\limits_{n=0}^{N-1} \sum\limits_{m=0}^{M-1} \hspace{-1mm} g(t - nT) \, X_{\mbox{\tiny{TF}}}[n,m] \, e^{j 2 \pi m \Delta f (t - nT)}, \nonumber \\\end{aligned}$$ where we have used (\[apintgeqn\]) in step (a). This completes the proof.
From (\[eqn44\]) we know that the TD signal $x_n(t) = \int_{-\infty}^{\infty} X_n(f) e^{j 2 \pi f t} df$ is the inverse Fourier transform of $X_n(f)$. However, in practical systems it is difficult to exactly implement the integral in this inverse Fourier transform. On the other hand, from (\[otfs11eqn2\]) it is clear that ${\Tilde x}_n(t), 0 \leq t < T$ can be easily computed by an OFDM modulator in existing 4G/5G modems with a sub-carrier spacing of $\Delta f$ and $M$ sub-carriers. For sufficiently large $M$, $x_n(t)$ is approximately time-limited to $[0 \,,\, T)$ and is approximately equal to ${\Tilde x}_n(t)$ in this interval (see (\[otfsderiveqn11\])). Therefore, for sufficiently large $M$, the DD domain modulation in (\[eqn44\]) is same as OTFS modulation, which can be practically implemented using the OFDM modulator in existing 4G/5G modems.
For sufficiently large $M$, we therefore expect the spectral efficiency (SE) performance of OTFS modulation to be same as that of the DD domain modulation in (\[eqn44\]). In the next section, we derive an expression for the SE achieved by the DD domain modulation in (\[eqn44\]). Numerical simulations in Section \[secsim\] reveal that, indeed the SE achieved by OTFS modulation is same as the SE achieved by the DD domain modulation in (\[eqn44\]).
Spectral Efficiency of DD Domain Modulation {#Zakrecv}
===========================================
In this section, we derive an expression for the spectral efficiency (SE) achieved by the DD domain modulation in (\[eqn44\]). For the derivation of the SE expression we consider a ZAK receiver [@Saif2020]. With $x(t)$ as the time-domain (TD) transmit signal, the received TD signal is given by [@Bello] $$\begin{aligned}
y(t) & = & \sum\limits_{i=1}^L h_i \, x(t - \tau_i) \, e^{j 2 \pi \nu_i (t - \tau_i)} \, + \, n(t)\end{aligned}$$ where $h_i, \tau_i, \nu_i, i=1,2,\cdots,L$ are the channel gain, delay and Doppler shift of the $i$-th channel path between the transmitter and the receiver. Also, $n(t)$ is the additive white Gaussian noise (AWGN) at the receiver. Further, we consider $0 < \tau_i < T, \, i=1,2,\cdots, L$. Since, the transmit TD signal $x(t)$ has most of its energy in the time-interval $[0 \,,\, NT)$ and $0 < \tau_i < T, i=1,2,\cdots, L$, the received signal $y(t)$ has most of its energy in the time-interval $[0 \,,\, (N+1)T)$.
In the ZAK receiver [@Saif2020], the ZAK representation of the received TD signal i.e., ${\mathcal Z}_y(\tau, \nu)$ is sampled at the discrete points $\left(\tau = \frac{l'T}{M}, \nu = \frac{k' \Delta f}{N} \right)$ in the DD domain. This sampled received DD domain signal is denoted by $Y[k',l'] {\stackrel {\Delta} {=} }{\mathcal Z}_y\left(\tau = \frac{l'T}{M}, \nu = \frac{k' \Delta f}{N} \right), k'=0,1,\cdots, N-1, l'=0,1,\cdots, M-1$. The DD domain information symbols $x[k,l], k=0,1,\cdots, N-1, l=0,1,\cdots, M-1$ are decoded from this sampled received DD domain signal. The expression for $Y[k',l']$ in terms of the information symbols $x[k,l]$ is given by the following theorem.
\[thm73\] The sampled received DD domain signal $Y[k',l']$ is given by (\[eqn67\]) (see top of next page), where ${\mathcal Z}_n\left(\tau, \nu \right)$ is the ZAK representation of AWGN $n(t)$.
See Appendix \[prfthm73\].
$$\begin{aligned}
\hline\end{aligned}$$
The following theorem gives the expression for the SE achieved by the DD domain modulation in (\[eqn44\]), with a ZAK receiver.
\[thm8\] Let $\rho$ denote the ratio of the average transmit power to the received noise power (in the communication bandwidth $M \Delta f$). With i.i.d. complex Gaussian distributed information symbols, the SE achieved by the DD domain modulation in (\[eqn44\]), with a ZAK receiver, is given by $$\begin{aligned}
\label{Ceqn}
C & {\stackrel {\Delta} {=} }& \frac{1}{MN} \, \log_2 \left\vert {\bf I} \, + \, \rho \, {\Tilde {\bf H}}^H {\Tilde {\bf K}}^{-1} {\Tilde {\bf H}} \right\vert,\end{aligned}$$ where ${\bf I}$ denotes the $MN \times MN$ identity matrix and the elements of ${\Tilde {\bf H}}, {\Tilde {\bf K}} \in {\mathbb C}^{MN \times MN}$ are given by
[ $$\begin{aligned}
{\Tilde H}[k'M + l'+1, kM+l +1] & \hspace{-3mm} {\stackrel {\Delta} {=} }\hspace{-3mm} & {\Tilde h}[k',l',k,l], \,\,\,\,\, (\mbox{\small{see (\ref{eqn67})}}) \nonumber \\
{\Tilde K}[{k'M + l' +1,kM + l +1}] & \hspace{-3mm} = & \hspace{-3mm} \begin{cases}
\left(1 + \frac{1}{N} \right) , k' = k \,,\, l' = l \\
\left( \frac{1}{N} \right) \hspace{6mm} , k' \ne k \,,\, l' = l \\
0 \hspace{11mm} \,, l' \ne l \\
\end{cases}, \nonumber \\
& & \hspace{-32mm} k',k = 0,1,\cdots, N-1 \,,\, l',l = 0,1,\cdots, M-1,\end{aligned}$$ ]{} where ${\Tilde h}[k',l',k,l]$ is given by (\[eqn67\]).
See Appendix \[prfthm8\].
Why is DD Domain Modulation Better Than OFDM ? {#secbetter}
==============================================
In this section, for the DD domain modulation in (\[eqn44\]) and also for OFDM, we study the impact of channel induced Doppler shift on inter-symbol interference. Let us consider a single path channel with delay $\tau'$, Doppler shift $\nu'$ and channel path gain $h'$. Since we want to study inter-symbol interference, let us consider the channel to be noise-free. From (\[eqn67\]), the received DD domain samples in the ZAK receiver are then given by
[ $$\begin{aligned}
\label{eqn91}
Y[k',l'] & = & \sqrt{MN} h' \sum\limits_{k=0}^{N-1} \sum\limits_{l=0}^{M-1} x[k,l] \, R_{(k,l)}[k',l'], \nonumber \\
R_{(k,l)}[k',l'] & {\stackrel {\Delta} {=} }& \frac{{\Tilde h}[k',l',k,l]}{h'}, \nonumber \\
& & \hspace{-6mm} k'=0,1,\cdots, N-1, \, l' = 0,1,\cdots, M-1.\end{aligned}$$ ]{} From (\[eqn91\]) it is clear that the information symbol $x[k,l]$ is received in the $(k',l')$-th DD domain sample $Y[k',l']$ through the coefficient $R_{(k,l)}[k',l']$. From (\[eqn67\]) it follows that
[ $$\begin{aligned}
\label{eqn75}
\left\vert R_{(k,l)}[k',l'] \right\vert^2 & = & {\Bigg [} \, \frac{\sin^2\left( \pi \left(k' - k - \frac{\nu'}{\Delta f/N} \right) \right)}{\sin^2\left( \frac{\pi}{N} \left(k' - k - \frac{\nu'}{\Delta f/N} \right) \right)} \nonumber \\
& & \,\,\,\, \frac{\sin^2\left( \pi \left(l' - l - \tau' M \Delta f \right) \right)}{\sin^2\left( \frac{\pi}{M} \left(l' - l - \tau' M \Delta f \right) \right)} {\Bigg ]}.\end{aligned}$$ ]{} From (\[eqn75\]) it is clear that if the path delay $\tau'$ is an integer multiple of $1/(M \Delta f) = T/M$ and if the Doppler shift $\nu'$ is an integer multiple of $\Delta f/N = 1/(NT)$, then $\left\vert R_{(k,l)}[k',l'] \right\vert^2 = 0$ for all $k' \ne \left[ k + \frac{\nu'}{\Delta f/N} \right]_{_N} $ and $l ' \ne \left[ l + \tau'M \Delta f \right]_{_M} $, i.e., $x[k,l]$ is only received in $Y\left[k' = \left[ k + \nu'NT \right]_{_N}, l' = \left[ l + \tau'M \Delta f \right]_{_M} \right]$ which is given by
[ $$\begin{aligned}
\label{eqn83}
Y\left[k' = \left[ k + \nu'NT \right]_{_N}, l' = \left[ l + \tau'M \Delta f \right]_{_M} \right] & & \nonumber \\
& & \hspace{-70mm} = \sqrt{MN} h' x[k,l] R_{(k,l)} \hspace{-1mm} \left[k' = \left[ k + \nu'NT \right]_{_N} \hspace{-1.5mm} , l' = \left[ l + \tau'M \Delta f \right]_{_M} \right], \nonumber \\\end{aligned}$$ ]{} i.e., there is no inter-symbol interference. However, when the delay and Doppler shift are not integer multiples of $1/(M \Delta f)$ and $\Delta f/N$, $\left\vert R_{(k,l)}[k',l'] \right\vert^2$ is not zero for all $(k',l') \ne \left( \left\lfloor \left[ k + \nu'NT \right]_{_N} \right\rfloor, \left\lfloor \left[ l + \tau'M \Delta f \right]_{_M} \right\rfloor \right)$, i.e., the energy of the DD domain basis signal carrying $x[k,l]$ leaks into other DD domain basis signals thereby creating inter-symbol interference. In the following we study the fraction of information symbols which can get significantly interfered by an information symbol.
\[fig5\]
Firstly, for a given $(k,l)$, from (\[eqn75\]) we notice that $\left\vert R_{(k,l)}[k',l'] \right\vert^2, k'=0,1,\cdots, N-1 ,l'=0,1,\cdots,M-1$ is a product of two terms, one which depends on the difference $(k' - k)$ along the Doppler domain and the other term which depends on the difference $(l' - l)$ along the delay domain. To understand the spread/leakage of energy along the Doppler domain, we consider the function $f_{_{\nu',N}}(u) {\stackrel {\Delta} {=} }\frac{1}{N^2} \frac{\sin^2\left( \pi \left(u - \frac{\nu'}{\Delta f/N} \right) \right)}{\sin^2\left( \frac{\pi}{N} \left(u - \frac{\nu'}{\Delta f/N} \right) \right)}$. Note that the term in $\left\vert R_{(k,l)}[k',l'] \right\vert^2$ which depends on $(k' - k)$ is given by $N^2 f_{_{\nu',N}}(u = k' - k)$. The function $f_{_{\nu',N}}(u)$ is periodic with period $N$ and has a peak at $u = \frac{\nu'}{\Delta f/N}$ (see Fig. \[fig5\]). From Fig. \[fig5\] and the expression for $f_{_{\nu',N}}(u)$ it is clear that the main lobe of $f_{_{\nu',N}}(u)$ is between $u = \frac{\nu'}{\Delta f/N} - 1$ and $u = \frac{\nu'}{\Delta f/N} +1$ (at both these values of $u$, $f_{_{\nu',N}}(u)$ is zero). Hence, the main lobe width of $f_{_{\nu',N}}(u)$ is two, which is independent of both $N$ and $\nu'$. Therefore, along the Doppler domain, for a given $k$, $\left\vert R_{(k,l)}[k',l'] \right\vert^2$ is mostly localized at $k' = \left\lfloor \left[ k + s + \frac{\nu'}{\Delta f/N} \right]_{_N} \right\rfloor, s=0,1$. Similarly, along the delay domain, for a given $l$, $\left\vert R_{(k,l)}[k',l'] \right\vert^2$ is mostly localized at $l'= \left\lfloor \left[ l + s + \tau'M \Delta f \right]_{_M} \right\rfloor, s=0,1$. Hence, it follows that irrespective of $(\tau', \nu')$, in the DD domain, $\left\vert R_{(k,l)}[k',l'] \right\vert^2, k'=0,1,\cdots, N-1 ,l'=0,1,\cdots,M-1$ is mostly localized at $(k',l') \in {\mathcal A}_{(k,l)}$, where
[ $$\begin{aligned}
{\mathcal A}_{(k,l)} & \hspace{-3mm} {\stackrel {\Delta} {=} }& \hspace{-3mm} {\Bigg \{} (k' , l')\, {\Bigg \vert } (k',l') \in {\mathcal S} \,\, \mbox{\small{and}}, k' = \left\lfloor \left[ k + s_1 + \frac{\nu'}{\Delta f/N} \right]_{_N} \right\rfloor \nonumber \\
& & \hspace{-1mm} \mbox{or} \,\, l'= \left\lfloor \left[ l + s_2 + \tau'M \Delta f \right]_{_M} \right\rfloor, \,\, s_1, s_2 = 0,1 {\Bigg \}}, \nonumber \\
& & \hspace{-12mm} {\mathcal S} {\stackrel {\Delta} {=} }\{ (k',l') \, | \, k' = 0,1,\cdots,N-1, l'=0,1,\cdots, M-1 \}.\end{aligned}$$ ]{} Therefore, the fraction of interfered symbols (i.e., fraction of the remaining $MN -1 $ information symbols which receive significant interference from $x[k,l]$) is roughly $\frac{\left\vert {\mathcal A}_{(k,l)} \right\vert - 1}{MN - 1} = \frac{(2M + 2N -5)}{MN -1}$. To be more precise, for a given $(k,l)$ and given $(\tau',\nu')$, let ${\mathcal B}_{(k,l)}$ denote the smallest cardinality set of $(k',l')$ pairs, such that the sum of $\left\vert R_{(k,l)}[k',l'] \right\vert^2$ for all $(k',l') \in {\mathcal B}_{(k,l)}$ is atleast $99 \%$ (i.e., $0.99$) of the total energy $\sum\limits_{k'=0}^{N-1} \sum\limits_{l'=0}^{M-1} \left\vert R_{(k,l)}[k',l'] \right\vert^2$, i.e.,
[ $$\begin{aligned}
{\mathcal B}_{(k,l)} & {\stackrel {\Delta} {=} }& \arg \hspace{-13mm} \min_{{\mathcal D} \subseteq {\mathcal S} \, {\Bigg \vert} \, \frac{\sum\limits_{(k',l') \in {\mathcal D} } \left\vert R_{(k,l)}[k',l'] \right\vert^2}{\sum\limits_{(k',l') \in {\mathcal S}} \left\vert R_{(k,l)}[k',l'] \right\vert^2 } \geq 0.99} \hspace{-4mm} \left\vert {\mathcal D} \right\vert.\end{aligned}$$ ]{}
\[fig6\]
In Fig. \[fig6\] we plot the fraction of interfered information symbols i.e. $\frac{\left(\left\vert {\mathcal B}_{(k,l)} \right\vert - 1 \right)}{MN - 1}$ as a function of increasing $\nu'/\Delta f$, for $N=23,46, 92$ and $M=45, l = 23$. For each $\nu'/\Delta f$, the fraction of interfered symbols is averaged with respect to $\tau' M \Delta f$ which is uniformly distributed in the interval $[ 0 \,,\, 0.5]$. Further, $k=11,23, 46$, respectively for $N=23, 46, 92$. We observe that, for a given $N$ the fraction of interfered symbols is upper bounded for all values of $\nu'/\Delta f$. Further, this upper bound decreases with increasing $N$. For $N=23$, this upper bound is roughly $11.6 \%$, which decreases to $7.6 \%$ and $5.1 \%$ respectively for $N=46$ and $N=92$. These values are close to our rough estimate of $\frac{(2M + 2N -5)}{MN -1}$, which is $12.67 \%$, $8.55 \%$ and $6.5 \%$ respectively for $N=23, 46, 92$.
Next we compare the fraction of interfered information symbols in DD domain modulation with that in CP-OFDM (Cyclic-Prefix OFDM with sub-carrier spacing $\Delta f = 1/T$ and bandwidth $M \Delta f$). In CP-OFDM it is well known that in the presence of channel induced Doppler shift, energy transmitted on a sub-carrier leaks into adjacent sub-carriers. The fraction of interfered information symbols in CP-OFDM is presented in Appendix \[cpofdm\].
\[fig7\]
In Fig. \[fig7\], the fraction of interfered symbols for both CP-OFDM and DD domain modulation is plotted. For CP-OFDM we have $M = 45$ and for DD domain modulation we have $M = 45, N = 46$, i.e., the bandwidth is $M \Delta f$ for both CP-OFDM and DD domain modulation. The time duration of a CP-OFDM symbol is $1/\Delta f = T$, during which $M$ information symbols are transmitted, while the time duration of a DD domain modulated signal is $NT$ during which $MN$ DD domain information symbols are transmitted. In Fig. \[fig7\], energy is transmitted on the $k=23$-rd sub-carrier in CP-OFDM, while in DD domain modulation, energy is transmitted on the $(k=23, l=23)$-th DD domain basis signal. From the figure it is clear that the fraction of interfered symbols can be very large in CP-OFDM when compared to that in DD domain modulation. For example, the maximum fraction of interfered symbols is $7.6 \%$ for DD domain modulation whereas it is about $48 \%$ for CP-OFDM. Due to the significantly lower fraction of interfered symbols in DD domain modulation, it is practically feasible to perform joint DD domain equalization of all $MN$ information symbols at the receiver.
Although a large $N$ helps in localizing the DD domain energy distribution of delay and Doppler shifted basis signals (which enables joint DD domain equalization), it also increases the duration of the modulated signal which is $NT$, i.e., $N$ times larger than the duration of an OFDM symbol. In other words, [*DD domain modulation achieves robustness to channel induced Doppler shift at the cost of increased latency*]{}.
Numerical Simulation {#secsim}
====================
In this section, we compare the spectral efficiency (SE) performance of the DD domain modulation in (\[eqn44\]) and OTFS modulation. We specifically consider communication of control and non-payload information between an Unmanned Aircraft System (UAS) and a Ground Station (GS) [@ICTCPaper]. A widely accepted model for the en-route scenario is the two-path model, with a direct and a reflected path. The gain for the direct path is $h_1 = \sqrt{K_f/(K_f+1)}$, while that for the reflected path is $h_2 \sim {\mathcal C}{\mathcal N}(0, 1/(K_f+1))$ [@Haas]. In [@Haas] it is mentioned that $K_f$ is typically $15$ dB. The delay between these two paths is $33 \, \mu s$ (i.e., $\tau_1=0, \tau_2 = 33 \, \mu s$). For an aircraft speed of $v$, the Doppler shift for the first path is $\nu_1 = v f_c/c$ and that for the second path is $\nu_2 = (v f_c/c) \cos\left(\pi - \theta \, {\mathcal U} \right)$, where $f_c$ is the carrier frequency, $c = 3 \times 10^8$ m/s is the speed of light, ${\mathcal U}$ is a random variable uniformly distributed in the interval $[0 \,,\, 1]$ and $\theta = 3.5^{\circ}$ is the Doppler beamwidth [@Haas]. The channel bandwidth is $90$ KHz, time duration is $23$ ms, and $f_c = 5.06$ GHz [@ICTCPaper]. We therefore choose $M = 45, N = 46$ and $\Delta f = 2 $ KHz ($T = 1/\Delta f = 0.5$ ms, $M \Delta f = 45 \times 2$ KHz = $90$ KHz, $N T = 46 \times 0.5$ ms = $23$ ms).
\[fig8\]
In Fig. \[fig8\] we plot the average SE achieved by OTFS modulation and that achieved by the DD domain modulation in (\[eqn44\]), as a function of increasing aircraft speed for different values of $\rho$. The SE achieved by the DD domain modulation is given by ${\mathbb E}[C]$ (where $C$ is given by (\[Ceqn\])) and that achieved by OTFS modulation is given by ${\mathbb E}[C_{\mbox{\tiny{Zak}}}]$ (where $C_{\mbox{\tiny{Zak}}}$ is given by equation $(24)$ in [@Saif2020]). Here, the expectation is w.r.t. the distribution of the channel path gains and the Doppler shift. From Fig. \[fig8\], it is clear that the SE performance of the DD domain modulation in (\[eqn44\]) is same as that of OTFS modulation. Further, for both these modulation schemes, for a given $\rho$, the SE performance remains constant with increasing aircraft speed (i.e., increasing Doppler shift).
Conclusion
==========
In this paper, using the ZAK representation of time-domain (TD) signals we have derived an orthonormal basis of time and bandwidth limited signals which are also localized in the delay-Doppler (DD) domain. We consider DD domain modulation based on this orthonormal basis and derive Orthogonal Time Frequency Space (OTFS) modulation. To the best of our knowledge, this paper is the first to rigorously derive OTFS modulation from first principles. We also show that with increasing time duration, the basis signals are increasingly localized in the DD domain, irrespective of the amount of Doppler shift. Increased localization reduces interference between information symbols modulated on different basis signals, thereby enabling joint DD domain equalization of all information symbols. Therefore, DD domain modulation achieves robustness to Doppler shift at the cost of increased latency.
Proof of Result \[zlem1\] {#prf_zlem1}
=========================
From (\[zdef1\]) it follows that $$\begin{aligned}
{\mathcal Z}_x(\tau + T, \nu) & \hspace{-3mm} = \hspace{-3.5mm} & \sqrt{T} \hspace{-1mm} \sum\limits_{k=-\infty}^{\infty} \hspace{-1mm} x(\tau + T + kT) \, e^{-j 2 \pi k \nu T } \nonumber \\
& \hspace{-3mm} = \hspace{-3.5mm} & e^{j 2 \pi \nu T} \sqrt{T} \hspace{-1mm} \sum\limits_{k=-\infty}^{\infty} \hspace{-2mm} x(\tau + (k+1)T ) e^{-j 2 \pi (k+1) \nu T } \nonumber \\
& \hspace{-3mm} = \hspace{-3.5mm} & e^{j 2 \pi \nu T} {\mathcal Z}_x(\tau,\nu).\end{aligned}$$ From (\[zdef1\]) it also follows that $$\begin{aligned}
{\mathcal Z}_x \left(\tau,\nu + \Delta f \right) & \hspace{-3mm} = & \hspace{-3mm} \sqrt{T} \sum\limits_{k=-\infty}^{\infty} \, x(\tau + kT) \, e^{-j 2 \pi k (\nu + \Delta f) T } \nonumber \\
& \hspace{-3mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-3mm} \sqrt{T} \sum\limits_{k=-\infty}^{\infty} \, x(\tau + kT) \, e^{-j 2 \pi k \nu T } \nonumber \\
& \hspace{-3mm} = & \hspace{-3mm} {\mathcal Z}_x \left(\tau,\nu \right) \end{aligned}$$ where step (a) follows from the fact that $T \Delta f = 1$.
Proof of Result \[zlem3\] {#prf_zlem3}
=========================
From (\[zdef1\]) it follows that $$\begin{aligned}
\label{zeqn8}
\hspace{-4mm} \sqrt{T} \int\limits_{0}^{\Delta f} \hspace{-1mm} {\mathcal Z}_x(t, \nu) d\nu & \hspace{-3mm} = & \hspace{-3mm} T \hspace{-2mm} \sum\limits_{k=-\infty}^{\infty} \hspace{-1mm} \left[ x(t + kT) \int\limits_{0}^{\Delta f} \hspace{-0.5mm} e^{-j 2 \pi \nu k T} d\nu \right].\end{aligned}$$ Since $\Delta f = 1/T$ and $k \in {\mathbb Z}$ we have $$\begin{aligned}
\label{zeqn9}
\int\limits_{0}^{\Delta f} \hspace{-0.5mm} e^{-j 2 \pi \nu k T} d\nu =
\begin{cases}
0 &, k \ne 0 \\
\frac{1}{T} &, k = 0 \\
\end{cases}.\end{aligned}$$ Using (\[zeqn9\]) in (\[zeqn8\]) we get (\[zeqn6\]). From (\[zdef1\]) it also follows that
[ $$\begin{aligned}
\label{zeqn10}
\frac{1}{\sqrt{T}} \hspace{-1mm} \int\limits_{0}^T \hspace{-1mm} {\mathcal Z}_x(\tau, f) e^{-j 2 \pi f \tau} d\tau & \hspace{-3.0mm} = & \hspace{-4.5mm} \sum\limits_{k=-\infty}^{\infty} \int\limits_{0}^{T} \hspace{-1mm} x(\tau + kT) e^{-j 2 \pi f (\tau + kT)} d\tau \nonumber \\
& & \hspace{-38mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}\sum\limits_{k=-\infty}^{\infty} \hspace{-2mm} \int\limits_{kT}^{(k+1)T} \hspace{-2mm} x(t) e^{-j 2 \pi f t} dt = \int\limits_{-\infty}^{\infty} \hspace{-1mm} x(t) e^{-j 2 \pi f t} dt \, = \, {\mathcal F}_x(f)\end{aligned}$$ ]{} where step (a) follows from the substitution $t = \tau + kT$ in the integral in the previous step.
Proof of Lemma \[plem1st\] {#prf_plem1st}
==========================
Using (\[zeqn6\]) from Result \[zlem3\], the TD signal $p_{(\tau_0, \nu_0)}(t)$ having ZAK representation ${\mathcal Z}_{(p, \tau_0, \nu_0)}(\tau, \nu)$ in (\[lceqn1\]) is given by
[ $$\begin{aligned}
p_{(\tau_0,\nu_0)}(t) & \hspace{-3mm} = & \hspace{-3mm} \sqrt{T} \int_{0}^{\Delta f} \hspace{-1mm} {\mathcal Z}_{(p, \tau_0, \nu_0)}(t, \nu) \, d\nu \nonumber \\
& \hspace{-32mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-15.5mm} \sqrt{T} \hspace{-2mm} \sum\limits_{n=-\infty}^{\infty} \hspace{-3mm} e^{j 2 \pi \nu_0 n T} \delta(t - \tau_0 - nT) \int_{0}^{\Delta f} \hspace{-3mm} \sum\limits_{m=-\infty}^{\infty} \hspace{-3mm} \delta (\nu - \nu_0 - m \Delta f) d\nu \nonumber \\
& \hspace{-32mm} = & \hspace{-15.5mm} \sqrt{T} \hspace{-2mm} \sum\limits_{n=-\infty}^{\infty} \hspace{-3mm} e^{j 2 \pi \nu_0 n T} \delta(t - \tau_0 - nT) \end{aligned}$$ ]{} where step (a) follows from the expression for ${\mathcal Z}_{(p, \tau_0, \nu_0)}(\tau, \nu)$ in (\[lceqn1\]).
Proof of Theorem \[thmbasis\] {#prf_thmbasis}
=============================
Firstly, we see that
[ $$\begin{aligned}
\label{prfceqn}
c_x(\tau_0, \nu_0) & \hspace{-3mm} = & \hspace{-3mm} \int_{-\infty}^{\infty} p^*_{(\tau_0, \nu_0)}(t) \, x(t) \, dt \nonumber \\
& \hspace{-3mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-3mm} \sqrt{T} \sum\limits_{n=-\infty}^{\infty} e^{-j 2 \pi \nu_0 n T} \int_{-\infty}^{\infty} \delta(t - \tau_0 - nT ) \, x(t) \, dt \nonumber \\
& \hspace{-3mm} = & \hspace{-3mm} \sqrt{T} \sum\limits_{n=-\infty}^{\infty} e^{-j 2 \pi \nu_0 n T} \, x(\tau_0 + nT) \, {\mathrel{\overset{\makebox[0pt]{{\tiny(b)}}}{=}}}\, {\mathcal Z}_x(\tau_0, \nu_0)\end{aligned}$$ ]{} where step (a) follows from the expression of $p_{(\tau_0, \nu_0)}(t)$ in (\[plemeqn2\]) and step (b) follows from (\[zdef1\]). Next
[ $$\begin{aligned}
\int_{0}^T \hspace{-2mm} \int_{0}^{\Delta f} \hspace{-4.5mm} c_x(\tau_0, \nu_0) p_{(\tau_0, \nu_0)}(t) d\tau_0 \, d\nu_0 & \hspace{-3mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-4mm} \int_{0}^T \hspace{-2mm} \int_{0}^{\Delta f} \hspace{-4.5mm} {\mathcal Z}_x(\tau_0, \nu_0) p_{(\tau_0, \nu_0)}(t) d\tau_0 d\nu_0 \nonumber \\
& \hspace{-94mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(b)}}}{=}}}& \hspace{-48mm} \hspace{-2mm} \sum\limits_{n_1 = -\infty}^{\infty} \sum\limits_{n_2 = -\infty}^{\infty} {\Bigg \{} \left[ \int_{0}^T \hspace{-1.5mm} x(\tau_0 + n_1 T) \, \delta(t - \tau_0 - n_2 T) \, d\tau_0 \right] \nonumber \\
& & \hspace{-7mm} \underbrace{\left[ T \int_{0}^{\Delta f} \hspace{-2mm} e^{j 2 \pi \nu_0 (n_2 - n_1) T} d\nu_0 \right]}_{= 1 \, \mbox{\tiny{if}} \, n_2 = n_1, 0 \, \mbox{\tiny{otherwise}} } {\Bigg \}} \nonumber \\
& \hspace{-94mm} = & \hspace{-48mm} \sum\limits_{n_1 = -\infty}^{\infty} \int_{0}^T \hspace{-1.5mm} x(\tau_0 + n_1 T) \, \overbrace{\delta(t - \tau_0 - n_1 T)}^{\mbox{\tiny{Dirac-delta at $\tau_0 = t - n_1 T$}}} \, d\tau_0 \nonumber \\
& \hspace{-94mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(c)}}}{=}}}& \hspace{-48mm} \int_{0}^T \hspace{-1.5mm} x\left(\tau_0 + \left\lfloor \frac{t}{T} \right\rfloor T \right) \, \overbrace{\delta\left(t - \tau_0 - \left\lfloor \frac{t}{T} \right\rfloor T \right)}^{\mbox{\tiny{Dirac-delta at $\tau_0 = t - \left\lfloor \frac{t}{T} \right\rfloor T$}}} \, d\tau_0 \, = \, x(t) \nonumber \\\end{aligned}$$ ]{} where step (a) follows from (\[prfceqn\]). Step (b) follows from substituting the expressions for ${\mathcal Z}_x(\tau_0, \nu_0)$ from (\[zdef1\]) and the expression for $p_{(\tau_0, \nu_0)}(t)$ from (\[plemeqn2\]) into the R.H.S. in step (a). Step (c) follows from the fact that the only non-zero term in the summation in the previous step is for $n_1 = \left\lfloor \frac{t}{T} \right\rfloor$.
ZAK Representation of Convolution of TD Signals {#prf_zlem5}
===============================================
This result states that the ZAK representation of the convolution of two time-domain (TD) signals is equivalent to convolution of their ZAK representations in the delay domain.
\[zlem5\] Let ${\mathcal Z}_a(\tau,\nu)$ and ${\mathcal Z}_b(\tau,\nu)$ be the ZAK representation of $a(t)$ and $b(t)$ respectively. Consider their TD convolution $c(t) = \int\limits_{-\infty}^{\infty} a(t') \, b(t - t') \, dt'$. The ZAK representation of $c(t)$ is given by
[ $$\begin{aligned}
\label{zeqn12}
{\mathcal Z}_c(\tau, \nu) & = & \frac{1}{\sqrt{T}} \int\limits_{0}^T {\mathcal Z}_a(\tau - \tau',\nu) \, {\mathcal Z}_b(\tau' , \nu) \, d\tau' \nonumber \\
& = & \frac{1}{\sqrt{T}} \int\limits_{0}^T {\mathcal Z}_a(\tau',\nu) \, {\mathcal Z}_b(\tau - \tau' , \nu) \, d\tau'.\end{aligned}$$ ]{}
Starting with the R.H.S. of (\[zeqn12\]) we have
[ $$\begin{aligned}
\frac{1}{\sqrt{T}} \int\limits_{0}^T {\mathcal Z}_a(\tau - \tau',\nu) {\mathcal Z}_b(\tau' , \nu) d\tau' & {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \nonumber \\
& & \hspace{-55mm} \sqrt{T} \hspace{-2mm} \sum\limits_{k_1=-\infty}^{\infty} \sum\limits_{k_2=-\infty}^{\infty} \hspace{-2mm} e^{- j 2 \pi (k_1 + k_2) \nu T} \hspace{-1mm} \int\limits_{0}^T a(\tau - \tau' + k_1 T) b(\tau' + k_2 T) d\tau' \nonumber \\
& \hspace{-98mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(b)}}}{=}}}& \hspace{-52.5mm} \sqrt{T} \hspace{-2mm} \sum\limits_{k_1=-\infty}^{\infty} \sum\limits_{k_2=-\infty}^{\infty} \hspace{-3mm} e^{- j 2 \pi (k_1 + k_2) \nu T} \hspace{-4.5mm} \int\limits_{k_2 T}^{(k_2 + 1)T} \hspace{-4.5mm} a(\tau + (k_1 +k_2) T - \tau'' ) b(\tau'') d\tau'' \nonumber \\
& \hspace{-98mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(c)}}}{=}}}& \hspace{-52.5mm} \sqrt{T} \hspace{-2mm} \sum\limits_{k_3=-\infty}^{\infty} \hspace{-3mm} e^{- j 2 \pi k_3 \nu T} \left[ \sum\limits_{k_2=-\infty}^{\infty} \hspace{-2.5mm} \int\limits_{k_2 T}^{(k_2 + 1)T} \hspace{-4.5mm} a(\tau + k_3 T - \tau'' ) b(\tau'') d\tau'' \right] \nonumber \\
& \hspace{-98mm} = & \hspace{-52.5mm} \sqrt{T} \hspace{-2mm} \sum\limits_{k_3=-\infty}^{\infty} \hspace{-3mm} e^{- j 2 \pi k_3 \nu T} \left[ \int\limits_{-\infty}^{\infty} \hspace{-1.5mm} a(\tau + k_3 T - \tau'' ) b(\tau'') d\tau'' \right] \nonumber \\
& \hspace{-98mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(d)}}}{=}}}& \hspace{-52.5mm} \sqrt{T} \hspace{-2mm} \sum\limits_{k_3=-\infty}^{\infty} \hspace{-3mm} e^{- j 2 \pi k_3 \nu T} c(\tau + k_3 T ) \, = \, {\mathcal Z}_c(\tau, \nu)\end{aligned}$$ ]{} where step (a) follows from (\[zdef1\]). In step (b) we have changed the integration variable to $\tau'' = \tau' + k_2 T$. Step (c) follows from replacing the summation variable $k_1$ by $k_3 = k_1 + k_2$ since in step (b), $k_1$ always appears as $(k_1 + k_2)$. Step (d) follows from the fact that $c(t)$ is the convolution of $a(t)$ and $b(t)$, i.e., $c(t) = \int\limits_{-\infty}^{\infty} a(t - t'') b(t'') dt''$. The last step follows from (\[zdef1\]). The second equality in the R.H.S. of (\[zeqn12\]) can be proved similarly.
ZAK Representation of Product of TD Signals {#prf_zlem6}
===========================================
This result states that the ZAK representation of the product of two time-domain (TD) signals is equivalent to convolution of their ZAK representations in the Doppler domain.
\[zlem6\] Let ${\mathcal Z}_a(\tau,\nu)$ and ${\mathcal Z}_b(\tau,\nu)$ be the ZAK representation of TD signals $a(t)$ and $b(t)$ respectively. Consider their product $c(t) = a(t) b(t)$. The ZAK representation of $c(t)$ is given by
[ $$\begin{aligned}
\label{zeqn13}
{\mathcal Z}_c(\tau, \nu) & = & \sqrt{T} \int\limits_{0}^{\Delta f} {\mathcal Z}_a(\tau,\nu - \nu') {\mathcal Z}_b(\tau , \nu') d\nu' \nonumber \\
& = & \sqrt{T} \int\limits_{0}^{\Delta f} {\mathcal Z}_a(\tau,\nu') {\mathcal Z}_b(\tau, \nu - \nu') d\nu'.\end{aligned}$$ ]{}
Starting with the R.H.S. in (\[zeqn13\]) we get
[ $$\begin{aligned}
\sqrt{T} \int\limits_{0}^{\Delta f} {\mathcal Z}_a(\tau,\nu - \nu') {\mathcal Z}_b(\tau , \nu') d\nu' & & \nonumber \\
& & \hspace{-44mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}\, \sqrt{T} \hspace{-2mm} \sum\limits_{k_1=-\infty}^{\infty} \sum\limits_{k_2=-\infty}^{\infty} \hspace{-1mm} {\Bigg [} a(\tau + k_1 T) b(\tau + k_2 T) e^{-j 2 \pi \nu k_1 T} \nonumber \\
& & \hspace{-6mm} \underbrace{\frac{1}{\Delta f} \int_{0}^{\Delta f} e^{j 2 \pi \nu' (k_1 - k_2) T} d\nu' }_{= 1 \, \mbox{\tiny{if}} \, k_1 = k_2, \, 0 \, \mbox{\tiny{otherwise}}} {\Bigg ]} \nonumber \\
& \hspace{-88mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(b)}}}{=}}}& \hspace{-46mm} \sqrt{T} \hspace{-2mm} \sum\limits_{k_1=-\infty}^{\infty} \hspace{-1mm} \underbrace{a(\tau + k_1 T) b(\tau + k_1T)}_{= c(\tau + k_1T)} e^{-j 2 \pi \nu k_1 T} \, {\mathrel{\overset{\makebox[0pt]{{\tiny(c)}}}{=}}}\, {\mathcal Z}_c(\tau,\nu)\end{aligned}$$ ]{} where step (a) follows from (\[zdef1\]). Step (b) follows from the fact that in the double summation in the previous step, only those terms are non-zero for which $k_2 = k_1$. Step (c) follows from (\[zdef1\]) and the fact that $c(t) = a(t) b(t)$.
Proof of Theorem \[thm21\] {#prf_thm21}
==========================
Using Result \[zlem6\] in Appendix \[prf\_zlem6\], for $0 \leq \tau < T$, the ZAK representation of $c(t) = p_{(\tau_0, \nu_0)}(t) \, q(t)$ is given by
[ $$\begin{aligned}
\label{zceqn23}
{\mathcal Z}_c(\tau, \nu) & = & \sqrt{T} \int_{0}^{\Delta f} {\mathcal Z}_{(p,\tau_0, \nu_0)}(\tau, \nu') \, {\mathcal Z}_q(\tau, \nu - \nu') \, d\nu' \nonumber \\
& {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \sqrt{T} \int_{0}^{\Delta f} \delta(\tau - \tau_0) \delta(\nu' - \nu_0) \, {\mathcal Z}_q(\tau, \nu - \nu') \, d\nu' \nonumber \\
& = & \sqrt{T} \, \delta(\tau - \tau_0) \, {\mathcal Z}_q(\tau, \nu - \nu_0)\end{aligned}$$ ]{} where step (a) follows from the fact that ${\mathcal Z}_{(p,\tau_0, \nu_0)}(\tau, \nu)$ is the ZAK representation of $p_{(\tau_0, \nu_0)}(t)$ (see Lemma \[plem1st\]), and also that, for $0 \leq \tau < T$ and $0 \leq \nu' < \Delta f$, ${\mathcal Z}_{(p, \tau_0, \nu_0)}(\tau, \nu') = \delta(\tau - \tau_0) \delta(\nu' - \nu_0)$ (see (\[lceqn1\])). Here ${\mathcal Z}_q(\tau, \nu)$ is the ZAK representation of $q(t)$, and is given by
[ $$\begin{aligned}
\label{zqeqn1}
{\mathcal Z}_q(\tau, \nu) & \hspace{-3mm} = & \hspace{-3mm} \sqrt{T} \hspace{-1mm} \sum\limits_{n=-\infty}^{\infty} \hspace{-2mm} q(\tau + nT) \, e^{-j 2 \pi \nu n T} \, {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}\, \sqrt{T} \hspace{-1mm} \sum\limits_{n=- \left\lfloor \frac{\tau}{T} \right\rfloor }^{N-1 - \left\lfloor \frac{\tau}{T} \right\rfloor} \hspace{-2.5mm} e^{-j 2 \pi \nu n T} \nonumber \\
& \hspace{-3mm} = & \hspace{-3mm} \sqrt{T} e^{j 2 \pi \nu \left\lfloor \frac{\tau}{T} \right\rfloor T } e^{-j \pi \nu (N-1) T} \, \frac{\sin\left( \pi \nu N T \right)}{\sin\left( \pi \nu T \right)}\end{aligned}$$ ]{} where step (a) follows from the fact that $q(t) = 1$ for $0 \leq t < NT$ and is zero otherwise (see (\[qteqn\])). From (\[basiseqn1\]) it follows that $\psi^{(q,s)}_{(\tau_0, \nu_0)}(t) = \left( p_{(\tau_0, \nu_0)}(t) \, q(t) \right) \star s(t) = c(t) \star s(t)$ and therefore using Result \[zlem5\] from Appendix \[prf\_zlem5\] it follows that
[ $$\begin{aligned}
{\mathcal Z}_{\psi, \tau_0, \nu_0}(\tau, \nu) & \hspace{-3mm} = & \hspace{-3mm} \frac{1}{\sqrt{T}} \int_{0}^T {\mathcal Z}_c(\tau', \nu) \, {\mathcal Z}_s(\tau - \tau' , \nu) \, d\tau' \nonumber \\
& \hspace{-14mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-9mm} \int_{0}^{T} \delta(\tau' - \tau_0) \, {\mathcal Z}_q(\tau', \nu - \nu_0) \, {\mathcal Z}_s(\tau - \tau' , \nu) \, d\tau' \nonumber \\
& = & {\mathcal Z}_q(\tau_0, \nu - \nu_0) \, {\mathcal Z}_s(\tau - \tau_0 , \nu)\end{aligned}$$ ]{} where step (a) follows from the expression of ${\mathcal Z}_c(\tau', \nu)$ in (\[zceqn23\]) for $0 \leq \tau' < T$. Here ${\mathcal Z}_s(\tau,\nu)$ is the ZAK representation of $s(t)$ which is given by
[ $$\begin{aligned}
{\mathcal Z}_s(\tau, \nu) & = & \sqrt{T} \sum\limits_{n=-\infty}^{\infty} s(\tau + nT) \, e^{-j 2 \pi \nu n T} \nonumber \\
& {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \sqrt{T} \int_{0}^{M \Delta f} \hspace{-2mm} e^{j 2 \pi f \tau } \underbrace{\left[ \sum\limits_{n=-\infty}^{\infty} \hspace{-2mm} e^{j 2 \pi (f - \nu) n T} \right]}_{= \frac{1}{T} \hspace{-2mm} \sum\limits_{m = - \infty}^{\infty} \delta(f - \nu - m \Delta f) } \, df \nonumber \\
& {\mathrel{\overset{\makebox[0pt]{{\tiny(b)}}}{=}}}& \frac{1}{\sqrt{T}} \sum\limits_{m=-\infty}^{\infty} \int_{0}^{M \Delta f} \hspace{-4mm} \underbrace{\delta(f - \nu - m \Delta f)}_{\mbox{\small{Dirac-delta at}} \, f = \nu + m \Delta f} \hspace{-3mm} e^{j 2 \pi f \tau } df. \nonumber \\
& {\mathrel{\overset{\makebox[0pt]{{\tiny(c)}}}{=}}}& \frac{1}{\sqrt{T}} e^{j 2 \pi \nu \tau} \sum\limits_{m= -\left\lfloor \frac{\nu}{\Delta f} \right\rfloor }^{(M-1) - \left\lfloor \frac{\nu}{\Delta f} \right\rfloor} e^{j 2 \pi m \Delta f \tau} \nonumber \\
& & \hspace{-20mm} = \frac{1}{\sqrt{T}} e^{j 2 \pi \nu \tau} e^{-j 2 \pi \left\lfloor \frac{\nu}{\Delta f} \right\rfloor \Delta f \tau} \, e^{j \pi (M-1) \Delta f \tau} \, \frac{\sin( \pi M \Delta f \tau)}{\sin(\pi \Delta f \tau)}\end{aligned}$$ ]{} where step (a) follows from the fact that $s(t) = \int_{0}^{M \Delta f} e^{j 2 \pi f t} df$ (see (\[steqn\])). Step (b) follows from the standard equation $\sum\limits_{n=-\infty}^{\infty} \hspace{-2mm} e^{j 2 \pi f n T} = \frac{1}{T} \hspace{-2mm} \sum\limits_{m = - \infty}^{\infty} \hspace{-3mm} \delta\left(f - \frac{m}{T} \right)$ and the fact that $T = 1/\Delta f$. Step (c) follows from the fact that $0 \leq \nu + m \Delta f < M \Delta f$ for $- \left\lfloor \frac{\nu}{\Delta f} \right\rfloor \leq m \leq (M-1) - \left\lfloor \frac{\nu}{\Delta f} \right\rfloor$. Indeed, using the R.H.S. above as the expression for ${\mathcal Z}_s(\tau, \nu)$, from (\[zeqn7\]) it can be checked that the Fourier transform of $s(t)$ i.e., ${\mathcal F}_s(f) = \frac{1}{\sqrt{T}} \int\limits_{0}^T {\mathcal Z}_s(\tau, f) e^{-j 2 \pi f \tau} \, d\tau$ is one when $f \in [0 \,,\, M \Delta f)$ and is zero otherwise.
Proof of Theorem \[lem18\] {#prf_lem18}
==========================
From the expression of $\alpha_{(k,l)}(t)$ in (\[eqn27\]) we have
[ $$\begin{aligned}
\int_{-\infty}^{\infty} \hspace{-3.5mm} \alpha_{(k_1,l_1)}(t) \, \alpha^*_{(k_2,l_2)}(t) \, dt & \hspace{-4mm} = & \hspace{-4mm} \frac{T}{MN} \sum\limits_{n_1 = 0}^{N-1} \sum\limits_{n_2 = 0}^{N-1} \hspace{-1mm} {\Bigg [} e^{j 2 \pi \frac{n_1 k_1 - n_2 k_2}{N} } \nonumber \\
& & \hspace{-7mm} (M \Delta f)^2 e^{j \pi M \Delta f \left(\frac{(l_2 - l_1)T}{M} + (n_2 - n_1)T \right)} \nonumber \\
& & \hspace{-43mm} \int_{-\infty}^{\infty} \hspace{-4mm} \mbox{\small{sinc}} \hspace{-1mm} \left( \hspace{-1mm} M \Delta f \hspace{-1mm} \left(\hspace{-1mm} t - \frac{l_1T}{M} - n_1 T \right)\hspace{-1mm} \right) \mbox{\small{sinc}} \hspace{-1mm} \left(M \Delta f \hspace{-1mm} \left( \hspace{-1mm} t - \frac{l_2T}{M} - n_2 T \right) \hspace{-1mm} \right) dt {\Bigg ]} \nonumber \\
& \hspace{-70mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-36mm} \frac{T}{MN} \sum\limits_{n_1 = 0}^{N-1} \sum\limits_{n_2 = 0}^{N-1} \hspace{-1mm} {\Bigg [} e^{j 2 \pi \frac{n_1 k_1 - n_2 k_2}{N} } e^{j \pi M \Delta f \left(\frac{(l_2 - l_1)T}{M} + (n_2 - n_1)T \right)} \nonumber \\
& & \hspace{-32mm} (M \Delta f) \mbox{\small{sinc}}\left(M \Delta f \left(\frac{(l_2 - l_1)T}{M} + (n_2 - n_1) T \right) \right) {\Bigg ]} \nonumber \\
& \hspace{-70mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(b)}}}{=}}}& \hspace{-36mm} \frac{1}{N} \sum\limits_{n_1 = 0}^{N-1} \sum\limits_{n_2 = 0}^{N-1} \hspace{-1mm} {\Bigg [} e^{j 2 \pi \frac{n_1 k_1 - n_2 k_2}{N} } e^{j \pi M \Delta f \left(\frac{(l_2 - l_1)T}{M} + (n_2 - n_1)T \right)} \nonumber \\
& & \delta[l_1 - l_2] \delta[n_1 - n_2] {\Bigg ]} \nonumber \\
& \hspace{-72mm} = & \hspace{-38mm} \delta[l_1 - l_2] \left( \frac{1}{N} \sum\limits_{n_1=0}^{N-1} e^{j 2 \pi \frac{n_1 (k_1 - k_2)}{N}} \right) \, = \, \delta[l_1 - l_2] \delta[k_1 - k_2]\end{aligned}$$ ]{} where step (a) follows from the fact that $\int_{-\infty}^{\infty} W^2 \mbox{\small{sinc}}(W(t - \tau_1)) \mbox{\small{sinc}}(W(t - \tau_2)) dt = W \mbox{\small{sinc}}(W(\tau_2 - \tau_1))$. Step (b) follows from the fact that $\mbox{\small{sinc}}\left(M \Delta f \left(\frac{(l_2 - l_1)T}{M} + (n_2 - n_1) T \right) \right) = \mbox{\small{sinc}}\left( l_2 - l_1 + M(n_2 - n_1)\right)$ which is one only when $l_1 = l_2$ and $n_1 = n_2$, and is otherwise zero. This is because, $l_1,l_2 =0,1,\cdots, M-1$ and therefore $(l_2 - l_1)$ can never be a non-zero integer multiple of $M$. This shows that $\boldsymbol \alpha$ is a basis with $MN$ orthonormal signals and is therefore $MN$-dimensional.
Proof of Theorem \[thm73\] {#prfthm73}
==========================
The ZAK representation of $y(t)$ is given by
[ $$\begin{aligned}
\label{eqn57}
\hspace{-1mm} {\mathcal Z}_y(\tau, \nu) & \hspace{-3mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-3mm} \sqrt{T} \hspace{-1mm} \sum\limits_{n'=0}^{N} \hspace{-1mm} y(\tau + n' T) e^{-j 2 \pi \nu n' T}, \,\, 0 \leq \tau < T \,,\, 0 \leq \nu < \Delta f \nonumber \\
& & \hspace{-15mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(b)}}}{=}}}\, \sum\limits_{i=1}^{L} h_i \, e^{j 2 \pi \nu_i (\tau - \tau_i)} \, {\mathcal Z}_x(\tau - \tau_i, \nu - \nu_i) \,\, + \,\, {\mathcal Z}_n(\tau, \nu)\end{aligned}$$ ]{} where step (a) follows from (\[zdef1\]) and step (b) follows from Result \[zlem52\]. Here ${\mathcal Z}_x(\tau, \nu)$ and ${\mathcal Z}_n(\tau, \nu)$ are the ZAK representations of $x(t)$ and $n(t)$ respectively. From (\[eqn44\]), the ZAK representation of $x(t)$ is given by $$\begin{aligned}
\label{eqn58}
{\mathcal Z}_x(\tau, \nu) & \hspace{-3mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-3mm} \frac{1}{\sqrt{MN}} \sum\limits_{k=0}^{N-1} \sum\limits_{l=0}^{M-1} x[k,l] {\mathcal Z}_{\psi, \frac{lT}{M}, \frac{k \Delta f}{N}}(\tau, \nu)\end{aligned}$$ where ${\mathcal Z}_{\psi,\frac{lT}{M}, \frac{k \Delta f}{N} }(\tau, \nu)$ is the ZAK representation of $\psi^{(q,s)}_{( lT/M, k /NT)}(t)$ (see (\[eqn43\])). Step (a) follows from the linearity of the ZAK representation (i.e., the ZAK representation of the sum of two TD signals is the sum of their ZAK representations), and the fact that $\alpha_{(k,l)}(t) = \frac{1}{\sqrt{MN}} \psi^{(q,s)}_{( lT/M, k /NT)}(t)$ (see (\[eqn27\])). From Theorem \[thm21\], the ZAK representation of $\psi^{(q,s)}_{( lT/M, k /NT)}(t)$ is given by
[ $$\begin{aligned}
\label{eqn52}
{\mathcal Z}_{\psi, \frac{lT}{M}, \frac{k \Delta f}{N}}(\tau, \nu) & \hspace{-3mm} = & \hspace{-3mm} {\mathcal Z}_q\left( \frac{lT}{M} , \nu - \frac{k \Delta f}{N} \right) \, {\mathcal Z}_s\left( \tau - \frac{lT}{M}, \nu \right) \nonumber \\
& & \hspace{-32mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}e^{-j \pi \left( \nu - \frac{k \Delta f}{N} \right) (N-1) T} \, \frac{\sin\left( \pi \left( \nu - \frac{k \Delta f}{N} \right) N T \right)}{\sin\left( \pi \left( \nu - \frac{k \Delta f}{N} \right) T \right)} e^{j 2 \pi \nu \left( \tau - \frac{lT}{M} \right)} \nonumber \\
& & \hspace{-32mm} e^{-j 2 \pi \left\lfloor \frac{\nu}{\Delta f} \right\rfloor \Delta f \left( \tau - \frac{lT}{M} \right) } e^{j \pi (M-1) \Delta f \left( \tau - \frac{lT}{M} \right)} \frac{\sin( \pi M \Delta f \left( \tau - \frac{lT}{M} \right) )}{\sin(\pi \Delta f \left( \tau - \frac{lT}{M} \right) )}\end{aligned}$$ ]{} where step (a) follows from the expressions of ${\mathcal Z}_q(\tau, \nu)$ and ${\mathcal Z}_s(\tau, \nu)$ in (\[expreqn1\]). Using (\[eqn52\]) in (\[eqn58\]) we get an expression for ${\mathcal Z}_x(\tau, \nu)$ in terms of the information symbols $x[k,l], k=0,1,\cdots, N-1, l=0,1,\cdots, M-1$. Using this expression in (\[eqn57\]) we then get an expression for ${\mathcal Z}_y(\tau, \nu)$ in terms of the information symbols $x[k,l]$. Using this expression, the sampled DD domain signal $Y[k',l'] = {\mathcal Z}_y\left( \tau = \frac{l'T}{M} , \nu = \frac{k' \Delta f}{N} \right)$ is given by (\[eqn67\]).
Proof of Theorem \[thm8\] {#prfthm8}
=========================
In (\[eqn67\]), $Z[k',l']$ is given by $$\begin{aligned}
\label{eqn79}
Z[k',l'] & = & {\mathcal Z}_n\left(\tau = \frac{l'T}{M} , \nu = \frac{k' \Delta f}{N} \right) \nonumber \\
& {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \sqrt{T} \sum\limits_{n'=0}^{N} n\left(n'T + \frac{l'T}{M} \right) \, e^{- j 2 \pi \frac{n' k'}{N} }\end{aligned}$$ where step (a) follows from the fact that the received TD signal $y(t)$ is limited to the interval $[0 \,,\, (N+1)T )$. Let the information symbols $x[k,l] \, \sim \, \mbox{\small{i.i.d.}} \, {\mathcal C}{\mathcal N}(0 , \rho)$. Then, since the signals $\alpha_{(k,l)}(t)$ belong to the orthonormal basis $\boldsymbol \alpha$ (see Theorem \[lem18\]), from the first equation in (\[eqn44\]) it follows that $$\begin{aligned}
{\mathbb E} \left[ \int_{-\infty}^{\infty} \left\vert x(t) \right\vert^2 \, dt \right] & = & MN \rho.\end{aligned}$$ Since the transmit signal $x(t)$ has most of its energy in the time-interval $[0 \,,\, NT)$, the average transmit power is $MN \rho/NT = M \rho/T$. Let the power spectral density (PSD) of AWGN be unity. Since the communication bandwidth is $M \Delta f$, the AWGN power at the receiver is $M\Delta f$. Therefore, the ratio of the transmit power to the receiver noise power is $$\begin{aligned}
\frac{M \rho / T}{M \Delta f } \, = \, \rho.\end{aligned}$$ Let the sampled DD domain signal $Y[k',l']= {\mathcal Z}_y\left(\tau = \frac{l'T}{M}, \nu = \frac{k' \Delta f}{N} \right), k'=0,1,\cdots, N-1, l'=0,1,\cdots, M-1$ be organized into a vector ${{\bf y}} \in {\mathbb C}^{MN \times 1}$, where the $(k'M + l' + 1)$-th element of ${{\bf y}}$ is $Y[k',l']$, $k'=0,1,\cdots,N-1, l'=0,1,\cdots, M-1$. Similarly, $x[k,l]$ is organized into the information symbol vector ${\bf x} \in {\mathbb C}^{MN \times 1}$ where the $(kM+l+1)$-th element of ${\bf x}$ is $x[k,l]$. Also, let ${\Tilde {\bf H}} \in {\mathbb C}^{MN \times MN}$ be the effective DD domain channel matrix whose element in its $(k'M + l' +1)$-th row and $(kM + l+1)$-th column is ${\Tilde h}[k',l',k,l]$ (${\Tilde h}[k',l',k,l]$ is defined in (\[eqn67\])). From (\[eqn67\]) it then follows that $$\begin{aligned}
\label{eqn71}
{ {\bf y}} & = & \sqrt{MN} \, {\Tilde {\bf H}} \, {\bf x} \, + \, {\bf z}\end{aligned}$$ where ${\bf z} \in {\mathbb C}^{MN \times 1}$ is the vector of noise samples, i.e., the $(k'M + l' + 1)$-th element of ${\bf z}$ is $Z[k',l']$. Since the communication bandwidth is $M \Delta f$ and the PSD of AWGN is unity, the noise samples $n\left(n'T + \frac{l'T}{M} \right), \, n'=0,1,\cdots, N, l'=0,1,\cdots, M-1$ are i.i.d. ${\mathcal C}{\mathcal N}(0, M \Delta f )$. The covariance matrix of the noise vector ${\bf z}$ is denoted by ${\bf K}_{\bf z}$. The expression for ${\bf K}_{\bf z}$ follows from equation $(22)$ in [@Saif2020], i.e., the element of ${\bf K}_{\bf z}$ in its $(k'M + l' +1)$-th row and $(kM + l +1)$-th column is given by
[ $$\begin{aligned}
\label{eqn81}
{K}_{\bf z}[{k'M + l' +1,kM + l +1}] & \hspace{-3mm} = & \hspace{-3mm} {\mathbb E} \left[ Z[k',l'] Z^*[k,l] \right]\nonumber \\
& \hspace{-12mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-7mm}
\begin{cases}
M N \left(1 + \frac{1}{N} \right) , k' = k \,,\, l' = l \\
M N \left( \frac{1}{N} \right) \hspace{6mm} , k' \ne k \,,\, l' = l \\
0 \hspace{11mm} \,, l' \ne l \\
\end{cases}, \nonumber \\
& & \hspace{-32mm} k',k = 0,1,\cdots, N-1 \,,\, l',l = 0,1,\cdots, M-1.\end{aligned}$$ ]{} where step (a) can be derived from the expression of $Z[k',l']$ in the R.H.S. of (\[eqn79\]). From (\[eqn71\]) it then follows that the spectral efficiency (SE) achieved by the DD domain modulation in (\[eqn44\]) is given by [@DTse] $$\begin{aligned}
C & = & \frac{1}{MN} \, \log_2 \left\vert {\bf I} \, + \, M N \rho \, {\Tilde {\bf H}}^H {{\bf K}_{{\bf z}}}^{-1} {\Tilde {\bf H}} \right\vert \nonumber \\
& = & \frac{1}{MN} \, \log_2 \left\vert {\bf I} \, + \, \rho \, {\Tilde {\bf H}}^H {\Tilde {\bf K}}^{-1} {\Tilde {\bf H}} \right\vert, \nonumber \\
{\Tilde {\bf K}} & {\stackrel {\Delta} {=} }& \frac{1}{MN} {{\bf K}_{{\bf z}}}\end{aligned}$$ where the elements of ${{\bf K}_{{\bf z}}}$ are given by (\[eqn81\]).
Fraction of Interfered Information Symbols in CP-OFDM {#cpofdm}
=====================================================
In OFDM, the $M$ orthonormal basis signals are
[ $$\begin{aligned}
\label{ofdmeqn1}
\phi_{k}(t) & = \begin{cases}
\frac{1}{\sqrt{T}} e^{j 2 \pi k \frac{t}{T}} \,&,\, 0 \leq t < T \\
0 \, &, \, \mbox{\small{otherwise}}
\end{cases} \nonumber \\
& & \hspace{-50mm} k=0,1,\cdots, M-1.\end{aligned}$$ ]{} With a cyclic prefix equal to the path delay $\tau'$, and information symbols $x[k], k=0,1,\cdots, M-1$, the transmit CP-OFDM signal is
[ $$\begin{aligned}
\label{ofdmeqn2}
x_{\mbox{\tiny{ofdm}}}(t) & = \begin{cases}
\sum\limits_{k=0}^{M-1} x[k] \, \phi_k(t) \, &, \, 0 \leq t < T \\
\sum\limits_{k=0}^{M-1} x[k] \, \phi_k(t+T) \, &, \, - \tau' \leq t < 0 \\
0 \, &, \, \mbox{\small{otherwise}}
\end{cases}.\end{aligned}$$ ]{} The received noise-free TD signal is given by $$\begin{aligned}
\label{ofdmeqn3}
y_{\mbox{\tiny{ofdm}}}(t) & = & h' \, e^{j 2 \pi \nu' (t - \tau')} \, x_{\mbox{\tiny{ofdm}}}(t - \tau') \end{aligned}$$ where $h'$ is the channel gain of the single channel path. The receiver removes the CP and computes
[ $$\begin{aligned}
\label{ofdmeqn4}
Y_{\mbox{\tiny{ofdm}}}[m] & \hspace{-3mm} {\stackrel {\Delta} {=} }& \hspace{-3mm} \int_{0}^{T} \hspace{-2mm} y_{\mbox{\tiny{ofdm}}}(t) \, \phi_{m}^*(t) \, dt \nonumber \\
& \hspace{-3mm} {\mathrel{\overset{\makebox[0pt]{{\tiny(a)}}}{=}}}& \hspace{-3mm} \sum\limits_{k=0}^{M-1} x[k] \, H_{\mbox{\tiny{ofdm}}}[m,k], \,\, m=0,1,\cdots, M-1, \nonumber \\
H_{\mbox{\tiny{ofdm}}}[m,k] & {\stackrel {\Delta} {=} }& h' \, \int_{0}^T \hspace{-1mm} e^{j 2 \pi \nu' (t - \tau')} \, \phi_k(t - \tau') \, \phi_m^*(t) \, dt \nonumber \\
& & \hspace{-22mm} = h' e^{-j 2 \pi \frac{\tau'}{T} \left({\nu' T} + k \right)} \, e^{j \pi \left( {\nu' T} + k - m \right)} \, \mbox{\small{sinc}}\left( {\nu' T} + k - m \right)\end{aligned}$$ ]{} where step (a) follows from substituting the R.H.S. of (\[ofdmeqn2\]) in (\[ofdmeqn3\]), and then using the resulting expression for $y_{\mbox{\tiny{ofdm}}}(t)$ in the integral of the first equation in (\[ofdmeqn4\]). From (\[ofdmeqn4\]) it is clear that due to the channel induced Doppler shift $\nu'$, an information symbol $x[k]$ transmitted on the $k$-th sub-carrier is received as $H[m,k] x[k]$ on the $m$-th sub-carrier. Since $\vert H_{\mbox{\tiny{ofdm}}}[m,k] \vert^2 = \vert h' \vert^2 \, \mbox{\small{sinc}}^2(\nu' T + k - m )$ (see (\[ofdmeqn4\])), it follows that most of the energy of $x[k]$ is received in and around the $m= \left\lfloor k + {\nu' T} \right\rfloor$-th sub-carrier. Next, for a given $k \in \{ 0,1,\cdots, M-1 \}$, let ${\mathcal G}_k$ denote the smallest cardinality set of sub-carrier indices such that the fraction of total energy of $x[k]$ received in the sub-carrier indices in ${\mathcal G}_k$ is at least $0.99$, i.e.,
[ $$\begin{aligned}
{\mathcal G}_k & \hspace{-3mm} {\stackrel {\Delta} {=} }& \hspace{-3mm} \arg \hspace{-9mm} \min_{{\mathcal D} \subseteq {\mathcal V} \, {\Big \vert} \, \frac{\sum\limits_{m \in {\mathcal D} } \left\vert H_{\mbox{\tiny{ofdm}}}[m,k] \right\vert^2}{\sum\limits_{m \in {\mathcal V}} \left\vert H_{\mbox{\tiny{ofdm}}}[m,k] \right\vert^2 } \geq 0.99} \hspace{-8mm} \left\vert {\mathcal D} \right\vert
\,\,\,\,\, = \,\,\,\,\, \arg \hspace{-9mm} \min_{{\mathcal D} \subseteq {\mathcal V} \, {\Big \vert} \, \frac{\sum\limits_{m \in {\mathcal D} } \mbox{\tiny{sinc}}^2\left( {\nu'T} + k - m \right)}{\sum\limits_{m \in {\mathcal V}} \mbox{\tiny{sinc}}^2\left( {\nu'T} + k - m \right)} \geq 0.99} \hspace{-8mm} \left\vert {\mathcal D} \right\vert, \nonumber \\
{\mathcal V} & {\stackrel {\Delta} {=} }& \{ 0, 1, \cdots, M-1 \}.\end{aligned}$$ ]{} The fraction of information symbols which are interfered by $x[k]$ is then given by
[ $$\begin{aligned}
\mbox{\small{Frac. of interfered symbols}} & = & \frac{\left\vert {\mathcal G}_k \right\vert - 1}{M-1}.\end{aligned}$$ ]{}
[1]{} IMT Vision - Framework and Overall Objectives of the Future Deployment of IMT for 2020 and beyond, Recommendation ITU-R M-2083-0, Sept. 2015 (www.itu.int). Fifth Generation New Radio (5G NR), Third Generation Partnership Project (3GPP), www.3gpp.org. R. Hadani, S. Rakib, M. Tsatsanis, A. Monk, A. J. Goldsmith, A. F. Molisch, “Orthogonal Time Frequency Space Modulation,” [*IEEE Wireless Comm. and Networking Conference (WCNC’17)*]{}, March 2017. A. Monk, R. Hadani, M. Tsatsanis and S. Rakib, “OTFS - Orthogonal Time Frequency Space: A Novel Modulation Meeting 5G High Mobility and Massive MIMO Challenges,” arXiv:1608.02993 \[cs.IT\] 9 Aug. 2016. R. Hadani and A. Monk, “OTFS: A New Generation of Modulation Addressing the Challenges of 5G,” arXiiv:1802.02623\[cs.IT\], www.arxiv. org, Feb. 2018. P. Raviteja, K. T. Phan, Y. Hong and E. Viterbo, “Interference Cancellation and Iterative Detection for Orthogonal Time Frequency Space Modulation,” [*IEEE Transactions on Wireless Communications*]{}, vol. 17, no. 10, Oct. 2018. M. K. Ramachandran and A. Chockalingam, “MIMO-OTFS in high-doppler fading channels: Signal detection and channel estimation,” [*in Proc. IEEE Global Commun. Conf. (GLOBECOM)*]{}, Kansas City, MO, USA, Dec. 2018, pp. 206–212. K. R. Murali and A. Chockalingam, “On OTFS Modulation for High-Doppler Fading Channels," [*in Proc. Information Theory and Applications Workshop (ITA)*]{}, Feb. 2018. P. Raviteja, K. T. Phan, Y. Hong, “Embedded Pilot-Aided Channel Estimation for OTFS in Delay-Doppler Channels," [*IEEE Transactions on Vehicular Technology*]{}, vol. 68, no. 5, May 2019. J. Zak, “Finite Translations in Solid State Physics," [*Phy. Rev. Lett.*]{}, 19, pp. 1385 - 1387,1967. A. J. E. M. Janssen, “The Zak Transform: A Signal Transform for Sampled Time-Continuous Signals," [*Philips J. Res.*]{}, 43, pp. 23-69, 1988. S. K. Mohammed, “Time-Domain to Delay-Doppler Domain Conversion of OTFS Signals in Very High Mobility Scenarios," submitted paper, www.arXiv.org (arXiv:2006.12413 \[cs.IT\]), June 2020. P. A. Bello, “Characterization of Randomly Time-Variant Linear Channels," [*IEEE Trans. Comm. Syst.*]{}, vol. 11, pp. 360-393, 1963. D. N. C. Tse and P. Vishwanath, [*Fundamentals of Wireless Communication*]{}, Cambridge, U.K.: Cambridge Univ. Press, 2005. H. W. Kim, K. Kang, K. Lim and J. Y. Ahn, “Performance Analysis for Terrestrial Radio waveform for Control and Non-Payload Communication for Unmanned Aircraft Systems," [*IEEE International Conf. on Information and Comm. Tech. Convergence (ICTC’ 2016)*]{}, pp. 759-761, Oct. 2016. E. Haas, “Aeronautical Channel Modeling," [*IEEE Transactions on Vehicular Technology*]{}, vol. 51, no. 2, pp. 254-264, March 2002.
[^1]:
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Hydrodynamics analysis of waves in two-dimensional degenerate electron gas with the account of separate spin evolution is presented. The transverse electric field is included along with the longitudinal electric field. The Coulomb exchange interaction is included in the analysis. In contrast with the three-dimensional plasma-like mediums the contribution of the transverse electric field is small. We show the decrease of frequency of both the extraordinary (Langmuir) wave and the spin-electron acoustic wave due to the exchange interaction. Moreover, spin-electron acoustic wave has negative dispersion at the relatively large spin-polarization. Corresponding dispersion dependencies are presented and analyzed.'
author:
- 'Pavel A. Andreev'
title: Extraordinary waves in two dimensional electron gas with separate spin evolution and Coulomb exchange interaction
---
\[sec:level1\] Introduction
===========================
Collective excitations in spin-polarized electron gas are considered as waves propagating in a plasma-like medium. The quantum hydrodynamic (QHD) method [@MaksimovTMP; @1999], [@MaksimovTMP; @2001] and its generalization for the separate spin evolution (SSE) [@Andreev; @PRE; @15; @SEAW] are applied for the theoretical modeling of this system.
The SSE-QHD method is developed for the spin-polarized degenerate electron gas [@Andreev; @PRE; @15; @SEAW]. Majority of papers on this subject are focused on three-dimensional electron gas existing in magnetically ordered metals (see for instance [@Andreev; @AoP; @15; @SEAW], [@Andreev; @1512; @surface]). However, it can be applied to the spin-polarized electron-hole liquid in the regime of degenerate carriers. It can be done similarly to the analysis of the electron-positron plasmas [@Andreev; @PRE; @16]. Including difference of the effective masses of the electrons and holes or (and) different concentrations of the electrons and holes, we find more similarity between electron-hole liquid and electron-positron-ion plasmas. Approximately, the SSE-QHD may be applied for non-degenerate electron-hole liquid, but it requires a modified equation of state and account of the scattering processes leading to the damping of excitations.
The SSE-QHD allows to consider linear [@Andreev; @AoP; @15; @SEAW], [@Andreev; @PRE; @16] and nonlinear [@Andreev; @PoP; @16; @exchange] bulk spin-electron acoustic waves (SEAWs). It also allows to consider the surface SEAWs [@Andreev; @1512; @surface] and SEAWs in two-dimensional electron gas [@Andreev; @EPL; @16].
Spectrum of the plasma-like mediums is highly affected by the transverse electric field at the wave propagation perpendicular to the external field. The transverse electric field leads to formation of the extraordinary waves instead of the Langmuir or upper hybrid waves for the electrostatic regime [@Landau; @v10]. Same effect was demonstrated for the SEAWs [@Trukhanova; @1603]. Therefore, we pay special attention to the transverse electric field in two-dimensional electron gas.
The collective excitations in the two-dimensional electron gas with no account of the spin polarization have been studied for a long time [@Stern; @PRL; @67], [@Grimes; @PRL; @76], including the magnetoplasmons existing in the two-dimensional electron gas located in the external magnetic field [@Oji; @PRB; @86], [@Batke; @PRB; @86]. The spin-polarized two-dimensional electron gas exists in the magnetic semiconductors. The study of SEAWs in two-dimensional electron gas appeared by analogy with bulk properties of the three-dimensional electron gas [@Andreev; @PRE; @15; @SEAW], [@Andreev; @AoP; @15; @SEAW]. However, there is a relatively long history of theoretical and experimental study of the spin plasmons in he spin-polarized two-dimensional electron gas (see for instance [@Ryan; @PRB; @91], [@Voitenko; @JLTP; @95]). The spin plasmons show dispersion and physical mechanism to the SEAWs. This similarity shows that described above linear and nonlinear, bulk and surface SEAWs can be considered as spin plasmons in three-dimensional electron gas.
It seems that all analysis of spin plasmons [@Ryan; @PRB; @91] is focused on the quantum wells in the magnetized semiconductors. Acoustic plasmon frequency is experimentally measured and the results are compared with the values calculated for spin polarized electrons and holes in p-type $A_{3}B_{5}$ semiconductors [@Voitenko; @JLTP; @95]. Spin flip waves and spin density fluctuations of a two-dimensional spin-polarized electron-gas in a semimagnetic $Cd_{1–0.008}Mn_{0.008}Te$ quantum well are considered in Ref. [@Perez; @PSS; @06]. A maximum value of the spin polarization degree of 35 percent is deduced from measurements. Collective excitations in the spin polarized quantum well were also considered in Refs. [@Marinescu; @PRB; @98], [@Yi; @Phys; @E; @00], [@Perez; @PRB; @09], [@Barate; @PRB; @10]. Recently, the contribution of Rashba spin-orbit coupling in the collective modes in two- and three-dimensional electron systems considered in Refs. [@Ullrich; @PRB; @02], [@Maiti; @PRB; @15].
Large contribution of the exchange interaction in the two-dimensional electron gas of GaAs microstructures is demonstrated experimentally [@Pinczuk; @PRL; @89] and theoretically within the local density approximation [@Bloss; @JAP; @89]. Hence, we consider the Coulomb exchange interaction contribution in spectrum of the Langmuir waves (plasmons) and the SEAWs (spin plasmons) in terms of the SSE-QHD. The influence of the exchange interaction on the collective effects in a quantum well was analyzed in Ref. [@Ryan; @PRB; @91; @b].
\[sec:level1\] Model
====================
We apply the SSE-QHD model developed in Ref. [@Andreev; @PRE; @15; @SEAW], [@Andreev; @AoP; @15; @SEAW] and adopted for two-dimensional systems in Ref. [@Andreev; @EPL; @16]. Since we are interested in analysis of the effects of Coulomb exchange interaction we use generalization of the SSE-QHD containing the exchange interaction [@Andreev; @PoP; @16; @exchange]. However, Ref. [@Andreev; @PoP; @16; @exchange] contains the exchange interaction for the three-dimensional electron gas. We need to substitute it in the similar way by the exchange interaction in two-dimensional electron gas in accordance with Refs. [@Datta; @JAP; @83], [@Andreev; @AoP; @14; @exchange]. We also use the full integral Maxwell equations for the expressions of the electric and magnetic fields to consider the contribution of the transverse electric field in the waves in magnetized spin-polarized electron gas. Overall we have the following set of four hydrodynamic equations: the continuity equations $$\label{2DExWsse cont eq electrons spin UP}
\partial_{t}n_{s}+\nabla(n_{s}\textbf{v}_{s})=\pm\frac{\gamma}{\hbar}(S_{x}B_{y}-S_{y}B_{x}),$$ and the Euler equations $$mn_{s}(\partial_{t}+\textbf{v}_{s}\nabla)\textbf{v}_{s}+\nabla p_{s}$$ $$\label{2DExWsse Euler eq spin UP with vel} =q_{e}n_{s}\biggl(\textbf{E}+\frac{1}{c}[\textbf{v}_{s},\textbf{B}]\biggr)+\textbf{F}_{SS,s},$$ with the force field of spin-spin interaction $$\textbf{F}_{SS,s}= \pm\gamma_{e}n_{s}\nabla B_{z} +\frac{\gamma_{e}}{2}(S_{x}\nabla B_{x}+S_{y}\nabla B_{y})$$ $$\label{2DExWsse} \pm\frac{m\gamma_{e}}{\hbar}[(\textbf{J}_{(M)x}-\textbf{v}_{s}S_{x})B_{y}-(\textbf{J}_{(M)y}-\textbf{v}_{s}S_{y})B_{x}],$$ containing the spin current $$\label{2DExWsse Spin current x} \textbf{J}_{(M)\alpha}=\frac{1}{2}(\textbf{v}_{u}+\textbf{v}_{d})S_{\alpha}-\varepsilon_{\alpha\beta z}\frac{\hbar}{4m} \biggl(\frac{\nabla n_{u}}{n_{u}}-\frac{\nabla n_{d}}{n_{d}}\biggr)S_{\beta},$$ and the effective pressure $$\label{2DExWsse Eq State partial} p_{s}=\pi\hbar^{2}n_{s}^{2}/m-\zeta \frac{8\beta}{3\pi \sqrt{\pi}}q_{e}^{2}n_{d}^{\frac{3}{2}}\delta_{sd},$$ where we introduce $\beta\equiv 24
\textrm{arsh}1=24\ln(1+\sqrt{2})=21.153$ and $$\label{EXCHANGE} \zeta=1-\frac{(1-\eta)^{3/2}}{(1+\eta)^{3/2}}.$$
The electromagnetic field is caused by the charges, electric currents and magnetic moments (spin) $\textbf{E}=-\nabla
\varphi-\partial_{t}\textbf{A}/c$, $\textbf{B}=\textbf{B}_{ext}+\nabla\times
\textbf{A}_{\textbf{j}}+\textbf{B}_{spin}$, where $$\label{2DExWsse } \varphi(\textbf{r},t)= \sum_{s=u,d} q_{e}\int\frac{n_{s}(\textbf{r}',t-|\textbf{r}-\textbf{r}'|/c)}{|\textbf{r}-\textbf{r}'|}d\textbf{r}',$$ $$\label{2DExWsse } \textbf{A}_{\textbf{j}}(\textbf{r},t)= \sum_{s=u,d} \frac{q_{e}}{c}\int\frac{\textbf{j}_{s}(\textbf{r}',t-|\textbf{r}-\textbf{r}'|/c)}{|\textbf{r}-\textbf{r}'|}d\textbf{r}',$$ and $$\label{2DExWsse} \textbf{B}_{spin}(\textbf{r},t)=\int [(\textbf{M}(\textbf{r}',t')\nabla)\nabla-\textbf{M}(\textbf{r}',t')\triangle]\frac{1}{\mid \textbf{r}-\textbf{r}'\mid}d\textbf{r}',$$ where $\textbf{M}(\textbf{r}',t')=\textbf{M}(\textbf{r}',t-|\textbf{r}-\textbf{r}'|/c)$ is the magnetization existed in point with coordinate $\textbf{r}'$ in an earlier moment of time $t'$, with $d\textbf{r}'=dx'dy'$ is the element of volume in 2D space.
\[sec:level1\] Linearized equations
===================================
We consider the propagation of plane waves along the Ox direction: $\textbf{k}=\{ k_{x}, 0, 0\}$. In the longitudinal waves (the electrostatic waves), the perturbation of electric field is parallel to the direction of wave propagation. However, we consider the longitudinally-transverse waves. Hence, we include the electric field perturbation along the Oy direction: $\textbf{E}=\{ E_{x}, E_{y}, 0\}$.
$$\label{2DExWsse cont lin 2 fl} -\imath\omega\delta n_{s}+\imath k_{x} n_{0s}\delta v_{sx}=0,$$
$$\label{2DExWsse Euler lin 2 fl} -\imath\omega mn_{0s}\delta \textbf{v}_{s}+\imath \textbf{k}\delta p_{s}=q_{e}n_{0s}\delta \textbf{E}+mn_{0s}\Omega_{e}[\delta \textbf{v}_{s}, \textbf{e}_{z}],$$
where $\Omega_{e}=q_{e}B_{0}/mc$ is the cyclotron frequency. We do not include the spin-spin interaction force field in the Euler equation as a small effect.
The perturbation of electric field has the following connection with the perturbation of concentration and the velocity field $$\delta \textbf{E}=-q_{e}\sum_{s=u,d}\biggl(\nabla\int\frac{\delta n_{s}(\textbf{r}',t-|\textbf{r}-\textbf{r}'|/c)}{|\textbf{r}-\textbf{r}'|}d\textbf{r}'$$ $$\label{2DExWsse } +\frac{n_{0s}}{c^{2}}\partial_{t}\int\frac{\delta\textbf{v}_{s}(\textbf{r}',t-|\textbf{r}-\textbf{r}'|/c)}{|\textbf{r}-\textbf{r}'|}d\textbf{r}' \biggr).$$
As it is demonstrated in Appendix for plane waves, the perturbation of the electric field can be presented in the following form: $$\label{2DExWsse } \delta \textbf{E}=\imath q_{e}\sum_{s=u,d}\biggl(-\Im \textbf{k}\delta n_{s}
+\Im \frac{\omega}{c^{2}}n_{0s}\delta
\textbf{v}_{s}\biggr).$$
![\[2DExWsse\_SF\_01\] (Color online) The figure shows the dimensionless dispersion dependence $\omega/\omega_{0}$ on the dimensionless wave vector $\kappa=k/\sqrt{n_{0}}$, where $\omega_{0}^{2}=2\pi e^{2}n_{0e}^{3/2}/m$ is a constant giving a characteristic frequency. Parameters are presented in the figure.](2DExWsse_SF_01.eps){width="8cm"}
![\[2DExWsse\_SF\_01\] (Color online) The figure shows a considerable change of the dispersion dependence at the small wave vectors due to the transverse electric field of wave. We present the dimensionless frequency $\xi$ as a function of the natural logarithm of the dimensionless wave vector $\ln\kappa$.](2DExWsse_SF_02.eps){width="8cm"}
As we will see below the result of application of equations (\[2DExWsse cont lin 2 fl\]) and (\[2DExWsse Euler lin 2 fl\]) is rather complicate. Thus, as an intermediate step we consider the extraordinary waves in 2DEG in terms of the single fluid model of electrons.
Single fluid model of electrons and extraordinary wave dispersion
=================================================================
Linearized set of quantum hydrodynamic equations for electrons considered as a single fluid is: $$\label{2DExWsse SF cont linear} -\imath\omega\delta n_{e}+\imath k_{x} n_{0e}\delta v_{ex}=0,$$ $$\label{2DExWsse } -\imath\omega n_{0e}\delta v_{ex}+\imath k_{x}U_{e}^{2}\delta n_{e}=\frac{q_{e}}{m}n_{0e}\delta E_{x}+n_{0e}\Omega_{e}\delta v_{ey},$$ $$\label{2DExWsse } -\imath\omega n_{0e}\delta v_{ey}=\frac{q_{e}}{m}n_{0e}\delta E_{y}-n_{0e}\Omega_{e}v_{ex},$$ where $$\label{2DExWsse } \delta E_{x}=q_{e}\Im\imath\biggl(-k_{x}\delta n_{e}+\frac{\omega}{c^{2}}n_{0e}\delta v_{x}\biggr),$$ $$\label{2DExWsse SF eq for Ey} \delta E_{y}=q_{e}\Im\imath \frac{\omega}{c^{2}}n_{0e}\delta v_{y},$$ and $U_{e}^{2}=(1+\eta^{2})\frac{\pi\hbar^{2}n_{0e}}{m_{e}^{2}}-(1+\eta)^{3/2}\zeta\frac{\beta\sqrt{2\pi}e^{2}}{\pi^{2}m_{e}}\sqrt{n_{0e}}$.
Equations (\[2DExWsse SF cont linear\])-(\[2DExWsse SF eq for Ey\]) lead to the following dispersion equation for the longitudinally-transverse waves in spin-polarized two-dimensional electron gas $$\omega^{2}\biggl(1+\frac{1}{\sqrt{1-\frac{\omega^{2}}{k^{2}c^{2}}}}\frac{\omega_{Le}^{2}}{k^{2}c^{2}}\biggr)-k^{2}U_{e}^{2}$$ $$\label{2DExWsse disp eq SF with TrF} -\frac{1}{\sqrt{1-\frac{\omega^{2}}{k^{2}c^{2}}}}\omega_{Le}^{2} =\frac{\Omega_{e}^{2}}{1+\frac{1}{\sqrt{1-\frac{\omega^{2}}{k^{2}c^{2}}}}\frac{\omega_{Le}^{2}}{k^{2}c^{2}}},$$ where we have used the two dimensional Langmuir frequency $$\label{2DExWsse Langmuir frq 2D}\omega_{Le}^{2}=\frac{2\pi e^2 k n_{0e}}{m}\sim k.$$
To drop the transverse electric field contribution we should consider the limit case $c\rightarrow\infty$. Thus, we find the dispersion dependence for the longitudinal waves in 2DEG in the single fluid model of electrons (see for instance [@Andreev; @AoP; @14; @exchange]) $$\label{2DExWsse disp Lang 2D with ex} \omega^{2}=\Omega_{e}^{2}+\omega_{Le}^{2}+U_{e}^{2}k^{2}.$$
In both regimes we find single branch of the dispersion dependence.
Considering the linear perturbations we can compare the contributions of the Fermi pressure and the exchange interaction. Their relative behavior can be described by the following dimensionless parameter: $\Lambda=\frac{\hbar^{2}}{me^{2}}\sqrt{n_{0e}}$ [@Andreev; @AoP; @14; @exchange]. More precisely, the Fermi pressure is larger then the Coulomb exchange interaction if the concentrations of electrons satisfy the following condition $\sqrt{n_{0d}}>\frac{2\zeta\sqrt{\pi}\beta}{\pi^{3}(1+\eta^{2})}\frac{me^{2}}{\hbar^{2}}$. It gives $n_{0d}(\eta=0.1)=6\times10^{16}$ cm$^{-2}$ or $n_{0d}(\eta=0.01)=7\times10^{15}$ cm$^{-2}$.
![\[2DExWsse\_SF\_01\] (Color online) The figure shows solutions of dispersion equation (\[2DExWsse Disp Eq two Fluids with Tr field\]) for regimes of two different concentrations at the fixed spin polarization. In each regime we find two solutions: the Langmuir wave and the SEAW.](2DExWsse_TF_01_diff_n.eps){width="8cm"}
![\[2DExWsse\_SF\_01\] (Color online) The figure shows dispersion dependencies arising as solutions of equation (\[2DExWsse Disp Eq two Fluids with Tr field\]) for regimes of two different spin polarizations at the same concentrations and external magnetic field. We assume that inner interaction with ions of the medium give dominant contribution to the spin polarization.](2DExWsse_TF_02_diff_h.eps){width="8cm"}
Numerical analysis shows that the change of the dispersion dependence due to the transverse electric field is very small, it arises in the regime of small wave vectors only. Formulae (\[2DExWsse disp Lang 2D with ex\]) and (\[2DExWsse disp eq SF with TrF\]) gives coinciding curves in Fig. 1. The small change of spectrum exists at the small wave vectors, as it is demonstrated in Fig. 2. Comparing two curves in Fig. 1 we see that the Coulomb exchange interaction gives considerable decrease of the dispersion dependence of the Langmuir waves or the extraordinary waves with the small contribution of the transverse electric field.
Dispersion dependence for two-fluid model of electrons
======================================================
After the calculation of electromagnetic potentials the linear SSE hydrodynamic equations (\[2DExWsse cont lin 2 fl\]), (\[2DExWsse Euler lin 2 fl\]) can be presented in a local form as a set of algebraic equations: $$\label{2DExWsse concentr perturb 2 fluids} \delta n_{s}=n_{0s}k_{x}\delta v_{xs}/\omega,$$ $$-\imath\omega mn_{0s}\delta v_{sx}+\imath k_{x} mU_{s}^{2}\delta n_{s}=q_{e}^{2}n_{0s}\imath\Im \biggl(-k_{x}\delta n_{u}-k_{x}\delta n_{d}$$ $$\label{2DExWsse v x perturb 2 fluids} +\frac{\omega}{c^{2}}n_{0u}\delta v_{ux}+\frac{\omega}{c^{2}}n_{0d}\delta v_{dx}\biggr)+mn_{0s}\Omega_{e}\delta v_{sy},$$ $$-\imath\omega mn_{0s}\delta v_{sy}=q_{e}^{2}n_{0s}\imath\Im \biggl(\frac{\omega}{c^{2}}n_{0u}\delta v_{uy}$$ $$\label{2DExWsse v y perturb 2 fluids} +\frac{\omega}{c^{2}}n_{0d}\delta v_{dy}\biggr) -mn_{0s}\Omega_{e}\delta v_{sx},$$ where $U_{s}^{2}=\frac{2\pi\hbar^{2}}{m^{2}}n_{0s}-\zeta\frac{4\beta
e^{2}}{\pi\sqrt{\pi}m_{e}}\sqrt{n_{0d}}\delta_{sd}$.
Substituting formula (\[2DExWsse concentr perturb 2 fluids\]) to equations (\[2DExWsse v x perturb 2 fluids\]) and (\[2DExWsse v y perturb 2 fluids\]) we obtain a set of four uniform algebraic equations. This set has a nonzero solution if its determinant is equal to zero. It leads to the following dispersion equation:
$$\biggl(1+\frac{\omega_{Ru}^{2}+\omega_{Rd}^{2}}{k^{2}c^{2}}\biggr)\biggl[(\omega^{2}-k^{2}U_{u}^{2})(\omega^{2}-k^{2}U_{d}^{2})
-[\omega_{Ru}^{2}(\omega^{2}-k^{2}U_{d}^{2})+\omega_{Rd}^{2}(\omega^{2}-k^{2}U_{u}^{2})]\biggl(1-\frac{\omega^{2}}{k^{2}c^{2}}\biggr)\biggr]$$ $$\label{2DExWsse Disp Eq two Fluids with Tr field} -\Omega_{e}^{2}\Biggl\{\biggl[\omega^{2}-k^{2}U_{u}^{2}-\omega_{Ru}^{2}\biggl(1-\frac{\omega^{2}}{k^{2}c^{2}}\biggr)\biggr]\biggl(1+\frac{\omega_{Ru}^{2}}{k^{2}c^{2}}\biggr)
+\biggl[\omega^{2}-k^{2}U_{d}^{2}-\omega_{Rd}^{2}\biggl(1-\frac{\omega^{2}}{k^{2}c^{2}}\biggr)\biggr]\biggl(1+\frac{\omega_{Rd}^{2}}{k^{2}c^{2}}\biggr)\Biggr\}+\Omega_{e}^{4}=0,$$
where $\omega_{Rs}^{2}=q_{e}^{2}n_{0s}k^{2}\Im/m=\omega_{Ls}^{2}/\sqrt{1-\omega^{2}/k^{2}c^{2}}$ is the retarding Langmuir frequency.
In the electrostatic limit ($c\rightarrow\infty$) we can drop the transverse electric field contribution in equation (\[2DExWsse Disp Eq two Fluids with Tr field\]) and find the following dispersion equation: $$\label{2DExWsse Disp Eq two Fluids longit} (\omega^{2}-\omega_{Lu}^{2}-\Omega_{e}^{2}-k^{2}U_{u}^{2})
(\omega^{2}-\omega_{Ld}^{2}-\Omega_{e}^{2}-k^{2}U_{d}^{2})
-\omega_{Lu}^{2}\omega_{Ld}^{2}=0.$$ It was derived in Ref. [@Andreev; @EPL; @16] (the dispersion equation is not presented in the paper, but its solution is presented by formula 7), it is also similar to results of earlier Ref. [@Ryan; @PRB; @91], starting from the SSE-QHD in the electrostatic regime.
Both equations (\[2DExWsse Disp Eq two Fluids with Tr field\]) and (\[2DExWsse Disp Eq two Fluids longit\]) have two solutions describing two branches of the wave dispersion dependence: the Langmuir (plasmon) mode and the spin-electron acoustic mode (spin-plasmon mode). Equation (\[2DExWsse Disp Eq two Fluids with Tr field\]) describes these waves in the regime of longitudinally-transverse waves. In this case, similarly to three-dimensional plasma-like mediums, we can call these wave the extraordinary waves.
At the account of the SSE we find numerically negligibly small contribution of the transverse electric field. Hence, equations (\[2DExWsse Disp Eq two Fluids with Tr field\]) and (\[2DExWsse Disp Eq two Fluids longit\]) give coinciding curves in Figs. 3, 4.
In Fig. 3 we see that at larger concentrations the dimensionless frequency of the Langmuir waves (the pair of upper curves) has smaller value and growth rate. The SEAWs (the pair of lower curves) have considerable smaller dimensionless frequencies at the larger concentration, but the growth rate becomes larger at the larger concentration.
As we see from Fig. 4 the increase of the spin polarization leads to the decrease of the frequencies of both waves. This effect reveals itself more in the SEAW spectrum. The exchange interaction decreases the SEAW frequency down to negative group velocity. Further increase of the spin polarization decreases the area of SEAW existence since its frequency goes down to zero value at the accessible wave vectors $\kappa <1$.
\[sec:level1\] Conclusion
=========================
We have considered waves in the magnetized two-dimensional electron gas located in the external constant uniform magnetic field directed perpendicular to the plane. We have payed attention to both the longitudinal and the transverse parts of the electric field of the wave perturbation propagating in the spin-polarized degenerate two-dimensional electron gas. Considering all electrons as the single fluid we find one wave solution: the extraordinary wave or the hybrid wave in the electrostatic limit. It was recently demonstrated that the account of the SSE in 2DEG with equilibrium spin polarization leads to the second branch: the SEAW. It was done in the longitudinal (the electrostatic limit) regime. Change of the dispersion dependence of the SEAWs at the account of the transverse electric field contribution in the wave propagation which forms the extraordinary SEAW has been demonstrated.
The influence of the Coulomb exchange interaction on the propagation of waves in 2DEG has been studied in both regimes: the single fluid model of electrons and the SSE-QHD.
Acknowledgements
================
Work of P.A. is supported by the Russian Foundation for Basic Research (grant no. 16-32-00886) and by the Dynasty foundation.
\[sec:level1\] Appendix: Integrals for potentials of electromagnetic field
==========================================================================
$$\label{2DExWsse } \delta \varphi=q_{e}\sum_{s=u,d}\int\frac{\delta n_{s}(\textbf{r}',t-|\textbf{r}-\textbf{r}'|/c)}{|\textbf{r}-\textbf{r}'|}d\textbf{r}'$$
$$\label{2DExWsse } \delta \textbf{A}=\frac{q_{e}}{c}\sum_{s=u,d}n_{0s}\int\frac{\delta \textbf{v}_{s}(\textbf{r}',t-|\textbf{r}-\textbf{r}'|/c)}{|\textbf{r}-\textbf{r}'|}d\textbf{r}'$$
We consider the plane wave perturbations $\delta f=F
e^{-\imath\omega t+\imath \textbf{k} \textbf{r}}$ which leads to $$\int\frac{\delta f(\textbf{r}',t-|\textbf{r}-\textbf{r}'|/c)}{|\textbf{r}-\textbf{r}'|}d\textbf{r}'=$$ $$=F\int\frac{e^{(-\imath\omega(t-|\textbf{r}-\textbf{r}'|/c)+\imath \textbf{k}\textbf{r}')}}{|\textbf{r}-\textbf{r}'|}d\textbf{r}'=$$ $$=F e^{-\imath\omega t+\imath \textbf{k}\textbf{r}} \int\frac{e^{(\imath\omega(|\textbf{r}-\textbf{r}'|/c)+\imath \textbf{k}(\textbf{r}'-\textbf{r}))}}{|\textbf{r}-\textbf{r}'|}d\textbf{r}'=$$ $$\label{2DExWsse }=\delta f \int e^{\imath\omega\xi/c+\imath k\xi\cos\varphi} d\varphi d\xi=\Im\delta f ,$$ where $$\label{2DExWsse Im def}\Im\equiv\int e^{\imath\omega\xi/c+\imath k\xi\cos\varphi} d\varphi d\xi.$$ As the result of calculation of integral (\[2DExWsse Im def\]) we find $$\label{2DExWsse }\Im=\frac{2\pi}{k}\frac{1}{\sqrt{1-\frac{\omega^{2}}{k^{2}c^{2}}}}.$$ Substituting our results in the expressions for $\delta\varphi$ and $\delta \textbf{A}$ we have $$\label{2DExWsse } \delta\varphi=q_{e}\Im (\delta n_{u}+\delta n_{d}),$$ $$\label{2DExWsse } \delta \textbf{A}=\frac{q_{e}}{c}\Im (n_{0u}\delta \textbf{v}_{u}+n_{0d}\delta \textbf{v}_{d}).$$
[17]{}
L. S. Kuz’menkov, S. G. Maksimov, Theor. Math. Phys. **118**, 227 (1999).
L. S. Kuz’menkov, S. G. Maksimov, V. V. Fedoseev, Theor. Math. Phys. **126**, 110 (2001).
P. A. Andreev, Phys. Rev. E **91**, 033111 (2015). P. A. Andreev, L. S. Kuz’menkov, Annals of Physics **361**, 278 (2015).
P. A. Andreev, L. S. Kuz’menkov, arXiv:1512.07940.
P. A. Andreev, Z. Iqbal, Phys. Rev. E **93**, 033209 (2016).
P. A. Andreev, Phys. Plasmas **23**, 012106 (2016).
P. A. Andreev, L. S. Kuz’menkov, EPL **113**, 17001 (2016).
L. Landau and E. M. Lifshitz, *Statistical Physics, part II* (Pergamon, New York, 1980).
P. A. Andreev, M. Iv. Trukhanova, arXiv:1603.07506.
F. Stern, Phys. Rev. Lett. **8**, 546 (1967).
C. C. Grimes and G. Adams, Phys. Rev. Lett. **36**, 146 (1976).
H. C. A. Oji and A. H. MacDonald, Phys. Rev. B **33**, 3810 (1986).
E. Batke, D. Heitmann, and C.W. Tu, P hys. Rev. B **34**, 6951 (1986).
J. C. Ryan, Phys. Rev. B **43**, 4499 (1991).
V. A. Voitenko, B. H. Bairamov, V. V. Toporov, V. K. Negoduiko, Journal of Low Temperature Physics **99**, 121 (1995).
F. Perez, B. Jusserand, D. Richards, and G. Karczewski, phys. stat. sol. (b) 243, 873 (2006).
D. C. Marinescu and J. J. Quinn, Phys. Rev. B **58**, 13762 (1998).
K. S. Yi, S. P. Hong, D. C. Marinescu, J. J. Quinn, Physica E: Low-dimensional Systems and Nanostructures **6**, 802 (2000).
Florent Perez, Phys. Rev. B **79**, 045306 (2009).
P. Barate, S. Cronenberger, M. Vladimirova, D. Scalbert, F. Perez, J. Gomez, B. Jusserand, H. Boukari, D. Ferrand, H. Mariette, J. Cibert, and M. Nawrocki, Phys. Rev. B **82**, 075306 (2010).
C. A. Ullrich and M. E. Flatte, Phys. Rev. B **66**, 205305 (2002).
Saurabh Maiti, Vladimir Zyuzin, and Dmitrii L. Maslov, Phys. Rev. B **91**, 035106 (2015).
A. Pinczuk, S. Schmitt-Rink, G. Danan, J. P. Valladares, L. N. Pfeiffer, and K. W. West, Phys. Rev. Lett. **63**, 1633 (1989).
W. L. Bloss, J. Appl. Phys. **66**, 3639 (1989).
J. C. Ryan, Phys. Rev. B **43**, 12406 (1991).
S. Datta and R. L. Gunshor, J. Appl. Phys. **54**, 4453 (1983).
P. A. Andreev, Annals of Physics **350**, 198 (2014).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We propose the use of ponderomotive forces to entangle the motions of different atoms. Two situations are analyzed: one where the atoms belong to the same optical cavity and interact with the same radiation field mode; the other where each atom is placed in own optical cavity and the output field of one cavity enters the other.'
address: |
$^1$INFM, Dipartimento di Fisica, Università di Milano, Via Celoria 16, I-20133 Milano, Italy\
$^2$Centre for Quantum Computation, Clarendon Laboratory, University of Oxford, Parks Road, Oxford OX1 3PU, England
author:
- Stefano Mancini$^1$ and Sougato Bose$^2$
title: Ponderomotive entangling of atomic motions
---
[2]{}
Introduction
============
The preparation of entangled atomic states is one of the goals of atomic physics and quantum optics. These states are the key ingredients for studying some fundamental issues of quantum mechanics [@BELL], as well as for certain applications related to quantum information [@BEDI]. Various methods have been recently proposed to engineer entanglement between atoms [@HAR; @ATOMS; @PAR]. They are based on achieving and controlling an effective interaction between the atoms that are to be entangled. Typically, these interactions are mediated by the electromagnetic field, but also involve transitions between internal atomic states.
On the other hand, atoms and radiation fields can interact via radiation pressure effects. The role of such [*ponderomotive effects*]{} in probing fundamental aspects of quantum theory has been pointed out in a number of recent papers [@OUR1; @OUR2]. Here we shall exploit ponderomotive effects to propose a method of entangling atomic motions. This is fundamentally important as radiation pressure effects are universal and scalable. For example, radiation pressure effects could also be fruitful in entangling massive particles (or even macroscopic objects). If such a scheme happens to be successful for atoms, one would acquire the confidence of trying out a similar scheme with more massive objects. Another advantage of a ponderomotive scheme becomes clear when we note that most existing mechanisms for entangling atoms, apart from a very few [@PAR], entangle internal atomic states. As such, the maximum degree of entanglement attainable is limited by the finite dimensionality of the Hilbert space of the internal states of the interacting atoms. Our scheme, on the other hand, entangles atomic motions and therefore entangles two continuous variable (infinite dimensional) systems. Moreover, as we shall demonstrate, our entangling mechanism, in contrast to most others, does not require carefully controlled switching on and off of external laser fields acting on the individual atoms.
As Hamiltonian model, we shall consider the case of large detuning of internal atomic transitions from the cavity field so that spontaneous emission can be neglected and the upper atomic level can be adiabatically eliminated [@WALMIL]. In this case, the atom-field interaction reduces to the product between the number of photons and the amplitude of atomic displacement. The atomic internal states are never involved in the interaction and the atom can always stay in a fixed internal ground state.
Atoms in the same optical cavity
================================
We first consider two trapped atoms $A$ and $B$ in the same optical cavity. When they are invested by the (off-resonant) laser light, the evolution in one spatial direction takes place according to the Hamiltonian (in natural units) $$\label{H1}
H=\chi c^{\dag}c \left(a+a^{\dagger}+b+b^{\dagger}\right) +\Omega
(a^{\dagger}a + b^{\dagger} b) \,$$ where $a,b$ and $c$ are annihilation operators for the vibrational motion of the two atoms and the cavity mode respectively. Furthermore, $\Omega$ is the vibrational frequency (assumed equal for the two atoms). The time evolution operator corresponding to the Hamiltonian (\[H1\]) can be put in the following form [@OUR1] $$\begin{aligned}
\label{U}
U(t)&=&\exp\left[2i\kappa^2\left(c^{\dag}c\right)^2
\left(t-\sin t\right)\right]
\nonumber\\
&&\times\exp\left[\kappa c^{\dag}c
\left(\eta(t)a^{\dag}-\eta^*(t)a\right)\right]
\nonumber\\
&&\times\exp\left[\kappa c^{\dag}c
\left(\eta(t)b^{\dag}-\eta^*(t)b\right)\right]
\nonumber\\
&&\times\exp\left[-it\left(a^{\dag}a+b^{\dag}b\right)\right]
\,,\end{aligned}$$ where $\eta(t)=(1-e^{-it})$, $\kappa=\chi/\Omega$ and the time is scaled accordingly to $\Omega t\to t$.
Let us now assume that initially both atoms are cooled down to their ground states and the cavity field is in coherent states, that is $$\label{Psiini}
|\Psi(0)\rangle= |0\rangle_a
\otimes |0\rangle_b
\otimes|\zeta\rangle_c\,,$$ then, the time evolution leads to $$\begin{aligned}
\label{Psit}
|\Psi(t)\rangle&=&e^{-|\zeta|^2/2}
\sum_{n=0}^{\infty}\frac{\zeta^n}{\sqrt{n!}}
e^{2i\kappa^2 n^2(t-\sin t)}
\nonumber\\
&&|n\kappa\eta(t)\rangle_a
\otimes |n\kappa\eta(t)\rangle_b
\otimes|n\rangle_c\,,\end{aligned}$$ where $|n\kappa\eta(t)\rangle_{a,b}$ are coherent states of the atoms.
Let us suppose to measure the quadrature $X=(c+c^{\dag})/\sqrt{2}$. Then the state after the measurement will be $$\label{Proj}
|{\widetilde\Psi}(t)\rangle={\cal N}\,
|x\rangle_c {}_c\langle x|\Psi(t)\rangle\,,$$ where ${\cal N}$ is a normalization constant, while $|x\rangle$ are the eigenvectors of the quadrature observable $X$. The inverse of normalization constant also gives the probability amplitude of the outcome $x$, that is $P(x)={\cal N}^{-2}$.
The joint state of the atoms after the measurement results $$\begin{aligned}
\label{Psitil}
|\widetilde{\Psi}(t)\rangle&=&{\cal N}e^{-|\zeta|^2/2}
\sum_{n=0}^{\infty}
\frac{\zeta^n}{\sqrt{n!}}e^{2i\kappa^2n^2(t-\sin t)}
{}_c\langle x|n\rangle_c
\nonumber\\
&&|n\kappa\eta(t)\rangle_a
\otimes |n\kappa\eta(t)\rangle_b\,,\end{aligned}$$ where $$\label{xn}
{}_c\langle x|n\rangle_c=\pi^{-1/4}(2^n n!)^{-1/2}H_n(x)\exp(-x^2/2)\,,$$ are the harmonic oscillator position eigenstates, with $H_n$ the Hermite polynomials.
The state (\[Psitil\]) depends on the time at which the measurement is performed and is conditioned to the result $x$ of the measurement. Let us choose a time $t_N=(2N+1)\pi$ with $N\in{\bf N}$, so that $\eta(t_N)=2$, and $\kappa\in{\bf N}$, thus Eq.(\[Psitil\]) becomes $$\label{PsitN}
|\widetilde{\Psi}(t_N)\rangle={\cal N}
e^{-|\zeta|^2/2}\sum_{n=0}^{\infty}
\frac{\zeta^{n}}{\sqrt{n!}}
\,{}_c\langle x|n\rangle_c\,
|4\kappa n\rangle_a \otimes |4\kappa n\rangle_b\,.$$ It is clear that the above state represent an entangled state for the atoms. In practice, since the radiation field mediates information between the two atoms, a measurement of its quadrature leaves the atoms correlated.
In order to quantify the degree of entanglement we calculate the linear entropy [@JEX] $$\label{Elin}
E=1-{\rm Tr}_a\{[{\rm Tr}_b(\,\widetilde\rho\,)]^2\}\,,$$ where $\widetilde\rho=
|\widetilde\Psi(t_N)\rangle \langle\widetilde\Psi(t_N)|$. However, since the formation of entangled states is conditioned to the measurement on the radiation field, it would be useful to define the efficiency of the entanglement procedure as $$\label{Ups}
\Upsilon(x)=E(x)\times P(x)\,.$$ Then, in Fig.\[fig1\] we show the efficiency $\Upsilon$ as function of quadrature outcome $x$. We clearly see that the efficiency increases as the radiation pressure (i.e., the amplitude of the radiation field) increases. The shape of $\Upsilon$ comes from the fact that $x=0$ is the most probable outcome of the measurement, but the entanglement has a minimum at that value.
Note that by increasing $\kappa$ sufficiently, the set of states $\{|4\kappa n\rangle\}$ in Eq.(\[PsitN\]) can be made to approach an orthonormal basis arbitrarily close. This means that we can approach the maximal entanglement possible in any $D\times
D$ system arbitrarily closely by setting the field amplitude $\zeta$ to a required value and increasing $\kappa$. For example, for a $10\times 10$ system, the maximum entanglement according to the measure of Eq.(\[Elin\]) is $0.9$. This is already exceeded for a field amplitude $\zeta=6$. The above fact clearly illustrates one of the advantages of entangling through our scheme in contrast to entangling the internal levels of two two-level atoms as in most existing schemes. The degree of entanglement achievable is not bounded from above by any fundamental constraint. Of course, the degree of entanglement one can practically produce will depend on parameters such as the Q-factor of the cavity in a specific experimental realization.
Atoms in distinct optical cavities
==================================
We now consider the two (trapped) atoms placed in separate cavities, and interacting sequentially with the same radiation field. That is, the outgoing field from the first cavity enters the second one. Therefore, the Hamiltonian (\[H1\]) should be modified as follows: $$\begin{aligned}
\label{H2}
H&=&\chi c_1^{\dag}c_1(a+a^{\dag})
+\chi c_2^{\dag}c_2(b+b^{\dag})
\nonumber\\
&&+\Omega(a^{\dag}a+b^{\dag}b)
\,,\end{aligned}$$ where we have assumed the same coupling constant and the same oscillatory frequency for the two atoms.
However, in such a case we have to consider photon losses, which we assume to occour at same rate $\gamma$ in the two cavities. We further assume a decay of the atomic motions at rate $\Gamma$. Thus, we can write down the quantum Langevin equations as $$\begin{aligned}
{\dot a}&=&-i\Omega a-i\chi c_1^{\dag}c_1-\frac{\Gamma}{2}a
+\sqrt{\Gamma}a^{in}\,,
\label{QLE1}
\\
{\dot b}&=&-i\Omega b-i\chi c_2^{\dag}c_2-\frac{\Gamma}{2}b
+\sqrt{\Gamma}b^{in}\,,
\label{QLE2}
\\
{\dot c_1}&=&-i\Delta_1 c_1-i\chi c_1(a+a^{\dag})
-\frac{\gamma}{2}c_1+\sqrt{\gamma}c_1^{in}\,,
\label{QLE3}
\\
{\dot c_2}&=&-i\Delta_2 c_2-i\chi c_2(b+b^{\dag})
-\frac{\gamma}{2}c_2+\sqrt{\gamma}c_2^{in}\,,
\label{QLE4}\end{aligned}$$ where all the input operators represent vacuum noise [@GAR]. $\Delta_1$ and $\Delta_2$ are the cavity detunings. The boundary condition reads $$\label{BOU}
c_2^{in}\equiv c_1^{out}=\sqrt{\gamma}c_1-c_1^{in}\,.$$
To solve the system of Eqs.(\[QLE1\]), (\[QLE2\]), (\[QLE3\]), (\[QLE4\]), we proceede by the linearization around the steady state. The latter is characterized by $$\begin{aligned}
\label{zeta12}
&&\zeta_{j}\left\{\frac{\gamma}{2}+\frac{\chi^2|\zeta_{j}|^2\Gamma}
{\frac{\Gamma^2}{4}+\Omega^2}\right.
\nonumber\\
&&\qquad\left.
+i\left[\Delta_{j}-2\frac{\chi^2|\zeta_{j}|^2\Omega}{\frac{\Gamma^2}{4}
+\Omega^2}\right]\right\}
=\sqrt{\gamma}\zeta_{j}^{in}\,;
j=1,2\,,\end{aligned}$$ with the relation $$\label{zetabou}
\zeta_2^{in}\equiv\zeta_1^{out}=\sqrt{\gamma}\zeta_1-\zeta_1^{in}\,.$$ Here, $\zeta$ is the steady state value of the operators $c$, and $\zeta_1^{in}$ is the amplitude of the input field (at the first cavity). Eq.(\[zeta12\]) shows a typical bistable behavior [@MT94]. Furthermore, the steady states of atomic operators $a$ and $b$ are given by $$\begin{aligned}
\alpha&=&-i\frac{\chi}{\frac{\Gamma}{2}+i\Omega}|\zeta_1|^2\,,
\label{al}
\\
\beta&=&-i\frac{\chi}{\frac{\Gamma}{2}+i\Omega}|\zeta_2|^2\,.
\label{be}\end{aligned}$$
The linearized system of equations can be written in the frequency domain as $$\label{matrixeq}
i\omega \tilde{v}={\cal M}\tilde{v}+\tilde{v}_{in}$$ where the transposed vectors $\tilde{v}^{T}$ and $\tilde{v}_{in}^{T}$ are given by $$\label{vvec}
\tilde{v}^{T}=(
\tilde{a},
\tilde{a}^{\dag},
\tilde{b},
\tilde{b}^{\dag},
\tilde{c}_1,
\tilde{c}_1^{\dag},
\tilde{c}_2,
\tilde{c}_2^{\dag})\,,$$ and $$\begin{aligned}
\label{vinvec}
\tilde{v}_{in}^{T}&=&(
\sqrt{\Gamma}\tilde{a}_{in},
\sqrt{\Gamma}\tilde{a}_{in}^{\dag},
\sqrt{\Gamma}\tilde{b}_{in},
\sqrt{\Gamma}\tilde{b}_{in}^{\dag},
\nonumber\\
&&\sqrt{\gamma}\tilde{c}^{in}_1,
\sqrt{\gamma}\tilde{c}_1^{in\,\dag},
-\sqrt{\gamma}\tilde{c}^{in}_2,
-\sqrt{\gamma}\tilde{c}_2^{in\,\dag})\,,\end{aligned}$$ where now all the operators represent small quantum fluctuations around steady state. Moreover, the $8\times 8$ matrix ${\cal M}$ is $$\label{calM}
{\cal M}=\left[
\begin{array}{cc}
M_{I}&M_{II}
\\
M_{III}&M_{IV}
\end{array}
\right]$$ with
$$\label{MI}
M_{I}=\left[
\begin{array}{cccc}
-\left(\frac{\Gamma}{2}+i\Omega\right)&0&0&0
\\
0&-\left(\frac{\Gamma}{2}-i\Omega\right)&0&0
\\
0&0&-\left(\frac{\Gamma}{2}+i\Omega\right)&0
\\
0&0&0&-\left(\frac{\Gamma}{2}-i\Omega\right)
\end{array}
\right]\,,$$
$$\label{MIIMIII}
M_{II}=i\chi\left[
\begin{array}{cccc}
-\zeta_1^*&-\zeta_1&0&0
\\
\zeta_1^*&\zeta_1&0&0
\\
0&0&-\zeta_2^*&-\zeta_2
\\
0&0&\zeta_2^*&\zeta_2
\end{array}
\right]\,;
\quad
M_{III}=i\chi\left[
\begin{array}{cccc}
-\zeta_1&-\zeta_1&0&0
\\
\zeta_1^*&\zeta_1^*&0&0
\\
0&0&-\zeta_2&-\zeta_2
\\
0&0&\zeta_2^*&\zeta_2^*
\end{array}
\right]\,,$$
$$\label{MIV}
M_{IV}=\left[
\begin{array}{cccc}
-\frac{\gamma}{2}
-i\left[\Delta_1+\chi\left(\alpha+\alpha^*\right)\right]&0&0&0
\\
0&-\frac{\gamma}{2}
+i\left[\Delta_1+\chi\left(\alpha+\alpha^*\right)\right]&0&0
\\
\gamma&0&-\frac{\gamma}{2}
-i\left[\Delta_2+\chi\left(\beta+\beta^*\right)\right]&0
\\
0&\gamma&0&-\frac{\gamma}{2}
+i\left[\Delta_2+\chi\left(\beta+\beta^*\right)\right]
\end{array}
\right]\,.$$
[2]{}
The solution of the Eq.(\[matrixeq\]) can formally be written as $$\label{matrixsol}
\tilde{v}(\omega)=
\left[i\omega{\cal I}-{\cal M}\right]^{-1} \tilde{v}_{in}(\omega)\,,$$ where ${\cal I}$ is the $8\times 8$ identity matrix. Then, the various frequency correlations can be easily calculated by using the correlations of the vacuum input noise [@GAR]. These should deserve to quantify the entanglement of atomic motions. Nevertheless, since we deal with non pure states, it is very difficult to quantify the degree of entanglement [@PLENIO]. To reach this goal we shall proceed as follows.
We first introduce the dimensionless atomic position and momentum variable $$\begin{aligned}
\label{qpab}
q_{a}&=&(a+a^{\dag})\,,q_{b}=(b+b^{\dag})\,,
\\
p_{a}&=&-i(a-a^{\dag})\,,p_{b}=-i(b-b^{\dag})\,,\end{aligned}$$ Now, if the atoms are entangled, one could [*infer*]{} position or momentum of one atom through position or momentum of the other [@REID]. The errors of these inferences are then quantified by the variances $\langle (q_a+q_b)^2 \rangle$ and $\langle
(p_a-p_b)^2 \rangle$. Once the product of these inference errors lies below the limit of the Heisenberg principle, i.e. $\langle
(q_a+q_b)^2 \rangle\dot\langle (p_a-p_b)^2 \rangle \le
|\langle[q_a,p_a]\rangle|^2/4$, an EPR-like paradox arises [@EPR]. This is a typical manifestation of the existence of purely quantum correlations between the two systems [@REID]. It is known that when $\langle (q_a+q_b)^2 \rangle\dot\langle
(p_a-p_b)^2 \rangle$ is less than $1$ [@cont], the state is entangled [*irrespective*]{} of whether it is pure or mixed. Though the criteria for the presence of entanglement in continuous variable systems is present [@cont; @simon], and we use this to prove the presence of entanglement in our case, there is as yet no rigorously proved measure of continuous variable entanglement. We shall use a quantity motivated from the above separability criterion to evaluate the degree of entanglement.
Now, given an operator $\tilde{\cal O}(\omega)$ in the frequency domain, we define the hermitian operator ${\cal R}_{\{\tilde{\cal O}\}}(\omega)=
[\tilde{\cal O}(\omega)+\tilde{\cal O}(-\omega)]/2$. Then, recalling the previous argument, we can define the degree of entanglement as $$\label{Epseudo}
E(\omega)=\frac{
\langle {\cal R}^2_{\{\tilde{q}_{a}+\tilde{q}_{b}\}}(\omega) \rangle
\,
\langle {\cal R}^2_{\{\tilde{p}_{a}-\tilde{p}_{b}\}}(\omega) \rangle}
{\frac{1}{4} \left|
\langle \left[{\cal R}_{\tilde{q}_a}(\omega),
{\cal R}_{\tilde{p}_a}(\omega)\right]
\rangle\right|^2}\,,$$ which can be considered as a signature of entanglement whenever it goes below one. Notice that this condition is much stronger than the simple entanglement requirement, in that $E<1$ requires EPR-type correlations.
In Fig.\[fig2\] we show the degree of entanglement (\[Epseudo\]) as function of the input field amplitude. The sharp decrement is due to the jump from one to the other branch of the bistable curve [@MT94]. Then, by increasing the radiation intensity, the entanglement tends to disappear because of the increasing radiation pressure noise. In case of small coupling constant, entanglement effect appears at higher amplitudes (dashed line). Notice that we have used nonzero detunings to establish correlations between $q$ and $p$ variables, while naturally only $q$ variables tend to couple as results from Hamiltonian (\[H2\]).
Conclusions
===========
In this paper we have discussed how to exploit radiation pressure effects to entangle the motions of two atoms. We have considered two scenarios: atoms in same and in separate cavities. As radiation pressure effects are very generic, the scheme should lead to those for more macroscopic objects. It also has the advantage of not involving either atomic internal states or various laser pulses being applied to the atoms. It is precisely this fact that offers the generality of our scheme, in the sense that it should not depend on the specific internal configuration of the atoms. Moreover, the entanglement between atoms in separate cavities is generated in the steady state ([*i.e.*]{} after all types of decoherence and dissipation have acted). This means that this type entanglement generating mechanism is robust in nature. Further generalizations of this scheme for entangling several atoms interacting with a common cavity field would be interesting and could potentially provide a simple way for generating multiparticle Schroedinger cat states.
Acknowledgements {#acknowledgements .unnumbered}
================
S. M. gratefully acknowledges financial support from Università di Camerino, Italy, under the Project ‘Giovani Ricercatori’.
J. S. Bell, Physics [**1**]{}, 195 (1965); J. F. Clauser, M. A. Horne, A. Shimony and R. A. Holt, Phys. Rev. Lett. [**23**]{}, 880 (1969).
C. H. Bennett, Phys. Today [**48**]{}(10), 24 (1995); D. P. DiVincenzo, Science [**270**]{}, 255 (1995).
E. Hagley, X. Mantre, G. Nogues, C. Wunderlich, M. Brune, J. M. Raimond and S. Haroche, Phys. Rev. Lett. [**79**]{}, 1 (1997); C. A. Sackett, D. Kielpinski, B. E. King, C. Langer, V. Meyer, C. J. Myatt, M. Rowe, Q. A. Turchette, W. M. Itano and D. J. Wineland, Nature [**404**]{}, 256 (2000).
J. I. Cirac and P. Zoller, Phys. Rev. A [**50**]{}, R2799 (1994); T. Pellizzari, S. Gardiner, J. I. Cirac and P. Zoller, Phys. Rev. Lett. [**75**]{}, 3788 (1995); J. I. Cirac and P. Zoller, Phys. Rev. Lett. [**74**]{}, 4091 (1995); J. F. Poyatos, J. I. Cirac and P. Zoller, Phys. Rev. Lett. [**81**]{}, 1322 (1998); C. Cabrillo, J. I. Cirac, P. Garcia-Fernandez and P. Zoller, Phys. Rev. A [**59**]{}, 1025 (1999); M.B. Plenio, S.F. Huelga, A. Beige and P.L. Knight, Phys. Rev. A [**59**]{} (1999) 2468; S. Bose, P. L. Knight, M. B. Plenio and V. Vedral, Phys. Rev. Lett. [**83**]{}, 5158 (1999); A. Kuzmich and E. S. Polzik, Phys. Rev. Lett. [**85**]{}, 5639 (2000); A. Beige, W. J. Munro and P. L. Knight, Phys. Rev. A [**62**]{} 2102, (2000); S. Bose and D. Home, quant-ph/0101093.
A. S. Parkins, J. Opt. B.: Quant. Semiclass. Opt. [**3**]{}, S18 (2001); A. S. Parkins and H. J. Kimble, Phys. Rev. A [**61**]{}, 052104 (2000).
S. Mancini, V. I. Man’ko and P. Tombesi, Phys. Rev. A [**55**]{}, 3042 (1997); S. Bose, K. Jacobs and P. Knight, Phys. Rev. A [**56**]{}, 4175 (1997).
S. Bose, K. Jacobs and P. Knight, Phys. Rev. A [**59**]{}, 3204 (1999); S. Mancini, D. Vitali and P. Tombesi, Phys. Rev. Lett. [**80**]{}, 688 (1998); S. Mancini, Phys. Lett. A [**279**]{}, 1 (2001); S. Mancini and A. Gatti, J. Opt. B.: Quant. Semiclass. Opt. [**3**]{}, S66 (2001); V. Giovannetti, S. Mancini and P. Tombesi, Europhys. Lett. (to appear).
D. F. Walls and G. J. Milburn, [*Quantum Optics*]{}, (Springer, Berlin, 1994), p.330.
G. Drobny, I. Jex and V. Buzek, Phys. Rev. A [**48**]{}, 569 (1993).
Gardiner, C. W. [*Quantum Noise*]{}, (Springer, Berlin, 1991).
S. Mancini and P. Tombesi, Phys. Rev. A [**49**]{}, 4055 (1994).
V. Vedral and M. B. Plenio, Phys. Rev. A [**57**]{}, 1619 (1998); S. Parker, S. Bose and M. B. Plenio, Phys. Rev. A [**61**]{}, 032305 (2000).
M. D. Reid and P. D. Drummond, Phys. Rev. Lett. [**60**]{}, 2731 (1988); M. Reid, Phys. Rev. A [**40**]{}, 913 (1989).
A. Einstein, B. Podolsky and N. Rosen, Phys. Rev. [**47**]{}, 777 (1935).
Lu-Ming Duan, G. Giedke, J. I. Cirac and P. Zoller, Phys. Rev. Lett. [**84**]{}, 2722 (2000).
R. Simon, Phys. Rev. Lett. 84, 2726 (2000).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A Theorem of Wang in [@Wa] implies that any holomorphic parallelism on a compact complex manifold $M$ is flat with respect to some complex Lie algebra structure whose dimension coincides with that of $M$. We study here rational parallelisms on complex manifolds. We exhibit rational parallelisms on compact complex manifolds which are not flat with respect to any complex Lie algebra structure.'
address:
- 'School of Mathematics, Tata Institute of Fundamental Research, Homi Bhabha Road, Mumbai 400005, India'
- 'Université Côte d’Azur, CNRS, LJAD, France'
author:
- Indranil Biswas
- Sorin Dumitrescu
title: Rational parallelism on complex manifolds
---
Introduction
============
A well-known result of Wang [@Wa] classifies compact complex manifolds $M$ with holomorphically trivial holomorphic tangent bundle $TM$.
Recall that a holomorphic trivialization of $TM$ is defined by a holomorphic one-form $\omega
\,\in\, \Omega^1 (M,\, V)$ with values in a complex vector space $V$ with $\dim
V\,=\,\dim_{\mathbb C} M$ such that $\omega_m \,:\, T_mM \,\longrightarrow\, V$ is a linear isomorphism at every $m \,\in\, M$. Hence $\omega$-constant vector fields define a parallelism on $M$ and trivialize $TM$. A compact complex manifold with trivial tangent bundle is called (holomorphically) [*parallelizable.*]{}
In this context Wang proved that there exists a complex Lie algebra structure $\mathfrak L$ on $V$ such that $\omega$ realizes a Lie algebra isomorphism between $TM$ and $\mathfrak L$ (see Theorem \[Wang\] here). Notice that this is equivalent to the assertion that $\omega$, seen as Cartan geometry modeled on $\mathfrak L$, is [*flat*]{} ([@Sh], Chapter 3 and Chapter 5), meaning that $\omega$ satisfies the Maurer-Cartan equation with respect to the Lie algebra structure $\mathfrak L$. In view of Darboux-Cartan Theorem, [@Sh Chapter 3, p. 116], this implies that $M$ inherits a $(L,\,L)$-structure in the sense of Ehresmann–Thurston [@Eh] with $L$ being a complex connected Lie group with Lie algebra $\mathfrak L$. The compactness assumption on $M$ ensures that the $(L,\,L)$-structure is complete and $M$ is biholomorphic to a quotient of $L$ by a lattice in it. This result was extended by Winkelmann to certain open complex manifolds [@Wi1; @Wi2].
We study here rational parallelisms on compact complex manifolds given by holomorphic one-forms $\omega\,\in\, \Omega^1 (M,\, V)$ such that $\omega_m \,:\, T_mM \,\longrightarrow\,
V$ is a linear isomorphism for $m$ lying in an open dense subset $U\,=\,M \setminus S$ (the subset $S$ is an analytic divisor, see Proposition \[divisor\]). In this case, $\omega$–constant vector fields are actually meromorphic on $M$ with poles lying in $S$ and they holomorphically trivialize $TU$. This definition corresponds to the particular case of a [ *branched Cartan geometry*]{}, whose model is a Lie group, as introduced and studied by the authors in [@BD]. We prove here (see Theorem \[flat\]) that on compact complex manifolds $M$ which are either in the Fujiki class $\mathcal C$ (e.g. $M$ is the meromorphic image of a Kähler manifold [@Fu]), or of complex dimension two or of algebraic dimension zero (e.g. all meromorphic functions on $M$ are constant), all such rational parallelisms are flat with respect to some complex Lie algebra structure on $V$ (meaning $\omega$ is a Lie algebra isomorphism). For such a situation we deduce that the fundamental group of $M$ is infinite.
Flat rational parallelisms were studied in particular in [@BC].
The main result proved here (Theorem \[non flat\]) exhibits examples of rational parallelisms on compact complex manifolds which are non-flat with respect to any complex Lie algebra structure.
Holomorphic parallelisms
========================
In this section we recall the proof of Wang’s classification Theorem of (holomorphically) parallelizable manifolds. The idea of this proof will be useful later on in the study of flatness for rational parallelisms.
Let us first recall the following:
A holomorphic trivialization (parallelization) of the holomorphic tangent bundle $TM$ of a compact complex manifold $M$ of (complex) dimension $m$ is a holomorphic one-form $\omega\,\in\,
\Omega^1(M, \,V)$ with values in a complex vector space $V$ of dimension $m$ such that $\omega_m
\,:\, T_mM \,\longrightarrow\, V$ is a linear isomorphism at every $m \,\in\, M$.
\[Wang\] Let $M$ be a compact complex manifold of dimension $m$ and $$\omega \,:\, TM \,\longrightarrow\, V$$ a holomorphic trivialization of its holomorphic tangent bundle. Then the universal cover of $M$ is biholomorphic to a complex Lie group $L$, and the pull-back of $\omega$ on $L$ coincides with the Maurer-Cartan form of $L$. Consequently, $d \omega + \frac{1}{2} \lbrack \omega,\, \omega \rbrack_{\mathfrak L} \,=\,0$, where $\mathfrak L$ is the Lie algebra of $L$. The manifold $M$ is biholomorphic to a quotient of $L$ by a lattice in it.
Moreover, $M$ is Kähler if and only if $L$ is abelian. In this case $M$ is a compact complex torus.
Consider a basis $(e_1,\, \cdots,\, e_m)$ of the complex vector space $V$, and let $X_{1},\, X_{2},\, \cdots, \,X_{m}$ be global holomorphic $\omega$-constant vector fields on $M$ such that $\omega (X_i)\,=\,e_i$ for all $i$. Consequently, the vector fields $X_{1},\, \cdots, \,X_{m}$ span $TM$.
Notice that for all $1\,\leq\, i,\,j \,\leq\, m$, we have $$\lbrack X_{i}, \,X_{j} \rbrack
\,=\,f_{1}^{ij} X_{1} + f_{2}^{ij} X_{2} +\ldots+ f_{m}^{ij}X_{m}$$ with $f_{k}^{ij}$ being holomorphic functions on $M$. Since $M$ is compact, these (holomorphic) functions are constant and, consequently, $X_{1},\, X_{2},\, \cdots, \,X_{m}$ generate a $m$-dimensional complex Lie algebra $\mathfrak L$. When $V$ is endowed with the Lie algebra structure of $\mathfrak L$, the form $\omega$ produces a Lie algebra isomorphism.
By Lie’s theorem, there exists a unique connected simply connected complex Lie group $L$ corresponding to $\mathfrak L$. The holomorphic parallelization of $M$ by $\omega$-constant holomorphic vector fields is locally isomorphic to the parallelization given by the left translation-invariant vector fields on the Lie group $L$.
Since $M$ is compact, all vector fields $X_{i}$ are complete and they define a holomorphic locally free transitive action of $L$ on $M$ (with discrete kernel). Hence $M$ is biholomorphic to a quotient of $L$ by a cocompact discrete subgroup $\Gamma$ in $L$.
The Lie-Cartan formula $$d\omega (X_{i},\,X_{j})\,=\,X_i \cdot \omega (X_j)-X_j \cdot \omega(X_i) -
\omega (\lbrack X_{i},\,X_{j} \rbrack)\,=\,-\omega (\lbrack X_{i},\,X_{j} \rbrack)\,=\,
-\lbrack \omega(X_i), \,\omega (X_j) \rbrack$$ shows that $\omega$ satisfies the Maurer-Cartan equation of the Lie group $L$, which can also be expressed more formally as $d \omega + \frac{1}{2} \lbrack \omega, \omega \rbrack_{\mathfrak L}
\,=\,0$.
Assume now that $M$ is Kähler. Then, any holomorphic form on $M$ is closed. The Maurer-Cartan formula shows that the one-forms composing the isomorphism $\omega$ are all closed if and only if $L$ is abelian and thus $M$ is a compact complex torus, which is the quotient of a complex vector space by a lattice.
With the terminology of Cartan geometries [@Sh], $\omega$ defines a flat Cartan geometry with respect to the Lie algebra structure $\mathfrak L$ on $V$. Indeed, the vanishing of Cartan’s curvature is equivalent to the fact that $\omega$ satisfies the Maurer-Cartan equation $d \omega + \frac{1}{2} \lbrack \omega, \omega \rbrack_{\mathfrak L}=0$.
Rational parallelisms
=====================
We study here rational parallelisms defined by a [*branched Cartan geometry*]{} modeled on a Lie group in the sense of [@BD]. Let us begin with the following:
A branched holomorphic co-parallelism on the holomorphic tangent bundle $TM$ of a compact complex manifold $M$ of (complex) dimension $m$ is a holomorphic one-form $\omega \,\in\,
\Omega^1(M,\, V)$ with values in a complex vector space $V$ of dimension $m$ such that $\omega_m \,:\, T_mM \to V$ is a linear isomorphism on an open dense subset $U$ in $M$ (which is necessarily the complement of a divisor in $M$, see Proposition \[divisor\]).
The branched holomorphic co-parallelism is [*flat*]{} in the sense of branched Cartan geometries [@BD], with respect to some Lie algebra structure $\mathfrak L$ on $V$, if and only if $\omega$ satisfies the Maurer-Cartan equation $d \omega + \frac{1}{2} \lbrack \omega, \,
\omega \rbrack_{\mathfrak L}\,=\,0$.
Notice that a basis $(e_1,\,\cdots,\, e_m)$ of $V$ uniquely defines a family $(X_1,\, \cdots,\,
X_m)$ of meromorphic vector fields on $M$ such that $\omega(X_i)\,=\,e_i$. This family of meromorphic vector fields $X_i$ holomorphically span $TM$ at the generic point in $M$: they form a [*rational parallelism*]{} of $TM$.
Moreover, the branched holomorphic co-parallelism is flat with respect to the Lie algebra structure $\mathfrak L$, if and only if $\lbrack X_i, \,X_j \rbrack \,= \,\lbrack e_i,\, e_j
\rbrack_{\mathfrak L}$, for all $i,\,j$.
\[1\] The pull-back of a (flat) holomorphic parallelism through a holomorphic map whose differential is invertible at the generic point gives rise to a (flat) branched holomorphic co-parallelism. Consequently, a blow-up (or a ramified cover) of a parallelizable manifold is endowed with a flat branched holomorphic co-parallelism.
The subset $M \setminus U$ where $\omega$ fails to be an isomorphism is called the [*branching locus*]{} of the co-parallelism.
Let us first prove:
\[divisor\] The branching locus of a branched co-parallelism $\omega$ is either empty, or it is an effective divisor in $M$ representing the canonical class. Consequently, if the canonical class of $M$ is trivial (or more generally, does not have an effective representative), any branched holomorphic co-parallelism on $TM$ has empty branching locus and therefore $M$ is a parallelizable manifold.
Choose a basis $(e_1,\, \cdots,\, e_m)$ of $V$ over $\mathbb C$ and consider the corresponding components $\omega_i$ of $\omega$ in this basis: $$\omega\,=\,(\omega_1, \,\cdots,\, \omega_m)
\,\in\, \Omega^1(M, \,\mathbb C^m)\, .$$
The branching locus of the co-parallelism $\omega$ coincides with the vanishing set of the $m$-form $\omega_1 \wedge \ldots \wedge \omega_m$, considered as a holomorphic section of the canonical bundle $K_M$. Consequently, the branching locus of $\omega$ is either empty or it coincides with an effective divisor representing the canonical class of $M$.
If the canonical class of $M$ does not admit an effective representative, then the co-parallelism $\omega$ has empty branching locus. It follows that $M$ is parallelizable manifold, and a quotient of a complex Lie group $L$ by a lattice in it (see Theorem \[Wang\]).
Let us now prove:
\[flat\] Let $M$ be a compact complex manifold endowed with a branched holomorphic co-parallelism $\omega \,\in\, \Omega^1(M,\,V)$.
\(i) If $M$ is either in the Fujiki class $\mathcal C$ or a complex surface, then $\omega$ is flat with respect to the abelian Lie algebra structure on $V$;
\(ii) If $M$ is of algebraic dimension zero (so all meromorphic functions on $M$ are constants), then $\omega$ is flat with respect to some complex Lie algebra structure $\mathfrak
L$ on $V$;
\(iii) If $\omega$ is flat with respect to some complex Lie algebra structure $\mathfrak L$ on $V$, then the fundamental group of $M$ is infinite.
To prove (i), let us first deal with the case where $M$ is in Fujiki class $\mathcal C$ (meaning that $M$ is the meromorphic image of a Kähler manifold [@Fu]). Fix a basis $(e_1,\, \cdots,\, e_m)$ of $V$ over $\mathbb C$ and consider the corresponding components $\omega_i$ of $\omega$ in this basis: $$\omega\,=\,(\omega_1, \,\cdots,\, \omega_m)
\,\in\, \Omega^1(M, \,\mathbb C^m)\, .$$
The manifold $M$ being in class $\mathcal C$, by a result of Varouchas, [@Va], it must be bimeromorphic to a Kähler manifold. Consequently, as for Kähler manifolds, all holomorphic one-forms on $M$ must be closed. This implies $d \omega_i \,=\,0$ and consequently $\omega$ is flat with respect to the abelian Lie algebra $\mathbb C^m$.
Moreover, since $M$ admits non-trivial closed holomorphic one-forms $\omega_i$, the abelianization of the fundamental group of $M$ is infinite. This gives the proof of statement (iii) when $\mathfrak L$ is abelian.
The same proof works for any compact complex surface, since any holomorphic one-form on a compact complex surface is closed (see, for example, [@Br], p. 644).
Proof of (ii): Let us assume that $M$ is of algebraic dimension zero and is endowed with a branched holomorphic co-parallelism $\omega \,\in\, \Omega^1(M,\,V)$. Fix a basis $(e_1,\, \cdots, \,
e_m)$ of $V$ over $\mathbb C$ and consider the corresponding meromorphic vector fields $X_i$ on $M$ defined by $\omega(X_i)\,=\,e_i$. They form a rational parallelism on $M$.
Notice that for all $1\,\leq\, i,\,j \,\leq\, m$, we have $$\lbrack X_{i}, \,X_{j} \rbrack \,=\,f_{1}^{ij} X_{1} + f_{2}^{ij} X_{2} +\ldots+ f_{m}^{ij}X_{m}$$ with $f_{k}^{ij}$ being meromorphic functions on $M$. Since $M$ is of algebraic dimension zero, the functions $f_{k}^{ij}$ are all constants and consequently $X_{1},\, X_{2},\, \cdots, \,X_{m}$ generate a $m$-dimensional complex Lie algebra $\mathfrak L$. When $V$ is endowed with the Lie algebra structure of $\mathfrak L$, the form $\omega$ produces a Lie algebra isomorphism. Hence the rational parallelism is flat with respect to the structure of the Lie algebra $\mathfrak L$.
The Darboux-Cartan Theorem ([@Sh], Chapter 3, p. 116) implies that the open dense subset $U\, \subset\, M$ where $\omega$ is an isomorphism, inherits a $(L,\,L)$-structure in the sense of Ehresmann-Thurston [@Eh], where $L$ is a complex connected Lie group with Lie algebra $\mathfrak L$.
Proof of (iii): To prove by contradiction, assume that the fundamental group of $M$ is finite. Then, up to replacing $M$ by its universal cover endowed with the pullback of $\omega$, we shall assume that $M$ is simply connected. Since $M$ is simply connected we get a holomorphic developing map $$d\,:\, M\,\longrightarrow\, L$$ which is a submersion on an open dense subset (outside the branching locus described in Proposition \[divisor\]) [@BD]. The manifold $M$ being compact, the image $d(M)$ must be closed (and open) in $L$, hence $L\,=\,d(M)$ is compact. But compact complex Lie groups are abelian. This implies $\mathfrak L$ is abelian, and we conclude as before in (i) that the components $\omega_i$ of $\omega$ are closed holomorphic one-forms. The complex manifold $M$ being simply connected, for each $i$ there exists a holomorphic function $h_i$ on $M$ such that $\omega_i\,=\,dh_i$. But holomorphic functions on compact manifolds are constant and, consequently, $\omega_i\,=\,dh_i\,=\,0$, for all $i$: a contradiction.
In contrast to the unbranched case (Theorem \[Wang\]) we exhibit the following non-flat examples, which are inspired by a construction of non-closed holomorphic one-forms in [@Br] (p. 648).
\[non flat\] There exists a branched holomorphic co-parallelism $\omega \,\in\,\Omega^1(P_E,\,V)$ on some compact (non-Kähler) principal elliptic bundle $P_E$, over the product of two Riemann surfaces $S_1$ and $S_2$ of genus $g\,\geq\, 2$, such that $\omega$ is non-flat with respect to any complex Lie algebra structure $\mathcal L$ on $V$.
We shall first construct a holomorphic two-form $\Omega$ on the product $S_1\times S_2$ of two-Riemann surfaces such that the periods of $\Omega$ belong to a lattice $\Lambda$ in $\mathbb C$.
Consider the standard elliptic curve $E\,=\,{\mathbb C}/\Lambda$, with $\Lambda
\,=\, \mathbb Z \oplus \sqrt{-1}\mathbb Z$. Let $z$ be the coordinate on $\mathbb C$ and $dz$ the associated standard (translation invariant) one-form on $E$. Notice that the periods of $dz$ form the lattice $\Lambda\,=\,{\mathbb Z}\oplus
\sqrt{-1}\mathbb Z$.
Choose two Riemann surfaces $S_1$ and $S_2$ (of genus $g \,\geq\, 2$) admitting holomorphic ramified covers $f_1 \,:\, S_1 \,\longrightarrow\, E$ and $f_2 \,:\, S_2 \,
\longrightarrow\, E$. Let us denote by $\Omega$ the holomorphic two-form on the complex surface $S_1 \times S_2$ defined by the pull-back: $$\pi_1^*(f_1^*dz) \wedge \pi_2^*(f_2^*dz)\, ,$$ where $\pi_1$ and $\pi_2$ are the projections of $S_1 \times S_2$ on the first and the second factor respectively. The periods of $\Omega$ belong to the lattice $\Lambda$ in $\mathbb C$.
Fix an open cover $\{U_i\}$ of $S_1 \times S_2$ such that all $U_i$ and all connected components of $U_i
\cap U_j$ are contractible. Then on each $U_i$ there exists a holomorphic one-form $\omega_i$ such that $\Omega_i\,=\,d \omega_i$, where $\Omega_i$ denotes the restriction of $\Omega$ to $U_i$. On any intersections $U_i \cap U_j$, we have $\omega_i-\omega_j\,=\,dF_{ij}$, where $F_{ij}$ is a holomorphic function defined on $U_i\cap U_j$.
On triple intersections $U_i \cap U_j \cap U_k$, the functions $F_{ij} +F_{jk}+F_{ki}$ form a locally constant two-cocycle which represents the class of $\Omega$ in $H^2(S_1 \times S_2,\, {
\mathbb C})$. Hence we can choose the forms $\omega_{i}$ and the associated functions $F_{ij}$ in such a way that $F_{ij}+F_{jk}+F_{ki }$ belongs to the lattice $\Lambda$ of periods for every triple $i,\, j,\, k$.
Consider then every holomorphic function $F_{ij}$ as taking values in the translations group $E =
{\mathbb C}/ \Lambda$. In this way we construct an associate one-cocycle with values in $E$. Let us form the corresponding holomorphic principal elliptic bundle $$\pi : P_E\, \longrightarrow\,
S_1 \times S_2$$ with $E$ as the structure group.
On each local trivialization $U_i \times E$ of $P_E$, consider the local one-form $p_1^*(\omega_i)+p_2^*dz $, where $p_1$ and $p_2$ are the projections of $U_i\times E$ on the first and the second factor respectively. By construction, these local one-forms glue to a global holomorphic one-form $\theta$ on the principal elliptic bundle $P_E$. Moreover we have $d\theta\,=\,\pi^*\Omega$. Since $\Omega$ does not vanish, $\theta$ is a non-closed holomorphic one-form on $P_E$, in particular the complex manifold $P_E$ is non-Kähler. In fact $\theta$ is a holomorphic connection form on the principal elliptic bundle $P_E$ whose curvature is $\pi^*\Omega$ [@At].
Consider a section $\omega_1$ of the canonical bundle $K_{S_1}$ of $S_1$ such that $\omega_1\,=
\,g_1 \cdot f_1^*(dz)$, where $g_1$ is a non-constant meromorphic function on $S_1$. Also consider a holomorphic section $\omega_2$ of the canonical bundle $K_{S_2}$ of $S_2$ such that $\omega_2\,=\,g_2\cdot f_2^*(dz)$ with $g_2$ being a non-constant meromorphic function on $S_2$.
Let $$\omega\,=\,(\theta, \,\pi^* \pi_1^*\omega_1, \,\pi^*\pi_2^*\omega_2) \,\in\,
\Omega^1(P_E,\, \mathbb C^3)$$ be the associated branched holomorphic co-parallelism on $P_E$. Then $$d \theta \,=\, \pi^*\Omega\,=\, (g \circ \pi) \cdot \pi^* (\pi_1^*(\omega_1) \wedge
\pi_2^*(\omega_2))\, ,$$ with $g= (g_1^{-1} \circ \pi_1) \cdot (g_2^{-1} \circ \pi_2)$ being a non-constant meromorphic function on $S_1 \times S_2$. Since $g \circ \pi$ is a non-constant (meromorphic) function on $P_E$, this implies that the branched holomorphic parallelism $\omega$ is non-flat for any Lie algebra structure on the vector space ${\mathbb C}^3$. Moreover, $\omega$ is not locally homogeneous on any nonempty open subset in $P_E$.
We have seen in Theorem \[flat\] (iii) that compact complex manifolds admitting flat branched holomorphic co-parallelisms have infinite fundamental group. We do not know if there exists compact complex simply connected manifolds bearing branched holomorphic co-parallelism (necessarily non-flat with respect to any Lie algebra structure, as that in Theorem \[non flat\]).
The manifold $P_E$ in Theorem \[non flat\] is a ramified cover of the parallelizable manifold $H/\Gamma$, where $H$ is the complex Heisenberg group of upper triangular unipotent $(3 \times 3)$ matrices with complex entries and $\Gamma$ is the lattice of matrices with Gaussian integers as entries. This quotient $H/\Gamma$ is biholomorphic to a principal elliptic bundle with fiber $E\,=\,{\mathbb C}/\Lambda$ (recall that $\Lambda\,=\,{\mathbb
Z}\oplus \sqrt{-1}{\mathbb Z}$) over the two-dimensional compact complex torus $E \times E$. The map $$f\,=\,(f_1\circ\pi_1,\, f_2\circ\pi_2) \,:\, S_1\times S_2 \,\longrightarrow\, E\times E$$ in the proof of Theorem \[non flat\] is a ramified cover. The bundle $P_E$ is the pull-back of the elliptic bundle $H/\Gamma \,\longrightarrow\, E \times E$ through $f$. Consequently, $P_E$ is a ramified cover of of $H/\Gamma$ and inherits a branched holomorphic co-parallelism which is flat with respect to the Lie algebra of $H$ (the three dimensional complex Heisenberg algebra) (see Remark \[1\]).
We do not know examples of compact complex manifolds admitting branched holomorphic co-parallelisms and which are not ramified covers of parallelizable manifolds. In particular, we do not know if all compact complex manifold admitting branched holomorphic co-parallelisms also admit flat branched holomorphic co-parallelisms (with respect to some Lie algebra structure).
Acknowledgements {#acknowledgements .unnumbered}
================
The second-named author wishes to thank T.I.F.R. Mumbai for hospitality. Both authors wishes to thank ICTS Bangalore and IISc Bangalore for hospitality.
This work has been supported by the French government through the UCAJEDI Investments in the Future project managed by the National Research Agency (ANR) with the reference number ANR2152IDEX201. The first author is partially supported by a J. C. Bose Fellowship.
[ZZZZ]{}
M. F. Atiyah, Complex analytic connections in fibre bundles, *Trans. Amer. Math. Soc.* **85** (1957), 181–207.
I. Biswas and S. Dumitrescu, Branched holomorphic Cartan geometries and Calabi-Yau manifolds, *Int. Math. Res. Not.*, DOI: 10.1093/imrn/rny003, arxiv:1706.04407.
D. Blázquez-Sanz and G. Casale, Parallelisms and Lie connections, S.I.G.M.A., **13 (86)**, (2017).
M. Brunella, On holomorphic forms on compact complex threefolds, *Comment. Math. Helv.* **74** (1999), 642–656.
C. Ehresmann, Sur les espaces localement homogènes, *L’Enseign. Math.* **35** (1936), 317–333.
A. Fujiki, On the structure of compact manifolds in $\mathcal C$, *Advances Studies in Pure Mathematics*, **1**, Algebraic Varieties and Analytic Varieties, (1983), 231–302.
R. W. Sharpe, [*Differential Geometry : Cartan’s Generalization of Klein’s Erlangen Program*]{}, Graduate Text Math., 166, Springer-Verlag, New York, Berlin, Heidelberg, 1997.
J. Varouchas, K[ä]{}hler spaces and proper open morphisms, *Math. Ann.* **283** (1989), 13–52.
H.-C. Wang, Complex Parallelisable manifolds, *Proc. Amer. Math. Soc.* **5** (1954), 771–776.
J. Winkelmann, On manifolds with trivial logarithmic tangent bundle, *Osaka Jour. Math.* **41** (2004), 473–484.
J. Winkelmann, On manifolds with trivial logarithmic tangent bundle: the non-Kähler case, *Trans. Groups* **13** (2008), 195–209.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Shawn M. Staudaher,$^{1,2}$[^1] Daniel A. Dale,$^{1}$[^2] Liese van Zee$^{3}$\
$^{1}$Physics and Astronomy Department, University of Wyoming, WY 82071\
$^{2}$Center for 21$^{\rm st}$ Century Universities, The Georgia Institute of Technology, GA 30332\
$^{3}$Astronomy Department, Indiana University, IN 47405
bibliography:
- 'refs.bib'
date: 'Accepted 2019 March 27. Received 2019 March 27; in original form 2018 September 25'
title: 'The Extended Disk Galaxy Exploration Science Survey: Description and Surface Brightness Profile Properties'
---
\[firstpage\]
body.tex gen\_table.tex mass\_table.tex
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The Hall coefficient of Gd-doped La$_{2/3}$Ca$_{1/3}$MnO$_3$ exhibits Arrhenius behavior over a temperature range from $2$ $T_c$ to $4$ $T_c$, with an activation energy very close to $2/3$ that of the electrical conductivity. Although both the doping level and thermoelectric coefficient indicate hole-like conduction, the Hall coefficient is electron-like. This unusual result provides strong evidence in favor of small-polaronic conduction in the paramagnetic regime of the manganites.
PACS: 75.50.Pp; 72.20.My; 71.38.+i
address: |
$^a$Department of Physics and Materials Research Laboratory\
University of Illinois at Urbana-Champaign. 1110 W. Green Street,\
Urbana IL, 61801-3080\
$^b$Center for Material Science, Los Alamos National Laboratory,\
Los Alamos, NM 87545\
$^c$U.S. Naval Research Laboratory, Washington, D.C. 20375-5000\
$^d$Sandia National Laboratories,Albuquerque, NM 87185
author:
- 'M. Jaime$^a$, H. T. Hardner$^a$, M.B. Salamon$^{a,b}$, M. Rubinstein$^c,$ P. Dorsey$^c,$ and D. Emin$^d$'
date: '9/12/96'
title: 'Hall-Effect Sign Anomaly and Small-Polaronic Conduction in (La$_{1-x}$Gd$_x$)$_{0.67}$Ca$_{0.33}$MnO$_3$'
---
A recent resurgence of interest in the transport properties of doped lanthanum manganites has resulted in the realization that electron-lattice interactions play a significant role. Studies of the archetypal system, (La$%
_{1-x}$R$_x$)$_{1-c}$Ca$_c$MnO$_3$, have demonstrated that the ferromagnetic transition temperature $T_c$ (and with it, the metal-semiconductor transition) are suppressed by the addition of rare-earth ions R whose smaller size further closes the Mn-O-Mn bond angle.[@font; @hwang] The temperature dependence of the resistivity above $T_c$ is remarkably unaffected by rare-earth substitutions, following a universal semiconductor-like curve. This supports the view that replacement of a trivalent rare-earth ion by Ca introduces a hole that is presumably associated with the Mn[* *]{}$e_g$ state. Theoretical attempts [@millis; @roder] to describe the large changes in resistivity and their sensitivity to magnetic fields in the vicinity of $T_c$ in the context of the double exchange model have led to the conclusion that strong electron-lattice effects are essential, and that the transition shares aspects of “polaron collapse” such as occurs in EuO. [@emin] The conductivity in the high temperature regime should be dominated by the hopping motion of self-trapped, small polarons. Indeed, quite recent experiments have shown the importance of electron-phonon interactions in CMR via the oxygen isotope effect. [@zhao]
A stable polaron in an ionic solid may be either a large (multisite) polaron that moves itinerantly, or a small (single-site) polaron that moves with a low ($<<$ $1$ cm$^2/$Vs) thermally assisted mobility. In the single-site limit, the self-trapped carrier’s energy is taken to depend only linearly on the displacement of atoms from their carrier-free positions. Consequently, the characteristic energy of the Seebeck coefficient E$_s$ is significantly smaller than the activation energy of the electrical conductivity E$_\sigma $; that E$_s<<$ E$_\sigma $ was demonstrated in earlier work, [@jaime] and taken as evidence of small-polaronic motion. E$_s$ measures the chemical potential of the self-trapped polaron.
Perhaps the most distinctive property of steady-state small-polaronic transport is its Hall mobility $\mu _H$. The activation energy of the Hall mobility is calculated to be always less than that for drift mobility E$_d.$ The simplest model predicts $\approx $E$_d/3,$ and this has been observed in, for example, oxygen-deficient LiNbO$_3.$ [@nagels] The sign of the Hall effect for small polaron hopping can be “anomalous.” A small polaron based on an electron can be deflected in a magnetic field as if it were positively charged and, conversely, a hole-based polaron can be deflected in the sense of a free electron. As first pointed out by Friedman and Holstein [@fh], the Hall effect in hopping conduction arises from interference effects of nearest neighbor hops along paths that define an Aharonov-Bohm loop. Sign anomalies arise when the loops involve an odd number of sites.[@emin1] In this Letter, we report the first measurement of the high-temperature Hall coefficient of manganite samples, finding that it exhibits Arrhenius behavior and a sign anomaly relative to both the nominal doping and the thermoelectric power. The results are discussed in terms of an extension of the Emin-Holstein (EH) theory of the Hall mobility in the adiabatic limit.
We exploit the sensitivity of these materials to rare-earth substitutions to lower the transition temperature from $\sim 260\;$K at $x=0$ to $\sim 130$ K, thereby extending the accessible temperature range to $\simeq 4$ $T_c.$ The samples used in this study were laser ablated from ceramic targets and deposited on LaAlO$_3$ substrates as described previously. [@jaime; @white] The ceramic target material (R = Gd and $x=0.25)$ has a resistivity maximum near $100$ K while the thermoelectric power drops abruptly to metallic values below $90K$. [@white] The effective rare-earth site radius is $%
\left\langle r_A\right\rangle =$ $0.113$ nm and, indeed, the properties of this sample are very similar to Y-substitute samples with the same $%
\left\langle r_A\right\rangle $ value. [@font] The laser deposited films show somewhat higher transition temperatures: the resistivity maximum is at $%
140$ K and the thermoelectric power becomes metal-like below $130$ K. X-ray data indicate that the in-plane lattice parameter of the laser ablated film is larger than that of the ceramic, presumably due to stress induced by the substrate. Fig. 1 shows resistivity and thermopower data for the $130$ K sample along with resistivity data for similar samples with $x=0$ and $%
x=0.5. $ We focus here on the $130$ K sample, for which $\rho $ drops by a factor $160$ between $0$ and $8$ Tesla at $142$ K at which field the peak in the thermopower is almost completely suppressed.
The most rapid motion of a small polaron occurs when the carrier hops each time the configuration of vibrating atoms in an adjacent site coincides with that in the occupied site. This regime is termed adiabatic with a conductivity given by $\sigma =\sigma _o\exp (-E_\sigma /$k$_B$T$),$ where $$\sigma _o=g_de^2\nu _0/ak_BT. \label{eq1}$$ Here, $\nu _0$ is a characteristic frequency and $a$ is the jump distance, which we take to be the Mn-Mn spacing, $0.39\;$nm. The factor $g_d$ depends on hopping geometry and has the value $g_d=3/2$ for the triangular lattice treated by EH and $g_d=1$ for nearest-neighbor hopping on a square planar lattice. A signature of the adiabatic limit is that the prefactor $\sigma _o$ approaches$\;e^2/\hbar a\simeq 7000\;$($\Omega $ cm)$^{-1}$ when $h\nu
_0\simeq k_BT$. If it is much smaller, the motion is termed nonadiabatic, and the prefactor contains an additional factor of T$^{-1/2}.$ In Fig. 2 we plot -$\log (\sigma $T$)$ and -$\log (\sigma $T$^{3/2})$, the adiabatic and nonadiabatic limits respectively, vs 1/T$.$ Close to $T_c$ there are significant deviations from Arrhenius behavior, as has been noted previously. [@white; @emin2] At higher temperatures, there is no significant difference between adiabatic and nonadiabatic fits; the fitting parameters are given in Table 1 along with those for Gd concentration $x=0$ and $x=0.5$. The adiabatic prefactor for the $x=0.25$ sample is $\approx
6000\;\Omega ^{-1}$cm$^{-1}$ at $300$ K, confirming that the hopping processes are adiabatic. The characteristic frequency $\nu _0$ in Eq. (1) is of order $10^{14}/g_d$ Hz. Evidence, discussed below, implies that $g_d$ is significantly larger than the value for pure nearest-neighbor hopping ($%
g_d=1$). This reduces $\nu _0$ to a value characteristic of optical phonons in transition-metal oxides.
Sections of these specimens were patterned by conventional lithographic methods into a five-terminal Hall geometry. Hall experiments were carried out in a high-temperature insert constructed for use at the $20$ Tesla superconducting magnet at the National High Magnetic Field Laboratory (Los Alamos, NM). Figure 3 shows the raw data (transverse and longitudinal voltages) obtained in for $x=0.25$ at $462$ K. Although the sample lithography resulted in negligible zero field transverse voltage in the metallic state, a transverse magnetoresistive signal is apparent above $T_c;$ see Fig. 3. This signal peaks around $T_c$ and then decreases with increasing temperature. Clearly, this indicates inhomogeneous current paths in the proximity of the metal-insulator transition, possibly due to local relaxation of epitaxial strain and resulting local variations in $T_c$. Consequently, we did not attempt to follow the behavior of the Hall constant below $\sim 2T_c.$ The transverse voltage data taken while sweeping the field from $-16$ Tesla to $+16$ Tesla, and that taken while sweeping back to $-16$ Tesla, were each fit to a second-order polynomial with the term linear in field attributed to the Hall effect. We verified in each case that the longitudinal magnetoresistance is completely symmetric in field. Fig. 4 shows the Hall coefficient derived from the linear term. Several points for the $x=0$ film are included. Due to the much higher $T_c$ of that sample, extraction of the Hall contribution leads to greater uncertainty. The data are, however, consistent with the Gd-substituted film. The line through the data points is an Arrhenius fit, giving the expression R$_H=-(3.8\times
10^{-11}\;$m$^3/$C$)\exp (91\;$meV$/$k$_B$T$).$ Note that the sign is negative, even though divalent dopants should introduce holes. That the Seebeck coefficient approaches a negative value at high temperatures has been attributed in part to the reduction in spin entropy produced when a hole converts a Mn$^{3+}$ ion to Mn$^{4+}.$ [@jaime]
Detailed expressions for the Hall effect in the adiabatic limit have been calculated by EH for the hopping of electrons with positive transfer integral $J$ on a triangular lattice, and results in a normal (electron-like) Hall coefficient. However, the sign of both the carrier and the transfer integral changes for hole conduction, leaving the sign of the Hall coefficient electron-like, and therefore anomalous. [@fh; @emin1] However, no anomaly would arise if the hopping involves 4-sided loops with vertices on nearest-neighbor Mn atoms. A sign anomaly, then, implies that hopping involves odd-membered Aharonov-Bohm loops. Such processes arise when next-nearest neighbor (nnn) transfer processes across cell face diagonals are permitted. If the Mn-O-Mn bonds were strictly colinear, the former processes would be disallowed by symmetry. However, the bond angles are substantially less than $180^{\circ }$, implying the presence of $\pi $-bond admixture, and opening a channel for diagonal hops. We have extended the triangular-lattice calculation of EH to the situation in which a hole on a Mn ion can hop to any of its four nearest neighbors in the plane normal to the applied field with transfer matrix element J $<0$ [*and* ]{}to its four next-nearest-neighbors (nnn) with a reduced transfer energy $\gamma $J$.$ We must also consider the effect of these diagonal hops (plus those in the plane containing both electric and magnetic fields) on the conductivity prefactor, Eq. (1). The Hall coefficient can be written as R$_H=$R$_H^o($T$%
)\exp (2E_\sigma /3k_BT),$ with $$R_H^o=-\frac{g_H}{g_d}\frac{F(|J|/k_BT)}{ne}\exp \left[ -[\epsilon
_0+(4|J|-E_s)/3]/k_BT\right] ; \label{eq2}$$ EH found that the factor $g_H=1/2$ for three-site hopping on a triangular lattice. In Eq.(2) we have expressed the carrier-density as $n\exp
(-E_s/k_BT),$ where $E_s$ is estimated to be 8 meV from the thermopower data. The quantity $\epsilon _0$ is the $J-$dependent portion of a carrier’s energy achieved when the local electronic energies of the three sites involved in an Aharonv-Bohm loop are equal. For the problem considered by EH, an electron hopping within a lattice composed of equilateral triangles, $%
\epsilon _0=-2|J|,$ and $g_H/g_d=1/3.$ Within the domain of validity of EH, the temperature dependence of $R_H$ arises primarily from the factor $\exp
(2E_\sigma /3k_BT)$ when $E_\sigma >>E_s.$ For holes hopping within a cubic lattice in which three-legged Aharonov-Bohm loops include face-diagonal transfer, we find $\epsilon _o=-|J|(\sqrt{8+\gamma ^2}-\gamma )/2$. In particular, $\epsilon _o$ varies from -$\sqrt{2}|J|$ to $-|J|$ as $\gamma $ increases from zero to unity [@woods], and the temperature dependence of $R_H$ remains dominated by the factor $\exp (2E_\sigma /3k_BT).$ Indeed, the energy characterizing the exponential rise of the Hall coefficient that we observe, E$_{Hall}=91\pm 5\;$meV, is about 2/3 the measured conductivity activation energy, E$_{Hall}/$E$_\sigma =0.64\pm 0.03,$ in excellent agreement with theory.
The geometrical factor $g_d$ depends on the ratio of the probability $%
P_{nnn} $ of nnn hops to $P_{nn},$ that of nn hops, through $%
g_d=(1+4P_{nnn}/P_{nn}).$ If these probabilities are comparable ($\gamma
\sim 1)$ $g_d=5,$ $g_H=2/5$ and the exponential factor in Eq. (2) becomes $%
\exp [(E_s-|J|)/3k_BT]\simeq 1.$ In the regime $|J|\gtrsim k_BT,$ the function $F(|J|/k_BT)$ is relatively constant with a value $\approx 0.2,$ and we find $R_H^o\simeq -0.02/ne=-3.8\times 10^{-11}$m$^3$/C. This yields an estimated carrier density $n=3.3\times 10^{27}$m$^{-3},$ quite close to the nominal level of $5.6\times 10^{27}$ m$^{-3}.$ Diagonal hopping also reduces the value of the attempt frequency $\nu _0$ required to fit the conductivity prefactor to $\approx 2\times 10^{13}$ Hz, as noted above.
In conclusion, we have measured the high-temperature Hall coefficient in manganite films and found that its temperature dependence is consistent with small-polaron charge carriers that move by hopping. Further, the magnitude of the conductivity prefactor indicates that the carrier motion is adiabatic. Finally, the sign anomaly in the Hall effect implies that small polarons hop not only among near-neighbor sites (making Aharonov-Bohm loops with an even number of legs) but must have a significant probability of traversing Hall-effect loops with odd numbers of legs. As such, the results indicate the occurrence of significant nnn transfer across face diagonals, and therefore a crucial role for deviations of the Mn-O-Mn bond angle from $%
180^{\circ }$. An interesting possibility, that may also relate to unusual high-temperature values observed for the Seebeck coefficient [@jaime], is that transport is a type of impurity conduction in which carriers remain adjacent to divalent cation dopants (i.e. Ca ions). The local distortions associated with the presence of the impurity may also increase the admixture of $\pi $-bonds, and enhance diagonal hopping.
We are pleased to acknowledge useful discussions with J.B. Goodenough and C.P. Flynn, and the assistance of A. Lacerda with high fields experiments. This work was supported by the Department of Energy, office of Basic Energy Sciences through Grant No. DEFG02-91ER45439 at the University of Illinois and under contract DE-AC04-94AL85000 at Sandia National Laboratory. MBS was supported, in part, as a Matthias Fellow at Los Alamos National Laboratory.
J. Fontcuberta, B. Martinez, A. Seffar, S. Piñol, J.L. García-Muñoz and X. Obradors, Phys. Rev. Lett. [**76**]{}, 1122 (1996); J. Fontcuberta, et. al. Appl. Phys. Lett. [**68**]{}, 2288 (1995).
H.Y. Hwang, S-W. Sheong, P.G. Radaelli, M. Marezio, and B. Batlogg, Phys. Rev. Lett. [**75**]{}, 914(1995); Note: the authors appear to have used incorrect coordination numbers in calculating $\left\langle
r_A\right\rangle .$
A.J. Millis, P.B. Littlewood, and B.I. Shraiman, Phys. Rev. Lett[*.*]{} [**74**]{}, 5144 (1995); A.J. Millis, et al. (preprint, 1996).
H. Röder, J. Zang, and A.R. Bishop, Phys. Rev. Lett[*.*]{} [**76**]{}, 1356 (1996).
D. Emin, M. Hillery, and N-L. H. Liu, Phys. Rev. B. [**33**]{}, 2933 (1986).
G. Zhao, K. Conder, H. Keller, K. A. Müller[*,* ]{}Nature[* *]{}[**381**]{}, 676 (1996).
D. Emin and T. Holstein, Ann. Phys. (N.Y.) [**53**]{}, 439 (1969).
M. Jaime, M.B. Salamon, K. Pettit, M. Rubinstein, R.E. Treece, J.S. Horowitz, and D.B. Chrisey, Appl. Phys. Lett. [**68,**]{} 1576 (1996); M. Jaime, et al. Phys. Rev. B[* *]{}(in press).
P. Nagels in [*The Hall Effect and its Applications*]{}, Edited by C.L. Chien and C.R. Westgate (Plenum, NY 1980) p. 253
L. Friedman and T. Holstein, Ann. Phys. (N.Y.) [**21**]{}, 494 (1963).
D. Emin, Phil. Mag[*.*]{} [**35**]{}, 1189 (1977).
P. White, M. Jaime, M.B. Salamon and M. Rubinstein, Bull. Am. Phys. Soc[*.*]{} [**41**]{}, 116 (1996).
D. Emin, Phys. Rev. Lett. [**32**]{}, 303 (1974).
C. Wood and D. Emin, Phys. Rev. B [**29**]{}, 4582 (1984).
---------------------------------------------------------------------------------------------------------------------------------------------------------------
(La$_{1-x}$Gd$_x$)$_{2/3}$Ca$_{1/3}$MnO$_3$ E$_\sigma ($m$e$V$)$ $\sigma E$_H($m$e$V$)$ R$_H^o(10^{-10}$ m$^3/$C$)$
_o(\Omega ^{-1}$cm$^{-1})$
--------------------------------------------- ---------------------- -------------------------------------- ---------------- ----------------------------------
sample A ($x=0$)
Nonadiabatic limit $102$ $2.5\times 10^7$ K$^{3/2}/$T$^{3/2}$ $\sim 69$
Adiabatic limit $85$ $7.7\times 10^5$ K$/$T $\sim 59$
sample B ($x=0.25$)
Nonadiabatic limit $158$ $5.1\times 10^7$ K$^{3/2}/$T$^{3/2}$ $112$ ($4.8\times 10^{-4}\;$K$^{-1})$T
Adiabatic limit $145$ $1.8\times 10^6$ K$/$T $91$ $0.38$
sample C ($x=0.5$)
Nonadiabatic limit $157$ $2.3\times 10^7$ K$^{3/2}/$T$^{3/2}$
Adiabatic limit $146$ $8.9\times 10^5$ K$/$T
---------------------------------------------------------------------------------------------------------------------------------------------------------------
: Parameters from adiabatic and non-adiabatic fits to the resistivity and Hall data \[t1\]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Modern Automatic Speech Recognition (ASR) systems can achieve high performance in terms of recognition accuracy. However, a perfectly *accurate* transcript still can be challenging to read due to grammatical errors, disfluency, and other errata common in spoken communication. Many downstream tasks and human readers rely on the output of the ASR system; therefore, errors introduced by the speaker and ASR system alike will be propagated to the next task in the pipeline. In this work, we propose a novel NLP task called ASR post-processing for readability (APR) that aims to transform the noisy ASR output into a readable text for humans and downstream tasks while maintaining the semantic meaning of the speaker. In addition, we describe a method to address the lack of task-specific data by synthesizing examples for the APR task using the datasets collected for Grammatical Error Correction (GEC) followed by text-to-speech (TTS) and ASR.
Furthermore, we propose metrics borrowed from similar tasks to evaluate performance on the APR task. We compare fine-tuned models based on several open-sourced and adapted pre-trained models with the traditional pipeline method. Our results suggest that fine-tuned models improve the performance on the APR task significantly, hinting at the potential benefits of using APR systems. We hope that the read, understand, and rewrite approach of our work can serve as a basis that many NLP tasks and human readers can benefit from.
author:
- |
[Junwei Liao]{}$^{1}$$^2$ $^2$ $^2$$^3$\
**Linjun Shou**^3^ **Hong Qu**^1^ **Michael Zeng**^2^\
$^1$[University of Electronic Science and Technology of China]{}\
$^2$[Microsoft Speech and Dialogue Research Group]{}\
$^3$[Microsoft STCA NLP Group]{}\
`[email protected]`\
`{seeskime, liyang.lu, yushi, migon, lisho, nzeng}@microsoft.com`\
`[email protected]`\
bibliography:
- 'acl2020.bib'
title: Improving Readability for Automatic Speech Recognition Transcription
---
Introduction
============
With the rapid development of speech-to-text technologies, ASR systems have achieved high recognition accuracy, even beating the performance of professional human transcribers on conversational telephone speech in terms of Word Error Rate (WER) [@xiong2018microsoft].
Automatic speech recognition systems bring convenience to users in many scenarios. However, colloquial speech is fraught with syntactic and grammatical errors, disfluency, informal words, and other noises that make it difficult to understand. While ASR systems do a great job in recognizing which words are said, its verbatim transcription creates many problems for modern applications that must comprehend the meaning and intent of what is said. Applications such as automatic subtitle generation and meeting minutes generation require automatic speech transcription that is highly readable for humans, while machine translation, dialogue systems, voice search, voice question answering, and many other applications require highly readable transcriptions to generate the best machine response. The existence of the defects in speech transcription will significantly harm the experience of the application users if the system cannot handle them well.
Inspired by the latest progress in natural language generation (NLG), grammatical error correction (GEC), machine translation, and transfer learning, we explore the idea of “understanding then rewriting" as a new ASR post-processing concept to provide conversion from raw ASR transcripts to error-free and highly readable text.
We propose ASR post-processing for readability (APR), which aims to transform the ASR output into a readable text for humans and downstream NLP tasks. Readability in this context refers to having proper segmentation, capitalization, fluency, and grammar, as well as properly formatted dates, times, and other numerical entities. Post-processing can be treated as a style transfer, converting informal speech to formal written language.
Due to the lack of relevant data, we constructed a dataset for the APR task using a GEC dataset as seed data. The GEC dataset is composed of pairs of grammatically incorrect sentences and corresponding sentences corrected by a human annotator. First, we used a text-to-speech (TTS) system to convert the ungrammatical sentences to speech. Then, we used an ASR system to transcribe the TTS output. Finally, we used the output of the ASR system and the original grammatical sentences to create the data pairs. By this means, we produced 1.1 million APR samples that are used for training and testing.
We investigated three mainstream Transformer-based sequence-to-sequence neural network architectures for the APR task. Specifically, we investigated MASS [@song2019mass], UniLM [@dong2019unified] and RoBERTa [@liu2019roberta], which are pre-trained models used for NLG and/or NLU tasks. We also attempted to leverage the advantages of both RoBERTa and UniLM by adapting the RoBERTa pre-trained model towards generative using a modified UniLM training approach (RoBERTa-UniLM).
We used several metrics to evaluate the four fine-tuned models on our APR dataset: readability-aware WER (RA-WER), BLEU, MaxMatch (M$^2$), and GLEU. The results show that the fine-tuned models outperform the baseline method significantly in terms of readability.
Our main contributions can be summarized as follows:
- We propose a novel task: ASR post-processing for readability (APR). It aims to solve the shortcomings of the traditional post-processing concept/methods by jointly performing error correction and readability improvements in one step.
- We describe a method to construct a dataset for the APR task.
- We experiment using state-of-the-art pre-trained models on the proposed APR dataset and achieved significant improvement on all metrics.
- We adapt RoBERTa as a generative model trained in the style of UniLM which shows its benefits on some metrics such as M$^2$ and BLEU.
Related Work
============
Automatic Speech Recognition (ASR)
----------------------------------
Traditional ASR systems take a pipelined approach, [@paulik2008sentence; @cho2012segmentation; @cho2015punctuation; @batista2008recovering; @gravano2009restoring; @vskodova2012discretion] relying on post-processing modules to improve the readability of the output text in two critical ways. First, a more robust language model is used to reduce word recognition errors via a second-pass rescoring of the output lattice or top recognition candidates. Then other sub-processes modify the sentence display format for readability using a series of steps, adding capitalization and punctuation, correcting simple grammatical errors, and formatting dates, times, and other numerical entities. We call these steps inverse text normalization (ITN). Originally, researchers mainly exploited handcrafted rules or statistical methods [@shugrina2010formatting; @anantaram2016repairing; @bohac2012post; @liyanapathirana2016using; @shivakumar2019learning; @cucu2013statistical; @bassil2012post] for post-processing. Recently, @guo2019spelling trained an LSTM-based sequence-to-sequence model to correct spelling errors. @hrinchuk2019correction investigated the use of Transfomer-based architectures for the correction of SR output into grammatically and semantically correct forms.
Traditional ASR post-processing methods offer improvements in readability; however, there are two important shortcomings. (1) Since the whole process is divided into several sub-processes, the mistakes in the previous steps will accumulate. For example, in the sentence, “Mary had a little lamb. It’s fleece was white as snow.”, if in the punctuation step, a period ‘.’ is added after the word “had,” the rule-based capitalization will capitalize the word‘*a*.’ (2) The traditional methods tend to transcribe the speech verbatim while ignoring the readability of the output text. It cannot detect and correct disfluency in spontaneous speech transcripts. For example, in an utterance such as “I want a flight ticket to Boston, uh, I mean to Denver on Friday”, the speaker means to communicate “I want a flight ticket to Denver on Friday.” The segment “*to Boston, uh, I mean*” in the transcript is not useful for interpreting the intent of the sentence and hinders human readability and the performance of many downstream tasks. Traditional methods optimized for recognition accuracy will keep these words, increasing the cognitive load of the reader.
Natural Language Processing (NLP)
---------------------------------
In NLP research, the most similar task to ours is automatic post-editing (APE) [@bojar2016findings], which has been extensively studied by the machine translation (MT) community (e.g., [@pal2016neural; @pal2017neural; @chatterjee2017multi; @hokamp2017ensembling; @tan2017neural]). These methods take input of the source language text, target language MT output, and target language post-editing (PE) for training. Based on our knowledge, there is no similar work in speech recognition field. Another similar task is the Grammatical Error Correction (GEC). GEC aims to correct different kinds of errors such as spelling, punctuation, grammatical, and word choice errors [@ge2018reaching; @zhang2019sequence; @napoles2019enabling; @napoles2017jfleg; @grundkiewicz2019neural; @choe2019neural]. The difference between our task and GEC is that the latter aims to correct written language, while our task aims to correct spoken language that contains noise introduced by ASR errors as well that introduced by the discrepancy between spoken and written formats of natural language. Due to the similarity between GEC and APR, we borrow some ideas from GEC and use GEC corpora as our seed corpus to synthesize our dataset and use GEC metrics, namely MaxMatch and GLEU, to evaluate APR performance.
Unsupervised Learning
---------------------
Pre-training approaches [@dai2015semi; @mccann2017learned; @howard2018universal] have drawn much attention recently, especially those that employ the Transformer [@Vaswani2017AttentionIA] architecture. The most successful approaches are variants of masked language models, which are denoising autoencoders trained to reconstruct text where a random subset of the words has been masked out. Among them, BERT [@devlin2018bert] and RoBERTa [@liu2019roberta] are single-stack Transformer encoders; GPT(-2) [@radford2018improving; @radford2019language] and XLNET[@yang2019xlnet] are single-stack Transformer decoders; UniLM [@dong2019unified] is a single-stack Transformer serving both encoder and decoder roles; and MASS [@song2019mass], BART [@lewis2019bart] and T5 [@raffel2019t5] are standard Tranformer-based neural machine translation architecture. We use RoBERTa, UniLM, and MASS as our base architectures and use their pre-trained models for the APR task.
------------ ---------------------------------------
**Input** *She see Tom is catched by policeman*
*in park at last night.*
**Output** *She saw Tom caught by a policeman*
*in the park last night.*
------------ ---------------------------------------
: \[gec-example\] A GEC data sample is shown. The input is a sentence with some grammatical errors. The output is a grammatically correct sentence.
Proposed Method
===============
{width="100.00000%"}
Dataset
-------
There exist a large amount of data that have been labeled for speech recognition. However, these data have two issues: first, they label the exact words that were spoken, including all disfluency. This is essential for HMM and hybrid acoustic model training but could hinder readability. Second, no capitalization and punctuation is present because spontaneous speech does not follow normal grammatical conventions. Similarly, entities, especially numerical entities, appear different in spoken language than when they appear in written form.
Due to these restrictions, we synthesize our data, simulating ASR errors by feeding sentences from a grammatical error correction (GEC) dataset into a text-to-speech (TTS) system and then transcribing this with an ASR model.
The GEC data samples contain grammatically correct and incorrect sentence pairs. A human corrects the grammatically incorrect sentence to obtain the target grammatically correct sentence. An example sentence pair from the seed corpus is shown in Table \[gec-example\]. Inspired by the GEC task, we simulated ASR errors using the GEC source sentences to obtain sentence pairs of which source sentences contain both grammatical errors and ASR errors. In the next section, we detail how we simulated the APR data. We discuss the simulated data statistics in \[sses:data\_statistic\].
### Dataset Synthesis {#ssec:data_synthesis}
We fed the grammatically incorrect sentences from the seed corpus into a neural-TTS system, which produced the audio files simulating human speakers. We used 320 different speaker voices for this simulation and split them into 220 for training, 50 for validation, and 50 for evaluation. Each sentence randomly selected one speaker, and all speakers have the same number of input sentences [@deng2018modeling]. Then these audio files are fed into the ASR system that outputs the corresponding transcript. The resulting text contains both the grammatical errors found initially in the GEC dataset and the TTS+ASR pipeline errors. We used original corrected sentences as our target. The whole process is illustrated in Figure \[fig-data\].
In addition to the mentioned simulation method, we tried using the top-k best output of the ASR system to augment our dataset ten-fold. However, we found that the augmented dataset is not beneficial, due to the lack of diversity in the resulting sentences, which often differ only in some characters (\[para:data-selection\]).
### Dataset Statistic {#sses:data_statistic}
Table \[table-stat\] shows dataset statistic of our data.
We used the data from the datasets provided by restricted tracks of BEA 2019 shared task [@Bryant2019TheBS] as our seed corpora for training. Specifically, we collected data from FCE [@Yannakoudakis2011AND], Lang-8 Corpus of Learner English [@Mizumoto2011MiningRL; @Tajiri2012TenseAA], and W&I+LOCNESS [@Bryant2019TheBS; @granger2014computer], totaling to around 1.1 million training samples.
Furthermore, we utilized CoNLL-2014 shared task dataset [@ng2014conll] and JFLEG [@napoles2017jfleg] test set as our evaluation seed corpora, to be aligned with the GEC literature [@ge2018reaching; @zhang2019sequence; @Kiyono2019AnES]. The CoNLL-2014 and JFLEG test sets contain 1,312 and 747 sentences, respectively.
In order to be consistent with the standard evaluation metrics in the GEC literature, we used MaxMatch (M$^2$) $F_{0.5}$ [@Dahlmeier2012BetterEF] for CoNLL-2014 and used GLEU [@Napoles2015GroundTF] for JFLEG evaluation. We used the CoNLL-2013 test set and JFLEG dev set as our development seed corpora for the CoNLL-2014 and JFLEG test sets, respectively.
Finally, through the process described in Section \[ssec:data\_synthesis\], we obtained the APR dataset illustrated in the right part of Table \[table-stat\].
Evaluation Metrics
------------------
Since our task is to improve the readability of automatic speech transcription, the word error rate (WER), a conventional metric that is widely used in speech recognition, is not suitable for our use case. As a part of our work, instead, we investigated the usefulness and consistency of different metrics directly or modified from that of related tasks such as speech recognition, machine translation, and grammatical error correction.
#### Speech Recognition Metric
First, we extended the conventional WER in speech recognition to readability-aware WER (RA-WER) by removing the text normalization before calculating Levenshtein distance. We treated all mismatches due to grammatical mistakes, disfluency, as well as improper formats of capitalization, punctuation, and written numerical entity as errors. If there are alternative references, we selected the closest one to the candidate.
#### Machine Translation Metric
The APR task can be treated as a translation problem from a spoken transcript to a more readable written text. In this case, we can take advantage of the BiLingual Evaluation Understudy (BLEU) [@Papineni2001BleuAM] score that is widely used in machine translation to measure the performance of the APR task. In BLEU, the precision score is computed over variable-length of n-grams with length penalty [@Papineni2002MachineTE] and optionally with smoothing [@Lin2004ORANGEAM].
#### Grammatical Error Correction Metrics
Syntax and grammatical errors can significantly impact the readability of speech transcription. To evaluate the correctness and fluency of the rewritten sentences, we used the most commonly used GEC metrics such as MaxMatch (M$^2$) and General Language Evaluation Understanding (GLEU) in our work. M$^2$ reports the F-score of edits over the optimal phrasal alignment between the candidate and the reference sentences [@Dahlmeier2012BetterEF]. GLEU captures fluency rewrites in addition to grammatical error corrections [@Napoles2015GroundTF]. It is an extension of BLEU [@Papineni2001BleuAM] by penalizing false negatives. Besides the candidates and references used in other metrics, GEC metrics also consider source sentences in order to detect the model edits. In all experiments, we used raw ASR transcription as the source sentence when calculating GEC metrics.
Baseline Setup
--------------
We used the production 2-step post-processing pipeline of our speech recognition system as the APR baseline, namely n-best LM rescoring followed by inverse text normalization (ITN). This pipeline works well for sequentially improving speech recognition accuracy and display format. We computed the values of the metrics between system output of every step and the reference grammatical sentence. As a comparison, we also evaluated the original ungrammatical sentences in the same corpora. Table \[baseline\] shows these baseline results on CoNLL-2014 and JFLEG test sets.
An interesting finding is that although the original ungrammatical sentences in JFLEG have more errors or are less smooth than the ones in CoNLL2014 according to the higher RA-WER (11.84 vs. 7.51) and lower BLEU (80.56 vs. 87.79), the situation is inverted after transforming the sentences to and back from speech (29.50 vs 31.15 in RA-WER and 62.42 vs 60.11 in BLEU). This result may indicate that: 1) JFLEG annotators focused more on fluency and formality of the rewriting rather than pure and token-level error corrections in CoNLL2014, and 2) the ASR system, due to the powerful language model, has the ability to regularize input errors and make the transcription appear more fluent and formal. Second, GEC metrics are more sensitive to correct edits than other metrics due to the consideration of the input source sentences. Third, ITN consistently shows a much more significant impact than LM rescoring, which demonstrates the importance of display format in readability and also raises the question of how to further emphasize the correctness for future work.
Models
------
In this work, we compare different Transformer [@Vaswani2017AttentionIA] architectures together with corresponding open-sourced pre-trained models.
### MASS[^1]
MASS [@song2019mass] adopts the encoder-decoder framework to reconstruct a sentence fragment given the remaining part of the sentence. This framework is ideally suited for our task.
Following the MASS setting, we tokenized the data using the Moses toolkit[^2] and used the same BPE codes and vocabulary from MASS.
We fine-tuned the model based on the weights pre-trained on English monolingual data. The model consists of a 6-layer encoder and a 6-layer decoder. The learning rate was $10^{-4}$ with linear warm-up beginning from $10^{-7}$ for the first 4K updates, followed by inverted squared decay. To fully utilize the GPU, we use dynamically sized mini-batches with 3000 tokens per batch.
### UniLM[^3]
UniLM [@dong2019unified] was pre-trained using the BERT-large [@devlin2018bert] architecture and three types of language modeling tasks: unidirectional, bidirectional, and sequence-to-sequence prediction. The unified modeling approach allows UniLM to be used for both discriminative and generative tasks.
Following the UniLM setting, we tokenize the training data using WordPiece [@Wu2016GooglesNM] with vocabulary size 28,996. The model is a 24-layer Transformer with around 340M parameters.
We fine-tuned the model for 4 epochs. The learning rate was $10^{-5}$, with linear warmup over the one-tenths of total steps and linear decay. The batch size, maximum sequence length and masking probability were set to 256, 192 and 0.7, respectively. We also used label smoothing [@szegedy2016rethinking] with a rate of 0.1. Following standard practice, we removed duplicate trigrams in beam search and tuned the maximum output length and length penalty on the development set [@Paulus2017ADR; @fan2017controllable].
### RoBERTa[^4]
RoBERTa [@liu2019roberta] is a robustly optimized BERT [@devlin2018bert] pre-training approach. Both BERT and RoBERTa have single Transformer stack and are pre-trained only using bidirectional prediction, which makes them more discriminative than generative. However, @hrinchuk2019correction demonstrated the effectiveness of transfer learning from BERT to sequence-to-sequence task by initializing both encoder and decoder with pre-trained BERT in their speech recognition correction work.
Inspired by this work and UniLM, we applied self-attention masks on the RoBERTa model to convert it into a sequence-to-sequence generation model. To achieve whole-sentence prediction rather than only masked-position prediction, we used an autoregressive approach during the fine-tuning. Another benefit from this approach is that the model can predict the end of sentence precisely; hence, there is no need to tune the maximum output length and length penalty as in UniLM fine-tuning.
Following the RoBERTa setting, the sentences were tokenized with a byte-level BPE tokenizer. The vocabulary size was 50,265. We fine-tuned the model based on RoBERTa-large pretrained weights.
### RoBERTa-UniLM
Besides using UniLM and RoBERTa, we also experimented to leverage the advantages of both works to further enhance the pre-trained model before fine-tuning it on APR task. We adapted RoBERTa-large model by training it longer on a combination of English Wikipedia[^5], Books[^6], and News-Crawl[^7] data, totaling to 66GB of uncompressed text. The training was similar to UniLM but also included autoregressive (both left-to-right and right-to-left) prediction. We kept next-sentence objective in the bidirectional masked LM (MLM). All predictions conducted by whole-word masking. The first three predictions also had bigram, trigram, and phrase masking each on about 10% of the training instances. We used Huggingface Transformers[^8] code for LM fine-tuning. The RoBERTa-UniLM model was trained for 10 days on 64 NVIDIA DGX-2 GPU cards for 7,200 steps with a batch size of 12,800. The learning rate was $10^{-4}$, with the same warmup and decay strategy with UniLM fine-tuning. The APR task fine-tuning and decoding were the same with the RoBERTa experiment.
In all fine-tuning experiments described above, checkpoints were selected on the development set, and the beam size for beam search was set to 5.
Results and Discussion
======================
Dataset Selection {#para:data-selection}
-----------------
As described in Section \[ssec:data\_synthesis\], we constructed the APR data using TTS and ASR systems. When an audio file synthesized by TTS is inputted to the ASR system, it will generate multiple candidate sentences from the beam search for re-ranking. These sentences have a few words that are different from the final output. At first, we used all of these sentences as our APR training data. When training our model on this data, we found that the loss converges very quickly. It usually takes only one-fourth epoch to converge. We inspected the data and found that top-K sentences produced by beam search lack in diversity, often differing only in a few characters. To further verify our assumption, we conducted an experiment on the MASS model and training data with different sizes.
**Data Size** **RA-WER** **BLEU** **M$^2$**
---------------- ------------ ----------- -----------
LARGE (18.6M) 18.96 74.90 71.05
MODERATE (16M) **17.15** **76.20** **71.76**
SMALL (1.1M) 18.56 74.51 71.43
ORIGIN (1.1M) 24.46 67.28 51.80
: \[dataset-comporison\] Evaluation of MASS model that is fine-tuned on different size training dataset is shown. The values are evaluated on the CoNLL test set. The numbers in the parentheses are the approximate number of sentence pairs. MASS trained on **MODERATE** data achieves the best scores on all metrics. MASS trained on **SMALL** data gets a comparable result to the highest scores with a significantly smaller dataset. **ORIGIN** is the original GEC sentence pairs, which are used as the seed corpus for the APR task.
Table \[dataset-comporison\] shows the results. **SMALL** data only includes the best output of the beam search and has 1.1M sentence pairs. **MODERATE** data contains top-k beams obtained with the beam search and has 16M sentences pairs. **LARGE** is the largest data, which also comprises original GEC pairs and TTS normalized data in addition to all data in **MODERATE**. **LARGE** has 18.6M sentence pairs. To demonstrate the difference between the GEC task and APR task, we also used the original GEC pairs as the training data denoted as **ORIGIN**.
In table \[dataset-comporison\], we can see **MODERATE** data get the best scores on all metrics. That means that including top-k beams indeed helped the APR task. However, by only using **SMALL** data, we still got a comparable result. The remaining data ($\approx$15M) yielded a 1.69 increase on BLEU. This result proves our assumption that the top-k beams obtained with the beam search are homogeneous, which is not very beneficial for the model to learn new patterns from the data. Given these results and efficient usage of computational resources, we used **SMALL** data in the remainder of our experiments. It is interesting that the **LARGE** data got a lower score than the **SMALL** data. We think the cause is the original GEC pairs, and TTS normalized data having different patterns with the ASR output data.
The last row of Table \[dataset-comporison\] is the **ORIGIN** data, which has the same target sentences with the **SMALL** data but differs in source sentences. The MASS model trained on GEC pairs only got 67.28 BLEU, which is much lower than any dataset with ASR output as the source. It shows that the GEC task is different from the APR task, and APR is a new task that deserves a dedicated research effort.
Model Comparison
----------------
In Table \[model\_comparison\], we report the experimental results of four fine-tuned models on **SMALL** dataset (1.1M sentence pairs) and compared them with the baseline method.
Compared with the 2-step pipeline baseline, all the fine-tuned model got better scores on almost all metrics, which proved the effectiveness of considering the APR as a sequence-to-sequence task and utilizing a pre-trained model. The only exception is that MASS got a higher RA-WER (18.37) than the baseline.
In four fine-tuned models, MASS had a lower performance compared to the other three. This is reasonable since MASS only has a 6-layer encoder and 6-layer decoder, which is equivalent to a 12-layer BERT base model with about 110M parameters, while the other three are all based on 24-layer BERT-large model with about 340M parameters. The result proved again that high capacity Transformer architecture had a positive impact on the ASR task. To compare the experimental results fairly, we will focus on three BERT-large based model in the following discussion.
RoBERTa and RoBERTa-UniLM model achieved better scores than UniLM in all metrics except GLEU on JFLEG test sets. Our experiments demonstrated that fine-tuned downstream tasks based on RoBERTa gave better performance than based on BERT, which is consistent with the RoBERTa paper [@liu2019roberta].
For CoNLL2014 test set, the three BERT-large based models got the comparable scores. While RoBERTa won in RA-WER and BLEU metrics, RoBERTa-UniLM won in M$^2$. CoNLL2014 test set includes minimal edits which correct the grammatical errors of a sentence but do not necessarily make it fluent or native-sounding. For JFLEG test set, an interesting finding is that although RoBERTa based model has fewer errors or more smooth than UniLM according to the lower RA-WER (13.88 vs. 17.10) and higher BLEU (80.77 vs. 76.58), the situation is opposite when considering fluency which is evaluated by GLEU (51.78 vs. 51.21). Unlike CoNLL, JFLEG does not restrict corrections to minimal error spans. Instead, it contains holistic sentence rewrites, which represent a diversity of potential corrections with four corrections of each sentence. Therefore, the less error and better alignment are not equivalent to the fluency or sound natural to a native English speaker, which means the metrics from ASR/MT/GEC may not be the best metrics for APR, and we need to find more suitable metrics for APR which considers more readability.
RoBERTa-UniLM did not show significant benefits than RoBERTa overall. One possible reason could be fewer LM training data. RoBERTa used 160GB text while RoBERTa-UniLM used 66GB. Another reason could be the training time. Due to the limitation of computation resources, we only updated the model for 7200 steps, and the training has not fully converged yet. However, even with limited training data and model updates, RoBERTa-UniLM demonstrated its potential in some metrics such as M$^2$ on CoNLL2014 and BLUE on JFLEG, while other metrics had very close scores to RoBERTa. This encouraged us to continue to improve the pre-trained model in the future. In table \[table-examples\], we show some examples generated by RoBERTa-UniLM. In these samples, the model correctly added punctuation, capitalization, changed the number formatting and corrected the recognition errors.
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
**Input:** *when we’re diagram honest out* with certain genetic *disease* are we supposed to disclose this result to our *relatives*
**Output:***When we are diagnosed* with *a* certain genetic *disease*, are we supposed to disclose this result to our *relatives*?
**Input:** *this* has been indicated in report done by *boston university* in *two thousand and twelve eighty percent* of information posted online is *true other* person may *like* three out of ten times in real *life*
**Output:** *This* has been indicated in *a* report done by *Boston University* in *2012*. *80%* of information posted online is *true, although a* person may *lie* three out of ten times in real *life*.
**Input:** *on* the other *hand* although in *turkey* the public transportation is getting *developed* especially in the two big *cities that is an anchor in istanbul* it is still nothing compared to *sweden*
**Output:** *On* the other *hand,* although in *Turkey* the public transportation is getting *developed,* especially in the two big *cities, Ankara and Istanbul,* it is still nothing compared to *Sweden.*
**Input:** *we* have a lot of *paper have* to *write*
**Output:** *We* have a lot of *papers* to *write.*
**Input:** *in* my *SC i* want to focus on how important is for students to learn *facts*
**Output:** *In* my *essay, I* want to focus on how important *it is* for students to learn *facts.*
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: \[table-examples\] Samples generated by the RoBERTa-UniLM model are shown. The italic font indicates the places of modifications.
Conclusion
==========
In this work, we proposed a new NLP task named ASR Post-processing for Readability (APR) that aims to correct grammar mistakes, disfluency, and readibility of the ASR output. We described our process for synthesizing a dataset for the APR task using the GEC datasets as our seed corpus by using TTS and ASR systems. We borrowed metrics from similar tasks and extended WER into readability-aware WER. We experimented with different dataset sizes and compared different models (MASS, UniLM, RoBERTa, RoBERTa-UniLM) with a traditional post-processing system. The results show that the fine-tuned models improved the readability of ASR output significantly, hinting at potential benefits of the APR task. We hope that our findings will encourage other researchers to work on improving readability in speech transcription systems. APR is an interesting research topic that can be considered as a style transfer from informal spoken language to a written, more formal language.
[^1]: [https://github.com/microsoft/MASS]{}
[^2]: [https://github.com/moses-smt/mosesdecoder]{}
[^3]: [https://github.com/microsoft/unilm]{}
[^4]: [https://github.com/pytorch/fairseq]{}
[^5]: [https://dumps.wikimedia.org/enwiki/latest/enwiki-latest-pages-articles.xml.bz2]{}
[^6]: [https://www.smashwords.com/books/category/1/downloads/0/free]{}
[^7]: [http://data.statmt.org/news-crawl/en-doc]{}
[^8]: [https://github.com/huggingface/transformers]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Thomas Decker[^1], Dominik Janzing, and Thomas Beth\
Institut f[ü]{}r Algorithmen und Kognitive Systeme, Universit[ä]{}t Karlsruhe,\
Am Fasanengarten 5, D-76131 Karlsruhe, Germany
date: 'August 19, 2003'
title: 'Quantum circuits for single-qubit measurements corresponding to platonic solids'
---
Introduction
============
A key postulate of textbook quantum mechanics is the assumption that measurements correspond to self-adjoint operators $A$ in such a way that the probability of each possible measurement outcome or set of possible outcomes can be computed from the spectral projections of $A$. If the corresponding system Hilbert space is finite dimensional $A$ can be written as $A=\sum_j \lambda_j P_j$ where $P_j$ is the projection on the eigenspace with eigenvalue $\lambda_j$. The probability of the outcome $\lambda_j$ is ${\rm tr}(\rho P_j)$ if the system is in a state with density matrix $\rho$. This type of measurement is called [*von-Neumann measurement*]{}, [*orthogonal measurement*]{} or [*projector-valued measurement*]{}.
Within the standard model of a quantum computer one can easily show that it is in principle possible to implement measurements for all self-adjoint operators $A$ acting on the Hilbert space $(\C^2)^{\otimes l}$, i.e., the state space of a quantum register with $l$ qubits. Since a universal quantum computer allows the implementation of each unitary transformation one could perform a unitary operation $U$ that diagonalizes $A$ with respect to the computational basis and measure with respect to this basis.
However, the description of measurements by self-adjoint operators is not general enough. Most general measurements are described by positive operator valued measures (POVMs). A POVM is defined as follows [@Davies2]. Let $\Omega$ be the set of possible outcomes and $\Sigma$ be a sigma-algebra of measurable subsets of $\Omega$. Let ${{\cal P}}$ be the set of positive operators acting on the Hilbert space ${{\cal H}}$. Then a POVM $A$ is a map $A: \Sigma \rightarrow
{{\cal P}}, m \mapsto A_m,$ with the following properties:
1. For all countable families $(m_j)$ of mutually disjoint sets $m_j$ one has $$A_{\cup_j m_j} = \sum_j A_{m_j}\,,$$ where the infinite sum converges in the weak operator topology.
2. $A_{\Omega}={\bf 1}$.
The probability for obtaining an outcome in the set $m$ is given by ${\rm tr}(\rho A_m)$. When the set $\Omega$ of possible outcomes is finite or countably infinite a POVM is uniquely given by a family $(A_j)$ of positive operators such that $p_j={\rm tr}(\rho A_j)$ is the probability for obtaining the outcome $j$. We only consider POVMs with a finite set $\Omega$ of outcomes. Furthermore, the considered POVMs have the following properties:
1. The family $(A_j)$ describes a single-qubit measurement, i.e., the system Hilbert space is $\C^2$.
2. Each $A_j$ is a rank-one operator, i.e., $A_j={|\Psi_j\rangle}{\langle \Psi_j|}$. The vectors ${|\Psi_j\rangle}$ have the same length. They are not necessarily normalized.
3. The operators $A_j$ correspond to symmetric points on the Bloch sphere. The symmetry groups are finite subgroups of $SO(3)$. The possible symmetry groups are the cyclic and dihedral groups and the symmetry groups of the platonic solids.
These properties show that we restrict our attention to a rather specific class of symmetric POVMs. The symmetry is fundamental in our constructions of the circuits implementing the POVMs. Specifically, we choose a cyclic subgroup of the symmetry group corresponding to a POVM. Under the action of the cyclic group the set of points on the Bloch sphere decomposes into several orbits. As shown in Section \[Section 2.1\] POVMs given by a single orbit can easily be implemented by a discrete Fourier transform. Since we have several orbits we have to use additional gates besides the Fourier transform to implement the POVM. This explains why the discrete Fourier transform plays a central role in all constructed circuits.
The intention of this paper is to show how the symmetry of a POVM can be used to construct a simple circuit for implementing the POVM. To our knowledge, there are no considerations of the implementation of POVMs besides [@Sasaki]. The investigation of the implementation and its complexity is motivated by the fact that there are examples where generalized measurements can extract more information about an unknown quantum state than projector-valued measurements. Symmetric POVMs may, for instance, be interesting when we want to distinguish between symmetric states [@Sasaki]. Furthermore, POVMs may perform better than orthogonal measurements with respect to appropriate information criteria (e.g. mutual information [@Davies] or the least square error [@Eldar]). Here we do neither consider these ”quality” criteria nor the post-measurement state. The post-measurement state may be relevant in order to understand information-disturbance trade-off relations [@Fuchs].
In the next section we describe the basic principles for implementing arbitrary POVMs. In Section \[Section 2\] we specify the correspondence of POVM operators to points on the Bloch sphere. Furthermore, we specify the symmetry of POVMs. In Sections \[Section 2.1\] and \[Section 2.2\] we consider the implementation of POVMs with a cyclic or dihedral symmetry group, respectively. These considerations are the basis of the implementations of POVMs corresponding to platonic solids. The implementation of these POVMs is discussed in Sections \[S 4\]–\[S ico\].
Orthogonal measurement of POVMs {#Sektion ortho}
===============================
In this section we briefly rephrase Neumark’s theorem describing the reduction of POVMs to orthogonal measurements [@Peres]. This theorem allows to implement POVMs by performing unitary transformations on the joint system consisting of the system to be measured and an ancilla register. The unitary transformations are followed by an orthogonal measurement in the computational basis.
Let $(A_j)$ with $j \in \{1, \ldots , n \}$ be a POVM with corresponding Hilbert space ${\mathbb C}^d$ where each $A_j={|\Psi_j\rangle} {\langle \Psi_j|} \in {\mathbb C}^{d \times d}$ is a positive operator of rank one. Due to the properties of POVMs we have $\sum_j A_j =I_d$ where $I_d$ denotes the identity matrix of size $d$. The choice of corresponding vectors ${|\Psi_j\rangle}$ is not unique since we can multiply each ${|\Psi_j\rangle}$ with a phase factor that is physically irrelevant. It is therefore reasonable to choose the phase factors in such a way that the implementation of the POVM is simplified. Our constructions in Sections 4–10 implicitly make use of this. For $n > d$ the vectors ${|\Psi_j\rangle}$ cannot be mutually orthogonal. As a simple example we consider a system with Hilbert space ${\mathbb C}^2$ and the following vectors: $${|\Psi_1\rangle} = \sqrt{\frac{1}{3}}
\left(\begin{array}{c}1\\1\end{array}\right), \;\; {|\Psi_2\rangle} =
\sqrt{\frac{1}{3}}\left(\begin{array}{c}1\\ \omega
\end{array}\right) \;{\rm and} \;\; {|\Psi_3\rangle}=\sqrt{\frac{1}{3}}
\left(\begin{array}{c}1 \\ \omega^2\end{array} \right).$$ Here is $\omega:= {\rm exp}(-2\pi i /3)$ a third root of unity. We therefore have $$A_1 = \frac{1}{3}\left( \begin{array}{cc}1 & 1 \\ 1 & 1
\end{array} \right), \; \; A_2=\frac{1}{3} \left( \begin{array}{cc}1 &
\omega^2 \\ \omega & 1 \end{array} \right)\; {\rm and} \;\; A_3=
\frac{1}{3}\left(
\begin{array}{cc}1 & \omega \\ \omega^2 & 1 \end{array} \right)$$ as POVM operators. In Section \[Section 2.1\] we consider a generalization of this POVM.
Assuming orthogonal measurements as basic measurements, we have to extend the system by at least $n-d$ dimensions to make a measurement with $n$ different measurement outcomes possible. In order to simplify notation, we consider the given system with $d$ dimensions as a subsystem of a system with $n$ dimensions. Since we are interested in quantum circuits we have to embed the system into a qubit register. This can be done by assuming that the POVM consists of $n=2^l$ operators. Note that this is no loss of generality since we can extend a given POVM by an appropriate number of zero operators $A_j = 0_d \in {\mathbb C}^{d \times d}$ where $0_d$ denotes the zero matrix of size $d$. This extension does not change the probability distribution of the POVM since $p_j = {\rm tr}(\rho 0_d) = 0$ for a zero operator $A_j=0_d$. In our example above we add the zero operator $A_4 = 0_2$ to the three POVM operators. We obtain a POVM that can be implemented on a register of two qubits.
The basic idea of Neumark’s theorem is to implement an orthogonal measurement $({\tilde A}_j)$ on the extended system with $n$ dimensions that corresponds to the POVM $(A_j)$ in the sense that it reproduces the correct probabilities $p_j$. We now consider the construction of the orthogonal measurement $({\tilde A_j})$. Let $\rho
\in {\mathbb C}^{d \times d}$ be the density matrix of the state to be measured. Then the state of the extended system with $n$ dimensions can be written as ${\tilde \rho} = \rho \oplus 0_{n-d} \in {\mathbb
C}^{n \times n}$. When we write the vectors ${|\Psi_j\rangle}$ as columns of the matrix $$M= \left( {|\Psi_1\rangle} \ldots {|\Psi_n\rangle} \right) \in
{\mathbb C}^{d \times n},$$ the operators ${\tilde A}_j = {|{\tilde
\Psi}_j\rangle} {\langle {\tilde \Psi}_j|} \in {\mathbb C}^{n \times n}$ are given by ${|{\tilde \Psi}_j\rangle} = {|\Psi_j\rangle} \oplus {|\Phi_j\rangle} \in
{\mathbb C}^n$. The extended vectors ${|{\tilde \Psi}_j\rangle}$ are the columns of the matrix $${\tilde M} = \left( \begin{array}{ccc} {|\Psi_1\rangle} & \ldots &
{|\Psi_n\rangle} \\ {|\Phi_1\rangle} & \ldots & {|\Phi_n\rangle} \end{array}
\right) \in {\mathbb C}^{n \times n},$$ that is an arbitrary unitary matrix containing $M$ as upper part of size $d \times n$. The extension of $M$ to a unitary matrix ${\tilde M}$ is always possible since the rows of $M$ are orthonormal. This is guaranteed by the fact that each POVM $(A_j)$ satisfies $\sum_j A_j =I_d$. The probability distribution ${\tilde p}_j = {\rm tr}({\tilde \rho} {\tilde A}_j)$ equals the distribution $p_j$ of the original POVM since $${\tilde p}_j = {\rm
tr}\left({\tilde \rho} {\tilde A}_j \right) = {\rm tr}\left(
\left(\rho \oplus 0_{n-d} \right) \left(\begin{array}{cc}
{|\Psi_j\rangle}{\langle \Psi_j|} & {|\Psi_j\rangle} {\langle \Phi_j|} \\ {|\Phi_j\rangle}
{\langle \Psi_j|} & {|\Phi_j\rangle} {\langle \Phi_j|}
\end{array} \right) \right) = {\rm tr}\left( \rho A_j \right) = p_j.$$ In our example, we have $$M= \sqrt{\frac{1}{3}} \left( \begin{array}{cccc} 1 &
1 & 1 & 0 \\ 1 & \omega & \omega^2 & 0 \end{array} \right) \in
{\mathbb C}^{2 \times 4}.$$ A possible unitary extension ${\tilde M}$ of this matrix is given by $${\tilde M} = \sqrt{\frac{1}{3}} \left(
\begin{array}{cccc} 1 & 1 & 1 & 0 \\ 1 & \omega & \omega^2 & 0 \\ 1 &
\omega^2 & \omega & 0 \\ 0 & 0 & 0 & \sqrt{3} \end{array} \right) \in
{\mathbb C}^{4 \times 4}$$ leading to the vectors ${| {\tilde
\Psi}_1\rangle} = \sqrt{1/3}(1,1,1,0)^T$, ${| {\tilde \Psi}_2\rangle} =
\sqrt{1/3}(1,\omega ,\omega^2,0)^T$, ${| {\tilde \Psi}_3\rangle} =
\sqrt{1/3}(1,\omega^2,\omega,0)^T$, and ${| {\tilde \Psi}_4\rangle} =
(0,0,0,1)^T$. With the state ${\tilde \rho} = \rho \oplus
0_2 \in {\mathbb C}^{4 \times 4}$, for instance, we obtain the probability $${\tilde p}_2 = {\rm tr}\left( \left( \begin{array}{cc|cc} \rho_{11}
& \rho_{12} & 0 & 0 \\ \rho_{21} & \rho_{22} & 0 & 0 \\ \hline0 & 0 &
0 & 0 \\ 0 & 0 & 0& 0 \end{array} \right) \left(
\begin{array}{cccc} 1 & \omega^2 & \omega & 0 \\ \omega & 1 & \omega^2
& 0 \\ \omega^2 & \omega & 1 & 0 \\ 0 & 0 & 0 & 0 \end{array} \right)
\right) = {\rm tr}\left( \rho \left( \begin{array}{cc} 1 &
\omega^2 \\ \omega & 1 \end{array} \right) \right) = p_2$$ for the second POVM operator. The probabilities for the other POVM operators are computed similarly.
The implementation of a POVM with corresponding matrix $M$ is obtained by the orthogonal measurement in the computational basis after performing the unitary transformation ${\tilde M}^\dagger$ on the system with initial state ${\tilde \rho}$. This unitary operation maps the vector ${| {\tilde \Psi}_j\rangle} = {|\Psi_j\rangle} \oplus
{|\Phi_j\rangle}$ to the computational basis vector ${|j\rangle}$. Therefore, the measurement in the computational basis after applying ${\tilde M}^\dagger$ corresponds to the measurement in the basis defined by the vectors ${|{\tilde \Psi}_j\rangle}$.
In summary, we are interested in constructing and implementing the matrix ${\tilde M}^\dagger$ for a given POVM corresponding to the matrix $M$. In the following sections the construction of the matrices ${\tilde M}^\dagger$ is considered for symmetric POVMs besides the decomposition of ${\tilde M}^\dagger$ into elementary (one- and two-qubit) gates. The symmetry leads to simple constructions and implementations based on Fourier transforms.
Symmetric POVMs on a single qubit {#Section 2}
=================================
As described in the previous section, the basis of the orthogonal measurement of a POVM with corresponding matrix $M$ is the implementation of ${\tilde M}^\dagger$. ${\tilde M}$ is a unitary extension of $M$. We can apply the algorithm in Section 4.5.1 of [@Nielsen] to obtain a quantum circuit for ${\tilde
M}^\dagger$. The algorithm decomposes the matrix ${\tilde M}^\dagger$ into a product of two-level matrices that can be translated into a sequence of elementary gates, i.e., each gate operates on one or two qubits. In general, the constructed circuit for ${\tilde M}^\dagger$ is of exponential size in the number $n$ of POVM operators. Intuitively, some symmetry properties of the considered POVMs may lead to algorithms constructing smaller circuits than the standard algorithm that works for arbitrary unitary matrices. To specify the symmetry of POVMs on a single qubit using geometric concepts, we use the correspondence of POVMs to points on the Bloch sphere as already mentioned in the introduction. Usually, each point on the Bloch sphere is considered as a pure state. Specifically, a pure state $\rho \in{\mathbb C}^{2\times 2}$ corresponds to the point $(x,y,z)^T \in {\mathbb R}^3$ on the Bloch sphere with $$\left(
\begin{array}{c}x\\y\\z \end{array} \right) = \left(
\begin{array}{c} {\rm tr}(\sigma_x \rho ) \\ {\rm tr}
(\sigma_y \rho) \\ {\rm tr}( \sigma_z \rho) \end{array} \right)$$ where $$\sigma_x=\left( \begin{array}{cc}0&1\\1&0\end{array}\right),
\sigma_y=\left(\begin{array}{cc}0&-i\\i&0\end{array}\right), \; {\rm
and} \; \sigma_z=\left( \begin{array}{cc}1&0\\0&-1\end{array}
\right)$$ denote the Pauli spin matrices. Conversely, the point $(x,y,z)^T \in {\mathbb R}^3$ on the Bloch sphere corresponds to the density matrix $$\frac{1}{2}\left(\begin{array}{cc}1+z & x-iy \\ x+iy & 1-z
\end{array} \right) \in {\mathbb C}^{2\times 2}.$$ For some special states the points on the Bloch sphere are shown in Figure
\[Figur Blochkugel\]. We now extend the Bloch sphere representation for states to a representation of POVM operators of rank one. Note that each pure state is a projection $\rho={|\Psi\rangle}{\langle \Psi|}$ of rank one. By rescaling a POVM operator $A_j =
{|\Psi_j\rangle}{\langle \Psi_j|}$ to a density matrix we can identify $A_j$ with a point on the Bloch sphere using the correspondence for pure states.
A symmetry of a POVM can be defined by a symmetry of the corresponding points on the Bloch sphere. We are interested in POVMs with a finite symmetry group for the points on the Bloch sphere, i.e., we consider finite subgroups of $SO(3)$ [@Sternberg]. There are two infinite families of finite subgroups, namely the cyclic groups $C_m$ and the dihedral groups $D_m$ for $m\geq 2$ as symmetry groups of an $m$-sided regular polygon (with the special case $m=2$). Furthermore, we have the symmetry groups of the five platonic solids (tetrahedron, cube, octa-, dodeca-, and icosahedron). As a restriction for the latter symmetry groups, we assume the points on the Bloch sphere of a POVM to coincide with the vertices of the platonic solid corresponding to the symmetry group.
The vertices of the regular polygons and the platonic solids depend on the orientation of the polygons and platonic solids in the Bloch sphere. Each orientation leads to another POVM. In order to simplify the constructions in the following sections we choose specific orientations of the regular polygons and platonic solids. To obtain the implementation of a POVM corresponding to the same polygon or platonic solid with another orientation, it suffices to implement a single qubit operation on the qubit to be measured.
Cyclic groups {#Section 2.1}
=============
The simplest finite symmetry groups of points on the Bloch sphere are the cyclic groups. For a fixed $m\geq 2$ we consider the rotations of an $m$-sided regular polygon with a common axis perpendicular to the face of the polygon. The rotations form a group that is isomorphic to the group $C_m=
\langle r \rangle$ with $r^m=1$. The implementation of POVMs corresponding to a single orbit of points under the action of a cyclic symmetry group is the basis of all constructions in the following sections.
In principle, we can choose an arbitrary orientation of the polygon corresponding to the cyclic symmetry group. For simplification, we choose the face of the regular polygon to be perpendicular to the $z$-axis in the Bloch sphere. In other words, the $z$-axis is the common axis of the rotations. For instance, the $5$-sided regular polygon is shown in Figure \[Figur cyclic vectors\].
The cyclic symmetry group of the polygon is generated by the $2\pi/m$ rotation about the $z$-axis. In the Hilbert space, this rotation of the Bloch sphere corresponds to the matrix ${\rm diag}(1,\omega) \in
{\mathbb C}^{2 \times 2}$ where $\omega := {\rm exp}(- 2 \pi i/m)$ is an $m$th complex root of unity. The diagonal form of the matrix is the reason for choosing the $z$-axis as common rotation axis. If we choose the vector $(1,1)^T \in {\mathbb C}^2$ and consider the orbit under the symmetry group then we get the vectors $(1, \omega^{j})^T
\in {\mathbb C}^2$ for $j \in \{0, \ldots, m-1\}$. Other vectors of the Bloch sphere do not lead to POVMs or to POVMs that correspond to a polygon with another rotation. The latter case is discussed at the end of the previous section. Due to the identity $$\sum_{j=0}^{m-1} \left(
\begin{array}{c} 1 \\ \omega^j \end{array} \right)
(1, \omega^{-j}) = \sum_{j=0}^{m-1} \left(
\begin{array}{cc}1&\omega^{-j}\\ \omega^{j}&1 \end{array} \right) = m
I_2$$ the elements ${|\Psi_j\rangle}= \sqrt{1/m}(1,\omega^{j-1} )^T \in
{\mathbb C}^2$ for $j \in \{1, \ldots, m\}$ define a POVM with $m=n$ operators on a qubit. Therefore, the unitary matrix $\tilde{M} \in
{\mathbb C}^{m \times m}$ is a unitary extension of the matrix $$M=\sqrt{\frac{1}{m}} \left( \begin{array}{cccc}
1&1&\ldots&1\\1&\omega& \ldots & \omega^{m-1} \end{array} \right) \in
{\mathbb C}^{2 \times m}.$$ $M$ corresponds to the first two rows of the discrete Fourier matrix $$F_m=\sqrt{\frac{1}{m}}\left(\omega^{jk}\right)_{j,k=0}^{m-1}
\in {\mathbb C}^{m \times m}$$ of size $m$. Consequently, by considering the qubit to be measured as a subsystem of an $m$-dimensional system the implementation of the inverse Fourier transform ${\tilde M}^\dagger = F_m^\dagger$ leads to the probability distribution of the cyclic POVM on the qubit.
For the construction of a quantum circuit we have to embed the system of dimension $m$ into a register with $l$ qubits. The register must have $r := 2^l \geq m$ dimensions. Following Section \[Sektion ortho\] we extend the cyclic POVM by an appropriate number of zero operators $A_{m+1}, \ldots , A_{r} = 0_2 \in {\mathbb C}^{2 \times
2}$. We therefore have $$M = \sqrt{\frac{1}{m}}\left(
\begin{array}{cccc|ccc} 1 & 1 & \ldots & 1 & 0 & \ldots & 0 \\ 1 &
\omega & \ldots & \omega^{m-1} & 0 & \ldots & 0 \end{array} \right)
\in {\mathbb C}^{2 \times r}.$$ A possible unitary extension ${\tilde
M}$ of this matrix is given by ${\tilde M} = F_m \oplus I_{r-m} \in
{\mathbb C}^{r \times r}$ where $I_{r-m}$ denotes the identity matrix of size $r-m$. Consequently, on a qubit register the cyclic POVM corresponding to the $m$-sided regular polygon can be implemented by performing the operation ${\tilde M}^\dagger = F_m^\dagger \oplus I_{r
- m}$. The circuit for implementing the cyclic POVM is schematically shown in Figure \[Figur DFT\]. Note that the embedding $\rho \mapsto
\rho \oplus 0_{r-m}$ corresponds to the use of initialized ancilla qubits.
The Fourier transform can be implemented efficiently if $m$ is a power of two [@Nielsen]. Furthermore, the embedding into a qubit register is straightforward since we do not need zero operators in this case. In summary, the cyclic POVM can be implemented efficiently on a qubit register if $m$ is a power of two. For instance, the quantum circuit for the implementation of the cyclic POVM is shown in Figure \[Figur DFT\] for $m=4$. The circuit of $F_4^\dagger$ is the standard circuit for Fourier transforms [@Nielsen]. Note that the first permutation of the qubits can be removed when we change the order of the input.
Dihedral groups {#Section 2.2}
===============
The cyclic symmetry group of an $m$-sided regular polygon which we considered in the previous section is a subgroup of the dihedral group. The dihedral group consists of all rotations which map the $m$-sided regular polygon onto itself. In contrast to the cyclic group we allow the rotations to have different axes. For a fixed $m\geq 2$, the dihedral group is isomorphic to $D_m=\langle r,s\rangle$ with $r^m=1$, $s^2=1$, and $srs=r^{-1}$. In order to use the results for the cyclic groups, we consider the same orientation of the regular polygon as in the previous section, i.e., the face of the polygon is orthogonal to the $z$-axis. Furthermore, we assume that at least one vertex is an element of the $x$-axis. Due to this orientation, the element $r$ corresponds to the $2 \pi/m$ rotation about the $z$-axis and the element $s$ corresponds to the $\pi$ rotation about the $x$-axis. In the Hilbert space these rotations correspond to the matrices $$\left(\begin{array}{cc}1 &0 \\ 0& \omega \end{array} \right) \; {\rm
and} \; \; \left( \begin{array}{cc}0&1\\1&0\end{array}\right)$$ where $\omega:={\rm exp}(-2\pi i/m)$ is a $m$th complex root of unity. We can define a projective representation of the group $D_m$ by mapping the element $r \in D_m$ to the first matrix and the element $s \in D_m$ to the second matrix.
We consider the orbit of a vector under the action of the dihedral group $D_m$. Let $(\alpha, \beta)^T \in {\mathbb C}^2$ with $|\alpha|^2 + |\beta|^2 =1$. Since a global phase factor of a vector is physically irrelevant we assume $\alpha \in {\mathbb R}$ without loss of generality. Under the action of the dihedral group the orbit contains the vectors $(\alpha,\beta \omega^j)^T$ and $(\beta, \alpha
\omega^{j})^T$ with $j \in \{0, \ldots , m-1\}$. An example of the orbit is shown in Figure \[Figur diederpunkte\].
We have at most $n=2m$ vectors. In the following, we assume that the orbit contains $2m$ elements. If the orbit of $D_m$ contains less than $2m$ points we have either the case that all points are on the $xy$-plane (and the POVM consists of a single orbit under the group $C_m$) or we have only the two points $(1,0)^T$ and $(0,1)^T$ defining an orthogonal measurement. Since $$\sum_j \left(
\begin{array}{c} \alpha \\ \beta \omega^j \end{array} \right) (\alpha,
{\overline \beta} \omega ^{-j}) + \sum_j \left(
\begin{array}{c} \beta \\ \alpha \omega^j \end{array} \right)
({\overline \beta}, \alpha \omega^{-j}) = m I_2,$$ we rescale $\alpha$ and $\beta$ with the factor $\sqrt{1/m}$ to obtain a POVM.
We now consider the implementation of the dihedral POVM. In order to analyze the structure, we do not consider the embedding of the constructed system into a qubit register in the first place. The orbit under the action of $D_m$ breaks into two orbits under the action of the subgroup $C_m$. The two orbits can be obtained by the action of $C_m$ on the vectors $(\alpha, \beta)^T$ and $(\beta,
\alpha)^T$. Therefore, we expect to obtain implementations of the dihedral POVMs which are similar to the implementations in the previous section. With an appropriate order of the vectors we have the matrix $$M= \left(
\begin{array}{cccc|cccc} \alpha & \alpha & \ldots & \alpha
& \beta & \beta & \ldots & \beta \\ \beta & \beta \omega & \ldots &
\beta \omega^{m-1} & \alpha & \alpha \omega & \ldots & \alpha
\omega^{m-1} \end{array} \right) \in {\mathbb C}^{2 \times 2m}.$$ For even $m$, this matrix can be extended to the unitary matrix $${\tilde M} = Q \left( \begin{array}{cccc|cccc} \alpha &
\alpha & \ldots & \alpha & \beta & \beta & \ldots & \beta \\ \alpha &
\alpha \omega & \ldots & \alpha \omega^{m-1} & -{\overline \beta} &
-{\overline \beta} \omega & \ldots & -{\overline \beta} \omega^{m-1} \\
\alpha & \alpha \omega^2 & \ldots & \alpha \omega^{2(m-1)} & \beta &
\beta \omega^2 & \ldots & \beta \omega^{2(m-1)} \\ \vdots &&& \vdots &
\vdots &&& \vdots \\ \alpha & \alpha \omega^{m-1} & \ldots & \alpha
\omega^{(m-1)(m-1)} & -{\overline \beta} & -{\overline \beta}
\omega^{m-1} & \ldots & -{\overline \beta}
\omega^{(m-1)(m-1)} \\ \hline &&&&&&\\ [-0.35cm]
{\overline \beta} & {\overline \beta} & \ldots &
{\overline \beta} & -\alpha & -\alpha & \ldots & -\alpha \\ \beta &
\beta \omega & \ldots & \beta \omega^{m-1} & \alpha & \alpha \omega
& \ldots & \alpha \omega^{m-1} \\ {\overline \beta} & {\overline
\beta} \omega^2 & \ldots & {\overline \beta} \omega^{2(m-1)} & -\alpha
& -\alpha \omega^2 & \ldots & -\alpha \omega^{2(m-1)} \\ \vdots &&&
\vdots & \vdots &&& \vdots \\ \beta & \beta
\omega^{m-1} & \ldots & \beta \omega^{(m-1)(m-1)} &
\alpha & \alpha \omega^{m-1} & \ldots & \alpha \omega^{(m-1)(m-1)}
\end{array} \right)$$ with a permutation matrix $Q \in {\mathbb C}^{n \times
n}$ fixing the first row and mapping the $(m+2)$nd row to the second row. For odd $m$, the extended matrix is similar. We only have to write $\beta \omega^{(m-1)j}$ instead of $- {\overline \beta}
\omega^{(m-1)j}$ in the $m$th row. In the last row we write ${\overline \beta} \omega^{(m-1)j}$ and $-\alpha \omega^{(m-1)j}$ instead of $\beta \omega^{(m-1)j}$ and $\alpha \omega^{(m-1)j}$, respectively. In order to simplify notation, we mainly consider the case of even $m$ in the following. The constructions for odd $m$ are similar.
We consider a decomposition of the matrix $Q^\dagger {\tilde M}$ to obtain a decomposition of ${\tilde M}$. The matrix $Q^\dagger {\tilde M}$ can be multiplied with $I_2 \otimes F_m^\dagger$ from the right leading to $$T= \sqrt{m}\left( \begin{array}{ll} {\rm diag}(\alpha, \alpha,
\alpha, \ldots, \alpha, \alpha ) & {\rm diag}(+\beta, -{\overline
\beta}, +\beta, \ldots, +\beta, -{\overline \beta})
\\ {\rm diag}({\overline \beta}, \beta, {\overline \beta}, \ldots,
{\overline \beta} , \beta) &
{\rm diag}(-\alpha,+ \alpha, -\alpha, \ldots, -\alpha, + \alpha )
\end{array} \right).$$ We now embed the system with $n=2m$ dimensions into a qubit register. We consider a register with $l$ qubits where $r := 2^l \geq n$. We replace the matrix $T$ with the matrix $T_r$ of the same structure but of size $r$. This is done by extending each of the four diagonal components to a diagonal matrix in ${\mathbb C}^{(r/2) \times (r/2)}$ while conserving the structure. For instance, the matrix $$T= \sqrt{3}\left(
\begin{array}{ccc|ccc} \alpha &&&\beta\\ &\alpha &&&-{\overline \beta}
\\&&\alpha &&& \beta \\ \hline &&&&\\ [-0.35cm] {\overline \beta} &&&
-\alpha \\ & \beta &&& \alpha \\ && {\overline \beta} &&& -\alpha
\end{array}\right)$$ is extended to the matrix $$T_8=\sqrt{3}\left( \begin{array}{cccc|cccc}
\alpha &&&&\beta\\ &\alpha &&&&-{\overline \beta} \\&&\alpha &&&&
\beta \\ &&& \alpha &&&& -{\overline \beta}
\\ \hline &&&&\\ [-0.35cm] {\overline \beta} &&&& -\alpha \\ & \beta
&&&& \alpha \\ && {\overline \beta} &&&& -\alpha\\ &&& \beta &&&& \alpha
\end{array}\right).$$ Furthermore, in the factorization $T=Q^\dagger {\tilde M} (I_2 \otimes
F_m^\dagger)$ the matrix $Q$ is replaced by a permutation matrix $Q_r
\in {\mathbb C}^{r \times r}$ that fixes the first row and maps the $(r/2+2)$nd row to the second row. In qubit notation, this permutation matrix can be described by ${|0 \ldots 0\rangle} \mapsto {|0 \ldots 0\rangle}$ and ${|10 \ldots 01\rangle} \mapsto {|00 \ldots 01\rangle}$. This permutation can be implemented by an XOR-gate on the first qubit controlled by the last qubit. Other implementations that satisfy the two constraints are also possible. The Fourier transform $F_m$ is replaced by $F_m \oplus
I_{r/2-m}$. In summary, we obtain a matrix ${\tilde M}_r$ that is defined by the equation $$\label{Gleichung}
{\tilde M}_r = Q_r T_r (I_2 \otimes (F_m \oplus I_{r/2-m})) \in {\mathbb
C}^{r \times r}.$$ This matrix is a unitary extension of the matrix $M$ corresponding to the dihedral POVM with some zero operators as discussed in Section \[Sektion ortho\]. Our example with $T_8$ leads to the matrix $$Q_8^\dagger {\tilde M}_8 = \left(
\begin{array}{cccc|cccc} \alpha & \alpha & \alpha & 0 & \beta & \beta
& \beta & 0 \\ \alpha & \alpha \omega & \alpha \omega^2 & 0 &
-{\overline \beta} & -{\overline \beta} \omega & -{\overline \beta}
\omega^2 & 0 \\ \alpha & \alpha \omega^2 & \alpha \omega & 0 & \beta &
\beta \omega^2 & \beta \omega & 0 \\ 0 & 0 & 0 & \sqrt{3} \alpha & 0 &
0& 0& -\sqrt{3} \, {\overline \beta} \\
\hline &&&& \\ [-0.35cm] {\overline \beta} &
{\overline \beta} & {\overline \beta} & 0 & - \alpha & -\alpha &
-\alpha & 0 \\ \beta & \beta \omega & \beta \omega^2 & 0 & \alpha &
\alpha \omega & \alpha \omega^2 & 0 \\ {\overline \beta} & {\overline
\beta} \omega^2 & {\overline \beta} \omega & 0 & - \alpha & - \alpha
\omega^2 & -\alpha \omega & 0 \\ 0 & 0 & 0 & \sqrt{3} \beta & 0 & 0&
0& \sqrt{3}\alpha \end{array}\right).$$ The matrix $Q_8$ maps the sixth row to the second row leading to the first two rows $$\left( \begin{array}{cccc|cccc} \alpha & \alpha &
\alpha & 0 & \beta & \beta & \beta & 0 \\ \beta & \beta \omega & \beta
\omega^2 & 0 & \alpha & \alpha \omega & \alpha \omega^2 & 0
\end{array} \right)$$ with two zero columns that do not change the POVM due to zero probability. For convenience, we shift the qubits according to the mapping ${|x_1 \ldots x_{l-1} x_l\rangle} \mapsto {|x_l x_1 \ldots x_{l-1}\rangle}$. We denote this permutation by $R$. After this reordering of qubits the matrix $T_r$ takes the simple form $$RT_rR^\dagger = A \otimes I_{r/4} :=
\sqrt{m}\left( \begin{array}{cccc} \alpha & \beta & 0 &
0 \\ {\overline \beta} & -\alpha & 0 & 0 \\ 0 & 0 &
\alpha & -{\overline \beta} \\ 0 & 0 & \beta & \alpha
\end{array} \right) \otimes
I_{r/4} .$$ By combining this equation with Equation (\[Gleichung\]) we get the factorization $${\tilde M}^\dagger_r = (I_2 \otimes (F_m^\dagger
\oplus I_{r/2-m})) R^\dagger (A^\dagger \otimes I_{r/4}) R
Q_r^\dagger.$$ Translating this equation into a quantum circuit, the decomposition of ${\tilde M}_r$ leads to the circuit scheme shown in Figure \[Figur schaltkreis dieder-2pt\].
The operation $A^\dagger$ is decomposed as $$A^\dagger= \sqrt{m} \left( \begin{array}{cccc}\alpha &
\beta & 0 & 0 \\ {\overline \beta} & - \alpha & 0 & 0 \\ 0 & 0 &
\sqrt{1/m} & 0 \\ 0 & 0 & 0 & \sqrt{1/m}
\end{array} \right)
\sqrt{m} \left( \begin{array}{cccc} \sqrt{1/m} & 0
& 0 &0 \\ 0 & \sqrt{1/m} & 0 & 0 \\ 0 & 0 & \alpha & {\overline \beta}
\\ 0 & 0 & - \beta & \alpha \end{array} \right)$$ corresponding to the second and third gates from the left in Figure \[Figur schaltkreis dieder-2pt\]. We do not have to implement the permutation $R$ explicitly if the controlled one-qubit operations are applied to appropriate qubit pairs. The given circuit can be slightly simplified by merging the first two gates from the left to a single controlled gate with the operation $$\sqrt{m}\left( \begin{array}{cc} \alpha & {\overline \beta} \\ -
\beta & \alpha\end{array} \right) \left( \begin{array}{cc}0&1\\1&0
\end{array} \right) = \sqrt{m} \left(
\begin{array}{cc}{\overline \beta} & \alpha \\ \alpha & -\beta
\end{array} \right).$$ As discussed in the previous section the Fourier transform $F_m$ can be implemented with a polylogarithmical number of gates if $m$ is a power of two. Consequently, the dihedral POVM can be implemented efficiently in these cases.
Tetrahedron {#S 4}
===========
The tetrahedron is the platonic solid with four faces. The symmetry group of the tetrahedron is isomorphic to the alternating group $A_4$. This group consists of the twelve permutations of four elements with positive signum. We consider the POVM corresponding to the vertices of the tetrahedron in the Bloch sphere. The tetrahedron is shown Figure \[Figur tetra-schraube\].
For instance, the vertex $1$ is given by the vector $(\sqrt{2/3},0,
\sqrt{1/3})^T \in {\mathbb R}^3$. The vertices of the tetrahedron correspond to the vectors $$\label{Gleichung 2}
\left( \begin{array}{c} \alpha \\ \beta \end{array} \right), \left(
\begin{array}{c} \alpha \\ - \beta \end{array} \right), \left(
\begin{array}{c} \beta \\ \alpha i\end{array} \right), \left(
\begin{array}{c} \beta \\ - \alpha i \end{array} \right) \in {\mathbb C}^2$$ with $\alpha = \sqrt{(3+\sqrt{3})/6}$ and $\beta = \sqrt{
(3-\sqrt{3})/6}$. The first pair of vectors corresponds to the vertices $1$ and $2$, the second pair corresponds to the vertices $3$ and $4$. Note the similarity of these vectors to the vectors $$\left( \begin{array}{c} \alpha \\ \beta \end{array} \right), \left(
\begin{array}{c} \alpha \\ - \beta \end{array} \right), \left(
\begin{array}{c} \beta \\ \alpha \end{array} \right), \left(
\begin{array}{c} \beta \\ - \alpha \end{array} \right).$$ These vectors result from the action of the dihedral group with $m=2$ as considered in previous section with the vector $(\alpha,
\beta)^T$. The factor $i$ in the second component of the last two vectors of Line (\[Gleichung 2\]) results from the $\pi/2$ rotation about the $z$-axis of the lower edge with vertices 3 and 4 relative to the upper edge with vertices 1 and 2. This rotation corresponds to the matrix ${\rm diag}(1,i) \in {\mathbb C}^{2 \times 2}$. Due to the equation $$\left(\begin{array}{c} \alpha \\ \beta \end{array} \right) (\alpha,
\beta) + \left(\begin{array}{c} \alpha \\ - \beta \end{array} \right)
(\alpha, -\beta) + \left(\begin{array}{c} \beta \\ \alpha i
\end{array} \right) (\beta, - \alpha i ) + \left(\begin{array}{c}
\beta \\ - \alpha i
\end{array} \right) (\beta, \alpha i) = 2 I_2$$ we have the matrix $$M=\left(
\begin{array}{cc|cc} \alpha & \alpha & \beta & \beta \\ \beta & -\beta
& \alpha i & - \alpha i \end{array} \right) \in {\mathbb C}^{2 \times
4}$$ with the rescaled elements $\alpha = \sqrt{
(3 + \sqrt{3})/12}$ and $\beta = \sqrt{ (3 -\sqrt{3})/12}$. This matrix can be extended to the unitary matrix $${\tilde M} = Q \left(
\begin{array}{cc|cc} \alpha & \alpha & \beta & \beta \\ \alpha & -
\alpha & - \beta \, i & \beta \, i \\ \hline \beta & \beta & - \alpha &
-\alpha \\ \beta & - \beta & \alpha \, i & -\alpha \, i \end{array}
\right) \in {\mathbb C}^{4 \times 4}$$ acting on a register of two qubits with the permutation matrix $$Q=\left(
\begin{array}{cccc} 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 1 \\ 0 & 0 & 1 & 0 \\
0 & 1 & 0 & 0 \end{array} \right) \in {\mathbb C}^{4 \times 4}.$$ The matrix $Q$ can be implemented by an XOR-gate on the first qubit controlled by the second qubit. We consider the decomposition of $Q^\dagger {\tilde M}$ to obtain a decomposition of ${\tilde
M}$. After multiplying $Q^\dagger {\tilde M}$ with $(I_2 \otimes F_2)
\in {\mathbb C}^{4 \times 4}$ we have $$Q^\dagger {\tilde M}(I_2
\otimes F_2 ) = \sqrt{2}\left( \begin{array}{cc|cc} \alpha & 0 & \beta
& 0 \\ 0 & \alpha & 0 & - \beta \, i \\ \hline \beta & 0 & -\alpha & 0
\\ 0 & \beta & 0 & \alpha \, i \end{array} \right) \in {\mathbb C}^{4
\times 4}$$ and after multiplying this matrix with ${\rm
diag}(1,1,1,i) \in {\mathbb C}^{4 \times 4}$
from the right we have the equation $$\label{Gleichung 3}
Q^\dagger {\tilde M} (I_2 \otimes F_2) \left(
\begin{array}{cccc} 1&0&0&0 \\ 0&1&0&0\\ 0&0& 1&0\\ 0&0&0&i
\end{array} \right) =
\left( \sqrt{2} \left( \begin{array}{cc} \alpha & \beta \\ \beta & - \alpha
\end{array} \right) \otimes I_2 \right) \in {\mathbb C}^{4 \times
4}.$$ The matrix ${\rm diag}(1,1,1,i)$ corresponds to a controlled phase gate ${\rm diag}(1,i)$ on the second qubit. Using Equation (\[Gleichung 3\]) we get the equation $${\tilde M}^\dagger =
(I_2 \otimes F_2) \left(
\begin{array}{cccc}1&0&0&0\\0&1&0&0\\0&0&1&0\\0&0&0&i\end{array}\right)
\left( \sqrt{2} \left(
\begin{array}{cc} \alpha & \beta \\ \beta & - \alpha \end{array}
\right) \otimes I_2 \right) Q^\dagger.$$ Consequently, the circuit in Figure \[Figur tetra-schraube-kreis\] implements the transformation ${\tilde M}^\dagger$ for the POVM corresponding to the tetrahedron.
Cube {#S 6}
====
The POVM associated with a cube in the Bloch sphere is a special case of the dihedral POVMs considered in Section \[Section 2.2\] with $m=4$. Nevertheless, we consider the implementation of the cubic POVM in this section since we can obtain a smaller circuit by using the special values of $\alpha$ and $\beta$. As in Section \[Section 2.2\] we rotate the cube in the Bloch sphere to obtain a face perpendicular to the $z$-axis. Furthermore, we can rotate the cube about this axis to get points corresponding to the vectors $$\left( \begin{array}{c} \alpha \\ \beta \end{array} \right), \left(
\begin{array}{c} \alpha \\ \beta i \end{array} \right), \left(
\begin{array}{c} \alpha \\ - \beta \end{array} \right), \left(
\begin{array}{c} \alpha \\ - \beta i \end{array} \right), \left(
\begin{array}{c} \beta \\ - \alpha \end{array} \right), \left(
\begin{array}{c} \beta \\ - \alpha i \end{array} \right), \left(
\begin{array}{c} \beta \\ \alpha \end{array} \right), \left(
\begin{array}{c} \beta \\ \alpha i\end{array} \right) \in {\mathbb
C}^2$$ with $\alpha=\sqrt{(3+\sqrt{3})/6}$ and $\beta = \sqrt{(3-\sqrt{3})/6}$. The first four vectors correspond to vertices 1–4 in Figure \[Figur cube\] and
the last four vectors correspond to vertices 5–8. For instance, the vertex $1$ corresponds to the Bloch point $
(\sqrt{2/3},0,\sqrt{1/3})^T \in {\mathbb R}^3$. Note that $\alpha$ and $\beta$ are real numbers. This allows us to use a more efficient construction than in Section \[Section 2.2\]. Since we have the equation $$\left(\begin{array}{c}\alpha\\ \beta \end{array} \right) ( \alpha,
\beta) + \ldots + \left(\begin{array}{c}\alpha\\ - \beta i\end{array}
\right) ( \alpha, \beta i) + \left(\begin{array}{c}\beta\\ -\alpha
\end{array} \right) ( \beta, -\alpha) + \ldots +
\left(\begin{array}{c}\beta\\ \alpha i \end{array} \right) ( \beta, -
\alpha i) = 4 I_2$$ the given vectors define a POVM when we rescale $\alpha$ and $\beta$ with $1/2$. The matrix $M$ corresponding to the POVM is given by $$M = \left(
\begin{array}{cccc|cccc}\alpha & \alpha & \alpha & \alpha & \beta &
\beta & \beta & \beta \\ \beta & \beta i & - \beta & - \beta i & -
\alpha & - \alpha i & \alpha & \alpha i \end{array} \right) \in
{\mathbb C}^{2 \times 8}.$$ In contrast to the construction of the dihedral POVM we use the fact that for even $m$ the element $-1$ is in the set $\{1, \omega,\ldots, \omega^{m-1} \}$ where $\omega:={\rm
exp}(-2\pi i/m)$ is an $m$-th complex root of unity. Therefore, we can reorder the vectors $({\overline \beta}, \alpha \omega^j)^T \in
{\mathbb C}^2$ considered in Section \[Section 2.2\] to obtain a matrix $M$ with the partial row $(-\alpha, -\alpha \omega, \ldots,
-\alpha \omega^{m-1})$ instead of $(\alpha, \alpha \omega, \ldots,
\alpha \omega^{m-1})$. This is besides $\beta \in {\mathbb R}$ the second reason that allows a more efficient construction compared to the construction for the dihedral POVM. Using the equation $\beta =
{\overline \beta}$ we can extend $M$ to the unitary matrix $${\tilde M} = Q \left( \begin{array}{cccc|cccc}
\alpha & \alpha & \alpha & \alpha & \beta & \beta & \beta & \beta \\
\alpha & \alpha i& -\alpha & -\alpha i& \beta & \beta i& -\beta & -\beta i\\
\alpha & -\alpha & \alpha &-\alpha & \beta & -\beta & \beta & -\beta \\
\alpha & -\alpha i& -\alpha & \alpha i& \beta & -\beta i& -\beta & \beta i\\
\hline \beta&\beta&\beta&\beta&-\alpha&-\alpha&-\alpha&-\alpha\\
\beta&\beta i&-\beta&-\beta i&-\alpha&-\alpha i&\alpha&\alpha i\\
\beta&-\beta&\beta&-\beta&-\alpha&\alpha&-\alpha&\alpha\\
\beta&-\beta i& -\beta&\beta i&-\alpha&\alpha i&\alpha&-\alpha i
\end{array}\right)$$ acting on a register of three qubits with a permutation $Q$ satisfying ${|000\rangle} \mapsto {|000\rangle}$ and ${|101\rangle} \mapsto {|001\rangle}$ in qubit notation. For instance, the permutation can be implemented by a single XOR-gate on the first qubit controlled by the third. We now consider the special structure of $Q^\dagger {\tilde M}$ to obtain a decomposition of ${\tilde
M}^\dagger$. More precisely, the matrix $Q^\dagger {\tilde M}$ can be written as the following tensor product $$Q^\dagger {\tilde M} =
\left( 2\left(
\begin{array}{cc}\alpha & \beta \\ \beta & - \alpha \end{array}
\right) \otimes F_4\right) \in {\mathbb C}^{8 \times 8}.$$ This leads to the identity $${\tilde M}^\dagger =
\left( 2 \left( \begin{array}{cc}\alpha & \beta \\ \beta & - \alpha
\end{array} \right) \otimes F_4^\dagger \right) Q^\dagger
\in {\mathbb C}^{8 \times 8}$$ defining the quantum circuit given in Figure \[Figur Cube-speciale-kreis\].
Compared to the general circuit in Section \[Section 2.2\] we are able to replace two controlled gates by a single uncontrolled gate.
Octahedron {#S Okta}
==========
The symmetry group of the octahedron is identical to the symmetry group of the cube since the octahedron is the dual polyhedron of the cube. The group is isomorphic to the symmetric group $S_4$. This group consists of all $24$ permutations of four elements. A simple implementation of the octahedral POVM can be obtained by the orientation of the octahedron as shown in Figure \[Figur Oktaeder\] where the upper face with vertices 1–3 and the lower face with vertices 4–6 are perpendicular to the $z$-axis. Vertex $1$ corresponds to the real vector $(\sqrt{2/3},0,\sqrt{1/3})^T \in
{\mathbb R}^3$. The complex vectors corresponding to the points 1–6 are given by
$$\left(\begin{array}{c} \alpha \\ \beta \end{array} \right),
\left(\begin{array}{c} \alpha \\ \beta \omega \end{array} \right),
\left(\begin{array}{c} \alpha \\ \beta \omega^2 \end{array} \right),
\left( \begin{array}{c} \beta \\ - \alpha \end{array}\right), \left(
\begin{array}{c} \beta \\ - \alpha \omega \end{array}\right), \left(
\begin{array}{c} \beta \\ - \alpha \omega^2 \end{array} \right) \in
{\mathbb C}^2$$ where $\omega:= {\rm exp}(-2\pi i /3)$ is a root of unity, $\alpha = \sqrt{(3+\sqrt{3})/6}$ and $\beta =
\sqrt{(3-\sqrt{3})/6}$. The first three vectors correspond to the upper three vertices 1–3 of the octahedron, the last three vectors to the lower three vertices 4–6. Despite the negative sign of the second component of the last three elements, these vectors are identical with the vectors $$\left(\begin{array}{c} \alpha \\ \beta \end{array} \right),
\left(\begin{array}{c} \alpha \\ \beta \omega \end{array} \right),
\left(\begin{array}{c} \alpha \\ \beta \omega^2 \end{array} \right),
\left( \begin{array}{c} \beta \\ \alpha \end{array}\right), \left(
\begin{array}{c} \beta \\ \alpha \omega \end{array}\right), \left(
\begin{array}{c} \beta \\ \alpha \omega^2 \end{array} \right)
\in {\mathbb C}^2.$$ The latter vectors are obtained by the vector $(\alpha, \beta)^T$ under the action of the dihedral group $D_3$ as discussed in Section \[Section 2.2\]. Similar to the factor $i$ in two vectors of the tetrahedral POVM in Section \[S 4\], the negative sign results from the $\pi$ rotation about the $z$-axis of the lower three vertices 4–6 relative to the upper three vertices 1–3. Since $$\left( \begin{array}{c} \alpha \\ \beta
\end{array} \right) (\alpha, \beta) + \ldots + \left(\begin{array}{c}
\beta \\ -\alpha \omega^2 \end{array} \right) (\beta, -\alpha \omega
)= 3 I_2$$ we rescale $\alpha$ and $\beta$ with the factor $\sqrt{1/3}$ to obtain a POVM. Therefore, we have $$\label{E 3}
M = \left(
\begin{array}{ccc|ccc}\alpha & \alpha & \alpha & \beta & \beta & \beta
\\ \beta & \beta \omega & \beta \omega^2& - \alpha & - \alpha \omega &
- \alpha \omega^2 \end{array}\right) \in {\mathbb C}^{2 \times 6}.$$ As already discussed, we have the negative signs in Equation (\[E 3\]) because the lower face is rotated relatively to the upper face. This is different from the cubic POVM where we have to reorder the operators (compared to the dihedral case) in order to get the negative signs in the second component of the last four vectors. To see this, we write these components as $-\alpha$, $-\alpha \omega$, $- \alpha
\omega^2$ and $-\alpha \omega^3$ with $\omega =i$ as in the dihedral case.
We now consider the extension of $M$ to a unitary matrix ${\tilde M}$. As in Section \[Section 2.2\] we do not embed the system with six dimensions into a qubit register in the first place. The matrix $M$ corresponds to the first two rows of the matrix $${\tilde M} = Q
\left( \begin{array}{ccc|ccc} \alpha & \alpha & \alpha & \beta &
\beta & \beta \\ \alpha & \alpha \omega & \alpha \omega^2 & \beta &
\beta \omega & \beta \omega^2 \\ \alpha & \alpha \omega^2 & \alpha
\omega & \beta & \beta \omega^2 & \beta \omega \\ \hline \beta & \beta
& \beta & -\alpha & - \alpha & - \alpha \\ \beta & \beta \omega &
\beta \omega^2 & -\alpha & - \alpha \omega & - \alpha \omega^2 \\
\beta & \beta \omega^2 & \beta \omega & -\alpha & - \alpha \omega^2 &
- \alpha \omega \end{array} \right)$$ where $Q$ is a permutation matrix that fixes the first row and maps the fifth row to the second. Similar to the previous section, this matrix can be written as $$\label{E 10}
{\tilde M} = Q \left( \sqrt{3}
\left( \begin{array}{cc}
\alpha & \beta \\ \beta & - \alpha \end{array} \right) \otimes
F_3\right) \in {\mathbb C}^{6 \times 6}.$$
We now translate the decomposition of ${\tilde M}$ into a circuit. We have to embed the system with six dimensions into a qubit register with at least three qubits. This can be done by replacing the Fourier matrix $F_3$ in Equation (\[E 10\]) with $F_3 \oplus I_1$ where $I_1 \in {\mathbb C}^{1\times 1}$ denotes the identity matrix of size one. This replacement leads to the matrix $${\tilde M}_8 = Q_8
\left( \begin{array}{cccc|cccc} \alpha & \alpha & \alpha
& 0 & \beta & \beta & \beta & 0 \\ \alpha & \alpha \omega & \alpha
\omega^2 & 0 & \beta & \beta \omega & \beta \omega^2 & 0 \\ \alpha &
\alpha \omega^2 & \alpha \omega & 0 & \beta & \beta \omega^2 & \beta
\omega & 0 \\ 0 & 0 & 0 & \sqrt{3} \alpha & 0 & 0 & 0 & \sqrt{3} \beta
\\ \hline \beta & \beta & \beta
& 0 & -\alpha & -\alpha & -\alpha & 0 \\ \beta & \beta \omega & \beta
\omega^2 & 0 & -\alpha & -\alpha \omega & -\alpha \omega^2 & 0 \\ \beta &
\beta \omega^2 & \beta \omega & 0 & -\alpha & -\alpha \omega^2 & -\alpha
\omega & 0 \\ 0 & 0 & 0 & \sqrt{3} \beta & 0 & 0 & 0 & - \sqrt{3} \alpha
\end{array} \right)$$ where $Q_8 \in {\mathbb C}^{8 \times 8}$ is a permutation matrix that fixes the first row and maps the sixth
row to the second row. In qubit notation, these constraints are given by ${|000\rangle} \mapsto {|000\rangle}$ and ${|101\rangle} \mapsto
{|001\rangle}$. For instance, this transformation can be implemented by an XOR-gate on the first qubit controlled by the last qubit. If we restrict ${\tilde M}_8$ to the first two rows we get the matrix $$\left( \begin{array}{cccc|cccc} \alpha & \alpha & \alpha & 0 &
\beta & \beta & \beta & 0 \\ \beta & \beta \omega & \beta \omega^2 &
0 & -\alpha & -\alpha \omega & - \alpha \omega^2 & 0 \end{array}
\right)$$ corresponding to the desired POVM. The POVM operator corresponding to the fourth and eighth column is $0_2 \in {\mathbb C}^{2 \times 2}$ leading to a zero probability for all states $\rho \in {\mathbb C}^{2 \times 2}$. In summary, we have the equation $${\tilde M}_8^\dagger = \left( \sqrt{3} \left(
\begin{array}{cc} \alpha & \beta \\ \beta & - \alpha \end{array}
\right) \otimes \left( F_3^\dagger \oplus I_1 \right) \right)
Q_8^\dagger$$ for the implementation of the octahedral POVM. This equation corresponds to the circuit shown in Figure \[Figur Kreis Okta\].
Dodecahedron
============
The dodecahedron is the platonic solid with twelve faces and twenty vertices. The symmetry group of the dodecahedron is isomorphic to the alternating group $A_5$. This group contains the sixty permutations of five elements with positive signum. The dodecahedron is shown in Figure \[Figur Dode\]. The upper face with vertices 1–5 and the lower face with vertices 6–10 are perpendicular to the $z$-axis. The point $$\left( \sqrt{\frac{10-2\sqrt{5}}{15}}\,, \;0 \,, \;
\sqrt{\frac{5+2\sqrt{5}}{15}}\right)^T \in {\mathbb R}^3$$ corresponds to vertex $1$. This orientation of the dodecahedron in the Bloch sphere leads to a simple construction of the dodecahedral POVM. With $\omega:={\rm exp}(-2\pi i /5)$, the points on the Bloch sphere correspond to the complex vectors
$$\label{E 11}
\left(\begin{array}{c}\alpha \\ \beta \omega^j \end{array} \right),
\left(\begin{array}{c}\beta \\ -\alpha \omega^j \end{array} \right),
\left(\begin{array}{c}\gamma \\ \delta \omega^j \end{array} \right),
\left(\begin{array}{c}\delta \\ -\gamma \omega^j \end{array} \right)
\in {\mathbb C}^2$$
where $j \in \{0,\ldots,4\}$. The vectors $(\alpha, \beta \omega^j)^T$ correspond to the points 1–5, the vectors $(\beta, -\alpha \omega^j)^T$ to 6–10, the vectors $(\gamma, \delta \omega^j)^T$ to 11–15, and the vectors $(\delta,
\gamma \omega^j)^T$ to the points 16–20. The parameters $\alpha,
\beta, \gamma,$ and $\delta$ are defined as follows: $$\alpha = \sqrt{ \frac{1}{2} +\frac{1}{30} \sqrt{75+30\sqrt{5}}}
\, , \; \beta = \sqrt{\frac{1}{2}-\frac{1}{30}\sqrt{75+30\sqrt{5}}}$$ and $$\gamma = \sqrt{\frac{1}{2} + \frac{1}{30}
\sqrt{75-30\sqrt{5}}} \, , \; \delta =
\sqrt{\frac{1}{2}-\frac{1}{30}\sqrt{75-30\sqrt{5}}} .$$ Due to the equation $$\left(
\begin{array}{c} \alpha \\ \beta \end{array} \right)(\alpha, \beta) +
\ldots + \left(\begin{array}{c}\delta \\ \gamma \omega^4 \end{array}
\right) (\delta, \gamma \omega^{-4} )= 10 I_2$$ we rescale the elements $\alpha, \beta, \gamma, \delta$ with the factor $\sqrt{1/10}$ to obtain a POVM. In contrast to the constructions of Sections \[Section 2.2\]–\[S Okta\] the points on the Bloch sphere decompose into four different orbits under the rotation about the $z$-axis. Note that there are two pairs of orbits. In Line (\[E 11\]) the vectors $(\alpha, \beta \omega^j)^T$ and $(\beta, -\alpha
\omega^j)^T$ are similar to the vectors $(\alpha, \beta \omega^j)^T$ and $(\beta, \alpha \omega^j)^T$. The latter vectors are the orbit of $(\alpha, \beta)^T$ under the dihedral group with $m=5$ as considered in Section \[Section 2.2\]. As in the previous section a $\pi$ rotation of one orbit relative to the other orbit causes the negative sign of the elements $(\beta, -\alpha \omega^j)^T$. Analogously, the second orbit under $D_5$ is defined by the third and fourth type of vectors in Line (\[E 11\]). In summary, the vertices of the dodecahedron correspond to two orbits under the dihedral group $D_5$ with a $\pi$ rotation about the $z$-axis of some points on each orbit. Consequently, we can expect to use a similar construction as in the previous sections.
We now consider the construction of the circuit for implementing the dodecahedral POVM. For convenience, we do not embed the system into a qubit register in the first place. We have the matrix $$M = \left(
\begin{array}{cccc|ccc|ccc|cccc}\alpha & \alpha & \ldots & \alpha &
\beta & \ldots & \beta & \gamma & \ldots & \gamma & \delta & \delta &
\ldots & \delta \\ \beta & \beta \omega & \ldots & \beta \omega^4 &
-\alpha & \ldots & - \alpha \omega^4 & \delta & \ldots & \delta
\omega^4 & - \gamma & - \gamma \omega & \ldots & - \gamma \omega^4
\end{array} \right).$$ This matrix corresponds to the first and second row of the unitary matrix ${\tilde M}$ defined by the equation $$\label{E d}
{\tilde M} = Q \left( A
\otimes F_5 \right) \in {\mathbb C}^{20\times 20},$$ where $Q \in {\mathbb C}^{20 \times 20}$ is a permutation matrix that fixes the first row and maps the seventh row to the second row. The matrix $A$ is defined by $$A = \sqrt{5} \left(
\begin{array}{cccc} \alpha & \beta & \gamma & \delta \\ \beta & -
\alpha & \delta & - \gamma \\ \gamma & - \delta & - \alpha & \beta \\
\delta & \gamma & - \beta & - \alpha \end{array} \right).$$ Now, we want to embed the extended system into a register with five qubits. Similar to the construction for the octahedron in Section \[S Okta\], we can do this by replacing the matrix $F_5$ in Equation (\[E d\]) by the matrix $(F_5 \oplus I_3) \in {\mathbb C}^{8 \times
8}$ where $I_3$ denotes the identity matrix of size three. The matrix $Q$ is replaced by a permutation matrix $Q_{32} \in {\mathbb C}^{32
\times 32}$ that satisfies ${|00000\rangle} \mapsto {|00000\rangle}$ and ${|01001\rangle} \mapsto {|00001\rangle}$ in the qubit notation. This permutation can be implemented by an XOR-operation on the second qubit controlled by the last qubit. In summary, the matrix ${\tilde
M}^\dagger_{32}$ is defined by the equation $${\tilde M}^\dagger_{32}
= \left( A^\dagger
\otimes \left( F_5^\dagger \oplus I_3 \right) \right) Q_{32}^\dagger.$$ The corresponding circuit is shown in Figure \[Figur Kreis Dodeka\].
Note that the matrix $A^\dagger$ can be written as product $A^\dagger = \left(I_2 \oplus (-\sigma_z) \right) \left( I_2 \otimes
B \right) R \left( I_2 \otimes C \right)$ with the matrices $$B= \left( \begin{array}{cc}u_- & -u_+ \\ u_+ &
u_-\end{array}\right),\;\; C= \left( \begin{array}{cc}
v_-&v_+\\ v_+&-v_-\end{array}\right)$$ and constants $$u_\pm = \sqrt{\frac{1}{2} \pm \sqrt{\frac{3+\sqrt{5}}{24}}}\; , \;\;
v_\pm = \mp \sqrt{\frac{1}{2} \pm \sqrt{\frac{\sqrt{5}-1}{8\sqrt{5}}}}.$$ The matrix $R$ is the product $$R= \sqrt{\frac{1}{2}} \left(
\begin{array}{cccc}1&0&-1&0\\ 0&\sqrt{2}&0&0 \\ 1&0&1&0 \\ 0&0&0&\sqrt{2}
\end{array}\right) \sqrt{\frac{1}{2}}
\left( \begin{array}{cccc}\sqrt{2}&0&0&0\\ 0&1&0&1 \\ 0&0&\sqrt{2}&0
\\ 0 & -1 & 0 & 1\end{array} \right).$$ In Figure \[Figur Kreis Dodeka\], the latter two matrices correspond to the two operations on the first qubit which are controlled by the second qubit.
Icosahedron {#S ico}
===========
The icosahedron is the dual polyhedron of the dodecahedron. Consequently, the symmetry groups of both platonic solids are identical. We assume the specific orientation of the icosahedron as shown in Figure \[Figur Ikos\] to obtain a simple construction of the icosahedral POVM. The upper face with vertices 1–3 and the lower face with vertices 4–6 are perpendicular to the $z$-axis. Vertex 1 is given by the vector $$\left( \sqrt{ \frac{10-2\sqrt{5}}{15}}\,,\; 0 \,,\;
\sqrt{\frac{5+2\sqrt{5}}{15}} \right)^T \in {\mathbb R}^3.$$
The vertices of the icosahedron in the Bloch sphere correspond to the complex vectors $$\label{E I}
\left(\begin{array}{c}\alpha \\ \beta \omega^j \end{array}\right),
\left(\begin{array}{c}\beta \\ - \alpha \omega^j \end{array}\right),
\left(\begin{array}{c}\gamma \\ \delta \omega^j \end{array}\right),
\left(\begin{array}{c}\delta \\ - \gamma \omega^j \end{array}\right)
\in {\mathbb C}^2$$ with $j \in \{0,1,2\}$, $\omega:={\rm exp}(- 2 \pi i /3)$ and $$\alpha=\sqrt{\frac{1}{2} + \frac{1}{30}
\sqrt{ 75 + 30 \sqrt{5}} } \; \; , \; \; \beta = \sqrt{ \frac{1}{2} -
\frac{1}{30} \sqrt{75 + 30 \sqrt{5}} }$$ and $$\gamma = \sqrt{ \frac{1}{2} - \frac{1}{30}\sqrt{75-30 \sqrt{5}}} \;
\; , \; \; \delta = \sqrt{ \frac{1}{2} + \frac{1}{30}\sqrt{75-30
\sqrt{5}}}.$$ The vectors in Line (\[E I\]) with $j=0$ correspond to the vertices $1$, $4$, $7$ and $10$ in the given order. As in the case of the dodecahedron we have four orbits under the rotations about the $z$-axis. Therefore, we can expect that a similar construction as in the previous section is possible. Due to the identity $$\left(
\begin{array}{c} \alpha \\ \beta \end{array} \right)(\alpha, \beta) +
\ldots + \left(\begin{array}{c}\delta \\ \gamma \omega^4 \end{array}
\right) (\delta, \gamma \omega^{-4} )= 6 I_2$$ we have the matrix $$M = \left( \begin{array}{ccc|ccc|ccc|ccc}
\alpha & \alpha & \alpha & \beta & \beta & \beta & \gamma & \gamma &
\gamma & \delta & \delta & \delta \\ \beta & \beta \omega & \beta
\omega^2 & -\alpha & - \alpha \omega & - \alpha \omega^2 & \delta &
\delta \omega & \delta \omega^2 & -\gamma & -\gamma \omega & - \gamma
\omega^2 \end{array} \right),$$ where we rescale $\alpha, \beta, \gamma$ and $\delta$ with the factor $\sqrt{1/6}$. The matrix $M$ consists of the first and second row of the matrix $$\label{EE I}
{\tilde M} = Q \left( A
\otimes F_3\right) \in {\mathbb C}^{12 \times 12},$$ where $Q \in {\mathbb C}^{12 \times 12}$ is a permutation matrix that fixes the first row and maps the fifth row to the second. Similar to the previous section, the matrix $A$ is given by $$A=\sqrt{3}\left( \begin{array}{cccc} \alpha & \beta &
\gamma & \delta \\ \beta & - \alpha & \delta & - \gamma \\ \gamma & -
\delta & - \alpha & \beta \\ \delta & \gamma & - \beta & - \alpha
\end{array} \right).$$
The embedding into a register with four qubits works analogously to the previous section. We replace $F_3$ in Equation (\[EE I\]) by $(F_3 \oplus I_1) \in {\mathbb C}^{4 \times 4}$ where $I_1$ denotes the identity matrix of size one. The matrix $Q$ is replaced by the matrix $Q_{16} \in {\mathbb C}^{16 \times 16}$ that can be described as ${|0000\rangle} \mapsto {|0000\rangle}$ and ${|0101\rangle} \mapsto {|0001\rangle}$ in the qubit notation. This permutation can be implemented by an XOR-operation on the second qubit controlled by the last qubit. Therefore, Equation (\[EE I\]) translates into $${\tilde M}_{16}^\dagger = \left( A^\dagger
\otimes \left( F_3^\dagger \oplus I_1
\right) \right) Q_{16}^\dagger.$$ The circuit corresponding to this decomposition of ${\tilde M}^\dagger_{16}$ is given in Figure \[Figur Kreis Ikosa\].
The matrix $A^\dagger$ can be translated into single- and two-qubit gates as shown in the previous section. In this translation we have to replace the constants $u_\pm$ and $v_\pm$ with $$u_\pm = \frac{1}{10} \sqrt{50\pm 5\sqrt{10(5+\sqrt{5})}}
\quad {\rm and} \quad v_\pm = \mp \frac{1}{2}
\sqrt{2 \pm \sqrt{5/3} \mp \sqrt{1/3}}.$$
Conclusions
===========
We have shown that all POVMs given by the vertices of platonic solids can be implemented using a discrete Fourier transform and a few other operations. The algorithms use the symmetry of the POVMs. A common feature of all constructions is the partition of the POVM operators into orbits under the action of a cyclic group. Since the Fourier transform allows to implement POVMs associated with an orbit under a cyclic group it is an essential part of all circuits. For most POVMs corresponding to a platonic solid, a tensor product of a Fourier transform and a specific low-dimensional matrix is a central building block of the circuit. The low-dimensional matrix represents in some sense the relations between the orbits.
The implementation of non-symmetric POVMs seems to be a non-trivial task. It would, for instance, be interesting to know which POVMs can be implemented efficiently, i.e., with a number of elementary gates which grows only polynomially in the number $n$ of POVM-operators. For the symmetric POVMs considered in this paper the question of efficiency makes only sense for the cyclic and dihedral POVMs since the size of the other POVMs is fixed. For $n=2^l$ the complexity of the Fourier transform $F_n$ is only polynomial in $l$. Therefore, the complexity of the circuits for the cyclic and dihedral POVMs grows only polylogarithmically in $n$.
The question of the efficiency of read-out mechanisms for a single bit has no counterpart in classical computer science. Complexity issues in quantum information theory deal not necessarily with the complexity of [*computational*]{} problems. They are also interesting in the context of measurements or state preparation procedures. However, there are some connections between a complexity theory of these non-computational quantum control problems and computational problems [@FQMA; @PSPACE]. Connections between the complexity of POVM measurements and other complexity issues may be subject of further research.
The authors acknowledge helpful discussions with M. Grassl and M. Rötteler. M. Rötteler brought the problem of implementing symmetric POVMs to our attention. This work was supported by grants of the BMBF project 01/BB01B.
E.B. Davies: [*Quantum theory of open systems*]{}, Academic Press, 1976.
M. Sasaki, S.M. Barnett, R. Jozsa, M. Osaki, O. Hirota: [*Accessible information and optimal strategies for real symmetrical quantum sources*]{}, Physical Review A, Vol. 59, No. 5, pp. 3325–3335, May 1999.
E.B. Davies: [*Information and Quantum Measurement*]{}, IEEE Transactions on Information Theory, Vol. IT-24, No. 5, September 1978.
Y.C. Eldar, G.D. Forney: [*On Quantum Detection and the Square-Root Measurement*]{}, IEEE Transactions on Information Theory, Vol. 47, pp. 858-872, Mar. 2001.
C. Fuchs: [*Information Gain vs. State Disturbance in Quantum Theory*]{}, LANL-preprint quant-ph/9611010.
A. Peres: [*Quantum Theory: Concepts and Methods*]{}, Kluwer Academic Publishers, 1993.
M.A. Nielsen, I.L. Chuang: [*Quantum Computation and Quantum Information*]{}, Cambridge University Press, 2000.
S. Sternberg: [*Group theory and physics*]{}, Cambridge University Press, 1994.
D. Janzing, P. Wocjan, Th. Beth: [*Cooling and Low Energy State Preparation for 3-local Hamiltonians are FQMA-complete*]{}, LANL-preprint quant-ph/0305050.
P. Wocjan, D. Janzing, Th. Decker, Th. Beth: [*Measuring 4-local n-qubit observables could probabilistically solve PSPACE*]{}, LANL-preprint quant-ph/0308011.
[^1]: e-mail: [{decker, janzing}@ira.uka.de]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recent results on jets and the hadronic final state from the HERA collaborations H1 and ZEUS are reviewed.'
address: 'Universität Hamburg, IExpPh, Luruper Chaussee 149, 22761 Hamburg, Germany'
author:
- 'T. Schörner-Sadenius on behalf of the H1 and ZEUS collaborations'
title: Jets and the hadronic final state at HERA
---
INTRODUCTION
============
Until the year 2000, the HERA experiments H1 and ZEUS have collected integrated luminosities of about 130 pb$^{-1}$. A large number of interesting results on jets and the hadronic final state from this ‘HERA 1’ data taking period have already been published, and many more analyses are still ongoing.
In 2001/2002 the HERA machine and also the experiments have undergone major changes with the aim of increasing the delivered luminosity by a factor of five. Due to technical problems, efficient data taking could not start before the end of 2003. Therefore, new results from the ‘HERA 2’ data taking period are only now starting to come out.
In the recent years the analysis work on the HERA 1 data has concentrated on making optimal use of the high statistics and either decreasing statistical (and also systematic) errors or performing more and more differential measurements. In addition, increased statistics allows for the discovery or first measurement of new or exotic phenomena.
This contribution will highlight new results concerning jets and the hadronic final state at HERA which were published or made preliminary since the EPS03 conference in Aachen.
![\[eventshapes\] ZEUS event shape variables: Simultaneous fit of $\alpha_S$ and the PC parameter $\overline{\alpha_0}$ to the mean of event shape variables using NLO+PC predictions.](figure1.eps){width="45.00000%"}
\[hadro\]EVENT SHAPE VARIABLES
==============================
Event shape variables like the thrust $T$, broadening $B$ or the $C$ parameter measure aspects of the topology of an event’s hadronic final state. These variables are inclusive and infrared safe so that hadronisation corrections can be estimated using power correction (PC) models.
ZEUS has performed measurements of event shape variables in the Breit frame in the range 80 $< Q^2 <$ 20480 ${\rm GeV^2}$ [@refshape]. The choice of the Breit frame assures that the current region, i.e. the direction of flight of the scattered parton, resembles a single hemisphere in $e^+e^-$ collisions, thus making the HERA results comparable to results from other colliders, especially LEP.
The PC model suggests that non-perturbative corrections to the predictions of observables can be parametrised with a universal parameter $\overline{\alpha_0}$ according to $\langle F \rangle = \langle F \rangle_{NLO} + \langle F \rangle_{pow}(\alpha_S,\overline{\alpha_0})$. So if the next-to-leading (NLO) prediction $\langle F \rangle_{NLO}$ exists, one can fit for $\alpha_S$ and for the universal parameter $\overline{\alpha_0}$.
In addition to the power corrections, next-to-leading log (NLL) resummations can help to improve the description of differential event shape distributions. The use of NLL sums requires a matching with the NLO prediction in order to avoid double-counting of terms.
Fitting the means of the event shape variables with NLO+PC predictions one obtains the results for $\alpha_S$ and $\overline{\alpha_0}$ shown in Figure \[eventshapes\]. The resulting $\alpha_S$ values are consistent with the current world average, $\alpha_S =$ 0.118. The data suggest $\overline{\alpha_0} =$ 0.5 consistently for all event shape variables except for the $C$ parameter, the fit of which is extremely sensitive to the fit range in the ZEUS analysis (in contrast to H1 measurements). This result for $\overline{\alpha_0}$ is roughly compatible with the results from other experiments.
\[prompt\]PROMPT PHOTON PRODUCTION
==================================
The ZEUS collaboration has for the first time measured prompt photon production in DIS in the full HERA 1 data sample of about 120 ${\rm pb^{-1}}$ [@zeusprompt]. Prompt photons are required to be well-isolated and are identified using calorimeter cluster shape information; neutral particle backgrounds are subtracted using Monte Carlo (MC) predictions.
![\[prompt1\]ZEUS prompt photon plus jet sample: photon pseudo-rapidity distribution. The data, which are shown together with the energy scale uncertainty band, are compared to the NLO calculations for which the renormalisation scale uncertainty is also indicated. Also shown is the contribution to the cross-section which is due to prompt photon emission from the quark line.](figure2.eps){width="35.00000%"}
Two analyses were applied to the data: In the first, inclusive prompt photon production is studied, and the data are compared to the leading-order (LO) MC models HERWIG and PYTHIA; none of the models is able to describe all features of the data, and both are far away from the data in terms of normalisation.
In the second analysis, a jet in the central rapidity region was required in addition to the isolated photon. Jets were reconstructed using a cone algorithm in the laboratory frame; their transverse energy had to be in excess of 6 GeV, and their pseudo-rapidity between -1.5 and 1.8. The measurements, which yield a total cross-section of 0.86 pb for prompt photon plus jets, are again compared to HERWIG and PYTHIA, but there exists also an NLO QCD calculation for this signature. PYTHIA and HERWIG manage to describe the shapes of the photon and jet transverse energy distributions, but they fail in describing both pseudo-rapidity distributions. In addition, their normalisation is again of by factors of about 2 (4) for PYTHIA (HERWIG).
The NLO QCD calculation for the prompt photon plus jet analysis predicts a cross-section of about 1.33 pb, which is supposed to be lowered by hadronisation corrections by about 30-40 $\%$. The normalisation of the calculation is therefore compatible with the data. As an example, Figure \[prompt1\] shows the prompt photon pseudo-rapidity distribution. The data are in agreement with the NLO predictions; however, the statistical uncertainty of the measurement is very large. The same is true for the jet transverse energy distribution. However, the calculations predict a too large cross-section at low photon transverse energies and at high (forward) jet pseudo-rapidities.
![\[shape2\]ZEUS $\alpha_S$ measurements based on the mean integrated jet shape $\langle \psi(r=0.5) \rangle$ in DIS events as a function of the jet transverse energy. The data are shown together with the combined $\alpha_S$ value and the world average.](figure3.eps){width="40.00000%"}
\[precise\]NEWS FROM JET PHYSICS
=================================
Determinations of $\alpha_s$ from jet shapes
--------------------------------------------
The ZEUS collaboration has measured the substructure dependence of jet cross-sections in photoproduction and DIS with the aim of extracting the strong coupling parameter $\alpha_S$ [@zeusshapes]. ZEUS used mainly the mean integrated jet shape $\langle \psi(r) \rangle$ which is defined as:
$$\langle \psi(r) \rangle = \frac{1}{N_{jets}}\Sigma_{jets}\frac{E_T(r)}{E_T^{jet}}.$$
Here, $N_{jets}$ is the total number of jets in the sample, $E_T(r)$ is the transverse energy within a cone of radius $r$ around the jet axis, and $E_T^{jet}$ is the transverse energy of the jet. The jet shape was measured, for example, as a function of the jet transverse energy in DIS events. For each data point, an $\alpha_S$ value is determined (Figure \[shape2\]); the resulting values are evolved to the mass of the $Z$, and the different values are combined into one $\alpha_S(M_Z)$ value: $\alpha_S = 0.1176\pm^{0.0091}_{0.0072}$, where only the dominating theoretical uncertainty is given.
Summary on $\alpha_S$
---------------------
Figure \[alpscollection\] gives an overview of various $\alpha_S$ measurements performed by the HERA collaborations in jet events. The measurements are in perfect agreement with the world average; the already very good precision of the measurements, which has been increased over the past few years, will be even more increased once the HERA 2 data are collected and analysed.
![\[alpscollection\]A summary of $\alpha_S$ measurements from H1 and ZEUS jet analyses.](figure4.eps){width="40.00000%"}
Jets in global QCD fits
-----------------------
Jet data from the HERA experiments can be used for more than just determinations of the strong coupling. Recently, the ZEUS collaboration has started to include jet data into their global QCD fits [@zeusglobal], thus supplementing the usual $F_2$ data. The aim of using jet data is to improve the precision of the PDF measurements, especially the gluon, at high values of $x$.
![\[zeusjets2\]ZEUS global QCD fit. The gluon density as a function of $x$ is shown for various $Q^2$ bins. Presented are the results of the ZEUS fit including jet data.](figure5.eps){width="45.00000%"}
The use of jet data in the fits, however, requires a fast evaluation of the NLO predictions for the jet cross-sections. During the minimisation procedure, this prediction has to be performed many thousand times. The usual typical 8 hours for sufficient statistics clearly is not acceptable here. The way around this problem is based on the following method. The jet cross-section can be written as the convolution of the hard partonic cross-section $\hat{\sigma}$ with the PDF $f_a$: $$\sigma = \Sigma_a \Sigma_{n=1}^{\infty}\alpha_S^n \int d\eta
f_a(\eta,\mu_F) \cdot \hat{\sigma}_n(\eta,\mu_F)$$ Here, the first sum runs over all parton flavours. Defining a sufficiently finely binned grid in $(\eta,\mu_F)$ and assuming that the PDF $f_a(\eta,\mu_F)$ is flat in every bin separately, one can write the cross-section in one bin of the $(\eta,\mu_F)$ grid as: $$\label{fl2}
\sigma(\eta,\mu_F) \sim \Sigma_a f_a(\eta,\mu_F) \cdot \Sigma_{n=0}^{\infty}
\alpha_S^n \int d\eta \hat{\sigma}_n(\eta,\mu_F)$$ where the integration is now only over the corresponding bin in $\eta$. The summation of all pieces $\sigma(\eta,\mu_F)$ leads again to the total cross-section. The benefit of this method is that the coefficients (the integrals in Equation \[fl2\]) have to be evaluated only once since they are independent of the PDFs; they can thus be stored in a grid and be convoluted with any PDF provided from the outside, for example from the fitting program. The convolution then takes only parts of a second, which allows for the use of this cross-section evaluation in fitting programs.
The effect of the use of jet data can be viewed in Figure \[zeusjets2\] which shows the results for the gluon density as achieved by the ZEUS fit using inclusive ($F_2$) and jet data. The data are shown as function of $x$ in different bins of $Q^2$. A clear decrease of the size of the uncertainty for the low-$Q^2$ region is found; the effect is less striking for higher $Q^2$ values.
For the future, ZEUS aims at including further jet data from photoproduction, heavy flavour production and also, if possible, from proton-antiproton collisions into the QCD fit.
\[problems\]FROM PHOTOPRODUCTION TO DIS
---------------------------------------
The transition region from photoproduction to DIS has since long been problematic in the theoretical description. The HERA collaborations have published new results on this topic [@transition1; @transition2; @transition3].
![\[resolvedzeus\]ZEUS dijet cross-section as a function of $Q^2$. Separately shown are the direct- and resolved-enriched samples and their description by the NLO QCD calculation. ](figure6.eps){width="40.00000%"}
![\[kamil\]H1 triple-differential dijet cross-section.](figure7.eps){width="45.00000%"}
H1 has analysed dijet cross-sections in the kinematic region 5 $< Q^2 <$ 100 ${\rm GeV^2}$ [@transition1]. The data, taken in the years 1996 and 1997, are compared to direct NLO calculations. Overall good agreement between data and predictions is observed. However, comparing the data to NLO predictions for dijet events with small azimuthal separation between the two hardest jets shows a clear excess of the data over the prediction. This discrepancy can be cured partly using a NLO three-jet calculation; however discrepancies remain at low $Q^2$ and $x$.
![\[h1forward\]H1 forward jet measurement as a function of $x$. The data are compared to (direct) NLO QCD calculations. Also indicated are the energy scale and renormalisation scale uncertainties.](figure8.eps){width="30.00000%"}
Figure \[resolvedzeus\] shows the ZEUS dijet cross-section as a function of $Q^2$ over a wide kinematic range from 0 ${\rm GeV^2}$ to 2000 ${\rm GeV^2}$ [@transition2]. The data are separated into regions of direct- and resolved-enriched events by means of the observable $x_{\gamma}$ which quantifies the fraction of the photon’s energy that entered the hard scattering. Shown is also the NLO QCD prediction. For the DIS regime, the NLO calculation, which does not include resolved contributions, clearly fails to describe the low-$x_{\gamma}$ component (independent of $Q^2$) and also the low-$Q^2$ region. However, the prediction is very sensitive to the renormalisation scale choice, and scale choices can be made such that the data are compatible with the predictions. For the photoproduction region, which also is dominated by low $x_{\gamma}$ values, the (direct+resolved) NLO calculation is able to describe the data.
H1 has measured the triple-differential dijet cross-section $d^3\sigma/d\overline{E}_TdQ^2dx_{\gamma}$ [@transition3] ($\overline{E}_T$ is the mean transverse energy of the two jets). Also here the conclusion is that at low $Q^2$ and also at low values of $\overline{E}_T$ NLO QCD calculations fail to describe the data at low values of $x_{\gamma}$. This is the case even if the calculations include the contributions from the resolved photon (Figure \[kamil\]). In the same analysis, H1 could also demonstrate the importance of the initial and final state parton showers which help considerably in describing the data with LO MC models.
\[problems2\]PARTON DYNAMICS AT LOW $x$
---------------------------------------
Another problematic issue in jet physics is the question of parton dynamics at low $x$. Earlier studies of so-called forward jets [@forwardjetsold] were designed to hunt for the break-down of DGLAP evolution and the onset of BFKL signatures. However, the results were never really conclusive, mainly because no NNLO DGLAP calculation and no BFKL prediction usable for experimentalists were (and are) available. Also the influence of resolved contributions to the predictions was unclear.
![\[forwardpi0\]H1 forward pion measurement as a function of $x$. The data are compared to various QCD calculations and models.](figure9.eps){width="45.00000%"}
There are new results by the H1 collaboration on forward jet and $\pi^0$ production in low $Q^2$ DIS [@h1forwardnew1; @h1forwardnew2]. The forward jet analysis, which was performed in the kinematic region 5 $< Q^2 <$ 85 ${\rm GeV^2}$, studied the triple-differential cross-section $d^3\sigma/d\overline{p}_T^2dQ^2dx$ and compared the data to NLO QCD calculations and various MC models [@h1forwardnew1]. One result is shown in Figure \[h1forward\], where the data are presented as a function of $x$ and are compared to the NLO QCD calculation which clearly fails to describe the data at low values of $x$. MC models implementing resolved photon contributions and models using the the colour-dipole model come close to the data, although a discrepancy between data and predictions remains at lowest values of $x$. A more differential study shows that the discrepancies are largest for low transverse energies and low $Q^2$.
For the case of forward $\pi^0$ production [@h1forwardnew2], most of the available predictions do a sufficiently good job. Figure \[forwardpi0\] shows the forward $\pi^0$ cross-section as a function of $x$ in three bins of $Q^2$; the good description of the data by most models (including the NLO calculation involving a convolution of matrix elements and fragmentation functions) is prominent.
\[pq\]STRANGE PENTAQUARKS
=========================
![\[pqresult\]The ZEUS $pK^0_S$ ($\overline{p}K^0_S$) invariant mass distribution with the peak at about 1522 MeV. Also shown are the results of a fit with a background function and two gaussians, the background function and the simulation of the ARIADNE Monte Carlo.](figure10.eps){width="40.00000%"}
The measurement of narrow baryonic resonances of positive strangeness in the $K^+n$ and $K^0_Sp$ decay channels with a mass of about 1530 MeV and a very narrow width has recently raised a lot of attention. The findings are consistent with five-quark bound states, so-called pentaquarks, $\Theta^+=uudd\overline{s}$.
The ZEUS collaboration started a search for the strange (anti)pentaquark in the $K^0_Sp(\overline{p})$ channel [@pqzeus], relying mainly on the precise tracking information from the central drift chamber. The analysis used all events with $Q^2 >$ 1 ${\rm GeV^2}$ of the full 1996-2000 data statistics of 121 ${\rm pb^{-1}}$, about 1.600.000 events.
The $K^0_Sp(\overline{p})$ invariant mass was obtained by combining the $K^0_S$ and (anti)proton candidates and fixing the kaon mass to the PDG value. Figure \[pqresult\] shows the resulting pentaquark candidate mass distribution together with the fit (a background function and two gaussians), the background function and the prediction of the ARIADNE Monte Carlo. A clear peak at 1522 MeV with a width of about 6 MeV is visible. The number of events ascribed to the signal is 221$\pm$48, of which 96$\pm$34 are in the antiproton channel. Thus, not only could the pentaquark be confirmed, but also the first observation of the antipentaquark with quark content $\overline{u}\overline{u}\overline{d}\overline{d}s$ was made.
The charmed pentaquark state that was observed by the H1 collaboration will be discussed in another proceedings contribution [@pqh1].
\[mix\]OTHER SELECTED TOPICS
============================
QCD instantons
--------------
Instantons are non-perturbative fluctuations of the gauge fields in non-abelian gauge theories. But although instantons are predicted by the standard model, they have not been observed so far. Especially QCD instantons, however, are supposed to have short-distance implications at sufficiently low energies; they are, for example, supposed to induce characteristic topologies in DIS $ep$ scattering events. Results on instanton searches have been reported on by the H1 collaboration before [@h1instanton]. Recently the ZEUS collaboration has published the results of a search for instanton events [@zeusinstanton] performed in about 40 ${\rm pb^{-1}}$ of data collected in the years 1996/97. The data were selected according to $Q^2 >$ 120 ${\rm GeV^2}$. For the search for instanton events, several discriminating techniques were applied which distinguish between standard NC DIS events and instanton events using multi-dimensional cuts. Instanton events were simulated using the QCDINS calculation which predicts an instanton cross-section of 8.9 pb. Under the assumption that all selected data events are instanton events, an upper limit on the instanton production cross-section of $\sigma <$ 26 pb was derived at 95 $\%$ C.L.
![\[bose\]Bose-Einstein correlations: ZEUS measurement of $R(Q_{12})$ for $Q^2 >$ 4 ${\rm GeV^2}$ together with two fits. The data points included in the fit are marked with circles.](figure11.eps){width="35.00000%"}
Bose-Einstein correlations
--------------------------
ZEUS has measured Bose-Einstein (BE) correlations in NC DIS over a wide kinematic range in $Q^2$ from 0.1 ${\rm GeV^2}$ to 8000 ${\rm GeV^2}$ [@zeusbe] in over 120 ${\rm pb^{-1}}$ of data collected from 1996 to 2000. As a measure for the correlation, the analysis used the double ratio $R(Q_{12}) = \zeta^{data}/\zeta^{MC,noBE}$, where $Q_{12}$ is the Lorentz-invariant momentum difference of the particles (which are assumed to be pions), and $\zeta^{data}$ ($\zeta^{MC,noBE}$) are the ratios of inclusive two-particle densities for like-sign and unlike-sign bosons, $\zeta =
\rho(++,--)/\rho(+-)$, for data and Monte Carlo, respectively. The Monte Carlo sample was generated without simulating Bose-Einstein effects, assuming that BE correlations can be factorised from other types of correlations and that non-BE correlations are well described by the MC models. Therefore $R(Q_{12})$ should be sensitive to BC correlations only.
Figure \[bose\] shows the double ratio $R(Q_{12})$ together with two fit models which aim at extracting the size $r$ of the source of final state pions and the coherence parameter $\lambda$. The data points marked with circles were actually included in the fit; as expected the parameters extracted from the gaussian fit show no dependence on $Q^2$. The radius of the production volume is found to be about $r \sim$ 0.65 fm, and the coherence parameter $\lambda \sim$ 0.47. A two-dimensional study of the Bose-Einstein effect shows that the pion-emitting region is elongated. All findings are compatible with measurements from other experiments and colliders.
Anti-Deuteron production
------------------------
The H1 collaboration has measured anti-deuteron production [@h1deuteron] and compared the production rate to results from central proton-proton (Au-Au) collisions measured at the CERN ISR (RHIC).
No significant signal can be found for positively charged tracks, but for negatively charged particles a clear signal of 45 anti-deuterons with an estimated background of about 1 event is observed, leading to a total anti-deuteron cross-section of about 2.7 nb.
The production of nuclei in particle collisions can be described in terms of the coalescence model in which the cross-section $\sigma_A$ for the production of a nucleus with $A$ nucleons can be related to the cross-section for the production of free nucleons in the same reaction, $\sigma_N$, by a so-called coalescence parameter $B_A$.
The (anti-)deuteron coalescence parameter value $B_2 \sim$ 0.01 found by ZEUS is compatible with values deduced at lower center-of-mass energies in $pp$ and $pA$ collisions, but much larger than observed in $Au-Au$ collisions at similar nucleon-nucleon center-of-mass energies. This is understandable if the size of the hard reaction ‘fireball’ is much smaller in $pp$ / $\gamma p$ than in heavy ion collisions.
\[conclusion\]CONCLUSION
========================
Jet and hadronic final state physics at HERA has been reviewed. Clear progress especially in jet physics in terms of statistical precision has been made over the past few years. Jets have even been used in global QCD fits with the aim of improving the knowledge of the gluon density at high values of $x$. In addition, exotic signatures like instantons or pentaquarks have been analysed. Nevertheless, HERA physicists are eagerly waiting for HERA 2 data which promise enough statistics to solve remaining questions and to come to firmer conclusions concerning $\alpha_S$ and the PDFs.
ACKNOWLEDGEMENT {#acknowledgement .unnumbered}
===============
I would like to thank my colleagues from H1 and ZEUS for their support and the conference organisers for a pleasant stay in Montpellier.
[29]{} ZEUS Coll., Abstract 5-0290, ICHEP 2004. ZEUS Coll., DESY-04-016. ZEUS Coll., DESY-04-072. ZEUS Coll., Abstract 5-0263, ICHEP 2004. H1 Coll., DESY-03-160. ZEUS Coll., DESY-04-053. H1 Coll., DESY-03-206. H1 Coll., Nucl. Phys. B538 (1999) 3;\
ZEUS Coll., Phys. Lett. B474 (2000) 1,233;\
ZEUS Coll., Eur. Phys. J. C6 (1999) 239. H1 Coll., H1prelim-04-033. H1 Coll., DESY-04-51. ZEUS Coll., DESY-04-164. K. Lipka, these proceedings. H1 Coll., Eur. Phys. J. C25 (2002) 495. ZEUS Coll., DESY-03-201. ZEUS Coll., DESY-03-176. H1 Coll., DESY-04-032.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Fyodorov, Hiary & Keating established an intriguing connection between the maxima of log-correlated processes and the ones of the Riemann zeta function on a short interval of the critical line. In particular, they suggest that the analogue of the free energy of the Riemann zeta function is identical to the one of the Random Energy Model in spin glasses. In this paper, the connection between spin glasses and the Riemann zeta function is explored further. We study a random model of the Riemann zeta function and show that its two-overlap distribution corresponds to the one of a one-step replica symmetry breaking (1-RSB) spin glass. This provides evidence that the local maxima of the zeta function are strongly clustered.'
address:
- 'Department of Mathematics, Baruch College and Graduate Center, City University of New York, New York, NY 10010.'
- 'Department of Mathematics, Graduate Center, City University of New York, New York, NY 10010.'
author:
- 'Louis-Pierre Arguin'
- Warren Tai
bibliography:
- 'bib\_rzf\_rsb.bib'
date: 'September 13, 2018'
title: |
Is the Riemann zeta function in a short interval\
a 1-RSB spin glass ?
---
Introduction and Main Result
============================
Background
----------
The Riemann zeta function is defined on ${\mathbb C}$ by $$\label{eqn: zeta}
\zeta(s)= \sum_{n\geq 1}\frac{1}{n^s}=\prod_{p\text{ primes}}\left(1-p^{-s}\right)^{-1} \text{ if ${{\rm Re}}\ s>1$,}$$ and can be analytically continued to the whole complex plane by the functional equation $$\zeta(s)=\chi(s) \zeta(1-s)\ , \qquad \chi(s)=2^s \pi^{s-1}\sin \left(\frac{\pi}{2}s\right)\Gamma(1-s)\ .$$ Trivial zeros are located at negative even integers where $\chi(s)=0$. The non-trivial zeros are restricted to the critical strip $0\leq {{\rm Re}}\ s\leq 1$. The Riemann hypothesis states that they all lie on the critical line ${{\rm Re}}\ s=1/2$. A weaker statement, yet with deep implications on the distribution of the primes, is the Lindelöf hypothesis which stipulates that the maximum of $\zeta$ on a large interval $[0,T]$ of the critical line grows slower than any power of $T$, i.e. $\zeta(1/2+{\mathrm{i}}T)$ is $\operatorname{O}(T^{\varepsilon})$ for any ${\varepsilon}>0$, see e.g. [@titchmarsh].
Mathematical physics has provided several important insights in the study of the Riemann zeta function over the years. We refer the reader to [@schumayer-hutchinson] for a broad discussion on this topic. We briefly highlight three contributions from statistical mechanics and probability. First, there are deep connections between the statistics of eigenvalues of random matrices and the zeros of zeta as exemplified by the Montgomery’s pair correlation conjecture, see for example [@bourgade-keating]. Second, the Riemann hypothesis can be recast in the framework of Ising models of statistical mechanics where it bears a resemblance to the Lee-Yang theorem. This perspective was investigated in details by Newman [@newman1; @newman2; @newman3]. It led to an equivalent reformulation of the Riemann hypothesis in terms of the exact value of the de Bruijn–Newman constant [@newman], see [@saouter_etal] for a numerical estimate of the constant and [@tao] for a proof that the constant is non-negative. Third, Fyodorov, Hiary & Keating [@fyodorov-hiary-keating] and Fyodorov & Keating [@fyodorov-keating] recently unraveled a striking connection between the local statistics of the large values of the Riemann zeta function on the critical line and the extremes of a class of disordered systems, the [*log-correlated processes*]{}, that includes among others branching Brownian motion and the two-dimensional Gaussian free field. This connection has also been extended recently to the theory of Gaussian multiplicative chaos by Saksman & Webb [@saksman-webb_1; @saksman-webb_2].
The Fyodorov-Hiary-Keating conjecture is as follows [@fyodorov-hiary-keating; @fyodorov-keating]: if $\tau$ is sampled uniformly on a large interval $[T,2T]$, then the maximum on a short interval, say $[0,1]$, around $\tau$ is $$\label{eqn: FHK}
\max_{h\in[0,1]}\log \left| \zeta(1/2+{\mathrm{i}}(\tau+h)\right|=\log\log T -\frac{3}{4}\log\log\log T+\mathcal M_{T} ,$$ where $(\mathcal M_T)$ is a sequence of random variables converging in distribution. The deterministic order of the maximum corresponds exactly to the one of a log-correlated process, such as a branching random walk and the two-dimensional Gaussian free field, see for example [@kistler; @arguin] for more background on this class of processes. The precise value of the leading order can be predicted heuristically since the process for $\log\zeta$ has effectively $\log T$ distinct values on $[0,1]$ (because there are on average $\log T$ zeros on $[0,1]$, see for example [@titchmarsh]), and the marginal distribution of $\log |\zeta(1/2+i(\tau+h))|$ should be close to Gaussian with variance $\frac{1}{2}\log\log T$ as predicted by Selberg’s Central Limit Theorem [@radziwill-soundararajan]. The log-correlations already appear at the level of the typical values from the multivariate CLT proved in [@bourgade]. The first order of the conjecture was proved recently in parallel: conditionally on the Riemann hypothesis in [@najnudel], and unconditionally in [@ABBRS]. The evidence in favor of the conjecture laid out by Fyodorov & Keating [@fyodorov-keating] suggests that the large values of the Riemann zeta function locally behaves like a disordered system of the spin-glass type characterized by an energy landscape with multiple minima, see Figure \[fig: landscape\].
\[fig: landscape\] ![The value of $-\log|\zeta(1/2+i(T+h))|$ for $T=10000$ and $h\in[0,50]$.](zeta_10000 "fig:"){height="5cm"}
In particular, by considering $-\log |\zeta(1/2+{\mathrm{i}}(\tau+h))|$ as the energy of a disordered system on the state space $[0,1]$, they predict that the analogue of the free energy is in the limit $$\label{eqn: free zeta}
\begin{aligned}
\lim_{T\to\infty} \frac{1}{\log \log T} \log \left(\log T \cdot \int_0^{1} |\zeta(\frac{1}{2}+\rm i(\tau+h))|^{\beta} {\rm d}h\right)=
\begin{cases}
1+\frac{\beta^2}{4} & \ \text{if $\beta<2$,}\\
\beta & \ \text{if $\beta\geq 2$,}
\end{cases}
\end{aligned}$$ similarly to a Random Energy Model (REM) with $\log T$ independent Gaussian variables of variance $\frac{1}{2}\log\log T$.
In this paper, we explore the connection with spin glasses further by providing evidence that $\log|\zeta|$ behaves locally like a spin glass with one-step replica symmetry breaking (1-RSB), cf. Theorem \[thm: main\]. More precisely, we study a simple random model introduced by Harper [@harper] for the large values of $\log |\zeta|$. We show that two points sampled from the Gibbs measure at low temperature have correlation coefficients (or [*overlap*]{}) $0$ or $1$ in the limit, similarly to a 1-RSB spin glass. We expect that part of our approach could be extended to prove a similar result for the Riemann zeta function itself as stated in Conjecture \[conj: zeta\] below.
The model and main result
-------------------------
Let $(U_p, p \text{ primes})$ be IID uniform random variables on the unit circle in ${\mathbb C}$. We write ${\mathbb{E}}$ for the expectation over the $U_p$’s. We study the stochastic process $$\label{eqn: X}
X_h=\sum_{ p\leq T} \frac{{{\rm Re}}(U_p p^{-ih})}{p^{1/2}}, \qquad h\in [0,1]\ .$$ We drop the dependence on $T$ in the notation for simplicity. The process $(X_h, h\in [0,1])$ is a good model for the large values of $\log |\zeta(1/2+{\mathrm{i}}(\tau+h))|$, $h\in [0,1]$, see [@arguin-belius-harper; @harper; @sound] for more details. For example, it is known that the deterministic order of $\max_{h\in[0,1]}X_h$ corresponds to the one in , as proved in [@arguin-belius-harper]. Roughly speaking, the process $X_h$ corresponds to the leading order of the logarithm of the Euler product with the identification $$\left(p^{-{\mathrm{i}}\tau}, p \text{ primes}\right) \longleftrightarrow \left(U_p, p \text{ primes}\right)\ .$$ It is easily checked by computing the joint moments that the above identification is exact as $T\to\infty$ in the sense of finite-dimensional distribution.
The covariance can be calculated using the explicit distribution of the $U_p$’s: $$\label{eqn: covariance1}
\begin{aligned}
{\mathbb{E}}[X_hX_{h'}]
&=\sum_{p\leq T}\int_0^{2\pi} \frac{1}{2}\left(e^{{\mathrm{i}}(\theta-h\log p)}+e^{-{\mathrm{i}}(\theta-h\log p)}\right)\cdot \frac{1}{2}\left(e^{{\mathrm{i}}(\theta-h'\log p)}+e^{-{\mathrm{i}}(\theta-h'\log p)}\right)\ \frac{{{\rm d}}\theta}{2\pi}\\
&= \frac{1}{2}\sum_{p\leq T} \frac{\cos(|h-h'|\log p)}{p}\ .
\end{aligned}$$ We are interested in the correlation coefficient or [*overlap*]{} (in the spin glass terminology): $$\label{eqn: rho}
\rho(h,h')
=\frac{{\mathbb{E}}[X_hX_{h'}]}{\sqrt{{\mathbb{E}}[X_h^2]\ {\mathbb{E}}[X_{h'}^2]}}\ , \qquad \text{ for a given pair $(h,h')$.}$$ Any sum over primes can be estimated using the Prime Number Theorem [@montgomery-vaughan], which gives the density of the primes up to very good errors, $$\label{eqn: PNT}
\#\{p\leq x: p \text{ prime}\}=\int_2^x \frac{1}{\log y} {{\rm d}}y+ \operatorname{O}(xe^{-c\sqrt{\log x}})\ .$$ (The error term, which is already more than sufficient for our purpose, is improved under the Riemann hypothesis.) In particular, this can be used to rewrite the covariances as (see Lemma \[lem: sum primes\] below for details), $$\label{eqn: covariance2}
{\mathbb{E}}[X_h^2]=\frac{1}{2}\sum_{p\leq T}p^{-1}=\frac{1}{2}\log\log T +\operatorname{O}(1) \qquad {\mathbb{E}}[X_hX_{h'}]= \frac{1}{2}\log |h-h'|^{-1}+\operatorname{O}(1)\ .$$ The process $(X_h)$ is said to be [*log-correlated*]{}, since the covariance decays approximately like the logarithm of the distance. The correlation coefficients as a function of the distance become $$\label{eqn: log cor}
\rho(h,h')=\frac{\log |h-h'|^{-1}}{\log\log T} +\operatorname{o}(1)\ , \text{ for $|h-h'|\geq (\log T)^{-1}$.}$$ Throughout the paper, we will use the notation $f(T)=\operatorname{o}(g(T))$ if $f(T)/g(T)\to 0$ and $f(T)=\operatorname{O}(g(T))$ if $f(T)/g(T)$ is bounded. We will sometimes use $f(T)\ll g(T)$ for short if $f(T)=\operatorname{O}(g(T))$ (the Vinogradov notation).
The main result of this paper is the limiting distribution of the correlation coefficient when $h$ and $h'$ are sampled from the Gibbs measure. This is referred to as the [*two-overlap distribution*]{} in the spin-glass terminology. We denote the the Gibbs measure by $$G_{\beta,T}(A)=\int_A \frac{e^{\beta X_h}}{Z_{\beta, T}} \ {{\rm d}}h \qquad Z_{\beta,T}=\int_0^1 e^{\beta X_h}\ {{\rm d}}h\ .$$
\[thm: main\] For every $\beta>2$ and for any interval $I\subseteq [0,1]$, $$\lim_{T\to \infty} {\mathbb{E}}\left[G_{\beta, T}^{\times 2}\{(h,h'): \rho(h,h')\in I\}\right]
= \frac{2}{\beta}{\mathbbm{1}}_I(0) + (1-\frac{2}{\beta}){\mathbbm{1}}_I(1)\ .$$ where ${\mathbbm{1}}_I$ is the indicator function of the set $I$. In other words, when $h,h'$ are sampled independently from the Gibbs measure $G_{\beta,T}$, the random variable $\rho(h,h')$ is Bernoulli-distributed with parameter $2/\beta$ in the limit $T\to\infty$.
The limit is exactly the two-overlap distribution of a 1-RSB spin glass. In view of the relation between the correlation coefficient and the distance $|h-h'|$, the result means that the large values of $X_h$ must lie at a distance $\operatorname{O}(1)$ or $\operatorname{O}((\log T)^{-1})$. The mesoscopic distances $(\log T)^{-\alpha}$, $0<\alpha<1$ are effectively ruled out. Similar results were obtained for the REM model [@derrida], and log-correlated processes [@derrida-spohn; @bovier-kurkova1; @bovier-kurkova2; @arguin-zindy1; @ABK; @arguin-zindy2; @jagannath; @ouimet2].
In the spirit of the Fyodorov-Hiary-Keating conjecture, Theorem \[thm: main\] suggests that $\log |\zeta|$ exhibits 1-RSB for $\beta$ large enough.
\[conj: zeta\] Consider $$\mathcal G_{\beta}(t)=|\zeta(1/2+{\mathrm{i}}t)|^\beta \qquad \mathcal Z_{\beta}(t)=\int_0^1 \mathcal G_{\beta}(t+h){{\rm d}}h\ .$$ For $\beta>2$, and any interval $I\subseteq [0,1]$, if $\tau$ is sampled uniformly on $[T,2T]$: $$\lim_{T\to \infty} {\mathbb{E}}\left[\int_{\{(h,h'):\rho(h,h')\in I\}} \frac{\mathcal G_{\beta}(\tau+h)\cdot \mathcal G_{\beta}(\tau+h')}{ Z_{\beta}(\tau)^2} \ {{\rm d}}h{{\rm d}}h'\right]
= \frac{2}{\beta}{\mathbbm{1}}_I(0) + (1-\frac{2}{\beta}){\mathbbm{1}}_I(1)\ \ .$$ In other words, points $h,h'$ whose $\zeta$-value is of the order of $\log\log T$ are at a distance of $\operatorname{O}(1)$ or $\operatorname{O}((\log T)^{-1})$.
The above conjecture implies a strong clustering of the high values of $\zeta$ at a scale $(\log T)^{-1}$ akin to the one observed in log-correlated process [@abk_genealogy]. In turns, this phenomenon has important consequences for the joint statistics of high values which should be Poissonian at a suitable scale as for log-correlated processes [@abk_poisson; @biskup-louidor]. In particular, it is expected that the statistics of the Gibbs weights is Poisson-Dirichlet [@arguin-zindy1; @arguin-zindy2], and that the Gibbs measure converges to an atomic measure on $[0,1]$, see [@vargas]. This perspective is studied in [@ouimet], and will be discussed further in a forthcoming paper.
[**Acknowledgements**]{}. L.-P. A. is supported by NSF CAREER 1653602, NSF grant DMS-1513441, and a Eugene M. Lang Junior Faculty Research Fellowship. W. T. is partially supported by NSF grant DMS-1513441. Both authors would like to thank Frédéric Ouimet for useful comments on a first version of the paper. L.-P. A. is indebted to Chuck Newman for his constant support and his scientific insights throughout the years.
Main Propositions and Proof of the Theorem \[thm: main\]
--------------------------------------------------------
The proof of Theorem \[thm: main\] is based on a method developed for log-correlated Gaussian processes by Arguin & Zindy [@arguin-zindy1; @arguin-zindy2]. It was adapted from a method of Bovier & Kurkova [@bovier-kurkova1; @bovier-kurkova2] for Generalized Random Energy Models (GREM’s). The main idea is to relate the distribution of the overlaps with the free energy of a perturbed process. In the present case, the process is not Gaussian and the method has to be modified. To this aim, consider the process at [*scale*]{} $\alpha$, for $0<\alpha<1$, where the sum over primes is truncated at $\exp((\log T)^\alpha)$, $$X_h(\alpha)=\sum_{\log p\leq (\log T)^\alpha} \frac{{{\rm Re}}(U_p \ p^{-ih})}{p^{1/2}}, \qquad h\in [0,1]\ .$$ Note that $X_h(1)=X_h$. For a small parameter $|u|<1$, we consider the [*free energy*]{} of the perturbed process $X_h+uX_h(\alpha)$ at scale $\alpha$: $$\label{eqn: f_T}
\begin{aligned}
F_T(\beta;\alpha, u)
&=
{\mathbb{E}}\left[\log \int_{0}^1 \exp\big(\beta (X_h+uX_{h}(\alpha)\big) {{\rm d}}h\right]\ .
\end{aligned}$$ The connection between the free energy and the distribution of the correlation coefficients is through Gaussian integration by parts. Of course, for the process $X_h$, this step is only approximate. It follows closely the work of Carmona & Hu [@carmona-hu] and Auffinger & Chen [@auffinger-chen] on the universality of the free energy and overlap distributions in the Sherrington-Kirkpatrick model.
\[prop: overlap free\] For any $0< \alpha< 1$, $$\left|\int_0^\alpha \ {\mathbb{E}}\left[G_{\beta, T}^{\times 2}\big\{(h,h'): \rho(h,h')\leq y \big\}\right]\ {{\rm d}}y
-\frac{2}{\beta^2\log\log T} \frac{\partial F_T}{\partial u}(\beta;\alpha, 0) \right|=\operatorname{o}(1)\ .$$
The free energy of the perturbed process is calculated using Kistler’s multiscale second moment method [@kistler]. The treatment is similar to the one of Arguin & Ouimet [@arguin-ouimet] for the perturbed Gaussian free field. The same result can be obtained by adapting the method of Bolthausen, Deuschel & Giacomin [@bolthausen-deuschel-giacomin] and Daviaud [@daviaud] to the model as was done in [@arguin-zindy1; @arguin-zindy2]. Kistler’s method is simpler and more flexible. The result is better stated by first defining $$\label{eqn: f}
f(\beta, \sigma^2)=
\begin{cases}
\beta^2\sigma^2/4 \ \ &\text{if $\beta\leq 2/\sigma$,}\\
\beta\sigma-1\ \ &\text{if $\beta\geq 2/\sigma$.}
\end{cases}$$
\[prop: free energy\] For every $\beta>0$ and $|u|<1$, the following limit holds $$\lim_{T\to\infty}\frac{1}{\log\log T} F_T(\beta;\alpha, u)
=
\begin{cases}
f\big(\beta, (1+u)^2\alpha+ (1-\alpha)\big) \ \ &\text{ if $u<0$,}\\
\alpha f(\beta, (1+u)^2) + (1-\alpha) f(\beta,1) \ \ &\text{ if $u\geq 0$.}
\end{cases}$$
The theorem follows from the above two propositions. They are proved in Sections 3 and 4 respectively. Estimates on the model needed for the proofs are given in Section 2.
We need to show that the distribution of $\rho(h,h')$ converges weakly to $\frac{2}{\beta}\delta_0+(1-\frac{2}{\beta})\delta_1$ where $\delta_a$ stands for the Dirac measure at $a$. Write $x_{\beta, T}(s)$ for ${\mathbb{E}}[G_{\beta, T}^{\times 2}\big\{(h,h'): \rho(h,h')\leq s \big\}]$. By compactness of the space of probability measures on $[0,1]$, we can find a subsequence of $(x_{\beta, T})$ that converges weakly to $x_\beta$ as $T\to\infty$. We show that the limit $x_\beta$ is unique and equals $x_\beta(s)=2/\beta$ for $0\leq s <1$, thereby proving the claimed convergence.
By definition of weak convergence, $x_{\beta,T}(s)$ converges to $x_\beta(s)$ at all points of continuity of $s$. Since $x_\beta$ is non-decreasing, this implies convergence almost everywhere. Thus, the dominated convergence theorem implies $$\label{eqn: subsequence}
\lim_{T\to\infty}\int_0^\alpha x_{\beta, T}(s) \ {{\rm d}}s=\int_0^\alpha x_{\beta}(s) \ {{\rm d}}s\ \ \text{ , for $0<\alpha<1$.}$$ The left-hand side can be rewritten using Proposition \[prop: overlap free\] as $$\label{eqn: derivative}
\lim_{T\to\infty}\int_0^\alpha x_{\beta, T}(s) \ {{\rm d}}s
= \lim_{T\to\infty}\frac{2}{\beta^2\log\log T} \frac{\partial F_T}{\partial u}(\beta;\alpha, 0)\ .$$ Since $\big(((\log\log T)^{-1} F_T(\beta;\alpha, u)\big)_T$ is a sequence of convex functions of $u$, the limit of the derivatives is the derivative of the limit at any point of differentiability. Here the limit of the expectation of the free energy is given by Proposition \[prop: free energy\], for $u$ small enough so that $\beta>2/\sigma$ whenever $\beta >2$, $$\label{eqn: derivative free}
\begin{aligned}
\lim_{T\to\infty}\frac{1}{\log\log T} F_T(\beta;\alpha, u)
&=
\begin{cases}
\beta\Big((1+u)^2\alpha+ (1-\alpha)\Big)^{1/2}-1 \ \ &\text{ if $u<0$,}\\
\alpha \beta (1+u) +(1- \alpha) \beta -1 \ \ &\text{ if $u\geq 0$.}
\end{cases}
\end{aligned}$$ In particular, the expected free energy is differentiable at $u=0$. Therefore, equations , and altogether imply $$\int_0^\alpha x_{\beta}(s) \ {{\rm d}}s= \alpha \frac{2}\beta\ \ \text{ , for $0<\alpha<1$.}$$ This means that for any $0<\alpha<\alpha'<1$ we have $$\frac{1}{\alpha'-\alpha}\int_\alpha^{\alpha'} x_{\beta}(s) \ {{\rm d}}s= \frac{2}\beta\ .$$ By taking $\alpha'-\alpha\to 0$, we conclude from the Lebesgue differentiation theorem that $x_{\beta}(s)=2/\beta$ almost everywhere. Since $x_\beta$ is non-decreasing and right-continuous, this implies that $x_\beta(s)=2/\beta$ for every $0\leq s<1$ as claimed.
Estimates on the model of zeta
==============================
In this section, we gather the estimates on the model of zeta needed for the proof of Propositions \[prop: overlap free\] and \[prop: free energy\]. Most of these results are contained in [@arguin-belius-harper]. We include them for completeness since we will need to deal with a perturbed version of the process $(X_h)$. It is is important to point out that most (but not all!) of these estimates can be obtained for zeta itself with some more work, see [@ABBRS].
The essential input from number theory for the model is the Prime Number Theorem . It shows that the density of the primes is approximately $1/\log p$. This implies, for example, that $\sum_{p} p^{-a}<\infty$ for $a>1$. The equation expressing the log-correlations for $h\neq h'$ is straightforward from the following lemma by taking $\Delta=|h-h'|$ and by splitting the sum into the ranges $\log p\leq |h-h'|^{-1}$ and $|h-h'|^{-1}<\log p\leq \log T$.
\[lem: sum primes\] Let $2\leq P<Q<\infty$. Then for $\Delta>0$, we have $$\label{eqn: cos p}
\begin{aligned}
\sum_{P\leq p\text{ primes}\leq Q} \frac{\cos (\Delta\cdot \log p)}{p}&= \int_P^Q \frac{\cos (\Delta\cdot \log v)}{v\log v} {{\rm d}}v+ \operatorname{O}(e^{-c\sqrt{\log P}})\\
&=
\begin{cases}
\log\log Q -\log\log P +\operatorname{O}(1) \ & \text{ for $\Delta\cdot \log Q\leq 1$,}\\
\operatorname{O}(\frac{1}{\Delta\cdot \log P})+ \operatorname{O}(e^{-c\sqrt{\log P}}) & \text{ for $\Delta\cdot\log P\geq 1$.}
\end{cases}
\end{aligned}$$
Denote by ${\rm Li}(x)=\int_2^x \frac{1}{\log y} {{\rm d}}y$ the logarithmic integal. Write $\mathcal E (x)$ for the function of bounded variation $\pi(x)-{\rm Li}(x)$ giving the error, and $f(x)$ for $\frac{\cos(\Delta\cdot \log x)}{x}$. Clearly, we have $$\sum_{P\leq p\leq Q} f(p )=\int_P^Q f(x) \pi(dx)= \int_P^Q \frac{f(x)}{\log x} dx + \int_P^Q f(x) \mathcal E(dx)\ .$$ It remains to estimate the error term. By integration by parts, $$\int_P^Q f(x) \mathcal E(dx)=
\mathcal E(Q ) f(Q )- \mathcal E(P ) f(P ) - \int_P^Q \mathcal E(x) f'(x) dx\ .$$ Note that $f(x)$ is of the order of $1/x$ and $f'(x)$ is of the order of $1/x^2$. Since $\mathcal E(x)=O(x e^{-c\sqrt{\log x}})$, the first claimed equality follows. For the dichotomy in the second equality, in the case $\Delta\cdot\log Q\leq 1$, we expand the cosine to get after the change of variable $y=\log x$ $$\int_{ P}^{Q} \frac{f(x)}{\log x} {{\rm d}}x=\int_{\log P}^{\log Q} \frac{\cos(\Delta\cdot y)}{y} {{\rm d}}y=\int_{\log P}^{\log Q} \left(\frac{1}{y}+\operatorname{O}(\Delta^2\cdot y)\right)\ {{\rm d}}y\ .$$ The result follows by integration. In the case $\Delta\cdot \log P \geq 1$, we integrate by parts to get $$\int_P^Q \frac{f(x)}{\log x} dx=\frac{\sin(\Delta\cdot y)}{\Delta\cdot y}\Big|_{\log P}^{\log Q}+\int_{\log P}^{\log Q} \frac{\sin(\Delta\cdot y)}{\Delta\cdot y^2} {{\rm d}}y\ .$$ Both terms are $\operatorname{O}(\frac{1}{\Delta\cdot \log P})$ as claimed.
Proposition \[prop: free energy\] gives an expression for the free energy of the perturbed process at scale $\alpha$. For simplicity, we denote this process by $$\label{eqn: X tilde}
\widetilde X_h= (1+u)X_h(\alpha) + X_h(\alpha,1) \ \ \text{for $X_h(\alpha,1)=X_h-X_h(\alpha)$, $h\in [0,1]$.}$$ Note that we recover $X_h$ at $u=0$. The finite-dimensional distributions of $(\widetilde X_h)$ can be explicitly computed. In fact, it is not hard to compute explicitly the moment generating function for any increment of $(X_h)$. We will only need the two-dimensional case.
\[prop: MGF\] Let $0\leq \alpha_1<\alpha_2\leq 1$. Consider $X_h(\alpha_1,\alpha_2)=X_h(\alpha_2)-X_h(\alpha_1)$. We have for $\lambda, \lambda'\in {\mathbb R}$ and $h,h'\in [0,1]$, $$\begin{aligned}
&{\mathbb{E}}\left[\exp\left(\lambda X_h(\alpha_1,\alpha_2) + \lambda'X_{h'}(\alpha_1,\alpha_2\right)\right]\\
&\hspace{2cm}=C(\lambda,\lambda')\cdot \exp\left(\frac{1}{2}\sum_{\substack{\log p\leq (\log T)^{\alpha_2}\\ \log p>(\log T)^{\alpha_1} }}
\frac{1}{2p}\Big(\lambda^2+{\lambda'}^2+2\lambda\lambda'\cos(|h-h'|\log p)\Big)\right)\ ,
\end{aligned}$$ where $C=C(\lambda, \lambda')$ is bounded if $\lambda$ and $\lambda'$ are bounded uniformly in $T$.
The expression can be evaluated explicitly as follows. Since the $U_p$’s are independent, we can first restrict the computation to a single $p$. Straightforward manipulations yield $${\mathbb{E}}\left[\exp\left(p^{-1/2}\lambda \cdot {{\rm Re}}(U_pp^{-{\mathrm{i}}h})+ p^{-1/2}\lambda'\cdot{{\rm Re}}(U_pp^{-{\mathrm{i}}h'})\right)\right]
={\mathbb{E}}\left[\exp(aU_p+\bar{a}\overline U_p)\right]$$ for $a=(2p^{1/2})^{-1}(\lambda p^{-{\mathrm{i}}h}+\lambda' p^{-{\mathrm{i}}h'})$. By expanding the exponentials and using the fact that $U_p$ is uniform on the unit circle, we get $$\label{eqn: mgf p}
\begin{aligned}
{\mathbb{E}}\left[\exp(aU_p+\bar{a}\overline U_p)\right]
&= \sum_{n=0}^\infty\sum_{k=0}^n \frac{a^k\bar{a}^{n-k}}{n!}{n\choose k}{\mathbb{E}}[U_p^k\overline U_p^{n-k}]\\
&=\sum_{m=0}^\infty \frac{1}{(m!)^2}\left(\frac{\lambda^2+{\lambda'}^2+2\lambda\lambda'\cos(|h-h'|\log p)}{4p}\right)^m\\
&= 1+ \left(\frac{\lambda^2+{\lambda'}^2+2\lambda\lambda'\cos(|h-h'|\log p)}{4p}\right)+\operatorname{O}(p^{-2})\ ,
\end{aligned}$$ where the $\operatorname{O}$-term depends on $\lambda,\lambda'$. The second equality follows from the fact that the expectation is non-zero only if $k=n/2$. It remains to take the product over the range of $p$. The claim then follows from the fact that the sum of $p^{-2}$ is finite by .
Proposition \[prop: MGF\] yields Gaussian bounds in the large deviation regime we are interested in. Indeed, by Chernoff’s bound (optimizing over $\lambda$), it implies that, for $\gamma>0$, $$\label{eqn: increment one point}
{\mathbb{P}}\left(X_h(\alpha_1,\alpha_2)>\gamma\log\log T\right)
\ll \exp\left(-\frac{\gamma^2 \log\log T}{(\alpha_2-\alpha_1)}\right)=(\log T)^{\frac{-\gamma^2}{\alpha_2-\alpha_1}}\ ,$$ where we used Lemma \[lem: sum primes\] to estimate the sum over primes. This supports the heuristic that $X_h(\alpha_1,\alpha_2)$ is approximately Gaussian of variance $\frac{\alpha_2-\alpha_1}{2}\log\log T$. This implies for $\widetilde X_h$ in $$\label{eqn: X tilde one-point}
{\mathbb{P}}\left(\widetilde X_h>\gamma\log\log T\right)
\ll \exp\left(-\frac{\gamma^2 \log\log T}{(1+u)^2\alpha+(1-\alpha)}\right)=(\log T)^{\frac{-\gamma^2}{(1+u)^2\alpha+(1-\alpha)}}\ .$$ The same can be done for two points $h,h'$. Using Lemma \[lem: sum primes\] again, we get $$\label{eqn: increment two points}
\begin{aligned}
&{\mathbb{P}}\left(X_h(\alpha_1,\alpha_2)>\gamma\log\log T,X_{h'}(\alpha_1,\alpha_2)>\gamma \log\log T \right)\\
&\hspace{3cm}\ll
\begin{cases}
\exp\left(-\frac{\gamma^2 \log\log T}{(\alpha_2-\alpha_1)}\right) \ &\text{ if $|h-h'|\leq (\log T)^{-\alpha_2}$,}\\
\exp\left(-2\frac{\gamma^2 \log\log T}{(\alpha_2-\alpha_1)}\right)\ &\text{ if $|h-h'|\geq (\log T)^{-\alpha_1}$.}\\
\end{cases}
\end{aligned}$$ This can be interpreted as follows. The increments are (almost) independent if the distance between the points is larger than the relevant scales of the increments, and are (almost) perfectly correlated if the distance is smaller than the scales.
It is important to note that if $\alpha_1>0$, then a stronger estimate than the one of Proposition \[prop: MGF\] holds. This is because the sum over primes in is then negligible since it is the tail of a summable series. This means that the constant $C(\lambda,\lambda')$ is then $1+\operatorname{O}(1)$. This gives a precise Gaussian estimate by inverting the moment generating function (or the Fourier transform if we pick $\lambda,\lambda'\in \mathbb C$). We omit the proof for conciseness and we refer to [@arguin-belius-harper] where this is done using a general version of the Berry-Esseen theorem.
\[prop: berry-esseen\] For $0<\alpha_1<\alpha_2\leq 1$ and $0<\gamma <1$, we have for $h\in [0,1]$, $${\mathbb{P}}(X_{h}(\alpha_1,\alpha_2)>\gamma \log\log T)
\gg\frac{1}{\sqrt{\log\log T}}\exp\left(-\frac{\gamma^2 \log\log T}{(\alpha_2-\alpha_1)}\right)=(\log T)^{\frac{-\gamma^2}{\alpha_2-\alpha_1}+\operatorname{o}(1)}\ .$$ Moreover, if $|h-h'|>(\log T)^{-\alpha_1}$, then $${\mathbb{P}}(X_{h}(\alpha_1,\alpha_2)>\gamma \log\log T, X_{h'}(\alpha_1,\alpha_2)>\gamma \log\log T)
=(1+\operatorname{o}(1)) {{\mathbb{P}}(X_{h}(\alpha_1,\alpha_2)>\gamma \log\log T)}^2\ .$$
Since the process $(X_h,h\in[0,1])$ is continuous and not discrete, we need a last estimate to control all values in an interval of length corresponding to the relevant scale. This is needed when proving rough bound on the maximum in Lemma \[lem: max\]. Heuristically, it says that the maximum of $X_h(\alpha_1,\alpha_2)$ over an interval of width smaller than $(\log T)^{-\alpha_2}$ behaves like a single value $X_h(\alpha_1,\alpha_2)$. This is done in [@arguin-belius-harper] by a chaining argument and we omit the proof for conciseness.
\[lem: smoothing\] Let $0\leq\alpha_1<\alpha_2\leq 1$. For every $h\in[0,1]$ and $\gamma>0$, we have $${\mathbb{P}}\left(\max_{|h-h'|\leq (\log T)^{-\alpha_2}}X_{h'}(\alpha_1,\alpha_2)> \gamma\log\log T\right) \ll (\log T)^{-\frac{\gamma^2}{\alpha_2-\alpha_1}}\ .$$ In particular, we have $${\mathbb{P}}\left(\max_{|h-h'|\leq (\log T)^{-1}}\widetilde X_{h'}> \gamma\log\log T\right) \ll (\log T)^{-\frac{\gamma^2}{\alpha(1+u)+(1-\alpha)}}\ .$$
Proof of Proposition \[prop: overlap free\]
===========================================
As mentioned in Section 1.3, the proof of Proposition \[prop: overlap free\] is based on an approximate Gaussian integration by parts as in [@carmona-hu] and [@auffinger-chen]. The following lemma is an adaptation for complex random variables of Lemma 4 in [@carmona-hu] .
\[lem: by parts\] Let $\xi$ be a complex random variable such that ${\mathbb{E}}[|\xi|^3]<\infty$, and ${\mathbb{E}}[\xi^2]={\mathbb{E}}[\xi]=0$. Let $F:\mathbb{C}\to\mathbb C$ be a twice continuously differentiable function such that for some $M>0$, $$\left\Vert\partial_z^2F\right\Vert_{\infty},
\left\Vert \partial^2_{\overline z}F\right\Vert_\infty <M\ ,$$ where $\left\Vert f\right\Vert_\infty=\sup _{z\in \mathbb{C}}\left\vert
f(z,\overline{z})\right\vert$. Then $$\left|{\mathbb{E}}\left[\xi F(\xi,\overline{\xi})\right] - {\mathbb{E}}[\left\vert\xi\right\vert^2]\
{\mathbb{E}}\left[ \partial _{\overline{z}}F(\xi,\overline{\xi})\right] \right|
\ll M\ {\mathbb{E}}[|\xi|^3].$$
Since ${\mathbb{E}}[\xi^2]={\mathbb{E}}[\xi]=0$, the left-hand side can be written as $$\label{eqn: to prove by parts}
{\mathbb{E}}\left[\xi \big(F(\xi,\overline{\xi}) -F(0,0)- \xi \partial_z F(0,0) - \overline \xi \partial_{\overline z} F(0,0)\big)\right]
- {\mathbb{E}}[\left\vert\xi\right\vert^2]\
{\mathbb{E}}\left[\Big(\partial _{\overline{z}}F(\xi,\overline{\xi})- \partial _{\overline{z}}F(0,0)\Big)\right]\ .$$ By Taylor’s theorem in several variables and the assumptions, the following estimates hold $$\begin{aligned}
&\left| F(\xi,\overline{\xi}) -F(0,0)- \xi \partial_z F(0,0) - \overline \xi \partial_{\overline z} F(0,0) \right|\ll M|\xi|^2\\
&\left| \partial _{\overline{z}}F(\xi,\overline{\xi})- \partial _{\overline{z}}F(0,0) \right|\ll M|\xi|\ .
\end{aligned}$$ Therefore the norm of gives $$\left|{\mathbb{E}}\left[\xi F(\xi,\overline{\xi})\right] - {\mathbb{E}}[\left\vert\xi\right\vert^2]\ {\mathbb{E}}\left[ \partial _{\overline{z}}F(\xi,\overline{\xi})\right] \right|
\ll
M ({\mathbb{E}}[|\xi|^3]+ {\mathbb{E}}[|\xi|^2]{\mathbb{E}}[|\xi|])\ .$$ The claim then follows by Hölder’s inequality.
As in [@carmona-hu], the lemma can be applied to relate the derivative of the free energy to the two-point correlations of the process.
\[prop: carmona-hu\] For any $p\leq T$, we have $$\begin{aligned}
&\frac{\partial}{\partial u} {\mathbb{E}}\left[ \log \int_0^1 \exp\big(\beta (X_h(T)+u \ {{\rm Re}}(U_pp^{-ih-1/2})\big) {{\rm d}}h\right]\Big|_{u=0}\\
&\hspace{2cm}= \frac{\beta}{2}{\mathbb{E}}\left[ \int_{[0,1]^2} \frac{1- \cos(|h-h'|\log p)}{p}\ {{\rm d}}G^{\times 2}_{\beta, T}(h,h')\right]
+\operatorname{O}(p^{-3/2})\ .
\end{aligned}$$
Write for short $\omega_p(h)=(2p^{1/2})^{-1}p^{-{\mathrm{i}}h}$. Direct differentiation yields at $u=0$ $$\label{eqn: U_p integral}
\int_0^1 U_p \omega_p(h)\ {{\rm d}}G_{\beta, T}(h) + \int_0^1 \overline{U_p} \overline{\omega_p}(h)\ {{\rm d}}G_{\beta, T}(h)\ .$$ We make the dependence on $U_p$ in the measure $G_{\beta, T}$ explicit. For this, define $$Y_p(h)=\beta \sum _{\substack{q \le T\\p\neq q}}{{\rm Re}}\left(\frac{U_q q^{-ih}}{q^{1/2}} \right)\ .$$ Clearly, $Y_p(h)$ is independent of $U_p$ by definition. Consider $$F_p(z,\overline{z})=\frac{\int_0^1 \omega_p(h)\ \exp(\beta \omega_p(h) z +\beta \overline{\omega_p}(h)
\overline{z}+ Y_p(h))\ {{\rm d}}h}
{\int_0^1\exp(\beta \omega_p(h') z +\beta \overline{\omega_p}(h')\overline z+ Y_p(h'))\ {{\rm d}}h'}\ .$$ Note that with this definition, the first integral in is $U_pF_p(U_p, \overline U_p)$ and the second is its complex conjugate. This shows that the derivative of the expectation at $u=0$ is $${\mathbb{E}}\left[U_p\cdot F_p(U_p, \overline U_p)+\overline{U_p}\cdot \overline{F_p}(U_p, \overline U_p)\right]\ .$$ It remains to apply Lemma \[lem: by parts\] with the function $F_p(z,\overline z)$ and $\xi=U_p$. Write for short for a function $H$ on $[0,1]$ $$\left\langle H\right\rangle^{(z,\overline{z})}=
\frac{\int_0^1 H(h)\ e^{\beta\left( z\ \omega_p(h) +\overline{z}\ \overline{\omega_p}(h)\right)+ Y_p(h)} {{\rm d}}h}
{\int_0^1 e^{\beta\left( z\ \omega_p(h) +\overline{z}\ \overline{\omega_p}(h)\right)+ Y_p(h)} {{\rm d}}h}\ .$$ Direct differentiation of the above yields $$\label{eqn: H deriv}
\begin{aligned}
\partial_{\overline z} \left\langle H\right\rangle^{(z,\overline{z})}=
\beta\left(\left\langle H \overline{\omega_p}\right\rangle^{(z,\overline{z})}-\left\langle
H\right\rangle^{(z,\overline{z})}\left\langle \overline{\omega_p}\right\rangle^{(z,\overline{z})}\right)\ .
\end{aligned}$$ In particular, for $H=\omega_p$, we get $$\label{eqn: F deriv}
\partial_{\overline z}F_p(z,\overline z)=\beta\left(\left\langle |\omega_p|^2\right\rangle^{(z,\overline{z})}-
|\left\langle \omega_p\right\rangle^{(z,\overline{z})}|^2\right)\ .$$ When evaluated at $z=U_p$, this is by definition of $\omega_p$ $$\label{eqn: F deriv 2}
\partial_{\overline z}F_p(U_p,\overline U_p)=\frac{\beta}{4} \int_{[0,1]^2}(p^{-1}-p^{-1}\cos(|h-h'|\log p)) \ {{\rm d}}G^{\times 2}_{\beta, T}(h,h')\ .$$ Clearly, $|\omega_p|\leq p^{-1/2}$. Therefore the second derivatives are easily checked to be bounded by $\operatorname{O}(p^{-3/2})$ by applying the formula to each term of . The statement of the lemma then follows from Lemma \[lem: by parts\] and , after noticing that the second term of is the conjugate of the first.
The proof of Proposition \[prop: overlap free\] is an application of Proposition \[prop: carmona-hu\] to a range of primes.
Recall the definition of $\rho(h,h')$ in equations and . On one hand, Fubini’s theorem directly implies that $$\label{eqn: fubini}
\begin{aligned}
\int_0^\alpha \ G_{\beta, T}^{\times 2}\big\{(h,h'): \rho(h,h')\leq y \big\}\ {{\rm d}}y
&=\int_{[0,1]^2} \left(\int_0^\alpha {{\mathbbm{1}}}_{\{\rho(h,h')\leq r\}} \ {{\rm d}}r\right) {{\rm d}}G_{\beta, T}(h,h')\\
&=\int_{[0,1]^2} \left(\alpha-\rho(h,h')\right){\mathbbm{1}}_{\{\rho(h,h')\leq \alpha\}} {{\rm d}}G_{\beta, T}(h,h')\ .
\end{aligned}$$ It remains to check on the other hand that the derivative in the proposition is close to the expectation of the above. Direct differentiation of at $u=0$ yields by Proposition \[prop: carmona-hu\] $$\label{eqn: deriv F}
\frac{\partial F_T}{\partial u}(\beta;\alpha, 0)
= \frac{\beta^2}{2} \int_{[0,1]^2} \sum_{\log p\leq (\log T)^{\alpha}} {\mathbb{E}}\big[p^{-1}(1- \cos(|h-h'|\log p))\ {{\rm d}}G^{\times 2}_{\beta, T}(h,h')\big]
+\operatorname{O}(\sum_{ p\leq e^{(\log T)^{\alpha}}}p^{-3/2})\ .$$ The error term is of order one by . Similarly, if $|h-h'|\leq (\log T)^{-\alpha}$, the sum in the integral is by $$\sum_{\log p\leq (\log T)^{\alpha}}p^{-1}(1- \cos(|h-h'|\log p))=
\alpha \log \log T - \alpha\log\log T +\operatorname{O}(1)=\operatorname{O}(1)\ .$$ On the other hand, if $|h-h'|>(\log T)^{-\alpha}$, the sum can be divided into three parts $$\sum_{\log p\leq (\log T)^{\alpha}}p^{-1}
-\sum_{\log p\leq |h-h'|^{-1}}p^{-1} \cos(|h-h'|\log p)
- \sum_{|h-h'|^{-1}<\log p\leq (\log T)^{\alpha}}p^{-1}\cos(|h-h'|\log p)\ .$$ When equation is applied to each of the parts, this equals $$\alpha\log \log T-\log |h-h'|^{-1} +\operatorname{O}(1)\ .$$ Furthermore, recall from that $\rho(h,h')\log\log T$ differs from $\log |h-h'|^{-1}$ by $\operatorname{o}(\log\log T)$. This implies that the conditions on $\log |h-h'|^{-1}$ can be replaced by $\rho(h,h')\log\log T$ at a cost of a term $\operatorname{o}(\log\log T)$ (since the sum would differ by a range of $\log p$ of at most $\operatorname{o}(\log T)$ primes). All these observations together imply $$\frac{1}{\log\log T}\sum_{\log p\leq (\log T)^{\alpha}}p^{-1}(1- \cos(|h-h'|\log p))=
\left(\alpha-\frac{\log |h-h'|^{-1}}{\log\log T}\right){\mathbbm{1}}_{\{\rho(h,h')\leq \alpha\}} +\operatorname{o}(1)\ .$$ We finally conclude by putting the right side back in the integral of and by using that $$\frac{2}{\beta^2\log\log T}\frac{\partial F_T}{\partial u}(\beta;\alpha, 0)
= \int_{[0,1]^2}\left(\alpha-\rho(h,h')\right){\mathbb{E}}\big[{\mathbbm{1}}_{\{\rho(h,h')\leq \alpha\}} {{\rm d}}G^{\times 2}_{\beta, T}(h,h')\big] +\operatorname{o}(1)\ .$$ This matches the first claim by an error $\operatorname{o}(1)$ thereby proving the proposition.
Proof of Proposition \[prop: free energy\]
==========================================
We write $\widetilde X_h= (1+u)X_h(\alpha) + X_h(\alpha,1)$ as in equation . The limit of the free energy of this process is obtained by Laplace’s method once the measure of high points is known, cf. Lemma \[lem: high points\]. The proof of Lemma \[lem: high points\] is based on a similar computation of [@arguin-ouimet] for the two-dimensional Gaussian free field based on Kistler’s multiscale second moment method [@kistler]. But first, we need an [*a priori*]{} restriction on the maximum of the process $(\widetilde X_h)$. The maximum depends on the value of the parameter $u$ as expected from GREM models. With this in mind, we define $$\label{eqn: gamma star}
\gamma^\star=
\begin{cases}
\Big((1+u)^2 \alpha + (1-\alpha)\Big)^{1/2}\ \ &\text{if $u\leq0$,}\\
(1+u)\alpha + (1-\alpha)\ \ &\text{if $u>0$.}
\end{cases}$$ Note that the two expressions are equal to $1$ at $u=0$ and that $\gamma^\star>1$ if $u> 0$, and $\gamma^\star<1$ if $u<0$. The next lemma bounds the maximum of $\widetilde X_h$.
\[lem: max\] For any ${\varepsilon}>0$, $$\lim_{T\to\infty}{\mathbb{P}}\left(\max_{h\in[0,1]} \widetilde X_h>(1+{\varepsilon})\gamma^\star \log\log T \right)=0\ .$$
This is a consequence of Lemma \[lem: smoothing\] which shows that the large values of $X_h(\alpha)$ are well approximated by points at a distance $(\log T)^{-\alpha}$. In the case $u\leq 0$, we use the lemma with $\alpha=1$. Without loss of generality, suppose that $\log T$ is an integer and consider $I_k$, $k\leq \log T$, a collection of intervals of length $(\log T)^{-1}$ that partitions $[0,1]$. Then a simple union bound yields $$\begin{aligned}
{\mathbb{P}}\left(\max_{h\in[0,1]} \widetilde X_h>(1+{\varepsilon})\gamma^\star \log\log T \right)
&\leq \sum_{k=1}^{\log T} {\mathbb{P}}\left(\max_{h\in I_k} \widetilde X_h >(1+{\varepsilon})\gamma^\star \log\log T \right)\ .
\end{aligned}$$ Lemma \[lem: smoothing\] applied to $\widetilde X_h$ then implies $$\begin{aligned}
{\mathbb{P}}\left(\max_{h\in[0,1]} \widetilde X_h>(1+{\varepsilon})\gamma^\star \log\log T \right)
&\ll (\log T) \exp\left(-\frac{(1+{\varepsilon})^2 \Big((1+u)^2\alpha + (1-\alpha)\Big)\log\log T)}{(1+u)^2\alpha + (1-\alpha)}\right)\\
&\leq (\log T)^{1-(1+{\varepsilon})^2}\ ,
\end{aligned}$$ which goes to $0$ as claimed.
In the case $u>0$, an extra restriction is needed since the large values of $X_h(\alpha)$ are themselves limited. Proceeding as above, without loss of generality, assume that $(\log T)^\alpha$, $(\log T)^{1-\alpha}$ and $\log T$ are integers. Consider the collection of intervals $J_j$, $j\leq (\log T)^{\alpha}$, that partitions $[0,1]$ into intervals of length $(\log T)^{-\alpha}$. Each $J_j$ is again partitioned into intervals $I_{jk}$, $k\leq (\log T)^{1-\alpha}$, of length $(\log T)^{-(1-\alpha)}$. Then Lemma \[lem: smoothing\] implies $$\label{eqn: max interval}
{\mathbb{P}}\left(\exists j: \max_{h\in J_{j}} X_h(\alpha)>(1+{\varepsilon})\cdot\alpha \log\log T \right)\to 0\ .$$ Therefore, the probability of the maximum of $\widetilde X_h$ can be restricted as follows: $$\begin{aligned}
&{\mathbb{P}}\left(\max_{h\in[0,1]} \widetilde X_h>(1+{\varepsilon})\gamma^\star \log\log T \right)\\
&= {\mathbb{P}}\left(\exists h\in [0,1]: \widetilde X_h>(1+{\varepsilon})\gamma^\star \log\log T,X_h(\alpha)\leq (1+{\varepsilon})\cdot \alpha \log \log T\right)+\operatorname{o}(1)\\
&\ll\sum_{j,k}
\sum_{q=0}^{(1+{\varepsilon})\cdot \alpha\log\log T}
{\mathbb{P}}\left(\max_{h\in J_j} X_h(\alpha)\in[q,q+1], \max_{h\in I_{jk}} X_h(\alpha,1)>(1+{\varepsilon})\gamma^\star \log\log T-(1+u)(q+1)\right)\ .
\end{aligned}$$ The last inequality is obtained by a union bound on the partition $(I_{jk})$ and by splitting the values of the maximum of $X_h(\alpha)$ on the range $[0,(1+{\varepsilon})\alpha \log\log T]$. (Note that $X_h(\alpha)$ is symmetric thus the maximum is greater than $0$ with large probability.) By independence between $(X_h(\alpha),h\in[0,1])$ and $(X_h(\alpha,1),h\in[0,1])$, Lemma \[lem: smoothing\] can be applied twice to get the following bound on the summand: $$\label{eqn: prob u<=0}
\ll \exp\left(\frac{-q^2}{\alpha \log\log T}\right)\cdot \exp\left(\frac{-\Big((1+{\varepsilon})\gamma^\star\log\log T-(1+u)(q+1)\Big)^2}{(1-\alpha) \log\log T}\right)\ .$$ On the interval $[0,(1+{\varepsilon})\alpha\log\log T]$, this is maximized at the endpoint $q=(1+{\varepsilon}) \alpha\log\log T$. (This is where the case $u<0$ differs, as the optimal $q$ there is within the interval. See Remark \[rem: dicho\] for more on this.) Putting this back in and summing over $j,k$, and $q$ finally give the estimate: $$\begin{aligned}
&{\mathbb{P}}\left(\max_{h\in[0,1]} \widetilde X_h>(1+{\varepsilon})\gamma^\star \log\log T \right)\\
&\ll (\log\log T)\cdot (\log T)^{\alpha} \cdot e^{\frac{-\Big((1+{\varepsilon})\alpha \log\log T\Big)^2}{\alpha \log\log T}}\cdot (\log T)^{1-\alpha} \cdot e^{\frac{-\Big((1+{\varepsilon})(1-\alpha) \log\log T\Big)^2}{(1-\alpha) \log\log T}}\\
&\ll (\log\log T)\cdot \log T^{1-(1+{\varepsilon})^2}=\operatorname{o}(1)\ .
\end{aligned}$$ This concludes the proof of the lemma.
Consider for $0<\alpha\leq 1$ and $|u|<1$ the (normalized) log-measure of $\gamma$-high points $$\label{eqn: log-measure}
\mathcal E_{\alpha,u}(\gamma;T)=
\frac{1}{\log\log T}\log {\rm Leb}\{h\in [0,1]: \widetilde X_h> \gamma \log \log T\}\ , 0< \gamma <\gamma^\star\ .$$ The limit of these quantities in probability can be computed following [@arguin-ouimet].
\[lem: high points\] The limit $\mathcal E_{\alpha,u}(\gamma)=\lim_{T\to \infty}\mathcal E_{\alpha,u}(\gamma;T)$ exists in probability. We have for $u<0$, $$\mathcal E_{\alpha,u}(\gamma)
=-\frac{\gamma^2}{(1+u)^2 \alpha+(1-\alpha)}\ ,$$ and for $u\geq 0$, $$\mathcal E_{\alpha,u}(\gamma)
=
\begin{cases}
-\frac{\gamma^2}{(1+u)^2 \alpha+(1-\alpha)} \ \ &\text{if $\gamma<\gamma_c$}\\
-\alpha -\frac{(\gamma-(1+u)\alpha)^2}{(1-\alpha)}\ \ &\text{if $\gamma\geq \gamma_c$.}
\end{cases}
\ \text{ for $\gamma_c=\frac{(1+u)^2\alpha+(1-\alpha)}{(1+u)}$.}$$
\[rem: dicho\] [ The dichotomy in the log-measure is due to the fact that for $h$ with values beyond $\gamma_c\log\log T$, the intermediate values $X_h(\alpha)$ is restricted by the maximal level $\alpha\log\log T$. More precisely, consider $$\label{eqn: M}
\begin{aligned}
\mathcal{M}_T = & {\rm
Leb}\{h\in [0,1]: \widetilde X_h> \gamma \log
\log T\}\\
\mathcal{M}'_T =&\mathrm{Leb}\{h\in[0,1]:(1+u) X_h(\alpha) \ge
\lambda \log \log T\}\\\mathcal{M} ''_T =&\mathrm{Leb}\{h\in[0,1]:(1+u)
X_h(\alpha) \ge
\lambda \log \log T,X_h(\alpha,1)\ge
(\gamma-\lambda)\log \log T\}\ .
\end{aligned}$$ Clearly, we must have $\mathcal{M}_T''\leq\mathcal{M}_T$. It turns out that $\mathcal{M}_T''$ and $\mathcal{M}_T$ are comparable for an optimal choice of $\lambda$ given by, when $u<0 $, $$\label{eqn: lambda1}
\lambda^\star =\frac{\gamma (1+u)^2\alpha}{(1+u)^2\alpha+ 1-\alpha},\qquad
\gamma<\gamma^\star,$$ and when $ u>0$, $$\label{eqn: lambda2}
\lambda^\star =\begin{cases}\frac{\gamma (1+u)^2\alpha}{(1+u)^2\alpha+
1-\alpha}
& \text{if }0<\gamma<\gamma_c\ , \\
(1+u)\alpha\ & \text{if }\gamma_c\le \gamma<\gamma^\star.
\end{cases}$$ One can see this at a heuristic level by considering first moments. Since the maximum of $X_{h}$ is well approximated by the maximum over lattice points spaced $(\log T)^{-1} $ apart, there should be $\gamma$-high points only if $$(\log T)\cdot \mathcal{M}_T''\ge 1.\label{eq:GammaHighPtsExist}$$ Moreover, we have that if $\mathcal{M}_T'=0$, then $\mathcal{M}_T''=0$. And the maximum of $X_{h}(\alpha)$ is well approximated by the maximum over lattice points spaced $(\log T)^{-\alpha} $ apart, so there should be $\gamma$-high points only if$$(\log T)^{\alpha}\cdot \mathcal{M}_T'\ge 1.\label{eq:GammaHighPtsExistAlpha}$$ Since $X_h(\alpha)$ and $X_h(\alpha,1)$ are approximately Gaussian with variance $\frac{1}{2}\alpha\log\log T$ and $\frac{1}{2}(1-\alpha)\log\log T$, the following should hold approximately $$\begin{aligned}
&\frac{\log {\mathbb{E}}[ (\log T)^\alpha\cdot \mathcal{M}_T']}{\log\log T}=\alpha-\frac{\lambda^2}{(1+u)^2\alpha}+\operatorname{o}(1)
\\
&\frac{\log {\mathbb{E}}[(\log T)\cdot \mathcal{M}_T'']}{\log\log T}=1-\frac{\lambda^2}{(1+u)^2\alpha}-\frac{(\gamma-\lambda)^2}{1-\alpha}+\operatorname{o}(1)\end{aligned}$$ Together with conditions (\[eq:GammaHighPtsExist\]) and (\[eq:GammaHighPtsExistAlpha\]), we obtain constraints on the value of $\lambda$: $$\begin{aligned}
& \alpha-\frac{\lambda^2}{(1+u)^2\alpha}\ge 0 \ ,
\label{eq:MaximizationConstraintAlpha}
\\
&1-\frac{\lambda^2}{(1+u)^2\alpha}-\frac{(\gamma-\lambda)^2}{1-\alpha}\ge0 \ .
\label{eq:MaximizationConstraint}\end{aligned}$$ By maximizing $\mathcal{M}_T''$, under the constraints (\[eq:MaximizationConstraintAlpha\]) and (\[eq:MaximizationConstraint\]), one gets the values and for $\lambda$. ]{}
With Remark \[rem: dicho\] in mind, we are ready to bound the log-measure.
*Upper bound on the log-measure.* For $0<\gamma<\gamma^\star$, consider $\mathcal{M}_T$ as in . We need to show that for ${\varepsilon}>0$ $$\label{eqn: to prove UB}
\lim_{T\to\infty} {\mathbb{P}}\left(\mathcal{M}_T > (\log T)^{\mathcal E_{\alpha,u}(\gamma)+{\varepsilon}}\right)=0\ .$$ We first prove the easiest cases where $u\ge 0$ and $\gamma< \gamma_c$, as well as $u\leq 0$. Let ${\varepsilon}>0$. And write $V=1-\alpha +(1+u)^2 \alpha$ for short. Observe that by Markov’s inequality and Fubini’s theorem $$\begin{aligned}
{\mathbb{P}}(\mathcal M_T>(\log T)^{-\frac{\gamma^2}{V}+{\varepsilon}})
&\leq (\log T)^{\frac{\gamma^2}{V}-{\varepsilon}} \int_0^1 {\mathbb{P}}(\widetilde X_h>\gamma\log\log T)\ {{\rm d}}h\\
& =(\log T)^{\frac{\gamma^2}{V}-{\varepsilon}} {\mathbb{P}}(\widetilde X_h>\gamma\log\log T)\ ,
\end{aligned}$$ where we used the fact that the variables $\widetilde X_h$, $h\in[0,1]$, are identically distributed. Since ${\mathbb{P}}(\widetilde X_h>\gamma\log\log T)\ll \exp(-\gamma^2\log\log T/V)$ by Equation , the claim follows.
The case $u> 0$, $\gamma> \gamma_c$ is more delicate as we need to control the values at scale $\alpha$. For ${\varepsilon}'>0$ to be fixed later, note that the same argument as for equation gives $$\label{eqn: measure scale alpha}
\begin{aligned}
&{\mathbb{P}}\left({\rm Leb}\{h\in [0,1]: X_h(\alpha)>(\alpha+{\varepsilon}')\log\log T\}>0 \right)\\
&\hspace{2cm} \leq {\mathbb{P}}\left(\exists h\in [0,1]: X_h(\alpha)>(\alpha+{\varepsilon}')\log\log T \right)\to 0\ .
\end{aligned}$$ The same hold by symmetry for $-X_h(\alpha)$. This implies $$\begin{aligned}
&{\mathbb{P}}\left(\mathcal{M}_T > (\log T)^{\mathcal E_{\alpha,u}(\gamma)+{\varepsilon}}\right)\\
&={\mathbb{P}}\left({\rm Leb}\{h: \widetilde X_h>\gamma\log\log T,|X_h(\alpha)|\leq (\alpha+{\varepsilon}')\log\log T\}>(\log T)^{\mathcal E_{\alpha,u}(\gamma)+{\varepsilon}}\right)
+\operatorname{o}(1)\ .
\end{aligned}$$ It remains to prove that the first term is $\operatorname{o}(1)$. As in the proof of Lemma \[lem: max\], we consider the partition of $[0,1]$ by intervals $J_j$, $j\leq (\log T)^{\alpha}$, and the sub-partition $I_{jk}$, $k\leq (\log T)^{1-\alpha}$. We also divide the interval $[-(\alpha+{\varepsilon}')\log\log T, (\alpha+{\varepsilon}')\log\log T]$ into intervals $[q,q+1]$. Then by Markov’s inequality and the additivity of the Lebesgue measure $$\label{eqn: leb decomp}
\begin{aligned}
&{\mathbb{P}}\left({\rm Leb}\{h: \widetilde X_h>\gamma\log\log T,|X_h(\alpha)|\leq (\alpha+{\varepsilon}')\log\log T\}>(\log T)^{\mathcal E_{\alpha,u}(\gamma)+{\varepsilon}}\right)\\
&\leq(\log T)^{-\mathcal E_{\alpha,u}(\gamma)-{\varepsilon}} \sum_{j,k}\sum_{|q|\leq (\alpha+{\varepsilon}')\log\log T} {\mathbb{E}}\left[{\rm Leb}\{h\in I_{jk}: \widetilde X_h>\gamma\log\log T,X_h(\alpha)\in[q,q+1] \}\right]\\
& \leq (\log T)^{-\mathcal E_{\alpha,u}(\gamma)-{\varepsilon}}\sum_{j,k}\sum_{|q|\leq (\alpha+{\varepsilon}')\log\log T} (\log T)^{-1} {\mathbb{P}}\left(X_h(\alpha,1)>\gamma\log\log T-(1+u)(q+1),X_h(\alpha)\geq q \}\right)\ .
\end{aligned}$$ The last line follows from Fubini’s theorem and the fact that ${\rm Leb}(I_{jk})=(\log T)^{-1}$. The probabilities can be bounded by the Gaussian bound $$\begin{aligned}
&{\mathbb{P}}\left(X_h(\alpha,1)>\gamma\log\log T-(1+u)(q+1),X_h(\alpha)\geq q \}\right)\\
&\ll \exp\left(\frac{-q^2}{\alpha \log\log T}\right)\cdot \exp\left(\frac{-\Big(\gamma\log\log T-(1+u)(q+1)\Big)^2}{(1-\alpha) \log\log T}\right)\ .
\end{aligned}$$ It is easily checked that the expression is maximized at $q>(\alpha+{\varepsilon}')\log\log T$ for ${\varepsilon}'$. Moreover, at the optimal $q=(\alpha+{\varepsilon}')\log\log T$ in the considered range, the probability equals $(1+\operatorname{o}(1))(\log T)^{\mathcal E_{\alpha,u}(\gamma)}$. Using this observation to bound the probability for each $q$ in , we get $$\begin{aligned}
&{\mathbb{P}}\left({\rm Leb}\{h: \widetilde X_h>\gamma\log\log T,|X_h(\alpha)|\leq (\alpha+{\varepsilon}')\log\log T\}>(\log T)^{\mathcal E_{\alpha,u}(\gamma)+{\varepsilon}}\right)\\
&\ll(\log T)^{-{\varepsilon}} \log\log T =\operatorname{o}(1)\ .
\end{aligned}$$ This finishes the proof of the upper bound.\
*Lower bound on the log-measure.* For ${\varepsilon}>0$, the goal is to show $$\label{eq: lower bound goal}
{\mathbb{P}}\left(\mathcal{M}_T> (\log T)^{\mathcal E_{\alpha,u}(\gamma)-{\varepsilon}} \right)\to 1 \ \text{as $T\to\infty$.}$$ This is done using the Paley-Zygmund inequality, which states that for a random variable $\mathcal{M}\ge 0$ and $0\le \eta_T\le
1$, $${\mathbb{P}}\left( \mathcal{M}>\eta _TE\left[ \mathcal{M} \right] \right)
\ge \left(1-\eta _T \right)^2\frac{{\mathbb{E}}\left[ \mathcal{M} \right]^2}{{\mathbb{E}}\left[ \mathcal{M}^2
\right]}.\label{eq:PaleyZygmund}$$ We will have $\eta_T\to 0$, so the main task will be to demonstrate $${\mathbb{E}}\left[ \mathcal{M}^2 \right]=\left( 1+\operatorname{o}\left( 1 \right) \right){\mathbb{E}}\left[
\mathcal{M} \right]^2\ .
\label{eq:ShowPaleyZygmund}$$ This cannot be achieved when $\mathcal{M}=\mathcal{M}_T$ because of the correlations in $X_h$. To overcome this problem, we define a modified version of $\mathcal{M}_T$ by *coarse graining* the field as described in [@kistler].
For $K\in {\mathbb N}$ (that will depend eventually on ${\varepsilon}$), assume without loss of generality that $\{0,\frac{1}{K},\frac{2}{K},\dots,
\frac{K-1}{K},1\}$ is a partition of $[0,1]$ that is a refinement of $\{0,\alpha,1\}$. Consider $\lambda<\lambda^\star$ as defined in and , and $\delta>0$ (that will depend on ${\varepsilon}$). Define the events for the $K$-level coarse increments:$$\label{eq: corase increments}
\mathcal{J}_h(m)=\begin{cases}\left\{ (1+u)X_h\left(\frac{m-1 }{K}
,\frac{m }{K}
\right)\ge(1+\delta)
\frac{\lambda\log \log T}{\alpha K}\right\} & \text{if }m=1,\dots ,\alpha K, \\
\left\{ X_h\left(\frac{m-1 }{K} ,\frac{m }{K}
\right)\ge(1+\delta)
\frac{(\gamma-\lambda)\log \log T}{(1-\alpha )K}\right\}&\text{if }m=\alpha K+1,\dots,
K. \end{cases}$$ Moreover, define the sets $$\label{eq: modified exceedance}
\mathcal A=\{h: \mathcal{J}_h(m) \text{ occurs } \forall m=2,\dots,K\} \qquad
\mathcal B=\{h: (1+u)X_h\left(\frac{1}{K}\right)\leq-\frac{\delta}{2}\log\log T\} \ .$$ Note that if $h\in \mathcal A $, by adding up the inequalities in $\mathcal{J}_h(m)$, we have for $K$ large enough, $$\label{eq: delta 2 lower bound}
\widetilde X_h-(1+u)X
_h\left(
\frac{1 }{K} \right)\ge(1+\delta)\left(\gamma -
\frac{\lambda}{\alpha K}\right)\log \log T\ge\left( \gamma+\frac{\delta}{2} \right)\log \log T\ .$$ Therefore this implies the inclusion $$\mathcal A \subset \left\{h\in [0,1]: \widetilde X_h \ge \gamma\log\log T \right\}\cup \mathcal B\ ,$$ so that $\mathcal{M}_T\ge \mathrm{Leb}(\mathcal A) -\mathrm{Leb}(\mathcal B)$. Equation and Fubini’s theorem shows that ${\mathbb{E}}[\mathrm{Leb}(\mathcal B)]\ll(\log T)^{-\frac{\delta^2K}{4(1+u)^2}}$. For $K$ large enough, Markov’s inequality then implies $${\mathbb{P}}\left(\mathrm{Leb}\left\{h\in [0,1]:(1+u)X_h\left(
\frac{1 }{K} \right)\le -\frac{\delta}{2}\log
\log T\right\}\le (\log T)^{\mathcal{E}_{\alpha,u}(\gamma)-\varepsilon} \right)\to 1.$$ The proof of is then reduced to show $$\label{eq: only need to show}
{\mathbb{P}}(\mathrm{Leb}(\mathcal A)>2(\log T)^{ \mathcal{E}_{\alpha,u}(\gamma)-\varepsilon})
=
{\mathbb{P}}(\mathrm{Leb}(\mathcal A)>\eta_T{\mathbb{E}}[\mathrm{Leb}(\mathcal A)])
\to 1\ ,$$ where $\eta_T $ is defined by $2(\log T)^{ \mathcal{E}_{\alpha,u}(\gamma)-\varepsilon}=\eta_T{\mathbb{E}}[\mathrm{Leb}(\mathcal A)]$.
Following , we first show $\eta_T\to0$. By (\[eq: modified exceedance\]), Fubini’s theorem, and independence, $$\label{eq: product of prob increments}
{\mathbb{E}}[\mathrm{Leb}(\mathcal A)]=\int_0^1\prod
_{m=2}^K{\mathbb{P}}\left( \mathcal{J}_h(m) \right) {{\rm d}}h
=\prod_{m=2}^K{\mathbb{P}}\left( \mathcal{J}_h(m) \right),$$ since the $X_h$’s are identically distributed. By Proposition \[prop: berry-esseen\], $$\label{eq: probability of increment}{\mathbb{P}}(\mathcal{J}_h(m))
\gg
\begin{cases}(\log
T)^{-
\frac{ (1+\delta)^2\lambda^2 }{\alpha^2 K(1+u)^2}
+\operatorname{o}(1)} & \text{when }m=1,\dots ,\alpha K, \\
(\log T)^{-\frac{(1+\delta)^2
(\gamma-\lambda)^2}{(1-\alpha )^2K}
+\operatorname{o}(1)} & \text{when }m=\alpha K+1,\dots,
K. \\
\end{cases}$$Thus, by (\[eq: product of prob increments\]) and (\[eq: probability of increment\]), we have $$\label{eq: expectation of modified}
{\mathbb{E}}[ \mathrm{Leb}(\mathcal A)]
\gg(\log
T)^{ - \frac{ (1+\delta)^2\lambda^2 }{\alpha (1+u)^2}-\frac{(1+\delta)^2
(\gamma-\lambda)^2}{(1-\alpha )}}(\log T)^{\frac{(1+\delta)^2\lambda^2 }{\alpha^2 K(1+u)^2}+\operatorname{o}(1)}\ .$$ We can take $\lambda$ close enough to $\lambda^\star$, $\delta$ small enough, and $K$ large enough so that $${\mathbb{E}}[ \mathrm{Leb}(\mathcal A)]
\gg(\log
T)^{ - \frac{ {\lambda^\star}^2 }{\alpha (1+u)^2}-
\frac{ (\gamma-\lambda^\star)^2}{(1-\alpha )}-\frac{{\varepsilon}}{2}}=(\log T)^{ \mathcal E_{\alpha,u}(\gamma)-\frac{{\varepsilon}}{2}}\ ,$$ where we replace the value of $\lambda^\star$ of and . This shows that $\eta_T\to 0$. Observe that, we also have the reverse inequality $$\label{eqn: reverse}
{\mathbb{E}}[ \mathrm{Leb}(\mathcal A)] \ll(\log T)^{ \mathcal E_{\alpha,u}(\gamma)+\frac{{\varepsilon}}{2}}\ ,$$ using instead of Proposition \[prop: berry-esseen\].
It remains to show (\[eq:ShowPaleyZygmund\]). By independence of increments and Fubini’s theorem, we have $$\label{eq: second moment}
{\mathbb{E}}[{\mathrm{Leb}(\mathcal A)}^2]=
\int _0^1\int _0^1\prod _{m=2}^{ K} {\mathbb{P}}(\mathcal{J}_h(m)\cap \mathcal{\mathcal{J}}_{h'}(m))\
{{\rm d}}h{{\rm d}}h'\ .$$ We split the integral into four integrals: I for $|h-h'|>(\log T)^{\frac{-1}{2K}}$, II for $(\log T)^{\frac{-1}{K}}\leq |h-h'|\leq(\log T)^{\frac{-1}{2K}}$, III for $(\log T)^{\frac{-r}{K}}<|h-h'|\leq(\log T)^{\frac{-(r-1)}{K}}$, $r=2,\dots K$, and IV for $|h-h'|\leq (\log T)^{-1}$. We will show that $\mathrm I={\mathbb{E}}\left[{\mathrm{Leb}(\mathcal A)} \right]^2\left( 1+\operatorname{o}(1) \right)$ and the others $\operatorname{o}({\mathbb{E}}[{\mathrm{Leb}(\mathcal A)} ]^2$).
- For $\mathrm{II}$, note that $\mathrm{Leb}^{\times 2}\{(h,h'):(\log T)^{\frac{-1}{K}}\leq |h-h'|\leq(\log T)^{\frac{-1}{2K}} \}\ll(\log T)^{\frac{-1}{2K}}$. Moreover, by and Proposition \[prop: berry-esseen\], we have ${\mathbb{P}}(\mathcal{J}_h(m)\cap \mathcal{\mathcal{J}}_{h'}(m))\ll {\mathbb{P}}(\mathcal{J}_h(m))^2$. This implies $\mathrm{II}= \operatorname{o}({\mathbb{E}}[{\mathrm{Leb}(\mathcal A)} ]^2)$.
- For $\mathrm{IV}$, note that clearly ${\mathbb{P}}(\mathcal{J}_h(m)\cap \mathcal{\mathcal{J}}_{h'}(m))\leq {\mathbb{P}}(\mathcal{J}_h(m))$. Thus, $\mathrm{IV}\ll (\log T)^{-1}{\mathbb{E}}[\mathrm{Leb}(\mathcal A)]$. Using and the fact that $1+ \mathcal{E}_{\alpha,u}(\gamma)>0$ for $\gamma<\gamma^\star$, one gets $\mathrm{IV}= \operatorname{o}({\mathbb{E}}[{\mathrm{Leb}(\mathcal A)} ]^2)$.
- For $\mathrm{I}$, note that $\mathrm{Leb}^{\times 2}\{(h,h'): |h-h'|>(\log T)^{\frac{-1}{2K}} \}=1+\operatorname{o}(1)$. Moreover, by Proposition \[prop: berry-esseen\], ${\mathbb{P}}(\mathcal{J}_h(m)\cap \mathcal{\mathcal{J}}_{h'}(m))=(1+\operatorname{o}(1)) {\mathbb{P}}(\mathcal{J}_h(m))^2$. This implies $\mathrm{I}= (1+\operatorname{o}(1)){\mathbb{E}}[{\mathrm{Leb}(\mathcal A)} ]^2)$.
- For $\mathrm{III}$, the integral is a sum over $r=2,\dots,K$ of integrals of pairs with $(\log T)^{\frac{-r}{K}}<|h-h'|\leq(\log T)^{\frac{-(r-1)}{K}}$. The measure of this set is $\ll (\log T)^{\frac{-(r-1)}{K}}$. For fix $r$, the integrand is $$\begin{aligned}
\prod_{m=2}^K {\mathbb{P}}(\mathcal{J}_h(m)\cap \mathcal{\mathcal{J}}_{h'}(m))
&\leq \prod_{m=2}^r {\mathbb{P}}(\mathcal{J}_h(m))\prod_{m=r+1}^K {\mathbb{P}}(\mathcal{J}_h(m)\cap \mathcal{\mathcal{J}}_{h'}(m))\\
&\ll \prod_{m=2}^r {\mathbb{P}}(\mathcal{J}_h(m))\prod_{m=r+1}^K {\mathbb{P}}(\mathcal{J}_h(m))^2\ ,
\end{aligned}$$ where the last line follows by and Proposition \[prop: berry-esseen\]. Putting all this together and factoring the square of the one-point probabilities, one gets $$\mathrm{III}
\ll{\mathbb{E}}[{\mathrm{Leb}(\mathcal A)} ]^2 \sum_{r=2}^K(\log T)^{\frac{-(r-1)}{K}} \prod_{m=2}^r \Big({\mathbb{P}}(\mathcal{J}_h(m))\Big)^{-1}\ .$$ We show $\prod_{m=2}^r \Big({\mathbb{P}}(\mathcal{J}_h(m))\Big)^{-1}<(\log T)^{\frac{(r-1)}{K}}$ uniformly in $T$. This finishes the proof since the sum is then the tail of a convergent geometric series. In the case $u<0$, since $\lambda<\lambda^\star$, and $(1+\delta)\gamma<\gamma^\star$ for $\delta$ small, we have by , $${\mathbb{P}}(\mathcal{J}_h(m))^{-1}
\ll
\begin{cases}
(\log T)^{\frac{{\lambda^\star}^2}{\alpha^2K(1+u)^2}} \ &\text{ if $m\leq \alpha K$}\\
(\log T)^{\frac{(\gamma^\star-\lambda^\star)^2}{(1-\alpha)^2K}} \ &\text{ if $m=\alpha K+1,\dots,K$.}\\
\end{cases}$$ By the definition of $\lambda^\star$ and $\gamma^\star=V^{1/2}$, this implies $$\prod_{m=2}^r \Big({\mathbb{P}}(\mathcal{J}_h(m))\Big)^{-1}
\ll
\begin{cases}
(\log T)^{\frac{(1+u)^2}{V}\frac{r-1}{K}} \ &\text{ if $r\leq \alpha K$}\\
(\log T)^{\frac{\alpha(1+u)^2}{V}+\frac{1}{V} \frac{r-\alpha K}{K}} \ &\text{ if $r=\alpha K+1,\dots,K$.}\\
\end{cases}$$ Since $u<0$, it is straightforward to check that the exponent is smaller than $\frac{r-1}{K}$ as claimed. The case $u\geq 0$ is done similarly by splitting into two cases $\gamma_c\leq\gamma<\gamma^\star$ and $0<\gamma<\gamma_c$. We omit the proof for conciseness.
We now have all the results to finish the proof of Proposition \[prop: free energy\] using Laplace’s method.
We first prove the limit in probability. The convergence in $L^1$, and in particular the convergence of the expectation, will be a consequence of Lemma \[lem: UI\] below. For fixed ${\varepsilon}>0$ and $M\in {\mathbb N}$, consider $$\gamma_j=\frac{j(1+{\varepsilon})}{M}\gamma^\star \qquad 0\leq j\leq M\ ,$$ and the event $$\begin{gathered}
A=\bigcap_{j=1}^M
\left\{(\log T)^{\mathcal E_{\alpha,u}(\gamma_j)-{\varepsilon}}\leq {\rm Leb}\{h: \widetilde X_h> \gamma_j \log \log T\} \leq (\log T)^{\mathcal E_{\alpha,u}(\gamma_j)+{\varepsilon}} \right\}\\
\bigcap \left\{ {\rm Leb}\{h: \widetilde X_h> \gamma_M\log \log T\}=0\right\}\ .\end{gathered}$$ By Lemma \[lem: max\] and Lemma \[lem: high points\], we have that ${\mathbb{P}}(A^c)\to 0$ as $T\to\infty$. It remains to prove that the free energy is close to the claimed expression on the event $A$. On one hand, the following upper bound holds on $A$: $$\begin{aligned}
\int_0^1 \exp \beta \widetilde X_h \ {{\rm d}}h&\leq \sum_{j=1}^M \int_0^1 \exp \beta \widetilde X_h\ {\mathbbm{1}}_{\{(\log T)^{\gamma_{j-1}}<e^{\widetilde X_h}\leq(\log T)^{\gamma_{j}}\}} \ {{\rm d}}h
+\int_0^1 \exp \beta \widetilde X_h\ {\mathbbm{1}}_{\{e^{\widetilde X_h}<1\}} \ {{\rm d}}h\\
& \leq \sum_{j=1}^M (\log T)^{\beta \gamma_{j}+\mathcal E_{\alpha,u}(\gamma_{j-1})+{\varepsilon}}+1\ .
\end{aligned}$$ On the other hand, we have the lower bound $$\begin{aligned}
\int_0^1 \exp \beta \widetilde X_h \ {{\rm d}}h&\geq \sum_{j=1}^M \int_0^1 \exp \beta \widetilde X_h\ {\mathbbm{1}}_{\{(\log T)^{\gamma_{j-1}}<e^{\widetilde X_h}\leq(\log T)^{\gamma_{j}}\}} \ {{\rm d}}h
& \geq \sum_{j=1}^M (\log T)^{\beta \gamma_{j-1}+\mathcal E_{\alpha,u}(\gamma_{j})-{\varepsilon}}\ .
\end{aligned}$$ Altogether, this implies $$\max_{1\leq j\leq M}\left\{\beta \gamma_{j-1}+\mathcal E_{\alpha,u}(\gamma_{j})-{\varepsilon}\right\} \leq \frac{\log \int_0^1 \exp \beta \widetilde X_h \ {{\rm d}}h}{\log\log T} \leq \max_{1\leq j\leq M}\left\{\beta \gamma_{j}+\mathcal E_{\alpha,u}(\gamma_{j-1})+{\varepsilon}\right\}+\operatorname{o}(1)\ .$$ In particular, by continuity of $\mathcal E_{\alpha,u}(\gamma)$, we can pick $M$ large enough depending on ${\varepsilon}$ and $T$ large enough so that $$\left|\frac{\log \int_0^1 \exp \beta \widetilde X_h \ {{\rm d}}h}{\log\log T} -\max_{\gamma\in [0,\gamma^{\star}]}\left\{\beta \gamma+\mathcal E_{\alpha,u}(\gamma)\right\} \right|
\leq 2{\varepsilon}\ .$$ As mentioned above, since ${\mathbb{P}}(A^c)\to 0$ as $T\to \infty$, this proves the convergence in probability $$\lim_{T\to\infty}\frac{\log \int_0^1 \exp \beta \widetilde X_h \ {{\rm d}}h}{\log\log T}= \max_{\gamma\in [0,\gamma^{\star}]}\left\{\beta \gamma+\mathcal E_{\alpha,u}(\gamma)\right\}\ .$$ It remains to check that the right side has the desired form. Let $V=(1+u)^2 \alpha+(1-\alpha)$. If $u<0$, the optimal $\gamma$ is $\beta V/2$ whenever $\beta V/2<\gamma^\star$, i.e., $\beta<2/V^{1/2}$. If $\beta\geq2/V^{1/2}$, then the optimal $\gamma$ is simply $\gamma^\star$. Therefore, we have $$\max_{\gamma\in [0,\gamma^{\star}]}\left\{\beta \gamma+\mathcal E_{\alpha,u}(\gamma)\right\}
=\begin{cases}
\frac{\beta^2V}{4} \ \ &\text{if $\beta<2/V^{1/2}$}\\
\beta V^{1/2}-1\ \ &\text{if $\beta\geq2/V^{1/2}$}\ .
\end{cases}$$ If $u\geq 0$, the optimal $\gamma$ is $\beta V/2$ if $\gamma <\gamma_c$, i.e., $\beta <2/(1+u)$. If $\gamma>\gamma_c$, then the optimal $\gamma$ is $(1+u)\alpha+\frac{\beta(1-\alpha)}{2}$ until it equals $\gamma^\star$. This happens at $\beta\geq 2$. Putting all this together, we obtain that $$\max_{\gamma\in [0,\gamma^{\star}]}\left\{\beta \gamma+\mathcal E_{\alpha,u}(\gamma)\right\}=
\begin{cases}
\frac{\beta^2\big((1+u)^2 \alpha+(1-\alpha)\big)}{4} \ \ &\text{if $\beta<\frac{2}{(1+u)}$}\\
\beta (1+u)\alpha -\alpha + \frac{\beta^2(1-\alpha)}{4}\ \ &\text{if $\frac{2}{(1+u)}\leq \beta<2$}\\
\beta \Big((1+u)\alpha+(1-\alpha)\Big)-1\ \ &\text{if $\beta\geq2$}\ .
\end{cases}$$ This corresponds to the expression in Proposition \[prop: free energy\] expressed in terms of .
\[lem: UI\] The sequence of random variables $$\Big(\frac{1}{\log\log T}\log \int_{0}^1 \exp\big(\beta (X_h+uX_{h}(\alpha)\big) {{\rm d}}h\Big)_{T>1}$$ is uniformly integrable. In particular, the convergence in probability of the sequence is equivalent to the convergence in $L^1$.
Write for short $$f_T
=(\log\log T)^{-1}\log \int_{0}^1 \exp \beta \widetilde X_h\ {{\rm d}}h\ .$$ We need to show that for any $\varepsilon>0$, there exists $C$ large enough so that uniformly in $T$, $${\mathbb{E}}[|f_T| {\mathbbm{1}}_{\{|f_T)|>C\}}]<\varepsilon \ .$$ It is easy to check that $$\label{eqn: UI}
{\mathbb{E}}[|f_T| {\mathbbm{1}}_{\{|f_T|>C\}}]=\int_C^\infty {\mathbb{P}}(f_T>y) \ {{\rm d}}y + C {\mathbb{P}}(f_T>C) + \int_{-\infty}^{-C} {\mathbb{P}}(f_T<y) \ {{\rm d}}y + C {\mathbb{P}}(f_T<-C)\ .$$ Therefore, it remains to get a good control on the right and left tail of $f_T$. For the right tail, observe that by Markov’s inequality $$\begin{aligned}
{\mathbb{P}}(f_T>y) ={\mathbb{P}}\Big(\int\exp\beta \widetilde X_h\ {{\rm d}}h>(\log T)^y\Big)
\leq (\log T)^{-y}\ {\mathbb{E}}\left[\int\exp\beta \widetilde X_h\ {{\rm d}}h\right]\ .
\end{aligned}$$ Using Proposition \[prop: MGF\] and Fubini’s theorem, we get $${\mathbb{P}}(f_T>y) \ll (\log T)^{((1+u)^2\alpha+(1-\alpha))\frac{\beta^2}{4}-y}\ .$$ This implies $$\int_C^\infty {\mathbb{P}}(f_T>y) \ {{\rm d}}y + C {\mathbb{P}}(f_T>C)
\ll\frac{(\log T)^{((1+u)^2\alpha+(1-\alpha))\frac{\beta^2}{4}-C}}{\log\log T}+C(\log T)^{((1+u)^2\alpha+(1-\alpha))\frac{\beta^2}{4}-C}\ .$$ It suffices to take $C>((1+u)^2\alpha+(1-\alpha))\frac{\beta^2}{4}$ for this to be uniformly small in $T$. The left tail is bounded the same way after noticing that by Markov’s and Jensen’s inequalities, $$\begin{aligned}
{\mathbb{P}}(f_T<-y) ={\mathbb{P}}\Big(\int \exp\beta \widetilde X_h\ {{\rm d}}h<(\log T)^{-y}\Big)
&\leq (\log T)^{-y}\ {\mathbb{E}}\left[\left(\int \exp\beta \widetilde X_h\ {{\rm d}}h\right)^{-1}\right]\\
&\leq (\log T)^{-y}\ {\mathbb{E}}\left[\int \exp-\beta \widetilde X_h\ {{\rm d}}h\right]\\
& \ll (\log T)^{((1+u)^2\alpha+(1-\alpha))\frac{\beta^2}{4}-y}\ .
\end{aligned}$$ These estimates imply that ${\mathbb{E}}[|f_T| {\mathbbm{1}}_{\{|f_T)|>C\}}]$ can be made arbitrarily small in by taking $C$ larger than $((1+u)^2\alpha+(1-\alpha))\frac{\beta^2}{4}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In the past years, the folding kinetics of many small single-domain proteins has been characterized by mutational $\Phi$-value analysis. In this article, a simple, essentially parameter-free model is introduced which derives folding routes from native structures by minimizing the entropic loop-closure cost during folding. The model predicts characteristic folding sequences of structural elements such as helices and $\beta$-strand pairings. Based on few simple rules, the kinetic impact of these structural elements is estimated from the routes and compared to average experimental $\Phi$-values for the helices and strands of 15 small, well-characterized proteins. The comparison leads on average to a correlation coefficient of 0.62 for all proteins with polarized $\Phi$-value distributions, and 0.74 if distributions with negative average $\Phi$-values are excluded. The diffuse $\Phi$-value distributions of the remaining proteins are reproduced correctly. The model shows that $\Phi$-value distributions, averaged over secondary structural elements, can often be traced back to entropic loop-closure events, but also indicates energetic preferences in the case of a few proteins governed by parallel folding processes.'
address: 'Max-Planck-Institut für Kolloid- und Grenzflächenforschung, 14424 Potsdam, Germany'
author:
- 'Thomas R. Weikl'
title: 'Loop-closure events during protein folding: Rationalizing the shape of $\Phi$-value distributions'
---
Introduction
============
Small single-domain proteins with less than 100 amino acids typically are two-state folders [@jackson98; @fersht99; @grantcharova01]. These proteins fold from the denatured to the native state without populating experimentally detectable intermediate states [@fersht99]. In recent years, the folding kinetics of many two-state proteins has been characterized by mutational $\Phi$-value analysis. A $\Phi$-value is a measure for the impact of a mutation on the folding kinetics, defined as $$\Phi=- \frac{R T\ln k'/k}{\Delta G' - \Delta G} \label{phi}$$ where $k$ and $\Delta G$ are the folding rate and stability of the wildtype protein, and $k'$ and $\Delta G'$ are the corresponding quantities of the mutant [@matouschek89; @fersht99]. For various two-state proteins, detailed $\Phi$-value distributions have been obtained by considering many single-residue mutations throughout the protein chains. A central question is why some proteins have polarized $\Phi$-value distributions, while others have diffuse distributions. In a polarized distribution, the $\Phi$-values for mutations in some of the secondary structural elements of the protein are significantly larger than the values in other secondary elements. In a diffuse distribution, the average $\Phi$-values for the secondary elements of the protein are rather similar.
Several results seem to indicate that the folding kinetics of two-state proteins is dominated by their native-state topology [@baker00]. Most importantly, the folding times of two-state folders have been found to correlate with the relative contact order (CO) of their native structures [@plaxco98; @plaxco00; @kuznetsov04; @dinner01; @ivankov03]. The relative CO of a protein is the average contact order, or ‘localness’, $|i-j|$ of all native contacts $(i,j)$, divided by the chain length of the protein. Here, $i$ and $j$ indicate the sequence positions of two residues in contact. The correlation holds for folding times over 6 orders of magnitude, from microseconds for $\alpha$-helical proteins with low relative CO to seconds for $\beta$-sheet-containing proteins with high relative CO. Comparable correlations with folding times have also been found for other measures of native-state topology [@gromiha01; @makarov02; @micheletti02; @micheletti03; @gong03; @nolting03].
However, reproducing detailed $\Phi$-value distributions in theoretical models which are based on native-state topology has been proven to be difficult. Several theoretical models derive folding routes or $\Phi$-values from native structures. [@alm99; @munoz99; @galzitskaya99; @guerois00; @ivankov01; @alm02; @bruscolini02; @karanicolas03; @clementi00; @hoang00; @li01; @karanicolas02; @ding02; @clementi03; @brown04; @portman01; @kameda03; @garbuzynskiy04]. Some of these models assume that the amino acid residues can be in either of two states, native-like folded or unfolded [@alm99; @munoz99; @galzitskaya99; @guerois00; @ivankov01; @alm02; @bruscolini02; @karanicolas03]. In this respect, the models are similar to the Zimm-Bragg model for helix-coil transitions where residues can either be in a helix or in a coil state [@zimm59], or to Ising models where particles can either have spin up or down. Other models use explicit chain representations of the proteins and simplified Go-type potential energies which impose the native structure by postulating favorable interaction energies only between pairs of residues that are in contact in the native structure [@clementi00; @hoang00; @li01; @karanicolas02; @ding02; @clementi03; @brown04]. The unfolding kinetics of proteins has also been considered in Molecular Dynamics simulations with all-atom models [@li96; @lazaridis97; @gsponer01; @dejong02; @paci02; @morra03]. Some of these models have been used to calculated $\Phi$-values either for a single protein or a small number of proteins [@alm99; @munoz99; @galzitskaya99; @clementi00; @guerois00; @li01; @karanicolas02; @clementi03; @brown04]. A systematic comparison for a set of 19 proteins has been performed by Alm et al. with an Ising-like model [@alm02]. For more than half of the proteins, Alm et al. obtain correlation coefficients $r$ from 0.41 to 0.88 between theoretical and experimental $\Phi$-values, and for 14 of the 19 proteins, the theoretical $\Phi$-values were better than random permutations of the experimental values. Kameda [@kameda03] has considered a Gaussian chain model with a Go-type interaction potential and obtains positive correlation coefficients $r$ between 0.12 and 0.65 for 7 out of 12 proteins. More recently, Garbuzynskiy et al. [@garbuzynskiy04] have reproduced the $\Phi$-value distributions of 17 proteins with an average correlation coefficient of 0.54.
The model presented here focuses on [*average*]{} $\Phi$-values for the secondary structural elements of a protein. The starting point of the model are native contact maps. The native contact map of a protein is a matrix in which element $(i,j)$ equals 1 if the two residues $i$ and $j$ are in contact in the native structure, and 0 otherwise. The contacts in the native contact map of a protein typically are arranged in clusters. These contact clusters correspond to structural elements such as $\alpha$-helices and $\beta$-strand pairings.
In the first step, the model derives folding routes from native contact maps. In this step, the model considers all sequences in which the contact clusters, or structural elements, can be formed. The key assumption of the model is that the dominant folding routes can be identified as those sequences of events which minimize the loop-closure cost and, hence, the entropic barriers during folding. The loop-closure cost of a folding sequence simply is defined as the sum of loop lengths for forming the clusters along that sequence. These loops lengths are estimated via the graph-theoretical concept of effective contact order ECO [@fiebig93; @dill93] (see Fig. \[eco\]). The ECOs and, thus, the loop-closure cost for forming nonlocal structural elements typically can be reduced by the previous formation of other, more local structural elements.
In the second step, the model estimates the kinetic impact of contact clusters and secondary structural elements from the folding routes. In the model, the kinetic impact of a contact cluster depends on how often the cluster appears on the folding routes to (other) nonlocal clusters, and on the ECOs of the cluster. The kinetic impact derived from the folding routes is compared to average experimental $\Phi$-values for the secondary structural elements. To test the model systematically, 15 proteins are considered which (i) are small in the sense that they have less than 10 contact clusters, and which (ii) are well-characterized in the sense that $\Phi$-values for at least 10 residue positions are available. The comparison between kinetic impact and average $\Phi$-values leads on average to a correlation coefficient of 0.62 for all 12 proteins with polarized $\Phi$-value distributions (see Fig. \[correlationCoefficients\]), and to an average correlation coefficient of 0.74 if three proteins with negative average $\Phi$-values are excluded. The three proteins have negative average $\Phi$-values below -0.1 in one of the secondary elements, which are difficult to interpret [@ozkan01; @li00]. The remaining proteins have diffuse $\Phi$-value distributions with similar average $\Phi$-values in the secondary elements. In agreement with the experiments, the distribution of kinetic impact for these proteins is also diffuse. The model thus shows that the polarized of diffuse shapes of most averaged $\Phi$-value distributions can be traced back to native-state topologies.
The minimum-ECO-routes defined here represent maybe the simplest possible topology-based modeling of protein folding routes. The prediction of these routes requires the definition of contact maps and contact clusters, but no parameter fitting since the routes are defined as minima of a loop-closure cost function in the space of possible folding sequences. Why can such a simple prediction of folding routes, in combination with a few rules for estimating the kinetic impact of structural elements, reproduce central aspects of mutational experiments? The reason seems to be that the barrier for protein folding is entropic. Furthermore, the relevant entropy here should be loop-closure entropy, since other entropic contributions like the entropy-loss for side-chain ‘freezing’ in contact clusters, once the loop is closed, should be rather independent of the specific route, i.e. of the sequences in which the clusters are formed. Clearly, this simple modeling has its limitations. The model is limited to average $\Phi$-values for secondary elements, since the realistic modeling of detailed $\Phi$-value distributions requires also energetic characteristics of the specific mutations [@merlo]. Another limitation is that the model can not address folding rates. The modeling of folding rates requires an additional estimate for the intrinsic, route-independent formation times of the contact clusters, besides loop-closure. In a previous related model, these intrinsic cluster formation times have been estimated via the number of steps required for ‘zipping up’, or ‘propagating’, a contact cluster after the initial loop-closure step [@weikl03a; @weikl03b]. This previous model has five parameters, which were fitted to the folding rates of 24 two-state folders, and considers a more complex set of partially formed zipping states of the clusters.
The model presented here is purely topology-based in the sense that it does not use any sequence-specific information. A central question is whether purely topology-based models can account for the experimentally observed differences in the folding kinetics between proteins with similar folds, or similar overall fold topology. A famous example are protein L and G, which both have a central $\alpha$-helix and two rather symmetric $\beta$-hairpins at the chain ends. Intriguingly, the structural symmetry is ‘broken’ in the $\Phi$-value distributions, and in each of these protein in a different way: protein L has the largest $\Phi$-values in the N-terminal hairpin [@kim00], and protein G in the C-terminal hairpin [@mccallister00]. In the model presented here, native-state topology is captured by the topology of the native contact maps. Protein L and G have very similar folds, but nonetheless small differences in their contact maps. This leads to different folding routes which reproduce the observed symmetry breaking for the two proteins. Other groups have used sequence-specific interaction energies in topology-based models to account for these differences between protein L and G [@karanicolas02; @clementi03; @kameda03; @brown04]. The present model traces the symmetry breaking of protein L and G back to native-state topology, but suggests that sequence-specific energetic contributions may affect the folding kinetics of Sso7d and CspB. According to the model, these proteins are governed by parallel folding processes with similar loop-closure cost. However, the experimental $\Phi$-value distributions seem to indicate that one of the parallel processes dominates the kinetics, presumably due to specific energetic interactions which are not considered in the model.
The model
=========
Folding routes
--------------
The starting point of the model are native contact maps. The native contact map of a protein is a matrix in which the element $(i,j)$ equals 1 if the two residues $i$ and $j$ are in contact in the native structure, and 0 otherwise. Here, two residues are taken to be in contact if the distance between their $C_\alpha$ or $C_\beta$ atoms is less than 6 Å, and if they are not nearest or next-nearest neighbors in the sequence. The native contacts are grouped into contact clusters (for details, see Methods section). These contact clusters correspond to the structural elements of the protein: helices, $\beta$-strand pairings, and tertiary interactions of helices or $\beta$-sheets. The contact maps and contact clusters of the 15 proteins considered here are shown in Fig. \[contactMaps\]. The contact clusters of a protein can be divided into local and nonlocal clusters. Local clusters contain at least one local contact $(i,j)$ with small contact order CO =$|i-j|<10$, whereas nonlocal clusters do not contain any such local contacts.
In the model, folding routes are derived from the loop-closure dependencies between the contact clusters. To determine the loop-closure relations, all possible sequences are considered in which the clusters can be formed. For a nonlocal cluster, the length of the loop which has to be closed to form the cluster contacts depends on these sequences of cluster formation. In other words, it depends on which other clusters have been formed previously. A simple example with only four contact clusters is CI2 (see Fig. \[contactMaps\]). The $\beta_1\beta_4$ cluster of CI2 consists of nonlocal contacts between the two chain ends. Forming these contacts from the fully unfolded state requires the closure of a relatively large loop, and hence costs a large amount of loop-closure entropy. However, forming one or several of the structural elements $\alpha$, $\beta_2\beta_3$, or $\beta_3\beta_4$ prior to $\beta_1\beta_4$ brings the two chain ends into closer spatial proximity and reduces the length of the loop which has to be closed to form $\beta_1\beta_4$.
The length of the loop which is closed to form a specific contact between two residues is estimated here via the concept of [*effective contact order*]{} (ECO) [@fiebig93; @dill93]. The ECO of the contact is the number of steps along the shortest path between these two residues. Each ‘step’ either is (i) a covalent bond between consecutive residues in the chain, or (ii) a previously formed noncovalent contact (see Fig. \[eco\]). In contrast, the contact order (CO) only takes into account steps of type (i) and hence measures the sequence separation of the two residues. Unlike the ECO, the CO is independent of the folding route, the sequence in which contacts are formed.
The key assumption of the model is that folding routes which involve only closures of relatively small loops dominate the folding process. These routes minimize the entropic loop-closure barriers during folding. To determine the minimum-entropy-loss routes, all possible sequences of cluster formation are considered. Sequences of cluster formation here are called [*folding sequences*]{}. The formation of each contact cluster in a folding sequence requires to close a loop. The length of this loop is estimated as the minimum ECO among all cluster contacts, the [*cluster ECO*]{}. The cluster ECO thus is an estimate for the length of the shortest loop that has to be closed to form the cluster in a given partially folded conformation[^1]. Suppose we have a sequence of clusters $C_1C_2 \ldots C_n$. Since no contacts have been formed prior to $C_1$, the ECO $\ell_1$ of this cluster simply is the minimum CO among the cluster contacts. For the other clusters $C_i$ in the folding sequence, the cluster ECO is the minimum ECO among the cluster contacts, given the contacts of the previously formed clusters $C_1, C_2, \ldots, C_{i-1}$. This leads to a sequence of cluster ECOs, or loop lengths, $\ell_1, \ell_2, \ldots, \ell_n$.
For each folding sequence $C_1C_2 \ldots C_n$, the total loop-closure cost can be defined as $s=\sum_{i=1}^n f(\ell_i)$ where $\ell_i$ are the cluster ECOs along the sequence, and $f(\ell_i)$ is a weighting function which increases with the loop length $\ell_i$. For simplicity, the linear weighting function $f(\ell_i)= \ell_i$ is used here [@weikl04]. This linear approximation for the free-energy cost of loop closure is not unreasonable since the range of relevant ECOs here only spans roughly one order of magnitude, from 2 to 20 or 30 (see Table 1). The total loop-closure cost then simply is the sum of ECOs $s=\sum_{i=1}^n \ell_i$ for all clusters along the sequence.
The [*minimum-ECO sequences*]{} to a given cluster $C_n$ are simply defined as local minima of the loop-closure cost $s$ in the space of all possible folding sequences to $C_n$. In this space, the neighbors of a given folding sequence $C_1C_2 \ldots C_n$ are those sequences which are obtained either by deleting one or several of the clusters from $C_1C_2 \ldots C_n$, or by adding one or several ‘new’ clusters somewhere in the sequence (see also Methods section). In principle, two neighboring folding sequences can have the same local minimum value of $s$. In this case, the longer sequence among the two is selected as the minimum-ECO sequence.
Finally, all minimum-ECO sequences which consist of the same set of clusters are taken to represent the same [*minimum-ECO route*]{}. These sequences have the same loop-closure cost $s$ and differ only by permutations from each other, which indicates parallel folding processes on the route. Suppose the ECO of the nonlocal cluster $C_3$ is only affected by the two local clusters $C_1$ and $C_2$. Since the ECOs of the local clusters $C_1$ and $C_2$ are independent of each other, the two sequences $C_1C_2C_3$ and $C_2C_1C_3$ then both are minimum-ECO sequences, representing the same minimum-ECO route. On this minimum-ECO route, the two local clusters $C_1$ and $C_2$ form in parallel, prior to the nonlocal cluster $C_3$.
Table 1 summarizes the loop-closure hierarchies on the minimum-ECO routes for the proteins considered here. For each nonlocal cluster of a protein, all clusters formed prior on the minimum-ECO route are shown. For some nonlocal clusters, there are multiple minimum-ECO routes. These multiple routes correspond to different local minima of the loop-closure cost $s$ in the space of folding sequences. However, local minima with a loop-closure cost $s$ which is by 10 or more larger than the global minimum are neglected. These local minima represent folding routes with significantly larger entropic barriers.
Kinetic impact of secondary structural elements
-----------------------------------------------
The most important kinetic data for two-state folders are $\Phi$-values, which reflect the impact of mutations on the folding kinetics (see eq. (\[phi\])). A $\Phi$-value distribution for a protein is obtained by considering many single-residue mutations throughout the protein chain. For comparison with the model, the experimental $\Phi$-value distributions here are averaged over whole secondary structural elements (helices or sheets). These average $\Phi$-values typically are positive and indicate the relative ‘kinetic importance’, or ‘kinetic impact’, of the secondary structural elements. For example, a relatively large average $\Phi$-value for a secondary element indicates that mutations in this element have a strong impact on the folding rate.
In order to compare with average experimental $\Phi$-values, the kinetic impact of secondary structural elements here is estimated from the loop-closure hierarchies summarized in Table 1. For this purpose, we first have to the consider the kinetic impact of the contact clusters. In a semi-quantitative approach, the kinetic impact of contact clusters and secondary structural elements here is divided into high (H), medium (M), or low (L).
First, it seems reasonable to assume that the kinetic impact of a cluster should be related to how often it appears on the minimum-ECO routes to other clusters. Suppose a local cluster appears on minimum-ECO routes to all non-local clusters. Mutations affecting the formation of this cluster then should strongly affect the overall folding kinetics. Hence, the cluster has a high kinetic impact. To quantify this notion, the [*occurrence number*]{} $n$ of a cluster is defined as the number of times it appears on all routes to all (other) nonlocal clusters. In other words, $n$ simply is the number of times the cluster occurs in the third column of Table 1. In terms of occurrence numbers, the first rule is:
1. The kinetic impact of a cluster is high (H) if its occurrence number $n$ on the minimum-ECO routes is larger than or equal to $\frac{2}{3} n_\text{max}$. Here, $n_{max}$ is the maximum value of $n$ among all clusters of the protein. The impact of the cluster is medium (M) for $\frac{1}{3}n_\text{max}\le n <\frac{2}{3}n_\text{max}$. The impact is low (L) for $n <\frac{1}{3}n_\text{max}$.
Second, the kinetic impact of nonlocal clusters should also be affected by the cluster ECO. Suppose a nonlocal cluster has a high cluster ECO on all minimum-ECO routes. This means that forming the cluster always involves the closure of a relatively large loop. It seems reasonable to assume that the kinetic impact of the cluster then is high, since the contacts of these clusters have to balance a relatively high loop-closure entropy. In other words, the formation of the cluster and, hence, the overall folding kinetics should be highly sensitive to mutations affecting the cluster contacts. The second rule is:
1. A nonlocal cluster has a high (H) kinetic impact if the ECO of this cluster is larger than 10 an all routes. The kinetic impact is medium (M) if the smallest cluster ECO has a value from 6 to 10, unless rule (1) specifies high impact.
According to the rules (1) and (2), the kinetic impact of a cluster thus is low if its occurrence number is small, and the cluster ECO is not larger than 5.
Finally, suppose a protein has two nonlocal clusters $C_1$ and $C_2$ which fold in parallel. This means that the cluster $C_1$ does not appear on the minimum-ECO routes to $C_2$, and vice versa. In general, the loop-closure cost for forming, e.g, $C_1$ can be significantly larger than the loop-closure cost for forming $C_2$. It seems reasonable that clusters appearing on the minimum-ECO routes to $C_1$ should then have a higher kinetic impact than clusters appearing only on minimum-ECO routes to $C_2$, since the entropic loop-closure barrier for forming $C_1$ is significantly larger. Therefore, the third rule is:
1. If two nonlocal clusters $C_1$ and $C_2$ do not occur on minimum-ECO routes to other clusters and have minimum loop-closure costs $s_1$ and $s_2$ with $s_1 > s_2 + 5$, the cluster occurrences on the routes to $C_2$ are not taken into account in rule (1). In particular, clusters which appear only on routes to $C_2$ have a low kinetic impact, independent of their ECO.
The rules (1), (2), and (3) define the kinetic impact of clusters. The translation into kinetic impact of secondary elements (strands or helices) is straightforward. The kinetic impact of a secondary element is high (H) if it has contacts in a cluster with high kinetic impact, and low (L) if it only has contacts in clusters with low kinetic impact. The kinetic impact of a secondary element is medium (M) if it has contacts in clusters with medium kinetic impact, but no contacts in clusters with high kinetic impact. As an example, the high kinetic impact of the clusters $\alpha_i$ and $\beta_k\beta_l$ of a protein results in a high kinetic impact of the secondary elements $\alpha_i$, $\beta_k$, and $\beta_l$. The relation between secondary elements and contact clusters is summarized in the cluster labels of Fig. \[contactMaps\].
Table 2 shows average experimental $\Phi$-values and kinetic impact for the strands and helices of the 15 proteins considered here. To illustrate the rules (1) and (2), consider for example the src SH3 domain. This protein has two nonlocal clusters, RT-$\beta_4$ and $\beta_1\beta_5$ (see Fig. \[contactMaps\]). The clusters $\beta_2\beta_3$ and $\beta_3\beta_4$ appear on the minimum-ECO routes to both nonlocal clusters (see Table 1) and, hence, have the occurrence number 2. The cluster RT only appears on the route to $\beta_1\beta_5$ and, hence, has occurrence number 1. According to rule (1), the kinetic impact of $\beta_2\beta_3$ and $\beta_3\beta_4$ thus is high (H), and the kinetic impact of RT is medium (M). According to rule (2), the kinetic impact of the cluster RT-$\beta_4$ is medium since it has the cluster ECO 10. Finally, the kinetic impact of $\beta_1\beta_5$ is low (L) since it has a small cluster ECO of 5 and occurrence number 0. Therefore, the kinetic impact of the strands $\beta_2$, $\beta_3$, and $\beta_4$ is high, the kinetic impact of RT is medium, and the kinetic impact of $\beta_1$ and $\beta_5$ is low, in perfect agreement with the average $\Phi$-values (see Table 2).
Rule (3) affects the proteins U1A and L23. In the case of U1A, the cluster $\alpha_1\alpha_2$ does not occur on the minimum-ECO routes to the two other nonlocal clusters $\beta_1\beta_3$ and $\beta_1\beta_4$ and has a significantly smaller loop-closure cost than $\beta_1\beta_4$. Therefore, $\alpha_2$ has a low kinetic impact, since it only appears on the minimum-ECO route to $\alpha_1\alpha_2$. In the case of L23, the nonlocal clusters t-$\alpha_2$ folds in parallel to $\beta_2\beta_4$, with significantly smaller loop-closure cost. As a consequence, the kinetic impact of $\beta_1$ and $\alpha_1$ is low since these secondary elements are only involved in the folding of t-$\alpha_2$.
Results and discussion
======================
To evaluate the model, is is useful to distinguish between proteins with polarized and diffuse $\Phi$-value distributions. Here, this distinction is based on the average $\Phi$-values for the helices and strands. A distribution is polarized if the $\Phi$-values for some of the secondary elements are significantly larger than for other secondary elements. To quantify this notion, a $\Phi$-value distribution here is defined as polarized if at least two average $\Phi$-values are by more than a factor 2.5 smaller than the maximum value of the distribution. A $\Phi$-value distribution is diffuse if this is not the case. In a diffuse distribution, all or all except one of the average $\Phi$-values are larger than 40% of the maximum among these values. An analogous definition can also be applied to the distribution of kinetic impact derived from the minimum-ECO routes. The distribution is diffuse if all or all except one of the secondary elements have high kinetic impact.
According to this definition, 3 among the 15 proteins considered here have a diffuse $\Phi$-value distribution. These proteins are CI2, S6, and FNfn10. In agreement with the experiments, the distribution of kinetic impact for the secondary structural elements of these proteins is also diffuse (see Table 2). The remaining 12 proteins have polarized $\Phi$-value distributions. Fig. \[correlationCoefficients\] shows the correlation coefficient $r$ between average ${\Phi}$-values and kinetic impact for each of these proteins. The calculate the correlation coefficients, the values 0, 1, and 2 are assigned to the kinetic impact L, M, and H.[^2] The correlation coefficient $r$ can attain values in the range -1 to 1 where 1 means ‘perfect’ correlation (proportionality), 0 means no correlation, and negative values mean anticorrelation.
Three of the lowest correlation coefficients are obtained for the $\alpha$-spectrin SH3 domain, protein G, and ACBP (see Fig. \[correlationCoefficients\]). These proteins have clearly negative average $\Phi$-values (smaller than -0.1) in one of the secondary elements. For the comparison with kinetic impact, the negative average $\Phi$-values were simply taken to be zero. However, excluding the helix $\alpha_2$ of ACBP with negative average $\Phi$-value from the correlation analysis leads to a correlation coefficient of 0.94, instead of 0.02. For the $\alpha$-spectrin SH3 domain, excluding the strand $\beta_2$ from the comparison leads to a correlation coefficient of 0.69 instead of 0.37. Thus, the relatively low correlation coefficients for these two proteins can be traced back directly to the secondary elements with negative $\Phi$-values.
Two other proteins with relatively low correlation coefficients in Fig. \[correlationCoefficients\] are Sso7d and CspB. These proteins have in common that the nonlocal clusters fold in parallel on the minimum-ECO routes. In the case of CspB, the nonlocal clusters are $\beta_1\beta_4$ and $\beta_3\beta_5$. Since the total loop-closure cost of the two parallel folding processes leading to these clusters are similar (see Table 1), the model takes them to be equally important for the kinetics. However, the experimental $\Phi$-values seem to indicate that the folding process leading the $\beta_1\beta_4$ has a larger impact on the kinetics than the parallel process leading to $\beta_3\beta_5$. The strands $\beta_1$ to $\beta_3$ of the two clusters $\beta_1\beta_2$ and $\beta_2\beta_3$ which are formed prior to $\beta_1\beta_4$ have relatively large average $\Phi$-values. In contrast, the strands of the cluster $\beta_4\beta_5$, which is formed prior to $\beta_3\beta_5$ on the parallel folding process, have significantly smaller average $\Phi$-values. In the case of Sso7d, the three nonlocal clusters $\alpha$-$\beta_3$, $\beta_1$-$\beta_5$ and $\alpha$-$\beta_1$ fold in parallel, with comparable loop-closure cost. Here, the experimental $\Phi$-values seem to indicate that the folding process leading to $\alpha$-$\beta_3$ dominates the folding kinetics. According to the model, the clusters formed prior to $\alpha$-$\beta_3$ are $\beta_3\beta_4$, $\beta_3\beta_5$, and $\alpha$. The secondary elements of these clusters have medium are large average $\Phi$-values, whereas the $\Phi$-values of the remaining secondary structural elements $\beta_1$, $\beta_2$, and $G_1$ are significantly smaller. In both proteins, specific energetic interactions, which are not taken into account in the model, may be responsible for the dominance of one the parallel folding processes with similar entropic loop-closure barriers.
For the remaining majority of proteins, the model reproduces the polarized $\Phi$-value distributions with relatively large correlation coefficients. This shows that $\Phi$-value distributions averaged over secondary elements are dominated by native-state topology. In the model, the native-state topology is captured by the topology of the native contact maps, or more precisely, by the ECO-dependencies between the contact clusters. Interestingly, the model is able to reproduce the experimentally observed differences in the $\Phi$-value distributions of protein L and G without sequence-specific information. These two proteins have very similar folds, but nonetheless small differences in their contacts maps. Whereas protein L has a small tertiary $\alpha\beta_1$ cluster, protein G has a tertiary $\alpha\beta_2$ cluster. This results in different folding routes and different distributions of kinetic impact (see Tables 1 and 2). In the case of protein L, the N-terminal hairpin $\beta_1\beta_2$ has higher kinetic impact the C-terminal hairpin $\beta_3\beta_4$, in agreement with the average $\Phi$-values. In the case of protein G, the kinetic impact and average $\Phi$-values are larger for the C-terminal hairpin $\beta_3\beta_4$. Other groups have used sequence-specific interaction energies to reproduce these differences between protein L and G [@karanicolas02; @clementi03; @kameda03; @brown04].
The folding routes of the model are hierarchic in the sense that the formation of nonlocal structural elements typically requires the prior formation of other, more local structural elements. It is important to note that the hierarchic folding routes do not contradict cooperative two-state folding with a characteristic single-exponential relaxation dynamics. We have recently developed a free-energy based model with similar loop-closure dependencies [@weikl04]. In this model, two-state folding cooperativity is reproduced when assuming that the local structural elements are unstable. On one hand, the nonlocal structural elements then stabilize the overall fold and, thus, also the local elements. On the other hand, the local structural elements reduce the loop-closure entropies for forming the nonlocal elements. [^3] On the energy landscapes, the formation of local structural elements then corresponds to uphill steps in free energy, and the formation of nonlocal structural elements to steps downhill in free energy, with characteristic barrier or ‘transition’ states in between. For an $\alpha$-helical protein, the hierarchy of local and nonlocal structural elements corresponds to a hierarchy of secondary and tertiary elements [@baldwin99], since the local structural elements are individual helices. However, this correspondence is not general: a $\beta$-hairpin, for example, is a local structural element, but involves both secondary and tertiary structure formation.
Conclusions
===========
The model presented here derives folding routes of proteins and the ‘kinetic impact’ of secondary structural elements from native structures. In a first step, minimum-entropy-loss routes are derived from the native contact maps. This step reveals characteristic loop-closure dependencies between local and nonlocal structural elements. In a second step, the model estimates the kinetic impact of secondary elements from the folding routes. In a systematic comparison for a large set of small and well-characterized proteins, relatively high correlation coefficients are obtained between kinetic impact and average experimental $\Phi$-values of the secondary elements. The model thus indicates that the shape of $\Phi$-value distributions is dominated by native-state topology.
Acknowledgements
================
I would like to thank Ken Dill for enjoyable and highly stimulating interactions and collaborations.
Methods
=======
Contact clusters {#contact-clusters .unnumbered}
----------------
The native contacts are grouped into contact clusters. In general, two contacts $(i,j)$ and $(k,l)$ are taken to be in the same cluster if they are close together on the contact map, according to the distance criterion $|i-k| + |j-l|\le 4$. However, peripheral contacts $(i,j)$ which have a minimum distance $|i-k| + |j-l|= 4$ to the other contacts in the cluster are discarded. For clusters corresponding to helices or $\beta$-strand pairings, also contacts $(i,j)$ which have a distance $|i-k| + |j-l|= 3$ to only one contact $(k,l)$ in the cluster and larger distances to the other contacts are defined as peripheral and discarded. The definition of peripheral contacts is more restrictive for these clusters since they are typically more compact than clusters corresponding to tertiary interactions of helices or sheets. A cluster has to contain at least three contacts. Isolated contacts or contact pairs are not taken into account. Discarding peripheral and isolated contacts helps to avoid an unreasonably large impact of individual contacts on the cluster ECOs, and hence on the model results.
The following PDB files have been used to determine the contact maps and contact clusters: CI2 (1COA); protein L (2PTL, residues 15 to 78); protein G (1PGB), src SH3 domain (1SRL); $\alpha$-spectrin SH3 domain (1SHG); ADA2h (1AYE); U1A (1URN, chain A); S6 (1RIS); TNfn3 (1TEN); FNfn10 (1FNF, residues 1416 to 1509); Titin (1TIT); CspB (1CSP); L23 (1N88); ACBP (2ABD).
Minimum-ECO sequences {#minimum-eco-sequences .unnumbered}
---------------------
As defined in section 2, the minimum-ECO sequences to a given cluster $C_n$ locally minimize the loop-closure cost $s$ in the space of folding sequences. Starting with the set of all possible folding sequences to $C_n$, the minimum-ECO sequences are obtained by applying the following two rules.
1. If two folding sequences $a$ and $b$ have the loop-closure costs $s_a<s_b$ and the set of clusters $\{C_1^{(a)},C_2^{(a)},\ldots C_{m}^{(a)}, C_n\}$ of sequence $a$ is a subset of the clusters $\{C_1^{(b)},C_2^{(b)},\ldots C_{k}^{(b)}, C_n\}$ of sequence $b$, then sequence $b$ is discarded.
2. If two folding sequences $a$ and $b$ have the loop-closure costs $s_a\ge s_b$ and the set of clusters $\{C_1^{(a)},C_2^{(a)},\ldots C_{m}^{(a)}, C_n\}$ of route $a$ is a proper subset of the clusters $\{C_1^{(b)},C_2^{(b)},\ldots C_{k}^{(b)}, C_n\}$ of sequence $b$, then sequence $a$ is discarded.
The rules (1) and (2) are best illustrated in a simple example. Suppose the folding sequence $C_1C_2C_3$ to cluster $C_3$ has the loop-closure cost $s_a$. Suppose now that the cluster $C_0$ is a cluster which does not affect any of the cluster ECOs of $C_1$, $C_2$, or $C_3$. If the cluster $C_0$ is, e.g., a local cluster with small cluster ECO the cost $s_b$ of the sequence $C_0C_1C_2C_3$ is only slightly larger than the cost $s_a$ of the sequence $C_1C_2C_3$. However, since there is no ECO-dependence between $C_0$ and the other three clusters, the sequence $C_0C_1C_2C_3$ is not a reasonable candidate for a minimum-ECO sequence to cluster $C_3$, and hence is discarded by rule (1).
On the other hand, let’s suppose that the sequence $C_1C_3$ has a larger cost $s_b$ than the sequence $C_1C_2C_3$. This means that the prior formation of $C_2$ affects the ECO of $C_3$. Therefore, the sequence $C_1C_3$ is discarded by rule (2) from the possible minimum-ECO sequences to $C_3$.
Secondary structure classification {#secondary-structure-classification .unnumbered}
----------------------------------
The calculation of average experimental $\Phi$-values for helices and strands requires secondary structure classifications. Where possible, the secondary structure definitions given in the PDB files (see above) have been used here. The PDB files of TNfn3 and the $\alpha$-spectrin SH3 domain do not contain secondary structure classifications. For TNfn3 and the structurally analogous protein FNfn10, secondary structure classifications have been taken from Hamill et al[@hamill00] and Cota et al[@cota01]. For the $\alpha$-spectrin SH3 domain, the secondary structure definition of the DSSP algorithm [@kabsch83] has been used. In the case of CspB, the first two substrands given in the PDB file are combined into strand $\beta_1$, the last two substrands into $\beta_5$, and the $3_{10}$ helix is defined from residues 30 to 33. The RT loop of the two SH3 domains is an irregular secondary structure defined here from residues 14 to 26.
[99]{}
Jackson SE. How do small single-domain proteins fold? Fold Des 1998;3:R81-R91.
Fersht AR. Structure and mechanism in protein science. New York: W. H. Freeman; 1999.
Grantcharova V, Alm EJ, Baker D, Horwich, AL. Mechanisms of protein folding. Curr Opin Struct Biol 2001;11:70-82.
Matouschek A, Kellis JT, Serrano L, Fersht, AR. Mapping the transition state and pathway of protein folding by protein engineering. Nature 1989;340:122-126.
Baker D. A surprising simplicity to protein folding. Nature 2000;405:39-42.
Plaxco KW, Simons, KT, Baker, D. Contact order, transition state placement and the refolding rates of single domain proteins. , J Mol Biol 1998;277:985-994.
Plaxco KW, Simons KT, Ruczinski I, Baker D. Topology, stability, sequence, and length: Defining the determinants of two-state protein folding kinetics. Biochemistry 2000;39:11177-11183.
Kuznetsov IB, Rackovsky S. Class-specific correlations between protein folding rate, structure-derived, and sequence-derived descriptors. Proteins 2004;54:333-341.
Dinner AR, Karplus M. The roles of stability and contact order in determining protein folding rates. Nat Struct Biol 2001; 8:21-22.
Ivankov DN, Garbuzynskiy SO, Alm E, Plaxco KW, Baker D, Finkelstein AV. Contact order revisited: Influence of protein size on the folding rate. Protein Sci 2003;12:2057-2062.
Gromiha MM, Selvaraj S. Comparison between long-range interactions and contact order in determining the folding rate of two-state proteins: Application of long-range order to folding rate prediction. J Mol Biol 2001;310:27-32.
Makarov DE, Keller CA, Plaxco KW, Metiu H. How the folding rate constant of simple, single-domain proteins depends on the number of native contacts. Proc Natl Acad Sci USA 2002;99:3535-3539.
Micheletti C, Lattanzi G, Maritan A. Elastic properties of proteins: Insight on the folding process and evolutionary selection of native structures. J Mol Biol 2002;321:909-921.
Micheletti C. Prediction of folding rates and transition-state placement from native-state geometry. Proteins 2003;51:74-84.
Gong HP, Isom DG, Srinivasan R, Rose GD. Local secondary structure content predicts folding rates for simple, two-state proteins. J Mol Biol 2003;327:1149-1154.
Nolting B, Schalike W, Hampel P, Grundig F, Gantert S, Sips N, Bandlow W, Qi PX. Structural determinants of the rate of protein folding. J Theo Biol 2003;223:299-307.
Alm E, Baker D. Prediction of protein-folding mechanisms from free-energy landscapes derived from native structures. Proc Natl Acad Sci USA 1999;96:11305-11310.
Muñoz V, Eaton WA. A simple model for calculating the kinetics of protein folding from three-dimensional structures. Proc Natl Acad Sci USA 1999;96:11311-11316.
Galzitskaya OV, Finkelstein AV. A theoretical search for folding/unfolding nuclei in three-dimensional protein structures. Proc Natl Acad Sci USA 1999;96:11299-11304.
Guerois R, Serrano L. The SH3-fold family: Experimental evidence and prediction of variations in the folding pathways. J Mol Biol 2000;304:967-982.
Ivankov DN, Finkelstein AV. Theoretical study of a landscape of protein folding-unfolding pathways. Folding rates at midtransition. Biochemistry 2001;40:9957-9961.
Alm E, Morozov AV, Kortemme T, Baker D. Simple physical models connect theory and experiment in protein folding kinetics. J Mol Biol 2002;322:463-476.
Bruscolini P, Pelizzola A. Exact solution of the Munoz-Eaton model for protein folding. Phys Rev Lett 2002;88:258101.
Karanicolas J, Brooks CL. The importance of explicit chain representation in protein folding models: An examination of Ising-like models. Proteins 2003;53:740-747.
Clementi C, Nymeyer H, Onuchic JN. Topological and energetic factors: What determines the structural details of the transition state ensemble and “en-route” intermediates for protein folding? An investigation for small globular proteins. J Mol Biol 2000;298:937-953.
Hoang TX, Cieplak M. Sequencing of folding events in Go-type proteins. J Chem Phys 2000;113:8319-8328.
Li L, Shakhnovich EI. Constructing, verifying, and dissecting the folding transition state of chymotrypsin inhibitor 2 with all-atom simulations. Proc Natl Acad Sci USA 2001;98:13014-13018.
Karanicolas J, Brooks CL. The origins of asymmetry in the folding transition states of protein L and protein G. Protein Sci 2002;11:2351-2361.
Ding F, Dokholyan NV, Buldyrev SV, Stanley HE, Shakhnovich EI. Direct molecular dynamics observation of protein folding transition state ensemble. Biophys J 2002;83:3525-3532.
Clementi C, Garcia AE, Onuchic JN. Interplay among tertiary contacts, secondary structure formation and side-chain packing in the protein folding mechanism: All-atom representation study of protein L. J Mol Biol 2003;326:933-954.
Brown S, Head-Gordon T. Intermediates and the folding of proteins L and G. Protein Sci 2004;13:958-970.
Portman JJ, Takada S, Wolynes PG. Microscopic theory of protein folding rates. I. Fine structure of the free energy profile and folding routes from a variational approach. J Chem Phys 2001;114:5069-5081.
Kameda T. Importance of sequence specificity for predicting protein folding pathways: Perturbed Gaussian chain model. Proteins 2003;53:616-628.
Garbuzynskiy SO, Finkelstein AV, Galzitskaya O. Outlining folding nuclei in globular proteins. J Mol Biol 2004;336:509-525.
Zimm BH, Bragg JK. Theory of the phase transition between helix and random coil. J Chem Phys 1959;31:526-535.
Li A, Daggett V. Identification and characterization of the unfolding transition state of chymotrypsin inhibitor 2 by molecular dynamics simulations. J Mol Biol 1996;257:412-429.
Lazaridis T, Karplus M. “New view" of protein folding reconciled with the old through multiple unfolding simulations. Science 1997;278:1928-1931.
Gsponer J, Caflisch A. Role of native topology investigated by multiple unfolding simulations of four SH3 domains. J Mol Biol 2001;309:285-298.
De Jong D, Riley R, Alonso DOV, Daggett V. Probing the energy landscape of protein folding/unfolding transition states. J Mol Biol 2002;319:229-242.
Paci E, Vendruscolo M, Karplus M. Native and non-native interactions along protein folding and unfolding pathways. Proteins 2002;47:379-392.
Morra G, Hodoscek M, Knapp EW. Unfolding of the cold shock protein studied with biased molecular dynamics. Proteins 2003;53:597-606.
Fiebig KM, Dill KA. Protein core assembly processes. J Chem Phys 1993;98:3475-3487.
Dill KA, Fiebig KM, Chan HS. Cooperativity in protein-folding kinetics. Proc Natl Acad Sci USA 1993;90:1942-1946.
Ozkan SB, Bahar I, Dill KA. Transition states and the meaning of $\Phi$-values in protein folding kinetics. Nature Struct Biol 2001;8:765-769.
Li L, Mirny AL, Shakhnovich EI. Kinetics, thermodynamics and evolution of non-native interactions in a protein folding nucleus. Nature Struct Biol 2000;7:336-342.
Merlo C, Dill KA, Weikl TR. $\Phi$-values in protein folding kinetics have energetic and structural components. Submitted
Weikl TR, Dill KA. Folding rates and low-entropy-loss routes of two-state proteins. J Mol Biol 2003;329:585-598.
Weikl TR, Dill KA. Folding kinetics of two-state proteins: Effect of circularization, permutation, and crosslinks J Mol Biol 2003;332:953-963.
Weikl TR, Palassini M, Dill KA. Cooperativity in two-state protein folding kinetics. Protein Sci 2004;13:822-829.
Kim DE, Fisher C, Baker D. A breakdown of symmetry in the folding transition state of protein L. J Mol Biol 2000;298:971-984.
McCallister EL, Alm E, Baker D. Critical role of $\beta$-hairpin formation in protein G folding. Nat Struct Biol 2000;7:669-673.
Karplus M, Weaver DL. Protein-folding dynamics. Nature 1976;260:404-406.
Karplus M, Weaver, DL. Protein-folding dynamics: The diffusion-collision model and experimental data. Protein Science 1994;3:650-668.
Kaya H, Chan HS. Contact order dependent protein folding rates: Kinetic consequences of a cooperative interplay between favorable nonlocal interactions and local conformational preferences. Proteins 2003;52:524-533.
Baldwin RL, Rose GD. Is protein folding hierarchic II. Folding intermediates and transition states. Trends Biochem Sci 1999;24:77-83.
Itzhaki LS, Otzen DE, Fersht AR. The structure of transition state for folding of chymotrypsin inhibitor 2 analysed by protein engineering methods: Evidence for a nucleation-condensation mechanism for protein folding. J Mol Biol 1995;254:260-288.
Riddle DS, Grantcharova VP, Santiago JV, Alm E, Ruczinski I, Baker D. Experiment and theory highlight role of native state topology in SH3 folding. Nature Struct Biol 1999;6:1016-1024.
Martinez JC, Serrano L. The folding transition state between SH3 domains is conformationally restricted and evolutionarily conserved. Nature Struct Biol 1999;6:1010-1016.
Villegas V, Martinez JC, Aviles FX, Serrano L. Structure of the transition state in the folding process of human procarboxypeptidase A2 activation domain. J Mol Biol 1998;283:1027-1036.
Ternström T, Mayor U, Akke M, Oliveberg M. From snapshot to movie: $\Phi$ analysis of protein folding transition states taken one step further. Proc Natl Acad Sci USA 1999;96:14854-14859.
Otzen DE, Oliveberg M. Conformational plasticity in folding of the split $\beta$-$\alpha$-$\beta$ protein S6: Evidence for burst-phase disruption of the native state. J Mol Biol 2002;317:613-627.
Hamill SJ, Steward A, Clarke J. The folding of an immunoglobulin-like greek key protein is defined by a common-core nucleus and regions constrained by topology. J Mol Biol 2000;297:165-178.
Cota E, Steward A, Fowler SB, Clarke J. The folding nucleus of a fibronection type III domain is composed of core residues of the immunoglobolin fold. J Mol Biol 2001;305:1185-1194.
Fowler SB, Clarke J. Mapping the folding pathway of an Immunoglobulin domain: Structural detail from $\Phi$ value analysis and movement of the transition state. Structure 2001;9:355-366.
Garcia-Mira MM, Böhringer D, Schmid FX. The folding transition state of the cold shock protein is strongly polarized. J Mol Biol 2004;339:555-569.
Hedberg L, Oliveberg M. Scattered Hammond plots reveal second level of site-specific information in protein folding: $\Phi'(\beta^\ddagger)$. Proc Natl Acad Sci USA 2004;101:7606-7611.
Kragelund BB, Osmark P, Neergaard TB, Schiodt J, Kristiansen K, Knudsen J, Poulsen FM. The formation of a native-like structure containing eight conserved hydrophobic residues is rate limiting in two-state protein folding of ACBP. Nature Struct Biol 1999;9:594-601.
Kabsch W, Sander C. Dictionary of protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features. Biopolymers 1983;22:2577-2637.
Table 1: Loop-closure events on minimum-ECO routes\
--------------------- --------------------- --------------------------------------------------------------------------------------------- --------------- --------------
nonlocal ECO for non- loop-closure
\[-0.05cm\] protein cluster clusters formed prior local cluster cost
CI2 $\beta_2\beta_3$ — 16 16
$\beta_1\beta_4$ $\alpha_1$, $\beta_3\beta_4$, $\beta_2\beta_3$, 7 27
protein L $\alpha\beta_1$ $\beta_1\beta_2$ 6 9
$\beta_1\beta_4$ $\beta_1\beta_2$, $\alpha$, $\beta_3\beta_4$, $\alpha\beta_1$ 9 22
protein G $\alpha\beta_2$ $\alpha$ 10 13
$\beta_1\beta_4$ $\beta_1\beta_2$, $\alpha$, $\beta_3\beta_4$ 9 18
or: $\alpha$, $\beta_3\beta_4$, $\alpha\beta_2$ 3 19
src SH3 RT-$\beta_4$ $\beta_2\beta_3$, $\beta_3\beta_4$ 10 17
$\beta_1\beta_5$ RT, $\beta_2\beta_3$, $\beta_3\beta_4$ 5 17
$\alpha$-spSH3 RT-$\beta_4$ $\beta_2\beta_3$ 7 10
$\beta_1\beta_5$ RT, $\beta_2\beta_3$, $\beta_3\beta_4$, G 5 17
or: RT, $\beta_2\beta_3$, G, RT-$\beta_4$ 3 17
Sso7d $\alpha$-$\beta_3$ $\beta_3\beta_4$, $\beta_4\beta_5$, $\alpha$ 7 16
$\beta_1$-$\beta_5$ $\beta_1\beta_2$, $\beta_3\beta_4$, $\beta_4\beta_5$ 9 18
$\alpha$-$\beta_1$ $\beta_1\beta_2$, $\beta_3\beta_4$, $\beta_4\beta_5$ 12 21
ADA2h $\beta_1\beta_3$ G 8 11
$\beta_1\beta_4$ $\alpha_1$, G, $\beta_2\beta_3$, $\alpha_2$ 9 21
or: G, $\alpha_2$, $\beta_1\beta_3$ 7 21
U1A $\alpha_1\alpha_2$ $\beta_2\beta_3$, $\alpha_2$ 3 9
$\beta_1\beta_3$ $\alpha_1\beta_2$, $\beta_2\beta_3$ 6 12
$\beta_1\beta_4$ $\alpha_1\beta_2$, $\beta_2\beta_3$, $T$, $\beta_1\beta_3$ 2 17
S6 $\beta_1\beta_3$ $\alpha_1$, $\beta_2\beta_3$ 14 20
$\beta_1\beta_4$ $\alpha_1$, $\beta_2\beta_3$, $\alpha_2$ 14 23
TNfn3 $\beta_2\beta_5$ $\beta_3\beta_4$ 9 15
$\beta_1$-$\beta_7$ $\beta_3\beta_4$, $\beta_6\beta_7$, $\beta_2\beta_5$ 2 20
$\beta_3\beta_6$ $\beta_1\beta_2$, $\beta_3\beta_4$, $\beta_6\beta_7$, $\beta_2\beta_5$, $\beta_1$-$\beta_7$ 4 27
\[0.1cm\] FNfn10 $\beta_2\beta_5$ $\beta_3\beta_4$ 9 17
$\beta_3\beta_6$ $T$ 22 25
$\beta_1\beta_6$ $\beta_1\beta_2$, $T$, $\beta_3\beta_6$ 2 30
or: $\beta_3\beta_4$, $\beta_6\beta_7$, $\beta_2\beta_5$, $\beta_1\beta_7$ 2 31
$\beta_1\beta_7$ $\beta_3\beta_4$, $\beta_6\beta_7$, $\beta_2\beta_5$ 9 29
or: $\beta_1\beta_2$, $T$, $\beta_6\beta_7$, $\beta_3\beta_6$, $\beta_1\beta_6$ 2 34
Titin $\beta_2\beta_5$ $T_1$, $T_2$ 14 20
$\beta_3\beta_6$ $\beta_4\beta_5$, $T_3$ 14 20
$\beta_1\beta_7$ $T_1$, $T_2$, $\beta_6\beta_7$, $\beta_2\beta_5$ 2 25
or: $\beta_1\beta_2$, $\beta_4\beta_5$, $T_3$, $\beta_3\beta_6$ 8 31
CspB $\beta_1\beta_4$ $\beta_1\beta_2$, $\beta_2\beta_3$, $T$ 4 19
$\beta_3\beta_5$ $\beta_4\beta_5$ 14 17
L23 $\beta_1\beta_2$ $\alpha_1$ 3 6
$\beta_3\beta_4$ – 12 12
$t$-$\alpha_2$ $\alpha_1$, $\alpha_2$, $\beta_1\beta_2$ 5 14
$\beta_2\beta_4$ $\alpha_2$, $\beta_3\beta_4$ 6 21
ACBP $t$-$\alpha_3$ $\alpha_3$ 5 8
$\alpha_2\alpha_3$ $\alpha_2$, $\alpha_3$, $t$-$\alpha_3$ 3 14
$t$-$\alpha_4$ $\alpha_2$, $\alpha_3$, $\alpha_4$, $t$-$\alpha_3$, $\alpha_2\alpha_3$ 6 23
$\alpha_1\alpha_4$ $\alpha_1$, $\alpha_2$, $\alpha_3$, $t$-$\alpha_3$, $\alpha_2\alpha_3$ 9 26
or: $\alpha_2$, $\alpha_3$, $\alpha_4$, $t$-$\alpha_3$, $\alpha_2\alpha_3$, $t$-$\alpha_4$ 5 28
--------------------- --------------------- --------------------------------------------------------------------------------------------- --------------- --------------
Table 2: Average $\Phi$-values and kinetic impact of secondary structural elements\
----------------------------------------- ------------- ------------- ------------- ------------- ------------ ------------- -------------
CI2 $\beta_1$ $\alpha$ $\beta_2$ $\beta_3$ $\beta_4$
$\bar{\Phi}_{\text{exp}}$ 0.23 (1) 0.32 (12) 0.15 (4) 0.32 (4) 0.03 (2)
kinimpact M H H H H
\[0.3cm\] protein L $\beta_1$ $\beta_2$ $\alpha$ $\beta_3$ $\beta_4$
$\bar{\Phi}_{\text{exp}}$ 0.36 (9) 0.46 (7) 0.15 (16) 0.18 (4) 0.14 (7)
kinimpact H H M M M
\[0.3cm\] protein G $\beta_1$ $\beta_2$ $\alpha$ $\beta_3$ $\beta_4$
$\bar{\Phi}_{\text{exp}}$ 0.36 (3) $-0.16$ (4) 0.13 (9) 0.63 (2) 0.27 (4)
kinimpact M M H H H
\[0.3cm\] src SH3 $\beta_1$ RT $\beta_2$ $\beta_3$ $\beta_4$ G $\beta_5$
$\bar{\Phi}_{\text{exp}}$ 0.02 (4) 0.10 (8) 0.46 (3) 0.53 (6) 0.43 (6) – $-0.04$ (2)
kinimpact L M H H H L L
\[0.3cm\] $\alpha$-spec SH3 $\beta_1$ RT $\beta_2$ $\beta_3$ $\beta_4$ G $\beta_5$
$\bar{\Phi}_{\text{exp}}$ 0.08 (2) 0.26 (3) $-0.20$ (1) 0.66 (3) 0.60 (2) 0.53 (1) 0.16 (1)
kinimpact L H H H M H L
\[0.3cm\] Sso7d $\beta_1$ $\beta_2$ $G_1$ $\beta_3$ $\beta_4$ $\beta_5$ $\alpha$
$\bar{\Phi}_{\text{exp}}$ $-0.03$ (2) 0.11 (2) $-0.03$ (2) 0.96 (2) 0.27 (4) 0.19 (5) $0.41$ (4)
kinimpact H H L H H H H
\[0.3cm\] ADA2h $\beta_1$ $\alpha_1$ G $\beta_2$ $\beta_3$ $\alpha_2$ $\beta_4$
$\bar{\Phi}_{\text{exp}}$ $0.42$ (3) 0.26 (3) – 0.06 (2) 0.29 (3) 0.49 (4) $0.14$ (2)
kinimpact M M H M M H M
\[0.3cm\] U1A $\beta_1$ $\alpha_1$ $\beta_2$ $\beta_3$ $\alpha_2$ $\beta_4$
$\bar{\Phi}_{\text{exp}}$ ($\beta=0.5$) $0.23$ (2) 0.38 (3) 0.73 (3) – 0.00 (1) $0.00$ (1)
$\bar{\Phi}_{\text{exp}}$ ($\beta=0.7$) $0.43$ (2) 0.63 (3) 0.98 (3) – 0.50 (1) $0.23 $ (1)
kinimpact M H H H L L
\[0.3cm\] S6 $\beta_1$ $\alpha_1$ $\beta_2$ $\beta_3$ $\alpha_2$ $\beta_4$
$\bar{\Phi}_{\text{exp}}$ 0.34 (4) 0.25 (4) 0.24 (1) 0.31 (5) 0.28 (2) 0.14 (2)
kinimpact H H H H M H
\[0.3cm\] TNfn3 $\beta_1$ $\beta_2$ $\beta_3$ $\beta_4$ $\beta_5$ $\beta_6$ $\beta_7$
$\bar{\Phi}_{\text{exp}}$ 0.12 (3) 0.27 (2) 0.36 (3) 0.55 (2) 0.47 (2) 0.42 (3) 0.11 (5)
kinimpact M H H H H H H
\[0.3cm\] FNfn10 $\beta_1$ $\beta_2$ $\beta_3$ $\beta_4$ $\beta_5$ $\beta_6$ $\beta_7$
$\bar{\Phi}_{\text{exp}}$ 0.3 (3) $-0.16$ (2) 0.55 (3) 0.35 (2) 0.29 (2) 0.44 (4) 0.73 (1)
kinimpact H H H H H H H
\[0.3cm\] Titin $\beta_1$ $\beta_2$ $\beta_3$ $\beta_4$ $\beta_5$ $\beta_6$ $\beta_7$
$\bar{\Phi}_{\text{exp}}$ 0.09 (3) 0.53 (4) 0.51 (2) 0.54 (2) 0.66 (3) 0.66 (3) 0.07 (3)
kinimpact M H H H H H M
\[0.3cm\] CspB $\beta_1$ $\beta_2$ $\beta_3$ G $\beta_4$ $\beta_5$
$\bar{\Phi}_{\text{exp}}$ 0.64 (6) 0.27 (4) 0.75 (1) $-0.06$ (2) 0.16 (2) 0.12 (2)
kinimpact H H H L H H
\[0.3cm\] L23 $\beta_1$ $\alpha_1$ $\beta_2$ $\alpha_2$ $\beta_3$ $\beta_4$ $\alpha_3$
$\bar{\Phi}_{\text{exp}}$ 0.08 (1) 0.03 (1) 0.20 (2) 0.34 (2) 0.10 (3) 0.29 (3) 0.02 (1)
kinimpact L L M H H H L
\[0.3cm\] ACBP $\alpha_1$ $\alpha_2$ $\alpha_3$ $\alpha_4$
$\bar{\Phi}_{\text{exp}}$ 0.34 (4) -0.19 (9) 0.57 (2) 0.21 (6)
kinimpact M H H M
----------------------------------------- ------------- ------------- ------------- ------------- ------------ ------------- -------------
Caption Table 2:\
The average $\Phi$-values have been calculated from data published in the following articles: CI2[@itzhaki95], protein L [@kim00], protein G [@mccallister00], src SH3 [@riddle99], $\alpha$-spectrin SH3 [@martinez99], Sso7d [@guerois00], ADA2h [@villegas98], U1A [@ternstrom99], S6 [@otzen02], TNfn3 [@hamill00], FNfn10 [@cota01], Titin [@fowler01], CspB [@garcia04], L23 [@hedberg04], ACBP [@kragelund99]. The number in brackets behind an average $\Phi$-value indicates the number of residues in the secondary element for which $\Phi$-values have been measured. Averages taken from many $\Phi$-values are more reliable. The kinetic impact of the secondary elements is derived from the results shown in Table 1 and can attain the values low (L), medium (M), or high (H).
[^1]: More precisely, the cluster ECO is an estimate for the length of the shortest loop that has to be closed to ‘initiate’ the cluster, i.e. to form the cluster contact(s) with minimum ECO. After ‘initiation’, the cluster is thought to be ‘zipped up’ in a series of small-loop-closure steps [@weikl03a; @weikl03b]. These zipping steps do not depend on the folding sequence. Therefore, they are not considered here.
[^2]: Any other ‘equidistant’ values $a$, $a+b$, and $a+2 b$ with $b>0$ for the kinetic impact L, M, and H result in the same correlation coefficients. Correlation coefficients are only given for proteins with polarized distributions since the do not reflect the quality of the modeling in the case of diffuse distributions with rather similar average $\Phi$-values for the secondary elements.
[^3]: Similar in spirit, the diffusion-collision model of Karplus and Weaver assumes that individual microdomains such as helices are unstable [@karplus76; @karplus94]. A direct, energetic local-nonlocal coupling has been recently used by Kaya and Chan [@kaya03] to obtain two-state cooperitivity in a simple lattice model.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Steven Dale Cutkosky
title: Generically finite morphisms
---
\[section\] \[Theorem\][Lemma]{} \[Theorem\][Corollary]{} \[Theorem\][Proposition]{} \[Theorem\][Remark]{} \[Theorem\][Remarks]{} \[Theorem\][Example]{} \[Theorem\][Examples]{} \[Theorem\][Definition]{} \[Theorem\] \[Theorem\][Conjecture]{} =21.5cm
[^1]
introduction
============
Suppose that $k$ is a field, and $f:Y\rightarrow X$ is a dominant, generically finite morphism of complete $k$-varieties. If $Y$ and $X$ are complete curves, then it is classical that $f$ is finite. If $Y$ and $X$ have dimensions $\ge 2$ $f$ need not be finite. The simplest example is the blowup of a nonsingular subvariety of a nonsingular projective variety.
It is however natural to ask the following question. Given a generically finite morphism $f:Y\rightarrow X$ as above, does there exist a commutative diagram $$\label{eqI1}
\begin{array}{lll}
Y_1&\stackrel{f_1}{\rightarrow}&X_1\\
\downarrow&& \downarrow\\
Y&\stackrel{f}{\rightarrow}&X
\end{array}$$ such that $f_1$ is finite, $Y_1$ and $X_1$ are nonsingular complete $k$-varieties, and the vertical arrows are birational? The answer to this question is no, as is shown by a theorem of Abhyankar (Theorem 11 [@Ab3]). This theorem, (as shown in Example \[Example2\] of this paper) proves that such a diagram cannot always be constructed even when $f:Y\rightarrow X$ is a $G$-equivariant morphism of complex projective surfaces, where the extension of function fields $k(X)\rightarrow k(Y)$ is Galois with Galois group $G$.
In the theory of resolution of singularities a modified version of this question is important. .2truein [**Question 1.**]{} With $f:Y\rightarrow X$ as above, is it possible to construct a diagram (\[eqI1\]) such that $f_1$ is finite, $Y_1$ and $X_1$ are complete $k$-varieties such that $Y_1$ is nonsingular, $X_1$ is normal and the vertical arrows are birational? .2truein This question has been posed by Abhyankar (with the further conditions that $Y_1\rightarrow Y$ is a sequence of blowups of nonsingular subvarieties and $Y$, $X$ are projective) explicitely on page 144 of [@Ab6], where it is called the “weak simultaneous resolution global conjecture” and implicitely in the paper [@Ab3].
As positive evidence for this conjecture, Abhyankar proves a local form of this conjecture for 2 dimensional function fields over an algebraically closed field of arbitrary characteristic [@Ab1], [@Ab4], (this is the two dimensional case of the “weak simultaneous resolution local conjecture” [@Ab6]).
An important case where Question 1 has a positive answer is for generically finite morphisms $f:Y\rightarrow X$ of projective varieties, over a field $k$ of characteristic zero, which induce a Galois extension of function fields. We give a simple proof in Theorem \[Theorem7\]. We can construct (with this Galois assumption) a diagram (\[eqI1\]) such that the conclusions of Question 1 hold, and $X_1$ has normal toric singularities. This is a relative version of Theorem 7, [@Ab3].
In Theorem \[Example1\] of this paper we give a counterexample to Question 1. The example is of a generically finite morphism $Y\rightarrow X$ of nonsingular, projective surfaces, defined over an algebraically closed field $k$ of characteristic not equal to 2. This counterexample is necessarilly then a counterexample to the “weak simultaneous resolution global conjecture”.
As the “weak simultaneous resolution local conjecture”, posed by Abhyankar on page 144 of [@Ab6] is true in characteristic 0 (we prove it in [@C3] as a corollary to the local monomialization theorem, Theorem 1.1 [@C2], and prove a stronger statement in Theorem \[Theorem1\] of this paper), Theorem \[Example1\] also gives a counterexample to the philosophy that a theorem which is true in valuation theory should also be true in the birational geometry of projective varieties. This is the philosophy which led to successful proofs of resolution of singularities for surfaces and 3-folds, in characteristic zero by Zariski ([@Z1], [@Z5]), and in positive characteristic by Abhyankar ([@Ab1],[@Ab4],[@Ab5]). Recently there has been progress on the important problem of local uniformization in positive characteristic in higher dimensions (c.f. [@HRW], [@Ku],[@Moh], [@T]). Ramification of morphisms of algebraic surfaces in positive characteristic is analyzed in [@CP] and [@CP2].
We prove in Theorem \[Theorem8\] that Question 1, and the “weak simultaneous resolution global conjecture” are almost true, as it always possible (over fields of characteristic zero) to construct a diagram (\[eqI1\]) where $f_1:Y_1\rightarrow X_1$ is a quasi-finite morphism of integral, finite type $k$-schemes, $Y_1$ is nonsingular, $X_1$ has normal toric singularities, the vertical morphisms are birational and every $k$-valuation of $k(X)$ has a center on $X_1$, every $k$-valuation of $k(Y)$ has a center on $Y_1$. That is, the answer to Question 1 becomes true if we weaken the condition that the vertical arrows are proper by not insisting that these morphisms be separated.
The essential technical result used in the proof of Theorem \[Theorem8\] is the Local Monomialization Theorem, Theorem 1.1 of [@C2]. Local monomialization is used to prove a strengthened version of the “weak simultaneous local conjecture”, Theorem \[Theorem1\] of this paper, which allows us to construct local solutions of the problem, which are patched in an arbitrary manner (this is where separatedness is lost) to construct $X_1$ and $Y_1$.
We will now give an overview of the proof of the construction of Theorem \[Example1\], which is the counterexample to Question 1 and the “weak simultaneous global conjecture”. We will use some of the notation explained in the following section on notations.
Suppose that $K^*$ is a finite extension of an algebraic function field $K$, defined over an algebraically closed field $k$ of characteristic zero, $\nu^*$ is a $k$-valuation of $K^*$ and $\nu$ is its restriction to $K$. By Theorem \[Theorem1\], there exists an algebraic regular local ring $S$ with quotient field $K^*$ dominated by $\nu^*$ and an algebraic normal local ring $R$ with quotient field $K$ such that $S$ lies over $R$ ($S$ is the localization at a maximal ideal of the integral closure of $R$ in $K^*$). This can be refined [@CP2] to show that there exist $S$ and $R$ as above such that the quotient of value groups $\Gamma_{\nu^*}/\Gamma_{\nu}$ acts faithfully on the power series ring $\hat S$ by $k$-algebra automorphisms so that $$\label{eqI2}
\hat R\otimes_{R/m_R}k'\cong (\hat S\otimes_{S/m_S}k')^{\Gamma_{\nu^*}/\Gamma_{\nu}}$$ where $k'$ is an algebraic closure of $S/m_S$. In some special cases, such as rational rank 2 valuations of algebraic function fields of dimension two, where $[K:K^*]$ is not divisible by $\text{char}(k)$, this construction is stable under quadratic transforms of $S$. We give a direct proof in this paper.
Let $k$ be an algebraically closed field, $\overline \nu$ be the rational rank 2 valuation on $\overline K=k(u,v)$ and $L_1$ be the $q$-cyclic extension of $\overline K$, where $q$ is a prime distinct from $\text{char}(k)$, constructed in Theorem 11 [@Ab3] of Abhyankar (the construction is recalled in Theorem \[Theorem10\] of this paper). $\overline R=k[u,v]_{(u,v)}$ is dominated by $\overline \nu$. The extension $\overline K\rightarrow L_1$ has the property that if $S_1$ is an algebraic regular local ring with quotient field $L_1$ which lies over an algebraic normal local ring $R_1$ with quotient field $\overline K$ such that $R_1$ is dominated by $\overline\nu$, and contains $\overline R$, then $R_1$ is singular.
Examining this example, we see that there is a unique extension $\nu_1$ of $\overline \nu$ to $L_1$, and the quotient of value groups $\Gamma_{\nu_1}/\Gamma_{\overline\nu}\cong{\bold Z}_q$. By (\[eqI2\]) we have $\hat R_1\cong\hat{S_1}^{{\bold Z}_q}$. Since $\hat R_1$ is singular and $\hat S_1$ is a power series ring in two variables over $k$, the algebraic fundamental group of $\hat R_1$ is $$\pi_1(\text{spec}(\hat R_1)-m_{\hat R_1})\cong{\bold Z}_q.$$
We now consider the extension $\nu_2$ of $\overline\nu$ to a particular $p$-cyclic extension $L_2$ of $\overline K$ where $p$ is a prime such that $p\ne q$ and $p\ne \text{char}(k)$. We have $\Gamma_{\nu_2}/\Gamma_{\overline\nu}\cong {\bold Z}_p$, and if $S_2$ is an algebraic regular local ring with quotient field $L_2$ which is dominated by $\nu_2$ which contains $\overline R$, and if there exists an algebraic normal local ring $R_2$ with quotient field $\overline K$ which lies below $S_2$, then $\hat R_2\cong\hat{S_2}^{{\bold Z}_p}$ by (\[eqI2\]). If $R_2$ is singular, the algebraic fundamental group of $\hat R_2$ is then $$\pi_1(\text{spec}(\hat R_2)-m_{\hat R_2})\cong{\bold Z}_p.$$
We then construct a morphism of projective nonsingular $k$-surfaces $\Phi:Y\rightarrow X$ such that $X$ has the function field $\overline K$ and $\overline R$ is the local ring of a point on $X$. $Y$ is constructed in such a way that $Y$ splits into two sheets over $\text{spec}(\overline R)$, and there are points on these two sheets which are formally the same as extensions $\overline R\rightarrow S_1$, $\overline R\rightarrow S_2$ into algebraic regular local rings with respective quotient fields $L_1$ and $L_2$ which are dominated by the respective valuations $\nu_1$ and $\nu_2$. We then use these formal embeddings to construct extensions $\overline\nu_1$ and $\overline \nu_2$ of $\nu$ to the function field $L_0$ of $Y$.
These extensions $\overline\nu_i$ have the property that if $Y_1$ is nonsingular and $Y_1\rightarrow Y$ is proper birational (so that it can be factored by blowups of points) then the map $Y_1\rightarrow X$ is formally isomorphic at the centers of the valuations $\overline \nu_1$ and $\overline \nu_2$ to the extensions of $\overline R$ by the corresponding sequences of quadratic transforms of the local rings $S_1$ and $S_2$ along the respective valuations $\nu_1$ and $\nu_2$.
Now suppose that we can construct a diagram $$\begin{array}{lll}
Y_1&\rightarrow&X_1\\
\downarrow&&\downarrow\\
Y&\rightarrow&X
\end{array}$$ such that $Y_1\rightarrow X_1$ is finite, $Y_1$ is nonsingular, $X_1$ is normal, and the vertical arrows are proper and birational. Let $R_1$ be the local ring of the center of $\overline \nu$ on $X_1$, $S(1)$ be the local ring of the center of $\overline\nu_1$ on $Y_1$, and let $S(2)$ be the local ring of the center of $\overline \nu_2$ on $Y_1$. Since $\overline \nu_1$ and $\overline \nu_2$ both extend $\overline \nu$, and $\Gamma_{\overline\nu_1}/\Gamma_{\overline\nu}\cong{\bold Z_q}$, $\Gamma_{\overline\nu_2}/\Gamma_{\overline\nu}\cong{\bold Z_p}$, we must have that $$\widehat{S(1)}^{{\bold Z}_q}\cong\widehat R_1\cong \widehat{S(2)}^{{\bold Z}_p}.$$ We then have that $R_1$ is singular, by our contruction of $\overline\nu_1$, and thus the algebraic fundamental group $\pi_1(\text{spec}(\hat{R}_1)-m_{\hat R_1})$ has simultaneously order $p$ and order $q\ne p$ which is impossible.
notations
=========
We will denote the maximal ideal of a local ring $R$ by $m_R$. We will denote the quotient field of a domain $R$ by $QF(R)$. Suppose that $R\subset S$ is an inclusion of local rings. We will say that $R$ dominates $S$ if $m_S\cap R=m_R$. Suppose that $K$ is an algebraic function field over a field $k$. We will say that a subring $R$ of $K$ is algebraic if $R$ is essentially of finite type over $k$. Suppose that $K^*$ is a finite extension of an algebraic function field $K$, $R$ is a local ring with $QF(K)$ and $S$ is a local ring with $QF(K^*)$. We will say that $S$ lies over $R$ and $R$ lies below $S$ if $S$ is a localization at a maximal ideal of the integral closure of $R$ in $K^*$. If $R$ is a local ring, $\hat R$ will denote the completion of $R$ at its maximal ideal.
Good introductions to the valuation theory which we require in this paper can be found in Chapter VI of [@ZS] and in [@Ab4]. A valuation $\nu$ of $K$ will be called a $k$-valuation if $\nu(k)=0$. We will denote by $V_{\nu}$ the associated valuation ring, which necessarily contains $k$. A valuation ring $V$ of $K$ will be called a $k$-valuation ring if $k\subset V$. The value group of a valuation $\nu$ will be denoted by $\Gamma_{\nu}$. If $X$ is an integral $k$-scheme with function field $K$, then a point $p\in X$ is called a center of the valuation $\nu$ (or the valuation ring $V_{\nu}$) if $V_{\nu}$ dominates ${\cal O}_{X,p}$. If $R$ is a subring of $V_{\nu}$ then the center of $\nu$ (the center of $V_{\nu}$) on $R$ is the prime ideal $R\cap m_{V_{\nu}}$.
Suppose that $R$ is a local domain. A monoidal transform $R\rightarrow R_1$ is a birational extension of local domains such that $R_1=R[\frac{P}{x}]_m$ where $P$ is a regular prime ideal of $R$, $0\ne x\in P$ and $m$ is a prime ideal of $R[\frac{P}{x}]$ such that $m\cap R=m_R$. $R\rightarrow R_1$ is called a quadratic transform if $P=m_R$.
If $R$ is regular, and $R\rightarrow R_1$ is a monodial transform, then there exists a regular sustem of parameters $(x_1,\ldots, x_n)$ in $R$ and $r\le n$ such that $$R_1=R\left[\frac{x_2}{x_1},\ldots,\frac{x_r}{x_1}\right]_m.$$
Suppose that $\nu$ is a valuation of the quotient field $R$ with valuation ring $V_{\nu}$ which dominates $R$. Then $R\rightarrow R_1$ is a monoidal transform along $\nu$ (along $V_{\nu}$) if $\nu$ dominates $R_1$.
We follow the notation of [@Ha]. In particular, we do not require that a scheme be separated.
A counterexample to global weak simultaneous resolution
=======================================================
In this section we construct the following example. This gives a counterexample to Question 1 stated in the introduction, as well as to the “weak simultaneous resolution global conjecture” stated by Abhyankar explicitely on page 144 [@Ab6] and implicitely in the paper [@Ab3]. As the “weak simultaneous resolution local conjecture” is true in characteristic 0 (We prove it in [@C3], and prove a stronger version in Theorem \[Theorem1\] of this paper), Theorem \[Example1\] also gives a counterexample to the philosophy that a theorem which is true in valuation theory should also be true in the birational geometry of projective varieties.
\[Example1\] Suppose that $k$ is an algebraically closed field of characteristic 0 or of odd prime characteristic. Then there exists a generically finite morphism $\Phi:Y\rightarrow X$ of projective nonsingular $k$-surfaces such that there does not exist a commutative diagram $$\begin{array}{lll}
Y_1&\stackrel{\Phi_1}{\rightarrow}&X_1\\
\downarrow&&\downarrow\\
Y&\stackrel{\Phi}{\rightarrow}&X
\end{array}$$ where the vertical arrows are birational and proper, $Y_1$ is nonsingular, $X_1$ is normal, and $\Phi_1$ is finite.
Throughout this section we will suppose that $k$ is an algebraically closed field.
\[Lemma3\] Suppose that $L$ is a 2 dimensional algebraic function field over $k$. Suppose that $R$ is an algebraic regular local ring with quotient field $L$ and maximal ideal $m_{R}=(u,v)$. Suppose that $\overline\nu$ is a rank 1, rational rank 2 valuation of $L$ such that $\overline \nu$ dominates $R$, $\overline\nu(u),\overline\nu(v)>0$ and $\overline \nu(u),\overline \nu(v)$ are rationally independent. Then
1. The value group of $\overline\nu$ is $\Gamma_{\overline \nu}={\bold Z}\overline\nu(u)+{\bold Z}\overline\nu(v)$.
2. Suppose that $R\rightarrow R_1$ is a sequence of quadratic transforms along $\overline \nu$. Then there exist regular parameters $(u_1,v_1)$ in $R_1$ and $a,b,c,d\in{\bold N}$ such that $$\begin{array}{ll}
u&=u_1^av_1^b\\
v&=u_1^cv_1^d
\end{array}$$ with $ad-bc=\pm 1$.
3. There exists a unique extension $\hat\nu$ of $\overline \nu$ to $\hat L=QF(\hat R)$ which dominates $\hat{R}$. The value group of $\hat \nu$ is $\Gamma_{\hat\nu}=\Gamma_{\nu}$.
4. If $\nu_1$ is a valuation such that $\nu_1$ is equivalent to $\overline \nu$, (and the value groups $\Gamma_{\overline\nu}$ and $\Gamma_{\nu_1}$ are embedded as subgroups of ${\bold R}$) then $$\frac{\nu_1(v)}{\nu_1(u)}=\frac{\overline \nu(v)}{\overline\nu(u)}.$$
Proof of 1. $f\in R$ implies there is an expression $f=\sum_{i+j=r}^{n-1}a_{ij}u^iv^j+h$ with $a_{ij}\in k$, $r=\text{ord}(f)$, $h\in (m_{R})^n$, where $n$ is such that $$n\overline\nu(m_{R})>\nu(\sum_{i+j=r}a_{ij}u^iv^j).$$ Thus since $\overline \nu(u)$ and $\overline \nu(v)$ are rationally independent, $$\overline \nu(f)=\overline\nu(\sum_{i+j=r}^{n-1}a_{ij}u^iv^j)=\text{min}\{\overline \nu(u^iv^j)\mid r\le i+j\le n-1,\,\,\,a_{ij}\ne 0\}.$$ Thus $\Gamma_{\overline\nu}={\bold Z}\overline\nu(u)+{\bold Z}\overline\nu(v)$.
Proof of 2. It suffices to prove this for a single quadratic transform. We either have that $\overline \nu(u)>\overline \nu(v)$ or $\overline \nu(v)>\overline \nu(u)$. In the first case have that $$R_1=R[\frac{u}{v},v]_{(\frac{u}{v},v)}$$ and $\overline \nu(\frac{u}{v})$, $\overline\nu(v)$ are linearly independent over ${\bold Q}$. In the second case we have that $$R_1=R[u,\frac{v}{u}]_{(u,\frac{v}{u})}$$ and $\overline\nu(u)$, $\overline\nu(\frac{v}{u})$ are linearly independent over $\bold Q$.
Proof of 3. Define an extension $\hat\nu$ of $\overline\nu$ to $\hat L$ by $$\hat\nu(f)=\text{min}\{i\overline\nu(u)+j\overline\nu(v)\mid a_{ij}\ne0\}$$ if $f\in\hat{R}$, and $f$ has the expression $f=\sum a_{ij}x^iy^j$ with $a_{ij}\in k$. $\hat\nu$ is a valuation since for $i,j,\alpha,\beta\in {\bold N}$, $$i\overline\nu(u)+j\overline\nu(v)=\alpha\overline\nu(u)+\beta\overline\nu(v)$$ implies $i=\alpha, j=\beta$. $\hat \nu$ dominates $\hat{R}$ and $\Gamma_{\hat\nu}=\Gamma_{\nu}$.
Suppose that $\tilde\nu$ is an extension of $\overline\nu$ to $\hat L$ which dominates $\hat R$. Suppose that $f\in\hat R$. Write $$f=\sum_{i+j=r}^{\infty}a_{ij}u^iv^j,$$ where $r=\text{ord}(f)$, $a_{ij}\in k$. There exists $n$ such that $$\overline \nu(\sum_{i+j=r}a_{ij}u^iv^j)<n\overline\nu(m_R).$$ Write $$f=\sum_{i+j=r}^{n-1}a_{ij}u^iv^j+g,$$ with $g\in m_R^n\hat R$. $$\tilde \nu(\sum_{i+j=r}^{n-1}a_{ij}u^iv^j)=\overline\nu(\sum_{i+j=r}^{n-1}a_{ij}u^iv^j)
=\text{min}\{\overline\nu(u^iv^j)\mid a_{ij}\ne 0, r\le i+j\le n-1\}
<n\overline\nu(m_R)$$ and $\tilde \nu(g)\ge n\overline\nu(m_R)$ so that $$\tilde\nu(f)=\text{min}\{\overline\nu(u^iv^j)\mid a_{ij}\ne 0\}=\hat\nu(f).$$
Proof of 4. As on page 653 [@Z1], we consider the convergent factions $\frac{f_p}{g_p}$ of $\tau=\frac{\overline\nu(v)}{\overline\nu(u)}$. Set $$\epsilon=f_{p-1}g_p-f_pg_{p-1}=\pm1.$$ $\epsilon,-\tau+\frac{f_{p-1}}{g_{p-1}},\tau-\frac{f_p}{g_p}$ have the same signs.
For arbitrary $p$, we can define $u_1,v_1\in L$ by $$u=u_1^{g_p}v_1^{g_{p-1}},
v=u_1^{f_p}v_1^{f_{p-1}}.$$ $$u_1^{\epsilon}=\frac{u^{f_{p-1}}}{v^{g_{p-1}}},
v_1^{\epsilon}=\frac{v^{g_p}}{u^{f_p}}.$$ $$\begin{array}{ll}
\epsilon \overline \nu(u_1)&=f_{p-1}\overline\nu(u)-g_{p-1}\overline\nu(v)\\
&=g_{p-1}\overline\nu(u)\left[\frac{f_{p-1}}{g_{p-1}}-\tau\right]
\end{array}$$ which implies that $\overline\nu(u_1)>0$. $$\begin{array}{ll}
\epsilon\overline\nu(v_1)&=g_p\overline\nu(v)-f_p\overline\nu(u)\\
&=\overline\nu(u)g_p\left[ \tau-\frac{f_p}{g_p}\right]
\end{array}$$ which implies $\overline \nu(v_1)>0$. Thus $\nu_1(u_1), \nu_1(v_1)>0$. $$f_{p-1}\nu_1(u)-g_{p-1}\nu_1(v)=\epsilon\nu_1(u_1),\,\,\,
-f_p\nu_1(u)+g_p\nu_1(v)=\epsilon\nu_1(v_1)$$ imply $$\frac{f_{p-1}}{g_{p-1}}>\frac{\nu_1(v)}{\nu_1(u)}>\frac{f_p}{g_p}$$ if $\epsilon=1$, $$\frac{f_{p-1}}{g_{p-1}}<\frac{\nu_1(v)}{\nu_1(u)}<\frac{f_p}{g_p}$$ if $\epsilon=-1$. Since this holds for all $p$, $$\frac{\nu_1(v)}{\nu_1(u)}=\frac{\overline \nu(v)}{\overline\nu(u)}.$$
\[Lemma7\] Suppose that $L$ is a 2 dimensional algebraic function field over $k$. Suppose that $R$ is an algebraic regular local ring with quotient field $L$ and maximal ideal $m_{R}=(u,v)$. Suppose that $\nu_1$ is a rank 1, rational rank 2 valuation of $\hat L=QF(\hat R)$ such that $\nu_1(u)$, $\nu_1(v)>0$ are rationally independent and which dominates $\hat{R}$. Then $\overline \nu=\nu_1\mid L$ is a rank 1, rational rank 2 valuation such that $$\Gamma_{\nu_1}={\bold Z}\overline \nu(u)+{\bold Z}\overline \nu(v)=\Gamma_{\overline \nu}.$$
By arguments as in the proof of 3. of Lemma \[Lemma3\], we see that if $f=\sum_{i+j=r}^{\infty}a_{ij}u^iv^j\in\hat R$ with $a_{ij}\in k$, then $\nu_1(f)=\text{min}\{\nu_1(u^iv^j)\mid a_{ij}\ne 0\}$. Thus $$\Gamma_{\nu_1} ={\bold Z}\overline\nu(u)+{\bold Z}\overline\nu(v)=\Gamma_{\overline\nu}.$$
Suppose that $R$ is an algebraic regular local ring with quotient field $K$. There exist many extensions of a given valuation $\nu$ of $K$ which dominates $R$ to $QF(\hat R)$ which do not dominate $\hat R$. Let $K=k(x)$, $\nu$ be the rank 1 discrete valuation with valuation ring $V_{\nu}=k[x]_{(x)}$ such that $\nu(x)=1$. Set $R=k[x]_{(x)}$.
Choose $f_1,\ldots,f_n\in \hat R=k[[x]]$ such that $x,f_1,\ldots,f_n$ are algebraically independent over $k$, and choose $\gamma_1,\ldots,\gamma_n\in {\bold R}$ such that $1,\gamma_1,\ldots,\gamma_n$ are linearly independent over ${\bold Q}$. $K_1=k(x,f_1,\ldots,f_n)$ is a rational function field in $n+1$ variables, so we can extend $\nu$ to a rank 1, rational rank $n+1$ valuation $\nu_1$ of $K_1$ by setting $\nu_1(f_i)=\gamma_i$, $1\le i\le n$. By Proposition 2.22 [@Ab4] or Theorem 5’, Section 4, Chapter VI [@ZS], $\nu_1$ extends (up to equivalence) to a valuation $\hat \nu$ of $QF(\hat R)$ which we can normalize so that it is an extension of $\nu$.
Write $f_i=x^{m_i}\lambda_i$ where $\lambda_i\in\hat R$ is a unit series. $\nu_1(\lambda_i)=\gamma_i-m_i\ne 0$. Since $\lambda_i$ and $\lambda_i^{-1}\in\hat R$, $\hat R$ contains elements of negative $\hat\nu$ value, and thus $\hat\nu$ does not dominate $\hat R$.
\[Lemma4\]Suppose that $K\rightarrow K^*$ is a finite extension of algebraic function fields over $k$ of dimension 2. Suppose that $\nu$ is a rank 1 rational rank 2 valuation of $K$, $\nu^*$ is an extension of $\nu$ to $K^*$. Suppose that $R_0$ is an algebraic regular local ring with quotient field $K$, maximal ideal $m_{R_0}=(u,v)$, $S$ is an algebraic regular local ring with quotient field $K^*$, maximal ideal $m_S=(x,y)$ and such that $S$ dominates $R$, $$\begin{array}{ll}
u&=x^ay^b\delta_1\\
v&=x^cy^d\delta_2
\end{array}$$ for some natural numbers $a,b,c,d$ and units $\delta_1,\delta_2\in S$, and such that the characteristic of $k$ does not divide $ad-bc$.
Suppose that $V_{\nu^*}$ dominates $S$ and $\nu(u),\nu(v)$ are rationally independent over $\bold
Q$. Then $$\Gamma_{\nu}={\bold Z}\nu(u)+{\bold Z}\nu(v)$$ and $\nu^*$ is a rank 1, rational rank 2 valuation of $K^*$ such that $\nu^*(x),\nu^*(y)$ are rationally independent over $\bold Q$, and $$\Gamma_{\nu^*}={\bold Z}\nu^*(x)+{\bold Z}\nu^*(y).$$ Suppose that $S\rightarrow S_1$ is a sequence of quadratic transforms along $\nu^*$. Then $S_1$ has regular parameters $(\overline x_1,\overline y_1)$ such that $$\begin{array}{ll}
x&=\overline x_1^{\overline a}\overline y_1^{\overline b}\\
y&=\overline x_1^{\overline c}\overline y_1^{\overline d}
\end{array}$$ with $\overline a\overline d-\overline b\overline c=\pm1$, and there exists a (unique) algebraic regular local ring $R_1$ with quotient field $K$ which lies below $S_1$. $\hat R_1\cong \hat{S_1}^{\Gamma_{\nu^*}/\Gamma_{\nu}}$, where $\Gamma_{\nu^*}/\Gamma_{\nu}$ acts faithfully on $\hat{S_1}$ by $k$-algebra automorphisms, by multiplication of $\overline x_1,\overline y_1$ by roots of unity in $k$.
$\nu^*$ has rational rank 2 and rank 1 since $K^*$ is finite over $K$ (Lemmas 1 and 2 of Section 11, Chapter VI [@ZS]). $\nu^*(x),\nu^*(y)$ are linearly independent over $\bold Q$, so Lemma \[Lemma3\] applies to $\nu$ and to $\nu^*$. We have an expression in $S_1$ $$\begin{array}{ll}
u&=\overline x_1^{\tilde a}\overline y_1^{\tilde b}\tilde \delta_1\\
v&=\overline x_1^{\tilde c}\overline y_1^{\tilde d}\tilde \delta_2
\end{array}$$ with natural numbers $\tilde a,\tilde b,\tilde c,\tilde d$ and units $\tilde\delta_1$ and $\tilde\delta_2$ in $S_1$ such that the characteristic of $k$ does not divide $\tilde a\tilde d-\tilde b\tilde c$. Let $$A=\left(\begin{array}{ll} \tilde a&\tilde b\\
\tilde c&\tilde d\end{array}\right),$$ $d=\mid\tilde a\tilde d-\tilde b\tilde c\mid$. There exist regular parameters $\tilde x_1,\tilde y_1$ in $\hat S_1$ such that $$u=\tilde x_1^{\tilde a}\tilde y_1^{\tilde b},\,\,\,\,\,
v=\tilde x_1^{\tilde c}\tilde y_1^{\tilde d}.$$ Let $\omega$ be a $d^{th}$ root of unity in $k$. ${\bold Z}^2/A{\bold Z}^2$ acts faithfully on $\hat S_1$ by $k$-algebra automorphisms. To $c\in {\bold Z}^2/A{\bold Z}^2$ the corresponding $k$-algebra automorphism $\sigma_c$ of $\hat S_1$ is defined by $$\sigma_c(\tilde x_1) =\omega^{<B_1,c>}\tilde x_1\,\,\,\,\,
\sigma_c(\tilde y_1)=\omega^{<B_2,c>}\tilde y_1$$ where $B_i$ is the $i^{th}$ row of $dA^{-1}=\pm\text{adj}(A)$. Since $k(u,v)\rightarrow
k(\tilde x_1,\tilde y_1)$ is Galois with Galois group ${\bold Z}^2/A{\bold Z}^2$, it follows that $\hat S_1^{{\bold Z}^2/A{\bold Z}^2}$ is the completion of a $k$-algebra generated by rational monomials $u^{\alpha_1}v^{\beta_1},\cdots,u^{\alpha_r}v^{\beta_r}$ (with $\alpha_i,\beta_i\in{\bold Z}$ for all $i$). $$R_0[u^{\alpha_1}v^{\beta_1},\ldots,u^{\alpha_r}v^{\beta_r}]\subset \hat S_1\cap K=S_1.$$ Let $R_1=R_0[u^{\alpha_1}v^{\beta_1},\ldots,u^{\alpha_r}v^{\beta_r}]_p$ where $p=R_0[u^{\alpha_1}v^{\beta_1},\ldots,u^{\alpha_r}v^{\beta_r}]\cap m_{S_1}$. $\hat R_1=\hat S_1^{{\bold Z}^2/A{\bold Z}^2}$ is normal, so $R_1=\hat R_1\cap K$ is normal. Since $\sqrt{m_{R_1}S_1}=m_{S_1}$, $R_1$ lies below $S_1$ by Zariski’s main Theorem (10.9 [@Ab5]). Uniqueness follows since the condition $R_1$ lies below $S_1$ implies $R_1=S_1\cap K$ by Proposition 1 (iv) [@Ab1].
The conclusion $\hat R_1\cong \hat{S_1}^{\Gamma_{\nu^*}/\Gamma_{\nu}}$ in Lemma \[Lemma4\] is a special case of a general result on ramification of valuations [@CP2].
\[Lemma5\] Suppose that $p$ is a prime such that $p$ is not the characteristic of $k$ and ${\bold Z}_p$ acts diagonally and faithfully on the powerseries ring $k[[x,y]]$. Set $R=k[[x,y]]^{{\bold Z}_p}$. Then $R$ is a normal local ring such that either
1. $R$ is regular and the algebraic fundamental group $$\pi_1(\text{spec}(R)-m_R)=0$$ or
2. $R$ is not regular, $R\rightarrow k[[x,y]]$ is unramified away from $m_R$ and the algebraic fundamental group $$\pi_1(\text{spec}(R)-m_R)\cong {\bold Z}_p.$$
Let $\omega$ be a primitive $p^{th}$ root of unity in $k$, $\sigma$ a generator of ${\bold Z}_p$. There exist integers $a,b$ with $0\le a,b<p$ such that $$\sigma(x)=\omega^ax,\,\,\,\,\,\sigma(y)=\omega^by.$$
Suppose that $a=0$. Then $R=k[[x,y^p]]$ is regular. If $b=0$ then $R=k[[x^p,y]]$ is regular. In both cases, $$\pi_1(\text{spec}(R)-m_R)=\pi_1(\text{spec}(R))=\pi_1(k)=0$$ by the purity of the branch locus (Theorems X 3.4, X 1.1 [@SGA2]).
Suppose that $a,b\ne 0$. Then for $1\le i\le p-1$ there exists a unique $j_i$ such that $bj_i\equiv ai\text{ mod }p$ with $0<j_i<p$. This implies that $x^{p-i}y^{j_i}$ is an invariant. Note that there exists an invariant of the form $x^{p-i_1}y$ for some $0<i_1<p$, so that $j_{i_1}=1$. We will show that $$R=k[[x^p,x^{p-1}y^{j_1},\ldots,xy^{j_{p-1}},y^p]].$$ We must show that any invariant monomial in $x$ and $y$ is a product of powers of these $p+1$ monomials.
Suppose that $x^iy^j$ is invariant. Then $ai+bj\equiv 0\text{ mod }p$. Write $i=\overline i+\lambda p$, $j=\overline j+\tau p$, with $0\le \overline i<p$, $0\le \overline j<p$. $$x^iy^j=x^{\overline i}y^{\overline j}x^{\lambda p}y^{\tau p}$$ $b\overline j\equiv -a\overline i\text{ mod }p$ implies $\overline i=\overline j=0$ or $\overline j=j_{p-\overline i}$.
Consider the finite map of normal local rings $$\Phi:Y=\text{spec}(k[[x,y]])\rightarrow X=\text{spec}(R).$$ The ramification locus of $\Phi$ is defined by the $2\times 2$ minors of $$J(\Phi)=
\left( \begin{array}{ll}
\frac{\partial(x^p)}{\partial x}& \frac{\partial(x^p)}{\partial y}\\
\frac{\partial (x^{p-1}y^{j_1})}{\partial x}&\frac{\partial(x^{p-1}y^{j_1})}{\partial y}\\
\vdots&\vdots\\
\frac{\partial(y^p)}{\partial x}&\frac{\partial(y^p)}{\partial y}
\end{array}\right).$$ $$\text{Det}\left(\begin{array}{ll}
y^{j_{p-1}}&j_{p-1}xy^{j_{p-1}-1}\\
0&py^{p-1}\\
\end{array}\right)=py^{p-1+j_{p-1}}$$ and $$\text{Det}\left(\begin{array}{ll} px^{p-1}&0\\
(p-i_1)x^{p-i_1-1}y&x^{p-i_1}\end{array}\right)=px^{2p-1-i_1}$$ implies $\sqrt{I_2(J(\Phi))}=(x,y)$. Thus $\Phi$ is unramified (and étale) away from $m_R$.
Suppose that $S$ is a complete normal local domain such that $S$ is finite over $R$, and $R\rightarrow S$ is étale away from $m_R$. Let $T$ be the normalization of the image of $S\otimes_Rk[[x,y]]$ in $QF(S)\otimes_{QF(R)}QF(k[[x,y,]])$. $k[[x,y]]\rightarrow T$ is étale away from $(x,y)$, so by the purity of the branch locus, and since $k$ is algebraically closed, $\text{spec}(T)$ is a disjoint union of copies of $\text{spec}(k[x,y]])$. A choice of one of these copies gives a factorization $$\text{spec}(k[[x,y]])\rightarrow\text{spec}(S)\rightarrow\text{spec}(R).$$ Thus $\pi_1(X-m_R)\cong {\bold Z}_p$.
Abhyankar constructs an example which shows that we cannot in general take $R$ to be regular in general in Corollary \[Corollary1\] (and thus we cannot take $R$ to be regular in Theorem \[Theorem1\]).
\[Theorem10\](Abhyankar) There exists a two dimensional algebraic regular local ring $\overline R$ with quotient field $\overline K$, a valuation $\overline \nu$ of $\overline K$ which dominates $\overline R$, and a finite extension $L_1$ of $\overline K$ such that if $\overline R_1$ is an algebraic regular local ring with quotient field $\overline K$ such that $\overline R\subset \overline R_1$ and $V_{\overline \nu}$ dominates $\overline R_1$, then there is a unique normal algebraic local ring $\overline S$ with quotient field $L_1$ lying over $\overline R_1$. $\overline S$ is not regular.
We give an outline of the construction, refering to Theorem 11 [@Ab3] for details.
Let $\overline K=k(u,v)$ be a rational function field in two variables. Let $q>3$ be a prime such that $q\ne \text{char}(k)$. Set $a=q-4$, Set $$\tau=a+\frac{1}{1+\frac{1}{a+\frac{1}{1+\frac{1}{a+\cdots}}}}\in{\bold R}-{\bold Q}.$$ Define a rank 1, rational rank 2 valuation $\overline \nu$ on $\overline K$ by setting $\overline \nu(u)=\tau$, $\overline\nu(v)=1$.
Set $\overline R=k[u,v]_{(u,v)}\subset V_{\overline\nu}$. Let $$L_1=\overline K[\overline z]/\overline z^q-uv^2.$$ Let $z$ be the image of $\overline z$ in $L_1$.
Abhyankar shows that if $\overline R_1$ is an algebraic regular local ring with quotient field $\overline K$ such that $\overline R\subset \overline R_1$, and $V_{\overline\nu}$ dominates $\overline R_1$, then there exists a unique normal algebraic local ring $\overline S$ with quotient field $L_1$ lying over $\overline R_1$ and $\overline S$ is not regular.
By Lemma \[Lemma3\], $\Gamma_{\overline\nu}={\bold Z}+\tau{\bold Z}$. We will show that there is a unique extension $\nu_1$ of $\overline \nu$ to $L_1$. First suppose that $\nu_1$ is a valuation of $L_1$ such that $V_{\nu_1}\cap \overline K=V_{\overline\nu}$. Since $\nu_1$ must have rank 1 and rational rank 2 (Lemmas 1 and 2, Section 11, Chapter VI [@ZS]), we can assume that the value group of $\nu_1$ is a subgroup of ${\bold R}$. We can then assume that $\nu_1$ is normalized so that $\nu_1(v)=1$. Since $\nu_1\mid \overline K$ is equivalent to $\overline\nu$, and $\nu_1(v)=1$, we have $\nu_1\mid \overline K=\overline\nu$ by Lemma \[Lemma3\]. Thus $\nu_1$ is an extension of $\overline\nu$. Since $\nu_1(z)=\frac{1}{q}(2+\tau)$, we have that $\Gamma_{\nu_1}/\Gamma_{\overline\nu}\cong{\bold Z}_q$, and $\nu_1$ is the unique extension of $\overline\nu$ to $L_1$, by corollary to Theorem 25, Section 12, Chapter VI [@ZS] and Lemma 2.18 [@Ab4].
Let $p$ be another prime such that $p\ne q$ and $p\ne\text{char}(k)$, and set $$L_2=\overline K[\overline w]/\overline w^p-uv^2.$$ Let $w$ be the image of $\overline w$ in $L_2$. By the same analysis as for $\nu_1$, there is a unique extension $\nu_2$ of $\overline \nu$ to $L_2$. $\nu_2(w)=
\frac{1}{p}(2+\tau)$ and $\Gamma_{\nu_2}/\Gamma_{\overline\nu}\cong {\bold Z}_p$.
We remark that $$\label{eq13}
\tau=q-4+\epsilon\text{ with }0<\epsilon<1.$$
Set $x_1(1)=\frac{v}{z}$, $y_1(1)=\frac{z^2}{v}\in L_1$. $$\begin{array}{ll}
\nu_1(x_1(1))&=\nu_1(v)-\nu_1(z)=\\
&=\frac{q-2-\tau}{q}=\frac{2-\epsilon}{q}>0
\end{array}$$ $$\begin{array}{ll}
\nu_1(y_1(1))&=2\nu_1(z)-\nu_1(v)=\frac{2}{q}(2+\tau)-1\\
&=\frac{q-4+2\epsilon}{q}>0.
\end{array}$$
Set $S_1=k[x_1(1),y_1(1)]_{(x_1(1),y_1(1))}$. $QF(S_1)=L_1$. $\overline R\subset S_1\subset V_{\nu_1}$. $$\label{eq18}
\begin{array}{ll}
u&=x_1(1)^{q-4}y_1(1)^{q-2}\\
v&=x_1(1)^2y_1(1).
\end{array}$$ $\Gamma_{\nu_1}=\nu_1(x_1(1)){\bold Z}+\nu_1(y_1(1)){\bold Z}$.
We will now impose the further condition that $5\le q<p<2q-4$. For example, we could take $q=11, p=13$ or $q=17, p=23$. Set $$x_1(2)=\frac{v}{w},\,\,\, y_1(2)=\frac{w^2}{v}\in L_2.$$ $$\begin{array}{ll}
\nu_2(x_1(2))&=\nu_2(v)-\nu_2(w)\\
&=\frac{p-2-\tau}{p}=\frac{(p-q)+2-\epsilon}{p}>0
\end{array}$$ $$\begin{array}{ll}
\nu_2(y_1(2))&=2\nu_2(w)-\nu_2(v)=\frac{2}{p}(2+\tau)-1\\
&=\frac{2q-p-4+2\epsilon}{p}>0.
\end{array}$$ Set $$S_2=k[x_1(2),y_1(2)]_{(x_1(2),y_1(2)}.$$ $QF(S_2)=L_2$. $\overline R\subset S_2\subset V_{\nu_2}$. $$\label{eq19}
\begin{array}{ll}
u&=x_1(2)^{p-4}y_1(2)^{p-2}\\
v&=x_1(2)^2y_1(2).
\end{array}$$
$\Gamma_{\nu_2}=\nu_2(x_1(2)){\bold Z}+\nu_2(y_1(2)){\bold Z}$.
We will now assume that $\text{char}(k)\ne 2$.
Let $k[x,y,z_1]$ be a polynomial ring in $x,y,z_1$, $$f=z_1^2-1+x^my^n\in k[x,y,z_1]$$ with $m,n$ odd and sufficiently large, as will be determined below. We will also assume that $m,n$ are not divisible by $\text{char}(k)$. Then $f$ is irreducible. Set $S_0=k[x,y,z_1]/(f)$. By abuse of notation, we will from now on identify $x,y,z_1$ with their equivalence classes in $S_0$. $S_0$ is smooth over $k$. Suppose that $a_1,b_1,c_1,d_1\in{\bold N}$ are such that $a_1d_1-b_1c_1$ is not divisible by $\text{char}(k)$ and $a_2,b_2,c_2,d_2\in{\bold N}$ are such that $a_2d_2-b_2c_2$ is not divisible by $\text{char}(k)$. We now impose the conditions $$m>\text{max}\{\mid a_1-a_2\mid,\mid c_1-c_2\mid\}$$ and $$n>\text{max}\{\mid b_1-b_2\mid,\mid d_1-d_2\mid\}.$$ Let $R=k[u,v]$ be a polynomial ring in two variables.
We define a $k$-algebra homomorphism $$R\rightarrow S_0$$ by $$\begin{array}{l}
u=x^{a_1}y^{b_1}(1-z_1)+x^{a_2}y^{b_2}(1+z_1)\\
v=x^{c_1}y^{d_1}(1-z_1)+x^{c_2}y^{d_2}(1+z_1)
\end{array}$$ Consider the prime ideals $P_1=(x,y,z_1+1)$ and $P_2=(x,y,z_1-1)$ in $S_0$. In the local ring $(S_0)_{P_1}$ we have $(P_1)_{P_1}=(x,y)$ since $$(z_1+1)=-(z_1-1)^{-1}x^my^n.$$ In the local ring $(S_0)_{P_2}$ we have $(P_2)_{P_2}=(x,y)$.
In $(S_0)_{P_1}$ we have $$\begin{array}{ll}
u&=x^{a_1}y^{b_1}(1-z_1-(z_1-1)^{-1}x^{a_2+m-a_1}y^{b_2+n-b_1})\\
&=x^{a_1}y^{b_1}\delta_1\\
v&=x^{c_1}y^{d_1}(1-z_1-(z_1-1)^{-1}x^{c_2+m-c_1}y^{d_2+n-d_1})\\
&=x^{c_1}y^{d_1}\delta_2
\end{array}$$ where $\delta_1,\delta_2\in (S_0)_{P_1}$ are units.
Since $a_1d_1-b_1c_1$ is not divisible by $\text{char}(k)$, we have regular parameters $x_1(1),y_1(1)\in\widehat{(S_0)_{P_1}}$ such that $$\label{eq16}
\begin{array}{ll}
u&=x_1(1)^{a_1}y_1(1)^{b_1}\\
v&=x_1(1)^{c_1}y_1(1)^{d_1}
\end{array}.$$ This implies that $R\subset S_0$ is an inclusion.
By a similar calculation, we have units $\epsilon_1,\epsilon_2\in (S_0)_{P_2}$ such that $$\begin{array}{ll}
u&=x^{a_2}y^{b_2}\epsilon_1\\
v&=x^{c_2}y^{d_2}\epsilon_2
\end{array}$$ and regular parameters $x_1(2),y_1(2)\in\widehat{(S_0)_{P_2}}$ such that $$\label{eq17}
\begin{array}{ll}
u&=x_1(2)^{a_2}y_1(2)^{b_2}\\
v&=x_1(2)^{c_2}y_1(2)^{d_2}
\end{array}.$$
With the notation introduced in Theorem \[Theorem10\], we have $\overline R=R_{(u,v)}$ and $\overline K=QF(R)$. Set $L_0=QF(S_0)$. $L_0$ is finite over $\overline K$ since $\overline K\rightarrow L_0$ is an inclusion of algebraic function fields of dimension 2.
In this construction, set $$a_1=q-4,b_1=q-2, c_1=2, d_1=1$$ and $$a_2=p-4,b_2=p-2,c_2=2,d_2=1,$$ where $p,q$ are the primes chosen in Theorem \[Theorem10\], and in the paragraph following Theorem \[Theorem10\].
Let $\Phi:Y\rightarrow X$ be a morphism of smooth projective surfaces over $k$ which extends our map $\text{spec}(S_0)\rightarrow \text{spec}(R)$. Such a map exists by resolution of singularities of surfaces in characteristic $\ge 0$ ([@Ab5], [@H2], [@L3]). For $i=1,2$ we have commutative diagrams: $$\label{eq14}
\begin{array}{lll}
\overline R=R_{(u,v)}&\rightarrow&(S_0)_{P_i}\\
\downarrow&&\downarrow\\
\hat{\overline R}=k[[u,v]]&\rightarrow &\widehat{(S_0)_{P_i}}=k[[x_1(i),y_1(i)]]
\end{array}$$ with (by (\[eq16\]) and (\[eq17\])) $$\begin{array}{ll}
u&=x_1(i)^{a_i}y_1(i)^{b_i}\\
v&=x_1(i)^{c_i}y_1(i)^{d_i}.
\end{array}$$ We further have commutative diagrams: $$\label{eq15}
\begin{array}{lll}
\overline R&\rightarrow&S_i\\
\downarrow&&\downarrow\\
\hat{\overline R}=k[[u,v]]&\rightarrow&\hat S_i=k[[x_1(i),y_1(i)]]
\end{array}$$ with (by (\[eq18\]) and (\[eq19\])) $$\begin{array}{ll}
u&=x_1(i)^{a_i}y_1(i)^{b_i}\\
v&=x_1(i)^{c_i}y_1(i)^{d_i}.
\end{array}$$ Diagrams (\[eq14\]) and (\[eq15\]) patch to give commutative diagrams for $i=1,2$ $$\label{eq20}
\begin{array}{lll}
R_{(u,v)}&\rightarrow &(S_0)_{P_i}\\
\downarrow&&\downarrow\\
\hat{\overline R}&\rightarrow &\hat S_i
\end{array}$$ Lemma \[Lemma3\] implies that for $i=1,2$, there exists a unique extension $\hat \nu_i$ of $\nu_i$ to $QF(\hat S_i)$ which dominates $\hat S_i$ and $\Gamma_{\hat\nu_i}=\Gamma_{\nu_i}$.
For $i=1,2$, let $\overline\nu_i=\hat\nu_i\mid L_0$ (under the inclusion $L_0\subset QF(\hat S_i)$ induced by (\[eq20\])). $\Gamma_{\overline\nu_i}\cong \Gamma_{\nu_i}$ by Lemma \[Lemma7\]. $\overline \nu_i\mid \overline K=\overline\nu$ for $i=1,2$ where $\overline\nu$ is the valuation of $\overline K$ introduced in Theorem \[Theorem10\].
Suppose that there exists a diagram $$\begin{array}{lll}
Y_1&\stackrel{\Phi_1}{\rightarrow}&X_1\\
\downarrow&&\downarrow\\
Y&\stackrel{\Phi}{\rightarrow}&X
\end{array}$$ such that the vertical arrows are birational and proper, $Y_1$ is nonsingular, $X_1$ is normal, and $\Phi_1$ is finite. Then $Y_1\rightarrow Y$ is a sequence of blowups of points (Theorem II.1.1 [@Z6]). There exist commutative diagrams
$$\label{eq21}
\begin{array}{lll}
R_1&\rightarrow&S(1)\\
\uparrow&&\uparrow \lambda_1\\
\overline R&\rightarrow&(S_0)_{P_1}
\end{array}$$
and $$\label{eq22}
\begin{array}{lll}
R_1&\rightarrow&S(2)\\
\uparrow&&\uparrow \lambda_2\\
\overline R&\rightarrow&(S_0)_{P_2}
\end{array}$$
where $R_1$ is the center of $\overline\nu$ on $X_1$, $S(1)$ is the local ring of the center of $\overline\nu_1$ on $Y_1$ and $S(2)$ is the local ring of the center of $\overline\nu_2$ on $Y_1$. $\lambda_1$ and $\lambda_2$ are products of quadratic transforms.
By Lemma \[Lemma4\] $$\label{eq23}
\hat R_1\cong \widehat{S(1)}^{\Gamma_{\overline\nu_1}/\Gamma_{\overline \nu}}.$$ and $$\label{eq24}
\hat R_1\cong \widehat{S(2)}^{\Gamma_{\overline\nu_2}/\Gamma_{\overline\nu}}.$$
By Theorem \[Theorem10\], (\[eq20\]) with $i=1$ and (\[eq21\]), $R_1$ is not regular. $$\Gamma_{\overline\nu_1}/\Gamma_{\overline \nu}\cong {\bold Z}_q\text{ and }
\Gamma_{\overline\nu_2}/\Gamma_{\overline \nu}\cong {\bold Z}_p$$ by our construction.
By Lemma \[Lemma5\] and (\[eq23\]), $$\pi_1(\text{spec}(\hat R_1)-m_{\hat R_1})
\cong{\bold Z}_q$$ and by (\[eq24\]) $$\pi_1(\text{spec}(\hat R_1)-m_{\hat R_1})
\cong{\bold Z}_p.$$ But $p\ne q$, so we have a contradiction.
Ramification of valuations in algebraic function fields
=======================================================
\[TheoremA\] (Monomialization; Theorem 1.1 [@C2]) Let $k$ be a field of characteristic zero, $K$ an algebraic function field over $k$, $K^*$ a finite algebraic extension of $K$, $\nu^*$ a $k$-valuation of $K^*$. Suppose that $S^*$ is an algebraic regular local ring with quotient field $K^*$ which is dominated by $\nu^*$ and $R^*$ is an algebraic regular local ring with quotient field $K$ which is dominated by $S^*$. Then there exist sequences of monoidal transforms $R^* \rightarrow R_0$ and $S^* \rightarrow S$ such that $\nu^*$ dominates $S$, $S$ dominates $R_0$ and there are regular parameters $(x_1, .... ,x_n)$ in $R_0$, $(y_1, ... ,y_n)$ in $S$, units $\delta_1,\ldots,\delta_n\in S$ and a matrix $A=(a_{ij})$ of nonnegative integers such that $\text{det}(A) \ne 0$ and $$\label{eq1*}
\begin{array}{lll}
x_1 &=& y_1^{a_{11}} \cdots y_n^{a_{1n}}\delta_1\\
&&\vdots\\
x_n &=& y_1^{a_{n1}} \cdots y_n^{a_{nn}}\delta_n.
\end{array}$$
The standard theorems on resolution of singularities allow one to easily find $R_0$ and $S$ such that (\[eq1\*\]) holds, but, in general, we will not have the essential condition $\text{det}(a_{ij}) \ne 0$. The difficulty in the proof of this Theorem is to achieve the condition $\text{det}(a_{ij})\ne0$.
Let $\alpha_i$ be the images of $\delta_i$ in $S/m_{S}$ for $1\le i\le n$. Let $C=(a_{ij})^{-1}$, a matrix with rational coefficients. Define regular parameters $(\overline y_1,\ldots,\overline y_n)$ in $\hat{S}$ by $$\overline y_i=\left(\frac{\delta_1}{\alpha_1}\right)^{c_{i1}}\cdots\left(\frac{\delta_n}{\alpha_n}\right)^{c_{in}}y_i$$ for $1\le i\le n$. We thus have relations $$\label{eq25}
x_i=\alpha_i\overline y_1^{a_{i1}}\cdots\overline y_n^{a_{in}}$$ with $\alpha_i\in S/m_S$ for $1\le i\le n$ in $$\hat R_0=R_0/m_{R_0}[[x_1,\ldots,x_n]]\rightarrow\hat S=S/m_S[[\overline y_1,\ldots,\overline y_n]].$$
\[RemarkT\] Suppose that $k'$ is a field, $A=(a_{ij})$ is an $n\times n$ matrix of natural numbers with $\text{det}(A)\ne 0$, and we have an inclusion of lattices $$N'={\bold Z}^n\stackrel{A}{\rightarrow} N={\bold Z}^n$$ with a corresponding inclusion of dual lattices $$M=\text{Hom}(N,{\bold Z})\rightarrow M'=\text{Hom}(N',{\bold Z}).$$ Let $\sigma$ be the cone generated by the rows of $A$ in $N\otimes {\bold R}$, $\sigma'$ be the cone generated by ${\bold N}^n$ in $N'\otimes{\bold R}$. Suppose that $\hat\sigma\cap M$ is generated by $e_1,\ldots,e_r$. By the theory of toric varieties (c.f. page 34 [@Fu]) we have inclusions of $k'$-algebras $$k'[\overline x_1,\ldots,\overline x_n]\rightarrow k'[\hat\sigma\cap M]=k'[\overline x^{e_1},
\ldots,\overline x^{e_r}]\rightarrow k'[\hat\sigma'\cap M']=k'[\overline y_1,\ldots,\overline y_n]$$ with $\overline x_i=\overline y_1^{a_{i1}}\cdots\overline y_n^{a_{in}}$ for $1\le i\le n$. $k'[\hat\sigma\cap M]$ is a normal ring with quotient field $k'(\overline x_1,\ldots,\overline x_n)$ and $k'[\hat\sigma'\cap M']$ is finite over $k[\hat\sigma\cap M]$.
If $k'={\bold C}$, this can be expressed in a particularly nice way. $N/N'\cong {\bold Z}^n/A{\bold Z}^n$ acts on ${\bold C}[\overline y_1,\ldots,\overline y_n]$ by associating to $c\in N/N'$ the ${\bold C}$-algebra automorphism $\sigma_c$ defined by $$\sigma_c(\overline y_i)=\text{exp}^{2\pi i<F_i,c>}\overline y_i$$ for $1\le i\le n$, where $$A^{-1}=\left(\begin{array}{l} F_1\\ \vdots\\ F_n\end{array} \right).$$
\[Theorem1\] Let $k$ be a field of characteristic zero, $K$ an algebraic function field over $k$, $K^*$ a finite algebraic extension of $K$, $\nu^*$ a $k$-valuation of $K^*$. Suppose that $S^*$ is an algebraic local ring with quotient field $K^*$ which is dominated by $\nu^*$ and $R^*$ is an algebraic local ring with quotient field $K$ which is dominated by $S^*$. Then there exists a commutative diagram $$\label{eq3}
\begin{array}{lllllll}
R_0&\rightarrow& R&\rightarrow& S&\subset&V_{\nu^*}\\
\uparrow&&&&\uparrow\\
R^*&&\rightarrow&&S^*
\end{array}$$ where $S^*\rightarrow S$ and $R^*\rightarrow R_0$ are sequences of monodial transforms along $\nu^*$ such that $R_0\rightarrow S$ have regular parameters of the form of the conclusions of Theorem \[TheoremA\], $R$ is an algebraic normal local ring with toric singularities, which is the localization of the blowup of an ideal in $R_0$, and the regular local ring $S$ is the localization at a maximal ideal of the integral closure of $R$ in $K^*$.
By resolution of singularities [@H1] (c.f. Theorem 2.6, Theorem 2.9 [@C2]), we first reduce to the case where $R^*$ and $S^*$ are regular, and then construct, by the local monomialization theorem, Theorem \[TheoremA\] a sequence of monodial transforms along $\nu^*$ $$\label{eq6}
\begin{array}{lllll}
R_0&\rightarrow &S&\subset &V_{\nu^*}\\
\uparrow&&\uparrow\\
R^*&\rightarrow &S^*
\end{array}$$ so that $R_0$ is a regular local ring with regular parameters $(x_1,\ldots, x_n)$, $S$ is a regular local ring with regular parameters $(y_1,\ldots, y_n)$, there are units $\delta_1,\ldots,\delta_n$ in $S$, and a matrix of natural numbers $A=(a_{ij})$ with nonzero determinant $d$ such that $$x_i=\delta_iy_1^{a_{i1}}\cdots y_n^{a_{in}}$$ for $1\le i\le n$.
Let $k'$ be an algebraic closure of $S/m_S$. With the notation of (\[eq25\]), set $\overline x_i=\frac{x_i}{\alpha_i}$, so that $$R_0\otimes_{R_0/m_{R_0}}k'\cong k'[[\overline x_1,\ldots,\overline x_n]]
\rightarrow S\otimes_{S/m_S}k'\cong k'[[\overline y_1,\ldots,\overline y_n]]$$ is defined by $$\overline x_i=\overline y_1^{a_{i1}}\cdots\overline y_n^{a_{in}},$$ $1\le i\le n$. With the notation of Remark \[RemarkT\], set $R=R_0[x^{e_1},\ldots,x^{e_r}]_m$ where $m=(x^{e_1},\ldots,x^{e_r})$. $$R_0[x^{e_1},\ldots,x^{e_r}]\subset \hat S\cap K^*=S$$ (by Lemma 2 [@Ab1]) and $m\subset m_{\hat S}\cap K^*=m_S$, so $S$ dominates $R$. $\hat R=R_0/m_{R_0}[[\overline x^{e_1},\ldots,x^{e_r}]]$ is integrally closed in its quotient field (by Remark \[RemarkT\] and Theorem 32, Section 13, Chapter VIII [@Z2]), so $R=\hat R\cap K$ is integrally closed in $K$. After possibly reindexing the $y_i$, we may assume that $d=\text{det}(A)>0$. Let $(b_{ij})$ be the adjoint matrix of $A$. Then $$\prod_{j=1}^nx_j^{b_{ij}}=\left(\prod_{j=1}^n\delta_j^{b_{ij}}\right)y_i^d\in R.$$ Thus $\sqrt{m_RS}=m_S$, so $R$ lies below $S$ by Zariski’s Main Theorem (10.9 [@Ab5]).
As an immediate consequence, we obtain a proof in characteristic zero of the “weak simultaneous resolution local conjecture”. which is stated explicitely on page 144 of [@Ab6], and is implicit in [@Ab3]. Abhyankar proves this for algebraic function fields of dimension two and any characteristic in [@Ab1] and [@Ab4]. In the paper [@C3], we have given a direct proof of this result, also as a consequence of Theorem \[TheoremA\] (Theorem 1.1 [@C2]).
\[Corollary1\](Corollary1)(Theorem 1.1 [@C3]) Let $k$ be a field of characteristic zero, $K$ an algebraic function field over $k$, $K^*$ a finite algebraic extension of $K$, $\nu^*$ a $k$-valuation of $K^*$, and $S^*$ an algebraic regular local ring with quotient field $K^*$ which is dominated by $\nu^*$. Then for some sequence of monodial transforms $S^*\rightarrow S$ along $\nu^*$, there exists a normal algebraic local ring $R$ with quotient field $K$, such that the regular local ring $S$ is the localization at a maximal ideal of the integral closure of $R$ in $K^*$.
There exists a normal algebraic local ring $R^*$ with quotient field $K$ such that $\nu^*$ dominates $R^*$ (take $R^*$ to be the local ring of the center of $\nu^*$ on a normal projective model of $K$). There exists a finite type $k$-algebra $T$ such that the integral closure of $R^*$ in $K^*$ is a localization of $T$, and $T$ is generated over $k$ by $g_1,\ldots,g_m\in V^*=V_{\nu^*}$ such that $\nu^*(g_i)\ge 0$ for all $i$. There exists a sequence of monoidal transforms $S^*\rightarrow S_1$ along $\nu^*$ such that $T\subset S_1$ (Theorem 2.7 [@C2]). $S_1$ dominates $R^*$. After replacing $S^*$ with $S_1$, we can assume that $S^*$ dominates $R^*$. Theorem \[Theorem1\] applies to this situation, so we can construct a diagram of the form (\[eq3\]).
When $K^*$ is Galois over $K$, it is not difficult to construct using Galois theory and resolution of singularities a regular local ring $S$ with quotient field $K^*$ and a normal local ring $R$ with quotient field $K$ such that $S$ lies over $R$ (Theorem 7 [@Ab3], Theorem \[Theorem7\]), although even in the Galois case the full statements of Theorem \[Theorem1\] and Corollary \[Corollary1\] do not follow from these results (Theorem 7 [@Ab3], Theorem \[Theorem7\]). The general case of non Galois extensions is much more subtle, and not as well behaved, as can be seen from Theorem \[Example1\].
Generically finite morphisms
============================
Suppose that $f:Y\rightarrow X$ is a dominant, generically finite morphism of complete $k$-varieties. In this section we construct a commutative diagram $$\begin{array}{lll}
\overline Y&\stackrel{f_1}{\rightarrow}&\overline X\\
\downarrow&&\downarrow\\
Y&\stackrel{f}{\rightarrow}&X
\end{array}$$ of the form of (\[eqI1\]) such that $\overline Y$ is nonsingular, $f_1$ is “close” to being finite, and the vertical arrows are birational and “close” to being proper.
Let $K$ be an algebraic function field over a field $k$ of characteristic zero. In his work on resolution of singularities, Zariski [@Z2] constructed for each $k$-valuation ring $V$ of $K$ a projective model $X_V$ of $K$ such that the center of $V$ is nonsingular on $X_V$. By the quasi-compactness of the Zariski-Riemann manifold of valuations of $K$, there exists a finite number of the $X_V$, $\{X_{V_1},\ldots,X_{V_n}\}$ such that every valuation ring $V$ of $K$ has a nonsingular center on at least one of the $X_{V_i}$.
In dimension $\le 3$, Zariski [@Z5] was able to patch the open nonsingular locus of appropriate birational transforms of the $X_{V_i}$ to produce a nonsingular projective model $X$ of $K$.
The only part of Zariski’s proof of the existence of a nonsingular model which does not extend to arbitrary dimension is the final step where nonsingular open subsets $U_i$ of $X_{V_i}$ are patched (after appropriate birational transforms) to produce a projective variety. Hironaka observes in Chapter 0, Section 6 of [@Ha] that we can always patch the nonsingular loci $U_i$ of the $X_{V_i}$ along open sets where they are isomorphic, to produce an integral finite type scheme $X$ such that $X$ is nonsingular and every valuation ring $V$ of $K$ has a center on $X$, but $X$ will in general not be separated. Hironaka calls such schemes “complete”.
If such an $X$ is separated, then the following Lemma shows that $X$ is in fact complete in the usual sense, that is $X$ is a proper $k$-scheme.
\[LemmaS\] Suppose that $X$ is a separated, integral, finite type $k$-scheme, such that every $k$-valuation of $k(X)$ has a center on $X$. Then $X$ is a proper $k$-scheme.
The proof of this Lemma is immediate from (ii) of Corollary II 7.3.10 [@EGA] (Recall that in the notation of [@EGA], a scheme is a separated pre-scheme), or can be deduced directly from the valuative criterion of properness (Theorem II 7.3.8 [@EGA] or Theorem II 4.7 [@Ha]).
One may expect a scheme $X$ which satisfies all of the conditions of the Lemma above except the separatedness condition to be universally closed. This is false, as is shown by the following example.
Suppose that $k$ is a field. There exists a (nonseparated) integral finite type $k$-scheme such that every $k$-valuation ring of $k(X)$ has a center on $X$, but $X$ is not universally closed over $k$.
Let $\phi$ be an imbedding of $\bold{P}^1$ in $\bold P^3$. Let $Z=\phi({\bold P}^1-\{\infty\})$, $x_0=\phi(\infty)$. Let $\pi:X_1\rightarrow{\bold P}^3$ be the blowup of $\overline Z=\phi({\bold P}^1)$, $X_2={\bold P}^3-\{x_0\}$. We can construct an integral finite type $k$-scheme $X$ by glueing $X_1$ to $X_2$ along the open sets $X_1-\pi^{-1}(\overline Z)$ and $X_2-Z$. By construction, every $k$-valuation of $k(X)$ has a center on $X$. $\phi$ induces an isomorphism of ${\bold P}^1-\{\infty\}$ with the closed subset $Z\subset X_1$. $Z$ is closed in $X$. Thus the induced morphism $\phi:\text{spec}(k({\bold P}^1)
\rightarrow X$ does not extend to a morphism $\text{spec}({\cal O}_{{\bold P}^1,\infty})
\rightarrow X$.
Suppose that $X$ is universally closed over $k$. Let $U=\text{spec}(k({\bold P}^1))$, $T=\text{spec}({\cal O}_{{\bold P}^1,\infty})$, with natural morphism $i:U\rightarrow T$. Let $t_1\in T$ be the generic point, $t_0\in T$ the special point.
Let $A$ be the closure of $\phi\times i(U)$ in $X\times T$. $\pi_2(A)$ is closed since the projection $\pi_2:X\times T\rightarrow T$ is closed by assumption. So there exists $y_0\in A$ such that $\pi_2(y_0)=t_0$. By Lemma II 4.4 [@Ha] we have an extension $T\rightarrow X\times T$ of $U\rightarrow X\times T$ which projects to an extension $\text{spec}(T)\rightarrow X$ of $U\rightarrow X$, and thus an extension of $\phi$ to $\text{spec}({\cal O}_{{\bold P}^1,\infty})\rightarrow X$, a contradiction.
\[Theorem8\] Suppose that $f:Y\rightarrow X$ is a dominant, generically finite morphism of complete varieties over a field $k$ of characteristic zero. Then there exists a commutative diagram of integral, finite type $k$-schemes $$\begin{array}{lll}
\overline Y &\rightarrow &\overline X\\
\downarrow&&\downarrow\\
Y&\rightarrow&X
\end{array}$$ such that $\overline Y$ is nonsingular, $\overline X$ has normal toric singularities, the vertical arrows are birational morphisms, $\overline Y\rightarrow \overline X$ is quasi-finite and every $k$-valuation ring of $k(X)$ has a center on $\overline X$, every $k$-valuation ring of $k(Y)$ has a center on $\overline Y$.
Theorem \[Example1\] (and Lemma \[LemmaS\]) show that we cannot take the vertical arrows in Theorem \[Theorem8\] to be proper.
Let $K=k(X)$ be the function field of $X$, $K^*=k(Y)$ be the function field of $Y$. By assumption, $k(Y)$ is finite over $k(X)$. By resolution of singularities [@H1], we may assume that $Y$ and $X$ are nonsingular. Let $V^*$ be a $k$-valuation ring of $K^*$ such that $\text{trdeg}_kV^*/m_{V^*}=0$, $V=V^*\cap K$. Let $p$ be the center of $V$ on $X$, $q$ the center of $V^*$ on $Y$.
By Theorem \[Theorem1\], there exists a sequence of the form (\[eq3\]), $$\begin{array}{lllll}
R_0&\rightarrow R\rightarrow&S&\subset&V^*\\
\uparrow&&\uparrow\\
{\cal O}_{X,p}&\rightarrow&{\cal O}_{Y,q}
\end{array},$$ such that $S$ is regular and $R$ has toric singularities.
Let $N={\bold Z}^n$, $\sigma$ be the cone generated by the rows of $A=(a_{ij})$ (with the notation of (\[eq1\*\]) in $N\otimes{\bold R}$. Let $M$ be the dual lattice of $N$, $\hat\sigma$ be the dual cone of $\sigma$. By the proof of Theorem \[Theorem1\] (and Remark \[RemarkT\]), if $\hat\sigma\cap M$ is generated by $e_1,\ldots, e_r$, then $R=R_0[x^{e_1},\ldots,x^{e_r}]_m$. There is a natural inclusion $$k[\hat\sigma\cap M]=
k[x^{e_1},\ldots,x^{e_r}]\rightarrow R.$$ $U_{\sigma}=\text{spec}(k[\hat\sigma\cap M])$ is a normal affine toric variety.
Thus there exist affine open sets $U_p$ of $p$ in $X$ and $\overline U_q$ of $q$ in $Y$ and affine rings $R_V$ with quotient fields $K$ and $S_{V^*}$ with quotient field $K^*$ with the following properties:
1. $R_V$ is normal and $S_{V^*}$ is regular.
2. If $p_1$ is the center of $V$ on $R_V$ and $q_1$ is the center of $V^*$ on $S_{V^*}$, then $(R_V)_{p_1}=R$, $(S_{V^*})_{q_1}=S$.
3. There is a commutative diagram $$\label{eq27}
\begin{array}{lll}
R_V&\rightarrow&S_{V^*}\\
\uparrow&&\uparrow\\
\Gamma(U_p,{\cal O}_X)&\rightarrow&\Gamma(\overline U_q,{\cal O}_Y)
\end{array}$$ such that $R_V\rightarrow S_{V^*}$ is quasi-finite, $$k[\hat\sigma\cap M]=k[x^{e_1},\ldots, x^{e_r}]\rightarrow R_V$$ is étale, so that $R_V$ has normal toric singularites, and the vertical arrows are birational morphisms.
Let $Z(K)$ be the Zariski-Riemann manifold of $k$-valuation rings of $K$, $Z(K^*)$ be the Zariski-Riemann manifold of $k$-valuation rings of $K^*$. These spaces have natural topologies with respect to which they are quasi-compact (Theorem 40 Section 17, Chapter VI, [@ZS]). There is a natural continuous map $\Phi:Z(K^*)\rightarrow Z(K)$ defined by $\Phi(V^*)=V^*\cap K$. For each $V^*\in Z(K^*)$ such that $\text{trdeg}_k V^*/m_{V^*}=0$, let $Y_{V^*}$ be a projective variety which contains $\text{spec}(S_{V^*} )$ as an open set, and let $X_V$ be a projective variety which contains $\text{spec}(R_V)$ as an open set and such that the birational rational maps $Y_{V^*}\rightarrow X_V$, $Y_{V^*}\rightarrow Y$ and $X_V\rightarrow X$ are morphisms. Then there are commutative diagrams of continuous maps $$\begin{array}{lll}
Z(K^*)&\rightarrow&Z(K)\\
\downarrow\pi_{V^*}&&\downarrow\pi_V\\
Y_{V^*}&\rightarrow&X_V\\
\downarrow&&\downarrow\\
Y&\rightarrow&X
\end{array}.$$ Let $\overline W_{V^*}=\pi_{V^*}^{-1}(\text{spec}(S_{V^*}))$, an open neighborhood of $V^*$ in $Z(K^*)$, $W_V=\pi_V^{-1}(\text{spec}(R_V))$, an open neighborhood of $V$ in $Z(K)$. Suppose that $V'\in Z(K^*)$ and $\text{trdeg}_k V'/m_{V'}>0$. Let $\pi:V'\rightarrow
V'/m_{V'}$ be the residue map. By corollary 3 to Theorem 5, Section 4, Chapter VI [@ZS], there exists a $k$-valuation ring $V_1$ with quotient field $V'/m_{V'}$ such that $\text{trdeg}_k V_1/m_{V_1}=0$. Set $V^*=\pi^{-1}(V_1)$, a $k$-valuation ring of $K^*$ such that $V^*$ is a specialization of $V'$ and $\text{trdeg}_kV^*/m_{V^*}=0$ (page 57 [@Ab4]). Since $V^*$ dominates a local ring $S$ of $W_{V^*}$, $V'$ must dominate a localization of $S$, which is the local ring of a point of $W_{V^*}$. Thus $V'\in \overline W_{V^*}$. $\{\overline W_{V^*}\}_{V^*\in Z(K^*)}$ is thus an open cover of $Z(K^*)$. $\{W_{V}\}_{V\in Z(K)}$ is also an open cover of $Z(K)$.
Since $Z(K^*)$ and $Z(K)$ are quasi-compact, there is a finite set of valuation rings $V_1^*,\ldots,V_m^*\in Z(K^*)$ with $\text{trdeg}_k V_i^*/m_{V_i^*}=0$ for all $i$ such that if $V_i=V_i^*\cap K$ then $\{\overline W_{V_1^*},\ldots,\overline W_{V_m^*}\}$ is an open cover of $Z(K^*)$ and $\{W_{V_1},\ldots,W_{V_m}\}$ is an open cover of $Z(K)$.
For $1\le i\le m$ we have commutative diagrams $$\begin{array}{rrr}
D_i:=\text{spec}(S_{V_i^*})&\rightarrow&C_i:=\text{spec}(R_{V_i})\\
b_i\downarrow&&a_i\downarrow\\
\overline U_{q_i}&\rightarrow&U_{p_i}
\end{array},$$ where $q_i$ is the center of $V_i^*$ on $Y$, $p_i$ is the center of $V_i$ on $X$. Let $A_i\subset X$ be nontrivial open sets where $a_i$ is an isomorphism and let $B_i\subset f^{-1}(A_i)$ be nontrivial open sets where $b_i$ is an isomorphism. Then define $\overline Y$ to be the finite type $k$-scheme obtained by patching $D_i$ to $D_j$ for $i\ne j$ along the nontrivial open set $B_i\cap B_j$, and define $\overline X$ to be the finite type $k$-scheme obtained by patching $C_i$ to $C_j$ for $i\ne j$ along the nontrivial open set $A_i\cap A_j$.
By construction, $\overline Y$ and $\overline X$ are integral, $\overline Y$ is nonsingular, $\overline X$ has normal toric singularities and $\overline f:\overline Y\rightarrow \overline X$ is quasi-finite.
Galois extensions
=================
For Galois extensions, Question 1 of the introduction has a positive answer. Suppose that $K\rightarrow K^*$ is a finite Galois extension of algebraic function fields of characteristic zero. The existence of a finite map of normal projective $k$-varieties $Y\rightarrow X$, where $Y$ is nonsingular, $k(X)=K$, $k(Y)=K^*$ has been proven by Abhyankar in Theorem 7 [@Ab3]. We prove a relative version of this result in Theorem \[Theorem7\].
\[Theorem7\] Suppose that $\Phi:Y\rightarrow X$ is a dominant morphism of projective varieties over an algebraically closed field $k$ of characteristic zero such that $k(Y)$ is a finite Galois extension of $k(X)$. Then there exists a commutative diagram $$\begin{array}{lll}
\overline Y&\rightarrow &\overline X\\
\downarrow&&\downarrow\\
Y&\rightarrow&X
\end{array}$$ such that $\overline Y\rightarrow Y$, $\overline X\rightarrow X$ are birational morphisms of projective $k$-varieties, $\overline Y$ is nonsingular, $\overline X$ is normal, $\overline Y\rightarrow \overline X$ is finite and $\overline X$ has normal toric singularities.
Let $X_0$ be the normalization of $X$ in $k(X)$, $Y_0$ be the normalization of $X_0$ in $k(Y)$. Let $G=\text{Gal}(k(Y)/k(X))$. $G$ acts on $Y_0$ and $Y_0/G=X_0$. Let $D_0$ be the branch locus of $Y_0\rightarrow X_0$. Let $\pi_1:X_1\rightarrow X_0$ be a resolution of singularities so that $D_1=\pi_1^{-1}(D_0)_{red}$ is a simple normal crossings divisor. If $X_1$ is the blowup of an ideal sheaf ${\cal I}_0$ in $X_0$, let $Y_1$ be the normalization of the blowup of ${\cal I}_0{\cal O}_{Y_0}$. $f_1:Y_1\rightarrow X_1$ is finite, and $G$ acts on $Y_1$. The branch locus of $f_1$ which is a divisor supported on $D_1$ (by the purity of the branch locus) has simple normal crossings, so by Abhyankar’s Lemma ([@Ab7], XIII 5.3 [@SGA]) $Y_1$ has normal toric singularities, and (by Lemma 7 [@Ab1]) if $p\in Y_1$ is a closed point, the stabilizer $$G^s(p)=\{\sigma\in G\mid \sigma(p)=p\}$$ is Abelian. Suppose that ${\cal I}_1\subset {\cal O}_{Y_1}$ is an ideal sheaf such that the blowup of ${\cal I}_1$ in $Y_1$ dominates $Y$. Let ${\cal J}=\prod_{g\in G}g({\cal I}_1)$, and $\tilde
Y_2$ be the blowup of ${\cal J}$. Then $G$ acts on $\tilde Y_2$ and the rational map $\tilde Y_2\rightarrow Y$ is a morphism. Let $Y_2\rightarrow \tilde Y_2$ be a $G$-equivariant resolution of singularities of $\tilde Y_2$ ([@BM], [@Vi]), with composed map $\pi_2:Y_2\rightarrow Y_1$. $G$ acts on $Y_2$ and for $q\in Y_2$, $G^s(q)<G^s(\pi_2(q))$ so that $G^s(q)$ is Abelian. Let $\overline Y=Y_2$, $\overline X=Y_2/G$, a projective $k$-variety (c.f. page 126, [@Har]), which has normal toric singularities (by Lemma 7, [@Ab1]).
\[Example2\] Even if $k(Y)$ is Galois over $k(X)$, with Galois group $G$, and $Y\rightarrow X$ is $G$-equivariant, we cannot take both $\overline X$ and $\overline Y$ to be nonsingular in Theorem \[Theorem8\] or in Theorem \[Theorem7\].
This is an immediate consequence of Abhyankar’s example, Theorem 11, [@Ab3] (restated in Theorem \[Theorem10\] of this paper.) Let $k$ be an algebraically closed field of characteristic zero. With the notations of Theorem \[Theorem10\], $L_1$ is a Galois extension of $\overline K=k(u,v)$ with Galois group ${\bold Z}_q$. ${\bold Z}_q$ acts on $S=k[u,v,z]/z^q-uv^2$ and its invariant ring is $R=k[u,v]$. With the notation of Theorem \[Theorem10\], $\overline R=R_{(u,v)}$. By equivariant resolution of singularites [@BM], [@Vi] (applied to the normalization of $X={\bold P}^2$ in $L_1$) there exists a dominant ${\bold Z}_q$ equivariant morphism of nonsingular projective $k$-surfaces $Y\rightarrow X$ such that $k(X)=\overline K$, $k(Y)=L_1$ and there exists a point $p\in X$ such that ${\cal O}_{X,p}=\overline R$.
Suppose that there exists a diagram $$\begin{array}{lll}
\overline Y&\stackrel{\overline \Phi}{\rightarrow} &\overline X\\
\downarrow&&\downarrow\\
Y&\rightarrow&X
\end{array}$$ as in the conclusions of Theorem \[Theorem8\] such that $\overline X$ is nonsingular. Let $\overline\nu$ be the valuation of $\overline K$ constructed in Theorem \[Theorem10\]. Let $\nu_1$ be the (unique) extension of $\overline\nu$ to $L_1$. Let $q\in\overline Y$ be a center of $\nu_1$, $p=\overline\Phi(q)$. $B={\cal O}_{\overline Y,q}$ dominates $A={\cal O}_{\overline X,p}$. Since $\overline\Phi$ is quasi-finite, $B$ lies over $A$ by Zariski’s Main Theorem (10.9 [@Ab5]). Since $A$ is regular and $\overline\nu$ dominates $A$, $B$ is not regular by Theorem \[Theorem10\], a contradiction.
[99]{}
, Local uniformization on algebraic surfaces over ground fields of characteristic $p\ne 0$, Annals of Math, 63 (1956), 491-526.
, Simultaneous resolution for algebraic surfaces, Amer. J. Math 78 (1956), 761-790.
, Ramification theoretic methods in algebraic geometry, Princeton University Press, 1959.
, Resolution of singularities of embedded algebraic surfaces, second edition, Springer, 1998.
, Resolution of singularities and modular Galois theory, Bulletin of the AMS 38 (2001), 131-171.
, Tame Coverings and fundamental groups of algebraic varieties, Amer. J. Math. 81. 1959, 46-94.
, Canonical desingularization in characteristic zero by blowing up the maximum strata of a local invariant, Invent. Math. 128, 1997, 207-302.
, Local Factorization and Monomialization of Morphisms, Astérisque 260, 1999.
, Simultaneous resolution of singularities, Proc. American Math. Soc. 128, (2000), 1905-1910.
, Monomial resolutions of morphisms of algebraic surfaces, Communications in Algebra 28 (2000), 5935-5959.
, Ramification of valuations, in preparation.
, Introduction to Toric Varieties, Princeton University Press, 1993.
Revêtements étales et groupe fondemental, Lecture Notes in math 224, Springer Verlag, Heidelberg (1971).
Cohomologie Locale des Faisceaux Cohérents et Théorèmes de Lefshetz locaux et Globaux, North-Holland, Amsterdam (1968).
Eléments de Géometrie Algébrique II, Étude globale élèmentaire de quelques classes de morphismes, Publ. Math. IHES 8 (1961).
, Algebraic Geometry, a first course, Springer Verlag, 1992.
, Algebraic Geometry, Springer-Verlag, 1977.
, Approximating discrete valuation rings by regular local rings, a remark on local uniformization, preprint.
, Resolution of singularities of an algebraic variety over a field of characteristic zero, Annals of Math, 79 (1964), 109-326.
, Desingularization of excellent surfaces, Advanced Science Seminar in Algebraic Geometry, Bowdoin College, Brunswick, Maine, 1967.
, Valuation theoretic and model theoretic aspects of local uniformization, in Resolution of Singularities, edited by H. Hauser, J. Lipman, F. Oort, A. Quirós, Springer-Verlag, 2000.
, Desingularization of two-dimensional schemes, Ann. Math. 107 (1978), 151-207.
, Quasi-Canonical uniformization of hypersurface singularities of characteristic zero, Comm. Algebra 20, 3207-3249 (1992).
, Valuations, Deformations and Toric Geometry, in Resolution of Singularities, edited by H. Hauser, J. Lipman, F. Oort, A. Quirós, Springer-Verlag, 2000.
, A course on constructive desingularization and equivariance, in Resolution of Singularities, edited by H. Hauser, J. Lipman, F. Oort, A. Quirós, Springer-Verlag, 2000.
, The reduction of the singularities of an algebraic surface, Annals of Math., 40 (1939) 639-689.
, Local uniformization of algebraic varieties, Annals of Math., 41, (1940), 852-896.
, Reduction of the singularities of algebraic three dimensional varieties, Annals of Math., 45 (1944) 472-542.
, Introduction to Minimal Models.
, Commutative Algebra II, Van Nostrand, Princeton (1960).
\
\
Steven Dale Cutkosky, Department of Mathematics, University of Missouri\
Columbia, MO 65211, USA\
cutkoskys@@missouri.edu
[^1]: partially supported by NSF
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'All the content consumed by mobile users, be it a web page or a live stream, undergoes some processing along the way; as an example, web pages and videos are transcoded to fit each device’s screen. The recent multi-access edge computing (MEC) paradigm envisions performing such processing [*within*]{} the cellular network, as opposed to resorting to a cloud server on the Internet. Designing a MEC network, i.e., placing and dimensioning the computational facilities therein, requires information on how much computational power is required to produce the contents needed by the users. However, real-world demand traces only contain information on how much data is downloaded. In this paper, we demonstrate how to [*enrich*]{} demand traces with information about the computational power needed to process the different types of content, and we show the substantial benefit that can be obtained from using such enriched traces for the design of MEC-based networks.'
author:
- 'Francesco Malandrino, Carla-Fabiana Chiasserini, Giuseppe Avino, Marco Malinverno, Scott Kirkpatrick, [^1]'
bibliography:
- 'refs.bib'
title: |
From Megabits to CPU Ticks:\
Enriching a Demand Trace in the Age of MEC
---
Introduction {#sec:intro}
============
Dynamic web pages, targeted advertisement, user-generated content have all increased the [*computation*]{} required to assemble the pages displayed by web browsers. Mobile services, and the mobile devices consuming them, have made this trend even more evident, and virtually all content mobile users see is the result of multiple steps of real-time, on-the-fly processing. Prominent examples include videos, that are transcoded to match the screen size and resolution of the device playing them, and social networks, which decide what to show to their users based on their identity and location.
[*Where*]{} such processing ought to be performed is a fundamental question. The traditional approach was to use ad hoc servers placed within each content provider’s network; cloud computing, with shared virtual servers placed in the Internet, provides a more efficient but substantially equivalent alternative [@5gcloud]. The recent multi-access edge computing (MEC) paradigm takes a different approach: it envisions moving services, i.e., the servers providing them, as close to mobile users as possible. Servers will not be localized in remote datacenters, but at different entities within the mobile core network itself, from core switches to base stations [@noi-pof; @noi-workshop]. Similarly to cloud scenarios, there is a measure of cooperation between content providers and network operators. The servers deployed by the network operators will run services developed by the content providers, usually coming as a set of [*virtual network functions*]{} (VNFs) running on virtual machines or containers.
Designing a MEC network requires making decisions on where, within the network edge, servers shall be placed, and how to dimension them. Both questions require a deep knowledge of the data demand the network will have to serve: more exactly, we need to know (i) how much data the users will require; (ii) the type of such data, e.g., the mobile apps requesting it, and (iii) how much computational power will be needed to produce these data. The first two items have been widely studied: using operator-provided [@orange-d4d; @Malandrino-TCS] and crowd-sourced [@noi-pof; @noi-workshop] traces, it is possible for researchers to obtain a fairly good picture of the data demand of mobile networks, including its time and space evolution, as well as the services contributing to it.
Estimating the [*processing power*]{} needed to serve a given network demand, on the other hand, is much more challenging. The processing required to generate one gigabit per second of video data is not the same as the processing needed for the same quantity of gaming updates or maps. Additionally, it makes a significant difference whether the data is consumed by a small number of users enjoying a high bandwidth, or through a larger number of lower-rate connections. [*None*]{} of the currently available real-world traces contain all the information needed to distinguish these cases.
We fill this gap by developing and demonstrating a methodology to [*enrich*]{} existing demand traces with computational power information, translating (so to say) megabits of downloaded data into CPU ticks consumed at the servers. Specifically, in this paper we (i) consider a large-scale, crowd-sourced cellular demand trace, already used in big data applications [@scott]; (ii) perform an extensive set of experiments, linking the quantity of downloaded data in different conditions (type of service, number of users,...) to the amount of CPU power required at the servers; (iii) assess the impact of enriching our traces on the resulting planning of the MEC network. Importantly, we make [*publicly available*]{} realistic user demand traces, as well as our experimental dataset linking user demand to computational burden [@html].
We stress that, although MEC was our original motivation and we use it as a test case, the problem we address is more fundamental and consistent with the high-level goal of big data research, that is, to turn data into [*information*]{}. In our case, data come from real-world measurements, and the information we seek has to do with next-generation networks based on MEC. In spite of all their (perceived) abundance and the (apparent) easiness with which they are processed, real-world measurements can only deal with present-day systems and, thus, they are not naturally well-suited to study future ones. In this paper, we demonstrate how this limitation can be overcome with the help of additional data, real-world experiments, and domain knowledge.
The remainder of the paper is organized as follows. We begin by reviewing related work in [Sec. \[sec:relwork\]]{}. Then we present the real-world demand traces we use as a starting point in [Sec. \[sec:traces\]]{}, and describe the experiments we perform to enrich them in [Sec. \[sec:enrich\]]{}. [Sec. \[sec:design\]]{} introduces and discusses our MEC network design strategy. After presenting our numerical results in [Sec. \[sec:results\]]{}, we conclude the paper in [Sec. \[sec:conclusion\]]{}.
Related work {#sec:relwork}
============
Our study is connected to three main categories of prior work: papers presenting real-world mobile traces and datasets; studies addressing MEC in general and MEC-based 5G networks specifically; works doing the latter using the first.
Many real-world traces come from volunteers, such as the MIT Reality Project [@mit-reality] and the Nokia Mobile Challenge. These traces include a great deal of valuable information; their main shortcoming is the limited number of participants (in the case of the Nokia Mobile Challenge, around two hundred). This scale is adequate to study, for example, user mobility or encounter patterns, but studying a whole cellular network requires information about many more users. Mobile operators are typically reluctant to release demand and deployment information to the scientific community. An exception is represented by the Data For Development dataset by Orange [@orange-d4d], including mobility information for 50,000 users in Ivory Coast, as well as CDR (call-detail record) information for phone calls and SMS messages. However, the Orange trace only includes voice calls and SMS, and is severely restricted by heavy anonymization – each ID encountered gets a new coded identity for each “ego site” to which they are a neighbor. In other cases [@Malandrino-TCS], mobile operators have released demand or deployment information to individual research teams under non-disclosure agreements; however, these traces typically include only one operator and/or only one city.
MEC has been recently introduced [@etsi-wp] as a way to move “the cloud”, i.e., the servers processing mobile traffic, closer to the end users, thus reducing latency and traffic load across the network infrastructure. Network Function Virtualization (NFV) is widely regarded to as an enabling technology for MEC (see, e.g., [@etsi-wp]). Recent works have studied the radio techniques needed to enable MEC [@mec-radio], its relationship to the Internet-of-things [@mec-iot] and context-aware, next-generation networks [@mec-5g]. Closer to our scenario, the authors of [@moving] study how caches and servers should be placed in the network as its load changes over time. Regarding MEC and caching, a prominent application is mobile video streaming. As an example, [@multiop-caching; @cdn1] account for layered video coding techniques, and address the problem of placing the right layers at the right cache – with [@multiop-caching] also accounting for cooperation between operators. Other works [@proactive-caching; @wons] aim at [*foreseeing*]{} the content demand, in order to proactively populate caches or serve users.
5G will significantly exploit the MEC paradigm, and a substantial body of research is devoted to the problem of placing VNFs across the network providers’ servers. As an example,[@AHirwe16; @TKuo16; @ABaumgartner15; @FJemaa16; @BAddis15] tackle the problems of VNF placement and routing from a network-centric viewpoint, i.e., they aim at minimizing the load of network resources. In particular, [@AHirwe16] seeks to balance the load on links and servers, while [@TKuo16] studies how to optimize routing to minimize network utilization. The above approaches formulate mixed-integer linear programming (MILP) problems and propose heuristic strategies to solve them. [@ABaumgartner15], [@FJemaa16] and [@BAddis15] formulate ILP problems, respectively aiming at minimizing the cost of used links and network nodes, minimizing resource utilization subject to QoS requirements, and minimizing bitrate variations through the VNF graph.
Not many works however exist that combine real-world traces and multi-access edge computing. Among the most recent ones, [@noi-pof] studies the price (in terms of additional infrastructure) of deploying caches within the cellular core network. Compared to our work, [@noi-pof] only focuses on caching and vehicular traffic, and it only considers the dataset for the city of Los Angeles.
Our earlier work [@noi-workshop] sets in the same scenario as this paper: the traces introduced in [Sec. \[sec:traces\]]{} were used to study the trade-offs between network latency and server utilization in MEC network design. The results presented in [@noi-workshop] use the quantity of data downloaded by users as a proxy metric for computational capabilities required at the servers, and their limited applicability represents one of the main motivations behind this paper.
Input data {#sec:traces}
==========
WeFi, now acquired by TruConnect Technologies, is an Android application providing location-specific information on the speed, security, and reliability of nearby Wi-Fi networks. At the same time, it collected data on the activity of its users, including location, mobile phone activity, and available connectivity options. For our study we use three datasets, coming from the U.S. cities of Atlanta, Los Angeles, and San Francisco, characterized by the features summarized in [Tab. \[tab:datasets\]]{}.
[1]{}[|c|X|X|X|]{} & Atlanta & Los Angeles & San Francisco\
Time of collection & Oct. 2015 & Oct. 2015 & Mar. 2015\
Covered area \[km$^2$\] & $55\times 66$ & $46\times 73$ & $14\times 11$\
Total traffic \[TB\] & 9.34 & 35.61 & 9.18\
Number of records & 13 million & 81 million & 60 million\
Unique users & 9,203 & 64,386 & 14,018\
Unique cells & 12,615 & 36,09 & 14,728\
Datasets are organized in [*records*]{}, each containing: time and GPS location; anonymized user identifier; current operator and cell identifier; active application on the smartphone and amount of data it downloads. New records are generated every time any of the above changes (e.g., the user moves or switches apps), or a one-hour period elapses.
The WeFi datasets represent a real-world, [*live*]{} snapshot of both mobile networks and their users, and have three features that make them especially relevant to our study. First, they contain information on several cities, different for traffic demand profile and network deployment. Furthermore, they include multiple mobile operators, with different deployment strategies, e.g., usage of micro- and macro-cells. Finally, they allow us to know the individual application generating each traffic flow.
The WeFi data has been already used in the field of [*big data*]{}, e.g., in the preliminary analysis [@scott] and follow-up works. The purpose therein was to identify patterns of living, commuting, fast food consumption, work activities and recreation for tens of thousands of people. That case study confirms that the content and quality of the WeFi datasets are sufficient not only for networking studies, but also for social observations and government actions.
[**Dataset available.**]{} While we cannot share the dataset we use, we produced a synthetic trace with the same time and space characteristic of its original counterpart, available for download from [@html]. For more details on the generation of such synthetic traces, the interested reader is referred to [@noi-tmc17].
Enriching the mobile data traces {#sec:enrich}
================================
As mentioned in [Sec. \[sec:intro\]]{}, our high-level goal is to use the [*demand*]{} information available in the datasets to understand the [*deployment*]{} we need, i.e., how much computational capacity shall be placed within the network, and where. To this end, we perform three main steps:
1. we design and perform a set of experiments measuring the computational load associated with different types of traffic, as detailed in [Sec. \[sec:sub-exp\]]{};
2. we use the experimental data to train a [*model*]{} connecting the two quantities, as described in [Sec. \[sec:sub-learn\]]{};
3. we exploit the model to estimate the computational capacity needed to serve the traffic we observe in our dataset, as discussed in [Sec. \[sec:sub-enrich\]]{}.
Experimental setup {#sec:sub-exp}
------------------
We focus on three different, highly representative types of traffic, namely, video streaming, gaming, and maps. For reproducibility purposes, we restrict ourselves to open-source programs, namely:
- for video, the FFserver server and the VLC client;
- for gaming, the Minecraft server and the corresponding Minecraft Pocket mobile client;
- for maps, the OpenMapTiles server [@tiles-server], based on data from OpenStreetMap [@osm].
For our experiments, we use a testbed, composed of two computers – a client and a server – connected via a Wi-Fi link. Both the client and the server are commodity, off-the-shelf computers equipped with Intel Core i7-7700T processors and running Ubuntu Linux 16.10. On the client machine, we run the Genymotion Android emulator, which in turn emulates a varying number of mobile clients. The server machine hosts one instance of the server application, containerized within Docker.
In streaming tests, we use three videos with different duration, resolution, and file size; for gaming, we employ three different moving patterns throughout the Minecraft world; for maps, we generate requests for randomly-chosen $20\times 20$km$^2$-areas. In both cases, we vary the number of clients between 1 and 8, and use the FRep program to drive the emulator, thus guaranteeing uniform, reproducible patterns of UI actions, e.g., taps and swipes.
In all experiments, we measure two quantities: the amount of data generated by the server program, and the computational capacity it consumes. The former simply corresponds to the traffic outgoing from the virtual interface created by Docker; the latter is obtained by polling the file relating to the server process. The FFserver and Minecraft servers are single-threaded, so there is only one process to account for. The OpenMapTiles server, on the other hand, includes several processes, including a database and a web server; in that case, we present the aggregate CPU consumption figures.
Results and model {#sec:sub-learn}
-----------------
[Fig. \[fig:scatter-video\]]{}, [Fig. \[fig:scatter-gaming\]]{} and [Fig. \[fig:scatter-maps\]]{} summarize our experimental results. Each blue dot therein corresponds to an experiment, e.g., a different combination of video file and number of clients; its position in the plot is determined by how much data was served in that experiment, and the corresponding load on the server. Computational load values are expressed in [*CPU ticks*]{}, the minimum unit of CPU scheduling in the operating system. In Linux, each tick represents 10 ms; as an example, saying that a process takes 200 ticks to complete means that the CPU was assigned to that process for 2 s [^2].
The first thing to notice is the scale of the two plots. As one might expect, serving video implies transferring massively more data than hosting a game server, and therefore video is rightfully regarded as one of the applications consuming most of the bandwidth offered by networks. Looking at the y-axis, on the other hand, a different aspect emerges: the amount of CPU resources associated with gaming and, to a lesser extent, maps, dwarfs the one required by video.
This disproportion makes intuitive sense: while streaming a video requires little more than reading a file from disk and feeding it into a network stream, game and map servers have to perform complex operations such as keeping track of the game status or rendering the maps. However, its scale is somehow unexpected, and serves as a clear reminder that designing and deploying the [*computation*]{} part of a network, i.e., where and how to place its computational power, using only the traffic demand as a guideline is likely to result in poor performance and low efficiency.
The vastly different relationships between served traffic and CPU load are also evident from the linear fit we obtain from the two sets of experiments, represented by red lines in [Fig. \[fig:scatter\]]{}. The fitted relationships are as follows: $$\label{eq:model}
\text{CPU}_{[\text{ticks}]}=\begin{cases}
0.25\times\text{traffic}_{[\text{Mbit}]}+6.76 & \text{for video,} \\
161.38\times\text{traffic}_{[\text{Mbit}]}+1675.03 & \text{for gaming,} \\
67.44\times\text{traffic}_{[\text{Mbit}]}-7.53 & \text{for maps.} \\
\end{cases}$$ Looking at [Fig. \[fig:scatter\]]{}, it also important to observe how the fit is much better for maps traffic than for gaming: the root mean square error (RMSE) for maps is 2% while it is equal to 6% for gaming. The reason is that the quantity of CPU required by gaming applications significantly depends on such factors as the actions performed by the players – a fact that leads to a larger variability.
The fact that all fitted relationships are linear is not especially relevant [*per se*]{}, nor particularly surprising; indeed, it makes intuitive sense that the amount of work needed to produce a certain type of data grows, more or less linearly, with the quantity of data to produce. What matters the most is the [*slope*]{} of the fitted lines, which changes by several orders of magnitude from one traffic category to another.
In the following, we will indicate with $\tau_k$ the number of CPU ticks needed to process one megabyte of traffic of category $k$; for instance, $\tau_\text{video}=0.25$ CPU ticks/Mbyte (i.e., the slope of the red line in [Fig. \[fig:scatter-video\]]{}), $\tau_\text{gaming}=161.38$ ticks/Mbyte (i.e., the slope of the red line in [Fig. \[fig:scatter-gaming\]]{}), and $\tau_\text{maps}=67.44$ ticks/Mbyte (i.e., the slope of the red line in [Fig. \[fig:scatter-maps\]]{}).
Enriching the dataset {#sec:sub-enrich}
---------------------
As a preliminary step, we need to decide, for each app we observe in the dataset, whether its traffic can be considered video-like, gaming-like, or map-like (or none of them). We perform such an assignment as follows:
- traffic coming from YouTube, Netflix, TimeWarner, ShowBox, Twitch, DirectTV, FoxSports, FoxNews, is tagged as video-like;
- traffic coming from Minecraft, World of Warcraft, Riptide, Grand Theft Auto, Rollercoaster Tycoon, This War of Mine, Titan Quest, Unkilled, is tagged as game-like;
- traffic coming from Google Maps and Waze is tagged as map-like.
As summarized in [Fig. \[fig:pies\]]{}(left), about 66% of the traffic we observe in our dataset can be classified as video-like, about 15% of it as game-like, while the amount of map-like traffic is much smaller. Much of the traffic that cannot be tagged comes from social networking applications, whose behavior cannot be easily studied due to the lack of open-source social network servers.
Once the assignment is made, we can use the relationship in [(\[eq:model\])]{}, learned in [Sec. \[sec:sub-learn\]]{}, to add a new column to our dataset. The column, called , expresses how much computational power is required to generate the traffic reported in each line of the dataset. Notice that relationships other than linear can be accounted for at no additional complexity.
[Fig. \[fig:pies\]]{} further highlights the gap between the quantity of traffic generated by different applications and the corresponding CPU load. Gaming and maps represent a small fraction of the total traffic ([Fig. \[fig:pies\]]{}(left)); however, by applying [(\[eq:model\])]{}, we obtain the results shown in [Fig. \[fig:pies\]]{}(right): gaming applications consume the vast majority, over 90%, of all computational resources. This further highlights the importance of accounting for the computational load of different types of traffic in network design.
![ Categories traffic demand belongs to (left); amount of computational power required to serve video-, and gaming- and maps-like traffic (right). \[fig:pies\] ](img/pie_traffic "fig:"){width=".23\textwidth"} ![ Categories traffic demand belongs to (left); amount of computational power required to serve video-, and gaming- and maps-like traffic (right). \[fig:pies\] ](img/pie_ticks "fig:"){width=".23\textwidth"}
Discussion {#sec:sub-discuss}
----------
Our trace enrichment strategy has some limitations: it does not account for all the traffic types, and might not perfectly model the behavior of all servers.
The first issue is evident from [Fig. \[fig:pies\]]{}(left): around 25% of all traffic cannot be classified as either video-like, gaming-like, or maps-like. On the other hand, our analysis is able to account for over 70% of all present-day traffic, and video, gaming and maps are the applications that are expected to grow the most in the near future.
The second issue has to do with the fact that we model the behavior of YouTube and Netflix using FFserver, of World of Warcraft through Minecraft, or of Google Maps using OpenStreetMap. On the one hand, this might sound a bit bold. On the other hand, it is true that the problems faced by different applications of the same type – and the solutions thereto – tend to overlap. This explains, as an example, the existence of gaming [*engines*]{} such as Unity or Microsoft XNA, providing the developers of a heterogeneous set of games with homogeneous solutions to a small set of common problems.
In summary, our analysis provides valuable guidelines highlighting the existence of a mismatch between the quantity of traffic and the corresponding CPU load, as well as its magnitude. Additionally, the methodology we employ is general, and works unmodified in cases where additional traffic types and/or applications are taken into account, or if larger-scale experiments can be performed.
[**Dataset available.**]{} The results of our experiments are available for download from [@html].
MEC design {#sec:design}
==========
We now show how the enhanced dataset we created can be used to devise a MEC design strategy. Our input data are represented by:
- a set of base stations;
- the expected/predicted traffic at each of them and its characteristics;
- the network connectivity.
Given the above, we have to dimension the network computational capabilities, i.e., to decide (i) where to place the MEC servers, and (ii) the capacity they should have.
We solve this problem accounting also for the specific applications we deal with, thus developing an [*application-aware deployment*]{}, accommodating the requirements of each application. Indeed, such requirements can be substantially different for different [*categories*]{} of applications. As an example, we might be willing to serve such (comparatively) delay-tolerant content as video-like and map-like traffic through a server located in the core network, while real-time services like mobile gaming will require their servers to be much closer to the end users, possibly at individual base stations. It is important to stress that, in pure NFV/MEC fashion, we allow the same servers to serve multiple applications concurrently, provided that the server capacity is not exceeded.
At last, note that our input data is represented by the traffic that is [*actually*]{} served by the base stations. Consistently with MEC design best practices [@mec-survey; @mec-ton], we do not need to explicitly account for access-network issues such as congestion and interference. Indeed, the coverage quality experienced by users influences the amount of data they are able to upload/download, and this aspect is captured by our system model and MEC design.
Network topology
----------------
The identity and positions of base stations $b\in{\mathcal{B}}$, as well as their demand $\delta(b,k,t)$ are readily available from the trace we describe in [Sec. \[sec:traces\]]{}. However, we have no information about the topology of the backhaul cellular network. Indeed, mobile operators are extremely reluctant to disclose this information, and virtually all works in the literature resort to synthetic topologies based on current best practices. Following [@softcell], we assume a [*fat-tree*]{} topology, where:
- base stations are grouped into [*rings*]{} of ten;
- every ten rings, there is an aggregation-level [*pod*]{};
- every ten pods, there is a core-level switch.
The topology has a fan-out of $2$, i.e., we connect every ring to the two closest pods, and every pod to the two closest core switches, while switches themselves are connected in a full mesh. The network topology we generate can be represented as a graph ${\mathcal{G}}=({\mathcal{N}},{\mathcal{E}})$.
We also generate a DAG ${\mathcal{D}}=({\mathcal{N}},{\mathcal{F}})$, to keep track of which nodes can serve each base station. In particular, we generate a (directed) edge $(n_1,n_2)\in{\mathcal{F}}$ between nodes such that:
- $n_1$ and $n_2$ are connected, i.e., $(n_1,n_2)\in{\mathcal{E}}$, and
- $n_1$ belongs to a network level immediately higher to the one of $n_2$, e.g., $n_1$ is a ring and $n_2$ is a base station.
We will say that node $n_1$ is a [*parent*]{} of $n_2$ if $(n_1,n_2)\in{\mathcal{F}}$. Base stations have no children, i.e., they are leaves in the DAG, while core switches have no parents, i.e., they are roots.
[**Notation.**]{} We denote by $n\in{\mathcal{N}}$ all the nodes of the cellular network, from base stations to core-level switches, and by $b\in{\mathcal{B}}\subset{\mathcal{N}}$ the base stations among them. The physical distance between any two nodes $n_1$ and $n_2$ is denoted by $d(n_1,n_2)$. Contents are specific to [*category*]{} $k\in{\mathcal{K}}$, e.g., video, gaming, or maps, and time is discretized into steps $t\in{\mathcal{T}}$. We denote by $\delta(b,k,t)$ the demand from users covered by base station $b$, for contents of category $k$ and during step $t$.
MEC design {#mec-design}
----------
Designing our MEC network means making two decisions, each corresponding to a binary variable:
- whether we should place a server at node $n\in{\mathcal{N}}$, expressed through variable $y(n)\in\{0,1\}$;
- whether the traffic coming from base station $b\in{\mathcal{B}}$ shall be served by the server placed at node $n\in{\mathcal{N}}$, expressed through variable $x(b,n)\in\{0,1\}$.
Note that none of the decision variables depends on the time step $t\in{\mathcal{T}}$; this reflects the fact that deployment decisions are made periodically, with a time period much longer than a single time step, accounting for the evolution of data demand over the previous period. Needless to say, we cannot serve anything on servers that do not exist, i.e., it must be: $$\nonumber
x(b,n)\leq y(n),\quad\forall b\in{\mathcal{B}},n\in{\mathcal{N}}.$$
The objective of the problem can be stated as deploying the smallest possible number of servers subject to delay constraints, i.e., $$\nonumber
\min_{x,y}\sum_{n\in{\mathcal{N}}}y(n).$$
\[line:input\] \[line:init\] \[line:while\] \[line:pairs1\] \[line:pairs2\] \[line:mindist\] \[line:call\]
\[line:consolidate1-begin\] \[line:consolidate1-end\] \[line:consolidate2-begin\] \[line:consolidate2-n3\] \[line:consolidate2-end\]
Optimally setting the binary $x$- and $y$-variables subject to constraints on the need to serve all traffic requires solving an ILP problem[^3], which is notoriously [@milp] impractical even for modestly-sized problem instances. We therefore devise a greedy design procedure, able to efficiently make good-quality deployment decisions.
### Greedy design procedure {#sec:algos}
Our greedy design procedure is inspired to hierarchical clustering and summarized in [Alg. \[alg:design\]]{}. It takes as input ([Line \[line:input\]]{}) the content category $k$ to consider, its maximum processing latency $L_{\max}$, its demand over time $\delta(b,k,t)$, and the sets ${\mathcal{B}}$ and ${\mathcal{N}}$ of, respectively, base stations and network nodes. This implies that [Alg. \[alg:design\]]{} makes [*per-category*]{} decisions, and is thus able to account for the fact that different categories of content, with different latency limits $L_{\max}$, might require different deployment strategies. Notice that with [*latency*]{} we indicate the time spent within the core network, which is itself a component of the total, end-to-end service time.
The algorithm starts ([Line \[line:init\]]{}) from a solution where each base station that ever serves at least one user requiring content category $k$ (i.e., $\max_{t\in{\mathcal{T}}}\delta(b,k,t)>0$) has its own server, i.e., $y(b)=1$. Then, as long as the latency resulting from our deployment does not exceed the threshold $L_{\max}$ ([Line \[line:while\]]{}), we select a pair of nodes $(n_1,n_2)$ to [*consolidate*]{} together, reducing the number of servers at the price of potentially increasing the service latency.
\[m\]\[c\][$n_1$]{} \[m\]\[c\][$n_2$]{} \[m\]\[c\][$n_3$]{} ![ The consolidation procedure. Circles represent nodes in $n\in{\mathcal{N}}$; black ones correspond to nodes that host a server, i.e., $y(n)=1$, white ones to nodes that do not, i.e., $y(n)=0$. If $n_1$ is $n_2$’s parent (top), the server at $n_2$ is removed and all base stations it served, are served by $n_2$’s parent $n_1$. If $n_1$ and $n_2$ are siblings (bottom), the servers at $n_1$ and $n_2$ are removed, and a new one is created at their parent $n_3$. \[fig:consolidate\] ](img/consolidate "fig:"){width=".22\textwidth"}
The consolidation procedure is depicted in [Fig. \[fig:consolidate\]]{}. It involves two nodes $n_1$ and $n_2$ such that either $n_2$ is $n_1$’s parent, or $n_1$ and $n_2$ are siblings, i.e., have a common parent $n_3$. In both cases, the server(s) at the child(ren) are removed, and all base stations they used to serve are served by the parent. Every time we perform the consolidation procedure, the number of servers deployed in the topology decreases by one unit, at the cost of potentially increasing the latency.
In [Line \[line:pairs1\]]{}–[Line \[line:pairs2\]]{} of [Alg. \[alg:design\]]{}, we construct a set ${\mathcal{P}}$ of pairs of nodes eligible for consolidation, i.e., such that (i) they both have a server, and either (ii) $n_1$ is $n_2$’s parent ([Line \[line:pairs1\]]{}), or (iii) both $n_1$ and $n_2$ have a common parent $n_3$. In [Line \[line:mindist\]]{}, we select the nodes $n_1^\star$ and $n_2^\star$ with the highest score, i.e., the most suitable to consolidate, and call the [**consolidate**]{} procedure with those nodes as an argument. As discussed in [Sec. \[sec:dist\]]{}, different definitions of score can be considered, leading to different deployment strategies.
[Alg. \[alg:consolidate\]]{} details the consolidation procedure, and takes as input the nodes $n_1$ and $n_2$ to consolidate. If $n_2$ is $n_1$’s parent, i.e., we are in the situation of [Fig. \[fig:consolidate\]]{}(top), the server at $n_2$ is removed and any base stations that were served by $n_2$ are served by $n_1$ ([Line \[line:consolidate1-begin\]]{}–[Line \[line:consolidate1-end\]]{} of [Alg. \[alg:consolidate\]]{}). If $n_1$ and $n_2$ are siblings, i.e., we are in the situation of [Fig. \[fig:consolidate\]]{}(bottom), then we first identify the node $n_3$ that is a parent to both $n_1$ and $n_2$ ([Line \[line:consolidate2-n3\]]{}). Afterwards, servers at both $n_1$ and $n_2$ are removed, a new server at $n_3$ is created, and all base stations that were served by $n_1$ or $n_2$ are served by $n_3$ ([Line \[line:consolidate2-begin\]]{}–[Line \[line:consolidate2-end\]]{} of [Alg. \[alg:consolidate\]]{}).
### Score definitions {#sec:dist}
A key feature of our approach is that it can support multiple deployment [*strategies*]{} through the one [*algorithm*]{} [Alg. \[alg:design\]]{}. We are able to do so by considering multiple definitions of the score associated with consolidating two network nodes, i.e., the implementation of function called in [Line \[line:mindist\]]{}. Recall that we are still focusing on a single content category $k$.
[**Location-based**]{} It is often desirable to consolidate nodes that are physically close to each other, as this typically translates into a shorter travel time between the users and the servers serving them. This corresponds to giving higher scores to pairs of nodes that are close to each other, i.e., $$\label{eq:dist-dist}
\texttt{score}(n_1,n_2)=-d(n_1,n_2).$$
[**Load-based**]{} An alternative deployment strategy takes into account the load of each server, and tries to avoid consolidating nodes whose demands have a similar time evolution, i.e., whose peak times tend to overlap. The rationale is that by doing so we can decrease the capacity requirements for the consolidated servers, which depend upon the [*peak*]{} of the combined load. More formally, let us define a serve vector $\vec{s}(n)$ for each node $n$. $\vec{s}(n)$ vectors have $|{\mathcal{T}}|$ elements, and each element $s(n)_{t}$ represents the total demand for contents of category $k$ by users at base stations that are served by node $n$ during time step $t$: $$\nonumber
s(n)_t=\sum_{b\in{\mathcal{B}}}x(b,n)\delta(b,k,t).$$ Given the $\vec{s}(n)$ values, we can define the score related to the $(n_1,n_2)$ pair as $$\begin{gathered}
\label{eq:dist-load}
\texttt{score}(n_1,n_2)=\max_{t\in{\mathcal{T}}} \tau_k s(n_1)_t+\max_{t\in{\mathcal{T}}} \tau_k s(n_2)_t+\\
-\max_{t\in{\mathcal{T}}}\tau_k\left (s(n_1)_t+s(n_2)_t\right).\end{gathered}$$ where the first two terms of the second member of [(\[eq:dist-load\])]{} represent the peak loads of nodes $n_1$ and $n_2$ before consolidation; the third term is the peak load of the combined server, after consolidation. A high-scoring consolidation operation will involve nodes with high peak loads (first two terms, with positive sign) that can be combined into a new, low-load server (third term, with negative sign).
Recall that the factor $\tau_k$ in [(\[eq:dist-load\])]{} expresses how many units of computational power (e.g., CPU ticks) are needed to generate one unit of traffic (e.g., one megabyte) of category $k$, as obtained in [Sec. \[sec:sub-enrich\]]{}. Using non-enriched traces corresponds to assuming $\tau_k=1,\,\forall k\in{\mathcal{K}}$.
![\[fig:fillnsrv-loc\]Location-based scores: number of servers (left) and amount of traffic processed (right) at the different network levels, for each iteration of [Alg. \[alg:design\]]{}. ](img/fillnsrv_mno3_location_mbit_nodelay "fig:"){width=".24\textwidth"} ![\[fig:fillnsrv-loc\]Location-based scores: number of servers (left) and amount of traffic processed (right) at the different network levels, for each iteration of [Alg. \[alg:design\]]{}. ](img/fillmbit_mno3_location_mbit_nodelay "fig:"){width=".24\textwidth"}
![\[fig:fillnsrv-dem\]Load-based scores: number of servers (left) and amount of traffic processed (right) at the different network levels, for each iteration of [Alg. \[alg:design\]]{}. ](img/fillnsrv_mno3_demand_mbit_nodelay "fig:"){width=".24\textwidth"} ![\[fig:fillnsrv-dem\]Load-based scores: number of servers (left) and amount of traffic processed (right) at the different network levels, for each iteration of [Alg. \[alg:design\]]{}. ](img/fillmbit_mno3_demand_mbit_nodelay "fig:"){width=".24\textwidth"}
### Multiple categories {#sec:multi}
As mentioned in [Sec. \[sec:algos\]]{}, both [Alg. \[alg:design\]]{} and [Alg. \[alg:consolidate\]]{} make decisions on a [*per-category*]{} basis, and are therefore able to reproduce the fact that different applications can require different deployments. In the following, we consider the presence of different application categories and describe how our approach can be easily leveraged to realize an application-aware deployment. We proceed as follows:
- we divide application into categories;
- we run our greedy deployment procedure separately for each category;
- we combine the resulting deployments.
Running [Alg. \[alg:design\]]{} separately for different categories also means evaluating [Line \[line:while\]]{} therein using the maximum latency values $L_{\max}$ of each category. Similarly, the demand-aware distance [(\[eq:dist-load\])]{} is computed separately for each category, only accounting for the contents falling in the current category.
Numerical results {#sec:results}
=================
In this section, we investigate how deployment strategies and score definitions ([Sec. \[sec:sub-strategy\]]{}), using enriched traces ([Sec. \[sec:sub-ticks\]]{}) and considering per-category latency limits ([Sec. \[sec:sub-latency\]]{}), impact the resulting MEC deployment and its effectiveness. For sake of brevity, we only present results for one of the three operators included in our trace. High-resolution versions of the plots for all operators, as well as the Matplotlib source code to generate them, are available from [@html].
Effect of the deployment strategy {#sec:sub-strategy}
---------------------------------
The first aspect we are interested in is the impact of the score definition we adopt, i.e., whether we use [(\[eq:dist-dist\])]{} or [(\[eq:dist-load\])]{} to implement the function in [Alg. \[alg:design\]]{}. To this end, we first assume no latency limit, i.e., $L_{\max}=\infty$, and study (i) how much traffic is processed at each level of the network – base station (BS), ring, aggregation, core – and (ii) how many servers are deployed therein.
[Fig. \[fig:fillnsrv-loc\]]{} and [Fig. \[fig:fillnsrv-dem\]]{} demonstrate how [Alg. \[alg:design\]]{} and the consolidation procedure work in the case of location-based and load-based scores, respectively. We start at iteration $0$ with one server at each BS; then, at each iteration, we reduce the total number of servers through consolidation, replacing BS servers with servers placed at the higher levels of the network topology. We can observe that, while [Fig. \[fig:fillnsrv-loc\]]{}(left) shows a significant number of servers at ring and aggregation nodes, [Fig. \[fig:fillnsrv-dem\]]{}(right) shows a tendency to process traffic either at BSs, or at very few, core-level servers. This is confirmed by [Fig. \[fig:fillnsrv-loc\]]{}(right) and [Fig. \[fig:fillnsrv-dem\]]{}(right): with location-based scores, servers are [*lazily*]{} moved from BSs to rings, and then further up, only when required. On the other hand, using load-based scores we aggressively process as much data as possible as high in the topology as possible, so as to smoothen the peaks.
Efficiency and latency: the importance of ticks {#sec:sub-ticks}
-----------------------------------------------
As we have seen, different deployment strategies (i.e., different scores) result in deeply different network planning decisions. In the following, we study (i) how such decisions impact the latency and efficiency of the resulting network, and (ii) if enriching the trace as discussed in [Sec. \[sec:enrich\]]{} impacts either metric.
We estimate the [*latency*]{} using [@ref-delay] as a reference: the connection between users (UEs in LTE terminology) and base stations (eNBs) requires 5 ms, while additional hops within the backhaul network (e.g., from BSs to rings) require about 2.3 ms each.
As for [*efficiency*]{}, we define it as the ratio between the average capacity that is actually required to process the traffic and the total capacity deployed throughout the network: $$\label{eq:eff}
\eta=\frac{
\frac{1}{|{\mathcal{T}}|}\sum_{t\in{\mathcal{T}}}\sum_{n\in{\mathcal{N}}}\sum_{k\in{\mathcal{K}}}\tau_k s(n)^k_t
}{
\sum_{n\in{\mathcal{N}}}\max_{t\in{\mathcal{T}}}\sum_{k\in{\mathcal{K}}} \tau_k s(n)^k_t
}.$$ A network where all servers are always fully utilized would have an efficiency of $1$, while a network where servers are almost never used would have almost-zero efficiency. Also notice that in [(\[eq:eff\])]{} we consider all content categories, and abuse the notation to indicate with $s(n)^k_t$ the amount of data of content of category $k$ that node $n$ must process at time $t$.
[Fig. \[fig:delay\]]{} shows that location-based scores are associated with a lower latency than load-based ones. On the other hand, as shown in [Fig. \[fig:eff\]]{}, load-based scores maximize the network efficiency – which, however, hardly exceeds 30%, due to time fluctuations of the demand. It is interesting to compare the solid and dashed lines in [Fig. \[fig:delay\]]{} and [Fig. \[fig:eff\]]{}, respectively obtained with the original trace and the enriched trace. Using the enriched trace, i.e., ticks instead of megabytes, provides both a lower latency [*and*]{} a higher efficiency. This is confirmed by [Fig. \[fig:delayeff\]]{}, depicting the latency/efficiency trade-offs that can be obtained using different scores and traces.
The difference between location- and load-based performance in [Fig. \[fig:delay\]]{}–[Fig. \[fig:delayeff\]]{} can be interpreted as the benefit obtained by using the enriched trace [*in lieu*]{} of the original one. By doing so, we can make better deployment decisions, which consistently result in better efficiency and lower latency.
The effect of latency limits {#sec:sub-latency}
----------------------------
We now consider finite latency limits; specifically, we set $L_{\max}=50$ ms for video and maps and, based on [@ref-maxdelay], $L_{\max}=10$ ms for gaming traffic. [Fig. \[fig:delay-split\]]{}, obtained with load-based scores, shows that different content categories now experience different latencies. In particular, gaming data are always served within its maximum latency limit $L_{\max}=10$ ms, while other content is served with a higher latency – but still lower than its maximum limit. This is due to the different locations within the network where content is processed, as shown in [Fig. \[fig:filled-split\]]{}. In the late iterations of the algorithm, we can observe that almost all gaming content (represented by patterned areas) is served at aggregation nodes, while almost all non-gaming content (solid areas) is served at core nodes. Indeed, based on [@ref-delay], going from UEs to core nodes entails a latency of 12 ms, which is not compatible with the latency limit for gaming. Furthermore, it is important to note that, even if latency requirements for video and maps would permit the deployment of their servers in the cloud, MEC is still an appealing solution as it avoids transferring large amounts of data over long distances, thus reducing bandwidth consumption.
Being unable to serve gaming content at core servers also has the potential to impair network efficiency; indeed, as it can be seen from [Fig. \[fig:eff\]]{} and [Fig. \[fig:delayeff\]]{}, the highest efficiency is reached when most content is processed there. By comparing [Fig. \[fig:eff-split\]]{} to [Fig. \[fig:eff\]]{}, we can indeed observe a decrease in the efficiency; however, such a decrease is lower than 10%, which confirms that MEC is able to deliver both low latency and high efficiency.
Finally, it is interesting to observe that the advantage of using enriched traces [*in lieu*]{} of the original is clearly visible, in terms of both latency and efficiency.
Conclusion and future work {#sec:conclusion}
==========================
One of the challenges we face in the design of next-generation networks is using real-world data about [*present-day*]{} technologies to predict their performance. In this work, we took cellular networks as a case study, and demonstrated how a real-world, large-scale dataset about data demand in LTE can be leveraged to design a next-generation, MEC-based network.
To this end, we [*enriched*]{} the dataset at our disposal, integrating it with information about the processing power needed to generate traffic of the two main categories of content. We obtained such information through a set of hands-on experiments, based on the Minecraft and FFserver open-source servers. Our results show that using the enriched trace instead of the original one results in better network design, allowing both lower latency [*and*]{} higher network efficiency.
[Francesco Malandrino]{} (M’15) earned his Ph.D. in 2012 from Politecnico di Torino, Italy, where he is currently an assistant professor. Before his current appointment, he held short-term positions at Trinity College, Dublin, and at the Hebrew University of Jerusalem as a Fibonacci Fellow. His interests focus on wireless and vehicular networks and infrastructure management.
[Carla-Fabiana Chiasserini]{} (M’98, SM’09, F’18) graduated in Electrical Engineering (summa cum laude) from the University of Florence in 1996. She received her Ph.D. from Politecnico di Torino, Italy, in 2000. She has worked as a visiting researcher at UCSD in 1998–2003, and she is currently an Associate Professor with the Department of Electronic Engineering and Telecommunications at Politecnico di Torino. Her research interests include architectures, protocols, and performance analysis of wireless networks. Dr. Chiasserini has published over 230 papers in prestigious journals and leading international conferences, and she serves as Associated Editor of several journals.
[Giuseppe Avino]{} obtained his master degree in Computer and Communication Network Engineering in 2016 from Politecnico di Torino. From January to October 2017, he carried out research on container virtualization at Politecnico di Torino, where he started his Ph.D. in November 2017. His interests focus on wireless network communication and container virtualization.
[Marco Malinverno]{} is a Ph.D. student at Politecnico di Torino. He obtained his master degree in Computer and Communication Networks Engineering in 2016. His research activities range from containerization techniques to vehicular wireless networks.
[Scott Kirkpatrick]{} (SM ’80, F ’91) has been a professor in the computer science department at the Hebrew University, in Jerusalem, Israel, since 2000. Before that he was at IBM Research, Yorktown Heights, where he did research in physics and engineering, developing simulated annealing and IBM’s first tablet computers. Professor Kirkpatrick is a Fellow of the APS, the AAAS, the IEEE, and the ACM, has written over 100 papers and holds more than 10 patents. He holds an AB from Princeton University and PhD (in physics) from Harvard University.
[^1]: This work is supported by the European Commission through the H2020 projects 5G-TRANSFORMER (Project ID 761536) and 5G-EVE (Project ID 815074).
[^2]: The real execution time might be longer, e.g., if the process is preempted by a higher-priority one during its execution.
[^3]: We skip the proof, based on a reduction from the SAT problem, in the interest of space.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We introduce a framework for ${\ensuremath{\mathbb{Z}}}$-gradings on cluster algebras (and their quantum analogues) that are compatible with mutation. To do this, one chooses the degrees of the (quantum) cluster variables in an initial seed subject to a compatibility with the initial exchange matrix, and then one extends this to all cluster variables by mutation. The resulting grading has the property that every (quantum) cluster variable is homogeneous.
In the quantum setting, we use this grading framework to give a construction that behaves somewhat like twisting, in that it produces a new quantum cluster algebra with the same cluster combinatorics but with different quasi-commutation relations between the cluster variables.
We apply these results to show that the quantum Grassmannians ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ admit quantum cluster algebra structures, as quantizations of the cluster algebra structures on the classical Grassmannian coordinate ring found by Scott. This is done by lifting the quantum cluster algebra structure on quantum matrices due to Geiß–Leclerc–Schröer and completes earlier work of the authors on the finite-type cases.
author:
- |
Jan E. Grabowski\
\
- |
Stéphane Launois\
\
bibliography:
- 'biblio.bib'
date: 10th January 2013
title: |
Graded quantum cluster algebras\
and an application to quantum Grassmannians
---
Introduction
============
Since cluster algebras were introduced by Fomin and Zelevinsky ([@FZ-CA1]), it has been recognised that cluster algebra structures on homogeneous coordinate rings on Grassmannians are among the most important classes of examples. The demonstration of the existence of such a structure is due to Scott [@Scott-Grassmannians] and one reason for the importance of this is that these cluster structures are typically of infinite mutation type but have combinatorics under tight control, because they are a realisation of certain aspects of Grassmannian combinatorics. We note for example that Fomin and Pylyavskyy ([@FominPylyavskyy]) have recently advocated further study of Grassmannian cluster structures for precisely these reasons.
Among those who study quantized coordinate rings, it is also widely acknowledged that Grassmannians have a special place. Again, the intricate geometric structures associated to Grassmannians, due in part to their Lie-theoretic origins, give a rich structure of their quantized coordinate rings, the quantum Grassmannians ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$. Linking these two points of view, it has long been expected that quantum Grassmannians should possess quantum cluster algebra structures, the definition of the latter being due to Berenstein and Zelevinsky ([@BZ-QCA]). In earlier work, the present authors showed that this is the case when the cluster structure was expected to be of finite type, namely the cases ${\mathbb{K}_{q}[\mathrm{Gr}(2,n)]}$ and ${\mathbb{K}_{q}[\mathrm{Gr}(3,n)]}$ for $n=6,7,8$. However a general proof was not given at that time: one aim of this paper was to give such a general proof and this is achieved in Theorem \[t:quotisotoGr\].
In the course of attempting to generalise our earlier work on quantum cluster algebra structures, it became apparent that new techniques would be required to handle the general case. The main tools needed were ways to transfer quantum cluster algebra structures between related algebras. In the commutative setting, many of these operations are easy to carry out but the noncommutative situation is considerably more delicate, as one must be sure not to destroy the property of quantum clusters consisting of pairwise quasi-commuting elements. (That is, variables in the same quantum cluster should commute up to a power of the deformation parameter $q$.)
Examination of these problems showed that the correct way to keep control of this problem in the quantum setting is, as often in quantum groups, to introduce gradings. While the definition and constructions here were originally formed with a view to solving the quantum Grassmannian problem, we believe that the framework we introduce here should be of significance to researchers interested in cluster algebras more generally. In particular, this framework applies to classical commutative cluster algebras as well as their quantum analogues and yields in a natural way statements about the homogeneity of (quantum) cluster variables in a variety of settings. We remark that these results sit among a surprisingly small number that deal with cluster algebras and properties of their cluster variables in infinite types alongside finite types. The definition of a graded cluster algebra also sits separately from the categorical setting (i.e. cluster categories and related constructions), though—as here—it is fully compatible with categorification and indeed the two can illuminate each other.
We therefore devote the first two main sections of this work to graded (quantum) cluster algebras, following some essential recollections. The first gives the definition of a graded cluster algebra: it is obtained from an initial seed by adding an additional piece of data, an assignment of integer degrees to each initial cluster variable, subject to a compatibility condition with the initial exchange matrix. (This compatibility is only with the exchange matrix, which is why the notion immediately extends to the quantum setting.) This initial data is propagated to the whole cluster algebra by mutation, the key point being that we can mutate the grading data in a natural way. An immediate consequence of the definition is that every (quantum) cluster variable is homogeneous for the resulting algebra grading. We note that Berenstein and Zelevinsky have introduced a similar notion of grading in [@BZ-QCA]. However this was not an algebra grading but only a module grading. Their definition helped inspire ours but the two are different.
We then give a number of constructions that use the grading to alter a given quantum cluster algebra structure; some of these constructions are trivial for commutative cluster algebras. For example, one may re-scale every initial quantum cluster variable for a graded quantum cluster algebra by $q^{1/2}$ to the power of its degree and obtain an isomorphic quantum cluster algebra. We also show how to naturally extend a quantum cluster algebra to a skew-Laurent extension of that algebra. Furthermore we can combine these ideas to “re-scale” a quantum cluster algebra structure using a skew-Laurent extension, namely Theorem \[t:rescaledQCA\]. This theorem is a twisting-like result, in that the new quantum cluster algebra structure so obtained has the same cluster combinatorics but different quasi-commutation relations. The existence of gradings is key for that result and in turn Theorem \[t:rescaledQCA\] is key to the application we describe below, to the quantum Grassmannian.
The problem of lifting the cluster algebra structure on the coordinate ring of the big cell of a partial flag variety has been solved by Geiß, Leclerc and Schröer in [@GLS-PFV §10]. The approach is straightforward: one can view the coordinate ring of the cell as the quotient of the coordinate ring of the partial flag variety by certain elements (minors), and this quotient allows one to lift certain distinguished elements from the quotient to homogeneous elements in a minimal (and hence unique) way. Then [@GLS-PFV Proposition 10.1] shows that every cluster variable has the required property and hence the lifting procedure for the whole cluster algebra structure is possible. Recalling that the cluster algebra structure on the coordinate ring of the cell is obtained categorically, one notes that the appropriate data for the lifting is also encoded categorically, giving rise to a hope that this may be used in the quantum setting also.
However it is not possible to follow the strategy of [@GLS-PFV] directly in the quantum setting. It is well-known that in the noncommutative setting, factors of rings by normal but not central elements can be “too small” and so the appropriate construction is localisation, in the form of noncommutative dehomogenisation (see [@KellyLenaganRigal §3], for example). Therefore our methods are necessarily different to, and indeed more technically complicated than, the approach of [@GLS-PFV].
This necessitated the introduction of the graded methods described above. As an application of these, we are able to give a noncommutative version of the lifting of Geiß–Leclerc–Schröer and prove that the quantum Grassmannians ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ admit graded quantum cluster algebra structures. The main prior results used are the existence of a quantum cluster algebra structure on ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$, as shown by Geiß–Leclerc–Schröer ([@GLS-QuantumPFV]), and the dehomogenisation isomorphism, due originally to Kelly–Lenagan–Rigal ([@KellyLenaganRigal]), though we use a version given by Lenagan–Russell ([@LenaganRussell]). The dehomogenisation isomorphism makes the key link between quantum matrices and the quantum Grassmannian and is the noncommutative expression of the former as the quantum coordinate ring of the big cell of the latter. That is, it is an isomorphism $$\alpha\colon {\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}[Y^{\pm 1}; \sigma] \to {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}\!\left[ [12\cdots k]^{-1} \right]$$ between a certain skew-Laurent extension of the quantum matrices and a certain localisation of the quantum Grassmannian. The main goal is to transfer a quantum cluster algebra structure through this isomorphism $\alpha$ and show that it can be lifted from the localisation to the quantum Grassmannian ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ itself.
As this construction is rather technical, for the benefit of the reader we give a detailed breakdown of the structure of the proof, as follows:
1. Analysis of the quantum cluster algebra structure on ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ given by [@GLS-QuantumPFV]. (§\[s:KqMatisQCA\])
1. Proof of the existence of a *graded* quantum cluster algebra structure on ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$.
------------------------------------------------------------------------
(Lemma \[lemma:degsumsequal\])
2. Observation of the existence of an “almost-grading” $\theta$, arising from the categorification of ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$. (p. –)
2. Identification of the image of the quantum cluster algebra structure on a skew-Laurent extension of ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ under the dehomogenisation isomorphism $\alpha$ described above, giving a quantum cluster algebra structure on a localisation of ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$. (§\[s:dehomog\])
1. \[B1\] Identification of the images of quantum cluster variables. (Theorem \[t:thetaisscalingpower\])
2. \[B2\] In particular, identification of the images of quantum minors. (Corollary \[c:PluckersAreQCVs\])
3. “Re-homogenisation” by transferring of the quantum cluster algebra structure from the localisation to the unlocalised algebra. (§\[s:QCAonKqGr\])
1. Alteration of the quantum cluster algebra structure on the localisation of ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ such that the “almost-grading” $\theta$ becomes an honest grading. (p. –)
2. Use of Theorem \[t:rescaledQCA\] to produce a new quantum cluster algebra with the same cluster combinatorics but whose quasi-commuting relations now match those of the quantum Grassmannian. The quantum cluster variables of the new algebra are shown to be products of elements of ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ multiplied by a power of a certain central element (that power being controlled by $\theta$). (Proposition \[p:rescaledLoctilde\], Lemma \[l:centralelt\])
3. The quotient that sets the above central element to 1 inherits a quantum cluster algebra structure (with the same cluster combinatorics). (Corollary \[c:quotisQCA\])
4. Demonstration that this quotient is isomorphic to ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$. (Theorem \[t:quotisotoGr\])
1. **(B1)**, **(B2)** and **(C2)** above collectively imply that there exists a surjective homomorphism.
2. The equalities of the Gelfand–Kirillov dimensions of the two algebras shows that this epimorphism is an isomorphism.
5. Finally, the powers of $q$ appearing in the expressions for the quantum cluster variables can be removed, as another consequence of the grading. (after Theorem \[t:quotisotoGr\])
This construction does indeed yield the same results as the authors’ earlier work ([@Gr2nSchubertQCA]) but now in arbitrary—and in particular far from finite—type, in a universal way. Other consequences of the approach taken here include the fact that every quantum cluster variable for this structure on ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ is homogeneous for the standard grading, with the quantum Plücker coordinates in degree one. It is hard to see how this property could be established in infinite type without the framework presented here. We note that we do not deduce this from any explicit formulæ; indeed, we have no such formulæ, though it would be interesting to understand the forms of the quantum cluster variables in tame types (${\mathbb{K}_{q}[\mathrm{Gr}(3,9)]}$ and ${\mathbb{K}_{q}[\mathrm{Gr}(4,8)]}$), now that the existence of the quantum cluster algebra structure is proved.
We also note that the above proof does not in fact rely on any particularly special properties of the quantum Grassmannian. Indeed, many of the steps outlined above have analogues for quantum coordinate rings of big cells of partial flag varieties in full generality; by the latter, we mean the algebras $A_{q}(\mathfrak{n}(w_{0}w_{0}^{K}))$ discussed in [@GLS-QuantumPFV §12.5]. In particular, Corollary 12.12 of that paper gives us a quantum cluster algebra structure to take as “input” to the process described above. Similarly, corresponding dehomogenisation isomorphisms in this more general case are known ([@Yakimov], [@Yakimov-2]).
However some adjustments may be needed in the general case. From the quantum Grassmannian ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ we localise by the Ore set $\{ [1\dotsm k]^{n} \mid n\in {\ensuremath \mathbb{N}}\}$, which naturally gives the localisation a ${\ensuremath{\mathbb{Z}}}$-grading. Then we obtain ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ as the degree zero part of this. In general, we will need to localise by powers of several elements, leading to multi-gradings on the localisation. Hence we will need to work with multi-gradings on the quantum cluster algebras $A_{q}(\mathfrak{n}(w_{0}w_{0}^{K}))$, to lift that structure to the quantized coordinate ring of the partial flag variety itself. We intend to return to this topic in more detail in future work.
Acknowledgements {#acknowledgements .unnumbered}
----------------
The second author is grateful for the financial support of EPSRC First Grant *EP/I018549/1*.
Preliminaries
=============
Quantum matrices and quantum Grassmannians {#ss:qGrassmannians}
------------------------------------------
Throughout, $\mathbb{C}$ denotes the field of complex numbers and $\mathbb{K}$ denotes the field $\mathbb{Q}(q^{1/2})$. Then in particular, the indeterminate $q$ is not a root of unity and has a square root in $\mathbb{K}$. Let $C$ be an $l{\times}l$ generalized Cartan matrix with columns indexed by a set $I$. Let $(H,\Pi,\Pi^{\vee})$ be a minimal realization of $C$, where $H {\ensuremath \cong}{\ensuremath \mathbb{C}}^{2\left| I \right|-\text{rank}(C)}$, $\Pi=\{ \alpha_{i} \mid i \in I \} \subset {\ensuremath {H}^{*}}$ (the simple roots) and $\Pi^{\vee}=\{ h_{i} \mid i \in I \} \subset H$ (the simple coroots). Then we say ${\ensuremath{\mathcal{C}}}=(C,I,H,\Pi,\Pi^{\vee})$ is a root datum associated to $C$. (Lusztig [@LusztigBook] has a more general definition of a root datum but this one will suffice for our purposes.)
If $G=G({\ensuremath{\mathcal{C}}})$ is a connected semisimple complex algebraic group associated to ${\ensuremath{\mathcal{C}}}$, $G$ has a (standard) parabolic subgroup $P_{J}$ associated to any choice of subset $J\subseteq I$. From this we can form $G/P_{J}$, a partial flag variety; the choice $J=\emptyset$ gives $G/P_{\emptyset}=G/B$, the full flag variety. We set $D=I\setminus J$.
The partial flag variety $G/P_{J}$ is a projective variety, via the well-known Plücker embedding $G/P_{J} {\hookrightarrow}\prod_{d\in D} \mathbb{P}(L(\omega_{d}))$. (Here, $L(\lambda)$ is the irreducible $G$-module corresponding to a dominant integral weight $\lambda$ and $\{\omega_{i}\}_{i\in I}$ are the fundamental weights.) Via the Plücker embedding, we may form the corresponding ${\ensuremath \mathbb{N}}^{D}$-graded multi-homogeneous coordinate algebra ${\ensuremath \mathbb{C}}[G/P_{J}]=\bigoplus_{\lambda \in {\ensuremath \mathbb{N}}^{D}} {\ensuremath {L(\lambda)}^{*}}$. The case we consider is that of the partial flag variety obtained from $G=G(A_{n})=SL_{n+1}({\ensuremath \mathbb{C}})$ and $J=I \setminus \{k\}$, namely $G/P_{J}=\mathrm{Gr}(k,n)$, the Grassmannian of $k$-dimensional subspaces in ${\ensuremath \mathbb{C}}^{n}$.
The coordinate ring ${\ensuremath \mathbb{C}}[G]$ has a quantum analogue, $\mathbb{K}_{q}[G]$ (see for example [@Brown-Goodearl], where this algebra is denoted ${\ensuremath{\mathcal{O}}}_{q}(G)$). Via this quantized coordinate ring, we can define a quantization $\mathbb{K}_{q}[G/P_{J}]$ of $\mathbb{K}[G/P_{J}]$.
We recall that the quantum matrix algebra ${\mathbb{K}_{q}[\mathrm{M}(k,n)]}$ is the $\mathbb{K}$-algebra generated by the set $\{ X_{ij} \mid 1\leq i\leq k,\ 1\leq j \leq n \}$ subject to the quantum $2{\times}2$ matrix relations on each $2{\times}2$ submatrix of $$\begin{pmatrix} X_{11} & X_{12} & \cdots & X_{1n} \\ \vdots & \vdots & \ddots & \vdots \\ X_{k1} & X_{k2} & \cdots & X_{kn} \end{pmatrix}.$$ The quantum $2{\times}2$ matrix relations on $\left( \begin{smallmatrix} a & b \\ c & d \end{smallmatrix} \right)$ are $$\begin{aligned}
ab & = qba & ac & = qca & bc & = cb \\ bd & = qdb & cd & = qdc & ad-da & = (q-q^{-1})bc. \end{aligned}$$
Recall that the $k{\times}k$ quantum minor $\Delta_{q}^{I}$ associated to the $k$-subset $I=\{ i_{1} < i_{2} < \cdots < i_{k} \}$ of $\{1, \dots , n\}$ is defined to be $$\Delta_{q}^{I} {\stackrel{\scriptscriptstyle{\mathrm{def}}}{=}}\sum_{\sigma \in S_{k}} (-q)^{l(\sigma)}X_{1i_{\sigma(1)}}\cdots X_{k\mspace{0.5mu}i_{\sigma(k)}}$$ where $S_{k}$ is the symmetric group of degree $k$ and $l$ is the usual length function on this. This defines the quantum minor $\Delta_{\{1,\ldots,k\}}^{I}$ but quantum minors of smaller degree or for different choices of the row set are defined analogously in the obvious way.
Our notation for a quantum minor with row set $I$ and column set $J$ will be ${\genfrac{[}{]}{0em}{1}{J}{I}}$; note that we suppress $q$, as all our minors will be quantum minors unless otherwise stated. For example, when $k=2$, we will write the quantum minor $\Delta_{q}^{ij}$ for $i<j$ as ${\genfrac{[}{]}{0em}{1}{ij}{12}}$; written in terms of the generators of ${\mathbb{K}_{q}[\mathrm{M}(2,2)]}$ this is equal to $X_{1i}X_{2j}-qX_{1j}X_{2i}$. Similarly, we will often denote the generator $X_{ij}$ of ${\mathbb{K}_{q}[\mathrm{M}(k,n)]}$ by $(ij)$.
Then we denote by ${\ensuremath{\mathcal{P}}}_{q}$ the set of all quantum Plücker coordinates, that is $${\ensuremath{\mathcal{P}}}_{q} = \{ \Delta_{q}^{I} \mid I \subseteq \{1,\ldots ,n\}, {\lvert I \rvert}=k \}.$$
\[def:KqGrkn\] The quantum Grassmannian ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ is the subalgebra of the quantum matrix algebra ${\mathbb{K}_{q}[\mathrm{M}(k,n)]}$ generated by the quantum Plücker coordinates ${\ensuremath{\mathcal{P}}}_{q}$.
When working with minors in the quantum Grassmannian ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$, where the row set is necessarily $\{ 1,\ldots ,k\}$, we will write $[I]$ for $\Delta_{q}^{I}$, [e.g. ]{}$[ij]$ for $\Delta_{q}^{ij}$ as above.
It is well-known that ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ is a Noetherian domain with Gelfand–Kirillov dimension $k(n-k)+1$. (Further ring-theoretic properties of quantum Grassmannians are established in [@Fioresi], [@KellyLenaganRigal], [@LaunoisLenaganRigal] and [@LenaganRigal].)
Cluster algebras
----------------
The construction of a cluster algebra of geometric type from an initial seed $(\underline{x},B)$, due to Fomin and Zelevinsky ([@FZ-CA1]), is now well-known. Here $\underline{x}$ is a transcendence base for a certain field of fractions of a polynomial ring and $B$ is a skew-symmetric integer matrix; often $B$ is replaced by the quiver $Q=Q(B)$ it defines in the natural way. We refer the reader who is unfamiliar with this construction to the survey of Keller ([@Keller-CASurvey]) and the recent book of Gekhtman, Shapiro and Vainshtein ([@GSV-Book]) for an introduction to the topic and summaries of the main related results in this area.
Quantum cluster algebras
------------------------
Berenstein and Zelevinsky ([@BZ-QCA]) have given a definition of a quantum cluster algebra. These algebras are non-commutative but not so far from being commutative. A quantum seed $(\underline{x},B,L)$ consists of $\underline{x}=(X_{1},\dotsc ,X_{r})$, simultaneously a transcendence base and generating set for the skew-field of fractions ${\ensuremath{\mathcal{F}}}$ of a quantum torus (over the field $\mathbb{K}$), a skew-symmetric integer matrix $B$ (the exchange matrix) and a second skew-symmetric integer matrix $L=(l_{ij})$ that determines the aforementioned quantum torus. That is, the matrix $L$ describes quasi-commutation relations between the variables in the cluster, where quasi-commuting means the existence of a relation of the form $X_{i}X_{j}=q^{l_{ij}}X_{j}X_{i}$.
There is also a mutation rule for these quasi-commutation matrices as well as a modified exchange relation that involves further coefficients that are powers of $q$ derived from $B$ and $L$, which we describe now. For $k$ a mutable index, set $$\begin{aligned}
\underline{b}_{k}^{+} & = -\underline{\boldsymbol{e}}_{k}+\sum_{b_{ik}>0}b_{ik}\underline{\boldsymbol{e}}_{i} \qquad \text{and} \\
\underline{b}_{k}^{-} & = -\underline{\boldsymbol{e}}_{k}-\sum_{b_{ik}<0}b_{ik}\underline{\boldsymbol{e}}_{i}\end{aligned}$$ where the vector $\underline{\boldsymbol{e}}_{i}\in {\ensuremath \mathbb{C}}^{r}$ ($r$ being the number of rows of $B$) is the $i$th standard basis vector. Note that the $k$th row of $B$ may be recovered as $B_{k}=\underline{b}_{k}^{+}-\underline{b}_{k}^{-}$.
Then given a quantum cluster $\underline{x}=(X_{1},\ldots,X_{r})$, exchange matrix $B$ and quasi-commutation matrix $L$, the exchange relation for mutation in the direction $k$ is given by $$X_{k}^{\prime}=M(\underline{b}_{k}^{+})+M(\underline{b}_{k}^{-})$$ with $$M(a_{1},\dotsc ,a_{r}) {\stackrel{\scriptscriptstyle{\mathrm{def}}}{=}}q^{\frac{1}{2}\sum_{i<j} a_{i}a_{j}l_{ji}}X_{1}^{a_{1}}\dotsm X_{r}^{a_{r}}.$$ By construction, the integers $a_{i}$ are all non-negative except for $a_{k}=-1$. The monomial $M$ (as we have defined it here) is related to the concept of a toric frame, also introduced in [@BZ-QCA]. The latter is a technical device used to make the general definition of a quantum cluster algebra. For our examples, where we start with a known algebra and want to exhibit a quantum cluster algebra structure on this, it will suffice to think of $M$ simply as a rule determining the exchange monomials.
The quantum cluster algebra ${\ensuremath{\mathcal{A}}}_{q}={\ensuremath{\mathcal{A}}}_{q}(\underline{x},B,L)$ defined by the initial data $(\underline{x},B,L)$ is the subalgebra of ${\ensuremath{\mathcal{F}}}$ generated by the set of all quantum cluster variables, that is, those elements of ${\ensuremath{\mathcal{F}}}$ obtained from the initial cluster variables by iterated mutation. We note that the presence of the factor $1/2$ in the quantum exchange relations is the reason for assuming that the element $q\in \mathbb{K}$ has a square root.
We will need to work with quantum cluster algebras with coefficients (also called frozen variables). That is, we designate some of the elements of the initial cluster to be mutable (i.e. we are allowed to mutate these) and the remainder to be non-mutable. We will also talk about the corresponding indices for the variables as being mutable or not; in [@BZ-QCA] the former are referred to as exchangeable indices. The rank of the quantum cluster algebra is the number of mutable variables in a cluster; we will refer to the total number of variables, mutable and not, as the cardinality of the cluster.
Note that we will adopt the convention that $B$ will be a *square* matrix—in the literature it is more common to let $B$ have as column indices just the mutable indices ([@GSV-Book] adopts the transpose convention, of having the rows indexed by the mutable indices). At some points—notably in the next paragraph—we will need the submatrix $B_{\text{mut}}$ of $B$ given by taking only the columns of $B$ with mutable indices. The matrix $B_{\text{mut}}$ is often referred to as the extended exchange matrix and *its* submatrix $B_{\text{mut}}^{\textrm{mut}}$ with row set also the mutable indices is what is usually called the principal part of $B$. Our square matrix $B$ is simply the “skew-symmetric extension” of $B_{\text{mut}}$, i.e. completing $B_{\text{mut}}$ to a square matrix in the unique way so that $B$ is skew-symmetric and so that if $i$ and $j$ are non-mutable indices then $b_{ij}=0$. (The latter choice accords with the convention that the exchange quiver has no arrows between frozen vertices.)
The natural requirement that all mutated quantum clusters also quasi-commute leads to a compatibility condition between $B$ and $L$, namely that ${{(B_{\text{mut}})}^{T}}L$ consists of two blocks, one a positive integer scalar multiple of the identity and one zero. The non-zero block must correspond exactly to the mutable (column) indices. However, these blocks need not be contiguous, depending on the ordering of the row and column labels. (Only one positive integer is required, as the principal part of $B$ is assumed to be skew-symmetric; if one assumes just skew-symmetrizability then the non-zero block is only required to be diagonal with positive integer diagonal entries.)
Let us say that an arbitrary skew-symmetric matrix $A$ is $d$-compatible with $B$ if ${{(B_{\text{mut}})}^{T}}A$ has the form described above for some non-negative integer $d$. It will suit our purposes later to allow $d=0$ (i.e. ${{(B_{\text{mut}})}^{T}}A=0$), even though $0$-compatibility is not permitted for the compatibility of a quasi-commutation matrix $L$ with $B$.
Importantly, Berenstein and Zelevinsky show that the exchange graph (whose vertices are the clusters and edges are mutations) remains unchanged in the quantum setting. That is, the matrix $L$ does not influence the exchange graph. It follows that quantum cluster algebras of finite type are classified by Dynkin types in exactly the same way as the classical cluster algebras.
Known or conjectured examples of quantum cluster algebras include
- quantum symmetric algebras (necessarily of cluster algebra rank 0);
- quantum Grassmannians of finite cluster algebra type ([i.e. ]{}${\mathbb{K}_{q}[\mathrm{Gr}(2,n)]}$ and ${\mathbb{K}_{q}[\mathrm{Gr}(3,6)]}$, ${\mathbb{K}_{q}[\mathrm{Gr}(3,7)]}$ and ${\mathbb{K}_{q}[\mathrm{Gr}(3,8)]}$) ([@Gr2nSchubertQCA]);
- Schubert cells of the quantum Grassmannians ${\mathbb{K}_{q}[\mathrm{Gr}(2,n)]}$ ([@Gr2nSchubertQCA],[@GLS-QuantumPFV]);
- the quantum coordinate ring of the unipotent subgroup $N(w)$ of a symmetric Kac–Moody group $G$ associated with a Weyl group element $w$ ([@GLS-QuantumPFV]), and hence as special cases of this
- the quantum coordinate ring of the big cell of a partial flag variety associated to $G$ and
- quantum matrices ${\mathbb{K}_{q}[\mathrm{M}(k,n)]}$; and
- conjecturally, quantum double Bruhat cells of semisimple Lie groups ([@BZ-QCA]).
We note that recently Goodearl and Yakimov ([@GoodearlYakimov]) have studied the existence of initial seeds in a large class of algebras, the so-called CGL extensions.
Graded seeds and graded quantum cluster algebras {#s:gradedQCAs}
================================================
Berenstein and Zelevinsky ([@BZ-QCA Definition 6.5]) have given a definition of graded quantum seeds, which give rise to module gradings but not algebra gradings. In what follows, we will have need of algebra gradings on quantum cluster algebras and so we now give a different definition of a graded seed, inspired by that of Berenstein and Zelevinsky but not equivalent to it.
A graded quantum seed is a quadruple $(\underline{x},B,L,G)$ such that
1. $(\underline{x}=(X_{1},\dotsc ,X_{r}),B,L)$ is a quantum seed of cardinality $r$ and
2. $G\in {\ensuremath{\mathbb{Z}}}^{r}$ is an integer (column) vector such that for all mutable indices $j$, the $j$th row of $B$, $B_{j}$, satisfies $B_{j}G=0$.
We will set $\deg_{G}(X_{i})=G_{i}$ for all $X_{i}$ belonging to the cluster $\underline{x}$. Then the second condition of the definition is equivalent to asking that every exchange relation (as encoded by the rows $B_{j}$) is homogeneous with respect to this degree, in the sense that the two Laurent monomials determining $X_{j}^{\prime}$ are of the same homogeneous degree. From the quiver perspective, this asks that the sum of the degrees of the variables with arrows to a given mutable vertex is equal to the sum of the degrees of the variables at the end of arrows leaving that vertex.
In contrast to the definition of Berenstein and Zelevinsky, the above can be extended to an algebra grading on the quantum torus associated to $(\underline{x},B,L)$, simply by setting $\deg_{G}(X_{i}^{-1})=-\deg_{G}(X_{i})$ and extending $\deg_{G}$ additively to all (Laurent) monomials.
We also need to be able to mutate our grading in a sensible fashion and it is clear what we ought to do. Let $(\underline{x}^{\prime},B^{\prime},L^{\prime})$ be the quantum seed given by mutation of $(\underline{x},B,L)$ in the direction $j$. We set $G^{\prime}_{i}=G_{i}$ for $i\neq j$ (i.e. the degrees of variables we are not mutating at this point remain the same). Then the homogeneity of the exchange relation $X_{j}^{\prime}=M(\underline{b}_{k}^{+})+M(\underline{b}_{k}^{-})$ implies that we should set $$G^{\prime}_{j} =\deg_{G^{\prime}}(X_{j}^{\prime}) =\deg_{G}(M(\underline{b}_{k}^{+})) = \deg_{G}(M(\underline{b}_{k}^{-})).$$
As discussed in [@BFZ-CA3] and [@BZ-QCA], the mutation operations can also be expressed in terms of row and column operations, or more concisely as corresponding matrix multiplications. To this end, we recall the definition of a matrix $E$ (denoted $E_{+}$ in [@BZ-QCA]) that encodes mutation of a seed with exchange matrix $B$ in the direction $j$ as follows: $$E_{rs}=
\begin{cases}
\delta_{rs} & \text{if}\ s\neq j; \\
-1 & \text{if}\ r=s=j; \\
\max(0,-b_{rj}) & \text{if}\ r\neq s=j.
\end{cases}$$ Then $B^{\prime}=EB{{E}^{T}}$ and $L^{\prime}={{E}^{T}}LE$. Our mutation of $G$ can be written in terms of $E$ similarly.
\[l:mutofgradinggivenbyE\] $G^{\prime}={{E}^{T}}G$.
This is straightforward to check directly.
We may also re-express this in terms of the vectors $\underline{b}_{k}^{\pm}$ defined above: $$G_{i}^{\prime} = \begin{cases} G_{i} & \text{if}\ i\neq k \\ \underline{b}_{k}^{-}\cdot G & \text{if}\ i=k \end{cases}.$$ Since $B_{k}=\underline{b}_{k}^{+}-\underline{b}_{k}^{-}$ and $G$ is a grading, so $B_{k}G=0$, we have $\underline{b}_{k}^{+}\cdot G=\underline{b}_{k}^{-}\cdot G$ so that we may use $\underline{b}_{k}^{+}$ instead of $\underline{b}_{k}^{-}$ in calculating $G^{\prime}$.
We also need to know that this mutation operation does indeed produce another graded seed.
\[l:mutofgrading\] For each mutable index $j$, $(B^{\prime})_{j}G^{\prime}=0$.
As noted in [@BZ-QCA Proposition 3.4], $E^{2}=1$ and so $$\begin{aligned}
(B^{\prime})_{j}G^{\prime} & = (EB{{E}^{T}})_{j}({{E}^{T}}G) \\
& = (EB({{E}^{T}})^{2}G)_{j} \\
& = (EBG)_{j} \\
& = E_{j}(BG) \\
& = 0 \end{aligned}$$ since $(BG)_{j}=B_{j}G=0$. Here, $(\ )_{j}$ refers to the $j$th row for matrices and column vectors as appropriate.
That is, $(\underline{x}^{\prime},B^{\prime},L^{\prime},G^{\prime})$ is again a graded seed. Furthermore, this mutation is involutive (cf. [@BZ-QCA Proposition 4.10]). Then we see that repeated mutation propagates a grading on an initial seed to every quantum cluster variable and hence to the associated quantum cluster algebra, as every exchange relation is homogeneous.
The quantum cluster algebra ${\ensuremath{\mathcal{A}}}_{q}(\underline{x},B,L,G)$ associated to an initial graded quantum seed $(\underline{x},B,L,G)$ is a ${\ensuremath{\mathbb{Z}}}$-graded algebra.
We note in particular that this construction—by definition—says that every quantum cluster variable of a graded quantum cluster algebra is homogeneous for this grading.
It is clear that all of the above is insensitive to replacing $G$ with $-G$, i.e. reversing the sign of every degree. Indeed for each graded quantum cluster algebra ${\ensuremath{\mathcal{A}}}_{q}(\underline{x},B,L,G)$ we have an isomorphic graded quantum cluster algebra ${\ensuremath{\mathcal{A}}}_{q}^{-}={\ensuremath{\mathcal{A}}}_{q}(\underline{x},B,L,-G)$ (where “isomorphic” here means as quantum cluster algebras, not just as algebras).
In none of the above have we used the quasi-commutation matrix $L$. Indeed all of the above goes through for classical cluster algebras too.
One consequence of the existence of a grading for a quantum cluster algebra is that this allow us to re-scale elements of the initial seed by powers of $q$ determined by the grading and obtain an isomorphic quantum cluster algebra, as follows.
\[p:rescaleQCAbypowersofq\] Let ${\ensuremath{\mathcal{A}}}_{q}={\ensuremath{\mathcal{A}}}_{q}(\underline{x}=(X_{1},\dotsc ,X_{r}),B,L,G)$ be a graded quantum cluster algebra. Let $\tilde{\underline{x}}=(\tilde{X}_{1},\dotsc ,\tilde{X}_{r})$ be defined by $\tilde{X}_{i}=q^{G_{i}/2}X_{i}$. Then there is an isomorphism of graded quantum cluster algebras between $\tilde{{\ensuremath{\mathcal{A}}}}_{q}={\ensuremath{\mathcal{A}}}_{q}(\tilde{\underline{x}},B,L,G)$ and ${\ensuremath{\mathcal{A}}}_{q}={\ensuremath{\mathcal{A}}}_{q}(\underline{x},B,L,G)$.
Firstly note that $\underline{x}$ and $\tilde{\underline{x}}$ determine the same quantum torus and hence the corresponding quantum cluster algebras ${\ensuremath{\mathcal{A}}}_{q}$ and $\tilde{{\ensuremath{\mathcal{A}}}}_{q}$ can be viewed as subalgebras of the same skew-field of fractions ${\ensuremath{\mathcal{F}}}$ of this quantum torus. We show that if $y$ and $\tilde{y}$ are quantum cluster variables for ${\ensuremath{\mathcal{A}}}_{q}$ and $\tilde{A}_{q}$ respectively that are obtained from their respective initial seeds by the same sequence of mutations, then $\tilde{y}=q^{\deg(y)/2}y$ where $\deg$ is the degree induced by the grading $G$.
More precisely, let $(\underline{x}=(X_{1},\dotsc ,X_{r}),B,L,G)$ be any graded quantum seed (not necessarily equal to the initial data for ${\ensuremath{\mathcal{A}}}_{q}$; we abuse notation here). Let $\tilde{\underline{x}}=(q^{G_{1}/2}X_{1},\dotsc ,q^{G_{r}/2}X_{r})$. We claim that if $X_{k}^{\prime}$ is the variable obtained from $\underline{x}$ by mutation in the direction $k$ then $\tilde{X}_{k}^{\prime}=q^{G_{k}^{\prime}/2}X_{k}^{\prime}$, where $G^{\prime}$ is obtained from $G$ by mutation in the direction $k$.
Firstly, recall the definition of the exchange monomial $M(a_{1},\dotsc ,a_{r})$ as $$M(a_{1},\dotsc ,a_{r}) {\stackrel{\scriptscriptstyle{\mathrm{def}}}{=}}q^{\frac{1}{2}\sum_{i<j} a_{i}a_{j}l_{ji}}X_{1}^{a_{1}}\dotsm X_{r}^{a_{r}}.$$ Then letting $\underline{a}=(a_{1},\dotsc ,a_{r})$ and $$\tilde{M}(\underline{a})=q^{\frac{1}{2}\sum_{i<j} a_{i}a_{j}l_{ji}}\tilde{X}_{1}^{a_{1}}\dotsm \tilde{X}_{r}^{a_{r}}$$ we have that $$\begin{aligned}
\tilde{M}(\underline{a}) & =q^{\frac{1}{2}\sum_{i<j} a_{i}a_{j}l_{ji}}q^{\frac{1}{2}\sum_{i=1}^{r} a_{i}G_{i}}X_{1}^{a_{1}}\dotsm X_{r}^{a_{r}} \\
& = q^{\frac{\underline{a}\cdot G}{2}}M(\underline{a}).\end{aligned}$$
Hence $$\begin{aligned}
\tilde{X}_{k}^{\prime} & =\tilde{M}(\underline{b}_{k}^{+})+\tilde{M}(\underline{b}_{k}^{-}) \\
& = q^{\frac{\underline{b}_{k}^{+} \cdot G}{2}}M(\underline{b}_{k}^{+})+q^{\frac{\underline{b}_{k}^{-} \cdot G}{2}}M(\underline{b}_{k}^{-}) \\
& = q^{\frac{\underline{b}_{k}^{+} \cdot G}{2}}\left( M(\underline{b}_{k}^{+})+M(\underline{b}_{k}^{-})\right) \\
& = q^{G_{k}^{\prime}/2}X_{k}^{\prime}\end{aligned}$$ as required. Here we have used that $G$ being a grading for $B$ implies that $\underline{b}_{k}^{+}\cdot G=\underline{b}_{k}^{-}\cdot G$ and the equality of both of these with the $k$th entry of the mutated grading $G^{\prime}$, as shown after Lemma \[l:mutofgradinggivenbyE\].
Now since the initial data for ${\ensuremath{\mathcal{A}}}_{q}$ and $\tilde{{\ensuremath{\mathcal{A}}}}_{q}$ differ only in the choice of initial cluster and, by the above, the subalgebras of ${\ensuremath{\mathcal{F}}}$ generated by the respective sets of quantum cluster variables are equal, it follows that these are isomorphic quantum cluster algebras.
We note that this construction does not have a counterpart in the classical setting of commutative cluster algebras. However it should have a semi-classical counterpart, for cluster algebras with compatible Poisson structures (in the sense of Gekhtman–Shapiro–Vainshtein).
Skew-Laurent extensions of quantum cluster algebras
===================================================
In this section, we consider two constructions that produce graded quantum cluster algebra structures on skew-Laurent extensions of a given graded quantum cluster algebra. The first simply adds the extending variable and its inverse as extra coefficients, while the second “re-scales” the original structure by multiplying each quantum cluster variable by a power of the extending variable. This second construction is similar to that in Proposition \[p:rescaleQCAbypowersofq\] but in general produces a new quantum cluster algebra not isomorphic to the first.
The first construction proceeds as follows.
\[p:skewLaurent-extra-coeffs\] Let ${\ensuremath{\mathcal{A}}}_{q}={\ensuremath{\mathcal{A}}}_{q}(\underline{x}=(X_{1},\ldots,X_{r}),B,L,G)$ be a graded quantum cluster algebra. Let $\sigma\colon {\ensuremath{\mathcal{A}}}_{q} \to {\ensuremath{\mathcal{A}}}_{q}$ be an algebra automorphism such that for each $1\leq i \leq r$, there exists $c_{i}\in {\ensuremath{\mathbb{Z}}}$ such that $\sigma(X_{i})=q^{c_{i}}X_{i}$; that is, $\sigma$ acts by multiplication by powers of $q$ on the elements of the initial cluster. Then the skew-Laurent extension ${\ensuremath{\mathcal{A}}}_{q}[y^{\pm 1};\sigma]$ is a graded quantum cluster algebra, with initial data $(\underline{x}',B',L',G')$ where
- $\underline{x}'=(X_{1},\ldots,X_{r},y,y^{-1})$, with $y$ and $y^{-1}$ additional coefficients,
- $Q(B')$ is equal to $Q(B) {\sqcup}\boxed{y\rule{0em}{0.9em}} {\sqcup}\boxed{y^{-1}\rule{0em}{0.9em}}$,
- $L'$ is determined by the quasi-commutation data in $L$ and the automorphism $\sigma$, and
- $(G')_{i}=G_{i}$ for $1\leq i \leq r$, $G'_{r+1}=1$ and $G'_{r+2}=-1$.
The matrices $B'$ and $L'$ are compatible, as $B'$ has only zero entries in the rows corresponding to $y$ and $y^{-1}$. The matrices $B'$ and $G'$ satisfy the grading condition for the same reason. Furthermore it is clear that the set of cluster variables for ${\ensuremath{\mathcal{A}}}_{q}[y^{\pm 1};\sigma]$ is equal to the disjoint union of the set of cluster variables for ${\ensuremath{\mathcal{A}}}_{q}$ and $\{ y,y^{-1} \}$, and so these generate all of ${\ensuremath{\mathcal{A}}}_{q}[y^{\pm 1};\sigma]$ which is therefore a graded quantum cluster algebra.
We note that we have chosen to put $y$ in degree $1$ (and its inverse in degree $-1$), to accord with the natural ${\ensuremath{\mathbb{Z}}}$-grading on this skew-Laurent extension. However neither the element $y$ nor its inverse interacts with the quantum cluster structure coming from ${\ensuremath{\mathcal{A}}}_{q}$ so this choice was essentially arbitrary.
Here we did not alter the original quantum cluster variables but in our second construction, we re-scale these by powers of an extending variable $z$. In its most general form, this re-scaling will involve two integer column vectors as (families of) parameters, $\underline{t}=(t_{1},\dotsc ,t_{r})$ and $\underline{u}=(u_{1},\dotsc ,u_{r})$. Given two such vectors we will denote by $\underline{t} \wedge \underline{u}$ the skew-symmetric matrix ${{\underline{t}}^{T}}\underline{u}-{{\underline{u}}^{T}}\underline{t}$, i.e. the $(i,j)$-entry of $\underline{t}\wedge \underline{u}$ is $t_{i}u_{j}-t_{j}u_{i}$.
Let ${\ensuremath{\mathcal{A}}}_{q}={\ensuremath{\mathcal{A}}}_{q}(\underline{x}=(X_{1},\ldots,X_{r}),B,L,G)$ be a graded quantum cluster algebra such that $L$ is $d$-compatible with $B$ with $d>0$, i.e. we have $({{B}^{T}}L)_{ij}=d\delta_{ij}$ for all mutable indices $i$ and any index $j$. (Such a $d$ certainly exists by the definition of compatibility between $B$ and $L$; we are simply naming it explicitly.)
Now let $\tau\colon {\ensuremath{\mathcal{A}}}_{q} \to {\ensuremath{\mathcal{A}}}_{q}$ be an algebra automorphism such that for each $1\leq i \leq r$, there exists $t_{i}\in {\ensuremath{\mathbb{Z}}}$ such that $\tau(X_{i})=q^{t_{i}}X_{i}$; that is, $\tau$ acts by multiplication by powers of $q$ on the elements of the initial cluster. We set $\underline{t}={{(t_{1},\dotsc ,t_{r})}^{T}}$. Also, let $\underline{u}={{(u_{1},\dotsc ,u_{r})}^{T}}\in {\ensuremath{\mathbb{Z}}}^{r}$ be such that $\underline{t} \wedge \underline{u}$ is $f$-compatible with $B$ for some $0\leq f<d$.
Then the following is a valid set of initial data for a graded quantum cluster algebra $\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}}={\ensuremath{\mathcal{A}}}_{q}(\underline{\tilde{x}},\tilde{B},\tilde{L},\tilde{G})$ where
- $\underline{\tilde{x}}=(X_{1}z^{u_{1}},\ldots,X_{r}z^{u_{r}})$,
- $\tilde{B}=B$,
- $\tilde{L}$ is determined by the quasi-commutation data in $L$ and the vectors $\underline{t}$ and $\underline{u}$, and
- $\tilde{G}=G$,
with $\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}}$ a subalgebra of the skew-field of fractions of the skew-Laurent extension ${\ensuremath{\mathcal{A}}}_{q}[z^{\pm 1};\tau]$ of ${\ensuremath{\mathcal{A}}}_{q}$.
We prove this validity via two lemmas, where we first establish more explicitly the matrix $\tilde{L}$ and then prove that it is compatible with $\tilde{B}=B$.
\[l:Lminuswedge\] We have $\tilde{L}=L-\underline{t}\wedge \underline{u}$.
Since $\tau(X_{i})=q^{t_{i}}X_{i}$ we have that $zX_{i}=\tau(X_{i})z=q^{t_{i}}X_{i}z$ for all $i$, by the definition of a skew-Laurent extension. For all $1\leq i,j \leq r$ we have $X_{i}X_{j}=q^{l_{ij}}X_{j}X_{i}$ and hence $$\begin{aligned}
\tilde{X}_{i}\tilde{X}_{j} & = (X_{i}z^{u_{i}})(X_{j}z^{u_{j}}) \\
& = X_{i}\left( (q^{t_{j}})^{u_{i}}X_{j}z^{u_{i}} \right)z^{u_{j}} \\
& = q^{t_{j}u_{i}}\left( q^{l_{ij}}X_{j}X_{i} \right) z^{u_{i}+u_{j}} \\
& = q^{l_{ij}+t_{j}u_{i}}X_{j}\left( X_{i}z^{u_{j}} \right) z^{u_{i}} \\
& = q^{l_{ij}+t_{j}u_{i}}X_{j}\left( (q^{-t_{i}})^{u_{j}}z^{u_{j}}X_{i} \right)z^{u_{i}} \\
& = q^{l_{ij}+t_{j}u_{i}-t_{i}u_{j}}(X_{j}z^{u_{j}})(X_{i}z^{u_{i}}) \\
& = q^{l_{ij}+t_{j}u_{i}-t_{i}u_{j}}\tilde{X}_{j}\tilde{X}_{i}.\end{aligned}$$ Therefore $\tilde{l}_{ij}=l_{ij}+t_{j}u_{i}-t_{i}u_{j}=l_{ij}-(\underline{t}\wedge \underline{u})_{ij}$. Note that we need $q$ not a root of unity at this point.
\[l:recaled-compatibility\] The matrices $\tilde{B}$ and $\tilde{L}$ are compatible.
Let $i$ be a mutable index and $j$ any index. Then $$\begin{aligned}
\left({{(\tilde{B}_{\text{mut}})}^{T}}\tilde{L}\right)_{ij} & = \left({{(B_{\text{mut}})}^{T}}(L-(\underline{t}\wedge \underline{u}))\right)_{ij} \\
& = \left({{(B_{\text{mut}})}^{T}}L\right)_{ij}-\left({{(B_{\text{mut}})}^{T}}(\underline{t}\wedge \underline{u})\right)_{ij} \\
& = d\delta_{ij}-f\delta_{ij} \\
& = (d-f)\delta_{ij}\end{aligned}$$ since $L$ is $d$-compatible with $B$ and, by assumption, $\underline{t}\wedge \underline{u}$ is $f$-compatible with $B$. Furthermore, $0\leq f<d$ so we see that $\tilde{L}$ is $(d-f)$-compatible with $B$, with $d-f$ a positive integer as required.
We note that the effect of this construction is to leave the initial exchange relations unchanged but to alter the quasi-commutation relations. That is, this construction can be thought of as a form of twisting of a quantum cluster algebra.
If $\underline{t}\wedge \underline{u}$ is $0$-compatible with $B$, the precise value of $d$ is irrelevant and we always obtain compatibility. We note some special cases.
Let ${\ensuremath{\mathcal{A}}}_{q}$, $\underline{t}$ and $\underline{u}$ be as above. Then
1. if $\underline{t}=0$ we have $\tilde{L}=L$, $f=0$ and hence ${\ensuremath{\mathcal{A}}}_{q}^{0,\underline{u}}{\ensuremath \cong}{\ensuremath{\mathcal{A}}}_{q}$;
2. if $\underline{u}=0$ we have $\tilde{\underline{x}}=\underline{x}$ and $\tilde{L}=L$ and hence ${\ensuremath{\mathcal{A}}}_{q}^{\underline{t},0}={\ensuremath{\mathcal{A}}}_{q}$; and
3. if $\underline{t}$ and $\underline{u}$ are ${\ensuremath{\mathbb{Z}}}$linearly dependent we have $\tilde{L}=L$, $f=0$ and hence ${\ensuremath{\mathcal{A}}}_{q}^{\underline{t},\underline{u}}{\ensuremath \cong}{\ensuremath{\mathcal{A}}}_{q}$.
[ ]{}
1. If $\underline{t}=0$ then $\underline{t}\wedge \underline{u}=0$, so that $\underline{t}\wedge \underline{u}$ is $0$-compatible with $B$. Then the remaining claims follow. Note that in this case the skew-Laurent extension induced by $\tau$ is a central (Laurent) extension.
2. If $\underline{u}=0$ then again $\underline{t}\wedge \underline{u}$ is $0$-compatible with $B$ and we see immediately that $\tilde{x}=x$, $\tilde{L}=L$ and hence ${\ensuremath{\mathcal{A}}}_{q}^{\underline{t},0}={\ensuremath{\mathcal{A}}}_{q}$. That is, when $\underline{u}=0$ the skew-Laurent extension has no interaction with the subalgebra ${\ensuremath{\mathcal{A}}}_{q}$ (therefore the choice of $\underline{t}$ is irrelevant).
3. If $\underline{u}=\lambda\underline{t}$ for some $\lambda\in {\ensuremath{\mathbb{Z}}}$ then as above, $\underline{t}\wedge \underline{u}=0$.
We observe that if $\underline{t}$ and $\underline{u}$ are gradings for $B$, so that $B_{i}\underline{t}=B_{i}\underline{u}=0$ for all mutable indices $i$, then $$B_{i}(\underline{t}\wedge \underline{u})={{(B_{\text{mut}})}^{T}}_{i}(\underline{t}\,{{\underline{u}}^{T}}-\underline{u}\,{{\underline{t}}^{T}})=0$$ so that $\underline{t}\wedge \underline{u}$ is $0$-compatible with $B$.
Under certain conditions, the above re-scaling in fact produces a quantum cluster algebra that is a subalgebra of the skew-Laurent extension itself, and not just of the skew-field of fractions of this. The next result gives just such a set of conditions. However in it we will need to introduce additional powers of $q$ to our re-scaling. The previous lemmas, as stated, still hold in this slightly more general setting: the proof of Lemma \[l:Lminuswedge\] requires a minor adjustment and Lemma \[l:recaled-compatibility\] then goes through verbatim.
\[t:rescaledQCA\] Let ${\ensuremath{\mathcal{A}}}_{q}={\ensuremath{\mathcal{A}}}_{q}(\underline{x}=(X_{1},\ldots,X_{r}),B,L,G)$ be a graded quantum cluster algebra. Also let $\underline{t}={{(t_{1},\dotsc ,t_{r})}^{T}}\in {\ensuremath{\mathbb{Z}}}^{r}$ and $\underline{u}={{(u_{1},\dotsc ,u_{r})}^{T}}\in {\ensuremath{\mathbb{Z}}}^{r}$ be such that $\underline{t}$ and $\underline{u}$ are gradings for $B$, i.e. $B_{i}\underline{t}=B_{i}\underline{u}=0$ for every mutable index $i$.
Then the following initial data determine a graded quantum cluster algebra $\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}}={\ensuremath{\mathcal{A}}}_{q}(\underline{\tilde{x}},\tilde{B},\tilde{L},\tilde{G})$ where
- $\underline{\tilde{x}}=(q^{\frac{t_{1}u_{1}}{2}}X_{1}z^{u_{1}},\ldots,q^{\frac{t_{r}u_{r}}{2}}X_{r}z^{u_{r}})$,
- $\tilde{B}=B$,
- $\tilde{L}=L-\underline{t}\wedge \underline{u}$, and
- $\tilde{G}=G+\underline{u}$,
with $\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}}$ a subalgebra of the skew-Laurent extension ${\ensuremath{\mathcal{A}}}_{q}[z^{\pm 1};\tau]$ of ${\ensuremath{\mathcal{A}}}_{q}$ whose automorphism $\tau$ is induced by $\underline{t}$, i.e. $\tau\colon {\ensuremath{\mathcal{A}}}_{q} \to {\ensuremath{\mathcal{A}}}_{q}$ is the algebra automorphism such that for each $1\leq i \leq r$, $\tau(X_{i})=q^{t_{i}}X_{i}$.
Our strategy for this proof will be to consider seeds augmented by the extra data assumed in the theorem and to show that these extended seeds behave appropriately under mutation. More precisely, consider as initial data the tuple $$(\underline{\tilde{x}}=(\tilde{X}_{1},\dotsc ,\tilde{X}_{r}),\tilde{B},\tilde{L},\tilde{G},\underline{t},\underline{u}),$$ where each component is as defined in, and satisfies the conditions of, the statement of the theorem and in particular $$\tilde{X}_{i}=q^{\frac{t_{i}u_{i}}{2}}X_{i}z^{u_{i}}.$$ We will call such a tuple a *re-scaled* seed and we first establish that this data is valid for defining a graded quantum cluster algebra.
We observed above that $\underline{t}$ and $\underline{u}$ being gradings implies that $\underline{t}\wedge \underline{u}$ is $0$-compatible and so we apply the lemmas to see that $\tilde{B}$ and $\tilde{L}$ are compatible. Since $\tilde{B}=B$ and $\tilde{G}=G+\underline{u}$ with both $G$ and $\underline{u}$ being gradings for $B$, the grading condition follows. That is, the first four components of the re-scaled initial seed are indeed valid data for the construction of a graded quantum cluster algebra. We note that the choice of $\tilde{G}=G+\underline{u}$ is natural, setting the degree of the re-scaled variable $q^{\frac{t_{i}u_{i}}{2}}X_{i}z^{u_{i}}$ to be the sum of the degree of $X_{i}$ (as described by $G$) and the power of $z$, namely $u_{i}$.
Mutation of re-scaled seeds is defined in the obvious way: the cluster $\tilde{\underline{x}}$ is mutated via the exchange relations determined by $\tilde{B}$ and $\tilde{L}$ as usual, $\tilde{B}$ and $\tilde{L}$ are mutated in the same way as for ungraded quantum cluster algebras via the corresponding matrix $E$ and the three gradings $\tilde{G}$, $\underline{t}$ and $\underline{u}$ are mutated as described in Section \[s:gradedQCAs\], namely by multiplication by ${{E}^{T}}$. Then Lemma \[l:mutofgrading\] assures us that the mutations of $\underline{t}$ and $\underline{u}$ are gradings for the corresponding mutation of $\tilde{B}$, so we see that the mutation of a re-scaled seed again satisfies the compatibilities and assumptions of the theorem, except that we need to see that the form of the mutated cluster variables is the same as that described above. That is, we wish to show that $\tilde{X}_{i}^{\prime}=q^{\frac{t_{i}^{\prime}u_{i}^{\prime}}{2}}(X_{i}^{\prime})z^{u_{i}^{\prime}}$.
In order to verify this we argue similarly to the proof of Proposition \[p:rescaleQCAbypowersofq\] and first consider the exchange monomials arising from a re-scaled seed, i.e. $$\tilde{M}(a_{1},\dotsc ,a_{r}) {\stackrel{\scriptscriptstyle{\mathrm{def}}}{=}}q^{\frac{1}{2}\sum_{i<j} a_{i}a_{j}\tilde{l}_{ji}}\tilde{X}_{1}^{a_{1}}\dotsm \tilde{X}_{r}^{a_{r}}.$$ We have $\tilde{X}_{i}=q^{\frac{t_{i}u_{i}}{2}}X_{i}z^{u_{i}}$ and it is straightforward to verify from the quasi-commutation relations between $z$ and the $X_{i}$ and the equation $\tilde{L}=L-\underline{t}\wedge \underline{u}$ that we have $$\tilde{M}(a_{1},\dotsc ,a_{r})=q^{\frac{1}{2}\sum_{i=1}^{r}\sum_{j=1}^{r} a_{i}a_{j}t_{i}u_{j}}M(a_{1},\dotsc ,a_{r})z^{\sum_{i=1}^{r} u_{i}a_{i}}.$$ Details may be found in the Appendix on page . Writing $\underline{a}=(a_{1},\dotsc ,a_{r})$, we may reformulate this as $$\tilde{M}(\underline{a})=q^{\frac{(\underline{a}\cdot \underline{t})(\underline{a} \cdot \underline{u})}{2}}M(\underline{a})z^{\underline{a}\cdot \underline{u}}.$$
Then for $k$ a mutable index, recalling that $\tilde{B}=B$ we see that mutation in the direction $k$ from the re-scaled seed yields $$\begin{aligned}
\tilde{X}_{k}^{\prime} & =\tilde{M}(\underline{b}_{k}^{+})+\tilde{M}(\underline{b}_{k}^{-}) \\
& = q^{\frac{(\underline{b}_{k}^{+}\cdot \underline{t})(\underline{b}_{k}^{+} \cdot \underline{u})}{2}}M(\underline{b}_{k}^{+})z^{\underline{b}_{k}^{+}\cdot \underline{u}}+q^{\frac{(\underline{b}_{k}^{-}\cdot \underline{t})(\underline{b}_{k}^{-} \cdot \underline{u})}{2}}M(\underline{b}_{k}^{-})z^{\underline{b}_{k}^{-}\cdot \underline{u}} \\
& = q^{\frac{{t}_{k}^{\prime}{u}_{k}^{\prime}}{2}}M(\underline{b}_{k}^{+})z^{{u}_{k}^{\prime}}+q^{\frac{{t}_{k}^{\prime}{u}_{k}^{\prime}}{2}}M(\underline{b}_{k}^{-})z^{{u}_{k}^{\prime}} \\
& = q^{\frac{{t}_{k}^{\prime}{u}_{k}^{\prime}}{2}}\left(M(\underline{b}_{k}^{+})+M(\underline{b}_{k}^{-})\right)z^{{u}_{k}^{\prime}} \\
& = q^{\frac{{t}_{k}^{\prime}{u}_{k}^{\prime}}{2}}(X_{k}^{\prime})z^{{u}_{k}^{\prime}} \end{aligned}$$ as desired. Here we have again used the fact that $\underline{b}_{k}^{+}\cdot \underline{v}=\underline{b}_{k}^{-}\cdot \underline{v}$ for any grading $\underline{v}$ for $B$ and the equality of both of these with the $k$th entry of the mutation of $\underline{v}$ (as noted after Lemma \[l:mutofgradinggivenbyE\]), as well as the fact that the corresponding exchange relation in the original quantum cluster algebra ${\ensuremath{\mathcal{A}}}_{q}$ is $X_{k}^{\prime}=M(\underline{b}_{k}^{+})+M(\underline{b}_{k}^{-})$. Note that the power of $z$ occurring is exactly the degree of $\tilde{X}_{k}^{\prime}$ (or equivalently of $X_{k}^{\prime}$) for the grading induced by $\underline{u}$.
That is, mutation of a re-scaled seed produces another re-scaled seed. Therefore iterated mutation from the re-scaled seed of the statement produces a graded quantum cluster algebra all of whose quantum cluster variables are contained in the skew-Laurent extension ${\ensuremath{\mathcal{A}}}_{q}[z^{\pm 1};\tau]$, i.e. no localisation of the latter is required.
From the proof of this theorem, we see the following.
\[c:formofrescaledQCV\] With notation as in the preceding Theorem, there is a bijection ${\varphi}$ between the sets of quantum cluster variables for the quantum cluster algebras ${\ensuremath{\mathcal{A}}}_{q}$ and $\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}}$. Furthermore, under this bijection, every quantum cluster variable $\tilde{X}$ of the quantum cluster algebra $\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}}$ is of the form $q^{a}Xz^{b}$ with $a,b\in {\ensuremath{\mathbb{Z}}}$ and $X={\varphi}^{-1}(\tilde{X})$ the corresponding quantum cluster variable in ${\ensuremath{\mathcal{A}}}_{q}$.
The quantum cluster algebra structure $\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}}$ from this theorem is a quantum cluster algebra structure on a proper subalgebra of the skew-Laurent extension ${\ensuremath{\mathcal{A}}}_{q}[z^{\pm 1};\tau]$; in general this subalgebra is not ${\ensuremath{\mathcal{A}}}_{q}$. However we could easily extend this to a quantum cluster algebra structure on the whole skew-Laurent extension by adding $z$ and $z^{-1}$ as coefficients in the same manner as in Proposition \[p:skewLaurent-extra-coeffs\], as the only issue here is the absence of the generators $z$ and $z^{-1}$.
We note that if $A$ is a $\mathbb{K}$-algebra and $\tau$ an automorphism of $A$ then there is an algebra isomorphism of $\left((A[z_{1}^{\pm 1};\tau])[z_{2}^{\pm 1};\tau]\right)/(z_{1}-z_{2})$ with $A[z_{1}^{\pm 1};\tau]$. That is, if we make a two-fold skew-Laurent extension of $A$ using first the automorphism $\tau$ and then using $\tau$ extended to $A[z_{1}^{\pm 1};\tau]$ as the identity on $z_{1}^{\pm 1}$, then taking the quotient that identifies the two variables yields an algebra isomorphic to a single extension using $\tau$.
The next lemma observes that one may reverse the above scaling procedure.
Let ${\ensuremath{\mathcal{A}}}_{q}={\ensuremath{\mathcal{A}}}_{q}(\underline{x},B,L,G)$ be a graded quantum cluster algebra and let $\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}} \subseteq {\ensuremath{\mathcal{A}}}_{q}[z^{\pm 1};\tau]$ be the quantum cluster algebra structure obtained from ${\ensuremath{\mathcal{A}}}_{q}$ by the construction of the preceding theorem. Then $(\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}})^{\underline{t},-\underline{u}}={\ensuremath{\mathcal{A}}}_{q}$, where the former is viewed as a subalgebra of ${\ensuremath{\mathcal{A}}}_{q}[z^{\pm 1};\tau]$ under the isomorphism described above.
The quantum cluster algebra $(\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}})^{\underline{t},-\underline{u}}$ is obtained from $\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}}$ by the construction of the theorem. Using the isomorphism of the quotient of the two-fold extension with the single extension described above we may abuse notation and write $z$ for both extending variables. We then see that $(\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}})^{\underline{t},-\underline{u}}$ has initial data
- $\tilde{\tilde{\underline{x}}}=(q^{\frac{-t_{1}u_{1}}{2}}(q^{\frac{t_{1}u_{1}}{2}}X_{1}z^{u_{1}})z^{-u_{1}},\dotsc ,q^{\frac{-t_{r}u_{r}}{2}}(q^{\frac{t_{r}u_{r}}{2}}X_{r}z^{u_{r}})z^{-u_{r}})=(X_{1},\dotsc ,X_{r})=\underline{x}$,
- $\tilde{\tilde{B}}=B$,
- $\tilde{\tilde{L}}=(L-\underline{t}\wedge \underline{u})-\underline{t}\wedge (-\underline{u})=L$ and
- $\tilde{\tilde{G}}=(G+\underline{u})+(-\underline{u})=G$.
That is, $(\tilde{{\ensuremath{\mathcal{A}}}}_{q}^{\underline{t},\underline{u}})^{\underline{t},-\underline{u}}$ has the same initial data as ${\ensuremath{\mathcal{A}}}_{q}$ so yields the same quantum cluster algebra.
This completes our general theory of graded quantum cluster algebras. Now we turn to our application, the existence of a quantum cluster algebra structure on the quantum Grassmannians.
The quantum cluster algebra structure on quantum matrices {#s:KqMatisQCA}
=========================================================
As noted above, the work of Geiß, Leclerc and Schröer ([@GLS-QuantumPFV Corollary 12.10]) has given a quantum cluster algebra structure on quantum matrices ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$. We use $j$ rather than $n$ here as this is the notation of [@GLS-QuantumPFV] and also we will want to consider ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ and its relationship with ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ subsequently; it will simplify the presentation in this section to use $j$ rather than $n-k$.
Our aim is to lift this to a quantum cluster algebra structure on the corresponding quantum Grassmannian ${\mathbb{K}_{q}[\mathrm{Gr}(k,k+j)]}$, in a similar fashion to [@GLS-PFV §10], so we record here the initial data provided by the construction in [@GLS-QuantumPFV]. (This section is an expansion of [@GLS-QuantumPFV Example 12.11], which describes the case $k=j=3$.)
Let $m=k+j-1$; for comparison with [@GLS-QuantumPFV §12.4], our parameter $m$ is their $n$. The construction of the quantum cluster algebra structure on ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$ is via the module category of the preprojective algebra $\Lambda=\Lambda(A_{m})$ associated to the Dynkin diagram $A_{m}$. For a description of the algebra $\Lambda$ and its representation theory, including Auslander–Reiten quivers for $2\leq m \leq 4$, we refer the reader to [@GLS-Semicanonical]. In what follows, we will state well-known properties of this algebra and its module category without proof.
We first need to construct the projective modules for $\Lambda$. A basis for the $i$th projective module $P_{i}$ is given by the set of paths leaving the vertex $i$ (modulo the preprojective relation). The $r$th Loewy layer of $P_{i}$ consists of the simple modules corresponding to the vertices at the ends of paths of length $r-1$ and so we see that $P_{i}$ has simple socle $i$ and simple top $m-i+1$. Here we use the common notation of having the vertex labels denote the simple modules for path algebras and quotients of these (see for example [@ASS-Book]); we will also use $S_{i}$ for this when this is clearer.
In general $P_{i}$ has the form of an $i$ by $m-i+1$ rectangle:
Note that if $k=j$ then $m=2k-1$ is odd and $P_{k}$ is self-dual—the rectangular shape depicted above is then a square. Hence for $k=j=3$, so $m=5$, the projective modules are
Now it is well-known that ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$ is isomorphic to the algebra $U_{q}(\mathfrak{n}(w))$ (also commonly denoted $U_{q}^{+}[w]$) associated to $\mathfrak{g}=\mathfrak{sl}_{m+1}$ where $w$ is the Weyl group word with reduced decomposition $$w=(s_{j}s_{j-1}\cdots s_{1})(s_{j+1}s_{j}\cdots s_{2})\cdots (s_{m}s_{m-1}\cdots s_{k}).$$ This may be found in [@MeriauxCauchon], for example. Let $$\underline{i}=(k,k+1,\ldots,m,k-1,k,\ldots,m-1,\ldots,1,2,\ldots,j)$$ be the sequence of indices for the above reduced decomposition for $w$; note that we have chosen the reverse order to that in [@GLS-QuantumPFV §12.4]. It is convenient to render this as a $k{\times}j$ matrix $(i_{(\alpha,\beta)})$ with $i_{(\alpha,\beta)}=k-\alpha+\beta$, for $1\leq \alpha \leq k$, $1 \leq \beta \leq j$, [i.e. ]{}$$\underline{i} = \begin{pmatrix} k & k+1 & \cdots & m-1 & m \\ k-1 & k & \cdots & m-2 & m-1 \\ \vdots & \vdots & \ddots & \vdots & \vdots \\ 2 & 3 & \cdots & j & j+1 \\ 1 & 2 & \cdots & j-1 & j \end{pmatrix}.$$
There is a natural total order on the set of indices of the matrix $\underline{i}$, given by $(\alpha,\beta)<(\alpha',\beta')$ if and only if $\alpha<\alpha'$ or $(\alpha=\alpha'$ and $\beta<\beta')$; that is, the ordering is lexicographical in each coordinate, taking the first coordinate first. We extend this to a total order on the set $\{ (\alpha,\beta) \mid 1\leq \alpha \leq k,\ 1\leq \beta \leq j\} {\mathrel{\cup}}\{0,kj+1\}$ by $0<(\alpha,\beta)<kj+1$ for all pairs $(\alpha,\beta)$. For an index $(\alpha,\beta)$, we define $$\begin{aligned}
(\alpha,\beta)^{-} & = & \max \left( \{ 0 \} {\mathrel{\cup}}\left\{ (\gamma,\delta)<(\alpha,\beta) \mid i_{(\gamma,\delta)}=i_{(\alpha,\beta)} \right\}\right) \\ & = & \begin{cases} 0 & \text{if}\ \alpha=1\ \text{or}\ \beta=1 \\ (\alpha-1,\beta-1) & \text{otherwise} \end{cases} \\
(\alpha,\beta)^{+} & = & \min\left(\{kj+1\} {\mathrel{\cup}}\left\{(\gamma,\delta)>(\alpha,\beta) \mid i_{(\gamma,\delta)}=i_{(\alpha,\beta)} \right\}\right) \\ & = & \begin{cases} kj+1 & \text{if}\ \alpha=k\ \text{or}\ \beta=j \\ (\alpha+1,\beta+1) & \text{otherwise} \end{cases}\end{aligned}$$ The frozen indices (that is, those indices that correspond to coefficients in the quantum cluster algebra structure) are exactly the $(\alpha,\beta)$ with $(\alpha,\beta)^{+}=kj+1$, [i.e. ]{}when $\alpha=k$ or $\beta=j$.
The initial seed is constructed from the module category as follows. The subcategory of $\mathrm{mod}(\Lambda)$ corresponding to the word $w$ above, which we denote by ${\ensuremath{\mathcal{C}}}_{w}$, is the subcategory generated by the projective module $P_{k}$. Certain quotients of $P_{k}$ give the modules corresponding to the standard generators $X_{ab}$ of ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$ and an iterated socle construction is used to produce modules in this subcategory that correspond to elements of the initial seed. More precisely, for each pair $(a,b)$ with $1\leq a \leq k$ and $1\leq b \leq j$, the module $P_{k}$ has a unique quotient $M_{(a,b)}$ whose dimension vector is $\underline{\boldsymbol{e}}_{a}+\underline{\boldsymbol{e}}_{a+1}+\cdots +\underline{\boldsymbol{e}}_{m-b+1}$ and this quotient corresponds to $X_{ab}$. From the above description of $P_{k}$ we see that these modules correspond to segments of the top edges of the rectangle describing $P_{k}$ that include the top (which is isomorphic to $S_{m-k+1}=S_{j}$). For $k=j=3$ ($m=5$) we have
To construct the modules corresponding to the initial seed, we need the following construction. Given a module $W$, we define
- $\mathrm{soc}_{(l)}(W){\stackrel{\scriptscriptstyle{\mathrm{def}}}{=}}{\displaystyle \sum_{\substack{U\leq W \\ U{\ensuremath \cong}S_{l}}} U}$ and
- $\mathrm{soc}_{(l_{1},l_{2},\ldots,l_{s})}(W){\stackrel{\scriptscriptstyle{\mathrm{def}}}{=}}W_{s}$ where the chain of submodules $0\subseteq W_{1} \subseteq W_{2} \subseteq \cdots \subseteq W_{s} \subseteq W$ is such that $W_{p}/W_{p-1} {\ensuremath \cong}\mathrm{soc}_{(l_{p})}(W/W_{p-1})$.
Then for $1\leq s \leq l(w)=kj$, we define $V_{s} {\stackrel{\scriptscriptstyle{\mathrm{def}}}{=}}\mathrm{soc}_{(i_{s},i_{s-1},\ldots,i_{1})}(P_{i_{s}})$. Thus for $k=j=3$ and our choice of reduced expression $\underline{i}$ above, $V_{1}=\mathrm{soc}_{(3)}(P_{3})=\mathrm{soc}(P_{3})=S_{3}$.
Similarly $V_{2}=\mathrm{soc}_{(4,3)}(P_{4})$ is defined by the chain $0\subseteq W_{1} \subseteq W_{2}=V_{2} \subseteq P_{4}$ with $W_{1}=\mathrm{soc}_{(4)}(P_{4})=S_{4}$ and $W_{2}/W_{1}=\mathrm{soc}_{(3)}(P_{4}/W_{1})=S_{3}$; that is, $V_{2}$ has two layers, a simple top and a simple socle isomorphic to $S_{3}$ and $S_{4}$ respectively. Arranging the modules $V_{s}$ in the same way as we did for the indices $i_{s}$, it is natural to re-number these as $V_{(\alpha,\beta)}$ for $1\leq \alpha \leq k$, $1\leq \beta \leq j$, and we see that the modules corresponding to the initial seed for this case are as follows:
To obtain the element of ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$ corresponding to the modules $V_{(\alpha,\beta)}$, we note that the construction of the $V_{(\alpha,\beta)}$ is such that $V_{(\alpha,\beta)}/V_{(\alpha,\beta)^{-}}=V_{(\alpha,\beta)}/V_{(\alpha-1,\beta-1)} {\ensuremath \cong}M_{(k-\alpha+1,n-\beta+1)}$ (the natural indexings of the $V_{(\alpha,\beta)}$ and the $M_{(a,b)}$ are opposed to each other, unfortunately). Then returning to our running example with $k=j=3$, $V_{(1,1)}/0=M_{(3,3)}$ so $V_{(1,1)}$ corresponds to $X_{33}$. A module $V_{(\alpha,\beta)}$ need not correspond to a generator: $V_{(2,2)}$ is an extension of $M_{(3,3)}$ by $M_{(2,2)}$ and corresponds to the quantum minor ${\genfrac{[}{]}{0em}{1}{23}{23}}$. Similarly $V_{(3,3)}=P_{3}$ is an extension of $V_{(2,2)}$ by $M_{(1,1)}$ and corresponds to the quantum minor ${\genfrac{[}{]}{0em}{1}{123}{123}}$.
We may describe the initial cluster coming from this construction, which we will call ${\ensuremath{\mathcal{M}}}(k,j)$, explicitly as follows.
\[d:qGLScluster\] For $1\leq r\leq k$ and $1\leq s \leq j$, define the sets $$\begin{aligned}
R(r,s) & = \{ k-r+1,k-r+2,\ldots ,k-r+s \} {\mathrel{\cap}}\{1,\ldots,k \} \\
C(r,s) & = \{ j-s+1,j-s+2,\ldots ,j-s+r \} {\mathrel{\cap}}\{1,\ldots,j \} \end{aligned}$$ Then we define ${\ensuremath{\mathcal{M}}}(k,j)=\{ {\genfrac{[}{]}{0em}{1}{C(r,s)}{R(r,s)}} \mid 1\leq r\leq k, 1\leq s\leq j\}$. It is natural to give ${\ensuremath{\mathcal{M}}}(k,j)$ as a $k{\times}j$ array (as we have for $\underline{i}$), where its $(r,s)$-entry, which we denote ${\ensuremath{\mathcal{M}}}_{kj}(r,s)$, is the quantum minor with row set $R(r,s)$ and column set $C(r,s)$. Should we need to consider ${\ensuremath{\mathcal{M}}}(k,j)$ as a sequence, its $\left((r-1)j+s\right)$-entry is ${\genfrac{[}{]}{0em}{1}{C(r,s)}{R(r,s)}}$.
The above association of modules to minors follows from well-known isomorphisms, such as the isomorphism of ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$ with $U_{q}(\mathfrak{n}(w))$ for the above $w$ as in [@MeriauxCauchon], and those in the paper [@GLS-QuantumPFV]. We note that for $\mathbb{K}_{q}[SL_{m+1}]$, the generalized quantum minors of [@BZ-QCA] are the usual quantum minors (analogous to the fact that the generalized minors of Fomin and Zelevinsky [@FZ-DoubleBruhat] coincide with the usual ones for $SL_{m+1}$). Then the unipotent quantum minors in the paper of Geiß, Leclerc and Schröer ([@GLS-QuantumPFV §5]) are generalized quantum minors divided by certain principal quantum minors. Following through the correspondence of these with dual PBW basis elements (in $U_{q}(\mathfrak{n}(w))$) and thence through the isomorphism of Mériaux and Cauchon ([@MeriauxCauchon]), we do indeed obtain the (usual) quantum minors in ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$.
The arrows in the exchange quiver for the initial seed are given by the combinatorial data associated to the reduced expression $\underline{i}$. Following through the definitions in [@GLS-QuantumPFV §9.4] in this case yields the following description of these:
- $(\alpha,\beta) \to (\alpha,\beta+1)$,
- $(\alpha,\beta) \to (\alpha+1,\beta)$ and
- $(\alpha,\beta) \to (\alpha-1,\beta-1)$,
where an arrow is defined only if both its start and end points are (thus there is no arrow $(1,1)\to (0,0)$, for example) and any arrows between indices for coefficients are suppressed. We note that these are exactly opposed to the natural inclusion and projection homomorphisms on the corresponding modules.
The quasi-commutation data is also encoded categorically: indexing by pairs as above, the matrix $L$ has entries $$l_{(\alpha,\beta),(\gamma,\delta)}=\dim \mathrm{Hom}_{\Lambda}(V_{(\alpha,\beta)},V_{(\gamma,\delta)})-\dim \mathrm{Hom}_{\Lambda}(V_{(\alpha,\beta)},V_{(\gamma,\delta)}).$$ Alternatively, this data can be obtained combinatorially ([@GLS-QuantumPFV Proposition 10.3]). The compatibility of the matrix corresponding to the arrows in the exchange quiver and the quasi-commutation matrix is shown in Proposition 10.1 of [@GLS-QuantumPFV].
Putting this all together, the initial cluster variables and their exchange quiver in ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$ are as illustrated in Figure \[fig:initialseed\]. In Figure \[fig:initialseed33\] we show this for our running example with $k=j=3$.
[ ]{}
[ ]{}
As is usual, vertices of the quiver that are frozen, [i.e. ]{}corresponding to elements that are coefficients and so not mutated, are boxed (a so-called ice quiver). We do not record here explicitly the quasi-commutation matrix. We denote this initial data for the quantum cluster algebra structure on ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$ as $({\ensuremath{\mathcal{M}}}(k,j),B(k,j),L(k,j))$, where ${\ensuremath{\mathcal{M}}}(k,j)$ is the initial cluster as above and $B$ and $L$ are the exchange and quasi-commutation matrices.
Then the main theorem of [@GLS-QuantumPFV], Theorem 12.3, tells us that in this case, with the above initial data, ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$ is a quantum cluster algebra. We note particularly Corollary 12.4 of [@GLS-QuantumPFV] which says that every relevant unipotent quantum minor occurs as a quantum cluster variable in this quantum cluster algebra structure. That is, in our particular case, every quantum minor in ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$ does indeed occur as a quantum cluster variable. Of course, outside the finite-type cases, we must have quantum cluster variables that are not quantum minors; we will say a little more about these below.
We observe that this quantum cluster algebra structure can be considered as a graded quantum cluster algebra structure, with respect to the natural choice of grading. We have that ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$ is an ${\ensuremath \mathbb{N}}$-graded algebra when we put all the matrix generators $X_{ij}$ in degree 1. Indeed, our choice of initial seed consists of homogeneous elements for this grading, as follows.
\[lemma:degofMrs\] We have ${\lvert R(r,s) \rvert}={\lvert C(r,s) \rvert}=\min(r,s)$, and so $\mathrm{deg}({\ensuremath{\mathcal{M}}}_{kj}(r,s))=\min(r,s)$.
So we set $G(k,j)$ to be the vector whose $(r,s)$-entry is equal to $\min(r,s)$. Furthermore, we see in the next lemma that the exchange quiver satisfies the required homogeneity property with respect to this grading.
\[lemma:degsumsequal\] At any mutable vertex $(\alpha,\beta)$, $$\sum_{(\gamma,\delta)\to (\alpha,\beta)} \deg({\ensuremath{\mathcal{M}}}_{kj}(\gamma,\delta))=\sum_{(\alpha,\beta)\to (\gamma,\delta)} \deg({\ensuremath{\mathcal{M}}}_{kj}(\gamma,\delta)).$$
For $\alpha=\beta=1$, we see that the two sums are equal to $2$.
Next assume that $\alpha=1$ and $\beta>1$. Then the vertices with arrows incoming to $(1,\beta)$ are $(1,\beta-1)$ and $(2,\beta+1)$, and the vertices with arrows outgoing from $(1,\beta)$ are $(1,\beta+1)$ and $(2,\beta)$. Since $\deg({\ensuremath{\mathcal{M}}}_{kj}(r,s))=\min(r,s)$, we see that the two sums are both equal to $1+2=3$, as $\beta>1$. Similarly the sums are equal (and equal to $3$) if $\alpha>1$ and $\beta=1$.
If $\alpha,\beta >1$, then $(\alpha,\beta)$ has six neighbours: $(\alpha,\beta-1)$, $(\alpha-1,\beta)$ and $(\alpha+1,\beta+1)$ with incoming arrows, and $(\alpha-1,\beta-1)$, $(\alpha,\beta+1)$ and $(\alpha+1,\beta)$ outgoing. Then if $\alpha=\beta$, the two sums are easily seen to be equal to $3\alpha-1$, or if $\alpha<\beta$ the sums are equal to $3\alpha$, or if $\beta<\alpha$ they are equal to $3\beta$.
Then by our earlier discussion, the quantum cluster algebra associated to the initial quantum seed $({\ensuremath{\mathcal{M}}}(k,j),B(k,j),L(k,j),G(k,j))$ is ${\ensuremath{\mathbb{Z}}}$-graded and in particular every quantum cluster variable is homogeneous with respect to this grading. Note that *a priori* we only deduce a ${\ensuremath{\mathbb{Z}}}$-grading.
This grading also has a categorical interpretation. As described in §9.6 of [@GLS-QuantumPFV], drawing on [@GLS-KacMoody [§10]{}], every module $X$ in ${\ensuremath{\mathcal{C}}}_{w}$ has a filtration $$0=X_{0} \subseteq X_{1} \subseteq X_{2} \subseteq \cdots \subseteq X_{r}=X$$ such that each subquotient $X_{i}/X_{i-1}$ is isomorphic to $M_{i}^{m_{i}}$ where $M_{i}$ is the module $M_{(k-\alpha+1,j-\beta+1)}$ corresponding to $V_{i}=V_{(\alpha,\beta)}$ (where $V_{k}$ was our original numbering for the $V$’s, coming from $\underline{i}$). Hence each module in ${\ensuremath{\mathcal{C}}}_{w}$ has an $M$-dimension vector, $\underline{m}(X)=(m_{1},\dotsc ,m_{r})\in {\ensuremath \mathbb{N}}^{r}$.
The general theory tells us that the modules $V_{(\alpha,\beta)}$ can be considered as being built up by repeated extensions of the modules $M_{(a,b)}$ (corresponding to the algebraic identification of $V_{(\alpha,\beta)}$ being a minor and thus a product of the matrix generators, to which the $M_{(a,b)}$ correspond). In the case at hand, we see that the minor corresponding to $V_{(\alpha,\beta)}$ is an $l{\times}l$ minor exactly when $V_{(\alpha,\beta)}$ has $\sum_{i=1}^{r} \underline{m}(V_{(\alpha,\beta)})_{i}=l$, i.e. when $V_{(\alpha,\beta)}$ has $l$ non-zero subquotients of the form described in the previous paragraph. We call this the sum of the entries of the $M$-dimension vector the $M$-dimension of the module.
We see that the initial exchange matrix (or quiver) has the necessary property to imply that this gives a grading by looking at the explicit description of the arrows. For example, the arrow $(\alpha,\beta)\to (\alpha-1,\beta-1)$ exactly corresponds to the inclusion homomorphism $V_{(\alpha-1,\beta-1)}\to V_{(\alpha,\beta)}$ for which $M_{(k-\alpha+1,j-\beta+1)}$ is the cokernel, thus $V_{(\alpha,\beta)}$ has $M$-dimension one greater than that of $V_{(\alpha-1,\beta-1)}$. One see that in the grid arrangement, $M$-dimension is constant along rows and increases by one on going down a row. Away from the boundary, every mutable vertex has the same number of arrows coming in from a given row as going out to it (either zero or one of each, in fact) and so the two sums of $M$-dimensions over arrows entering or leaving the vertex are equal. It is straightforward to check that the boundary cases also have the required property.
Indeed the fact that every module in ${\ensuremath{\mathcal{C}}}_{w}$ has a filtration with subquotients the modules $M_{(a,b)}$ makes it clear that this grading is the usual ${\ensuremath \mathbb{N}}$-grading on ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$, for we have these modules $M_{(a,b)}$ in degree one as for the matrix generators. Thus we can view the above statement as saying that the quantum cluster algebra structure is compatible with the natural graded algebra structure of ${\mathbb{K}_{q}[\mathrm{M}(k,j)]}$. This will be important for us later. Again, we see that this is a property of the category that does not rely on being in the quantum case: this grading is present whether one considers ${\ensuremath{\mathcal{C}}}_{w}$ to be categorifying the commutative or the quantum coordinate ring.
Finally,\[page:theta-start\] we note one more grading-like datum associated to the category ${\ensuremath{\mathcal{C}}}_{w}$. Namely, following [@GLS-PFV §10], to each module $M$ in ${\ensuremath{\mathcal{C}}}_{w}$ we may associate the natural number given by $\theta(M)=\dim \mathrm{Hom}_{\Lambda}(M,S_{j})$, where $j=m-k+1$. Then $\theta(M)$ is the multiplicity of $S_{k}$ in the top of the module $M$ and we see from the above that $\theta(V_{(\alpha,\beta)})=\theta(M_{(\alpha,\beta)})=1$ for all $1\leq \alpha \leq k$, $1\leq \beta \leq j$. However $\theta$ is not always equal to 1: in the example of ${\mathbb{K}_{q}[\mathrm{M}(3,3)]}$, mutating $V_{(2,2)}$ yields a module $W$ with $\theta(W)=2$.
This data has the property that it is compatible with mutation, in the following sense: if $M^{\prime}$ is the module obtained by mutating $M$, so that there exist two exact sequences $$0 \to M \to U \to M^{\prime} \to 0 \qquad \text{and} \qquad 0 \to M^{\prime} \to W \to M \to 0$$ where $U$ and $W$ correspond to the exchange monomials, then $$\dim \mathrm{Hom}_{\Lambda}(M^{\prime},S_{j})=\max \{ \dim \mathrm{Hom}_{\Lambda}(U,S_{j}), \dim \mathrm{Hom}_{\Lambda}(W,S_{j})\}-\dim \mathrm{Hom}_{\Lambda}(M,S_{j}).$$
An analogue of this formula is stated as [@GLS-PFV Proposition 10.1], for socles as opposed to tops. The paper [@GLS-PFV], in which the classical version of the topic of this paper is considered, works with a category of submodules whereas the quantum version in [@GLS-QuantumPFV] uses a category of factor modules. Consequently, to fit with [@GLS-QuantumPFV] we need to look at tops here as opposed to socles.
More compactly, in the above notation, $\theta(M^{\prime})=\max\{ \theta(U),\theta(W) \}-\theta(M)$. That is, given the values of $\theta$ on a collection of modules associated to an initial cluster, one may calculate the values on all modules associated to cluster variables. Notice that because $\theta$ is a dimension, it necessarily takes natural number values; that the formula $\theta(M^{\prime})=\max\{ \theta(U),\theta(W) \}-\theta(M)$ produces this is not *a priori* clear.
Note also that this data is not a grading for the (quantum) cluster algebra structure above. Indeed, at the vertex indexed by $(1,1)$ we have two outgoing arrows to modules each of which has a 1-dimensional top but only one incoming arrow, from a module that also has a 1-dimensional top. That is, $\theta(U)\neq \theta(W)$ in this case, although the formula does tell us that the mutated module also has a 1-dimensional top. At all other mutable vertices for the cluster ${\ensuremath{\mathcal{M}}}(k,j)$ we do have homogeneity with respect to this function $\theta$, however.\[page:theta-end\]
The dehomogenisation isomorphism and the image of the cluster structure under this {#s:dehomog}
==================================================================================
In work of Kelly, Lenagan and Rigal ([@KellyLenaganRigal]), a noncommutative dehomogenisation of an ${\ensuremath \mathbb{N}}$-graded algebra is defined and their Corollary 4.1 describes an isomorphism of the localisation of the quantum Grassmannian at the minor $[(n-k+1)\dotsm n]$ with a skew-Laurent extension of a quantum matrix algebra. In [@LenaganRussell], a dehomogenisation isomorphism $\rho$ involving ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ localised at the consecutive minor $[\widetilde{a}\ (\widetilde{a+1})\ \cdots\ (\widetilde{a+k-1})]$ is constructed, where “$\,\widetilde{\ \ }\,$” indicates that values are taken modulo $n$ and from the set $\{1,\ldots,n\}$. In order to match conventions already fixed, we will need the map corresponding to the special case of the map $\rho$ of [@LenaganRussell] for the value $a=1$, the original work of [@KellyLenaganRigal] being the case $a=n-k+1$.
This map is key to the lifting procedure to obtain the quantum cluster algebra structure on the quantum Grassmannian, and we recall its definition. (Here, $\widehat{a}$ denotes an omitted index.)
\[p:dhom\] Let $\sigma$ be the automorphism of ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ defined by $\sigma(X_{ij})=qX_{ij}$. The map $$\alpha\colon {\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}[Y^{\pm 1}; \sigma] \to {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}\!\left[ [12\cdots k]^{-1} \right]$$ defined by $$\alpha(X_{ij}) = [1\cdots \widehat{k-i+1} \cdots k\ (j+k)][1\cdots k]^{-1}, \quad \alpha(Y) =[12\cdots k]$$ is an algebra isomorphism.
This map allows us to transport the quantum cluster algebra structure on quantum matrices, as described in the previous section, to the above localisation of the quantum Grassmannian. Set $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})={\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}\!\left[ [12\cdots k]^{-1} \right]$.
From generalities on noncommutative dehomogenisations (see [@KellyLenaganRigal §3]), the ${\ensuremath \mathbb{N}}$-grading on ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ that has all the generating Plücker coordinates $[I]$ in degree one gives rise to a ${\ensuremath{\mathbb{Z}}}$-grading on the localisation. Let ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}_{i}$ denote the degree $i$ homogeneous component of ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ for the aforementioned grading. Then $$\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})={\bigoplus}_{l\in {\ensuremath{\mathbb{Z}}}} \mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})_{l}$$ with $$\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})_{l}=\sum_{j\geq 0} {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}_{l+j}[1\dotsm k]^{-j}.$$ The noncommutative dehomogenisation of ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ is defined to be the degree 0 part of $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ with respect to this ${\ensuremath{\mathbb{Z}}}$-grading and the results of [@KellyLenaganRigal] and [@LenaganRussell] show that this degree 0 part is isomorphic to the quantum matrices ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$, via the map $\alpha$ of Proposition \[p:dhom\]. In particular, the map $\alpha$ sends an element of the quantum matrices to a $\mathbb{K}$-linear combination of elements of the form $m[1\dotsm k]^{-j}$ where $m$ is a homogeneous polynomial of degree $j$ in ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$.
We note that ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ is a subalgebra of $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ but that ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ has non-trivial intersection with every homogeneous component of $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$. More precisely, ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}_{l}\subseteq \mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})_{l}$ for every $l$; less formally, ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ is “spread out” across all the components of $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ and this is at the root of the technical difficulties that must be overcome in order to deduce our main result.
Now Proposition \[p:dhom\] describes the algebra $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ in terms of a skew-Laurent extension of the quantum matrices and so we can use Proposition \[p:skewLaurent-extra-coeffs\] to deduce the following.
The localisation $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ is a graded quantum cluster algebra.
We have seen in Section \[s:KqMatisQCA\] that ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ is a graded quantum cluster algebra with the initial data provided by the theorem of Geiß, Leclerc and Schröer and the grading being the standard grading on ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$. In particular, the initial cluster consists of homogeneous elements and the automorphism $\sigma$ in Proposition \[p:dhom\] therefore acts on the initial cluster variable ${\ensuremath{\mathcal{M}}}_{k(n-k)}(r,s)={\genfrac{[}{]}{0em}{1}{C(r,s)}{R(r,s)}}$ by multiplication by $q^{\deg({\ensuremath{\mathcal{M}}}_{k(n-k)}(r,s))}=q^{\min(r,s)}$. Thus the required conditions for applying Proposition \[p:skewLaurent-extra-coeffs\] hold and the skew-Laurent extension ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}[Y^{\pm 1};\sigma]$ induced by $\sigma$ is a graded quantum cluster algebra. Note that we have as additional coefficients $Y$ and $Y^{-1}$.
Since the map $\alpha$ in Proposition \[p:dhom\] is an algebra isomorphism, this structure is transported to $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$. The extra coefficients $Y$ and $Y^{-1}$ are mapped under $\alpha$ to $[1\dotsm k]$ and $[1\dotsm k]^{-1}$ respectively. We choose to place the coefficient $[1\dotsm k]$ at position $(1,k+1)$ and $[1\dotsm k]^{-1}$ at $(2,k+1)$; these choices are arbitrary.
We will denote by ${\ensuremath{\mathcal{L}}}(k,n)$ the union of the image of ${\ensuremath{\mathcal{M}}}(k,n-k)$ under $\alpha$ with the set $\{ [1\dotsm k],[1\dotsm k]^{-1} \}$. The corresponding exchange matrix will be denoted $B^{\mathrm{Loc}}(k,n)$; this matrix is obtained from the exchange matrix $B(k,n-k)$ by adding rows and columns indexed as $(1,k+1)$ and $(2,k+1)$ consisting of zeroes, corresponding to the two extra coefficients. The quasi-commutation matrix will be denoted $L^{\mathrm{Loc}}(k,n)$ and, as described in Proposition \[p:skewLaurent-extra-coeffs\], this is determined by $L(k,n-k)$ and the automorphism $\sigma$.
For our grading, described by $G^{\mathrm{Loc}}(k,n) \in {\ensuremath{\mathbb{Z}}}^{k(n-k)+2}$, we take the corresponding entry from $G(k,n-k)$ for elements of the image of ${\ensuremath{\mathcal{M}}}(k,n-k)$ under $\alpha$ and take $G^{\mathrm{Loc}}_{(1,k+1)}=1$ and $G^{\mathrm{Loc}}_{(2,k+1)}=-1$, in accordance with the natural choice. We note that this is not the natural grading on $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ described above: that would of course have the image of ${\ensuremath{\mathcal{M}}}(k,n-k)$ in degree 0, though it would agree with $G^{\mathrm{Loc}}_{(1,k+1)}=1$ and $G^{\mathrm{Loc}}_{(2,k+1)}=-1$. In making the choice of $G^{\mathrm{Loc}}(k,n)$ described here we are explicitly choosing to retain the grading associated to the pre-image under $\alpha$, in ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$.
Thus $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ has a quantum cluster algebra structure with initial data $$({\ensuremath{\mathcal{L}}}(k,n),B^{\mathrm{Loc}}(k,n),L^{\mathrm{Loc}}(k,n),G^{\mathrm{Loc}}(k,n)). \qedhere$$
However, this does not show that ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ is a quantum cluster algebra. We see from Proposition \[p:dhom\] that $\alpha$ maps each matrix generator $X_{ij}$ to the product of a quantum Plücker coordinate and the element $[1\cdots k]^{-1}$, and these are not elements of ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$, viewed as a subalgebra of the localisation in the obvious way. So although ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ is a subalgebra of a quantum cluster algebra, it is not a cluster subalgebra: the cluster variables are not elements of the subalgebra.
For later use, we record a special case of a companion result of Lenagan and Russell ([@LenaganRussell Proposition 3.3]) that describes the image of a quantum minor in ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ under the map $\alpha$.
First, for indexing sets $I=\{ i_{1},\ldots, i_{t}\}$ and $J=\{ j_{1},\ldots ,j_{t}\}$, define $$Q_{1}(I,J) =\{ \widetilde{j_{1}+k},\widetilde{j_{2}+k},\ldots,\widetilde{j_{t}+k}\} {\sqcup}\left( \{ 1,\ldots,k \} \setminus \{ k-i_{1}+1,k-i_{2}+1,\ldots k-i_{t}+1 \} \right)$$ It is straightforward to verify that this is a subset of $\{1,\ldots, n\}$ of cardinality $k$.
\[l:alphaofminor\] Let ${\genfrac{[}{]}{0em}{1}{J}{I}}$ denote the quantum minor in ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ with row and column indexing sets $I$ and $J$. Then $$\alpha({\genfrac{[}{]}{0em}{1}{J}{I}})=[Q_{1}(I,J)][1\cdots k]^{-1}.$$
Note that the sets $Q_{1}(I,J)$ describe column sets of maximal minors, whose row set is necessarily $\{1,\dotsc,k\}$.
This lemma has the following two consequences for the quantum cluster algebra structure on $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$. First, we see that we can rephrase the lemma in terms of the components of the noncommutative dehomogenisation as described above. Recall that $\alpha$ restricts to an isomorphism of ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ with the degree 0 part of $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$; the lemma tells us more.
\[c:imageofICunderalpha\] The image of ${\ensuremath{\mathcal{M}}}(k,n-k)$ under $\alpha$ in $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ is contained in the subspace ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}_{1}[1\cdots k]^{-1}$ of $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})_{0}=\sum_{j\geq 0} {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}_{j}[1\dotsm k]^{-j}$.
From Definition \[d:qGLScluster\], we may compute the image of ${\ensuremath{\mathcal{M}}}(k,n-k)$ under $\alpha$ explicitly.
Let ${\genfrac{[}{]}{0em}{1}{C(r,s)}{R(r,s)}} \in {\ensuremath{\mathcal{M}}}(k,n-k)$. Then $$\alpha({\genfrac{[}{]}{0em}{1}{C(r,s)}{R(r,s)}}) = \left[\{ \widetilde{1-s},\widetilde{2-s},\ldots ,\widetilde{r-s} \} {\sqcup}\left(\{ 1,\ldots,k\} \setminus \{r,\ldots ,r+s\}\right)\right][1\cdots k]^{-1}.$$
We illustrate this for our running example in Figure \[fig:initialseedLoc36\].
[ ]{}
These results tell us about our initial cluster variables but in order to complete the lifting of the whole quantum cluster algebra structure to the quantum Grassmannian, we need a stronger statement on the images of *all* quantum cluster variables. This is achieved by the following theorem, which uses a cluster algebra argument, as opposed to direct calculation of the sort that gives the above results. It also emphasises the relevance of the categorification.
\[t:thetaisscalingpower\] Let $v$ be a quantum cluster variable for the quantum cluster algebra structure on ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ constructed from the initial data ${\ensuremath{\mathcal{M}}}(k,n-k)$. Let $M(v)$ be the module in ${\ensuremath{\mathcal{C}}}_{w}$ corresponding to $v$. Then $\alpha(v) \in {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}_{\theta(M(v))}[1\dotsm k]^{-\theta(M(v))}$, where $\theta(M(v))$ is equal to the dimension of the top of the module $M(v)$.
We argue by induction on the length of mutation sequences. Firstly, we see that the claim holds for elements of the initial cluster ${\ensuremath{\mathcal{M}}}(k,n-k)$ by the observation that $\theta(V_{(\alpha,\beta)})=1$ for all $\alpha$ and $\beta$ and by Corollary \[c:imageofICunderalpha\].
Note that the localisation at hand, $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$, is constructed from ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ by localisation at an Ore set that consists of (positive integer) powers of a single element, $[1\cdots k]$. Since in any localisation a finite set of elements has a common denominator, for any finite set of elements $A_{1},\ldots,A_{r}$ of $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ there exists a positive integer $m$ such that there exist $B_{1},\ldots,B_{r} \in {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ with $A_{i}=B_{i}[1\cdots k]^{-m}$. Furthermore, we can choose $m$ to be the (unique) least such positive integer, so that we may speak of lowest common denominators in this localisation.
In particular, this holds for $r=1$, so that every element $A$ of $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ has a unique expression as $B[1\cdots k]^{-d(A)}$ with $B\in {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ and $d(A)$ minimal. Equivalently, there exists a unique smallest $d(A)$ such that $A$ is an element of the subspace ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}_{d(A)}[1\cdots k]^{-d(A)}$ of $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$.
Since $[1\cdots k]\in {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$, we see that this implies that $A\in {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}_{j}[1\cdots k]^{-j}$ for all $j\geq d(A)$ since $$A=B[1\cdots k]^{-d(A)}=(B[1\cdots k]^{j-d(A)})[1\cdots k]^{-j}.$$ This is why the decomposition $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})=\sum_{j\geq 0} {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}_{j}[1\cdots k]^{-j}$ is not a direct sum decomposition.
Now, assume that the claim holds for some cluster ${\ensuremath{\mathcal{N}}}$ mutation-equivalent to ${\ensuremath{\mathcal{M}}}(k,n-k)$. Let $X_{1},\dotsc ,X_{r}$ be the quantum cluster variables appearing in ${\ensuremath{\mathcal{N}}}$. Then the mutation of $X_{i}$, say, is computed by taking the sum of the two relevant exchange monomials. More precisely, recall that the exchange relations take the form $$X_{i}^{\prime}=M(\underline{b}_{i}^{+})+M(\underline{b}_{i}^{-})$$ with $$M(a_{1},\dotsc ,a_{r}) {\stackrel{\scriptscriptstyle{\mathrm{def}}}{=}}q^{\frac{1}{2}\sum_{u<v} a_{u}a_{v}l_{vu}}X_{1}^{a_{1}}\dotsm X_{r}^{a_{r}}$$ and the integers $a_{i}$ are all non-negative except for $a_{k}=-1$. It is convenient to observe, however, that since $X_{i}$ quasi-commutes with every other element of the cluster, so does its inverse and we can re-write the exchange relation in the form $$X_{i}^{\prime}X_{i}=N_{+}+N_{-}$$ by quasi-commuting $X_{i}^{-1}$ to the right-hand side of each monomial $M(\underline{b}_{i}^{+})$ and $M(\underline{b}_{i}^{-})$ and multiplying through. Of course, this changes the powers of $q$ appearing in front of the monomials but for the present argument this does not matter. Now we see that we may use the inductive hypothesis and the existence of lowest common denominators to first write $N_{\pm}=S_{\pm}[1\cdots k]^{-d(N_{\pm})}$ with $S_{\pm}\in {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ and $d(N_{\pm})$ positive integers and then to write $N_{+}+N_{-}$ as a product of an element of ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ with some power $m$ of $[1\cdots k]^{-1}$. Furthermore it is clear that the minimal such $m$ is equal to the maximum of the powers $d(N_{+})$ and $d(N_{-})$. Hence the exchange relation tells us that when we write $X_{i}=T_{i}[1\cdots k]^{-d(X_{i})}$ and $X_{i}^{\prime}=T^{\prime}_{i}[1\cdots k]^{-d(X^{\prime}_{i})}$ with $T_{i},T_{i}^{\prime}\in {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$, $d(X^{\prime}_{i})$ is equal to $\max\{ d(N_{+}),d(N_{-})\}-d(X_{i})$.
However, we have seen this formula previously: it is precisely the formula determining the values of $\theta$ by repeated mutation. Since the integers given by $\theta$ (the dimension of the top of the corresponding module) and $d$ (the minimal positive integer described above) take the same initial values—this being the base case for our induction—and since they mutate by identical formulæ, we see that they agree on all quantum cluster variables. This proves the theorem.
A more direct argument can be made for the quantum cluster variables for ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ whose image under $\alpha$ is a quantum Plücker coordinate multiplied by some power of $[1\cdots k]^{-1}$. For we may observe that as we range over all possible indexing sets $I$ and $J$ of quantum minors in ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$, the collection of sets $Q_{1}(I,J)$ ranges over *all* $k$-subsets of $\{1,\dotsc ,n\}$. Then since Corollary 12.4 of [@GLS-QuantumPFV] tells us that every quantum minor does occur as a quantum cluster variable in the quantum cluster algebra structure for ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$, we see the following.
\[c:PluckersAreQCVs\] For each $k$-subset $I$ of $\{1,\dotsc ,n\}$, we have that $[I][1\cdots k]^{-1}$ is a quantum cluster variable in the above quantum cluster algebra structure on $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$.
Here, $[I]$ is the quantum Plücker coordinate corresponding to $I$. Note that this is consistent with the above general theorem, since the corresponding modules have simple tops; the latter follows from the fact that this is true for the modules $V_{(\alpha,\beta)}$, this itself being a feature of being in type $A$.
A quantum cluster algebra structure on the quantum Grassmannian {#s:QCAonKqGr}
===============================================================
The\[page:fix-theta-start\] final step is to use Theorem \[t:rescaledQCA\] to re-scale the quantum cluster variables appearing in the above quantum cluster algebra structure on $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ to eliminate the inverse of the minor $[1\cdots k]$ that appears. By doing so, we will see that all the re-scaled quantum cluster variables in fact lie in ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$, which together with Corollary \[c:PluckersAreQCVs\] will imply that we have a (graded) quantum cluster algebra structure on ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$, since the Plücker coordinates generate ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$.
From the previous section, notably Theorem \[t:thetaisscalingpower\], we know that the power of $[1\cdots k]$ appearing in any quantum cluster variable is exactly given by $\theta$, the dimension of the top of the corresponding module in the categorification. We would like to apply our re-scaling theorem, Theorem \[t:rescaledQCA\], but we cannot do so directly because as we noted before $\theta$ is not a grading. Therefore our first task is to fix this.
More concretely, we will alter slightly the initial data in order to correct the inhomogeneity at the position $(1,1)$. For we observe that at every mutable index $(\alpha,\beta)$ except the top-left, i.e. except at $(1,1)$, the exchange quiver has the *same* number of incoming and outgoing arrows. In other words, the exchange matrix $B=B^{\mathrm{Loc}}(k,n)$ admits a grading by the vector $\underline{a}=(1,\dotsc ,1,-1,1)\in {\ensuremath{\mathbb{Z}}}^{k(n-k)+2}$ except for at the index $(1,1)$. We would like to use this grading $\underline{a}$ as one of our input data to Theorem \[t:rescaledQCA\] but we need to homogenise $B$ at $(1,1)$ in order to do so. We note that $\underline{a}$ ends with the values $-1$ and $1$ in order to reflect the natural grading on $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ in terms of the power of $[1\dotsm k]^{-1}$ occurring in our expressions for the quantum cluster variables.
We observe that the grading $\underline{a}$ does correspond to $\theta$, at least away from the coefficients $[1\cdots k]^{\pm 1}$. This is as expected, for this categorical data $\theta$ is exactly what is used in [@GLS-PFV §10] to make the lifting work classically. The classical version uses quotients rather than localisations, as we must, but we can see that reinterpreting [@GLS-PFV §10] in terms of localisations gives rise to the analogue of what we do here.
The underlying reason for the choice we will make below, and its classical analogue in [@GLS-PFV §10], is geometric. This is explained in the discussion after Theorem 4.14 of [@GSV-Book]. That theorem describes the passage from the corresponding classical (commutative) cluster algebra structure on $\mathbb{K}[\mathrm{Mat}(k,n-k)]$ to one on $\mathbb{K}[\mathrm{Gr}(k,n)]$ in more concrete terms than those used in [@GLS-PFV §10], where the general result was the focus. (We note that we do not directly rely on the classical result but do indirectly, in that results we use from [@GLS-QuantumPFV] rely on the existence of the classical cluster algebra structure on $\mathbb{K}[\mathrm{Mat}(k,n-k)]$.)
Hence we define a new initial datum as follows. Let $$({\ensuremath{\mathcal{L}}}={\ensuremath{\mathcal{L}}}(k,n),B=B^{\mathrm{Loc}}(k,n),L=L^{\mathrm{Loc}}(k,n),G=G^{\mathrm{Loc}}(k,n))$$ denote the initial data for the quantum cluster algebra structure on $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ described in the previous section. Here ${\ensuremath{\mathcal{L}}}=\{ {\ensuremath{\mathcal{L}}}_{(r,s)} \mid 1\leq r\leq k,\ 1\leq s \leq n-k \} {\mathrel{\cup}}\{ {\ensuremath{\mathcal{L}}}_{(1,k+1)},{\ensuremath{\mathcal{L}}}_{(2,k+1)}\}$. We add an additional coefficient (i.e. a non-mutable variable) ${\ensuremath{\mathcal{L}}}_{(0,0)}$ to the initial cluster ${\ensuremath{\mathcal{L}}}$, namely ${\ensuremath{\mathcal{L}}}_{(0,0)}=[1\dotsm k][1\dotsm k]^{-1}$. Let us denote by $\hat{{\ensuremath{\mathcal{L}}}}$ the set ${\ensuremath{\mathcal{L}}} {\mathrel{\cup}}\{ {\ensuremath{\mathcal{L}}}_{(0,0)} \}$. Of course, this additional element is simply the identity for the algebra $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ and as such it certainly quasi-commutes with every element of ${\ensuremath{\mathcal{L}}}$, giving that the corresponding quasi-commutation matrix $\hat{L}$ is constructed from $L$ by setting $\hat{L}_{(0,0),(0,0)}=0$, $\hat{L}_{(0,0),(r,s)}=\hat{L}_{(r,s),(0,0)}=0$ for all $(r,s)\in \{ 1\leq r \leq k, 1\leq s \leq n-k \} {\mathrel{\cup}}\{ (1,k+1),(2,k+1) \}$ and $\hat{L}_{(r_{1},s_{1}),(r_{2},s_{2})}=L_{(r_{1},s_{1}),(r_{2},s_{2})}$ whenever $(r_{1},s_{1}),(r_{2},s_{2})\neq (0,0)$.
Next we add an extra arrow to the exchange quiver, from $(0,0)$ to $(1,1)$, or equivalently define $$\hat{B}_{(r_{1},s_{1}),(r_{2},s_{2})} {\stackrel{\scriptscriptstyle{\mathrm{def}}}{=}}\begin{cases} 0 & \text{if}\ (r_{1},s_{1})=(0,0)\ \text{and}\ (r_{2},s_{2})=(0,0), \\ & \qquad \text{or}\ (r_{1},s_{1})=(0,0)\ \text{and}\ (r_{2},s_{2})\neq (1,1), \\ & \qquad \text{or}\ (r_{1},s_{1})\neq (1,1)\ \text{and}\ (r_{2},s_{2})=(0,0) \\ 1 & \text{if}\ (r_{1},s_{1})=(0,0)\ \text{and}\ (r_{2},s_{2})=(1,1) \\ -1 & \text{if}\ (r_{1},s_{1})=(1,1)\ \text{and}\ (r_{2},s_{2})=(0,0) \\ B_{(r_{1},s_{1}),(r_{2},s_{2})} & \text{otherwise.} \end{cases}$$
For a grading $\hat{G}$ we take $\hat{G}=(-1,\dotsc ,-1,1,-1)\in {\ensuremath{\mathbb{Z}}}^{k(n-k)+3}$. Our reason for doing so is that $\hat{G}$ records the power of $[1\dotsm k]$ occurring in the initial cluster variables as shown in the previous section, namely $-1$ except for the coefficient $[1\dotsm k]$. This is indeed still a grading, as is easily checked. We will not forget the data in the original grading $G$: it will also be used when we apply Theorem \[t:rescaledQCA\], which takes *two* gradings among its inputs.
Now it is straightforward to check that $(\hat{{\ensuremath{\mathcal{L}}}},\hat{B},\hat{L},\hat{G})$ is valid initial data for a graded quantum cluster algebra structure on $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$. Indeed, compatibility of $\hat{B}$ with $\hat{L}$ is immediate, since $\hat{L}$ contains only zeroes in the row and column indexed by $(0,0)$. The grading condition holds by construction.
Furthermore, the quantum cluster variables obtained by iterated mutation from the initial seed $(\hat{{\ensuremath{\mathcal{L}}}},\hat{B},\hat{L},\hat{G})$ are equal to those obtained from $({\ensuremath{\mathcal{L}}},B,L,G)$, since the new variable is simply the identity in $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ and as such has no effect whatsoever on any exchange monomials it appears in. It is to this altered graded quantum cluster algebra structure on $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$, with initial seed $(\hat{{\ensuremath{\mathcal{L}}}},\hat{B},\hat{L},\hat{G})$ that we will apply Theorem \[t:rescaledQCA\].\[page:fix-theta-end\]
In Figure \[fig:initialseedLoc36-homogenised\], we give this homogenised initial cluster for our running example with $k=3$ and $n=6$.
[ ]{}
As in Proposition \[p:dhom\], let $\sigma$ be the automorphism of ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ defined by $\sigma(X_{ij})=qX_{ij}$. Then there is an automorphism $\hat{\sigma}$ of ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}[Y^{\pm 1}; \sigma]$ defined by $\hat{\sigma}|_{{\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}}=\sigma$ and $\hat{\sigma}(Y)=Y$.
One easily sees that $\hat{\sigma}$ respects the relations $YX_{ij}=\sigma(X_{ij})Y$ in the skew-Laurent extension.
\[c:automtau\] There is an automorphism of $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n-k)]})$ defined by $\tau=\alpha \circ \hat{\sigma} \circ \alpha^{-1}$, where $\alpha$ is the dehomogenisation isomorphism.
Denote by $\hat{\underline{t}}$ the vector with $\hat{\underline{t}}_{(0,0)}=0$ and $\hat{\underline{t}}_{(r,s)}=G_{(r,s)}$. Then the vector $\hat{\underline{t}}$ described a grading for $\hat{B}$ above, since $G$ was a grading for $B$. The choice of $\hat{\underline{t}}_{(0,0)}=0$ is the unique one such that $\hat{\underline{t}}$ is indeed an extension of the grading $G$ to a grading for $\hat{B}$ but is also consistent with the natural degree of $\hat{L}_{(0,0)}=[1\dotsm k][1\dots k]^{-1}=1$ being zero.
Recall that the grading $G$ describes precisely the degree of the initial quantum cluster variables for the quantum cluster algebra structure on ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$, where “degree” means as a homogeneous polynomial in the matrix generators. Indeed we saw that the quantum cluster algebra structure on ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$ is precisely graded by this natural grading. Then it is clear from its definition that $\sigma$ acts by multiplication by $q$ to this degree, as we noted previously. Passing this through the isomorphism $\alpha$, we see that $\tau$ is exactly the automorphism induced by $\hat{\underline{t}}$, or equivalently that $\hat{\underline{t}}$ may be recovered from $\tau$.
Now we apply Theorem \[t:rescaledQCA\].
\[p:rescaledLoctilde\] Let $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ have the graded quantum cluster algebra structure induced by the initial seed $(\hat{{\ensuremath{\mathcal{L}}}},\hat{B},\hat{L},\hat{G})$. Then there exists a graded quantum cluster algebra structure on a subalgebra $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ of the skew-Laurent extension $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})[Z^{\pm 1};\tau]$ with initial data
- $\tilde{{\ensuremath{\mathcal{L}}}}=\{ [1\dotsm k][1\dotsm k]^{-1}Z \} {\mathrel{\cup}}\{ \tilde{{\ensuremath{\mathcal{L}}}}_{(r,s)}=q^{\hat{\underline{t}}_{(r,s)}/2}\hat{{\ensuremath{\mathcal{L}}}}_{(r,s)}Z \mid 1\leq r\leq k,\ 1\leq s\leq n-k \} {\mathrel{\cup}}\{ q^{1/2}[1\dotsm k]Z^{-1}, q^{-1/2}[1\dotsm k]^{-1}Z \}$
- $\tilde{B}=\hat{B}$,
- $\tilde{L}$ satisfies $\tilde{L}_{(r_{1},s_{1}),(r_{s},s_{2})}=\hat{L}_{(r_{1},s_{1}),(r_{2},s_{2})}+\hat{\underline{t}}_{(r_{2},s_{2})}-\hat{\underline{t}}_{(r_{1},s_{1})}$, and
- $\tilde{G}=0$.
Here the automorphism $\tau$ is as described in the previous lemma, inducing the grading $\hat{\underline{t}}$.
We apply Theorem \[t:rescaledQCA\] with $\underline{t}=\hat{\underline{t}}$ and $\underline{u}=-\hat{G}$. As noted above, both are gradings for $\hat{B}$. Then one easily checks that $(\underline{t}\wedge \underline{u})_{ij}=t_{j}-t_{i}$, giving the form of $\tilde{L}$ as stated, and we have $\tilde{G}=\hat{G}+(-\hat{G})=0$.
Indeed, applying Corollary \[c:formofrescaledQCV\] to this setting, we have that the quantum cluster variables for this quantum cluster algebra structure on $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ are in bijection with those of $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ and furthermore the former have the form of a product of a power of $q$, a quantum cluster variable for $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ and a power of $Z$.
Thus for our running example $k=3$ and $n=6$ we have as initial cluster for the quantum cluster algebra $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(3,6)]})$ that shown in Figure \[fig:initialseedLoc36-homog-and-rescaled\]. The values of $\hat{\underline{t}}$ giving the powers of $q$ appearing are derived by reading off from Figure \[fig:initialseed33\] the degrees of the corresponding variables as homogeneous polynomials in the matrix generators. So for example $\hat{\underline{t}}_{(2,2)}=2$ since $[356][123]^{-1}=\alpha({\genfrac{[}{]}{0em}{1}{23}{23}})$ and the latter has degree 2.
[ ]{}
\[l:centralelt\] The element $q^{-1/2}[1\dotsm k]^{-1}Z$ is central in $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})[Z^{\pm 1};\tau]$ and hence is a central coefficient in the quantum cluster algebra $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$.
We chose $\tau$ so that $$\tau([1\dotsm k])=\alpha(\sigma(\alpha^{-1}([1\dotsm k])))=\alpha(\sigma(Y))=\alpha(Y)=[1\dotsm k],$$ and hence $Z$ commutes with $[1\dotsm k]$, and so that $\alpha(Y)=[1\dotsm k]$ and $Z$ satisfy the same quasi-commutation relations with $\alpha({\mathbb{K}_{q}[\mathrm{M}(k,n-k)]})$. It follows that the stated element is central.
\[c:quotisQCA\] The quotient algebra $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})/(q^{-1/2}[1\dotsm k]^{-1}Z-1)$ inherits a graded quantum cluster algebra structure.
It is straightforward to see that the quotient of a quantum cluster algebra by a central coefficient is again a quantum cluster algebra, with the natural quotient data. Since the element $q^{-1/2}[1\dotsm k]^{-1}Z$ has degree 0 (as $\tilde{G}=0$), this quantum cluster algebra is again graded by the grading $-\hat{G}$, suitably restricted; indeed we see that this grading is equal to $\underline{1}=(1,\dotsc ,1)\in {\ensuremath{\mathbb{Z}}}^{k(n-k)+1}$.
In particular, in this quotient the two coefficients $q^{1/2}[1\dotsm k]Z^{-1}$ and $q^{-1/2}[1\dotsm k]^{-1}Z$ are both identified with the identity and as such may be deleted from the quantum cluster algebra data with no effect, which we do.
\[t:quotisotoGr\] The quotient algebra $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})/(q^{-1/2}[1\dotsm k]^{-1}Z-1)$ is isomorphic to ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$. Hence the quantum Grassmannian ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ admits a graded quantum algebra structure.
We proceed in two steps. First, we show the existence of a surjective homomorphism from ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ to the quotient $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})/(q^{-1/2}[1\dotsm k]^{-1}Z-1)$. Then we show that the latter has the same Gelfand–Kirillov dimension as ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$, from which it follows that the two algebras are isomorphic.
Let us denote by $z$ the (central) element $q^{-1/2}[1\dotsm k]^{-1}Z \in \mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})[Z^{\pm 1};\tau]$. Then let $f\colon {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]} \to \mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})[Z^{\pm 1};\tau]$ be the linear map defined on the generating quantum Plücker coordinates $[I]$ of ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ by $f([I])=[I]z$, extended to first to monomials multiplicatively and then extended linearly. Notice that $f$ is a map whose codomain is the whole skew-Laurent extension defined in Proposition \[p:rescaledLoctilde\]. Since ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ is spanned as a vector space by monomials in the quantum Plücker coordinates and since the defining (generalised) quantum Plücker relations in ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ are homogeneous ([@Kolb Remark 3.3]), the centrality of $z$ implies that this yields a well-defined algebra homomorphism.
Next we establish that the image of $f$ lies in $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$. This follows from Theorem \[t:thetaisscalingpower\]: we defined the elements ${\ensuremath{\mathcal{L}}}_{(r,s)}$ to be the images under the dehomogenisation isomorphism $\alpha$ of the initial cluster variables for the quantum cluster algebra structure on ${\mathbb{K}_{q}[\mathrm{M}(k,n-k)]}$. In particular, as noted after Theorem \[t:thetaisscalingpower\], every quantum cluster variable of $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ is equal to an element of ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ multiplied by some power of $z$. For this is exactly why the choices in Proposition \[p:rescaledLoctilde\] were made: Theorem \[t:thetaisscalingpower\] gives that every quantum cluster variable $v$ in $\mathrm{Loc}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ has a unique expression as an element of ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ multiplied by $([1\dotsc k]^{-1})^{\theta(\alpha^{-1}(v))}$ and we applied the re-scaling construction of Theorem \[t:rescaledQCA\] using the data from $\theta$ (now “fixed” to be a genuine grading), re-scaling every quantum cluster variable precisely by the power of $Z$ needed to ensure that we obtain as quantum cluster variables elements of ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ multiplied by powers of $z$.
Furthermore, since the quantum cluster variables by definition generate $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$, $f$ is a surjective homomorphism onto $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$. Composing $f$ with the natural projection of $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ onto the quotient by the ideal generated by $z-1$, we have a surjective homomorphism $g \colon {\mathbb{K}_{q}[\mathrm{Gr}(k,n)]} \to \widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})/(z-1)$, as we wanted.
Next, we show the equality of the Gelfand–Kirillov dimensions of the domain and codomain of $g$. The GK-dimension of ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ is well-known to be $k(n-k)+1$ so we compute the GK-dimension of the quotient that is the codomain. By construction, $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ contains the quantum affine space $\mathbb{A}_{q}$ whose generators are precisely the elements in the initial seed of Proposition \[p:rescaledLoctilde\] except $z^{-1}$. Since $\tilde{L}\setminus \{z^{-1} \}$ is a quasi-commuting set, being a quantum cluster, this is clear.
Now $\mathbb{A}_q$ has $k(n-k)+2$ generators, so $k(n-k)+2=\mathrm{GKdim}\ \mathbb{A}_q \leq \mathrm{GKdim} \widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$. On the other hand, by the quantum Laurent phenomenon ([@BZ-QCA Corollary 5.2]), $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ is contained in the quantum torus $\mathbb{T}_q$ associated to $\mathbb{A}_q$. So $\mathrm{GKdim}\ \widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}) \leq \mathrm{GKdim}\ \mathbb{T}_q =k(n-k)+2$.
Hence, $\mathrm{GKdim}\ \widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})=k(n-k)+2$ and so it follows from [@KrauseLenagan Proposition 3.15] that $\mathrm{GKdim}\ \widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})/(z-1)\leq (k(n-k)+2)-1=k(n-k)+1$ since $z-1$ is certainly regular.
On the other hand, $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})/(z-1)$ contains (a copy of) the quantum affine space generated by the images of the indeterminates in the initial seed $\tilde{{\ensuremath{\mathcal{L}}}}$ of $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ except $z$ and $z^{-1}$. So $\mathrm{GKdim}\ \widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})/(z-1) \geq k(n-k)+1$ and hence we in fact have equality.
Now since ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ and $\widetilde{\mathrm{Loc}}({\mathbb{K}_{q}[\mathrm{Gr}(k,n)]})$ are domains, it follows that these are isomorphic, as every epimorphism of domains of the same Gelfand–Kirillov dimension is an isomorphism ([@KrauseLenagan Proposition 3.15], as previously).
In Figure \[fig:initialseedKqGr36\] we show the initial data for the quantum cluster algebra structure thus obtained on ${\mathbb{K}_{q}[\mathrm{Gr}(3,6)]}$. We note that due to the powers of $q$ that are present, this is not identical to the quantum cluster algebra structure obtained in the authors’ earlier work ([@Gr2nSchubertQCA]). However, to conclude, we may apply Proposition \[p:rescaleQCAbypowersofq\], for we see that the power of $q$ appearing in the expression for the $(r,s)$ variable is exactly $q^{\frac{\hat{\underline{t}}_{(r,s)}+1}{2}}$. The corresponding vector ${{(\hat{\underline{t}}_{(0,0)}+1,\dotsc ,\hat{\underline{t}}_{(k,n-k)}+1)}^{T}}$ is a grading for the relevant exchange matrix, since both $\hat{\underline{t}}$ and $(1,\dotsc ,1)$ are. Then we may apply Proposition \[p:rescaleQCAbypowersofq\] with the negative of this grading, to obtain an isomorphic quantum cluster algebra without these powers of $q$. In this way, we recover exactly the quantum cluster algebra structure obtained in [@Gr2nSchubertQCA]. Figure \[fig:finalQCAstrKqGr36\] shows this final re-scaled quantum cluster algebra structure on ${\mathbb{K}_{q}[\mathrm{Gr}(3,6)]}$.
[ ]{}
[ ]{}
The preceding Corollary establishes that ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$ is a *graded* quantum cluster algebra and we have seen that this grading is the standard grading on ${\mathbb{K}_{q}[\mathrm{Gr}(k,n)]}$, with the quantum Plücker coordinates in degree one. Then in particular it follows from the general theory of graded quantum cluster algebras that every quantum cluster variable is homogeneous with respect to this graded, a phenomenon observed in the authors’ earlier work ([@Gr2nSchubertQCA]) (and only experimentally for ${\mathbb{K}_{q}[\mathrm{Gr}(3,n)]}$, $n=6,7,8$).
We note that we have worked throughout over the field $\mathbb{Q}(q)$, that is, with $q$ transcendental over $\mathbb{Q}$. This assumption is necessary because this is the context in which the main theorem of Geiß–Leclerc–Schröer is proved ([@GLS-QuantumPFV Theorem 12.3]), for a number of technical reasons. Since that theorem provides the starting point for our lifting, namely the quantum cluster algebra structure on quantum matrices, we must make this assumption too. However our methods here only use that $q$ is not a root of unity, so that if the aforementioned result is extended, our conclusion will also follow immediately without need for modification.
So, in line with Conjecture 12.7 of [@GLS-QuantumPFV], we conjecture that the above quantum cluster algebra structure on the quantum Grassmannian can be realised on an integral form, i.e. over $\mathbb{Q}[q,q^{-1}]$. Indeed the explicit descriptions of the quantum cluster variables in the authors’ earlier work suggest that this structure may even be defined over ${\ensuremath{\mathbb{Z}}}[q,q^{-1}]$. However, many of the constructions we have applied, notably Theorem \[t:rescaledQCA\], involve powers of $q^{1/2}$ and it appears to be a delicate matter to see that $q^{1/2}$ does not enter into the final quantum cluster algebra structure.
As noted in the introduction, we expect that the methods presented here—or generalisations of them—can be used to establish the existence of (graded) quantum cluster algebra structures on the quantized coordinate rings of arbitrary partial flag varieties. The relationships between the latter and their localisations that give the quantized coordinate rings of the big cells of the corresponding partial flag variety are well-understood and dehomogenisation isomorphisms such as that used here are known. Modifications of the constructions here may be necessary, however. For example, multi-gradings may be needed where the coordinate rings of the big cells involve localisation at several elements.
Appendix
========
Here we gather the aforementioned details of the calculations used in the proof of Theorem \[t:rescaledQCA\]. Notation is as in that proof.
$\tilde{X}_{i}^{a_{i}}=q^{\frac{1}{2}a_{i}^{2}t_{i}u_{i}}X_{i}^{a_{i}}z^{a_{i}u_{i}}$
If $a_{i}\geq 0$ then $$\begin{aligned}
\tilde{X}_{i}^{a_{i}} & = \left(q^{\frac{t_{i}u_{i}}{2}}X_{i}z^{u_{i}}\right)^{a_{i}} \\
& = q^{\frac{a_{i}t_{i}u_{i}}{2}}\left( X_{i}z^{u_{i}}\right)^{a_{i}} \\
& = q^{\frac{a_{i}t_{i}u_{i}}{2}}q^{\left(\sum_{j=1}^{a_{i}-1} j\right)t_{i}u_{i}}X_{i}^{a_{i}}z^{a_{i}u_{i}} \\
& = q^{\frac{a_{i}t_{i}u_{i}}{2}}q^{\left(\frac{a_{i}(a_{i}-1)}{2}\right)t_{i}u_{i}}X_{i}^{a_{i}}z^{a_{i}u_{i}} \\
& = q^{\frac{1}{2}a_{i}^{2}t_{i}u_{i}}X_{i}^{a_{i}}z^{a_{i}u_{i}} \end{aligned}$$ since $z^{u_{i}}X_{i}=q^{t_{i}u_{i}}X_{i}z^{u_{i}}$. The sum $\left(\sum_{j=1}^{a_{i}-1} j\right)t_{i}u_{i}$ arises from moving $z^{u_{i}}$’s to the right the required number of times to rearrange the product as shown. It is straightforward to check that the claim is also correct for $a_{i}\leq 0$ by similar means.
Set $\beta_{i}=q^{\frac{1}{2}a_{i}^{2}t_{i}u_{i}}$.
$z^{a_{i}u_{i}}\left(\prod_{j=i+1}^{r} X_{j}^{a_{j}}z^{a_{j}u_{j}}\right)=\left(\prod_{j=i+1}^{r} q^{a_{i}a_{j}t_{j}u_{i}}\right)\left(\prod_{j=i+1}^{r} X_{j}^{a_{j}}z^{a_{j}u_{j}}\right)z^{a_{i}u_{i}}$
This follows from the defining quasi-commutation relation $zX_{j}=q^{t_{j}}X_{j}z$ and noting that $z$ commutes with itself.
Set $\alpha_{i}=\prod_{j=i+1}^{r} q^{a_{i}a_{j}t_{j}u_{i}}=\prod_{i<j} q^{a_{i}a_{j}t_{j}u_{i}}$ so that $$z^{u_{i}a_{i}}\prod_{i<j} X_{j}^{a_{j}}z^{a_{j}u_{j}}=\alpha_{i}\left(\prod_{i<j} X_{j}^{a_{j}}z^{a_{j}u_{j}} \right)z^{u_{i}a_{i}},$$ by the preceding lemma.
$\prod_{i=1}^{r} \tilde{X}_{i}^{a_{i}} = \left(\prod_{i=1}^{r-1} \alpha_{i}\right)\left( \prod_{i=1}^{r} \beta_{i} \right)X_{1}^{a_{1}}\dotsm X_{r}^{a_{r}}z^{\left(\sum_{i=1}^{r} a_{i}u_{i}\right)}$
$$\begin{aligned}
\prod_{i=1}^{r} \tilde{X}_{i}^{a_{i}} & = \prod_{i=1}^{r} \beta_{i}X_{i}^{a_{i}}z^{a_{i}u_{i}} \\
& = \left(\prod_{i=1}^{r} \beta_{i}\right)\prod_{i=1}^{r} X_{i}^{a_{i}}z^{a_{i}u_{i}} \\
& = \left(\prod_{i=1}^{r-1} \alpha_{i}\right)\left( \prod_{i=1}^{r} \beta_{i} \right)X_{1}^{a_{1}}\dotsm X_{r}^{a_{r}}z^{\left(\sum_{i=1}^{r} a_{i}u_{i}\right)}\end{aligned}$$
by using the above lemmas repeatedly.
[ ]{}
1. $\prod_{i=1}^{r-1} \alpha_{i}=q^{\sum_{i<j} a_{i}a_{j}t_{j}u_{i}}$
2. $\prod_{i=1}^{r} \beta_{i} = q^{\frac{1}{2}\sum_{i=1}^{r} a_{i}^{2}t_{i}u_{i}}$
These equalities are immediate from the definitions of $\alpha_{i}$ and $\beta_{i}$ respectively.
$\tilde{M}(a_{1},\dotsc ,a_{r})=q^{\frac{1}{2}\sum_{i=1}^{r} \sum_{j=1}^{r} a_{i}a_{j}t_{i}u_{j}}M(a_{1},\dotsc ,a_{r})z^{\left(\sum_{i=1}^{r} a_{i}u_{i}\right)}$
$$\begin{aligned}
\tilde{M}(a_{1},\dotsc ,a_{r}) & = q^{\frac{1}{2}\sum_{i<j} a_{i}a_{j}\tilde{l}_{ji}}\left(\prod_{i=1}^{r} \tilde{X}_{i}^{a_{i}} \right) \\
& = q^{\frac{1}{2}\sum_{i<j} a_{i}a_{j}(l_{ji}+t_{i}u_{j}-t_{j}u_{i})}\left(\prod_{i=1}^{r} \tilde{X}_{i}^{a_{i}} \right) \\
& = q^{\frac{1}{2}\sum_{i<j} a_{i}a_{j}(l_{ji}+t_{i}u_{j}-t_{j}u_{i})}\left(\prod_{i=1}^{r-1} \alpha_{i}\right)\left( \prod_{i=1}^{r} \beta_{i} \right)X_{1}^{a_{1}}\dotsm X_{r}^{a_{r}}z^{\left(\sum_{i=1}^{r} a_{i}u_{i}\right)} \\
& = q^{\frac{1}{2}\sum_{i<j} a_{i}a_{j}(l_{ji}+t_{i}u_{j}-t_{j}u_{i})}q^{\sum_{i<j} a_{i}a_{j}t_{j}u_{i}}q^{\frac{1}{2}\sum_{i=1}^{r} a_{i}^{2}t_{i}u_{i}}X_{1}^{a_{1}}\dotsm X_{r}^{a_{r}}z^{\left(\sum_{i=1}^{r} a_{i}u_{i}\right)} \\
& = q^{\frac{1}{2}\sum_{i<j} a_{i}a_{j}(l_{ji}+t_{i}u_{j}-t_{j}u_{i})}q^{\sum_{i<j} a_{i}a_{j}t_{j}u_{i}}q^{\frac{1}{2}\sum_{i=1}^{r} a_{i}^{2}t_{i}u_{i}}q^{-\frac{1}{2}\sum_{i<j} a_{i}a_{j}l_{ji}}M(a_{1},\dotsc ,a_{r})z^{\left(\sum_{i=1}^{r} a_{i}u_{i}\right)} \\
& = q^{\frac{1}{2}\left((\sum_{i<j} a_{i}a_{j}(t_{i}u_{j}+t_{j}u_{i}))+(\sum_{i=1}^{r} a_{i}^{2}t_{i}u_{i})\right)}M(a_{1},\dotsc ,a_{r})z^{\left(\sum_{i=1}^{r} a_{i}u_{i}\right)} \\
& = q^{\frac{1}{2}\sum_{i=1}^{r} \sum_{j=1}^{r} a_{i}a_{j}t_{i}u_{j}}M(a_{1},\dotsc ,a_{r})z^{\left(\sum_{i=1}^{r} a_{i}u_{i}\right)}\end{aligned}$$
This is the equality as claimed in the proof of Theorem \[t:rescaledQCA\].
\[references\]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Level-spacing distributions of the Gaussian Unitary Ensemble (GUE) of random matrix theory are expressed in terms of solutions of coupled differential equations. Series solutions up to order $50$ in the level spacing are obtained, thus providing a very good description of the small-spacing part of the level-spacing distribution, which can be used to make comparisons with experimental or numerical data. The level-spacing distributions can be obtained by solving the system of differential equations numerically.'
address: 'Applied Mathematics Department, The Open University, Walton Hall, Milton Keynes MK7 6AA, UK'
author:
- 'Uwe Grimm[^1]'
title: 'Level-spacing distributions of the Gaussian unitary random matrix ensemble'
---
Introduction
============
Ever since the pioneering works of Wigner and Dyson [@Wigner; @Dyson1; @Dyson2; @Dyson3; @Dyson4; @DM63], random matrix theory has been a major tool in the investigation of complex systems in physics, see [*e.g.*]{} Refs. [@Haake; @Guhr]. In particular, the three “classical” random matrix ensembles, the Gaussian Orthogonal Ensemble (GOE), the Gaussian Unitary Ensemble (GUE) and the Gaussian Sympletic Ensemble (GSE) have been studied in much detail [@Mehta] and have proven of relevance in various applications, not alone in physics, see [*e.g.*]{} [@TW] for examples.
One of the characteristic features that is frequently used in investigations is the distribution of differences between eigenvalues. These level-spacing distributions are universal quantities, in the sense that they correctly describe the level-spacing distributions in many complex systems. In many applications, experimentally observed or numerically calculated discrete data, such as energy levels of large nuclei, are analyzed and their spacing distribution is compared to that of the appropriate random matrix ensemble, where the choice of the appropriate ensemble is dictated by the symmetry of the system. Besides the spacing between neighboring levels, also the spacing between levels separated by a fixed number of levels, [*i.e.*]{}, next-nearest neighbors, next-next-nearest neighbors, [*etc.*]{}, has been considered. A recent example of an astonishingly precise agreement between energy levels of a Hamiltonian and random matrix distributions was found in a simple tight-binding model of an electron moving on a planar quasiperiodic graph [@ZGRS; @SGRZ1; @SGRZ2; @GRSZ], which is used to model electrons in quasicrystals [@GS1; @GS2]. But the distribution functions of random matrix theory are even of relevance in pure mathematics; the paradigm of an application in number theory is the distribution of zeros of Riemann’s zeta function on the critical line [@Odlyzko].
However, there exists no simple closed form for the level-spacing distributions of the above-mentioned ensembles of random matrix theory. Thus, in many cases, the empirical data are actually compared to a so-called Wigner surmise, which corresponds to the spacing distribution of neighboring levels obtained from two-by-two random matrices, rather than from the limiting distribution for matrices of infinite size. For example, in the GOE case, the difference between the Wigner surmise and the true GOE spacing distribution is rather small, but the example of the quasiperiodic tight-binding model [@ZGRS; @SGRZ1; @SGRZ2; @GRSZ] shows that it may indeed be detected in physical applications. For the GUE case, the situation is somewhat better as the complete small spacing expansion for the nearest neighbor spacing distribution is known [@Haake]. Numerical values, series expansions and a Padé analysis for the nearest-neighbor spacing distributions can be found in [@Mehta; @Haake; @DiHa]. However, level spacings beyond the nearest-neighbor distributions, the spacing distributions of levels separated by a number of other levels, have scarcely been considered. However, explicit power series expressions are available [@Mehta92], although the coefficients in the series still involve determinants that need to be evaluated.
Here we employ the approach of Ref. [@BTW], which is based on the relation to differential equations of Painlevé type discovered by Jimbo [*at al.*]{} [@JMMS], to obtain series expansions or numerical estimates of these functions, compare also [@WFC; @FW2; @FW3; @F] for connections between random matrices and Painlevé transcendents. In this paper, we generalize this approach to calculate series expansions for small spacings up to order $50$ for the higher level-spacing distributions of the GUE. The expansions provide a detailed account of the small-spacing behavior of these distributions, and thus can be applied to analyze data. The exact coefficients involve ratios of large integer numbers, which makes them somewhat difficult to present in a paper. Therefore, the analytic expressions of the expansion coefficients are not included; they are provided in form of [Mathematica]{} [@Wolfram] code on the author’s homepage [@code]. The other two classical ensembles, the GOE and the GSE, can be treated in a similar fashion, although the calculation is somewhat more involved [@BTW]. The results for these two universality classes will be published separately.
Why are we interested in calculating these expansion to such high order in $s$? There are several reasons. First of all, the data suggest that the series have infinite radius of convergence, thus the level-spacing distributions appear to be analytic functions of $s$. This means that, the more coefficients we know, the larger the interval where the truncated series gives a good approximation of the spacing functions. Secondly, the higher level spacing distributions have leading terms of increasing power in $s$, so in order to at least obtain the leading terms of the higher spacing function we have to expand all function during the calculation up to that given order in $s$. In our case, expanding up to order $s^{50}$ just suffices to gives us the leading terms of the spacing distribution of levels separated by five other levels. Thirdly, although it is, in principle, possible to integrate the system of coupled differential equations numerically and to obtain “numerically exact” results for the spacing functions, this does require the knowledge of at least the leading terms in the expansions as we cannot integrate directly from $s=0$. Finally, it turns out to be very difficult to obtain a numerically exact distribution function in this way, because the distribution functions arise as differences between functions that grow exponentially with $s$, and thus requires high precision numerics. In fact, the author was not able to integrate numerically the set of equations for the spacing functions with four or five intermediate levels, even though the algebraic computer package [Mathematica]{} [@Wolfram] in principle allows for arbitrary precision arithmetics.
The paper is organized as follows. After this introduction, the method used to calculate the spacing distribution is briefly summarized, following closely the discussion of Ref. [@BTW]. After that, the series expansion data are presented and their applicability is discussed. We end with some concluding remarks.
Level-spacing distributions
===========================
The basic objects that usually are considered are the probabilities $E_{n}(s)$ that an interval of length $s$ contains exactly $n$ eigenvalues. Here, $s$ denotes the energy-level spacing in units of the mean level spacing. The gap probability $E_{0}(s)$ that an interval of length $s$ contains no eigenvalue at all can be expressed as a Fredholm determinant of a certain integral operator $K$ with a sine kernel, see [@TW] and references therein. Jimbo [*et al.*]{} [@JMMS] showed that this Fredholm determinant can be expressed as $$D(s;\lambda) \; := \; \det\left( I - \lambda K\right) \; = \;
\exp\left( \int\limits_{0}^{\pi s}\! dx\; \frac{\sigma(x;\lambda)}{x}
\right)
\label{eq:d}$$ where the function $\sigma(x;\lambda)$, considered as a function of the variable $x$ whereas $\lambda$ is regarded as a parameter, is a solution of a Painlevé V differential equation $$\Big(x\:\frac{d^{2}\!\sigma}{d x^{2}}\Big)^{2} \: + \;
4\Big( x\:\frac{d\sigma^{}}{d x} -
\sigma \rule{0pt}{20pt}\Big)
\Big[ x\:\frac{d\sigma}{d x} -\sigma +
\Big( \frac{d\sigma}{d x}\Big)^{\! 2}
\Big] \; = \; 0.
\label{eq:pv}$$ The parameter $\lambda$ is introduced via the boundary condition $$\sigma(x;\lambda) \; = \; -\frac{\lambda}{\pi}\: x \; + \; o(x)
\label{eq:ic}$$ for $x\rightarrow 0$. For the GUE, $D(s;\lambda)$ is the generating function of the probabilities $E_{n}(s)$, thus $$E_{n}(s) = D_{n}(s),
\label{eq:en}$$ where we defined $$\begin{aligned}
D_{0}(s) &=& D(s;1) \label{eq:d0}\\
D_{n}(s) &=&
\left.\frac{(-1)^n}{n!}
\frac{\partial^n D(s;\lambda)}{\partial\lambda^n}\right|_{\lambda=1},
\quad n>0.
\label{eq:dn}\end{aligned}$$ The level-spacing distributions $P_{n}(s)$, which are the probability distributions to find two energy levels at a distance $s$ with $n$ levels lying in-between, are then obtained as $$P_{n}(s) \; = \; \frac{d^2}{ds^2} \; \sum_{m=0}^{n}\: (n-m+1)\: E_{m}(s)
\label{eq:pn}$$ where $P_{0}(s)$ is the usual nearest-neighbour spacing distribution.
Now, it is rather straightforward to derive series expansions for the distributions $P_{n}(s)$ by making a power-series ansatz for the solution $\sigma(x;\lambda)$ of the differential equation (\[eq:pv\]) with the appropriate boundary condition (\[eq:ic\]). The distributions $P_{n}(s)$ were computed with the commercial algebraic computer program [Mathematica]{} [@Wolfram]. Whereas it is rather easy to calculate the first few series coefficients in this way, a straightforward implementation of the equations in an algebraic programming language like [Mathematica]{} will generally not succeed to calculate higher orders. In order to achieve this, the calculation was split into smaller tasks and an iterative scheme was used, calculating the expansions coefficients one by one. In this way, it was possible to obtain the series expansions for the spacing functions $P_{n}(s)$ up to order $s^{50}$.
Series expansions
=================
Small-$s$ expansions $$P_{n}(s) = \sum_{k=0}^{\infty}p_{n;k}s^k
\label{eq:ps}$$ of the level-spacing distributions $P_{n}(s)$ (\[eq:pn\]) can be obtained as follows. As a first step, we define functions $\sigma_{n}(x)$ by $$\begin{aligned}
\sigma_{0}(x) &=& \sigma(x;1),\\
\sigma_{n}(x) &=& \left.
\frac{\partial^n \sigma(x;\lambda)}{\partial\lambda^n}\right|_{\lambda=1}.\end{aligned}$$ The function $\sigma_{0}(x)$ satisfies the differential equation (\[eq:pv\]) and behaves as $$\sigma_{0}(x) \; = \; -\frac{1}{\pi}\: x \; + \; o(x)
\label{eq:ic0}$$ for small $x$. This follows from Eq. (\[eq:ic\]), which also determines the small-$x$ behavior of $\sigma_{n}(x)$ for $n>0$, $$\begin{aligned}
\sigma_{1}(x) & = & -\frac{1}{\pi}\: x \; + \; o(x),
\label{eq:ic1}\\
\sigma_{n}(x) & = & o(x)\qquad n>1.
\label{eq:icn}\end{aligned}$$ For any $N$, the functions $\sigma_{n}(x)$, $0\le n\le N$ satisfy a coupled set of differential equations that are obtained by taking the derivative of the original Painlevé equation (\[eq:pv\]) with respect to $\lambda$ and putting $\lambda=1$. For instance, for $N=2$ the set of differential equations becomes $$\begin{aligned}
0 & = & (x \sigma''_{0})^{2} +
4(x\sigma'_{0} - \sigma_{0})(x\sigma'_{0} -\sigma_{0} +{\sigma'_{0}}^{2}),
\label{eq:pv0}\\
0 & = & x^2\sigma''_{1}\sigma''_{0} +
2(x\sigma'_{1}-\sigma_{1})(x\sigma'_{0}-\sigma_{0}+{\sigma'_{0}}^{2})
\nonumber\\
& & +
2(x\sigma'_{0}-\sigma_{0})(x\sigma'_{1}-\sigma_{1}+2\sigma'_{1}\sigma'_{0}),
\label{eq:pv1}\\
0 & = & x^2\sigma''_{2}\sigma''_{0} + (x \sigma''_{1})^{2}\nonumber\\
& & +
2(x\sigma'_{2}-\sigma_{2})(x\sigma'_{0}-\sigma_{0}+{\sigma'_{0}}^{2})
\nonumber\\
& & +
4(x\sigma'_{1}-\sigma_{1})(x\sigma'_{1}-\sigma_{1}+2\sigma'_{1}\sigma'_{0})
\nonumber\\
& & +
2(x\sigma'_{0}-\sigma_{0})(x\sigma'_{2}-\sigma_{2}+2\sigma'_{2}\sigma'_{0}
+2{\sigma'_{1}}^{2}).
\label{eq:pv2}\end{aligned}$$ In order to obtain series expansions for the level-spacing distributions $P_{n}(s)$ (\[eq:pn\]) with $0\le n\le N$, we first compute series expansions for the functions $\sigma_{n}(x)$ for $0\le n\le N$. This is done by inserting polynomials $$\sigma_{n}(x) = \sum_{k=0}^{K}c_{n;k} \left(\frac{x}{\pi}\right)^{k}$$ of order $K$ into the set of differential equations obtained from (\[eq:pv\]), and implementing the initial conditions (\[eq:ic0\])–(\[eq:icn\]) by setting $c_{n;0}=0$ for $0\le n\le
N$, $c_{0;1}=c_{1;1}=-1$ and $c_{n;1}=0$ for $2\le n\le N$. The initial conditions (\[eq:ic0\]) and (\[eq:ic1\]) motivate the convenient choice of the expansion variable $x/\pi$ rather than $x$.
From the expansions of the functions $\sigma_{n}(x)$, $0\le n\le N$, we can derive expansions for $D_{n}(s)$ with $0\le n\le N$ by using Eqs. (\[eq:d\]), (\[eq:d0\]), and (\[eq:dn\]). For example, we have $$\begin{aligned}
D_{0}(s) & = &
\exp\left(
\int\limits_{0}^{\pi s}\! \frac{\sigma_{0}(x)\,dx}{x}\right)
\label{eq:d0exp}\\
D_{1}(s) & = &
D_{0}(s)\:\int\limits_{0}^{\pi s}\! \frac{\sigma_{1}(x)\, dx}{x},
\\
D_{2}(s) & = &
D_{0}(s)\:\left[
\int\limits_{0}^{\pi s}\! \frac{\sigma_{2}(x)\, dx}{x} +
{\left(\int\limits_{0}^{\pi s}\! \frac{\sigma_{1}(x)\, dx}{x}\right)}^{2}
\right],\end{aligned}$$
------ ------ --------------------------- ------ --------------------------- ------ --------------------------- ------ --------------------------- ------ --------------------------- ------ ---------------------------
$0$ $0$ $0$ $0$ $0$ $0$ $0$
$1$ $0$ $0$ $0$ $0$ $0$ $0$
$2$ $3$ $.290$ $0$ $0$ $0$ $0$ $0$
$3$ $0$ $0$ $0$ $0$ $0$ $0$
$4$ $-4$ $.329$ $0$ $0$ $0$ $0$ $0$
$5$ $0$ $0$ $0$ $0$ $0$ $0$
$6$ $3$ $.052$ $0$ $0$ $0$ $0$ $0$
$7$ $-2$ $.374\!\times\! 10^{-1}$ $2$ $.374\!\times\! 10^{-1}$ $0$ $0$ $0$ $0$
$8$ $-1$ $.339$ $0$ $0$ $0$ $0$ $0$
$9$ $2$ $.104\!\times\! 10^{-1}$ $-2$ $.104\!\times\! 10^{-1}$ $0$ $0$ $0$ $0$
$10$ $4$ $.004\!\times\! 10^{-1}$ $0$ $0$ $0$ $0$ $0$
$11$ $-9$ $.088\!\times\! 10^{-2}$ $9$ $.088\!\times\! 10^{-2}$ $0$ $0$ $0$ $0$
$12$ $-8$ $.685\!\times\! 10^{-2}$ $0$ $0$ $0$ $0$ $0$
$13$ $2$ $.491\!\times\! 10^{-2}$ $-2$ $.491\!\times\! 10^{-2}$ $0$ $0$ $0$ $0$
$14$ $1$ $.446\!\times\! 10^{-2}$ $-3$ $.450\!\times\! 10^{-4}$ $1$ $.725\!\times\! 10^{-4}$ $0$ $0$ $0$
$15$ $-4$ $.827\!\times\! 10^{-3}$ $4$ $.827\!\times\! 10^{-3}$ $0$ $0$ $0$ $0$
$16$ $-1$ $.974\!\times\! 10^{-3}$ $ 2$ $.625\!\times\! 10^{-4}$ $-1$ $.313\!\times\! 10^{-4}$ $0$ $0$ $0$
$17$ $7$ $.038\!\times\! 10^{-4}$ $-7$ $.038\!\times\! 10^{-4}$ $0$ $0$ $0$ $0$
$18$ $2$ $.411\!\times\! 10^{-4}$ $-9$ $.932\!\times\! 10^{-5}$ $4$ $.966\!\times\! 10^{-5}$ $0$ $0$ $0$
$19$ $-8$ $.042\!\times\! 10^{-5}$ $8$ $.042\!\times\! 10^{-5}$ $0$ $0$ $0$ $0$
$20$ $-2$ $.884\!\times\! 10^{-5}$ $2$ $.496\!\times\! 10^{-5}$ $-1$ $.248\!\times\! 10^{-5}$ $0$ $0$ $0$
$21$ $7$ $.418\!\times\! 10^{-6}$ $-7$ $.418\!\times\! 10^{-6}$ $0$ $0$ $0$ $0$
$22$ $3$ $.527\!\times\! 10^{-6}$ $-4$ $.714\!\times\! 10^{-6}$ $2$ $.357\!\times\! 10^{-6}$ $0$ $0$ $0$
$23$ $-5$ $.659\!\times\! 10^{-7}$ $5$ $.678\!\times\! 10^{-7}$ $-2$ $.805\!\times\! 10^{-9}$ $9$ $.351\!\times\! 10^{-10}$ $0$ $0$
$24$ $-4$ $.301\!\times\! 10^{-7}$ $7$ $.180\!\times\! 10^{-7}$ $-3$ $.590\!\times\! 10^{-7}$ $0$ $0$ $0$
$25$ $3$ $.684\!\times\! 10^{-8}$ $-3$ $.816\!\times\! 10^{-8}$ $1$ $.983\!\times\! 10^{-9}$ $-6$ $.610\!\times\! 10^{-10}$ $0$ $0$
$26$ $4$ $.991\!\times\! 10^{-8}$ $-9$ $.240\!\times\! 10^{-8}$ $4$ $.620\!\times\! 10^{-8}$ $0$ $0$ $0$
$27$ $-2$ $.210\!\times\! 10^{-9}$ $2$ $.677\!\times\! 10^{-9}$ $-7$ $.013\!\times\! 10^{-10}$ $2$ $.338\!\times\! 10^{-10}$ $0$ $0$
$28$ $-5$ $.357\!\times\! 10^{-9}$ $1$ $.038\!\times\! 10^{-8}$ $-5$ $.189\!\times\! 10^{-9}$ $0$ $0$ $0$
$29$ $1$ $.484\!\times\! 10^{-10}$ $-2$ $.587\!\times\! 10^{-10}$ $1$ $.655\!\times\! 10^{-10}$ $-5$ $.517\!\times\! 10^{-11}$ $0$ $0$
$30$ $5$ $.265\!\times\! 10^{-10}$ $-1$ $.040\!\times\! 10^{-9}$ $5$ $.198\!\times\! 10^{-10}$ $0$ $0$ $0$
$31$ $-1$ $.363\!\times\! 10^{-11}$ $3$ $.320\!\times\! 10^{-11}$ $-2$ $.936\!\times\! 10^{-11}$ $9$ $.788\!\times\! 10^{-12}$ $0$ $0$
$32$ $-4$ $.738\!\times\! 10^{-11}$ $9$ $.430\!\times\! 10^{-11}$ $-4$ $.715\!\times\! 10^{-11}$ $0$ $0$ $0$
$33$ $1$ $.532\!\times\! 10^{-12}$ $-4$ $.317\!\times\! 10^{-12}$ $4$ $.178\!\times\! 10^{-12}$ $-1$ $.393\!\times\! 10^{-12}$ $0$ $0$
$34$ $3$ $.915\!\times\! 10^{-12}$ $-7$ $.816\!\times\! 10^{-12}$ $3$ $.908\!\times\! 10^{-12}$ $-1$ $.072\!\times\! 10^{-16}$ $2$ $.680\!\times\! 10^{-17}$ $0$
$35$ $-1$ $.700\!\times\! 10^{-13}$ $5$ $.011\!\times\! 10^{-13}$ $-4$ $.966\!\times\! 10^{-13}$ $1$ $.655\!\times\! 10^{-13}$ $0$ $0$
$36$ $-2$ $.980\!\times\! 10^{-13}$ $5$ $.956\!\times\! 10^{-13}$ $-2$ $.979\!\times\! 10^{-13}$ $7$ $.275\!\times\! 10^{-17}$ $-1$ $.819\!\times\! 10^{-17}$ $0$
$37$ $1$ $.703\!\times\! 10^{-14}$ $-5$ $.084\!\times\! 10^{-14}$ $5$ $.071\!\times\! 10^{-14}$ $-1$ $.690\!\times\! 10^{-14}$ $0$ $0$
$38$ $2$ $.095\!\times\! 10^{-14}$ $-4$ $.191\!\times\! 10^{-14}$ $2$ $.098\!\times\! 10^{-14}$ $-2$ $.473\!\times\! 10^{-17}$ $6$ $.184\!\times\! 10^{-18}$ $0$
$39$ $-1$ $.516\!\times\! 10^{-15}$ $4$ $.542\!\times\! 10^{-15}$ $-4$ $.539\!\times\! 10^{-15}$ $1$ $.513\!\times\! 10^{-15}$ $0$ $0$
$40$ $-1$ $.365\!\times\! 10^{-15}$ $2$ $.732\!\times\! 10^{-15}$ $-1$ $.372\!\times\! 10^{-15}$ $5$ $.620\!\times\! 10^{-18}$ $-1$ $.405\!\times\! 10^{-18}$ $0$
$41$ $1$ $.205\!\times\! 10^{-16}$ $-3$ $.613\!\times\! 10^{-16}$ $3$ $.612\!\times\! 10^{-16}$ $-1$ $.204\!\times\! 10^{-16}$ $0$ $0$
$42$ $8$ $.260\!\times\! 10^{-17}$ $-1$ $.657\!\times\! 10^{-16}$ $8$ $.379\!\times\! 10^{-17}$ $-9$ $.603\!\times\! 10^{-19}$ $2$ $.401\!\times\! 10^{-19}$ $0$
$43$ $-8$ $.622\!\times\! 10^{-18}$ $2$ $.586\!\times\! 10^{-17}$ $-2$ $.586\!\times\! 10^{-17}$ $8$ $.620\!\times\! 10^{-18}$ $0$ $0$
$44$ $-4$ $.659\!\times\! 10^{-18}$ $9$ $.384\!\times\! 10^{-18}$ $-4$ $.824\!\times\! 10^{-18}$ $1$ $.316\!\times\! 10^{-19}$ $-3$ $.291\!\times\! 10^{-20}$ $0$
$45$ $5$ $.600\!\times\! 10^{-19}$ $-1$ $.680\!\times\! 10^{-18}$ $1$ $.680\!\times\! 10^{-18}$ $-5$ $.600\!\times\! 10^{-19}$ $0$ $0$
$46$ $2$ $.459\!\times\! 10^{-19}$ $-4$ $.994\!\times\! 10^{-19}$ $2$ $.648\!\times\! 10^{-19}$ $-1$ $.509\!\times\! 10^{-20}$ $3$ $.772\!\times\! 10^{-21}$ $0$
$47$ $-3$ $.324\!\times\! 10^{-20}$ $9$ $.972\!\times\! 10^{-20}$ $-9$ $.972\!\times\! 10^{-20}$ $3$ $.324\!\times\! 10^{-20}$ $-1$ $.483\!\times\! 10^{-26}$ $2$ $.966\!\times\! 10^{-27}$
$48$ $-1$ $.221\!\times\! 10^{-20}$ $2$ $.516\!\times\! 10^{-20}$ $-1$ $.407\!\times\! 10^{-20}$ $1$ $.487\!\times\! 10^{-21}$ $-3$ $.717\!\times\! 10^{-22}$ $0$
$49$ $1$ $.814\!\times\! 10^{-21}$ $-5$ $.441\!\times\! 10^{-21}$ $5$ $.441\!\times\! 10^{-21}$ $-1$ $.814\!\times\! 10^{-21}$ $9$ $.814\!\times\! 10^{-27}$ $-1$ $.963\!\times\! 10^{-27}$
$50$ $5$ $.735\!\times\! 10^{-22}$ $-1$ $.211\!\times\! 10^{-21}$ $7$ $.343\!\times\! 10^{-22}$ $-1$ $.287\!\times\! 10^{-22}$ $3$ $.217\!\times\! 10^{-23}$ $0$
------ ------ --------------------------- ------ --------------------------- ------ --------------------------- ------ --------------------------- ------ --------------------------- ------ ---------------------------
: Numerical values of expansion coefficients (\[eq:ps\]) of the level-spacing distributions $P_{0}$, $P_{1}$, $P_{2}$, $P_{3}$, $P_{4}$, and $P_{5}$.\[tab1\]
and so forth. Finally, this translates into an expansion for the level-spacing distributions $P_{n}(s)$ via Eqs. (\[eq:en\]) and (\[eq:pn\]).
As Eq. (\[eq:pn\]) involves two derivatives with respect to the variable $s$, we need to obtain the expansions of the functions $\sigma_{n}(x)$ to two additional orders in $x$; so in order to achieve a result for $P_{n}(s)$ that is correct to order $50$ in $s$ we have to compute the expansions of $\sigma_{n}(x)$ to order $52$ in $x/\pi$. Furthermore, we need to expand the exponential function in $D_{0}(s)$ (\[eq:d0exp\]), whereas the integration on $x$ is trivially performed on polynomials. As a result, the expansion coefficients $p_{n;k}$ (\[eq:ps\]) are expressions involving rather lengthy rational numbers and powers of $\pi$. The expansions for $P_{n}(s)$ with $0\le n\le 5$ are given in Appendix \[app1\]; for higher $n>5$ the lowest order in the small-$s$ expansion of $P_{n}(s)$ is larger than $50$. The numerical values of the expansion coefficients are given in Table \[tab1\], compare also the numerical values of the coefficients $p_{0;k}$ for $k\le 32$ given in Ref. [@Haake Table 4.1]. The modulus of the coefficients $p_{n;k}$ appears to decrease relatively rapidly with increasing order $k$, so we may expect the series to converge on a rather large domain around the origin.
Numerical integration
=====================
To give an account of the accuracy of the expansions, we compare them to the result of a numerical integration of the coupled set of differential equations (\[eq:pv0\])–(\[eq:pv2\]) derived from the original Painlevé equation (\[eq:pv\]). These are also obtained using [Mathematica]{} [@Wolfram]. In practice, we use initial values for the functions $\sigma_{n}(x)$ for small, but non-zero $x$ which are obtained from their expansions, and then integrate the system to larger values of $x$.
{width="50.00000%"} \[fig:p0\]
{width="50.00000%"} \[fig:p1\]
In fact, it turns out that this is not as easily done as stated. The reason is that some of the functions involved become very small or very large as $s$ is increased, and the computation of $P_{n}(s)$ essentially involves the cancellation of these large terms. In particular, the functions $\sigma_{n}(x)$ behave asymptotically like [@BTW] $$\begin{aligned}
\sigma_{0}(x)&\sim& -\frac{1}{4}x^{2}\\
\sigma_{n}(x)&\sim& -\frac{n!}{(8\pi)^{n/2}} \frac{\exp(nx)}{x^{n/2-1}}\end{aligned}$$ for large argument $x$. So, for $n>0$, $\sigma_{n}(x)$ grows exponentially with $x$, whereas $P_{n}(s)$ is at most of order one, and eventually decreases as $\exp(-\pi^2 s^2/8)$ for large $s$, see Appendix \[app2\] where the asymptotic behavior of the functions $P_{n}(s)$ for large spacings $s$ is discussed.
{width="50.00000%"} \[fig:p2\]
Results for $n=0$, $n=1$ and $n=2$ are shown in Figs. \[fig:p0\], \[fig:p1\] amd \[fig:p2\], respectively. Apparently, the numerical solutions agree well with the series expansions of Appendix \[app1\] up to about $s\lesssim 2.5$. For $n=0$, the distribution function is well reproduced by the asymptotic form of Eq. (\[eq:p0dys\]) down to $s\gtrsim 1.5$. Similarly, as shown in Figs. \[fig:p1\] and \[fig:p2\], the small-$s$ expansions of Eqs. \[eq:pu1\] and \[eq:pu2\] describe the functions $P_{1}(s)$ and $P_{2}(s)$ for $s\lesssim 2.5$. The asymptotic forms of Eqs. (\[eq:enbtw\])–(\[eq:pnas\]) reproduce the functions $P_{1}(s)$ and $P_{2}(s)$ very well for $s\gtrsim 3$ and $s\gtrsim
4.5$, respectively. In principle, it should be possible to extend the $s^{-1}$ expansion of Eq. (\[eq:enbtw\]) as well, at least for fixed values of $n$, and thus improve the situation considerably.
Concluding remarks
==================
We presented series expansions for the spacing distributions $P_{n}(s)$ of the GUE. The expansion was calculated up to order $50$ in the spacing $s$, and for all distributions $P_{n}(s)$. In practice, this means that for $0\le n\le 5$ the leading terms up to order $s^{50}$ were obtained, whereas for $n>5$ the leading order is larger. The expansions agree with the complete series solutions obtained by Mehta [@Mehta92], which were derived in a different way.
The explicit expansions give a very precise account of the small-spacing part of the spacing distributions, and might prove useful for comparisons with experimental or numerical data. The leading asymptotic behavior for large spacing was obtained in Ref. [@BTW] by similar means.
Within the same framework, albeit slightly more involved, one can derive analogous series expansions for the spacing distributions of the GOE and the GSE.
It is conceivable that other forms of expansions of the level spacing distributions $P_{n}(s)$ for small $s$ might improve the convergence. A candidate might be to start from a Wigner surmise type form as discussed in Ref. [@FW1].
The author is indebted to C. A. Tracy who pointed out Ref. [@BTW] and provided him with the corresponding [Mathematica]{} programs. These established a starting point and inspired the structure of the programs that were used to calculate the expansions presented in this paper. The author also acknowledges useful discussions with P. J. Forrester and R. A. Römer. He also thanks N. S. Witte for spotting some misprints in a couple of coefficients in a previous version of this manuscript, and for bringing Refs.[@DiHa] and [@Mehta92] to his attention.
Leading terms of series expansions {#app1}
==================================
Here we list the leading terms of the small-$s$ expansions for $P_{n}(s)$. The complete expansions can be downloaded as Mathematica code from [@code]. The numerical values of the coefficients up to order $s^{50}$ are given in Table \[tab1\]. $$\begin{aligned}
P_{0}(s) &=& \frac{\pi^{2}\,s^{2}}{3}
-\frac{2\,\pi^{4}\,s^{4}}{45} +
\frac{\pi^{6}\,s^{6}}{315} - \frac{\pi^{6}\,s^{7}}{4050} -
\frac{2\,\pi^{8}\,s^{8}}{14175} +
\frac{11\,\pi^{8}\,s^{9}}{496125} +
\frac{2\,\pi^{10}\,s^{10}}{467775} -
\ldots
\label{eq:pu0}\\
P_{1}(s) &=& \frac{\pi^{6}\,s^{7}}{4050} -
\frac{11\,\pi^{8}\,s^{9}}{496125} +
\frac{13\,\pi^{10}\,s^{11}}{13395375} -
\frac{4586\,\pi^{12}\,s^{13}}{170188239375} -
\frac{\pi^{12}\,s^{14}}{2679075000} +
\ldots
\label{eq:pu1}\\
P_{2}(s) &=& \frac{\pi^{12}\,s^{14}}{5358150000} -
\frac{17\,\pi^{14}\,s^{16}}{1181472075000} +
\frac{1577\,\pi^{16}\,s^{18}}{2859162421500000} -
\ldots
\label{eq:pu2}\\
P_{3}(s) &=& \frac{\pi^{20}\,s^{23}}{9378525331350000000} -
\frac{13\,\pi^{22}\,s^{25}}{1702202347640025000000} + \ldots
\label{eq:pu3}\\
P_{4}(s) &=& \frac{\pi^{30}\,s^{34}}{30645402510264863844600000000000} -
\ldots
\label{eq:pu4}\\
P_{5}(s) &=&
\frac{\pi^{42}\,s^{47}}{255963589608666174754500410100972300000000000000} -
\ldots
\label{eq:pu5}\end{aligned}$$ For the level-spacing distributions $P_{n}(s)$ with $n>5$ the leading terms are of higher order than $s^{50}$.
Asymptotic behavior {#app2}
===================
The large-$s$ asymptotics of the level-spacing distributions $P_{n}(s)$ are also known. For $n=0$, Dyson’s asymptotic result [@Dyson5; @Haake] reads $$E^{\text{(a)}}_{0}(s) =
\left(\frac{2}{\pi s}\right)^{1/4}
\exp\left(\frac{\ln 2}{12} + 3\zeta'(-1)-\frac{\pi^2 s^2}{8}\right)
\label{eq:e0dys}$$ and thus $$\begin{aligned}
P^{\text{(a)}}_{0}(s)
&=& \frac{d^2}{ds^2}E^{\text{(a)}}_{0}(s)\nonumber\\
&=& \frac{\pi^4}{16}\left(s^2 - \frac{2}{\pi^2}
+ \frac{5}{\pi^4 s^2}\right) E^{\text{(a)}}_{0}(s).
\label{eq:p0dys}\end{aligned}$$ Here, $\zeta'(-1)\approx -0.165421$ denotes the derivative of Riemann’s $\zeta$-function evaluated at $-1$.
For $n>0$, the asymptotics are given in Ref. [@BTW]. For the GUE, this gives $$\begin{aligned}
\lefteqn{\frac{E^{\text{(a)}}_{n}(s)}{E^{\text{(a)}}_{0}(s)}
= \frac{B_{n}\exp(n\pi s)}{s^{n^2/2}}\,
\Big(1 + \frac{(2n^{2}+7)n}{8\,\pi\, s} }
\nonumber\\
&& \hphantom{\times\Big(}
+ \frac{(4n^{4} + 48n^{2} + 229)n^{2}}{128\,(\pi\, s)^{2}}
+ O[(\pi\, s)^{-3}]\Big)
\label{eq:enbtw}\end{aligned}$$ where $$B_{n} = 2^{-\frac{n^{2}+2n}{2}}\,\pi^{-\frac{n^{2}+n}{2}}\,
\prod_{m=1}^{n-1} m! \; .$$ From this, we can easily calculate the corresponding asymptotic behavior $P^{\text{(a)}}_{n}(s)$ as $$P^{\text{(a)}}_{n}(s) \; = \;
\frac{d^2}{ds^2} \; \sum_{m=0}^{n}\: (n-m+1)\: E^{\text{(a)}}_{m}(s),
\label{eq:pnas}$$ which is just Eq. (\[eq:pn\]) applied to the asymptotic expressions. The resulting terms are rather involved; however, it is apparent from Eqs. (\[eq:e0dys\]), (\[eq:enbtw\]), and (\[eq:pnas\]) that the leading behavior in all cases is $\exp(-\pi^{2}s^{2}/8)$.
[10]{}
E. P. Wigner, On the statistical distribution of the widths and spacings of nuclear resonance levels, Proc. Camb. Philos. Soc. **47**, 790–798 (1951).
F. J. Dyson, Statistical theory of energy levels of complex systems, I, J. Math. Phys. **3**, 140–156 (1962).
F. J. Dyson, Statistical theory of energy levels of complex systems, II, J. Math. Phys. **3**, 157–165 (1962).
F. J. Dyson, Statistical theory of energy levels of complex systems, III, J. Math. Phys. **3**, 166–175 (1962).
F. J. Dyson, Brownian motion model for the eigenvalues of a random matrix, J. Math. Phys. **3**, 1199–1215 (1962).
F. J. Dyson and M. L. Mehta, Statistical theory of energy levels of complex systems, IV, J. Math. Phys. **4**, 701–712 (1963).
F. Haake, Quantum Signatures of Chaos (Springer, Berlin,1991).
T. Guhr, A. -Groeling and H. A. ", Random-matrix theories in quantum physics: common concepts, Phys. Rep. **299**, 189–425 (1999).
M. L. Mehta, Random Matrices, second ed. (Academic Press, Boston, 1991).
C. A. Tracy and H. Widom, Universality of the distribution functions of random matrix theory, in: Statistical Physics on the Eve of the 21st Century: In Honour of J. B. McGuire on the Occasion of His 65th Birthday, edited by M. T. Batchelor and L. T. Wille (World Scientific, Singapore, 1999)
J.-X. Zhong, U. Grimm, R. A. and M. Schreiber, Level-spacing distributions of planar quasiperiodic tight-binding models, Phys. Rev. Lett. **80**, 3996–3999 (1998); cond-mat/9710006.
M. Schreiber, U. Grimm, R. A. and J.-X. Zhong, Application of random matrix theory to quasiperiodic systems, Physica A **266**, 477–480 (1999); cond-mat/9809370.
M. Schreiber, U. Grimm, R. A. and J.-X. Zhong, Energy levels of quasiperiodic Hamiltonians, spectral unfolding, and random matrix theory, Comp. Phys. Commun. **121–122**, 499–501 (1999); cond-mat/9811359.
U. Grimm, R. A. , M. Schreiber and J.-X. Zhong, Universal level-spacing statistics in quasiperiodic tight-binding models, Mat. Sci. Eng. A **294–296**, 564–567 (2000); cond-mat/9908063.
U. Grimm and M. Schreiber, Aperiodic Tilings on the Computer, in: Quasicrystals: An Introduction to Structure, Physical Properties and Applications, edited by J.-B. Suck, M. Schreiber and P. Häussler, Springer Series in Materials Science, vol. 55 (Springer, Berlin, 2002) pp. 49–66; cond-mat/9903010.
U. Grimm and M. Schreiber, Energy spectra and eigenstates of quasiperiodic tight-binding Hamiltonians, in: Quasicrystals: Structure and Physical Properties, edited by H.-R. Trebin, (Wiley-VCH, Berlin, 2003) pp. 210–235; cond-mat/0212140.
A. M. Odlyzko, On the distribution of spacings between zeros of the zeta function, Math. Comp. **48**, 273–308 (1987).
B. Dietz and F. Haake, Taylor and Pade analysis of the level spacing distributions of random-matrix ensembles, Z. Phys. B **80**, 153–158 (1990).
M. L. Mehta, Power series for level spacing functions of random matrix ensembles, Z. Phys. B **86**, 285–290 (1992).
N. S. Witte, P. J. Forrester and C. M. Cosgrove, Integrability, random matrices and Painlevé transcendents, Preprint math-ph/0008033.
P. J. Forrester and N. S. Witte, Application of the $\tau$-function theory of Painlevé equations to random matrices: PIV, PII and the GUE, Preprint math-ph/0103025.
P. J. Forrester and N. S. Witte, Discrete Painlevé equations and random matrix averages, Preprint math-ph/0304020 (2003).
P. J. Forrester, Growth models, random matrices and Painlevé transcendents, Nonlinearity **16**, R27–R49 (2003); math-ph/0309058.
E. L. Basor, C. A. Tracy and H. Widom, Asymptotics of level-spacing distributions for random matrices, Phys. Rev. Lett. **69**, 5–8 (1992).
M. Jimbo, T. Miwa, Y. and M. Sato, Density matrix of an impenetrable Bose gas and the fifth transcendent, Physica D **1**, 80–158 (1980).
S. Wolfram, Mathematica: A System for Doing Mathematics by Computer, second ed. (Addison-Wesley, Reading MA, 1991).
U. Grimm, LevelSpacingGUE.m, [http://mcs.open.ac.uk/ugg2/RMT/LevelSpacingGUE.m]{}
P. J. Forrester and N. S. Witte, Exact Wigner surmise type evaluation of the spacing distribution in the bulk of the scaled random matrix ensembles, Lett. Math. Phys. **53**, 195–200 (2000); math-ph/0009023.
F. J. Dyson, Fredholm determinants and inverse scattering problems, Commun. Math. Phys. **47**, 171–183 (1976).
[^1]: Corresponding author: e-mail: [[email protected]]{}, Phone: +441908659991, Fax: +441908652140
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Yuta Hamada,'
- Tatsuo Kobayashi
- and Shohei Uemura
title: |
1.7cm **Flavor structure in D-brane models:\
Majorana neutrino masses0.5cm**
---
Introduction
============
The Standard Model has been confirmed by the discovery of the Higgs scalar and other precision measurements. However, it has various mysteries still. One of them is the mystery on the flavor structure. Why are there three generations ? Why are quark and lepton masses hierarchical ? Which mechanism determines their mixing angles ? Indeed, the Yukawa sector has most of free parameters in the Standard model. Discrete flavor symmetries would be important to understand fermion masses and mixing angles [@Altarelli:2010gt; @Ishimori:2010au; @King:2013eh]. For example, the mixing matrix in the lepton sector, the PMNS matrix, can be approximated by the tri-bimaximal mixing matrix in the limit $\theta_{13}=0$ [@Harrison:2002er]. In field-theoretical model building, one starts with a large flavor symmetry. Then, one assumes that the flavor symmetry breaks properly into $Z_3$ and $Z_2$ subsymmetries in the charged lepton or the neutrino masses, such that the tri-bimaximal mixing can be realized.
Superstring theory is a promising candidate for unified theory of all of the interactions including gravity and all of the matter fields and Higgs field(s) (see for a review [@Ibanez]). It is found that superstring theory on six-dimensional compact space leads to interesting flavor structures. In particular, certain types of four-dimensional superstring models with rather simple six-dimensional compact spaces such as tori and orbifolds lead to definite discrete flavor symmetries. For example, intersecting D-brane models and magnetized D-brane models are among interesting model building in superstring theory [@Bachas:1995ik; @Berkooz:1996km; @Blumenhagen:2000wh; @Aldazabal:2000dg; @Angelantonj:2000hi; @Ibanez:2001nd] (see for review [@Blumenhagen:2006ci; @Ibanez] and references therein). These intersecting/magnetized D-brane models can lead to discrete flavor symmetries such as $D_4$, $\Delta(27)$, $\Delta(54)$ [@Abe:2009vi; @Abe:2009uz; @BerasaluceGonzalez:2012vb].[^1] Similar discrete flavor symmetries can be derived in heterotic string theory on orbifolds [@Kobayashi:2004ya].[^2] In these models, we can calculate explicitly Yukawa couplings and higher order couplings [@Hamidi:1986vh; @Cvetic:2003ch; @Cremades:2004wa]
However, such discrete flavor symmetries may be broken by non-perturbative effects. From such a viewpoint, anomalies of discrete symmetries [@Araki:2008ek; @Araki:2007ss; @Nilles:2013lda; @Bizet:2013wha; @Honecker:2013hda] are important because anomalous symmetries may be broken by non-perturbative effects. Even anomaly-free U(1) gauge symmetries can be broken when axions couple with U(1) gauge bosons and they become massive. Furthermore, as concrete non-perturbative effects, D-brane instanton effects have been studied [@Blumenhagen:2006xt] (see also for a review [@Blumenhagen:2009qh] and references therein). From the viewpoint of flavor physics, one of important points is that D-brane instanton effects can generate right-handed Majorana neutrino masses [@Ibanez:2006da; @Ibanez:2007rs; @Cvetic:2007ku]. Then, it is also important to investigate patterns of right-handed Majorana neutrino mass matrices derived by D-brane instanton effects and study whether such effects break some or all of discrete flavor symmetries and which symmetries remain unbroken.
In this paper, we study the flavor structure in intersecting D-brane models as well as magnetized D-brane models. We study anomalies of discrete flavor symmetries derived in intersecting D-brane models. We also study right-handed Majorana neutrino mass matrices, which can be generated by D-brane instanton effects. We show which types of Majorana mass matrices can be derived and which flavor symmetries remain unbroken even with right-handed Majorana neutrino mass matrices generated by D-brane instanton effects.
This paper is organized as follows. In section 2, we review briefly the discrete flavor symmetries derived from intersecting D-brane models as well as magnetized D-brane models. In section 3, we study anomalies of these discrete flavor symmetries[.]{} In section 4, we study right-handed Majorana masses generated by D-brane instanton effects. Section 5 is devoted to conclusion and discussion. [In Appendix A, we show the computation to integrate non-vanishing Wilson line phase.]{}
Discrete flavor symmetries
==========================
In this section, we review briefly discrete flavor symmetries appearing in intersecting D-brane models as well as magnetized D-brane models [@Abe:2009vi; @BerasaluceGonzalez:2012vb]. For concreteness, we consider IIA D6-brane models on $T^6=T^2_1\times T^2_2 \times T^2_3$, where each D6-brane wraps one-cycle of each $T^2$ of $T^6=T^2_1\times T^2_2 \times T^2_3$. That is, our setup is as follows. We consider $N_a$ stacks of D6-branes, which lead to $U(N_a)$ gauge symmetry, and they have winding numbers $(n_a^i,m_a^i)$ along the $x_i$ and $y_i$ directions on $T^2_i$, where we use orthogonal coordinates $(x_i,y_i)$ on $T^2_i$. When we denote the basis of one-cycles on $T^2_i$ by $[a_i]$ and $[b_i]$, which correspond to the $x_i$ and $y_i$ directions, the three-cycle, along which this set of D6-brane winds, is represented by $$\begin{aligned}
[\Pi_a] = \prod^3_{i=1} (n^i_a [a_i] + m^i_a[b_i]).\end{aligned}$$
Here, we consider two sets of D-branes, one set is $N_a$ stacks of D6-branes and another is $N_b$ stacks of D6-branes. These lead to $U(N_a)\times U(N_b)$ gauge groups. Suppose that these two stacks of D6-branes intersect each other on $T^2_i$. Their intersecting number on $T^2_i$ is obtained by $$\begin{aligned}
I_{ab}^{(i)}= (n^i_a m_b^i - m_a^in_b^i),\end{aligned}$$ and their total intersecting number on $T^6$ is obtained by $$\begin{aligned}
[\Pi_a]\cdot[\Pi_b] = I_{ab} = \prod_{i=1}^3 I_{ab}^{(i)}.\end{aligned}$$ Then, chiral matter fields with bi-fundamental representations $(N_a,\bar N_b)_{(1,-1)}$ under $U(N_a)\times U(N_b)$ appear at intersecting points on $T^2_i$, where the index $(1,-1)$ denotes $U(1)^2$ charges inside $U(N_a)$ and $U(N_b)$. There appear $I_{ab}$ families of bi-fundamental matter fields. When $I_{ab}$ is negative, there appear $|I_{ab}|$ families of matter fields with the conjugate representation $(\bar N_a, N_b)_{(-1,1)}$.
The total flavor symmetry is a direct product of flavor symmetries appearing on one of $T^2_i$. Thus, we concentrate on the flavor symmetry realized on one of $T^2_i$. Then, we denote $I_{ab}^{(i)}=g$. Theses modes on $T^2_i$ have definite $Z_g$ charges and $Z_g$ transformation is represented by $$\begin{aligned}
Z = \left(
\begin{array}{ccccc}
1 & & & & \\
& \rho & & & \\
& & \rho^2 & & \\
& & & \ddots & \\
& & & & \rho^{g-1}
\end{array}
\right),\end{aligned}$$ where $\rho = e^{2\pi i/g}$. In addition, there is a cyclic permutation symmetry $Z_g^{(C)}$ among these modes, i.e. $$\begin{aligned}
C= \left(
\begin{array}{cccccc}
0 & 1 & 0 & 0 & \cdots & 0 \\
0 & 0 & 1 & 0 & \cdots & 0 \\
& & & & \ddots & \\
1 & & & & \cdots & 0
\end{array}
\right).\end{aligned}$$ Furthermore, these elements do not commute each other, $$\begin{aligned}
CZ = \rho Z C.\end{aligned}$$ Thus, this flavor symmetry includes another $Z_g'$ symmetry, which is represented by $$\begin{aligned}
Z' = \left(
\begin{array}{ccccc}
\rho & & & & \\
& & & \ddots & \\
& & & & \rho
\end{array}
\right).\end{aligned}$$ Then, these would generate the non-Abelian flavor symmetry, $(Z_g \times Z_g') \rtimes Z_g^{(C)}$.
For example, when $g=2$ and $g=3$, the symmetries correspond to $D_4$ and $\Delta(27)$. In addition, when the totally D-brane system has the $Z_2$ reflection symmetry $P$ between $i$-th mode and $(g-i)$-th mode the $\Delta(27)$ symmetry for $g=3$ is enhanced into $\Delta(54)$ [@Abe:2009vi].
Similarly, we can discuss models with more than two sets of D-branes. For example, suppose that we add $N_c$ stacks of D-branes to the above system, and that their intersecting numbers satisfy $G.C.D.(I_{ab}^{(i)},I_{ac}^{(i)},I_{bc}^{(i)})=d$. Then, this model has the discrete flavor symmetry $(Z_d \times Z_d') \rtimes Z_d^{(C)}$.
The above result is applicable to intersecting D-brane models on orientifolds through simple extension. Also, we can extend our discussions to orbifold cases [@Abe:2009vi; @Abe:2009uz; @BerasaluceGonzalez:2012vb].
Since magnetized D-brane models are T-duals to intersecting D-brane models, the magnetized D-brane models also have the same discrete flavor symmetries. For example, we start with $(N_a + N_b)$ stacks of D9-branes on $T^6$. Then, we introduce the magnetic flux on $T_i^2$ along $U(1)_a$ and $U(1)_b$ directions in $U(N_a + N_b)$ as $$\begin{aligned}
F^{(i)} = 2\pi \left(
\begin{array}{cccccc}
M_a^{(i)} & & & & & \\
& \ddots & & & & \\
& & M_a^{(i)} & & & \\
& & & M_b^{(i)} & & \\
& & & & \ddots & \\
& & & & & M_b^{(i)} \\
\end{array} \right),
\end{aligned}$$ where $M_a^{(i)}$ and $M_b^{(i)}$ are integers. This magnetic flux background breaks the gauge group $U(N_a + N_b)$ into $U(N_a) \times U(N_b)$. The gaugino fields in the off-diagonal part correspond to the $(N_a,\bar N_b)$ bi-fundamental matter fields under the unbroken $U(N_a) \times U(N_b)$ gauge symmetry. Zero-modes with such representation appear in this model, and the number of zero-modes on $T^2_i$ is equal to $M_a^{(i)}-M_b^{(i)}$. When we denote $M_a^{(i)}-M_b^{(i)}=g$, this magnetized D-brane model leads to the same discrete flavor symmetry, $(Z_g \times Z_g')\rtimes Z_g^{(C)}$ as the above intersecting D-brane model.
Discrete anomalies
==================
In this section, we study anomalies of discrete flavor symmetries.
$U(1)$ anomalies
----------------
Before studying anomalies of discrete flavor symmetries, it is useful to review anomalies of $U(1)$ gauge symmetries. In this subsection, we give a brief review on $U(1)$ anomalies [@Aldazabal:2000dg; @Cvetic:2001nr] (see also [@Blumenhagen:2006ci; @Ibanez]).
First of all, we consider the torus compactification. A D6-brane has a charge of RR 7-form $C_7$. The total charge should vanish in a compact space. That leads to the following tadpole cancellation condition, $$\begin{aligned}
\sum_a N_a [\Pi_a] =0.\end{aligned}$$
The $SU(N_a)^3$ anomaly coefficient is calculated in the intersecting D-brane models by $$\begin{aligned}
A_a = \sum_b I_{ab} N_b,\end{aligned}$$ because there are $I_{ab}$ matter fields with $(N_a,\bar N_b)_{(1,-1)}$ for $I_{ab} >0$ and $|I_{ab}|$ matter fields with $(\bar N_a,N_b)_{(-1,1)}$ for $I_{ab} <0$. However, the tadpole cancellation condition leads to $$\begin{aligned}
[\Pi_a]\cdot \sum_b N_b [\Pi_b] = 0.\end{aligned}$$ That implies that $A_a = 0$, that is, anomaly free.
The $U(1)_a \times SU(N_b)^2$ mixed anomaly coefficient is obtained by $$\begin{aligned}
A_{ab} = N_a I_{ab}.\end{aligned}$$ This anomaly is not always vanishing. However, this anomaly can always be canceled by the Green-Schwarz mechanism, where an axion shifts under the $U(1)$ gauge transformation and the anomalous U(1) gauge boson becomes massive.
The $U(1)$-gravity$^2$ anomaly coefficient is obtained by $$\begin{aligned}
A_{a-{\rm grav}} = N_a \sum_b I_{ab}N_b.\end{aligned}$$ This anomaly is always vanishing when the tadpole cancellation condition is satisfied.
Next, we review on anomalies for the orientifold compactification. That is, we introduce $O6$-branes along the direction $\prod_i [a_i]$. The system must be symmetric under the $Z_2$ reflection, $y_i \rightarrow -y_i$. In this case, we have to introduce a mirror $D6_{a'}$-branes with the winding number $(n_a^i,-m_a^i)$ corresponding to $(n_a^i,m_a^i)$. The O6-brane has (–4) times as RR charge as a D6-brane. Then, the RR-tadpole cancellation condition requires $$\begin{aligned}
\sum_a N_a([\Pi_a] + [\Pi_{a'}]) - 4 [\Pi_{O6}] =0.\end{aligned}$$ $$\begin{aligned}
\sum_{a \neq b}N_a [\Pi_b]\cdot ([\Pi_a] + [\Pi_{a'}])
+ N_b [\Pi_b]\cdot [\Pi_{b'}] - 4[\Pi_b]\cdot [\Pi_{O6}] =0.\end{aligned}$$
In addition to $I_{ab}$ families of $(N_a,\bar N_b)_{(1,-1)}$ matter fields, there appear $I_{ab'}$ families of $(N_a,N_b)_{(1,1)}$ matter fields. Moreover, there appear matter fields with symmetric and asymmetric representations under $U(N_a)$ with charge 2. Their numbers are obtained by $$\begin{aligned}
& & \#_{a, {\rm asymm}} = \frac12 ([\Pi_a] \cdot [\Pi_{a'}]-
[\Pi_a]\cdot [\Pi_{O6}]) +[\Pi_a]\cdot [\Pi_{O6}], \\
& & \#_{a, {\rm symm}} = \frac12 ([\Pi_a] \cdot [\Pi_{a'}]-
[\Pi_a]\cdot [\Pi_{O6}]).\end{aligned}$$ In this case, we can show that the $SU(N_a)^3$ anomaly coefficient always vanishes when the RR-tadpole cancellation condition is satisfied, similarly to in the torus compactification. Also, the $U(1)_a-SU(N_b)^2$ anomaly coefficient is not always vanishing, but such anomaly can be canceled by the Green-Schwarz mechanism.
Finally, the $U(1)_a-$gravity$^2$ anomaly coefficient is obtained by $$\begin{aligned}
A_{a-{\rm grav}} &=& \prod_{b \neq} N_a N_b ([\Pi_{a}]\cdot [\Pi_b] +
[\Pi_a]\cdot [\Pi_{b'}]) +2 \frac{N_a(N_a-1)}{2} \#_{a ,{\rm asymm}} \nonumber \\
& &
+2 \frac{N_a(N_a+1)}{2} \#_{a ,{\rm symm}} \nonumber \\
&=& 3N_a[\Pi_a]\cdot[\Pi_{O6}].\end{aligned}$$ This does not always vanish, but such anomaly can be canceled by the Green-Schwarz mechanism.
Discrete anomalies
------------------
In the gauge theory with the gauge group $G$ and the Abelian discrete symmetry $Z_N$, the $Z_N-G^2$ mixed anomaly coefficient is calculated by [@Ibanez:1991hv; @Banks:1991xj; @Araki:2008ek; @Ishimori:2010au], $$\begin{aligned}
A_{Z_N-G^2} =\sum_m q^{(m)} T_2({\bf R}^{(m)}),\end{aligned}$$ where the summation of $m$ is taken over fermions with $Z_N$ charges $q^{(m)}$ and the representation ${\bf R}^{(m)}$ under $G$. Here, $T_2({\bf R}^{(m)})$ denotes the Dynkin index and we use the normalization such that $T_2=1/2$ for the fundamental representation of $SU(N)$. When the following condition is satisfied [@Ibanez:1991hv; @Banks:1991xj; @Araki:2008ek; @Ishimori:2010au], $$\begin{aligned}
\sum_m q^{(m)} T_2({\bf R}^{(m)}) =0 ~~({\rm mod}~~N/2),\end{aligned}$$ the $Z_N$ symmetry is anomaly-free. Similarly, we can calculate the $Z_N$-gravity$^2$ anomaly coefficient by Tr$ q^{(m)}$. If Tr$ q^{(m)} = 0$ (mod $N/2$), $Z_N$ is anomaly-free. For example, $Z_2$ symmetry is always anomaly-free.
Each generator of non-Abelian discrete symmetries corresponds to an Abelian symmetry. Thus, if each Abelian generator of non-Abelian discrete flavor symmetry satisfies the above anomaly-free condition, the total non-Abelian symmetry is anomaly-free. When some discrete Abelian symmetries are anomalous, the total non-Abelian discrete symmetry is broken, and the subgroup, which does not include anomalous generators, remains unbroken.
In the non-Abelian discrete symmetry, there appear multiplets and each generator is represented by a matrix, $M$. When $\det M=1$, the corresponding Abelian discrete symmetry is always anomaly-free. Only multiplets with $\det M \neq 1$ can contribute on anomalies. Since we have $\det Z' = 1$, the corresponding $Z_g'$ symmetry is always anomaly-free. On the other hand, we find $\det Z = \det C= 1$ for $g=$ odd and $\det Z = \det C= -1$ for $g=$ even. That means that the discrete flavor symmetry $(Z_g \times Z_g') \rtimes Z_g^{(C)}$ is always anomaly-free for $g=$ odd, but $Z_g$ and $Z_g^{(C)}$ can be anomalous for $g=$ even. In particular, their $Z_2$ parts are anomalous. One has to check the anomaly-free condition for such $Z_2$ part for $Z_g$ and $Z_g^{(C)}$. For example, the $\Delta(27)$ flavor symmetry for $g=3$ is always anomaly-free. However, $Z_2$ subgroups of $D_4$ for $g=2$ corresponding to the following elements, $$\begin{aligned}
\left(
\begin{array} {cc}
1 & 0 \\
0 & -1
\end{array}
\right),\qquad
\left(
\begin{array} {cc}
0 & 1 \\
1 & 0
\end{array}
\right),\end{aligned}$$ can be anomalous.
First, we discuss the torus compactification. For simplicity, we concentrate on the flavor symmetry appearing the first torus $T^2_1$ and we assume that all of intersecting numbers on $T^2_1$, $I_{ab}^{1}$, are even. Thus, the total flavor symmetry includes the $Z_2$ symmetry as well as $Z_2^{(C)}$, which can be anomalous. Also, we assume that there appears a trivial symmetry from the other $T^2_2 \times T^2_3$. Now, let us examine the $Z_2-SU(N_a)^2$ anomaly. There are $I_{ab}$ bi-fundamental matter fields with the representation $(N_a,\bar N_b)$. A half of $I_{ab}$ matter fields have even $Z_2$ charge and the others have odd $Z_2$ charge. The anomaly coefficient of $Z_2-SU(N_a)^2$ anomaly can be written by $$\begin{aligned}
\sum_b \frac{I_{ab}}{2} N_b \frac12.\end{aligned}$$ It vanishes because the tadpole cancellation condition, $\sum_b I_{ab}N_b=0$. Thus, this $Z_2$ symmetry is anomaly-free on the torus compactification. Since only this $Z_2$ symmetry can be anomalous and the others are always anomaly-free, the non-Abelian flavor symmetries $(Z_g \times Z_g') \rtimes Z_g^{(C)}$ are always anomaly-free in the torus compactification.
Next, we study the orientifold compactification. Similarly, we can calculate the $Z_2-SU(N_a)^2$ anomaly coefficient, $$\begin{aligned}
& & \sum_{b \neq a} \left( \frac{I_{ab}}{2} N_b + \frac{I_{ab'}}{2} N_{b'}\right) \frac12
+ \frac{N_a-2}{4} \#_{a, {\rm asymm}}
+ \frac{N_a+2}{4} \#_{a, {\rm symm}} \nonumber \\
&=& {\frac{[\Pi_a]\cdot[\Pi_{O6}]}{2}}\end{aligned}$$ That is not always vanishing, but it is proportional to the $U(1)_a$-grav$^2$ anomaly. Thus, this anomaly could be canceled when one requires the axion shift under the $Z_2$ transformation, which is related with the axion shift under $U(1)_a$. In addition, when D6$_a$ branes are parallel to the O6-branes, $Z_2-SU(N_a)^2$ anomaly coefficient is always vanishing.
Majorana neutrino masses
========================
In the previous section, we have studied on anomalies of discrete flavor symmetries. Certain symmetries are anomaly-free. For example, the $\Delta(27)$ flavor symmetry is anomaly-free. Anomalous symmetries can be broken by non-perturbative effects. There is no guarantee that anomaly-free symmetries are not broken by stringy non-perturbative effects. In this section, we consider D-brane instanton effects as concrete non-perturbative effects. We study which form of right-handed Majorana neutrino mass matrix can be generated by D-brane instanton effects. Indeed, following [@Blumenhagen:2006xt; @Blumenhagen:2009qh; @Ibanez:2006da], we study the sneutrino mass matrix assuming that the neutrino mass matrix has the same form and supersymmetry breaking effects are small.
Neutrino mass matrix
--------------------
Here, we study right-handed Majorana neutrino masses, which can be generated by D-brane instanton effects. We assume that $g$ families of right-handed neutrinos $\nu_R^a$ appear by intersections between D6$_c$-brane and D6$_d$ branes, and that their intersecting numbers are equal to $I_{cd}^{(i)} = g$ for the i-th $T^2$ and $I_{cd}^{(j)} =1$ for the other tori. For the moment, let us concentrate on the three-generation model, $I_{cd}=3$, which can be obtained by $(I^{(1)}_{cd},I^{(2)}_{cd},I^{(3)}_{cd}) = (\underline{3,1,1})$, where the underline denotes all the possible permutations. We consider D2-brane instanton, which wraps one-cycle of each $T^2$ of $T^6=T^2\times T^2\times T^2$. We call it $D2_M$-brane. It intersects with D6$_{c}$ brane and D6$_{d}$ brane. At these intersecting points, zero-modes $\alpha_i$ and $\gamma_j$ appear and their numbers are obtained by $I_{Mc}$ and $I_{dM}$. Only if there are two zero-modes for both $\alpha_i$ and $\gamma_j$ the neutrino masses can be generated by D2-brane instanton effect [@Blumenhagen:2006xt; @Blumenhagen:2009qh; @Ibanez:2006da], $$\begin{aligned}
& & M\int d^2\alpha d^2 \gamma ~e^{-d^{ij}_a \alpha_i \nu_R^a \gamma_j}
= Mc_{ab}, \nonumber \\
& & c_{ab} =
\nu^a_R \nu^b_R (\varepsilon_{ij} \varepsilon_{k \ell}d^{ik}_a d^{j \ell}_b),\end{aligned}$$ where the mass scale $M$ would be determined by the string scale $M_{st}$ and the instanton world volume $V$ as $M = M_{st}e^{-V}$. Here, $d^{ij}_a$ is the 3-point coupling coefficient among $\alpha_i$, $\nu_R^a$ and $\gamma_j$ [@Cvetic:2003ch], which we show explicitly in the next subsection. The 3-point coupling coefficient $d_a^{ij}$ can be written by $d_a^{ij}=d_{a1}^{ij}d_{a2}^{ij}d_{a3}^{ij}$, where $d_{ak}^{ij}$ for $k=1,2,3$ is the contribution from the $k$-th torus. In addition, when $\alpha_i$, $\gamma_j$, or $\nu^a$ are localized at a single intersecting point on the $k$-th torus, we omit the indexes such as $d_{ak}^{j}$, $d_{ak}^{i}$, or $d_{k}^{ij}$.
We have to take into account all of the possible $D2_M$-brane configurations, which can generate the above neutrino mass terms. One can obtain two zero-modes of $\alpha_i$ and $\gamma_j$ for the $D2_M$-brane set corresponding to Sp(2) or U(2) gauge group with the intersecting numbers $|I_{Mc}|=|I_{Md}|=1$ [@Ibanez:2007rs] or a single $D2_M$-brane with the intersecting numbers, $|I_{Mc}|=|I_{Md}|=2$.
When the $D2_M$-brane set corresponds to the Sp(2) or U(2) brane, the zero-modes, $\alpha_i$ and $\gamma_j$, are doublets and the gauge invariance allows the certain couplings, say $\alpha_i$ and $\gamma_i$, but not $\alpha_i$ and $\gamma_j$ for $i \neq j$. When $I_{Mc}=I_{dM}=1$, the following form of the Majorana mass is generated, $$\begin{aligned}
\int d^2\alpha d^2 \gamma ~e^{-d^{11}_a \alpha_1 \nu_R^a \gamma_1-
d^{22}_a \alpha_2 \nu_R^a \gamma_2}
= \nu^a_R \nu^b_R d^{11}_a d^{22}_b.\end{aligned}$$ More explicitly, the following form of mass matrix is obtained [@Ibanez:2007rs], $$\begin{aligned}
M c_{ab} = \left(
\begin{array}{ccc}
d_{1}^{11} d_{1}^{22} & d_{1}^{11}d_{2}^{22} &d_{1}^{11}d_{3}^{22} \\
d_{2}^{11} d_{1}^{22} & d_{2}^{11}d_{2}^{22} &d_{2}^{11}d_{3}^{22} \\
d_{3}^{11} d_{1}^{22} & d_{3}^{11}d_{2}^{22} &d_{3}^{11}d_{3}^{22} \\
\end{array}
\right).\end{aligned}$$ This Majorana mass matrix has the rank one. However, we have to take into account all of the $D2_M$-brane configurations, that is, the position of $D2_M$-brane sets. Thus, we integrate over the position of the $D2_M$-brane sets. Such integration over the $D2_M$-brane position would recover the cyclic permutation symmetry, $Z_{g=3}^{(C)}$. Then, we would obtain the following form of Majorana neutrino mass matrix, $$\begin{aligned}
\label{eq:n-mass-33}
M = \left(
\begin{array}{ccc}
A & B & B\\
B & A & B\\
B & B & A\\
\end{array}
\right).\end{aligned}$$ We will show this form by an explicit calculation in the next subsection. As a result, there remains the cyclic permutation symmetry, $Z_{g=3}^{(C)}$, unbroken, but $Z_{g=3}$ and $Z_{g=3}'$ symmetries are broken by D-brane instanton effects, which generate the Majorana neutrino masses. This form also has the $Z_2$ reflection symmetry $P$. Thus, if the full D-brane system has the $Z_2$ reflection symmetry, the symmetry is enhanced into $S_3$.
Similarly, we can study a single $D2_M$-brane with the intersecting numbers, $|I_{Mc}|=|I_{Md}|=2$. There are two types of $D2_M$-brane instanton configurations leading to $|I_{Mc}|=|I_{Md}|=2$. In one type, we have the configuration with $|I_{Mc}^{(j)}| = |I_{Md}^{(k)}| =2$ for $j \neq k$, and in the other type we have the configuration with $|I_{Mc}^{(j)}| = |I_{Md}^{(j)}| =2$.
In the first case with $|I_{Mc}^{(j)}| = |I_{Md}^{(k)}| =2$ for $j \neq k$, let us set e.g. $j=1$ and $k=2$. Then, the Yukawa coupling $d^{ij}_a$ can be written by $d^{ij}_a=d^i_{a1}d^j_{a2}d_{a3}$. Also we assume that $I_{cd}^{(1)}=3$ and $I_{cd}^{(2)}=I_{cd}^{(3)}=1$. Then, the neutrino mass can be written by $$\varepsilon_{ij} \varepsilon_{k\ell} d_{a}^{ik} d_{b}^{jl}=\varepsilon_{ij} \varepsilon_{kl} d_{a1}^{i}d_{2}^k
d_{3} d_{b1}^{j}d_{2}^\ell d_{3}.$$ However, this vanishes identically [@Ibanez:2006da]. We obtain the same result for $|I_{Mc}^{(j)} |= |I_{Md}^{(k)} |=2$ with $j \neq k$, when $(I^{(1)}_{cd},I^{(2)}_{cd},I^{(3)}_{cd}) = (\underline{3,1,1})$.
On the other hand, if a single $D2_M$-brane configuration with $|I_{Mc}^{(j)}| = |I_{Md}^{(j)}| =2$ is possible, we obtain the non-vanishing neutrino mass matrix $Mc_{ab}$. Then, when we integrate over the position of the $D2_M$-brane instanton, we would obtain the same results as Eq.(\[eq:n-mass-33\]). Thus, the cyclic permutation symmetry $Z_{g=3}^{(C)}$ is recovered.
This result can be extended for models with $g$ flavors of neutrinos. When we take into account all of the possible D-brane instanton configurations, we would realize the neutrino mass matrix $Mc_{ab}$ with the cyclic permutation symmetry $Z_g^{(C)}$, i.e. $$\begin{aligned}
c_{ab}=c_{a'b'} {\rm~~~for~~~}a'=a+1,~b'=b+1.\end{aligned}$$ Also the mass matrix is symmetric, i.e. $c_{ab} = c_{ba}$. For example, we obtain $$\begin{aligned}
c_{ab}=\left(
\begin{array}{cc}
A & B \\
B & A
\end{array}
\right),\end{aligned}$$ for $g=2$ and $$\begin{aligned}
c_{ab}=\left(
\begin{array}{cccc}
A & B & B' & B\\
B & A & B & B'\\
B' & B & A & B \\
B & B' & B & A
\end{array}
\right),\end{aligned}$$ for $g=4$. It is found that the D-brane instantons break $Z_g'$ into $Z_2$ if $g$ is even. Otherwise, the $Z_g'$ symmetry as well as the $Z_g$ symmetry is completely broken. However, the cyclic permutation symmetry remains.[^3]
We have studied the neutrino mass matrix by assuming that the neutrino and sneutrino have the same mass matrix and supersymmetry breaking effect is small [@Blumenhagen:2006xt; @Blumenhagen:2009qh; @Ibanez:2006da]. The important point to derive our result is the cyclic permutation symmetry. Thus, we would obtain the same result if the D-brane instatons do not break such a symmetry but supersymmetry is broken.
Explicit computation
--------------------
Here, we discuss the Majorana neutrino mass matrix by computing explicitly the three-generation models. We consider the D2-brane instanton corresponding to Sp(2) or U(2) gauge symmetry. Suppose that D6$_c$ and D6$_d$ branes have the intersecting number, $I_{cd}=3$, and at three intersecting points there appear three generations of right-handed neutrinos. We set $(I^{(1)}_{cd},I^{(2)}_{cd},I^{3)}_{cd})=(3,1,1)$, and $I_{Mc}=I_{dM}=1$. Because the right-handed neutrinos are localized at different points from each other on the first torus, only the first torus is important for the flavor symmetry. Thus, we concentrate on the first torus for computations on Yukawa couplings and Majorana masses. We also omit the index corresponding to the $k$-torus. In the following computations, we set Wilson line moduli zero because it dose not affect the flavor structure ([see]{} appendix A for more detail).
There are three generations of $\nu_a$ and we here label their flavor index as $a=0,1,2$. Also there are two-zero modes, $\alpha_i$ and $\gamma_j$ $(i,j=1,2)$, but note that these indexes $i,j$ correspond to the doublets under $Sp(2)$ or $U(2)$ and the intersecting numbers, $I_{cM}$, and $I_{Md}$, are equal to one, $I_{cM}=I_{Md}=1$.
Suppose that there are three fields $\phi_a$, $\chi_{i'}$ and $\chi_{j'}$ with the “flavor numbers”, $a=0,\cdots,I_{cd}-1$, $i'=0,\cdots,I_{dM}-1$, and $j'=0,\cdots,I_{Mc}-1$, where $I_{cd}$, $I_{dM}$, and $I_{Mc}$ are the corresponding intersecting numbers on the torus. In this case, the 3-point couplings $d_a^{i'j'}$ among three fields can be calculated by [@Cvetic:2003ch] $$d_{a}^{i'j'}=C \sum_{\ell \in {Z}} {\rm exp}\left( \frac{-A_{ai'j'}(\ell)}{2\pi \alpha'} \right),$$ where $C$ is a flavor-independent constant due to quantum contributions and $$A_{ai'j'}(\ell)=\frac{1}{2} A |I_{cd} I_{dM} I_{Mc}|\left( \frac{a}{I_{cd}}
+ \frac{i'}{I_{dM}}+ \frac{j'}{I_{Mc}}+\frac{\varepsilon}{I_{dM}I_{Mc}} +\ell \right)^2,$$ and $A$ denotes the area of the first torus. Here, $\varepsilon$ denotes the position of $D2_M$-brane on the first torus and we normalize $\varepsilon$ such that $\varepsilon$ varies $[0,1]$ on the torus. Note that this coupling corresponds to the contribution on the first torus, which determines the flavor structure, but we have omitted the index corresponding to the first torus.
By using the $\vartheta$-function, $$\vartheta \left[
\begin{array}{c}
a \\
b \\
\end{array}\right] (\nu,\tau) = \sum_{\ell \in {Z}} {\rm exp} \left[ \pi i (a+\ell)^{2} \tau + 2 \pi i (a+\ell)(\nu + b)
\right],$$ we can write $$d_{a}^{i'j'} = C \vartheta \left[
\begin{array}{c}
\frac{a}{I_{cd}} + \frac{i'}{I_{dM}} + \frac{j'}{I_{Mc}}+\frac{\varepsilon}{I_{dM} I_{Mc}} \\
0 \\
\end{array}
\right] \left( 0,\frac{iA{\textcolor{black}}{|}I_{cd} I_{dM} I_{Mc}|}{4\pi^{2}\alpha'} \right) .
\label{eq:3pcoupling}$$
Our model corresponds to $a=0,1,2$, $I_{cd}=3$, $i'=j'=0$, $I_{dM}=I_{Mc}=1$. In the above model, the 3-point couplings among $\nu_a$, $\alpha_i$, and $\gamma_j$ are written by $$d_{a}^{ij} = \delta_{ij} \vartheta \left[
\begin{array}{c}
-\frac{a}{3} + \varepsilon \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{4\pi^{2}\alpha'} \right).$$ Recall again that the indexes $i$ and $j$ of $\alpha_i$ and $\gamma_j$ are doublet indexes under Sp(2) or U(2).
Using this, the matrix $c_{ab}$ is written by the integration of the position $\varepsilon$ over $[0,1]$, $$\begin{aligned}
c_{ab} & = & \int_{0}^{1} d\varepsilon \vartheta \left[
\begin{array}{c}
-\frac{a}{3} + \varepsilon \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{4\pi^{2}\alpha'} \right)
\vartheta \left[
\begin{array}{c}
-\frac{b}{3} + \varepsilon \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{4\pi^{2}\alpha'} \right) \nonumber \\
& = &
\int_{0}^{1} d\varepsilon \sum_{m=1}^{2}
\vartheta \left[
\begin{array}{c}
-\frac{a}{6} - \frac{b}{6} + \varepsilon + \frac{m}{2}\\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{2\pi^{2}\alpha'} \right) \\
& & \times \vartheta \left[
\begin{array}{c}
-\frac{a}{6} +\frac{b}{6} +\frac{m}{2} \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{2\pi^{2}\alpha'} \right). \nonumber\end{aligned}$$ We obtain $$\begin{aligned}
&&
\int_{0}^{1} d\varepsilon
\vartheta \left[
\begin{array}{c}
-\frac{a}{3} + \varepsilon + \frac{m}{2}\\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{2\pi^{2}\alpha'} \right) \nonumber \\
& & =
\int_{0}^{1} d\varepsilon
\sum_{l \in {Z}} {\rm exp}\left[\pi i (-a/3 + \varepsilon +m/2 +\ell)^{2}\left(\frac{3 i A}{2 \pi^{2} \alpha'} \right) \right] \nonumber \\
& &=
\int_{-\infty}^{\infty} dx {\rm exp} \left[-\frac{3A}{2\pi \alpha'} ( x - a/3 +m/2 )^{2} \right]\\
& & =
\sqrt{\frac{2\pi^{2}\alpha'}{3A}}. \nonumber \end{aligned}$$ Using it, the matrix elements $c_{ab}$ can be computed as follows. It is found that the diagonal elements $c_{aa}$ do not depend on $a$ and they are written by $$c_{aa} = \sqrt{\frac{2\pi^{2}\alpha'}{3A}}
\left(\vartheta \left[
\begin{array}{c}
0 \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{2\pi^{2}\alpha'} \right)
+
\vartheta \left[
\begin{array}{c}
\frac{1}{2} \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{2\pi^{2}\alpha'} \right)
\right).$$ Similarly, the off-diagonal elements are written by $$c_{01} = \sqrt{\frac{2\pi^{2}\alpha'}{3A}}
\left(\vartheta \left[
\begin{array}{c}
\frac{1}{6} \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{2\pi^{2}\alpha'} \right)
+
\vartheta \left[
\begin{array}{c}
\frac{2}{3} \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{2\pi^{2}\alpha'} \right)
\right),$$ $$c_{02} = \sqrt{\frac{2\pi^{2}\alpha'}{3A}}
\left(\vartheta \left[
\begin{array}{c}
\frac{1}{3} \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{2\pi^{2}\alpha'} \right)
+
\vartheta \left[
\begin{array}{c}
\frac{5}{6} \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{2\pi^{2}\alpha'} \right)
\right),$$ $$c_{12} = \sqrt{\frac{2\pi^{2}\alpha'}{3A}}
\left(\vartheta \left[
\begin{array}{c}
\frac{1}{6} \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{2\pi^{2}\alpha'} \right)
+
\vartheta \left[
\begin{array}{c}
\frac{2}{3} \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{2\pi^{2}\alpha'} \right)
\right).$$
However, we have the following formula of the $\vartheta$-function $$\vartheta \left[
\begin{array}{c}
a \\
b \\
\end{array}
\right] ( \nu ,\tau)
=
\vartheta \left[
\begin{array}{c}
a +1 \\
b \\
\end{array}
\right] (\nu,\tau),$$ $$\vartheta \left[
\begin{array}{c}
-a \\
0 \\
\end{array}
\right] ( 0 ,\tau)
=
\vartheta \left[
\begin{array}{c}
a \\
0 \\
\end{array}
\right] (0,\tau).$$
Then, we see that all of the off-diagonal elements are the same, $$\begin{aligned}
c_{01}=c_{12}=c_{20}.
$$ That is, we can realize the form (\[eq:n-mass-33\]) by explicit calculations. Figure \[fig:B/A\] shows the ratio $B/A=c_{12}/c_{aa}$ in (\[eq:n-mass-33\]) by varying the area $3A/2\pi^2\alpha'$.
Phenomenological implication
----------------------------
Here we discuss phenomenological implication of our result. The Majorana mass matrix with the form (\[eq:n-mass-33\]) can be diagonalized by the following matrix, $$\begin{aligned}
\label{eq:mixing}
\left(
\begin{array}{ccc}
\sqrt{2/3}c & 1/\sqrt{3} & -\sqrt{2/3}s \\
-1/\sqrt{6}c-1/\sqrt{2}s & 1/\sqrt{3} & 1/\sqrt{6}s-1/\sqrt{2}c \\
-1/\sqrt{6}c+1/\sqrt{2}s & 1/\sqrt{3} & 1/\sqrt{6}s +1/\sqrt{2}c \\
\end{array}
\right),\end{aligned}$$ where $c=\cos \theta$ and $s= \sin \theta$, and the eigenvalues are $A-B$, $A+2B$ and $A-B$. That is, two eigenvalues are degenerate. This is because the mass matrix (\[eq:n-mass-33\]) has the additional $Z_2$ reflection symmetry $P$ and the symmetry is enhanced into $S_3$. At any rate, this form of the mixing matrix is interesting, although the mass eigenvalues may be not completely realistic.
Suppose that the Dirac neutrino Yukawa couplings and charged lepton mass matrix are almost diagonal.[^4] Then, the lepton mixing matrix is obtained as the above matrix (\[eq:mixing\]). That is the trimaximal matrix.
When $s=0$, the above matrix becomes the tri-bimaximal mixing matrix. In field-theoretical model building, the tri-bimaximal mixing matrix can be obtained as follows [@Altarelli:2010gt; @Ishimori:2010au; @King:2013eh]. We start with a larger flavor symmetry and break by vacuum expectation values of scalar fields. However, one assumes that $Z_3$ and $Z_2$ subsymmetries remain in the charged lepton or neutrino mass terms. Then, the tri-bimaximal mixing matrix can be realized. In our string theory, such a $Z_3$ symmetry is realized by geometrical symmetry of the cyclic permutation $Z_3^{C}$, which can not be broken by the D-brane instanton effects, although other symmetries are broken.
We may need some corrections to realize the experimental values of neutrino masses.[^5] At least, the above results show that we can realize non-trivial mixing in the lepton sector even though our assumption above the Dirac masses can not be realized.
Conclusion and discussion
=========================
We have studied the flavor structure in intersecting D-brane models. We have discussed the anomalies of flavor symmetries. Certain symmetries are anomaly-free, and anomaly coefficients of discrete symmetries have the specific feature. We have studied the Majorana neutrino masses, which can be generated by D-brane instanton effects. It is found that the mass matrix form with the cyclic permutation symmetry can be realized by integrating over the position of D-brane instanton. That would lead to the interesting form of mixing angles. It is interesting to apply our results for more concrete models. We would study numerical analyses elsewhere.
In some models, there appear more than one pair of Higgs fields. Their masses would be generated by D-brane instanton effects. It would be important to study the form of such Higgs mass matrix. Also, some of Yukawa couplings may be generated by D-brane instanton effects. Thus, it would be important to extend our analysis to Higgs mass matrix and Yukawa matrices.
Acknowledgement {#acknowledgement .unnumbered}
---------------
The work of Y. H. is supported in part by the Grant-in-Aid for Japan Society for the Promotion of Science (JSPS) Fellows No.25$\cdot$1107. The work of T.K. is supported in part by the Grants-in-Aid for Scientific No. 25400252 from the Ministry of Education, Culture,Sports, Science and Technology of Japan.
Integration of Wilson line moduli
=================================
Here we integrate Wilson line moduli of the $D2_{M}$-brane. Non-zero Wilson line varies the 3 point coupling (\[eq:3pcoupling\]) to [the following form.]{}
$$d_{a}^{i'j'} = C \vartheta \left[
\begin{array}{c}
\frac{a}{I_{cd}} + \frac{i'}{I_{dM}} + \frac{j'}{I_{Mc}}+\frac{\varepsilon}{I_{dM} I_{Mc}} \\
0 \\
\end{array}
\right] \left( \phi ,\frac{iA |I_{cd} I_{dM} I_{Mc}|}{4\pi^{2}\alpha'} \right) .
\label{eq:3pcoupling_WL}$$
Here, $\phi$ is Wilson line phase. The matrix $c_{ab}$ is written by the integration of the position $\varepsilon$ and Wilson line moduli $\phi$.
$$\begin{aligned}
c_{ab} & = & \int d\phi \int_{0}^{1} d\varepsilon \vartheta \left[
\begin{array}{c}
-\frac{a}{3} + \varepsilon \\
0 \\
\end{array}
\right] \left( \phi ,\frac{3iA}{4\pi^{2}\alpha'} \right)
\vartheta \left[
\begin{array}{c}
-\frac{b}{3} + \varepsilon \\
0 \\
\end{array}
\right] \left( \phi ,\frac{3iA}{4\pi^{2}\alpha'} \right) \nonumber \\
& = &
\int d\phi \int_{0}^{1} d\varepsilon \sum_{m=1}^{2}
\vartheta \left[
\begin{array}{c}
-\frac{a}{6} - \frac{b}{6} + \varepsilon + \frac{m}{2}\\
0 \\
\end{array}
\right] \left( 2\phi ,\frac{3iA}{2\pi^{2}\alpha'} \right) \\
& & \times \vartheta \left[
\begin{array}{c}
-\frac{a}{6} +\frac{b}{6} +\frac{m}{2} \\
0 \\
\end{array}
\right] \left( 0,\frac{3iA}{2\pi^{2}\alpha'} \right). \nonumber\end{aligned}$$
We get $$\begin{aligned}
&&\int d\phi \int_{0}^{1} d\epsilon
\vartheta \left[
\begin{array}{c}
-\frac{a}{6} - \frac{b}{6} + \epsilon + \frac{m}{2}\\ \nonumber
0 \\
\end{array}
\right] \left( 2\phi ,2i \frac{3A}{2\pi^{2} \alpha'} \right)\\ \nonumber
& &=
\int d\phi \int_{0}^{1} d\epsilon \sum_{l\in \bf{Z}} e^{- \frac{3A}{\pi \alpha'}\left(-\frac{a}{6}-\frac{b}{6}+\epsilon +\frac{m}{2} +l \right)^{2}+4\pi i \left(-\frac{a}{6}-\frac{b}{6}+\epsilon +\frac{m}{2} + l \right)\phi}\\ \nonumber
& &=
\int d\phi \int_{0}^{1} d\epsilon \sum_{l\in \bf{Z}} e^{- \frac{3A}{\pi \alpha'} \left(-\frac{a}{6}-\frac{b}{6}+\epsilon +\frac{m}{2} + l + i\frac{ 2\pi^{2} \alpha'}{3A}\phi \right)^2 - \frac{4\pi^{3} \alpha' \phi^{2}}{3A}}\\
& &=
\sqrt{\frac{\pi^{2} \alpha'}{3A}}\int d\phi e^{- \frac{4\pi^{3} \alpha' \phi^{2}}{3A}}.\end{aligned}$$ [This factor is independent of flavor index, but universal. Thus, the integration of Wilson line moduli does not affect flavor structure.]{}
[99]{}
G. Altarelli and F. Feruglio, Rev. Mod. Phys. [**82**]{}, 2701 (2010) \[arXiv:1002.0211 \[hep-ph\]\].
H. Ishimori, T. Kobayashi, H. Ohki, Y. Shimizu, H. Okada and M. Tanimoto, Prog. Theor. Phys. Suppl. [**183**]{}, 1 (2010) \[arXiv:1003.3552 \[hep-th\]\]; Lect. Notes Phys. [**858**]{}, pp.1 (2012); Fortsch. Phys. [**61**]{}, 441 (2013).
S. F. King and C. Luhn, Rept. Prog. Phys. [**76**]{}, 056201 (2013) \[arXiv:1301.1340 \[hep-ph\]\]. P. F. Harrison, D. H. Perkins and W. G. Scott, Phys. Lett. B [**530**]{}, 167 (2002) \[hep-ph/0202074\]; P. F. Harrison and W. G. Scott, Phys. Lett. B [**535**]{}, 163 (2002) \[hep-ph/0203209\]; Phys. Lett. B [**557**]{}, 76 (2003) \[hep-ph/0302025\]. L. E. Ibanez and A. M. Uranga, “String theory and particle physics: An introduction to string phenomenology,” Cambridge University Press (2012).
C. Bachas, hep-th/9503030.
M. Berkooz, M. R. Douglas and R. G. Leigh, Nucl. Phys. B [**480**]{}, 265 (1996) \[hep-th/9606139\]. R. Blumenhagen, L. Goerlich, B. Kors and D. Lust, JHEP [**0010**]{}, 006 (2000) \[hep-th/0007024\].
G. Aldazabal, S. Franco, L. E. Ibanez, R. Rabadan and A. M. Uranga, J. Math. Phys. [**42**]{}, 3103 (2001) \[hep-th/0011073\]; JHEP [**0102**]{}, 047 (2001) \[hep-ph/0011132\].
C. Angelantonj, I. Antoniadis, E. Dudas and A. Sagnotti, Phys. Lett. B [**489**]{}, 223 (2000) \[hep-th/0007090\].
L. E. Ibanez, F. Marchesano and R. Rabadan, JHEP [**0111**]{}, 002 (2001) \[hep-th/0105155\].
R. Blumenhagen, B. Kors, D. Lust and S. Stieberger, Phys. Rept. [**445**]{}, 1 (2007) \[hep-th/0610327\].
H. Abe, K. -S. Choi, T. Kobayashi and H. Ohki, Nucl. Phys. B [**820**]{} (2009) 317 \[arXiv:0904.2631 \[hep-ph\]\]. H. Abe, K. -S. Choi, T. Kobayashi and H. Ohki, Phys. Rev. D [**80**]{} (2009) 126006 \[arXiv:0907.5274 \[hep-th\]\]; Phys. Rev. D [**81**]{} (2010) 126003 \[arXiv:1001.1788 \[hep-th\]\].
M. Berasaluce-Gonzalez, P. G. Camara, F. Marchesano, D. Regalado and A. M. Uranga, JHEP [**1209**]{}, 059 (2012) \[arXiv:1206.2383 \[hep-th\]\]; F. Marchesano, D. Regalado and L. Vazquez-Mercado, JHEP [**1309**]{}, 028 (2013) \[arXiv:1306.1284 \[hep-th\]\].
T. Kobayashi, S. Raby and R. J. Zhang, Nucl. Phys. B [**704**]{}, 3 (2005) \[arXiv:hep-ph/0409098\]; T. Kobayashi, H. P. Nilles, F. Ploger, S. Raby and M. Ratz, Nucl. Phys. B [**768**]{}, 135 (2007) \[arXiv:hep-ph/0611020\]; P. Ko, T. Kobayashi, J. h. Park and S. Raby, Phys. Rev. D [**76**]{}, 035005 (2007) \[Erratum-ibid. D [**76**]{}, 059901 (2007)\] \[arXiv:0704.2807 \[hep-ph\]\]. T. Higaki, N. Kitazawa, T. Kobayashi and K. -j. Takahashi, Phys. Rev. D [**72**]{}, 086003 (2005) \[hep-th/0504019\].
S. Hamidi and C. Vafa, Nucl. Phys. B [**279**]{}, 465 (1987); L. J. Dixon, D. Friedan, E. J. Martinec and S. H. Shenker, Nucl. Phys. B [**282**]{}, 13 (1987); T. T. Burwick, R. K. Kaiser and H. F. Muller, Nucl. Phys. B [**355**]{}, 689 (1991); J. Erler, D. Jungnickel, M. Spalinski and S. Stieberger, Nucl. Phys. B [**397**]{}, 379 (1993) \[hep-th/9207049\]; K. -S. Choi and T. Kobayashi, Nucl. Phys. B [**797**]{}, 295 (2008) \[arXiv:0711.4894 \[hep-th\]\]. T. Kobayashi, S. L. Parameswaran, S. Ramos-Sanchez and I. Zavala, JHEP [**1205**]{}, 008 (2012) \[Erratum-ibid. [**1212**]{}, 049 (2012)\] \[arXiv:1107.2137 \[hep-th\]\].
M. Cvetic and I. Papadimitriou, Phys. Rev. D [**68**]{}, 046001 (2003) \[Erratum-ibid. D [**70**]{}, 029903 (2004)\] \[hep-th/0303083\]; S. A. Abel and A. W. Owen, Nucl. Phys. B [**663**]{}, 197 (2003) \[hep-th/0303124\]; D. Cremades, L. E. Ibanez and F. Marchesano, JHEP [**0307**]{}, 038 (2003) \[hep-th/0302105\]; S. A. Abel and A. W. Owen, Nucl. Phys. B [**682**]{}, 183 (2004) \[hep-th/0310257\]. D. Cremades, L. E. Ibanez and F. Marchesano, JHEP [**0405**]{}, 079 (2004) \[hep-th/0404229\]; H. Abe, K. -S. Choi, T. Kobayashi and H. Ohki, JHEP [**0906**]{}, 080 (2009) \[arXiv:0903.3800 \[hep-th\]\]; Y. Hamada and T. Kobayashi, Prog. Theor. Phys. [**128**]{}, 903 (2012) \[arXiv:1207.6867 \[hep-th\]\].
M. Berasaluce-Gonzalez, L. E. Ibanez, P. Soler and A. M. Uranga, JHEP [**1112**]{}, 113 (2011) \[arXiv:1106.4169 \[hep-th\]\]. L. E. Ibanez, A. N. Schellekens and A. M. Uranga, Nucl. Phys. B [**865**]{}, 509 (2012) \[arXiv:1205.5364 \[hep-th\]\]. M. Berasaluce-Gonzalez, P. G. Camara, F. Marchesano and A. M. Uranga, JHEP [**1304**]{}, 138 (2013) \[arXiv:1211.5317 \[hep-th\]\]. P. Anastasopoulos, M. Cvetic, R. Richter and P. K. S. Vaudrevange, JHEP [**1303**]{}, 011 (2013) \[arXiv:1211.1017 \[hep-th\]\].
G. Honecker and W. Staessens, JHEP [**1310**]{}, 146 (2013) \[arXiv:1303.4415 \[hep-th\]\]. N. G. Cabo Bizet, T. Kobayashi, D. K. Mayorga Pena, S. L. Parameswaran, M. Schmitz and I. Zavala, JHEP [**1305**]{}, 076 (2013) \[arXiv:1301.2322 \[hep-th\]\]. H. P. Nilles, S. ulRamos-Sanchez, M. Ratz and P. K. S. Vaudrevange, Phys. Lett. B [**726**]{}, 876 (2013) \[arXiv:1308.3435 \[hep-th\]\]. N. G. C. Bizet, T. Kobayashi, D. K. M. Pena, S. L. Parameswaran, M. Schmitz and I. Zavala, arXiv:1308.5669 \[hep-th\].
T. Araki, T. Kobayashi, J. Kubo, S. Ramos-Sanchez, M. Ratz and P. K. S. Vaudrevange, Nucl. Phys. B [**805**]{}, 124 (2008) \[arXiv:0805.0207 \[hep-th\]\]. T. Araki, K. -S. Choi, T. Kobayashi, J. Kubo and H. Ohki, Phys. Rev. D [**76**]{}, 066006 (2007) \[arXiv:0705.3075 \[hep-ph\]\].
R. Blumenhagen, M. Cvetic and T. Weigand, Nucl. Phys. B [**771**]{}, 113 (2007) \[hep-th/0609191\]. R. Blumenhagen, M. Cvetic, S. Kachru and T. Weigand, Ann. Rev. Nucl. Part. Sci. [**59**]{}, 269 (2009) \[arXiv:0902.3251 \[hep-th\]\].
L. E. Ibanez and A. M. Uranga, JHEP [**0703**]{}, 052 (2007) \[hep-th/0609213\].
L. E. Ibanez, A. N. Schellekens and A. M. Uranga, JHEP [**0706**]{}, 011 (2007) \[arXiv:0704.1079 \[hep-th\]\]; S. Antusch, L. E. Ibanez and T. Macri, JHEP [**0709**]{}, 087 (2007) \[arXiv:0706.2132 \[hep-ph\]\].
M. Cvetic, R. Richter and T. Weigand, Phys. Rev. D [**76**]{}, 086002 (2007) \[hep-th/0703028\].
M. Cvetic, G. Shiu and A. M. Uranga, Nucl. Phys. B [**615**]{}, 3 (2001) \[hep-th/0107166\].
L. E. Ibanez and G. G. Ross, Phys. Lett. B [**260**]{}, 291 (1991). T. Banks and M. Dine, Phys. Rev. D [**45**]{}, 1424 (1992) \[hep-th/9109045\].
[^1]: See also [@Higaki:2005ie].
[^2]: See for recent works on other discrete stringy symmetries, e.g. [@BerasaluceGonzalez:2011wy; @Ibanez:2012wg; @BerasaluceGonzalez:2012zn; @Anastasopoulos:2012zu; @Honecker:2013hda; @Bizet:2013gf; @Nilles:2013lda; @Bizet:2013wha] .
[^3]: These forms also have the $Z_2$ reflection symmetry.
[^4]: The $\Delta(27)$ flavor symmetry as well as $\Delta(54)$ flavor symmetry may be useful to realize such a form.
[^5]: To resolve the degeneracy between two mass eigenvalues, it may be important to break the $Z_2$ reflection symmetry $P$. The full D-brane system, i.e. the full Lagrangian of the low-energy effective field theory, may not have such $Z_2$ symmetry and the above degeneracy may be resolved by radiative corrections.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The skew of a binary string is the difference between the number of zeroes and the number of ones, while the length of the string is the sum of these two numbers. We consider certain suffixes of the lexicographically-least de Bruijn sequence at natural breakpoints of the binary string. We show that the skew and length of these suffixes are enumerated by sequences generalizing the Fibonacci and Lucas numbers, respectively.'
author:
- Joshua Cooper
- 'Christine E. Heitsch'
bibliography:
- 'citepoints.bib'
title: ' Generalized Fibonacci recurrences and the lex-least De Bruijn sequence'
---
Introduction {#intro}
============
Let $w = a_{1} a_{2} \ldots a_{l}$ be a word over the alphabet $\{0,1\}$ of length $|w| = l$. When $|w| = 2^{n}$ and the indices of $w$ are interpreted cyclically, the word is said to be a *binary de Bruijn sequence of order $n$* if it contains each of the $2^{n}$ distinct binary strings of length $n$ as a subword. The string $00010111$ is a binary de Bruijn sequences of order $3$.
A *binary necklace* is an equivalence class of binary words under rotation. The representative element for the equivalence class is chosen to be the lexicographically least one. A binary string is a *Lyndon word* if it is an aperiodic necklace representative. The binary Lyndon words of length $\leq 4$ are $0$, $1$, $01$, $001$, $011$, $0001$, $0011$, $0111$.
De Bruijn sequences and Lyndon words are related via the “Ford sequence,” denoted here [$\mathcal{F}_{n}$]{}, which is the lexicographically least binary de Bruijn sequence of order $n$. Fredricksen proved [@fredricksen-82] that [$\mathcal{F}_{n}$]{} is obtained by concatenating all Lyndon words of lengths dividing $n$ in lexicographic order. For instance, $\mathcal{F}_{4} = 0000100110101111$. We note that this result generalizes to constructing the lexicographically-least de Bruijn sequence over an arbitrary alphabet [@fredricksen-maiorana-78; @moreno-04].
The Ford sequence is also generated by applying a greedy strategy to the production of a binary de Bruijn sequence. The algorithm constructs [$\mathcal{F}_{n}$]{} one bit at a time, preferring $0$’s to $1$’s whenever possible. Given this, it is reasonable to expect that initial segments of [$\mathcal{F}_{n}$]{} contain many more zeros than ones. In fact, previous work [@ford1] shows that the maximum difference (called the discrepancy) of [$\mathcal{F}_{n}$]{} is $\Theta(2^n \log n/n)$.
The discrepancy is the maximum possible “skew” over all prefixes of [$\mathcal{F}_{n}$]{}. The [*skew*]{} of a binary string $w$ of length $l$, denoted $\operatorname{sk}(w)$, is the difference between the number of zeros and the number of ones. Since the length of $w$ is the sum of these two numbers, we have that $$\operatorname{sk}(w) = \sum_{i = 1}^{l} (-1)^{a_{i}}
\mbox{ and }
|w| = \sum_{i=1}^{l} (1)^{a_{i}}.$$
Figure \[fslabels\] illustrates the discrepancy for $n = 4,5,6,7$ by graphing the skew of all prefixes of [$\mathcal{F}_{n}$]{}. As illustrated on the graphs, there are natural breakpoints in [$\mathcal{F}_{n}$]{}following the occurrence of the subword $0^i 1^{n-i}$ for $1 \leq i \leq n -1$. The cases where $i = 0$ and $i = n$ are the final $1$ and initial $0$, respectively.
\
In this article, we prove that the skew at these breakpoints gives sequences of values which are Fibonacci-like. Our results are given in terms of suffixes of the Ford sequence, which are directly related to prefixes by $\operatorname{sk}({\ensuremath{\mathcal{F}_{n}}}) = 0$ and $|{\ensuremath{\mathcal{F}_{n}}}| = 2^n$. As stated precisely in Theorem \[main\] below, we show that the skew and the length of these breakpoint suffixes of [$\mathcal{F}_{n}$]{} are enumerated by sequences generalizing the Fibonacci and Lucas numbers, respectively.
Preliminaries and statement of main result {#prelim}
==========================================
Let $\ell_{0} = 1$ and, for $1 \leq i \leq n - 1$, let $\ell_{i}$ be the subword of [$\mathcal{F}_{n}$]{} which begins immediately after the Lyndon word $0^{i+1}1^{n-i-1}$ and ends with the string $0^{i} 1^{n-i}$. Hence, $\ell_{i}$ consists of all Lyndon words of length $d \mid n$, in lexicographic order, which contain the substring $0^i$ but not $0^{i+1}$. For technical reasons, if $i > n - 1$, then we define $\ell_{i} = \varepsilon$, the empty string.
Let $\mathcal{L}_{m}$ be the concatenation of $\ell_{m} \ell_{m-1} \ldots \ell_{1}$ for a given [$\mathcal{F}_{n}$]{}. Hence, $\mathcal{L}_{m}$ is the substring of the $n$-th Ford sequence which contains the Lyndon words of length $d > 1$ with at least one $0$ and at most $m$ consecutive $0$’s. Let [$\mathcal{K}_{m}$]{} be the proper suffix of [$\mathcal{F}_{n}$]{} consisting of the Lyndon words of length $d \mid n$ containing at most $m$ consecutive $0$’s; $${\ensuremath{\mathcal{K}_{m}}}= \mathcal{L}_{m} 1 = \ell_{m} \ell_{m-1} \ldots \ell_{1} \ell_{0} \mbox{.}$$ If $m \geq n - 1$, then ${\ensuremath{\mathcal{K}_{m}}}$ is $\mathcal{F}_{n}$ except for the initial $0$. Also, $\mathcal{K}_0 = \ell_0 = 1$.
The *Fibonacci numbers* are defined by the recurrence $F_{n} = F_{n-1} + F_{n-2}$ with initial conditions $F_{0} = 0$ and $F_{1} = 1$. The *Lucas numbers* are defined by the recurrence $L_{n} = L_{n-1} + L_{n-2}$ with initial conditions $L_{0} = 2$ and $L_{1} = 1$. The (ordinary) generating functions for these sequences are $x / (1 - x - x^{2})$ and $(2-x) / (1 - x - x^{2})$, respectively. For a detailed treatment of generating functions for recurrence relations, we refer the reader to [@wilf-94].
Let $${\ensuremath{d_{m}(x)}} = 1 - x - x^{2} -\ldots - x^{m} = 1 - x \sum_{i = 0}^{m-1} x^{i}.$$
Let ${\ensuremath{G^{m}_{n}}}$ be the integer sequence defined by the $m$-th order recurrence ${\ensuremath{G^{m}_{n}}} = \sum_{i = 1}^{m} {\ensuremath{G^{m}_{n-i}}}$ with initial conditions ${\ensuremath{G^{m}_{0}}} = {\ensuremath{G^{m}_{1}}} = \ldots = {\ensuremath{G^{m}_{m-1}}} = 1$. This is a generalization of the Fibonacci numbers, and when $m = 2$ we recover ${\ensuremath{G^{2}_{n}}} = F_{n+1}$. It is straightforward to see that the sequence ${\ensuremath{G^{m}_{n}}}$ has the generating function $$\frac{1 - \sum_{i = 2}^{m-1} (i-1) x^{i}}{{\ensuremath{d_{m}(x)}}}.$$ There are many possible generalizations of the Fibonacci numbers, depending on how the initial conditions $F_{0} = 0$ and $F_{1} = 1$ (and, in this case, $F_{2} = 1$) are extended.
There are likewise different generalizations of the Lucas numbers. Let ${\ensuremath{H^{m}_{n}}}$ be the sequence defined by the $m$-th order recurrence ${\ensuremath{H^{m}_{n}}} = \sum_{i = 1}^{m} {\ensuremath{H^{m}_{n-i}}}$ with initial conditions ${\ensuremath{H^{m}_{0}}} = m$ and ${\ensuremath{H^{m}_{i}}} = 2^{i} - 1$ for $1 \leq i \leq m - 1$. So ${\ensuremath{H^{2}_{n}}} = L_{n}$, and it is straightforward to see that the generating function for the sequence ${\ensuremath{H^{m}_{n}}}$ is $$\frac{m - \sum_{i = 1}^{m-1}(m - i) x^{i}}{{\ensuremath{d_{m}(x)}}}.$$
In this article, we prove the following.
\[main\] For [$\mathcal{F}_{n}$]{} with $n > 0$ and $m \geq 0$, $$\operatorname{sk}({\ensuremath{\mathcal{K}_{m}}}) = -{\ensuremath{G^{m + 1}_{n - 1}}}
\mbox{ and } |{\ensuremath{\mathcal{K}_{m}}}| = {\ensuremath{H^{m+1}_{n}}} .$$
Fibonacci, Lucas, De Bruijn, and Lyndon
=======================================
We first prove Theorem \[main\] for the special case when $m = 1$. That the result holds when $m = 0$ follows directly from the definitions.
We define a Lyndon word $w$ to be a *primitive of order $i$* if $w = 0^{i} 1^{j}$ with $i + j = |w|$, $i, j \geq 1$. If a Lyndon word is not primitive, we say it is *composite*.
Let $w$ be a Lyndon word of length $d \mid n$ which occurs in $\ell_{1}$. Then $w$ may be uniquely parsed into primitives of order $1$ as $$w = 0 1^{j_{1}} 0 1^{j_{2}} \ldots 0 1^{j_{k}}
\mbox{, where $j_{l} \geq 1$ and $\sum_{l = 1}^{k} (1 + j_{l}) = d$.}$$ Let $\phi$ be a mapping from primitives of order 1 into the integers where $\phi(0 1^j) = 1 + j$ and let $\Phi(n)$ be the multiset obtained by applying $\phi$ to the $0 1^{j}$ subwords of $\ell_{1}$ from [$\mathcal{F}_{n}$]{}. For instance, in $\mathcal{F}_{6}$ we have $\ell_{1} = 01010111011011111$ and $\Phi(6) = \{2, 2, 3, 4, 6\}$.
Let $c(n,k)$ be the number of integers $k \geq 2$ in the multiset $\Phi(n)$. Since each $\ell_{1}$ from [$\mathcal{F}_{n}$]{} with $n > 1$ contains exactly one primitive of order 1 and length $n$, we have that $c(n, k) = 1$ when $n = k$. Also, $c(n, k) = 0$ for $n < k$. As we show below, the distribution for other $(n,k)$ is Fibonacci-like.
Recall that a *composition* of an integer $n$ into $k$ (positive) parts is an ordered sum of integers $$n = x_{1} + x_{2} + \ldots + x_{k} \mbox{ where $x_{i} \geq 1$.}$$ We denote such a composition of $n$ as an ordered $k$-tuple $x = (x_{1}, x_{2}, \ldots, x_{k})$.
Since primitives of order $1$ have length $k \geq 2$, in the proof below we consider compositions of $n - k$ with parts greater than $1$. There are $F_{n - k - 1}$ distinct $x = (x_{1}, \ldots, x_{j})$, with $\sum_{i = 1}^{j} x_{i} = n - k$ and $x_{i} \geq 2$, a fact easily obtained by induction on $n-k$. We show that there are an equal number of distinct $01^{k-1}$ primitives in the substring $\ell_{1}$ of [$\mathcal{F}_{n}$]{}.
\[fibcounts\] $$c(n, k) = F_{n-k-1} \mbox{ for } 2 \leq k \leq n -1\mbox{.}$$
Let $k \geq 2$ be fixed. For $n = k + 1$, there cannot be a primitive of order $1$ and length $k$ in $\ell_1$ of $\mathcal{F}_{k+1}$, so $c(k+1, k) = 0$.
Let $n \geq k + 2$. Consider the compositions of $n - k$ with parts greater than 1. For each such $x$, let $\omega(x)$ be the binary string $$0 1^{k-1} 0 1^{x_{1} - 1} \ldots 0 1^{x_{j} - 1}.$$
Suppose that $\omega(x)$ is aperiodic, and let $\lambda(x)$ be the Lyndon word which is the representative element for the equivalence class under rotation of $\omega(x)$. If $k$ does not occur in $x$, then $\lambda(x)$ contributes exactly one integer $k$ to the multiset $\Phi(n)$.
Otherwise, $\lambda(x)$ contributes $m + 1$ times to the count of $c(n,k)$, where there are $m \geq 1$ parts of $x$ which equal $k$. In this case, there are $m$ additional compositions $x^{j_{1}}, \ldots, x^{j_{m}}$ of $n - k$ such that $\omega(x^{j_{1}}), \ldots, \omega(x^{j_{m}})$ all belong to the equivalence class of $\omega(x)$ under rotation. Hence, there are $m + 1$ compositions of $n - k$ which are associated with the same Lyndon word of length $n$ from $\ell_{1}$.
Suppose now that $\omega(x)$ is periodic with period $p$. Let $\lambda(x)$ be the Lyndon word of length $p$ such that $(\lambda(x))^{n/p}$ is an element of the rotational equivalence class of $\omega(x)$. Then there are $q(n/p) - 1$ parts of $x$ which equal $k$ for some $q \geq 1$. Hence, $\lambda(x)$ contributes $q$ integers $k$ to the multiset $\Phi(n)$. Observe that if $q = 1$, then $x$ is the only composition of $n - k$ associated with $\lambda(x)$. Otherwise, there are $q - 1$ other instances of $01^{k-1}$ in $\lambda(x)$ and $q - 1$ distinct compositions of $n - k$ which are associated with $\lambda(x)$.
Recall that the skew of a binary string $w$ is the difference between the number of $0$’s in $w$ and the number of $1$’s, denoted here ${\ensuremath{\epsilon(w)}}$ and ${\ensuremath{\delta(w)}}$ respectively. Hence, $$\operatorname{sk}(w) = {\ensuremath{\epsilon(w)}} - {\ensuremath{\delta(w)}}.$$
For [$\mathcal{F}_{n}$]{} with $n \geq 2$, ${\ensuremath{\epsilon(\ell_{1})}} = F_{n-1}$ and ${\ensuremath{\delta(\ell_{1})}} = F_{n+1} - 1$.
The result follows from Theorem \[fibcounts\], and the identities $$\sum_{i = 1}^{n} F_{i} = F_{n+2} - 1$$ and $$\sum_{i = 1}^{n} i \cdot F_{n - i} = F_{n+3} - (n + 2).$$ Each primitive of order $1$ and length $k$ contributes a zero to ${\ensuremath{\epsilon(\ell_{1})}}$ and $k - 1$ ones to ${\ensuremath{\delta(\ell_{1})}}$. Hence, $$\begin{aligned}
{\ensuremath{\epsilon(\ell_{1})}} & = & \sum_{k = 2}^{n} c(n, k) \\
& = & 1 + \sum_{k = 2}^{n-1} F_{n-k-1} \\
& = & F_{n-1} \mbox{.} \end{aligned}$$ Likewise, we have that $$\begin{aligned}
{\ensuremath{\delta(\ell_{1})}} & = & \sum_{k = 2}^{n} (k - 1) \cdot c(n, k) \\
& = & (n-1) - \sum_{k = 2}^{n-1} F_{n-k-1} + \sum_{k = 2}^{n-1} k \cdot F_{n-k-1} \\
& = & (n-1) - (F_{n-1} - 1) - F_{n - 2} + \sum_{k = 1}^{n-1} k \cdot F_{n - 1 - k} \\
& = & n - F_{n} + F_{n + 2} - (n + 1) \\
& = & F_{n + 1} - 1. \end{aligned}$$
Because the Lucas and Fibonacci numbers are related as $L_{n} = F_{n-1} + F_{n+1}$ for $n \geq 1$, we have the following result.
For [$\mathcal{F}_{n}$]{} with $n \geq 2$, $\operatorname{sk}(\ell_{1}) = - F_{n} + 1$ and $|\ell_{1}| = L_{n} - 1$.
Since $\mathcal{K}_{1} = \ell_{1}1$ for [$\mathcal{F}_{n}$]{} with $n \geq 1$, we know that $\operatorname{sk}(\mathcal{K}_1) = -{\ensuremath{G^{2}_{n-1}}}$ and $|\mathcal{K}_1| = {\ensuremath{H^{2}_{n}}}$. Hence, Theorem \[main\] holds for $m = 0, 1$.
Generalizing to higher orders
=============================
We generalize compositions of an integer $n$ into parts greater than $1$ to accommodate Lyndon word primitives $0^{i} 1^{j}$ of order $i \geq 1$. We use the notation $x^{(y)} = (x^{(y-1)})'$ where $x^{(0)} = x$, so $x^{(1)} = x'$, $x^{(2)} = x''$, etc. We say that $$n = x_{1}^{(y_{1})} + x_{2}^{(y_{k})} + \ldots + x_{k}^{(y_{k})}$$ is an *$m$-colored composition of $n$ into $k$ parts greater than $1$* if
- $\sum_{i = 1}^{k} x_{i} = n$ with $x_{i} \geq 2$,
- and $0 \leq y_{i} \leq \min\{x_{i} - 2, m - 1\}$ for $1 \leq i \leq k$.
Note that $x_{i}^{(y_{i})} = x_{j}^{(y_{j})}$ if and only if $x_{i} = x_{j}$ and $y_{i} = y_{j}$. For instance, $5$, $5'$, $2 + 3$, $3 + 2$, $3' + 2$, and $2 + 3'$ are the six $2$-colored compositions of $5$.
For $m \geq 2$, let ${\ensuremath{P^{m}_{n}}}$ be the sequence defined by the $m$-th order recurrence ${\ensuremath{P^{m}_{n}}} = \sum_{i = 1}^{m} {\ensuremath{P^{m}_{n - i}}}$ with initial conditions ${\ensuremath{P^{m}_{0}}} = \ldots = {\ensuremath{P^{m}_{m-3}}} = 0$, ${\ensuremath{P^{m}_{m-2}}} = 1$, and ${\ensuremath{P^{m}_{m-1}}} = 0$.
The sequence ${\ensuremath{P^{m}_{n}}}$ is another generalization of the Fibonacci numbers. For $m = 2$, we have that ${\ensuremath{P^{2}_{n}}} = F_{n-1}$ when $n \geq 1$. In general, an inductive argument from the definition yields the following identities.
Let $m \geq 2$. For $m \leq n \leq 2m - 2$, ${\ensuremath{P^{m}_{n}}} = 2^{n-m}$. Also, ${\ensuremath{P^{m}_{2m-1}}} = 2^{m-1} - 1$.
Let $d(m, n)$ be the number of $m$-colored compositions of $n$ into parts greater than $1$ for $m, n \geq 1$.
$$d(m, n) = {\ensuremath{P^{m + 1}_{n + m - 1}}}.$$
The $m$-colored compositions of $n$ satisfy an $(m+1)$-th order recursion as follows. Let $$x = (x_{1}^{(y_{1})}, x_{2}^{(y_{2})}, \ldots, x_{k}^{(y_{k})})$$ for $\sum_{i = 1}^{k} x_{i} = n$ with integers $x_{i} \geq 2$ and $0 \leq y_{i} \leq \min\{x_{k} - 2, m - 1\}$.
Suppose $n \geq m + 2$. If $y_{k} < x_{k} - 2$, then $$(x_{1}^{(y_{1})}, x_{2}^{(y_{2})}, \ldots, (x_{k} - 1)^{(y_{k})})$$ is an $m$-colored composition of $n - 1$. Otherwise, $2 \leq x_{k} \leq m + 1$ and $$(x_{1}^{(y_{1})}, x_{2}^{(y_{2})}, \ldots, x_{k-1}^{(y_{k-1})})$$ is an $m$-colored composition of $n - x_{k}$. For $n = m + 2$, the recurrence has only $m$ terms since there is no $m$-colored composition of $1$ with parts $\geq 2$.
For initial conditions, we consider $m$-colored compositions of integers $n$ with $1 \leq n \leq m + 1$. We have that $d(m, 1) = 0 = {\ensuremath{P^{m+1}_{m}}}$ and $d(m, 2) = 1 = {\ensuremath{P^{m+1}_{m + 1}}}$ for all $m$. We claim that $d(m, n) = 2^{n-2}$ for $2 \leq n \leq m + 1$.
When $2 \leq n \leq m$, the only symbols that can occur in the $m$-colored composition of $n$ are $\{2, 3, 3', 4, 4', 4'', 5, \ldots, n^{(n-2)}\}$ and $d(m, n) = d(m - 1, n)$.
Consider $d(m, m + 1)$. There is exactly one $m$-colored composition of $m + 1$ which is not an $(m-1)$-colored composition of $m + 1$, namely $x = ((m+1)^{(m-1)})$. Hence $d(m, m + 1) = 1 + d(m-1, m+1)$. Inductively, then, $$d(m, m + 1) = 1 + \sum_{i = 1}^{m} d(m - 1, i) =
1 + 0 + \sum_{i = 2}^{m} 2^{i-2} = 2^{m - 1}.$$
It is again straightforward to see that the sequence ${\ensuremath{P^{m}_{n}}}$ has the generating function $$p_{m}(x) = \frac{x^{m-2}(1 - x)}{{\ensuremath{d_{m}(x)}}}$$ where $${\ensuremath{d_{m}(x)}} = 1 - x - x^{2} -\ldots - x^{m} = 1 - x \sum_{i = 0}^{m-1} x^{i}$$ as in the generating functions for ${\ensuremath{G^{m}_{n}}}$ and ${\ensuremath{H^{m}_{n}}}$ from Section \[intro\].
Let $Z = \{2, 3, 3', 4, 4', 4'', \ldots\}$ be the set of colored integers. Let $\psi$ be a mapping from binary strings $0^{i} 1^{j}$ for $i, j \geq 1$ to $Z$ where $\psi(0^{i} 1^{j}) = (i + j)^{(i-1)}$.
Recall that $\mathcal{L}_{m}$ is the concatenation of $\ell_{m} \ell_{m-1} \ldots \ell_{1}$ from [$\mathcal{F}_{n}$]{}, where $\ell_{i} = \varepsilon$ for $i > n - 1$. Let $\Psi(m, n)$ be the $m$-colored multiset obtained by applying $\psi$ to the primitives of order $1 \leq i \leq m$ from $\mathcal{L}_{m}$. Let $c(m, n, k)$ be the number of integers $k^{(0)} = k \geq 2$ in $\Psi(m, n)$.
\[cmnk\] For $m \geq 1, k \geq 2$, the count $c(m, n, k)$ is the coefficient of $x^{n}$ in $$x^{k - m + 1} p_{m+1}(x) = \frac{x^k (1 - x)}{{\ensuremath{d_{m+1}(x)}}}.$$
The argument is essentially the same as the proof of Theorem \[fibcounts\], except that we consider $m$-colored compositions $x$ of $n - k$ with parts greater than $1$ and give the result in terms of generating functions. The rotational symmetries of $\omega(x)$, where the definition is extended to higher order primitives, depend on which parts of $x$ are $k^{(0)} = k$.
By exchanging the colors of $k^{(0)}$ and some $k^{(i)}$ with $0 < i \leq \min\{k - 2, m - 1\}$ occurring in the $m$-colored compositions of $n - k$, we see that the number of occurrences of $k^{(i)}$ is also $c(m, n, k)$.
Consider [$\mathcal{F}_{n}$]{} and $\mathcal{L}_{m}$ with $n \geq 0$ and $m \geq 1$. Then ${\ensuremath{\epsilon(\mathcal{L}_{m})}}$ is the coefficient of $x^{n}$ in $$\frac{x^{2} \sum_{i = 0}^{m-1} (i+1) x^{i}}{{\ensuremath{d_{m+1}(x)}}}$$ and ${\ensuremath{\delta(\mathcal{L}_{m})}}$ is the coefficient of $x^{n}$ in $$\frac{x^{2} \sum_{i=0}^{m-1} x^{i}}{(1-x) {\ensuremath{d_{m+1}(x)}}} .$$
Let $1 \leq i \leq m$ be fixed. Each primitive of order $i$ contributes $i$ zeros to ${\ensuremath{\epsilon(\mathcal{L}_{m})}}$, and the number of colored integers $k^{(i-1)}$ in $\Psi(m,n)$ is the sum of $c(m, n, k)$ for $i + 1 \leq k \leq n$. According to Theorem \[cmnk\], this is the sum of the first $n - (i + 1) + 1$ terms of the sequence $a_{0}, a_{1}, a_{2}, \ldots$ whose generating function is $$\frac{(1-x)}{{\ensuremath{d_{m+1}(x)}}},$$ which is the coefficient of $x^{n - i - 1}$ in the series $$\frac{1}{{\ensuremath{d_{m+1}(x)}}}.$$ The result for ${\ensuremath{\epsilon(\mathcal{L}_{m})}}$ follows by a weighted summation over all $1 \leq i \leq m$, with the exponents adjusted appropriately.
Each primitive of order $i$ and length $k$ contributes $(k - i)$ ones to ${\ensuremath{\delta(\mathcal{L}_{m})}}$. To calculate the contribution for a given $i$, we again sum over the first $n - i$ terms of the sequence $a_{0}, a_{1}, a_{2}, \ldots$, except that now each term $a_{j}$ is weighted by $(n - i - j)$. This is the coefficient of $x^{n - i - 1}$ in the series $$\frac{1}{(1 - x) {\ensuremath{d_{m+1}(x)}}}.$$ Summing over the possible $i$’s yields ${\ensuremath{\delta(\mathcal{L}_{m})}}$.
Recall that ${\ensuremath{\mathcal{K}_{m}}}$ is the proper suffix of [$\mathcal{F}_{n}$]{} consisting of the Lyndon words of length $d \mid n$ containing at most $m$ consecutive $0$’s; $${\ensuremath{\mathcal{K}_{m}}}= \mathcal{L}_{m} 1.$$
Consider [$\mathcal{F}_{n}$]{} and [$\mathcal{K}_{m}$]{} with $n \geq 1$ and $m \geq 0$. Then $\operatorname{sk}({\ensuremath{\mathcal{K}_{m}}})$ is the coefficient of $x^{n}$ in $$\frac{- x + \sum_{i = 3}^{m+1} (i - 2)x^{i}}{{\ensuremath{d_{m+1}(x)}}}$$ and $|{\ensuremath{\mathcal{K}_{m}}}|$ is the coefficient of $x^{n}$ in $$\frac{\sum_{i = 1}^{m+1} i x^{i}}{{\ensuremath{d_{m+1}(x)}}}.$$
We have $\operatorname{sk}({\ensuremath{\mathcal{K}_{m}}}) = {\ensuremath{\epsilon({\ensuremath{\mathcal{K}_{m}}})}} - {\ensuremath{\delta({\ensuremath{\mathcal{K}_{m}}})}}$ where ${\ensuremath{\epsilon({\ensuremath{\mathcal{K}_{m}}})}} = {\ensuremath{\epsilon(\mathcal{L}_{m})}}$ and ${\ensuremath{\delta({\ensuremath{\mathcal{K}_{m}}})}} = {\ensuremath{\delta(\mathcal{L}_{m})}} + 1$. Adding $\frac{x}{1-x}$ to the generating function for ${\ensuremath{\delta(\mathcal{L}_{m})}}$ yields $$\frac{x (1 - x^{m+1})}{(1-x) {\ensuremath{d_{m+1}(x)}}}.$$ Taking the difference with ${\ensuremath{\epsilon({\ensuremath{\mathcal{K}_{m}}})}}$ gives $$\frac{-x + \sum_{i = 2}^{m+1} x^{i} - (m-1)x^{m+2}}{(1-x){\ensuremath{d_{m+1}(x)}}}$$ which simplifies to the desired result. Similarly, adding the two series yields $$\frac{x + \sum_{i = 2}^{m+1} x^{i} - (m+1)x^{m+2}}{(1-x){\ensuremath{d_{m+1}(x)}}}$$ which again simplifies.
Offsetting the sequence ${\ensuremath{G^{m}_{n}}}$ by an initial zero, and recalculating the generating function with the initial ${\ensuremath{H^{m}_{0}}} = m$ replaced by a zero, we have the result stated in Theorem \[main\].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The photon energy spectrum in inclusive weak radiative $\bar B\to X_s\gamma$ decay is computed to order $\alpha_s^2 \beta_0$. This result is used to extract a value for the HQET parameter $\bar\Lambda$ from the average $\langle
1-2E_\gamma / m_B \rangle$, and a value of the parameter $\lambda_1$ from $\langle (1-2E_\gamma / m_B)^2 \rangle$. An accurate measurement of $\langle
1-2E_\gamma / m_B \rangle$ can determine the size of the nonperturbative contributions to the $\Upsilon(1S)$ mass which cannot be absorbed into the $b$ quark pole mass.
address: |
$^a$Theory Group, Fermilab, P.O. Box 500, Batavia, IL 60510\
$^b$Department of Physics, University of Toronto,\
60 St. George Street, Toronto, Ontario, Canada M5S 1A7\
$^c$Department of Physics, University of California at San Diego,\
9500 Gilman Drive, La Jolla, CA 92093–0319\
$^d$California Institute of Technology, Pasadena, CA 91125
author:
- 'Zoltan Ligeti,${}^{a}$ Michael Luke,${}^{b}$ Aneesh V. Manohar,${}^c$ and Mark B. Wise${\,}^d$'
title: 'The $\bar B\to X_s \gamma$ Photon Spectrum'
---
Comparison of the measured weak radiative $\bar B\to X_s\gamma$ decay rate with theory is an important test of the standard model. In contrast to the decay rate itself, the shape of the photon spectrum is not expected to be sensitive to new physics, but it can nevertheless provide important information. First of all, studying the photon spectrum is important for understanding how precisely the total rate can be predicted in the presence of an experimental cut on the photon energy [@CLEO], which is important for a model independent interpretation of the resulting decay rate. Secondly, moments of the photon spectrum may be used to measure the heavy quark effective theory (HQET) parameters which determine the quark pole mass and kinetic energy [@FLS; @AZ], much like the shape of the lepton energy [@gremmetal] or hadronic invariant mass [@FLSmass] spectrum in semileptonic $\bar B\to
X_c\,\ell\,\bar\nu$ decay. The main purpose of this paper is to present the order $\alpha_s^2\beta_0$ piece of the two-loop correction to the photon spectrum, and to study its implications. A calculation to this order is required for a meaningful comparison of the HQET parameters extracted from $\bar B\to X_s\gamma$ with those from other processes.
To leading order in small weak mixing angles the effective Hamiltonian is $$H_{\rm eff} = - \frac{4G_F}{\sqrt2}\, V_{tb} V_{ts}^*
\sum_{i=1}^8 C_i(\mu)\, O_i \,,$$ where $G_F$ is the Fermi constant, $V_{ij}$ are elements of the Cabibbo–Kobayashi–Maskawa matrix, $C_i (\mu)$ are Wilson coefficients evaluated at a subtraction point $\mu$, and $O_i$ are the dimension six operators $$\label{ops}
%%\renewcommand{\arraystretch}{1.5}
\begin{array}{rclrcl}
O_1 &=& (\bar c_{L\beta} \gamma^\mu b_{L\beta}) (\bar s_{L\alpha} \gamma_\mu
c_{L\beta}) \,, &
O_2 &=& (\bar c_{L\alpha} \gamma^\mu b_{L\alpha}) (\bar s_{L\beta} \gamma_\mu
c_{L\beta}) \,, \\
O_3 &=& \displaystyle (\bar s_{L\alpha} \gamma^\mu b_{L\alpha})
\sum_q (\bar q_{L\beta} \gamma_\mu q_{L\beta}) \,,\qquad &
O_4 &=& \displaystyle (\bar s_{L\alpha} \gamma^\mu b_{L\beta})
\sum_q (\bar q_{L\beta} \gamma_\mu q_{L\alpha}) \,, \\
O_5 &=& \displaystyle (\bar s_{L\alpha} \gamma^\mu b_{L\alpha})
\sum_q (\bar q_{R\beta} \gamma_\mu q_{R\beta}) \,, &
O_6 &=& \displaystyle (\bar s_{L\alpha} \gamma^\mu b_{L\beta})
\sum_q (\bar q_{R\beta} \gamma_\mu q_{R\alpha}) \,, \\
O_7 &=& \displaystyle {e\over 16\pi^2}\,
m_b \bar s_{L\alpha} \sigma^{\mu\nu} b_{R\alpha} F_{\mu\nu} \,, &
O_8 &=& \displaystyle {g\over16\pi^2}\,
m_b \bar s_{L\alpha} \sigma^{\mu\nu}
T_{\alpha\beta}^a b_{R\beta} G_{\mu\nu}^a \,.
\end{array}$$ In Eq. (\[ops\]), $e$ is the electromagnetic coupling, $g$ is the strong coupling, $m_b$ is the $b$ quark mass, $F_{\mu\nu}$ is the electromagnetic field strength tensor, $G_{\mu\nu}^a$ is the strong interaction field strength tensor, and $T^a$ is a color $SU(3)$ generator. The sums over $q$ include $q =
u,d,s,c,b$ and the subscripts $L,R$ denote left and right handed fields. The Wilson coefficients have been calculated to next-to-leading order (NLO) [@Misiak; @match; @fourquark]. Using $\alpha_s(m_Z)=0.12$, and the convention that the covariant derivative is $D_\mu = \partial_\mu +
igA_\mu^aT^a + ie\,QA_\mu$ (where $Q$ is the fermion’s electric charge), the values we need are $C_2(m_b)=1.13$, $C_7(m_b)=-0.306$, $C_8(m_b)=-0.168$ [@Misiak].
For the photon energy, $E_\gamma$, not too close to its maximal value, the photon spectrum $d\Gamma/dE_\gamma$ for weak radiative $B$ decay has a perturbative expansion in the strong interaction fine structure constant $\alpha_s$. It is known at order $\alpha_s$ and the main purpose of this letter is to present the order $\alpha_s^2 \beta_0$ (so-called BLM [@BLM]) contribution. It is well known that the part of the order $\alpha_s^2$ piece proportional to the one-loop beta function, $\beta_0 = 11-2n_f/3$ usually provides a reliable estimate of the full order $\alpha_s^2$ piece. This part of the order $\alpha_s^2$ contribution is straightforward to compute using the method of Smith and Voloshin [@SmVo].
Using the dimensionless variable[^1], $x_b = 2E_\gamma/m_b$, the photon energy spectrum in $\bar
B\to X_s \gamma$ takes the form $$\label{spectrum}
\frac1{\Gamma_0}\, \frac{d\Gamma}{dx_b} \bigg|_{x_{b} < 1}
= A_0 (x_b) + \frac{\alpha_s (m_b)}{\pi} A_1 (x_b)
+ \bigg( \frac{\alpha_s (m_b)}{\pi} \bigg)^2 \beta_0\, A_2(x_b) + \ldots \,,$$ where $$\Gamma_0 = {G_F^2\,|V_{tb}V_{ts}^*|^2\,\alpha_{\rm em}\,C_7^2\over32\pi^4}\,
m_b^5 \,,$$ is the contribution of the tree level matrix element of $O_7$ to the $B\to X_s\gamma$ decay rate, and $$\label{Apdef}
A_p(x_b) = \sum_{i\leq j} a_p^{ij} (x_b)
\Bigg[ \frac{C_i (m_b) C_j (m_b)}{C_7 (m_b)^2} \Bigg] \,.$$ The sums over $i,j$ in Eq. (\[Apdef\]) give the contributions of the various operators in Eq. (\[ops\]) to the photon energy spectrum.
It is important to note that since the coefficients in $H_{\rm eff}$ are known only to NLO accuracy, the BLM calculation of the $O_1-O_8$ contribution to the photon spectrum is only meaningful away from the endpoint. At the endpoint, order $\alpha_s^2$ contributions to the matrix elements are the same order as the unknown NNLO running \[where $\alpha_s\ln(m_W/m_b)$ is counted as ${\cal
O}(1)$\]. Neglecting the small contribution to $A_0$ from $O_1-O_6$ discussed in the next paragraph, at least one gluon must be in the final state to populate the spectrum for $x_b<1$, so it is consistent to combine the $\alpha_s^2$ matrix elements with the NLO Wilson coefficients. (Strictly speaking, we should for consistency only use the $\beta_0$ part of the NLO running of the operators with the BLM calculation, but for simplicity we will use the full NLO result. The difference between these two approaches is small.) Thus powers of $\alpha_s$ in Eq. (\[spectrum\]) and elsewhere reflect the perturbation expansion of the matrix elements only, and not of the Wilson coefficients.
At zeroth order in the strong coupling, the spectrum for $x_b<1$ arises from matrix elements of the four-quark operators $O_1-O_6$ in Eq. (\[ops\]). Of these $O_1$ and $O_2$ include two charm quarks in the final state, and therefore they contribute to the photon spectrum only for lower values of $x_b$ than what we consider in this paper. These contributions are divergent in perturbation theory, and the divergence can be absorbed into the definition of the quark to photon fragmentation function, $D^{q\to\gamma}(x)$, which depends on an infrared scale $\Lambda$. $D^{q\to\gamma}(x)$ is calculable in the leading logarithmic approximation [@EW; @decayfn]. There is some data on $D^{q\to \gamma}(x)$, however, the experimental errors are still quite large [@ALEPH]. This fragmentation contribution to the coefficients $a_0^{ij} (x)$ vanishes as $x_b\to 1$, and it is small in the region of large $x_b$, $0.65<x_b$, which we consider in this paper.
A very important $B$ decay background to the $\bar B\to X_s\gamma$ photon spectrum is from nonleptonic $b\to c\bar ud$ and $b\to u\bar ud$ decays, where a massless quark in the final state radiates a photon. Such backgrounds due to the operators $(\bar c_L \gamma^\mu b_L) (\bar d_L \gamma_\mu u_L)$ and $(\bar
u_L \gamma^\mu b_L) (\bar d_L \gamma_\mu u_L)$ are shown in Fig. 1 (using $|V_{ub}/V_{cb}|=0.1$). We used the Duke–Owens parameterization of the fragmentation function [@DO], setting $\Lambda=1.3\,$GeV and $Q^2=m_b^2$. (This value of $\Lambda$ is motivated by a fit to the ALEPH data [@ALEPH].) The uncertainty of this result is sizable, since the $\Lambda$-dependence is large and $m_b$ may not be large enough to justify keeping only the leading logarithms. Close to maximal $x_b$ the resummed fragmentation function may predict too large a suppression of the photon spectrum, since the lightest exclusive final states dominate there. The background from $b\to c \bar u d$ ($b\to u \bar u d$) is more than 50% of the 77 contribution to $(1/\Gamma_0)d\Gamma/dx_b$ below $x_b\sim0.75$ ($x_b\sim0.65$).[^2] Therefore, we will concentrate on the region $x_b>0.65$; to measure the $B\to X_s\gamma$ photon spectrum at lower values of $x_b$ would not only require excluding final states with charm with very good efficiency, but also demanding a strange quark in the final state. Note that for $B\to X_d\gamma$, the fragmentation contribution from $b\to u\bar ud$ is larger than the short distance piece unless $x_b$ is very close to 1.
Neglecting the strange quark mass, $a_1^{88}$ is also divergent in perturbation theory. This divergence can also be absorbed into the definition of fragmentation functions. In the leading logarithmic approximation [@KLP] $$a_1^{88} (x) = \left(\frac{4\pi}{3\alpha_{\rm em}}\right)
[ D^{s\to \gamma}(x) + D^{g\to \gamma}(x) ] \,,$$ where $D^{s\to\gamma}(x)$ and $D^{g\to\gamma}(x)$ are the strange quark to photon and gluon to photon fragmentation functions, which have large uncertainties. In the region $x_b>0.65$, the $a_1^{88}$ contribution to the photon spectrum $(1/\Gamma_0)d\Gamma/dx_b$ is less than 0.01. Given the uncertainty in $a_1^{88}$, and its small magnitude, it does not appear useful to calculate $a_2^{88}$.
Experimentally, because of backgrounds, only $\bar B\to X_s\gamma$ photons with large energies can be detected. The present experimental cut is $E_\gamma >
2.1\,$GeV at CLEO [@CLEO], which corresponds to $x_b>0.875$ with $m_b=4.8\,$GeV. In the large $x_b$ region the most important contribution to the sum in Eq. (\[Apdef\]) come from the 77 term, with moderate corrections from the 22, 78, and 27 terms. The other contributions (88, 28, and the ones involving $O_1$ and $O_3 - O_6$) are very small, and will be neglected in this paper.
Simple analytic expressions for $a_1^{77}$ and $a_1^{78}$ are available, $$\begin{aligned}
a_1^{77} (x) &=& \frac{(2x^2 - 3x - 6) x + 2(x^2 - 3) \ln(1-x)}{3(1-x)} \,,
\label{a177} \\
a_1^{78} (x) &=& \frac{8}{9} \bigg[ \frac{4 + x^2}{4} + \frac{1-x}{x}
\ln(1-x) \bigg] \,. \label{a178}\end{aligned}$$ Neglecting the small $A_0$ term in Eq. (\[spectrum\]), we can calculate the shape of the photon spectrum away from $x=1$ to order $\alpha_s^2\beta_0$ accuracy knowing the effective Hamiltonian to order $\alpha_s$ (NLO) only. At order $\alpha_s^2\beta_0$, we find that $a_2^{77}$ and $a_2^{78}$ are given by $$\begin{aligned}
a_2^{77}(x) &=& \frac{1}{18} \Bigg[ \frac{38x^3 - 93x^2 + 6x - 36}{4(1-x)}
- \frac{6x^4 - 31 x^3 + 24x^2 - 30x + 18}{2x (1-x)} \ln (1-x) \nonumber\\*
&& + 3 (3 - x^2) \frac{3\ln^2 (1 - x) + 2L_2 (x)}{2 (1 - x)}\Bigg] \,,
\label{a277} \\
a_2^{78}(x) &=& \frac{1}{9} \Bigg[ \frac{19x^2 - 24x + 88}{12}
- \frac{3x^3 - 12x^2 + 56x - 32}{6x} \ln(1-x) \nonumber\\*
&& - (1 - x) \frac{3\ln^2 (1 - x) + 2L_2 (x)}{x} \Bigg] \,, \label{a278}\end{aligned}$$ where $L_2(z) = -\int_0^z dt \ln(1-t)/t$ is the dilogarithm. The strange quark mass is neglected throughout this paper; it only enters the final results quadratically, as .
The functions of $a_1^{22}$ and $a_1^{27}$ are known in the literature [@AG; @Pott], and we agree analytically with those results. The order $\alpha_s^2\beta_0$ contributions, $a_2^{22}$ and $a_2^{27}$, are computed numerically. We find it most useful to present simple approximations to these functions $$\begin{aligned}
\label{ai22}
a_1^{22}(x) &\simeq& -0.0842 + 0.3333x - 0.2005x^2 + 0.0227x^3 \nonumber\\*
&&\phantom{} + \bigg({m_c\over m_b} - {1.4\over4.8}\bigg)
(-0.454 + 0.061x) \,, \nonumber\\*
a_2^{22}(x) &\simeq& -0.1272 + 0.3957x - 0.3227x^2 + 0.0952x^3
- 0.0180\ln(1-x) \nonumber\\*
&&\phantom{} + \bigg({m_c\over m_b} - {1.4\over4.8}\bigg)
[-0.155 - 0.106x + 0.106\ln(1-x)] \,, \end{aligned}$$ and $$\begin{aligned}
\label{ai27}
a_1^{27}(x) &\simeq& -0.1064 + 0.4950x - 0.4361x^2 + 0.0373x^3 \nonumber\\*
&&\phantom{} + \bigg({m_c\over m_b} - {1.4\over4.8}\bigg)
(-1.207 + 2.901x) \,, \nonumber\\*
a_2^{27}(x) &\simeq& -0.0156 + 0.0463x + 0.3467x^2 - 0.3045x^3
+ 0.0027\ln(1-x) \nonumber\\*
&&\phantom{} + \bigg({m_c\over m_b} - {1.4\over4.8}\bigg)
[-1.523 + 2.538x - 0.448\ln(1-x)] \,.\end{aligned}$$ These approximations are accurate to within 1% in the region $x_b>0.6$ for $m_c/m_b = 1.4/4.8$. The 27 contribution is very sensitive to $m_c/m_b$. Changing $m_c/m_b$ from $1.4/4.8$ to $1.2/4.6$ or $1.6/5.0$ modifies $a_1^{27}$ and $a_2^{27}$ dramatically. The 22 contribution only changes in the previously mentioned range of $m_c/m_b$ by $\pm(20-25)\%$. The 22 contribution is also accurate to within 1% when $m_c/m_b$ changes by $\pm0.03$. However, the 27 contribution is only accurate at the 20% level when $m_c/m_b$ changes in this range. Note that the perturbation series in $\alpha_s$ is particularly badly behaved for the 27 contribution. Roughly 2/3 of the 22 contribution is from absorptive parts corresponding to real intermediate states.
The coefficients $a_2^{ij}$ are determined by calculating the order $\alpha_s^2
n_f$ piece and making the identification, $-2n_f/3 \to \beta_0$. There is a subtlety in applying this method to weak radiative $B$ decay. There is a contribution of order $\alpha_s^0 n_f$ from the tree level $b \to s\gamma q
\bar q$ matrix elements of $O_3 - O_6$, coming from Feynman diagrams where the photon couples to the bottom or strange quarks. It is not associated with a term of order $\alpha_s^0 \beta_0$. To avoid adding an analogous spurious order $\alpha_s^2 \beta_0$ contribution to $a_2^{27}$ and $a_2^{22}$, only diagrams where the photon couples to the charm quark were included in the calculation of the matrix element of $O_2$.
Part of the $\bar B\to X_s\gamma$ matrix element of $O_2$ is not adequately calculated in perturbation theory. It corresponds to the process $\bar B\to
J/\psi X_s$ followed by the decay $J/\psi\to \gamma + {\rm (light\ hadrons)}$. There will be large corrections to the part of the charm quark loop where the $c\bar c$ are almost on-shell and have the same velocity. In this region there are large “Coulombic QCD corrections” that produce the $J/\psi$ state. However, cutting this small part of the $c\bar c$ phase space out of our calculation of the matrix element of $O_2$ has a negligible effect. Hence, at the order of perturbation theory to which we are working, calculating the $c\bar c$ loop while removing $J/\psi$’s from the data would be a consistent approximation.
The sum of the 77, 22, 78, and 27 contributions is plotted in Fig. 2 in the region $0.65<x_b<0.9$ (using $\alpha_s(m_b)=0.22$ and $\beta_0=25/3$). For very large $x$, other effects that we have not calculated become important. There are both nonperturbative and perturbative terms that are singular as $x\to 1$. They sum into a shape function that modifies the spectrum in this region [@shapefn]. Unfortunately, at the present time, it is not possible to make a model independent estimate of these effects. Therefore, we do not plot the perturbation theory predictions for $x_b>0.9$. In the plotted region, the 22, 78, and 27 terms make a moderate correction to the dominant 77 contribution to $(1/\Gamma_0)d\Gamma/dx$, which is shown in Fig. 2 with the thin curves.
The $b$ quark mass can be eliminated in favor of the $B$ meson mass by a change of variables to $$x_B = 2E_\gamma / m_B \,.$$ Using $m_b = m_B - \bar\Lambda + (\lambda_1+3\lambda_2)/(2m_b) + \ldots$, the photon spectrum becomes $$\frac{d\Gamma}{dx_B} =
\bigg(1 + {\bar\Lambda\over m_B} + \ldots \bigg)\,
\frac{d\Gamma}{dx_b} \bigg|_{x_b = x_B(1 + \bar\Lambda/m_B + \ldots)} \,.$$ For $x_B$ within a region of order $\Lambda_{\rm QCD}/m_B$ of unity (its maximal value) nonperturbative effects are very important. However, for integrals of $x_B$ over a large enough range these nonperturbative effects are small.
An important integral of this type is $$\label{moment1}
\overline{(1 - x_B)} \Big|_{x_B > 1 - \delta} =
{\displaystyle \int_{1-\delta}^1 dx_B\, (1-x_B)\, \frac{d\Gamma}{dx_B} \over
\displaystyle \int_{1-\delta}^1 dx_B\, \frac{d\Gamma}{dx_B} } \,.$$ The parameter $\delta=1-2E_\gamma^{\rm min}/m_B$ has to satisfy $\delta > \Lambda_{\rm QCD}/m_B$; otherwise nonperturbative effects are not under control. It is straightforward to show that $$\label{beauty}
\overline{(1 - x_B)} \Big|_{x_B > 1-\delta} =
{\bar\Lambda\over m_B} + \bigg(1-{\bar\Lambda\over m_B}\bigg)\,
\langle 1 - x_b \rangle \Big|_{x_b > 1-\delta}
- {\bar\Lambda\over m_B}\, \delta(1-\delta)\, \frac1{\Gamma_0}\,
\frac{d\Gamma}{dx_b} \bigg|_{x_b = 1-\delta} + \ldots \,,$$ where $$\label{momentq}
\langle 1 - x_b \rangle \Big|_{x_b > 1 - \delta} = \int_{1-\delta}^1 dx_b\,
(1-x_b)\, \frac1{\Gamma_0}\, \frac{d\Gamma}{dx_b} \,.$$ Note that all terms but the first one in Eq. (\[beauty\]) have perturbative expansions which begin at order $\alpha_s$. The ellipses denote contributions of order $(\Lambda_{\rm QCD}/m_B)^3$, $\alpha_s(\Lambda_{\rm QCD}/m_B)^2$, and $\alpha_s^2$ terms not enhanced by $\beta_0$, but it does not contain contributions of order $(\Lambda_{\rm QCD}/m_B)^2$ or additional terms[^3] of order $\alpha_s(\Lambda_{\rm QCD}/m_B)$. Terms in the operator product expansion proportional to $\lambda_{1,2}/m_b^2$ enter precisely in the form so that they are absorbed in $m_B$ in Eq. (\[beauty\]) [@AZ]. There are also nonperturbative corrections suppressed by $(\Lambda_{\rm QCD}/m_c)^2$ instead of $(\Lambda_{\rm
QCD}/m_b)^2$ [@1/mc2]. These do not contribute to Eq. (\[beauty\]).
Using our results, $\langle 1-x_b \rangle |_{x_b>1-\delta}$ in Eq. (\[momentq\]) is known to order $\alpha_s^2\beta_0$. Writing $$\langle 1 - x_b \rangle \Big|_{x_b > 1-\delta} =
B_0(\delta) + \frac{\alpha_s(m_b)}{\pi}\, B_1(\delta)
+ \bigg(\frac{\alpha_s(m_b)}{\pi}\bigg)^2 \beta_0\, B_2(\delta) + \ldots \,,$$ $B_p$ have decompositions analogous to Eq. (\[Apdef\]), $$B_p(\delta) = \sum_{i\leq j} b_p^{ij} (\delta)
\Bigg[ \frac{C_i(m_b) C_j(m_b)}{C_7^2(m_b)} \Bigg] \,.$$ Neglecting $B_0(\delta)$, Eqs. (\[a177\]) and (\[a277\]) yield for the dominant 77 contribution $$\begin{aligned}
b_1^{77} (\delta) &=& \frac{\delta}{54} \left[- 9\delta^3 + 14\delta^2
+ 72\delta - 54 + 12 (\delta^2 - 3\delta - 6) \ln\delta \right] , \\
b_2^{77} (\delta) &=& \frac{1}{2592} \bigg[ -369\delta^4 + 116\delta^3 +
1800\delta^2 - 3852\delta\nonumber \\
&+& 408\pi^2 + 12\delta (9\delta^3 + 34\delta^2 - 102\delta + 66) \ln\delta
\nonumber \\
&-& 216\delta (\delta^2 -3\delta -6) \ln^2 \delta - 144 (\delta^3 - 3\delta^2
- 6\delta + 17) L_2 (1-\delta)\bigg] \,. \label{b277}\end{aligned}$$ Our prediction for $\langle 1-x_b \rangle |_{x_b>1-\delta}$ is shown in Fig. 3 as a function of $\delta$, both at order $\alpha_s$ and $\alpha_s^2\beta_0$. The bad behavior of the perturbation expansion would improve somewhat by evaluating the strong coupling at a smaller scale than $m_b$, such as $m_b\sqrt\delta$, the maximal available invariant mass of the hadronic final state. This bad behavior may also be related to the renormalon ambiguity [@renormalon] in $\bar\Lambda$.
A determination of $\bar\Lambda$ is straightforward using Eq. (\[beauty\]). The left hand side is directly measurable, while $\langle 1-x_b \rangle
|_{x_b>1-\delta}$ and $(1/\Gamma_0)d\Gamma/dx_b|_{x_b>1-\delta}$ in the second and third terms on the right hand side can be read off from Figs. 3 and 2, respectively. Using the CLEO data in the region $E_\gamma >
2.1\,$GeV [@CLEO], we obtain the central values $\bar\Lambda_{\alpha_s^2\beta_0} \simeq 270\,$MeV and $\bar\Lambda_{\alpha_s}
\simeq 390\,$MeV. We have indicated the order kept in the perturbation expansion to determine $\bar\Lambda$, since a value of $\bar\Lambda$ extracted from data can only be used consistently in predictions valid to the same order in $\alpha_s$. These values are consistent with the ones obtained from a fit to the $\bar B\to X_c\,\ell\,\bar\nu$ lepton spectrum [@gremmetal], and from the CLEO fit [@CLEO2] to the $\bar B\to X_c\,\ell\,\bar\nu$ hadron mass distribution [@FLSmass].
At the present time this extraction of $\bar\Lambda$ has large uncertainties. The potentially most serious one is from both nonperturbative and perturbative terms that are singular as $x\to 1$ and sum into a shape function that modifies the spectrum near the endpoint. A model independent determination of these effects is not available at the present time, however, it may be possible to address this issue using lattice QCD [@latticeshape]. For sufficiently large $\delta$ these effects are not important. It has been estimated that they may be significant even if the cut on the photon energy is lowered to around $E_\gamma = 2\,$GeV [@KaNe; @Bauer], but this is based on phenomenological models. We have implicitly neglected these effects throughout our analysis. The validity of this can be tested experimentally by checking whether the value of $\bar\Lambda$ extracted from Eq. (\[beauty\]) is independent of $\delta$ in some range. This would also improve our confidence that the total decay rate in the region $x_B>1-\delta$ can be predicted in perturbative QCD without model dependence.
The value of $\bar\Lambda$ at order $\alpha_s$ has a sizable scale dependence: lowering the scale such that $\alpha_s$ changes from 0.22 to 0.3 reduces the value of $\bar\Lambda_{\alpha_s}$ by about $40\,$MeV. At order $\alpha_s^2\beta_0$ this scale dependence is much smaller. Uncertainties due to the unknown order $(\Lambda_{\rm QCD}/m_B)^3$ terms in the OPE [@Bauer] are largely uncorrelated to those in the analyses of the lepton energy or hadron mass spectra in $\bar B\to X_c\,\ell\,\bar\nu$ [@GK]. The effect of the boost from the $B$ rest frame into the $\Upsilon(4S)$ is small for $\overline{(1-x_B)} |_{x_B>1-\delta}$ [@KaNe].
The upsilon expansion [@upsexp] yields parameter free predictions for $\overline{(1 - x_B)} |_{x_B > 1-\delta}$ in terms of the $\Upsilon(1S)$ meson mass. The analog of Eq. (\[beauty\]) is $$\label{upsbeauty}
\overline{(1 - x_B)} \Big|_{x_B > 1-\delta} = 1 -
{m_\Upsilon\over 2m_B} \left[ 1 + 0.011\epsilon + 0.019(\epsilon^2)_{\rm BLM}
- \langle 1-x_b \rangle \Big|_{x_b > (2m_B/m_\Upsilon)(1-\delta)} \right] ,$$ where $\epsilon\equiv1$ denotes the order in the upsilon expansion. For $E_\gamma>2.1\,$GeV this relation gives 0.111, whereas the central value from the CLEO data is around 0.093.[^4] In Fig. 4 we plot the prediction for $\overline{(1 - x_B)} |_{x_B > 1-\delta}$ as a function of $\delta$, both at order $\epsilon$ and $(\epsilon^2)_{\rm
BLM}$. The perturbation expansion is much better behaved than the one shown in Fig. 3. The most important uncertainty in this approach is the size of nonperturbative contributions to the $\Upsilon(1S)$ mass other than those which can be absorbed into the $b$ quark mass. These have been neglected in Eq. (\[upsbeauty\]). If the nonperturbative contribution to the $\Upsilon(1S)$ mass, $\Delta_\Upsilon$, were known, it could be included by replacing $m_\Upsilon$ by $m_\Upsilon-\Delta_\Upsilon$. For example, $\Delta_\Upsilon = +300\,$MeV increases $\overline{(1 - x_B)}$ by 21%, so measuring $\overline{(1 - x_B)}$ with such accuracy will have important implications for the physics of quarkonia as well as for $B$ physics.
The variance of the photon energy distribution can be used to determine $\lambda_1$ [@AZ; @Bauer]. The analog of Eq. (\[beauty\]) in this case is $$\begin{aligned}
\label{beauty2}
\overline{(1 - x_B)^2} \Big|_{x_B > 1-\delta} -
\left[ \overline{(1 - x_B)} \Big|_{x_B > 1-\delta} \right]^2 =
-{\lambda_1\over 3m_B^2} &+& {\beta^2\over3} +
\bigg(1-{2\bar\Lambda\over m_B}\bigg)\,
\langle (1 - x_b)^2 \rangle \Big|_{x_b > 1-\delta} \nonumber\\*
&-& {\bar\Lambda\over m_B}\, \delta^2(1-\delta)\, \frac1{\Gamma_0}\,
\frac{d\Gamma}{dx_b} \bigg|_{x_b = 1-\delta} + \ldots \,,\end{aligned}$$ where $\beta\simeq0.064$ is the magnitude of the velocity of the $B$ meson in the $\Upsilon(4S)$ rest frame, and only the leading $\beta$-dependence has been kept. The ellipses denote terms of order $(\Lambda_{\rm QCD}/m_B)^3$, $\alpha_s(\Lambda_{\rm QCD}/m_B)^2$, and $\alpha_s^2$ terms not enhanced by $\beta_0$. Our prediction for $\langle (1-x_b)^2 \rangle |_{x_b>1-\delta}$ is shown in Fig. 5. Note that unlike the case of $\overline{(1 - x_B)} |_{x_B >
1-\delta}$, the effect of the boost is very important in Eq. (\[beauty2\]). Using the CLEO data in the region $E_\gamma > 2.1\,$GeV, we obtain the central value $\lambda_1 \simeq -0.1\,{\rm GeV}^2$, with large experimental errors. The uncertainty in this value of $\lambda_1$ due to $\bar\Lambda$ is small. Nonperturbative effects from the cut on $E_\gamma$ [@Bauer], and the unknown higher order contributions to Eq. (\[beauty2\]) are expected to have a larger impact on the determination of $\lambda_1$ than the corresponding effects have on the determination of $\bar\Lambda$ from Eq. (\[beauty\]).
In summary, we calculated order $\alpha_s^2\beta_0$ corrections to the shape of the photon energy spectrum in weak radiative $\bar B\to X_s\gamma$ decay. The dominant 77 contribution is given by simple analytic formulae in Eqs. (\[a177\]) and (\[a277\]). The other terms relevant in the region $x_b>0.65$ are the 22 and 27 contributions given in Eqs. (\[ai22\]) and (\[ai27\]), and the 78 term given in Eqs. (\[a178\]) and (\[a278\]). The HQET parameter $\bar\Lambda$ can be extracted from the average $\langle 1 -
2E_\gamma/m_B \rangle$ using Eq. (\[beauty\]), and it can also be used to test whether the nonperturbative contribution to the Upsilon mass is small. The CLEO data in the region $E_\gamma > 2.1\,$GeV implies the central values $\bar\Lambda_{\alpha_s} \simeq 390\,$MeV and $\bar\Lambda_{\alpha_s^2\beta_0}
\simeq 270\,$MeV at order $\alpha_s$ and $\alpha_s^2\beta_0$, respectively. Possible contributions to the total decay rate from physics beyond the standard model are unlikely to affect this determination of $\bar\Lambda$. In the future, checking the $\delta$-independence of the extracted value of $\bar\Lambda$, and comparing the experimental and theoretical shapes of the photon spectrum for $x_b<0.9$ can provide a check that nonperturbative effects and backgrounds are under control. This would also improve our confidence that the total decay rate in the region $x_B > 1-\delta$ can be predicted model independently, and used to search for signatures of new physics with better sensitivity.
We thank Mikolaj Misiak for several useful discussions. M.L. thanks the Caltech Theory Group for hospitality while part of this work was completed. This work was supported in part by the Department of Energy under Grant Nos. DE-FG03-92-ER40701 and DOE-FG03-97ER40546. M.L. was supported in part by the Natural Sciences and Engineering Research Council of Canada and the Sloan Foundation. Fermilab is operated by Universities Research Association, Inc., under DOE contract DE-AC02-76CH03000.
S. Glenn [*et al.*]{}, CLEO Collaboration, CLEO CONF 98-17.
A.F. Falk, M. Luke, and M.J. Savage, Phys. Rev. D49 (1994) 3367.
A. Kapustin and Z. Ligeti, Phys. Lett. B355 (1995) 318.
M. Gremm, A. Kapustin, Z. Ligeti, and M.B. Wise, Phys. Rev. Lett. 77 (1996) 20;\
M. Gremm and I. Stewart, Phys. Rev. D55 (1997) 1226.
A.F. Falk, M. Luke, and M.J. Savage, Phys. Rev. D53 (1996) 2491; Phys. Rev. D53 (1996) 6316; A.F. Falk and M. Luke, Phys. Rev. D57 (1998) 424.
K. Chetyrkin, M. Misiak, and M. Munz, Phys. Lett. B400 (1997) 206;\
M. Misiak and M. Munz, Phys. Lett. B344 (1995) 308.
K. Adel and Y.P. Yao, Phys. Rev. D49 (1994) 4945;\
C. Greub and T. Hurth, Phys. Rev. D56 (1997) 2934.
A.J. Buras, M. Jamin, M.E. Lautenbacher, and P.H. Weisz, Nucl. Phys. B370 (1992) 69.
S.J. Brodsky, G.P. Lepage, and P.B. Mackenzie, Phys. Rev. D28 (1983) 228.
B.H. Smith and M.B. Voloshin, Phys. Lett. B340 (1994) 176.
E. Witten, Nucl. Phys. B120 (1977) 189.
C.H. Llewellyn Smith, Phys. Lett. B79 (1978) 83;\
K. Koller, T.F. Walsh, and P.M. Zerwas, Z. Phys. C2 (1979) 197.
D. Buskulic [*et al.*]{}, ALEPH Collaboration, Z. Phys. C69 (1996) 365.
D.W. Duke and J.F. Owens, Phys. Rev. D26 (1982) 1600.
A. Kapustin, Z. Ligeti, and H.D. Politzer, Phys. Lett. B357 (1995) 653.
A. Ali and C. Greub, Z. Phys. C49 (1991) 431; Phys. Lett. B259 (1991) 182; Phys. Lett. B287 (1992) 191.
N. Pott, Phys. Rev. D54 (1996) 938.
M. Neubert, Phys. Rev. D49 (1994) 4623; D49 (1994) 3392; I.I. Bigi, M.A. Shifman, N.G. Uraltsev and A.I. Vainshtein, Int. J. Mod. Phys. A9 (1994) 2467; R.D. Dikeman, M. Shifman, and N.G. Uraltsev, Int. J. Mod. Phys. A11 (1996) 571.
M.B. Voloshin, Phys. Lett. B397 (1997) 275;\
Z. Ligeti, L. Randall, and M.B. Wise, Phys. Lett. B402 (1997) 178;\
A.K. Grant, A.G. Morgan, S. Nussinov, R.D. Peccei, Phys. Rev. D56 (1997) 3151;\
G. Buchalla, G. Isidori, S.J. Rey, Nucl. Phys. B511 (1998) 594.
I.I. Bigi, M.A. Shifman, N.G. Uraltsev, A.I. Vainshtein, Phys. Rev. D50 (1994) 2234;\
M. Beneke and V.M. Braun, Nucl. Phys. B426 (1994) 301.
J. Bartelt [*et al.*]{}, CLEO Collaboration, CLEO CONF 98-21.
U. Aglietti [*et al.*]{}, Phys. Lett. B441 (1998) 371.
A.L. Kagan and M. Neubert, CERN-TH-98-99 \[hep-ph/9805303\].
C. Bauer, Phys. Rev. D57 (1998) 5611.
M. Gremm and A. Kapustin, Phys. Rev. D55 (1997) 6924.
A.H. Hoang, Z. Ligeti, and A.V. Manohar, Phys. Rev. Lett. 82 (1999) 277; Phys. Rev. D59 (1999) 074017 \[hep-ph/9811239\].
[^1]: Later we will introduce a dimensionless photon energy variable normalized by the $B$ meson mass, $x_B = 2
E_\gamma/m_B$.
[^2]: Note that these backgrounds are steeply falling functions of $x_b$, and are indeed negligible in the present CLEO region of $E_\gamma>2.1\,$GeV. The tree level contribution of the operators $O_3-O_6$ in Eq. (\[ops\]) to the photon spectrum is about a fifth of the $b\to u \bar u d$ background.
[^3]: There are actually additional contributions formally of order $\alpha_s(\Lambda_{\rm QCD}/m_B)$ coming from the expansion of $m_c/m_b$ in the 22 and 27 terms. Although the 27 term is very sensitive to the value of $m_c/m_b$, this $\bar\Lambda$-dependence is negligible for $\overline{(1-x_B)}
|_{x_B>1-\delta}$.
[^4]: It is interesting to note that including the CLEO data point in the $1.9\,{\rm GeV} < E_\gamma < 2.1\,$GeV bin, the experimental central value of $\overline{(1 - x_B)}$ over the region $E_\gamma > 1.9\,$GeV is 0.117, whereas the upsilon expansion predicts 0.120.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We propose that the anomalously bright white dwarf luminosity function observed in NGC 6791 (Bedin et al 2005) is the consequence of the formation of $0.5 M_{\odot}$ white dwarfs with Helium cores instead of Carbon cores. This may happen if mass loss during the ascent of the Red Giant Branch is strong enough to prevent a star from reaching the Helium flash. Such a model can explain the slower white dwarf cooling (relative to standard models) and fits naturally with scenarios advanced to explain Extreme Horizontal Branch stars, a population of which are also found in this cluster.'
author:
- 'Brad M. S. Hansen'
title: 'WHITE DWARFS IN NGC 6791: AVOIDING THE HELIUM FLASH'
---
Introduction
============
The open cluster NGC 6791 is an interesting object for several reasons. It is amongst the richest of the Galactic disk clusters, possibly the oldest, and possesses a significantly supersolar metallicity. These attributes have made it a target of several detailed studies. Nevertheless, there is still some uncertainty as to the true age. Stetson et al (2004) give an estimate $\sim 12$ Gyr, while studies such as that of Carney, Lee & Dodson (2005) or King et al (2005) yield estimates closer to $\sim 8$ Gyr. The age uncertainties are covariant with uncertainties in the cluster metallicity ($\rm \left[ Fe/H \right] \sim 0.25$–$0.5$) and distance.
The above estimates are all arrived at by measurements of the main sequence turnoff and giant sequences. The detection of a significant white dwarf population in NGC 6791 (Bedin et al 2005) is thus of considerable interest, since it offers the potential to measure cluster parameters by entirely independent means. However, fitting standard white dwarf models to the observed luminosity function yields a cluster age of $\sim 3$ Gyr. In this paper we examine some non-standard white dwarf models and propose an explanation for this puzzling discrepancy.
In § \[Models\] we review the results of Bedin et al and summarise the various possible explanations they were able to rule out. Thereafter we describe two additional possibilities not considered in that paper (residual nuclear burning and the production of massive white dwarfs with Helium cores). In § \[LF\] we compare these models to the data and demonstrate that the latter may indeed provide an explanation for the NGC 6791 luminosity function. In § \[Discuss\] we examine some of the consequences of this model and possible predictions.
The White Dwarf Luminosity Function {#Models}
===================================
The luminosity function derived by Bedin et al has the characteristic shape expected for the white dwarf population drawn from a burst of star formation – strongly peaked near a limiting value followed by a sharp drop at lower luminosities. The problem is that the location of the peak is far brighter than was anticipated. The expected absolute magnitude for a $0.5 M_{\odot}$ white dwarf of age 8 Gyr is $M_{606} \sim 16$ which, combined with a distance modulus of $\mu_{606}=13.44$, yields an expected peak location of $F606W=29.5$. Yet the luminosity function of Bedin et al peaks at approximately $F606W\sim 27.5$. They present compelling evidence that the peak lies well above their completeness limit, so that one is forced to consider ways to generate a peak in the luminosity function at much brighter magnitudes than expected.
Bedin et al reviewed several possible explanations for their luminosity function. Simply decreasing the distance or extinction is not viable, as it would drive the age derived from the main sequence turnoff to unacceptably large values. Small changes in the white dwarf models, such as changing the Hydrogen layer thickness or the ratio of Carbon and Oxygen in the core, do not produce a big enough effect. White dwarfs produced by truncated stellar evolution in binaries (Kippenhahn, Kohl & Weigert 1967) do cool more slowly, but generally have a range of masses $\sim 0.3$–$0.4 M_{\odot}$ and so would be redder than the observed white dwarf cooling track. Indeed, any explanation involving binary stars needs to account for the narrow mass range and the lack of bright (relative to a white dwarf) companions at the present time.
There is one potential explanation in the literature, that actually predates the Bedin et al observation. Bildsten & Hall (2001) and Deloye & Bildsten (2002) discuss the retardation of white dwarf cooling that is made possible by the sedimentation of $^{22}$Ne during the cooling process. Although the contribution is small for most white dwarfs, it gets larger if there is more $^{22}$Ne – which is expected from more metal-rich systems. Thus, Deloye & Bildsten in fact predicted that the effects of sedimentation would be largest in such a system as NGC 6791. At first glance, the NGC 6791 white dwarfs might appear to be a stunning confirmation of the prediction of ‘sedimentars’, but the size of the predicted effect is somewhat smaller than that needed to explain the Bedin et al result. Figure 12 and 13 of Deloye & Bildsten can be converted into a prediction that the bulk of the white dwarfs should be found between $F606W=28$–$29$, depending on the value of the assumed diffusion coefficient (the brightest value being found for a rate ten times faster than the nominal but uncertain value). Nevertheless, the agreement may be improved by further calculations in progress (L.Bildsten, personal communication).
At present, however, the observations still require an explanation and so we wish to re-examine two issues touched on by Bedin et al, but not explored to the fullest.
Thick Hydrogen Layers and Nuclear Burning
-----------------------------------------
Bedin et al did examine the effects of changing the Hydrogen layer mass on the cooling, but as far as can be told from the paper, considered only the effect it has on heat transport from the core to the surface. For a large enough Hydrogen layer, the pressure and temperature at the base are high enough to maintain some level of residual nuclear burning by the pp process, i.e. an additional heat source. This mechanism has been invoked in previous instances of anomalously young white dwarfs with low mass (e.g. Alberts et al 1996; Driebe et al 1998). Could residual nuclear burning be the explanation for the bright white dwarfs in NGC 6791? Unfortunately, this seems unlikely. The effects of nuclear burning are only significant for masses $M<0.25 M_{\odot}$ (Driebe et al 1999; Althaus, Serenelli & Benvenuto 2001; Hansen, Kalogera & Rasio 2003).
To confirm this, we calculate several new white dwarf cooling sequences using the code described in Hansen (1999). The Helium layer is taken to be $q_{He} = 10^{-2}$. All models have total mass $0.5 M_{\odot}$. We use the same C/O profile as in the previous models, but consider a variety of Hydrogen layer thicknesses, ranging from $q_{H}=10^{-4}$ to $q_{H} = 5 \times 10^{-3}$. The lower end of the range is the canonical value expected for most white dwarfs, but larger values are possible given the uncertain nature of the mass loss history of evolved stars.
In the initial stages (central temperatures $>10^8$ K) of the approach to the cooling sequence, the evolution is driven by neutrino cooling in the core and the cooling is similar for all models. However, as the star shrinks towards a more compact, degenerate configuration, the pressure and temperature at the base of the Hydrogen layer increase and nuclear burning is possible for some range of Hydrogen layer masses. For even the most massive layers considered, the models remain brighter than the fiducial ($q_H = 10^{-4}$) models only for $\sim 2$ Gyr. This is because, if one increases the mass of the Hydrogen layer, the pressure and density at the base of the envelope are higher and so the rate of burning is higher. In the end the effect of a larger reservoir is balanced by a higher rate of consumption, limiting the delay one can achieve. After 2 Gyr, no model has a Hydrogen layer larger than $q_H \sim 5 \times 10^{-4}$. Thus, after $\sim 8$ Gyr, the white dwarf will have approximately the same luminosity regardless of the initial Hydrogen mass. This is demonstrated in Figure \[LT\]. The white dwarf cooling time (neglecting the main sequence lifetime of the progenitor for now) for white dwarfs at $F606W \sim 27.3$ is 1.6 Gyr for a standard C/O model. Even the thickest Hydrogen envelope considered ($q_H \sim 5 \times 10^{-3}$) only lengthens the cooling time to this luminosity by 0.6 Gyr.
Helium Cores {#HeCore}
------------
The notion that the NGC 6791 white dwarfs have cores composed of Helium, rather than Carbon or Oxygen, was touched on briefly by Bedin et al. However, they ruled this out as Helium core white dwarfs are believed to result from binary evolution and the majority have masses that range $\sim 0.3$–$0.4 M_{\odot}$ and would thus be too red to fit the observations.
We believe this to be an overly conservative restriction. In this section we examine models for white dwarfs of mass $0.45$–$0.55 M_{\odot}$ (models with this mass are consistent with the colour and magnitude of the upper cooling sequence) which nevertheless possess cores composed purely of Helium. We will leave issues of provenance to § \[LF\].
For white dwarfs hot enough to be far from the strongly coupled regime in the core, the heat content (and so the cooling time) of the star is inversely proportional to the mass number of the core constituent. Thus, if we consider two $0.5 M_{\odot}$ white dwarfs, one composed of Carbon and the other of Helium, both with an age 8 Gyr, then the latter will be considerably brighter. In fact, it will have approximately the same luminosity as the Carbon core model has at an age $\sim 8/3 \sim 2.7$ Gyr. This is approximately the age inferred by Bedin et al. Figure \[LT\] shows that the Helium core model fares considerably better than the C/O models, yielding a cooling time of $\sim 4$ Gyr to $M_V\sim 27.3$. While this is still somewhat below the 8 Gyr age of the cluster, we must recall that the cluster age is the sum of the time spent on both the cooling sequence and on the main sequence. Thus, the difference between the cluster age and the Helium-core cooling time indicates the mass of the main sequence progenitors of this anomalous population.
We use the analytic stellar evolution formulae encoded in the [SSE]{} package (Hurley, Pols & Tout 2000) to calculate the main sequence lifetimes of stars with $Z=0.035$. For the low end of the cluster age range ($\sim 8$ Gyr), the progenitor mass for Helium core white dwarfs at the observed cutoff is $\sim 1.45 M_{\odot}$, while for the high end of the age range ($\sim 12$ Gyr), the progenitor mass is $\sim 1.17 M_{\odot}$. Exploring the range of applicable white dwarf masses ($0.45 M_{\odot}$ to $0.55 M_{\odot}$) does not change the conclusions significantly.
Next we will try to model the luminosity function. Before doing this we need to consider in more detail the possible origins of such a population.
Avoiding the Helium flash in NGC 6791 {#LF}
=====================================
There are good reasons to believe most white dwarfs have C/O cores. Upon reaching the tip of the Red Giant Branch (RGB), a star ignites core Helium burning (under degenerate conditions for lower mass progenitors – which leads to a thermonuclear runaway, a.k.a. the Helium flash), and the star moves onto the Horizontal Branch. It undergoes extended core Helium burning followed by shell burning on the Asymptotic Giant Branch before becoming a white dwarf. In order to avoid converting the Helium core to Carbon and Oxygen the star has to find some way to circumvent these later stages of stellar evolution.
Stars in binaries sometimes manage this feat by losing their envelope on the ascent of the RGB due to Roche lobe overflow i.e., mass transfer to a companion. However, this leads to a range of masses, most of which are considerably lower than the $\sim 0.5 M_{\odot}$ needed to fit the NGC 6791 cooling sequence. Bedin et al have already considered and rejected this possibility. In order for our model to work we have to hypothesize that many single stars in NGC 6791 have managed to lose their envelopes before reaching the Helium flash, so that they move directly from the RGB to the white dwarf cooling sequence.
In fact, this hypothesis dovetails quite naturally with several theories for the origins and nature of Extreme Horizontal Branch (EHB) stars. This term is used to describe a class of stars, found both in the field and in clusters, which appear to be related to traditional Horizontal Branch stars but which are hotter/bluer. It has been suggested (Faulkner 1972; Sweigart, Mengel & Demarque 1974) that such stars are core Helium burning stars with particularly thin Hydrogen envelopes, possibly as a result of extreme mass loss on the RGB. Subsequent work has developed this picture even further. Castellani & Castellani (1993) and Castellani, Luridiana & Romaniello (1994) report the formation of what they call “Red Giant Stragglers” in models for globular cluster evolution. In some cases, the stars make it to the white dwarf cooling sequence before igniting Helium. D’Cruz et al (1996) considered models with a range of mass loss rates on the RGB and found that, in addition to the formation of EHB stars, they formed so-called “flash-manque” stars, which lose so much mass that they never ignite Helium and simply go directly to Helium core white dwarfs. In fact, given the large amount of mass loss required to form EHB stars, especially the very bluest ‘blue hook stars’ (Brown et al 2001; Cassisi et al 2003; Moehler et al 2004), it would require extreme fine tuning to produce EHB stars without also producing some Helium core white dwarfs. This is important because NGC 6791 possesses a significant population of such EHB stars (Kaluzny & Udalski 1992; Liebert, Saffer & Green 1994). Thus, if one takes the above models at face value, a substantial population of Helium core white dwarfs is expected wherever EHB stars are found and thus should be found in NGC 6791.
We now try to model the observed luminosity function using this framework. We note that not all the white dwarfs in NGC 6791 can have Helium cores. The models discussed above that successfully avoid the Helium flash do so with progenitors that begin their lives with masses slightly larger than $1 M_{\odot}$. Thus, in our models we shall impose a critical mass $m_{crit}$, above which stars always produce standard C/O core white dwarfs. Furthermore, NGC 6791 does possess normal Helium-burning stars (the EHB stars make up $\sim 15\%$ of the Helium-burning stars according to Liebert et al 1994) and even the EHB stars are Helium-burning, so that clearly some C/O white dwarfs are being produced. It seems likely that stellar evolution in this cluster explores all three post-RGB avenues discussed above. One may estimate the branching ratios as follows.
King et al (2005) note that two of the EHB candidates from Kaluzny & Udalski (1992) lie within the HST field. Models suggest that this evolutionary stage lasts for $\sim 10^8$ years. The offspring of EHB stars are C/O white dwarfs and so, using the cooling time of C/O models (to $F606W=28$), we estimate that $\sim 2 \times 10^9/10^8 \times 2 = 40$ of the white dwarfs in this field brighter than $F606W=28$ should have come through this channel. In addition, C/O white dwarfs are also the end product of normal Helium-burning stars in the cluster. Although we do not have a strict count of HB stars in this field, Liebert et al (1994) estimate that $\sim 15$% of all NGC 6791 Helium-burning stars are EHB stars. Thus, we estimate that the total number of C/O white dwarfs produced in this field with $F606W<28$ is $\sim 40/0.15\sim 270$. The total number of white dwarfs observed by Bedin et al in the same field is $\sim 600$. The difference between these two numbers is the number of Helium core dwarfs that went directly from RGB to cooling sequence. Thus, the ratio of C/O core white dwarfs to Helium core dwarfs is estimated to be $\sim 270:330 \sim 5:6$. If we break this up further into EHB/normal HB/direct He core we estimate ratios $\sim 1:6:8$. The principal uncertainty in this procedure is the degree to which internal cluster dynamical evolution violates the implicit closed box assumption. Although EHB stars and white dwarfs should have very similar masses, the normal HB stars may have slightly higher masses and thus may be slightly underepresented (since the field is away from the cluster core) in the census just described. In light of this uncertainty, we adopt a ratio of C/O to Helium cores of 1:1 below.
Figure \[LF1\] shows the luminosity function of Bedin et al (now binned in 0.5 magnitude bins) compared to sample 8 Gyr model luminosity functions. The best fit value of $m_{crit}=1.6 M_{\odot}$ is slightly larger than the $1.45 M_{\odot}$ quoted in § \[HeCore\] because the peak of the realised Monte Carlo number distribution lies slightly below the cutoff. The solid line shows a histogram for a model in which all stars with $m<m_{crit}$ form Helium core white dwarfs. The dotted line shows the C/O core luminosity function for the same parameters. The dashed line corresponds to a model in which 50% of the stars with $m<m_{crit}$ form Helium cores and the rest form C/O cores. We have included the effects of incompleteness, using the results from Bedin et al, kindly provided by Ivan King. The truncation of the C/O core luminosity function shows where this becomes important. The truncation of the Helium core luminosity function occurs at significantly brighter magnitudes. The uncertain age of the cluster means that there is some flexibility in the choice of parameters. Figure \[LF2\] shows a fit using an age of 12 Gyr and $m_{crit}=1.25 M_{\odot}$. The solid, dotted and dashed lines once again indicate luminosity functions for Helium core, C/O core and a 1:1 combination, as in Figure \[LF1\].
There are several points to note about these fits. We have not performed a detailed parameter scan in fitting the data because we do not possess a proper photometric uncertainty map for this data (which indicates the probabilistic relation between intrinsic and observed magnitude as a function of model magnitude – see Hansen et al (2004) for a more detailed discussion of this question in the context of the globular cluster M4). The structure in the CMD in Figure 1 of Bedin et al suggests this is likely to be an important consideration in performing a true fit. We also need to assume something about the distribution of progenitor masses. We have assumed top heavy mass functions, with a slope $x<-1$ (where Salpeter is $x=1.35$). This is necessary to obtain luminosity functions as peaked as those seen here, and is also broadly consistent with the main sequence mass function observed by King et al (2005) for this cluster (the number of main sequence stars per magnitude bin increases with decreasing magnitude/increasing mass near the turnoff). For an open cluster as old as NGC 6791, there has likely been significant dynamical evolution and so one should be cautious when interpreting this as a true mass function.
Discussion {#Discuss}
==========
In this paper we propose a model to explain the origin of the peak in the NGC 6791 luminosity function at luminosities somewhat brighter than those expected on the basis of traditional white dwarf cooling models. Our proposal is that these are the result of the formation of Helium core white dwarfs resulting from strong mass loss on the RGB, thereby avoiding an episode of core Helium burning on the Horizontal Branch. In support of this model we point to the substantial population of EHB stars in NGC 6791, which are proposed to be stars which lost enough mass to almost make it to the white dwarf cooling track with their Helium cores intact but finally ignited Helium burning during the final contraction stage. In many models, EHB stars and Helium core white dwarfs are expected to form together as they are consequences of the same phenomenon – strong mass loss.
The scenario discussed here does make a definite prediction – that the white dwarf luminosity function for NGC 6791 should be bimodal, with a second peak at fainter magnitudes resulting from traditional C/O white dwarfs which did, indeed, pass through the Horizontal Branch phase. Figure \[LF3\] shows the expected luminosity function for our simple 8 Gyr model with $m_{crit}=1.6 M_{\odot}$. The solid, dotted and dashed histograms show the same luminosity functions as in Figure \[LF1\] but now we have not modelled the effects of incompleteness. We see that the second peak is expected to lie at magnitudes $F606 \sim 29$–$30$. The exact value will vary depending on the mass of the white dwarfs at the faint end. The value shown here is probably a little optimistic since we used only $0.5 M_{\odot}$ white dwarfs.
The narrowness of the white dwarf cooling sequence suggests that, however these Helium-core white dwarfs form, it must be by a process that strongly favors core masses $>0.4 M_{\odot}$. If heavy mass loss on the RGB is indeed the cause, this mass limit suggests that the progenitors must get within $\sim$ 1 magnitude of the tip of the RGB before their evolution is truncated (based on metal-rich models using the formulae of Hurley et al 2000). There is some observational support for this notion. Origlia et al (2002) used ISOCAM to search for infrared excesses suggestive of mass loss in the cores of five globular clusters. They found evidence for significant mass loss only close to the RGB tip. Similarly, Ita et al (2002) found many variable stars near the tip of the RGB in the Large Magellanic Cloud, possibly an indication of pulsation driven mass loss. While these results are encouraging, they only indicate that most of the mass loss on the RGB is indeed likely to occur near the tip. There is clearly still much work required to verify that the amount of mass loss on the NGC 6791 RGB is sufficient to justify our hypothesis.
Finally, the referee has raised the question whether the existence of this anomalous white dwarf population poses a problem for the determination of population ages from the white dwarf cooling sequence (Fontaine, Brassard & Bergeron 2001, Hansen 2004 and references therein). Within the model proposed here, this is not the case, because the difference between the Helium core and Carbon core white dwarfs is discrete in nature and any anomaly is readily apparent. Just as someone cannot be “a little bit pregnant”, a star must belong to one population or the other. If the model of Deloye & Bildsten is correct, however, then the anomaly is “tunable” via the abundance of $^{22}$Ne and thereby does indeed introduce an additional uncertainty to take into account. Thus, it is of considerable interest to distinguish between these two possibilities with a deeper observation of this cluster.
The work described here is supported by NASA grant ATP03-000-0084 and the Alfred P. Sloan Foundation. The author acknowledges helpful discussions with Jason Kalirai, Mike Rich, Harvey Richer, Ivan King, Francesco Ferraro and Lars Bildsten. He thanks the referee for a thoughtful and helpful referee report.
Alberts, F., Savonije, G. J., van den Heuvel, E. P. J. & Pols, O. R., 1996, Nature, 380, 676 Althaus, L. G., Serenelli, A. M. & Benvenuto, O. G., 2001, ApJ, 554, 1110 Bedin, L. R., Salaris, M., Piotto, G., King, I. R., Anderson, J., Cassissi, S. & Momany, Y., 2005, ApJ, 624, L45 Bergeron, P., Wesemael, F. & Beauchamp, A., 1995, PASP, 107, 1047 Bildsten, L. & Hall, D. M., 2001, ApJ, 549, L219 Brown, T. M., Sweigart, A. V., Lanz, T., Landsman, W. B. & Hubeny, I., 2001, ApJ, 562, 368 Carney, B. W., Lee, J.-W. & Dodson, B., 2005, AJ, 129, 656 Cassisi. S., Schlattl, H., Salaris, M. & Weiss, A., 2003, ApJ, 582, L43 Castellani, M. & Castellani, V., 1993, ApJ, 407, 649 Castellani, V., Luridiana, V., & Romaniello, M., 1994, ApJ, 428, 633 D’Cruz, N.L., Dorman, B., Rood, R. T. & O’Connell, R. W., 1996, ApJ, 466, 359 Deloye, C. J. & Bildsten, L., 2002, ApJ, 580, 1077 Driebe, T., Schönberner, D., Blöcker, T. & Herwig, F., 1998, A&A, 339, 123 Driebe, T., Blöcker, T., Schönberner, D., & Herwig, F., 1999, A&A, 350, 89 Faulkner, J., 1972, ApJ, 173, 401 Fontaine, G., Brassard, P. & Bergeron, P., 2001, PASP, 113, 409 Hansen, B., 2004, Phys. Rep., 399, 1 Hansen, B., 1999, ApJ, 520, 680 Hansen, B. et al., 2004, ApJS, 155, 551 Hansen, B., Kalogera, V. & Rasio, F. A., 2003, ApJ, 586, 1364 Hurley, J. R., Pols, O. R. & Tout, C. A., 2000, MNRAS, 315, 543 Ita, Y. et al., 2002, MNRAS, 337, L31 Kaluzny, J. & Udalsky, A., 1992, Acta Astron, 42, 29 King, I. R., Bedin, L. R., Piotto, G., Cassissi, S. & Anderson, J., 2005, astro-ph/0504627 Kippenhahn, R., Kohl, K. & Weigert, A., 1967, Z. Astr., 66, 58 Liebert, J., Saffer, R. A. & Green, E. M., 1994, ApJ, 107, 1408 Moehler, S., Sweigart, A. V., Landsman, W. B., Hammer, N. J. & Dreizler, S., 2004, A&A, 415, 313 Origlia, L., Ferraro, F. R., Fusi Pecci, F. & Root, R. T., 2002, ApJ, 571, 458 Stetson, P. B., Bruntt, H. & Grundahl, F.,2003, 115, 413 Sweigart, A. V., Mengel, J. G. & Demarque, P., 1974. A&A, 30, 13
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We investigate how the integrability of the derivatives of Orlicz-Sobolev mappings defined on open subsets of ${\ensuremath{\mathbb{R}}}^n$ affect the sizes of the images of sets of Hausdorff dimension less than $n$. We measure the sizes of the image sets in terms of generalized Hausdorff measures.'
address: 'Department of Mathematics and Statistics, University of Jyväskylä, P.O. Box 35, Fin-40014 University of Jyväskylä, Finland'
author:
- Tapio Rajala
- Aleksandra Zapadinskaya
- Thomas Zürcher
bibliography:
- 'jyv7.bib'
title: Generalized Hausdorff dimension distortion in Euclidean spaces under Sobolev mappings
---
Introduction
============
In this paper, we continue the study of Orlicz-Sobolev mappings by extending planar dimension distortion results to their $n$-dimensional analogues. We work with mappings that belong to the class $W^{1,1}_{\mathrm{loc}}$, meaning that the components of the mappings have locally integrable distributional derivatives. In order to obtain interesting distortion results, we make further assumptions on the mappings.
The distortion is measured on the level of generalized Hausdorff measure. The *generalized Hausdorff measure* of a set $A\subset {\ensuremath{\mathbb{R}}}^n$ is defined as $$\mathcal{H}^h(A)=\lim_{\delta\to0}\mathcal{H}^h_\delta(A),$$ where $$\mathcal{H}^h_\delta(A)=\inf\Bigl\{\sum\limits_{i=1}^{\infty}
h(\operatorname{diam}U_i)\colon A\subset\bigcup\limits_{i=1}^{\infty} U_i,\,
\operatorname{diam}U_i\leq\delta\Bigr\}$$ and $h\colon [0,\infty[ \to [0,\infty[$ is a dimension gauge: $\lim_{t\to0+}h(t)=h(0)=0$ and $h$ is non-decreasing. In the special case where $h(t)=t^\alpha$ with some $\alpha\geq0$, we have the usual *Hausdorff $\alpha$-dimensional measure*, which we simply denote by $\mathcal{H}^\alpha$. The *Hausdorff dimension* $\dim_{\mathcal{H}}A$ of a set $A \subset {\ensuremath{\mathbb{R}}}^n$ is the smallest $\alpha_0\geq0$ such that $\mathcal{H}^\alpha(A)=0$ for any $\alpha>\alpha_0$.
Let $\Omega\subset{\ensuremath{\mathbb{R}}}^n$ be an open set and $f\colon\Omega\to{\ensuremath{\mathbb{R}}}^n$ a continuous mapping. We obtain distortion estimates of the following form: If the mapping $f$ is assumed to be in an appropriate Orlicz-Sobolev class, then we have for every $E\subset\Omega$ $$\dim_{\mathcal H}(E) < n \Longrightarrow \mathcal H^{h_\gamma}(f(E)) = 0,$$ where $h_\gamma(t) = t^n \log^\gamma(1/t)$. Estimates of this kind were established in the plane in [@KZZ1; @KZZ; @Tapio]. The parameter $\gamma$ in the dimension gauge is chosen according to the integrability of the differential of the mapping $f$. The dependence between the parameter $\gamma$ and the threshold for the integrability of the differential varies with the global assumptions on the mapping $f$.
We make the strongest integrability assumptions for mappings that are only assumed to be continuous. We may then relax the assumptions if we require that in addition to being continuous, the mappings are also monotone. Recall that a real valued function $f\colon\Omega\to{\ensuremath{\mathbb{R}}}$ is said to be *monotone* if for every ball $B\subset\Omega$ we have $$\sup_{\partial B}f=\sup_Bf\,\,\,\,\,\,\,\,\text{ and }\,\,\,\,\,\,\,
\inf_{\partial B}f=\inf_Bf.$$ A mapping $f\colon\Omega\to{\ensuremath{\mathbb{R}}}^n$ is called monotone, if all its component functions are monotone. Further relaxation may be made if we assume that the mappings are homeomorphisms.
We list our results for the dimension distortion under Orlicz-Sobolev mappings under the three different assumptions in the following theorem.
\[os\] Let $\Omega\subset{\ensuremath{\mathbb{R}}}^n$ be an open set and $f:\Omega\to{\ensuremath{\mathbb{R}}}^n$ a continuous map in $W^{1,1}_{\mathrm{loc}}(\Omega,{\ensuremath{\mathbb{R}}}^n)$ with $|Df|^n \log^\lambda(e + |Df|)\in
L^1_{\mathrm{loc}}(\Omega)$ for some$\lambda\in{\ensuremath{\mathbb{R}}}$. Then $$\dim_{\mathcal H}(E) < n \Longrightarrow \mathcal
H^{h_\gamma}(f(E)) = 0,$$ if one of the following cases occurs:
- $\lambda > n-1$ and $\gamma < \lambda - n + 1$.
- $f$ is monotone, $\lambda > 0$, and $\gamma \leq \lambda$.
- $f$ is a homeomorphism, $f^{-1}\in W_{\mathrm{loc}}^{1,p}(f(\Omega),{\ensuremath{\mathbb{R}}}^n)$ for some , $\lambda > -1$, and $\gamma \leq
\lambda+1$.
Theorem \[os\] will be proved in the remaining sections of the paper. The claims (*ii*) and (*iii*) are the analogues of the results known in the planar case: [@KZZ Theorem 2] and [@Tapio Theorem 1.1]. Their proofs follow the planar ones. However, we have changed the presentation of the proof of claim (*ii*) in order to emphasize the common key elements of the proofs of claim (*i*) and claim (*ii*). Both of these claims follow from a tailored version of a Rado-Reichelderfer condition. To obtain the condition for claim (*i*), we rely on an auxiliary result from [@KKM99].
We do not know if the estimates in (*i*) and (*ii*) are sharp. However, by [@HK03 Proposition 5.1], see Section 2 in [@KZZ1] as well, given any $\lambda>0$, we may find a homeomorphism $f\colon[0,1]^n\to[0,1]^n$ such that $|Df|^n\log^s(e+|Df|)$ is integrable for all $s<\lambda-1$, mapping a set of Hausdorff dimension strictly less than $n$ onto a set of positive generalized Hausdorff measure with the gauge function $h(t)=t^n\log^\lambda(1/t)$. This example demonstrates the sharpness of the estimate in (*iii*).
Theorem \[os\] implies dimension distortion results for mappings of finite distortion. Recall that a continuous mapping $f \in
W^{1,1}_{\mathrm{loc}}(\Omega, {\ensuremath{\mathbb{R}}}^n)$ is called a *mapping of finite distortion*, if its Jacobian $J_f$ is locally integrable and there exists a measurable function $K \colon \Omega \to [1,\infty[$ such that $$|Df(x)|^n \le K(x)J_f(x)$$ at almost every point $x \in \Omega$. If we assume $K$ to be bounded, we obtain the class of *quasiregular mappings*. However, weaker assumptions already imply distortion estimates. For example, the assumption that the function $\exp(\lambda K)$ is locally integrable for some parameter $\lambda>0$. The mappings for which this is true are called *mappings of $\lambda$-exponentially integrable distortion*. See [@Geh-Väi; @gehring; @BMT] for dimension distortion results for quasiconformal mappings and [@HK03; @KZZ] for generalized dimension distortion estimates for mappings of exponentially integrable distortion in the plane.
We follow the approach taken in [@KZZ] and prove dimension distortion results for mappings of $\lambda$-exponentially integrable distortion using the higher regularity of the weak derivatives of the mappings. We obtain the higher regularity from [@FKZ05]. In the plane, the sharp regularity is known, see [@eero]. A straightforward combination of Theorem \[os\] and [@FKZ05 Theorem 1.1] gives the following result.
\[fd\] Let $\Omega\subset{\ensuremath{\mathbb{R}}}^n$ be a domain and $\lambda>0$. There exist positive constants $c_1$ and $c_2$ depending only on $n$ such that if $f:\Omega\to{\ensuremath{\mathbb{R}}}^n$ is of $\lambda$-exponentially integrable distortion and satisfies
- $\lambda > 1/c_1$ and $\gamma \leq c_1\lambda - 1$ or
- $f$ is homeomorphic and $\gamma \leq c_2\lambda$,
the following implication is true: $$\dim_{\mathcal H}(E) < n \Longrightarrow \mathcal
H^{h_\gamma}(f(E)) = 0.$$
By [@FKZ05 Theorem 1.1] there exists a constant $c_1$ depending only on $n$ such that $$\label{eq:corollary}
|Df|^n\log^{c_1\lambda-1}(e+|Df|)\in L^1_{\mathrm{loc}}(\Omega)$$ for each mapping $f$ satisfying the assumptions of the corollary.
By a result in [@vodopis] saying that mappings of finite distortion of the class $W^{1,n}_{\mathrm{loc}}(\Omega,{\ensuremath{\mathbb{R}}}^n)$ are continuous and monotone, claim (*i*) follows from Theorem \[os\] (*ii*).
For the proof of (*ii*), we note that implies that $|Df|$ is in the Lorentz space $L^{n-1,1}_\textnormal{loc}(\Omega)$, see [@Lorentz V.3]. The assumptions of Theorem 4.1 in [@HeKoMa] are fulfilled in our settings, giving us $f^{-1}\in W^{1,n}_{\mathrm{loc}}(f(\Omega),\Omega)$. This together with allows us to apply (*iii*) of Theorem \[os\], which concludes the proof.
The estimate in Corollary \[fd\] (*ii*) is sharp modulo the constant. Indeed, using again [@HK03 Proposition 5.1], for any given $0<\varepsilon<\lambda$, we find a homeomorphism $f\colon[0,1]^n\to[0,1]^n$, having $\nobreak{(\lambda-\varepsilon)}$-exponentially integrable distortion and mapping a set of Hausdorff dimension strictly less than $n$ onto a set of positive generalized Hausdorff measure with the gauge function $h(t)=t^n\log^\lambda(1/t)$. This example makes us believe that (*ii*) in Corollary \[fd\] holds with $c_2=1$. However, we do not know if the estimate in (*i*) is sharp.
In the next section, we prove a result that we will use for proving the cases (*i*) and (*ii*) in Theorem \[os\]. Afterwards, three more sections follow — one for each case of Theorem \[os\].
Rado-Reichelderfer condition
============================
A version of the following inequality was first mentioned by Rado and Reichelderfer, [@RadoReichelderfer].
\[RR\] Suppose that $\Omega\subset
{\ensuremath{\mathbb{R}}}^n$ is an open set, $f\colon \Omega\to{\ensuremath{\mathbb{R}}}^n$ a continuous mapping, and $\gamma>0$. Assume further that there is a function $\rho \colon \Omega \to [0,\infty]$ and constants $C\ge1$ and $r_0>0$ such that $$\rho\log^\gamma(e+\rho)\in L^1_\textnormal{loc}(\Omega)$$ and $$\label{RReq}
(\operatorname{diam}f(B(x,r)))^n\leq \int_{B(x,r)} \rho$$ for all $0<r<r_0$ and $x\in \Omega$ taken so that $B(x,Cr) \subset \Omega$. Then $$\dim_{\mathcal{H}}(E)<n\Longrightarrow
\mathcal{H}^{h_\gamma}(f(E))=0.$$
Let $E\subset \Omega$ be such that $\dim_{\mathcal{H}}(E)<n$. By the $\sigma$-additivity of the Hausdorff measure, we may assume without loss of generality that $E \subset\subset \Omega$. Fix $0<\varepsilon<\min\{1,e^{-\lambda n}\}$. By the absolute continuity of the integral, we may find such that $$\label{intbound}
\int_{A}\rho\log^\gamma(e+\rho)<\varepsilon<\min\{1,e^{-\lambda n}\},$$ whenever $A\subset\Omega$ is such that $\mathcal L^n (A)<\delta$, which is for example the case if $A$ is a ball of radius less than $\delta$. Note that $$\rho(x) \le \rho(x)\log^\gamma(e + \rho(x)),$$ so we obtain from $$\label{smallballs}
\operatorname{diam}f(B(x,r)) \le \left(\int_{B(x,r)}\rho \right)^{1/n}<\varepsilon^{1/n} < e^{-\lambda}$$ for $B(x,r)\subset\Omega$ with $0<r<\delta$.
Let us describe how to choose a suitable cover for $E$ that gives rise to an eligible cover of $f(E)$. First, we take $\alpha\in]\max\{\frac{n}{2},\dim_{\mathcal{H}}E\},n[$. Notice that there exists $t_0>0$ such that $$\label{logestimate}
t^n \le t^\alpha \log^\gamma\left(\frac{1}{t}\right) \le t^{(\alpha +\dim_{\mathcal{H}}E)/2}$$ for all $0<t<t_0$. Since $\mathcal{H}^{(\alpha
+\dim_{\mathcal{H}}E)/2}(E) = 0$, we may find for every$\varepsilon'>0$ a covering of $E$ with balls $\{B_i'\}_{i=1}^\infty$ of diameter less than$\min\{t_0,\delta, \frac{1}{C}\operatorname{dist}(E, \mathbb{R}^n \setminus \Omega)\}$ so that $$\sum_{i=1}^\infty(\operatorname{diam}B_i')^{(\alpha +\dim_{\mathcal{H}}E)/2} < \varepsilon'.$$ From the cover $\{B'_i\}$, we move to a more suitable cover. This is done by defining a new collection of balls $$\mathcal{B}' = \big\{B(x,r) : x \in B_i'\cap E \text{ and }2r = \operatorname{diam}B_i'\text{ for some }i \in {\ensuremath{\mathbb{N}}}\big\}$$ and applying the Besicovitch covering theorem to it. This gives a constant $0<N<\infty$, depending only on $n$, and a covering of the set $E$ with balls $\mathcal{B}=\{B_j\}_{j=1}^\infty \subset \mathcal{B}'$ of diameters less than $\delta$ so that $\sum_j \chi_{B_j}(x) \le N$ for every $x \in \Omega$.
Now, for any ball $B'_i$ from the original cover, all the balls $B(x,r)$ in $\mathcal{B}$ with $x \in B'_i$ and $2r = \operatorname{diam}B_i'$ contain the center of the ball $B'_i$, hence there are at most $N$ such balls. Therefore $$\sum_{j=1}^\infty(\operatorname{diam}B_j)^{(\alpha +\dim_{\mathcal{H}}E)/2} \le N\sum_{i=1}^\infty(\operatorname{diam}B_i')^{(\alpha +\dim_{\mathcal{H}}E)/2} < N\varepsilon'.$$ Now, by taking $\varepsilon'$ small enough, we have $\mathcal L^n(\bigcup_j B_j)<\delta$ and by $$\sum_{j=1}^\infty (\operatorname{diam}B_j)^\alpha\log^\gamma\Bigl(\frac{1}{\operatorname{diam}B_j}\Bigr)<\varepsilon.$$
Let us show that $\mathcal{H}_\delta^{h_\gamma}(f(E))=0$. We use the monotonicity of $t\log^\lambda(1/t)$ for $t\in
]0,e^{-\lambda}[$ and estimate together with the Rado-Reichelderfer condition to obtain $$\begin{aligned}
\label{StartEq}
(\operatorname{diam}f(B))^n\log^\gamma\Bigl(\frac{1}{(\operatorname{diam}f(B))^n}\Bigr)\leq
\int_B \rho\cdot \log^\gamma\Bigl(\frac{1}{\int_B \rho}\Bigr).\end{aligned}$$ We consider two cases. Let $$\begin{aligned}
\mathcal{B}_1:=\Big\{B_j\in \mathcal{B}:\, \int_{B_j}\rho\leq (\operatorname{diam}B_j)^\alpha\Big\},\\
\mathcal{B}_2:=\Big\{B_j\in \mathcal{B}:\, \int_{B_j}\rho > (\operatorname{diam}B_j)^\alpha\Big\}.\end{aligned}$$ For $B\in \mathcal{B}_1$, we use again the monotonicity of $t\log^\lambda(1/t)$ to obtain from inequality $$\begin{aligned}
(\operatorname{diam}f(B))^n\log^\gamma\Bigl(\frac{1}{(\operatorname{diam}f(B))^n}\Bigr)\leq
\alpha^\gamma(\operatorname{diam}B)^\alpha \log^\gamma\Bigl(\frac{1}{\operatorname{diam}B}\Bigr).\end{aligned}$$ In case that $B \in \mathcal{B}_2$, we conclude from that $$\begin{aligned}
(\operatorname{diam}f(B))^n\log^\gamma\Bigl(\frac{1}{(\operatorname{diam}f(B))^n}\Bigr)
\leq
\alpha^\gamma\int_B \rho\cdot \log^\gamma\Bigl(\frac{1}{\operatorname{diam}B}\Bigr).\end{aligned}$$ We split the integral in two parts. In order to continue, we set $$A_B:=\Big\{y\in B:\, \rho(y)\le\frac{1}{(\operatorname{diam}B)^{n-\alpha}}\Big\}$$ and obtain $$\begin{aligned}
\int_{A_B} \rho(y)&\cdot \log^\gamma\Bigl(\frac{1}{\operatorname{diam}B}\Bigr)\, dy\\
&\leq
\frac{\omega_n}{(\operatorname{diam}B)^{n-\alpha}}(\operatorname{diam}B)^n\log^\gamma\Bigl(\frac{1}{\operatorname{diam}B}\Bigr)\\
&= \omega_n(\operatorname{diam}B)^\alpha\log^\gamma\Bigl(\frac{1}{\operatorname{diam}B}\Bigr),\end{aligned}$$ where $\omega_n$ denotes the $n$-dimensional Lebesgue measure of an $n$-dimensional unit ball. Now, we focus on $B\setminus A_B$: $$\begin{aligned}
\int_{B\setminus A_B}&\rho(y)\log^\gamma\Bigl(\frac{1}{(\operatorname{diam}B)^\alpha}\Bigr)\, dy\\
=&\int_{B\setminus A_B} \rho(y)\Bigl(\frac{\alpha}{n-\alpha}\Bigr)^\gamma
\log^\gamma\Bigl(\frac{1}{(\operatorname{diam}B)^{n-\alpha}}\Bigr)\, dy \\
\leq& \Bigl(\frac{\alpha}{n-\alpha}\Bigr)^\gamma\int_{B\setminus A_B}
\rho(y)
\log^\gamma(e+\rho(y))\, dy.\end{aligned}$$ Hence, keeping in mind, we obtain the following upper bound: $$\begin{aligned}
\sum\limits_{B\in\mathcal
B_2}&\int_{B\setminus A_B}\rho(y)\log^\gamma\Bigl(\frac{1}{(\operatorname{diam}B)^\alpha}\Bigr)\, dy\\
&\leq\Bigl(\frac{\alpha}{n-\alpha}\Bigr)^\gamma N\int_{\bigcup_i B_i}\rho(y)
\log^\gamma(e+\rho(y))\, dy\\
&<N\Bigl(\frac{\alpha}{n-\alpha}\Bigr)^\gamma
\varepsilon.\end{aligned}$$ Thus, we estimate $$\begin{aligned}
\mathcal H^h_{\varepsilon^{1/n}}(f(E))\leq&n^{-\gamma}\sum_{j=1}^\infty (\operatorname{diam}f(B_j))^n\log^\gamma\Bigl(\frac{1}{(\operatorname{diam}f(B_j))^n}\Bigr)\\
\leq&\omega_n
n^{-\gamma}\alpha^\gamma\Bigl(1+N\Bigl(\frac{\alpha}{n-\alpha}
\Bigr)^\gamma \Bigr)\varepsilon.\end{aligned}$$ Letting $\varepsilon$ go to zero concludes the proof.
Proof of the continuous case
============================
Suppose that $\varphi$ is a positive function on the interval $]0,\infty[$. We write $$F_\varphi(s)=
\begin{cases}
s\varphi^{\frac{1}{n}-1}(s) & s>0,\\
0 & s=0.
\end{cases}$$
Our proof of Theorem \[os\] (i) is based on [@KKM99 Theorem 3.2], which, in the case of continuous mappings, states the following.
\[Lorentz\] Let $u\in W^{1,1}_{\mathrm{loc}}(\Omega,{\ensuremath{\mathbb{R}}})$ be a continuous function. Let further $\varphi$ be a positive, non-increasing function on $]0,\infty[$. Suppose that $$\int_\Omega F_\varphi(|Du(x)|)\, dx<\infty$$ and $$\int_0^\infty \varphi^{1/n}(t)\, dt<\infty.$$ Then $$(\operatorname{osc}_{B(x,r)} u)^n = \int_{B(x,r)}
\frac{2^{n(n+2)}}{n\omega_n}\biggl(\int_0^\infty
\varphi^{1/n}(t)\, dt\biggr)^{n-1} F_\varphi(|Du(x)|)\, dx.$$
Let us fix $\gamma\in]0,\lambda+1-n[$ and prove that $\mathcal
H^h(f(E))=0$ with $h(t)=t^n\log^\gamma(1/t)$. We want to verify the Rado-Reichelderfer condition by applying Theorem \[Lorentz\]. Hence, we need to choose a function $\varphi\colon
]0,\infty[\to {\ensuremath{\mathbb{R}}}$ so that $F_\varphi$ matches the integrability condition specified in Theorem \[os\]. We choose and set $$\varphi(s)=
\begin{cases}
\log^\frac{\tilde{\lambda} n}{1-n}(e+1) & 0<s<1,\\
s^{-n}\log^\frac{\tilde{\lambda} n}{1-n}(e+s) & s\geq 1.
\end{cases}$$ This is a suitable choice for $\varphi$ since then $F_\varphi$ is defined as $$F_\varphi(s)=
\begin{cases}
0 & s=0,\\
s\log^{\tilde{\lambda}}(e+1) & 0<s\leq 1,\\
s^n\log^{\tilde{\lambda}}(e+s) & 1<s
\end{cases}$$ giving $$\int_\Omega F_\varphi(|Df(x)|)\, dx <\infty$$ by assumption. Clearly, $\varphi$ satisfies the requirements in Theorem \[Lorentz\].
We conclude from Theorem \[Lorentz\] that there exists a constant $C>0$ depending only on $n$ and $\tilde{\lambda}$ so that for every $x \in \Omega$ and $r > 0$ with $B(x,r) \subset \Omega$, the estimate $$\begin{aligned}
\label{mainineq}
(\operatorname{diam}f(B(x,r)))^n\leq& \Bigl(\sum\limits_{i=1}^n\operatorname{osc}^2_{B(x,r)}f_i\Bigr)^{n/2}\\
\leq&
\Bigl(\sum\limits_{i=1}^n\Bigl(C\int_{B(x,r)}F_\varphi(|Df_i|)\Bigr)^{2/n}\Bigr)^{n/2
}\notag\\
\leq& \,Cn^{n/2}\int_{B(x,r)}F_\varphi(|Df|)\notag\end{aligned}$$ holds. We define $\rho$ as $Cn^{n/2}F_\varphi(|Df|)$. We further notice that $$\rho\log^\gamma(e+\rho)$$ is locally integrable in $\Omega$ due to the choice of parameters and the given integrability of $|Df|$; indeed, behaves asymptotically like $s\log^{\tilde{\lambda}+\gamma}(e+s)$ when $s$ is large, and we have . Having verified the Rado-Reichelderfer condition as specified in Proposition \[RR\], we conclude the proof.
Proof of the continuous and monotone case
=========================================
We use the symbol $A(x,r,R)$ to denote the closed annulus with center at $x$ and radii $r$ and $R$: $$A(x,r,R)=\{y\in{\ensuremath{\mathbb{R}}}^n\colon r\leq|x-y|\leq R\}.$$ The proof we present here repeats the main steps of its planar analogue [@KZZ Theorem 2]. We start by proving an estimate on the sizes of the images of balls.
\[lem:uboj\] Let $\Omega$ be a domain in ${\ensuremath{\mathbb{R}}}^n$ and $f\colon\Omega\to
f(\Omega)$ be a continuous and monotone mapping of the Sobolev class $W^{1,p}_{\mathrm{loc}}(\Omega, {\ensuremath{\mathbb{R}}}^n)$, where $n-1<p<n$. Then there exists a constant $C$ depending only on $n$ and $p$ such that for every ball $B(x,2r)\subset\subset\Omega$, the following estimate holds $$\label{estimateAnnulus}
\operatorname{diam}f(B(x,r))\leq Cr^{1-\frac{n}{p}}
\Bigl(\int_{A(x,r,2r)}|Df|^p\Bigr)^{1/p}.$$
Fix a ball $B(x,r)$. Without loss of generality, we may assume that $B(x,2r)\subset\subset\Omega'$ for some domain $\Omega'$ in $\Omega$ and that $f\in W^{1,p}(\Omega',{\ensuremath{\mathbb{R}}}^n)$. Using the monotonicity of $f$, we obtain $$\begin{aligned}
\operatorname{diam}f(B(x,r))&\leq(\sum_{i=1}^{n}(\operatorname{osc}_{B(x,r)} f_i)^2)^{1/2}\\
&\leq\sum\limits_{i=1}^{n}\operatorname{osc}\limits_{B(x,r)} f_i\leq\sum\limits_{i=1}^{n}\operatorname{osc}\limits_{S(x,t)} f_i \end{aligned}$$ for all $t\in]r,2r[$. Since $f_i$ is in $W^{1,p}(\Omega',
{\ensuremath{\mathbb{R}}})$, $i=1,\ldots,n$, Theorem 5.16 on p. 121 in [@Degree] gives us $f_i\in W^{1,p}(S(x,t))$, $i=1,\ldots,n$, for almost every $t\in]r,2r[$. See the same page for the definition of the class $W^{1,p}(X)$ for a $C^\infty$ paracompact manifold $X$ such as a sphere and for Theorem 5.15, which gives us $$\operatorname{osc}\limits_{S(x,t)} f_i\leq
Ct^{1-\frac{n-1}{p}}\Bigl(\int_{S(x,t)}|Df_i|^p\Bigr)^{1/p}$$ for almost every $t\in [r,2r]$, $i=1,\ldots,n$. Hence $$\operatorname{diam}f(B(x,r))\leq Ct^{1-\frac{n-1}{p}}\sum\limits_{i=1}^{n}\Bigl(\int_{S(x,t)}|Df_i|^p\Bigr)^{1/p}$$ for almost every $t\in[r,2r]$. Integrating the last inequality over $[r,2r]$ with respect to $t$ and using Hölder’s inequality, we obtain $$\begin{aligned}
r\operatorname{diam}&f(B(x,r))\leq C\int_{[r,2r]}t^{1-\frac{n-1}{p}}
\sum\limits_{i=1}^{n}\Bigl(\int_{S(x,t)}|Df_i|^p\Bigr)^{1/p}dt\\
&\leq C\sum\limits_{i=1}^{n}\Bigl(\int_{[r,2r]}\int_{S(x,t)}|Df_i|^p\,dt\Bigr)^{1/p}
\Bigl(\int_{[r,2r]}t^{\frac{p-n+1}{p-1}}\Bigr)^{\frac{p-1}{p}}\\
&\leq Cr^{2-\frac{n}{p}}\sum\limits_{i=1}^{n}\Bigl(\int_{A(x,r,2r)}|Df_i|^p\Bigr)^{1/p}
\\&\leq Cnr^{2-\frac{n}{p}}\Bigl(\int_{A(x,r,2r)}|Df|^p\Bigr)^{1/p}.\end{aligned}$$
In comparison with the Rado-Reichelderfer condition , we integrate in over the annulus $A(x,r,2r)$ instead of the ball $B(x,r)$. Our aim is to bound the integral over the annulus by an integral over the ball. In order to do so, we have to change the integrand. More precisely, we replace it by a maximal operator.
Assume that $\Omega\subset{\ensuremath{\mathbb{R}}}^n$ is a cube and $h\colon \Omega \to {\ensuremath{\mathbb{R}}}$ is a non-negative and integrable function. The maximal operator $\mathcal{M}_\Omega$ at a point $x \in \Omega$ is defined by $$\mathcal{M}_\Omega h(x)=\sup\Bigl\{{\Xint-}_{Q} h\, dy\colon\, x\in Q\subset
\Omega\Bigr\},$$ where the supremum is taken over all subcubes of $\Omega$ containing the point $x$.
In order to continue, we use the following lemma. It was proved in [@KZZ Lemma 2] in the planar case. However, the same proof works in higher dimensions.
\[lemma2\] Let $\Omega\subset {\ensuremath{\mathbb{R}}}^n$ be a cube, $f\colon\Omega\to {\ensuremath{\mathbb{R}}}^n$ be a mapping of the Sobolev class $W^{1,p}_{\mathrm{loc}}(\Omega, {\ensuremath{\mathbb{R}}}^n)$, where $n-1<p<n$. Then there exists a constant $C$ depending only on $n$ such that the inequality $$\int_{A(x,r,2r)}|Df|^p\leq
C\int_{B\left(x,r\right)}\mathcal{M}_\Omega|Df|^p$$ holds for all $x\in\Omega$ and $r>0$, such that $B(x,4\sqrt{n}r)\subset\subset\Omega$.
We will also use the following auxiliary result, which was proved in [@Limits Lemma 5.1].
\[lemmafromLimits\] Suppose $A \colon [0,\infty[ \to [0,\infty[$ is increasing and $\Phi(t) = A(t)t^p$ for some $p > 1$. Then there exists a constant $C>0$ depending only on $n$ and $p$ such that $$\int_\Omega \Phi(M_\Omega h) \le C \int_\Omega \Phi(Ch).$$
Now, we are ready to combine the preceding 3 lemmas.
\[prop:1\] Let $\Omega\subset {\ensuremath{\mathbb{R}}}^n$ be a cube and $f\colon \Omega\to
f(\Omega)$ be a continuous and monotone mapping in $W^{1,1}\left(\Omega,{\ensuremath{\mathbb{R}}}^n\right)$ satisfying $$|{Df}|^n\log^\lambda\left(e+|{Df}|\right)\in
L^1\left(\Omega\right)$$ for some $\lambda>0$. Then there exists a function $\rho\in L^1(\Omega)$ such that $$\label{intOfRho}
\rho\log^\lambda(e+\rho)\in L^1(\Omega)$$ and $$\label{RRConclusion}
(\operatorname{diam}f(B\left(x,r\right)))^n\leq \int_{B\left(x,r\right)}\rho,$$ for all $x\in \Omega$ and $r>0$ such that $B(x,4\sqrt{n}r)\subset\subset\Omega$.
The proof is similar to the one in [@KZZ]. Fix $x\in \Omega$ and $r>0$ such that $B(x,4\sqrt{n}r)\subset\subset\Omega$. Fix also $p\in]n-1,n[$. In Lemma \[lem:uboj\], we deduced the inequality $$\operatorname{diam}f(B(x,r))\leq Cr^{1-\frac{n}{p}}
\Bigl(\int_{A(x,r,2r)}|Df|^p\Bigr)^{1/p}$$ with a constant $C$ depending only on $n$ and $p$. In what follows, the constant $C$, still depending only on $n$ and $p$, may change its value from occurrence to occurrence. The combination of the inequality above with the upper bound found in Lemma \[lemma2\] implies $$\operatorname{diam}f(B(x,r))\leq Cr^{1-\frac{n}{p}}
\Bigl(\int_{B(x,r)}\mathcal{M}_\Omega(|Df|^p)\Bigr)^{1/p}.$$ We continue the estimation with the Hölder inequality: $$\begin{aligned}
\label{eq:keyineq}
(\operatorname{diam}f(B(x,r)))^n&\leq Cr^{n-\frac{n^2}{p}}
\Bigl(\int_{B(x,r)}\mathcal{M}_\Omega(|Df|^p)\Bigr)^{n/p}\\
&\leq C\int_{B(x,r)}\mathcal{M}^{n/p}_\Omega(|Df|^p).\notag\end{aligned}$$ In order to finish, we have to verify that $$\rho(x):=\mathcal M^{n/p}_\Omega(|Df|^p)(x)$$ fulfills . As in Lemma 3 in [@KZZ], we prove it as an application of Lemma \[lemmafromLimits\]. We let $h=|Df|^p$ and $\Phi(t)=t^{n/p}\log^\lambda(e+t^{n/p})$ and use the fact that $n/p>1$. We get $$\begin{aligned}
\int_\Omega&\mathcal{M}_\Omega^{n/p}(|Df|^p)
\log^\lambda\bigl(e+\mathcal{M}_\Omega^{n/p}(|Df|^p)\bigr)\\
&\leq C\int_\Omega|Df|^n
\log^\lambda(e+C|Df|^n)\\
&\leq C\int_\Omega|Df|^n
\log^\lambda(e+|Df|)<\infty.\end{aligned}$$
By the $\sigma$-additivity of the Hausdorff measure, we may assume that $\Omega$ is a cube. In Proposition \[prop:1\], we have verified the Rado-Reichelderfer condition . In Proposition \[RR\], we have shown that this condition is sufficient under the given integrability of $|Df|$ to conclude the proof.
Proof of the homeomorphic case
==============================
For a set $V\subset{\ensuremath{\mathbb{R}}}^n$ and a number $\delta>0$, $V+\delta$ denotes the set $\{y\in{\ensuremath{\mathbb{R}}}^n\,:\operatorname{dist}(y,V)<\delta\}$.
Without loss of generality, we assume for the rest of this section that $\Omega$, in addition to being open, is connected. The first step is an analogue of [@KZZ1 Lemma 3.1].
\[lemma:inverse\] Let $f\colon\Omega\to f(\Omega)\subset{\ensuremath{\mathbb{R}}}^n$ be a homeomorphism such that $f^{-1}\in
W^{1,p}_{\mathrm{loc}}(\Omega,{\ensuremath{\mathbb{R}}}^n)$ for some $p\in]n-1,n[$. Then there exists a set $F\subset
f(\Omega)$ such that $\mathcal{H}^{n-\frac{p}{2}}(F)=0$ and for all $y\in
f(\Omega)\setminus F$ there exist constants $C_y>0$ and $r_y>0$ such that $$\label{ineq:desired}
\operatorname{diam}(f^{-1}(B(y,r)))\leq C_yr^{1/2},$$ for all $0<r<r_y$.
Let us apply Lemma \[lem:uboj\] to the mapping $f^{-1}$. We obtain $$\operatorname{diam}f^{-1}(B(y,r))\leq Cr^{1-\frac{n}{p}}
\sum\limits_{i=1}^{n}\Bigl(\int_{B(y,2r)}|Df^{-1}_i|^p\Bigr)^{1/p}$$ for all $y\in f(\Omega)$ and $r>0$ such that $B(y,2r)\subset\subset f(\Omega)$, where the constant $C$ depends only on $n$ and $p$. So, the desired inequality holds for all $y\in f(\Omega)$ for which $$\label{eq:bad}
r^{\frac{p}{2}-n}\int_{B(y,2r)}|Df^{-1}_i|^p<M_y$$ is valid for all $i=1,\ldots,n$, for all small enough $r>0$ and for some constant $M_y$, depending on $y$. Let $F_1$ be the set of those $y$ for which does not hold with $i=1$. Let $K\subset f(\Omega)$ be a compact set and fix some $\delta>0$ such that $\operatorname{dist}(K,\partial f(\Omega))>\delta$. For every $k\in\mathbb{N}$ and every $y$ in $F_1\cap K$, there exists $r_{k,y}<\delta/10$ such that $\int_{B(y,2r_{k,y})}|Df^{-1}_1|^p\geq
k(r_{k,y})^{n-\frac{p}{2}}$. Consider the collection of balls $\mathcal{B}_k=\{B(y,2r_{k,y})\colon y\in F_1\cap K\}$ for every $k\in{\ensuremath{\mathbb{N}}}$. Using Vitali’s covering theorem, we obtain for every $k\in\mathbb{N}$ a countable subcollection of disjoint balls $B_{k,j}$, $j=1,2,\ldots$, centered in $F_1\cap K$, having radii $2r_{k,j}<\delta/5$, and with $\bigcup\limits_{j=1}^\infty5B_{k,j}$ covering $F_1\cap K$. We have $$\begin{aligned}
\mathcal{H}^{n-\frac{p}{2}}_\delta(F_1\cap
K)&\leq\sum\limits_{j=1}^{\infty}(20r_{k,j})^{n-\frac{p}{2}}\leq\frac{20^{n-\frac{p}{2}}}{k}
\sum\limits_{j=1}^{\infty}\int_{B_{k,j}}|Df^{-1}_1|^p\\&
\leq\frac{20^{n-\frac{p}{2}}}{k}\int_{K+\delta/5}|Df^{-1}_1|^p\end{aligned}$$ for all $k\in\mathbb N$. Letting $k\to\infty$ and $\delta\to0$, we obtain .
Since $f$ is a homeomorphism in $W^{1,p}_{\mathrm{loc}}(\Omega,{\ensuremath{\mathbb{R}}}^n)$ for all $0<p<n$, we know by [@STH] that the Jacobian of $f$ is either non-positive or non-negative almost everywhere (we assumed $\Omega$ to be connected). Thus, we may assume that $J_f\geq 0$ a.e. in $\Omega$. The rest of the proof goes as in [@Tapio] (Lemma 2.2 and the proof of Theorem 1.1).
Let us outline the proof. First, we cover the set $E$ with a countable collection of sets $E_j$ so that in each set the constants $C_y$ and $r_y$ of Lemma \[lemma:inverse\] are fixed. On each of the covering sets, we use the fact that the dimension of the set is less than $n$ to obtain a nice covering by small balls. Next, we estimate the size of the image of the sets $E_j$ by the integral of the Jacobian of $f$. For this purpose, we enlarge the covering balls, use the $5r$-covering theorem and Lemma \[lemma:inverse\]. The claim follows from the higher integrability of the Jacobian of $f$, which is obtained from the assumption on the integrability of $|Df|$ by [@ivaver Corollary 9.1].
Acknowledgments {#acknowledgments .unnumbered}
---------------
The authors thank Pekka Koskela for suggesting this problem.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Gas in galactic disks is collected by gravitational instabilities into giant atomic-molecular complexes, but only the inner, molecular parts of these structures are able to collapse to form stars. Determining what controls the ratio of atomic to molecular hydrogen in complexes is therefore a significant problem in star formation and galactic evolution. In this paper we use the model of H$_2$ formation, dissociation, and shielding developed in the previous paper in this series to make theoretical predictions for atomic to molecular ratios as a function of galactic properties. We find that the molecular fraction in a galaxy is determined primarily by its column density and secondarily by its metallicity, and is to good approximation independent of the strength of the interstellar radiation field. We show that the column of atomic hydrogen required to shield a molecular region against dissociation is $\sim 10$ ${M_{\odot}}$ pc$^{-2}$ at solar metallicity. We compare our model to data from recent surveys of the Milky Way and of nearby galaxies, and show that the both the primary dependence of molecular fraction on column density and the secondary dependence on metallicity that we predict are in good agreement with observed galaxy properties.'
author:
- 'Mark R. Krumholz'
- 'Christopher F. McKee'
- Jason Tumlinson
title: |
The Atomic to Molecular Transition in Galaxies.\
II: HI and H$_2$ Column Densities
---
Introduction
============
The formation of molecular hydrogen is a critical step in the transformation of interstellar gas into new stars. The neutral atomic interstellar medium (ISM) in galaxies is generally segregated into cold clouds embedded in a warm inter-cloud medium [@mckee77; @wolfire03], and the inner parts of some of these cold atomic clouds harbor regions where the gas is well-shielded against dissociation by the interstellar radiation field (ISRF). In these regions molecules form, and once they do star formation follows.
A full theory of star formation requires as one of its components a method for expressing in terms of observables the fraction of a galaxy’s ISM that is in the molecular phase [e.g. @krumholz05c]. No models published to date satisfy this requirement, but observations have yielded a number of empirical rules for galaxies’ molecular content. Based on [H<span style="font-variant:small-caps;">i</span> ]{}and CO mapping of nearby galaxies @wong02 [hereafter WB02] and @blitz04 [@blitz06b hereafter BR04 and BR06] infer that the molecular to atomic surface density ratio $R_{\rm H_2}=\Sigma_{\rm H_2}/\Sigma_{\rm HI}$ in a galaxy varies with the interstellar pressure $P$ needed for hydrostatic balance in the ISM as $R_{\rm H_2}\propto P^{0.92}$, and that the atomic surface density saturates at a maximum value of $\sim 10$ ${M_{\odot}}$ pc$^{-2}$. The observed saturation and a similar dependence of molecular fraction on pressure are also seen in newer surveys such as the HERA CO-Line Extragalactic Survey (HERACLES) that cover a broader range of galaxy properties at higher spatial resolution [@walter08a; @bigiel08; @leroy08 hereafter L08]. However the physical origin of these patterns is unclear. The samples on which they are based are composed solely of nearby galaxies with a limited range of properties, and in the absence of a physical model it is uncertain how far they can safely be extrapolated to regimes of metallicity, surface density, or other properties not represented in samples of nearby galaxies.
Theoretical treatments of the problem to date do not yet make such an extrapolation possible. A number of authors have considered the microphysics of H$_2$ formation and the structure of photodissociation regions in varying levels of detail [e.g. @vandishoeck86; @black87; @sternberg88; @elmegreen89; @draine96; @neufeld96; @spaans97; @hollenbach99; @liszt00; @liszt02; @browning03; @allen04], but none of these treatments address the problem of atomic to molecular ratios on galactic scales. @wyse86a and @wang90a [@wang90b] present models for cloud formation in galactic disks, but these both rely on prescriptions for the rate of conversion of atomic to molecular gas that are based either on rates of cloud collisions or on Schmidt laws, not on physical models of H$_2$ formation and dissociation. @elmegreen93 gives a theory of the molecular fraction in galaxies that does include a treatment of the H$_2$ formation and self-shielding. However, his model neglects dust shielding, an order unity effect, and it also requires knowledge of a galaxy’s ISRF strength, which cannot easily be determined observationally, and its interstellar pressure, which can only be inferred indirectly based on arguments about hydrostatic balance. This makes the model difficult to test or to apply as part of a larger theory of star formation. @schaye04 considers the conditions necessary to form a cold atomic phase of the ISM. The existence of such a phase is a necessary but not sufficient condition for molecule formation, so while @schaye04’s model provides a minimum condition for star formation, in makes no statements about what fraction of the ISM goes into the molecular phase able to form stars, and thus no statement on the star formation rate that is achieved once the minimum condition is met.
Numerical models are in a similar situation. @hidaka02a and @pelupessy06 simulate galaxies using subgrid models for H$_2$ formation similar to those presented by @elmegreen93, and show that they can reproduce some qualitative features of the H$_2$ distribution in galaxies. @robertson08 show that a simulation of a galaxy’s ISM that includes radiative heating and cooling in the ionized and atomic phases, coupled with an approximate treatment of H$_2$ formation on grains and dissociation by the ISRF, can reproduce the observed molecular content of galaxies. This suggests that the simulations contain the necessary physical ingredients to explain the observations, but the simulations do not by themselves reveal how these ingredients fit together to produce the observed result. Moreover, like the observed empirical rules, the simulations are based on a very limited range of galaxy properties, and in the absence of a model we can use to understand the origin of the simulation results, it is unclear how to extrapolate. Extending the simulations to cover the full range of galaxy parameters in which we are interested would be prohibitively expensive in terms of both computational and human time.
Our goal in this paper is to remedy this lack of theoretical understanding by providing a first-principles theoretical calculation of the molecular content of a galactic disk in terms of direct observables. In @krumholz08c [hereafter Paper I], we lay the groundwork for this treatment by solving the idealized problem of determining the location of the atomic to molecular transition in a uniform spherical gas cloud bathed in a uniform, isotropic dissociating radiation field. In this paper we apply our idealized model to atomic-molecular complexes in galaxies as a way of elucidating the underlying physical processes and parameters that determine the molecular content. We refer readers to Paper I for a full description of our solution to the idealized problem, but here we repeat a central point: for a spherical cloud exposed to an isotropic dissociating radiation field, if we approximate the transition from atomic to molecular as occurring in an infinitely thin shell separating gas that is fully molecular from gas of negligible molecular content, the fraction of a cloud’s radius at which this transition occurs is solely a function of two dimensionless numbers: $$\begin{aligned}
\chi & = & \frac{{f_{\rm diss}}{\sigma_{\rm d}}c {E^{*}}_0}{n{\mathcal{R}}} \\
{\tau_{\rm R}}& = & n {\sigma_{\rm d}}R.\end{aligned}$$ Here ${f_{\rm diss}}\approx 0.1$ is the fraction of absorptions of a Lyman-Werner (LW) band photons that produce H$_2$ dissociation rather than simply excitation and radiative decay to a bound state, ${\sigma_{\rm d}}$ is the dust absorption (not extinction) cross-section per hydrogen nucleus in the LW band, ${E^{*}}_0$ is the free-space number density of LW photons (i.e. far outside our cloud), $n$ is the number density of hydrogen nuclei in the atomic shielding layer, ${\mathcal{R}}$ is the H$_2$ formation rate coefficient on dust grain surfaces, and $R$ is the cloud radius.
The quantity ${\tau_{\rm R}}$ is simply a measure of the size of the cloud. It is the dust optical depth that a cloud would have if its density throughout were equal to its density in the atomic region. We may think of $\chi$ as a dimensionless measure of the intensity of the dissociating radiation; formally it is equal to the ratio of the rate at which LW photons are absorbed by dust grains to the rate at which they are absorbed by hydrogen molecules in a parcel of predominantly atomic gas in dissociation equilibrium in free-space. This is a measure of the strength of the radiation field because in strong radiation fields the gas contains very few molecules, so most LW photons are absorbed by dust and $\chi$ is large. In weak radiation fields the gas contains more molecules, which due to their large resonant cross-section dominate the absorption rate, making $\chi$ small.
Over the remainder of this paper we apply the model of Paper I to atomic-molecular complexes in galaxies. In § \[giantclouds\] we begin by considering giant clouds, which we may approximate as slabs, and in § \[finite\] we extend our treatment to clouds of finite size. In § \[obscomparison\] we compare the model predictions of the previous two sections to observations of atomic and molecular gas. Finally in § \[conclusion\] we summarize and discuss conclusions.
The Atomic Envelopes of Giant Clouds {#giantclouds}
====================================
In this section we specialize to the case of giant clouds, which we define as those for which $e^{{\tau_{\rm R}}}\gg 1$. For these clouds we show in that the dust optical depth from the cloud surface to the atomic-molecuar transition, ${\tau_{\rm HI}}$, is a function of $\chi$ alone. We therefore begin our analysis with an estimate of $\chi$.
The Normalized Radiation Field {#normrad}
------------------------------
Of the quantities that enter into the normalized radiation field $\chi$, ${f_{\rm diss}}$ is the most certain, because it is only a very weak function of the spectrum of the dissociating radiation. We therefore take it to have a constant value ${f_{\rm diss}}\approx 0.1$ independent of environment (@draine96 [@browning03]; ). Similarly, ${\sigma_{\rm d}}$ and ${\mathcal{R}}$ are functions of the properties of dust grains. These are both measures of the total surface area of dust grains mixed in the atomic shielding envelope around a molecular cloud; the former measures the area available for absorbing photons, while the latter measures the area available for catalyzing H$_2$ formation. There will of course be an additional dependence of these quantities on the optical and chemical properties of grains, but these effects likely provide only a small fraction of the total variation. To first order, therefore, we expect the ratio ${\sigma_{\rm d}}/{\mathcal{R}}$ to vary little with galactic environment, and we can simply adopt the value from the solar neighborhood. This is $$\label{sigmar}
\frac{{\sigma_{\rm d}}}{{\mathcal{R}}} = 3.2\times 10^{-5} \frac{\sigma_{\rm d, -21}}{{\mathcal{R}}_{-16.5}} \mbox{ s cm}^{-1}$$ where $\sigma_{\rm d,-21}={\sigma_{\rm d}}/10^{-21}$ cm$^2$, ${\mathcal{R}}_{-16.5}={\mathcal{R}}/10^{-16.5}$ cm$^3$ s$^{-1}$, and our best estimates for the solar neighborhood give $\sigma_{\rm d,-21} \approx {\mathcal{R}}_{-16.5} \approx 1$ [@draine96; @wolfire08].
Unfortunately, $n$ and ${E^{*}}_0$ are considerably harder to determine, since we cannot easily make direct measurements of the atomic density around a molecular cloud or the dissociating radiation field to which it is subjected, particularly for clouds in extragalactic space. (It is possible to determine these quantities for PDRs being produced by individual star clusters – see @smith00a and @heiner08a [@heiner08b] – but these methods are generally not able to determine mean radiation fields around giant clouds.) However, we can still gain considerable insight into the ratio ${E^{*}}_0/n$ that enters into $\chi$ if we realize that $n$ is not free to assume any value for a given ${E^{*}}_0$. The atomic gas in a galaxy generally comprises regions of both cold and warm gas (cold neutral medium and warm neutral medium, or CNM and WNM, respectively) in approximate pressure balance [e.g. @mckee77; @wolfire03]. Molecular clouds form in regions where the gas is primarily cold. This is because the the effective opacity to LW photons provided by the small population of molecules found in a given element of predominantly atomic gas varies as $n^2$, so the cold phase, due to its higher density, is far more effective at shielding from LW photons than the warm phase. Thus the $n$ we are concerned with is not the mean density of a galaxy’s atomic ISM, it is the density in the cold phase only. In the presumably dense gas in the vicinity of a molecular cloud, most of the mass is likely in the cold rather than the warm phase in any event.
We can estimate the CNM density by using the condition of pressure balance between the cold and warm phases. @wolfire03 show that, for a given ambient FUV radiation intensity $G_0$ (given in units of the @habing68 field, corresponding to a number density ${E^{*}}_0\approx 4.4 \times 10^{-4}$ LW photons cm$^{-3}$), ionization rate from EUV radiation and x-rays $\zeta_t$, abundance of dust and polycyclic aromatic hydrocarbons $Z_d$, and gas phase metal abundance $Z_g$, the minimum number density $n_{\rm min}$ at which CNM can exist in pressure balance with WNM is well-approximated by $$\label{nmin1}
n_{\rm min} \approx 31 G_0' \frac{Z_d'/Z_g'}{1 + 3.1 (G_0'
Z_d'/\zeta_t')^{0.365}}\mbox{ cm}^{-3},$$ where the primes denote quantities normalized to their values in the solar neighborhood. @wolfire03 obtain this expression by constructing a temperature-density relation, determined by balancing the rate of grain photoelectric heating against cooling by the fine structure lines of C<span style="font-variant:small-caps;">ii</span> and O<span style="font-variant:small-caps;">i</span>. Once they have constructed the $T-n$ curve, they identify the temperature at which the pressure is minimized. This is the warmest temperature at which the CNM can be in pressure balance, and thus the corresponding density is the lowest possible CNM density. The primary uncertainty in this expression arises from the abundance, size distribution, and reaction properties of polycyclic aromatic hydrocarbons (PAHs), but changes in PAH properties generally change $n_{\rm min}$ only at the factor of $\sim 2$ level (cf. Figure 8 of @wolfire03).
In a galaxy where young stars provide the dominant source of radiation and the IMF is constant, the FUV heating rate and the EUV/x-ray ionization rate are likely to be proportional to the star formation rate, and therefore to each other. We therefore assume that $\zeta_t'
= G_0'$. Furthermore, if the physics of dust formation does not vary strongly from galaxy to galaxy, then the dust and gas phase metal abundances are likely proportional to the total metallicity $Z$, so we adopt $Z_d'=Z_g'=Z'$. With these approximations the minimum CNM density becomes $$\label{nmin}
n_{\rm min} \approx 31 \frac{G_0'}{1 + 3.1 Z'^{0.365}}\mbox{ cm}^{-3}.$$ We caution at this point that both the assumptions that $\zeta_t' = G_0'$ and $Z_d'=Z_g'=Z'$ are unlikely to hold in elliptical galaxies, where young stars are not the dominant sources of EUV or x-ray radiation, and where the amount of dust per unit metallicity is known to be different than in spirals. Thus, equation (\[nmin\]) is unlikely to hold in ellipticals.
The CNM can exist in pressure balance at densities higher than $n_{\rm min}$, so we take the typical CNM density to be $$\label{ncnm}
n_{\rm CNM} = {\phi_{\rm CNM}}n_{\rm min}.$$ We adopt ${\phi_{\rm CNM}}\approx 3$ as our fiducial value, which gives a CNM density of $n_{\rm CNM}=22$ cm$^{-3}$ and (using @wolfire03’s $T-n$ relation, equation \[wolfiretn\] of this paper) a temperature of $T_{\rm CNM}=105$ K, consistent with observations of typical CNM properties in the solar neighborhood. (Near a GMC, we expect $G_0'\sim 10$ rather than $G_0'=1$, due to the proximity of sites of star formation – Wolfire, Hollenbach, & McKee 2008, in preparation – but this does not affect our results, since we only care about the ratio $G_0'/n$.) In practice ${\phi_{\rm CNM}}$ cannot be much larger than this, because pressure balance between the CNM and WNM is possible only over a limited range of CNM densities. If the CNM densities exceeds $n_{\rm min}$ by more than a factor of $\sim 10$ the CNM and WNM again cannot be in pressure balance because it is impossible for the warm phase to have a high enough pressure.
Using equation (\[sigmar\]) for ${\sigma_{\rm d}}/{\mathcal{R}}$ and equation (\[ncnm\]) for $n$, and noting that the LW photon number density in the solar neighborhood is roughly $7.5\times 10^{-4}$ cm$^{-3}$ (@draine78; ), we find a total estimate for the dimensionless radiation field strength $$\label{chieqn}
\chi = 2.3 \left(\frac{\sigma_{\rm d,-21}}{{\mathcal{R}}_{-16.5}}\right) \frac{1+3.1 Z'^{0.365}}{{\phi_{\rm CNM}}}.$$ Note that all explicit dependence on the dust properties, the radiation field, and the atomic gas density have cancelled out of this expression.
Dependence on the dust properties has dropped out for the simple reason explained above: ${\sigma_{\rm d}}$ and ${\mathcal{R}}$ are both measures of the dust surface area, so their ratio is nearly constant. We can understand the somewhat more subtle reason that dependence on radiation field and the atomic gas density cancel by examining the physics behind expression (\[nmin1\]). As noted already, the minimum possible density in the cold atomic phase of the ISM corresponds to the density and temperature at which the pressure reaches a local minimum. Because the dependence of the cooling rate on gas temperature is determined almost entirely by the quantum mechanical constants and element abundances that determine the shapes of the C<span style="font-variant:small-caps;">ii</span> and O<span style="font-variant:small-caps;">i</span> cooling curves, and the photoelectric heating rate is essentially independent of temperature, the temperature at which this minimum pressure occurs is nearly fixed at $\approx 240$ K, and does not depend on the background radiation field (c.f. equation 34 and Appendix C of @wolfire03). Thus the density minimum will simply be the density at which the temperature reaches $\approx 240$ K. Since the heating rate varies as $n {E^{*}}_0$ and the cooling rate as $n^2$, it immediately follows that the density at which a fixed temperature is reached varies as $n\propto {E^{*}}_0$. This explains why $n/{E^{*}}_0$ is nearly constant in the CNM. There is only a weak dependence on metallicity, which arises because the heating rate depends on the charge state of polycyclic aromatic hydrocarbons (PAHs), and this in turn depends weakly on metallicity.
Before moving on, we should note that we have neglected the role of internal radiation in determining where a cloud changes from atomic to molecular. This is justified because most stars that contribute significant amounts of dissociating radiation are born in molecular clouds, but they do not stay internal to those clouds for very long. Most dissociating photons come from massive stars born in clusters that burrow their way out of their parent molecular clouds via their winds and H<span style="font-variant:small-caps;">ii</span> regions in only a few Myr. Thus, most of the dissociating radiation to which a molecular cloud is subjected is delivered externally rather than internally, even if it comes from stars born in that cloud.
This does, however, raise another cautionary point. We have computed the CNM density based on an implicit assumption of pressure balance, and we must consider under what circumstances pressure balance might not hold. One situation in which gas might not reach pressure balance is if it is subjected to hydrodynamic perturbations such as supernova shocks that create rapid and substantial changes in pressure, pushing gas into the unstable regime of density and pressure. Such gas is subject to an instability in which pockets of stable CNM condense within it, leaving behind a lower density ambient medium that expands to become stable WNM [e.g. @audit05a]. Thus the typical gas density will be significantly different than the value $n_{\rm CNM}$ that we have estimated only if the time between successive shocks that drive gas into instability is small compared to the time required for this instability to operate. This is of order the cooling time scale, which @wolfire03 estimate to be $$t_{\rm cool} \approx 7.7 \left(\frac{T}{10^4\mbox{ K}}\right)^{1.2} \left(\frac{nT}{3000\mbox{ K cm}^{-3}}\right)^{-0.8}\mbox{ Myr},$$ where $n$ and $T$ are the gas number density and temperature. For typical CNM conditions near the Solar circle this is $\sim 40$ kyr, while for typical WNM conditions it is slightly under 10 Myr. Higher values of $G_0'$, as are expected near GMCs, reduce these to a few kyr and a few Myr, respectively. Thus perturbations that produce velocities ${\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <}
{\sim}\:$}}5$ km s$^{-1}$, i.e. not fast enough to induce shocks in the WNM, are essentially ineffective at driving the gas out of equilibrium, since any CNM gas they disturb will re-equilibrate very quickly. Stronger perturbations such as supernova blast waves that drive WNM unstable can keep substantial amounts of gas out of pressure balance only if they recur on time scales of a few Myr or less. For both supernova blast waves and the shocks induced by the ubiquitous turbulence in the atomic ISM the recurrence time is $\sim 10$ Myr, larger (although not hugely so) than the equilibration time [@wolfire03]. We can therefore conclude that the typical atomic envelope around a molecular cloud is likely to be close to pressure balance between CNM and WNM. Some envelopes will have been subjected to a strong shock in the last few Myr, and these may have CNM densities substantially different than $n_{\rm CNM}$, but they will be in the minority.
Alternately, one could consider a galaxy in which the average galactic environment is so extreme that no two-phase equilibrium is possible, i.e. the pressure is so low that only WNM is stable, or the pressure is so high that only CNM is stable. @wolfire03 and @schaye04 both find that a pressure so low that only WNM exists is consistent with vertical hydrostatic balance only in the very diffuse outer parts of galactic disks. For the Milky Way, @wolfire03 find that a CNM phase can exist everywhere the mean ISM density is ${\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle >}
{\sim}\:$}}0.2$ cm$^{-3}$, which is everywhere in the Milky Way inside $\sim 15$ kpc in Galactocentric radius; @schaye04 estimates that hydrostatic balance requires the existence of a CNM phase any time the local gas surface density exceeds $3-10$ ${M_{\odot}}$ pc$^{-2}$, which is true over a similar region. Thus, we can conclude that a pure WNM is unlikely to exist anywhere except in the far outer regions of galactic disks; in such regions, equation (\[chieqn\]) underestimates the dimensionless radiation field strength, and thus we will overestimate the molecular fraction. The converse possibility is a galaxy in which the pressure is so high that no WNM is present, only CNM; however, @wolfire03 find that that WNM can be present any time the mean ISM density ${\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <}
{\sim}\:$}}70$ cm$^{-3}$, unless the CNM is entirely confined by a surrounding hot ionized medium. The case where no WNM exists because $n{\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle >}
{\sim}\:$}}70$ cm$^{-3}$ corresponds to an entire galaxy whose mean density matches that of a typical molecular cloud in the Milky Way, and such conditions are only found in starburst systems where the molecular fraction is essentially unity. In such cases our models will provide only an upper limit to the amount of [H<span style="font-variant:small-caps;">i</span> ]{}present, but even then our model may apply if the atomic shielding layer occurs far from the galactic midplane where the density is lower. The latter possibility, of an ISM consisting solely of cold atomic gas and hot ionized gas, appears not be be realized in nature. Thus, in summary, we expect our two-phase model to apply everywhere in galaxies except in their far outer parts, where the density is so low that no cold phase exists, and at the midplanes of starburst galaxies, where the density is so high that no warm phase can exist.
The Shielding Column {#shieldcolumn}
--------------------
The normalized radiation field $\chi$ is the primary factor controlling the size of the atomic gas column that is needed to shield a molecular cloud against dissociation, and thus in determining fraction of the gas in a galaxy is molecular. This quantity in turn determines what fraction of a galaxy’s ISM is molecular, and therefore available for star formation, and what fraction is atomic, and thus without star formation. The invariance of $\chi$ across galactic environments has important consequences for atomic-molecular complexes. First, $\chi$ measures the relative importance of dust shielding and H$_2$ self-shielding, so our results that $\chi\sim 1$ across galaxies implies that dust- and self-shielding contribute nearly equally in essentially all galactic environments. We do not expect to find clouds where either dust- or molecular-shielding completely dominate except near strong sources of dissociating radiation where the atomic ISM is not in pressure balance or in regions of such low or high density that the atomic ISM does not have two phases.
Second, the dust optical depth through the [H<span style="font-variant:small-caps;">i</span> ]{}shielding layer, ${\tau_{\rm HI}}$, is primarily a function of $\chi$; the dependence on ${\tau_{\rm R}}$ arises from geometric effects, and is greater than order unity only if the molecular region inside a cloud is a small fraction of its size. (Strictly speaking this is the optical depth of the CNM, not of all the atomic gas. As noted above most of the atomic gas around a molecular cloud is probably CNM, but there could be significant amounts of WNM along the line of sight that is not associated with that cloud, and which does not contribute to shielding it.) This implies that the dust optical depth through the atomic envelopes of molecular clouds should be roughly constant across galactic environments, at least as long as this optical depth is not close to that of the entire atomic-molecular complex. The only significant variation will be a weak increase in ${\tau_{\rm HI}}$ with metallicity. Correspondingly, the total [H<span style="font-variant:small-caps;">i</span> ]{}gas column will decrease with metallicity to a power less than unity, since ${\tau_{\rm HI}}$ increases slightly with metallicity, but the column of [H<span style="font-variant:small-caps;">i</span> ]{}required to achieve this dust optical depth decreases with metallicity. To illustrate for these effects we solve for ${\tau_{\rm HI}}$ as a function of $Z$ in the limit ${\tau_{\rm R}}\rightarrow\infty$ using the formalism of , and plot the result in Figure \[taushieldlarge\].
As the plot shows, four our fiducial model ${\phi_{\rm CNM}}=3$, we predict that the [H<span style="font-variant:small-caps;">i</span> ]{}layer around a molecular cloud in the Milky Way, $Z'=1$, should have an absorption optical depth of ${\tau_{\rm HI}}=0.40$ to LW photons, corresponding to an [H<span style="font-variant:small-caps;">i</span> ]{}column $N_{\rm {H\textsc{i} }}=4.0\times 10^{20}$ cm$^{-2}$ (mass column density $\Sigma=4.5$ ${M_{\odot}}$ pc$^{-2}$), assuming a dust absorption cross section per H nucleus of ${\sigma_{\rm d}}=10^{-21}$ cm$^{-2}$. It is important to note that all of these values represent the absorption column on *one side* of a giant cloud. A 21-cm observation would detect the shielding column on both sides for a cloud exposed to the ISRF on both sides, so the detected [H<span style="font-variant:small-caps;">i</span> ]{}column would be *double* the values given in Figure \[taushieldlarge\]. As a shall see in § \[finite\], the column is somewhat larger for a cloud of finite size.
It is important to note that the shielding column we have calculated is somewhat different than the atomic-to-molecular transition column density reported for the Milky Way by @savage77 [$\log N(H)=20.7$] and for the LMC and SMC by @tumlinson02a [$\log N(H)\ge 21.3$ and $\ge 22$, respectively]. These values are the total column densities along pencil-beam lines of sight at which the fraction of the gas column in the form of H$_2$ reaches about $10\%$ of the total. In contrast, in the two-zone approximation we adopt in , we assume that the atomic-to-molecular transition is infinitely sharp, and under this approximation the shielding column we report is the column at which the gas goes from fully atomic to fully molecular. Were the transition truly infinitely sharp as we have approximated it to be, the ratio of H$_2$ to total column density would be zero at our computed shielding column. Comparing our theoretical shielding columns to the detailed numerical radiative transfer models we present in shows that in reality, for conditions typical of the Milky Way, the ratio of H$_2$ column to total column at our calculated transition column $N_{\rm HI}$ is roughly 20%. Since this is a factor of 2 larger than the 10% ratio used in the observationally-defined transition column, and in our simple model the H$_2$ fraction increases linearly with total column density once we pass our predicted transition point, we expect our shielding column to be a factor of $\sim 2$ larger than the values reported by @savage77 and @tumlinson02a We compare our model predictions to these data sets in more detail in § \[galobs\].
Using the extinction and absorption curves of @draine03a [@draine03b; @draine03c], the ratio of visual extinction to 1000 Å absorption is $A_V/{\tau_{\rm HI}}=0.48$ for @draine03a’s $R_V=4.0$ model, so the visual extinction corresponding to ${\tau_{\rm HI}}=0.40$ is $A_V=0.19$. Adopting the $R_V=5.5$ curve instead, appropriate for denser clouds, gives $A_V=0.28$, while $R_V=3.1$, for diffuse regions, gives $A_V=0.13$. Our estimates for the LW dust optical depth and visual extinction vary little with metallicity, changing by only a factor of $2.7$ for a metallicity ranging from $10^{-2} Z_{\odot}$ to $10^{0.5} Z_{\odot}$.
The variation between the curves with ${\phi_{\rm CNM}}=1,3,10$ show the full plausible range of variation in shielding column arising from our uncertainty about the true density in the atomic envelopes of molecular clouds. The ${\phi_{\rm CNM}}=1$ and $10$ curves are both within a factor of $2.6$ of the fiducial model, so this is an upper bound on our uncertainty. The actual error is likely to be smaller than this, since ${\phi_{\rm CNM}}=1$ and $10$ correspond to the extreme assumptions that the CNM assumes is minimum or maximum possible equilibrium densities.
We can obtain a quick approximation to the results shown in Figure \[taushieldlarge\] simply by noting that at solar metallicity our fiducial normalized radiation field is $\chi=3.1$, and we show in that for a giant cloud with $\chi<4.1$ (corresponding to $Z'<2.5$ for our fiducial parameters) the LW dust optical depth through the atomic shielding layer is $${\tau_{\rm HI}}= \frac{\psi}{4},$$ where $$\label{psidefn}
\psi=\chi\frac{2.5+\chi}{2.5+\chi e}.$$ The dust-adjusted radiation field $\psi$ is a function only of metallicity; for our fiducial parameters ${\phi_{\rm CNM}}=3$ and $\sigma_{\rm d,-21}/{\mathcal{R}}_{-16.5}=1$, and Milky Way metallicity, $Z'=1$, we obtain $\psi=1.6$. Moreover, the dependence on metallicity is weak: at $Z'=1/10$, $\psi=1.0$, while at $Z'=1/100$, $\psi=0.77$. Because $\psi$ depends only on metallicity, we can also express the characteristic [H<span style="font-variant:small-caps;">i</span> ]{}shielding column on one side of a giant cloud solely as a function of $Z'$: $$\begin{aligned}
\label{sigmalarge}
\Sigma_{\rm {H\textsc{i} }} & = & \frac{{\mu_{\rm H}}}{{\sigma_{\rm d}}} {\tau_{\rm HI}}(Z',\phi_{\chi}) \\
& = & 4.5\,{M_{\odot}}\mbox{ pc}^{-2} \frac{f(Z', \phi_{\chi})}{\sigma_{\rm 0,-21} Z'}\end{aligned}$$ where ${\mu_{\rm H}}=2.34 \times 10^{-24}$ g is the mean mass per hydrogen nucleus and $\sigma_{\rm 0,-21}$ is the dust absorption cross-section at Milky Way metallicity ($\log Z'=0$) in units of $10^{-21}$ cm$^2$. The function $f(Z',\phi_{\chi})$ is given by $$\begin{aligned}
f(Z',\phi_{\chi}) & = & 0.54 \left(0.32+Z'^{0.385}\right) \phi_{\chi}^{-1} \cdot {}
\nonumber \\
& & \qquad
\left(\frac{1.05 \phi_{\chi}+0.42+Z'^{0.385}}{0.39\phi_{\chi}+0.42+Z'^{0.385}}\right),\end{aligned}$$ where $$\phi_{\chi} \equiv \left(\frac{{\phi_{\rm CNM}}}{3}\right) \left( \frac{{\mathcal{R}}_{-16.5}}{\sigma_{\rm d,-21}}\right),$$ $f(1,1)=1$, and for our fiducial parameters $\phi_{\chi} = 1$. The numerical factors that appear in $f(Z',\phi_{\chi})$ are derived simply by substituting equation (\[chieqn\]) for $\chi$ into equation (\[psidefn\]) and thence into equation (\[sigmalarge\]). These equations, and therefore the numerical values in the function $f$, depend solely on microphysical constants that describe the properties of molecular hydrogen and the chemistry of its formation on grain surfaces (which set ${f_{\rm diss}}$ and ${\sigma_{\rm d}}/{\mathcal{R}}$) and the shapes of the C<span style="font-variant:small-caps;">ii</span> and O<span style="font-variant:small-caps;">i</span> cooling curves (which set the ratio ${E^{*}}_0/n$). We have therefore calculated the shielding column to good approximation solely in terms of microphysical constants.
The Atomic Envelopes of Finite Clouds {#finite}
=====================================
Formulation of the Problem
--------------------------
To account for the fact that clouds have finite sizes and column densities, and that these can be quite small in dwarf galaxies or other low-pressure environments, we must examine the second dimensionless number that characterizes H$_2$ formation and shielding: ${\tau_{\rm R}}=n{\sigma_{\rm d}}R$. Consider a cloud of known, fixed column density ${\Sigma_{\rm comp}}$. If atomic-molecular complexes were of uniform density then we could find ${\tau_{\rm R}}$ simply by multiplying ${\Sigma_{\rm comp}}$ by the dust cross section per unit mass ${\sigma_{\rm d}}/{\mu_{\rm H}}$, where ${\mu_{\rm H}}\approx 2.34\times 10^{-24}$ g is the mean mass per H nucleus. However, the atomic region is warmer and has a lower mean mass per particle than the molecular one, and thus has a correspondingly lower density. This reduces the dust optical depth through it. Since it is the density and dust optical depth through the atomic shielding layer that matters, we must estimate ${\tau_{\rm R}}$ using the value of $n$ appropriate for the atomic gas rather than the mean density in the complex. In other words, the quantity we want is $${\tau_{\rm R}}={n_{\rm CNM}}{\sigma_{\rm d}}R,$$ and we define $${\phi_{\rm mol}}\equiv \frac{n_{\rm mol}}{{n_{\rm CNM}}}$$ as the ratio of densities. Here $n_{\rm mol}$ and ${n_{\rm CNM}}$ are the number densities of hydrogen nuclei in the molecular and CNM phases of the ISM, respectively.
In the Milky Way, typical molecular cloud densities are $n_{\rm mol} \approx 100$ cm$^{-3}$ [@mckee07b], while observations of the giant [H<span style="font-variant:small-caps;">i</span> ]{}clouds around these molecular regions find typical densities ${n_{\rm CNM}}\approx 10$ cm$^{-3}$ [@elmegreen87], suggesting ${\phi_{\rm mol}}\approx 10$. We do not expect this ratio to vary strongly between galaxies, so we should generally find ${\phi_{\rm mol}}\approx 10$.
We can can make an independent argument for ${\phi_{\rm mol}}\approx 10$ by considering thermal pressure balance across the atomic-molecular interface. This argument only applies to gas near an atomic-molecular transition surface, which may or may not include the bulk of the gas in a cloud, but it does provide an estimate for the density ratio near the interface. Pressure balance requires that $$\label{phimoleqn}
{\phi_{\rm mol}}= 1.8 \frac{T_{\rm CNM}}{T_{\rm mol}},$$ where $T_{\rm mol}$ and $T_{\rm CNM}$ are the temperatures in the molecular and cold neutral atomic media, respectively, and the factor of $1.8$ accounts for the difference in mean number of particles per H nucleus in the two phases. Across a very wide range of galactic environments the temperature in the molecular phase of the ISM is $T_{\rm mol} \approx 10-20$ K, as a result of the balance between grain photoelectric heating and CO cooling. Since are interested in gas at the edge of the molecular region, we adopt $T_{\rm mol} = 20$ K as typical. (Temperatures are somewhat higher in starburst systems, but in these galaxies the molecular fraction is essentially unity in any event.) Using the model of @wolfire03 for the atomic medium, and making the approximations $Z'_d=Z'_g=Z'$ and $\zeta'_t=G'_0$ as in § \[normrad\], the density-temperature relation in the atomic gas is $$\label{wolfiretn}
{n_{\rm CNM}}\approx \frac{20\, G_0' {T_{\rm CNM,2}}^{-0.2} e^{1.5/{T_{\rm CNM,2}}}}{1+2.6 ({T_{\rm CNM,2}}^{1/2} Z')^{0.365}},$$ where ${T_{\rm CNM,2}}={T_{\rm CNM}}/(100\mbox{ K})$. Combining this with equations (\[nmin\]) and (\[ncnm\]) enables us to write an implicit equation for the CNM temperature in terms of ${\phi_{\rm CNM}}$ and $Z'$: $$\label{tcnmeqn}
\frac{20 \, {T_{\rm CNM,2}}^{-0.2} e^{1.5/{T_{\rm CNM,2}}}}{1+2.6 ({T_{\rm CNM,2}}^{1/2} Z')^{0.365}} = {\phi_{\rm CNM}}\frac{31}{1+3.1Z'^{0.365}}.$$ Substituting the solution to this equation in equation (\[phimoleqn\]) immediately gives us ${\phi_{\rm mol}}$, the ratio of the number densities of H nuclei in the molecular and CNM gas. For our fiducial ${\phi_{\rm CNM}}=3$, we find ${\phi_{\rm mol}}=9.6$ at $Z'=1$, varying by only a few percent for metallicities in the range $Z'=10^{-2}-10^{1}$. Given the encouraging agreement between this and the value ${\phi_{\rm mol}}\approx 10$ we find observationally, we adopt the value of ${\phi_{\rm mol}}$ given by equations (\[phimoleqn\]) and (\[tcnmeqn\]) as our standard value for the remainder of this work.
A spherical cloud that consists of a molecular core of number density $n_{\rm mol}$ and an outer atomic envelope of number density $n_{\rm CNM}$ has a mean column density $$\label{sigmacompeq}
{\Sigma_{\rm comp}}= \frac{4}{3} {\mu_{\rm H}}n_{\rm CNM} R \left[1+\left({\phi_{\rm mol}}-1\right){x_{\rm H_2}}^3\right],$$ where ${x_{\rm H_2}}$ is the fraction of the cloud radius at which it transitions from molecular to atomic, i.e. ${x_{\rm H_2}}=1$ corresponds to a cloud that is molecular throughout and ${x_{\rm H_2}}=0$ to one that is atomic throughout (see Paper I). It is convenient to rewrite this in terms of an optical depth $$\begin{aligned}
\label{taucompdefn}
{\tau_{\rm c}}& \equiv & \frac{3}{4}\left(\frac{{\Sigma_{\rm comp}}{\sigma_{\rm d}}}{{\mu_{\rm H}}}\right) \\
& \rightarrow & 0.067\, Z' \Sigma_{\rm comp,0},\end{aligned}$$ where $\Sigma_{\rm comp,0}={\Sigma_{\rm comp}}/(1\,{M_{\odot}}\mbox{ pc}^{-2})$ and the arrow in the second step indicates that we have used our fiducial ${\sigma_{\rm d}}=10^{-21} Z'$ cm$^2$. Equations (\[sigmacompeq\]) and (\[taucompdefn\]) imply $$\label{xmtaur}
{\tau_{\rm c}}= {\tau_{\rm R}}\left[1+\left({\phi_{\rm mol}}-1\right) {x_{\rm H_2}}^3\right].
$$ Note that neither ${\tau_{\rm R}}$ nor ${\tau_{\rm c}}$ is the true center-to-edge dust optical depth of the complex; ${\tau_{\rm R}}$ is the optical depth the complex would have if its density were $n_{\rm CNM}$ throughout, and ${\tau_{\rm c}}$ is the optical depth it would have if its atomic and molecular gas were mixed uniformly rather than spatially segregated.
We are now in a position to compute the shielding column and the atomic and molecular fractions in finite clouds. If we consider a complex of a given column density ${\Sigma_{\rm comp}}$ and metallicity $Z'$, and we take the dust opacity to be given by its Milky Way value adjusted for metallicity, ${\sigma_{\rm d}}= \sigma_{\rm d,MW} Z'$ with $\sigma_{\rm d,MW}=10^{-21}$ cm$^{-2}$, then equations (\[taucompdefn\]) and (\[xmtaur\]) give one constraint on the unknowns ${\tau_{\rm R}}$ and ${x_{\rm H_2}}$ from pressure balance between the atomic and molecular phases. Dissociation-formation equilibrium, as computed in , gives a second constraint. We show in how to compute the value of ${x_{\rm H_2}}$ for a given ${\tau_{\rm R}}$ and $\chi$: its value is given implicitly by the solution to equations (33) and (37), or (43) and (44), of that paper. Since we have already computed $\chi$ in terms of the metallicity (equation \[chieqn\]), a choice of ${\Sigma_{\rm comp}}$ and $Z'$ fully determine the two unknowns ${\tau_{\rm R}}$ and ${x_{\rm H_2}}$.
Numerical Solution
------------------
We can either solve this system of nonlinear algebraic equations numerically, or approximate the solution analytically. We first show the results of a numerical calculation for a variety of values of ${\Sigma_{\rm comp}}$ and $Z'$ in Figure \[finitecloud\]. Rather than giving ${\tau_{\rm R}}$ and ${x_{\rm H_2}}$ directly, which are not particularly interesting because we cannot measure them directly, we plot three derived quantities of interest which in at least some circumstances we can observe: the LW optical depth and visual extinction from the cloud surface to the atomic-molecular transition surface along a radial path, the [H<span style="font-variant:small-caps;">i</span> ]{}column averaged over the entire cloud, and the total H$_2$ mass fraction over the entire cloud.
It is important to point out that the value of $A_V$ we report is measured differently than the [H<span style="font-variant:small-caps;">i</span> ]{}column density $\Sigma_{\rm HI}$, or than the column densities we will use in § \[obscomparison\]. The column density is measured by averaging the mass per unit area over the entire complex, while $A_V$ is measured along a single pencil beam from the surface of the cloud to the atomic-molecular transition along a radial trajectory. The former quantity is more analogous to what is measured in an observation using a telescope beam that only marginally resolves or does not resolve a complex, while the latter is more closely analogous to a measurement of the extinction of a background point source through a cloud. We also caution that, for reasons we discuss in § \[geomerror\], our predictions are only accurate for molecular mass fractions $>1/2$. (This is in the worst case of very low metallicity and intermediate ${\phi_{\rm CNM}}$; our accuracy range expands as metallicity increases toward solar and as ${\phi_{\rm CNM}}$ gets smaller or larger than $3$.) Below this limit our calculations yield only upper limits on the molecular fraction, not firm predictions. This confidence limit is shown in the Figure \[finitecloud\].
The plots immediately yield a number of interesting results. First, our prediction of nearly constant $A_V$ through the atomic shielding layers around molecular clouds continues to hold whenever there is a significant molecular fraction, even for finite clouds. Our prediction of a characteristic $A_V\approx 0.2$ through atomic shielding envelopes of molecular clouds therefore continues to apply.
We also find that there is a saturation in the [H<span style="font-variant:small-caps;">i</span> ]{}column density at roughly $6$ ${M_{\odot}}$ pc$^{-2}$ for solar metallicity, which rises by a factor of a somewhat less then ten for every decade by which the metallicity declines. The [H<span style="font-variant:small-caps;">i</span> ]{}column saturates simply because once ${\Sigma_{\rm comp}}$ is large enough, the cloud column densities become so large that they are effectively in the infinite cloud limit. At this point the shielding column is geometry-independent, and is determined solely by the normalized strength of the radiation field, $\chi$, a value that does not vary much from galaxy to galaxy. Once ${\Sigma_{\rm comp}}$ is sufficiently large to put a complex in the large cloud limit, adding additional mass simply increases the size of the shielded molecular layer, so the H$_2$ fraction just rises smoothly. The saturation value of $6$ ${M_{\odot}}$ pc$^{-2}$ at solar metallicity is set by a combination of the fundamental constants describing H$_2$ formation and dissociation, the shape of the C<span style="font-variant:small-caps;">ii</span> and O<span style="font-variant:small-caps;">i</span> cooling curves (which determine the CNM density and temperature), and the properties of interstellar dust grains, which set the ratio ${\sigma_{\rm d}}/{\mathcal{R}}$.
Geometric Uncertainties for Finite Clouds {#geomerror}
-----------------------------------------
In § 4.7 of we show that our method for determining the molecular abundance in finite clouds suffers from a systematic uncertainty arising from our imperfect knowledge of the opacity along rays that pass through the atomic envelope of a cloud. For our fiducial model we take the opacity along these rays due to molecules mixed into the atomic gas to be set by the value of the dissociation radiation field at the surface of the zone where molecules dominate the opacity. Our results depend on this approximation very little except at low molecular volume fraction, ${x_{\rm H_2}}^3 {\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <}
{\sim}\:$}}0.2$; in that case the uncertainty about this approximation means that our model enables us to predict only an upper limit on the molecular fraction, not an exact value.
In this paper we are concerned with molecular mass rather than volume fractions, so we must quantify that uncertainty. To do so, we proceed as in § 4.7 of : we adopt the opposite assumption, that opacity along rays passing through the region of the cloud where molecules dominate the opacity but still constitute a small fraction of all H nuclei is infinite. We then repeat the calculations of § \[finite\] following this assumption: for a given ${\Sigma_{\rm comp}}$ and $Z'$, rather than solve the system of equations formed by equation (\[xmtaur\]) of this paper and equations (33) and (37) or (43) and (44) of , we instead solve equation (\[xmtaur\]) together with equations (69) and (70) or (71) of . Doing so gives a lower bound on the molecular content for a given ${\Sigma_{\rm comp}}$ and $Z'$. By comparing the results in this case to our fiducial calculations as presented in § \[finite\], we obtain an estimate of the uncertainty of our results.
In Figure \[hocomparison\] we show the results of this exercise. On the $y$-axis we show the ratio of the H$_2$ mass fraction predicted using our maximum opacity assumption, which represents the minimum possible molecular content, divided by the value produced by our fiducial model, which represents the maximum. This gives an estimate of our uncertainty. The $x$-axis indicates the H$_2$ mass fraction predicted using the fiducial assumption we make elsewhere in the paper. As the plot shows, the two calculations differ most at low $Z'$ and intermediate ${\phi_{\rm CNM}}$. In this case the calculations differ by a factor of a few for H$_2$ mass fractions below $\sim 0.5$. If we adopt a factor of $3$ as an accuracy goal, this means that for cases where we predict an H$_2$ mass fraction below $0.5$, and at low metallicity, our predictions should be taken only as upper limits. For solar metallicity or higher our confidence range extends down to molecular mass fractions around $0.4$, and we attain upper limits below this.
Analytic Approximation
----------------------
We can gain additional insight into the behavior of the solution by constructing an analytic approximation. The ratio of the molecular mass $M_{\rm H_2}$ to the total complex mass $M$ is $$\label{molfrac}
{f_{\rm H_2}}\equiv \frac{M_{\rm H_2}}{M} = \frac{{\phi_{\rm mol}}{x_{\rm H_2}}^3}{1+\left({\phi_{\rm mol}}-1\right){x_{\rm H_2}}^3}.$$ We wish to obtain an approximation for this in terms of the known quantities $\psi$ (given in terms of metallicity by equations \[chieqn\] and \[psidefn\]) and ${\tau_{\rm c}}$ (given in terms of complex column density by equation \[taucompdefn\]). We show in that for $\psi{\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <}
{\sim}\:$}}3$ and molecular volume fractions ${x_{\rm H_2}}^3 {\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle >}
{\sim}\:$}}0.15$, a range in parameter space that includes most of our models for realistic cloud parameters, the molecular volume is well-approximated by $$\label{xmapprox}
{x_{\rm H_2}}^3 \approx 1 - \frac{3\psi}{4{\tau_{\rm R\Delta}}},$$ where for convenience we have defined $${\tau_{\rm R\Delta}}\equiv {\tau_{\rm R}}+ a\psi,$$ and $a=0.2$ is a numerical parameter that is optimized for agreement between the approximate and numerical solutions. Substituting this approximation into the condition for pressure balance, equation (\[xmtaur\]), gives $$\label{taucapproxeqn}
{\tau_{\rm c}}= {\tau_{\rm R}}\left[{\phi_{\rm mol}}+\frac{3\psi}{4{\tau_{\rm R\Delta}}}\left(1-{\phi_{\rm mol}}\right)\right].$$ As in , our approach to obtaining an analytic solution is to perform a series expansion in $a$. We therefore define $${\tau_{\rm c\Delta}}\equiv {\tau_{\rm c}}\left(1+\frac{a\psi}{{\tau_{\rm R}}}\right),$$ so that ${\tau_{\rm c\Delta}}/{\tau_{\rm c}}= {\tau_{\rm R\Delta}}/{\tau_{\rm R}}$. Using this definition of ${\tau_{\rm c\Delta}}$ together with equation (\[taucapproxeqn\]) for ${\tau_{\rm c}}$ implies that $$\label{taucdeqn}
{\tau_{\rm c\Delta}}= {\phi_{\rm mol}}\left({\tau_{\rm R\Delta}}- \frac{3}{4}\psi\right) + \frac{3}{4}\psi.$$ If we now use our approximation (\[xmapprox\]) and rewrite the result in terms of ${\tau_{\rm c\Delta}}$ using equation (\[taucdeqn\]), we obtain $$\label{mh2taucd}
{f_{\rm H_2}}= 1 - \frac{3\psi}{4 {\tau_{\rm c\Delta}}}.$$ We must now express ${\tau_{\rm c\Delta}}$ in terms of $\psi$, ${\tau_{\rm c}}$, and $a$ alone. Thus $$\begin{aligned}
{\tau_{\rm c\Delta}}& = & {\tau_{\rm c}}+ a\psi\left(\frac{{\tau_{\rm c}}}{{\tau_{\rm R}}}\right) \\
\label{taucdapprox}
& = & {\tau_{\rm c}}+ a\psi\left(\frac{{\tau_{\rm c\Delta}}}{{\tau_{\rm R\Delta}}}\right).\end{aligned}$$ The second term on the RHS still involves the unknowns ${\tau_{\rm c\Delta}}$ and ${\tau_{\rm R\Delta}}$, but because they are already multiplied by $a$ we now need only determine them to zeroth order in $a$. Solviing equation (\[taucdeqn\]) for ${\tau_{\rm R\Delta}}$ gives $$\begin{aligned}
\frac{{\tau_{\rm R\Delta}}}{{\tau_{\rm c\Delta}}} & = & \frac{1}{{\phi_{\rm mol}}}+\left(1-\frac{1}{{\phi_{\rm mol}}}\right)\frac{3\psi}{4{\tau_{\rm c\Delta}}} \\
& \approx & \frac{1}{{\phi_{\rm mol}}}+\left(1-\frac{1}{{\phi_{\rm mol}}}\right)\frac{3\psi}{4{\tau_{\rm c}}},
\label{taucdapprox1}\end{aligned}$$ where in the second step we have dropped a term of order $a$ to obtain an expression that is accurate to zeroth order in $a$. Substituting this into equation (\[taucdapprox\]), and thence into equation (\[mh2taucd\]), gives our final expression for the molecular mass fraction, accurate to first order in $a$: $$\label{molfracanalyt}
{f_{\rm H_2}}= 1 - \frac{3\psi}{4{\tau_{\rm c}}} \left[1 + \frac{4a\psi{\phi_{\rm mol}}}{4{\tau_{\rm c}}+3({\phi_{\rm mol}}-1)\psi}\right]^{-1}.$$ Comparison of this approximate expression with the numerical solution illustrated in Figure \[finitecloud\] shows that for our fiducial ${\phi_{\rm CNM}}=3$ and metallicities from $Z'=10^{-2}-10$, it is accurate to better than 30% whenever the approximation analytic solution gives ${f_{\rm H_2}}> 0.25$, but that it goes to zero too sharply at low molecular fraction. We can improve the approximation by forcing the H$_2$ fraction to approach zero smoothly rather than sharply at low column density. Experimentation shows that the expression $${f_{\rm HI}}^{-3} = 1+\left\{\left(\frac{4{\tau_{\rm c}}}{3\psi}\right)\left[1 + \frac{4a\psi{\phi_{\rm mol}}}{4{\tau_{\rm c}}+3({\phi_{\rm mol}}-1)\psi}\right]\right\}^{3}
\label{hifracanalyt}$$ matches the numerical result for ${f_{\rm HI}}\equiv M_{\rm HI}/M$ for ${\phi_{\rm CNM}}=3$ to better than 20% for all $Z'<10$ regardless of the value of ${f_{\rm HI}}$. (However note that, as we show in § \[geomerror\], for ${f_{\rm H_2}}{\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <}
{\sim}\:$}}1/2$ our estimate of the molecular content is only an upper limit, and this is true of equation \[hifracanalyt\] as well.) Using equation (\[taucompdefn\]) to replace ${\tau_{\rm c}}$ with ${\Sigma_{\rm comp}}$, and substituting in our fiducial values ${\phi_{\rm CNM}}=5$, $a=0.2$, and ${\sigma_{\rm d}}=10^{-21}Z'$ cm$^{-2}$, equation (\[hifracanalyt\]) becomes $$\label{hifracanalyt1}
{f_{\rm HI}}\rightarrow \left[1+\left(\frac{s}{11}\right)^3\left(\frac{125+s}{96+s}\right)^3\right]^{-1/3}$$ where $$s \equiv \frac{\Sigma_{\rm comp,0} Z'}{\psi}$$ and $\Sigma_{\rm comp,0} = {\Sigma_{\rm comp}}/(1\,{M_{\odot}}\mbox{ pc}^{-2})$. Note that our result indicates that to good approximation the molecular content of an atomic-molecular complex depends only on the combination of input parameters $Z'\Sigma_{\rm comp,0}/\psi$; the numerator $Z'\Sigma_{\rm comp,0}$ is simply the dust column density of the complex up to a scaling factor, while the denominator $\psi$ is the dimensionless radiation field, which equations (\[chieqn\]) and (\[psidefn\]) give solely as a function of metallicity.
From (\[hifracanalyt1\]), it also immediately follows that the H$_2$ to [H<span style="font-variant:small-caps;">i</span> ]{}ratio ${R_{\rm H_2}}\equiv {f_{\rm H_2}}/{f_{\rm HI}}$ is $${R_{\rm H_2}}\approx \left[1+\left(\frac{s}{11}\right)^3\left(\frac{125+s}{96+s}\right)^3\right]^{1/3}-1.$$ For ${R_{\rm H_2}}>1$, which is the regime for which our models apply with high confidence, we can use an even simpler expression $${R_{\rm H_2}}\approx 0.08s = 0.08 \frac{\Sigma_{\rm comp,0} Z'}{\psi},$$ where is accurate to $\sim 30\%$.
Similarly, we show in that the dust absorption optical depth through the atomic layer for a finite cloud is well-approximated by $${\tau_{\rm HI}}\approx \frac{\psi}{4}\left[\frac{1}{1 - (a'/4)(\psi/{\tau_{\rm R}})}\right],$$ where $a'=\frac{3}{2}-4a=0.7$. If we treat $a'/4$ as a small parameter and perform a series expansion around it, then we need only approximate ${\tau_{\rm R}}$ to zeroth order in $a$. We can do this simply by using equation (\[taucdapprox1\]) with $a=0$, which allows us to set ${\tau_{\rm R}}={\tau_{\rm R\Delta}}$ and ${\tau_{\rm c}}={\tau_{\rm c\Delta}}$. Thus to zeroth order in $a$ we have $${\tau_{\rm R}}= {\tau_{\rm R\Delta}}\left[\frac{1}{{\phi_{\rm mol}}}+\left(1-\frac{1}{{\phi_{\rm mol}}}\right)\frac{3\psi}{4{\tau_{\rm c}}}\right],$$ and to first order in $a$ or $a'$ we have $$\label{tauhianalyt}
{\tau_{\rm HI}}= \frac{\psi}{4}\left[1-\frac{a'\psi{\phi_{\rm mol}}}{4{\tau_{\rm c}}+3({\phi_{\rm mol}}-1)\psi}\right]^{-1}.$$ As with approximation (\[molfracanalyt\]) for the molecular mass fraction, this expression works well whenever the molecular fraction is not too low, and may be improved by forcing the optical depth to approach the total cloud optical depth smoothly when the column density becomes low. The expression $${\tau_{\rm HI}}^{-2} = {\tau_{\rm c}}^{-2}+\frac{16}{\psi^2}\left[1-\frac{a'\psi{\phi_{\rm mol}}}{4{\tau_{\rm c}}+3({\phi_{\rm mol}}-1)\psi}\right]^{2}$$ is accurate to better than 35% for all $Z'<10$, and to better than 25% for $Z'<1$.
It is also convenient to invert our analytic expressions to determine column density as a function of molecular content and metallicity. The term $(125+s)/(96+s)$ in equation (\[hifracanalyt1\]) is generally close to unity except at extremely high column densities, so if we neglect second-order corrections to the difference between this term and unity, we can solve equation (\[hifracanalyt1\]) for ${f_{\rm HI}}$ to obtain $$\label{sinveqn}
\frac{\Sigma_{\rm comp,0} Z'}{\psi} \rightarrow 11 \left({f_{\rm HI}}^{-3}-1\right)^{1/3} \frac{8.7+ \left({f_{\rm HI}}^{-3}-1\right)^{1/3}}{11+ \left({f_{\rm HI}}^{-3}-1\right)^{1/3}}$$ This expression matches the numerical solution at the $\sim 20\%$ level for ${f_{\rm HI}}<0.75$. Note that this result implies that the column density ${\Sigma_{\rm comp}}$ at which a given [H<span style="font-variant:small-caps;">i</span> ]{}fraction is reached depends on metallicity both explicitly through the $Z'$ term in the numerator, representing the effect of metallicity on dust content, and implicitly through $\psi$ (equations \[chieqn\] and \[psidefn\]), representing the effect of metallicity on the ratio of radiation intensity to CNM density. Since $\psi$ is an increasing function of metallicity, for a given molecular fraction ${\Sigma_{\rm comp}}$ has a weaker than linear dependence on metallicity. For example, evaluating (\[sinveqn\]) with ${f_{\rm HI}}=0.5$ at solar metallicity $Z'=1$ indicates that we expect the gas to be half molecular for complexes with ${\Sigma_{\rm comp}}=27$ ${M_{\odot}}$ pc$^{-2}$. (The exact numerical solution is ${\Sigma_{\rm comp}}=25.5$ ${M_{\odot}}$ pc$^{-2}$.) At one-third solar metallicity, $Z'=1/3$, half molecular content is reached at ${\Sigma_{\rm comp}}=67$ ${M_{\odot}}$ pc$^{-2}$ (using equation \[sinveqn\]; numerically ${\Sigma_{\rm comp}}=55.3$ ${M_{\odot}}$ pc$^{-2}$), somewhat less than a factor of $3$ higher.
Comparison to Observations {#obscomparison}
==========================
Our model makes strong predictions for the relative fractions of [H<span style="font-variant:small-caps;">i</span> ]{}and H$_2$ as a function of total surface density and metallicity, and in this section we compare these results to a variety of galactic and extragalactic observations.
Extragalactic Observations
--------------------------
### Data Sets
[lccc]{} Solar (Milky Way) & 8.76 & B & 7\
DDO154 & 7.67 & L & 3\
HOI & 7.54 & L & 6\
HOII & 7.68 & L & 6\
IC10 & 8.26 & B & 1\
IC2574 & 7.94 & L & 6\
NGC0598 & 8.49 & B & 5\
NGC0628 & 8.51 & L & 5\
NGC0925 & 8.32 & L & 5\
NGC2403 & 8.39 & L & 5\
NGC2841 & 8.81 & L & 5\
NGC2976 & 8.30 & L & 8\
NGC3077 & 8.64 & L & 8\
NGC3184 & 8.72 & L & 5\
NGC3198 & 8.42 & L & 5\
NGC3351 & 8.80 & L & 5\
NGC3521 & 8.49 & BL & 5\
NGC3627 & 9.25 & BL & 4\
NGC4214 & 8.22 & L & 2\
NGC4321 & 8.71 & B & 5\
NGC4414 & ... & B & ...\
NGC4449 & 8.31 & L & 2\
NGC4501 & 8.78 & B & 5\
NGC4736 & 8.50 & BL & 5\
NGC5033 & 8.68 & B & 5\
NGC5055 & 8.68 & BL & 5\
NGC5194 & 8.75 & BL & 5\
NGC5457 & 8.44 & B & 5\
NGC6946 & 8.53 & L & 5\
NGC7331 & 8.48 & BL & 5\
NGC7793 & 8.34 & L & 5\
We use three extragalactic data sets for comparison to our models. Two are recent surveys that have mapped nearby galaxies in 21 cm [H<span style="font-variant:small-caps;">i</span> ]{}and 2.6 mm CO($1\rightarrow0$) emission at overlapping positions, and therefore provide an ideal laboratory in which to test our model. The first of these is the work of , , and , who report [H<span style="font-variant:small-caps;">i</span> ]{}and H$_2$ surface densities on a pixel-by-pixel basis in 14 nearby galaxies (including the Milky Way). The H$_2$ surface densities are inferred from CO observations taken as part of the BIMA SONG survey [@regan01; @helfer03], while the [H<span style="font-variant:small-caps;">i</span> ]{}observations are from the VLA. and supplement these data with stellar surface density measurements from 2MASS [@jarrett03], which together with the equations given in can be used to derive a pressure in each pixel if one assumes that the gas in in hydrostatic balance, that the stellar scale height greatly exceeds the gas scale height, and that the gas velocity dispersion has a known value. The galaxies in the sample are all molecule-rich spirals with metallicities within $0.5$ dex of solar.
The second extragalactic data set we use is compiled by , who a combine [H<span style="font-variant:small-caps;">i</span> ]{}measurements from the THINGS survey with CO data partly taken from BIMA SONG and partly from the ongoing HERACLES survey. The authors also include 2MASS stellar surface densities in their compilation, and give mean pressures. Unlike the sample the data reported are averages over galactocentric rings rather than individual pixels, although point-by-point maps at sub-kpc resolution are in preparation (F. Walter, 2008, private communication). The sample includes 23 galaxies, of which roughly half are large spirals and roughly half are low-mass, H<span style="font-variant:small-caps;">i</span>-dominated dwarfs. The galaxies in the data set partly overlap with those in the sample of , but extend over a wider range of metallicities and molecule fractions.
We summarize the galaxies in the samples in Table \[metaltab\]. We also report metallicities for each galaxy where these are available in the literature. We do not include NGC4414 in the analysis, because no gas phase metallicity is available for it. In the comparison that follows, we neglect the presence of metallicity gradients within these galaxies, because gradients are only available for some of them. On top of this, we note that the metallicities themselves are probably uncertain at levels from hundreths to tenths of a dex, depending on the galaxy and the analysis technique used, which adds additional scatter on top of that already introduced by our neglect of metallicity gradients.
In both data sets the uncertainty in the [H<span style="font-variant:small-caps;">i</span> ]{}column densities is generally $\sim 10\%$. In the CO data the formal uncertainties are generally $\sim 20-30\%$, but the dominant uncertainty is probably systematic: the X factor used to convert observed CO luminosities into H$_2$ column densities. This is uncertain at the factor of $\sim 2$ level [@blitz07a], and almost certainly varies with metallicity [e.g. @bolatto08a]. Thus, although for clarity we will suppress error bars in the plots that follow, recall that the molecular data is uncertain at the factor of $\sim 2$ level.
The third data set we use, in § \[smcsec\], is the [*Spitzer*]{} Survey of the Small Magellanic Cloud (S$^3$MC) [@bolatto07a; @leroy07a]. This survey differs from the SONG and THINGS data sets in that those surveys infer the presence of H$_2$ via CO emission, whereas S$^3$MC measures molecular hydrogen using measurements of dust from [*Spitzer*]{} combined with [H<span style="font-variant:small-caps;">i</span> ]{}measurements by @stanimirovic99a and @stanimirovic04a. The basic idea behind the technique is that one determines the dust-to-gas ratio in a low-column density region where molecules are thought to contribute negligibly to the total column. Then by comparing the [H<span style="font-variant:small-caps;">i</span> ]{}and dust column density maps, one can infer the presence of H$_2$ in pixels where the total dust column exceeds what one would expect for the observed [H<span style="font-variant:small-caps;">i</span> ]{}column and a fixed dust-to-gas ratio. The reason for using this technique is that, at the low metallicity of the SMC ($\log[\mbox{O}/\mbox{H}]+12=8.0$, @dufour84a, i.e. $0.76$ dex below Solar), CO may cease to be a reliable tracer of molecular gas. The SMC represents the lowest metallicity galaxy for which we have H$_2$ detections rather than upper limits; the sample does not contain any galaxies with metallicities as low as the SMC, and no CO was detected in any of the galaxies with metallicities comparable to or lower than the SMC. Thus, the SMC represents a unique opportunity to test our models at very low metallicity.
Before comparing to these data sets, it is worth commenting briefly on one additional extragalactic data set to which we will not compare our models: observations of H$_2$ column densities along sightlines in the LMC and SMC using [*FUSE*]{} [@tumlinson02a]. We do not use this data set for comparison because it includes only sightlines with low column densities that are strongly dominated by atomic gas; the highest reported H$_2$ fraction is below 10%. Sightlines with significantly higher molecular content than this absorb too much background starlight to allow [*FUSE*]{} to make a reliable measurement of the H$_2$ column. The low column densities of the clouds that [*FUSE*]{} can observe in these galaxies place almost all of them into the regime where our theory yields only upper limits. Those limits are generally consistent with the data, but the comparison is not particularly illuminating.
### Column Density and Metallicity Dependence {#sigmazdepend}
In this section we compare our models predictions of the [H<span style="font-variant:small-caps;">i</span> ]{}and H$_2$ mass fractions as a function of total column density and metallicity to our two extragalactic data sets. To perform this comparison, for convenience we first break the sample into four metallicity bins: $-1.25 < \log Z' < -0.5$, $-0.5 < \log Z' < -0.25$, $-0.25 < \log Z' < 0$, and $0 < \log Z' < 0.5$, where $Z' = Z / Z_{\odot}$ and we adopt $\log(\mbox{O}/\mbox{H})+12 = 8.76$ as the value corresponding to solar metallicity [@caffau08]. These bins roughly evenly divide the data; we do not include the galaxies for which metallicities are not available in the literature.
Next we must consider how finite spatial resolution will affect our comparison. Given the beam sizes in the observed data, which range from $\sim 0.3$ kpc for the nearest galaxies to a few kpc for the most distant, each pixel (for ) or ring (for ) in the observed data set is likely to contain multiple atomic-molecular complexes. The atomic and molecular observations are convolved to the same resolution, so this does not bias the measurement of the atomic to molecular ratio, although it does mean that this ratio is measured over an averaging scale set by the beam size. Since individual atomic molecular complexes presumably represent peaks of the galactic column density, however, the total ([H<span style="font-variant:small-caps;">i</span> ]{}+ H$_2$) observed gas surface density ${\Sigma_{\rm obs}}$ reported for each pixel or ring represents only a lower limit on ${\Sigma_{\rm comp}}$, the total surface density of individual complexes. In our model the fraction of the cloud in molecular form is a strictly increasing function of ${\Sigma_{\rm comp}}$ (and of metallicity $Z$), while the atomic fraction is a strictly decreasing function. Because we have only the observed column density ${\Sigma_{\rm obs}}$ available, not the column density of an individual complex ${\Sigma_{\rm comp}}$, we have no choice but to take ${\Sigma_{\rm comp}}= {\Sigma_{\rm obs}}$ as our best estimate. The error in this approximation probably ranges from tens of percent for nearby galaxies where the beam size is not much larger than the size of a complex up to an order of magnitude or more for ring-averages in distant galaxies. As a result, though, we expect to overestimate the atomic fraction and underestimate the molecular fraction. Physically we may think of this as a clumping effect: the clumpier the gas is, the better able it is to shield itself against dissociating radiation. Since our observations smooth over scales larger than the characteristic gas clumping scale, we will miss this effect.
There are also two effects which go in the other direction, however. First, as noted above, shielding comes primarily from cold atomic gas, not warm gas. Although most of the gas in the immediate vicinity of a single molecular region is probably cold, observations that average over many molecular cloud complexes are likely to include a fair amount of WNM gas as well. Since we have only the total [H<span style="font-variant:small-caps;">i</span> ]{}column densities including both phases, we are prone to overestimate ${\Sigma_{\rm comp}}$ and therefore the molecular fraction, because the WNM raises the [H<span style="font-variant:small-caps;">i</span> ]{}column but does not provide much shielding. This effect is not important at moderate to high column densities, where the atomic gas does not dominate the total mass budget, but it could be significant at lower densities.
Second, there may be gas along our line of sight through a galaxy that is not associated with an atomic-molecular complex along that line of sight. This effect will increase in severity as the galaxy comes close to edge-on, since this will increase the path length of our line of sight through the galaxy. As with WNM gas, this extra material contributes to ${\Sigma_{\rm obs}}$ but not to the complex surface density ${\Sigma_{\rm comp}}$, and it therefore leads us to overestimate the molecular fraction. We can perform a very simple calculation to estimate the size of this effect. Consider a simple self-gravitating gas disk characterized by the standard vertical density profile $n \propto {\mbox{sech}}^2(0.88 z/h)$, where $h$ is the half-height of the gas. At some point in this disk is an atomic-molecular complex, centered at the midplane. Since complexes are found preferentially at the midplane, unless the galaxy is very close to edge-on then we need not consider the possibility of our line of sight intersecting multiple independent complexes. Since the complex formed from a large-scale gravitational instability in the disk, its characteristic size is $\sim h$, and we therefore consider gas to be “associated" with the complex if it is within a distance $h$ of it in the plane of the disk. Suppose this galaxy has an inclination $i$. Making the worst-case assumption, that there is no density enhancement due to the presence of the self-gravitating complex, we can then compute what the fraction of the gas we see along our line of sight is not associated with it. This is simply $$\frac{\int_{h\cot i}^{\infty} {\mbox{sech}}^2 (0.88 z/h) \, dz}{\int_0^{\infty} {\mbox{sech}}^2 (0.88 z/h) \,dz}
= 1 - \tanh(1.13 \cot i).$$ This is less than $0.5$ for all inclinations less than $64^{\circ}$. Only a handful of the galaxies in the SONG and THINGS surveys have inclinations larger than this, so even in the worst case scenario where complexes do not represent any enhancement of the gas density, we expect non-associated gas to produce an error smaller than a factor of $\sim 2$ for the great majority of the galaxies to which we are comparing.
Given the limitations imposed by finite resolution, we proceed as follows. Using the method described in § \[finite\], we compute the molecular mass fraction $$f_{\rm H_2}({\Sigma_{\rm comp}}, Z) \equiv \frac{M_{\rm H_2}}{M},$$ as a function of complex column density ${\Sigma_{\rm comp}}$ and metallicity $Z$. (For these and all subsequent predictions we use our fiducial value of ${\phi_{\rm CNM}}=3$.) We then generate predicted [H<span style="font-variant:small-caps;">i</span> ]{}and H$_2$ column densities for each metallicity bin and each observed column density $\Sigma$ via $$\begin{aligned}
\label{h2mfraceqn}
\Sigma_{\rm H_2,predicted} & = & f_{\rm H_2}({\Sigma_{\rm obs}}, Z_{\rm min}) {\Sigma_{\rm obs}}\\
\Sigma_{\rm HI,predicted} & = & [1-f_{\rm H_2}({\Sigma_{\rm obs}}, Z_{\rm min})] {\Sigma_{\rm obs}},\end{aligned}$$ where $Z_{\rm min}$ is the minimum metallicity for that bin. Since $f_{\rm H_2}({\Sigma_{\rm comp}}, Z)$ is a strictly increasing function of ${\Sigma_{\rm comp}}$ and $Z$, and we know that ${\Sigma_{\rm obs}}< {\Sigma_{\rm comp}}$ and $Z_{\rm min} < Z$, we expect $f_{\rm H_2}({\Sigma_{\rm obs}}, Z_{\rm min})$ to be a lower limit on the true molecular fraction. We therefore expect $\Sigma_{\rm H_2,predicted}$ to be a lower limit on the observed data, and $\Sigma_{\rm HI,predicted}$ to be an upper limit. The possible exception to this statement is at low column densities, where a significant fraction of the [H<span style="font-variant:small-caps;">i</span> ]{}column may be in the form of WNM.
We plot the data against our theoretical prediction for the upper envelope of ${\Sigma_{\rm HI}}$ in Figure \[hicomp\], and we show the corresponding predicted lower envelopes for $\Sigma_{\rm H_2}$ and $R_{\rm H_2}\equiv \Sigma_{\rm H_2}/{\Sigma_{\rm HI}}$ in Figures \[molcomp\] and \[rmolcomp\]. For the data set, rather than plotting the tens of thousands of individual pixels it contains, for each galaxy we show the data averaged over 20 logarithmically-spaced column density bins running from the minimum to the maximum value of ${\Sigma_{\rm obs}}$ reported for that galaxy. As the plots show, our model predictions for the upper envelope of the [H<span style="font-variant:small-caps;">i</span> ]{}surface density and the corresponding lower envelope of the H$_2$ surface density as a function of total surface density and metallicity agree very well with the data. The data fill the space up to our predicted envelopes but for the most part do not cross them, even when the predicted [H<span style="font-variant:small-caps;">i</span> ]{}envelope becomes flat for ${\Sigma_{\rm obs}}{\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle >}
{\sim}\:$}}10$ ${M_{\odot}}$ pc$^{-2}$. Moreover, our model recovers not only the primary dependence of ${\Sigma_{\rm HI}}$ on the total observed column density ${\Sigma_{\rm obs}}$, but also the secondary dependence on metallicity. For example, rings in the lowest metallicity bin in the data set reach total mean column densities of almost $20$ ${M_{\odot}}$ pc$^{-2}$, but still show no detectable molecular component. On the other hand, in the highest metallicity bin the molecular fraction is close to $70\%$ in rings with ${\Sigma_{\rm obs}}\approx 20$ ${M_{\odot}}$ pc$^{-2}$. Our models reproduce this effect: at a metallicity of $\log Z'=-0.5$, we predict that the gas will be $94\%$ atomic even at a surface density of $20$ ${M_{\odot}}$ pc$^{-2}$, whereas for $\log Z'=0.5$ we predict an atomic fraction of only $30\%$ at that column density, in agreement with the data.
We caution that we can only predict upper limits on the molecular content in regions of parameter space when our predicted molecular fraction falls below $\sim 1/2$, for the reasons discussed in § \[geomerror\]. We have indicated the regions where our model predictions convert to upper limits in Figures \[hicomp\] – \[rmolcomp\]. Alternately, one can express this uncertainty as giving a minimum column density at which we can predict a value rather than an upper limit for molecular content. For reference, at the metallicities of $\log Z'=-1.25$, $-0.5$, $-0.25$, $0$, and $0.5$ which define the edges of our metallicity bins, the minimum column densities for which we can predict numerical values to better than factor of few confidence are $250$, $58$, $38$, $25$, and $11$ ${M_{\odot}}$ pc$^{-2}$, respectively.
Finally, we note that the future HERACLES / THINGS data set represents an opportunity to perform an even stronger test of our model. In Figure \[hicomp\] a significant fraction of the data points fall below our predicted upper limits, and correspondingly these points are above our lower limits in Figures \[molcomp\] and \[rmolcomp\]. We hypothesize that these data points represent rings or pixels within which the gas is significantly clumped, so that the averaged column density ${\Sigma_{\rm obs}}$ seen in the observation is significantly lower than the column density at which most of the molecular gas in that beam or ring is found, and our calculation for a complex with ${\Sigma_{\rm comp}}={\Sigma_{\rm obs}}$ overestimates the [H<span style="font-variant:small-caps;">i</span> ]{}column. In reality these regions probably consist of patches of high column density where most of the molecules reside, embedded in a lower density ambient medium that has a lower molecular fraction than we determine by averaging over large scales. If we could observe these regions at higher resolution, in Figure \[hicomp\] the high column, high molecular content points will lie to the right of and slightly above the low resolution points, since both the total and [H<span style="font-variant:small-caps;">i</span> ]{}column densities will be higher than for the lower resolution observation, but the increase in the total column will be larger than in the [H<span style="font-variant:small-caps;">i</span> ]{}column. Conversely, the low column, low molecule patches will lie to the left and slightly downward from the low resolution points, since both the total and [H<span style="font-variant:small-caps;">i</span> ]{}columns will decline, but the [H<span style="font-variant:small-caps;">i</span> ]{}by less, since the atomic fraction rises. These changes will bring the data points closer to our model curves. Indeed, the existing data already hint that such an effect is present: the single beam-averaged observations of the data set scatter away from our limit lines noticeably less than the ring-averaged observations from . Similarly, we can divide the point-by-point data from into a “near" sample, consisting of galaxies for which the resolution is smaller than 1 kpc, and a “far" sample, consisting of galaxies with larger resolutions. We plot a version of Figure \[hicomp\] using only this divided data set from in Figure \[hicomp\_dist\]. The comparison is quite noisy, but the data in that bin do at least seem consistent with the hypothesis that the near data fall closer to the model lines than the far data.
The full HERACLES survey, currently underway, will report measurements of the molecular surface density for individual patches $\sim 0.5$ kpc in size, generally smaller than the beam patches the BIMA / SONG survey. While this is still considerably larger than the ${\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <}
{\sim}\:$}}0.1$ kpc-size of a typical atomic-molecular complex in the Milky Way, at the higher spatial resolution of the full HERACLES data set beam-smearing effects should be reduced and the column densities reported for each patch should be closer to the true column densities ${\Sigma_{\rm comp}}$ of the individual atomic-molecular complexes. We therefore predict that in the full HERACLES data set the $\Sigma_{\rm {H\textsc{i} }} - {\Sigma_{\rm obs}}$ relation should be closer to our theoretical upper and lower limit curves than the lower resolution or azimuthally-averaged data shown in Figures \[hicomp\] – \[rmolcomp\].
### Pressure Dependence
and find that the molecular fraction in a galaxy correlates with the mid-plane gas pressure as $${R_{\rm H_2}}\equiv \frac{\Sigma_{\rm H_2}}{{\Sigma_{\rm HI}}} = \left(\frac{P/{k_{\rm B}}}{3.5\pm0.6\times 10^4\mbox{ K cm}^{-3}}\right)^{0.92\pm 0.07},
\label{br06fit}$$ while find $${R_{\rm H_2}}= \left(\frac{P/{k_{\rm B}}}{2.0\times 10^4\mbox{ K cm}^{-3}}\right)^{0.8}
\label{l08fit}$$ for their sample. Although our model does not directly give a prediction for the dependence of the molecular fraction on pressure, and we argue that surface density and metallicity are the physical variables that directly control the molecular fraction, we wish to check whether our model is consistent with the observed correlation.
To avoid introducing any bias in performing this check, we must determine pressures from observable quantities in the same way that and do, following an approximation introduced by .[^1] The BR04 approximation treats the galaxy as an infinite thin disk of uniform gas and stars in vertical hydrostatic balance. For such a disk the pressure is related to surface density by $$\label{psigma}
P = \frac{\pi}{2} G {\Sigma_{\rm g}}\left({\Sigma_{\rm g}}+ \Sigma_*\frac{{v_{\rm g}}}{v_*}\right),$$ where ${\Sigma_{\rm g}}$ is the total ([H<span style="font-variant:small-caps;">i</span> ]{}plus H$_2$) gas surface density, $\Sigma_*$ is the stellar surface density, and ${v_{\rm g}}$ and $v_*$ are the vertical velocity dispersions of the gas and stellar components, respectively. The gas velocity dispersion is roughly constant, ${v_{\rm g}}\approx 8$ km s$^{-1}$, for the galaxies in the and samples. The term $ \Sigma_*/v_*$ presents more difficulty, however. It varies by orders of magnitude from galaxy centers to edges, and between galaxies, so we cannot pick a single value for it. Since in our model the molecular fraction is a function only of metallicity and ${\Sigma_{\rm g}}$, this means that we do not predict a single-valued relationship between ${R_{\rm H_2}}$ and $P$ that we can directly compare to the empirical fits.
However, we can still compare our model to the data in two ways. First, we can pick a range of values of $\Sigma_*/v_*$ consistent with the range in the observed sample, and demonstrate that the resulting range of predictions for ${R_{\rm H_2}}$ covered by our model is consistent with the observational data. For this purpose we follow in adopting a constant value ${v_{\rm g}}=8$ km s$^{-1}$ and using simplified version of equation (\[psigma\]), which follows from assuming that both the surface density and scale height of the stars are much greater than those of the gas: $$\begin{aligned}
\frac{P}{{k_{\rm B}}} & = & 272\mbox{ cm}^{-3}\mbox{ K} \left(\frac{{\Sigma_{\rm obs}}}{{M_{\odot}}\mbox{ pc}^{-2}}\right)
\left(\frac{\Sigma_*}{{M_{\odot}}\mbox{ pc}^{-2}}\right)^{0.5}
\nonumber
\\
& & \qquad {} \times \left(\frac{{v_{\rm g}}}{\mbox{km s}^{-1}}\right) \left(\frac{h_*}{\mbox{pc}}\right)^{-0.5},
\label{presdefn}\end{aligned}$$ where ${\Sigma_{\rm obs}}$ is the observed value of the total gas surface density ${\Sigma_{\rm g}}$, $h_*$ is the stellar scale height, and for convenience we define $\rho_*=\Sigma_*/h_*$. We consider values of $\log \rho_*$ from $-2.5$ to $1.5$ in units of ${M_{\odot}}$ pc$^{-3}$, a range that covers almost all of the samples of and . For a given value of $\rho_*$, equation (\[presdefn\]) gives a single-valued relationship between $P$ and ${\Sigma_{\rm g}}$, so for each $P$ and a choice of metallicity we can use our model to make a prediction for $\Sigma_{\rm HI}$ and thus for ${R_{\rm H_2}}$. As with our calculations in § \[sigmazdepend\], we expect to systematically underpredict ${R_{\rm H_2}}$ because the observed column density ${\Sigma_{\rm obs}}$ is an underestimate of the true column density of atomic-molecular complexes that are not resolved by the telescope beam.
We show the results of this computation in Figure \[pres1\]. As expected, the model covers an area consistent with the observed data. First note that there is a systematic offset between the results from and ; this is likely because some of the structural parameters that are used in equation (\[presdefn\]) are quite uncertain, particularly $h_*$, and the values adopted by the two surveys are not necessarily the same. We should keep this uncertainty in mind as we proceed, since it suggests an upper limit on the level of agreement we can expect.
Nonetheless, the correlation between ${R_{\rm H_2}}$ and $P$ shown in both observational surveys, ${R_{\rm H_2}}\propto P^{0.8-0.9}$, is somewhat flatter than any of our model curves for a particular choice of $\rho_*$, which all approach ${R_{\rm H_2}}\propto P^1$ at large $P$. However this is to be expected: low pressures and gas column densities are found preferentially in the outer parts of galaxies where $\rho_*$ is small, so at low $P$ we expect to be closer to the higher model curves, which have low $\rho_*$. Conversely, high pressures and values of ${\Sigma_{\rm g}}$ are found preferentially in regions with higher stellar densities $\rho_{*}$, so high $P$ values should be closer to the lower model curves, which have larger $\rho_*$. This covariance between $P$ and $\rho_*$ results in a flattening of the ${R_{\rm H_2}}\propto P$ relation that we would predict if all galaxies had fixed $\rho_*$.
We can eliminate this covariance effect and produce a stronger test than that shown in Figure \[pres1\] by using our model to generate predictions for molecular content directly, and plotting those against the inferred pressure. To do this, for each galactocentric ring in the data set we take the observed total gas surface density ${\Sigma_{\rm obs}}$ and the metallicity given in Table \[metaltab\] (which we treat as constant in each the galaxy) and use our model to predict ${R_{\rm H_2}}$. Similarly, for the data set we take the observed gas surface density ${\Sigma_{\rm obs}}$ in each pixel and use our model to generate a prediction for ${R_{\rm H_2}}$. Since the pressure in each ring or pixel is known, by this procedure we generate a synthetic set of data points $(P,{R_{\rm H_2}})$ which we can compare to the observations, and from which we can generate a fit for ${R_{\rm H_2}}$ as a function of $P$.
Before making the comparison, however, we must also model the effects of finite telescope sensitivity, which make it impossible to detect molecules below a certain minimum column density. Since, as Figure \[pres1\] shows, our model predicts that the slope of ${R_{\rm H_2}}$ versus $P$ varies with $P$, this can affect the fit we generate from our synthetic data. We include this effect using a procedure nearly identical to that of . These authors first estimate the minimum value of ${R_{\rm H_2}}$ as a function of pressure that they could detect in each galaxy based on their telescope noise limits. For our synthetic data we adopt a minimum $R_{\rm H_2,min}=0.05$ for all galaxies, slightly below the lowest average ${R_{\rm H_2}}$ that report for any of the galaxies they analyze. For each galaxy we then identify the lowest pressure $P_{\rm min}$ for which all the pixels / rings have ${R_{\rm H_2}}>R_{\rm H_2,min}$. At pressures above $P_{\rm min}$, the sample should be nearly complete, in the sense that no pixels or rings will have molecular non-detections that could bias a fit of ${R_{\rm H_2}}$ versus $P$. At pressures below $P_{\rm min}$ the data are incomplete, and we therefore drop all pixels or rings whose pressures are below $P_{\rm min}$. This produces a sensitivity-corrected set of synthetic data. Finally, following we then bin the pixels by pressure, using bins 0.1 dex wide, and for each pressure bin in each galaxy we compute an average value of ${R_{\rm H_2}}$. We do not bin the rings in the sample.
We overplot the synthetic and real data in Figure \[pres2\]. Fitting our synthetic data to a powerlaw function for ${R_{\rm H_2}}$ versus $P$ gives $${R_{\rm H_2}}= \left(\frac{P/{k_{\rm B}}}{9.0\times 10^3\mbox{ K cm}^{-3}}\right)^{0.81},
\label{theoryfit}$$ if we include our models for both the and data sets. Using only one or the other gives $${R_{\rm H_2}}=
\left\{
\begin{array}{ll}
\left(\frac{P/{k_{\rm B}}}{2.2\times 10^2\mbox{\small\ K cm}^{-3}}\right)^{0.74}, \quad & \mbox{(BR06)} \\
\left(\frac{P/{k_{\rm B}}}{2.8\times 10^5\mbox{\small\ K cm}^{-3}}\right)^{0.85}, \quad & \mbox{(L08)}
\end{array}
\right..$$ As expected, we obtain a slope shallower than unity as a result of the systematic increase of $\rho_*$ with $P$. The scatter in the real and synthetic data are also comparable. As the Figure shows, our model predictions overlap with the observed data reasonably well, particularly for the data; the best fit for the data set is reasonably close to the observational best fit, while for the theoretical predictions and the best fit to them give a slope similar to the observed value, but are systematically shifted to lower ${R_{\rm H_2}}$. The displacement between our model and the data probably exists for the same reason that our predictions for $\Sigma_{\rm H_2}$ and ${R_{\rm H_2}}$ in § \[sigmazdepend\] are lower limits: spatial averaging leads to an underestimate of the true column densities of the atomic-molecular complexes in a galaxy, which in turn leads us to slightly underpredict the molecular content. The averaging is significantly worse if done over rings than over individual beam pointings, so the offset is noticeably larger for the ring-averaged data set.
It is also worth cautioning that in performing this fit we have included model points where we predict ${R_{\rm H_2}}<1$, a region of parameter space where the predictions of our theory should be taken as upper limits. If we exclude these points from both our model predictions and from the observed data, all the slopes become significantly more shallow, but they remain consistent with one another. In this case our model predictions give a best fit of ${R_{\rm H_2}}= [(P/{k_{\rm B}})/(83\mbox{ K cm}^{-3})]^{0.47}$ for the combined BR06 and L08 samples, while the and data sets give $[(P/{k_{\rm B}})/(120\mbox{ K cm}^{-3})]^{0.51}$ and $[(P/{k_{\rm B}})/(300\mbox{ K cm}^{-3})]^{0.53}$, respectively. These fits are performed without weighting by the errors or properly including the effects of upper limits, so they should only be taken as general indications, but the results do show that even if we limit our fit to the part of parameter space where we can apply our theory with high confidence, we obtain good agreement between model predictions and observations.
We can further check the level of agreement between the model and the data, and see how well our theory compares to the purely empirical fit, using the sample, where there is no (or less) systematic offset due to unresolved clumping. In Figure \[pres\_single\] we plot the same data as in Figure \[pres2\], but show the data and model predictions for the galaxies one by one. As the plots show, our model not only fits the general trend between ${R_{\rm H_2}}$ and $P$, for most galaxies we obtain a good match on a bin-by-bin basis. Even in those galaxies where the agreement is poor, our model still gives a correct lower limit for ${R_{\rm H_2}}$. The level of agreement between our predicted curves and the data is as good as the purely empirical fit between ${R_{\rm H_2}}$ and pressure in , without the need for any free parameters. It is not clear why one model or the other does better for particular galaxies. Our model appears to provide a substantially better fit for NGC3627, NGC5033, and NGC5055, while doing noticeably worse for IC 10, NGC0598, and NGC4736; for the rest of the sample the empirical fit and our model calculations are nearly equally good fits. It is not clear what galaxy properties favor one model or the other.
### The Small Magellanic Cloud {#smcsec}
As noted above, the SMC’s metallicity of $12+\log(\mbox{O}/\mbox{H})=8.0$ [@dufour84a] makes it the lowest metallicity galaxy in our sample for which we have detections rather than upper limits on the H$_2$ content. To analyze this data set, we take the maps of [H<span style="font-variant:small-caps;">i</span> ]{}and H$_2$ column determined by @leroy07a and break the data into 20 bins in total gas column (including He). In each bin we compute the mean [H<span style="font-variant:small-caps;">i</span> ]{}and H$_2$ column density. It is worth noting that the technique used to infer H$_2$ column densities in this data set have $\sim 50\%$ systematic uncertainties, so we must proceed with this caution in mind.
To test our theoretical model, we generate a prediction for the H$_2$ mass fraction as a function of total gas column density ${\Sigma_{\rm comp}}$ as in equation (\[h2mfraceqn\]), using a metallicity relative to Solar of $Z'\approx 0.2$. Before comparing the model curve to the data, we face the difficult problem of inclination correction. The SMC is a triaxial structure with an aspect ratio of roughly $1:2:4$, with $4$ representing the direction along the line of sight [@crowl01a]. The large extent along the line of sight means that the inclination correction is very significant, but the triaxiality of the galaxy means that no single number describes the inclination, as is the case for a disk. We therefore compare the model and the data using two different inclination corrections, one assuming an inclination $i=76^{\circ}$ (corresponding to a $1:4$ aspect ratio) and one assuming an inclination $i=63^{\circ}$ (corresponding to a $1:2$ aspect ratio). These should bracket the true inclination correction. Another caution we should make here is that, due to these large inclinations, there may be significant amounts of gas along a given line of sight that are not associated with whatever atomic-molecular complex it intersects, so the [H<span style="font-variant:small-caps;">i</span> ]{}column densities may be overestimates (see § \[sigmazdepend\]).
Figure \[smcplot\] shows a comparison of the data and the model curve. As expected, the data for the two different inclination corrections bracket the model curve quite well. We should be cautious about reading too much into the agreement at low $\Sigma_{\rm obs}$ (${\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <}
{\sim}\:$}}30$ ${M_{\odot}}$ pc$^{-2}$), where both the data and our model are quite uncertain. Indeed, ${R_{\rm H_2}}$ is such a sharp function of $\Sigma_{\rm obs}$ at low column that the factor of $2$ variation in the total column introduced by the uncertainty in the inclination translates to a factor of $\sim 50$ difference in the value of ${R_{\rm H_2}}$ at a fixed $\Sigma_{\rm obs}$. Thus, our model need only be correct within a factor of 50 to lie in between the two curves for the different inclination corrections. At higher $\Sigma_{\rm obs}$ (${\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle >}
{\sim}\:$}}80$ ${M_{\odot}}$ pc$^{-2}$), on the other hand, the inclination correction only produces a factor of $\sim 2$ uncertainty in ${R_{\rm H_2}}$, and the ability of our model to match the data here is significant. It shows that we are capable of predicting the total column at which ${R_{\rm H_2}}\sim 1$ to better than a factor of 2 accuracy even in a galaxy with metallicity $\sim 1/10$ Solar.
Galactic Observations {#galobs}
---------------------
### Data Sets
We next compare our model to observations of the molecular content of clouds in the Milky Way, as measured by the *Copernicus* [@spitzer75] and the *Far Ultraviolet Spectroscopic Explorer (FUSE)* [@moos00; @sahnow00] missions. These satellites measured absorption of ultraviolet light from background stars, and in some cases AGN, in the Lyman-Werner bands, enabling them to estimate the population of molecular hydrogen along a given line of sight.
@savage77 and @bohlin78 (for *Copernicus*) and @rachford02 [@rachford08a] (for *FUSE*) report measurements of the molecular hydrogen columns in the Milky Way disk for lines of sight at low galactic latitude ($b{\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <}
{\sim}\:$}}10^{\circ}$), using stars as background sources. They combine these with [H<span style="font-variant:small-caps;">i</span> ]{}observations along the same lines of sight to determine molecular hydrogen fractions for clouds these sightlines. @gillmon06a report *FUSE* observations of sightlines at high galactic latitude ($b{\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle >}
{\sim}\:$}}20^{\circ}$) using AGN as background sources. These observations probe clouds lying above or below the galactic plane, which are presumably illuminated only on one side by stars in the disk. As with the low-latitude samples, @gillmon06a combine these observations with measurements of the [H<span style="font-variant:small-caps;">i</span> ]{}column along the same sightlines to determine molecular fractions.
These observations are different than the extragalactic ones in that they are pencil-beam measurements rather than averages over atomic-molecular complexes. We possess little information about the geometry of the clouds these beams probe, and this ignorance complicates comparison of the data to our model. A given observed total (atomic plus molecular) column density $\Sigma_{\rm obs}$ might be the result of a line of sight passing through the center of a small cloud, or might be the result of a beam that passes only tangentially through a much larger cloud. Our models do not predict the same H$_2$ content in these two cases, so in the absence of additional information there is no way to map a given $\Sigma_{\rm obs}$ to a unique prediction for the molecular fraction. Similarly, there is good evidence in the data sets from the rotational excitation of the H$_2$ that at least some lines of sight probe multiple clouds that are separated in space [e.g. @browning03]. Our model predicts a lower molecular fraction for this case than for the case of a single cloud with the same $\Sigma_{\rm obs}$. Attempts to remove these effects statistically are complicated by the fact that the *FUSE* and *Copernicus* lines of sight are not an unbiased sample. Each of the samples we use was chosen specifically to probe a certain range in column density or other properties, and all the selections are biased against high column densities and high molecular fractions, both of which produce high extinctions of LW photons that make determining a molecular column very costly or altogether impossible. Below we discuss how we deal with the problems of geometric uncertainty and selection bias for the galactic plane and high-latitude samples.
A second complication for this data set is that the assumption we make in our model that clouds are subjected to a relatively uniform background dissociating radiation field [cf. @krumholz08c] may not be appropriate in this context. Treating the ISRF as uniform is reasonable for the giant cloud complexes with masses $\ga 10^4$ ${M_{\odot}}$ and sizes $\ga 100$ pc probed by extragalactic measurements; these sample the radiative output of many stars and star clusters. In contrast, the entire sample of high latitude clouds observed by @gillmon06a has a total mass $\sim 3000$ ${M_{\odot}}$ spread out over a $\sim 100$ pc$^2$ area [@gillmon06b], making them tiny in comparison to giant complexes. The galactic clouds probed by the *Copernicus* and *FUSE* observations are of unknown size, but are likely to be similarly small. This is because any line of sight passing through a true giant cloud complex would be completely opaque in the Lyman-Werner bands, thus preventing *Copernicus* or *FUSE* from making any measurement of the H$_2$ column. The dissociating flux in the vicinity of such small clouds may be dominated by a single nearby star or star cluster.
Due to the effects of uncertain geometry and non-uniform radiation fields, we expect this data set to show significantly more scatter than the extragalactic one.
### The Galactic Plane Data
Given the limitations of pencil-beam measurements, we must compare our model to the data in a way that accounts for our geometric uncertainty. The most straightforward comparison we can make is by considering a limiting case. Consider an observation that measures a total column density ${\Sigma_{\rm obs}}$. According to our model, the fraction of gas in molecular form will depend on the geometry of the complex; if the line of sight probes the outer parts of a very large cloud, always staying within the atomic region, then in principle there could be no molecules present no matter how large ${\Sigma_{\rm obs}}$ might be. The reverse is not true, however. Even if the line of sight passes exactly through the center of an atomic-molecular complex, it must still pass through atomic shielding layer, and there will therefore be some [H<span style="font-variant:small-caps;">i</span> ]{}along that line of sight. In our model, therefore, there is a minimum amount of [H<span style="font-variant:small-caps;">i</span> ]{}that we predict must be present for a given column of H$_2$; this minimum corresponds to the case of a line of sight that passes directly through the center of a complex. This implies that we can make a prediction as follows: we consider a complex of mean column density ${\Sigma_{\rm comp}}$, and use our model to predict the atomic, molecular, and total column densities, $\Sigma_{\rm HI,cen}$, $\Sigma_{\rm H_2,cen}$, and $\Sigma_{\rm cen}$, that one would see along a line of sight passing through the center of the complex. The curves $\Sigma_{\rm HI,cen}$ versus $\Sigma_{\rm cen}$ and $\Sigma_{\rm H_2,cen}$ versus $\Sigma_{\rm cen}$ that we generate through this procedure should then be lower and upper limits, respectively, on the observed distributions of $\Sigma_{\rm HI}$ versus ${\Sigma_{\rm obs}}$ and $\Sigma_{\rm H_2}$ versus ${\Sigma_{\rm obs}}$.
Instead of computing a limit, we can also make an estimate for the “typical" atomic and molecular fractions we should see at a given ${\Sigma_{\rm obs}}$. If we knew the true distribution of column densities for atomic-molecular complexes in the Milky Way, we could do this by integrating all lines of sight through that distribution and computing the mean atomic and molecular columns at a given total column. However, we do not know the true distribution, and even if we did this procedure would not address effects of observational bias in selecting sightlines. Instead, we make a much rougher calculation. Using our model we can compute the radius ${x_{\rm H_2}}({\Sigma_{\rm comp}})$ at which the cloud whose true mean column density is ${\Sigma_{\rm comp}}$ transitions from atomic to molecular for solar metallicity ($Z=1$). The atomic and molecular columns along a sightline that strikes the cloud of radius $R$ at a distance $\beta R$ from its center are $$\begin{aligned}
\Sigma_{\rm HI} & = &
\frac{2{\tau_{\rm R}}}{{\sigma_{\rm d}}}
\left\{
\begin{array}{ll}
\sqrt{1-\beta^2}-\sqrt{{x_{\rm H_2}}^2-\beta^2}, \quad & \beta < {x_{\rm H_2}}\\
\sqrt{1-\beta^2}, & \beta > {x_{\rm H_2}}\end{array}
\right. \\
\Sigma_{\rm H_2} & = &
\frac{2{\tau_{\rm R}}{\phi_{\rm mol}}}{{\sigma_{\rm d}}}
\left\{
\begin{array}{ll}
\sqrt{{x_{\rm H_2}}^2-\beta^2}, \quad & \beta<{x_{\rm H_2}}\\
0, & \beta > {x_{\rm H_2}}\end{array}
\right.,\end{aligned}$$ and the H$_2$ fraction averaged over all pencil beams passing through the cloud is, $$f_{\rm H_2,beams}({\Sigma_{\rm comp}}) =
2\int_0^{1} \beta \frac{\Sigma_{\rm H_2}}{\Sigma_{\rm H_2}+\Sigma_{\rm HI}}
d\beta.$$ To estimate a “typical" atomic or molecular content for a given observed total gas column ${\Sigma_{\rm obs}}$, we can simply take $\Sigma_{\rm HI,obs} \approx [1-f_{\rm H_2,beams}({\Sigma_{\rm obs}})]{\Sigma_{\rm obs}}$ and $\Sigma_{\rm H2,obs} \approx f_{\rm H_2,beams}({\Sigma_{\rm obs}}){\Sigma_{\rm obs}}$. This is equivalent to saying that our sightlines do indeed probe random impact parameters on cloud complexes, and that when we observe a sightline of column density ${\Sigma_{\rm obs}}$, most of the time we are observing a complex whose mean column density is also about ${\Sigma_{\rm obs}}$. This assumption could fail if low column density clouds were rare compared to high column density ones, so that most of the sightlines that produce a given ${\Sigma_{\rm obs}}$ were really tangential paths through high column clouds rather than beams passing close to the the centers of low column clouds. However, there is no evidence that high-column density atomic-molecular complexes outnumber low-column ones, and in general in the ISM more mass tends to be in diffuse than dense structures, so we proceed with our assumption that ${\Sigma_{\rm comp}}\approx {\Sigma_{\rm obs}}$.
We plot both our limits on the [H<span style="font-variant:small-caps;">i</span> ]{}and H$_2$ columns and our estimates for their typical values against the *Copernicus* and *FUSE* data sets in Figure \[galcomp\]. As the plot shows, both our limits and our typical columns match the observed data reasonably well. As with the extragalactic case, we recover the overall trend that the gas is mostly [H<span style="font-variant:small-caps;">i</span> ]{}until a total column of a few ${M_{\odot}}$ pc$^{-2}$, and then mostly molecular thereafter.
### The High-Latitude Data
The comparison to the high-latitude data set is somewhat less straightforward, because the clouds at high latitude are only illuminated from one side. Indeed, @gillmon06a find that these clouds become molecular at column densities a factor of $\sim 2$ lower than do clouds in the galactic plane, probably as a result of this one-sided illumination. We have not considered the case of clouds of finite size illuminated over only half their surfaces. However, we can obtain a reasonable approximation to this configuration using the case of *semi-infinite* clouds subjected to an isotropic, uniform dissociation radiation field, which we considered in § \[giantclouds\]. A semi-infinite cloud blocks dissociating radiation over $2\pi$ sr, and thus the depth of the atomic shielding layer at its surface is determined only by photons that arrive from the “front side". This is therefore a close analogy to the case of a high-latitude cloud, although it differs in the cloud geometry.
In Paper I we show that using a semi-infinite slab in place of a finite cloud tends to produce errors in estimating the exact location of a transition from atomic to molecular in a finite cloud, but that the semi-infinite calculation does give a good estimate of the column at which clouds transition from mostly atomic to mostly molecular. We therefore expect to obtain roughly the right transition column and thus the right column of [H<span style="font-variant:small-caps;">i</span> ]{}for clouds that are mostly molecular, but not particularly accurate predictions for the exact molecular column. For a semi-infinite cloud of solar metallicity, as we show in § \[giantclouds\] we predict a molecular to atomic ratio of $${R_{\rm H_2}}=
\left\{
\begin{array}{ll}
0, & \Sigma < 4.5\,{M_{\odot}}\mbox{ pc}^{-2} \\
(\Sigma-4.5\,{M_{\odot}}\mbox{ pc}^{-2})/\Sigma, & \Sigma > 4.5\,{M_{\odot}}\mbox{ pc}^{-2}
\end{array}
\right.,$$ where $\Sigma$ is the total column density. In other words, for a semi-infinite cloud of solar metallicity, the first $4.5$ ${M_{\odot}}$ pc$^{-2}$ are atomic, and the rest are molecular. We show this prediction overplotted with the *FUSE* high-latititude data in Figure \[hlcomp\]. As the plot shows, we indeed do not get very accurate predictions for the exact ratio of [H<span style="font-variant:small-caps;">i</span> ]{}to H$_2$, but our calculation agrees quite well with the general value of $\Sigma$ for which the transition from molecular to atomic occurs. In particular, our one-sided model recovers the observational result that a cloud illuminated from one side shows a lower atomic-to-molecular transition column than an isotropically-illuminated cloud.
Summary and Conclusions {#conclusion}
=======================
We present a first-principles calculation of the molecular gas content of galactic disks in terms of the observable properties of those galaxies. Our calculations build on the simple model for photodissociation fronts in finite clouds presented in , in which we show that the amount of atomic material required to shield a molecular cloud against dissociation by the interstellar radiation field (ISRF) can be characterized by two parameters: $\chi$, a radiation field strength normalized by the gas density and the properties of dust grains, and ${\tau_{\rm R}}$, a measure of the dust optical depth of a cloud. We show that, due to the way the density in the cold phase of the atomic ISM varies with ISRF, the normalized radiation field strength takes a characteristic value $\chi\approx 1$ in all galaxies where a two-phase atomic ISM is present, with only a weak dependence on metallicity.
The existence of a characteristic normalized radiation field strength, and its weak dependence on metallicity, has a number of important consequences. First, it enables us to give a simple analytic approximation (equations \[hifracanalyt\] or \[hifracanalyt1\]) for the fraction of mass in an atomic-molecular complex that will be in the atomic or molecular phases solely in terms of the column density of the complex and the metallicity of the gas. This makes it easy to test our calculations against observations. Second, we show that as a consequence of $\chi$ assuming a nearly fixed value, the atomic envelopes of molecular clouds have a characteristic visual extinction $A_V\approx 0.2$ at solar metallicity. Similarly, the transition from atomic to molecular gas occurs at a characteristic shielding column of $\Sigma_{\rm HI} \approx 10$ ${M_{\odot}}$ pc$^{-2}$. These quantities both vary sub-linearly metallicity, with $A_V$ declining and $\Sigma_{\rm HI}$ increasing as metallicity does. We calculate these values and their metallicity-dependence solely in terms of the microphysical constants that describe the properties of H$_2$ formation and dissociation, the cooling curves of C<span style="font-variant:small-caps;">ii</span> and O<span style="font-variant:small-caps;">i</span>, and the properties of interstellar dust grains. Our model does not depend on unobservable parameters such as the gas volume density or the ISRF strength in a galaxy.
Our model compares favorably with observations of the atomic and molecular content of clouds both in the Milky Way and in nearby galaxies. We are able to reproduce both the characteristic column density at which clouds transition from being primarily atomic to primarily molecular, and the way that this characteristic column depends on metallicity and on whether clouds are illuminated on one side or on both sides. We are also able to reproduce the observed correlation between molecular content and interstellar pressure.
The development of a predictive model for the molecular content of galaxies that does not rely on unknown and generally unobservable quantities such as the radiation fields or gas volume densities in those galaxies opens up new possibilities to advance our understanding of galactic evolution. In low-density dwarfs or high-redshift galaxies containing few metals, the formation of molecular clouds may be the rate limiting step in star formation. On the other hand, previous work has shown that, once molecular gas forms, it converts itself into stars at a rate of a few percent of the mass per free-fall time independent of its density or environment [@krumholz05c; @krumholz07e; @mckee07b]. Thus a theory of molecule formation creates the possibility of developing a theory of the star formation rate capable of making predictions that can be applied not only in the relatively molecule-rich nearby galaxies for which empirical star formation laws have been determined [e.g. @kennicutt98a], but also in the more distant and lower metallicity universe where these laws are known to break down [e.g. @wolfe06a]. We plan to explore such a theory in future work.
We thank B. Rachford and A. Leroy for providing copies of their data, and F. Bigiel, S. Faber, N. Gnedin, A. Leroy, A. Sternberg, and F. Walter for helpful discussions and comments on the manuscript. We thank the referee, E. Rosolowsky, for providing comments that improved the quality of the paper. Support for this work was provided by: NASA through Hubble Fellowship grant \#HSF-HF-01186 awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS 5-26555 (MRK); NASA, as part of the Spitzer Space Telescope Theoretical Research Program, through a contract issued by the Jet Propulsion Laboratory, California Institute of Technology (MRK); and by the National Science Foundation through grants AST-0807739 (to MRK), AST-0606831 (to CFM), and PHY05-51164 (to the Kavli Institute for Theoretical Physics, where MRK, CFM, and JT collaborated on this work).
[^1]: It is important to note that the midplane pressure which the observers attempt to estimate is not the same as the pressure in the molecular gas or the CNM which we balanced in § \[normrad\]; the total midplane pressure includes contributions from all phases of the ISM rather than just the CNM or molecular clouds, as well as contributions from magnetic fields, cosmic rays, and bulk motions. The pressure we use in § \[normrad\] includes thermal pressure in the CNM or molecular gas only.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We have performed three-flavor Boltzmann neutrino transport radiation hydrodynamics simulations covering a period of 3 s after the formation of a protoneutron star in a core-collapse supernova explosion. Our results show that a treatment of charged-current neutrino interactions in hot and dense matter as suggested by @Reddy.Prakash.Lattimer:1998 \[Phys. Rev. D **58**, 013009 (1998)\] has a strong impact on the luminosities and spectra of the emitted neutrinos. When compared with simulations that neglect mean field effects on the neutrino opacities, we find that the luminosities of all neutrino flavors are reduced while the spectral differences between electron neutrino and antineutrino are increased. Their magnitude depends on the equation of state and in particular on the symmetry energy at sub-nuclear densities. These modifications reduce the proton-to-nucleon ratio of the outflow, increasing slightly their entropy. They are expected to have a substantial impact on the nucleosynthesis in neutrino-driven winds, even though they do not result in conditions that favor an r-process. Contrarily to previous findings, our simulations show that the spectra of electron neutrinos remain substantially different from those of other (anti)neutrino flavors during the entire deleptonization phase of the protoneutron star. The obtained luminosity and spectral changes are also expected to have important consequences for neutrino flavor oscillations and neutrino detection on Earth.'
author:
- 'G. Mart[í]{}nez-Pinedo'
- 'T. Fischer'
- 'A. Lohs'
- 'L. Huther'
title: 'Charged-current weak interaction processes in hot and dense matter and its impact on the spectra of neutrinos emitted from proto-neutron star cooling'
---
Supernova explosions of massive stars are related to the birth of neutron stars due to the collapse of the stellar core at the end of stellar evolution [@Janka.Langanke.ea:2007]. The detection of neutrinos from SN1987A [@Hirata.Kajita.ea:1987; @*Koshiba:1992] confirmed that the $\approx 3\times10^{53}$ ergs of gravitational energy gained by the core collapse are emitted as neutrino radiation on time scales of tens of seconds, during which the central protoneutron star (PNS) cools, deleptonizes and contracts. In the delayed neutrino-heating explosion mechanism [@Bethe.Wilson:1985; @Janka.Langanke.ea:2007], neutrinos also transport energy from the nascent PNS to the stalled bounce shock. This mechanism remains the most viable scenario to explain supernova explosions as confirmed by recent two-dimensional simulations [@Mueller.Janka.Marek:2012]. Once the explosion sets in, the continuous emission of neutrinos from the PNS drives a low-mass outflow known as neutrino-driven wind [@duncan.shapiro.wasserman:1986] that is currently considered the favored site for the productions of elements heavier than iron (e.g. [@Qian:2003]). As neutrinos travel through the stellar mantle, they can suffer flavor oscillations [@Duan.Fuller.Qian:2010], contribute to the nucleosynthesis of several rare isotopes [@Woosley.Hartmann.ea:1990; @*Heger.Kolbe.ea:2005] and even drive an $r$ process in the He-shell of metal-poor stars [@Banerjee.Haxton.Qian:2011] before they are finally detected on Earth.
Accounting for all aspects discussed above requires the knowledge of the spectra of the neutrinos emitted during the cooling phase of the PNS. Due to their low energies $\nu_{\mu,\tau},\bar\nu_{\mu,\tau}$ interact only via neutral-current reactions. Hence, together with the neutron-richness of the PNS surface one expects the following neutrino-energy hierarchy: $\langle E_{\nu_{\mu,\tau}} \rangle>\langle
E_{\bar{\nu}_e} \rangle>\langle E_{\nu_e}
\rangle$ [@keil.raffelt.janka:2003; @Fischer.Martinez-Pinedo.ea:2012], with $\langle E \rangle$ the average energy of the neutrino spectrum. Early supernova models [@Woosley.Wilson.ea:1994; @*Takahashi.Witti.Janka:1994] predicted large energy differences between $\bar{\nu}_e$ and $\nu_e$ that resulted in neutron-rich ejecta as required by $r$-process nucleosynthesis [@hoffman.woosley.qian:1997]. However, as the treatment of neutrino transport and weak interaction processes improved, the computed energy difference between $\bar{\nu}_e$ and $\nu_e$ decreased and the early wind ejecta became proton rich [@Liebendoerfer.Mezzacappa.ea:2001a; @*Buras.Rampp.ea:2006]. More recently, it has been possible to perform supernova simulations based on three-flavor Boltzmann neutrino transport for time scales of several tens of seconds [@Huedepohl.Mueller.ea:2010; @Fischer.Whitehouse.ea:2010], covering the whole deleptonization of the PNS. These simulations predict a continuous decrease of the energy difference between neutrinos and antineutrinos of all flavors that became practically indistinguishable after $\approx 10$ s. The exact value depends on the progenitor. The proton-richness of the ejecta increases continuously with time and leaves the $\nu p$ process [@Froehlich.Martinez-Pinedo.ea:2006; @*Pruet.Hoffman.ea:2006; @*Wanajo:2006] as the only mechanism for producing elements heavier than iron in neutrino-driven winds.
The simulations of ref. [@Fischer.Whitehouse.ea:2010] have been recently analyzed, showing that the convergence of neutrino and antineutrino spectra at late times is due to the suppression of charged-current processes at high densities [@Fischer.Martinez-Pinedo.ea:2012]. This analysis was based on a set of neutrino opacities that assume a non-interacting gas of nucleons and nuclei. This approximation may be valid during the accretion phase prior to the onset of the supernova explosion when the region from where neutrinos decouple, the neutrinospheres, is located at relatively low densities, $\sim 10^{11}$ g cm$^{-3}$. However, as the PNS cools the neutriospheres move to increasingly higher densities where the non-interacting gas approximation breaks down. The nuclear interaction is treated at the mean-field level in equations of state (EoS) commonly used in core-collapse supernova simulations [@Lattimer.Swesty:1991; @Shen.Toki.ea:1998a]. However, its influence on weak interaction processes is often neglected. In this Letter, we show that a treatment of neutrino-matter interactions that is consistent with the underlying EoS has a strong impact on the spectra and luminosities of the emitted neutrinos. We discuss the relevance for nucleosynthesis, neutrino oscillation studies and neutrino detection.
Our study is based on the work of Ref. [@Reddy.Prakash.Lattimer:1998] where corrections to the opacities due to strong interactions are considered at the mean-field level. Effects of many-body correlations [@Reddy.Prakash.ea:1999; @*Burrows.Sawyer:1998; @*Burrows.Sawyer:1999] will be considered in a forthcoming publication. They are expected to affect the neutrino spectra at later times [@Roberts.Shen.ea:2012] than considered in the present study. We focus on charged-current (anti)neutrino absorption processes on neutrons and protons and the inverse reactions: $e^- + p \rightleftarrows n + \nu_e$ and $e^+ + n
\rightleftarrows p + \bar{\nu}_e$, which are those mainly affected by mean-field corrections.
EoS commonly used in core-collapse supernova simulations, see e.g. Refs. [@Lattimer.Swesty:1991; @Shen.Toki.ea:1998a], treat neutrons and protons as a gas of quasi-particles that move in a mean-field single-particle potential $U$. Assuming non-relativistic nucleons, which is a good approximation for densities $\rho \leq 5
\rho_0$ where $\rho_0 \approx 2.5\times 10^{14}$ g cm$^{-3}$, the energy momentum relation closely resembles the non-interacting case [@Reddy.Prakash.Lattimer:1998]: $$\label{eq:1}
E_i(\bm{p}_i) = \frac{\bm{p}^2_i}{2m^*_i} + m_i + U_i, \quad i = n,p,$$ with particle rest-masses $m_i$. Both the single-particle mean-field potentials and the (Landau) effective masses, $m^*_i$ depend on density, temperature and proton-to-nucleon ratio, $Y_e$. Importantly, due to the extreme neutron-rich conditions the mean-field potentials for neutron and protons can be very different with their relative difference $U_n - U_p$ directly related to the nuclear symmetry energy [@Reddy.Prakash.Lattimer:1998].
In order to quantify the impact of the mean field potentials, let us consider (anti)neutrino absorption on neutrons (protons). Assuming zero momentum transfer, i.e. $\bm{p}_n \approx \bm{p}_p$ (elastic approximation), the electron(positron) and (anti)neutrino energies are related by:
\[eq:2\] $$\begin{aligned}
\label{eq:3}
E_{\nu_e} &=& E_{e^-} - (m_n - m_p) - (U_n - U_p), \\
\label{eq:4}
E_{\bar{\nu}_e} &=& E_{e^+} + (m_n - m_p) + (U_n - U_p).
\end{aligned}$$
Eqs. (\[eq:3\]) and (\[eq:4\]) show that the contribution of the mean field potential reduces (increases) the energy of the emitted neutrinos (antineutrinos). The energy difference between neutrinos and antineutrinos is increased by an amount $2(U_n - U_p)$. The opacity, or inverse mean-free path, for (anti)neutrino absorption also changes. In the elastic approximation and assuming extreme relativistic electrons, the opacity for neutrino absorption becomes [@Bruenn:1985; @Reddy.Prakash.Lattimer:1998]: $$\begin{aligned}
\label{eq:5}
\chi(E_{\nu_e}) &=&
\frac{G_F^2 V_{ud}^2}{\pi (\hbar c)^4} (g_V^2+3 g_A^2) \cdot
\nonumber \\
&& E_{e}^2 [1-f_e(E_{e})] \frac{n_n - n_p}{1-e^{\beta(\eta_p -U_p - \eta_n+U_n)}},\end{aligned}$$ with $E_{\nu_e}$ and $E_e$ related by equation (\[eq:3\]). The emissivity, $j(E_{\nu_e})$, can be obtained from the detailed balance condition $j(E_{\nu_e}) = \exp(-\beta(E_{\nu_e} -
\mu^{\text{eq}}_\nu)) \chi(E_{\nu_e})$, with $\mu^{\text{eq}}_\nu =
\mu_e - (\mu_n - \mu_p)$ the equilibrium neutrino chemical potential, $\mu$ the chemical potential including rest mass and $\beta$ the inverse temperature. The opacity and emissivity for antineutrino absorption are obtained exchanging neutron and proton and using equation (\[eq:4\]) to relate the positron and antineutrino energies. In equation (\[eq:5\]), $G_F$ is the Fermi coupling constant, $V_{ud}$ is the up-down entry of the Cabibbo-Kobayashi-Maskawa matrix, $g_V$ and $g_A$ are vector and axial-vector coupling constants and $n_{p,n}$ the number density of protons or neutrons. $f$ is the Fermi-Dirac distribution function and $\eta$ is the chemical potential (without rest mass). The quantity $\varphi = \eta - U$ represents the chemical potential for a non-interacting gas of nucleons, that is related to the nucleon number density by the relation: $$\label{eq:6}
n = 2\int \frac{d^3\bm{p}}{(2\pi\hbar c)^3}
\frac{1}{e^{\beta\left(\frac{p^2}{2m} -\varphi\right)}+1}.$$ Eq. (\[eq:6\]) provides a method of determining the mean-field potential, $U$, when using an EoS that does not provide this quantity, e.g., the EoS of ref. [@Shen.Toki.ea:1998a].
![(Color online) Opacity and emissivity for neutrino (left panels) and antineutrino (right panels), evaluated at conditions $\rho = 2.1\times10^{13}$ g cm$^{-3}$, $T= 7.4$ MeV and $Y_e=0.035$.[]{data-label="fig:opacity"}](emissivity.eps){width="\linewidth"}
Figure \[fig:opacity\] shows neutrino and antineutrino opacities and emissivities evaluated at conditions found at the antineutrinosphere for the 18 M$_\odot$ model of ref. [@Fischer.Martinez-Pinedo.ea:2012] at 1 s after bounce. The curves labeled RMF ($U_n,U_p$) include the contribution of the mean-field potentials $U_n = -7.6$ MeV and $U_p =
-14.2$ MeV [@Shen.Toki.ea:1998a], while the curves labeled ($U_n=0,U_p=0$) assume a non-interacting gas of nucleons, i.e. neglect the contribution of the potentials but still use chemical potentials as given by the EoS. Due to the presence of the mean-field potentials the effective $Q$-value for electron capture increases with respect to the free case producing neutrinos with substantially lower energy. For the inverse process, neutrino absorption, the opacity is enhanced due to the fact that the produced electron gains an energy $U_n - U_p$ reducing the final-state Pauli blocking of the electron. The situation is completely analogous to (anti)neutrino emission and absorption on heavy neutron-rich nuclei [@Langanke.Martinez-Pinedo:2003]. Using Eq. (\[eq:5\]), it can be shown that the opacity for the non-interacting gas, $\chi_{\text{ni}}$, is related to the mean-field opacity, $\chi_{\text{mf}}$ by $\chi_{\text{mf}}(E) = \chi_{\text{ni}}
(E+U_n-U_p)$. This relationship produces a large enhancement of the neutrino opacity at high densities, $\rho \approx 10^{14}$, where $U_n
- U_p \approx 50$ MeV when compared with the non-interacting approximation used in ref. [@Fischer.Martinez-Pinedo.ea:2012]. For antineutrino absorption, due to the fact that positrons follow Boltzmann statistics, the non-interacting emissivity and mean-field emissivities are related by $j_{\text{mf}}(E)=j_{\text{ni}}(E-U_n+U_p)$. The mean-field antineutrino opacity is larger at high densities as final-state Pauli blocking of the neutrons becomes less efficient.
In the following, we explore the impact that a description of opacities consistent with the EoS has on the spectra and luminosities of the emitted neutrinos. We have performed core-collapse supernova simulations based on spherically symmetric radiation hydrodynamics with three-flavor Boltzmann neutrino transport. Since our goal is to explore the differences in neutrino energies and luminosities due to the inclusion of mean-field potentials, we have used a low resolution transport scheme with 12 energy bins and allowed only for radially in and outgoing neutrinos. Despite of its limited resolution, it reproduces the absolute values of luminosities and average energies predicted by higher resolution simulations [@Huedepohl.Mueller.ea:2010; @Fischer.Whitehouse.ea:2010]. Table 1 of [@Fischer.Martinez-Pinedo.ea:2012] list the weak processes considered in our simulations. We use the baryonic high-density EoS from Shen *et al.* [@Shen.Toki.ea:1998a] for matter in nuclear statistical equilibrium (NSE) at temperatures above 0.45 MeV. As the tabulation of Shen *et al.* does not provide the mean-field potentials, we have computed them using eq. (\[eq:6\]). In the non-NSE regime, we use the EoS of ref. [@Timmes.Arnett:1999], which we also use for electrons, positrons and photons in the NSE regime. The simulations are based on the 15 M$_\odot$ progenitor of ref. [@Woosley.Heger.Weaver:2002]. Because spherically symmetric simulations do not result in explosions for such a massive iron-core progenitor, we enhance the neutrino heating rates in the gain region following the scheme of ref. [@Fischer.Whitehouse.ea:2010]. It results in the onset of explosion at about 350 ms post bounce. The simulations are evolved from core collapse, through the explosion up to more than 3 seconds after bounce. During core collapse and post-bounce accretion phases, the mean-field potentials, $U_n,U_p$, are only on the order of several 100 keV in the region of neutrino decoupling, which is located at intermediate densities on the order of $10^{11}$ g cm$^{-3}$. Hence, the inclusion of mean-field potentials does not affect the supernova dynamics prior to the explosion.
![(Color online) The left panel shows the neutrino luminosity (upper) and average energy (lower) evolution. The right panel shows the evolution of the proton-to-nucleon ratio, $Y_e$, and the entropy per nucleon for several mass elements ejected from the PNS surface. The curves shown in blue use neutrino opacities computed using the mean-field potentials of the EoS [@Shen.Toki.ea:1998a] while they are neglected on the red curves. \[fig:lumin.tracer\]](tracer-s15.eps){width="\linewidth"}
After the onset of the explosion the neutrinospheres move to increasingly higher densities reaching values of the order of $10^{13}$ g cm$^{-3}$. The left panels of Fig. \[fig:lumin.tracer\] show the evolution of the luminosity and average neutrino energy for all neutrino flavors. These observables are sampled in a co-moving reference frame at a distance of 1000 km. Using charged-current neutrino opacities that include the mean-field potentials slightly reduces the luminosities for all neutrino flavors. Moreover, as expected from the discussion above, it enhances the differences in luminosities and average energies between neutrinos and antineutrinos.
![(Color online) Neutrino spectra for all flavors (solid lines left panel: $\nu_e$, dashed lines left panel: $\nu_{\mu/\tau}$, solid lines right panel: $\bar\nu_e$, dashed lines left panel: $\bar\nu_{\mu/\tau}$), including the mean-field potentials (blue) and without (red). \[fig:spectra\]](spect-alpha.eps){width="\linewidth"}
Fig. \[fig:spectra\] shows the different neutrino spectra for all flavors at 3 s after bounce at a distance of 30 km outside the neutrinospheres. At this distance neutrinos can be considered free streaming but they have not yet been subject to collective neutrino flavor oscillations [@Duan.Fuller.Qian:2010]. These may result in spectral swaps [@Dasgupta.Dighe.ea:2009] that occur in regions near to spectral crossings. We expect substantially different oscillation patterns for the spectra obtained with opacities consistent with the EoS.
The changes in electron (anti)neutrino spectra and luminosities have important consequences for nucleosynthesis in neutrino-driven winds. The increased difference between average energies of $\nu_e$ and $\bar{\nu}_e$ spectra impacts the $Y_e$ of the ejected matter resulting in slightly neutron-rich conditions for the early ejecta (see lower-right panel of Fig. \[fig:lumin.tracer\]) while at later times the ejecta become proton rich. In the simulation that neglects the contributions of the mean-field potentials the ejecta are always proton-rich. The decrease in $Y_e$ is not large enough to favor an r-process but may help in the production of isotopes like $^{92}$Mo that are only made under slightly neutron-rich conditions [@Hoffman.Woosley.ea:1996; @Wanajo:2006]. Also relevant for nucleosynthesis is the slight increase in entropy per nucleon of the ejected material (see upper-right panel Fig. \[fig:lumin.tracer\]) that can be related to the reduced neutrino luminosities [@Qian.Woosley:1996].
We have shown that a treatment of the charged-current (anti)neutrino opacities, that is consistent with the EoS as suggested by [@Reddy.Prakash.Lattimer:1998], has important consequences for the neutrino-spectra evolution during the PNS cooling phase. The most relevant finding is an increased difference between average energies of $\nu_e$ and $\bar{\nu}_e$ that persist during the whole simulation time of 3 seconds after the onset of the explosion. The changes on neutrino spectra are expected to have important consequences for nucleosynthesis, flavor oscillations and neutrino detection on Earth. Our results imply that not only the evolution of the neutrino luminosities [@Roberts.Shen.ea:2012] but also the spectral differences between $\nu_e$ and $\bar{\nu}_e$ are sensitive to the symmetry energy of nuclear matter. Our simulations are based on neutrino opacities computed using the elastic approximation that neglects momentum exchange between nucleons. They need to be extended to consider the full kinematics [@Reddy.Prakash.Lattimer:1998] of the reaction and many-body correlations [@Reddy.Prakash.ea:1999; @*Burrows.Sawyer:1998; @*Burrows.Sawyer:1999] that are expected to become important at later times than those considered in the present study. Furthermore, it is important to explore the sensitivity of the results to different EoS and in particular to EoS that are consistent with recent constrains on the nuclear symmetry energy [@Lattimer.Lim:2012].
G.M.P. is partly supported by the Deutsche Forschungsgemeinschaft through contract SFB 634, the Helmholtz International Center for FAIR within the framework of the LOEWE program launched by the state of Hesse and the Helmholtz Association through the Nuclear Astrophysics Virtual Institute (VH-VI-417). T.F. is supported by the Swiss National Science Foundation under project no. PBBSP2-133378. A.L. is supported by the Helmholtz International Center for FAIR and GSI Helmholtzzentrum für Schwerioneneforschung. L.H. is supported by the Deutsche Forschungsgemeinschaft through contract SFB 634. We thank fruitful discussions with B. Friman, M. Hempel, H.-Th. Janka, K. Langanke, J. M. Lattimer, M. Liebendörfer, B. Müller, F.-K. Thielemann, and S. Typel.
[38]{}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 10.1103/PhysRevD.58.013009) [****, ()](\doibase 10.1016/j.physrep.2007.02.002) @noop [****, ()]{} [****, ()](\doibase 10.1016/0370-1573(92)90083-C) [****, ()](\doibase 10.1086/163343) @noop [ ()]{}, [****, ()](\doibase 10.1086/164587) [****, ()](\doibase 10.1016/S0146-6410(02)00178-3) [****, ()](\doibase 10.1146/annurev.nucl.012809.104524) [****, ()](\doibase 10.1086/168839) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.106.201104) [****, ()](\doibase 10.1086/375130) [****, ()](\doibase 10.1103/PhysRevD.85.083003) [****, ()](\doibase
10.1086/174638) @noop [****, ()]{} [****, ()](\doibase 10.1086/304181) [****, ()](\doibase 10.1103/PhysRevD.63.103004) [****, ()](\doibase 10.1051/0004-6361:20053783) [****, ()](\doibase 10.1103/PhysRevLett.104.251101) [****, ()](\doibase 10.1051/0004-6361/200913106) [****, ()](\doibase
10.1103/PhysRevLett.96.142502) [****, ()](\doibase
10.1086/503891) [****, ()](\doibase 10.1086/505483) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevC.59.2888) [****, ()](\doibase 10.1103/PhysRevC.58.554) [****, ()](\doibase 10.1103/PhysRevC.59.510) [****, ()](\doibase
10.1103/PhysRevLett.108.061103) [****, ()](\doibase 10.1086/191056) @noop [****, ()]{} [****, ()](\doibase 10.1086/313271) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.103.051105) [****, ()](\doibase 10.1086/176986) [****, ()](\doibase 10.1086/177973) @noop [ ()]{},
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We establish renormalizability of the full spectral action for the Yang–Mills system on a flat 4-dimensional background manifold. Interpreting the spectral action as a higher-derivative gauge theory, we find that it behaves unexpectedly well as far as renormalization is concerned. Namely, a power counting argument implies that the spectral action is superrenormalizable. From BRST-invariance of the one-loop effective action, we conclude that it is actually renormalizable as a gauge theory.'
address: 'Institute for Mathematics, Astrophysics and Particle Physics, Radboud University Nijmegen, Heyendaalseweg 135, 6525 AJ Nijmegen, The Netherlands'
author:
- 'Walter D. van Suijlekom'
date: 8 February 2011
title: |
Renormalization of the spectral action\
for the Yang–Mills system
---
Introduction
============
One of the great successes of noncommutative geometry [@C94] is in its application to high-energy physics. Replacing the spacetime manifold by a noncommutative manifold, one puts the full Standard Model of elementary particles on equal geometrical footing as Einstein’s General theory of Relativity. This is worked out in full detail in [@CCM07] (see also [@CM07] and the companion [@CC10]), including the physical predictions that are a consequence of this description.
Being a geometrical description of the Standard Model that is comparable to General Relativity makes it immediate that its quantization comes with the usual problems, actually typical for the latter theory. At the moment, one works with the noncommutative manifold as setting the [*classical*]{} starting point – indeed allowing for a derivation of the full Standard Model Lagrangian at the classical level. Then, one adopts the physics textbook perturbative quantization of it, and arrive at physical predictions via the known Standard Model RG-equations. It needs no stressing that the situation around its quantization should be improved, and in the present letter we intend to take a first step in this direction.
We start with the full asymptotic expansion of the spectral action of Chamseddine and Connes [@CC96; @CC97] in the case of the Yang–Mills system on a flat background manifold. By naive power counting we show – after a suitable gauge-fixing – that the full spectral action is superrenormalizable as a higher-derivative gauge theory [@Sla71; @Sla72b] (cf. [@FS80 Section 4.4]). Then, we demonstrate that the needed counterterms are gauge invariant polynomials that can safely be added to the spectral action. This shows renormalizability of the full spectral action for the Yang–Mills action, compatibly with gauge invariance.
The Yang–Mills system
=====================
The object of study in this paper is the spectral action for the Yang–Mills (YM) system on a flat background manifold. It is given by the relatively simple formula: $$S[A] := \operatorname{Tr}f(D_A/\Lambda).$$ This [*spectral action*]{} has firm roots in the noncommutative geometrical description of the Yang–Mills system, we refer to [@CCM07] for more details. For our purposes, it suffices to know that $D_A$ is a Dirac operator with coefficients in a $SU(N)$-vector bundle equipped with a connection $A$. That is, locally we have $$D_A = i \gamma^\mu \nabla_\mu + \gamma^\mu A_\mu.$$ with $\nabla_\mu$ the spin connection on a Riemannian spin manifold $M$. For simplicity, we take $M$ to be flat ([*i.e.*]{} vanishing Riemann curvature tensor) and 4-dimensional. Furthermore, we will assume that $f$ is a Laplace transform: $$f(x) = \int_{t>0} e^{-tx^2} g(t) dt,$$ even though this assumption could be avoided by using spectral densities instead ([@EGV98] and also [@Var06 Section 8.4])
In the above notation, there is an asymptotic expansion (as $\Lambda \to \infty$): $$\label{sa-eym}
S[A] \sim \sum_{m \geq 0} \Lambda^{4-m} f_{4-m} \int_M a_m (x,D_A^2),$$ in terms of the Seeley–De Witt invariants of $D_A^2$. The coefficients are defined by $f_k := \int t^{-k/2} g(t)dt$.
Recall that the Seeley–De Witt coefficients $a_m(x,D_A^2)$ are gauge invariant polynomials in the fields $A_\mu$. Indeed, the Weitzenböck formula gives $$D_A^2 =- (\partial_\mu - i A_\mu) (\partial^\mu - i A^\mu)+ i\sum_{\mu < \nu}\gamma_\mu \gamma_\nu F_{\mu\nu}$$ in terms of the curvature $F_{\mu\nu} = \partial_\mu A_\nu - \partial_\nu A_\mu -i [A_\mu,A_\nu]$ of $A_\mu$. Consequently, a Theorem by Gilkey [@Gil84 Theorem 4.8.16] shows that (in this case) $a_m$ are polynomial gauge invariants in $F_{\mu\nu}$ and its covariant derivatives. The [*order*]{} $\operatorname{ord}$ of $a_m$ is $m$, where we set on generators: $$\operatorname{ord}A_{\mu_1; \mu_2\cdots \mu_k} = k.$$ Consequently, the curvature $F_{\mu\nu}$ has order $2$, and $F_{\mu_1 \mu_2; \mu_3 \cdots \mu_k}$ has order $k$. For example, $a_4(x,D_A^2)$ is proportional to $\operatorname{Tr}F_{\mu\nu}F^{\mu\nu}$ and more generally: $$a_{4+2k}(x,D_A^2) =c_k \operatorname{Tr}F_{\mu\nu} \Delta^k_A (F^{\mu\nu}) + \cO(F^3)$$ for some constants $c_k$ and the Laplacian $\Delta_A= -(\partial_\mu - i A_\mu)^2$ (see also [@Avr99] and references therein). The remainder is of third and higher order in $F$, plus its covariant derivatives, adding up to an order equal to $4+2k$.
It is the term $a_4$ that gives rise to the Yang–Mills action functional, the higher-order terms are usually ignored (being proportional to an inverse power of the ‘cut-off’ $\Lambda$). More recently, also the higher-order terms, or even the full spectral action were studied in specific cases in [@CC11; @MPT10] and from a more general point of view in [@Sui10].
The quadratic term $S_0[A]$ in $S[A]$ is given asymptotically (as $\Lambda \to \infty$) by $$S_0[A] \sim \sum_{k \geq 0} \Lambda^{-2k} f_{-2k} c_k \int \operatorname{Tr}\hat F_{\mu\nu} \Delta^{k} (\hat F^{\mu\nu})$$ where we have set $\hat F_{\mu\nu} = \partial_\mu A_\nu - \partial_\nu A_\mu$ and $\Delta= -\partial^\mu \partial_\mu$.
We assume that the first term is the usual (free part of the) Yang–Mills action, that is, we adjust the positive function $f$ so that $f_0 c_0 =-1/4$. For the other coefficients, we have the following neat expression.[^1]
The coefficients $f_{-2k}$ are related to the $2k$’th derivatives of $f$ at zero: $$f_{-2k} = \frac{(-1)^k f^{(2k)}(0)}{(2k-1)!!}.$$
With $f(x) = \int e^{-tx^2} g(t)dt$ we derive for its derivatives: $$f^{(2k)}(x) = \int_{t>0} e^{-tx^2/2} H_{2k}(\sqrt t x) t^{k} g(t)dt$$ in terms of the Hermite polynomials $H_{n}(x) \equiv (-1)^n e^{x^2/2} (d/dx)^n e^{-x^2/2}$. Evaluating both sides at zero gives the desired result, using in addition that $H_{2k}(0)= (-1)^k (2k-1)!!$.
We end this section by introducing a formal expansion $
{\varphi}_\Lambda(\Delta) = (f_0c_0)^{-1} \sum_{k \geq0} \Lambda^{-2k} f_{-2k} c_k \Delta^k
$ (starting with $1$) so that we can write more concisely $$S_0[A] \sim -\frac{1}{4} \int \operatorname{Tr}\hat F_{\mu\nu} {\varphi}_\Lambda(\Delta) (\hat F^{\mu\nu})$$ This form motivates the interpretation of $S_0[A]$ (and of $S[A]$) as a higher-derivative gauge theory. As we will see below, this indeed regularizes the theory in such a way that $S[A]$ defines a superrenormalizable field theory.
Gauge fixing in the YM-system
=============================
We add a gauge-fixing term of the following higher-derivative form: $$\label{sa-gf}
S_\gf[A] \sim - \frac{1}{2 \xi} \int \partial_\mu A^\mu {\varphi}_\Lambda(\Delta) \left( \partial_\nu A^\nu \right)$$ We derive the [*propagator*]{} by inverting the non-degenerate quadratic form given by $S_0[A] + S_\gf[A]$: $$D_{\mu\nu}^{ab}(p; \Lambda) = \left[ g_{\mu\nu} - (1-\xi) \frac{p_\mu p_\nu}{ (p^2 +i \eta)}\right] \frac{\delta^{ab}}{(p^2 +i \eta) {\varphi}_\Lambda(p^2)}$$ which for the moment is a formal expansion in $\Lambda$. We will come back to it in more detail in the next section.
As usual, the above gauge fixing requires a Jacobian, conveniently described by a Faddeev–Popov ghost Lagrangian: $$\label{sa-gh}
S_\gh[A,\bar C,C] \sim - \int \partial_\mu \bar C {\varphi}_\Lambda(\Delta) \left( \partial^\mu C + [A^\mu,C] \right)$$ Here $C,\bar C$ are the Faddeev–Popov ghost fields and their propagator is $$\tilde D^{ab}(p; \Lambda) = \frac{\delta^{ab}}{(p^2 + i \eta) {\varphi}_\Lambda(p^2)}.$$
The sum $S[A] + S_\gf[A] + S_\gh[A,\bar C, C]$ is invariant under the BRST-transformations: $$\begin{gathered}
\label{brst}
sA_\mu = \partial_\mu C + [A_\mu,C];\qquad s C = -\half [C,C]; \qquad s \bar C = \xi^{-1} \partial_\mu A^\mu.\end{gathered}$$
First, $s(S)=0$ because of gauge invariance of $S[A]$. We compute $$s(S_\gf) = -\frac{1}{\xi} \int ( \partial_\mu A^\mu) {\varphi}_\Lambda(\Delta) \left( \partial_\nu \partial^\nu C + \partial_\nu( [A^\nu,C] \right)$$ On the other hand, $$s(S_\gh) = - \frac{1}{\xi} \int (\partial_\mu \partial^\nu A_\nu ){\varphi}_\Lambda(\Delta) \left( \partial^\mu C+ [A^\mu,C] \right)$$ which modulo vanishing boundary terms is minus the previous expression.
Note that $s^2 \neq 0$, which can be cured by standard homological methods: introduce an auxiliary (aka Nakanishi-Lautrup) field $h$ so that $\bar C$ and $h$ form a contractible pair in BRST-cohomology. In other words, we replace the above transformation in on $\bar C$ by $s \bar C = --h$ and $s h = 0$. If we replace $S_\gf + S_\gh$ by $s \Psi$ with $\Psi$ an arbitrary [*gauge fixing fermion*]{}, it follows from gauge invariance of $S$ and nilpotency of $s$ that $S + s \Psi$ is BRST-invariant. The above special form of $S_\gf+ S_\gh$ can be recovered by choosing $$\Psi =- \int {\varphi}_\Lambda(\Delta) (\bar C) \left( \half \xi h + \partial_\mu A^\mu \right).$$
One might wonder what gauge fixing condition is implemented by $S_\gf$ as in , given the presence of the term ${\varphi}_\Lambda(\Delta)$. Under suitable conditions on the function $f$, the function $x \mapsto {\varphi}_\Lambda(x)$ is positive, turning the bilinear form $$(\omega_1,\omega_2) := - \int \operatorname{Tr}\omega_1 \wedge \ast ({\varphi}_\Lambda(\Delta) \omega_2)$$ into an inner product. On the Lagrangian level, we can equally well implement the Lorenz gauge fixing condition $\partial \cdot A = 0$ using this inner product instead of the usual $L^2$-inner product. This gives rise to $S_\gf[A] = ( \partial \cdot A, \partial \cdot A)/2\xi$. Similarly, $S_\gf$ is given by the inner product $(\bar C, \partial_\mu C + [A_\mu,C])$.
Renormalization of the spectral action for the YM-system
========================================================
As said, we consider the spectral action for the Yang–Mills system as a higher-derivative field theory. This means that we will use the higher derivatives of $F_{\mu\nu}$ that appear in the asymptotic expansion as natural regulators of the theory, similar to [@Sla71; @Sla72b] (see also [@FS80 Sect. 4.4]). However, note that the regularizing terms are already present in the spectral action $S[A]$ and need not be introduced as such. Let us consider the expansion up to order $n$ (which we assume to be at least $8$), [*i.e.*]{} we set $f_{4-m} = 0$ for all $m > n$. Also, assume a gauge fixing of the form and .
Then, we easily derive from the structure of ${\varphi}_\Lambda(p^2)$ that the propagators of both the gauge field and the ghost field behave as $|p|^{-n+2}$ as $|p| \to \infty$. Indeed, in this case: $${\varphi}_\Lambda(p^2) = \sum_{k=0}^{n/2-2} \Lambda^{-2k} f_{-2k} c_k p^{2k}.$$
Moreover, the weights of the interaction in terms of powers of momenta is given by: $$\begin{array}{|c|c|c|}
\hline
\text{vertex} & \text{valence} & \max \# \text{ der}\\
\hline
\hline
\parbox{2cm}{\vspace{1mm}\includegraphics[scale=.2]{./v3.eps}} & 3 & n-3\\[2mm]
\parbox{2cm}{\includegraphics[scale=.2]{./v4.eps}} & 4 & n-4\\[2mm]
\vdots& \vdots & \vdots\\
\parbox{2cm}{\includegraphics[scale=.2]{./vn.eps}} & n & 0 \\[2mm]
\parbox{2cm}{\includegraphics[scale=.2]{./tildev.eps}} & 3 & n-3\\[2mm]
\hline
\end{array}$$ We will use $v_k$ to indicate the number of gauge interaction vertices of valence $k$, and with $\tilde v$ the number of ghost-gauge vertices.
Let us now find an expression for the [*superficial degree of divergence*]{} $\omega$ of a graph consisting of $I$ internal gauge edges, $\tilde I$ internal ghost edges, $v_k$ valence $k$ gauge vertices and $\tilde v$ ghost-gauge vertices. In 4 dimensions, we find at loop order $L$: $$\omega = 4L - I(n-2) - \tilde I (n-2) + \sum_{i=3}^n v_i (n-i) + \tilde v (n-3).$$
Let $E$ and $\tilde E$ denote the number of external gauge and ghost edges, respectively. The superficial degree of divergence of the graph equals $$\omega = (4-n)(L-1) + 4 - (E+\tilde E).$$
We use the relations $$2 I + E = \sum_i i v_i + \tilde v; \qquad
2 \tilde I + \tilde E = 2\tilde v$$ where $E$ and $\tilde E$ are the number of external gauge and ghost legs, respectively. Indeed, these formulas count the number of half (gauge/ghost) edges in a graph in two ways: from the number of edges and from the valences of the vertices. We use them to substitute for $2I$ and $2\tilde I$ in the above expression for $\omega$ so as to obtain $$\omega = 4L - In - \tilde I n + n \left(\sum_i v_i + \tilde v \right) - (E + \tilde E)$$ from which the result follows at once from Euler’s formula $L= I + \tilde I - \sum_i v_i - \tilde v -1$.
As a consequence, $\omega < 0$ if $L \geq 2$ (provided $n \geq 8$): all Feynman graphs are finite at loop order greater than 1. If $L=1$, then there are finitely many graphs which are divergent, namely those for which $E+ \tilde E \leq 4$. We conclude that the spectral action for the Yang–Mills system is superrenormalizable.
Of course, the spectral action being a gauge theory, there is more to renormalizability than just power counting: we have to establish gauge invariance of the counterterms. We already know that the counterterms needed to render the perturbative quantization of the spectral action finite are of order $4$ or less in the fields and arise only from one-loop graphs. The key property of the effective action at one loop is that it is BRST-invariant: $$s(\Gamma_{1}) = 0.$$ In particular, assuming a regularization compatible with gauge invariance, the divergent part $\Gamma_{1,\infty}$ is BRST-invariant. Results from [@Dix91; @DTV85; @DTV85b; @BDK90; @DHTV91] on BRST-cohomology for Yang–Mills type theories ascertain that the only BRST-closed functional of order 4 or less in the fields is represented by $$\delta Z \int F_{\mu\nu}F^{\mu\nu}$$ for some constant $\delta Z$. The counterterm $\Gamma_{1,\infty}$ can thus be added to $S$ and absorbed by a redefinition of the fields and coupling constant: $$\begin{gathered}
A_0 = \sqrt{1+\delta Z} A ; \qquad g_0 = \frac{g}{\sqrt{1+ \delta Z}}\end{gathered}$$ Equivalently, one could leave $A$ and $g$ invariant, and redefine $f_0 \mapsto (1+\delta Z) f_0$, leaving all other coefficients $f_{4-m}$ invariant. Intriguingly, renormalization of the spectral action for the YM-system can thus be accomplished merely by shifting the function $f$ in such a way that $f(0) \mapsto (1+\delta Z) f(0)$, whilst leaving all its higher derivatives at $0$ invariant.
The above form for $\Gamma_{1,\infty}$ can actually be established by an explicit computation in dimensional regularization following [@PS96; @PS97]. We intend to present the full details elsewhere.
Acknowledgements {#acknowledgements .unnumbered}
================
I would like to thank Klaas Landsman for useful discussions and remarks. I am grateful to Dmitri Vassilevich for useful comments. NWO is acknowledged for support under VENI-project 639.031.827.
[10]{}
I. Avramidi. . 11 (1999) 947–980.
F. Brandt, N. Dragon, and M. Kreuzer. Lie algebra cohomology. B332 (1990) 250.
A. H. Chamseddine and A. Connes. Universal formula for noncommutative geometry actions: [U]{}nifications of gravity and the standard model. 77 (1996) 4868–4871.
A. H. Chamseddine and A. Connes. The spectral action principle. 186 (1997) 731–750.
A. H. Chamseddine and A. Connes. . 58 (2010) 553–600.
A. H. Chamseddine and A. Connes. . To appear.
A. H. Chamseddine, A. Connes, and M. Marcolli. . 11 (2007) 991–1089.
A. Connes. . Academic Press, San Diego, 1994.
A. Connes and M. Marcolli. . AMS, Providence, 2008.
J. A. Dixon. . 139 (1991) 495–526.
M. Dubois-Violette, M. Talon, and C. M. Viallet. algebras: Analysis of the consistency equations in gauge theory. 102 (1985) 105.
M. Dubois-Violette, M. Talon, and C. M. Viallet. Results on [BRS]{} cohomology in gauge theory. B158 (1985) 231.
M. Dubois-Violette, M. Henneaux, M. Talon, and C.-M. Viallet. . B267 (1991) 81–87.
R. Estrada, J. M. Gracia-Bond[í]{}a, and J. C. Várilly. On summability of distributions and spectral geometry. 191 (1998) 219–248.
L. Faddeev and A. Slavnov. . Benjaming Cummings, 1980.
P. B. Gilkey. , volume 11 of [*Mathematics Lecture Series*]{}. Publish or Perish Inc., Wilmington, DE, 1984.
M. Marcolli, E. Pierpaoli, and K. Teh. , 1005.2256.
P. I. Pronin and K. V. Stepanyantz. . 109 (1996) 215–231.
P. I. Pronin and K. V. Stepanyantz. . B414 (1997) 117–122.
A. A. Slavnov. . B31 (1971) 301–315.
A. A. Slavnov. . 13 (1972) 174–177.
W. D. van Suijlekom. Perturbations and operator trace functions. To appear in [*J. Funct. Anal.*]{}
J. C. Várilly. . European Math. Soc. Publishing House (EMS Series of Lectures in Mathematics), 2006.
[^1]: The coefficients $f_{2k}$ for positive $k$ were found to be the $k+1$’th moments of $f$, cf. [@CM07 Sect. 1.11] for more details.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'To explain the multi-wavelength light curves (from radio to X-ray) of HST-1 in the M87 jet, we propose an hourglass model that is a modified two-zone system of Tavecchio & Ghisellini (hereafter TG08)$\colon$ a slow hourglass-shaped or Laval nozzle-shaped layer connected by two revolving exponential surfaces surrounding a fast spine, through which plasma blobs flow. Based on the conservation of magnetic flux, the magnetic field changes along the axis of the hourglass. We adopt the result of TG08—the high-energy emission from GeV to TeV can be produced through inverse Compton by the two-zone system, and the photons from radio to X-ray are mainly radiated by the fast inner zone system. Here, we only discuss the light curves of the fast inner blob from radio to X-ray. When a compressible blob travels down the axis of the first bulb in the hourglass, because of magnetic flux conservation, its cross section experiences an adiabatic compression process, which results in particle acceleration and the brightening of HST-1. When the blob moves into the second bulb of the hourglass, because of magnetic flux conservation, the dimming of the knot occurs along with an adiabatic expansion of its cross section. A similar broken exponential function could fit the TeV peaks in M87, which may imply a correlation between the TeV flares of M87 and the light curves from radio to X-ray in HST-1. The Very Large Array (VLA) 22 GHz radio light curve of HST-1 verifies our prediction based on the model fit to the main peak of the VLA 15 GHz radio light curve.'
author:
- 'Wen-Po Liu, Guang-Yao Zhao, Yong Jun Chen, Chun-Cheng Wang, and Zhi-Qiang Shen'
title: 'AN HOURGLASS MODEL FOR THE FLARE OF HST-1 IN M87'
---
INTRODUCTION AND OBSERVATION CONSTRAINTS
========================================
As is well known, HST-1 is the innermost knot of the M87 jet, located $\sim$80 pc from the core (Biretta et al. 1999). The multi-wavelength light curves of HST-1 have been previously studied using radio data (Chang et al. 2010), optical and UV data (Perlman et al. 2003; Madrid 2009) and X-ray data (Harris et al. 2003, 2006, 2009). Chen et al. (2011) investigated the radio polarization and spectral variability of HST-1, and Perlman et al. (2011) researched the optical polarization and spectral variability of the M87 jet. Cheung et al. (2007) argued that HST-1 may be the site of the flaring TeV gamma-ray emission reported by the H.E.S.S. (Aharonian et al. 2006). A study of particular note was by Abramowski et al. (2012), who released 10 yr of multi-wavelength observations of M87 and the very high energy $\gamma$-ray flare of 2010. The quasi-simultaneous spectrum (from the radio to X-ray band; e.g., Marshall et al. 2002; Waters & Zepf 2005; Perlman & Wilson 2005; Harris et al. 2006; Cheung et al. 2007) and polarization observations (e.g., Perlman et al. 1999, 2011; Chen et al. 2011) of the M87 knots demonstrate the nature of synchrotron radiation.
From the multi-wavelength light curves of HST-1 (Fig. 1), we could obtain some physical constrains on HST-1: the light curves show two big flares, with the main peaks of the light curves around the year 2005.30 and the second ones around the year 2007.
Here, we discuss a pure (or single) process (Doppler effect). For synchrotron emission in the case of a moving sphere, with observed fluxes $I_{\nu,\mbox{\scriptsize obs}}=\delta^{3+\alpha}I_{\nu}$ (Dermer 1995; $\delta$ is the Doppler factor of HST-1 and $\alpha=(p-1)/2$, where $p$ is the spectral index of the particles), the change of the Doppler factor may explain the change in the light curve. However, Harris et al. (2006) suggested a modest beaming synchrotron model with a Doppler factor of three or four, while Wang & Zhou (2009) obtained the Doppler factor of HST-1 to be $3.57\pm0.51$ through a synchrotron model fitting. The HST-1 complex could be model fitted with multiple components (e.g., Cheung et al. 2007; Giroletti et al. 2012); in this case, the speed of one component in HST-1 is obtained through a longer monitoring of the same one, but the velocity estimate usually has fairly large uncertainties. Giroletti et al. (2012) reported the apparent velocities of very long baseline interferometry (1.7 GHz Very Long Baseline Array, VLBA and 5 GHz European VLBI Network, EVN) components in HST-1 with high precision ($<$$2\%$, this is an unprecedented accuracy for determining the apparent velocities of components in HST-1) during the decay period of the HST-1 flare, which implies that components in HST-1 have uniform motion with high precision. Hence, the variation range of the Doppler factor may be very small and the change in the Doppler factor in HST-1 may not explain the order of the flux change.
Excluding the aforementioned single mechanism, we believe that the flare of HST-1 may be correlated with some complex processes which include a changing magnetic field strength.
In Section 2, we describe in detail our model for HST-1. In Section 3, we present and discuss the fitting results of this model to the main peak in the multi-wavelength light curves of HST-1. A summary is provided in Section 4.
The Model
=========
Tavecchio & Ghisellini (2008, TG08) suggested a two-zone scenario in subparsec-scale jets to explain the TeV emissions in M87: a slow hollow cylindrical layer (the velocity relative to the M87 core is $v_l$) surrounds a fast cylindrical zone (the velocity relative to the M87 core is $v_s$, $v_s\gg v_l$, and the velocity of the inner zone relative to the cylindrical layer is $v$). The high energy from GeV to TeV could be produced through inverse Compton by the two-zone, and the photons from radio to X-ray are mainly radiated by the fast inner zone. If this subparsec-scale structure is located within HST-1, HST-1 may be a TeV emission source. However, the model of TG08 could not explain the multi-wavelength light curves (or flare) from radio to TeV. We found that a modified scheme of the TG08 model could achieve this; here we only discuss the light curves of the inner blob from radio to X-ray. We believe that the slow layer may be an hourglass-shaped or Laval nozzle-shaped layer connected by two revolving exponential surfaces (Fig. 2). Considering magnetic field conservation, we believe that the magnetic field will change along the axis of the hourglass. We assumed that the length of the inner blob along the jet axis may be smaller than the radius $R_s$ of its cross section, which may be subparsec-scale, and may remain unchanged (the blob may be constrained in a series of shocks along the jet axis). If the axis coordinate and radius of the hourglass nozzle are $x_n$ and $R_n$, respectively, then the layer radius $R$ (we assumed that $R\sim R_s$) as a broken exponential function of the axis coordinate is $$R=\left\{
\begin{array}{ll}
R_n e^{k(x_n-x)},&\mbox{~$x<x_n$;}\\
R_n e^{k^{'}(x-x_n)},&\mbox{~$x>x_n$,}\\
\end{array}
\right.$$
where $k$ and $k^{'}$ are constant.
Considering magnetic field conservation, $B\propto R^{-2}$, the magnetic field $B$ along the axis of the hourglass will be $$B=\left\{
\begin{array}{ll}
B_n e^{-2k(x_n-x)},&\mbox{~$x<x_n$;}\\
B_n e^{-2k^{'}(x-x_n)},&\mbox{~$x>x_n$,}\\
\end{array}
\right.$$ where $B_n$ is the magnetic field of the hourglass nozzle.
When a blob travels down the axis of the first bulb in the hourglass, because of magnetic field conservation, its cross section experiences an adiabatic compression (compression velocity $u=-(dR/dt)=R_n k v e^{k(x_n-x)}$), which results in particle acceleration and the brightening of HST-1 (this may explain why Perlman et al. 2011 found no evidence for the motion of the flaring blob of HST-1 in their optical data from *Hubble Space Telescope* (*HST*) observations but Giroletti et al. 2012 found no prominent stationary components in their radio data from VLBA and EVN observations). When the blob moves into the second bulb of the hourglass, considering magnetic field conservation, the dimming of the knot will occur along with an adiabatic expansion of its cross section ($u=-(dR/dt)=-R_n k^{'} v e^{k^{'}(x-x_n)}$).
When a blob approaches the hourglass nozzle, the energy gains of the particles in the blob are $(\partial E/\partial t)=\alpha_2 E$, where $\alpha_2=(2/3)(u/R)=(2/3)kv>0$ (Pacholczyk 1970). The compression timescale $\tau=(R/u)=(2/3\alpha_2)$.
A modified continuous injection model, which considers that source particles (a broken power law which could be interpreted as partial loss or escape of high-energy particles in the acceleration region) are continuously injected into an adjacent radiation region from an acceleration region, can be fit to the spectral energy distribution of the outer jet in M87 (Liu & Shen 2007; Sahayanathan 2008 presented a similar model with two spectral indices). We suppose that the radiation mechanism of the inner jet is the same as that of the outer jet. When the source particles (a broken power law) have been injected for the time interval of $t_1$, the radiation region of a blob in HST-1 moves into the hourglass layer. Then, the initial condition of the particle spectra (this is a relic of a past process and may not be ignored for the next process) in HST-1 at the beginning of the compression process is (Liu & Shen 2007; Sahayanathan 2008)
$$N^{*}(E, \theta, 0)=\left\{
\begin{array}{ll}
q_1^{*} t_1 E^{-p_1},&\mbox{~$E\ll E_b$;}\\
q_2^{*} t_1 E^{-(p_2+1)},&\mbox{~$E_b\ll E\ll \frac{1}{\beta_0 t_1}$,}\\
\end{array}
\right.$$
where $N^{*}$, $q_1^{*}$, and $q_2^{*}$ refer to the entire radiation region in the blob taken as an entity (but $N$, $q_1$, and $q_2$ are applied to a unit volume); $\theta$ is the pitch angle between the magnetic field and the particle; $E_b$ is the break energy of a broken power law (here, we assumed that $E_b<(1/\beta_0 t_1)$; Sahayanathan 2008); $p_1$ and $p_2$ denote particle spectrum indices that may be different (e.g., Sahayanathan 2008); $\beta_0=bB^2_{\perp}$; $b$ is a constant; and $B_{\perp}$ represents the component of the magnetic field perpendicular to the velocity of the particle.
Because the acting timescale of the compression process is shorter than that of the particle injection process, which may be comparable with the kinetic timescale of HST-1 relative to the M87 core, we can ignore the influence of particle injection in the compression process. Next, we consider synchrotron radiation and an adiabatic compression of the radiation region in the blob, and the kinetic equation is (Kardashev 1962)
$$\frac{\partial N^{*}}{\partial t}=-\alpha_2 \frac{\partial}{\partial
E}(EN^{*})+\beta \frac{\partial}{\partial E}(E^2 N^{*}),$$
Under the initial conditions given in Equation (3), we have
$$N^{*}(E, \theta, t)=
q^{*}t_1 E^{-\lambda }[1-E e^{-a_2}\int_{0}^{t} \beta e^{a_2} dt]^{\lambda-2}e^{(\lambda-1)a_2},$$
where $a_2=\int_{0}^{t} \alpha_2 dt$, $q^{*}$ represents $q_1^{*}$ or $q_2^{*}$, $\lambda$ is a substitute for $p_1$ or $p_2+1$, and we have assumed that $E\ll (e^{a_2}/\int_{0}^{t} \beta e^{a_2} dt)$ and $\alpha_2=$const in the formula (5). Hence,
$$N^{*}(E, \theta, t)=\left\{
\begin{array}{ll}
q_1^{*} t_1 E^{-p_1}e^{(p_1-1)\alpha_2 t}[1-\frac{\beta}{7\alpha_2 }E(1-e^{-7\alpha_2 t})]^{p_1-2},&\mbox{~$E\ll E_b$;}\\
q_2^{*} t_1 E^{-(p_2+1)}e^{p_2\alpha_2 t}[1-\frac{\beta}{7\alpha_2}E(1-e^{-7\alpha_2 t})]^{p_2-1},&\mbox{~$E_b\ll E\ll \frac{7\alpha_2}{\beta (1-e^{-7\alpha_2 t})}$.}\\
\end{array}
\right.$$
Considering the conservation of magnetic flux, $B\propto R^{-2}$, $R=R_0 e^{-k v t}=R_0 e^{-(3/2)\alpha_2 t}$, where $R_0$ is the initial radius of the blob. Hence, $\beta=\beta_0 e^{6\alpha_2 t}$. The factors $[1-(\beta/7\alpha_2)E(1-e^{-7\alpha_2 t})]^{p_1-2}$ and $[1-(\beta/7\alpha_2)E(1-e^{-7\alpha_2 t})]^{p_2-1}$ in Formula (6) are close to 1, because of the following reasons. (1) $E\ll (7\alpha_2/\beta (1-e^{-7\alpha_2 t}))$, and so $1-(\beta/7\alpha_2)E(1-e^{-7\alpha_2 t})\rightarrow 1$. (2) $p_1, p_2\sim 2$ in the M87 jet (Perlman & Wilson 2005; Liu & Shen 2007; Perlman et al. 2011), hence $p_1-2\sim 0$ and $p_2-1\sim 1$.
Formula (6) is thus close to $$N^{*}(E, \theta, t)\approx \left\{
\begin{array}{ll}
q_1^{*} t_1 E^{-p_1}e^{(p_1-1)\alpha_2 t},&\mbox{~$E\ll E_b$;}\\
q_2^{*} t_1 E^{-(p_2+1)}e^{p_2\alpha_2 t},&\mbox{~$E_b\ll E\ll \frac{7\alpha_2}{\beta (1-e^{-7\alpha_2 t})}$,}\\
\end{array}
\right.$$
and when this is converted to unit volume (entire volume $V\propto R^2 \propto e^{-3\alpha_2 t}$), $$N(E, \theta, t)=\frac{N^{*}}{V}\approx \left\{
\begin{array}{ll}
q_1 t_1 E^{-p_1}e^{(p_1+2)\alpha_2 t},&\mbox{~$E\ll E_b$;}\\
q_2 t_1 E^{-(p_2+1)}e^{(p_2+3)\alpha_2 t},&\mbox{~$E_b\ll E\ll \frac{7\alpha_2}{\beta (1-e^{-7\alpha_2 t})}$.}\\
\end{array}
\right.$$
Next, we consider the increase in magnetic field strength and the time dependence of blob length along the line of sight ($\propto e^{-(3/2)\alpha_2 t}$). If the distributions of particles are isotropic, we can derive the flux formula of the synchrotron model: $$I_{\nu}\propto\left\{
\begin{array}{ll}
e^{(5p_1+4)\alpha_2 t/2}\nu^{-(p_1-1)/2},&\mbox{~$\nu\ll \nu_{B1}$;}\\
e^{(5p_2+9)\alpha_2 t/2}\nu^{-p_2/2},&\mbox{~$\nu_{B1}\ll \nu\ll \nu_{B2}$,}\\
\end{array}
\right.$$ where $\nu_{B1}$ and $\nu_{B2}$ are break frequencies.
When the radiation region of the blob in HST-1 moves away from the hourglass and into the adjacent bulb, deceleration of particles in the blob may occur. Then, $\alpha_2 ^{'}=-(2/3)k^{'}v<0$, the expansion timescale $\tau^{'}=-(2/3\alpha_2 ^{'})$, and the formula for the particle spectrum will become $$N^{'}(E, \theta, t)\approx \left\{
\begin{array}{ll}
q_1 t_1 E^{-p_1}e^{(p_1+2)\alpha_2 t_2}e^{(p_1+2)\alpha_2 ^{'} (t-t_2)},&\mbox{~$E\ll E_b$;}\\
q_2 t_1 E^{-(p_2+1)}e^{(p_2+3)\alpha_2 t_2}e^{(p_2+3)\alpha_2 ^{'} (t-t_2)},&\mbox{~$E_b\ll E\ll \frac{7\alpha_2^{'}}{\beta^{'} [1-e^{-7\alpha_2^{'}} (t-t_2)]}$,}\\
\end{array}
\right.$$
where $t_2$ is the acting timescale of the compression process, $\beta^{'}\propto e^{6 \alpha_2^{'} (t-t_2)}$.
Now, we consider the decrease of magnetic field strength and the time dependence of blob length along the line of sight ($\propto e^{-(3/2)\alpha_2^{'} (t-t_2)}$). With this consideration, the flux expression would be changed to $$I_{\nu}^{'}\propto\left\{
\begin{array}{ll}
e^{(5p_1+4)\alpha_2 ^{'} t/2}\nu^{-(p_1-1)/2},&\mbox{~$\nu\ll \nu_{B1}$;}\\
e^{(5p_2+9)\alpha_2 ^{'} t/2}\nu^{-p_2/2},&\mbox{~$\nu_{B1}\ll \nu\ll \nu_{B2}^{'}$,}\\
\end{array}
\right.$$
where $\nu_{B1}$ and $\nu_{B2}^{'}$ are break frequencies. According to Formulae (9) and (11), the flux $f$ changes exponentially with time in our model, i.e., for low frequency, $$f\propto\left\{
\begin{array}{ll}
e^{(5p_1+4) \alpha_2 t/2},&\mbox{~$t<t_{\mbox{\scriptsize peak}}$;}\\
e^{(5p_1+4) \alpha_2 ^{'} t/2},&\mbox{~$t>t_{\mbox{\scriptsize peak}}$,}\\
\end{array}
\right.$$
where $t_{\mbox{\scriptsize peak}}$ denotes the peak time of the light curve.
Further, for high frequency, $$f\propto\left\{
\begin{array}{ll}
e^{(5p_2+9) \alpha_2 t/2},&\mbox{~$t<t_{\mbox{\scriptsize peak}}$;}\\
e^{(5p_2+9)\alpha_2 ^{'} t/2},&\mbox{~$t>t_{\mbox{\scriptsize peak}}$.}\\
\end{array}
\right.$$
FITTING RESULTS AND DISCUSSION
==============================
Now, we apply the above hourglass model to the multi-wavelength light curves of HST-1 in the M87 jet. Here, the beaming factor of a blob is assumed to be constant.
The data we used are plotted in Fig. 1. The units of the radio data are 1 Jy. VLBA 15 GHz radio data from Chang et al. (2010), in which relative uncertainty is assumed to be $5\%$, are plotted as up triangles and down triangles that show the upper limits of fluxes. Very Large Array (VLA) 15 GHz radio data from Harris et al. (2009) and Abramowski et al. (2012) are plotted as squares. The units of the UV data are 1 mJy (Madrid 2009); these data are plotted as circles. For the X-ray data, we use the flux density integrated from 0.2 to 6 keV (Harris et al. 2006, 2009), and the units are $10^{-11}$ erg cm$^{-2}$ s$^{-1}$; these data are plotted as diamonds. The X-ray data after 2005 August 6 were estimated by assuming that the correction factor for the (unknown) effective area (Harris et al. 2006) is the same as that for 2005 August 6. Harris et al. (2006) estimate that the resulting uncertainties are of the order of $15\%$. Note that because of the unknown effective area, the uncertainties of the X-ray fluxes are larger than those in the counts, which may reduce (or smooth) the individual oscillations of the X-ray fluxes. In our model, as the time dependence of the effective area is considered, the resulting uncertainties of the X-ray fluxes may be larger than $15\%$.
Based on the shape trends of the radio, UV, and X-ray light curves in HST-1, as shown in Fig. 1, we select a common shape section (from about the year 2003 to 2006.60), which may be the main peak in our fitting area. We used the weighted least-squares method to fit our model to this main peak, with Equation (12) corresponding to the radio and UV light curves and Equation (13) corresponding to the X-ray light curve. Although there is a peak time in our model, its exact position was unknown in the fitting of our model to the light curve. Hence, we first arbitrarily divided the light curve data during the chosen period into two groups to perform the least-squares method, in which the sum of the reduced chi square $\chi^2_{\nu}$ for the two parts is the least. Then we can calculate the corresponding least $\chi^2_{\nu}$ by changing the division of the two groups. All reasonable combinations of the two groups (e.g., the fitting peak time should lie between the two groups) are considered before we obtain the best fit with a minimal $\chi^2_{\nu}$. These best-fit parameters for each light curve are shown in Fig. 1 by blue solid lines. Note that the reduced chi square value depends on uncertainties of the data.
Our model can well fit the main peak in the multi-wavelength light curves of HST-1 as shown in Fig. 1. It satisfies the aforementioned first constraint for the multi-wavelength light curves of HST-1: the main peak time is around the year 2005.30, as shown in Fig. 1 and Table 1. The stratified effect of radiation regions in the outer knots of M87 has been verified by Perlman et al. (1999), Marshall et al. (2002), and Perlman $\&$ Wilson (2005). The flattened peak section in the radio light curve can be explained by the greater length of the radio radiation region along the jet axis than the UV and X-ray region. Abramowski et al. (2012) showed that the VLA 22 GHz radio light curve of HST-1 is consistent with the 15 GHz light curve within the error range, which implies that our model can also explain the VLA 22 GHz light curve. In other words, the observation verifies our prediction based on the model fit to the main peak of the VLA 15 GHz light curve.
The aforementioned case is a single component; however, a complex structure that contains multiple components may be more reasonable, as in this case, each peak corresponds to a component passing through the hourglass nozzle. We also use the aforementioned method to fit our model to the possible secondary peak of the multi-wavelength light curves. The best fits for our model are plotted in Fig. 1 in red dotted and dot-dashed lines.
A similar broken exponential function could fit the TeV peaks in M87 (Abramowski et al. 2012), which may imply a correlation between the TeV flares of M87 and the light curves from radio to X-ray in HST-1. Further, the maximum of the TeV flares of M87 was coincident with the peak of light curves from radio to X-ray in HST-1 observed in 2005 (Cheung et al. 2007); this may imply that the observable TeV flux density was produced through inverse Compton as a blob passed through the hourglass nozzle. The detailed generation process of the TeV flare may be very complicated.
The estimates of $p_1$ and $p_2$ in HST-1 require simultaneous broad observational data, which are scarce. Liu & Shen (2007) found that the averaged spectral index of outer knots in M87 is about 2.36 through model fitting. If we assumed that the spectral index of HST-1 is similar to that of outer knots (i.e., $p_1, p_2\approx 2.36$), we could derive some physical parameters of HST-1 from our model fits, as shown in Table 2. Based on the timescale of $\sim$ 5.6 yr, the local size of the component is smaller than a parsec. The derived $\alpha_2$ agrees well with the derived $\alpha_2^{'}$, and the common value is around 0.12, which implies that the hourglass-shaped layer may be symmetrical with respect to the nozzle.
CONCLUSION
==========
We propose a modified two-zone system of TG08: a fast blob passes through a slow hourglass-shaped layer that is connected by two revolving exponential surfaces. Mainly, the emissions from radio to X-ray are radiated by the fast blob. Because of magnetic flux conservation, the brightening and dimming of HST-1 could be explained as adiabatic compression and expansion, respectively, of a blob passing through the outer layer. The observable TeV flux density may be produced through inverse Compton as a blob passes through the nozzle. The VLA 22 GHz radio light curve of HST-1 was used to verify our model.
ACKNOWLEDGMENT
==============
We are grateful to Prof. D. E. Harris for his help in the VLA 15 GHz radio light curve of HST-1. We acknowledge the support from the National Natural Science Foundation of China (NSFC) through grants U1231106 and 11273042, the Science and Technology Commission of Shanghai Municipality (12ZR1436100), and the Scientific Research Foundation of the Civil Aviation University of China (09QD15X). This work is partly supported by the China Ministry of Science and Technology under the State Key Development Program for Basic Research (2012CB821800), NSFC (grants 10625314, 11121062, 11173046, 11033007, 10973012, and 11073019), the CAS/SAFEA International Partnership Program for Creative Research Teams, the Strategic Priority Research Program on Space Science, the Chinese Academy of Sciences (Grant No. XDA04060700), and 973 program 2007CB815405.
[crrrrrrrr]{}
X-ray & $\cdots$ & $1.33\pm0.09$ & $\cdots$ & $-1.33\pm0.07$ & $2005.26^{+0.04}_{-0.14}$ & $1.35$ (26)\
UV & $0.88\pm0.04$ & $\cdots$ & $-0.85\pm0.12$ & $\cdots$ & $2005.37^{+0.10}_{-0.02}$ & $8.99$ (17)\
Radio (VLA 15 GHz) & $0.80\pm0.06$ & $\cdots$ & $-1.14\pm0.17$ & $\cdots$ & $2005.67^{+0.46}_{-0.34}$ & $0.66$ (5)\
Radio (VLBA 15 GHz) & $1.02\pm0.13$ & $\cdots$ & $-0.92\pm0.51$ & $\cdots$ & $2005.04^{+0.26}_{-0.12}$ & $40.80$ (2)\
[crrrrrrrrrr]{} X-ray & $0.13\pm 0.01$ & $5.21\pm 0.35$ & $-0.13\pm0.01$ & $5.21\pm 0.27$\
UV & $0.11\pm 0.01$ & $5.98\pm 0.27$ & $-0.11\pm0.02$ & $6.20\pm 0.87$\
Radio (VLA 15 GHz) & $0.10\pm 0.01$ & $6.58\pm 0.49$ & $-0.14\pm0.02$ & $4.62\pm 0.69$\
Radio (VLBA 15 GHz) & $0.13\pm 0.02$ & $5.16\pm 0.66$ & $-0.12\pm0.06$ & $5.72\pm 3.17$\
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We live in a modern world supported by large, complex networks. Examples range from financial markets to communication and transportation systems. In many realistic situations the flow of physical quantities in the network, as characterized by the loads on nodes, is important. We show that for such networks where loads can redistribute among the nodes, intentional attacks can lead to a [*cascade*]{} of overload failures, which can in turn cause the entire or a substantial part of the network to collapse. This is relevant for real-world networks that possess a highly heterogeneous distribution of [*loads*]{}, such as the Internet and power grids. We demonstrate that the heterogeneity of these networks makes them particularly vulnerable to attacks in that a large-scale cascade may be triggered by disabling a [*single*]{} key node. This brings obvious concerns on the security of such systems.'
author:
- 'Adilson E. Motter'
- 'Ying-Cheng Lai'
title: 'Cascade-based attacks on complex networks'
---
\[Phys. Rev. E [**66**]{}, 065102 (2002)\]
Complex networks are an essential part of a modern society [@Strogatz:2001; @AB:2002]. It has been shown that many networks, such as the world-wide web (WWW), the Internet, and electrical power grids, present a surprisingly small average distance between nodes and a highly organized distribution of links per node [@WS:1998; @BA:1999; @AJB:2000]. Generally, the average distance will not be affected by the removal of a random subset of nodes, but it will increase significantly if the removed nodes are among the most connected ones [@AJB:2000] (see also Refs. [@cohen:2000; @callaway:2000; @cohen:2001]). The existence of a giant connected component in the network, however, does not depend on the presence of highly connected nodes. For instance, the WWW has homepages with many thousands of hyperlinks and can remain well connected after the removal of all homepages with five or more hyperlinks [@Broder:2000]. In addition, the giant component itself is typically a [*small-world network*]{} [@Watts:book] even after the removal of all highly connected nodes [@Alessandro:2002]. These pioneering studies on network security address mainly static properties, $i.e.$, the effect of different network architectures. They suggest that the network connectivity, and hence its functionability, is robust against random failure of nodes [@AJB:2000; @cohen:2000; @callaway:2000] and to some extent is even robust against intentional attacks [@Broder:2000; @Alessandro:2002]. Here we show that for many physical networks, the removal of nodes can have a much more devastating consequence when the intrinsic [*dynamics*]{} of flows of physical quantities in the network is taken into account. In a power transmission grid, for instance, each node (power station) deals with a load of power. The removal of nodes, either by random breakdown or intentional attacks, changes the balance of flows and leads to a global redistribution of loads over all the network. This can trigger a cascade of overload failures [@Watts:2002; @Moreno:2002], as the one that happened on August 10, 1996 in the western United States power grid [@Carreras:2000; @Sachtjen:2000]. Another example is the Internet [@pastor:2001; @willinger:2002; @goh:2002], where the load represents the amount of information a node (router) is requested to transmit per unit of time, and overloads correspond to congestion [@congestions]. Internet collapses caused by congestion have been reported since its very beginning [@Jacobson:1998]. In this Rapid Communication, we introduce a model for cascading failure in complex networks and show that it is applicable to realistic networks such as the Internet and power grids.
For a given network, suppose that at each time step one unit of the relevant quantity, which can be information, energy, etc., is exchanged between every pair of nodes and transmitted along the shortest path connecting them. The load at a node is then the total number of shortest paths passing through the node [@Newman:2001; @Goh:2001; @HK:2002]. The capacity of a node is the maximum load that the node can handle. In man-made networks, the capacity is severely limited by cost. Thus, it is natural to assume that the capacity $C_j$ of node $j$ is proportional to its initial load $L_j$, $$C_j=(1+\alpha)L_j,\;\;\; j=1,2,...N,
\label{capacity}$$ where the constant $\alpha\geq 0$ is the [*tolerance*]{} parameter, and $N$ is the initial number of nodes. When all the nodes are on, the network operates in a free-flow state insofar as $\alpha\geq 0$. But, the removal of nodes in general changes the distribution of shortest paths. The load at a particular node can then change. If it increases and becomes larger than the capacity, the corresponding node fails. Any failure leads to a new redistribution of loads and, as a result, subsequent failures can occur. This step-by-step process is what we call a [*cascading failure*]{}, or a cascade. It can stop after a few steps but it can also propagate and shutdown a considerable fraction of the whole network [@comment]. A fundamental question is: under what conditions can such a global cascade take place?
Here we focus on cascades triggered by the removal of a single node. If a node has a relatively small load, its removal will not cause major changes in the balance of loads, and subsequent overload failures are unlikely to occur. However, when the load at the node is relatively large, its removal is likely to affect significantly loads at other nodes and possibly starts a sequence of overload failures. Our result is the following: global cascades occur if (1) the network exhibits a highly heterogeneous distribution of loads; (2) the removed node is among those with higher load. Otherwise, cascades are not expected. The distribution of loads is in turn highly correlated with the distribution of links: networks with heterogeneous distribution of links are expected to be heterogeneous with respect to load so that on average, nodes with larger number of links will have higher load [@Goh:2001]. This result confirms the robust-yet-fragile property of heterogeneous networks, which was first observed in Ref. [@AJB:2000] for the attack on [*several*]{} nodes. The cascade effect is important, however, because a large damage can be caused in this case by the attack on a [*single*]{} node. While a network with more links can be more resistant against cascading failures, in practice the number of links is limited by cost.
Now we provide evidence for our result. We study cascades triggered by random breakdown and by intentional attacks. To simulate the former, we choose a trigger at random among all the nodes of the network, as can occur in networks such as power grids [@Carreras:2000]. In the case of attack the targeted node is selected from those with highest loads or largest [*degrees*]{} (number of links at a node). We consider heterogeneous networks with algebraic (scale-free) distribution $P$ of links, as observed in real systems [@Redner:1998; @FFF:1999; @BA:1999; @AB:2002], $$P(k)\sim k^{-\gamma},
\label{distribution}$$ where $k$ denotes the degree and $\gamma$ the scaling exponent, and compare them with an equivalent homogeneous configuration. These networks are generated according to the procedure in Refs. [@NSW:2001; @Motter:2002], where the nodes are connected randomly for a given degree distribution, and self- and repeated links are forbidden. The damage caused by a cascade is quantified in terms of the relative size $G$ of the largest connected component, $$G=N'/N,
\label{size}$$ where $N$ and $N'$ are the numbers of nodes in the largest component before and after the cascade, respectively.
Figure \[fig1\] shows the relative size $G$ of the largest component after cascading, as a function of the tolerance parameter $\alpha$, for a scale-free network. We can see that on average $G$ remains close to unity in the case of random breakdowns but it is significantly reduced under intentional attacks, even for $\alpha$ unrealistically large. Indeed, the size of the largest component is reduced by more than $20\%$ for $\alpha=1$, $i.e.$, for a capacity as large as two times the capacity required for the system to operate when all the nodes function normally. This result is in agreement with intuition, because in the case of random breakdown the trigger is probably one of the many nodes with small load, while in the case of intentional attack it is a node with very large load. The damage is larger for smaller values of $\alpha$, as it is for load-based attacks when compared with degree-based attacks. For instance, in the load-based attack for $\alpha=0.2$, more than $60\%$ of the nodes are affected. For the 5000-node networks used in our simulations, it means that a cascade triggered by the attack on a single node shuts down and disconnects more than 3000 others!
Figure \[fig2\] shows the corresponding results for a homogeneous network with the same number of nodes and exactly 3 links per node. To make a meaningful comparison we display in the inset results for an algebraic network with about the same average degree (actually larger, which strengthens our conclusions). The homogeneous network does not experience cascading failures due either to random breakdown or to intentional attacks for $\alpha$ as small as $0.05$. For the heterogeneous (scale-free) network, for the same value of $\alpha$, cascades triggered by the attack on a key node can reduce the largest connected component to less than $10\%$ of the original size, as shown in the inset. Therefore, homogeneous networks appear to be more robust against attacks than the heterogeneous ones. This conclusion does not rely on the particularities of these models, as the same was also observed for classes of networks with exponential and Poisson-like distributions of degrees ($e.g.$, the Erdös-Rényi model [@ER]): their homogeneity makes them relatively resistant to cascades triggered by attacks. The networks corresponding to the inset of Fig. \[fig2\] are generated according to the same scaling distribution of those in Fig. \[fig1\], except that in this case the minimal number of links at a node is set to be 2. Therefore, this inset shows that the fragility of scale-free networks is due to their heterogeneity and does not rely on the presence of nodes with degree one, which are easily disconnectable. Naturally, the increase of the average degree reduces the damage of the cascade, as can be seen from a comparison between Fig. \[fig1\] and the inset of Fig. \[fig2\].
Many real-world networks are heterogeneous and as such are expected to undergo large-scale cascades if some vital nodes are attacked, but rarely in the case of random breakdown. As an example we consider the Internet at autonomous system level [@Internet], which displays an algebraic distribution of links [@AJB:2000]. The damage caused by triggers of higher load or degree is much larger than that by random breakdown, as shown in Fig. \[fig3\]. The cascading failures are rarely triggered by random breakdown for $\alpha > 0.05$, but more than $20\%$ of the nodes can be disconnected with the intentional attack on only one node for $\alpha\leq 0.4$. We have also considered the electrical power grid of the western United States [@power-grid]. The degree distribution in this network is consistent with an exponential [@Amaral:2000] and is thus relatively homogeneous. The distribution of loads, however, is more skewed than that displayed by semirandom networks [@NSW:2001; @Motter:2002] with the same distribution of links, indicating that the power grid has structures that are not captured by these models. As a result, global cascades can be triggered by load-based intentional attacks but not by random or degree-based removal of nodes, as shown in Fig. \[fig4\]. We see that the attack on a single node with large load reduces the largest connected component to less than a half of its initial size, even when the network is highly tolerant ($e.g.$, $\alpha=1$).
Our result is thus that real networks are naturally evolved to be quite resistant to random failure of nodes, but the presence of a few nodes with exceptionally [*large load*]{}, which is known to be ubiquitous in natural and man-made networks, has a disturbing side effect: the attack on a single important node (one of those with high load) may trigger a cascade of overload failures capable of disabling the network almost entirely. Such an event has dramatic consequences on the network performance, because the functionability of a network relies on the ability of the nodes to communicate efficiently with each other. What is the use, say, of having a phone if you cannot call anybody?
We conclude with some thoughts on the meaning of our results for security. An effective attack relies on identifying vulnerabilities and is far from being random. Our society is geographically distributed in a way that natural hazards are by no means random [@editorial]. An example is the crowding of people, communication, transportation, and financial centers around seismic areas, like the Pacific Rim. Natural disasters and intentional attacks can then have devastating consequences on the complex networks underlying the society. These consequences will be more severe if the damage on one or few nodes is capable of spreading over the entire network. In this sense a cascade-based attack can be much more destructive than any other strategies of attack previously considered [@AJB:2000; @Broder:2000; @Motter:2002; @callaway:2000; @cohen:2001; @sole:2001; @jeong:2001; @HKYH:2002].
The authors thank Réka Albert and Duncan J. Watts for providing the Internet and power-grid data, respectively. This work was supported by AFOSR under Grant No. F49620-98-1-0400 and by NSF under Grant No. PHY-9996454.
S.H. Strogatz, Nature (London) [**410**]{}, 268 (2001).
R. Albert and A.-L. Barabási, Rev. Mod. Phys. [**74**]{}, 47 (2002).
R. Albert, H. Jeong, and A.-L. Barabási, Nature (London) [**406**]{}, 378 (2000).
D.J. Watts and S.H. Strogatz, Nature (London) [**393**]{}, 440 (1998).
A.-L. Barabási and R. Albert, Science [**286**]{}, 509 (1999).
R. Cohen, K. Erez, D. ben-Avraham, and S. Havlin, Phys. Rev. Lett. [**85**]{}, 4626 (2000).
D.S. Callaway, M.E.J. Newman, S.H. Strogatz, and D.J. Watts, Phys. Rev. Lett. [**85**]{}, 5468 (2000).
R. Cohen, K. Erez, D. ben-Avraham, and S. Havlin, Phys. Rev. Lett. [**86**]{}, 3682 (2001).
A. Broder, R. Kumar, F. Maghoul, P. Raghavan, S. Rajagopalan, R. Stata, A. Tomkins, and J. Wiener, Comput. Netw. [**33**]{}, 309 (2000).
D.J. Watts, [*Small Worlds: The Dynamics of Networks between Order and Randomness*]{} (Princeton University Press, Princeton, 1999).
A.P.S. de Moura, Y.-C. Lai, A.E. Motter, and P. Dasgupta (unpublished).
D.J. Watts, Proc. Natl. Acad. Sci. U.S.A. [**99**]{}, 5766 (2002).
Y. Moreno, J.B. Gómez, and A.F. Pacheco, Europhys. Lett. [**58**]{}, 630 (2002).
B.A. Carreras, D.E. Newman, I. Dolrou, and A.B. Poole, in Proceedings of Hawaii International Conference on System Sciences, Maui, Hawaii, 2000.
M.L. Sachtjen, B.A. Carreras, and V. E. Lynch, Phys. Rev. E [**61**]{}, 4877 (2000).
R. Pastor-Satorras, A. Vázquez, and A. Vespignani, Phys. Rev. Lett. [**87**]{}, 258701 (2001).
W. Willinger, R. Govindan, S. Jamin, V. Paxson, and S. Shenker, Proc. Natl. Acad. Sci. U.S.A. [**99**]{}, 2573 (2002).
K.-I. Goh, B. Kahng, and D. Kim, Phys. Rev. Lett. [**88**]{}, 108701 (2002).
R. Guimerà, A. Arenas, A. Días-Guilera, and F. Giralt, Phys. Rev. E [**66**]{}, 026704 (2002).
V. Jacobson, Comput. Commun. Rev. [**18**]{}, 314 (1988).
M.E.J. Newman, Phys. Rev. E [**64**]{}, 016132 (2001).
K.-I. Goh, B. Kahng, and D. Kim, Phys. Rev. Lett. [**87**]{}, 278701 (2001).
P. Holme and B.J. Kim, Phys. Rev. E [**65**]{}, 066109 (2002).
A different model and mechanism for overload breakdown in growing networks has been considered by Holme and Kim in Ref. [@HK:2002]. These authors focus on overloads caused by the growth of the network. Their model assigns the same capacity to every node in the network. In their analysis, when a node is overloaded, the links to that node are removed, but the node itself is not removed and can be reconnected in the future. Their conclusion is that, to avoid overloads, the capacity must grow with the size of the network. Our model is different from the model in Ref. [@HK:2002] as we assume the capacity to be [*node dependent*]{} and the failed nodes to be [*permanently removed*]{} from the network. More importantly, we address the issues of [*intentional attack*]{} and [*random breakdown*]{}, and we study how the network collapses under overload failures induced by them. We assume that the time scale for these events is much smaller than the time scale in which the network grows.
S. Redner, Eur. Phys. J. B [**4**]{}, 131 (1998).
M. Faloutsos, P. Faloutsos, and C. Faloutsos, Comput. Commun. Rev. [**29**]{}, 251 (1999).
M.E.J. Newman, S.H. Strogatz, and D.J. Watts, Phys. Rev. E [**64**]{}, 026118 (2001).
A.E. Motter, T. Nishikawa, and Y.-C. Lai, Phys. Rev. E [**66**]{}, 065103 (2002).
P. Erdös and A. Rényi, Publ. Math. Inst. Hung. Acad. Sci. [**5**]{} , 17 (1960).
http://moat.nlanr.net/AS/Data/ASconnlist.20000102.946809601
ftp://ftp.santafe.edu/pub/duncan/power$\_$unweighted
L.A.N. Amaral, A. Scala, M. Barthélémy, and H.E. Stanley, Proc. Natl. Acad. Sci. U.S.A. [**97**]{}, 11149 (2000).
D. Kennedy, Science [**295**]{}, 405 (2002).
R.V. Solé and J.M. Montoya, Proc. R. Soc. London, Ser. B [**268**]{}, 2039 (2001).
H. Jeong, S.P. Mason, A.-L. Barabási, and Z.N. Oltvai, Nature (London) [**411**]{}, 41 (2001).
P. Holme, B.J. Kim, C.N. Yoon, and S.K. Han, Phys. Rev. E [**65**]{}, 056109 (2002).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We investigate primordial magnetogenesis and the evolution of the electromagnetic field through a quantum bounce [@Peter:2006hx], in a model that starts in the far past from a contracting phase where only dust is present and the electromagnetic field is in the adiabatic quantum vacuum state. By including a coupling between curvature and electromagnetism of the form $RF_{\mu \nu}F^{\mu \nu}$, we find acceptable magnetic field seeds within the current observational constraints at 1 mega-parsec (Mpc), and that the magnetic power spectrum evolves as a power-law with spectral index $n_B=6$. It is also shown that the electromagnetic backreaction is not an issue in the model under scrutiny.'
author:
- 'E. Frion$^{1}$ [^1], N. Pinto-Neto$^{1,2}$ [^2], S. D. P. Vitenti$^{3}$ [^3] and S. E. Perez Bergliaffa$^{4}$ [^4]'
bibliography:
- 'bibliography.bib'
title: Primordial Magnetogenesis in a Bouncing Universe
---
Introduction
============
The existence of magnetic fields in a variety of scales in the Universe (see for instance [@Durrer2013; @beck2012magnetic; @Beck:2013bxa]) calls the question of their origin. In particular, there are several observations consistent with weak $\sim 10^{-16}$ Gauss fields in the intergalactic medium, coherent on Mpc scales: the 21-cm hydrogen line [@Minoda:2018gxj], the anisotropy of ultra-high energy cosmic rays [@Bray:2018ipq], CMB distortions [@Ade:2015cva; @Chluba:2019kpb], B-mode polarization measurements [@Zucca:2016iur; @Pogosian:2018vfr], magnetic reheating [@Saga:2017wwr], Big Bang Nucleosynthesis (BBN) [@Kawasaki:2012va], and $\gamma$-rays [@Barai:2018msb], among others. Since such fields remained largely undisturbed during the cosmological evolution (as opposed to those in the presence of structure), they offer a window to their origin, which is generally assumed to be primordial.
Primordial seed fields (which may be amplified later by the dynamo mechanism [@Brandenburg2004]) are generated before structure formation, for instance out of the expansion of the universe, either during inflation [@Martin:2007ue; @Ratra:1991bn; @Davis:1995mv; @Berera:1998hv; @Kandus:1999st; @Bassett:1999ta; @Battaner:2000kf; @Davis:2000zp; @Tornkvist:2000js; @Davis:2005ih; @Anber:2006xt; @Jimenez:2010hu; @Das:2010ywa; @BeltranJimenez:2011vn; @Bonvin:2011dt; @Elizalde:2012kz; @Bonvin:2013tba; @Caprini:2014mja; @Choudhury:2015jaa; @Sharma:2017eps; @Caprini:2017vnn; @Sharma:2018kgs; @Kamarpour:2018ckk; @Savchenko:2018pdr; @Sobol:2018djj; @Subramanian:2019jyd; @Patel:2019isj; @Sobol:2019xls; @Fujita:2019pmi; @Shakeri:2019mnt; @Kobayashi:2019uqs; @Shtanov:2019civ; @Sharma:2019jtb], or in cosmological models with a bounce [@Battefeld:2004cd; @Salim:2006nw; @Membiela2013; @Sriramkumar:2015yza; @Chowdhury2016; @Qian2016; @Koley:2016jdw; @Chen:2017cjx; @Leite:2018bbo; @Chowdhury:2018blx].[^5] However, since minimally-coupled electromagnetism is conformally invariant, the expansion cannot affect its vacuum state. Hence such invariance must be broken in order to generate seed magnetic fields.
Conformal invariance can be broken in several ways: through the addition of a mass term [@Enqvist:2004yy], by coupling the electromagnetic (EM) field to a massless charged scalar field [@Emami2009] or the axion [@Adshead:2016iae], and by a non-minimal coupling with gravity. The last option has been widely studied in the case of inflationary models (see [@Turner1988; @Bamba2006; @Campanelli2008; @Kunze2009; @Kunze2012; @Savchenko2018], among others). However, inflationary magnetogenesis is not free of problems. Among these, we can mention an exponential sensitivity of the amplitude of the generated magnetic field with the parameters of the inflationary model [@Subramanian:2015lua], the strong coupling problem [@Demozzi2009], and the limits in the magnetic field strength coming from the gravitational backreaction of the electric fields that are produced simultaneously with the magnetic fields [@Green2015]. Hence, instead of an inflationary model, a nonsingular cosmological model (see [@Novello2008] for a review) in conjunction with a coupling of the type $RF_{\mu\nu}F^{\mu\nu}$ will be used here to study the production of seed magnetic fields. Nonsingular models are likely to ease both the problem of the exponential sensitivity of the result and the strong coupling problem, since they expand slower than inflationary models. Moreover, we shall see below that backreaction is not an issue for the model chosen here.
It is worth remarking that magnetogenesis in nonsingular cosmological models has been studied before, always in the presence of a scalar field. The models already studied may be divided into two classes, depending on whether the coupling of the EM field with the scalar field is fixed on theoretical grounds (see for instance [@Salim2004; @Salim2006]), or chosen in a convenient way in terms of the expansion factor (see [@Membiela2013; @Qian2016; @Chowdhury2016] ). The coupling between the Ricci scalar and the EM field to be adopted in this work, which is theoretically motivated by the vacuum polarization described quantum electrodynamics (QED) in a curved background [@Drummond:1979pp], introduces a mass scale to be fixed by observations.
We shall start in Sec. \[backg\] with a brief summary of the background model that will be used in what follows. In Sec. \[emsector\], the equations governing the behavior of the perturbations of the electromagnetic field in a curved background will be reviewed. We show in Sec. \[Aresults\] the analytic solutions for the gauge field and its momentum. These results are used to understand the numerical solutions in Sec. \[results\]. The comparison of the results with observations is presented in Sec. \[discussion\]. The fact that backreaction does not affect the background dynamics is shown in Sec. \[backreaction\]. Also, we show in appendix \[sec:adiab\] how to obtain appropriate initial conditions for the electromagnetic field from an adiabatic vacuum.
The background {#backg}
==============
Cosmological models displaying a bounce solve the singularity problem by construction. They join a contracting phase, in which the Universe was initially very large and almost flat, to a subsequent expanding phase. In such models, the curvature scale tends to infinity in the asymptotic past. As a consequence, vacuum initial conditions for cosmological perturbations can be imposed in the dust-dominated contracting phase [^6], leading to a scale invariant spectrum [@Peter:2008qz]. The bounce can be either generated classically (see *e.g.* [@Wands:2008tv; @Ijjas:2016tpn; @Cubero:2019lxw; @Galkina:2019pir]) or by quantum effects (see *e.g.* [@Peter:2008qz; @Almeida:2018xvj; @Bacalhau:2017hja; @Frion:2018oij]).
The cosmological model that will be used here as background was obtained in [@Peter:2006hx] by solving the Wheeler-deWitt equation in the presence of a single perfect fluid. The solution was obtained in the minisuperspace approximation, and in the framework of the theory of de Broglie and Bohm (dBB) [@Bohm:1951xw; @Bohm:1951xx]. The reason behind this choice is that the dBB interpretation is very well suited for cosmology, since it needs no external classical apparatus, as opposed to the Copenhagen interpretation.
The expression of the scale factor in the case of a flat spatial section obtained in [@Peter:2006hx] is given by $$\label{adet}
a(T)=a_b\left(1+\frac{T^2}{T_b^2}
\right)^{\frac{1}{3(1-w)}} \;,$$ where $w$ is the equation of state of the fluid.[^7] All quantities appearing hereafter with the subscript $b$ are evaluated at the bounce (with the exception of $T_b$, which fixes the bounce timescale), while quantities with the subscript $0$ are evaluated today. The spacetime geometry associated with is nonsingular, and the scale factor tends to the classical evolution for $\vert
T\vert \gg T_b$. The relation between $T$ and the cosmic time $t$ is given by $$\begin{aligned}
dt=a^{3w}dT \;.\end{aligned}$$
From now on, we shall set $w=0$, leading to a scale invariant spectrum for the curvature perturbations, and allowing us to set $t=T$. It will also be useful to express the scale factor as $a(t)\equiv a_0 Y(t)$, with $$\begin{aligned}
Y(t)= \frac{1}{x_b} \left( 1+\frac{t^2}{t_b^2}
\right)^{1/3} \;,
\label{yt}\end{aligned}$$ where we have defined $x\equiv a_0/a$ and $t_b \equiv 2\ell_b$, with $\ell_b$ the curvature scale at the bounce ($\ell_b \equiv 1/\sqrt{\vert R(0)\vert}$ where $R$ is the four-dimensional Ricci scalar) satisfying $10^3\:t_{\rm Planck}<t_b<10^{40}\:t_{\rm Planck}$.[^8]
For the subsequent calculations, it is convenient to define parameters that are directly related to observations. Let us first write down the Friedmann equation $$\begin{aligned}
H^2 = \frac{8\pi G}{3} \frac{\rho_m}{a^3} \;,
\label{friedrho}\end{aligned}$$ with $\rho_m$ the dark matter density energy. The ratio between Eq. at some time $t$ and the same equation evaluated today leads to $$\begin{aligned}
H^2 = H^{2}_{0} \Omega_m x^3 \;,
\label{fried}\end{aligned}$$ with $\Omega_m$ the dimensionless dark matter density today. Note that at $x=1$ we have $H^2 = H_0^2\Omega_m$, this means that in the contraction phase, at the same scale as today $a = a_0$, the Hubble factor is $-H_0\sqrt{\Omega_m}$ due to the lack of other matter components. Then, from the expansion of $a(t)$ for large values of $t$, it follows that $$\begin{aligned}
H^2\approx \frac{4}{9t_b^2}\left( \frac{x}{x_b}\right)^3 \;.
\label{larget}\end{aligned}$$ Now, using $H_0=70 \; \textup{km\;s}^{-1}\textup{\;Mpc}^{-1}$ and the lower bound on $t_b$, it is straightforward to derive an upper limit on $x_b$ by equating Eqs. and , $$\begin{aligned}
\Omega_m = \frac 4 9 \frac{1}{t_b^2 x_b^3 H_0^2} \; \implies {x}_b < \frac{10^{38}}{\Omega_m^{1/3}} \; .
\label{limit xb}\end{aligned}$$
For later convenience, we define $R_{H_0} \equiv H_0^{-1}$, $t_s \equiv {t}/{R_{H_0}}$, and $
\alpha \equiv {R_{H_0}}/{t_b}$, and rewrite $Y(t)$ as $$\begin{aligned}
Y(t_s) = \frac{1}{x_b} \left(1+\alpha^2t_s^2\right)^{1/3} \;,\end{aligned}$$ with $$\begin{aligned}
\alpha = \frac 3 2 \sqrt{\Omega_mx_b^3} \;.
\label{alpha}\end{aligned}$$ We will see in the next section how to relate the previous quantities to the electromagnetic power spectrum, and what constraints can be derived on the parameters of the model.
The electromagnetic sector {#emsector}
==========================
To describe electromagnetism we shall adopt the Lagrangian $$\label{lagr}
{\cal L}= -f
F_{\mu\nu}F^{\mu\nu} \;,$$ where $$\begin{aligned}
f \equiv \frac{1}{4}+\frac{R}{m_{\star}^2} \;,
\label{coupling}\end{aligned}$$ and $m_{\star}$ is a mass scale to be determined by observations. As mentioned in the Introduction, the nonminimal coupling in this Lagrangian breaks conformal invariance, and paves the way to the production of primordial electromagnetic fields.
The equations of motion for the electromagnetic field that follow from Eq. are $$\label{eqmov}
\partial_\mu(\sqrt{-g}\:f\:F^{\mu\nu})=0 \;,$$ where the field $F_{\mu\nu}$ is expressed in terms of the gauge potential $A_\mu$ as $F_{\mu\nu}=\partial_\mu A_\nu-\partial_\nu A_\mu$. To quantize the electromagnetic field, we expand the operator associated to the spatial part of the vector potential as $$\label{decomp}
\hat{A}_i(t,\mathbf{x})=\sum_{\sigma=1,2}\int \frac{d^3k}{(2\pi)^{3/2}}\left[\epsilon_{i,\sigma}(\mathbf{k})\hat{a}_{\mathbf{k},\sigma} A_{k,\sigma} (t)e^{i\mathbf{k}\cdot\mathbf{x}}+H.C.\right] \;,$$ where $\epsilon_{i,\sigma}(\mathbf{k})$ are two orthonormal and transverse vectors which are constant across spatial sheets (they have zero Lie derivative with respect to the spatial foliation vector field) and $H.C.$ stands for the Hermitian conjugate. The operators $\hat{a}_{\mathbf{k},\sigma}$ and $\hat{a}^\dagger_{\mathbf{k},\sigma}$ are respectively the annihilation and creation operators. They satisfy $[\hat{a}_{\mathbf{k},\sigma}, \hat{a}^\dagger_{\mathbf{k'},\sigma'}] =
\delta_{\sigma\sigma'}\delta(\mathbf{k} - \mathbf{k'})$, $[\hat{a}_{\mathbf{k},\sigma}, \hat{a}_{\mathbf{k'},\sigma'}]=0$, and $[\hat{a}^\dagger_{\mathbf{k},\sigma},
\hat{a}^\dagger_{\mathbf{k'},\sigma'}]=0$. Note that in the equations above we adopted the Coulomb gauge with respect to the cosmic time foliation ($A_0=0$ and $\partial_i
A^i=0$). The time-dependent coefficients $A_{k,\sigma}(t)$ and their associated momenta $\Pi_{k,\sigma} \equiv
4afA^\prime_{k,\sigma}(t)$ must satisfy $$\label{vacuumnorma}
A_{k,\sigma}(t)\Pi^*_{k,\sigma}(t)
-A^*_{k,\sigma}(t)\Pi_{k,\sigma}(t) = i,$$ for each $k$ and $\sigma$. It should be emphasized that the quantization of the gauge-fixed electromagnetic field in the absence of charges is equivalent to that of two free real scalar fields. Consequently, the choice of vacuum for each polarization $\sigma$ corresponds to the choice of vacuum of each scalar degree of freedom. However, using the fact that we are dealing with an isotropic background, there is no reason to make different choices of vacuum for different polarizations. For this reason, we choose a single time-dependent coefficient to describe both polarizations, i.e., $A_{k,1} = A_{k,2} \equiv{A_{k}} $. Therefore, the same vacuum is chosen for both polarizations. Now, inserting this decomposition in Eq. , we get the equation governing the evolution of the modes $A_k(t)$ $$\begin{aligned}
\label{pot}
\ddot{A}_k+\left(\frac{\dot a}{a}+\frac{\dot f}{f}
\right)\dot{A}_k +\frac{k^2}{a^2}A_k=0 \;.\end{aligned}$$ Defining $$\begin{aligned}
k_s \equiv kR_H,\;\;
\;\;A_{sk}(t_s) \equiv \frac{A_k(t_s)}{\sqrt{x_bR_{H_0}}} \;,\end{aligned}$$ where $R_H=R_{H_0}/a_0$ is the comoving Hubble radius today, the differential Eq. can be written as $$\begin{aligned}
A_{sk}^{\prime\prime}+\left(\frac{ Y^\prime}{Y}+\frac{f^\prime}{f}
\right){A}_{sk}^\prime +\frac{k_s^2}{Y^2}A_{sk}=0 \;,
\label{diffeq}\end{aligned}$$ where a prime denotes the derivative with respect to $t_s$. The coupling then takes the form $$\begin{aligned}
f=\frac 1 4 \left[ 1+C^2\frac{\alpha^2 t_s^2+3}{(\alpha^2t_s^2+1)^2}
\right] \;; \;\;\;\;\; \textup{with}\;\; C^2 \equiv \frac{4}{3} \frac{\ell_*^2}{t_b^2} \;,
\;\;\;\;\;\ell_* \equiv \frac{1}{m_*} \;.
\label{cc}\end{aligned}$$ An upper limit on $C$ can be straightforwardly derived from Eq. . Since any contribution to the usual Maxwell’s equations at BBN must be negligible, we impose the second term in Eq. to be smaller than $10^{-2}$ at BBN. Together with the fact that $\alpha^2 t_s^2 \gg 1$ at this time, we get $$\begin{aligned}
C < 10^{-19} x_b^{3/2} \;.
\label{limit c}\end{aligned}$$ The energy densities of the electric and magnetic fields are respectively given by $$\begin{aligned}
\rho_E &=\frac{f}{8\pi}g^{ij}A_i^\prime A_j^\prime\;, \\
\rho_B &=\frac{f}{16\pi}g^{ij}g^{lm}(\partial_j A_m-\partial_m A_j) (\partial_i A_l-\partial_l A_i) \;,\end{aligned}$$ where $g^{ij}=\delta^{ij}/a^2$ are the spatial components of the inverse metric. To find the spectral energy densities, we first insert expansion into $\rho_E$ and $\rho_B$. The resulting operators $\hat{\rho}_E$ and $\hat{\rho}_B$ upon quantization are $$\begin{aligned}
\hat{\rho}_B &= \frac{f}{2\pi^2 R_{H_0}^4 Y^{4}} \int {\mathrm{d}}{\ln{k}} \; \vert A_{sk}\vert^2 k^5 \;, \label{magnetic spectral density} \\
\hat{\rho}_E &= \frac{f}{2\pi^2 R_{H_0}^4 Y^{2}} \int {\mathrm{d}}{\ln{k}} \; \vert A^{\prime}_{sk}\vert^2 k^3 \;. \label{electric spectral density}\end{aligned}$$ We now evaluate the expectation value of the two densities in vacuum, defined by $\hat{a}_{\mathbf{k},\sigma} {\left\vert{0}\right\rangle}=0$, and define the spectra as $$\begin{aligned}
{{\cal P}_{i}} \equiv \frac{\textup{d}{\left\langle0\right\vert} \hat{\rho}_i {\left\vert{0}\right\rangle}}{{\mathrm{d}}{\ln{k}}} \;, \quad i=E,B \;.\end{aligned}$$ This yields the magnetic and electric spectra, respectively $$\begin{aligned}
{\cal P_{B}} &\equiv B^2_\lambda = \frac{f}{2\pi^2 R_{H_0}^4}\frac{\vert A_{sk}\vert^2}{Y^4}k^5 \;,
\label{magpow} \\
{\cal P_{E}} &\equiv E_{\lambda}^2=\frac{f}{2\pi^2 R_{H_0}^4}\frac{\vert A_{sk}^\prime\vert^2}{Y^2}k^3 = \frac{1}{2\pi^2 R_{H_0}^4}\frac{\vert \Pi_{sk}\vert^2}{fY^4}k^3 \;.
\label{elecpow}\end{aligned}$$ In the last line, we also expressed ${\cal P_{E}}$ in terms of the momentum canonically conjugate to the gauge field $\Pi_{sk} = Yf A_{sk}^\prime$ (see Appendix \[sec:adiab\]), which is nothing but the electric field mode itself.
Finally, we can express the magnetic and electric fields, $B_{\lambda}$ and $E_{\lambda}$, using $H_0^2=1.15\times 10^{-64}$ G $$\begin{aligned}
\label{est}
B_\lambda &=\sqrt{\frac{f}{2\pi^2}}
\frac{\vert A_{sk}\vert }{Y^2}k^{5/2}\:1.15\times 10^{-64} {\rm G} \;, \\
E_\lambda &=\sqrt{\frac{1}{2 \pi^2 f}}
\frac{\vert \Pi_{sk}\vert }{Y^2}k^{3/2}\:1.15\times 10^{-64} {\rm G}.\end{aligned}$$
Analytical results {#Aresults}
==================
In this section, we obtain analytically the time behavior and spectra of $A_{k}$ satisfying Eq. (from now on the index $s$ on the time variable and wavenumber will be omitted), and its canonical momentum $\Pi_k$, in the different stages of the cosmic evolution. In the sequel, this analysis will be compared with the numerical results.
As shown in Appendix \[sec:adiab\], the adiabatic vacuum is a consistent choice for the EM field initial conditions. The modes in vacuum are $$\begin{split}
\vert A_k\vert &= \sqrt{\frac{2}{k}} +\dots \;, \\
\vert \Pi_k\vert &= \sqrt{\frac{k}{8}} +\dots. \;,
\end{split}$$ and both the field and its canonical momentum are constant in this regime. Now that the initial conditions for the EM field have been defined, we can move on to the analysis of the evolution of the electric and magnetic modes from the far past up to the present day.
Three important characteristic times related to the evolution of the modes are worthy of note . The first is the time limit of the adiabatic regime, $\vert t_c\vert $, defined in Eq. . The second one is the time where quantum effects leading to the bounce take place, *i.e.* $\vert t_b\vert = 1/\alpha$. Consequently, the bounce phase takes place for $t$ such that $-1/\alpha < t < 1/\alpha$. The third one is the characteristic time when the evolution of $f$ becomes important. Examining Eq. , one gets the time $\vert t_f\vert = C/\alpha$, up to $\vert t_b\vert $, which means that the evolution of $f$ is important when $-C/\alpha < t < -1/\alpha$, and $1/\alpha < t < C/\alpha$. The domain of physically allowed parameters imposes that $$\label{times}
\vert t_c\vert \gg \vert t_f\vert \gg \vert t_b\vert \;.$$
For $\vert t\vert <\vert t_c\vert $, the solution leaves the frequency-dominated region. In this case, one can perform the usual expansion in $\nu^2$ derived from the Hamilton Eqs. through iterative substitutions: $$\label{iterations1}
\begin{split}
\Pi_k(t) &= -\int^t m(t_1)\nu^2(t_1) A_k(t_1){\mathrm{d}}t_1 + A_2(k) = mA^\prime_k(t)\Rightarrow\\
A_k(t) &= -\int^t \frac{{\mathrm{d}}t_2}{m(t_2)}\int^{t_2}m(t_1)\nu^2(t_1) A_k(t_1) {\mathrm{d}}t_1 + A_2(k)\int^t \frac{{\mathrm{d}}t_1}{m(t_1)} + A_1(k)\Rightarrow\\
A_k(t) &= A_1(k)\left(1-\int^t \frac{{\mathrm{d}}t_2}{m(t_2)}\int^{t_2}m(t_1)\nu^2(t_1) {\mathrm{d}}t_1\right) + \\
& A_2(k)\left(\int^t \frac{{\mathrm{d}}t_1}{m(t_1)} - \int^t \frac{{\mathrm{d}}t_2}{m(t_2)}\int^{t_2}m(t_1)\nu^2(t_1){\mathrm{d}}t_1\int^{t_1} \frac{{\mathrm{d}}t_3}{m(t_3)}\right) +\dots \; ,
\end{split}$$ where $A_1(k)$ and $A_2(k)$ are constants in time depending only on $k$, leading to the momentum expression $$\label{iterations2}
\Pi_k(t) = -A_1(k)\int^{t}m(t_1)\nu^2(t_1) {\mathrm{d}}t_1 + A_2(k)\left(1 - \int^t m(t_1)\nu^2(t_1){\mathrm{d}}t_1\int^{t_1} \frac{{\mathrm{d}}t_2}{m(t_2)}\right)+\dots \;.$$ We can now evaluate the time evolution and spectra in the different phases of the cosmic evolution.
The contracting phase and the bounce
------------------------------------
In the case of $A_k(t)$, all time-dependent terms are decaying in the contracting era up to the end of the bounce. As a consequence, $A_k(t)=A_1(k)$ is constant during all this phase. By continuity with the adiabatic phase, we conclude that $$\label{A1}
A_1(k)\propto k^{-1/2} \;.$$
The time-dependent terms of the momentum $\Pi_k(t)$ are also decaying, except for the one multiplying $A_1(k)$, which grows as $t^{-5/3}$ for $-C/\alpha < t < -1/\alpha$, since $f\propto 1/t^2$ in this region. Then, for $t< -C/\alpha$, $\Pi_k(t)=A_2(k)$ which, by continuity with the adiabatic phase, implies that $$\label{A2}
A_2(k)\propto k^{1/2}.$$ In the period $-C/\alpha < t < -1/\alpha$, the term multiplying $A_1(k)$ eventually surpasses the constant mode at a time $t_\pi$, and $\Pi_k(t)$ grows.
At the bounce itself $Y$ and $f$ are almost constant, therefore the modes will not evolve during this phase.
The expanding phase
-------------------
In the expanding phase, the most important growing function related to $A_k(t)$ is the first one multiplying $A_2(k)$, which grows as fast as $t^{7/3}$ starting from some time $t_A$ in the interval $1/\alpha < t < C/\alpha$, and as $t^{1/3}$ for $C/\alpha < t < t_c$.
In the case of $\Pi_k(t)$, as the integral multiplying $A_1(k)$ strongly decreases as $t^{-5/3}$ when $1/\alpha < t < C/\alpha$, the value of $\Pi_k(t)$ saturates in the value it gets by the end of the bounce, $t\approx 1/\alpha$. Also, $\Pi_k(t)$ acquires a $k^2$ dependence through the $\nu^2$ term. Combined with the $k$ dependence of $A_1(k)$, we obtain $\Pi_k(t) \propto k^{3/2}$.
After $t_c$, both $A_k(t)$ and $\Pi_k(t)$ begin to oscillate.
Summary
-------
For the $A$-field, the spectra and time dependence in the different cosmic evolution phases is: $$\label{historyA}
\begin{split}
-\infty < t < t_A \; &: \vert A_k(t)\vert \propto k^{-1/2} \;, \\
t_A < t < C/\alpha \; &: \vert A_k(t)\vert \propto k^{1/2}t^{7/3} \;, \\
C/\alpha < t < k^{-3} \; &: \vert A_k(t)\vert \propto k^{1/2}t^{1/3} \;, \\
t > k^{-3} \; &: \vert A_k(t)\vert \propto k^{1/2} \; \times {\textrm{(oscillatory factors)}},
\end{split}$$ where $t_A \in (1/\alpha,C/\alpha)$.
For the $\Pi$-field, we have: $$\label{historyP}
\begin{split}
-\infty < t < t_\pi \; &: \vert \Pi_k(t)\vert \propto k^{1/2} \;,\\
t_\pi < t < -1/\alpha \; &: \vert \Pi_k(t)\vert \propto k^{3/2} t^{-5/3} \;,\\-1/\alpha< t < k^{-3} \; &: \vert \Pi_k(t)\vert \propto k^{3/2} \;,\\
t > k^{-3} \; &: \vert \Pi_k(t)\vert \propto k^{3/2} \; \times {\textrm{(oscillatory factors)}},
\end{split}$$ where $t_\pi \in (-C/\alpha,-1/\alpha)$.
Note that both the final spectrum of ${\cal P_{B}}$ and ${\cal P_{E}}$ (given in Eqs. and ) go as $k^6$.
After these analytical considerations, let us now turn to the numerical calculations, which confirm the behaviors presented in this section, and allow the calculation of the amplitudes.
Numerical results {#results}
=================
We start this section by showing in Fig. \[ymf\] the time behavior of the coupling $f$ given in Eq. , the scale factor $Y=a/a_0$ from Eq. , and the mass $m=Yf$. From the definition of $|t_f|$ and $|t_b|$ in the previous section, and choosing $C=10^{23}$ and $x_b=10^{30}$, we obtain respectively $|t_f|\simeq 10^{-22}$ and $|t_b|\simeq 10^{-45}$. This is consistent with the behavior shown in the figure.
![Evolution of the coupling $f$, the scale factor $Y$ normalised today, and the mass $m=af$ with time. We have used $C=10^{23}$ and $x_b=10^{30}$.[]{data-label="ymf"}](MagDustYMF_cc23_xb30.pdf)
The numerical evolution of the gauge field $A_{k}$ and its momentum $\Pi_k$ is shown next. In Fig. \[mode evolution cc\], the influence of the parameter $C$ on the evolution of the modes is shown explicitly for $C=10^{19}$ and $C=10^{23}$ with $x_b=10^{30}$, while the influence of $x_b$ is shown in Fig. \[mode evolution xb\] for $x_b=10^{30}$ and $C=10^{36}$ with $C=10^{23}$.[^9] Note that in these figures, as well as in the following ones, we performed the computation for $1<k<4000$, since $k=4000$ implies a physical wavelength of about 1 Mpc (remember that $k$ is in units of Hubble radius). One can verify in these figures all time and $k$ dependence described in Sec. \[Aresults\], summarized in Eqs. and .
![Evolution of the absolute values of the magnetic modes ($A_{k}$) and their momentum ($\Pi_k $) through the bounce in a dust background for $C=10^{19}$ and $x_b=10^{30}$ (top), and for $C=10^{23}$ and $x_b=10^{30}$ (bottom). The same colour for the gauge field and its momentum evolution is chosen for a given $k_s$. We see that larger values of $C$ lead to a higher final amplitude.[]{data-label="mode evolution cc"}](MagDustModeEvol_cc19_xb30.pdf "fig:") ![Evolution of the absolute values of the magnetic modes ($A_{k}$) and their momentum ($\Pi_k $) through the bounce in a dust background for $C=10^{19}$ and $x_b=10^{30}$ (top), and for $C=10^{23}$ and $x_b=10^{30}$ (bottom). The same colour for the gauge field and its momentum evolution is chosen for a given $k_s$. We see that larger values of $C$ lead to a higher final amplitude.[]{data-label="mode evolution cc"}](MagDustModeEvol_cc23_xb30.pdf "fig:")
![Same as Fig. \[mode evolution cc\] for $C=10^{23}$ and $x_b=10^{30}$ (top), and for $C=10^{23}$ and $x_b=10^{36}$ (bottom). We see that larger values of $x_b$ lead to a quicker evolution of the modes.[]{data-label="mode evolution xb"}](MagDustModeEvol_cc23_xb30.pdf "fig:") ![Same as Fig. \[mode evolution cc\] for $C=10^{23}$ and $x_b=10^{30}$ (top), and for $C=10^{23}$ and $x_b=10^{36}$ (bottom). We see that larger values of $x_b$ lead to a quicker evolution of the modes.[]{data-label="mode evolution xb"}](MagDustModeEvol_cc23_xb36.pdf "fig:")
Now that the evolution of the modes has been described, we can use the shape of the spectra that follows from the results in Figs. \[mode evolution cc\] and \[mode evolution xb\], and Eqs and , the last one expressed in terms of the momentum, to fathom the time evolution of the magnetic and electric power spectra shown in Figs \[spectra evolution cc\] and \[spectra evolution xb\]. At the beginning of the evolution, modes are not excited. Only vacuum fluctuations are present, with the usual $k^4$ spectrum, increasing as $Y^{-4}$ due to contraction. When the coupling $f$ becomes relevant, the magnetic field power spectrum begins to increase faster, since $f$ is a growing function in the contracting phase, while the electric field power spectrum presents a slower increment, up to the time when $\Pi_k$ also begins to increase. After the bounce the situation is reversed, because $f$ is a decaying function of time in the expanding phase: the electric power spectrum decreases much slower than the magnetic one. Using Eq. , one can see that the decay is mild, going as $t^{-2/3}$, when $1/\alpha < t < C/\alpha$, opening a window in time where the electric spectrum has a significantly higher contribution than the magnetic one.
![Evolution of the magnetic (dashed lines) and electric (continuous line) power spectra for $C=10^{19}$ and $x_b=10^{30}$ (top), and for $C=10^{23}$ and $x_b=10^{30}$ (bottom). We see that with larger $C$’s, the decrease of the electric contribution at late times happens later, and the total electromagnetic power spectrum is more important.[]{data-label="spectra evolution cc"}](MagDustPBPE_cc19_xb30.pdf "fig:") ![Evolution of the magnetic (dashed lines) and electric (continuous line) power spectra for $C=10^{19}$ and $x_b=10^{30}$ (top), and for $C=10^{23}$ and $x_b=10^{30}$ (bottom). We see that with larger $C$’s, the decrease of the electric contribution at late times happens later, and the total electromagnetic power spectrum is more important.[]{data-label="spectra evolution cc"}](MagDustPBPE_cc23_xb30.pdf "fig:")
![Same as Fig. \[spectra evolution cc\] for $C=10^{23}$ and $x_b=10^{30}$ (top), and for $C=10^{23}$ and $x_b=10^{36}$ (bottom). Higher values of $x_b$ imply an overall stronger total electromagnetic power spectrum, but with a stronger decrease rate at late times.[]{data-label="spectra evolution xb"}](MagDustPBPE_cc23_xb30.pdf "fig:") ![Same as Fig. \[spectra evolution cc\] for $C=10^{23}$ and $x_b=10^{30}$ (top), and for $C=10^{23}$ and $x_b=10^{36}$ (bottom). Higher values of $x_b$ imply an overall stronger total electromagnetic power spectrum, but with a stronger decrease rate at late times.[]{data-label="spectra evolution xb"}](MagDustPBPE_cc23_xb36.pdf "fig:")
Another interesting aspect of the magnetic and electric power spectra is their dependence in terms of $k$, shown in Fig. \[powerspectrum\]. As predicted in Sec. \[Aresults\], we obtain the spectral index $n_B=6$. This is typical of non-helicoidal and causally generated magnetic fields, as noted by Caprini and Durrer [@Caprini:2001nb; @Durrer:2003ja].
![Behavior of the magnetic power spectrum today from (blue) for $C=10^{23}$ and $x_b=10^{30}$. It is perfectly compatible with a power-law (top figure) with spectral index $n_B=6$ (orange). Note that $P_{B0}\equiv P_B(kR_H=1)$. We also show that the electric power spectrum behaves in the same fashion (bottom).[]{data-label="powerspectrum"}](MagDustPS_cc19_xb30.pdf "fig:") ![Behavior of the magnetic power spectrum today from (blue) for $C=10^{23}$ and $x_b=10^{30}$. It is perfectly compatible with a power-law (top figure) with spectral index $n_B=6$ (orange). Note that $P_{B0}\equiv P_B(kR_H=1)$. We also show that the electric power spectrum behaves in the same fashion (bottom).[]{data-label="powerspectrum"}](MagDustPEPS_cc19_xb30.pdf "fig:")
From the power spectrum, we are able to get the amplitude of the magnetic field as a function of the scale, which is shown in Figs. \[magfield\] and \[magfield2\]. Fig. \[magfield\] shows that a larger $x_b$, or equivalently a lower scale factor at the bounce ($a_b$), results in a lower amplitude of the field. Thus a deeper bounce tends to generate weaker magnetic fields. This is because electric and magnetic fields are generated when $f$ effectively changes in time, which happens for $-C/\alpha < t < C/\alpha$ (except for the short period of the bounce). Since $\alpha \propto
x_b^{3/2}$, a larger $x_b$ implies a shorter period in which the non-minimal coupling is effective. For the same reason, a larger value of $C$ leads to a larger amplitude of the magnetic field.
![Magnetic field amplitude for $C=2.6 \times 10^{26}$ and $x_b=10^{38}$ (top), and $C=6.5 \times 10^{25}$ and $x_b=10^{38}$ (bottom). For these values, the seed field is sufficient to trigger the dynamo mecanism at large scales. The amplitude today is larger at all scales for larger values of $C$.[]{data-label="magfield"}](MagDustB_cc2-6e26e38_lim.pdf "fig:") ![Magnetic field amplitude for $C=2.6 \times 10^{26}$ and $x_b=10^{38}$ (top), and $C=6.5 \times 10^{25}$ and $x_b=10^{38}$ (bottom). For these values, the seed field is sufficient to trigger the dynamo mecanism at large scales. The amplitude today is larger at all scales for larger values of $C$.[]{data-label="magfield"}](MagDustB_cc6-5e25_xb38.pdf "fig:")
![Magnetic field amplitude for $C=6.5 \times 10^{25}$ and $x_b=10^{38}$ (top) and for $C=6.5 \times 10^{25}$ and $x_b=10^{36}$ (bottom). The amplitude today is bigger at all scales when $x_b$ is smaller.[]{data-label="magfield2"}](MagDustB_cc6-5e25_xb38.pdf "fig:") ![Magnetic field amplitude for $C=6.5 \times 10^{25}$ and $x_b=10^{38}$ (top) and for $C=6.5 \times 10^{25}$ and $x_b=10^{36}$ (bottom). The amplitude today is bigger at all scales when $x_b$ is smaller.[]{data-label="magfield2"}](MagDustB_cc6-5e25_xb36.pdf "fig:")
In the next section, we discuss how observations and theoretical limits can be used to constrain the parameters of our models.
Discussion
==========
We now wish to confront the results of the previous section with observational and theoretical limits found in the literature. Limits coming from several physical processes can be invoked, as recalled in the introduction. However, it is worth noting that many of them focus on specific models with considerable uncertainties, or use specific priors leading to confusion on the possible upper and lower bounds.[^10] Since there is no unanimously accepted limit on the spectral index, we will focus on the bounds derived considering $n_B$ as a free parameter. Thus, we shall consider an upper bound around $B_{\lambda}<10^{-9}G$,[^11] and a first lower bound of around $B_{\lambda}>10^{-17}G$.[^12] The second lower limit we consider concerns the minimum seed field in galaxies that would be amplified via dynamo mechanism [@subramanian1994], namely $B_{\lambda}>10^{-21}G$.
These theoretical and observational limits are used in Figure \[parameter space 1Mpc\] to constrain the region in parameter space for which consistent values of magnetic seed fields,[^13] evaluated today, are obtained at 1 Mpc. The upper value $x_b\lesssim10^{38}$ comes from Eq. reflecting the earliest possible time for the bounce to occur. It is denoted “Planck Scale” in the graph. There is another limit set to preserve nucleosynthesis denoted “BBN”. This can be derived by plugging Eq. into Eq. presented below, giving $m_{\star}=10^{-19}m_e$.
In order to infer the allowed mass scales for the minimal coupling, one can use, for instance, the relation between $C$ and $m_{\star}$ coming from Eq. to show $$\label{mass}
\frac{m_{\star}}{m_e} = \frac{\alpha}{10^{38} C},$$ where $m_e$ is the electron mass. The maximum mass allowed in this model is then $0.1 m_e$. Therefore, the value of the electron mass for $m_{\star}$ is not allowed by our model, a feature shared with power-law inflationary models [@Turner1988; @Campanelli2008].
![Parameter space with magnetic field amplitudes consistent with current limits at 1Mpc. The blue region represents the allowed values to initiate the dynamo effect, with the blue line a theoretical lower limit [@subramanian1994; @Martin:2007ue]. The orange region represents allowed values by observations at large scales in voids, with the orange line a lower limit derived by blazars observations [@Taylor:2011bn] and the green line an upper limit derived using Ultra-High-Energy Cosmic Rays, Ultra-Faint Dwarf galaxies, 21-cm hydrogen lines, etc. [@Bray:2018ipq; @Safarzadeh:2019kyq; @Minoda:2018gxj]. Note the orange and blue regions are overlapping. The grey shaded region represents excluded values of the magnetic field. Each oblique grey line gives an amplitude for the magnetic field a hundred times higher than the lower line.[]{data-label="parameter space 1Mpc"}](parameter_space_1Mpc_xbm.pdf)
Backreaction
============
When dealing with primordial magnetogenesis, a recurrent issue one must be aware of is the backreaction of the electromagnetic contribution on the background. When the electromagnetic energy density becomes higher than the background energy density, the background dynamics is modified and anisotropies can appear [@Kanno:2009ei].
We define the matter and radiation energy densities, respectively, as $$\begin{aligned}
\rho_m \equiv \frac{\Omega_m}{Y^3}\;, \quad \rho_r \equiv \int {\mathrm{d}}{\ln{k}}\; \left(\mathcal{P}_{E,0}+\mathcal{P}_{B,0}\right) \left(\frac{Y_0}{Y}\right)^{4} \;.\end{aligned}$$ As pointed out in previous works on magnetogenesis in bouncing models, see *e.g.* [@Sriramkumar:2015yza], the vanishing of the Hubble rate at the bounce leads, via the Friedmann Eq. , to $\rho_m=0$. However, this is not the case here. The classical Friedmann equations are not valid around the bounce, which is dominated by quantum cosmological effects, and $\rho_m \propto Y^{-3}$ always. However, this does not guarantee that the model is free from backreaction. Let us examine this point in more detail in this section.
As the electromagnetic power goes as $Y^{-4}$, and $\rho_m \propto Y^{-3}$, the first obvious critical point to investigate the issue of backreaction is at the bounce itself. As shown in the previous section, we have near the bounce that $\vert A_k \vert \propto k^{-1/2}$ and $\vert \Pi_k\vert \propto k^{3/2}$. Furthermore, $\vert A_k\vert $ does not depend on $x_b$ and $C$, and $\vert \Pi_k\vert \propto C^2/\sqrt{x_b}$. This can be seen by inspecting the integral appearing in the first term of Eq. , where after integration, and evaluating at the bounce, we get the constants $C^2 x_b/\alpha = C^2/x_b^{1/2}$.
After integrating the magnetic and electric energy densities at the bounce, see Eqs. and , and denoting the cut-off scale as $k_f$ (which we will refer to galactic scales, where this simple treatment may cease to be valid due to short range interactions leading to dissipation and other effects), we obtain $$\begin{aligned}
\rho_{B,b} = \frac{3C^2 x_b^4}{32 \pi^2 R_{H_0}^4} k_f^4 \;, \quad \rho_{E,b} = \frac{C^2 x_b^3}{9 \pi^2 R_{H_0}^4} k_f^6 \;.\end{aligned}$$ The ratio of magnetic energy density over electric energy density is then simply $$\begin{aligned}
\frac{\rho_{B,b}}{\rho_{E,b}} \approx \frac{x_b}{k_f^2} \;,\end{aligned}$$ and the magnetic field is dominant when $\rho_{B,b} \gg \rho_{E,b}$, or $\sqrt{x_b}\gg k_f$. As $x_b\gg 1$, this condition is always satisfied.
In units of Hubble radius, the matter energy density reads $$\begin{aligned}
\rho_{m,b} = \frac{7.8 \Omega_m 10^{120}}{R_{H_0}^4 Y^3} \;.\end{aligned}$$ At the bounce, the matter energy density is given by $$\begin{aligned}
\rho_{m,b} = \frac{7.8 \Omega_m 10^{120} x_b^3}{R_{H_0}^4} \;.\end{aligned}$$ Then, comparing the magnetic density to the matter density, and requiring the ratio be small enough gives $$\begin{aligned}
\frac{\rho_{B,b}}{\rho_{m,b}} < 10^{-4} \quad \implies C^2 x_b k_f^4 < 10^{118} \;.
\label{magnetic backreaction}\end{aligned}$$ Choosing the galactic scale (tens of kiloparsecs), $k\approx 10^5$, gives $C^2 x_b < 10^{98}$. The values given in Fig. \[parameter space 1Mpc\] all respect this constraint. In conclusion, there is no electromagnetic backreaction at the bounce.
As we have seen in Figs. \[spectra evolution cc\] and \[spectra evolution xb\], and discussed when commenting them, the electric density overcomes the magnetic density after the bounce for some time during the period $1/\alpha < t < C/\alpha$. The coupling behaves as $$\begin{aligned}
f \propto t^{-2} \;, \quad \frac{1}{\alpha} < t < \frac{C}{\alpha} \;,\end{aligned}$$ and the scale factor as $Y \propto t^{\frac{2}{3}}$ in this region. This can be also be seen in Fig. \[ymf\]. Then, the electric density goes as $\rho_E \propto t^{-2/3}$. This is to be compared to the matter density $\rho_m \propto t^{-2}$, giving the ratio evolution $$\begin{aligned}
\frac{\rho_E}{\rho_m} \propto t^{\frac{4}{3}} \;.\end{aligned}$$ To get an estimate of the electric backreaction, let us evaluate the initial conditions at the bounce and evolve this ratio in the considered time range. Performing a procedure similar to the one leading to , we obtain $$\begin{aligned}
\frac{\rho_{E,b}}{\rho_{m,b}} = 10^{-122}\:C^2 k_f^6 . \end{aligned}$$ Then, the ratio will evolve as $$\begin{aligned}
\frac{\rho_{E}}{\rho_{m}} = 10^{-122}\:C^2 k_f^6 \left(\frac{t_f}{t_i} \right)^{\frac{4}{3}} \;.\end{aligned}$$ Choosing the initial time $t_i\equiv1/\alpha$ and the final time $t_f\equiv C/\alpha$ and imposing once again that the backreaction be small, we finally obtain $$\begin{aligned}
\frac{\rho_{E}}{\rho_{m}} < 10^{-4} \quad \implies C^{\frac{10}{3}} k_f^6 < 10^{118} \;. \end{aligned}$$ Once again, $k_f \approx 10^{5}$ is compatible with the maximum value $C\approx10^{26.3}$ allowed in Fig. \[parameter space 1Mpc\]. Again, there is no backreaction problem in our model [^14].
Conclusions
===========
We presented in this work the generation of primordial magnetic fields in the context of a cosmological bounce, through a coupling between curvature and electromagnetism, predicted by QED in curved spacetimes [@Drummond:1979pp]. A homogeneous and isotropic background filled with pressureless (dark) matter in the contracting phase, followed by a bounce and an expanding phase has been considered. The bounce is produced by quantum effects described in the de Broglie-Bohm interpretation of quantum mechanics, motivated by the inconsistency of using standard quantum mechanics in quantum cosmology [@Pinto-Neto:2013toa]. Moreover, one of the advantages of bouncing magnetogenesis is the absence of the strong coupling problem. The model is characterized by three parameters, namely the presureless (dark) matter density today, $\Omega_m$, the scale factor at which the bounce happens, $x_b$, and the mass scale of the coupling $m_{\star}$.
We showed that an adiabatic vacuum can be defined as initial condition for the electromagnetic field in the far past of the contracting phase. Having defined the vacuum, we were able to explain analytically the behavior of the electric and magnetic modes, summarized in Eqs. and . We then confronted these analytical results with a numerical integration of the modes, given in Figs. \[mode evolution cc\] and \[mode evolution xb\], and presented in Figs. \[spectra evolution cc\] and \[spectra evolution xb\] the time evolution of the magnetic and electric power spectra. We illustrated the scale dependence of both spectra in Fig \[powerspectrum\], finding they behaved as a power-law with the same spectral index $n_E=n_B=6$. This result is reminiscent of non-helicoidal, causally generated magnetic fields from phase transitions in the early Universe [@Durrer:2003ja]. In Figs. \[magfield\] and \[magfield2\], we showed the amplitude of the magnetic field today was found to be strong enough on a wide range of scales to pass the current limits from observations.[^15]At the scale of 1 Mpc, we have derived constraints on $x_b$ and $m_{\star}$, summarized in Fig. \[parameter space 1Mpc\]. Finally, we also demonstrated that backreaction is not a problem in our model.
Though the results from our analysis are quite promising, we have omitted several possible effects that could constrain our results further. First, the presence of an electromagnetic energy density in spacetime should induce a stochastic background of gravitational waves, even moreso since the magnetic fields generated have a very blue spectrum. Thus, the inclusion of theoretical limits on gravitational waves production [@Caprini:2001nb] will be investigated in the future. This will be even more relevant with the upcoming detections from LISA [@Caprini:2009yp; @Caprini:2018mtu; @Saga:2018ont; @Pol:2019yex].
A second point of interest would be to take into account other possible backreaction effects. It has been shown recently that the vacuum polarisation in a dielectric medium, the so-called Schwinger effect, increases the medium conductivity and subsequently stops the magnetic field production [@Sobol:2018djj; @Sobol:2019xls; @Sharma:2017eps; @Sharma:2018kgs; @Shakeri:2019mnt]. This would lead to weaker magnetic fields than expected, and could constrain further our model.
As a possible extension of our work, other non-minimal couplings between the electromagnetic and the gravitational field (involving the Ricci and Riemann tensors) could be considered in the generation of primordial magnetic fields.[^16] Also of importance is the parity-violating coupling $RF\tilde F$ [^17] , which may be associated to the generation of helical magnetic fields. We leave these open questions for future work.
The authors wish to thank Samuel Colin for his important contributions in earlier stages of this work, particularly for helping them to discard a previous proposal for the coupling. EF acknowledges support by the *Coordenação de Aperfeiçoamento de Pessoal de Nível Superior* - Brazil (CAPES) - Finance Code 001. EF thanks the *Centro Brasileiro de Pesquisas Físicas* (CBPF) for its hospitality. NPN acknowledges support of CNPq of Brazil under grant PQ-IB 309073/2017-0.
Adiabatic vacuum initial conditions {#sec:adiab}
===================================
First, we need to impose initial conditions for the EM field. To this end, we follow the adiabatic vacuum prescription implemented in Ref. [@sandro_vacuum]. Even though we are dealing with vector degrees of freedom, since the time-dependent coefficient $A_k(t)$ satisfies the normalization condition , it follows that $A_k(t)$ has a behavior similar to the one of the coefficient one would obtain when quantizing a single free scalar field. Let us then consider the Hamiltonian $$\label{hamiltonian}
\mathcal{H} = \frac{\Pi_k^2}{2m} + \frac{m\nu^2 A_k^2}{2} \;,$$ where $m$ and $\nu$ can be functions of time. The Hamilton equations of motion $$\label{eq-hamiltonian}
A^\prime_k=\frac{\Pi_k}{m}; \quad \Pi^\prime_k = -m\nu^2 A_k$$ lead to Eq. if one identifies $m = Yf$ and $\nu = k / Y$.
A convenient choice is to express $A_{k}$ and $\Pi_k$ as the components of a particular eigenvector of the complex structure matrix (see Ref. [@sandro_vacuum] for the mathematical and physical reasons to implement this choice), $$\label{v:components}
\begin{split}
A_k &\equiv \frac{1}{2}\exp{(-\gamma_k/2)} \left[\exp{(\chi_k/2)}-{\mathsf{i}}\exp{(-\chi_k/2)}\right] \;,\\
\Pi_k &\equiv -\frac{1}{2}\exp{(\gamma_k/2)} \left[\exp{(\chi_k/2)}+{\mathsf{i}}\exp{(-\chi_k/2)} \right)] \;.
\end{split}$$ The variables $\chi_k$ and $\gamma_k$ are real time-dependent functions, and can be used to represent the aforementioned matrix as $$\label{M:components}
M_a{}^b = \left( \begin{array}{cc} \sinh{\chi_k} & \cosh{\chi_k} \exp({-\gamma_k}) \\ -\cosh{\chi_k} \exp({\gamma_k}) & -\sinh{\chi_k} \end{array} \right).$$ Latin indices ($a,b,c, $ …) refer to the phase space vector components defined by $v_a \equiv \left(A_k,\Pi_k\right)$, which are raised and lowered using the symplectic matrix as defined in Ref. [@sandro_vacuum]. The phase space vectors $v_a$ satisfying the normalization condition (modulo a global time-dependent phase) have an one-to-one correspondence with matrices of the form shown in Eq. , and consequently with a pair $(\chi_k,\;\gamma_k)$. For this reason, we will denote interchangeably $\left(A_k,\Pi_k\right)$ and $\left(\chi_k,\;\gamma_k\right)$ with the same symbol $v_a$.
The Hamilton Eqs. induce the dynamics of the matrix $M_a{}^b$, which reads $$\label{set:equations}
\begin{split}
\chi_k^\prime&= -2\nu\sinh(\gamma_k-\xi) \;,\\
\gamma_k^\prime &= +2\nu\cosh(\gamma_k-\xi)\tanh(\chi_k) \;,
\end{split}$$ where $\xi \equiv \ln (m\nu)$. The complex structure matrix satisfies $$\label{M2:identity}
M_a{}^c M_c{}^b = -\delta_a{}^b \;,$$ and, the comparison of two different vacuum definitions, given respectively by $v_a$ and $u_a$, yields the Bogoliubov coefficients $$\label{beta-M}
\vert \beta_{v,u}\vert^2 = -\frac{1}{4} {\rm Tr} \left[{\rm\bf I} + {\bf M}(v) {\bf M}(u)\right] \;,$$ with ${\rm Tr}$ the trace operator, ${\rm\bf I}$ the identity matrix and ${\bf M}(v)$ (${\bf M}(u)$) is the matrix associated with the components $M_a{}^b(v)$ ($M_a{}^b(u)$) defined by the vector components $v_a$ ($u_a$). In this framework, a vacuum choice translates into a choice of functions $v^V_a \equiv \left(\chi_k^V(t),\;\gamma_k^{V}(t)\right)$ defined locally (with a finite number of time derivatives of the background variables), which do not necessarily satisfy the equations of motion but give an approximation close enough to a solution. Moreover, the vacuum must be fixed by choosing a time $t_0$ where the variables satisfy $$v_a(t_0) = v^V_a(t_0), \quad\implies\quad \left(\chi_k(t_0), \gamma_k(t_0)\right) = \left(\chi^V_k(t_0), \gamma^V_k(t_0)\right).$$ In other words, if $v_a^V$ is stable in the sense that $$\Delta v_a \equiv \left(\delta\chi_k,\;\delta\gamma_k \right) = \left(\chi_k(t) - \chi^V_k(t), \gamma_k(t) - \gamma^V_k(t)\right)$$ remains small for a finite time interval, then particle creation will also be small in this interval. This characterizes the so-called adiabatic vacuum. Hence, we find the adiabatic vacuum by finding the critical points of the system . When $\xi$ is constant in time, the critical points of the system are obvious: $\chi^V_k = 0$ and $\gamma^V_k = \xi$, a choice satisfying the condition of being locally defined in terms of the background. Then, substituting into Eq. , and using it as initial conditons for the system , yields the following solution $$\label{WKB}
\begin{split}
A_k &= \frac{e^{-{\mathsf{i}}\pi/4}}{\sqrt{2m\nu}}
\exp\left[-{\mathsf{i}}\int_{t_0}^t\nu{{\mathrm{d}}{}t}\right] \;,\\
\Pi_k &= -{\mathsf{i}}e^{-{\mathsf{i}}\pi/4}\sqrt{\frac{m\nu}{2}}
\exp\left[-{\mathsf{i}}\int_{t_0}^t\nu{{\mathrm{d}}{}t}\right] \;.
\end{split}$$ In this case, the vacuum is perfectly stable, there is never particle production because $\chi_k(t) = \chi^V_k(t) = 0$ and $\gamma_k(t) = \gamma^V_k(t) = \xi$ for any time $t$, and consequently $\vert\beta_{v,v^V}\vert^2=0$, see Eq. . We have a perfect adiabatic vacuum, which coincides with the WKB solution.
In the case where $\xi$ changes in time, there is one well-known situation where adiabatic vacua can be defined: when the mode frequencies dominate the dynamics. Let us define $$F_n \equiv \left(\frac{1}{2\nu}\frac{{\mathrm{d}}}{{\mathrm{d}}t}\right)^n \xi \;,$$ where $F_0 = \xi$, the function $F_1$ is the ratio between the time derivative of $\xi$ and $\theta \equiv \int2\nu{\mathrm{d}}t$, $F_2$ the ratio between the time derivative of $F_1$ and $\theta$ and so forth. Then, in the case $1 \gg F_1 \gg\dots\gg F_n>\dots$, which means that $\xi$ slowly varies in cosmic time when compared with the variation of $\theta$, one can still find approximate critical points (*i.e.* adiabatic vacua), which can be reached through successive approximations, as explained in Ref. [@sandro_vacuum]. Up to second order, the approximate critical points read $$\begin{aligned}
\label{variables2}
\chi^V_k&= F_1 \;,\nonumber \\
\gamma^V_k&= F_0 -F_2\;.\end{aligned}$$ If they are inserted in Eq. , they lead to the usual WKB expansion (modulo a time-dependent phase). As discussed in [@sandro_vacuum], around these functions, the variables $\Delta v_a$ satisfy a forced harmonic oscillator equation of motion with force of order $\mathcal{O}(F_3)$.
In our case, we have $m\nu=kf$. In the far past of the contracting phase one gets, for $f$ given in Eq. , $$\label{xi1}
\left\vert\frac{{\mathrm{d}}\xi}{{\mathrm{d}}\theta}\right\vert \approx \frac{C^2}{x_b^3k\vert t\vert^{7/3}} \ll 1 \;,$$ which implies that $$\vert t\vert \gg \vert t_a\vert \equiv\left(\frac{C^2}{x_b^3k}\right)^{3/7}.$$ As the physically relevant parameter space we consider satisfies $C^2/x_b^3 \ll 1$, then $\vert t_a\vert \ll 1$, and this condition is easily satisfied.
However, the other adiabaticity conditions impose a more stringent constraint on $\vert t\vert$. Indeed,
$$\label{xi2}
\left\vert\frac{{\mathrm{d}}^2\xi}{{\mathrm{d}}\theta^2}\right\vert \ll \left\vert\frac{{\mathrm{d}}\xi}{{\mathrm{d}}\theta}\right\vert \Rightarrow \vert t\vert \gg \vert t_c\vert \equiv
\left(\frac{7}{3 k}\right)^{3}\approx k^{-3} \;.$$
One can easily verify that all other conditions yield, apart numerical factors of order $1$,[^18] the same condition . Hence, the adiabaticity condition reads
$$\label{adiabatic}
\vert t\vert \gg \vert t_c\vert \approx k^{-3}.$$
This means that modes with the size of the Hubble radius today leave (enter) the adiabatic regime in the contracting (expanding) phase for times of the order the Hubble time today, independently of the parameters $x_b$ and $C$. Smaller wavelengths leave (enter) the adiabatic regime later (earlier) than the present Hubble time, following the rule $k^{-3}$.
To summarize, one can impose adiabatic vacuum initial conditions for the electromagnetic field in the contracting phase of the present bouncing model when $\vert t\vert \gg \vert t_c\vert \approx k^{-3}$. In this regime, the modes read, at leading order, $$\label{initial-vacuum}
\begin{split}
A_k &= \frac{e^{-{\mathsf{i}}\pi/4}}{\sqrt{2kf}}\exp\left(-{\mathsf{i}}k\eta\right)+\dots \;,\\
\Pi_k &= -{\mathsf{i}}e^{-{\mathsf{i}}\pi/4}\sqrt{\frac{kf}{2}}
\exp\left(-{\mathsf{i}}k\eta\right)+\dots \;,
\end{split}$$ where $\eta$ is the conformal time $dt = Yd\eta$.
Since $f\approx 1/4$ for $\vert t\vert \gg \vert t_c\vert $ , it follows that $$\label{initial-modulus}
\begin{split}
\vert A_k\vert &= \sqrt{\frac{2}{k}} +\dots \;, \\
\vert \Pi_k\vert &= \sqrt{\frac{k}{8}} +\dots. \;,
\end{split}$$ and both the field and its canonical momentum are constant in this regime.
[^1]: <[email protected]>
[^2]: <[email protected]>
[^3]: <[email protected]>
[^4]: [[email protected]](mailto:[email protected])
[^5]: Cosmological magnetic fields may also be produced during phase transitions, see for instance [@Grasso:2000wj], or through the generation of vortical currents [@Carrilho:2019qlb].
[^6]: In models in which radiation is important initially, thermal fluctuations may dominate over quantum fluctuations, see [@Cai2008; @Cai2009; @Bhattacharya2013].
[^7]: Note that a scale factor of this form was introduced by hand in [@Sriramkumar:2015yza] to generate scale invariant magnetic fields, while it emerges naturally from quantum effects here.
[^8]: The lower bound is set by imposing the validity of the Wheeler-DeWitt equation, *i.e.*, by restricting the curvature to values such that possible discreteness of the spacetime geometry is negligible, while quantum effects are still relevant [@Peter:2006hx]. Since $t_\text{Planck}\simeq 10^{-44}s$ and recalling that BBN happened around $10^{4}s$, the upper bound simply reflects the latest time at which the bounce can occur.
[^9]: We choose the values of $C$ and $x_b$ to be well inside the allowed parameter space at 1 Mpc, as can be seen in Fig. \[parameter space 1Mpc\]. We will use the same set of values throughout this section, except for Figs \[magfield\] and \[magfield2\].
[^10]: For instance, see [@Zucca:2016iur; @Pogosian:2018vfr] for a discussion about the suppressed apparent limit on the magnetic spectral index $n_B$, when assuming a different prior from Planck 2015 [@Ade:2015cva]
[^11]: See [@Bray:2018ipq] and [@Safarzadeh:2019kyq] for recent limits using ultra-high-energy cosmic rays anisotropy and ultra-faint dwarf galaxies, respectively. See also [@Broderick:2018nqf] for a stronger upper limit of $B_{\lambda}<10^{-15}G$, putting detections of intergalactic magnetic fields with $\gamma$-ray under pressure.
[^12]: This limit comes from the non-detection of secondary GeV $\gamma$-rays around TeV blazars. However, there is still an ongoing debate on whether this lower limit should be trusted. See for example [@Broderick:2011av; @Subramanian:2019jyd].
[^13]: Within the commonly invoked limits, as discussed earlier.
[^14]: To discuss the (absence of) backreaction in our model, we have shown that the electromagnetic energy density is always smaller than the matter energy density. In other models of bounce, such as those based on the Lee-Wick theory [@Cai2008], there are mechanisms preventing *ab initio* the uncontrolled growth of the electromagnetic energy density.
[^15]: It is worthwhile noting that the same coupling, when considered in the context of power-law inflation, does not generate large enough magnetic fields, see [@Campanelli2008].
[^16]: Such couplings were considered in the framework of power-law inflation in [@Kunze2009].
[^17]: The authors wish to thank the referee for calling their attention to this possibility.
[^18]: It starts around $1$ and grows slowly with $n$, this is a natural feature of an asymptotic expansion. In other words, for a fixed time and mode $k$ there is a maximum order $n$ from which the series starts to be a bad approximation.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Mohammad Javad Abdi [^1]'
bibliography:
- 'Paper3.bib'
title: 'COMPARISON OF SEVERAL REWEIGHTED $l_1$-ALGORITHMS FOR SOLVING CARDINALITY MINIMIZATION PROBLEMS'
---
0.22in
**Abstract.** Reweighted $l_1$-algorithms have attracted a lot of attention in the field of applied mathematics. A unified framework of such algorithms has been recently proposed in [@1]. In this paper we construct a few new examples of reweighted $l_1$-methods. These functions are certain concave approximations of the $l_0$-norm function. We focus on the numerical comparison between some new and existing reweighted $l_1$-algorithms. We show how the change of parameters in reweighted algorithms may affect the performance of the algorithms for finding the solution of the cardinality minimization problem. In our experiments, the problem data were generated according to different statistical distributions, and we test the algorithms on different sparsity level of the solution of the problem. Our numerical results demonstrate that the reweighted $l_1$-method is one of the efficient methods for locating the solution of the cardinality minimization problem.
Introduction
============
The cardinality minimization problem over a convex set $C\subset \mathbb{R}^n$ can be written as $$\begin{aligned}
\begin{split}\label{CMPMAIN}
\mbox{Minimize } &\|x\|_0& \\
\mbox {s.t. } & x\in C.&\\
\end{split}\end{aligned}$$ This problem is to minimize the number of non-zero components of a vector satisfying certain constraints. In other words, cardinality minimization problem is looking for the sparsest vector in a given feasible set.
In this paper, we suppose that $C$ is defined by an undetermined system of linear equations, i.e, $$C=\{x; Ax=b\},~\mbox{where}~A\in \mathbb{R}^{m\times n}~(m<n)~,b\in \mathbb{R}^{m}.$$ These linear systems have infinite many solutions, and the purpose of cardinality minimization problem(CMP) is to find the sparsest one, which can be stated as follows: $$\begin{aligned}
\begin{split}\label{CMP}
\mbox{Minimize } &\|x\|_0& \\
\mbox {s.t. } & Ax=b.&\\
\end{split}\end{aligned}$$
The problem ($\ref{CMP}$) is closely related to compressed sensing which is dealing with the reconstruction of sparse signals from a limited number of linear measurements [@34; @35; @36; @37]. Also problems with cardinality constraints have a wide range of applications, especially in portfolio optimization problems [@39; @40], and principal component analysis and model reduction [@41; @43]. The more generalized version of cardinality minimization problems is so called rank minimization problems which have been considered in recent years [@8; @9].
The $l_0$-norm function is discontinuous, so the main idea for solving the problem ($\ref{CMP}$) is to approximate the $l_0$-norm function by some other continuous functions which are easier to deal with. For example $l_p$-norm function($0<p<1$) is one of the approximations of the $l_0$-norm. $l_p$ minimization $(0<p<1)$ has been studied in [@44; @46; @26]. Figure (1) represents the graph of $\sum_{i=1}^n|x_i|^p$, for $p=1$, $p=0.6$, and $p=0.2$. Note that as $p$ goes to zero, $\sum_{i=1}^n|x_i|^p$ approaches to the $l_0$-norm function.
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{NormpPlot.eps}
\label{FIG1}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{lognormplot.eps}
\label{FIG2}
\end{array}$
As seen in the Figure (1), the closest convex approximation of $\|x\|_0$ is the well known $l_1$-norm function. So it is unavoidable to use non-convex functions, especially concave functions, in order to have a better approximation of $\|x\|_0$. In [@1] Zhao and Li introduced the following function which is a combination of $l_p$-norm$(0<p<1)$ and the $\log$ function to approximate $\|x\|_0$,(see Figure (2)): $$F_{\epsilon}(x)=\sum_{i=1}^n \log(|x_i|+\epsilon)+\sum_{i=1}^n(|x_i|+\epsilon)^p,~0<p<1.$$ We will discuss this kind of functions later. Finite successive linear approximation algorithms have also been used and are still used to get an approximated solution of the concave approximation problems [@18; @19; @20; @60].
$l_1$-minimization methods have been used to solve the problem ($\ref{CMP}$). This is motivated by the main idea of replacing $\|x\|_0$ function with its local convex envelop, the $l_1$-norm function, and then solve the resulting linear program [@3; @4; @48]. Under certain conditions the $l_1$-minimization method is able to obtain the exact solution of the problem ($\ref{CMP}$) for very sparse solutions of the system $Ax=b$. In the literature, several conditions have been introduced and discussed for the equivalence between the $l_1$-minimization and the $l_0$-minimization. The outstanding ones are spark [@4], mutual coherence [@5; @6], restricted isometry property(RIP)[@3; @10; @11], and null space property(NSP) [@12; @13]. For large optimization problems, the unconstrained version of the problem has been investigated in the literature, that may be referred as Lasso-type problems [@52].
Numerical experiments show that weighted approaches are very affective in locating an exact solution of the problem ($\ref{CMP}$), and it can outperform other methods, in many situations [@1; @22; @23; @24; @25; @26].\
Candes, Wakin, and Boyd [@23] proposed a weighted $l_1$-algorithm as follows: $$\begin{aligned}
\begin{split}\label{WP1}
\mbox{Minimize } &\sum_{i=1}^n\omega_i|x_i|& \\
\mbox {s.t. } & Ax=b.&\\
\end{split}\end{aligned}$$ By introducing a diagonal matrix $W=\mbox{diag}(\omega_1,\omega_2,...,\omega_n)$, the problem above can be written as $$\begin{aligned}
\begin{split}\label{WP2}
\mbox{Minimize } &\|Wx\|_1& \\
\mbox {s.t. } & Ax=b.&\\
\end{split}\end{aligned}$$ The weight can be interpreted as penalties for the components of the vector $x$. Larger penalties, ($\omega_i$)s, apply to smaller component of the vector $x$, for example one may choose the weights as $\frac{1}{|x_i|},~i=1,...,n$. However to avoid having infinity penalties, one may add a parameter like $\epsilon>0$ to define the following weight [@23] $$\omega_i=\frac{1}{|x_i|+\epsilon},~i=1,...,n.$$
Choosing a proper $\epsilon$ to have a more efficient algorithm is one of the challenges. Very small or very large $\epsilon$ might lead to improper weights which may cause the failure of the algorithms. Since very small $\epsilon$ might result in infinity penalties for small component of $x$, and with very big $\epsilon$ the penalty might not recognize the difference between the small components of $x$ and the large ones. We discuss the choice of $\epsilon$ for our algorithms in the numerical experiment later.\
In an iterative reweighted $l_1$-algorithm, weights can be defined from the iteration in the previous step. Suppose the solution at the step $l$ is $x^l$, then the weight at the next step $l+1$, can be given as $\omega^{l+1}=\frac{1}{|x^l|+\epsilon}$. This was introduced by Candes, Wakin, and Boyd, and we refer to their algorithm as CWB, in this paper.\
In [@1], Zhao and Li introduced a unified framework for the reweighted $l_1$-minimization. The main idea is to define a merit function which is a certain concave approximation of the cardinality function, and to construct different types of weights through the linearization techniques. Based on the class of merit functions defined by Zhao and Li [@1], we identify several new specific merit functions which are used to define the weights of the reweighted $l_1$-algorithms. The main purpose of this paper is to study these merit functions, and to test the success probability of the reweighted algorithms associated with these merit functions for locating the sparsest solution of linear systems, where the matrices, A, are generated based on different statistical distributions. Note that most of the previous experiments in the literature use normally distributed matrices. Also, we demonstrate how the parameters used in the algorithms may affect the performance of these methods. Furthermore, we evaluate what choice of $\epsilon$ may make our algorithms work better. In section 2, we discuss different types of merit functions and the associated reweighted $l_1$-algorithms. In section 3, we present and discuss our numerical results, and provide comparison between these algorithms.
Merit functions and reweighted algorithms
=========================================
Merit functions have been used frequently in the field of optimization. Recently Zhao and Li [@1] has used merit functions to approximate $l_0$-norm. The merit function is defined as follows.
**Merit function:** For any $\epsilon>0$, a merit function $F_{\epsilon}(x):\mathbb{R}^n\rightarrow \mathbb{R}$ for approximating the $l_0$-norm, is strictly concave, separable, coercive, strictly increasing and twice differentiable, with the following properties:
1. $\lim_{\epsilon\rightarrow 0}\frac{F_{\epsilon}(x)}{g(\epsilon)}=\|x\|_0+C,~g(\epsilon)>0~ \mbox{is a function of }\epsilon, ~\mbox{and}~C~\mbox{is a constant},$\
2. $F_{\epsilon}(x)=F_{\epsilon}(|x|),~~\forall x\in \mathbb{R}^n,$\
3. $\lim_{(x_i,\epsilon)\rightarrow (0,0)}[\nabla F_{\epsilon}(x)]_i=\infty,~~\forall x\geq 0,~\forall\epsilon>0, $\
4. $\lim_{\epsilon \rightarrow 0}[\nabla F_{\epsilon}(x)]_i=c_i,~~\forall x_i>0,~\mbox{where each}~c_i~\mbox{is a positive constant}.$
After replacing the $\|x\|_0$ by a merit function the problem ($\ref{CMP}$) can be written as $$\begin{aligned}
\begin{split}\label{Merit}
\mbox{Minimize } &F_{\epsilon}(x)& \\
\mbox {s.t. } & Ax=b.&\\
\end{split}\end{aligned}$$
Note that $F_{\epsilon}(x)$ is a concave function. One of the usual methods to solve concave optimization problems is to apply the linearization method, which in this case is a special type of Majorization-Minimization(MM) method. For more illustration see that by applying the Taylor expansion of $F_{\epsilon}(x)$ around a point $u$, we conclude $$F_{\epsilon}(x)\leq F_{\epsilon}(u)+\langle\nabla F_{\epsilon}(u),x-u\rangle.$$ The right hand side of the inequality above is a linear function. Hence the problem ($\ref{Merit}$) would be reduced to the following linear program: $$\begin{aligned}
\begin{split}\label{Merit1}
\mbox{Minimize } &\langle \nabla F_{\epsilon}(u),x\rangle&\\
\mbox {s.t. } & Ax=b.&\\
\end{split}\end{aligned}$$ So in our iterative reweighted algorithm, we solve at step $l$ the following optimization problem: $$\begin{aligned}
\begin{split}\label{Merit2}
\mbox{Minimize } &\langle\nabla F_{\epsilon}(x^l),x\rangle&\\
\mbox {s.t. } & Ax=b,&\\
\end{split}\end{aligned}$$ where $x^l$ is the solution of the previous iteration, and $\nabla F_{\epsilon}(x^l)$ are the weights. The reweighted $l_1$-algorithm can be defined as follows:
- Set $l$ as an index which counts the iterations, and choose a small enough $\epsilon>0$.
- Step 0: Choose a starting point $x^{1}$. This can be obtained by solving the $l_1$-minimization problem.
- Step $l$: Set $\omega^{l}=\nabla F_{\epsilon}(x^l)$, and Solve $$x^{l+1}=\mbox{argmin}\{\langle \omega^{l},x\rangle:Ax=b\}.$$
- Step $l+1$: If some termination criteria holds, stop. Otherwise, set $l\leftarrow l+1$, and go to step $l$.
An additional step can be added to the above algorithm concerning the choice of $\epsilon$. In this paper our updating rule is $\epsilon_{l+1}=0.5\epsilon_l$. In CWB algorithm, $\epsilon$ is updated as $\epsilon_{l+1}=\max\{|x^l|_{(i_0)},0.001\}$, where $i_0=\frac{m}{[4\log(\frac{n}{m})]}$, and $|x|_{i_0}$ is the biggest $i_0$ elements of $x$.\
It is quit challenging to prove that under a mild condition, the reweighted $l_1$-algorithm converges to the sparsest solution of problem $(\ref{CMP})$. This is still an open questions in this field. However some progress have been made in this area [@1; @57; @26; @24]. Mangasarian [@57] introduced a successive linearization algorithm(SLA) to find the solution of general complementarity problems, and proved that SLA algorithm terminates in finite number of iterations, and creates decreasing objective function values at each iteration. Furthermore he proved these values converge to a stationary point. Chen and Zhou in [@26] proved that the sequence generated by reweighted $l_1$-algorithm converges to a stationary point of a kind of truncated $l_p$-minimization problem ($0<p<1$). Similar results can also be found in [@22]. Recently, Zhao and Li [@1] defined a range space property(RSP) for matrices, under which he proved that the reweighted $l_1$-algorithm converges to certain sparse solution of the problem.
Following the framework of the reweighted $l_1$-algorithm in [@1], we discuss some new merit functions. Before we go ahead, let’s consider the following merit function $$\label{merit1} F_{\epsilon}(x)=\sum_{i=1}^n \log(|x_i|+\epsilon)+\sum_{i=1}^n(|x_i|+\epsilon)^p,$$ where $0<p<1$, which is mentioned in [@1], based on which we will construct new merit functions. To verify that the above function is a merit function, one should check all the defined properties are satisfied. First, let’s verify that this function is an approximation of $l_0$-norm function.\
Indeed, it is easy to check that $$\lim_{\epsilon\rightarrow 0} \left (n-\frac{\sum_{i=1}^n\log(|x_i|+\epsilon)+\sum_{i=1}^n(|x_i|+\epsilon)^p}{\log\epsilon}\right )=\|x\|_0.$$ Note that $$\lim_{x\rightarrow \infty}F_{\epsilon}(x)=\infty,$$ which means that the function is coercive. It is clear that $F_{\epsilon}(x)=F_{\epsilon}(|x|)$, and the function is increasing. In $R_+^n$, we have $$\nabla F_{\epsilon}(x)=\left(\frac{1+(x_1+\epsilon)^p p}{x_1+\epsilon},...,\frac{1+(x_n+\epsilon)^p p}{x_n+\epsilon}\right)^T.$$ Also, for every $i=1,...,n$, we have $$\lim_{(x_i,\epsilon)\rightarrow (0,0)}[ \nabla F_{\epsilon}(|x|)]_i=\lim_{(x_i,\epsilon)\rightarrow (0,0)} \frac{1+(|x_i|+\epsilon)^p p}{|x_i|+\epsilon}=\infty,~i=1,...,n,$$ and $\nabla F_{\epsilon}(x)$ is bounded when $\epsilon\rightarrow 0$. Since $0<p<1$ and $x_i>0$, we have $$p(x_i+\epsilon)^p>p^2(x_i+\epsilon)^p,~i=1,...,n,$$ so $$\frac{-1+(x_i+\epsilon)^pp^2-(x_i+\epsilon)^pp}{x_i+\epsilon}<0,~i=1,...,n,$$ and hence $$\nabla^2F_{\epsilon}(x)=\mbox{diag}\left(\frac{-1+(x_i+\epsilon)^pp^2-(x_i+\epsilon)^pp}{x_i+\epsilon}\right)\prec 0,~i=1,...,n.$$ As seen, in $R_+^n$ the Hessian of the above merit function is negative definite, so the function $F_{\epsilon}(x)$ is strictly concave. From the above discussion one can define the following weights for the reweighted $l_1$-algorithm: $$\omega_i=[ \nabla F_{\epsilon}(|x|)]_i=\frac{1+(|x_i|+\epsilon)^p p}{|x_i|+\epsilon},~i=1,...,n.$$ Note that the item (2) of the definition of a merit function implies that $[\nabla F_{\epsilon}(x^l)]_i\rightarrow \infty$ as $(x_i^l,\epsilon)\rightarrow (0,0)$, which means larger penalties(weights) for the smaller elements of $x$, at each iteration.\
Now, we start to define a new merit function as follows $$\label{merit2}F_{\epsilon}(x)=\sum_{i=1}^n \log(\log(|x_i|+\epsilon+(|x_i|+\epsilon)^p)).$$ We verify this function is a merit function. Clearly, this function is an approximation of $l_0$-norm function. Because $$\lim_{x\rightarrow 0}\frac{\log(\log(x_i+\epsilon+(x_i+\epsilon)^p))}{\log(\log(\epsilon))}=0,~\mbox{for}~ x_i\neq 0,$$ and $$\lim_{x\rightarrow 0}\frac{\log(\log(x_i+\epsilon+(x_i+\epsilon)^p))}{\log(\log(\epsilon))}=1,~\mbox{for}~ x_i=0,$$ we conclude that $$\lim_{x\rightarrow 0}\frac{\sum_{i=1}^n \log(\log(|x_i|+\epsilon+(|x_i|+\epsilon)^p))}{\log(\log(\epsilon))}=n-\|x\|_0.$$ In $R_+^n$, the gradient of $F_{\epsilon}(x)$ is given by $$[\nabla F_{\epsilon}(x)]_i=\frac{1+\frac{(x_i+\epsilon)^pp}{x_i+\epsilon}}{(x_i+\epsilon+(x_i+\epsilon)^p)(\log(x_i+\epsilon+(x_i+\epsilon)^p))},$$ and since $\lim_{x_i\rightarrow0} x_i\log(x_i)=0$, we have $$\lim_{(x_i,\epsilon)\rightarrow (0,0)} [\nabla F_{\epsilon}(x)]_i=\infty.$$ Note that for every $x_i>0$, $$\lim_{\epsilon\rightarrow 0}[\nabla F_{\epsilon}(x)]_i=\frac{1+px_i^{p-1}}{(x_i+x_i^p)\log(x_i+x_i^p)}=c_i,~i=1,...,n,$$ where $c_i$ is positive and bounded for every $i=1,...,n$. Also in $R_+^n$, $\nabla^2 F_{\epsilon}(x)$ is a diagonal matrix with the following entries on its diagonal, $$[\nabla^2F_{\epsilon}(x)]_{ii}=\frac{\frac{(x_i+\epsilon)^pp^2}{(x_i+\epsilon)^2}-\frac{(x_i+\epsilon)^pp}{(x_i+\epsilon)^2}}{(x_i+\epsilon+(x_i+\epsilon)^p)\log(x_i+\epsilon+(x_i+\epsilon)^p)}$$ $$-\frac{\left(1+\frac{(x_i+\epsilon)^pp}{x_i+\epsilon}\right)^2}{\left(x_i+\epsilon+(x_i+\epsilon)^p\right)^2\log(x_i+\epsilon+(x_i+\epsilon)^p)}$$ $$-\frac{\left(1+\frac{(x_i+\epsilon)^pp}{x_i+\epsilon}\right)^2}{\left(x_i+\epsilon+(x_i+\epsilon)^p\right)^2\log\left(x_i+\epsilon+(x_i+\epsilon)^p\right)^2},~i=1,...,n.$$ Since, for every $i=1,...,n$, $[\nabla^2F_{\epsilon}(x)]_{ii}<0$, we have $$\nabla^2 F_{\epsilon}(x)\prec 0.$$ So the function $(\ref{merit1})$ is strictly concave, and it is a merit function.
Based on $(\ref{merit1})$, the reweighted $l_1$-algorithm choose the following weights: $$\omega_i=\frac{1+\frac{(|x_i|+\epsilon)^pp}{|x_i|+\epsilon}}{(|x_i|+\epsilon+(|x_i|+\epsilon)^p)(\log(|x_i|+\epsilon+(|x_i|+\epsilon)^p))},~i=1,...,n.$$
In this paper, we refer to $W_1$ as the reweighted algorithm with the above weights. The Figure (15) shows the probability of success of $W_1$ algorithm via different choices of $\epsilon$. This figure demonstrates that $\epsilon=0.01$ works very good to locate the exact solution of the problem $(\ref{CMP})$, when the sparsity is 15. Clearly the above weights are related to the parameter $p$, so we tested the performance of $W_1$ algorithm for different sparsity of the solution, i.e, $k=5,10,15,20$, via different choices of $p$. Thirteen different values of $p$ have been tested (matrix $A$ has been normally distributed), and the result is summarized in Figure (14). Obviously the probability of success is higher when the sparsity of the solution is lower. This can be seen in Figure (14).\
Another new merit function can be defined as follows $$\label{Merit2} F_{\epsilon}(x)=\frac{1}{p}\sum_{i=1}^n \left(\log(|x_i|+\epsilon+(|x_i|+\epsilon)^q)\right)^p,$$ where $0<p,q<1$, which is an approximation of $\|x\|_0$. In fact $$n-\lim_{\epsilon\rightarrow0}\frac{\frac{1}{p}\sum_{i=1}^n \left(\log(|x_i|+\epsilon+(|x_i|+\epsilon)^q)\right)^p}{\frac{1}{p}\left(\log(\epsilon+\epsilon^q)\right)^p}=\|x\|_0.$$ In $R_+^n$, the gradient of $F_{\epsilon}(x)$ is given by $$[\nabla F_{\epsilon}(x)]_i=\frac{\log\left(x_i+\epsilon+(x_i+\epsilon)^q\right)^p\left(1+\frac{(x_i+\epsilon)^qq}{x_i+\epsilon}\right)}{\left(x_i+\epsilon+(x_i+\epsilon)^q\right)\log\left(x_i+\epsilon+(x_i+\epsilon)^q\right)}.$$ Note that $\lim_{(x_i,\epsilon)\rightarrow (0,0)}[\nabla F_{\epsilon}(x)]_i=\infty$, and $\lim_{\epsilon\rightarrow 0}[\nabla F_{\epsilon}(x)]_i$ is bounded, for every fixed $x>0$. In $R_+^n$, the Hessian is a diagonal matrix with the following diagonal elements $$[\nabla^2F_{\epsilon}(x)]_{ii}=\frac{\log\left(x_i+\epsilon+(x_i+\epsilon)^p\right)^q}{\left(x_i+\epsilon+(x_i+\epsilon)^q\right)\log\left(x_i+\epsilon+(x_i+\epsilon)^q\right)}$$$$.\left(\frac{(p-1)\left(1+\frac{(x_i+\epsilon)^qq}{x_i+\epsilon}\right)^2}{\left(x_i+\epsilon+(x_i+\epsilon)^q\right)\log\left(x_i+\epsilon+(x_i+\epsilon)^q\right)}\right.+$$ $$\left.\frac{(x_i+\epsilon)^qq^2-(x_i+\epsilon)^qq}{(x_i+\epsilon)^2}-\frac{\left(1+\frac{(x_i+\epsilon)^qq}{x_i+\epsilon}\right)^2}{x_i+\epsilon+(x_i+\epsilon)^q}\right),~i=1,...,n.$$ where $0<p,q<1$. Clearly, $\nabla^2F_{\epsilon}(x)\prec0$, which implies the function is strictly concave. Thus, the function ($\ref{Merit2}$) is a merit function, so we may choose the following weights in our algorithm: $$\omega_i=\frac{\log\left(|x_i|+\epsilon+(|x_i|+\epsilon)^q\right)^p\left(1+\frac{(|x_i|+\epsilon)^qq}{|x_i|+\epsilon}\right)}{\left(|x_i|+\epsilon+(|x_i|+\epsilon)^q\right)\log\left(|x_i|+\epsilon+(|x_i|+\epsilon)^q\right)},~i=1,...,n.$$
We refer to $W_2$ as the reweighted algorithm with the weights above. The Figures (11),(12),(13) show the performance of $W_2$ algorithm for finding the exact solution of the problem $(\ref{CMP})$ for different choices of the weights parameters, $p$ and $q$, and for different fixed sparsity of the solution, i.e., $k=5,10,15,20$.\
**Remark.**We see from above that the $\log$ function plays a vital rule in constructing a merit function. As pointed in [@1; @62], the $\log$ function can enhance the concavity of a given function without affecting its coercivity and monotonicity. For the convergency analysis of the reweighted $l_1$-algorithms based on the class of merit functions that defined at the beginning of this chapter, one may refer to the Theorems 3.9 and 3.11 in [@1], where it has been shown that under the so-called RSP condition, the algorithm may converge to a solution of problem ($\ref{CMP}$) with certain level of sparsity.\
Numerical Experiments
=====================
In this section, we compare the performance of the algorithms above for finding the exact solution of the problem $(\ref{CMP})$ through the numerical tests. We compare the following algorithms in our numerical experiments.\
$l_1$-min: $$\begin{aligned}
\begin{split}\label{l1}
\mbox{Minimize } &\|x\|_1& \\
\mbox {s.t. } & Ax=b,&\\
\end{split}\end{aligned}$$ CWB(Candes, Wakin, Boyd): $$\begin{aligned}
\begin{split}\label{CWB}
x^{l+1}=\mbox{argmin } &\sum_{i=1}^n\frac{1}{|x_i^l|+\epsilon^l}|x_i|& \\
\mbox {s.t. } & Ax=b,&\\
\end{split}\end{aligned}$$ $W_1$: $$\begin{aligned}
\begin{split}\label{LW1}
x^{l+1}=\mbox{argmin } &\sum_{i=1}^n\frac{1+\frac{(|x_i^l|+\epsilon^l)^pp}{|x_i^l|+\epsilon^l}}{(|x_i^l|+\epsilon^l+(|x_i^l|+\epsilon^l)^p)(\log(|x_i^l|+\epsilon^l+(|x_i^l|+\epsilon^l)^p))}|x_i|& \\
\mbox {s.t. } & Ax=b,&\\
\end{split}\end{aligned}$$ $W_2$: $$\begin{aligned}
\begin{split}\label{LW2}
x^{l+1}=\mbox{argmin } &\sum_{i=1}^n\frac{\log\left(|x_i^l|+\epsilon^l+(|x_i^l|+\epsilon^l)^q\right)^p\left(1+\frac{(|x_i^l|+\epsilon^l)^qq}{|x_i^l|+\epsilon^l}\right)}{\left(|x_i^l|+\epsilon+(|x_i^l|+\epsilon^l)^q\right)\log\left(|x_i^l|+\epsilon^l+(|x_i^l|+\epsilon^l)^q\right)}|x_i|& \\
\mbox {s.t. } & Ax=b,&\\
\end{split}\end{aligned}$$ where $0<p,q<1$, $A\in \mathbb{R}^{50\times 200},~b\in \mathbb{R}^{50},~\mbox{and}~x\in\mathbb{R}^{200}$.\
In our numerical works, we randomly generated the matrix $A\in \mathbb{R}^{50\times 200}$, and for a fixed sparsity, we randomly generated the solution vector $x\in \mathbb{R}^{200}$. We tested 100 randomly generated marices, $A$, for different level of $k$-sparsity of the solution, i.e., $k=1,2,...,26$. The matrix $A$ (the problem data) was randomly generated based on different statistical distributions. Most of the previous numerical experiments in the literature usually use normally distributed matrices.\
The distributions that we considered were Normal $(N(\mu, \sigma))$ with the parameters $\mu=0$ and $\sigma=1$, Poisson $(Pois(\lambda))$ with the parameter $\lambda=2$, Exponential $(Exp(\mu))$ with the parameter $\mu=5$, F-distribution $(F(\alpha,\beta))$ with the parameters $\alpha=1$ and $\beta=6$, Gamma distribution $(Gam(a,b))$ with parameters $a=5$ and $b=10$, and Uniform distribution $(U(N))$ with the parameter $N=10$. The probability of success of the 4 algorithms mentioned above, i.e, $l_1$-min, $CWB$, $W_1$, $W_2$ have been compared via different sparsity of the solution, and through all the above differently distributed matrices $A$. On a laptop with a Core 2 Duo CPU (2.00 GHz, 2.00GHz) and 4.00 GB of RAM memory, each comparing figure took approximately 14-hours time (in average).\
The updating rule $\epsilon^{l+1}=\frac{1}{2}\epsilon^l$ was used, at each iteration $l$. The choice of $\epsilon$ is crucial for reweighted $l_1$-algorithms. Hence, we have also tested the algorithms by applying Candes, Wakin, Boyd(CWB) updating rule for $\epsilon$, and also a fixed $\epsilon=0.01$. These figures demonstrate how these choices of $\epsilon$ may affect the performance of the algorithms.\
As seen, the weights in $W_1$ and $W_2$ vary for different values of $p$ and $p,q$. Therefore, we have tried different choices of $p$ and $q$ to find out how they may affect the success probability for $W_1$ and $W_2$ algorithms.\
In the Figure (3), the matrix A has been generated from $Exp(\mu)$, with $\mu=5$. We set $p=0.05$ in $W_1$, and $p=q=0.05$ in $W_2$. As shown, all of the algorithms are very successful when $\|x\|_0<7$. When $7<\|x\|_0<11$, CWB, $W_1$ and $W_2$ almost perform the same as each other, but when $\|x\|_0>11$, $W_1$ and $W_2$ outperform the CWB algorithm. All of the algorithms fail when the cardinality of the solution is above 25, i.e, $\|x\|_0>25$.\
In Figure (4), the matrix A has been generated from $Exp(\mu)$, with $\mu=5$, as in Figure (3). However, in this case we used different values for $p$ and $q$. We chose $p=q=0.4$, which is much larger than $0.05$. As expected, both $W_1$ and $W_2$ perform significantly worse than the case of $p=q=0.05$. Even for lower sparsity, both algorithms fail to locate the exact sparse solution with a high probability.\
In Figure (5), the matrix $A$ has been generated from $F(\alpha, \beta)$, with $\alpha=1$ and $\beta=6$, and we set $p=q=0.05$. As shown, all of the algorithms start failing when the cardinality of the solution is higher than 4, i.e, $\|x\|_0>4$. $W_1$ and $W_2$ perform better than $CWB$ for higher cardinality of the solution, and $W_2$ is slightly better than $W_1$ in general.\
In Figure (6), the matrix $A$ has been generated from $Gam(a,b)$, with $a=5$ and $b=10$, and we set $p=q=0.05$. For lower cardinality of the solution, $CWB$ and $W_1$ perform slightly better than $W_2$ when $\|x\|_0<8$. Also $CWB$, $W_1$, and $l_1$-min are completely successful for locating the exact solution, when $\|x\|_0<8$. But for $8<\|x\|_0<10$, only $CWB$ and $W_1$ are successful. $l_1$-min, $CWB$, $W_1$ and $W_2$ fail when $\|x\|_0>17$, $\|x\|_0>21$, $\|x\|_0>24$, $\|x\|_0>26$, respectively. Therefore $W_1$ and $W_2$ perform significantly better for higher cardinality of the solution.\
In Figure (7), the matrix $A$ has been generated from $N(\mu,\sigma)$, with $\mu=0$ and $\sigma=1$, and we set $p=q=0.05$. As shown, $l_1$-min, $CWB$, and $W_1$ are very successful for finding the sparsest solution of the system when $\|x\|_0<8$. $W_1$ and $W_2$ perform better than the other two algorithms for higher cardinality of the solution.\
In Figure (8), the matrix $A$ has been generated from $N(\mu,\sigma)$, with $\mu=0$ and $\sigma=1$ as in the Figure(7). However, we chose bigger values for $p$ and $q$, i.e, $p=q=0.4$. For large values of $p$ and $q$, $W_2$ starts failing for $\|x\|_0>4$, and performs much worst than $W_1$, $CWB$, and $l_1$-min. Also, for higher cardinality of the solution $CWB$ performs better than $W_1$ and $W_2$. Hence, from this figure and the Figure (4), one may conclude that smaller values for $p$ and $q$ should be chosen in order to achieve better results. Note that for large values of $p$ and $q$ the merit functions in $W_1$ and $W_2$ are not good concave approximations of $l_0$-norm.\
In Figure (9), the matrix $A$ has been generated form $U(N)$, with $N=10$, and we set $p=q=0.05$. All of the algorithms except $W_2$ are successful for finding the sparsest solution of the system when $\|x\|_0<9$. For $9<\|x\|_0<12$, $CWB$ performs slightly better than $W_1$ and $W_2$. But for higher cardinality of the solution, $W_1$ and $W_2$ outperform $l_1$-min and $CWB$.\
In Figure(10), the matrix $A$ has been generated from $Pois(\lambda)$, with $\lambda=5$, and we set $p=q=0.05$. All of the algorithms except $W_2$ are successful for finding the sparsest solution of the system when $\|x\|_0<9$. For higher cardinality of the solution, $W_1$ and $W_2$ outperform the other algorithms.\
Clearly, for small values of $p$, the best algorithm is $W_1$ in general, i.e. for different cardinality of the solution and for different tested distributions. For all of the different tested distributions, both $W_1$ and $W_2$ (for small choices of $p$ and $q$) outperform $CWB$ when the cardinality of the solution is higher.\
In the Figures (11), (12), (13), we focused on the performance of $W_2$ algorithms for different values of $p$ and $q$ via different fixed cardinality of the solution. In the Figure (11), we fixed $p=0.08$ and set different values of $q$. We examined the probability of success of $W_2$ for different fixed sparsity of 5,10,15,20. As expected, when cardinality of the solution is lower the success probability of $W_2$ is higher. As seen, the probability of success for fixed sparsity of 5 is the highest, and the probability of success for fixed sparsity of 20 is the lowest. The Figures (12) and (13) show the same results for fixed $p=0.4$ and $p=0.8$, respectively.\
In the Figure (14), the performance of $W_1$ has been tested using different choices of $p$ and different fixed sparsity of the solution. As seen, in Figure (14), when $p$ increases from 0.04 to 1, the probability of success of the algorithm becomes lower(except some jumps). As shown, for different fixed sparsity of 5,10,15,20 the highest probability of success was achieved when $p=0.04$. Looking back to the merit function defined for the $W_1$ algorithm, one may see that for smaller values of $p$ the function is a better concave approximation of $l_0$-norm.\
As we have discussed before, the choice of $\epsilon$ for the reweighted $l_1$-algorithm is important. Either very small or very big $\epsilon$ may result in improper weights, which may cause the failure of the algorithms. In Figure (15), we fixed the sparsity of the solution ($k=15$) and set $p=0.05$. Different choices of $\epsilon$ have been tested to suggest what $\epsilon$ might be the good one for which $W_1$ performs better. The matrix $A$ has been generated from $N(0,1)$. As shown, when $\epsilon$ tends to zero (e.g. $\epsilon\approx 0.0001$), or when $\epsilon$ is big (e.g. $\epsilon \approx 0.1$), the probability of success decreases. Our numerical experiments, in Figure (15), show that $\epsilon=0.01$ is a good choice for the weights in $W_1$ algorithm.\
In Figure (16), we fixed $\epsilon=0.01$(with no updating rule) and compared the performance of $l_1$-min, $CWB$, $W_1$, $W_2$. Like Figure (7), the matrix has been generated from $N(0,1)$, and we set $p=q=0.05$. Our numerical experiment show that $W_1$ and $W_2$ significantly outperform $CWB$ especially for higher cardinality of the solution. Comparing Figure (16) and Figure (7), one may conclude that even for a fixed $\epsilon$, if chosen correctly, both $W_1$ and $W_2$ algorithms may perform very well to find a sparse solution.\
Again to show that how important the choice of $\epsilon$ is, we compare the performance of $l_1$-min, $CWB$, $W_1$, $W_2$ based on the Candes updating rule. As seen in Figure(17), $CWB$ outperforms both $W_1$ and $W_2$ for lower cardinality of the solution, i.e, when $\|x\|_0<12$ in our numerical experiments.\
Conclusion
==========
We introduced a few concave approximations for the function $\|x\|_0$. These approximations can be employed to define new weights for the reweighted $l_1$-algorithms, which are used to locate the sparse solution of a linear system of equations. Through numerical experiments, we compared the performance of these reweighted algorithms and some existing reweighted algorithms when applied to linear systems with different statistically distributed matrices $A$, and with different sparsity of the solution. We have also explained when the new reweighted algorithms outperform some existing algorithms in different situations. We have also illustrated that how different choices of $p$ and $q$ may affect the performance of the algorithms. In addition, we have shown that the choices of $\epsilon$ may remarkably affect the performance of these algorithms as well. \
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{exponential.eps}
\label{FIG2}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{exponential4.eps}
\label{FIG3}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{FDistribution.eps}
\label{FIG4}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{gamma.eps}
\label{FIG5}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{normal.eps}
\label{FIG6}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{normal4.eps}
\label{FIG7}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{uniform.eps}
\label{FIG8}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{Poisson.eps}
\label{FIG9}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{p08qdifferent.eps}
\label{FIG10}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{W2p4.eps}
\label{FIG11}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{p8qdifferw2.eps}
\label{FIG12}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{W1.eps}
\label{FIG9}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth, totalheight=.25\textheight]{RecoveryEpsilonNW1.eps}
\label{FIG14}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth, totalheight=.25\textheight]{FixedAlpha.eps}
\label{FIG14}
\end{array}$
$\begin{array}{c}
\includegraphics[width=0.5\textwidth,totalheight=0.25\textheight]{CWBupdate.eps}
\label{FIG13}
\end{array}$
[^1]: School of Mathematics, University of Birmingham, Edgbaston B15 2TT, United Kingdom. Email: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider two-sided matching markets, and study the incentives of agents to circumvent a centralized clearing house by signing binding contracts with one another. It is well-known that if the clearing house implements a stable match and preferences are known, then no group of agents can profitably deviate in this manner. We ask whether this property holds even when agents have *incomplete information* about their own preferences or the preferences of others. We find that it does not. In particular, when agents are uncertain about the preferences of others, *every* mechanism is susceptible to deviations by groups of agents. When, in addition, agents are uncertain about their *own* preferences, every mechanism is susceptible to deviations in which a single pair of agents agrees in advance to match to each other.'
author:
- 'Nick Arnosti[^1], Nicole Immorlica [^2], Brendan Lucier'
bibliography:
- '../UnravelingBibliography.bib'
date: 'February 16, 2015'
title: 'The (Non)-Existence of Stable Mechanisms in Incomplete Information Environments'
---
Introduction
============
Related Work {#sec:related}
============
Model and Notation
==================
Results
=======
Discussion
==========
[0.8]{}
[^1]: Dept. of Management Science & Engineering, Stanford University. Work conducted at Microsoft Research.
[^2]: Microsoft Research.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Finding novel atomically-thin heterostructures and understanding their characteristic properties are critical for developing better nanoscale optoelectronic devices. In this study, we investigate the electronic and optical properties of GaS-Ca(OH)$_2$ heterostructure using first-principle calculations. The band gap of the GaS-Ca(OH)$_2$ heterostructure is significantly reduced when compared with those of the isolated constituent layers. Our calculations show that the GaS-Ca(OH)$_2$ heterostructure is a type-II heterojunction which can be used to separate photoinduced charge carriers where electrons are localized in GaS and holes in the Ca(OH)$_2$ layer. This leads to spatially indirect excitons which are important for solar energy and optoelectronic applications due to their long lifetime. By solving the Bethe-Salpeter equation on top of single shot GW calculation (G$_0$W$_0$) the dielectric function and optical oscillator strength of the constituent monolayers and the heterostructure are obtained. The oscillator strength of the optical transition for GaS monolayer is an order of magnitude larger than Ca(OH)$_2$ monolayer. We also found that the calculated optical spectra of different stacking types of the heterostructure show dissimilarities, although their electronic structures are rather similar. This prediction can be used to determine the stacking type of ultra-thin heterostructures.'
author:
- 'E. Torun'
- 'H. Sahin'
- 'F. M. Peeters'
title: 'Optical Properties of GaS-Ca(OH)$_2$ bilayer heterostructure'
---
Introduction
============
The successful synthesis of graphene was a milestone in condensed matter physics and materials science. [@graphene1; @graphene2; @graphene3] Due to this remarkable achievement, a new field of quasi-two-dimensional (2D) materials has emerged which changed the perspective of materials research. Since then, the attention of material science and condensed matter physics has been widened towards new single layer structures such as silicene [@guzman; @seymur], germanene [@seymur; @houssa; @sahin], transition-metal dichalcogenides (TMDs) [@neto; @mak; @splendiani; @wang], alkaline-earth-metal hydroxide (AEMHs) [@port] and post-transition metal chalcogenides (PTMCs).[@ma; @ptmcs; @ptmcs2]
These individual monolayers posses various significant electronic and optical properties which make them promising candidates for the next generation of nanoscale devices.[@ataca; @miro; @chhowalla] For instance, monolayer TMDs are direct band gap semiconductors unlike their bulk counterparts, [@mak; @splendiani; @wang] and exhibit large exciton binding energies in the order of 0.1-1.0 eV which results in exciton resonances at room temperature. [@he; @klots; @chernikov; @ugeda] In addition, having strong coupling between the spin and the valley degrees of freedom opens up the possibility of valleytronic devices. [@heinz; @sallen; @xu]
In spite of the large amount of research on graphene-like structures and ultra-thin TMDs, studies on single layer AEMHs (i.e. Ca(OH)$_2$ and Ni(OH)$_2$) and PTMCs (i.e. GaS and GaSe) are sparse and have only very recently started. In the recent study of Aierken *et al.* it was shown that Ca(OH)$_2$ can be isolated in monolayer form and it is a direct band gap semiconductor independent of the number of layers. [@port] Similar to Ca(OH)$_2$, monolayer GaS and GaSe have been synthesized recently and it has been shown that both monolayers are indirect gap semiconductors and they are suitable candidates for use in field-effect transistors and nanophotonic devices. [@ptmcs; @late1; @late2; @hu]
Another important aspect of the mentioned layered structures is their usage as building blocks for novel multi-layer heterostructures. Recently, a new field of research in materials science has emerged that deals with the stacking of two or more different monolayers on top of each other, namely van der Waals (vdW) heterostructures.[@geim] This widens considerably the diversity of possibilities for new functionalities and increases the range of different electronic and opto-electronic applications. For instance, it has been shown that a p-n junction based on the hBN-WSe$_2$ heterostructure exhibit tunable electroluminescence [@ross]. In addition, TMD heterostructures show a tunable photovoltaic effect [@furchi] and can be used as tunnel diodes and transistors [@roy]. Long-lived interlayer excitons were observed by photoluminescence excitation spectroscopy in MoSe$_2$-WSe$_2$ heterostructure where electrons and holes are localized in different layers. [@rivera] It has been shown that spatially indirect excitons can also be found in MoS$_2$-WSe$_2$ heterostructures which is a type II heterojuction. [@fang] Very recently, Calman *et al.* showed that the excitons in MoS$_2$ and hexagonal boron nitride (hBN) vdW heterostructure can be controlled by a gate voltage, temperature, and the intensity and the helicity of the optical excitation. [@calman]
So far, the building blocks of these multi-layer heterostructures are usually graphene (and graphene-like monolayers) or TMD monolayers. Whereas, in this work we propose a new kind of heterostructure whose building blocks are GaS and Ca(OH)$_2$ which are relatively new 2D crystals and they are members of PTMCs and AEMHs, respectively. We show that the GaS and Ca(OH)$_2$ monolayers have similar lattice parameters and they are wide band gap semiconductors. When they are stacked on top of each other they form a type II heterojuction which has a smaller band gap than the constituent layers. The calculated optical oscillator strength of the GaS monolayer is $\sim$10 times larger than the one for monolayer Ca(OH)$_2$.
This paper is organized as follows: We first provide the computational methodology in Sec. II. Then, we investigate the geometric, electronic and optical properties of the isolated GaS and Ca(OH)$_2$ monolayers in Sec. III and the GaS-Ca(OH)$_2$ heterostructure in Sec. IV. In Sec. V, we show that the optical spectra of the heterostructure can be used to characterize their stacking type. Finally, we conclude our results in Sec. VI.
Computational Methodology
=========================
First-principle calculations are performed using the frozen-core projector augmented wave (PAW) [@paw1] potentials as implemented in the Vienna Ab-initio Simulation Package (VASP).[@vasp1] The electronic exchange-correlation potential is treated within the generalized gradient approximation (GGA) of Perdew-Burke-Ernzerhof (PBE).[@pbe1] A plane-wave basis set with kinetic energy cutoff of 500 eV is used. A vacuum spacing of more than 12 Å is taken to prevent interaction between adjacent images. A set of 20$\times$20$\times$1 $\Gamma$ centered **k**-point sampling is used for the primitive unit cells. The structures are relaxed until self-consistency for ionic relaxation reached 10$^{-5}$ eV between two consecutive steps. Pressures on the lattice unit cell are decreased to values less than 1.0 kBar in all directions. Atomic charge transfers are calculated using Bader’s charge analysis. [@bader1] We used the DFT-D2 method of Grimme as implemented in VASP for the van der Waals correction in all the calculations. [@grimme]
The dielectric function and the optical oscillator strength of the individual monolayers and the heterostructure are calculated by solving the Bethe-Salpeter equation (BSE) on top of the single shot GW (G$_{0}$W$_{0}$) calculation which is performed on top of standard DFT calculations including spin-orbit coupling (SOC). During this process we used $6\times6\times1$ $\Gamma$ centered **k**-point sampling. The cutoff for the response function was set to 200 eV. The number of bands used in our calculation is 320. The cutoff energy for the plane-waves was chosen to be 400 eV. We include 4 valence and 4 conduction bands into the calculations in the BSE step.
Once the dielectric function is obtained other linear optical spectral quantities such as layer dependent optical absorbance ($A(\omega)$), transmittance ($T(\omega)$) and frequency dependent reflectivity at normal incidence ($R(\omega)$) can be calculated using the formula;
$$\begin{split}
A(\omega)&=\frac{\omega}{c} L \text{Im}\varepsilon(\omega), \\
T(\omega)&=1-A(\omega),\\
R(\omega)&=\left|\frac{\sqrt{\varepsilon(\omega)}-1}{\sqrt{\varepsilon(\omega)}+1}\right|^2
\end{split}$$
where c, L, $\omega$ and $\varepsilon(\omega)$ are the speed of light, the length of the cell in layer-normal direction, frequency of light and complex dielectric function, respectively.
GS and C(OH)$_2$ monolayers
===========================
The optimized atomic structures of single layers of GaS and Ca(OH)$_2$ are shown in Fig. \[fig1\]. The optimized lattice parameters for isolated monolayers are almost equal to each other, 3.58 Å and 3.59 Å for GaS and Ca(OH)$_2$, respectively. In the GaS monolayer, every Ga atom is covalently bonded to three S and one Ga atom which creates a trigonal prismatic symmetry. The distance between two nearest neighbor Ga atoms is 2.44 Å and the Ga-S distance 2.36 Å is slightly shorter. In the Ca(OH)$_2$ monolayer, the Ca atom has 6 ionic bonds with O atoms with a bond length of 2.36 Å. Each O atom in the primitive cell has one bond with a H atom with bond length 0.96 Å.
![\[fig1\] The optimized atomic structures of (a) GaS and (c) Ca(OH)$_2$ monolayers and their electronic structures and PDOS (right panel) with SOC (b) and (d), respectively. The gallium, sulfur, calcium, oxygen and hydrogen atoms are shown in green, yellow, blue, red and white, respectively. (e) The calculated band alignment of the monolayers where the vacuum level of both materials is set to 0 eV.](fig1.pdf){width="13cm"}
The band structures including SOC of the individual monolayers are shown in Fig. \[fig1\]. Both electronic structures are basically similar to the ones reported earlier. [@port; @ma] The GaS monolayer is an indirect band gap semiconductor with a gap of 2.59 eV, as seen in Fig. \[fig1\](b). The valence band minimum (VBM) of the compound resides along the $M-\Gamma$ direction but the conduction band minimum (CBM) is at the $M$ point in the Brillouin zone (BZ). The states in the vicinity of the Fermi level are composed of $p$ orbitals, however the VBM is mostly made up of the $s$ orbitals of the Ga and S atoms. Our Bader analysis showed that the Ga-S bonds have mostly a covalent character and the Ga atoms donate 0.7$e$ whereas each S atom gains 0.7$e$ charge.
Contrary to the GaS monolayer, the Ca(OH)$_2$ monolayer is a direct band gap semiconductor with a band gap of 3.66 eV. The VBM and CBM of the compound are located at the $\Gamma$ point in the BZ. The states at the vicinity of the Fermi level originate from the $p_{x}$ and $p_{y}$ orbitals of the O atom, however the VBM is mostly from the $s$ and $p_{z}$ orbitals of the Ca and O atoms. The Ca atom and H atoms donate, 1.6$e$ and 0.6$e$ charge respectively, and each O atom receive 1.4$e$. The bonds in Ca(OH)$_2$ have mostly an ionic character.
Our calculation revealed that the vacuum level of the isolated monolayers are different from each other. When the vacuum level of the monolayers is set to 0 eV as shown in Fig. \[fig1\](e), the obtained heterostructure is a type II heterojunction which opens up the possibility of using it as an electron-hole separator under photo-excitation. As shown in the figure, the excited electrons and the holes pile up in the GaS and Ca(OH)$_2$ monolayer, respectively. Since the electrons and the holes of the heterostructure reside in different layers, the created excitons are spatially indirect and the recombination occurs through the staggered gap of the heterojuction. It has been shown that spatially indirect excitons in MX$_2$ heterobilayers have a long lifetime ($\sim$ 20 - 30 ns at room temperature) which is important for applications in optoelectronics and photovoltaics. [@palummo]
In Figs. \[fig2\](a) and (b) the dielectric function and the oscillator strength of the optical transitions of isolated GaS and Ca(OH)$_2$ monolayers are shown, respectively. For GaS, the first peak of the dielectric function is composed of four optical transitions which are from valence band edge to conduction band edges at M, $\Gamma$ and K points in the Brillouin zone (BZ) which are very close in energy. For the Ca(OH)$_2$ monolayer, the first peak is composed of 2 optical transitions from $\Gamma$ to $\Gamma$ in the BZ. These two peaks are split by about 40 meV due to the SOC. The first peak of the dielectric function is considered as the optical band gap of the compound which is at $\sim$ 3.60 eV and $\sim$ 4.48 eV for GaS and Ca(OH)$_2$, respectively with exciton binding energy of 1.12 eV and 2.10 eV, respectively.
![\[fig2\] Imaginary part of the dielectric function and the oscillator strength of the optical transitions of (a) GaS monolayer (b) Ca(OH)$_2$ monolayer and (c) GaS-Ca(OH)$_2$ heterostructure. ](fig2.pdf){width="13cm"}
Heterostructure
===============
Having similar lattice constants opens up the possibility of using these monolayers for building blocks of vdW heterostructures. For this purpose we place GaS monolayer on top of Ca(OH)$_2$ monolayer. In order to find the minimum energy configuration of the heterostructure we shift the GaS monolayer along three different directions. The minimum energy configuration is obtained when the Ga and S atoms are on top of the O (and H) and Ca atom of the Ca(OH)$_2$ monolayer, respectively, as shown in Fig. \[fig3\]. The geometric structure of the constituent layers of the heterostructure does not change when compared with their isolated form. The distance between the two layers is 1.98 Å and the binding energy ($E_{B}$) of the heterostructure is 0.12 eV per unit cell (Table. \[table\]).
![\[fig3\] (a) The optimized atomic structures of GaS-Ca(OH)$_2$ heterostructure and (b) its electronic band structures together with the PDOS (right panel) with SOC. The gallium, sulfur, calcium, oxygen and hydrogen atoms are shown in green, yellow, blue, red and white, respectively.](fig3.pdf){width="13cm"}
Although the interaction between these two layers are weak, there is a significant decrease in the band gap when these two single layers are stacked on top of each other (see in Fig. \[fig3\](b)). This dramatic modification in the band gap is due to the difference in vacuum level of the constituent monolayers. Our calculations show that the heterostructure has an indirect band gap of 1.45 eV where the VBM is at the $\Gamma$ and the CBM is at the $M$ point in the BZ. The $\Gamma$ to $\Gamma$ gap is 1.63 eV. The advantage of having an indirect band gap heterostructure is that it results in long-lived excitons due to the small overlap of electron-hole wavefunctions. This opens up the possibility of using the GaS-Ca(OH)$_2$ heterojuction for excitonic solar cells so that the electron-hole pairs can be split relatively easy.
As shown in Table \[table\], the calculated workfunction of the heterostructure is found as 6.21 eV and 5.06 eV for GaS and Ca(OH)$_2$ sides of the heterojunction. Due to the interaction between two layers the workfunction values diminished 0.1 and increased 0.21 eV compared with the values for the isolated GaS and Ca(OH)$_2$ monolayers, respectively.
The partial density of states (PDOS) of the GaS-Ca(OH)$_2$ heterostructure is shown in the right panel of Fig. \[fig3\](b). As expected from the band alignment the valence band and the conduction band of the heterostructure are from GaS and Ca(OH)$_2$ monolayers, respectively. The resulting band structure of the heterostructure is shown in Fig. \[fig3\](b). As seen the band structure of the heterostructure is almost an overlap of the band structures of the isolated monolayers. This is due to the weak interaction between layers so that the dispersion of the bands does not change.
In order to investigate only the spatially indirect excitons in the heterostructure, 4 valence and 4 conduction bands were taken into account for the optical transitions in the BSE step. So, the optical transitions between these bands correspond to the interlayer recombination of the electrons and holes through the staggered gap. This means that the exciton peaks shown in Fig. \[fig2\](c) correspond only to spatially indirect excitons in the heterojunction. Consistent with the prediction of gap closing in the heterostructure, the first peak in the dielectric function appears at lower energy ($\sim$ 3.10 eV) than the one of the constituent monolayers as seen in the figure. The exciton binding energy of the heterostructure is calculated as 0.90 eV.
---------------- --------- --------- ---------------- --------- ----------- -- -- -- -- -- -- -- -- -- -- --
$a$=$b$ $E_{B}$ $E_{coh}$/atom $E_g$ $\Phi$
(Å) (eV) (eV) (eV) (eV)
GaS 3.58 — 3.83 2.59(i) 6.31
Ca(OH)$_2$ 3.59 — 4.59 3.66(d) 4.85
GaS-Ca(OH)$_2$ 3.58 0.12 4.26 1.45(i) 6.21,5.06
---------------- --------- --------- ---------------- --------- ----------- -- -- -- -- -- -- -- -- -- -- --
: \[table\] Ground state properties of GaS and Ca(OH)$_2$ monolayers and the heterostructure composed of these two monolayers. Calculated lattice parameters $a$ and $b$, interlayer binding energy $E_{B}$ (per unit cell), the band gap $E_g$ and the workfunction $\Phi$ (for the heterstructure the first value is for the GaS and the second one is for the Ca(OH)$_2$ side, respectively) of the structures.
Our calculations also reveal that the oscillator strengths of the GaS monolayer is an order of magnitude larger than that of the Ca(OH)$_2$ monolayer and the GaS-Ca(OH)$_2$ heterostructure. However, the oscillator strength and the dielectric function of the heterostructure resemble the properties of Ca(OH)$_2$. Similarly, as shown by Fang *et al.*[@fang], in case of the MoS$_2$-WSe$_2$ heterostructure, which is also a type II heterojunction, its optical properties resembles the one of MoS$_2$. Creation of such nanoscale type II heterojuctions leads to formation of spatially indirect excitons which are vital elements for exciton solar cells and optoelectronics devices due to the relatively longer life times of the excitons.
Optical Characterization of Stacking Type
=========================================
In this section, we will show that the optical spectra of ultra-thin heterostructures can be used to determine the stacking type of them. In Fig. \[fig4\], the optimized atomic configuration, the electronic structure and the optical transmittance and the reflectivity of the two lowest energy configurations of the GaS-Ca(OH)$_2$ heterostructure are shown. The atomic structures of these two lowest energy configurations are completely different from each other, although the energy difference between them is quite small $\sim$ 5 meV.
The higher energy state (Fig. \[fig4\](b)) is obtained by 180$^{\circ}$ rotation of the ground state configuration (Fig. \[fig4\](a)), so it will be called as “rotated state” from now on. In the rotated state the Ga and S atoms are on top of the Ca and O (and H) atom of the Ca(OH)$_2$ monolayer, respectively and in both cases the interlayer spacing is found as 1.98 Å. As shown in Figs. \[fig4\](c) and (d), the electronic structure of the ground and the rotated state are similar and the dispersion of the bands are identical. The only difference is that the band gap of the rotated state is 3 meV larger.
Although, the electronic structure of the ground and rotated states are very close to each other, optical transmittance and the reflectivity profiles of their band edges are distinct as shown in Figs. \[fig4\](e) and (f). The main differences in the optical spectra of these two stacking cases originate from the different strength of the optical transitions. It is predicted that the energies of these optical transitions are similar but their intensities are distinct. This is due to the different orientation of the atoms in these two stackings. For instance, the two peaks derived from the band edges are unique for both structures. As seen in the optical transmittance profile of the ground state (Fig. \[fig4\](e)) these two peaks can be easily identified because of their almost equal oscillator strength. In the rotated case however, the oscillator strength of the first peak is much smaller than the second one so that it can not be identified in the transmittance profile. On the other hand, the relative intensity of the first two peaks in the reflectivity profile are also unique for each stacking case. The intensity of the second peak is comparable but much smaller as compared to the first peak in the spectrum of the ground state.
This prediction is rather crucial to determine the stacking types of the heterostructures in experiments. The heterojunctions which are produced by growth technique might contain more than one stacking types in it, if the energy difference between different stacking types is small. Therefore, it is important to find a powerfull and sensitive tool to identify different stacking types in the produced heterostructures. In the light of above discussion, we suggest that the optical spectra (i.e. transmittance and reflectivity) can be used to identify different stacking types in a heterostructure.
![\[fig4\] The optimized atomic structures of the two lowest energy configurations of GaS-Ca(OH)$_2$ heterostructure; (a) ground and (b) rotated state which has 5 meV higher energy. Their electronic structure ((c) and (d)) and optical transmittance and reflectivity ((e) and (f)).](fig4.pdf){width="13cm"}
Conclusions
===========
We proposed a new kind of heterostructure in which the constituent monolayers are GaS and Ca(OH)$_2$ which are recently synthesized. These two monolayers have similar lattice constants and they are wide band gap semiconductors. When they are stacked on top of each other the electronic band gap of the obtained heterostructure decreases significantly. Our calculations revealed that the obtained heterojuction is a type II semiconductor in which the conduction band and the valence band are from GaS and Ca(OH)$_2$, respectively. This leads to localization of the electrons and holes in different layers which results in spatially indirect excitons. Our BSE-G$_0$W$_0$ calculations showed that the optical oscillator strength of the GaS monolayer is an order of magnitude larger than the one of monolayer Ca(OH)$_2$. We found that the band edge profiles of the optical spectrum of different stacking types of the heterojunction show differences although their electronic structures are rather similar. This prediction opens up the possibility of using the optical spectrum for the characterization of the stacking type of the ultra-thin heterostructures.
Acknowledgments
===============
This work was supported by the Flemish Science Foundation (FWO-Vl) and the Methusalem foundation of the Flemish government. Computational resources were provided by TUBITAK ULAKBIM, High Performance and Grid Computing Center (TR-Grid e-Infrastructure), and HPC infrastructure of the University of Antwerp (CalcUA) a division of the Flemish Supercomputer Center (VSC), which is funded by the Hercules foundation. H.S. is supported by a FWO Pegasus long Marie Curie Fellowship.
[99]{}
K. S. Novoselov, A. K. Geim, S. V. Morozov, D. Jiang, Y. Zhang, S. V. Dubonos, I. V. Grigorieva, and A. A. Firsov, Science **306**, 666 (2004).
K. S. Novoselov, D. Jiang, F. Schedin, T. J. Booth, V. V. Khotkevich, S. V. Morozov, and A. K. Geim, Proc. Natl. Acad. Sci. USA **102**, 10451 (2005).
A. K. Geim and K. S. Novoselov, Nat. Mater. **6**, 183 (2007).
G. G. Guzmàn-Verri and L. C. Lew Yan Voon, Phys. Rev. B **76**, 075131 (2007).
S. Cahangirov, M. Topsakal, E. Akturk, H. Sahin, and S. Ciraci, Phys. Rev. Lett. **102**, 236804 (2009).
M. Houssa, G. Pourtois, V. V. Afanasév, and A. Stesmans, Appl. Phys. Lett. **96**, 082111 (2010).
H. Sahin, S. Cahangirov, M. Topsakal, E. Bekaroglu, E. Akturk, R. T. Senger, and S. Ciraci, Phys. Rev. B **80**, 155453 (2009).
A. H. C. Neto and K. Novoselov, Rep. Prog. Phys. **74**, 082501 (2011).
K. F. Mak, C. Lee, J. Hone, J. Shan, and T. F. Heinz, Phys. Rev. Lett. **105**, 136805 (2010).
A. Splendiani, L. Sun, Y. Zhang, T. Li, J. Kim, C.-Y. Chim, G. Galli, and F. Wang, Nano Lett. **10**, 1271 (2010).
Q. H. Wang, K. Kalantar-Zadeh, A. Kis, J. N. Coleman, and M. S. Strano, Nat. Nanotechnol. **7**, 699 (2012).
Y. Aierken, H. Sahin, F. Iyikanat, S. Horzum, A. Suslu, B. Chen, R. T. Senger, S. Tongay, and F. M. Peeters, Phys. Rev. B **91**, 245413 (2015).
Y. Ma, Y. Dai, M. Guo, L. Yu, and B. Huang, Phys. Chem. Chem. Phys. **15**, 7098 (2013).
X. Meng, A. Pant, H. Cai, J. Kang, H. Sahin, B. Chen, K. Wu, S. Yang, A. Suslu, F. M. Peeters, and S. Tongay, Nanoscale **7**, 17109 (2015).
T. Cao, Z. Li, and S. G. Louie, Phys. Rev. Lett. **114**, 236602 (2015).
C. Ataca, H. Sahin, and S. Ciraci, J. Phys. Chem. C **116**, 8983 (2012).
P. Miró, M. Audiffred, and T. Heine, Chem. Soc. Rev. **43**, 6537 (2014).
M. Chhowalla, H. S. Shin, H. S. Shin, L.-J. Li, K. P. Loh, and H. Zhang, Nat. Chemistry **5**, 263 (2013).
K. He, N. Kumar, L. Zhao, Z. Wang, K. F. Mak, H. Zhao, and J. Shan, Phys. Rev. Lett. **113**, 026803 (2014).
A. R. Klots, A. K. M. Newaz, B. Wang, D. Prasai, H. Krzyzanowska, J. Lin, D. Caudel, N. J. Ghimire, J. Yan, B. L. Ivanov, K. A. Velizhanin, A. Burger, D. G. Mandrus, N. H. Tolk, S. T. Pantelides, and K. I. Bolotin, Scientific Reports **4**, 6608 (2014).
A. Chernikov, T. C. Berkelbach, H. M. Hill, A. Rigosi, Y. Li, O. B. Aslan, D. R. Reichman, M. S. Hybertsen, and T. F. Heinz, Phys. Rev. Lett. **113**, 076802 (2014).
M. M. Ugeda, A. J. Bradley, S.-F. Shi, F. H. da Jornada, Y. Zhang, D. Y. Qiu, W. Ruan, S.-K. Mo, Z. Hussain, Z.-X. Shen, F. Wang, Steven G. Louie, and Michael F. Crommie, Nat. Mat. **13**, 1091 (2014).
K. F. Mak, K. He, J. Shan, and Tony F. Heinz, Nat. Nanotechnol. **7**, 494 (2012).
G. Sallen, L. Bouet, X. Marie, G. Wang, C. R. Zhu, W. P. Han, Y. Lu, P. H. Tan, T. Amand, B. L. Liu, and B. Urbaszek, Phys. Rev. B **86**, 081301(R) (2012).
X. Xu, W. Yao, D. Xiao, and T. F. Heinz, Nat. Phys. **10**, 343 (2014).
D. J. Late, B. Liu, H. S. S. R. Matte, C. N. R. Rao, and V. P. Dravid, Adv. Funct. Mater. **22**, 1894 (2012).
D. J. Late, B. Liu, J. Luo, A. Yan, H. S. S. R. Matte, M. Grayson, C. N. R. Rao, and V. P. Dravid, Adv. Funct. Mater. **24**, 3549 (2012).
P. A. Hu, Z. Z. Wen, L. F. Wang, P. H. Tan, and K. Xiao, ACS Nano **6**, 5988 (2012).
A. K. Geim and I. V. Grigorieva, Nature (London) **499**, 419 (2013).
J. S. Ross, P. Klement, A. M. Jones, N. J. Ghimire, J. Yan, D. G. Mandrus, T. Taniguchi, K. Watanabe, K. Kitamura, W. Yao, D. H. Cobden, and X. Xu, Nat. Nanotechnol. **79**, 268 (2014).
M. M. Furchi, A. Pospischil, F. Libisch, J. Burgdörfer, and T. Mueller, Nano Lett. **14**, 4785 (2014).
T. Roy, M. Tosun, X. Cao, H. Fang, D.-H. Lien, P. Zhao, Y.-Z. Chen, Y.-L. Chueh, J. Guo, and A. Javey, ACS Nano **9**, 2071 (2015).
P. Rivera, J. R. Schaibley, A. M. Jones, J. S. Ross, S. Wu, G. Aivazian, P. Klement, K. Seyler, G. Clark, N. J. Ghimire, J. Yan, D. G. Mandrus, W. Yao, and X. Xu, Nat. Commun. **6**, 6242 (2015).
H. Fang, C. Battaglia, C. Carraro, S. Nemsak, B. Ozdol, J. S. Kang, H. A. Bechtel, S. B. Desai, F. Kronast, A. A. Unal, G. Conti, C. Conlon, G. K. Palsson, M. C. Martin, Andrew M. Minor, C. S. Fadley, E. Yablonovitch, R. Maboudian, and A. Javey, Proc. Natl. Acad. Sci. USA **111**, 6198 (2014).
E. V. Calman, C. J. Dorow, M. M. Fogler, L. V. Butov, S. Hu, A. Mishchenko, and A. K. Geim, arXiv:1510.04410v1.
G. Kresse and D. Joubert, Phys. Rev. B **59**, 1758 (1999).
G. Kresse and J. Hafner, Phys. Rev. B **47**, 558(R) (1993).
J. P. Perdew, K. Burke, and M. Ernzerhof, Phys. Rev. Lett. **77**, 3865 (1996).
R. F. W. Bader, *Atoms in Molecules - A Quantum Theory* (Oxford University Press, Oxford, UK, 1990).
S. Grimme, J. Comput. Chem. **27**, 1787 (2006).
M. Palummo, M. Bernardi, and J. C. Grossman, Nano Lett. **15**, 2794 (2015).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We numerically study the dual field theory evolution of five-dimensional asymptotically anti-de Sitter solutions of supergravity that develop cosmological singularities. The dual theory is an unstable deformation of the ${\cal N}=4$ gauge theory on $ \mathbb{R} \times S ^{3} $, and the big crunch singularity in the bulk occurs when a boundary scalar field runs to infinity. Consistent quantum evolution requires one imposes boundary conditions at infinity. Modeling these by a steep regularization of the scalar potential, we find that when an initially nearly homogeneous wavepacket rolls down the potential, most of the potential energy of the initial configuration is converted into gradient energy during the first oscillation of the field. This indicates there is no transition from a big crunch to a big bang in the bulk for dual boundary conditions of this kind.'
author:
- |
Lorenzo Battarra[^1] and Thomas Hertog[^2]\
\
\
\
bibliography:
- 'AdSCrunch7.bib'
nocite: '[@Hertog:2004ns; @Maldacena:1997re; @Elitzur:2005kz; @Gasperini1993]'
title: '**Particle Production near an AdS Crunch**'
---
Introduction
============
One of the main goals of quantum gravity is to advance our understanding of the nature of the big bang. A particularly important issue is whether the big bang represents the beginning of the universe, or whether semiclassical evolution essentially continues further back in time, possibly involving a transition from a big crunch to a big bang [@Gasperini1993; @Khoury2002].
In recent years this problem has been studied in the context of the AdS/CFT correspondence [@Maldacena:1997re], which allows one to map toy model cosmologies with anti-de Sitter (AdS) boundary conditions to a dual quantum field theory living on the boundary of spacetime. Explicit examples of collapsing $AdS$-cosmologies were constructed in $\N=8$ supergravity in four [@Hertog:2004rz; @Hertog:2005hu] and five [@Craps2007] dimensions, using well-defined generalizations of the usual boundary conditions on some of the negative mass squared scalars which allow smooth, asymptotically $AdS$ initial data to evolve into a spacelike singularity that reaches the spacetime boundary in finite time. The properties of these solutions and their duals were further explored in [@Elitzur:2005kz; @Barbon2010; @Bernamonti2009; @Asnin2009]. However, a clear understanding of the quantum nature of cosmological singularities is still lacking.
The dual description of the $AdS$-cosmologies found in [@Hertog:2004rz; @Hertog:2005hu; @Craps2007] involves field theories with scalar potentials which are unbounded below that drive certain boundary scalars to infinity in finite time. Here we concentrate on the five-dimensional solutions, where the dual field theory is a deformation of ${\cal N}=4$ Super-Yang-Mills (SYM) theory on $ \mathbb{R} \times S ^{3} $ by an unbounded double trace potential $-f {\cal O}^2/2$, with ${\cal O}$ a trace operator quadratic in the adjoint Higgs scalars. This deformation is renormalizable and the coupling $f$ that governs the instability is asymptotically free [@Craps2007].
The big crunch singularity in the bulk occurs when the boundary scalar ${\cal O}$ diverges. Consistent quantum evolution requires that one imposes boundary conditions at infinite scalar field, [*i.e.*]{} a self-adjoint extension of the system. With a self-adjoint extension the boundary evolution is unitary and predicts a perfectly homogeneous wave packet rolls down the negative potential and bounces back. In the bulk this behavior would correspond to a quantum transition from a big crunch to a big bang, as envisioned e.g. in ekpyrotic cosmology.
However, it has been argued [@Hertog:2004rz] that the full field theory evolution is likely to be very different. This is because when a homogeneous wave packet rolls down, the negative effective mass term in the potential amplifies long wavelength quantum fluctuations hereby converting the potential energy of the initial configuration into gradient energy. In order for the boundary theory to predict a transition from a collapsing phase in the bulk to an approximately homogeneous, semiclassical expanding cosmology after the singularity there must be a parameter regime where ${\cal O}$ returns close to its original value after one oscillation. However, this requires the backreaction of the inhomogeneities on the evolution of the homogeneous wave packet to be sufficiently small.
Here we quantify the tachyonic amplification of fluctuations by numerically evolving the field theory on a lattice. To study the question of a possible big crunch big bang transition, one is interested in the dynamics on rather short timescales and in particular in the first oscillation of the field. Since the occupation numbers of the relevant perturbation modes become rapidly very large, the system behaves classically during these first stages of the evolution. The scalar potential is unbounded below, but the self-adjoint extension essentially acts as a brick wall at infinity. We therefore model this by adding a steep regularization term to the unbounded potential and considering the limit in which the UV regulator $\gamma$ introduced this way is taken to be very small. We find that in the $\gamma \rightarrow 0 ^{+}$ limit, the tachyonic amplification of perturbations is a very efficient mechanism to produce particles and prevents the homogeneous mode from rolling back up the potential. We conclude, therefore, that for ‘brick-wall’ boundary conditions at infinity and for the class of models considered here, the dual description predicts no transition from a big crunch to a big bang in the bulk.
At the same time, however, our simulations provide some insight in what kind of models might exhibit a different behavior near the singularity. We discuss this briefly at the end of this paper.
AdS Cosmology and its Dual Description
======================================
AdS Cosmology
-------------
Our starting point is $ \mathcal{N} = 8$ gauged supergravity in five dimensions, which is thought to be a consistent truncation of ten-dimensional type IIB supergravity on $S ^5$. This admits a consistent truncation to gravity coupled to a single $SO(5)$-invariant scalar $\varphi$. The action then reduces to $$\label{lagr}
S = \int d ^5x \, \sqrt{-g} \left( \frac{1}{2} \mathrm{R}^{(2)} - \frac{1}{2} (\partial \varphi) ^2 +
\frac{1}{4R _{AdS}^2} \left( 15 e^{2 \gamma \varphi} + 10 e^{-4 \gamma \varphi} - e^{-10 \gamma \varphi} \right) \right)$$ where $ \gamma \equiv \sqrt{2/15}$ and where we have chosen units in which the five-dimensional Planck mass is unity. The maximum of the potential at $\varphi=0$ corresponds to the $AdS_5$ vacuum solution. Small fluctuations around this have $m^2=-4/ R_{AdS}^{2}$, which saturates the Breitenlohner-Freedman (BF) bound in five dimensions.
In all asymptotically AdS solutions, the scalar field $\varphi$ decays at large radial coordinate $r$ as $$\label{asscalar}
\varphi \sim\frac{ \alpha (t,\Omega) \ln{r}}{r ^2} + \frac{ \beta (t,\Omega)}{r ^2}$$ where $(t,\Omega)$ are the time and angular coordinates of the boundary $\mathbb{R} \times S ^{3} $ of the $AdS$ cylinder. To define the theory one ought to specify boundary conditions at $r=\infty$ on the metric and scalar field. This amounts to specifying a relation between $\a$ and $\b$ in (\[asscalar\]). The usual boundary conditions correspond to taking $\a=0$, and leaving $\b$ unspecified. Alternatively, one can adopt boundary conditions of the form [@Hertog:2004ns] \[genbc\] = -, where $W(\b)$ is an essentially arbitrary real, smooth function. Boundary conditions of the form (\[genbc\]) are invariant under global time translations. The conserved energy associated with this depends on $W$, but is well-defined and finite. In this paper we adopt scalar field boundary conditions of the form $\alpha = f \beta$, with $f$ an arbitrary constant. The corresponding asymptotic form of the metric and the expression of the conserved mass can be found in [@Hertog:2004dr; @Henneaux:2004zi].
For small, positive values of $f$, solutions were found [@Hertog:2004rz] in which smooth, spherically symmetric initial data of approximately zero mass evolve into a big crunch – a spacelike singularity that reaches the boundary of AdS in finite global time. These solutions can be viewed as a five-dimensional version of open FLRW universes in which $\varphi$ rolls down the negative potential, causing the scale factor $a(t)$ to vanish in finite time. In terms of global time $t$ it was found that $\b \sim\b(t=0)/(\cos t/R _{AdS})^2$. Hence $\b \rightarrow \infty$ as $t/R _{AdS} \rightarrow \pi/2$, which is when the singularity hits the boundary [@Hertog:2004rz].
Dual CFT description
--------------------
**
### Classical Evolution {#classical-evolution .unnumbered}
With the usual $\a=0$ boundary conditions, the dual field theory is $\N=4$ super Yang-Mills theory. The bulk scalars that saturate the BF bound in $AdS$ correspond in the gauge theory to the operators $\frac{1}{N} \,{{\rm Tr}}[\Phi^i \Phi^j - (1/6) \delta^{ij} \Phi^2]$, where $\Phi^i$ are the six scalars in $\N=4$ super Yang-Mills. The $SO(5)$-invariant bulk scalar $\varphi$ that we have kept in (\[lagr\]) couples to the operator \[operator\] [O]{}= [[Tr]{}]{}.
In general, imposing nontrivial boundary conditions $\a (\b)$ in the bulk corresponds to adding a multi-trace interaction $W({\cal O})$ to the CFT action [@Witten:2001ua; @Berkooz:2002ug], such that after formally replacing ${\cal O}$ by its expectation value $\b$ one obtains (\[genbc\]). Hence the boundary conditions $\alpha = f \beta$ that we have adopted correspond to adding a double trace term to the field theory action \[Ocubed\] S = S\_0 - W([O]{}) = S\_0+ \^2. The operator ${\cal O}$ has dimension two, so the extra term is marginal and preserves conformal invariance, at least classically.
We have taken the constant $f$ to be small and positive in the bulk. The term we have added to the CFT action, therefore, corresponds to a negative potential. Hence it is plausible that the dual field theory admits negative energy states and has a spectrum that is unbounded below. This would mean that the usual vacuum must be unstable, and that there are (non-gravitational) instantons which describe its decay. After the tunneling, the field rolls down the potential and becomes infinite in finite time. This corresponds to the big crunch singularity in the bulk and provides a qualitative dual explanation for the fact that the function $\beta$ of the asymptotic bulk solution (\[asscalar\]) diverges as $t/R \rightarrow \pi/2$, when the big crunch singularity hits the boundary. Indeed, $\beta$ is interpreted as the expectation value of ${\cal O}$ in the dual CFT, and the bulk analysis predicts that to leading order in $1/N$, $\langle {\cal O}\rangle$ diverges in finite time.
For our purposes it suffices to concentrate on the steepest negative direction of the potential. Fluctuations in orthogonal directions in field space acquire a positive mass and are suppressed. For the $SO(5)$-invariant operator we consider, the most unstable direction comes from the $-\Phi_1^4$ term in [(\[Ocubed\])]{}. From now on we focus on the dynamics of $\Phi_1$ as it rolls along a fixed direction in $SU(N)$, i.e. $\Phi_1(x)=\phi(x)U$, with $U$ a constant Hermitian matrix satisfying ${\rm Tr} \,U^2=1$, so that $\phi$ is a canonically normalized scalar field. The action for this scalar is then given by $$\label{scaction}
S = \int_{ \mathbb{R} \times S _{3}} d ^4x \sqrt{-g} \left( - \frac{1}{2}(\partial \phi) ^2 - \frac{1}{12} \mathrm{R} ^{(2)} \phi ^2 + \frac{1}{4} \lambda \phi ^4 \right)$$ where $\lambda = f/N^2$, $ \mathrm{R} ^{(2)} = 6/R^2$ is the Ricci scalar of the 3-sphere, and the coefficient of the $ \phi ^2$ term is fixed by the conformal symmetry.
Although this is clearly a huge simplification of the field theory, at the classical level it captures the bulk behavior in a surprisingly quantitative way. In particular, it admits the following exact homogeneous classical solution \[clsol\] (t) = . which is analogous to the background evolution of the AdS cosmology considered in [@Hertog:2004rz]. Indeed, since $\phi^2$ is identified with $\b$ on the bulk side, the time dependence of this solution agrees with the supergravity prediction[^3], including the fact that the field diverges at $t/R \rightarrow \pi/2$. Although conformal invariance is broken quantum mechanically, it has been shown that quantum corrections do not turn around the potential [@Craps2007].
### Quantum Evolution {#quantum-evolution .unnumbered}
We have seen that evolution ends in finite time in the classical description of a simplified version of the dual field theory, in agreement with the supergravity result. It has been noted, however, that this conclusion changes dramatically if one considers the quantum mechanics of the dual model (\[scaction\]). That is, one again concentrates on the homogeneous mode $\phi (t)=x(t)$ only, but one now treats this quantum mechanically.
In quantum mechanics, to ensure unitarity one constructs a self-adjoint extension of the Hamiltonian. This is done by carefully specifying its domain [@Reed:1975uy; @Carreau90]. One finds the center of a wave packet follows essentially the classical trajectory and still reaches infinity in finite time, but it bounces off infinity and reappears as a left-moving wave packet [@Carreau90]. Hence a quantum mechanical treatment of the homogeneous mode of the dual model appears to suggest that there is an immediate transition from a big crunch to a big bang in the bulk.
However, it has been argued [@Hertog:2004rz] that the full field theory evolution is likely to be very different. Indeed, a quantum mechanical analysis obviously neglects the possibility of particle production. Calculations in similar scalar field theories indicate that this significantly affects the evolution of the homogeneous mode. In particular, in many theories where the scalar field rolls down from the top of its effective potential, particles are produced in great numbers [*while the field is rolling down*]{}. This phenomenon is often called tachyonic preheating [@Kofman:2001rb; @Felder:2001kt]. It happens essentially because the negative effective mass term in the potential amplifies long wavelength quantum fluctuations.
Tachyonic preheating is so efficient that in some theories most of the initial potential energy density is converted into gradient energy well before the field reaches the true minimum. In theories of this kind a prolonged stage of oscillations of the homogeneous component of the scalar field around the true minimum of the potential does not exist in spontaneous symmetry breaking.
One expects that the tachyonic amplification of fluctuations similarly affects the evolution of the dual field theory description of the AdS cosmologies, in which the supergravity initial data correspond to a homogeneous field theory configuration high up the potential. In this case the scalar potential does not have a global minimum. However, the quantum mechanical self-adjoint extension – assuming for now it can be implemented in field theory – essentially acts as a brick wall at infinity. From a cosmological perspective, one is interested in the dynamics at intermediate times, namely whether a wave packet rolling down the potential bounces back one or several times before the system settles down. To answer this one must quantify the effect of the tachyonic instability, and the non-linear growth of fluctuations that comes with it, on the background evolution. This is the subject of the remainder of this paper.
Dual Model
==========
‘Brick Wall’ Regularization
---------------------------
We have seen that a self-adjoint extension in a quantum mechanical treatment of (\[scaction\]) acts as a brick wall at infinity. Perhaps the simplest way to try to implement this at the field theory level is to add a steep regularization term to the potential[^4], $$\label{eq:potentialReg}
V( \phi) = \frac{1}{2} \phi ^2 - \frac{1}{4} \lambda \phi ^4 + \epsilon \phi ^6$$ and to consider the dynamics in the limit $ \epsilon \rightarrow 0 ^{+}$. It is useful to introduce the parameter $\gamma \equiv 32 \epsilon /
\lambda ^{2}$, which can be thought of as the square of a UV regularization length. In terms of $\gamma$ the potential (\[eq:potentialReg\]) can be written as $$\begin{aligned}
\label{eq:potentialRescaling}
V( \phi) & = & \frac{1}{2} \phi ^2 - \frac{1}{4} \lambda \phi ^4 +
\frac{1}{32} \lambda ^2 \gamma \phi ^6 \\
& \equiv & \lambda ^{-1} V _1(\sqrt{\lambda} \phi)\end{aligned}$$ where $V _1$ is defined as $V$ with $\lambda = 1$. For $ \gamma \leq 1$, the potential has two positive roots $ \phi _{ \pm}$ (see Fig \[fig:potential\]). The global minimum of V on the positive axis is located at $ \phi _{m}$, and we will write $V _{m} \equiv V( \phi _{m})$.
For $ \gamma \rightarrow 0 ^{+}$, to which we will refer as the *scaling limit*, $ \phi _{-}$ is insensitive to the regularization and tends to $ \sqrt{2 / \lambda}$. On the contrary, the large $ \phi$ parameters $ \phi _{m}$, $ \phi _{+}$ and $V _{m}$ follow the scaling given by $ \gamma$: $$\lambda \phi _{+} ^{2} \simeq 8 \gamma ^{-1}, \quad \lambda \phi _{m} ^2 \simeq \frac{16}{3 } \gamma ^{-1} , \quad\lambda V _{m} \simeq -\frac{64}{27} \gamma ^{-2}.$$
![\[fig:potential\] Regularized scalar potential in the boundary model with $ \lambda = 1$ and $ \gamma = 0.4$. ](potential.pdf){width="8cm"}
In order to model the background evolution of the AdS cosmologies we consider homogeneous field theory solutions $\bar \phi (t)$ that start at rest at $ \bar{ \phi} = \phi_{-}$. If one neglects fluctuations, the homogeneous field oscillates back and forward between the positive roots of the potential, just as in the quantum mechanical treatment of the unbounded potential discussed above. The period $T$ of the homogenous field oscillations is independent of the potential parameters and given by $$\label{eq:period}
T = \sqrt{2} \int _{\phi _{-}} ^{\phi _{+}} \frac{d \phi}{ \sqrt{- V(\phi)}} = \pi$$ Hence in the limit $\gamma \rightarrow 0 ^{+}$ the field reaches infinity exactly at $t = \pi/2$, which corresponds to the time at which the big crunch singularity in the bulk reaches the boundary. This suggests that for small $\gamma$ the $\phi ^{6}$ regularization acts very much as a brick wall. Indeed, the interval of time $T_{+}$ during which $ \bar{\phi} \ge \phi _{m}$ in one oscillation scales as $ \sqrt{\gamma}$, T \_[+]{} = \_[\_[m]{}]{} \^[\_[+]{}]{} .
However, the regularization of the boundary potential affects the evolution in the bulk, because it changes the bulk boundary conditions on the timelike boundary. According to the general prescription (\[genbc\]) these now read $$\alpha = \lambda \beta - 3 \epsilon \beta ^2.$$ For small $\epsilon$ this can only significantly affect the evolution near where the singularity hits the boundary, where $\b$ is large. Nevertheless this suffices to turn the cosmological singularity into a large, stable black hole with scalar hair [@Hertog:2005hu], where the bulk scalar field turned on is dual to the operator (\[operator\]) in the boundary theory. These hairy black holes, which do not exist for $\a= \lambda \b$ boundary conditions, have been interpreted in the dual theory as thermal excitations about the global negative minimum $V_m$ [@Hertog:2005hu]. Hence they are the natural end state of evolution in the bulk corresponding to wave packets rolling down a regularized potential (\[eq:potentialReg\]). In the limit $\epsilon \rightarrow 0$ where the global minimum goes to minus infinity the hairy black holes become infinitely large and one recovers the original cosmological solutions [@Hertog:2005hu].
Hence in the black hole context there is convincing evidence that the dual system will eventually thermalize. However here we are more interested in the dynamics at intermediate times. Indeed, whether the dual theory predicts a transition from a big crunch to a big bang depends on whether the backreaction of the inhomogeneities prevents an initially homogeneous wave packet rolling down the potential from returning close to its original value after the [*first*]{} oscillation.
Initial Conditions of Perturbations
-----------------------------------
Following [@Kofman:2001rb; @Desroche2005] we set the fluctuations initially in their instantaneous adiabatic vacuum, specified by $$\begin{aligned}
\label{eq:initialSpectrum}
\langle \delta \phi(\vec{k})\, \delta \phi(\vec{k}') \rangle & \equiv & \frac{(2 \pi) ^{3}}{2\,\omega _{0} (k)} \delta ^{3}( \vec{k} + \vec{k}')\nonumber \\
\delta \dot{\phi}(\vec{k}) & = & \pm i\, \omega _{0}(k) \delta \phi(\vec{k})\end{aligned}$$ Because of the unstable $ - \phi^4$ term in the potential, one has $m _{0} ^{2} \equiv V_{,\phi \phi}(\phi _{-}) < 0$. We therefore set \[omega\] \_[0]{} (k) \^[2]{} = {
[cr]{} k \^[2]{} & k < |m \_[0]{}|\
k \^[2]{} + m \^[2]{}\_[0]{} & k |m \_[0]{}|
. to avoid that $\omega _{0}(k)$ is imaginary for $k<|m _{0}|$. The initialization of the frequency [(\[omega\])]{} for the modes with $ k < |m _0|$ is to some extent arbitrary. However, in the limit $\gamma \rightarrow 0$ these modes, which are tachyonic at $t = 0$, constitute only a small part of the ensemble of modes that become tachyonic at some point during the evolution. As part of our numerical study, we have verified that changing the initialization of the $k < |m _0|$ modes does not significantly affect the evolution.
We are initializing random *classical* fluctuations to reproduce the quantum vacuum fluctuations. The consistency of this approach relies on the fact that, as we will show, in the subsequent evolution the occupation numbers become rapidly very large so that the quantum evolution behaves classically, at least during the first stages of the evolution.
Finally we note that for fixed $\gamma$, changing $\lambda$ is equivalent to rescaling the width of the initial wavepacket. Indeed, defining $\chi \equiv \sqrt{\lambda} \phi$, the field equation for $ \chi$ is given by $$\label{eq:eomReduced}
\ddot{\chi} - \Delta \chi + \frac{d V _1}{d \chi} = 0$$ which is the field equation for $\phi$ with $\lambda = 1$. The homogeneous mode $\bar \chi (t)$ also takes the correct initial value, since $V _1 ( \sqrt{ \lambda} \phi _{-}) = 0$. By contrast the amplitude of the fluctuations $\delta \chi$ is rescaled by a factor $\sqrt{\lambda}$ with respect to the $\lambda = 1$ model (with $ \hbar = c = 1$): $$\langle \delta \chi(\vec{k})\, \delta \chi(\vec{k}') \rangle = \lambda \frac{(2 \pi) ^{3}}{2\,\omega _{0} (k)} \delta ^{3}( \vec{k} + \vec{k}').$$ Hence to study the evolution for a range of values of $\lambda$ one can equivalently change the width of the initial wavepacket, and then go back to the physical variable $ \phi$.
Tachyonic instability
---------------------
The effective mass $m ^{2} = V''(\phi)$ of $\phi$ is negative in a range of field values of width $\Delta \phi \sim 1/ \sqrt{ \lambda \gamma}$, extending roughly from $ \phi = \phi _{-} \simeq \sqrt{2/ \lambda}$ to $ \phi \simeq \sqrt{3/ \lambda \gamma}$. The lowest value of $m ^{2}$ is $$\label{eq:tachScale}
m ^{2} _{min} = 1 - \frac{12}{5 \gamma} \stackrel{ \gamma \ll 1}{ \simeq} - \frac{12}{5 \gamma}.$$ This value is $ \lambda$-independent, as can be simply deduced from the fact that $V_{,\phi \phi} = {V _1}_{,\chi \chi}$. Modes with momenta $k^2$ between $0$ and $k _{T}^2 \equiv -m ^{2} _{min}$ have an imaginary frequency at some point during the evolution and hence are tachyonic. Decreasing $\gamma$ not only widens the tachyonic band but also amplifies more strongly the modes within the tachyonic range $x \equiv k/k _{T} < 1$. This provides a mechanism for converting potential energy into particles, acting as a friction term for the motion of the homogeneous component of the scalar field.
Eq.(\[eq:tachScale\]) shows that $\gamma$ acts as a UV regularization length squared. Indeed, the tachyonic instability produces particles with momenta $$k \lesssim k _{T} \sim \gamma ^{-1/2}$$ irrespective of the value of $\lambda$. However, the details of the spectrum will generally be $ \lambda $ dependent at times sufficiently large for the non-linearities to play a significant role.
Systems of this kind have been examined in the context of cosmological preheating and dynamical symmetry breaking [@Kofman:2001rb; @Desroche2005; @Felder:2001kt]. The general conclusion is that the tachyonic amplification is a very efficient mechanism to convert potential energy into particles so that it typically takes only a few oscillations before the homogeneous mode stabilizes around the global minimum. A sample evolution of the homogeneous field in our regularized model (\[eq:potentialReg\]) of the dual theory is given in Fig \[fig:meanPhi\], for $ \lambda/4 = 10 ^{-6}$ and $ \gamma = 0.3$. One sees that with these parameter values the field returns twice to close its original value before backreaction suddenly kicks in and dramatically changes the evolution. Once the homogeneous mode, which is obtained by averaging the field value over the lattice sites, has settled around the global minimum the system thermalizes on a much longer timescale through the gradual transfer of energy into higher harmonic modes. The central question we wish to address here is whether the first oscillation survives in the $\gamma \rightarrow 0 ^{+}$ limit.
To answer this one needs to take in account the evolution of the fluctuations at the non-linear level. This can be seen as follows. The linearized perturbation equation reads $$\label{linperteq}
\ddot{ \phi _{k}} + (k^2 + m^2(t)) \phi _{k} = 0$$ where $m ^2(t) \equiv V _{, \phi \phi}( \bar{ \phi}(t))$. In the absence of the regularization term, one finds that where $ \tau \equiv \pi/2 - t$ is the time away from the singularity. Hence for $ \tau \ll 1/k$, becomes $$\ddot{ \phi _{k}} - \frac{6}{ \tau^2}\phi _{k} = 0$$
![\[fig:meanPhi\] Evolution of the homogeneous component $ \bar{ \chi}(t)$, calculated by averaging the field value over the lattice sites, for $ \lambda/4 = 10 ^{-6}$ and $ \gamma = 0.3$, taking in account the growth of fluctuations $\delta \chi$. The colored band represents the $ \bar{ \chi} \pm \sqrt{ \langle \delta \chi ^2 \rangle}$ zone.](oscillations.pdf){width="8cm"}
which is solved by $$\label{eq:linearizedGrowth}
| \phi _{k} ( \tau) | \simeq \frac{| \phi _{k} ( t = 0) | }{ (k \tau )^2}.$$ The IR divergence in (\[eq:linearizedGrowth\]) obviously comes from using the near-singularity behaviour of the classical solution and has no physical meaning. It can be dealt with by imposing an IR cutoff $k _{IR}$ which mimics the actual solution. At a given $ \tau$, one has $$\langle \delta \phi ^2 \rangle ( \tau) \gtrsim 4 \pi \int _{k _{IR}}^{1/ \tau} dk \, k^2 \frac{ \langle | \phi _{k} (t=0) |^2 \rangle}{ (k \tau)^4} \sim \tau ^{-4} \sim \lambda^2 \bar{ \phi}^4$$ where we have restricted the range of integration to the modes that are tachyonic at time $\tau$. This indicates that while the field rolls down, non-linearities become important when $ \bar{ \phi} \sim 1/ \lambda$.
Lattice Evolution
=================
To study the dynamics in the presence of the brick wall regularization, we have developed a numerical code that solves the non-linear Klein–Gordon equation (\[eq:eomReduced\]) on a 3–dimensional cubical lattice with periodic boundary conditions. We employ the same second order algorithm used in [@Felder:2000hq]. Significant improvements in the numerical performance were made with respect to previous codes. In this section we first describe the numerical setup in more detail and then discuss our results for the classical evolution on the lattice.
Numerical Setup
---------------
### IR cutoff {#ir-cutoff .unnumbered}
The scalar theory (\[scaction\]) is defined on $ \mathbb{R} \times S ^{3} $. The conformal coupling to the curvature of the is crucial to determine the initial conditions for the evolution. On the other hand, in the scaling regime, the tachyonic scale $l _{T}^2 = k _{T} ^{-2} \propto \gamma$ becomes much smaller than $R$. Hence one expects the IR cutoff given by the radius of the sphere to be unimportant in the $\gamma \rightarrow 0$ limit, so that one can safely work on a toroidal lattice. To verify this is indeed the case we have compared the evolution with the lattice size adjusted so that the momentum gap between the homogeneous mode and the lowest $k$-mode is the same as on the sphere, to the evolution on a very large lattice. It was found that in the scaling regime, the finite volume of the 3–sphere has no effect[^5].
Finally we have verified that, again in the scaling regime, the statistical dependence on the (pseudo–)random initial conditions is negligible if one considers averaged quantities. This provides a consistency check for the semi–classical approach we employ.
### UV convergence {#uv-convergence .unnumbered}
We are particularly interested in the early stages of the evolution, such as the dependence of the first oscillation of the field on the parameter values. It turns out that, as expected, in this regime only tachyonic modes with $x \lesssim 1$ are important. This is illustrated in (right panel) where we plot the mode occupation number as a function of $x$ after one oscillation. One sees the occupation numbers fall sharply for $x \geq 1$. Hence one expects that a lattice cutoff at $x _{UV} \simeq 2 $ should be sufficient to ensure the UV convergence of the lattice evolution. This turns out to be the case: in Fig \[fig:UVconv1\] we show that the homogeneous field displacement after one oscillation becomes cutoff independent for $x \gtrsim 2$.
When the first oscillation ends, the spectrum is essentially cut at $x \simeq 1$. However at later times, when the field relaxes around the minimum, the gradient energy is transferred to higher $x$ modes. This is also shown in Fig \[fig:deltaphiDef\], right panel. This is a slow process of which only the first stages are given by classical evolution. However, whereas this is important to understand the ultimate equilibrium state of the field theory, it is not relevant for the ‘cosmological’ question whether there is a bounce in the $\gamma \rightarrow 0$ limit. For this, the field displacement $\Delta \bar \chi$, shown in Fig \[fig:deltaphiDef\] (left panel), turns out to be a useful variable.
![\[fig:UVconv1\] The ‘homogeneous’ field displacement after the first oscillation, as a function of the lattice UV cutoff, for $ \lambda = 4 \cdot 10^{-3}$, $ \gamma = 10^{-3}$.](UVconvergence.pdf){width="8cm"}
\[eq:deltaPhiDef\] ![ \[fig:deltaphiDef\] The homogeneous component $ \bar{\chi}$ as a function of time (left), and the mode occupation numbers as a function of $x \equiv k / k _{T}$, at $t \simeq 1.6$ when the first oscillation is completed (dashed) and at $t =2$ (solid). Here $ \lambda /4 = 10 ^{-4}$, $ \gamma = 10 ^{-4}$.](deltaPhiDef.pdf "fig:"){width="7.7cm"}
![ \[fig:deltaphiDef\] The homogeneous component $ \bar{\chi}$ as a function of time (left), and the mode occupation numbers as a function of $x \equiv k / k _{T}$, at $t \simeq 1.6$ when the first oscillation is completed (dashed) and at $t =2$ (solid). Here $ \lambda /4 = 10 ^{-4}$, $ \gamma = 10 ^{-4}$.](occupationNumbers.pdf){width="7.7cm"}
Numerical results
-----------------
To evaluate the particle production and in particular its backreaction, we adopt as an indicator the field displacement $ \Delta \bar{ \phi} \equiv \bar{ \phi} ^{(1)} - \phi _{-}$ between the homogeneous field value after the first oscillation, $ \bar{ \phi} ^{(1)}$, and its initial value $ \phi _{-}$ (see Fig \[fig:deltaphiDef\]). By homogenous field we mean the average of the field value over the lattice sites. The growth of inhomogeneities can be quantified by evaluating the energy density $ \rho _{g}$ injected in spatial gradients after a single oscillation. These two quantities should naturally be compared to $ (\phi _{m} -\phi _{-})$ and $|V _{m}|$, which determine their upper bounds in the presence of the regularization.
To compare theories with different values of $ \lambda$ it is useful to work with the rescaled quantities $ \Delta \bar{ \chi} = \sqrt{ \lambda} \,\Delta \bar{ \phi}$ and $ \rho _{ \chi} = \lambda \rho _{g}$. The advantage of these variables is that the reference quantities $ (\chi _{m} -\chi _{-})$ and $V _{m,1}$ are $ \lambda$–independent.
Our study covers a wide range of parameters, namely from $ \lambda \simeq 10 ^{-1}$ to $\lambda \simeq 10 ^{-10}$, and from $ \gamma \simeq 10 ^{-1}$ to $ \gamma \simeq 10 ^{-6}$. The lower bounds are imposed by numerical limitations: small values of $\lambda$ require a higher numerical precision to resolve the fluctuations, whereas smaller values of $\gamma$ lead to a more violent and “tachyonic” first oscillation that requires a better discretization.
### AdS crunch limit: $\gamma$ dependence {#ads-crunch-limit-gamma-dependence .unnumbered}
![ Homogeneous field displacement (left) and amount of gradient energy density (right) after one oscillation, as functions of $ \gamma ^{-1}$. The black solid lines represent respectively $ (\chi _{m} -\chi _{-})$ and $V _{m,1}$. From top to bottom, $ \lambda /4 = 10 ^{-1}$, $10 ^{-2}$, $10 ^{-3}$, $10 ^{-4}$, $10 ^{-6}$. []{data-label="fig:conv"}](deltaPhiGamma.pdf){width="7.8cm"}
![ Homogeneous field displacement (left) and amount of gradient energy density (right) after one oscillation, as functions of $ \gamma ^{-1}$. The black solid lines represent respectively $ (\chi _{m} -\chi _{-})$ and $V _{m,1}$. From top to bottom, $ \lambda /4 = 10 ^{-1}$, $10 ^{-2}$, $10 ^{-3}$, $10 ^{-4}$, $10 ^{-6}$. []{data-label="fig:conv"}](rhoGradGamma.pdf){width="7.8cm"}
For large values of $\gamma$ near $1$, simulations have shown that a few oscillations occur before the homogeneous field reaches its equilibrium value around the true minimum of the potential (see Fig \[fig:meanPhi\]). At first the fluctuations are in the linear regime, in which their amplitude is essentially multiplied by a constant (which depends on the wavenumber) at each oscillation. However, since the growth of perturbations is exponential, at some point the backreaction becomes important, rapidly dampening the oscillation amplitude.
When one decreases $\gamma$, the potential becomes steeper and deeper. This means that more modes become tachyonic during the first stages of the evolution, and further that the low k modes grow faster. We find the energy converted in particles is *not* restored back in the homogeneous modes when the field rolls up again. As a consequence the displacement of the homogeneous field after one oscillation increases when $\gamma$ decreases.
Since for sufficiently small values of $\gamma$ the only relevant scale in the problem is set by $\phi _{m}$, one expects that $\Delta \bar{\phi} \propto \phi _{m}$ and $ \rho _{g} \propto V _{m}$. This is confirmed by the numerical simulations, as shown in Fig \[fig:conv\]. Hence $ \Delta \bar{ \phi}$ as well as $ \rho _{g}$ follow essentially the scaling that is imposed by the brick wall. The scaling limit is manifest for $ \Delta \bar{\phi} \gtrsim \phi _{-}$. In this regime one has $$\label{eq:gammaScaling}
\Delta \bar{\phi} \propto \gamma ^{- \frac{1}{2}}, \qquad \rho _{g} \propto \gamma ^{-2}$$ The proportionality constants in (\[eq:gammaScaling\]) depend on $\lambda$ and will be discussed below.
The subsequent evolution depends on the details of the model. For intermediate values of $\gamma$ it takes some time for $ \bar{ \phi}$ to stabilize around $ \phi _{m}$. This process involves a slow cascading effect whereby the energy is transferred from the tachyonic band to a wider band of modes extending up to wavenumbers of a few times $k _{T}$.
However for small values of $\gamma$ the evolution proceeds in a much more turbulent manner. In this regime the non-linearities are important and even at times $t \leq T/2$ we find patches with $ \phi \simeq \phi _{m}$ form. The inhomogeneities behave in a local manner till $t \sim T/2$, in line with the findings in [@Craps2007], but for $t \geq T/2$ the evolution proceeds via the expansion and collision of approximately homogeneous bubbles (see also [@Barbon2010]). This produces a violent backreaction on $ \bar{ \phi}$ which transfers energy rapidly to high k modes. Consequently, no further large oscillations in $ \bar{ \phi}$ occur. The bubble dynamics for $t \geq T/2$ is heavily dependent on the regularization, and therefore on the kind of boundary conditions we impose at infinity.
The simulations have also shown that, in the small $\gamma$ regime, a certain amount of tunneling takes place to the true minimum of the potential located at negative $ \phi$. In fact, for small $ \gamma$ the potential barrier is small compared to the scale of the true minimum of the potential, so the potential is essentially flat around $ \phi = 0$. It is therefore not surprising that local energy overdensities let bubbles where the field is negative form. We have checked carefully that this phenomenon is not due to numerical violations of energy conservation. As a matter of fact, this feature of the evolution does not seem crucial for our study. Indeed, an infinite potential barrier at $ \phi = 0$ does not significantly modify the scaling behavior of (\[eq:gammaScaling\]).
![ \[fig:lambdaDependence\]Homogeneous field displacement (left) and amount of gradient energy density injected (right) after one oscillation, as functions of $ \lambda$. From top to bottom, $ \gamma = 10^{-6},10^{-5},10^{-4},10^{-3},10^{-2},5 \cdot 10^{-2}$. The curves with the two largest values of $\gamma$ exhibit the linear regime for small $\lambda$. The straight dashed lines are, respectively, the $\sqrt{ \lambda}$ and constant fits to the linear analysis discussed in the text. For smaller values of $\gamma$ the backreaction is always large and one obtains the scaling given in Eq (\[eq:powerLaws\]).](deltaPhiLambda.pdf){width="7.8cm"}
![ \[fig:lambdaDependence\]Homogeneous field displacement (left) and amount of gradient energy density injected (right) after one oscillation, as functions of $ \lambda$. From top to bottom, $ \gamma = 10^{-6},10^{-5},10^{-4},10^{-3},10^{-2},5 \cdot 10^{-2}$. The curves with the two largest values of $\gamma$ exhibit the linear regime for small $\lambda$. The straight dashed lines are, respectively, the $\sqrt{ \lambda}$ and constant fits to the linear analysis discussed in the text. For smaller values of $\gamma$ the backreaction is always large and one obtains the scaling given in Eq (\[eq:powerLaws\]).](rhoGradLambda.pdf){width="7.8cm"}
### Non–linearities: $\lambda$ dependence {#nonlinearities-lambda-dependence .unnumbered}
For $\lambda \ll \gamma$ and fixed $ \gamma <1$, the fluctuations are small compared to $ \bar{ \phi}$ and behave linearly at least during the first oscillation[^6]. At the linear order, both the initial conditions and the evolution of the $ \phi$ inhomogeneities are $ \lambda$ - independent. Consequently we expect no $ \lambda$–dependence for $ \rho _{g} $ in this regime. One should also have $$\label{eq:linear}
\Delta \bar{ \phi} \simeq \frac{ \rho _{g} }{ V _{, \phi}( \phi _{-})} \propto \sqrt{ \lambda}.$$ By contrast, for larger values of $\lambda$ the first oscillation may already involve significant backreaction, with $ \Delta \bar{\phi} \gg \phi _{-}$. The homogeneous field displacement is then related to $ \rho _{g}$ as $$\label{eq:relation}
\rho _{g} \sim \lambda ( \Delta \bar{ \phi} ) ^4.$$ Numerical simulations have confirmed this qualitative behavior. In Fig \[fig:lambdaDependence\] we show that for $ \lambda \ll \gamma$ the gradient energy $ \rho _{g}$ becomes independent of $\lambda$ and $ \Delta \bar{ \phi}$ scales as $ \sqrt{ \lambda}$, as expected from the linear analysis (\[eq:linear\]). For larger ratios of $\lambda /\gamma$ both the homogeneous field displacement and the energy density in fluctuations follow a rather precise scaling with $ \lambda$ which appears to be universal for sufficiently small $ \gamma$. The equal spacing between the data sets in Fig \[fig:lambdaDependence\] corresponds, of course, to the scaling with $\gamma$ given in (\[eq:gammaScaling\]).
To summarize, in the scaling limit $ \gamma \rightarrow 0 ^{+}$ and well beyond the linear regime, the quantities of interest behave as $$\label{eq:powerLaws}
\Delta \bar \phi \propto \frac{ \lambda ^{ \alpha}}{ \gamma ^{ \frac{1}{2}}}, \qquad
\rho _{g} \propto \frac{ \lambda ^{ \beta}}{ \gamma ^{ 2}}$$ with $ \alpha \simeq \beta \simeq -0.33$, which agrees with (\[eq:relation\]) to a good accuracy. These exponents are non–trivial and the result of a highly non–linear evolution involving an interplay between tachyonic growth of individual modes and interactions.
The ‘cosmological’ limit corresponds to taking $ \gamma \rightarrow 0 ^{+}$ for *fixed* $ \lambda$. In this limit, it follows from (\[eq:powerLaws\]) that $\Delta \phi ^{(1)}$ and $\rho _{g}$ both diverge. Consequently, there is no meaningful first oscillation back to a finite value of the field, and therefore no evidence in our dual model for a transition in the bulk from a big crunch to a big bang.
### Potential renormalization {#potential-renormalization .unnumbered}
A one-loop computation shows that the coupling $ \lambda$ of the unstable $-\phi^4$ potential runs logarithmically. For large field values the Coleman-Weinberg potential is given by $$V _{r} \sim - \frac{ \lambda}{ 4 \log{ \frac{ \phi}{ M}}} \phi ^4$$ where $M$ is the renormalization scale. The logarithmic running effectively decreases the parametric divergence of the tachyonic scale with $ \gamma$. At leading order in the logarithmic corrections one finds $$k _{T} ^2 \propto \frac{ \gamma ^{-1}}{ \log ^2 \gamma}.$$ Modeling the self-adjoint extension again by the addition of a $\phi^6$ regularization term we obtain, as a toy model, $$\label{reglog}
V(\phi) = \frac{1}{2} \phi ^2 - \frac{\lambda}{2 \log \left( \frac{ \phi^2 + \phi _0^2}{M ^2} \right) } \phi ^4+ \frac{1}{32} \lambda^2 \gamma \phi^6$$ where $ \phi_0$ is added to produce an essentially constant $ \lambda$ for small values of $ \phi$.
Intuitively, one expects that the running of $\lambda$ leads to logarithmic corrections to the scalings given in (\[eq:powerLaws\]). A numerical study of (\[reglog\]) confirms this: the effect of the logarithmic running appears to add only finite $ \gamma$ corrections to the scalings. This means the energy in particles is still determined by the tachyonic scale, and therefore our conclusions remain unchanged.
Conclusion
==========
Using the AdS/CFT correspondence, the approach to a big crunch singularity in five-dimensional asymptotically anti-de Sitter spacetimes can be described in a dual field theory by an initially nearly homogeneous wave packet that rolls down an unbounded potential direction of a double trace deformation $-f {\cal O}^2/2$ of ${\cal N}=4$ Super-Yang-Mills (SYM) theory on $ \mathbb{R} \times S ^{3} $, with ${\cal O}$ a trace operator quadratic in the adjoint Higgs scalars. The big crunch singularity in the bulk occurs when the boundary scalar runs to infinity in finite time.
Concentrating on the steepest negative direction of the potential, we have numerically solved for the dual evolution. Consistent quantum evolution requires that one considers a self-adjoint extension of the system. Since this essentially acts as a brick wall at infinity, we have modeled this by adding a steep regularization term to the unbounded potential. This introduces a UV regulator $\gamma$, and the cosmological limit in the bulk corresponds to taking $\gamma \rightarrow 0^{+}$.
While the wave packet rolls down the potential, the negative effective mass of the potential amplifies quantum perturbations which therefore behave classically, at least during the first stages of the evolution. We find that for small $\gamma$ this converts most of the potential energy of the initial configuration into gradient energy during the first oscillation of the wave packet. In particular, for small $\gamma$ the system tends to a universal ‘scaling limit’ in which the mean field displacement $\Delta \bar{\phi}$ and the gradient energy $\rho _{g}$ after one oscillation scale as (an appropriate power of) the UV regulator itself. This means that, in the cosmological $\gamma \rightarrow 0^{+}$ limit, the boundary scalar does not return close to its original value after the first oscillation. Instead the gradient energy diverges, indicating there is no transition in the bulk from a big crunch to a big bang in the models considered here. In the boundary theory, our simulations show a sudden phase of bubble formation at large field, when some regions of space bounce back while others are still going down.
A different way to implement self-adjoint boundary conditions in field theory, based on an extension of the domain of the scalars involved, has been explored in [@Craps2007] and more recently in [@Turok2010]. There is some evidence [@Hertog] this allows for a smoother dynamical evolution across the singularity and in particular that this avoids the breakdown of the local behavior of the perturbations near the singularity. Hence a violent phase of bubble dynamics like what we observed in our simulations may not occur with boundary conditions of this kind. We note, however, that in the models we have considered we find that, in the scaling regime, perturbations are in the non-linear regime already while the wave packet rolls down. It seems plausible that this is largely independent of the nature of the boundary conditions at infinity. One will therefore have to take in account non-linearities in order to determine conclusively whether the boundary scalar can return close to its original value after the first oscillation with different boundary conditions at infinity.
A possible exception are models where the coupling $\lambda$ governing the potential instability is significantly smaller than the UV regularization scale $\gamma$ for large field values. In this case our simulations indicate that the perturbations remain in the linear regime during the entire first oscillation. The backreaction of particle production can then be limited, at least in a certain range of parameter values. In the models discussed here, we have seen it is not possible to maintain $\lambda < \gamma$ in the cosmological regime involving small $\gamma$. Indeed, although the coupling of the unstable double trace deformation runs logarithmically and is asymptotically free, we find that in order to limit the backreaction $\lambda$ would have to go to zero in the UV faster than this. Whether this can be realized in a quantum treatment of the three-dimensional models discussed in [@Craps2009] remains to be seen.
Even though in the regularized models, the expectation value of ${\cal O}$ rapidly stabilizes around the global minimum of the potential, the spectrum of produced particles at early times is essentially limited to the low tachyonic band of momenta. Subsequently the system thermalizes, due to the self-interaction of the field. However this takes place on much longer time scales and cannot be described by the classical simulations discussed here. The approach to a thermal final state in the boundary theory nevertheless provides an interesting testing ground of the AdS/CFT correspondence in the presence of multitrace deformations. Some aspects of this will be discussed elsewhere [@Battarra10b].
Acknowledgments {#acknowledgments .unnumbered}
===============
TH thanks Ben Craps and Neil Turok for stimulating discussions and collaboration on related problems. We also thank Ben Craps for helpful comments on a draft of this paper. This work is supported in part by the ANR (France) under grant ANR-09-BLAN-0157.
[^1]: [email protected]
[^2]: [email protected]
[^3]: It follows from Fig 14 (b) in [@Craps2007] that for sufficiently small values of $f$ the supergravity analysis predicts $\b \sim 1/\lambda$. The field theory solution [(\[clsol\])]{} reproduces this scaling.
[^4]: From here onwards we set the AdS radius equal to one, so that the dual field theory lives on $S^3\times R$ where the sphere also has unit radius.
[^5]: We have also verified that the energy density correlation length (see [@Frolov2008]) is much smaller than the IR cutoff.
[^6]: We emphasize that, since the tachyonic instability is an extremely efficient mechanism to produce particles, this regime typically corresponds to values of $ \lambda$ that are very small compared to (an appropriate power of) $ \gamma$. For this reason the regime in which the perturbations behave linearly for the entire duration of the first oscillation has turned out to be numerically inaccessible for $ \gamma \lesssim 10^{-3}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'An optimal control problem with a time-parameter is considered. The functional to be optimized includes the maximum over time-horizon reached by a function of the state variable, and so an $L^\infty$-term. In addition to the classical control function, the time at which this maximum is reached is considered as a free parameter. The problem couples the behavior of the state and the control, with this time-parameter. A change of variable is introduced to derive first and second-order optimality conditions. This allows the implementation of a Newton method. Numerical simulations are developed, for selected ordinary differential equations and a partial differential equation, which illustrate the influence of the additional parameter and the original motivation.'
author:
- 'Sébastien Court[^1]'
- 'Karl Kunisch[^2]'
- Laurent Pfeiffer
bibliography:
- 'Max-max\_ref.bib'
title: 'Optimal Control for a Class of Infinite Dimensional Systems Involving an $L^\infty$-term in the Cost Functional'
---
[**Keywords: Optimality conditions, PDE-constrained optimization, Transversality conditions, hybrid optimal control problem.**]{}\
\
[**AMS subject classifications (2010): 49K20, 49K15, 49K21, 93C30, 90C46.**]{}
Introduction
============
We consider optimal control problems with the objective of optimizing a function of the state variable at some time $\tau$ during its time evolution. The time $\tau$ itself is free to move within the time-span $(0,T)$ of the whole experiment. As a very first thought, one might think of the prototype of a new vehicle and ask what is the highest reachable speed on the race track. Our original motivation for the class of problems under consideration stems from the domain of cardiac electro-physiology. The problem of fibrillation of a part of the heart can be mechanically tackled with the use of electric shocks acting on this muscle, leading to the re-oxygenation of the ill area, and by this means forcing this area to recover a healthy electric activity. The efficiency of the is known to be related to the maximum reached over time by the derivative of a pressure in the heart. This quantity can be mathematically formulated as the maximum taken by a function of the state variable of the problem. While we continue to work on this challenging problem, we focus in this article on the methodology enabling to derive optimality conditions for such a problem, for a simpler class of partial differential equations.
The optimal control problem that we shall investigate can be formulated as follows: $$\max_{\tau \in (0,T),\, u \in {\mathrm{L}}^2(0,T;U)} \int_0^T \ell(y(t),u(t)) {\mathrm{d}}t + \phi_1(y(\tau)) + \phi_2(y(T)),
\quad \text{subject to: } \quad
\dot{y}= f(y,u), \quad y(0)=y_0,$$ where the state variable $y$ is the solution of an evolution equation controlled by $u$. In this problem, $\ell$ denotes the cost functional, $\phi_1$ the functional we want to maximize at some time $\tau$, and $\phi_2$ is the terminal cost. The analytical framework will be specified later. The specificity of this kind of problem lies in the fact that a time parameter, namely $\tau$, can be optimized. Not only do we maximize $\phi_1\circ y$ with the help of the control $u$, but we also optimize the time $\tau$ for which the maximum is reached. Note that, when $\phi_1$ is nonnegative, the problem can be equivalently formulated in the following way: $$\max_{u \in {\mathrm{L}}^2(0,T;U)} \int_0^T \ell(y(t),u(t)) {\mathrm{d}}t + \| \phi_1(y(\cdot)) \|_{{\mathrm{L}}^{\infty}(0,T;Y)} + \phi_2(y(T)), \quad
\text{subject to: } \quad \dot{y}= f(y,u), \quad y(0)= y_0.$$ In this fashion, the cost function incorporates the ${\mathrm{L}}^\infty$-norm of a given function of the state variable.
$$\begin{aligned}
\left\{\begin{array} {l}
\displaystyle \max_{
\begin{subarray}{c} 0= \tau_0 < \tau_1 < ... < \tau_K= T \\ u \in {\mathrm{L}}^2(0,T;U)
\end{subarray}
} \ \sum_{k=1}^K \ \int_{\tau_{k-1}}^{\tau_k} \ell_k(y(t),u(t)) {\mathrm{d}}t + \phi_k(y(\tau_k)), \\
\text{subject to: } \quad
\dot{y}(t)= f_k(y(t),u(t)), \text{ for a.\,e.\,} t \in (\tau_{k-1},\tau_k), \quad
y(0)= y_0.
\end{array} \right. \label{eqHybrid}\end{aligned}$$ The above problem incorporates time parameters (or switching times) $\tau_1$,...,$\tau_{K-1}$ which can be optimized. The term [*hybrid*]{} refers here to the fact that at the switching times $\tau_k$, the system move from a given regime (described here by the dynamics $f_k$) to another one (described by the dynamics $f_{k+1}$). At the switching times, the integral cost changes and a function of the state is also incorporated into the cost function. There are many ways to generalize problem (see e.g. [@GP05]). For example, one can consider a formulation of the problem for which the dynamics $f_k$ (used during the interval $(\tau_{k-1},\tau_k)$) can be itself chosen into a finite set of functions. Such generalizations are beyond the scope of the paper.
In the optimal control literature, many problems include time parameters that have to be optimized. Among them, time-optimal control problems have probably attracted the most the attention. These problems basically consist in minimizing the time needed to reach a given target. We refer to the early reference [@HL69] on this topic. Time-optimal control problems have been studied for various models: See for example [@KW13] for the wave equation, [@KR15; @KPR16] for the monodomain system, [@Bar97] for the Navier-Stokes equations, [@MRT12] for the heat equation. We also mention the time-crisis management problem studied in [@BR16]; For such a problem, the time spent out of a certain closed domain must be minimized.
For our problem, the first-order optimality conditions consist of a weak maximum principle for the control control, and a transversality condition for the optimality of the time parameter. The derivation of the transversality condition is difficult, in so far as the Lagrangian of the problem is not differentiable with respect to the time-parameters. In e.g. [@IK10; @RZ99; @RZ00], a change of variables (in time) is performed to derive the transversality condition in the case of a time-optimal control problem. This is the approach that we adopt here. It also enables us to derive second-order optimality conditions, as in [@KPR16] and [@LiGaoWang] for finite-dimensional hybrid control problems. In this last reference, a numerical test based on a Riccati equation is provided to check the sufficient second-order conditions. In [@GP05], specific needle variations are designed, for hybrid systems. Other approaches can also be considered. In [@BR16], the time-crisis management problem is regularized and optimality conditions are derived by $\Gamma$-convergence.
One novelty of our work lies in the fact that this class of hybrid problems is addressed in infinite dimensions, theoretically as well as for numerical applications (with an example dealing with a partial differential equation). Another novelty is the derivation of second-order optimality conditions, which is – as far as we know – addressed in few articles (except for instance [@LiGaoWang]).
The paper is organized as follows: In section \[sec2\] the functional framework is specified, and the problem is transformed with a change of variables. Section \[sec3\] is devoted to the derivation of first and second-order optimality conditions. The abstract framework proposed here is shown to be satisfied, as an example, by the Navier-Stokes equations in dimension 2. The issue of numerical resolution is addressed in section \[sec4\]: We use the theoretical expressions of the optimality condition in order to perform an algorithm which solves the problem, mixing Barzilai-Borwein gradient steps with newton’s steps. As illustrations, we consider two classical examples of ordinary differential systems, and one example dealing with the Burgers equation.
#### Notation.
First and second-order partial derivatives are denoted with the use of indexes. When there is no , the time variable is sometimes omitted, or written only once (e.g. $H(y,u,p)(s)$, instead of $H(y(s),u(s),p(s))$. The first and second-order derivatives with respect to all variables (except the adjoint state, in the case of the Hamiltonian) are denoted by $D$ and $D^2$, respectively. When a function $h$ is left- and right-continuous at a given time $t$, the left- and right-limits are denoted by $h(t^-)$ and $h(t^+)$, respectively, and the jump is denoted by $[h]_t$.
Formulation of the problem {#sec2}
==========================
Setting {#secset}
-------
Let $X$ be a real Hilbert space, and $Y$ be a reflexive Banach space forming with $X$ a Gelfand triple $Y \subseteq X \equiv X' \subseteq Y'$, with $Y$ densely contained in $X$. We denote $Z = Y'$, so that $Z'$ is isomorphic to $Y$. Let $U$ be the Banach space of controls. Further let $\ell:X\times U \rightarrow {\mathbb{R}}$ and $f:Y\times U \rightarrow Y'$ be two twice Fr[é]{}chet-differentiable mappings. For $\tilde{u} \in {\mathrm{L}}^2(0,T;U)$, the state is governed by the autonomous control system $$\begin{aligned}
\label{mainsys0}
\left\{ \begin{array} {l}
\dot{\tilde{y}} = f(\tilde{y},\tilde{u}) \quad \text{on } (0,T),\\
\tilde{y}(0) = y_0,
\end{array} \right. \end{aligned}$$ where the initial condition $y_0 \in X$ is given. We assume that for all $y_0 \in X$ and $\tilde{u}\in {\mathrm{L}}^2(0,T;U)$ this system has a unique solution $\tilde{y}$ in the space: $$\begin{aligned}
W(0,T;Y) := {\mathrm{L}}^2(0,T;Y) \cap {\mathrm{W}}^{1,2}(0,T;Y').\end{aligned}$$ Recall the continuous embedding $W(0,T;Y) \hookrightarrow \mathcal{C}([0,T];X)$, so that in particular the initial condition makes sense in $X$. Let $\phi_1$ and $\phi_2: X \rightarrow {\mathbb{R}}$ be two twice continuously Fréchet-differentiable mappings. We consider the following problem: $$\begin{aligned}
& & \left\{ \begin{array} {l}
\displaystyle \max_{\tau\in(0,T), \ \tilde{u} \in {\mathrm{L}}^2(0,T;U)} \ \int_0^T \ell(\tilde{y}(t),\tilde{u}(t)) {\mathrm{d}}t + \phi_1(\tilde{y}(\tau)) + \phi_2(\tilde{y}(T)) \vspace{2mm} \\
\displaystyle \text{subject to: }\quad
\dot{\tilde{y}} = f(\tilde{y},\tilde{u}) \text{ in } Y',\quad \tilde{y}(0) = y_0 \text{ in } X.
\end{array} \right. \label{pbmain}\end{aligned}$$
Transformation of the problem
-----------------------------
The formulation of problem does not enable us to derive optimality conditions. Indeed, the control is not continuous at the optimal time $\tau$, in general. Therefore, the trajectory is not differentiable at $\tau$ and the cost is not differentiable with respect to $\tau$. This difficulty can be by introducing the following change of variables, $\pi(\cdot,\tau) : [0,2] \rightarrow [0,T]$, for all $\tau \in (0,T)$, $$\begin{aligned}
\pi(s,\tau) & = &
\left\{ \begin{array} {ll}
\tau s & \text{if } s\in [0,1], \\
(T-\tau) s + 2\tau -T & \text{if } s\in [1, 2].
\end{array} \right.\end{aligned}$$ Observe that $\pi(1,\tau)= \tau$. For future reference we introduce the time-derivative $\dot{\pi}$ of $\pi$ (with respect to $s$), as well its the partial derivative $\dot{\pi}_\tau$ with respect to $\tau$: $$\dot{\pi}(s,\tau) =
\left\{ \begin{array} {ll}
\tau & \text{if } s\in [0,1], \\
(T-\tau) & \text{if } s\in [1, 2],
\end{array} \right.
\qquad
\dot{\pi}_\tau(s,\tau) =
\left\{ \begin{array} {ll}
1 & \text{if } s\in [0,1], \\
-1 & \text{if } s\in [1, 2].
\end{array} \right.$$ Observe that $\dot{\pi}_\tau$ is actually independent of $\tau$. To the , we will simply write $\dot{\pi}_\tau(s)$. Given $\pi(\cdot,\tau)$ we introduce the following change of unknowns $$\begin{aligned}
y:s \mapsto \tilde{y} \circ \pi(s,\tau), \quad u:s \mapsto \tilde{u} \circ \pi(s,\tau),
\quad s \in [0,2]. \label{eqChangeOfVar}\end{aligned}$$ Then, for $(u,\tau) \in {\mathrm{L}}^2(0,2;U) \times (0,T)$, we are lead to consider the following system $$\begin{aligned}
\label{mainsys}
\left\{ \begin{array} {l}
\dot{y} = \dot{\pi}(\cdot,\tau) f(y,u) \quad \text{on } (0,2),\\
y(0) = y_0,
\end{array} \right.\end{aligned}$$ and the following reformulated optimal control problem: $$\begin{aligned}
& & \left\{ \begin{array} {l}
\displaystyle \max_{\tau\in(0,T), \ u \in {\mathrm{L}}^2(0,2;U)} \ J(u,\tau):= \int_0^2 \dot{\pi}(s,\tau) \ell(y(s),u(s)) {\mathrm{d}}s + \phi_1(y(1)) + \phi_2(y(2)) \vspace{2mm} \\
\displaystyle \text{subject to: }\quad
\dot{y} = \dot{\pi}(\cdot,\tau) f(y,u) \text{ in } Y',\quad y(0) = y_0 \text{ in } X.
\end{array} \right. \label{pbmain2}\end{aligned}$$
The equivalence of problems and is straightforward because, on one hand, the time-derivative of $\tilde{y} \circ \pi(\cdot,\tau)$ is expressed with the chain rule as $\dot{\pi}(\cdot,\tau) \dot{\tilde{y}} \circ \pi(\cdot,\tau) = \dot{\pi}(\cdot,\tau)f(\tilde{y}\circ \pi(\cdot,\tau), \tilde{u}\circ \pi(\cdot,\tau))$. On the other hand, the integral on $(0,T)$ can be split on $(0,\tau) \cup (\tau ,T)$, and the change of variables $\pi(\cdot,\tau)$ is used for transforming the integrals on $(0,\tau)$ and $(\tau,T)$ into the integrals on $(0,1)$ and $(1,2)$, respectively. To sum up, if $\tilde{u} \in {\mathrm{L}}^2(0,T;U)$, $\tau \in \text{\textcolor{black}{$(0,T)$}}$ and $u= \tilde{u}\circ \pi(\cdot,\tau)$, the pair $(\tilde{u},\tau)$ is an optimal solution of the original problem if and only if the pair $(u,\tau)$ is an optimal solution of the reformulated problem .
First and second-order optimality conditions {#sec3}
============================================
In this section we derive necessary optimality conditions for the reformulated problem . Throughout this section, $\bar{u} \in {\mathrm{L}}^2(0,2;U)$ is a fixed value of the control and $\bar{\tau} \in (0,T)$ a fixed value of the variable $\tau$. Note that we do not follow up the special cases which arise for $\bar{\tau}=0$ and $\bar{\tau}=T$. We denote by $\bar{y} \in W(0,2;Y)$ the corresponding state variable, which is the solution of for $(u,\tau)=(\bar{u}, \bar{\tau})$.
The approach we use for deriving optimality conditions is classical, as described in [@HinzeKK01; @Hinze], for instance. However, due to the time transformation and the additional optimization variable, special attention is required. Considering the state equation as a constraint of the optimization problem , our approach mainly consists in computing the first- and second-order derivatives of the associated Lagrangian.
Linearization of the system
---------------------------
We introduce the control-to-state mapping , where $S(u,\tau)$ is the solution of . We have denoted $\bar{y} = S(\bar{u},\bar{\tau})$. The differentiability properties for this mapping $S$ derive from assumptions $\mathbf{A1'}$–$\mathbf{A3'}$.
\[lemma\_diffSys\] The mapping $S$ is twice continuously Fréchet-differentiable on ${\mathrm{L}}^\infty(0,2;U) \times (0,T)$. For $v \in {\mathrm{L}}^2(0,2;U)$, the derivatives $z=S_u(\bar{u},\bar{\tau}).v $ and $w= S_\tau(\bar{u},\bar{\tau})$ are the the following systems: $$\label{syssensuv}
\left\{ \begin{array} {l}
\dot{z} = \dot{\pi}(\cdot,\bar{\tau})f_y(\bar{y},\bar{u}).z + \dot{\pi}f_u(\bar{y},\bar{u}).v \quad \text{on } (0,2), \\
z(0) = 0,
\end{array} \right.
\quad
\left\{ \begin{array} {l}
\dot{w} = \dot{\pi}(\cdot,\bar{\tau})f_y(\bar{y},\bar{u}).w + \dot{\pi}_{\tau}f(\bar{y},\bar{u}) \quad \text{on } (0,2), \\
w(0) = 0.
\end{array} \right.$$
Consider the mapping $$\begin{aligned}
\begin{array} {rccl}
e: & W(0,2;Y) \times {\mathrm{L}}^2(0,2;U) \times (0,T) & \rightarrow & {\mathrm{L}}^2(0,2;Y')\times X, \\
& (y,u,\tau) & \mapsto & (\dot{y} - \dot{\pi}(\cdot,\tau)f(y,u) , y(0) - y_0).
\end{array}\end{aligned}$$ Since we have the identity $e(S(u,\tau),u,\tau) = 0$, assumptions $\mathbf{A1'}$–$\mathbf{A3'}$ enable us to apply the implicit function theorem. In fact, surjectivity of the mapping $e_y(y,u,\tau)$ is a consequence of $\mathbf{A3'}$, and the required smoothness conditions follow from $\mathbf{A1'}$. The result then follows.
Comments on the functional framework
------------------------------------
### Example
The variational framework $W(0,T;Y)$ is applicable, for instance, to the Burgers system (see section \[secnumPDE\]), or to the unsteady Navier-Stokes system (see [@HinzeKK01; @HinzeKK]). Assumptions $\mathbf{A1}$–$\mathbf{A3}$ can be easily verified for the Burgers system, if we refer to the analysis of e.g. [@Volkwein].\
Let us detail the verification for the Navier-Stokes system: $$\begin{aligned}
\dot{y} + (y\cdot \nabla) y - \nu \Delta y + \nabla p = Bu & \quad & \text{in } \Omega \times (0,T), \\
\operatorname{div}y = 0 & \quad & \text{in } \Omega \times (0,T), \\
y = 0 & \quad & \text{on } {\partial}\Omega\times (0,T), \\
y(0) = y_0 & \quad & \text{in } \Omega .
$$ In the system above, $\Omega$ is a bounded domain of ${\mathbb{R}}^2$, with smooth boundary, and $B\in \mathcal{L}(U,X)$ is a linear operator, for instance $B = \mathds{1}_{\omega}$ where $\omega \subset \Omega$, and $U = [L^2(\omega)]^2$. We consider the solenoidal spaces $$\begin{aligned}
Y = \left\{ y \in [{\mathrm{H}}^1_0(\Omega)]^2\mid \ \operatorname{div}y = 0\right\},
\quad X = \left\{ y \in [{\mathrm{L}}^2(\Omega)]^2\mid \ \operatorname{div}y = 0\right\},
\quad Y' \supset [{\mathrm{H}}^{-1}(\Omega)]^2, \end{aligned}$$ and $y_0 \in X$. In the variational formulation given below, the pressure $p$ disappears: $$\begin{aligned}
& & \text{Find $y \in W(0,T;Y)$ such that for all $\varphi \in {\mathrm{L}}^2(0,T;Y)$ we have almost everywhere in $(0,T)$:} \\
& & \left\{\begin{array} {l}
\langle \dot{y} + (y\cdot \nabla)y; \varphi \rangle_{Y';Y}
+ \nu \langle \nabla y ; \nabla \varphi \rangle_{[{\mathrm{L}}^2(\Omega)]^2} = \langle Bu ; \varphi \rangle_{Y';Y},\\
\langle y(0);\varphi\rangle_X = \langle y_0;\varphi\rangle_X.
\end{array} \right.\end{aligned}$$ For details we refer to [@Temam]. Here we have $f(y,u) = \nu \Delta y - (y\cdot \nabla) y + Bu$. In order to verify assumption $\mathbf{A1}$, a short consideration shows that first and second order directional differentiability and continuity of the derivatives (and hence Fréchet differentiability) of $f$ will follow from the continuity of the Oseen–type term $(z_1, z_2, z_3) \mapsto (z_1 \cdot \nabla)z_2 + (z_3 \cdot \nabla)z_1$ from $[W(0,T;Y)]^3$ to ${\mathrm{L}}^2(0,T;Y')$, which will be given below. Since $\operatorname{div}z_1 = \operatorname{div}z_2 = \operatorname{div}z_3 = 0$, we have $$\begin{aligned}
(z_1 \cdot \nabla)z_2 + (z_3 \cdot \nabla)z_1 & = & \operatorname{div}( z_2 \otimes z_1 + z_1 \otimes z_3), \\
\| (z_1 \cdot \nabla)z_2 + (z_3 \cdot \nabla)z_1 \|_{Y'} & \leq &
\| (z_1 \cdot \nabla)z_2 + (z_3 \cdot \nabla)z_1 \|_{[{\mathrm{H}}^{-1}(\Omega)]^2} \\
& \leq & \| z_1 \otimes z_2 + z_3 \otimes z_1 \|_{[{\mathrm{L}}^{2}(\Omega)]^2}\leq
\| z_1 \otimes z_2\|_{[{\mathrm{L}}^{2}(\Omega)]^4} + \|z_3 \otimes z_1 \|_{[{\mathrm{L}}^{2}(\Omega)]^4}.\end{aligned}$$ In view of the symmetric role played by $z_1$, $z_2$ and $z_3$, we only estimate the first term. From [@Grubb] (Appendix B, Proposition B.1), in dimension $2$ there exists a constant $C>0$ independent of $z_1$ and $z_2$ such that $$\begin{aligned}
\| z_1 \otimes z_2\|_{[{\mathrm{L}}^{2}(\Omega)]^4} & \leq & C\| z_1\|_{[{\mathrm{H}}^{1/2}(\Omega)]^2}
\| z_2\|_{[{\mathrm{H}}^{1/2}(\Omega)]^2}.\end{aligned}$$ By interpolation we deduce $$\begin{aligned}
\| z_1 \otimes z_2\|_{[{\mathrm{L}}^{2}(\Omega)]^4} & \leq &
C\| z_1\|^{1/2}_{[{\mathrm{L}}^{2}(\Omega)]^2} \| z_1\|^{1/2}_{[{\mathrm{H}}^{1}(\Omega)]^2}
\| z_2\|^{1/2}_{[{\mathrm{L}}^{2}(\Omega)]^2}\| z_2\|^{1/2}_{[{\mathrm{H}}^{1}(\Omega)]^2}.\end{aligned}$$ By integrating in time, this yields $$\begin{aligned}
\| z_1 \otimes z_2\|_{{\mathrm{L}}^2(0,T;[{\mathrm{L}}^{2}(\Omega)]^4)} & \leq &
C\| z_1\|^{1/2}_{{\mathrm{L}}^{\infty}(0,T;X)}\| z_2\|^{1/2}_{{\mathrm{L}}^{\infty}(0,T;X)}
\| z_1\|^{1/2}_{{\mathrm{L}}^{2}(0,T;Y)}
\| z_2\|^{1/2}_{{\mathrm{L}}^{2}(0,T;Y)} \\
& \leq & \tilde{C}\| z_1\|_{W(0,T;Y)}\| z_2\|_{W(0,T;Y)},\end{aligned}$$ for some constant $\tilde{C} >0$. Thus, assumption $\mathbf{A1}$ is satisfied. Assumptions $\mathbf{A2}$ and $\mathbf{A3}$ are due to [@HinzeKK01] (Proposition 2.1 page 928, and Proposition 2.4 page 929, respectively).
### Additional regularity
In other partial differential equations examples, the function space framework $W(0,T;Y)$ can be too restrictive. This is due to the fact that, in the linearized state equation given in assumption $\mathbf{A3}$, the term $f_y(\bar{y},\bar{u})$ appears as a coefficient. Hence the solution of this linearized system may not be well-defined, unless $f$ satisfies appropriate growth bounds. For this purpose an alternative functional framework can be appropriate. For instance, let $\hat{Y} \subset Y \subset X$ be three Hilbert spaces endowed with a chain of continuous embeddings. Duality is understood with respect to $X \equiv X'$. We set $$\begin{aligned}
\hat{W}(0,T;\hat{Y}) & = & \{ y\in {\mathrm{L}}^2(0,T;\hat{Y}) \mid \ \dot{y} \in {\mathrm{L}}^2(0,T;X) \}.
\end{aligned}$$ We still have $\hat{W}(0,T;\hat{Y}) \hookrightarrow C([0,T];X)$, but in general we do not have $\hat{W}(0,T;\hat{Y}) \hookrightarrow C([0,T];Y)$, except for instance when we have the interpolation $Y = [X;\hat{Y}]_{1/2}$. When we do not have this embedding, we can add $y\in \mathcal{C}([0,T];Y)$ into the definition of $\hat{W}(0,T;\hat{Y})$ above. In applications, for a given smooth domain $\Omega$, we may think of $X = {\mathrm{L}}^2(\Omega)$, and $Y$, $\hat{Y}$ subspaces of ${\mathrm{H}}^1(\Omega)$, ${\mathrm{H}}^2(\Omega)$, respectively. The mapping $e$ introduced in the proof of Lemma \[lemma\_diffSys\] is then modified to be $$\begin{aligned}
\begin{array} {rccl}
e: & \hat{W}(0,2;\hat{Y}) \times {\mathrm{L}}^2(0,2;U) \times {\mathbb{R}}& \rightarrow & {\mathrm{L}}^2(0,2;X)\times Y.
\end{array}
\end{aligned}$$ The assumptions utilized in this proof are implied by $\mathbf{A1}$–$\mathbf{A3}$. In order to transpose the assumptions $\mathbf{A1}$–$\mathbf{A3}$ and to derive differentiability in the framework considered here, we introduce the mapping $\hat{f}:\hat{Y}\times U \rightarrow X$ associated with $f$ as a restriction. We would assume:
$\mathbf{B1}$
: The mapping $\hat{f}: W(0,T;\hat{Y}) \times {\mathrm{L}}^2(0,T;U) \times {\mathbb{R}}\rightarrow {\mathrm{L}}^2(0,T;X)$ is twice continuously Fréchet-differentiable.
$\mathbf{B2}$
: For all $(u,\tau) \in {\mathrm{L}}^2(0,T;U)\times {\mathbb{R}}$, system admits a unique solution in $\hat{W}(0,T;\hat{Y})$.
$\mathbf{B3}$
: For all $\xi \in {\mathrm{L}}^2(0,T;X)$, $z_0 \in Y$, there exists a unique $z \in \hat{W}(0,T;\hat{Y})$ solution to the following system $$\begin{aligned}
\left\{ \begin{array} {l}
\dot{z} = \hat{f}_y(\bar{y},\bar{u}).z + \xi \quad \text{on } (0,T), \\
z(0) = z_0.
\end{array} \right.
\end{aligned}$$
This framework, where strong regularity is considered, is needed when the nonlinearity of the mapping $f$ cannot be handled in the context of weak solutions. For instance, polynomial nonlinearities, as those occurring in the Schlögel and FitzHugh-Nagumo systems (see [@CRT]), require the notion of strong solutions. In the context of parabolic equations, the time-dependent operator $\hat{f}_y(\bar{y},\bar{u})$ can be studied with the approaches of [@Bardos] or [@Troltzsch] for instance.\
For a sake of being specific, in the rest of the paper we will keep the framework of the function space $W(0,T;Y)$, with assumptions $\mathbf{A1}$–$\mathbf{A3}$. Analogous results as those obtained for $W(0,T;Y)$ also hold for $\hat{W}(0,T;\hat{Y})$.
Vectorial formalism
-------------------
We define the functional space $\mathcal{Y}$ and its dual space as follows: $$\begin{aligned}
\mathcal{Y}= X \times X \times {\mathrm{L}}^2(0,2;Y), & &
\mathcal{Y}'= X \times X \times {\mathrm{L}}^2(0,2;Y').\end{aligned}$$ Next we introduce the mapping $\mathbf{S}$ as follows: $$\begin{aligned}
\mathbf{S}: (u,\tau) \in {\mathrm{L}}^2(0,2;U) \times (0,T) & \mapsto & (S(u,\tau)(1), S(u,\tau)(2), S(u,\tau)) \in \mathcal{Y}.\end{aligned}$$ As a consequence of Lemma \[lemma\_diffSys\], the mapping $\mathbf{S}$ is twice differentiable. Its first-order derivatives are given by: $$\begin{aligned}
\label{eqDerivativesS}
\mathbf{S}_u(u,\tau)= \big(S_u(u,\tau)(1), S_u(u,\tau)(2), S_u(u,\tau) \big), & &
\mathbf{S}_\tau(u,\tau)= \big(S_\tau(u,\tau)(1), S_\tau(u,\tau)(2), S_\tau(u,\tau) \big).\end{aligned}$$ Let us define the operator $\mathcal{K} \in \mathcal{L}\left({\mathrm{L}}^2(0,2;Z);\mathcal{Y} \right)$ by $$\begin{aligned}
\begin{array} {rrcl}
\mathcal{K} : &
\xi & \mapsto & (z(1),z(2),z)
\end{array} \end{aligned}$$ where $z \in W(0,2;Y)$ is defined – in virtue of assumption $\mathbf{A3}$ – as the solution of $$\begin{aligned}
\left\{ \begin{array} {l}
\dot{z} = \dot{\pi}(\cdot,\bar{\tau})f_y(\bar{y},\bar{u}).z + \xi \quad \text{on } (0,2), \\
z(0) = 0.
\end{array} \right. \label{syslemmaS}\end{aligned}$$
\[lemma-adjS\] The adjoint $\mathcal{K}^\ast \in \mathcal{L} \left( \mathcal{Y}' ; {\mathrm{L}}^2(0,2;Z')\right)$ of $\mathcal{K}$ is given by $\mathcal{K}^{\ast}(a,b,w) = q$, where $q$ is the solution of $$\begin{aligned}
\left\{ \begin{array} {l}
-\dot{q} = \dot{\pi}(\cdot,\bar{\tau})f_y(\bar{y},\bar{u})^\ast.q + w \quad \text{on } (0,1)\cup (1,2), \\
q(2) = b,\\
q(1^+)-q(1^-) + a = 0.
\end{array} \right. \label{syslemmaStar}
\end{aligned}$$
Let $\xi \in {\mathrm{L}}^2(0,2;Z)$ and let $z$ be the solution of system corresponding to $\xi$. Let $(a,b,w) \in \mathcal{Y}'$ and denote by $q$ the solution of system corresponding to $(a,b,w)$. We calculate by integration by parts $$\begin{aligned}
\langle (a,b,w);\mathcal{K}(\xi)\rangle_{\mathcal{Y}';\mathcal{Y}} & = & \langle w;z \rangle_{{\mathrm{L}}^2(0,2;Y');{\mathrm{L}}^2(0,2;Y)}
+ \langle a;z(1)\rangle_{X} + \langle b;z(2)\rangle_{X} \\
& = & \int_0^{1} \langle -\dot{q} -\dot{\pi} f_y^\ast(\bar{y},\bar{u}).q; z\rangle_{Y';Y}{\mathrm{d}}s + \int_{1}^2 \langle -\dot{q} -\dot{\pi} f_y^\ast(\bar{y},\bar{u}).q; z\rangle_{Y';Y}{\mathrm{d}}s \\
& & + \, \langle a;z(1)\rangle_{X} + \langle b;z(2)\rangle_{X}\\
& = & \int_0^{1} \langle q;\dot{z}-\dot{\pi} f_y(\bar{y},\bar{u}).z \rangle_{Z';Z}{\mathrm{d}}s -\langle q(1^-);z(1)\rangle_X + \int_{1}^2 \langle q;\dot{z}-\dot{\pi} f_y(\bar{y},\bar{u}).z \rangle_{Z';Z}{\mathrm{d}}s\\
& & + \, \langle q(1^+);z(1)\rangle_X - \langle q(2);z(2)\rangle_X + \langle a;z(1)\rangle_X + \langle b;z(2)\rangle_X \\
& = & \int_0^{1} \langle q; \xi \rangle_{Z';Z} {\mathrm{d}}s + \int_{1}^2 \langle q;\xi \rangle_{Z';Z}{\mathrm{d}}s,
\end{aligned}$$ which leads to $\langle (a,b,w);\mathcal{K}(\xi)\rangle_{\mathcal{Y}';\mathcal{Y}} = \langle q ; \xi \rangle_{{\mathrm{L}}^2(0,2;Z');{\mathrm{L}}^2(0,2;Z)}$ and thus completes the proof.
Lemma \[lemma-adjS\] enables us to conveniently express the adjoint operators of $\mathbf{S}_u(\bar{u},\bar{\tau})$ and $\mathbf{S}_\tau(\bar{u},\bar{\tau})$.
The adjoint operators $\mathbf{S}_u(\bar{u},\bar{\tau})^{\ast} \in \mathcal{L}\left(\mathcal{Y}';{\mathrm{L}}^2(0,2;U')\right)$ and $\mathbf{S}_\tau(\bar{u},\bar{\tau})^{\ast} \in \mathcal{L}\left(\mathcal{Y}';{\mathbb{R}}\right)$ are given by $$\begin{aligned}
\mathbf{S}_u(\bar{u},\bar{\tau})^{\ast}.(a,b,w) = \dot{\pi}(\cdot,\bar{\tau}) f_u^\ast(\bar{y},\bar{u}) \mathcal{K}^\ast(a,b,w),
& &
\mathbf{S}_\tau(\bar{u},\bar{\tau})^{\ast}.(a,b,w) = \int_0^2\dot{\pi}_\tau \langle f(\bar{y},\bar{u}) ;\mathcal{K}^\ast(a,b,w)\rangle_{Y';Y} {\mathrm{d}}s.
\end{aligned}$$
As a consequence of Lemma \[lemma\_diffSys\] and , we have $\mathbf{S}_u(\bar{u},\bar{\tau}) = \dot{\pi}(\cdot,\bar{\tau}) \mathcal{K} \circ f_u(\bar{y},\bar{u})$ and $\mathbf{S}_\tau(\bar{u},\bar{\tau}) = \dot{\pi}_\tau \mathcal{K} \circ f(\bar{y},\bar{u})$. The result now follows from Lemma \[lemma-adjS\].
Lagrangian formulation
----------------------
We introduce the Hamiltonian: $$\begin{aligned}
\begin{array}{rrcl}
H:& Y \times U \times Z' & \mapsto & \mathbb{R} \\
& (y,u,p) & \rightarrow & \ell(y,u) + \langle p; f(y,u) \rangle_{Z';Z}.
\end{array}\end{aligned}$$ Given, $(\bar{y}, \bar{u}, \bar{\tau}) \in W(0,2;Y) \times {\mathrm{L}}^2(0,2;U) \times (0,T)$, the adjoint state $\bar{p}$ is assumed to satisfy $$\begin{aligned}
\bar{p}_{|(0,1)} \in W(0,1;Z'), & & \bar{p}_{|(1,2)} \in W(1,2;Z')\end{aligned}$$ (and thus $\bar{p} \in W(0,2;Z')$), by being defined as the solution of the following linear system: $$\begin{aligned}
\label{sysadj}
\left\{ \begin{array} {l}
-\dot{p} = \dot{\pi}(\cdot,\bar{\tau})H_y(\bar{y}, \bar{u}, p)\quad \text{on } (0,1) \cup (1,2),\\
p(2) = D \phi_2(\bar{y}(2)) , \\
p(1^+) - p(1^-) + D\phi_1(\bar{y}(1)) = 0.
\end{array} \right. \end{aligned}$$ Recall the continuous embeddings $W(I;Z') \hookrightarrow \mathcal{C}(\overline{I};X)$, for $I = (0,1)$ and $I = (1,2)$. In order to solve this backward system, we first consider $\bar{p}(2) = D\phi_2(\bar{y}(2))$ as the initial condition in $X$, next compute $\bar{p}$ on $(1,2)$ according to the first equation of , deduce $\bar{p}(1^-)$ from $\bar{p}(1^+)$ with the transmission condition in $X$, and finally compute $\bar{p}$ on $(0,1)$ as previously. From a more abstract point of view, since the affine mapping $p \mapsto H_y(\bar{y},\bar{u},p) = f_y(\bar{y},\bar{u})^*.p + \ell_y(\bar{y},\bar{u})$ is in the form of the right-hand-side in system , and thus Lemma \[lemma-adjS\] allows the existence and uniqueness of a solution to system .\
We define the Lagrangian as follows: $$\begin{aligned}
\begin{array} {rcl}
& L : &
\big(X\times X\times W(0,2;Y)\big) \times {\mathrm{L}}^2(0,2;U) \times (0,T) \times W(0,2;Y) \rightarrow \mathbb{R} \\
& & \big( \mathbf{y}= (a_1,a_2,y), u, \tau, p \big) \mapsto
\phi_1(a_1)+ \phi_2(a_2) + \displaystyle{\int_0^2} \Big(\dot{\pi}(s,\tau) H(y,u,p)(s)-\langle p(s),\dot{y}(s) \rangle_{Z';Z} \Big){\mathrm{d}}s \\
& & \hspace*{105pt}
- \langle p(0),y(0)-y_0 \rangle_X
+ \langle p(2),y(2)-a_2 \rangle_X
- \langle [p]_{1}, y(1)-a_1 \rangle_X,
\end{array} \end{aligned}$$ where we denote $[p]_{1} = p(1^+) - p(1^-)$. Note also that the Lagrangian $L$ is twice differentiable. The following lemma will enable us to calculate the first and second-order derivatives of the cost function $J$ in a convenient way.
\[lemma\_Ly\] The following identity holds: $$\label{lagrangianIsCost}
J(u,\tau)= L(\mathbf{S}(u,\tau),u,\tau,p), \quad
\forall (u,\tau,p) \in {\mathrm{L}}^2(0,2;U)\times (0,T)\times W(0,2;Y).$$ Moreover, setting $\bar{\mathbf{y}}=(\bar{y}(1),\bar{y}(2),\bar{y}) = \mathbf{S}(\bar{u},\bar{\tau}) \in \mathcal{Y}$ and $\bar{p}$ as the solution of the adjoint state equation corresponding to $(\bar{y},\bar{u},\bar{\tau})$, we have $L_{\mathbf{y}}(\bar{\mathbf{y}},\bar{u},\bar{\tau}, \bar{p})= 0
$ in $\mathcal{Y}'$.
Identity follows directly from the definitions of ${J}$, $\mathbf{S}$, and $L$. We decompose $L_{\mathbf{y}}(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p}) \in \mathcal{Y}'$ into $L_{a_1}(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p}) \in X$, $L_{a_2}(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p}) \in X$ and $L_{y}(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p})\in Y'$. From the definition of the adjoint state, we obtain the following identities in $X$: $$\begin{aligned}
L_{a_1}(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p})= D\phi_1(\bar{y}(1)) + [\bar{p}]_{1} =0, & &
L_{a_2}(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p})= D\phi_2(\bar{y}(2)) - \bar{p}(2)= 0.\end{aligned}$$ Moreover, for all $\delta y \in {\mathrm{W}}(0,2;Y)$, we obtain by integration by parts $$\begin{aligned}
\langle L_y(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p}); \delta y) \rangle_{Y';Y} & = &
\int_0^2 \dot{\pi}(s,\bar{\tau})\langle H_y(\bar{y},\bar{u},\bar{p})(s) ;\delta y(s)\rangle_{Y';Y}{\mathrm{d}}s
- \int_0^2 \langle \bar{p}(s);\delta \dot{y}(s) \rangle_{Y;Y'}{\mathrm{d}}s \\
& & \qquad -\, \langle \bar{p}(0), \delta y(0) \rangle_X
+ \langle \bar{p}(2), \delta y(2) \rangle_X
- \langle [\bar{p}]_{1}, \delta y(1) \rangle_X \\
& =& 0,\end{aligned}$$ which concludes the proof.
First and second-order optimality conditions {#first-and-second-order-optimality-conditions}
--------------------------------------------
Throughout this subsection $\bar{p}$ denotes the adjoint state corresponding to the triplet $(\bar{y},\bar{u},\bar{\tau})$, with $\bar{y} = S(\bar{u},\bar{\tau})$. We further denote by ${\mathrm{I}}$ the identity mapping on ${\mathrm{L}}^2(0,2;U')$.
\[propJ1\] The first-order derivatives of $J$ are given by: $$\begin{aligned}
{J}_u(\bar{u},\bar{\tau}).v= L_u(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p}).v
= \int_0^2 \dot{\pi}(s,\bar{\tau}) H_u(\bar{y},\bar{u},\bar{p})(s).v(s) {\mathrm{d}}s, \quad
{J}_\tau(\bar{u},{\bar\tau})= L_\tau(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p}) =
\int_0^2 \dot{\pi}_\tau H(\bar{y},\bar{u},\bar{p})(s) {\mathrm{d}}s.\end{aligned}$$ If the pair $(\bar{u},\bar{\tau})$ is optimal, then $D{J}(\bar{u},\bar{\tau})=0$.
The proposition follows directly from Lemma \[lemma\_Ly\]. Applying the chain rule to , we obtain: $$\begin{aligned}
D{J}(\bar{u},\bar{\tau}) & = & \begin{pmatrix} L_{\mathbf{y}}(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p})\mathbf{S}_u(\bar{u},\bar{\tau}) +
L_u(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p}) \\
L_{\mathbf{y}}(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p})\mathbf{S}_\tau(\bar{u},\bar{\tau}) +
L_\tau(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p}) \end{pmatrix}.
$$ Since from Lemma \[lemma\_Ly\] we have $L_{\mathbf{y}}(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p})=0$, the result follows.
Note that even in case where the Hamiltonian is strongly uniformly convex with respect to $u$, the optimal control will not be continuous at time $1$ (in general), because of the jump of the adjoint state.
\[propJ2\] The second-order derivative of ${J}$ is given by: $$\begin{aligned}
D^2 {J}(\bar{u},\bar{\tau}) & = &
\begin{pmatrix}
\mathbf{S}_u^\ast(\bar{u},\bar{\tau}) & {\mathrm{I}}& 0 \\
\mathbf{S}_\tau^\ast(\bar{u},\bar{\tau}) & 0 & 1
\end{pmatrix}
D^2 L(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p})
\begin{pmatrix}
\mathbf{S}_u(\bar{u},\bar{\tau}) & \mathbf{S}_\tau(\bar{u},\bar{\tau}) \\
{\mathrm{I}}& 0 \\
0 & 1
\end{pmatrix}.\end{aligned}$$ The second-order derivatives read, in a more explicit form: $$\begin{aligned}
D^2J(\bar{u},\bar{\tau}).\big( (v,\theta), (\hat{v},\hat{\theta}) \big)
& = & D^2 \phi_1(\bar{y}(1)). \big( z(1), \hat{z}(1) \big)
+ D^2 \phi_2(\bar{y}(2)). \big( z(2), \hat{z}(2) \big) \\
& & + \int_0^2 \dot{\pi}(s,\bar{\tau}) D^2 H(\bar{y},\bar{u},\bar{p}).
\big( (z,v),(\hat{z},\hat{v}) \big)(s) {\mathrm{d}}s \\
& & + \theta \int_0^2 \dot{\pi}_\tau (s) DH(\bar{y},\bar{u},\bar{p})\big( \hat{z}, \hat{v} \big)(s) {\mathrm{d}}s
+\hat{\theta} \int_0^2 \dot{\pi}_\tau (s) DH(\bar{y},\bar{u},\bar{p})\big( z, v \big)(s) {\mathrm{d}}s,\end{aligned}$$ where $z= S_u(\bar{u},\bar{\tau}).v+ S_\tau(\bar{u},\bar{\tau}).\theta$ and $\hat{z}= S_u(\bar{u},\bar{\tau}).\hat{v}+ S_\tau(\bar{u},\bar{\tau}).\hat{\theta}$.
Once again, the proposition follows directly from Lemma \[lemma\_Ly\]. Applying the chain rule to , we obtain: $$\begin{aligned}
D^2{J}(\bar{u},\bar{\tau}) & = &
\begin{pmatrix}
\mathbf{S}_u^\ast(\bar{u},\bar{\tau}) & {\mathrm{I}}& 0 \\
\mathbf{S}_\tau^\ast(\bar{u},\bar{\tau}) & 0 & 1
\end{pmatrix}
D^2 L(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p})
\begin{pmatrix}
\mathbf{S}_u(\bar{u},\bar{\tau}) & \mathbf{S}_\tau(\bar{u},\bar{\tau}) \\
{\mathrm{I}}& 0 \\
0 & 1
\end{pmatrix}
+
L_{\mathbf{y}}(\bar{\mathbf{y}},\bar{u},\bar{\tau}) D^2 \mathbf{S}(\bar{u},\bar{\tau}).\end{aligned}$$ The term involving $D^2\mathbf{S}(\bar{u},\bar{\tau})$ vanishes, since $L_{\mathbf{y}}(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p})= 0$. The explicit form follows from the compact relation below, where the notation of the variable $(\bar{\mathbf{y}},\bar{u},\bar{\tau},\bar{p})$ has been omitted, for a sake of clarity: $$\begin{aligned}
D^2 L & = &
\left(
\begin{array}{cc|ccc}
D^2 \phi_1(\bar{y}(1)) & 0 & 0 & 0 & 0 \\
0 & D^2 \phi_2(\bar{y}(2)) & 0 & 0 & 0 \\ \hline
0 & 0 & L_{yy} & L_{yu} & L_{y\tau} \\
0 & 0 & L_{uy} & L_{uu} & L_{u\tau} \\
0 & 0 & L_{\tau y} & L_{\tau u} & 0
\end{array}
\right).\end{aligned}$$ Denoting $\mathbf{z} = (z(1),z(2),z)$ and $\hat{\mathbf{z}} = (\hat{z}(1),\hat{z}(2),\hat{z})$, the partial derivatives above are formally given by: $$\begin{aligned}
\left\langle
\begin{pmatrix}
L_{yy} & L_{yu} \\
L_{uy} & L_{uu}
\end{pmatrix} .
\begin{pmatrix}
z \\
v
\end{pmatrix} ; \begin{pmatrix} \hat{z} \\ \hat{v} \end{pmatrix} \right\rangle_{{\mathrm{L}}^2(0,2;Y)';{\mathrm{L}}^2(0,2;Y)}
& = & \int_0^2 \dot{\pi}(s,\bar{\tau}) D^2 H(\bar{y},\bar{u},\bar{p})(s).
\big( (z,v),(\hat{z},\hat{v}) \big) (s) {\mathrm{d}}s \\
\left\langle
\begin{pmatrix}
L_{\tau y} \\
L_{\tau u}
\end{pmatrix};
\begin{pmatrix}
z \\
v
\end{pmatrix}
\right\rangle_{{\mathrm{L}}^2(0,2;Y)';{\mathrm{L}}^2(0,2;Y)}
& = & \int_0^2 \dot{\pi}_\tau DH(\bar{y},\bar{u},\bar{p})(s). ( z, v ) (s) {\mathrm{d}}s.\end{aligned}$$ So the proof is complete.
If $(\bar{u},\bar{\tau})$ is a solution to problem , then for all $(v,\theta) \in {\mathrm{L}}^2(0,2;U)\times {\mathbb{R}}$, $$\begin{aligned}
D^2J(\bar{u},\bar{\tau}).(v,\theta)^2 & \leq & 0.\end{aligned}$$ Conversly, if $DJ(\bar{u},\bar{\tau})= 0$, and if there exists $\alpha > 0$ such that for all $(v,\theta) \in {\mathrm{L}}^2(0,2;U) \times {\mathbb{R}}$ $$\begin{aligned}
\label{eqCS2}
D^2J(\bar{u},\bar{\tau}).(v,\theta)^2 & \leq & -\alpha \big( \| v \|_{L^2(0,2;U)}^2 + \theta^2), $$ then $(\bar{u},\bar{\tau})$ is a local solution to problem . More precisely, for all $\beta \in (0,\alpha)$, there exists $\varepsilon>0$, such that for all $(u,\tau) \in {\mathrm{L}}^2(0,2;U) \times (0,T)$, $$\begin{aligned}
\label{eqGrowth}
\| u - \bar{u} \|_{L^2(0,T;U)} + |\tau - \bar{\tau}| \leq \varepsilon
& \Rightarrow & J(u,\tau) \leq J(\bar{u},\bar{\tau}) - \frac{\beta}{2} \big( \| u-\bar{u} \|_{L^2(0,2;U)}^2 + (\tau-\bar{\tau})^2).\end{aligned}$$
If $(\bar{u},\bar{\tau})$ is a solution to , then for $\varepsilon>0$ small enough the pair $(\bar{u}+v,\bar{\tau}+ \varepsilon \theta)$ is feasible and therefore $$0 \geq \lim_{\varepsilon \downarrow 0} \frac{J(\bar{u}+\varepsilon v, \bar{\tau}+ \varepsilon \theta)-J(\bar{u},\bar{\tau})}{\varepsilon^2}= \frac{1}{2}D^2 J(\bar{u},\bar{\tau})(v,\theta)^2,$$ since $DJ(\bar{u},\bar{\tau})=0$. Conversly, assume that $DJ(\bar{u},\bar{\tau})=0$ and that holds. Let $\varepsilon>0$ be such that for all $(u,\tau) \in {\mathrm{L}}^2(0,2;U) \times (0,T)$ we have $$\begin{aligned}
\label{eqContHessian}
\| u-\bar{u} \|_{{\mathrm{L}}^2(0,2;U)} + |\tau-\bar{\tau}| \leq \varepsilon & \Rightarrow &
| (D^2J(u,\tau) - D^2J(\bar{u},\bar{\tau})).(v,\theta)^2 | \leq (\alpha-\beta)( \| v \|_{{\mathrm{L}}(0,2;Y)}^2 + \theta^2).\end{aligned}$$ By Taylor’s Theorem, for all $(u,\tau) \in {\mathrm{L}}^2(0,2;U) \times (0,T)$ such that $\| u-\bar{u} \|_{{\mathrm{L}}^2(0,2;U)} + |\tau-\bar{\tau}| \leq \varepsilon$, there exists $\gamma \in [0,1]$ such that $$\begin{aligned}
J(u,\tau) & = & J(\bar{u},\bar{\tau}) + \frac{1}{2}D^2J(\bar{u}+ \gamma v, \bar{\tau}+ \gamma \theta).(v,\theta)^2,\end{aligned}$$ where $v= u-\bar{u}$ and $\theta= \tau-\bar{\tau}$. Combined with , we obtain , which completes the proof.
Numerical realization {#sec4}
=====================
Method {#secpractice}
------
The numerical realization is based on a Newton method for the reduced formulation. In order to reach the attraction area for the Newton’s method, first some Barzilai-Borwein gradient steps are performed (see [@BB] for instance). In these gradient steps, only two inner products are computed for each step. Next, when the norm of the gradient is small enough, we switch to the full-step Newton’s algorithm, which is faster (quadratic convergence), but demands one linear system solve at each iteration. Switching to the gradient step is monitored by the norm of the gradient given by $$\begin{aligned}
||| (J_u,J_\tau) ||| : = \| (J_u,J_\tau) \|_{{\mathrm{L}}^2(0,2;{\mathbb{R}}^m) \times
{\mathbb{R}}}^2 & = & \int_0^2 |J_u|^2_{{\mathbb{R}}^m}(s){\mathrm{d}}s + |J_\tau|_{{\mathbb{R}}}^2.\end{aligned}$$ The integral above is approximated by the trapezoidal rule. The algorithm then performed is the following:
Initialization:
: $u = 0$, $\tau_0 = T/2$, for $1\leq i\leq N$, $s(i) = 2i/N$.
Gradient:
: Compute $(J_u,J_\tau)$.
Barzilai-Borwein steps:
: While $||| (J_u,J_\tau) ||| > 1.e-4$, do gradient steps.
Newton steps:
: While $||| (J_u,J_\tau) ||| > 1.e-12$, do:\
$\bullet$ Compute $(\delta u, \delta \tau)$ by solving system .\
$\bullet$ Update the unknowns: $u_{k+1} = u_k + \delta u$, $\tau_{k+1} = \tau_k + \delta \tau$.\
$\bullet$ Update the gradient.
Post-processing:
: for $1\leq i\leq N$, $t(i) := \pi(s(i),\tau)$.
Recall that $J$ is defined in . The derivatives of $J$ are provided by Proposition and Proposition . For solving one Newton step, we use the Gmres algorithm, calling only the evaluation of the mapping $$\begin{aligned}
\left(\begin{matrix} \delta u \\ \delta \tau \end{matrix} \right) & \mapsto &
\left(\begin{matrix}J_{uu} & J_{u\tau} \\ J_{\tau u} & J_{\tau\tau} \end{matrix} \right) .\left(\begin{matrix} \delta u \\ \delta \tau \end{matrix} \right), \label{sysNewton}\end{aligned}$$ which is in particular, for $(u,\tau)$ and $y = S(u,\tau)$ given, the evaluation of the following quantities: $$\begin{aligned}
J_{uu}.(\delta u, \cdot) & = & \dot{\pi}(\cdot,\tau) \Big( S_u^\ast H_{yy} S_u.\delta u + H_{uy}S_u.\delta u + S_u^\ast H_{yu}.\delta u + H_{uu}.\delta u \Big) \\
& & + S_u^\ast D^2\phi_1(y(1))S_u.\delta u + S_u^\ast D^2\phi_2(y(2))S_u.\delta u , \\
J_{\tau u} & = & \dot{\pi}(\cdot,\tau) \Big( S_u^\ast H_y + H_u + H_{uy}S_\tau + S^\ast_u H_{yy}S_\tau \Big) + S_u^\ast D^2\phi_1(y(1))S_\tau + S_u^\ast D^2\phi_2(y(2))S_\tau.\end{aligned}$$ Note, in particular, that we do not need any evaluation of the adjoint operator $S_\tau^{\ast}$.
Illustrations
-------------
In all the tests below, for the cost function we choose $$\begin{aligned}
\ell(y,u) = -\frac{\alpha}{2} |u|_{{\mathbb{R}}^m}^2,\end{aligned}$$ for a cost parameter $\alpha >0$. The examples dealing with ordinary differential systems, considered below, are inspired by [@Trelat].
### The Lotka-Volterra prey-predator system
Consider the following differential system: $$\begin{aligned}
\left( \begin{matrix} \dot{y}_1 \\ \dot{y}_2 \end{matrix} \right) & = &
\left( \begin{matrix} (y_1(a-by_2) + \text{\textcolor{black}{$u_1$}}y_1)(1-c_1y_1) \\ (y_2(qy_1-r)+ \text{\textcolor{black}{$u_2$}}y_2)(1-c_2y_2) \end{matrix} \right).
\end{aligned}$$ In this system the variables $y_1$ and $y_2$ represent the densities of population of preys and predators, respectively. The multiplicative terms of type $(1-c_iy_i)$ are considered in order to limit the values of these densities to $1/c_i$. We assume that we can control the birth rates and death rates of both species, through the bilinear control made of $u_1$ and $u_2$. The numerical method given in section \[secpractice\] is applied to this system, with $$\begin{aligned}
T = 30.0, \quad y_0 = (1.0,2.0)^T, \quad \alpha = 10.0, \quad a=0.3, \quad b=0.1, \quad r=0.2, \quad q=0.1, \quad c_1 =c_2 = 0.05.\end{aligned}$$ We do not consider any terminal cost, namely $\phi_2 \equiv 0$, and the functional we maximize at some time $\tau$ is $\phi_1(y) = y_2$. It means that we want to maximize the density of population of predators. The tests presented in Figures \[figLV\] and \[figULV\] are obtained with a Crank-Nicholson time-discretization, with $N = 3000$ time steps.
\
\
At the beginning, the action of the control leads to diminution of the populations of both preys and predators, then leads to introduction of preys in order to feed the predators, and then favor their reproduction. The maximum of predators is reached at $\tau \approx 20.57$. As expected, the time-derivative of the second-component of the state has a jump. Note that in this example the first-order optimality condition gives in particular the equalities $$\begin{aligned}
\alpha u_1 = y_1(1-c_1y_1)p_1, & & \alpha u_2 = y_2(1-c_2y_2)p_2.\end{aligned}$$ Without any terminal cost, the control is indeed null for $t > \tau$, as expected in system defining the adjoint-state $(p_1,p_2)^T$.
#### With terminal cost.
In order to avoid the extinction of population of the preys after having maximized the population of predators, we consider the terminal cost functional $$\begin{aligned}
\phi_2(y) & = & -\beta \log \left( \left|\frac{y_1}{y_{\mathrm{des}}}\right| \right)^2,\end{aligned}$$ where $\beta > 0$ is a coefficient chosen large enough and $y_{\mathrm{des}}$ is the desired value for the density of preys at time $t=T$. The idea is to penalize the extinction of this population (the case $y_1(T) = 0$ is forbidden), and to force to reach the desired value by choosing $\beta$ large enough. With the same coefficients chosen as previously, the same time-discretization, and with $\beta = 25.0$ and $y_{\mathrm{des}} = 1.0$, we obtain the results presented in Figures \[figterm1\] and \[figterm2\].\
\
\
The optimal time for the maximum of predators is $\tau \approx 14.87$. As expected, the maximum is smaller than the one reached without terminal cost, because the control has to be activated for $t >\tau$ in order to take into account the functional $\phi_2$.
### The simple damped pendulum
Consider the model of a simple pendulum, as described in Figure \[figPendulum\] below.
The weight is assumed to satisfy the Newton’s law, and so in particular the angle $\theta$ has to to satisfy the equation $\ddot{\theta} +\lambda\dot{\theta} +\mu\sin \theta = 0$, where $\lambda >0$ is a damping term, and $\mu$ is a coefficient depending on the gravity field and the length of the taut rope. the corresponding differential system is $$\begin{aligned}
\left( \begin{matrix} \dot{y}_1 \\ \dot{y}_2 \end{matrix} \right) & = &
\left( \begin{matrix} y_2 \\ -\lambda y_2 -\mu \sin y_1 \end{matrix} \right) + \left(\begin{matrix} 0 \\ u \end{matrix}\right).\end{aligned}$$ where $y = (y_1,y_2)^T := (\theta,\dot{\theta})^T$, and where the control $u$ represents some additional horizontal force in the Newton’s law. The numerical method developed in section \[secpractice\] is performed, with $$\begin{aligned}
T = 25.0, \quad y_0 = (-1.0,0.0)^T, \quad \alpha = 10.0, \quad \lambda =0.03, \quad \mu =1.0.\end{aligned}$$ Here again, we do not consider any terminal cost, namely $\phi_2 \equiv 0$. The functional we maximize at some time $\tau$ is $\phi_1(y) = y_1$, namely the angle and thus the height of the weight. The results presented in Figures \[figP\] and \[figUP\] are obtained with a Crank-Nicholson time-discretization, with $N = 2500$ time steps.\
\
\
The maximum is reached at $\tau \approx 17.22$. As we can see, the control can be kept activated on several pseudo-periods, and it can compensate the damping effect. But numerically, the difficulty lies in considering an appropriate initialization for $\tau$ (which is not $T/2$ in that case). Indeed, our numerical approach enables us only to get a critical point, and thus without a correct initialization, we can also catch some local maximum or minimum, instead of the desired maximum.
### A partial differential equation {#secnumPDE}
Consider the following Burgers-type system $$\begin{aligned}
\left\{ \begin{array} {ll}
\dot{y} = \nu y_{xx} - \beta yy_x + \mathds{1}_{\omega} u, & (x,t) \in (0,1) \times (0,T), \\
y(0,t) = y(1,t) = 0, & t \in (0,T), \\
y(x,0) = 10.(1 - e^{-(1-x)})(e^{-(1-x)} - e^{-1}), & x \in (0,1),
\end{array} \right.\end{aligned}$$ where $\nu$ and $\beta$ are positive constant. Here the unknown $y$ is considered in the space $W(0,T; {\mathrm{H}}^1_0(0,1))$ (defined in section \[secset\]), corresponding to the Gelfand triplet ${\mathrm{H}}^1_0(0,1) \hookrightarrow {\mathrm{L}}^2(0,1) \hookrightarrow {\mathrm{H}}^{-1}(0,1)$. Recall that ${\mathrm{H}}^1(0,1) \hookrightarrow \mathcal{C}([0,1])$. The question of well-posedness for such a system is addressed in [@Volkwein]. The control $u$ is considered in ${\mathrm{L}}^2(\omega)$, with $\omega = [0.00 ; 0.25]$, and the cost term is given by $\ell(y,u) = -\frac{\alpha}{2} \|u\|^2_{{\mathrm{L}}^2(\omega)}$. We want to maximize, at some optimal time $\tau \in (0,T)$, the following quantity $$\begin{aligned}
\phi_1(y(\cdot,\tau)) & = & \frac{1}{2}\int_D |y(x,\tau)|^2{\mathrm{d}}x\end{aligned}$$ with $D = [0.25 ; 0.30]$, and without considering any terminal cost: $\phi_2 \equiv 0$. The space discretization is done with finite P1-elements, with $n = 101$ degrees of freedom for the state variable, and hence $m=26$ unknowns for the control variable. The time discretization is done with a Crank-Nicholson scheme, with $N = 1000$ time steps. We consider the following parameters $$\begin{aligned}
T = 10.0, \quad \alpha = 2.10^{-9}, \quad \nu = 2.10^{-4}, \quad \beta = 0.05.\end{aligned}$$ The evolutions of the state and the control through the time are represented in Figure \[figBurgers\] below.\
\
$ t = 0.0777 $
\
$ t = 4.7111 $
\
$ t = 4.7839 $
\
$ t = 4.8325 $
\
\
$ t = 4.8519 $
\
$ t = 4.8568 \approx \tau $
\
$ t = 5.0111 $
\
$ t = 10.0000 $
\
\
The maximum is reached for $\tau \approx 4.86$. In view of the profile of the initial condition, without control the solution is transported to the left part of the domain. The simulation shows that the control seems to wait for the so transported energy, before being mainly active during a small period before $t = \tau$, operating a bumping effect. Its influence on the state leads to transport some energy from $\omega$ to $D$. For $t > \tau$, the energy obtained on $D$ is diffused into the profile (the viscosity $\nu$ is chosen here very small, so that the diffusion due to the term $\nu y_{xx}$ is almost not noticed). Note that some delay is encoded into the model, that is to say when the maximum is reached for the state, at time $\tau$, and even a little bit before this moment, the control is no longer active.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors gratefully acknowledge support by the Austrian Science Fund (FWF) special research grant SFB-F32 “Mathematical Optimization and Applications in Biomedical Sciences”, and by the ERC advanced grant 668998 (OCLOC) under the EU’s H2020 research program.
[^1]: Institute for Mathematics and Scientific Computing, Karl-Franzens-Universität, Heinrichstr. 36, 8010 Graz, Austria, email: [[email protected]]{}, [[email protected]]{}.
[^2]: Institute for Mathematics and Scientific Computing, Karl-Franzens-Universität, Heinrichstr. 36, 8010 Graz, Austria, and Radon Institute, Austrian Academy of Sciences, email: [[email protected]]{}.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Optimal packing of spheres in $\mathbb R^d$ is studied by optimization of the energy $E$ (effective conductivity) of composites with ideally conducting spherical inclusions. It is demonstrated that the minimum of $E$ over locations of spheres is attained at the optimal packing. The energy is estimated in the framework of structural approximations. This method yields upper bounds and sometimes exact values for the maximal concentrations of spheres in $\mathbb R^d$. A constructive algorithm for the optimal locations of spheres associated to the classes of the Delaunay graphs is constructed.'
author:
- |
Vladimir Mityushev\
Pedagogical University,\
ul. Podchorazych 2, Krakow 30-084, Poland
title: 'Optimal packing of spheres in $\mathbb R^d$ and extremal effective conductivity'
---
Keyword: Optimal packing of spheres; discrete energy; effective conductivity, Voronoi tessellation; Delaunay graph
MSC: 52C17, 05B40, 74Q15
Introduction
============
Packing problems refers to geometrical optimization problems. Various methods from different topics of mathematics were applied to these problems. Extended reviews can be found in the books [@Bezdek] [@Bezdek2], [@CS], [@Hales], [@Toth]. One of the most popular problem is the optimal packing of spheres in $\mathbb R^d$ [@CS]. Its complete solution for 2D is given in [@Toth] and for 3D in [@Hales].
It was noted in [@Mit2012] that solution to the physical problem of the optimal effective conductivity in 2D implies solution to the geometrical problem of the packing disks. The physical problem can be stated as follows. Given ideal conductors (having infinite conductivity coefficient) of the shape $D_i \subset \mathbb R^d$ ($i=1,2,\ldots$). To locate $D_i$ ($i=1,2,\ldots$) in the host medium of a finite conductivity in such a way that the homogenized medium is macroscopically isotropic and its effective conductivity attains the minimal value. Rigorous mathematical statements of the physical problem with fixed geometry can be found in [@BP] and in other works devoted to homogenization.
Recent results in structural approximations [@LB], [@Kolpakov] shown that densely packed composites can be investigated by the functional associated to the discrete energy. It has the following structure $$\min_{t_1, t_2, \ldots}\;\mathop{{\sum}}_{k,j} g_{kj}^{(0)} |t_k-t_j|^2,
\label{eq:en1}$$ where the minimum is taken over the values $t_j$ prescribed to the inclusion $D_j$. The value $g_{kj}^{(0)}$ expresses the main term of the interparticle flux between the neighbor inclusions $D_j$ and $D_k$ when the distance $\delta_{jk}$ between them tends to zero. Usually, $g_{kj}^{(0)}$ has a simple form. In the case of linear conductivity, $g_{kj}^{(0)}$ not always has a singularity as $\delta_{jk} \to 0$ (see 3D examples in [@Kolpakov]). For instance, the flux between two spheres in $\mathbb R^d$ for $d>3$ is always regular for linear conductivity. Fortunately, $g_{kj}^{(0)}$ is always singular for non-linear conductivity governed by the $p$-Laplacian for $p>\frac{d+1}2$. This fact enables us to consider the minimum not only in $t_j$ but also in the locations of the spherical inclusions $D_j$.
This paper is devoted to study the minimum and its application to the sphere packing problem. Sec.\[sec2\] shortly presents the structural approximation theory following [@Keller], [@LB], [@Kolpakov] in $\mathbb R^d$ for $d=2,3$. Sec.\[sec3\] is devoted to extension of the theory to the general space $\mathbb R^d$ and construction of the corresponding discrete energy. Estimations of energy are performed in Sec.\[sec4\]. Sec.\[sec5\] contains concluding remarks and discussion.
Structural approximation in $\mathbb R^d$ {#sec2}
=========================================
Let $\boldsymbol{\nu}_j$ ($1,2, \ldots, d$) be the fundamental translation vectors in the space $\mathbb R^d$ ($d\geq 2$) which form a lattice $\mathcal Q=\{\sum_{j=1}^d m_j \boldsymbol{\nu}_j:m_j \in \mathbb Z \}$. The fundamental parallelotope $Q_{0}$ is defined by its $2^d$ vertices $\frac 12 \sum_{j=1}^d (\pm\boldsymbol{\nu}_j)$. Tow points $\mathbf a, \mathbf b \in \mathbb R^d$ are identified if their difference $\mathbf a- \mathbf b=\sum_{j=1}^d m_j \boldsymbol{\nu}_j$ belongs to the lattice $\mathcal Q$. Hence, such a topology is introduced on $Q_{0}$ that the opposite faces are glued. In the case $\mathbb R^2$, the fundamental parallelogram $Q_{0}$ can be considered as the classical flat torus. We will call the introduced topology on $Q_{0}$ in $\mathbb R^d$ also by toroidal. The distance $\|\mathbf a-\mathbf b\|$ between two points $\mathbf a,\mathbf b \in Q_0$ is introduced as $$\|\mathbf a-\mathbf b\|:=\min_{m_1,\ldots,m_d \in \mathbb Z} \left|\mathbf a-\mathbf b+\sum_{j=1}^d m_j \boldsymbol{\nu}_j\right|,
\label{eq:sam1}$$ where the modulus means the Euclidean distance between the points $\mathbf a$ and $\mathbf b$ in $\mathbb R^d$.
![Periodicity cell $Q_{0}$. White and gray balls (disks) are identified in the torus topology.[]{data-label="fig:Figure-1"}](Figure-1.eps){width=".7\textwidth"}
Consider $n$ non-overlapping balls $D_k = \{\mathbf x \in \mathbb R^d: |\mathbf x - \mathbf a_k|<r \}$ of radius $r$ with the centers $\mathbf a_k$ in the cell $Q_{0}$ (see Fig.\[fig:Figure-1\]). Let $D_0$ be the complement of all closure balls $D_k \cup \partial D_k$ to the domain $Q_{0}$. Following [@LB], [@Kolpakov] and works cited therein we now shortly present the state of art in the non-linear homogenization of periodic composites governed by the $p$-Laplace equation ($p\geq 2$) $$\nabla \cdot |\nabla u|^{p-2} \nabla u = 0, \quad \mathbf x \in D_{0}.
\label{eq:p1}$$ The scalar function $-u(\mathbf x)$ is called the potential (sometimes the potential is taken as $u(\mathbf x)$), the vector function $\mathbf J=|\nabla u|^{p-2} \nabla u$ is called the flux. The $p$-Laplace equation can be smoothly continued into the perforated domain $D_{0}+\sum_{j=1}^d \sum_{m_j\in \mathbb Z} m_j \boldsymbol{\nu}_j$. Introduce a unit vector $\boldsymbol{\xi} =(\xi_1,\xi_2,\ldots,\xi_d)\in \mathbb R^d$. The external potential is determined by the linear function $u_0(\mathbf x) =\boldsymbol{\xi} \cdot \mathbf x$ up to an arbitrary additive constant. The vector $\boldsymbol{\Omega}_{m_1,\ldots,m_d}=\sum_{j=1}^d m_j \xi_j \boldsymbol{\nu}_j$ with $m_j\in \mathbb Z$ corresponds to the potential jumps with respect to the lattice $\mathcal Q$, i.e., the potential is quasi-periodic: $$u\left(\mathbf x+\sum_{j=1}^d m_j \boldsymbol{\nu}_j\right)=u(\mathbf x)+\boldsymbol{\Omega}_{m_1,\ldots,m_d},\quad \mathbf x \in D_{0}\; (\forall m_j\in \mathbb Z).
\label{eq:p3}$$ The flux is periodic: $$\mathbf J\left(\mathbf x+\sum_{j=1}^d m_j \boldsymbol{\nu}_j\right)=\mathbf J(\mathbf x), \quad \mathbf x \in D_{0}\; (\forall m_j\in \mathbb Z).
\label{eq:p3J}$$
The potential satisfies the boundary conditions $$u(\mathbf x)=t_k, \quad |\mathbf x - \mathbf a_k|=r\; (k=1,2,\ldots,n),
\label{eq:p4}$$ where $t_k$ are undetermined constants. The total normal flux through each sphere vanishes: $$\int_{\partial D_k}\mathbf J(\mathbf x)\cdot \mathbf n\; ds =0, \quad k=1,2,\ldots,n,
\label{eq:p5}$$ where $\mathbf n$ denotes the outward unit normal vector to the sphere $\partial D_k$.
The problem - describes the field in the periodic composite when the inclusions $D_k$ are occupied by perfect conductor and the host conductivity is governed by equation . In electrostatics, $\mathbf E= \nabla u$ denotes the electric field, $\mathbf J= |\nabla u|^{p-2} \nabla u$ the electric current density. Energy passing through the cell per unit volume (effective conductivity) is calculated by formula (6.1.6) from [@LB] $$\lambda=\frac 1{2|Q_0|} \int_{D_{0}} \mathbf J \cdot \mathbf E\; d\mathbf x =\frac 1{2|Q_0|} \int_{D_{0}} |\nabla u|^{p} d\mathbf x,
\label{eq:p2}$$ where $|Q_0|$ stands for the volume of $Q_0$. It is assumed that the centers $\mathbf a_k$ ($k = 1,2,\ldots,n$) are distributed in $Q_{0}$ in such a way that the corresponding composite is isotropic in macroscale, i.e., the effective conductivity of the composite is expressed by a scalar $\lambda$ and does not depend on the unit external flux $\boldsymbol{\xi}$.
The energy $\lambda$ can be found as the minimum of the functional [@LB]: $$\lambda=\min_{v \in V} \frac 1{2|Q_0|} \int_{D_{0}} |\nabla v|^{p} d\mathbf x,
\label{eq:p2e}$$ where the space $V$ consists of the quasi-periodic functions from the Sobolev space $W^{1,p}(Q_0)$: $$V=\{v \in W^{1,p}(Q_0): v(\mathbf x)=t_k\; \mbox{on}\; D_k\; (k=1,2,\ldots,n)\}.
\label{eq:p2v}$$ Here, quasi-periodicity means that the conditions and are fulfilled for $v$. Though the definition of the space $V$ depends on the external flux $\boldsymbol{\xi}$, the energy does not depend because of isotropy.
The discrete network is a graph $\Gamma$ on the cell $Q_0$ with the vertices at $\mathbf a_k$ ($k=1,2,\ldots,n$) and the edges correspond to the necks between neighbors. Neighbors are defined as balls (disks) that share a common edge of the Voronoi tessellation of $Q_0$ with respect to their centers in toroidal topology. For each fixed $\mathbf a_k$, introduce the set $J_k$ of indexes for neighbor vertices and their total number $N_k=\# J_k$. The Voronoi tessellation in finite domains and the Delaunay graph $\Gamma$ are precisely described in [@LB], [@Kolpakov]. The evident modification to periodic structures can be taken over. We use the term the Delaunay graph following [@LB], [@Kolpakov] because it slightly differs from the Delaunay triangulation in degenerate cases. For example, consider a square and its four vertices. The traditional Delaunay triangulation has four sides of the square and one of the diagonals. In our approach, the Delaunay graph has only four sides (see Example \[ex4.3\] in Sec.\[sec4\]).
The discrete network model is based on the justification that the flux is concentrated in the necks between closely spaced inclusions (see [@LB] for non-linear equation and [@LB], [@Kolpakov] in the linear case $p=2$). First, we discuss the linear case when $p=2$. For two balls $D_k$ and $D_j$ the computation of the relative interparticle flux $g_{kj}$ (transport coefficient in terms of [@Kolpakov] and capacity in [@Keller]) relies on Keller’s formulae [@Keller] $$g_{kj}=g_{kj}^{(0)} +O(\delta_{jk}^0), \quad \delta \to 0,
\label{eq:keller0}$$ where in 3D $$g_{kj}^{(0)}=-\pi r \ln \delta_{kj}
\label{eq:keller1}$$ and in 2D $$g_{kj}^{(0)}=\pi \sqrt{\frac{r}{\delta_{kj}}}.
\label{eq:keller2}$$ Here, $\delta_{kj}$ denotes the gap between the balls $D_k$ and $D_j$: $$\delta_{kj}=\|\mathbf a_k-\mathbf a_j\|-2r.
\label{eq:keller3}$$ To each edge of the graph $\Gamma$ the flux or is assigned.
In the non-linear case for $p>2$, we have the following formulae due to [@LB] in 3D $$g_{kj}^{(0)}=\frac{\pi}{(p-2) r^{p-3}} \left(\frac{r}{\delta_{kj}} \right)^{p-2}
\label{eq:keller4}$$ and in 2D $$g_{kj}^{(0)}=\frac{(2p-5)!!}{(2p-4)!!} \;\frac{\pi^{\frac 32}}{r^{p-2}} \; \left(\frac{r}{\delta_{kj}} \right)^{p-\frac 32}.
\label{eq:keller5}$$ It is worth noting that $g_{kj}^{(0)}$ depends only on geometry of the problem, i.e., on $r$ and on the gap $\delta_{kj}$ given by .
The explicit coefficient $\frac{(2p-5)!!}{(2p-4)!!}$ is introduced in instead of the coefficient of the Taylor series for $(1-x)^{-\frac 12}$ in [@LB]. Formulae and are taken form [@LB] with slight corrections (see equations , and a remark below them).
Let for shortness, $\mathbf t= (t_1,t_2, \ldots,t_n) \in \mathbb R^n$ and $\mathbf a = (\mathbf a_1,\mathbf a_2, \ldots, \mathbf a_n) \in \mathbb R^{n} \times \mathbb R^{d}$. Introduce the double sum $$\mathop{{\sum}'}_{k,j} = \sum_{k=1}^n \sum_{j\in J_k}.
\label{eq:sa2d}$$ In order to formulate the main asymptotic result of [@LB] introduce the maximal length of edges $\delta = \max_k \max_{j \in J_k} \delta_{jk}$ of the graph $\Gamma$. Following [@LB], [@Kolpakov] we consider the class $\mathcal D$ of macroscopically isotropic composites with densely packed inclusions. The term “densely packed inclusions” means that any such a location of balls has a percolation chain for $\delta=0$. More precisely, consider a set of non-overlapping balls in $Q_0$ with $\delta \geq 0$ endowed with the toroidal topology. Change of $\delta$ means that the centers of balls are fixed but their radii change and remain equal. Let for $\delta =0$ there exists a chain of touching balls connecting the opposite faces of the parallelotope $Q_0$. Such a chain is called a percolation chain[^1]. Macroscopic isotropy implies that each pair of the opposite faces of $Q_0$ posses a percolation chain for $\delta =0$ (see Fig.\[fig:Figure-2\]).
![Percolation chains marked by black.[]{data-label="fig:Figure-2"}](Figure-2.eps){width=".7\textwidth"}
Any macroscopically isotropic location not belonging to the class $\mathcal D$ of densely packed inclusions can be replaced by an element of $\mathcal D$ having higher concentration. It is possible to do it by parallel translations of non-touching groups of balls to make them touched.
Introduce the main term of the discrete energy [@LB] $$\sigma = \min_{\mathbf t}\;\frac 1{2|Q_0|} \mathop{{\sum}'}_{k,j} g_{kj}^{(0)} |t_k-t_j|^p,
\label{eq:sa1d}$$ where $g_{kj}^{(0)}$ is given by or by .
\[thLB\] The continuous energy and the discrete energy for $p \geq 2$ in the spaces $\mathbb R^d$ ($d=2,3$) tend to $+\infty$ as $\delta \to +0$. Moreover, $\lambda$ can be approximated by $\sigma$ for sufficiently small $\delta$: $$\lambda=\sigma +O(\delta^0), \;\mbox{as} \; \delta \to 0.
\label{eq:sa2app}$$
Th.\[thLB\] justifies the theory of structural approximation [@LB], [@Kolpakov] based on a mesoscopic discretization when edges and vertices of the graph $\Gamma$ correspond to inclusions. This theorem was applied in [@LB], [@Kolpakov], [@ryl2008] to systematic study of densely packed composites with given locations of inclusions.
Discrete network in $\mathbb R^d$ {#sec3}
=================================
Though the theory of structural approximation [@LB], [@Kolpakov], [@ryl2008] was constructed and justified in $\mathbb R^2$ and in $\mathbb R^3$, its main results hold in the general space $\mathbb R^d$ under the condition that the interparticle flux $g_{kj}$ tends to infinity as $\delta_{kj} \to 0$. It can be established by a formal repetition of the arguments from [@LB], [@Kolpakov] in $\mathbb R^d$. In the present section, we partially fill this gap and calculate the main term $g_{kj}^{(0)}$ of $g_{kj}$ up to a constant order term $O(\delta_{jk}^0)$.
Following Keller [@Keller] and explanations by Kolpakov [@Kolpakov p. 18-25] we proceed to estimate the interparticle flux $g_{jk}$ for two balls $D_j$ and $D_k$ with the centers located at the points $(0,0,\ldots,0,\pm \frac a2)$ of radius $r<\frac a2$ in $\mathbb R^d$ ($d =2,3, \ldots$). Let the potential $u(\mathbf x)$ is equal to constants $\pm C$ on the spheres $(x_d\mp \frac a2)^2+R^2=r^2$ where $R = \sqrt{x_1^2+x_2^2+\ldots +x_{d-1}^2}$. Near the gap the spheres can be approximated by the paraboloids [@LB], [@Kolpakov] $$x_n=\pm \left(\frac{\delta_{jk}}2+\frac{R^2}{2r}\right),
\label{eq:sph1a}$$ where $\delta_{jk}=a-2r$. The potential can be approximated up to $O(\delta_{jk}^0)$ by the function (see (6.4.12) from [@LB] for 2D and 3D cases) $$u_0(\mathbf x) = \frac{x_d}{H(\mathbf x_{d-1})},
\label{eq:sa1a}$$ where $H(\mathbf x_{d-1})=\delta +\frac{R^2}{r}$ is the distance between the paraboloids. Following [@LB] we can approximate the local gradient by $$\nabla u_0(\mathbf x) =\left(0,\ldots,0, \frac{1}{\delta_{jk} +\frac{R^2}{r}}\right)+O(\delta_{jk}^0)
\label{eq:saf1}$$ and the local flux by $$|\nabla u_0(\mathbf x)|^{p-2} \nabla u_0(\mathbf x) =\left(0,\ldots,0, \frac{1}{\left(\delta_{jk} +\frac{R^2}{r}\right)^{p-1}}\right)+O(\delta_{jk}^0) .
\label{eq:saf2}$$ Then, $g_{jk}$ is approximated up tp $O(\delta_{jk}^0)$ by the integral over the $(d-1)$-dimensional ball $B$ on the hyperplane $x_d=0$ defined by inequality $R<r$: $$g_{jk}^{(0)} = \int_{B}\frac{d\mathbf x_{d-1}}{\left(\delta_{jk} +\frac{R^2}{r}\right)^{p-1}},
\label{eq:sa2a}$$ where the differential $d\mathbf x_{d-1} = dx_1\;dx_2\; \cdots dx_{d-1}$.
First, we consider partial cases of discussed in the previous works [@Keller], [@LB], [@Kolpakov]. Let $p=d=2$. Then, becomes $$g_{jk}^{(0)} = \int_{-r}^r\frac{dR}{\delta +\frac{R^2}{r}} =\pi \sqrt{\frac{r}{\delta_{jk}}}.
\label{eq:p2d2}$$ Let $p=2$ and $d=3$. Then, gives $$g_{jk}^{(0)} = 2 \pi \int_0^r\frac{R\;dR}{\delta_{jk} +\frac{R^2}{r}} = \pi r \ln{\frac{r}{\delta_{jk}}}.\label{eq:p2d3}$$ Formulae - coincide with the corresponding formulae from [@Keller], [@LB], [@Kolpakov] (see also -).
We now proceed to investigate the general case. The integral in the spherical coordinates $(R,\phi_1,\phi_2, \ldots, \phi_{d-2})$ becomes $$\begin{array}{lll}
g_{jk}^{(0)} =
\\
\int_0^{2\pi}d\phi_{d-2} \int_0^{\pi}d\phi_{d-3} \cdots \int_0^{\pi}d\phi_{1} \int_0^{r} \frac{\sin^{d-3} \phi_1 \sin^{d-4} \phi_2 \cdots \sin \phi_{d-3}}{\left(\delta +\frac{R^2}{r}\right)^{p-1}}R^{d-2} dR.
\end{array}
\label{eq:sa3a}$$ It can be calculated analogously to the volume of the $d$-dimensional ball. First, calculate $$\int_0^{2\pi}d\phi_{d-2} \int_0^{\pi}d\phi_{d-3} \cdots \int_0^{\pi} \sin^{d-3} \phi_1 \sin^{d-4} \phi_2 \cdots \sin \phi_{d-3} \;d\phi_{1}=
\frac{2\pi^{\frac{d-1}2}}{\Gamma \left(\frac{d-1}2\right)},
\label{eq:si1}$$ where the $\Gamma$-function is used. The integral in $R$ is calculated by formula $$\int_0^{r} \frac{R^{d-2} dR}{\left(\delta_{jk} +\frac{R^2}{r}\right)^{p-1}} = \frac{r^{d-1}\;
{}_2 F_1\left(\frac{d-1}{2},p-1,\frac{d+1}{2},-\frac{r}{\delta_{jk}}\right)}{\delta_{jk} ^{ p-1}(d-1) },
\label{eq:si2}$$ where the hypergeometric function ${}_2 F_1$ is used. Therefore, $$g_{jk} = \frac{2\pi^{\frac{d-1}2} r^{d-1}}{\Gamma \left(\frac{d-1}2\right)}
\frac{
{}_2 F_1\left(\frac{d-1}{2},p-1,\frac{d+1}{2},-\frac{r}{\delta_{jk}}\right)}
{\delta_{jk}^{ p-1}(d-1) } .
\label{eq:sa4a}$$ Hereafter, we assume that $p$ is a natural number greater than 2 and $$p>\frac{d+1}2.
\label{eq:pd}$$ The following asymptotic formula takes place for $Z \to \infty$ $${}_2F_1\left(\frac{d-1}{2},p-1,\frac{d+1}{2},-Z\right)=\frac{1}{Z^{\frac{d-1}{2}}}
\frac{\Gamma \left(\frac{d+1}2\right)\Gamma \left(p-\frac{d+1}2\right)}{\Gamma \left(p-1\right)}+O\left(\frac{1}{Z^{q}}\right),\label{eq:sa5a}$$ where $q>\frac{d-1}{2}$. Formula is got by use of the package $Mathematica^{\circledR}$ Applying to we obtain the main asymptotic term of $g_{jk}$ as $\delta_{jk} \to 0$ $$g^{(0)}_{jk} =\frac{1}{\delta_{jk}^{p-\frac{d+1}{2}}}
\frac{2 (\pi r)^{\frac{d-1}{2}}
\Gamma \left(\frac{d+1}{2}\right)
\Gamma \left(p-\frac{d+1}{2}\right)}
{(d-1) \Gamma \left(\frac{d-1}{2}\right)
\Gamma (p-1)}.
\label{eq:sa6a}$$ Let $d$ be an odd number. Then, becomes $$g^{(0)}_{jk} =\frac{1}{\delta_{jk}^{p-\frac{d+1}{2}}}
\frac{(\pi r)^{\frac{d-1}{2}}\left(p-\frac{d+3}{2}\right)!}
{(p-2)! }.
\label{eq:sa8a}$$ Let $d$ be an even number. Using formula $$\Gamma \left(\frac{m}2\right)=\frac{(m-2)!! \sqrt{\pi}}{2^{\frac{m-1}2}}, \quad
m \in \mathbb N,
\label{eq:sa7a}$$ we rewrite in the form[^2]
$$g^{(0)}_{jk} =\frac{1}{\delta_{jk}^{p-\frac{d+1}{2}}}
\frac{\sqrt{\pi} (\pi r)^{\frac{d-1}{2}}(2p-d-3)!!}
{2^{p-\frac{d+1}{2}}(p-2)}.
\label{eq:sa9a}$$
Introduce the functional associated with energy $$E(\mathbf t, \mathbf a)= \frac 1{2|Q_0|} \mathop{{\sum}'}_{k,j} (t_k-t_j)^p
f(\|\mathbf a_k-\mathbf a_j\|)
\label{eq:sa1f}$$ and the main term of the discrete energy $$\sigma= \min_{\mathbf t}\; \frac 1{2|Q_0|} \mathop{{\sum}'}_{k,j} (t_k-t_j)^p
f(\|\mathbf a_k-\mathbf a_j\|).
\label{eq:sa1g}$$ Here, the main term of the interparticle flux $g^{(0)}_{jk}$ is written through the function $f(x) = c (x-2r)^{-\left(p-\frac{d+1}{2}\right)}$ of one variable $x \geq 0$. The constant $c$ depends on $r$, $p$, $d$ and can be explicitly written by use of , . Th.\[thLB\] can be extended to the general space $\mathbb R^d$ as follows.
\[thLB2\] Consider the class $\mathcal D$ of densely packed balls in $\mathbb R^d$. The continuous energy (see also ) and the discrete energy for $p \in \mathbb N$ satisfying tend to $+\infty$ as $\delta \to +0$. Moreover, $\lambda$ can be approximated by $\sigma$ for sufficiently small $\delta$: $$\lambda=\sigma +O(\delta^0), \;\mbox{as} \; \delta \to 0.
\label{eq:sa1app}$$
Proof of the theorem repeats the proof of Th.\[thLB\] from [@LB] by its extension to the periodicity cell $Q_0$ in the space $\mathbb R^d$.
\[th3\] Consider the class $\mathcal D$ of densely packed balls in $\mathbb R^d$. Let $\sigma = \sigma(\mathbf a)$ attains the global minimum at a location $\mathbf a_*(\delta)$ for sufficiently small $\delta$. Then, the optimal packing is attained at $\mathbf a_*(0)$.
Proof. Let $\sigma(\mathbf a)$ attain the global minimum at $\mathbf a_*(\delta)$ for sufficiently small $\delta$ and $\phi_*$ denote the concentration of balls for the location $\mathbf a_*(0)$. The function $\sigma(\mathbf a)$ continuously depends on concentration [@LB], [@Kolpakov]. Hence, the function $\sigma(\mathbf a_*(\delta))$ is continuous in $0<\delta<\delta_0$ for sufficiently small $\delta_0$ and $\sigma(\mathbf a_*(0))=+\infty$.
Let the optimal packing be attained at another location $\mathbf a^*$ for which the concentration $\phi^*>\phi_*$. The location $\mathbf a^*$ contains a percolation chain. Take such a radius $r_0<r$ for which the concentration is reduced to $\phi_*$. Then, all the balls in this location $\mathbf a^*$ with the radius $r_0$ are separated from each other, hence the corresponding conductivity $\sigma(\mathbf a^*)$ is a finite number. But the minimal conductivity $\sigma(\mathbf a_*(\delta))$ tends to infinity as $\delta\to 0$. This yields a contradiction. The theorem is proved.
Extremal energy {#sec4}
===============
Let all the periodic Delaunay graphs $\Gamma$ with $n$ vertices are divided onto the equivalence classes of isomorphic graphs. A class of graphs will be denoted by $\mathcal G$. We are looking for the global minimum of the functional in $\mathbf t$ and $\mathbf a$ in the torus topology: $$\min_{\mathbf t, \mathbf a} E(\mathbf t, \mathbf a)= \mathop{{\sum}'}_{k,j}
|t_k-t_j|^p f(\|\mathbf a_k-\mathbf a_j\|)
\label{eq:sa4}$$ in a fixed class $\mathcal G$.
It is evident that the minimum exists since the continuous function $f(x)$ decreases and $0 \leq f(\|\mathbf a_k-\mathbf a_j\|) \leq +\infty$ for all $\|\mathbf a_k-\mathbf a_j\| \geq 2r$. The function $f(x)$ as a convex function for $2r \leq x <+\infty$ satisfies Jensen’s inequality $$\sum_{i=1}^M p_i f(x_i) \geq f \left(\sum_{i=1}^M p_i x_i \right),
\label{eq:sa5}$$ where the sum of positive numbers $p_i$ is equal to unity. Equality holds if and only if all $x_i$ are equal. Let the finite sum $\mathop{{\sum}'}_{k,j}$ in is arranged in such a way that $x_i=\|a_k-a_j\|$ and $p_i=\frac 1T(t_k-t_j)^p$, where $T=\mathop{{\sum}'}_{k,j} (t_k-t_j)^p$.
Application of yields $$\mathop{{\sum}'}_{k,j} (t_k-t_j)^p f\left(\|\mathbf a_k-\mathbf a_j\|\right)
\geq T f \left(\frac 1T \mathop{{\sum}'}_{k,j} (t_k-t_j)^p \|\mathbf a_k-\mathbf a_j\| \right).
\label{eq:sa6}$$ Hölder’s inequality states that for non-negative $a_i$ and $b_i$ $$\sum_{i=1}^M a_i b_i \leq \left(\sum_{i=1}^M a_i^2 \right)^{\frac 12} \left(\sum_{i=1}^M b_i^2 \right)^{\frac 12}.
\label{eq:sa7}$$ This implies that $$\mathop{{\sum}'}_{k,j} (t_k-t_j)^p \|\mathbf a_k-\mathbf a_j\| \leq \left[ \mathop{{\sum}'}_{k,j} (t_k-t_j)^{2p} \right]^{\frac 12} \left[ \mathop{{\sum}'}_{k,j} \|\mathbf a_k-\mathbf a_j\|^2 \right]^{\frac 12}.
\label{eq:sa8}$$ The function $f(x)$ decreases, hence and give $$\mathop{{\sum}'}_{k,j} (t_k-t_j)^p f\left(\|a_k-a_j\|\right)
\geq T f \left(\frac 1T \left[ \mathop{{\sum}'}_{k,j} (t_k-t_j)^{2p} \right]^{\frac 12} \left[ \mathop{{\sum}'}_{k,j} \|a_k-a_j\|^2 \right]^{\frac 12} \right).
\label{eq:sa9}$$ The minimum of the right hand part of on $\mathbf a = (\mathbf a_1,\mathbf a_2, \ldots, \mathbf a_n)$ is achieved independently on $t_k$ for $$\max_{\mathbf a} g(\mathbf a)=\mathop{{\sum}'}_{k,j} \|\mathbf a_k-\mathbf a_j\|^2.
\label{eq:sa10}$$
\[lemm1\] For any fixed $n$, any local maximizer of $g(\mathbf a)$ is the global maximizer which fulfils the system of linear algebraic equations $$\mathbf a_k=\frac 1{N_k} \sum_{j\in J_k} \mathbf a_j+\frac 1{N_k} \sum_{\ell=1}^d s_{j\ell}\boldsymbol{\nu}_{\ell},\quad k=1,2,\ldots,n,
\label{eq:sa11}$$ where $s_{j\ell}$ can take the values $0,\pm 1$ in accordance with the fixed Voronoi tessellation (the fixed graph $\Gamma$). The system has always a unique solution $\mathbf a = (\mathbf a_1,\mathbf a_2, \ldots, \mathbf a_n)$ up to an arbitrary additive constant vector.
Proof. It follows from the definition and the properties of the Voronoi tessellation that $$\|\mathbf a_k-\mathbf a_j\|=\left|\mathbf a_k-\mathbf a_j+\sum_{\ell=1}^d s_{j\ell}\boldsymbol{\nu}_{\ell}\right|
\label{eq:sam1a}$$ for some $s_{j\ell}$ which can take the values $0,\pm 1$. The extremal points of can be found from the system of equations $$\nabla_k g(\mathbf a)=0,
\quad k=1,2,\ldots,n,
\label{eq:sa12}$$ where $\mathbf a_k= (x_1^{(k)}, x_2^{(k)}, \ldots,x_d^{(k)})$ and $$\nabla_k = \left( \frac{\partial }{\partial x_1^{(k)}}, \frac{\partial }{\partial x_2^{(k)}}, \ldots, \frac{\partial }{\partial x_d^{(k)}}\right)$$ The parallelotope $Q_0$ is a compact manifold without boundary, hence all the extremal points of satisfy this system. Equations can be written in the equivalent form .
One can see that the sum of all equations gives an identity, hence, they are linearly dependent. Moreover, if $\mathbf a = (\mathbf a_1,\mathbf a_2, \ldots, \mathbf a_n)$ is a solution of , then $(\mathbf a_1+\mathbf c,\mathbf a_2+\mathbf c, \ldots, \mathbf a_n+\mathbf c)$ is also a solution of for any $\mathbf c \in \mathbb R^d$. Consider the homogeneous system corresponding to $$\mathbf a_k=\frac 1{N_k} \sum_{j\in J_k} \mathbf a_j, \quad
k=1,2,\ldots,n,
\label{eq:sa13}$$ where $\mathbf a_k$ belong to the cell $Q_0$. The system can be decoupled by coordinates onto independent systems $$x_p^{(k)}=\frac 1{N_k} \sum_{j\in J_k} x_p^{(j)},\;
k=1,2,\ldots,n \quad (p=1,2,\ldots, d).
\label{eq:sa14}$$ Each $p$th system has only constant solutions $$x_p^{(1)} = x_p^{(2)} = \cdots = x_p^{(n)}.
\label{eq:sa16}$$ This follows from the consideration of the quadratic form $$X_p=\frac 12 \mathop{{\sum}'}_{k,j} (x_p^{(k)}-x_p^{(j)})^2.
\label{eq:sa15}$$ It is symmetric and positive semi-definite. Therefore, the quadratic form has a global minimum attained at a linear set of $\mathbb R^n$. All the local minima of coincide with the global minimum. One can see that the quadratic form attains the global minimum at the set . Therefore, all solutions of , hence, of are only constants. Then, the system has only one condition of solvability for the right hand part which is fulfilled. Therefore, the system always has a unique solution up to an arbitrary additive constant.
The lemma is proved. We now proceed to summarize the obtained results concerning the geometrical optimal packing problem of equal spheres stated for periodic packing in $\mathbb R^d$. We consider the periodicity cell $Q_0$ represented by a parallelotope with glued opposite faces which contains an arbitrary number $n$ of spheres. The centers of spheres $\mathbf a = \{\mathbf a_1,\mathbf a_2, \ldots, \mathbf a_n \}$ lie in $Q_0$, satisfy the non-ovelapping restriction $\|\mathbf a_k -\mathbf a_j\| \geq 2r$ and the corresponding Delaunay graph $\Gamma$ belong to a fixed class of graphs $\mathcal G$. Lemma \[lemm1\] gives a bound for the packing problem in the fixed class $\mathcal G$. This bound yields the exact global minimum when the inequality becomes an equality. Let some of the differences $|t_k-t_j|$ vanish and the rest of the differences $|t_k-t_j|$ are equal. If for the equal differences $|t_k-t_j|$, the differences $\|\mathbf a_k -\mathbf a_j\|$ are also equal we have got this exact global minimum. Actually, such a situation frequently met, for instance for graphs corresponding to laminated lattices.
Consider the class of graphs $\mathcal A_2$ in $\mathbb R^2$ with $n=m^2$ ($m\in \mathbb N$) vertices containing the hexagonal lattice $A_2$. This class is determined by the kissing number $N_k=6$ for all the disks. The first step in the proof of optimality of the hexagonal lattice is the proof of equality $N_k=6$. Here, we refer to [@Toth]. Hence, if we restrict ourselves by the class $\mathcal A_2$, we do not go out a set of graphs containing the optimal packing.
The second step consists in the direct check that the regular hexagonal lattice satisfies . After, a simple external flux has to be applied and the corresponding $|t_k-t_j|$ and $\|\mathbf a_k -\mathbf a_j\|$ have to be calculated. Consider the regular hexagonal lattice generated by the vectors $$\boldsymbol{\nu}_1=\left(m,0 \right), \quad\boldsymbol{\nu}_2=\left(m\cos \frac{\pi}3, m\sin \frac{\pi}3 \right).
\label{eq:bas}$$ The radius of disks holds $r=\frac 12$. The periodicity cell $Q_0$ is determined by the vectors . It is convenient to consider $A_2$ as a laminated lattice with layers perpendicular to the axis $x_2$. Introduce the number of the layer $q \in \mathbb Z$, where the $q$th layer consists of the disks whose centers have the $x_2$-coordinate equal to $\frac{\sqrt{3}}2 q$ (see Figure 3).
![Hexagonal lattice as a laminated lattice with layers $q \in \mathbb Z$. The external flux is perpendicular to the axis $x_1$.[]{data-label="fig:Figure-3"}](Figure-3.eps){width=".7\textwidth"}
Let the external flux be directed along the axis $x_2$ and it is determined by the external potential $u_0(\mathbf x)=\frac{\sqrt{3}}2 x_2$. Then the continuous and discrete potentials take the value $q$ on all the disks of the $q$th layer. Therefore, the difference $|t_k-t_j|$ is equal to zero if $\mathbf a_k$ and $\mathbf a_j$ belong the same layer and it is equal to unity if $\mathbf a_k$ and $\mathbf a_j$ belong neighbor layers. The difference $\|\mathbf a_k -\mathbf a_j\|$ takes the same value $2r=1$ for touching disks from neighbor layers. The inequality becomes an equality in this case. It follows from Lemma \[lemm1\] that the global minimum on the class of graphs $\mathcal A_2$ is attains at the regular hexagonal graph. The optimal location $\mathbf a_*$ satisfies , hence depend on the basic vectors $\boldsymbol{\nu}_1$ and $\boldsymbol{\nu}_2$. If we take other basic vectors different from , we arrive at the optimal location $\mathbf a_*$ which is not macroscopically isotropic.
This example gives an alternative “physical” proof of the optimal 2D packing attained for the hexagonal array but the first step of the proof contains equation $N_k=6$ established by geometrical arguments.
\[ex4.3\] Consider the class of graphs $\mathcal Z^2$ in $\mathbb R^2$ with $n=m^2$ ($m\in \mathbb N$) vertices containing the square lattice $\mathbb Z^2$. This class is determined by the contact number $N_k=4$ for all the disks.
Consider the regular square lattice generated by the vectors $\boldsymbol{\nu}_1=\left(m,0 \right)$ and $\boldsymbol{\nu}_2=\left(0, m \right)$. The periodicity cell $Q_0$ is determined by these basic vectors. In accordance with the structural approximation theory the corresponding Delaunay graph[^3] consists of the edges parallel to the axes $x_1$ and $x_2$. The lattice $\mathbb R^2$ is laminated with layers perpendicular to the axis $x_1$. The $q$th layer consists of the disks whose centers have the $x_2$-coordinate equal to $q$. The external flux is determined by the external potential $-u_0(\mathbf x)= -x_2$. Then the continuous and discrete potentials take the value $q$ on all the disks of the $q$th layer. The further arguments repeat the previous example. Lemma \[lemm1\] implies that the global minimum on the class of graphs $\mathcal Z^2$ is attains at the regular square graph.
It is worth noting that the regular square lattice $\mathbb Z^2$ gives the optimal packing in the class $\mathcal Z^2$. One can see that the class $\mathcal Z^2$ can be considered as unstable. It contains only one element $\mathbb Z^2$, since any perturbation of a vertex yields flipping and changes the structure of $\mathcal Z^2$.
Consider the class of graphs $\mathcal A_3$ in $\mathbb R^3$ with $n=m^3$ ($m\in \mathbb N$) vertices containing the regular face-centered cubic lattice $A_3$. Since this lattice has the laminated structure, we can consider the external flux directed perpendicular to the layers. Then, the potential is constant in each layer and the differences $\|\mathbf a_k -\mathbf a_j\| =2r$ are constant where the neighbor points $\mathbf a_k$, $\mathbf a_j$ belong to the neighbor layers. This implies that the lattice $A_3$ reaches the optimal packing in the class $\mathcal A_3$. It is worth noting that the optimal location $\mathbf a_*$ in the class $\mathcal A_3$ depends on the basic vectors prescribed to the fcc lattice $A_3$ since $\mathbf a_*$ satisfies the system whose the right hand side depends on the basic vectors. The class of graphs containing the lattice corresponding to the hexagonal close-packing determines another optimal location, of course the hcp.
A class of graphs containing $A_3$ with arbitrary fixed basic vectors determines its optimal location $\mathbf a_*$ through the system . This location $\mathbf a_*$ depends on the basic vectors and must determine a macroscopically isotropic structure. The latter condition implies that the basic vectors have to correspond only to the fcc and hcp structures.
Perhaps, the kissing number $N_k=12$ determines the class containing the optimal packing. This argument could give an alternative proof of the optimal packing for the regular fcc lattice [@Hales].
These examples can be extended to the general space. Consider the class of graphs $\mathcal A_d$ in $\mathbb R^d$ containing a regular laminated lattice $\Lambda_d$. Then, the regular lattice $\Lambda_d$ yields the optimal packing in the class $\mathcal A_d$.
Discussion {#sec5}
==========
The most important fact used in this paper that the solution to the physical problem of the minimal energy (effective conductivity) implies solution to the geometrical problem of the optimal packing. Lemma \[lemm1\] yields upper bounds for the classical optimal packing problem. For laminated structures these bounds are exact, hence solve the optimal packing problems within the considered classes. Lemma \[lemm1\] yields not only upper bounds but an effective algorithm for their computations, since the linear system is decomposed onto $d$ independent systems with $n$ equations.
Therefore, we solve the optimal packing problem in any fixed class of graphs $\mathcal G$. Let us discuss these classes of graphs. Though the number of spheres per cell $n$ is arbitrary and can tend to infinity, the periodicity cell is fixed at the beginning, hence the optimal location $\mathbf a_*$ depends on the basic vectors. It follows also from the observation that the right hand part of the system contains the basic vectors. For sufficiently large $n$, the majority of equations of are homogeneous and only “boundary” equations are inhomogeneous. Their number is of order $\sqrt{n}$. It is interesting to investigate the asymptotic dependence of $\mathbf a_*$ on the basic elements as $n \to \infty$. This will show, for instance, the effectiveness of the hexagonal type packing in the cubic cell in $\mathbb R^d$.
We do not know a priory from the cell $Q_0$ and structure of $\mathcal G$ that the solution $\mathbf a_*$ of the system will form a macroscopically isotropic structure. However, a simple necessary condition of isotropy can be checked: the location $\mathbf a_*$ must contain $d$ percolation chains connecting the opposite face of the parallelotope $Q_0$ (see Fig.\[fig:Figure-2\]).
A straight forward verification which class $\mathcal G$ yields optimal packing can be proposed for a fixed $n$. First, for clarity consider 1D case. Then there exists exactly one class of graph $\mathcal Z$ including the regular 1D lattice $\mathbb Z$. A graph from $\mathcal Z$ consists of sequentially located edges along the axis. It follows from the system that each point $a_k$ lies in the middle of the segment $(a_{k-1}, a_{k+1})$ that immediately implies the points $a_k$ form the lattice $\mathbb Z$. The following extension of the above scheme to $\mathbb R^d$ can be proposed. Let $x_p^{(k)}$ denote the $p$th coordinate of the point $\mathbf a_k$ ($p=1,2,\ldots,d$). All the points $x_p^{(k)}$ ($k=1,2,\ldots,n$) are located on the real axis $\mathbb R$ (coincidence is permitted) and satisfy the system $$x_p^{(k)}=\frac 1{N_k} \sum_{j\in J_k} x_p^{(j)}+\frac 1{N_k} \sum_{\ell=1}^d s_{j\ell}\nu_p^{(\ell)},\quad
k=1,2,\ldots,n,
\label{eq:sa14b}$$ where $\nu_p^{(\ell)}$ denotes the $p$th coordinate of the basic vector $\boldsymbol{\nu}_{\ell}$. The class of graphs $\mathcal G$ determines the sets $J_k$, i.e., connections between the points $x_p^{(k)}$ and $s_{j\ell}$. In particular, $\mathcal G$ determines $N_k$ restricted by the kissing number. It is important to solve the system keeping the parameters $\nu_p^{(\ell)}$ in symbolic form. It can be done by separate numerical solutions to the systems $$x_p^{(k \ell)}=\frac 1{N_k} \sum_{j\in J_k} x_p^{(j \ell)}+ \frac{s_{j\ell}}{N_k} ,\quad
k=1,2,\ldots,n,
\label{eq:sa14c}$$ and by the linear combination $$x_p^{(k)}=\sum_{\ell=1}^d x_p^{(k \ell)}\nu_p^{(\ell)},\quad
k=1,2,\ldots,n.
\label{eq:sa14d}$$ All the computations can be made separately for $p=1,2,\ldots,d$. As a result we obtain expressions for the optimal $\mathbf a_k$ in the class $\mathcal G$ $$\mathbf a_k=\sum_{\ell=1}^d \mathbf a_k^{(\ell)}\boldsymbol{\nu}_{\ell},\quad
k=1,2,\ldots,n,
\label{eq:sa14e}$$ where the vectors $\mathbf a_k^{(\ell)}$ are given numerically. The next numerical problem consists in determinations of such bases $\{\boldsymbol{\nu}_{\ell}\}_{\ell=1}^d$ which possess $d$ percolation chains connecting the opposite faces of the parallelotope $Q_0$. Ultimately, one has to perform this procedure for all the classes of graphs and take the best packing. The above algorithm is rather an idea how to apply solution to the packing problem within classes $\mathcal G$ to the traditional sphere packing [@CS], [@Hales].
The physical problem is considered in the class of periodic structures. General non-periodic composites were discussed in literature (see a review in [@Kolpakov] and references therein). The main question consists in extension of the structural approximation theory [@LB], [@Kolpakov] developed for densely packed composites to the non-periodic case. We do not discuss the case of different radii. We suppose that a similar work can be done, since simple formulae for $g^{(0)}_{jk}$ are known [@LB], [@Kolpakov] and can be extended to $\mathbb R^d$ following Sec.\[sec3\].
Bakhvalov, N.S., Panasenko, G.P.: Homogenization: Averaging processes in periodic media. Nauka, Moscow (1984) \[in Russian\]; English transl. Kluwer, Dordrecht /Boston /London (1989)
Berlyand, L., Kolpakov, A.G., Novikov, A.: Introduction to the Network Approximation. Method for Materials Modeling. Cambridge University Press, Cambridge (2013) Bezdek, K.: Lectures on Sphere Arrangements - the Discrete Geometric Side. Springer, New York etc (2013).
Bezdek, K., Deza, A., Ye, Yinyu (Eds.): Discrete Geometry and Optimization. Springer, New York etc (2013). Conway, J., Sloane, N.J.A.: Sphere Packings, Lattices and Groups. Springer, New York etc (1999)
Keller, J.B.: Conductivity of a Medium Containing a Dense Array of Perfectly Conducting Spheres or Cylinders or Nonconducting Cylinders. J. Appl. Phys. 34, 991-993 (1963)
Lagarias, J.C. (Ed.): The Kepler Conjecture: The Hales-Ferguson Proof. Springer, New York etc (2011)
Kolpakov, A.A., Kolpakov, A.G.: Capacity and Transport in Contrast Composite Structures: Asymptotic Analysis and Applications. CRC Press, Boca Raton (2010)
Mityushev V., Rylko N.: Optimal distribution of the non-overlapping conducting disks. Multiscale Model. Simul. 10, 180-190 (2012)
Rylko N.: Structure of the scalar field around unidirectional circular cylinders. Proc. R. Soc. A464, 391-407 (2008)
Tóth, L. Fejes: Lagerungen in der Ebene auf der Kugel und im Raum. Springer-Verlag, Berlin, New York (1953)
[^1]: In the percolation theory, a percolation chain of balls is usually defined as a chain of overlapping open balls.
[^2]: Our formulae , and slightly diverge with the corresponding formulae from [@LB] (see for instance (6.2.13)) by multipliers. It can be related to the error power $p$ taken in [@LB] (e.g. (6.4.15) from [@LB] for 2D and 3D cases) instead of the correct $p-1$ from .
[^3]: It does not formally produce a Delaunay triangulation in the commonly used sense.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Observations of the PSRB1259$-$63/SS2883 binary system using the CANGAROO-II Cherenkov telescope are reported. This nearby binary consists of a 48msec radio pulsar in a highly eccentric orbit around a Be star, and offers a unique laboratory to investigate the interactions between the outflows of the pulsar and Be star at various distances. It has been pointed out that the relativistic pulsar wind and the dense mass outflow of the Be star may result in the emission of gamma rays up to TeV energies. We have observed the binary in 2000 and 2001, $\sim$47 and $\sim$157 days after the October 2000 periastron. Upper limits at the 0.13–0.54 Crab level are obtained. A new model calculation for high-energy gamma-ray emission from the Be star outflow is introduced and the estimated gamma-ray flux considering Bremsstrahlung, inverse Compton scattering, and the decay of neutral pions produced in proton-proton interactions, is found to be comparable to the upper limits of these observations. Comparing our results with these model calculations, the mass-outflow parameters of the Be star are constrained.'
author:
- 'A. Kawachi, T. Naito, J.R. Patterson, P.G. Edwards, A. Asahara, G.V. Bicknell, R.W. Clay, R. Enomoto, S. Gunji, S. Hara, T. Hara, T. Hattori, Sei. Hayashi, Shin. Hayashi, C. Itoh, S. Kabuki, F. Kajino, H. Katagiri, T. Kifune, L. Ksenofontov, H. Kubo, J. Kushida, Y. Matsubara, Y. Mizumoto, M. Mori, H. Moro, H. Muraishi, Y. Muraki, T. Nakase, D. Nishida, K. Nishijima, M. Ohishi, K. Okumura, R.J. Protheroe, K. Sakurazawa, D.L. Swaby, T. Tanimori, F. Tokanai, K. Tsuchiya, H. Tsunoo, T. Uchida, A. Watanabe, S. Watanabe, S. Yanagita, T. Yoshida, and T. Yoshikoshi'
title: 'A Search for TeV Gamma-ray Emission from the PSR B1259$-$63/SS2883 Binary System with the CANGAROO-II 10-m Telescope'
---
Introduction {#sect:intro}
============
PSR B1259$-$63 (($\alpha,\delta$)(J2000) $=$ (13$^h$02$^m$47$^s$.68, $-$635008.6)) is a 48msec radio pulsar discovered in a 1500MHz radio survey of the southern Galactic plane[@johnston92a] which was subsequently found to be in a highly eccentric orbit with a 10th magnitude main-sequence star, SS2883[@johnston92b; @johnston94]. With an orbital eccentricity of 0.87, the separation of the stars varies in the range 0.97$\sim$14.0$\times {\rm 10}^{13}$cm during the orbital period of 1236.72 days. The periastron epoch is MJD 48124.35[@wex98]. SS2883 is of spectral type B2e[@johnston94], with a mass $M_*$ of $\sim$10$M_{\sun}$ and a radius $R_*$ of $\sim$6$R_{\sun}$. The luminosity and radius of the B2e star correspond to an effective temperature $T_{\rm eff}$ of $\sim$27,000 K at the star surface[@tavani97]. Its characteristic emission disc extends to at least 20$R_*$, similar to the distance between the pulsar and the Be star at periastron. Here we assume a distance of 1.5kpc to the binary system, which has been estimated from optical photometric observations of SS2883[@johnston94].
The periastron passages have been closely observed at radio frequencies [@johnston99; @connors02]. No pulsed emission was detected for about five weeks centered on periastron and the pulsed emission was depolarized for $\sim$200 days also centered on periastron. Timing measurements have shown that the disc of the Be star is likely to be inclined with respect to the orbital plane [@melatos95; @wex98], which has been suggested in @kaspi95 [@tavani97]. In @connors02, the unpulsed light curves are discussed with an assumption of two short-time crossings of the pulsar and the disc, before \[($\tau - $18 d) $\sim$ ($\tau - $8 d)\] and after \[($\tau +$12 d) $\sim$ ($\tau +$22 d)\] periastron ($\tau$).
A weak X-ray signal was first detected by [*ROSAT*]{} which observed the system just after apastron in September 1992[@cominsky94]. Through 1994–1996, unpulsed X-ray emission with a single power-law spectrum was detected at the six different orbital phases observed by [*ASCA*]{} [@hirayama96; @hirayama99]. The photon index of the X-ray spectrum is about $-$1.6 in the post-periastron to apastron period, steepening towards periastron where the steepest index of $-$1.96 was observed. The 1–10 keV band luminosity varies by about an order of magnitude, from $\sim$10$^{34}$ ergs s$^{-1}$ around periastron to $\sim$10$^{33}$ ergs s$^{-1}$ at apastron. The maximum luminosity was detected at $\tau -$12 d, with the intensity decreasing at periastron, then increasing again. The column density was low and constant ($\rm{6}\times\rm{10}^{21}\rm{cm}^{-2}$) at all orbital phases. The periastron passage in January 1994 was monitored by a multi-wavelength campaign including observations in the X-ray and gamma-ray bands with [*ROSAT*]{}, [*ASCA*]{} and [*CGRO*]{} [@grove95]. The power-law spectrum (photon index $\sim -$2.0) extended to the 200keV energy region of [*OSSE*]{}, with no pulsations being detected. No emission in the energy range of 1MeV–3GeV was detected down to the observational limits. [*OSSE*]{} failed to detect signals at the apastron passage in 1996, however, its upper limit does not conflict with the extrapolation of the [*ASCA*]{} spectrum[@hirayama99]. Several TeV observations of the binary system were performed in 1994 and 1997 using the CANGAROO 3.8-m ground-based Cherenkov telescope, resulting in a marginally significant suggestion of gamma-ray signals [@sako97].
The multi-wavelength spectrum from the 1994 periastron strongly implies that the hard X-ray emission up to 200keV originates from synchrotron radiation of non-thermal electrons[@tavanikaspi94]. Electrons released in the pulsar wind may be accelerated in a shock wave generated in the region where the relativistic pulsar wind interacts with dense mass flow from the Be star. Adjusting the pressure balance between the flows, @tavani97 have interpreted the measured hard X-ray spectrum on the basis of accelerated particles in the pulsar-side shock, using an approximated approach to the Klein-Nishina effect for emission. They conclude that the energy loss of electrons due to inverse Compton scattering is dominant, and that the Lorentz factor of accelerated electrons is $\Gamma_{\rm e} = {\rm 10}^6-{\rm 10}^7$. Accretion onto the neutron star is unlikely to be significant, as there is an absence of X-ray/gamma-ray pulsations, an absence of the day–scale fluctuations in X-rays, a relatively low X-ray luminosity, and negligible absorption as a result of the low column density. The consistency of the X-ray luminosities at the same orbital phases in different years supports the idea that the observed time variability is due to binary modulation [@hirayama99; @kaspi97]. Recently, another model for the pulsar-side shock considering the Klein-Nishina effect in the emission and cooling process has been proposed. @shibazaki02 note that the inverse Compton cooling dominated spectrum is flatter, since the Klein-Nishina effect suppresses the cooling of higher-energy electrons. It is argued that synchrotron cooling, instead of the inverse Compton scattering discussed by @tavani97, is the dominant process for the energy loss of electrons in the pulsar wind on the account of the steepening of the X-ray spectral index observed around the 1994 periastron.
The light curves of the radio unpulsed emission around periastron have been recently modeled by the adiabatic expansion of synchrotron bubbles formed in the pulsar and the Be star disc interaction [@connors02]. They assume short-time interactions of the pulsar and the disc, as the pulsar should cross the disc twice in the orbital period. When the pulsar enters the disc, electrons are accelerated in the contact surface of the pulsar wind and the disc material, but after the pulsar leaves the disc, the pulsar-wind bubble remains behind, moves in the disc-flow, and decays through synchrotron losses. The model successfully explains the radio data, however, it does not appear to describe the X-ray data well; for example, the weak unpulsed emissions in the X-ray region was observed after these modeled bubbles should have decayed by adiabatic expansion as moving outwards, and, the constant spectral index in the radio region is inconsistent with the steepening observed in the X-ray spectrum.
Electrons accelerated in a pulsar-side shock to Lorentz factors of $\Gamma_{\rm e} ~\gtrsim \rm{10}^6$ in the radiative environment of the binary system may produce high energy gamma rays. The suggestion that detectable levels of gamma-ray emission may arise in the shocked pulsar wind via inverse Compton scattering [@kirkball99] provided the initial motivation for the observations described here. Subsequently, inverse Compton emission from the un-shocked region of the pulsar wind has been considered [@ballkirk00; @balldodd01]. The integrated contribution from the un-shocked pulsar wind may increase the gamma-ray flux around periastron for some conditions. The maximum level of emission in the TeV energy range is estimated to be $\sim{\rm 4} \times {\rm 10}^{-5}{\rm MeV cm}^{-2}{\rm s}^{-1}$ in the integrated energy flux with a wind Lorentz factor of 10$^7$[@ballkirk00] which may raise the TeV gamma-ray flux above our detector’s sensitivity of typically ${\rm 10}^{-11}-{\rm 10}^{-12}{\rm TeV cm}^{-2}{\rm s}^{-1}$. Further studies have considered the effect of the termination of the wind [@balldodd01], which, depending on its assumped location, may act to decrease the inverse Compton flux compared to previous predictions. On the other hand, @shibazaki02 predict a maximum energy flux from the pulsar wind of $\sim {\rm 10}^{-13} {\rm erg \,s}^{-1}{\rm cm}^{-2}$ at TeV energies around periastron, which would require a very deep observation to detect.
At the contact surface of the pulsar and Be star flows, ions and electrons in the Be star outflow may be accelerated to high energies via the first-order Fermi mechanism. In the dense outflow from the Be star there is a lot of target material for proton-proton interactions and Bremsstrahlung emission. The Be star also provides target photons for upscattering by the inverse Compton mechanism, in addition to the 2.7K microwave background radiation. The densities of these targets increase as the contact surface gets closer to the Be star, and so does the total energy of accelerated particles. In this paper, a new model calculation for gamma-ray emission from the accelerated particles in the Be star outflow, taking into consideration Bremsstrahlung, the inverse Compton mechanism, and proton-proton interactions, is applied to the binary system and discussed along with our observational results.
Observations
============
The CANGAROO ($\underline{C}$ollaboration between $\underline{A}$ustralia and $\underline{N}$ippon for a $\underline{GA}$mma-$\underline{R}$ay $\underline{O}$bservatory in the $\underline{O}$utback) Cherenkov telescopes are located near Woomera, South Australia (13647E, 3106S, 160m a.s.l.). The 3.8-m CANGAROO telescope, used for the previous observations of PSRB1259$-$63, was operated from 1991 to 1998. The CANGAROO-II telescope was constructed in 1999 initially with a 7-m diameter dish, which was upgraded to 10-m in 2000[@mori01]. Cherenkov photons from extensive air showers which are initiated by primary gamma-rays/cosmic rays are collected with a parabolic reflector and detected by an imaging camera placed in the prime focal plane. The 10-m reflector consists of 114 spherical plastic mirrors of 0.8m diameter[@kawachi01] to make a composite parabolic shape ($f$/0.8). The camera has an array of 552 photomultiplier tubes (PMTs) of half-inch diameter (Hamamatsu, R4124UV) covering a field-of-view of 2.8, with groups of 16 PMTs using the same high voltage line. In the observations described in this paper, accepted events were required to meet two conditions within the central region of the camera (1$\arcdeg$ diameter); at least three PMTs above a $\sim$3-photoelectron threshold, and at least one preamplifier unit with an analogue sum of signals above a $\sim$10-photoelectron threshold. The typical raw trigger rates of these observations were in the 10–80 Hz range. The charge and timing information of each PMT signal was recorded for each event. Details of the telescope are given in @mori01 and @tanimori_NIM.
The PSRB1259$-$63/SS2883 system was observed with the CANGAROO-II Cherenkov telescopes at two different orbital phases; for several days in December 2000 (hereafter $Obs.\,A$; MJD 51881.4 in average of the observation time) and in March 2001 ($Obs.\,B$; MJD 51991.5), about 47 and 157 days after the periastron of October 2000 (MJD 51834.51). The orbital phase of $Obs.\,A$ is similar to the [*ASCA*]{} Obs.4 in @hirayama99. We have no data closer to periastron as observing conditions were not suitable during June–November 2000. The orbital phases of the observations are schematically shown in Figure\[fig:obs\_orbit\], and details of the observations are summarized in Table\[table:obs\_summary\]. In all observations the telescope was pointed so that the binary system was at the tracking center.
The target and offset region(s) were observed for equal amounts of time on each night under moonless and fine sky conditions. In a typical observation, the target (ON-source) region was observed for a few hours including source culmination, and an offset (OFF-source) run was carried out before and/or after the ON-source run to cover the same track, in the elevation and azimuthal angles, as that of the ON-source run. The average zenith angles were 58$\arcdeg$ for $Obs.\,A$ and 34$\arcdeg$ for $Obs.\,B$, corresponding to differences in observing seasons. The observations closer to periastron, $Obs.\,A$, were performed at larger zenith angles, thus at a higher energy threshold, than the optimum observing condition for the source. The zenith angle of 58$\arcdeg$ is similar to the observing conditions for the Crab nebula ($\sim$55$\arcdeg$ in @tanimori_crab) from the CANGAROO site, thus the data has been analyzed in a similar manner to the recent analysis of the Crab data [@itoh03].
Analysis and Results
====================
Data Analysis
-------------
The digitized counts of PMT signal charges have been calibrated after pedestal subtraction. The gains of the pixels have been normalized using a blue LED located at the center of the telescope to illuminate the camera uniformly. The LED is driven by a fast pulse generator. Accidental events, caused mainly due to the night sky or environmental background light, have been removed in the analysis by requiring at least five neighboring PMTs to exceed the $\sim$3.3 photoelectron threshold within $\pm$35nsec. A small number of pixels affected by noise or the passage of bright stars have been excluded from the analysis. The count rate of events which satisfy these criteria is stable (about 0.6Hz and 1.9Hz in $Obs.\,A$ and $Obs.\,B$ data sets, respectively), and differences in the raw event rates between ON- and OFF-source runs, ratios of about 1:4 in $Obs.\,A$ and 3:2 in $Obs.\,B$, respectively, have been resolved. Data under cloudy or unstable conditions have been rejected by checking deviations of the counting rate. The effective observation time and event numbers at this stage of the analysis are summarized in Table\[table:data\_summary\].
In order to discriminate gamma-ray events from cosmic-ray induced events, a likelihood analysis has been applied [@enomoto_sim2; @itoh03] to the characteristic light images recorded by the camera. In this image-analysis, we also require the total charge contained in an image to be greater than $\sim$35 photoelectrons, in order to improve the efficiency of background rejection. We have used Monte Carlo simulations [@enomoto_sim1] for a gamma-ray ($\gamma$) event set and the OFF-source events for the background ($BG$) event set. The average zenith angle, the discarded pixels, and the cut parameters in the analysis have been taken into account in the simulations for the different conditions of $Obs.\,A$ and $Obs.\,B$. The following results are based on simulations assuming a power-law energy spectrum with an index of $-$2.5 for the generated gamma-rays. The “hit-map” of triggered pixels, weighted by size of the signal, has been fitted with an ellipse parametrized by the r.m.s. spread of light along the minor/major axis of the image ($width/length$), and the distance between the image centroid and the source position ($distance$) [@hillas82] to characterize each event. Images truncated by the camera edge, or too concentrated at the camera center, have been omitted by putting a loose limit on $distance$ (0.35$\le distance \le$ 1.2). Each of the image parameters have been plotted against the total signal in the image for the $\gamma$ and $BG$ event sets to produce probability density functions (PDF) of the “$\gamma$-like” and “$BG$-like” events, respectively. Thus the dependences of the image parameters on the event size have been taken into the analysis. Finally, a single parameter, $R_{prob} = Prob(\gamma)/[Prob(\gamma)+Prob(BG)]$, is calculated for each event, where $Prob(\gamma)$\[$Prob(BG)$\] is the probability of the event being due to a $\gamma$ \[$BG$\], calculated from the two-dimensional PDFs. We use the PDFs of three parameters with equal weight. A selection criteria ($R =$0.4; $R_{prob} \ge R$) has been chosen considering the acceptance of $\gamma$ events and the figure-of-merit of $\gamma$ events to the $BG$ events. The event numbers of the selected data are listed in Table\[table:data\_summary\].
After the image selection, gamma-rays from the pulsar should have the image orientation parameter $alpha$ [@punch92] less than 20$\arcdeg$ for the $Obs.\,A$ data set, and 15$\arcdeg$ for the $Obs.\,B$ set. A broader $alpha$ distribution for gamma-rays at larger zenith angle is expected from simulations[@okumura_mrk421]. Figure\[fig:alpha\_distribution\] shows the distributions of $alpha$ after all the other cuts have been applied. The OFF-source (“OFF") $alpha$ distributions are normalized to that of the ON-source (“ON") by the number of events with $alpha \ge 40\arcdeg$. The normalization factor is consistent with that deduced from the effective observation times, within statistical errors. No statistically significant excess of the “ON" over the “OFF" is seen in either of the $alpha$ histograms. Subtracting the normalized “OFF" counts from the “ON" data results in 31 and 47 events within the $alpha$ selection criteria, corresponding to the significances (assuming Poisson fluctuations only) of $+$1.0$\sigma$ and $+$0.60$\sigma$, for $Obs.\,A$ and $Obs.\,B$ respectively.
Energy Threshold and Upper Limit Flux
-------------------------------------
The gamma-ray acceptance for the cuts used in the analysis has been estimated based on the simulations. The energy threshold is defined as the peak of the acceptance multiplied by the generated energy spectrum, and thresholds of 3.6TeV and 0.78TeV are derived for a $E^{-2.5}$ spectrum for the $Obs.\,A$ and $Obs.\,B$ data sets, respectively. The corresponding effective areas are 3.6$\times$10$^9$ cm$^{2}$ and 1.3$\times$10$^9$ cm$^{2}$. The 2$\sigma$ upper limits of our result are
> $F(\geq 3.6TeV) \leq 1.4\times10^{-12}{\rm cm}^{-2}{\rm s}^{-1}$ ($Obs.\,A$)
and
> $F(\geq 0.78TeV) \leq 3.1\times10^{-12}{\rm cm}^{-2}{\rm s}^{-1}$ ($Obs.\,B$).
Figure\[fig:integral\_flux\] shows the upper limits of the two observations. The integral flux of the Crab nebula is also shown [@tanimori_crab] as a reference. The systematic error in the energy scale determination has been estimated to be 15 percent[@itoh03] and is shown in the figure as errors in the abscissa.
The whole procedure of the analysis has been performed changing the simulated spectral index from $-$2.5 to $-$2.0. The significance levels of the gamma-ray signals are unchanged, and the corresponding threshold energies and the gamma-ray acceptances increase by about 20 percent. Thus the upper limits assuming a spectral index of $-$2.0 are; $F(\geq 4.0TeV) \leq 1.2\times10^{-12}{\rm cm}^{-2}{\rm s}^{-1}$ ($Obs.\,A$) and $F(\geq 0.88TeV) \leq 2.7\times10^{-12}{\rm cm}^{-2}{\rm s}^{-1}$ ($Obs.\,B$), respectively, as shown in Figure\[fig:integral\_flux\].
Discussion
==========
The observational results are compared with some model calculations. A new model of gamma-ray emissivity is introduced, considering the particles accelerated in the Be star outflow.
Models of the two flows
-----------------------
Figure\[fig:binary\_system\_image\] schematically illustrates the assumed configuration of the system: the pulsar and its relativistic pulsar wind, the Be star and its polar and disc-like outflows, and the shock composed of three surfaces: pulsar-side shock, contact surface, and Be-star-side shock. Particles are assumed to be accelerated by the shock at the pressure balance between the flows of the two stars. In the figure, the contact discontinuity between the pulsar wind and the equatorial disc of the Be star is illustrated. The alignment of the Be star disc to the orbital plane, and its effect, will be discussed later in calculating light curves over orbital phase.
For the pulsar wind, we adopt the model of @kennel84 for the synchrotron nebula around the Crab pulsar. For the Be star mass-flow, the simple model of @waters86 is used, which represents radiations from Be stars using the IR, optical, and UV observational results. The parameters are chosen so as to be consistent with the observational results of Be stars in general [@cote87] and of the PSR B1259$-$63/SS2883 binary . We fully consider the Klein-Nishina effect in the calculations of the emission processes via electrons. Provided the pulsar wind is driven by the spin down luminosity ($\dot{E}_{\rm rot}$) of the pulsar, a fraction ($f_{\rm pw} =$ 0.1) of the wind luminosity is assumed to be enhanced in the equatorial plane. Both kinetic and electromagnetic energies are included in $\dot{E}_{\rm rot}$ [@kennel84]. The radial distribution of the wind pressure, $P_{\rm pw}$, is given by $$\label{eq:p_pw}
P_{\rm pw}(r) = \frac{\dot{E}_{\rm rot}}{f_{\rm pw}4\pi r^2 c}~~,$$ where $r$ is the distance from the pulsar and $c$ is the speed of light.
For the mass-flow of the Be star, we consider a high-density, slow, equatorially orbiting disc-like flow [@waters86], and a low-density, fast, polar component (stellar wind) [@waters88; @dougherty94] as well. The density profile, $\rho$, is assumed to depend on the distance from the center of the Be star, $R$, as $\rho(R) = \rho_0 (R/R_*)^{-n}$ with a power-law index $n$, where $R_*$ is the star radius and $\rho_0$ is the density of the outflow at the surface of the star. The flow speed $v(R) = v_0(R/R_*)^{n-2}$ is obtained from conservation of mass flux, where $v_0$ is speed of the outflow at the surface of the star. Then the momentum flux of the flow, $P_{\rm Be}$, is $$\label{eq:p_be}
P_{\rm Be}(R) = \rho v^2 = \rho_0 v_0^2 (\frac{R}{R_*})^{n-4}.$$ In our calculation, indices $n$ of 2.5 and 2 are chosen in outflows of disc and polar wind, respectively.
The location of the shock regime is determined by the balance between pressures of the pulsar wind (Eq.\[eq:p\_pw\]) and of the Be star outflow (Eq.\[eq:p\_be\]).
We introduce a new parameter, $x$, defined as $$x = \frac{\rho_0}{{\rm 10}^{-12} {\rm g cm}^{-3}}\frac{v_0}
{{\rm 10}^6 {\rm cm s}^{-1}}.$$ When $x$ is larger, the location of the pressure balance becomes further from the Be star. If we assume that the opening angle $\theta_{\rm disc}$ of the disc outflow is 15[@johnston96], the parameter $x$ is related to the $\Upsilon$ of @tavani97 by $\Upsilon
\equiv (\dot{M}/{\rm 10}^{-8}{M}_{\sun}\ {\rm yr}^{-1})\
(v_{0}/{\rm 10}^6 {\rm cm\ s}^{-1})
= {\rm 0.90} \times x
\times (v_{0}/{\rm 10}^6 {\rm cm\ s}^{-1})$, where $\dot{M}$ is the mass loss rate. The parameter $x$ depends on $v_0$ and $\rho_0$, which are obtained directly from UV/optical observations, independent of the disc opening angle. As shown later, the gamma-ray emission is approximately proportional to $x^2$ in our model.
Particle acceleration and gamma-ray spectrum {#sect:discussion-2}
--------------------------------------------
First, we deduce the flux $j_{\rm i}$ (i $=$ e,p) as a function of energy $E_{\rm i}$ of the particles in the Be star outflow on the basis of Fermi acceleration, where $e$ denotes electrons and $p$ denotes protons. For simplicity, we assume that all ions are protons. Secondly, we calculate the energy flux from the binary system induced from the emission mechanisms of Bremsstrahlung, inverse Compton (electrons) and proton-proton collisions.
In general terms, the momentum spectrum $dN/dP$ of the shock-accelerated particles is expressed as $N_{0} P^{-\alpha}$ and 4$\pi j$ is obtained from $v dN/dE_{\rm k}$, where $E_{\rm k}$ is the kinetic energy, described as $dN/dE_{\rm k} = dP/dE_{\rm k} dN/dP$. The constant $N_{\rm 0,i}$ is evaluated with the following integration regarding of the energy balance at the shock location in the Be star flow; $$\label{eq:energy_balance-general}
\int^{E^{\rm max_{\rm i}}}_{m_{\rm i} c^2}{\frac{dN_{\rm i}}{dE_{\rm i}}
(E_{\rm i}) dE_{\rm i}}
= f_{\rm acc, i} P_{\rm Be} (R_{\rm shock}), ({\rm i} = {\rm e,p}),$$ where $R_{\rm shock}$ is the distance to the contact surface from the Be star center. We assume $f_{\rm acc, i} = $ 0.001 and 0.1 is assumed for $i = e$ and $p$, respectively, as the efficiency of the acceleration to be consistent with an e/p ratio in cosmic ray observations (e.g. @mullerproc [@baring99]). The variation of $R_{\rm shock}$ causes the orbital modulation in the light curve. The orbital inclination to the line of sight has not been included in our calculation, as this effect is less significant in the Be star emission models than in the pulsar wind emission models. Anisotropy of the optical photons from the Be star is neglected for simplicity. Assuming Fermi acceleration, a power-law index of $\alpha = -$2.0 is taken for the proton momentum spectrum $dN_{\rm p}/dP_{\rm p}$ with an assumed compression ratio of 4.0. The spectral index of the electron momentum spectrum at the shock front does not vary much from the canonical $\alpha = -$2.0 for plausible values of pulsar wind parameters, because inverse Compton cooling in the higher energy electrons are reduced by the Klein-Nishina effect [@shibazaki02]. In addition, synchrotron cooling does not affect the spectral index, since the magnetic-field strength on the Be-star side should be weak. We therefore assume that the electron spectral index has a constant value of $\alpha = -$2.0.
The integration is performed from the threshold energy or the particle mass, $m_{\rm e,p} c^2$, to the maximum energy of the accelerated particle $E^{\rm max}_{\rm e,p}$, which we assume here to be $\sim$10$^{15}$ eV. Applying the obtained $j_{\rm e,p}(E_{\rm e,p})$, the gamma-ray spectrum from the source at the distance $D$ is calculated. For the proton-proton collision emission mechanisms, the spectrum is calculated as $$\label{eq:gamma_flux-protons}
F_{\gamma}^{\rm pp}(E_{\gamma}) = \frac{1}{D^2}\int{n_{\rm target} dV}
\int\int{dE_\pi dE_{\rm p}} \frac{2}{p_\pi} j_{\rm p}(E_{\rm p})
\frac{d\sigma_{\rm pp\rightarrow\pi}(E_{\pi},E_{\rm p})}{dE_{\pi}}$$ where $n_{\rm target}$ stands for $\rho/m_{\rm p}$, and $E_{\rm p,\pi}$ and $p_{\rm p,\pi}$ denote the energy and momentum of protons or pions, respectively. Full descriptions of the integral limit and $\sigma_{\rm pp\rightarrow\pi}$ are given in @naitotakahara94. The contributions of the inverse Compton (IC) and Bremsstrahlung are calculated from $j_{\rm e,p}(E_{\rm e,p})$ as $$\label{eq:gamma_flux-electrons}
F_{\gamma}^{\rm IC, Brem}(E_{\gamma}) = \frac{1}{D^2}\int{n_{\rm target} dV}
\int^{E^{\rm max}_{\rm e}}_{m_{\rm e} c^2} dE_{\rm e} j_{\rm e}(E_{\rm e})
\frac{d\sigma}{dE_{\gamma}},$$ where $n_{\rm target} = n_{\rm photon}$ and $\frac{d\sigma}{dE_{\gamma}}$ is a cross section which includes the Klein-Nishina effect for inverse Compton emission, and $n_{\rm target}$ of $\rho/m_{\rm p}$ and the cross section $\frac{d\sigma}{dE_{\gamma}}$ of electron-proton and electron-electron interaction are used for Bremsstrahlung emission [@gaisser98; @sturner97]. For $n_{\rm photon}$, we adopt 2.7 K CMB and $T_{\rm eff} =$ 27,000 K black body radiation from the Be star. In the spatial integration, we assume that the accelerated particles extend into the Be star outflow downstream of the shock. The contributions of different emission mechanisms are calculated with $x_{\rm disc} =$1500 for the phase of periastron and their differential energy spectra are shown in Fig.\[fig:spectrum\_comparison\] (the disc and the pulsar wind are assumed to interact at periastron in the calculation). The total gamma-ray flux is deduced as $F_{\gamma}(E_{\gamma}) = F_{\gamma}^{\rm Brem}(E_{\gamma}) +
F_{\gamma}^{\rm IC}(E_{\gamma}) + F_{\gamma}^{\rm pp}(E_{\gamma})$ and the dominant contribution is of $F_{\gamma}^{\rm pp}(E_{\gamma})$. The inverse Compton flux in the sub-TeV energy region, expected from the pulsar-wind side[@shibazaki02], is comparable to $F_{\gamma}^{\rm IC}$ from the Be star outflows, except that the former has a break $\sim$400 GeV due to the stronger magnetic field in the pulsar wind side. After the spatial integration, the total flux is approximately expressed as $$\label{eq:gamma_flux-propt}
F_{\gamma}(E_{\gamma}, x)
\propto x^2 \frac{1}{n-1} \frac{1}{R_{\rm shock}(x,n,v_0)}~~.$$ $R_{\rm shock}(x,n,v_0)$ for the same orbital phase does not vary much within the parameter range discussed in the following.
The adopted model parameters are summarized in Table\[table:model\_parameters\]. Now we discuss the possible ranges of two parameters, $x$ and the density profile index $n$. For the polar component, the value $x_{\rm polar}$ is set to be proportional to the disc component, $x_{\rm disc}$. The factor is estimated using the following two equations, $\dot{M}_{\rm polar}
=4 \pi R_*^2~\rho_{\rm polar, 0}~v_{\rm polar, 0}~(1-\sin \theta_{\rm disc})$ and $\dot{M}_{\rm disc}
=4 \pi R_*^2~\rho_{\rm disc, 0}~v_{\rm disc, 0}~\sin \theta_{\rm disc}$, where the surface density, initial velocity, and mass loss rate of the disc \[polar wind\] flow are denoted as $\rho_{\rm disc[polar], 0}, v_{\rm disc[polar], 0},$ and $\dot{M}_{\rm disc[polar]}$, respectively, and $\theta_{\rm disc}$ denotes the opening angle of disc outflow. The ratio of two $x$ parameters is $$\frac{x_{\rm polar}}{x_{\rm disc}} =
3.49 \times 10^{-1}
\frac{\dot{M}_{\rm polar}}{\dot{M}_{\rm disc}},$$ assuming $\theta_{\rm disc}$ = 15[@johnston96]. From the observed intensities of the UV line (due to the polar wind) and of the IR radiation (from the disc), @lamers87 deduce the mass loss ratio of the two flow components as $\frac{\dot{M}_{\rm polar}}{\dot{M}_{\rm disc}}$ of 10$^{-1}$–10$^{-4}$. We take $x_{\rm polar}$ of 10$^{-1} \times x_{\rm disc}$ as a rather optimistic value. The thick disc-like flow is an effective site for the production of gamma-ray emission, and makes the dominant contribution to the total intensity.
For $x_{\rm disc}$, early studies [@waters86; @waters88; @dougherty94] have estimated possible ranges of ${\rm 10}^5 < v_{\rm disc, 0} < {\rm 10}^7 {\rm cm s}^{-1}$ and ${\rm 10}^{-13} < \rho_{\rm disc, 0} < {\rm 10}^{-9} {\rm g cm}^{-3}$. Thus we investigate 500 $\le x_{\rm disc} \le$ 5000 in this model analysis.
In Eq.\[eq:p\_be\], $n = $2 corresponds to a constant speed and $n = $4 corresponds to a constant ram pressure. We can approximate the polar wind with $n_{\rm polar} =$2, since the polar wind is generally thought to reach a terminal speed within a few stellar radii. In contrast, the disc density falls rapidly with radius as the rotating material gradually accelerates outward. @ball99 describes the disc of SS2883 with $n_{\rm disc} \lesssim$ 4, while @waters88 gives 2 $< n_{\rm disc} <$ 3.25 from a general consideration of Be star discs. We adopt $n_{\rm disc} = $2.5 here. Changing $n_{\rm disc}$ to 4 and keeping other parameters fixed reduces the emission by a factor of about 2 (Eq. \[eq:gamma\_flux-propt\]).
There are no fixed limits for the orbital phases in which the Be star disc outflow interacts with the pulsar wind. We consider three possibilities in calculating the light curves; (i) aligned disc to the orbital plane and interaction throughout the orbit, (ii) mis-aligned disc and interaction in the $\sim$200-day period around periastron ($\tau$), during which the radio emission is depolarized, or (iii) mis-aligned disc and interaction in two short periods, \[($\tau - $18 d) $\sim$ ($\tau - $8 d)\] and \[($\tau +$12 d) $\sim$ ($\tau +$22 d)\], as discussed in @connors02. Eq.\[eq:gamma\_flux-propt\] with $x_{\rm polar}$ of 10$^{-1} \times x_{\rm disc}$, suggests that the contribution from the polar-wind–pulsar-wind interaction is a factor 1.5$\times {\rm 10}^{-2}$ of that from the disc–pulsar-wind interaction. The polar wind is generally assumed to interact with the pulsar wind at all orbital phases. When the disc and pulsar-wind interaction diminishes. the estimated intensity from the system is only of the polar-wind contribution, and is reduced by a factor of $\sim {\rm 10}^{-2}$. We take account of (i), implying the maximum effect of the disc-pulsar wind interaction, though the disc material becomes dilute at larger distances (Eq.\[eq:p\_be\]). In (ii) and (iii), we consider emissions from the pulsar-wind bubble formed in the disc flow, after the pulsar leaves the disc[@connors02]. The bubble moves at the velocity $v_{\rm bubble}$ in the outflow and shock acceleration of particles in the flow proceeds in the contact discontinuity between the bubble and the outflow material. Emissions from the moving bubble are calculated along its trace referring to the material and momentum density profiles of the flow, by replacing $R_{\rm shock}(x,n,v_0)$ with $R_{\rm shock}(t=t_0, x,n,v_0) + v_{\rm bubble}(t-t_0)$ in Eq.\[eq:gamma\_flux-propt\], where $t_0$ denotes the time when the pulsar moves out of the disc flow. We assume an initial value of $v_{\rm bubble} =$ 100${\rm km}{\rm s}^{-1}$ which is larger than the value used in @connors02, 15${\rm km}{\rm s}^{-1}$, but is similar to the model in @paredes91 as well as to the typical velocity of the disc flow. The adiabatic expansion, which is mainly important for synchrotron emission, does not affect much the emission mechanism mentioned here. The rise time of bubble emission is assumed to be $\sim$ 1 day.
Comparison with the Results
---------------------------
The observational upper limits are compared with light curves calculated from the model. The energy thresholds of our results have been scaled to 1TeV assuming a $E^{-2.0}$ spectrum. The spectra calculated with the model assumption (i) in disc-pulsar wind interaction, are integrated (E $\ge$ 1TeV) for four different mass outflow parameters, $x_{\rm disc} =$ 500, 1000, 1500, and 5000 (Fig.\[fig:model\_comparison\_x\]). The outflow parameter is constrained by our results to $x_{\rm disc} \le $1500. The light curves with the different model assumptions (i)–(iii) for the fixed mass outflow parameter $x_{\rm disc}$ of 1500 are shown in Fig.\[fig:model\_comparison\_interact\]. As discussed in the previous subsection, the light curve is reduced by a factor of $\sim$ 10$^{-2}$ outside the assumed disc–pulsar-wind interaction period since the polar-wind becomes the only counterpart of the pulsar-wind. In addition, contribution from the wind-bubble formed in the disc–pulsar-wind interaction, remains while the bubble is moving in the disc. Thus, for the model assumption (iii) where the disc and the pulsar-wind interact twice in the orbit, the emission peak after periastron consists of the “second” disc–pulsar-wind interaction, of the disc-wind-bubble interaction where the bubble is the outcome of the “first” interaction, and of the polar-wind–pulsar-wind interaction. For all three assumptions the constraint from the observations, mainly from $Obs.\,A$, is similar. With this relatively small outflow pressure, the Be star wind may not be able to overwhelm the pulsar wind pressure to produce accretion onto the pulsar, as has been suggested by the X-ray observations.
Besides our emission models based on the Be-star outflows, we discuss the light curve shown in Fig.5 of @ballkirk00 as the optimum case for TeV emission from the pulsar wind side, using rather ideal model of the inverse Compton scattering on the un-shocked pulsar wind with a wind Lorentz factor of 10$^7$. Our upper limits are modified into units of integral energy flux using approximated spectral indices in Fig.4 of @ballkirk00, but the obtained limit of $\sim$1$\times {\rm 10}^{-5} {\rm MeV} {\rm cm}^{-2}{\rm s}^{-1}$, does not strongly constrain the model since the light curve quickly declines from $\sim$5 at the periastron epoch to 0.2 in units of ${\rm 10}^{-5} {\rm MeV} {\rm cm}^{-2}{\rm s}^{-1}$. The integrated flux greater than 1 TeV is obtained from another model calculation of pulsar wind emission using the spectra in Fig.7 of @shibazaki02. They argue for the dominance of synchrotron cooling in the energy loss of the pulsar wind electrons. Assuming the distance of 1.5kpc, the predicted flux of ${\rm 10}^{-14} {\rm cm}^{-2} {\rm s}^{-1}$ is about two orders of magnitude smaller than our limit.
Recently, new projects of ground based Cherenkov telescopes have begun operations[@review-latest]. With the improved sensitivity and the lower energy threshold, they will offer a better opportunity to observe the PSRB1259$-$63 binary system in the high-energy band. For projects such as CANGAROO-III or H.E.S.S., located in the southern hemisphere, a 50-hour observation of the binary system gives a typical sensitivity of $\sim {\rm 10}^{-11} {\rm cm}^{-2} {\rm s}^{-1}$ with the energy threshold of $\sim$100GeV[@hess-performance]. The calculated spectra of our models are integrated again, for the energy greater than 100GeV, for comparison with this sensitivity. In Fig.\[fig:future\_comparison\], the sensitivity levels of 20-hour, 10-hour and 5-hour (statistically scaled) observations, respectively, are drawn over the calculated light curves. A day-scale light curve might be detectable for the model with $x_{\rm disc} \ge \sim$700 along the periastron passage. @balldodd01 estimate the $\sim$100 GeV emission from the pulsar wind with a Lorentz factor of 10$^6$, and their light curves are compared with these expected sensitivities after modification of the unit into the integral energy flux (${\rm MeV} {\rm cm}^{-2}{\rm s}^{-1}$), assuming the spectral shape (Fig.4 of @balldodd01). The light curves in Fig.\[fig:future\_comparison\] $(right)$ are taken from Fig.5 of @balldodd01 showing terminated (solid line) and un-terminated (dashed line) shock models in the pulsar wind emissions. Both model predictions are comparable with the detectable flux, at least, around the periastron epoch. From @shibazaki02, integrations $E \ge$ 100 GeV are performed resulting in fluxes of $\sim$4$\times {\rm 10}^{-12}
{\rm cm}^{-2} {\rm s}^{-1}$ at periastron and $\sim$1$\times {\rm 10}^{-12} {\rm cm}^{-2} {\rm s}^{-1}$ at apastron, which is still below the improved sensitivity of ground based detectors.
Summary
=======
The PSRB1259$-$63/SS2883 binary system has been observed at TeV energies using the CANGAROO-II 10-m telescope. The observations were performed at two different orbital phases following the October 2000 periastron. Upper limits on the integrated TeV gamma-ray flux are obtained. A new model for gamma-ray emission from the Be star outflow has been introduced, and contributions from Bremsstrahlung, inverse Compton scattering, and proton-proton interactions, are calculated, with possible variations in parameters considered. The light curves are calculated with different assumptions on the disc-pulsar wind interaction. The estimated light curves are discussed and compared with our observational results to constrain the disc-like outflow density $\rho_{0,-12} = \rho_{0}/{\rm 10}^{-12}$ g cm$^{-3}$ and its flow speed $v_{0, 6} = v_{0}/10^{6}$ cm s$^{-1}$ by $x_{\rm disc} = \rho_{0,-12}~v_{0, 6} \le$ 1500. The next periastron will occur in March 2004 when the condition will favorable for small zenith angle observations and hence low energy thresholds for ground based Cherenkov telescopes. Further observations of PSRB1259$-$63 system during the periastron passage (including the weeks before and after the periastron, respectively) are encouraged to provide valuable information.
The authors greatly thank Dr. L. Ball for introducing his model calculations to us and for promoting the first idea of these observations. Dr.T. Terasawa gave useful comments in the early discussion of the model. The research was supported by a Grant-in-Aid for Scientific Research of the Ministry of Education, Culture, Science, Sports and Technology of Japan and by the Australian Research Council. AA, SH, JK, LK, KO, KS, and KT were supported by Research Fellowship and Postdoctoral Fellowships of Japan Society of Promotion of Science.
[999]{} Asahara, A. et al. 2003, Nucl. Inst. Meth., submitted Ball, L., Melatos, A., Johnston, S. & Skjaeraasen O. 1999, , 514, L39 Ball, L. & Kirk, J. G. 2000, Astropart. Phys., 12, 335 Ball, L. & Dodd, J. 2001, Publ. Astron. Soc. Aust., 18, 98 Baring, M. G., Ellison, D. C., Reynolds, S. P., Grenier, I. A., & Goret, P. 1999, , 513, 311 Cominsky, L., Roberts, M., & Johnston, S. 1994, , 427, 978 Connors, T. W., Johnston, S., Manchester, R. N., & McConnell, D. 2002, , 336, 1201 Cote, J. & Waters, L. B. F. M. 1987, , 176, 93 Dougherty, S. M., Waters, L. B. F. M., Burki, G., Cote, J., Cramer, N., van Kerkwijk, M. H., & Taylor A. R., 1994, , 290, 609 Enomoto, R. et al. 2002, Astropart. Phys., 16, 235 Enomoto, R. et al. 2002, Nature, 416, 823; Gaisser, T. K., Protheroe, R. J. & Stanev, T., 1998, , 492, 219 Grove, J. E., Tavani, M., Purcell, W. R., Johnson, W. N., Kurfess, J. D., Strickman, M. S. & Arons, J., 1995, , 447, L113 Hillas, A. M., 1982, J. Phys. G 8, 1475 Hirayama, M., Nagase, F., Tavani, M., Kaspi, V. M., Kawai, N. & Arons, J. 1996, , 48, 833 Hirayama, M., Cominsky, L. R., Kaspi, V. M., Nagase, F., Tavani, M., Kawai, N. & Grove, J. E. 1999, , 521, 718 Itoh, C. et al., 2003, , 402, 443 Johnston, S., Lyne, A. G., Manchester, R. N., Kniffen, D. A., D’Amico, N., Lim, J. & Ashworth, M. 1992, , 255, 401 Johnston, S., Manchester, R. N., Lyne, A. G., Bailes, M., Kaspi, V. M., Guojun, Qiao & D’Amico, N. 1992, , 387, L37 Johnston, S., Manchester, R. N., Lyne, A. G., Nicastro, L. & Spyromilio, J. 1994, , 268, 430 Johnston, S., Manchester, R. N., Lyne, A. G., D’Amico, N., Bailes, M., Gaensler, B. M. & Nicastro, L. 1996, , 279, 1026 Johnston, S., Manchester, R. N., McConnell, D. and Campbell-Wilson, D. 1999, , 302, 277 Kaspi, V. M. et al., 1995, , 453, 424 Kaspi, V. M. 1997, in Proc. Neutron Stars and Pulsars, eds.: Shibazaki, N., Kawai, N., Shibata, S. & Kifune, T. (Tokyo: UAP), 519 Kawachi, A. et al. 2001, Astropart. Phys., 14, 261 Kennel, C. F. & Coroniti, F. V. 1984, , 283, 694 Kirk, J. G., Ball, L. & Skjaeraasen, O. 1999, Astropart. Phys., 10, 31 Konopelko, A. 1999, Astropart. Phys., 11 263 Lamers, H. J. G. L. M. & Waters, L. B. F. M. 1987, , 182, 80 Melatos, A., Johnston, S. & Melrose, D. B. 1995, , 275, 381 Müller, D. et al., 1995, in Proc. 24th Internat. Cosmic Ray Conf., (Rome), vol.3 13 Murata, K., Tamaki, H., Maki, H. & Shibazaki, N. 2003, , 55, 467 Mori, M. et al., 2001, in Proc. 27th Internat. Cosmic Ray Conf., (Hamburg), vol.5 2831 Naito, T. & Takahara, F. 1994, J. Physics. G, 20, 477 Okumura, K. et al., 2002, , 579, L9 Ong, R. A. 2003, in Proc. The Universe Viewed in Gamma-Rays, eds.: Enomoto, R., Mori, M., & Yanagita, S. (Chiba: UAP), 587 Paredes, J. M., Mart, J., Estallela, R. & Sarrate, J. 1991, , 248, 124 Punch, M. et al., 1992, Nature 358, 477 Sako, T. et al. 1997, in Proc. 25th Internat. Cosmic Ray Conf., (Durban), vol.3 193 Sturner, S. J., Skibo, J. G., Dermer, C. D. & Mattox, J. R., 1997, 490, 619 Tanimori, T. et al. 1998, , 492, L33 Tavani, M., Arons, A. & Kaspi, V. M. 1994, , 433, L37 Tavani, M. & Arons, J. 1997, , 477, 439 Waters, L. B. F. M. 1986, , 162, 121 Waters, L. B. F. M., Taylor, A. R., van den Heuvel, E. P. J., Habets, G. M. H. & Persi, P. 1988, , 198, 200 Wex, N., Johnston, S., Manchester, R. N., Lyne, A. G., Stappers, B. W. & Bailes, M. 1998, , 298, 997
[lcccc]{}\
Observation & & True Anomaly & Separation\
& (UT) & (MJD) & (degree) & (10$^{13}$cm)\
Obs.A & & &\
& 2000/12/01 & 51879.72–51879.76 & &\
& 2000/12/03 & 51881.69–51881.74 & &\
& 2000/12/04 & 51882.71–51882.76 & &\
& & &\
& & 125 & 3.6\
Obs.B & & &\
& 2001/03/19 & 51987.60–51987.68 & &\
& 2001/03/21 & 51989.62–51989.75 & &\
& 2001/03/22 & 51990.62–51990.79 & &\
& 2001/03/24 & 51992.59–51992.73 & &\
& 2001/03/25 & 51993.54–51993.70 & &\
& 2001/03/26 & 51994.53–51994.60 & &\
& & &\
& & 153 & 8.1\
\
[lccccccc]{}\
Data & $\theta_{zen}$ & $E_{th}$ & &\
& & & Real & Effective & Recorded & Noise & Image\
& & & & & & Reduction & Selected\
& (degree) & (TeV) & (min) & (min) & & &\
$Obs.\,A$ & & & & & & &\
ON: & 58.9 & 3.6 & 202 & 196 & 1.17E5 & 6.62E3 & 1.85E3\
OFF: & 59.2 & & 193 & 160 & 4.29E5 & 5.86E3 & 1.62E3\
$Obs.\,B$ & & & & & & &\
ON: & 34.0 & 0.78 & 1078 & 623 & 4.43E6 & 7.06E4 & 1.81E4\
OFF: & 34.3 & & 1001 & 645 & 2.63E6 & 7.21E4 & 1.87E4\
[lr]{}\
Parameters & Adopted Values\
&\
Fraction to the equatorial plane, f$_{\rm pw}$ & 0.1\
&\
Radius, $R_*$ & 6$R_{\sun} = {\rm 4.17} \times {\rm 10}^{11}{\rm cm}$\
Distance, $D$ & 1.5kpc\
Opening angle of the disc outflow, $\theta_{\rm disc}$ & 15 degree\
Power law index of density profile of the disc, $n_{\rm disc}$ & 2.5\
Power law index of density profile of the polar wind, $n_{\rm polar}$ & 2\
Efficiency of the acceleration, $f_{\rm acc}$ & 0.1 (proton)\
& 0.001 (electron)\
Power law index of the proton/electron energy flux, $\alpha$ & $-$2.0\
Maximum energy of the accelerated particles, $E^{\rm max}_{\rm e,p}$ & 10$^{15}$ eV\
Outflow parameter of the disc, $x_{\rm disc}$ & 500 – 5000\
Outflow parameter of the polar wind, $x_{\rm polar}$ & $10^{-1} \times x_{\rm disc}$\
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Weyl fermions are two-component chiral fermions in $(3+1)$-dimensions. When coupled to a gauge field, the Weyl fermion is known to have an axial anomaly, which means the current conservation of the left-handed and right-handed Weyl fermions cannot be preserved separately. Recently, Weyl fermions have been proposed in condensed matter systems named as “Weyl semi-metals". In this paper we propose a Weyl semi-metal phase in magnetically doped topological insulators, and study the axial anomaly in this system. We propose that the magnetic fluctuation in this system plays the role of a “chiral gauge field" which minimally couples to the Weyl fermions with opposite charges for two chiralities. We study the anomaly equation of this sytem and discuss its physical consequences, including one-dimensional chiral modes in a ferromagnetic vortex line, and a novel plasmon-magnon coupling.'
author:
- 'Chao-Xing Liu$^1$, Peng Ye$^2$ and Xiao-Liang Qi$^3$'
title: 'Chiral gauge field and axial anomaly in a Weyl semi-metal'
---
[*Introduction -* ]{} In the quantum field theory, a (3+1)-dimensional massless Dirac fermion is decomposed to two independent two-component fermions known as the Weyl fermions. Weyl fermion has a definite chirality, lefthanded or righthanded, determined by the sign of its spin polarization along the momentum direction.[@peskin1995] Classically, the lefthanded and righthanded Weyl fermions are decoupled and can be coupled independently to two gauge fields, leading to a separate charge conservation. The gauge field that couples differently to Weyl fermions with two chiralities is called a chiral gauge field. For example the SU(2) gauge field in the Standard Model is a chiral gauge field. It is well-known that the chiral charge conservation is violated in a quantum theory of Weyl fermions in a background gauge field, which is known as the axial anomaly[@adler1969; @bell1969; @nielsen1983].
Recently, Weyl fermions are also introduced into condensed matter physics. The Weyl fermions are shown to be the topologically robust boundary states of $(4+1)$-d time-reversal (TR) invariant topological insulators (TI)[@qi2008a], and the axial anomaly corresponds to a topological response of the $(4+1)$-d TI. This approach is related to the domain wall fermion approach[@kaplan1992] and Callan-Harvey effect[@callan1985] in high energy physics. By dimensional reduction, the $(4+1)$-d topological insulator is reduced to the $(3+1)$-d TI[@qi2011; @hasan2010; @moore2010] and the Weyl fermion is reduced to $(2+1)$-d surface states of the TI. Weyl fermions also appear directly in $(3+1)$-d gapless electron systems, which are named as “Weyl semi-metals"[@volovik2003; @wan2011; @xu2011; @burkov2011a; @burkov2011b; @cho2011; @fang2011; @jiang2011; @hosur2012; @yang2011; @nielsen1983; @aji2011]. Since a system with both TR and parity (P) symmetries have all energy bands doubly degenerate, the Weyl semi-metal state can only be realized in a system breaking TR and/or P symmetry.
A natural question is whether the chiral gauge field can be realized in the Weyl semi-metals, and if yes, what is the physical consequence. In this letter, we address these questions in TR breaking Weyl semi-metals. We show that generically a ferromagnetic moment couples to the Weyl fermions as a chiral gauge field. As an explicit example system, we study a model of magnetically doped topological insulator, which can be driven into the Weyl semi-metal phase with strong enough magnetic moments. The presence of the chiral gauge field leads to an anomaly equation satisfied by the charge current, which leads to new topological phenomena such as chiral one-dimensional states in a magnetic vortex, and a topological coupling between spin fluctuation and plasmons. [*Chiral gauge field and anomaly equation -* ]{} We start with a general discussion of Weyl fermions in condensed matter physics. In a weakly interacting crystalline material, Weyl fermion states generically appear when two energy bands cross at a generic point $\vec{K}_0$ in the Brillouin zone. The low energy physics around $\vec{K}_0$ is described by a two-component Hamiltonian $H_W=\hbar \sum_{i,j=x,y,z}v_{ij}k_i\sigma_j$, with $k_i$ the momentum away from $\vec{K}_0$, and $\sigma_j$ the Pauli matrices. The matrix $v_{ij}$ describes the generic linear coupling between momentum and spin degree of freedom described by $\sigma_j$. By rotating the basis one can always diagonalize $v_{ij}$, and the three diagonal components are anisotropic velocities. Without losing generality, we restrict our discussion on isotropic Weyl fermions with the simple Hamiltonian $H=\hbar v_f{\vec{ \sigma}\cdot \vec{k}}$. Our results on anomaly and chiral gauge field is insensitive to the anisotropy in the velocity. The sign of the Fermi velocity $v_f$ determines the chirality of Weyl fermion. A single Weyl fermion is topologically stable as long as translation symmetry is preserved[@wan2011], thus for a Weyl semi-metal with translation symmetry, any local perturbation can only move the nodal point $\vec{K}_0$ in the momentum space.
According to the Nielsen-Ninomiya theorem [@nielsen1981a; @nielsen1981b], in a lattice model the number of Weyl fermions with opposite chiralities must be equal. Consequently, the minimum number of Weyl fermions in a Brillioun zone is 2. Moreover, because TR symmetry preserves the chirality of Weyl fermion, in TR invariant system the minimum number of Weyl fermions is 4[@burkov2011b]. In the following, we focus on the “minimal Weyl semi-metal" which break TR but preserves P, with two Weyl fermions of opposite chiralities at wavevectors $\vec{K}_0$ and $-\vec{K}_0$, related to each other by spatial inversion. $\vec{K}_0$ is a generic point in BZ away from TR invariant momenta.
We consider an arbitrary perturbation to the system of two Weyl fermions. As long as the perturbation is so smooth that the momentum transfer is much smaller than $2|\vec{K}_0|$, the two Weyl fermions remains decoupled. The effective Hamiltonian of the lefthanded Weyl fermion under perturbation is $H_L=\hbar v_f\vec{\sigma}\cdot\vec{k}+\delta H_L$, with $\delta H_L$ a generic $2\times 2$ Hermitian matrix. To the leading order one can ignore the $k$ dependence and consider $\delta H_L$ as a constant term. Then $\delta H_L$ can always be expanded to the form $\delta H_L=\hbar v_f\vec{\sigma}\cdot \vec{a}_L+a_{0L}$ with the last term proportional to identity. Adding this to the Weyl fermion Hamiltonian we find $H_L=\hbar v_f(\vec{k}+\vec{a}_L)\cdot\vec{\sigma}+a_{0L}$ with $a_{\mu L}=(a_{0L},\vec{a}_L)$ behaving as a gauge field. Similarly one can define the gauge field $a_{\mu R}$ minimally coupled to the righthanded Weyl fermions with the Hamiltonian $H_R=-\hbar v_f(\vec{k}+\vec{a}_R)\cdot\vec{\sigma}+a_{0R}$. The two Weyl fermions can be described together by a $4\times 4$ Hamiltonian: $$\begin{aligned}
&&H
=\hbar v_f\left( (\vec{k}+\vec{A})\cdot\vec{\sigma}\tau_z+\vec{a}\cdot\vec{\sigma}\right)+a_0\tau_z +A_0
\label{eq:CGF_Ham1}\end{aligned}$$ with $A_\mu=(a_{\mu L}+a_{\mu R})/2$ behaving like the electromagnetic gauge field, and $a_\mu=(a_{\mu L}-a_{\mu R})/2$ the chiral gauge field. $A_\mu$ and $a_\mu$ have different properties under P and TR. For example if the perturbation we consider is a fluctuation of a ferromagnetic moment, only $\vec{a}$ will be induced which is TR odd and P even.
As known from the quantum field theory, when a Weyl fermion is coupled to a gauge field, the charge conservation is broken at the quantum field level, leading to the axial anomaly[@peskin1995; @zee2010], which can be described by the anomaly equation $\partial_{\mu}j^{\mu L(R)}=(-)\frac{1}{32\pi^2}\epsilon^{\lambda\rho\mu\nu}f^{L(R)}_{\lambda\rho}f^{L(R)}_{\mu\nu}$ where $f^{L(R)}_{\mu\nu}=\partial_{\mu}a_{\nu L(R)}-\partial_{\nu}a_{\mu L(R)}$. Since the gauge field $a_{\mu L(R)}$ of lefthanded (righthanded) Weyl fermion is related to the gauge field $A_{\mu}$ and $a_{\mu}$, the anomaly equation can also be rewritten as $\partial_\mu j^{\mu L(R)}=(-)\frac{1}{32\pi^2}\epsilon^{\lambda\rho\mu\nu}(F_{\lambda\rho}+(-)f_{\lambda\rho})(F_{\mu\nu}+(-)f_{\mu\nu})$, where $F_{\mu\nu}=\partial_{\mu}A_{\nu}-\partial_{\nu}A_{\mu}$ is electromagnetic field strength and $f_{\mu\nu}=\partial_{\mu}a_{\nu}-\partial_{\nu}a_{\mu}$ is chiral gauge field strength. Let’s define the charge current as $j^{\mu}=j^{\mu L}+j^{\mu R}$ and the axial current as $j^{\mu 5}=j^{\mu R}-j^{\mu L}$. When $\vec{A}$ and $\vec{a}$ coexist, we find both the axial current and charge current are nonconserved with the anomaly equations $$\begin{aligned}
&&\partial_\mu j^{\mu 5}=-\frac{1}{16\pi^2}\epsilon^{\lambda\rho\mu\nu}(F_{\alpha\beta}F_{\mu\nu}+f_{\lambda\rho}f_{\mu\nu}), \label{eq:CGF_Anomalyaxial} \\
&&\partial_\mu j^{\mu}=\frac{1}{8\pi^2}\epsilon^{\lambda\rho\mu\nu}f_{\lambda\rho}F_{\mu\nu}.
\label{eq:CGF_Anomaly}\end{aligned}$$ The equation (\[eq:CGF\_Anomalyaxial\]) is the axial current anomaly[@nielsen1983; @aji2011] but with the additional term induced by chiral gauge field, while the equation (\[eq:CGF\_Anomaly\]) indicates the conservation of charge current is also broken due to the combination effect of chiral gauge field and electromagnetic field, which is the main focus of this paper. At the first glance, the breaking of the charge conservation seems something unphysical. Here we emphasize that the Weyl fermion description is only a low energy effective theory and the high energy part is not taken into account. Let’s define $j_b^{\mu}=-\frac{1}{2\pi^2}\epsilon^{\mu\nu\lambda\rho}a_{\nu}\partial_{\lambda}A_{\rho}$, and the right-hand side of equation (\[eq:CGF\_Anomaly\]) can be written as a total derivative of $j^{\mu}_b$, and the charge conservation law $\partial_{\mu}\left( j^{\mu}+j_b^{\mu} \right)=0$ is recovered if $j^{\mu}_b$ is regarded as a current from the high energy part which is neglected in our description. Actually we notice that the spatial component of $j^{\mu}_b$ is given by $\vec{j}_b=-\frac{1}{2\pi^2}\vec{a}\times \vec{E}$ with the electric field $\vec{E}$, exactly corresponding to the anomalous Hall response of Weyl fermion, as first derived in Ref [@yang2011]. To make our discussion concrete, we first propose a realization of Weyl fermions and chiral gauge field in magnetically doped topological insulators, before discussing the physical consequence of this anomaly equation.
![ (a) “Chiral magnetic field” can be generated by the magnetic vortex configuration in a topological insulator cylinder. Here the vector $\vec{b}$ indicates the direction of the “chiral magentic field”. The Landau level spectrum of a massless Dirac fermion is plotted (b) for a uniform magnetic field $\vec{B}$ and (c) for a uniform “chiral magentic field” $\vec{b}$. []{data-label="fig:vortex"}](vortex.eps){width="3.5in"}
![ (a) The energy dispersion as a function of $k_z$ along a ferromagnetic vortex line in a topological insulators with the total angular momentum $J_z=\frac{1}{2}$. (b) The corresponding radial wave function for the two chiral modes at $k_z=0$. Here red line is for the wave function near the vortex core (r=0) and the black line for the one at r=R. (c) The energy dispersion of the two zero modes for different $J_z=n+\frac{1}{2}$. The gap is observed for large $|n|$, due to the finite size effect, which gives a cut-off of the total number of chiral modes. In (d), the wave function of the inner chiral mode moves outwards, hybridizing with the outer chiral mode (Here we take $n=3$ and $k_z=0.06$ 1/Å). The parameters of the four band model are taken to be $M_0=0$, $M_1=0.342eV\cdot$Å$^2$, $M_2=18.25eV\cdot$Å$^2$, $B_0=1.33eV\cdot$Å, $A_0=2.82eV\cdot$Å, $U_0=0.1eV$ and $W_0=0.06eV$. []{data-label="fig:chiral"}](chiralmode.eps){width="3.5in"}
[*Material realization -*]{} It is first suggested that Weyl fermions can be realized in pyrochlore iridates[@wan2011], and later another material HgCr$_2$Se$_4$ is also proposed[@xu2011]. However both the materials include multiple Weyl fermions with the number larger than 2, making the system complicated, therefore it is desired to have a system with the minimal number of Weyl fermions, which actually can be achieved by magentically doped topological insulators[@burkov2011a; @cho2011]. By substituting the atoms, it is possible to tune the band gap of topological insulators, and even induce the phase transition between trivial and non-trivial phases, which has been realized in TlBi(S$_{1-\delta}$Se$_{\delta}$)$_2$ recently[@yan2010; @lin2010; @xu2011a]. Near the transition point, the bulk gap is minimized and can be overcomed by the exchange coupling from magnetic doping. The ferromagnetism in the Cr or Fe doped Bi$_2$Te$_3$ and Sb$_2$Te$_3$ has been observed in experiment[@chen2010; @chang2011; @wray2011], therefore the magnetically doped Bi$_2$Se$_3$ and TlBiSe$_2$ family of materials are the suitable platform for the realization of minimal number of Weyl fermions. Here we adopt the four band model[@zhang2009; @liu2010] with general mass terms, to describe these materials, $$\begin{aligned}
&&H=H_0+H_1\label{eq:MR_Ham1}\\
&&H_0=\epsilon(\vec{k})+\mathcal{M}(\vec{k})\Gamma_5+B_0k_z\Gamma_4+A_0(k_y\Gamma_1-k_x\Gamma_2)\nonumber\\
&&H_1=\sum_{ij}m_{ij}\Gamma_{ij}\nonumber\end{aligned}$$ where $\epsilon_{\bold{k}}=C_0+C_1k_z^2+C_2k^2_\parallel$, $\mathcal{M}(\bold{k})=M_0+M_1k_z^2+M_2k^2_\parallel$. The $\Gamma$ matrices are defined as $\Gamma_{1,2,3}=\sigma_{x,y,z}\tau_x$, $\Gamma_4=\tau_y$, $\Gamma_5=\tau_z$, and $\Gamma_{ab}=[\Gamma_a,\Gamma_b]/2i$ ($a,b=1,\dots,5$). Ferromagnetism breaks T but preserves P, therefore by inspecting the symmetry property of $\Gamma$ matrices (eg. the table III in the reference [@liu2010]), we immediately find only two sets of $\Gamma$ matrices are allowed in $H_1$: $\Gamma_{ij}=\varepsilon_{ijk}\sigma_k$ and $\Gamma_{i4}=\sigma_i\tau_z$ ($i,j,k=x,y,z$). Generally $\Gamma_{12}$ and $\Gamma_{34}$ can be induced by z-direction magnetization, while $(\Gamma_{14},\Gamma_{24})$ and $\left( \Gamma_{23},\Gamma_{31}\right)$ originate from in-plane magnetization. It is shown that $\Gamma_{14}$, $\Gamma_{24}$ and $\Gamma_{12}$ induce two Weyl fermions while $\Gamma_{23}$, $\Gamma_{31}$ and $\Gamma_{34}$ yield a nodal ring[@burkov2011b]. Since now we are interested in the Weyl fermion regime, we focus on the simple case with $H_1=U_{0}\Gamma_{12}$, yielding the energy dispersion $E_{st}=\epsilon(\vec{k})+s\sqrt{A_0^2(k_x^2+k_y^2)+(\sqrt{\mathcal{M}^2+B_0^2k_z^2}+t|U_0|)^2}$, where $s,t=\pm 1$. The two bands in the middle with $t=-1$ touches when the conditions $\mathcal{M}^2+B^2_0k^2_z=U_0^2$ and $k_x=k_y=0$ are satisfied. If we neglect the quadratic term in $\mathcal{M}$ for simplicity, the bulk gap is closed, realizing Weyl fermions, at the momentum $k_z=\pm K_0$ with $K_0=\frac{1}{B_0}\sqrt{U_0^2-M_0^2}$ if $|U_0|>|M_0|$.
![ (a) The experiment setup and the diagram for the corresponding physical process. Here the red line denotes the electromagnetic field, and the blue dashed line is for $a_z$ field. (b) The Landau levels for the four band model (\[eq:MR\_Ham1\]) in the uniform magnetic field. (c) The imaginary part of the correlation function $\langle a_z(q)a_z(-q)\rangle$. Here we take $\omega_0/\omega_p=0.1$, $v_s=0$. []{data-label="fig:setup"}](setup.eps){width="3.5in"}
Next we consider the perturbation around the gapless points $\vec{k}=(0,0,\pm K_0)$ with the perturbed Hamiltonian $H'=B_0 \delta k_z\tau_y+A_0\left( \delta k_y\sigma_x\tau_x- \delta k_x\sigma_y\tau_x \right)
+ \sum_{i=x,y,z}\left( \mu_i\sigma_i+ \nu_i\sigma_i\tau_z \right)$, where $\vec{\mu}$ and $\vec{\nu}$ denote the magnetic fluctuation and $\delta \vec{k}$ is the momentum expanded around $(0,0,\pm K_0)$. We project the Hamiltonian $H'$ into the subspace expanded by $|-,-\rangle$ and $|+,-\rangle$ by perturbation theory, obtaining the effective Hamiltonian $H_{eff}=\hbar v_{fz}(\delta k_z\sigma_z\tau_z+a_z\sigma_z)+
\hbar v_{f\parallel}\sum_{i=x,y}\left( \delta k_i\sigma_i\tau_z+a_i\sigma_i \right)$, where $\sigma$ denotes spin, $\tau$ represents two Dirac cones at $(0,0,\pm K_0)$, the Fermi velocity $\hbar v_{fz}=-\frac{B_0^2K_0}{U_0}$, $\hbar v_{f\parallel}=-A_0$ and the chiral gauge potential $\vec{a}$ is given by $$\begin{aligned}
&&\hbar v_{fz} a_z=\mu_z-\frac{1}{U_0}\left( \nu_z^2-\mu_x^2-\mu_y^2-\frac{M_0^2}{U_0^2}\sum_i\nu_i^2 \right)\label{eq:MR_chiral1}\\
&&\hbar v_{f\parallel} a_x=\frac{B_0K_0}{U_0}\nu_x-\frac{B_0K_0}{U_0^2}\nu_z\mu_x\label{eq:MR_chiral2}\\
&&\hbar v_{f\parallel} a_y=\frac{B_0K_0}{U_0}\nu_y-\frac{B_0K_0}{U_0^2}\nu_z\mu_y.
\label{eq:MR_chiral3}\end{aligned}$$
[*Physical consequence -*]{} Let’s introduce “chiral magnetic field” $\vec{b}=\nabla\times\vec{a}$ and “chiral electric field” $\vec{e}=\frac{\partial\vec{a}}{\partial t}$, then equation (\[eq:CGF\_Anomaly\]) can be rewritten as $$\begin{aligned}
\frac{\partial\rho}{\partial t}+\vec{\nabla}\cdot\vec{j}=\frac{1}{2\pi^2}\left( \vec{b}\cdot\vec{E}+\vec{e}\cdot\vec{B} \right),
\label{eq:PC_Anomaly}\end{aligned}$$ with $\rho$ and $\vec{j}$ the charge density and current, respectively. In the following we study the physical consequences of the two terms on the righthand side of this equation.
The first term $\vec{b}\cdot\vec{E}$ describes the effect of a chiral magnetic field parallel to the electric field. The “chiral” magnetic field can also induce Landau levels, similar to the Landau levels by magnetic field. For a uniform field $\vec{b}=b_0\hat{e}_z$, the Landau level spectrum is given by $E_{\pm,\alpha}(n)=\pm\hbar v_f\sqrt{k_z^2+2b_0n}$ with $n=1,2,\dots$ and $\alpha=\pm$ denote two Dirac cones. In addition, there are two zeroth Landau levels, both with the dispersion $E_{\alpha}(0)=-\hbar v_fk_z$ ($\alpha=\pm$), as shown in Fig.\[fig:vortex\] (c). It should be noticed that the two zeroth Landau levels are one-dimensional modes with the same chirality, in contrast to the case of an ordinary magnetic field $\vec{B}=B_0\hat{e}_z$ shown in Fig. \[fig:vortex\] (b). In other words, the low energy dynamics of the system is described by two chiral fermions in each area with one flux quanta of $\vec{b}$. In this case, the anomaly equation (\[eq:PC\_Anomaly\]) reduces to the chiral anomaly of 1D chiral fermions[@johnson1963; @jackiw1985].
A key difference of the chiral gauge field from the electromagnetic gauge field is that the gauge vector potential $\vec{a}$ is physical and thus has to be single valued. A uniform $\vec{b}$ field corresponds to a $\vec{a}$ linearly increasing towards the boundary of the system, which is unphysical. Also the Landau level spectrum above does not consider the coupling between the two Weyl fermions. To obtain a more complete understanding to this problem, we consider the four band model (\[eq:MR\_Ham1\]) with the magnetic vortex configuration $m_{14}=-W_0\sin\theta$, $m_{24}=W_0\cos\theta$ ($\theta$ is the angular coordinate), $m_{12}=U_0$ and all the other $m_{ij}=0$. Assuming a cylinder shape geometry as shown in Fig. (\[fig:vortex\]) (a), the system has translation symmetry in $z$ direction and rotation symmetry according to $z$ axis. Therefore the momentum $k_z$ and total angular momentum $J_z$ are good quantum numbers, and the Schordinger equation can be solved numerically for each sector of $k_z,J_z$ by introducing a discretization, as described in the appendix in detail. Fig \[fig:chiral\] (a) and (b) show the band dispersion along z direction for the angular momentum $J_z=1/2$, and the corresponding wave function. There are two gapless modes with opposite chirality, different from the uniform field Landau levels in Fig. \[fig:vortex\] (c). The wavefunction shows that the two chiral modes are spatially separated, with one wave function around $r=0$ and the other one located at the boundary $r=R$, as shown in Fig \[fig:chiral\] (a) and (b). From this result we see that the chiral modes in the zeroth Landau level are compensated by modes with opposite chirality on the boundary, which is expected since in such a finite system the number of left and right moving 1D states must be equal. With increasing the angular momentum $J_z$, the wavefunction of the inner chiral mode moves towards larger radius, as shown in Fig. \[fig:chiral\] (d). For a finite system, the wave functions of the two chiral states near $r=0$ and $r=R$ will overlap with each other for large angular quantum $J_z$ (Fig \[fig:chiral\] (d)), leading to a gap opening, as shown in Fig \[fig:chiral\] (c). Consequently we obtain a finite number of chiral modes in the zeroth Landau level. As expected, the number of chiral modes is determined by the total flux of $\vec{a}$ in the system, just like the case of uniform $\vec{b}$ field. In such a configuration, the consequence of the anomaly equation (\[eq:PC\_Anomaly\]) is actually a quantum Hall effect[@yang2011]. In an electric field $\vec{E}=E\hat{z}$ parallel to $\vec{b}$, the anomaly equation describes a charge generation around the center of the system, while the charge on the boundary is annihilated. This is a consequence of a Hall current flowing along the radial direction towards the center, which can be measured in transport experiments.
The second term on the right hand side of the anomaly equation (\[eq:PC\_Anomaly\]) describes the combination effect of magnetic field and “chiral” electric field. To understand this term, consider a uniform magnetic field $\vec{B}=B_0\hat{z}$ and a uniform vector potential $\vec{a}=a_z(t)\hat{z}$ changing adiabatically in time. The anomaly equation leads to $\delta\rho=\frac{G}{2\pi}\delta a_z$, with $G=\frac{eB_0}{h}$ the Landau level degeneracy. Therefore the change of $a_z$ leads to a charge density modulation proportional to it. To understand this equation we consider the Landau level spectrum for the four band model (\[eq:MR\_Ham1\]), as is shown in Fig \[fig:setup\] (b). It is important to note that the two zeroth Landau levels with opposite chirality have the same spin polarization. Consequently, the exchange coupling of $\vec{a}=a_z\hat{z}$ with the zeroth Landau level states is equivalent to a scalar potential, which shifts the chemical potential and leads to the change of charge density.
Since this term couples charge density and magnetization, it leads to an interesting physical consequence of the hybridization between the plasmon and magnon modes. The effective action of the present system can be given by $S_{A}+S_{a}+S_{RPA}$, where $S_{A}=\int dkd\omega A_0(q)G_{A0}^{-1} A_0(-q)$, $S_a=\int dkd\omega a_z(q)G_a^{-1}a_z(-q)$ and $S_{RPA}=\int dkd\omega \left[ (A_0(q)+a_z(q))\Pi(A_0(-q)+a_z(-q)) \right]$ with $ q=(\omega,k)$. Here $a_z$ has been rescaled to have the same dimension as $A_0$. $S_A$ describes the dynamics of the scalar potential $A_0$ with $G_A^{-1}\sim v_c^2k^2$ with photon velocity $v_c$, $S_a$ describes the dynamics of $a_z$ field with $G_a^{-1}\sim \omega^2-v_s^2k^2-\omega_0^2$ where $\omega_0$ gives the excitation gap and $v_s$ is the magnon velocity, and $S_{RPA}$ gives the effective action after integrating out the interacting fermions with the random phase approximation, where $\Pi\sim -\frac{\omega_p^2}{\omega^2}v_c^2k^2$[@nagaosa1999] ($\omega_p$ is the plasmon frequency). Such a hybridization is shown schematically in Fig \[fig:setup\] (a). To see the effect on the dynamics of the magnon described by $a_z$, we integrate out $A_0$ field and obtain $S_{eff}=\int dkd\omega a_z(q)\left( G^{-1}_a+\frac{\Pi G_A^{-1}}{\Pi+G_A^{-1}} \right)a_z(-q)$. The corresponding correlation function is given by $\langle a_z(q)a_z(-q)\rangle\sim \left( \omega^2-v_s^2k^2-\omega_0^2-\frac{\omega_p^2v_c^2k^2}{\omega^2-\omega_p^2+i\eta}+i\eta \right)^{-1}$, which corresponds to the spin susceptibility. As plotted in Fig \[fig:setup\] (c), the correlation function has two poles, of which one corresponds to the intrinsic magnon excitation with the frequency around $\omega_0$, while the other one only appears for finite $k$ with the intensity proportional to $k^2$ and is induced by the plasmons with frequency around $\omega_p$. The plasmon frequency can be estimated as $\sim35meV$ for Weyl fermions [@dassarma2009] with dielectric constant $\sim100$, Fermi velocity $\sim6.85\times 10^5 m/s$, and electron density $\sim10^{19}$ cm$^{-3}$. Such an additional mode in magnon spectrum can be observed in neutron scattering experiments and compared with the plasmon frequency $\omega_p$ determined by reflection spectroscopy. The authors would like to thank J. Jain, K. Sun, C.K. Xu, Y.S. Wu and S.C. Zhang for helpful discussions. This work is supported by Tsinghua Education Foundation North America (P.Y.), and the Defense Advanced Research Projects Agency Microsystems Technology Office, MesoDynamic Architecture Program (MESO) through the contract number N66001-11-1-4105 (X.L.Q). This work was supported in part by the National Science Foundation under Grant No. NSF PHY11-25915 when the authors participated in the KITP program Topological Insulators and Superconductors. We thank KITP for hospitality.
Numerical method for the calculation of the chiral mode in the ferromagnetic vortex core
========================================================================================
In the appendix, we describe our numerical method for the calculation of energy dispersion and eigen wavefunction for the ferromagnetic vortex configuration. We start from the four band model (4) in the main text and in the cylinder coordinate $(r,\theta,z)$, the Hamiltonian takes the form of $$\begin{aligned}
&&H=H_0+H_1\label{eq:AP_Ham1}\end{aligned}$$ $$\begin{aligned}
&&H_0=\mathcal{M}(\vec{k})\tau_z+B_0k_z\tau_y+A_0(k_y\sigma_x\tau_x-k_x\sigma_y\tau_x)\nonumber\end{aligned}$$ $$\begin{aligned}
&&=\left(
\begin{array}{cccc}
\mathcal{M}(\vec{k})&0&-iB_0k_z&iA_0k_-\\
0&\mathcal{M}(\vec{k})&-iA_0k_+&-iB_0k_z\\
iB_0k_z&iA_0k_-&-\mathcal{M}(\vec{k})&0\\
-iA_0k_+&iB_0k_z&0&-\mathcal{M}(\vec{k})
\end{array}
\right)\nonumber\end{aligned}$$ $$\begin{aligned}
&&H_1=-W_0\sin\theta\sigma_x\tau_z+W_0\cos\theta\sigma_y\tau_z+U_0\sigma_z\nonumber\\
&&=\left(
\begin{array}{cccc}
U_0&-iW_0e^{-i\theta}&0&0\\
iW_0e^{i\theta}&-U_0&0&0\\
0&0&U_0&iW_0e^{-i\theta}\\
0&0&-iW_0e^{i\theta}&-U_0
\end{array}
\right)\nonumber
\label{}\end{aligned}$$ where we have $\partial_x=\cos\theta\partial_r-\frac{\sin\theta}{r}\partial_\theta$ and $\partial_y=\sin\theta\partial_r+\frac{\cos\theta}{r}\partial_\theta$, therefore $k_-=k_x-ik_y=-i\partial_x-\partial_y=-i\left( \cos\theta\partial_r-\frac{\sin\theta}{r}\partial_\theta \right)
-\left( \sin\theta\partial_r+\frac{\cos\theta}{r}\partial_\theta \right)=-ie^{-i\theta}\partial_r
-\frac{e^{-i\theta}}{r}\partial_\theta$, $k_+=-i\partial_x+\partial_y=-i\left( \cos\theta\partial_r-\frac{\sin\theta}{r}\partial_\theta \right)
+\left( \sin\theta\partial_r+\frac{\cos\theta}{r}\partial_\theta \right)=-ie^{i\theta}\partial_r
+\frac{e^{i\theta}}{r}\partial_\theta$ and $k_x^2+k_y^2=-\left( \frac{\partial^2}{\partial r^2}+\frac{1}{r}\frac{\partial}{\partial r}+\frac{1}{r^2}\frac{\partial}{\partial \theta^2}\right)$. The above Hamiltonian has in-plane rotation symmetry along z axis and the corresponding total angular momentum can be defined as $J_z=L_z+\frac{1}{2}\sigma_z$ where $L_z=-i\frac{\partial}{\partial\theta}$ and the Pauli matrix $\sigma_z$ denotes the spin part. With the in-plane rotation symmetry, the wavefunction ansatz can be taken as $\tilde{\psi}(r,\theta)=\left[ e^{in\theta}f_1(r),e^{i(n+1)\theta}f_2(r),e^{in\theta}f_3(r),e^{i(n+1)\theta}f_4(r) \right]^T$ where the total angular momentum $J_z=n+\frac{1}{2}$. The Hamiltonian is changed to
$$\begin{aligned}
\tilde{H}=\left(
\begin{array}{cccc}
\tilde{\mathcal{M}}(n)+U_0&-iW_0&-iB_0k_z&A_0\left( \partial_r+\frac{n+1}{r} \right)\\
iW_0&\tilde{\mathcal{M}}(n+1)-U_0&A_0\left( -\partial_r+\frac{n}{r} \right)&-iB_0k_z\\
iB_0k_z&A_0\left( \partial_r+\frac{n+1}{r} \right)&-\tilde{\mathcal{M}}(n)+U_0&iW_0\\
A_0(-\partial_r+\frac{n}{r})&iB_0k_z&-iW_0&-\tilde{\mathcal{M}}(n+1)-U_0
\end{array}
\right)
\label{eq:H1}\end{aligned}$$
where $\tilde{\mathcal{M}}(n)=M_0+M_1k_z^2-M_2\left( \frac{\partial^2}{\partial r^2}+\frac{1}{r}\frac{\partial}{\partial r}-\frac{n^2}{r^2} \right)$ and the wave function is now given by $\tilde{\psi}=\left[ f_1(r),f_2(r),f_3(r),f_4(r) \right]^T$. Let’s introduce the new wave function $\psi$ as $\tilde{\psi}=\frac{1}{\sqrt{r}}\psi$, then the normalization relation $\int rdrd\theta |\tilde{\psi}|^2=1$ is changed to $\int drd\theta |\psi|^2=1$, and the effective Hamiltonian is rewritten as
$$\begin{aligned}
&&H=\left(
\begin{array}{cccc}
\mathcal{M}(n)+U_0&-iW_0&-iB_0k_z&A_0\left( \partial_r+\frac{n+1/2}{r} \right)\\
iW_0&\mathcal{M}(n+1)-U_0&A_0\left( -\partial_r+\frac{n+1/2}{r} \right)&-iB_0k_z\\
iB_0k_z&A_0\left( \partial_r+\frac{n+1/2}{r} \right)&-\mathcal{M}(n)+U_0&iW_0\\
A_0(-\partial_r+\frac{n+1/2}{r})&iB_0k_z&-iW_0&-\mathcal{M}(n+1)-U_0
\end{array}
\right)
\label{eq:H2}\end{aligned}$$
where $\mathcal{M}(n)=M_0+M_1k_z^2-M_2\left( \frac{\partial^2}{\partial r^2}-\frac{n^2-1/4}{r^2} \right)$. This Hamiltonian can be written in a compact form
$$\begin{aligned}
&&H=\left[ M_0+M_1k_z^2-M_2\left( \frac{\partial^2}{\partial r^2}-\frac{(n+1/2)^2}{r^2} \right) \right]\tau_z-M_2\frac{n+1/2}{r^2}\sigma_z\tau_z\nonumber\\
&&+B_0k_z\tau_y+iA_0\frac{\partial}{\partial r}\sigma_y\tau_x+A_0\frac{n+1/2}{r}\sigma_x\tau_x+W_0\sigma_y\tau_z+U_0\sigma_z.
\label{eq:H3}\end{aligned}$$
We can discretize the Hamiltonian (\[eq:H2\]) and solve the eigenstate problem for the radial equation numerically. The corresponding result is shown in Fig 2 of the main text. For $n=0$, we indeed find two gapless modes with the opposite velocities along z direction, and these two gapless modes are spatially separated with one wave function mainly staying at $r=0$ and the other one at $r=R$, as shown by the red and black lines in Fig 2 (a) and (b) of the main text. However with increasing n, a gap is opened between the two low energy modes, as shown in Fig 2 (c) in the main text. To get more analytical understanding of the radial equation, we consider the $r\rightarrow \infty$ limit with $U_0=0$, $k_z=0$, where the radial Hamiltonian is simplified as $H=\left( M_0-M_2\frac{\partial^2}{\partial r^2}\right)\tau_z
+W_0\sigma_y\tau_z+iA_0\sigma_y\tau_x\frac{\partial}{\partial r}$. With the wave function ansatz $\psi\sim e^{\lambda r}\phi$, we obtain the equation $A_0\lambda\phi=\left[ \left( M_0-M_2\lambda^2 \right)\sigma_y\tau_y+W_0\tau_y \right]\phi$ for the zero modes. Since $\left[ \sigma_y\tau_y,\tau_y \right]=0$, we can take the common eigen-states of $\sigma_y\tau_y$ and $\tau_y$ for $\phi$, $\sigma_y\tau_y\phi_{ts}=t\phi_{ts}$ and $\tau_y\phi_{ts}=s\phi_{ts}$, then the wave function can be expressed as $\psi=\sum_{\alpha,t,s}c_{\alpha,ts}e^{\lambda_{\alpha}(t,s)r}\phi_{t,s}$, with $\lambda$ given by $\lambda_{\alpha}(t,s)=\frac{-tA_0+\alpha\sqrt{A_0^2+4M_2\left( tsW_0+M_0 \right)}}{2M_2}$. The existence of the edge mode requires $\lambda_+(+,+)
\lambda_-(+,+)>0$ or $\lambda_+(+,-)\lambda_-(-,-)>0$, leading to the following different regimes: in the normal regime $M_0M_2>0$, the system has no zero mode when $|W_0|<|M_0|$ and one zero mode when $|W_0|>|M_0|$, while in the inverted regime $M_0M_2<0$, the system has one zero mode when $|W_0|>|M_0|$ and two zero modes when $|W_0|<|M_0|$. Taking into account the $k_z$ dependent term, it turns out that one zero mode case corresponds to the 1D chiral state and two zero modes case is the 1D helical state. However since time reversal is broken in the present system, the helical state is not protected and can be gapped. Therefore the only robust state is the chiral state when $|W_0|>|M_0|$. We emphasize that the transition at $|W_0|=|M_0|$ exactly corresponds to the condition for the appearance of the gapless Weyl fermions for the uniform magnetization. For the finite r, the terms proportional to $\frac{1}{r}$ and $\frac{1}{r^2}$ will push the chiral mode around r=0 outwards, thus with increasing the angular momentum number n, the wave function of the chiral mode near $r=0$ extends to the large r region and mixes with the chiral mode at $r=R$, opening a gap.
[36]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ** (, , ).
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ** (, , ).
, , , , ****, ().
, , , , , ****, ().
, ****, ().
, , , ****, ().
, ().
, , , , ().
, ().
, , , ****, ().
, , , ****, ().
, ().
, ****, ().
, ****, ().
, ** (, , ).
, , , , , , , ****, ().
, , , , , , ****, ().
, , , , , , , , , , , ****, ().
, , , , , , , , , , , ****, ().
, , , , , , , , , , , ().
, , , , , , , , , , pp. ().
, , , , , , ****, ().
, , , , , , ****, ().
, ****, ().
, **, .
, ** (, , ).
, ****, ().
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We adduce the results of the condensate fraction calculation for liquid helium-4. The method is derived from the first principles and involves minimum assumptions. The only experimental quantity we need in our calculations is the static structure factor which is easily measurable. We use the approximation in which expressions contain one summation in the wave vector space (or one integration it the radius vector space) in order to demonstrate the validity of our method. The computed values of the condensate fraction lie within $8.8\div14\%$ depending on the model potential for the short-range interaction. This result agrees with recent experimental measurements and numerical estimations.
[**Keywords:**]{} liquid helium, Bose-condensate, condensate fraction
PACS: 64.40.–w, 67.40.Kh
author:
- |
A. A. Rovenchak[^1] and I. O. Vakarchuk[^2]\
Department for Theoretical Physics,\
Ivan Franko Lviv National University,\
12 Draghomanov Str., Lviv, Ukraine, UA–79005
title: 'Calculation of the Condensate Fraction in Liquid Helium-4'
---
Introduction
============
The phenomenon of Bose–Einstein condensation and superfluidity in Bose-systems is a subject of great interest. Superfluidity in liquid helium-4 can be easily observed now but the measurements on the condensate fraction remain a difficult problem because of the strong interactions in this fluid. Researchers, both experimentalists and theorists, now generally concur that $\sim10\%$ of atoms are in the lowest (zero momentum) state at zero temperature. The relative number of atoms with zero momentum is called condensate fraction, here we denote it as $f$.
The problem of the calculation of condensate fraction in liquid helium-4 was first considered in the classical paper by Penrose and Onsager [@PenroseOnsager]. The authors used a very rough approximation of hard spheres for the ground-state wavefunction and calculated this quantity as approximately 0.08 or 8%.
During recent twenty years different methods for measurements and theoretical calculations of the condensate fractions were used. Neutron scattering is used to draw the information on the momentum distribution and hence on the condensate fraction. Sears [*et al*]{} [@SearsSvensson82] obtained the value 13.9%. Later, similar value of 13.3% was obtained by Sears [@Sears83] using the data on the temperature variations of the average kinetic energy.
Numerical study based on the variational method were carried out in the series of works by Manousakis, Pandharipande [*et al*]{} [@Manousakis85; @Manousakis85a; @Manousakis91] by means of Jastrow wavefunction with three-body corrections. Their results vary within $8.2\div10.3\%$.
The numerical estimations of the condensate fractions were made also in the series of works by Vakarchuk [*et al*]{} and Vakarchuk [@Vakarchuk87; @Vakarchuk90UFZh; @Vakarchuk90TMF] leading to the results of $3.7\div8.8\%$. The methods used in these papers are based on the mean-spherical approximation for the structure functions of liquid helium-4 at zero temperature [@Vakarchuk87], direct quantum-mechanical computations from the first principles [@Vakarchuk90UFZh; @Vakarchuk90TMF].
The value of approximately 10% was obtained for $f$ by Sokol and collaborators by means of deep-inelastic neutron scattering at high-momentum transfer [@Sokol89; @Sokol90; @Sokol91].
Monte Carlo (MC) methods were also used to study the question of the condensate fraction. The result of Whitlock and Panoff [@Whitlock87] is $\simeq9\%$ (Green Function MC). Ceperley and Pollock [@Ceperley87] obtained value $\simeq7\%$ using Path Integral MC technics at 1.18 K. Also, Moroni [*et al*]{} [@Moroni97] gives 7.2% from Diffusion MC simulations.
Approximately in the same time the experimental result of $\simeq6\%$ was obtained by Azuah [*et al*]{} [@Azuah97].
Series of papers by Mayers [*et al*]{} [@Mayers96; @Mayers97] was devoted to both calculation of the condensate fraction based on phenomenological assumptions [@Mayers96] (result is 9.9%) and measurements on the high-energy scattering [@Mayers97] ($f=15\pm4\%$ at 1.3 K).
Recent high precision measurements of the dynamic structure factor allowed Glyde [*et al*]{} [@Glyde00] to derive the information on the condensate fraction at zero temperature and its temperature dependence. They report $f=7.25\pm0.75\%$ at 0 K.
A semi-phenomenological method for extracting the condensate fraction data was proposed recently by Rinat and Taragin [@RinatTaragin01]. Authors obtained $f$ at different temperatures and extrapolated their results to 0 K as $f=9.0\pm0.3\%$.
The aim of this work is to show a possibility of calculation of the condensate fraction without drawing any special huge computational efforts.
The method we use was worked out in [@Vakarchuk90TMF]. The condensate fraction is calculated directly from the single-particle distribution function $F_1(R)$ as its long-range limit: $$f=N_0/N=\lim_{R\to\infty} F_1(R)$$ where $N_0$ is the number of particles in zero momentum state and $N$ is the total number of particles in the system.
Our technics do not involve any phenomenological assumptions. For the computations we use as input only one experimental quantity, the static structure factor of helium at zero temperature. It is the advantage comparing to those methods for which such quantities as the dynamic structure factor is necessary. Its precise measurement is much more complicated problem comparing with that of the static structure factor.
The method is fully controlled, i. e., we always know which effects are neglected, and it is not difficult to take them into account in the next approximation (in principle). Here we present only the first approximation (RPA) but the expressions are easily extendable in order to take subtler effects like many-particle interactions.
The only key problem arising while extending our method further is the knowledge of the static three-particle structure factor. Unlike usual structure factor, it is not an easily measurable quantity already. One of the ways for its calculation we see in such a scheme: expressions for the three-particle distribution function $F_3$ (beyond superposition approximation) containing pair distribution function $\to$ 6-dimensional Fourier transformation of $F_3$ leading to the three-particle structure factor. A separate paper will be devoted to this problem.
We consider three different approaches for the short-range repulsive interactions. In the first one, they are not taken it into account explicitly at all. In the second approach this repulsion is modeled by Meyer’s function $e^{(A/R)^n}-1$, $R$ is the radius-vector absolute value. We chose $A=2.1$ Å and $n=12$. This potential is referred here as ‘almost hard spheres’ (AHS) [@Ours2001]. The third model is the potential of hard spheres (HS) with diameter 2.1 Å.
Calculating procedure
=====================
Principal calculating procedure consists of the following steps:
1. [**Input**]{}: Experimental data on the structure factor [@SvenssonSears80] \[ProcInput\] (should be converted into the zero temperature [@Ours2000]);
2. [**Input**]{}: Model for the short-range interactions \[ProcModel\] (HS, AHS or none) [@Ours2001];
3. [**Output**]{}: Structure factor of the model system corresponding to the short-range repulsion; \[Proc\_Sqsr\]
4. [**Output**]{}: Fourier transformation of the interatomic potential \[Proc\_nuq\] (effective) [@Ours2000; @Ours2001];
5. [**Output**]{}: Pair distribution function $F_2(R)$ \[Proc\_F2\] [@SvenssonSears80];
6. [**Result**]{}: Condensate fraction. \[Proc5\]
In this work, we propose a simple method for the calculation of the condensed fraction using formulae obtained earlier in [@Vakarchuk90UFZh]: $$\begin{aligned}
\label{Initial}
&&f=\exp\left(I_{1A}+\Delta_1J+\ldots\right),\\ \nonumber
&&I_{1A}=-{1\over4N}\sum_{{\bf q}\neq0}(\alpha_q-1)^2/\alpha_q,\\
\nonumber
&&\Delta_1J=\varrho \int d{\bf R} \left[2 h^*(R)-h(R)+h^2(R)/4 \right],\nonumber\\
&&h(R)=F_2(R)-1, \qquad h^*(R)=\sqrt{F_2(R)}-1,\nonumber\end{aligned}$$ dots denote terms having more than one integration over wave vector or coordinate, $F_2(R)$ is the pair distribution function (PDF), $N$ is the number of atoms, and $$\alpha_q=\sqrt{1+2\varrho\nu_q\biggl/{\hbar^2 q^2\over2m}},$$ $\varrho$ is helium density, $\varrho=0.02185$ Å$^{-3}$, $m$ is the mass of helium atom, $m=4.0026$ a. m. u., $q$ is the wave vector.
The expression for PDF within the accepted approach was obtained in [@Vakarchuk90UFZh]: $$\label{F2def}
F_2(R)=\exp\left\{
{1\over N}\sum_{{\bf q}\neq0}{2a_2({\bf q})\over1-2a_2({\bf q})}
e^{i{\bf qR}}+\dots
\right\},$$ while the self-consistent description of short- and long-range correlations in liquid helium-4 results in such a result for the function $a_2$ [@Vakarchuk79]: $$\label{a2def}
a_2({\bf q})={1\over2}\left( {1\over S^{\rm s.r.}_q} - {1\over S_q} \right)$$ with $S^{\rm s.r.}_q$ being the structure factor of the model system (it equals 1 if the short-range repulsion is not given explicitly), superscript ‘s. r.’ corresponds to ‘short-range’. $S_q$ is the experimental structure factor (at 0 K). These expressions contain $\nu_q$ being the Fourier image of the interatomic potential in helium that has an effective nature. It is obtained from the first principles using the collective variables formalism in the Schrödinger equation, as described in [@Ours2000].
On the step (\[Proc\_nuq\]) of the calculation procedure we encountered problems while considering the HS short-range potential. The reason is clear: a weak damping of the structure functions at large values of wave vector or radius vector requires much more careful numerical computations comparing to those in two other cases.
Some usual problems appear on the step (\[Proc\_F2\]) of the calculating procedure where we smoothed $F_2(R)$ calculated from (\[F2def\]) and truncated it at the distances $R<2.2$ Å. We put $F_2(R<2.2\ {\rm\AA})=0$ in order to obtain a correct PDF at small distances [@Robkoff81]. We present the calculated PDFs in comparison with the experimental (being more precise, derived from the measurements on the liquid structure factor) one [@Robkoff81] in the Fig. \[F2fig\].
![Pair distribution function. The dashed line corresponds to the experimental data by Robkoff and Hallock [@Robkoff81] at 1.38 K. The dotted line represents the smoothed data in the case when no explicit short-range repulsion is given. The dashed-dotted and solid lines are PDFs for HS and AHS short-range potentials respectively.[]{data-label="F2fig"}](F2.eps){width="75mm"}
Results and discussion
======================
We have calculated the condensate fraction at zero temperature using different models for the short-range repulsive part of the potential.
The calculated values of $f$ are:
- No explicit short-range potential $f=14\%$;
- AHS short-range potential $f=11\%$;
- HS short-range potential $f=8.8\%$.
We expect the real value being about $10\div30\%$ less than the calculated one due to the contribution of the higher-order terms, as described in [@Vakarchuk90UFZh]. But in the first approximation we consider it to be a good accuracy for $f$ in liquid helium-4.
Our result for the condensate fraction is within the range of theoretical, numerical and experimental estimations mentioned here in the introductory part of the paper. Since our expressions are derived from the first principles, one can use the number for $f$ calculated in this way as a test for the condensate fraction information indirectly extracted from the experimental measurements. This becomes possible after the next one and two approximations are computed.
We found that the short-range repulsive part of the potential has a significant influence on the results for the condensate fraction. Namely, when the short-range interactions are included into the consideration explicitly, the results appear to be essentially dependent on the “hardness” of the core. A proper model for this interaction will be found when more terms are taken into account in the expressions (\[Initial\]), (\[F2def\]). The reason is that in principle the summation of the whole series corresponds to the correct calculation of the short-range interactions [@Vakarchuk79]. On the other hand, the condensate fraction in liquid helium-4 is a very sensitive (with respect to the order of approximation) quantity. Thus, we can expect that having found such a model we obtain a possibility to use simple RPA-like expressions instead of higher-order ones. By now, the AHS ($n=12$, $A=2.1$ Å) potential seems to be the most suitable for this purpose.
[9]{}
O. Penrose, L. Onsager, Phys. Rev. [**104**]{}, 576 (1956). V. F. Sears, E. C. Svensson, P. Martel, A. D. B. Woods, Phys. Rev. Lett. [**49**]{}, 279 (1982). V. F. Sears, Phys. Rev. B [**28**]{}, 5109 (1983). E. Manousakis, V. R. Pandharipande, Q. N. Usmani, Phys. Rev. B [**31**]{}, 7022 (1985). E. Manousakis, V. R. Pandharipande, Phys. Rev. B [**31**]{}, 7029 (1985). E. Manousakis, V. R. Pandharipande, Q. N. Usmani, Phys. Rev. B [**43**]{}, 13587 (1991). I. A. Vakarchuk, V. M. Migal, Ukr. Fiz. Zhurn [**32**]{}, 786 (1987). I. A. Vakarchuk, Ukr. Fiz. Zhurn. [**35**]{}, 1261 (1990). I. A. Vakarchuk, Theor. Math. Phys. [**80**]{}, 983 (1989); [**82**]{}, 308 (1990). T. R. Sosnick, W. M. Snow, P. E. Sokol, R. N. Silver, Europhys. Lett. [**9**]{}, 707 (1989). T. R. Sosnick, W. M. Snow, P. E. Sokol, Phys. Rev. B [**41**]{}, 707 (1990). T. R. Sosnick, W. M. Snow, R. N. Silver, P. E. Sokol, Phys. Rev. B [**43**]{}, 216 (1991). P. A. Whitlock, R. M. Panoff, Can. J. Phys. [**65**]{}, 1409 (1987). D. M. Ceperley, E. L. Pollock, Can. J. Phys. [**65**]{}, 1416 (1987). S. Moroni, G. Senatore, S. Fantoni, Phys. Rev. B [**55**]{}, 1040 (1997). R. T. Azuah, W. G. Stirling, H. R. Glyde, M. Boninsegni, P. E. Sokol, S. M. Bennington, Phys. Rev. B [**56**]{}, 14620 (1997). J. Mayers, Rutherford Appleton Laboratory Technical Report RAL–TR–96–031 (1996). J. Mayers, C. Andreani, D. Colognesi, J. Phys.: Condens. Matter [**9**]{}, 10639 (1997). H. R. Glyde, R. T. Azuah, W. G. Stirling, Phys. Rev. B [**62**]{}, 14337 (2000). A. S. Rinat, M. F. Taragin, J. Low. Temp. Phys. [**123**]{}, 139 (2001). I. O. Vakarchuk, A. A. Rovenchak, J. Phys. Stud. (2001) (to be published); see also cond-mat/0104008. E. C. Svensson, V. F. Sears, A. D. B. Woods, P. Martel, Phys. Rev. B [**21**]{}, 8 (1980). I. O. Vakarchuk, V. V. Babin, A. A. Rovenchak, J. Phys. Stud. [**4**]{}, 16 (2000). I. A .Vakarchuk, I. R. Yukhnovskii, Theor. Math. Phys. [**40**]{}, 626 (1979). H. N. Robkoff, R. B. Hallock, Phys. Rev. B [**24**]{}, 159 (1981).
[^1]: [email protected]
[^2]: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We prove that the Shilov boundary is invariant under proper holomorphic mappings between some classes of domains (containing among others quasi-balanced domains with the continuous Minkowski functionals). Moreover, we obtain an extension theorem for proper holomorphic mappings between quasi-circular domains.
Using these results we show that there are no non-trivial proper holomorphic self-mappings in the tetrablock. Another important result of our work is a description of Shilov boundaries of a large class of domains (containing among other the symmetrized polydisc and the tetrablock).
It is also shown that the tetrablock is not $\mathbb C$-convex.
address: |
Instytut Matematyki\
Uniwersytet Jagielloński\
Łojasiewicza 6, 30-348 Kraków\
Poland
author:
- Łukasz Kosiński
title: 'Geometry of quasi-circular domains and applications to tetrablock'
---
Introduction and statement of results
=====================================
In the paper we will use the notion of *quasi-circular* domains. Let $m_1,\ldots,m_n$ be relatively prime natural numbers. Recall that a domain $D\subset\mathbb C^n$ is said to be $(m_1,\ldots,m_n)$-circular (shortly *quasi-circular*) if $$\label{defcir} (\lambda^{m_1}x_1,\ldots,\lambda^{m_n}x_n) \in D\quad \text{for any}\quad |\lambda|=1,\ x=(x_1,\ldots,x_n)\in D.$$ If the relation (\[defcir\]) holds with $|\lambda|\leq 1,$ then $D$ is said to be $(m_1,\ldots,m_n)$-balanced (shortly *quasi-balanced*).
Let $\mathcal R_{II}$ denote the classical Cartan domain of the second type, i.e. $$\mathcal R_{II}=\{z\in\mathcal{M}_{2\times2}(\mathbb C):\ z=z^t,\ ||z||<1\},$$ where $||\cdot||$ is the operator norm and $\mathcal M_{2\times 2}(\mathbb C)$ denotes the space of $2\times 2$ complex matrices. Put $$\Pi:\mathcal{M}_{2\times2}(\mathbb C)\ni z=(z_{i,j})\rightarrow (z_{1,1},z_{2,2},\det z)\in\mathbb C^3.$$ We define $\mathbb E:=\Pi(\mathcal R_{II}).$ The domain $\mathbb E$ is called the *tetrablock*.
The tetrablock is a $(1,1,2)$-balanced domain in $\mathbb C^3$ appearing in control engineering and produces problems of a function-theoretic character. Its geometric properties have been investigated in several papers (see e.g. [@You0], [@Zwo], [@You] and references contained there). Recall here that in [@You] the author using Kaup’s theorem obtained a description of the group of automorphisms of this domain. In the paper we prove an Alexander-type theorem for the tetrablock showing that every proper holomorphic self-map of the tetrablock is an automorphism.
\[main\] Let $\varphi:\mathbb E\to\mathbb E$ be a proper holomorphic mapping. Then $\varphi$ is an automorphism.
As a side effect we obtain a natural correspondence between automorphisms of the tetrablock and of the classical domain of the second type indicated in Lemma \[lemma2\]. This correspondence gives much easier and more elementary method of deriving the explicit formulas for automorphisms of the tetrablock. In particular, we extend results from [@You] and simplify their proofs.
The methods used in the paper rely upon the investigation of proper holomorphic mappings between quasi-circular domains. We start with generalizing the Bell’s extension result (see [@Bel]). Next we analyze the behavior of the Shilov boundary under proper holomorphic mappings. We have the following
\[shpr\]Let $D$ and $G$ be bounded domains in $\mathbb C^n$ and let $f:D\to G$ be a proper holomorphic mapping extending continously to $\overline D.$ Assume that there is an increasing family of domains $\{G_m\},$ $G_m\Subset G_{m+1},$ such that $\bigcup G_m=G$ and $\overline{(\bigcup_m \partial_s G_m)}\cap\partial G=\partial_s G.$
Then $f(\partial_s D)=\partial_s G.$
Note that in general the Shilov boundary is not invariant even under biholomorphic polynomial mappings - see Example \[example\].
Based on the former idea we also obtain the following result:
\[lempr\]Let $f:D\to G$ be a proper holomorphic mappings between domains in $\mathbb C^n.$ Let $L$ be a domain relatively compact in $G.$ Put $K=f^{-1}(L).$ Then $$f(\partial_s K)=\partial_s L\quad \text{and}\quad f(\partial_b K)=\partial_b L.$$
As a consequence of our considerations we show that any proper holomorphic mapping between quasi-balanced bounded domains preserves the Shilov boundary. Namely, we have the following
\[cor\] *a)* Let $D\subset\mathbb C^n$ be a bounded domain and let $G$ be a bounded quasi-balanced domain in $\mathbb C^n.$ Assume that the Minkowski functional associated to $G$ is continuous and for any open, relatively compact subset $K$ of $D$ there is an open neighborhood $U$ of $\overline D$ such that $k_D(z,\overline w)$ extends holomorphically on $U\times \tilde K.$
Then every proper holomorphic mapping $f:D\to G$ maps $\partial_s D$ onto $\partial_s G.$
b\) Let $D$ and $G$ be bounded quasi-balanced domains. If the Minkowski functionals of $D$ and $G$ are continuous, then any proper holomorphic mapping between $D$ and $G$ preserves the Shilov boundary.
As we indicate in the sequel the results obtained in the paper give immediately a description of the Shilov boundaries of many domains like the symmetrized polydisc (see [@Edi-Zwo]), the tetrablock (see [@You0]) etc. Moreover, they exclude the existence of proper holomorphic mappings between some domains. For example, the well known theorem stating that there is no proper holomorphic mapping between the polydisc and the Euclidean ball is a direct consequence of our results.
As a by-product of our considerations we obtain in Lemma \[bih\] an extension of the main result from [@Tum-Hen].
In this paper in Remark \[rem\] it is also shown that the tetrablock is not $\mathbb C$-convex. Recall that a consequence of the Lempert theorem is the fact that the Carathéodory pseudodistance and the Lempert function of a $\mathbb C$-convex domain with $\mathcal C^2$ boundary coincide (see [@Jac]). Since results obtained in [@You0] (see also [@Zwo]) suggest that the equality between the Carathéodory pseudodistance and the Lempert function holds in the tetrablock, the tetrablock is the candidate for the first bounded pseudoconvex domain non-biholomorphically equivalent to a $\mathbb C$-convex domain for which the equality between mentioned above functions holds.
It also seems to be interesting whether the tetrablock may be exhausted by domains biholomorphic to $\mathbb C$-convex domains.
Here is some notation. Throughout the paper $\mathbb D$ denotes the unit disc in the complex plane. The unit Euclidean ball in $\mathbb C^n$ is denoted by $\mathbb B_n.$ Moreover $\prop(D,G)$ is the set of proper holomorphic mappings between domains $D$ and $G$. The Shilov and Bergman boundary is denoted respectively by $\partial_s$ and $\partial_b.$
Now I would like to thank professor Włodzimierz Zwonek for reading the manuscript, many remarks and fruitful discussions.
Extension of proper holomorphic mappings between quasi-balanced domains
=======================================================================
We start this section with recalling basic properties of circular domains and the Bergman projection which will be useful in the sequel. By $k_D$ we shall denote the Bergman kernel associated to a domain $D.$ Let moreover $P_D$ denote the Bergman projection for $D.$ We use the notation $$k_D^{\alpha}(z,w)=\partial^{\alpha}k_D(z,w)\quad \text{and}\quad k_D^{\overline{\alpha}}(z,w)=\partial^{\overline{\alpha}}k_D(z,w),$$ where $\partial^{\alpha}$ stands for $\frac{\partial^{|\alpha|
}}{z^{\alpha}}$ and $\partial^{\overline{\alpha}}$ stands for $\frac{\partial^{|\alpha|}}{\overline w^{\alpha}}.$
For a given $(m_1,\ldots,m_n)$-balanced domain $D$ in $\mathbb C^n,$ where $m_1,\ldots,m_n$ are relatively prime natural numbers, we define the *Minkowski functional* $$\mu_D(x):=\inf\{\lambda>0:\ (\lambda^{-m_1}x_1,\ldots,\lambda^{-m_n}x_n)\in D\},\quad x=(x_1,\ldots,x_m)\in\mathbb C^n.$$ This function has similar properties as the standard Minkowski functional for balanced domains. Some of them may be found in [@Nik]. In particular, $$\begin{aligned}
&\mu_D(\alpha^{m_1}x_1,\ldots,\alpha^{m_n}x_n)=|\alpha|\mu_D(x),\quad x\in\mathbb C^n,\ \alpha\in\mathbb C,\\
&D=\{x\in\mathbb C^n:\ \mu_D(x)<1\}.\end{aligned}$$
For a subset $K$ of $\mathbb C^n$ we put $\tilde K:=\{\overline x:\ x\in K\}.$
\[poznauwaga\]Let $D$ be an $(m_1,\ldots,m_n)$-balanced bounded domain whose Minkowski functional is continuous. Put $D_r:=\{x\in\mathbb C^n:\ \mu_D(x)<r\},$ $r>0.$ Since $$k_D((r^{m_1}z_1,\ldots, r^{m_n} z_n),w)=k_D(z,(r^{m_1}w_1,\ldots, r^{m_n}w_n))$$ for $z,w\in D,$ $r\in [0,1],$ we easily find that the function $(z,w)\to k_D(z,\overline w)$ may be extended holomorphically to $D_{1/r}\times \tilde{D_r}$ for any $0<r\leq 1.$
It follows from [@Bel0] that if $f:D\to G$ is a proper holomorphic mapping between bounded domains $D,$ $G$ in $\mathbb C^n,$ then for any $\Phi\in L^2(G)$ we have $$\label{row}P_{D}(\det[f']\cdot(\Phi\circ f))=\det[f'] \cdot ((P_G\Phi)\circ f).$$
Assume additionally that $G$ is an $(m_1,\ldots,m_n)$-circular domain containing the origin. Choose $\delta>0$ such that $\delta\overline{\mathbb B_n}\subset G.$ Let $\theta$ be a radial function in $\mathcal{C}^{\infty}_0(\delta\mathbb B_n)$ such that $\theta\geq 0$ and $\int_{\delta\mathbb B_n}\theta=1$. Since holomorphic functions assume their average values we find that $$\label{3}\partial^{\alpha}h(0) = \int_{G}(\partial^{\alpha}h)\theta d\lambda^{2n}=\int_{G} h(-1)^{|\alpha|} \partial^{\alpha}\theta d\lambda^{2n}$$ for every $h\in\mathcal O(G)\cap L^2(G).$
On the other hand $h(z)=\int_{G}k_{G}(z,w)h(w) d\lambda^{2n}(w),$ $z\in G.$ Since $k_G(z,\cdot)$ extends holomorphically to a neighborhood of $\overline G$ provided that $z$ is sufficiently close to $0$, one may differentiate this formula at $z=0$ to get that $$\label{ker0}\partial^{\alpha}h(0)=\int_{G}\partial^{\alpha}k_{G}(0,w)h(w) d\lambda^{2n}(w),\quad h\in\mathcal O(G) \cap L^2(G).$$ This relation together with (\[3\]) gives $$\label{ker}P_G((-1)^{|\alpha|}\overline{\partial}^{\alpha}\theta)=k_G^{\overline{\alpha}}(\cdot,0).$$
The next lemma has been proved by S. Bell in the case when $D$ and $G$ are bounded circular domains and $0\in G$ (see [@Bel]). It is interesting that after minor modifications the methods used by Bell yield a stronger result. We present the whole proof for the sake of completeness.
\[text\]Let $D,G$ be bounded domains in $\mathbb C^n.$ Suppose that $G$ is $(m_1,\ldots,m_n)$-circular and contains the origin. Assume moreover that the domain $D$ satisfies the following property: for any open, relatively compact subset $K$ of $D$ there is an open set $U$ containing $\overline D$ such that $(z,w)\to k_D(z,\overline w)$ extends holomorphically to $U\times \tilde K.$
Then any proper holomorphic mapping $f:D\to G$ extends holomorphically to a neighborhood of $\overline{D}.$
Let $m=(m_1,\ldots,m_n).$ Properties of the Bergman kernel and a standard argument imply that the equation $$k_G((\lambda^{m_1}z_1,\ldots, \lambda^{m_n} z_n),w)=k_G(z,(\overline{\lambda^{m_1}}w_1,\ldots, \overline{\lambda^{m_n}}w_n))$$ holds for any $z,w\in G$ and $|\lambda|$ sufficiently close to $1.$ Differentiating this formula several times with respect to $\overline w_i$ and putting $w=0$ we find that $$\frac{\partial^{\alpha} k_G}{\partial\overline w^{\alpha}}((\lambda^{m_1}z_1,\ldots, \lambda^{m_n} z_n),0)=\lambda^{\langle \alpha,m\rangle}\frac{\partial^{\alpha} k_G}{\partial\overline w^{\alpha}}(z,0)$$ for $\alpha\in\mathbb N^n,$ $z\in G$ and $|\lambda|$ sufficiently close to $1.$
Whence a standard argument shows that there are $c_{\beta}\in\mathbb C$ such that $$\label{z1}k_G^{\overline{\alpha}} (z,0)= \sum c_{\beta}z^{\beta},\ z\in G,$$ where the sum is taken over $\beta\in\mathbb N^n$ satisfying the relation $\langle\beta, m \rangle= \langle \alpha, m \rangle.$ Therefore, the linear independence of $k_G^{\overline{\alpha}}(z,0)$ (see (\[ker0\])) implies that for every $\beta\in\mathbb N^n$ there are $\tilde c_{\alpha}$ such that $$\label{gen}z^{\beta}=\sum \tilde c_{\alpha} k_G^{\overline{\alpha}} (z,0),$$ where the sum is taken over $\alpha\in\mathbb N^n$ satisfying the relation $\langle\alpha, m \rangle= \langle \beta, m \rangle.$
Now (\[gen\]) together with (\[ker\]) provide us with the function $\phi_{i,k}\in\mathcal C_0^{\infty}(\delta\mathbb B_n)$ such that $$z_i^k=P_G(\Phi_{i,k}),\quad i=1,\ldots,n,\ k\in\mathbb N.$$ Making use of the above relations we infer that $$\label{ext}\det[f'(z)]f_i^k(z)=\det[f'(z)](z_i^k\circ f(z))=\int_D k_D(z,w)\det[f'(w)]\Phi_{i,k}(f(w)) d\lambda^{2n} (w),$$ for $i=1,\ldots,n,$ and $k\in\mathbb N.$ From these relations and the assumption on $k_D$ we easily conclude that all the functions appearing in the left side of (\[ext\]) extend holomorphically to some open, connected neighborhood $U$ of $\overline D.$
We will briefly show that $f_i$ extends holomorphically to the domain $U.$ Putting $u=\det[f']$ we have the following situation $$u\in\mathcal O(U),\quad u\not\equiv 0 \quad\text{and}\quad uf_i^k\in\mathcal O(U),\ k\in\mathbb N.$$ Fix any point $x\in U$ such that $u(x)=0.$ Changing, if necessary, the coordinates system we may assume that both $u$ and $uf_i$ satisfy the assumptions of Weierstrass Preparation Theorem near $x.$ Since $uf_i^k$ is holomorphic on $U$, the Weierstrass polynomial associated to $u$ divides the Weierstrass polynomial associated to $uf_i.$ This, in particular, means that $f_i$ is locally bounded near the analytic set $\{u=0\},$ so the assertion follows from the Riemann’s removable singularity theorem.
Note that the continuity of the Minkowski functional of a bounded quasi-balanced domain $D$ is equivalent to the fact that for every $0<r<1$ the domain $D$ is relatively compact in $D_{1/r}.$ Therefore any quasi-balanced domain fulfils the assumptions of Lemma \[text\].
\[pext\]Any proper holomorphic mapping $f:\mathcal R_{II}\to\mathbb E$ may be extended holomorphically to a neighborhood of $\overline{\mathcal R_{II}}.$
Proofs of Theorems \[shpr\], \[lempr\] and the applications
===========================================================
We start this section with the following
The technical assumption occurring in the Theorem \[shpr\] seems to be very natural. Observe that $x\in \overline{(\bigcup_m \partial_s G_m)}\cap\partial G$ if and only if there is a subsequence $(n_k)$ and there are $x_{n_k}\in \partial_s G_{n_k}$ such that $x_{n_k}\to x.$
One may very easily show that for any bounded domain $G$ and any increasing family of domains $\{G_m\}$ such that $\bigcup G_m=G,$ the Shilov boundary of $G$ is contained in $\overline{(\bigcup_m \partial_s G_m)}\cap\partial G.$
\[example\] Note that Theorem \[shpr\] does not remain valid if we remove the assumption $\overline{(\bigcup_m \partial_s G_m)}\cap\partial G=\partial_s G$ even in the case when $f$ is a proper polynomial mapping. As an example one may take $D=\mathbb D\cap\{z\in\mathbb D:\ \im z>0\},$ $G=\mathbb D\setminus [0,1)$ and $f(z)=z^2.$
The inclusion $\partial_s G\subset f(\partial_s D)$ follows immediately from the definition of the Shilov boundary. We shall prove that $\partial_s D\subset f^{-1}(\partial_s G).$ Assume the contrary i.e. there is a $\psi\in\mathcal O(D)\cap \mathcal C(\overline D)$ such that $$\label{our}|\psi(x_0)|>\max\{|\psi(x)|:\ x\in f^{-1}(\partial_s G)\},$$ for some $x_0\in\partial D.$ Note that $$\limsup_{m\to\infty}\max\{|\psi(x)|:\ x\in D\cap f^{-1}(\partial_s G_m)\}\leq\max\{|\psi(x)|:\ x\in \partial D\cap f^{-1}(\partial_s G)\}.$$ Actually, otherwise there would exist a subsequence $(m_k)\subset\mathbb N,$ $\epsilon>0$ and $x_{m_k}\in D\cap f^{-1}(\partial_s G_{m_k})$ such that $$|\psi(x_{m_k})|>\max\{|\psi(x)|:\ x\in \partial D\cap f^{-1}(\partial_s G)\}+\epsilon.$$ Passing, if necessary, to a subsequence we can assume that $x_{m_k}$ converges to some $x_0.$ Using the assumptions on the domain $G$ and the mapping $f$ we infer that $f(x_0)\in\partial_s G.$ Thus $$|\psi(x_0)|\geq\max\{|\psi(x)|:\ x\in \partial D\cap f^{-1}(\partial_s G)\}+\epsilon\quad\text{and}\quad x_0\in\partial D\cap f^{-1} (\partial_s G),$$ which gives an obvious contradiction.
Therefore we may take $m$ big enough and replace $x_0$ by a point $x_0'\in f^{-1}(\overline G_m)$ sufficiently close to $x_0$ at which the mapping $f$ is non-degenerate so that $$\label{aa}|\psi(x_0')|>A:=\max\{|\psi(x)|:\ x\in D\cap f^{-1}(\partial_s G_m)\},\quad \#f^{-1}(f(x_0'))=k,$$ where $k$ denotes the multiplicity of $f.$
Let $h_j,\ j=1,\ldots, k,$ be holomorphic mappings in the neighborhood of $f(x_0')$ given by $f^{-1}=\{h_j:\ j=1,\ldots,k\}.$ Making use of (\[aa\]) together with the Kronecker Theorem (see e.g. [@Har]) one may show the existence of a natural number $d$ such that $$\label{aaa}|\psi(h_1(f(x_0')))^d+\ldots +\psi(h_k(f(x_0')))^d|>kA^d.$$ To prove it put $a_j=\psi(h_j(f(x_0'))),$ $j=1,\ldots,k.$ Change, if necessary, the order of $a_j$ so that $|a_1|=\ldots=|a_l|$ and $|a_j|<|a_1|$ for $j=l+1, \ldots, n.$ Dividing all $a_j$ by $\psi(h_1(f(x_0')))$ we reduce ourselves to the following situation: $$a_j=e^{i\theta_j},\ j=1\ldots,l,\quad |a_j|<1,\ j=l+1,\ldots,n\quad \text{and}\quad A<1,$$ where $\theta_j\in \mathbb R,$ $j=1,\ldots,l.$
Changing the order once again we may assume that $1,\theta_1,\ldots ,\theta_{l_1}$ are $\mathbb Q$-linearly independent, $l_1\leq l,$ $l_1\in\mathbb N\cup \{0\}$ and $$\theta_j=\frac{q_{j,0}}{N}+\sum_{\iota=1}^{l_1} \frac{q_{j,\iota}}{N} \theta_{\iota},$$ $j=l_1+1,\ldots,l,$ where $q_{j,\iota}\in\mathbb Z,$ $\iota=0,\ldots,l_1,$ and $N\in\mathbb N.$ Put $M=\max\{|q_{j,\iota}|,N\}.$
According to the Kronecker Theorem (see e.g. [@Har]) there is a sequence of natural numbers $(\tilde d_{\mu})$ such that $-\frac{1}{(\mu+1) k M} < \arg (e^{2\pi i\tilde d_{\mu} \theta_j}) <\frac{1}{(\mu+1) k M},$ $j=1,\ldots,l_1,$ $\mu\in\mathbb N.$ In particular, $-\frac{1}{\mu+1}< \arg(e^{2\pi i d_{\mu}\theta_j})<\frac{1}{\mu+1}$ for $j=1,\ldots,l,$ $\mu\in\mathbb N,$ where $d_{\mu}:=N\tilde d_{\mu}.$
Properties of $(d_{\mu})$ guarantee that $|a_1^{d_{\mu}}+\ldots +a_l^{d_{\mu}}| \to l$ as $\mu\to \infty.$ Since $d_{\mu}\to\infty,$ we find that $kA^{d_{\mu}}\to 0$ and $|a_{l+1}^{d_{\mu}}+\ldots +a_k^{d_{\mu}}| \to 0,$ $\mu\to \infty.$ Therefore $|a_1^{d_{\mu}} +\ldots+a_k^{d_{\mu}}|-kA^{d_{\mu}}\to l>0,$ which obviously proves the existence of a natural number $d$ fulfilling (\[aaa\]).
Put $$\zeta(x)=x_1^d+\ldots+ x_k^d,\quad \text{for}\quad x=(x_1,\ldots,x_k)\in\mathbb C^k.$$ A well known argument shows that the formula $\varphi=\zeta\circ(\psi\times\ldots\times\psi)\circ f^{-1}$ defines a holomorphic function on $G.$ It follows from (\[aa\]) and (\[aaa\]) that $$|\varphi(f(x_0'))|>\max\{|\varphi(x)|:\ x\in \partial_s G_m\};$$ a contradiction.
\[glupiauwaga\]It is clear that the proof remains valid if the assumption $\overline{(\bigcup_m \partial_s G_m)}\cap\partial G=\partial_s G$ occurring in Theorem \[shpr\] is replaced by a weaker condition $\overline{(\bigcup_m \partial_b G_m)}\cap\partial G=\partial_s G,$ where $\partial_b$ denotes the Bergman boundary.
The following result is a direct consequence of Theorem \[shpr\]:
Let $D$ and $G$ be bounded domains in $\mathbb C^n$ and let $f:D\to G$ be a proper holomorphic mapping extending holomorphically to a neighborhood of $\overline D.$ Assume that there is an increasing family of domains $\{G_m\},$ $G_m\Subset G_{m+1},$ such that $\bigcup G_m=G$ and $\overline{(\bigcup_m \partial_b G_m)}\cap\partial G=\partial_b G.$
Then $f(\partial_b D)=\partial_b G.$
It follows from Remark \[glupiauwaga\] that $\partial_s G=\partial_b G.$ Thus, Theorem \[shpr\] together with Remark \[glupiauwaga\] gives: $$\partial_b D\subset \partial_s D\subset f^{-1}(\partial_s G)=f^{-1}(\partial_b G).$$ The inclusion $\partial_b G\subset f(\partial_b D)$ may be shown as in the proof of Theorem \[shpr\].
The inclusions $\partial_s L\subset f(\partial_s K)$ and $\partial_b L\subset f(\partial_b K)$ are clear.
It remains to show that $\partial_s K\subset f^{-1}(\partial_s L)$ and $\partial_b K\subset f^{-1}(\partial_b L).$ We will prove both inclusion simultaneously. Assume a contrary, i.e. there is a function $\psi\in \mathcal O(K)\cap \mathcal C(\overline G)$ (respectively $\psi\in \mathcal O(\overline K)$) such that $|\psi|$ does not attain its maximum on $f^{-1}(\partial_b L),$ i.e. $$|\psi(x_0)|>A:=\max\{ |\psi(x)|:\ x\in f^{-1}(\partial_s L)\}$$ (resp. $|\psi(x_0)|>A:=\max\{ |\psi(x)|:\ x\in f^{-1}(\partial_b L)\}$) for some $x_0\in K.$ Obviously $f(x_0)$ may be assumed to be a regular value of $f.$
Let $k$ denote the multiplicity of the mapping $f:D\to G.$ Clearly, $f|_K:K\to L$ is also of multiplicity $k.$
Write $f^{-1}=\{h_1, \ldots, h_k\}$ in a neighborhood of $a$, where $h_i$ are holomorphic functions. One may repeat the argument used in the proof of Theorem \[shpr\] to show the existence of a natural number $d$ such that $$\label{aaa2}|\psi(h_1(f(x_0)))^d+\ldots +\psi(h_k(f(x_0)))^d|>kA^d.$$ Define $\zeta(x)=x_1^d+\ldots+ x_k^d,$ for $x=(x_1,\ldots,x_k)\in\mathbb C^k.$ A function $\varphi$ given by the formula $\varphi=\zeta\circ(\psi\times\ldots\times\psi)\circ f^{-1}$ is holomorphic $L$ and continuous on $\overline{f^{-1}(L)}=f^{-1}(\overline L)$ (resp. $\varphi$ is holomorphic in an open neighborhood of $\overline L$). It follows from (\[aaa2\]) that $$|\varphi(f(x_0))|>\max\{|\varphi(x)|:\ x\in \partial_s L\};$$ a contradiction.
*a)* Define $$G_{m}:=\left\{x\in \mathbb C^n:\ \mu_G(x)<1-\frac{1}{m}\right\},\quad m=2,3\ldots.$$ It is clear that the family $\{G_m\}_m$ satisfies the assumptions of Theorem \[shpr\]. So applying Lemma \[text\] we reduce the situation to the one occurring in Theorem \[shpr\].
*b)* It is a direct consequence of *a)* and Remark \[poznauwaga\]
Note that Theorem \[lempr\] and Corollary \[cor\] allow us to determine the Shilov boundary of some classes of domains containing the symmetrized polydisc (see [@Edi-Zwo]) and the tetrablock. For example $\partial_s\mathbb E=\Pi(\partial_s\mathcal R_{II})=\Pi(\mathcal U),$ where $\mathcal U$ consists of unitary symmetric matrices (see also [@You], where the author using elementary methods computed the Shilov boundary of the tetrablock).
It is also interesting that Theorem \[shpr\] may be used for showing the non-existence of proper holomorphic mappings between some domains. For instance, using Corollary \[cor\] we immediately see that $\prop(\mathbb D^n,\mathbb B_n)$ and $\prop(\mathbb B_n,\mathbb D^n)$ are empty for $n\geq 2$ (see also [@Nar]). As an other example of the application of this result, observe that the theorem showing that there are no proper holomorphic mappings between $\mathbb B_n\times\mathbb B_m$ and $\mathbb B_{n+m}$ follows directly from Corollary \[cor\].
Applications to the tetrablock
==============================
The next result has been proved in [@Rud] for the Euclidean ball in $\mathbb C^n.$ We would like to mention here that for our purposes a much weaker result of Tumanov and Henkin proved in [@Tum-Hen] is sufficient. However, it seems to be interesting that after some modifications the Rudin’s idea may be applied to the symmetric domains.
First recall a well known classical result.
\[lemrud\] Suppose that $\Omega_1$ and $\Omega_2$ are balanced domains in $\mathbb C^n$ and $\mathbb C^m$ respectively. Suppose moreover that $\Omega_2$ is convex and bounded and $F:\Omega_1\to\Omega_2$ is holomorphic. Then $F'(0)$ maps $\Omega_1$ into $\Omega_2.$
If additionally $F(0)=0,$ then $F(\lambda\Omega_1)\subset \lambda\Omega_2,$ $0\leq\lambda\leq1.$
\[bih\] Let $a_0,$ $b_0$ be any unitary symmetric matrices. Let $U,$ $V$ be open neighborhoods of $a_0$ and $b_0$ respectively. Let $\varphi:U\cap\mathcal R_{II}\to V\cap\mathcal R_{II}$ be a biholomorphic mapping. If $\varphi(a_k)\to b_0$ for some $a_k\to a_0,$ then $\varphi$ extends to an automorphism of $\mathcal R_{II}.$
A direct computation shows that for any symmetric unitary matrix $a$ there is a unitary matrix $u$ such that $uu^t=a.$ Since any of the mappings $\mathcal R_{II}\ni x\to uxu^t\in\mathcal R_{II},$ where $u$ is unitary, is an automorphism of $\mathcal R_{II},$ we may assume that $a_0=b_0=1.$
Recall that (see e.g. [@Hua]) for every $a\in\mathcal R_{II}$ the mapping $$\label{wzornaaut}\varphi_a(x)=-a+(1-aa^*)^{1/2}x(1-a^*x)^{-1}(1-a^*a)^{1/2}$$ is an automorphism of $\mathcal R_{II}$, and $\varphi_a(0)=-a$ and its inverse is given by $\varphi_a^{-1}=\varphi_{-a}$.
Put $b_k=\varphi(a_k)$ and define $G_k:=\varphi_{b_k}\circ\varphi\circ\varphi_{-a_k}:\varphi_{a_k}(U\cap \mathcal R_{II})\to \varphi_{b_k}(V\cap\mathcal R_{II}),\ k\in\mathbb N.$ Note that $G_k$ is a biholomorphic mapping, $G_k(0)=0.$ Clearly $\varphi_{-a}(x)\to 1$ locally uniformly whenever $a\to1,$ so a compactness argument gives the existence of $\delta_k>0$ such that $\delta_k\to 1,$ as $k\to\infty,$ and both $\varphi_{a_k}(U\cap \mathcal R_{II}),$ $\varphi_{b_k}(V\cap\mathcal R_{II})$ contain a domain $\delta_k\mathcal R_{II}.$ Properly scaled Lemma \[lemrud\] implies that $\delta_k^3\leq |\det G_k'(0)|\leq \delta_k^{-3}.$
Since $G_k(0)=0,$ it follows that there exists a subsequence of $\{G_k\}$ (also denote by $\{G_k\}$) converging locally uniformly to $G:\mathcal R_{II}\to\mathcal R_{II}.$ Clearly $|\det G'(0)|=1$ and $G(0)=0,$ so by Lemma \[lemrud\] the domain $\mathcal R_{II}$ is mapped by $G'(0)$ into $\mathcal R_{II}.$ Since $|\det G'(0)|=1,$ the mapping $G'(0)$ preserves the volume. Hence $G'(0)$ maps $\mathcal R_{II}$ onto $\mathcal R_{II},$ in particular it is a unitary operator. Compose $G$ with $(G'(0))^{-1}$ and then apply the Cartan theorem in order to find that $G$ is also unitary.
Let $\mathcal N=\{z\in\mathcal M_{2\times 2}(\mathbb C):\ z=z^t,\ ||z||\neq\rho(z)\},$ where $\rho$ denotes the spectral radius. Note that $\mathcal N\cap\mathcal R_{II}$ is open and dense in $\mathcal R_{II}.$ Moreover $\lambda z\in\mathcal N$ for any $z\in\mathcal N$ and $\lambda\in\mathbb C\setminus\{0\}.$ For $z\in\mathcal R_{II}$ define $D_z=\{\lambda z:\ ||\lambda z||<1\}\subset\mathcal R_{II}.$
Let $K$ be any compact subset of $\mathcal N\cap \mathcal R_{II}.$ Observe that $$\label{deb}\bigcup\{D_z:\ z\in K\}\subset\varphi_{a}(\mathcal R_{II}\cap U)$$ for $a\in\mathcal R_{II}$ sufficiently close to $1.$ Indeed, otherwise there would exist sequences $(\lambda_n)\subset\mathbb C,$ $(z_n)\subset K$ and $(a_n)\subset\mathcal R_{II}$ such that $a_n\to 1$ and $\lambda_n\to \lambda_0\in\mathbb C,$ $z_n\to z_0\in K$ and $\lambda_n z_n \not\in \varphi_{a_n}(\mathcal R_{II}\cap U)$ (pass to subsequences, if necessary). If $\lambda_0=0,$ then the contradiction is obvious. In the other case $\lambda_0 z_0\in\mathcal N\cap\overline{\mathcal R_{II}}.$ It follows that $\det(1-\lambda_0z_0)\neq 0$ (otherwise $\rho(\lambda_0 z_0)\geq1\geq||\lambda_0 z_0||$). This in particular means that $\varphi_{a_n}^{-1}(\lambda_n z_n)$ converges to $1$ (use the formula (\[wzornaaut\])). Whence $\lambda_n z_n\in \varphi_{a_n}(\mathcal R_{II}\cap U)$ for large $n$; a contradiction.
Since $G$ is unitary, $G^{-1}(\mathcal N)\cap\mathcal N$ is open and dense in $\mathcal R_{II}.$ Let $B=\{z\in\mathcal M_{2\times 2}(\mathbb C):\ z=z^t,\ ||z-p||<2c\},$ where $p=p^t$ and $c>0$ are chosen such that $$B\Subset G^{-1}(\mathcal N) \cap\mathcal N\quad \text{and}\quad ||z||<1-c\quad \text{for}\quad z\in B.$$
Property (\[deb\]) yields the existence of an $n$ such that $$\label{dgz}D_z\subset\varphi_{a_n}(U\cap\mathcal R_{II})\quad \text{and}\quad D_{G(z)}\subset\varphi_{b_n}(V\cap\mathcal R_{II}),\quad z\in B.$$ We may assume that for such chosen $n:$ $||G_n(z)-G(z)||<c$ whenever $||z||\leq 1-c,$ $z=z^t.$
Then for $z=z^t$ such that $||z-p||<c,$ the set $D_z$ is contained in $\varphi_{a_n}(U\cap\mathcal R_{II}).$ Since $||G^{-1}(G_n(z))-p||=||G_n(z)-G(p)||<2c$ we see that $G^{-1}(G_n(z))\in B.$ Therefore, making use of (\[dgz\]) we get that $D_{G_n(z)}\subset \varphi_{b_n}(V\cap\mathcal R_{II}).$
Thus we may use a standard argument to the mapping $G_n:D_z\to\mathcal R_{II},$ where $z=z^t$ is such that $||z-p||<c,$ in order to find that $||G_n(z)||\leq ||z||.$ The same argument applied to $G_n^{-1}:D_{G_n(z)}\to \mathcal R_{II}$ together with the previous inequality gives $||G_n(z)||=||z||$ for $||z-p||<c$, $z=z^t.$ Obviously this equality remains validate on the whole $\varphi_{a_n}(\mathcal R_{II} \cap U).$
Choose $r$ such that a ball $r\mathcal R_{II}$ is contained in $\varphi_{a_n}(\mathcal R_{II}\cap U)\cap\varphi_{b_n}(\mathcal R_{II}\cap V)$ for a large $n$. The restriction of $G_k$ to $r\mathcal R_{II}$ is an automorphism of $r\mathcal R_{II}$ fixing $0.$ So we conclude from the description of the group of automorphism of classical Cartan domain of the second type that $G_k$ is unitary. From this piece of information we immediately get the assertion.
We are ready to show the correspondence between proper holomorphic self-mappings of the tetrablock and the Cartan domain of the second type.
\[lemma2\] Let $\varphi:\mathbb E\to\mathbb E$ be a proper holomorphic mapping. Then, there is $\psi\in\aut(\mathcal R_{II})$ such that $$\varphi\circ\Pi=\Pi\circ\psi$$
First observe that $\Pi^{-1}({\mathbb E})=\mathcal R_{II},$ so it is very easy to see that $\Pi:\mathcal R_{II}\to\mathbb E$ is proper. Put $f:=\varphi\circ\Pi.$ By Corollary \[pext\] the mapping $f$ extends to an open neighborhood $\Omega_1$ of $\overline{\mathcal R_{II}}.$ Define $$\mathcal J:=\{x\in\Omega_1:\ \det[f'(x)]\neq 0\ \text{and}\ f_1(x)f_2(x)\neq f_3(x)\}.$$ Since every proper holomorphic mapping is non-degenerate, properties of the Shilov boundary show that the intersection of sets $\mathcal J$ and $\partial_s\mathcal R_{II}$ is non-empty. Take any $x_0\in\mathcal J\cap \partial_s\mathcal R_{II}.$
Fix a $y_0$ such that $\Pi(y_0) = f(x_0).$ The choice of $x_0$ and properties of covering maps allow us to choose open neighborhoods $U,$ $V$ of $x_0,$ $y_0$ respectively and a biholomorphic mapping such that $$f=\Pi\circ \psi\quad \text{on}\ U.$$ We find from Corollary \[cor\] that $f(x_0)$ lies in the Shilov boundary of the tetrablock. So $\psi(x_0)$ is unitary.
Lemma \[bih\] and the identity principle finish the proof.
Now we are able to prove an Alexander-type theorem for the tetrablock.
It suffices to apply Lemma \[lemma2\] to get that the mapping $\varphi\circ\Pi$ has multiplicity $2.$ Since $\Pi$ also has multiplicity $2$ we infer that $\varphi$ is an automorphism.
\[rem\]Note that the tetrablock is not $\mathbb C$-convex. Actually, let $$\gamma(x)=|x_1-\overline{x_2}x_3|+|x_1x_2-x_3|+|x_3|^2,\quad \text{for}\quad
x=(x_1,x_2,x_3)\in\mathbb C^3.$$ As shown in [@You0], $x\in\mathbb E$ if and only if $\gamma(x)<1.$
For $\zeta\in\mathbb C$ put $$\varphi(\zeta):=\left(\frac{1-i}{2}\zeta+\frac{1+i}{2},\frac{1+i}{2}\zeta+\frac{i-1}{2},i\zeta\right).$$ Obviously $\varphi(1),\ \varphi(-1)\in\overline{\mathbb E}.$ Moreover $\varphi(i\zeta)=\left(\frac{1+i}{2}(\zeta+1),\frac{i-1}{2}(\zeta+1),-\zeta\right).$ An easy computation shows that for any $\zeta\in\mathbb R:$ $$\begin{aligned}
\gamma(\varphi(i\zeta))&=\left|\frac{1+i}{2}(\zeta+1)-\frac{i+1}{2}(\zeta+1)\zeta\right|+\left|-1/2(\zeta+1)^2+\zeta\right|+\zeta^2=\\
&=\frac{\sqrt{2}}{2}|1-\zeta^2|+|\frac{\zeta^2+1}{2}|+\zeta^2=\frac{\sqrt{2}}{2}|1-\zeta^2|+\frac{3}{2}\zeta^2+1/2.\end{aligned}$$ In particular $\gamma(\varphi(z))>1$ for any $z\in\{x\in\mathbb C:\ \rea x=0\},$ so $\mathbb E\cap\varphi(\mathbb C)$ is not connected.
[999999999]{} <span style="font-variant:small-caps;">A. Abouhajar, M. White and N. Young</span>, *A Schwarz lemma for a domain related to mu-synthesis*, Journal of Geometric Analysis **17** (2007), no. 4, 717–750. <span style="font-variant:small-caps;">S. Bell</span>, *Analytic hypoellipticity of the $\bar \partial $-Neumann problem and extendability of holomorphic mappings*, Acta Math. **147** (1981), no. 1-2, 109–116. <span style="font-variant:small-caps;">S. Bell</span>, *Proper holomorphic mappings between circular domains*, Comment. Math. Helv. **57** (1982), no. 4, 532–538. <span style="font-variant:small-caps;">A. Edigarian and W. Zwonek</span>, *Geometry of the symmetrized polydisc*, Arch. Math. **84** (2005) 364–-374. <span style="font-variant:small-caps;">A. Edigarian and W. Zwonek</span>, *Schwarz lemma for the tetrablock*, Bulletin of the London Mathematical Society, **41** (2009), no. 3, 506-514. <span style="font-variant:small-caps;">G. H. Hardy and E. M. Wright</span>, *An Introduction to the Theory of Numbers*, Oxford Science Publ., 1978. <span style="font-variant:small-caps;">G.M. Henkin and R. Novikov</span>, *Proper mappings of classical domains*, Lecture Notes in Math. **1043**, (1984), 625–627. <span style="font-variant:small-caps;">L.K. Hua</span>, *Harmonic Analysis of Functions of Several Complex Variables in the Classical Domains*, AMS. (1963), Providence. <span style="font-variant:small-caps;">D. Jacquet</span> *$\mathbb C$-convex domains with $C\sp 2$ boundary*, Complex Var. Elliptic Equ. **51** (2006), no. 4, 303–312. <span style="font-variant:small-caps;">L. Lempert</span>, *La métrique de Kobayashi et la représentation des domaines sur la boule*, Bull. Soc. Math. France **109** (1981), 427–474. <span style="font-variant:small-caps;">N. Mok</span>, *Nonexistence of proper holomorphic maps between certain classical bounded symmetric domains*, Chinese Annals of Mathematics - *Series B* **29**, No.2 (2008), 135–146. <span style="font-variant:small-caps;">R. Narasimhan</span>, *Several Complex Variables,* Chicago Lectures in Mathematics, 1971. <span style="font-variant:small-caps;">N. Nikolov</span>, *The symmetrized polydisc cannot be exhausted by domains biholomorphic to convex domains*, Ann. Polon. Math., **88** (2006), 279–283. <span style="font-variant:small-caps;">W. Rudin</span>, *Fuction theory in the unit ball of $\mathbb C^n$*, Grundlehren der mathematischen Wissenschaften 241 (1980). <span style="font-variant:small-caps;">W. Rudin</span>, *Holomorphic maps that extend to automorphisms of a ball*, Proc. Amer. Math. Soc. **81** (1981), no. 3, 429–432. <span style="font-variant:small-caps;">Z.-H. Tu</span>, *Rigidity of proper holomorphic mappings between equidimensional bounded symmetric domains*, Proc. Amer. Math. Soc. **130** (2002), 1035–1042. <span style="font-variant:small-caps;">Z.-H. Tu</span>, *Rigidity of proper holomorphic mappings between nonequidimensional bounded symmetric domains*, Math. Z. **240** (2002), 13–35. <span style="font-variant:small-caps;">Z.-H. Tu</span>, *Rigidity of proper holomorphic mappings between bounded symmetric domains*, Geometric function theory in several complex variables, pp.310-316, World Sci. Publishing, River Edge, NJ, 2004. <span style="font-variant:small-caps;">A.E. Tumanov and G.M. Khenkin</span>, *Local characterization of holomorphic automorphisms of Siegel domains*, (Russian) Funktsional. Anal. i Prilozhen. **17**0 (1983), no. 4, 49–61. <span style="font-variant:small-caps;">N. Young</span>, *The automorphism group of the tetrablock*, Journal of the London Mathematical Society , **77** (2008), no. 3, 757-770.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Michael C. Mackey[^1]'
- 'Moisés Santillán[^2]'
title: 'Mathematics, Biology, and Physics: Interactions and Interdependence'
---
Introduction {#introduction .unnumbered}
============
Modern science means that which has a solid conceptual framework, and which considers experimental results as the ultimate litmus test against which to validate any theoretical construct. Its birth can be traced back to the 16th and 17th Centuries. The work of people like Nicholaus Copernicus, Galileo Galilei, Johannes Kepler, William Harvey, Vesalius, and others was seminal to this development. Before the so-called Scientific Revolution, natural philosophers (the forefathers of scientists as we know them today) did not perform experiments as manual labor was considered a lower class activity. This attitude, inherited from the Greeks, changed between the 16th and the 18th Centuries, as merchants and craftsmen gained economic and political power. As a result economics, politics, and science went through significant changes. In that period democracy, capitalism, and modern science were founded and emerged as the cornerstones of a new era.
During the Enlightenment, in the latter part of the 18th and early part of the 19th Centuries, scientific disciplines started to be hierarchically classified. This classification works well in some instances, and without it dealing with the rapidly growing body of knowledge of the past 200 years would have been difficult. However, it fails to fairly represent the interdisciplinary work which has been, and continues to be, highly important. In this paper we give a taste of the rich historical relation between physics, mathematics, and the biological sciences. We argue that this will continue to play a very important role in the future, based on historical examples and on a brief review of the current situation.
The 18th and 19th Centuries {#the-18th-and-19th-centuries .unnumbered}
===========================
Electrophysiology is the science that studies the interaction between electromagnetic fields and biological tissues. This includes the generation of electric or magnetic fields and electric currents in some specialized organs, the intrinsic electric and magnetic properties of tissue, the response of specialized cells (like neurons and muscle cells) to stimulation, etc. Up to the middle of the 19th Century, the historical development of electrophysiology paralleled that of electromagnetism. The first electric generating machines and the Leyden jar were constructed to produce static electricity for a specific purpose: to “electrify" and to stimulate humans. The Voltaic pile was developed with the idea of galvanic (i.e. direct current, as opposed to faradic or alternating current) stimulation. Bioelectric and biomagnetic measurements were the incentive for the development of sensitive measurement instruments, like the galvanometer and the capillary electrometer. Thus, it is no surprise that some scientists of the time made important contributions to the development of both the biological and the physical sciences. In the following paragraphs we present a brief review of the work of some of these interdisciplinary workers. We do not attempt to present a detailed review of the history of electrophysiology, as our purpose is only to exemplify the rich interdisciplinary interactions of the 18th and 19th Centuries.
The essential invention necessary for the application of a stimulating electric current was the Leyden jar (a capacitor formed by a glass bottle covered with metal foil on the inner and outer surfaces), independently invented in Germany (1745) and The Netherlands (1746). With it, Benjamin Franklin’s experiments allowed him to deduce the concept of positive and negative electricity in 1747. Franklin also studied atmospheric electricity with his famous kite experiment in 1752 (many American school children have heard the apocryphal stories of Franklin flying kites during thunderstorms strings soaked in salt water).
The most famous experiments in neuromuscular stimulation of the time were performed by Luigi Galvani, professor of anatomy at the University of Bologna. His first important finding is dated January 26, 1781. A dissected and prepared frog was lying on the same table as an electric machine. When his assistant touched the femoral nerve of the frog with a scalpel, sparks were simultaneously discharged in the nearby electric machine, and violent muscular contractions occurred. (It has been suggested that the assistant was Galvani’s wife Lucia, who is known to have helped him with his experiments). This is cited as the first documented experiment in neuromuscular electric stimulation.
Galvani continued the stimulation studies with atmospheric electricity on a prepared frog leg. He connected an electric conductor between the side of the house and the nerve innervating the frog leg. Then he grounded the muscle with another conductor in an adjacent well. Contractions were obtained simultaneous with the occurrence of lightning flashes. In September 1786, Galvani was trying to obtain contractions from atmospheric electricity during calm weather. He suspended frog preparations from an iron railing in his garden by brass hooks inserted through the spinal cord. Galvani happened to press the hook against the railing when the leg was also in contact with it. Observing frequent contractions, he repeated the experiment in a closed room. He placed the frog leg on an iron plate and pressed the brass hook against the plate, and muscular contractions occurred. Systematically continuing these experiments, Galvani found that when the nerve and the muscle of a frog were simultaneously touched with a bimetallic strip of copper and zinc, a contraction of the muscle was produced. This experiment is often cited as the classic study to demonstrate the existence of Òanimal electricity. Galvani did not understand the mechanism of the stimulation with the bimetallic strip. His explanation for this phenomenon was that the bimetallic strip was discharging the animal electricity existing in the body.
Galvani’s investigations intrigued his friend and colleague Alessandro Volta (professor of physics in Pavia), who eventually came up with a totally different (and correct) explanation for the phenomena that Galvani was trying to explain. In the process, Galvani and Volta maintained their friendship (in spite of their differences of scientific opinion), and Volta developed the ideas that eventually led to the invention of the Voltaic pile in 1800 (forerunner of the modern battery), a battery that could produce continuous electric current. Incidentally, Volta only completed the equivalent of his doctoral dissertation when he was 50 years old!
All of these contributions to electrophysiology were experimental. The first significant theoretical contributions were made by the German scientist and philosopher Hermann Ludwig Ferdinand von Helmholtz. A physician by education and, in 1849, appointed professor of physiology at Königsberg, he moved to the chair of physiology at Bonn in 1855 and, in 1871, was awarded the chair of physics at the University of Berlin. Helmholtz’s fundamental experimental and theoretical scientific contributions in the field of electrophysiology included the demonstration that axons are extensions of the nerve cell body, the establishment of the law of conservation of energy (the First Law of Thermodynamics), the invention of the myograph, and the first measurement of the action potential conduction velocity in a motor nerve axon. Besides these, the contributions of Helmholtz to other fields of science include fundamental work in physiology, acoustics, optics, electrodynamics, thermodynamics, and meteorology. He invented the ophthalmoscope and was the author of the theory of hearing from which all modern theories of resonance are derived. Another important contribution to the development of biophysics was Helmholtz’s philosophical position in favor of founding physiology completely on the principles of physics and chemistry at a time when physiological explanations were based on vital forces which were not physical in nature.
The 20th Century {#the-20th-century .unnumbered}
================
In the 18th and 19th Centuries interdisciplinary research bridging physics, mathematics and biology was carried out by scientists educated as physicians. The 20th Century witnessed a reversal of this trend with major contributions to biology from people with solid backgrounds in physics and mathematics. There are two of these disciplines in which the contributions by physicists and mathematicians were particularly important: electrophysiology (following the tradition of Galvani, Volta, Helmholtz, etc.) and molecular biology.
Electrophysiology {#electrophysiology .unnumbered}
-----------------
The growth of biophysics owes much to A. V. Hill, whose work on muscle calorimetry was essential to our understanding of the physiology of muscle contraction. Hill received an undergraduate degree in physics and mathematics, and a doctorate in physiology, all from Cambridge. Besides his work on muscle contraction, Hill also addressed problems related to the propagation of the nervous impulse, the binding of oxygen by hemoglobin, and on calorimetry of animals. He discovered that heat is produced during the nerve impulse. Hill’s original papers reveal an elegant mixture of biological concepts and experiments together with physical and mathematical theory and insight. His discoveries concerning the production of heat in muscle earned him the Nobel Prize in 1922, and his research gave rise to an enthusiastic following in the field of biophysics. He was instrumental in establishing an extremely successful interdisciplinary school in Cambridge, whose investigators received a number of Nobel prizes.
A few years later Bernard Katz, working at University College London with his student Paul Fatt, made a major advance in our understanding of the chemical and quantal nature of synaptic transmission in the papers “An analysis of the end-plate potential recorded with an intra-cellular electrode" and “Spontaneous subthreshold activity at motor nerve endings", which were marvels of experimental investigation combined with mathematical modelling of stochastic processes. Katz was one of the recipients of the 1970 Nobel Prize for “discoveries concerning the humoral transmittors in the nerve terminals and the mechanism for their storage, release and inactivation."
Jumping back a few decades, the German physical chemist Walter Nernst was interested in the transport of electrical charge in electrolyte solutions. His work intrigued another physicist, Max Planck, one of the fathers of modern quantum theory, who extended Nernst’s experimental and theoretical work, eventually writing down a transport equation describing the current flow in an electrolyte under the combined action of an electric field and a concentration gradient. This work lay largely forgotten until the 1930s, when it was picked up by the physicist Kenneth S. Cole at Columbia University and his graduate student David Goldman (originally trained in physics). They realized that the work of Nernst and Planck (in the form of the Nernst-Planck equation) could be used to describe ion transport through biological membranes and did so with great effect. Their work resulted in the development of the Goldman equation, which describes the membrane equilibrium potential in terms of intra- and extracellular ionic concentrations and ionic permeabilities. This background theoretical work of Nernst and Planck was also instrumental in helping Cole to experimentally demonstrate that there was a massive increase in membrane conductance during an action potential.
Two of the most distinguished alumni of Hill’s Cambridge interdisciplinary school were A. L. Hodgkin and A. F. Huxley. Both studied physics, mathematics, and physiology at Trinity College, Cambridge, whose high table included, at that time, an astonishing array of scientific talent with people like J. J. Thomson, Lord Rutherford, F.W. Aston, A.S. Eddington, F.G. Hopkins, G. H. Hardy, F.J.W. Roughton, W.A.H. Rushton, A.V. Hill, and E.D. Adrian. Hodgkin and Huxley developed a long-lasting collaboration, interrupted only by the outbreak of World War II.
In 1938 Hodgkin spent the summer with Cole at Woods Hole, and they demonstrated the overshoot of the action potential which had significant implications in terms of potential ionic mechanisms. It seems reasonable to suppose that Cole and Hodgkin discussed the possible meanings of these discoveries and what types of experiments were needed to determine exactly what was going on. Because of their training they would have seen that some means must be found to bring under experimental control the variable (either membrane current or membrane voltage) that is responsible for the all-or-nothing behavior of the action potential. Hence taming the action potential required either controlling the current or the voltage. They undoubtedly realized that space clamping was necessary for both current and voltage clamping and since both knew cable theory, they knew that space clamping was best done by drastically reducing internal resistance (space clamping) so the space constant was much longer than the length of axon under study.
The Second World War interrupted these investigations and Cole, like hundreds of other scientists, was caught up in the war effort. Cole moved from Columbia to the Manhattan Project in Chicago and worked on radiation dosimetry and radiation damage in tissues during the war. After the war he was at the University of Chicago for a few years. When the war was over, one of the positive outcomes was the existence of high-input impedance vacuum tubes that had been developed for the amplifiers in radar receivers. Cole, working with Marmont in Chicago, used these new electronic advances to build a feedback circuit that allowed them to space clamp axons. These axons developed an all-or-none action potential when sufficiently depolarized, and the implication was that voltage clamping was necessary to tame the axon to measure the dependence of membrane current on membrane voltage.
Shortly after the war (1948), Hodgkin visited the United States and Cole’s laboratory in Chicago, and realized that the results of the space clamp experiments meant that voltage clamping was the way to go. On his return to England he teamed up with Huxley to really measure what was going on during the generation of an action potential in the squid giant axon. This work was published in a brilliant series of five papers in the Journal of Physiology in 1952. The final one is an intellectual tour de force combining both experimental data analysis and mathematical modelling (the Hodgkin-Huxley equations) that eventually won Hodgkin and Huxley the Nobel Prize in 1963, along with J.C. Eccles, “for their discoveries concerning the ionic mechanisms involved in excitation and inhibition in the peripheral and central portions of the nerve cell membrane." Huxley the mathematician/physiologist was not content to stop there, however, and went on to publish his celebrated review of muscle contraction data and its synthesis into the mathematically formulated cross bridge theory in 1957, a theory that still stands in its essential ingredients today.
The Hodgkin-Huxley model for excitability in the membrane of the squid giant axon is complicated and consists of one nonlinear partial differential equation coupled to three ordinary differential equations. In the early 1960s Richard FitzHugh applied some of the techniques that he had learned from the Russian applied mathematics literature to an analysis of the Hodgkin-Huxley equations. That reduction of the Hodgkin-Huxley equations later became known as the FitzHugh-Nagumo model and has given us great insight into the mathematical and physiological complexities of the excitability process. Another consequence of the Hodgkin-Huxley model, taken to its interpretational extreme, was the implication that there were microscopic “channels" in the membrane through which ions would flow and which were controlled by membrane potential. There were strong experimental data also leading to the same conclusion including the binding of tetrodotoxin (TTX) to nerve membranes to block sodium currents, titration studies indicating that there were about 20 TTX binding sites per square micrometer, and membrane noise measurements. However, it was left to the German physicist Erwin Neher, in conjunction with the physiologist Bert Sakmann, to develop the technology and techniques that eventually allowed them to demonstrate the existence of these ion channels. They were awarded the Nobel Prize in 1991 for this work. Modifications of the Hodgkin-Huxley equations were soon proposed for cardiac tissue as well as a myriad of other excitable cells.
Extensions of the work of Hodgkin and Huxley soon followed. For example, J. W. Woodbury (a physicist turned physiologist) and his student W. E. Crill found that current injected into one cell in a sheet of heart muscle changed the membrane voltage in nearby cells in an anisotropic manner. This showed that there must be low resistance connections between abutting cells in heart tissue and paved the way for the discovery and characterization of gap junctions between the cells (in a variety of tissues such as epithelia). Woodbury also showed that Eyring reaction rate theory, learned from his famous foster Doctorvater Henry Eyring, can be used to explain the linear current-voltage relationship of open sodium channels by choosing the appropriate electrochemical potential profile encountered by a sodium ion while traversing a Na ion channel. This, together with other lines of types of experimental evidence mentioned above, established the feasibility of the ion channel concept before single channel conductances were directly measured by Neher.
One of the most remarkable individuals interested in the dynamic behavior of simple nervous systems was H. K. Hartline of the John Hopkins University. Hartline was trained as a physiologist, and following receipt of his M.D. spent an additional two years at Hopkins taking mathematics and physics courses. For some unaccountable reason he was still not satisfied with his training and obtained funding to study for a further year in Leipzig with the physicist Werner Heisenberg and a second year in Munich with Arthur Sommerfeld. Armed with this rather formidable training in the biological, mathematical, and physical sciences he then devoted the majority of his professional life at Hopkins to the experimental study of the physiology of the retina of the horseshoe crab Limulus. His papers are a marvel of beautiful experimental work combined with mathematical modelling designed to explain and codify his findings, and his life work justly earned him the Nobel Prize in 1967 (with George Wald) “for his discoveries concerning the primary physiological and chemical visual processes in the eye." As an aside we should point out that FitzHugh (of the FitzHugh-Nagumo reduction of the Hodgkin-Huxley model) received his Ph.D. in biophysics (where he learned mathematics, physics, and chemistry) under Hartline after completing his biological studies at the University of Colorado.
One can hardly underestimate the impact that this work in excitable cell physiology has had on the biological sciences since the impact is so broad and pervasive. The Notices of the American Mathematical Society (December 1999) has a very nice article by Nancy Kopell with some of the mathematical side of the story, and Nature Neuroscience (November 2000) featured some of this from a biological perspective in an interesting and lively series of survey articles.
Molecular biology {#molecular-biology .unnumbered}
-----------------
Genetics started in 1866, when Gregor Mendel first deduced the basic laws of inheritance. However, modern genetics, with its capacity to manipulate the very essence of living things, only came into being with the rise of molecular investigations culminating in the breakthrough discovery of the structure of DNA; for which Francis Crick, James D. Watson, and Maurice Wilkins received the Nobel prize in 1962. The contribution of physics and physicists to this, what Watson calls Act 1 of molecular biology’s great drama, was seminal. Here we review the work of some of the physicists who helped shape molecular biology into the exciting science it currently is.
Max Delbrück received his doctorate in theoretical physics from the University of Göttingen, and then spent three postdoctoral years in England, Switzerland, and Denmark. His interest in biology was aroused during his stay in Denmark by Niels Bohr’s speculation that the complementarity principle of quantum mechanics might have wide applications to other scientific fields, and especially to the relation between physics and biology. Back in Berlin, Delbrück initiated an interdisciplinary collaboration with Nikolai W. Timofeeff and Karl G. Zimmer on biologically inspired problems. Based on X-ray induced mutagenesis experiments and applying concepts from quantum mechanics, they suggested that chromosomes are nothing more than large molecules and that mutations can be viewed as ionization processes. These results were published in 1935. Schrödinger’s little book “What is Life?" (1944) was in part inspired by this paper.
In 1937, Delbrück moved from Germany to the United States, and decided to remain after the start of World War II. At that time he initiated a fruitful collaboration with Salvador Luria on the genetic structure of bacteriophage (bacteria-infecting viruses) and on the genetic mechanism of DNA replication. After the outbreak of the war, Delbrück and Luria were classified as ’enemy aliens’ by the American government despite their open opposition to the Nazi and Fascist regimes. This classification fortuitously allowed them to pursue their own investigations without having to join any military project. For “their discoveries concerning the replication mechanism and the genetic structure of viruses," Delbrück and Luria were awarded the Nobel Prize in 1969, along with Alfred D. Hershey. In the early 1950’s Delbrück’s research interests shifted from molecular genetics to sensory physiology, with the goal of clarifying the molecular nature of the primary transduction processes of sense organs. Delbrück was also involved in setting up an institute of molecular genetics at the University of Cologne. It was formally dedicated on June 22nd, 1962, with Niels Bohr as the principal speaker. His lecture entitled “Light and Life Revisited" commented on his original one of 1933, which had been the starting point of Delbrück’s interest in biology. It was to be Bohr’s last formal lecture. He died before completing the manuscript of this lecture for publication.
Erwin Schrödinger is regarded as one of the fathers of quantum mechanics. However, his interests went far beyond physics. He was particularly interested in philosophy and biology. Early in his career, he made substantial contributions to the theory of color vision. Schrödinger’s personal life was tumultuous. He participated as an officer in World War I on the Italian front. For a variety of reasons, Schrödinger moved constantly, holding positions in Austria, Switzerland, Germany, England, and then Austria again. Soon after he took up this last position in Graz, Austria fell into the hands of the Nazis, and Schrödinger escaped to Ireland since his initial departure from Berlin when the National Socialists took power was considered an unfriendly act.
In Ireland, Schrödinger joined the Institute for Advanced Studies in Dublin. His contract required him to give a yearly series of public lectures. In 1943, he elected to discuss whether the events in space and time which take place within the spatial boundary of a living organism can be accounted for by physics and chemistry in light of the most recent developments in quantum mechanics and its application to genetics. These lectures were published in book form in 1944 under the title “What is Life?" After discussing how thermodynamics plays a role in the processes of life and reviewing the not-so-recent results on mutagenesis by Delbrück et al., Schrödinger argued in “What is Life?" that life could be thought of in terms of storing and transmitting information. Chromosomes were thus simply bearers of information. Because so much information had to be packed into every cell, Schrödinger argued it must be compressed into what he called a ‘hereditary code-script’ embedded in the molecular fabric of chromosomes. To understand life, then, it was necessary to identify these molecules, and crack their code. Schrödinger’s book had the very positive effect of popularizing the Delbrück paper and of rephrasing some important questions derived from it in a language accessible to the non-expert. The book’s publication could not have been better timed, and it was tremendously influential. Many of those who would play major roles in the development of molecular biology were drawn to this field after reading “What is is Life?" Schrödinger’s recruits included Francis Crick, James D. Watson, Maurice Wilkins, Seymour Benzer, and François Jacob.
Francis Crick studied physics at University College, London. After graduating, he started research for a doctorate, but this was interrupted by the outbreak of World War II. During the war he worked as a scientist for the British Admiralty, mainly on magnetic and acoustic mines. When the war ended, Crick had planned to stay in military research but, on reading Schrödinger’s book, he joined the Medical Research Council Unit in Cambridge to study biology. In 1951, Crick started a collaboration with James D. Watson, who came to Cambridge as a postdoctoral fellow. Watson had originally considered being a naturalist, but he was also hooked on gene research by Schrödinger’s book. Linus Pauling had discovered the alpha helix protein structure by making scale models of the different parts of the molecule, and working out possible 3-dimensional schemes to infer which type of helical fold would be compatible with the underlaying chemical features of the polypeptide (amino acid) chain. Following Pauling’s approach, Watson and Crick started to look for the structure of DNA, which in 1944 had been discovered to be the substance making up the chromosomes. They finally succeeded in the Spring of 1953. Not only did they determine the structure of DNA, but they also proposed a scheme for its replication.
Essential for the work of Watson and Crick were the experimental results of Rosalind Franklin and Maurice Wilkins. Franklin had a background in chemistry while Wilkins was a physicist. During World War II, Wilkins worked in the Manhattan Project. For him, as for many other of the scientists involved, the actual deployment of the bombs in Hiroshima and Nagasaki, the culmination of all their work, was profoundly disillusioning. He considered forsaking science altogether to become a painter in Paris. However, he too had read Schrödinger’s book and biology intervened. Franklin, working in Wilkins’ lab, recorded the DNA X-Ray diffraction patterns that allowed Watson and Crick to beat Pauling in the race to determine the structure of DNA. Crick, Watson, and Wilkins received the Nobel Prize in 1962 “for their discoveries concerning the molecular structure of nuclear acids and its significance for information transfer in living material." Rosalind Franklin had died at an early age a few years before, and was not recognized for her essential contributions.
Knowing the structure of DNA was only the start. Next it was necessary to find the sequence of genes and chromosomes, to understand the molecular machinery used to read the messages in DNA, and to understand the regulatory mechanisms through which the genes are controlled. These questions were answered by a second generation of molecular biologists like Seymour Benzer, Sydney Brenner, François Jacob, Jacques Monod, and Walter Gilbert. Seymour Benzer and Walter Gilbert had also been educated as physicists, but were attracted to the excitement of the new science. Seymour Benzer also heeded the clarion call of the Schrödinger book. He was a pioneer of gene sequencing. Among other things, Benzer was the first to produce a map of a single bacteriophage gene, rII, showing how a series of mutations (all errors in the gene script) were laid out linearly along the viral DNA.
Walter Gilbert received his doctorate in theoretical physics and, after becoming professor at Harvard, worked on particle physics and quantum field theory for a number of years. Then his interests shifted. In 1960 Gilbert joined James Watson and François Gros in a project to identify messenger RNA. After a year of work on this problem, Gilbert returned to physics only to re-return to molecular biology shortly afterwards. Some of the more important contributions of Gilbert and his collaborators to this field are: the discovery that a single messenger molecule can service many ribosomes at once and that the growing proteinic chain always remains attached to a transfer RNA molecule; the isolation of the lactose repressor, the first example of a genetic control element; the invention of the rolling circle model, which describes one of the two ways DNA molecules duplicate themselves; the isolation of the DNA fragment to which the lac repressor binds; and the development of rapid chemical DNA sequencing and of recombinant DNA techniques. Walter Gilbert and Frederick Sanger received the Nobel Prize in 1980, “for their contributions concerning the determination of base sequences in nucleic acids."
Present and future perspectives {#present-and-future-perspectives .unnumbered}
===============================
What we have described so far have been a few of the significant advances made in the study of systems in which there was a certain clear and obvious physics and mathematics component to the research being carried out. The advances made in the biological understanding were often quite dependent on the application of physical and mathematical principles, or the development of the physics and the mathematics was clearly driven by observations in biology. This strong interdependence is mirrored in the highlighting of biologically oriented problems in the new millennium (January, 2000) issues of Physics Today and the Notices of the American Mathematical Society as well as the special November, 2000 Nature Neuroscience issue “Computational Approaches to Brain Function." The Notices of the American Mathematical Society have on several occasions focussed on problems involving biomathematics (September, 1995) or molecular biology (April and May, 2002).
Many major universities in the world have at least one research group working in these fields. However, listing them all is beyond the scope or the intent of this article. Our purpose has only been to illustrate how widespread and important biophysics and biomathematics have been in the past few centuries and the increase in their importance in the past few decades.
Darwin’s theory states that, given the environmental conditions, the fittest individuals are the ones that survive and reproduce. However, it is impossible to identify the current fittest individuals whose genes are going to pass to the next generation. They can be pinpointed only after they have survived. Thus, according to some, Darwinism is tautological since it only predicts the survival of the survivors. In trying to foresee the future of science, we face the same problem. It is not possible to identify the current areas of scientific research that will play a relevant role in the development of science and technology. Even though we acknowledge this problem, it is our belief that, given the fruitful historical relation and the present blooming of biological, physical, and mathematical interdisciplinary sciences, they are going to be so important in the near future that the avant garde biological scientists will be those with a strong background in both the biological and the physical-mathematical sciences.
The mathematical and computational modelling of biological systems is a subject of increasingly intense interest. The accelerating growth of biological knowledge, in concert with a growing appreciation of the spatial and temporal complexity of events within cells, tissues, organs, and populations, threatens to overwhelm our capacity to integrate, understand, and reason about biology and biological function. The construction, analysis, and simulation of formal mathematical models is a useful way to manage such problems. Metabolism, signal transduction, genetic regulation, circadian rhythms, and various aspects of neurobiology are just a subset of the phenomena that have been successfully treated by mathematical modelling. What are the likely areas of advancement for the future? Predicting the future has fascinated and confounded man for centuries, probably for as long as he has been able to articulate the concept of the future. For example, some relatively recent predictions were:
> Physics is finished, young man. It’s a dead-end street.\
> -Unknown teacher of Max Planck, late 19th century\
> I believe that the motion picture is destined to revolutionize our educational system and that in a few years it will supplant largely, if not entirely, the use of textbooks.\
> -Thomas Edison, 1922\
> It is probable that television drama of high caliber and produced by first-rate artists will materially raise the level of dramatic taste of the nation.\
> -David Sarnoff, 1939
Being aware of the almost certain folly of trying to predict the future, as illustrated by these quotations, we nevertheless take the leap and mention several areas in which we feel that significant advances are likely to take place over the present century.
- The sequencing of human and other genomes has provided a spectacular amount of data which needs to be organized and analyzed before its significance becomes clear. The mathematical techniques necessary to do so are still to be developed. This has opened a whole new area of research known as bio-informatics, which is rapidly growing and, presumably, will keep on growing at an accelerated pace in the next few years. However, we are of the opinion that the sequence analysis component of bio-informatics will quickly evolve to become a mere tool widely and easily used by scientific practitioners (in analogy with the transition from scientific computing being done on large mainframe computers a few decades ago, and now being almost exclusively carried out on inexpensive workstations).
- The classification aspects of bio-informatics will be rapidly replaced by efforts to understand the regulation of gene networks using established and new techniques from non-linear dynamics. Mathematical modelling and analysis of the mechanisms of gene regulation will continue at an ever accelerating pace. This, in conjunction with the already established ability to produce “designer" molecular circuits, will be instrumental in the targeted treatment of disease through gene therapy.
- Attempts to understand the noisy interactions in gene regulation and expression at the single cell level will lead to the development of new mathematical techniques for dealing with chemical reactions in which the law of large numbers cannot be invoked.
- The Herculean efforts of countless neurobiologists over the past century have given us much insight into the functioning of single neurons as well as the behavior of simple neural circuits and some extremely simple sensory and motor systems. This progress will continue and lead to the efficient treatment of many neuron-related diseases, to a better design of protheses, and perhaps, to a deeper understanding of the relation between brain and mind. Shall we, at some time, be able to really understand phenomena like cognition and memory? Maybe, maybe not. Perhaps, as some philosophers maintain, the human mind is unable to understand itself. However, we firmly believe that the neurophysiological sciences will thrive in the near future, with physics and mathematics playing a central role in such progress. Examples are the use of vagal stimulation to abort epileptic seizures, and deep brain stimulation to control the tremor of Parkinson disease.
- Biophysical advances in determining the structure and dynamic properties of membrane channels and receptors have proceeded at a rapid pace over the past decade. There is every reason to anticipate that this will only accelerate in the future. The accumulated knowledge, in conjunction with modelling and production of designer molecules, will enable the efficient development and production of drugs specifically targeted to the elimination of disease symptoms if not the disease itself.
- The accelerated rhythm at which technology is progressing makes us believe that, in the near future, it will be possible to combine knowledge and techniques from biology, chemistry, biochemistry, computer science, engineering and physics to engineer designer molecules for specific medical and industrial purposes.
- Interdisciplinary work focused in the development of bio-materials, bio-electronic devices, and bio-mechanical systems will improve the design of artificial organs, protheses, and implants through the development of hybrid animate-inanimate devices.
- Epidemiological research aided by mathematical modelling and statistical analysis will help us understand the dynamics of disease transmission and to design more efficacious treatment and vaccination strategies.
- The difficulty in collecting high resolution temporal and spatial data from ecological and meteorological systems has limited the success of mathematical modelling approaches in these fields. The availability of more sophisticated geographic information systems and massive parallel computational power will alleviate these problems.
Summary {#summary .unnumbered}
=======
There has been a long and rich tradition of fruitful interdisciplinary interplay between the physical and biological sciences extending over several centuries, as we have illustrated with a few examples. Many other examples could have been offered to illustrate the point, and would simply serve to highlight the rich interactions between apparently disparate branches of science. We expect that these interactions and interdependence will continue and become even stronger in the future.
Acknowledgments {#acknowledgments .unnumbered}
===============
We are grateful to N. Anderson-Mackey, R. FitzHugh and J. Walter Woodbury for extensive comments on this article, which is largely based on a lecture with the same title given by MCM 4 May, 2001, at the University of Oxford as the Leverhulme Professor of Mathematical Biology for the 2001 academic year. This work was supported by COFAA-IPN (M[é]{}xico), EDI-IPN (M[é]{}xico), MITACS (Canada), the Natural Sciences and Engineering Research Council (NSERC grant OGP-0036920, Canada), and Le Fonds pour la Formation de Chercheurs et l’Aide ˆ la Recherche (FCAR grant 98ER1057, Qu[é]{}bec).
Bibliography {#bibliography .unnumbered}
============
- Alessandro Volta web page at University of Pavia. URL: http: //ppp.unipv.it/Volta
- Asimov, I. Asimov’s Biographical Encyclopedia of Science and Technology: The Lives and Achievements of 1510 Great Scientists from Ancient Times to the Present Chronologically Arranged. Garden City, N.Y.: Doubleday, 1982
- Bernal, J. D. Science in History. Cambridge, Mass.: M.I.T. Press, 1971
- Koenigsberger, L. Herman von Helmholtz. New York, N.Y.: Dover Publications, 1965.
- Moore, W. Schrödinger, Life and Thought. Cambridge; New York: Cambridge University Press, 1989
- MacTutor History of Mathematics. URL: http: //www-gap.dcs.st-and.ac.uk/history
- Malmivuo, J. and Plonsey, R. Bioelectromagnetism. New York, Oxford: University Press, 1995. URL: http: //butler.cc.tut.fi/malmivuo/bem/bembook/index.htm
- Nobel e-Museum. URL: http: //www.nobel.se
- Scrödinger, E. What is Life?: The Physical Aspects of the Living Cell. Cambridge: University Press, 1944
- Watson, J. DNA: The Secret of Life. New York: Alfred A. Knopf, 2003
[^1]: e-mail: [email protected], Departments of Physiology, Physics & Mathematics and Centre for Nonlinear Dynamics, McGill University, 3655 Promenade Sir William Osler, H3G 1Y6 Montreal, QC, CANADA
[^2]: e-mail: [email protected], Permanent address: Depto. de Física, Esc. Sup. de Física y Matemáticas, Inst. Politécnico Nal. 07738 México D. F., MÉXICO
|
{
"pile_set_name": "ArXiv"
}
|
---
---
8.75in 0.5in
**Discovery of the Acoustic Faraday Effect in Superfluid $^3$He-B**
Y. Lee, T.M. Haard, W.P. Halperin and J.A. Sauls
*Department of Physics and Astronomy, Northwestern University, Evanston, IL 60208, USA*
Superfluidity in $^3$He results from the binding of the $^3$He particles with nuclear spin $s=1/2$ into molecules called “Cooper pairs” with binding energy, $2\Delta$. [@leg75; @vol90; @and75; @whe75] The pairs undergo a type of Bose-Einstein condensation having a close analogy to the Bardeen-Cooper-Schrieffer condensation [@bar57] phenomenon associated with superconductivity in metals. One important difference is that the pairs that form the condensate in $^3$He have total spin $S=1$ and an orbital wave function with relative angular momentum $L=1$ ($p$-$wave$). This is in contrast to superconductors which are formed with Cooper pairs of electrons having $S=0$ and $L=0$ ($s$-$wave$) or, as is the case of high temperature superconductors, $S=0$ and $L=2$ ($d$-$wave$). In superfluid $^3$He, the spin and orbital angular momentum vectors are locked at a fixed angle to one another. This is called [*broken relative spin-orbit symmetry*]{}.[@leg75; @vol90] The equilibrium superfluid state is described as a condenstae of Cooper pairs with a total angular momentum, $J=L\oplus S=0$. In addition, the Cooper pairs can be resonantly excited by sound waves to quantum states with total angular momentum $J=2$. [@mak74] This is reminiscent of diatomic molecules which have similar excited states. The above description applies to the B-phase of superfluid $^3$He, the most stable phase at low pressure. The acoustic Faraday effect occurs in $^3$He-B as a consequence of spontaneously broken relative spin-orbit symmetry.[@moo93] An applied field magnetically polarizes the spins of the Cooper pairs which, through coupling to their orbital motion, rotates the polarization of transverse sound. The rotational excitations of Cooper pairs are essential [@moo93] to our observation of the propagation of transverse acoustic waves in $^3$He-B since they significantly increase the sound velocity making the sound mode much easier to detect; the closer the sound energy is to the energy of the Cooper pair excited state, the stronger is this effect. Furthermore the Cooper pair excited states have a linear Zeeman splitting with magnetic field.[@ave80; @sch81] Of the five ($2J+1$) Zeeman sub-states there is one, $m_J=+1$, which couples to right circularly polarized transverse sound, and a second, $m_J=-1$, couples to left circularly polarized sound. Thus the speeds of these two transverse waves are different in a magnetic field. We call this property acoustic birefringence. It leads to the acoustic Faraday effect where the magnetic field rotates the polarization direction of linearly polarized sound. Our measurements show that the rotation angle can be as large as $1.4\times10^7$ deg/cm-Tesla, much larger than the ususal magneto-optical Faraday effect. [@ben65]
Excitation and detection of transverse sound is provided by a high Q ($\approx 3000$) AC-cut, quartz transducer with a fundamental resonance frequency of 12 MHz. It generates and detects shear waves with a specific linear polarization. The detection method is based on measurement of the electrical impedance of the transducer using a frequency-modulated cw-bridge spectrometer. [@lee96] All measurements were performed at 82.26 MHz, the $7^{th}$ harmonic of the transducer, with frequency modulation at 400 Hz and an amplitude of 3 kHz. The electrical impedance of the transducer is a direct measure of the acoustic impedance of the surrounding liquid $^3$He in the acoustic cavity that is shown in Fig. 1. Linearly polarized waves are excited by the transducer, reflected from the opposite surface of the acoustic cell, and detected by the same transducer. Under conditions of high attenuation there is no reflected wave, and the acoustic response is determined by the bulk acoustic impedance, $Z_a=\rho\omega/q$ where $\rho$ is the density of the liquid, $\omega$ is the sound frequency, $q=k+i\alpha$ is the complex wave number, $\alpha$ is the attenuation, and $2\pi/k$ is the wavelength. A change in either the attenuation or the phase velocity, $C_{\phi}=\omega/k$, produces a change in the impedance, $Z_a$. On cooling into the superfluid the acoustic response shown in Fig. 2 varies smoothly with temperature in this highly attenuating region. If the attenuation is low, there is interference between the source and reflected waves which modulates the local acoustic impedance detected by the transducer. Consequently the acoustic response oscillates as the phase velocity changes with temperature. The oscillations in Fig. 2 at low temperatures correspond to interference between outgoing and reflected waves, and so they indicate the existence of some form of propagating wave. Each period of the oscillations corresponds to a change in velocity sufficient to increase, or decrease, by unity the number of half wavelengths in the cavity. The amplitude of the oscillations increases as the temperature is reduced indicating that attenuation of the sound mode decreases with decreasing temperature.
The features labeled $A$ and $B$ in Fig. 2 are identified with known physical processes for sound absorption in superfluid $^3$He-B. [@vol90; @wol78; @mck90] Feature $A$ corresponds to onset of the dissociation of Cooper pairs by sound where $\hbar\omega=2\Delta(T_A)$. In the temperature range between $T_A$ and the superfluid transition temperature, $T_c$, the attenuation of the liquid is extremely high owing to this mechanism. The point $B$ corresponds to resonant absorption of sound at $\hbar\omega=
1.5\Delta(T)$ by the excited Cooper pairs with angular momentum $J=2$. [@moo93] Transverse sound is extinguished below this temperature. Early attempts to observe transverse sound in the normal phase of $^3$He were inconlusive, [@roa76a; @flo76; @flo78] and furthermore, it was originally expected that the transverse mode would be suppressed in the superfluid phase. [@com76; @mak77] More recent theoretical work [@moo93] clarified the role of the Cooper pair excitations showing that they increase the transverse sound speed which results in a more robust propagating transverse acoustic wave at low temperatures in superfluid $^3$He-B. The first experimental evidence for this can be found in the acoustic impedance measurements of Kalbfeld, Kucera, and Ketterson. [@kal93]
The proof that the impedance oscillations correspond to a propagating [*transverse*]{} sound mode is given in Fig. 3. In Fig. 3[**a**]{} we show data sets at a pressure of $4.42\, \rm{bar}$ in magnetic fields of $52\,\rm{G}$, $101\,\rm{G}$ and $152\,\rm{G}$. The principal feature is that the magnetic field modulates the zero field oscillations shown in Fig. 2. Our detector is only sensitive to linearly polarized transverse sound having a specific direction. Application of a field of $52\,\rm{G}$ in the direction of wave propagation suppresses the oscillations near $T=0.465\,T_c$ that were present in zero field. This corresponds to a $90^{\circ}$ rotation of the polarization of the first reflected transverse sound wave making the polarization orthogonal to the detection direction. Doubling the magnetic field restores the transverse sound oscillations at this temperature. The oscillations are suppressed once again by tripling the field to $152\,\rm{G}$. Also note that near the points labeled $90^{\circ}$ and $270^{\circ}$, there are smaller amplitude impedance oscillations with shorter period than the primary oscillations. These come from interference of doubly reflected waves within the cavity. We demonstrate this fact with a simple, but powerful, simulation of the acoustic impedance oscillations shown in Fig. 3[**b**]{}.
In zero field, superfluid $^3$He-B is non-magnetic and non-birefringent. Linearly polarized transverse sound is the superposition of two circularly polarized waves having the same velocity and attenuation. Application of a magnetic field gives rise to acoustic [*circular*]{} birefringence through the Zeeman splitting of the excited states of the Cooper pairs that couple to the transverse sound modes; thus, right- and left-circularly polarized waves propagate with different speeds, $C_{\pm}=C_{\phi}\pm\delta C_{\phi}$. For magnetic fields well below $1\,\mbox{kG}$ the difference in propagation speeds is linear in the magnetic field, $\delta C_{\phi}\propto H$. This implies that a linearly polarized wave generated by the transducer undergoes Faraday rotation of its polarization as it propagates. Upon reflection from the opposite wall of the cavity the linearly polarized wave with ${\bf q}\parallel{\bf H}$ reverses direction. The reflected wave propagates with the polarization rotating with the [*same handedness relative to the direction of the field*]{}, i.e. the rotation of the polarization accumulates after reflection from a surface. The spatial period for rotation of the polarization by $360^{\circ}$ is
$$\label{Period}
\Lambda=
4\pi{\left(C_{\phi}\over\omega\right)}{\left|C_{\phi}\over\delta
C_{\phi}\right|}
\,.$$
The Faraday effect produces a sinusoidal modulation of the impedance oscillations as a function of magnetic field with a period that is inversely proportional to the field, i.e. $\Lambda\propto 1/H$. The constant of proportionality in magneto-optics is called the Verdet constant, $\mbox{V}=2\pi/H\Lambda$.
In Fig. 3[**b**]{} we show the result of our numerical calculation of the sound wave amplitude in the direction detected by the transducer. The oscillations shown in the figure come from interference between the source wave and multiply reflected waves. The calculation uses the attenuation and phase velocity measured in zero field. The Verdet constant is obtained from the measurement at $52\,\mbox{G}$. The simulation reproduces all the observed features of the impedance as a function of temperature including the maximum in the modulation at $T/T_c=0.415\,,\,H=101\,\mbox{G}$ and the minimum at $T/T_c=0.415\,,\,H=152\,\mbox{G}$, which confirms that the Faraday period is proportional to $1/H$. The simulation also produces the fine structure oscillations in the impedance near the points labeled $90^{\circ}$ and $270^{\circ}$. The fine structure is observed when the polarization rotates by an odd multiple of $90^{\circ}$ upon a single round trip in the cell. Then waves that traverse the cell twice are $180^{\circ}$ out of phase relative to the source wave, and consequently the period of the impedance oscillations is halved. The amplitude of the oscillations is substantially reduced because of attenuation over the longer pathlength. This structure provides proof that impedance oscillations are modulated by the Faraday effect for propagating transverse waves.
The impedance data from our experiments were analyzed to obtain the spatial period for the rotation of the polarization and were found to be in agreement with the theoretical prediction [@moo93] for the Faraday rotation period. The theoretical results for the period can be expressed in the form,
$$\label{Lambda-theory}
\Lambda=K{\sqrt{T/T_{+}-1}\over gH}.$$
for fields $H \ll 1\,\mbox{kG}$ and temperatures above and near the extinction point $B$. The temperature, $T_{+}$, corresponds to the extinction of transverse sound by resonant excitation of Cooper pairs with $J=2, m_J=+1$, at a slightly higher temperature than the $B$ extinction point in zero field as shown in the inset to Fig. 2 ($e.g.$ at $H=100\,\mbox{G}$, $T_{+}-T_B
\approx 1~{\mu}K$). The magnitude of the Faraday rotation period depends on accurately known superfluid properties, contained in the parameter $K$, as well as one parameter that is not well-established, the Landé g-factor, $g$, for the Zeeman splitting of the $J=2$, Cooper pair excited state.
Movshovich,[*et al.*]{}[@mov88] analyzed the splitting of the $J=2$ multiplet in the absorption spectrum of longitudinal sound to find a value of $g=0.042$. In that experiment it was not possible to resolve the splitting except for fields above $2\,\mbox{kG}$. At these high fields the non-linear field dependence due to the Paschen-Back effect[@sch83; @shi83] becomes comparable to the linear Zeeman splitting,[@hal90; @mov91] which makes it difficult to determine the Landé g-factor accurately. We have analyzed our measurements of the acoustic Faraday effect to determine the g-factor with high accuracy at low fields, which eliminates the complication of the high-field Paschen-Back effect. We find $g=0.020\pm 0.002$. Our significantly smaller value of the Landé g-factor has the interpretation that there are important $L=3$ ($f$-$wave$) pairing correlations in the superfluid condensate, about $7\%$ of the dominant $p$-$wave$ interactions.[@sau82]
[10]{}
Landau, L. D. Oscillations in a Fermi liquid. , 101-108 (1957).
Leggett, A. J. A theoretical interpretation of the new phases of liquid $^3$[H]{}e. , [**47**]{}, 331-414 (1975).
Vollhardt, D. and Wölfle, P. . (Taylor & Francis, New York, 1990).
Anderson, P. W. and Brinkman, W. F. Theory of anisotropic superfluidity in $^3$[H]{}e. in [*[H]{}elium Liquids*]{}. ed. J. G. M. Armitage and I. E. Farquhar, p. 315 (Academic Press, New York, 1975).
Wheatley, J. C. Experimental properties of superfluid $^3$[H]{}e. , 415-470 (1975).
Bardeen, J. , Cooper, L. N., and Schrieffer, R. heory of [S]{}uperconductivity. , 1175-1204 (1957).
Maki, K. Propagation of zero sound in the [B]{}alian-[W]{}erthamer state. , 465-477 (1974).
Moores, G. F., and Sauls, J. A. Transverse [W]{}aves in [S]{}uperfluid . ,13-37 (1993).
Avenel, O., Varoquaux, E., and Ebisawa, H. Field splitting of the new attenuation peak in $^3$[H]{}e-B. , [**45**]{}, 1952-1955 (1980).
Schopohl, N. and Tewordt, L. Landé factors of collective mode multiplets in $^3$[H]{}e-B and coupling strengths to sound waves. , 67-90 (1981).
Bennett, H. S. and Stern, E. A. Faraday and Kerr effects in ferromagnetics. , A448-461 (1965).
Lee, Y., Hamot, P. J., Meisel, M. W., Sprague, D. T., Haard, T. M., Kycia, J. B., Rand, M. R., and W.P. Halperin High frequency acoustic measurements in liquid $^3$[H]{}e near the transition temperature. , 265-272 (1996).
Wölfle, P. and Einzel, D. Transport and relaxation properties of superfluid $^3$[H]{}e. II. , 39-56 (1978).
Mc[K]{}enzie, R. H. and Sauls, J. A. ollective [M]{}odes and [N]{}onlinear [A]{}coustics in [S]{}uperfluid [[$^3{H}e$-[B]{}]{}]{}. in [*[H]{}elium Three*]{}. ed. W. P. Halperin and L. P. Pitaevskii, p. 255 (Elsevier Science Publishers, Amsterdam, 1990).
Roach, P. R. and Ketterson, J. B. Observation of transverse zero sound in normal $^3$[H]{}e. , [**36**]{}, 736-740 (1976).
Flowers, E. G., Richardson, R. W., and Williamson, S. J. Transverse zero sound in normal $^3$[H]{}e. , 309-311 (1976).
Flowers, E. G. and Richardson, R. W. Transverse acoustic impedance of normal $^3$[H]{}e. ,1238-1248 (1978).
Combescot, M. and Combescot, R. Transverse zero sound propagation in superfluid $^3$[H]{}e. ,181-182 (1976).
Maki, K., and Ebisawa, H. Transverse zero sound in superfluid $^3$[H]{}e. , 627-636 (1977).
Kalbfeld, S., Kucera, D. M., and Ketterson, J. B. Observation of an evolving standing-wave pattern involving a transverse disturbance in superfluid $^3$[H]{}e. , 2264-2267 (1993).
Movshovich, R., Varoquaux, E., Kim, N., and Lee, D. M. Splitting of the squashing collective mode of superfluid $^3$[H]{}e-[B]{} by a magnetic field. , 1732-1735 (1988).
Schopohl, N., Warnke, M., and Tewordt, L. Effect of gap distortion on the field splitting of collective modes in superfluid $^3$[H]{}e-[B]{}. , [**50**]{}, 1066-1069 (1983).
Shivaram, B. S., Meisel, M. W, Sarma, B. K., Halperin, W. P., and Ketterson, J. B. Nonlinear Zeeman shifts in the collective-mode spectrum of $^3$[H]{}e-[B]{}. , [**50**]{}, 1070-1072 (1983).
Halperin, W. P. and Varoquaux, E. Order [P]{}arameter [C]{}ollective [M]{}odes in [S]{}uperfluid $^3$[H]{}e. In W. P. Halperin and L. P. Pitaevskii, editors, [*[H]{}elium Three*]{}, p. 353 (Elsevier Science Publishers, Amsterdam, 1990).
Movshovich, R., Varoquaux, E., Kim, N., and Lee, D. M. Fivefold splitting of the squashing mode of superfluid $^3$[H]{}e-[B]{} by a magnetic field. , 332-340 (1991).
Sauls, J. A. and Serene, J. W. Interaction effects on the [Z]{}eeman splitting of collective modes in superfluid $^3$[H]{}e-[B]{}. , 1183-1186 (1982).
=0.75
=0.75
=0.75
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Исследуются погружения графов в проективную плоскость. Получена классификация погружений с точностью до регулярной гомотопности. Построен полный инвариант погружений с точностью до регулярной гомотопности. Случай погружений графов в любую компактную поверхность, отличную от проективной плоскости, был известен. [**Ключевые слова:**]{} регулярное погружение графа, регулярная гомотопия погружений, движения Рейдемейстера, число самопересечения.'
author:
- 'M.A. Ivashkovskii'
title: Immersions of graphs to the projective plane
---
[**Abstract.**]{} Immersions of graphs to the projective plane are studied. A classification of immersions up to regular homotopy is obtained. A complete invariant of immersions up to regular homotopy is constructed. The case of graphs immersions to any compact surface different from the projective plane was known.
[**Key words:**]{} regular immersion of a graph, regular homotopy of immersions, Reidemeister moves, self-intersection number.
ПОГРУЖЕНИЯ ГРАФОВ В ПРОЕКТИВНУЮ ПЛОСКОСТЬ\
М.А. Ивашковский
Пусть дан связный граф $G$ (возможно, имеющий петли и кратные ребра) с выделенной на нем вершиной $v$. Рассмотрим погружение $\gamma: G \looparrowright M$ графа $G$ в связное компактное гладкое двумерное многообразие $M$ (определение 1 ниже). Требуется получить классификацию всех возможных погружений с точностью до регулярной гомотопности (определение 2 ниже). В настоящей заметке излагается решение \[1\] этой задачи в случае, когда $M=\mathbb R P^2$ — проективная плоскость, и построен полный инвариант погружений графа $G$ в проективную плоскость в терминах индекса самопересечения кривых по модулю 2. Полный инвариант погружений графа $G$ в любую поверхность $M\ne\mathbb R P^2$ был построен Д.А. Пермяковым \[2\] в терминах чисел вращения кривых.
Перейдем к точным формулировкам. Будем предполагать, что граф $G$ состоит из одной вершины и $n$ ребер (петель) $e_i$, $i=1,\dots,n$, т.е. является букетом $n$ окружностей. Случай произвольного графа легко сводится к этому случаю \[2\]. Выберем и фиксируем параметризацию на каждом ребре $e_i$ графа $G$.
Отображение $\gamma:G \to M$ назовем [*погружением*]{} графа $G$ в поверхность $M$ (и обозначим через $\gamma:G \looparrowright M$), если его ограничение на любое замкнутое ребро $\overline{e_i}$ является регулярной кривой (относительно параметризации $t\mapsto(\gamma|_{e_i})(t)$, $t\in[0,1]$, отвечающей заданной параметризации ребра $e_i$), и кроме того $2n$ касательных векторов $\frac{d}{dt}|_{t=0}(\gamma|_{\overline{e_i}}),-\frac{d}{dt}|_{t=1}(\gamma|_{\overline{e_i}})$, $i=1,\dots,n$, в точке $\gamma(v)$ попарно несонаправлены, где $v$ — вершина, в которой сходятся петли графа $G$.
Семейство погружений $\gamma_u: G\looparrowright M$, $u\in[0,1]$, графа в поверхность назовем [*регулярной гомотопией*]{}, если оно является гомотопией в обычном смысле (т.е. отображение $\Gamma:G\times[0,1]\to M$, $\Gamma(x,u)=\gamma_u(x)$, непрерывно), и ограничение этой гомотопии на любое замкнутое ребро $\overline{e_i}$ задается $C^\infty$-гладким отображением $[0,1]\times[0,1]\to M$, $(t,u)\mapsto(\gamma_u|_{\overline{e_i}})(t)$, $t,u\in[0,1]$. Два погружения $\gamma_0,\gamma_1$ назовем [*регулярно гомотопными*]{} (обозначение $\gamma_0 \overset{\text{\it reg}}{\sim} \gamma_1$), если их можно соединить регулярной гомотопией $\gamma_u: G\looparrowright M$, $u\in[0,1]$.
Опишем построение полного инварианта регулярной гомотопности погружений.
Обозначим $\gamma|_{e_i}=:e'_1$, $i=1,\dots,n$. Из вершины $v':=\gamma(v)$ графа $G':=\gamma(G)$ выходят полуребра $e'_1,\dots,e'_n$ и в нее входят полуребра $(e'_1)^{-1},\dots,(e'_n)^{-1}$. Пусть $or$ — локальная ориентация касательной плоскости $T_{\gamma(v)}M$ в точке $\gamma(v)$. Выберем обход вершины $v'$, согласованный с ориентацией $or$. Будем выписывать полуребра $e'_i,(e'_i)^{-1}$ в порядке их появления во время обхода. Рассмотрим соответствующий набор из $e_i,e_i^{-1}$ (т.е. соответствующее слово в алфавите $\{e_1,\dots,e_n,e_1^{-1},\dots,e_n^{-1}\}$) с точностью до циклической перестановки элементов набора назовем [*циклическим порядком полуребер в вершине*]{}, отвечающим погружению $\gamma$ и локальной ориентации $or$. При изменении локальной ориентации получим другой циклический порядок; полученную неупорядоченную пару циклических порядков назовем [*циклическим порядком полуребер в вершине*]{} для погружения $\gamma$.
Имеем биекцию между множеством всевозможных циклических порядков для оснащенных погружений и множеством перестановок $\Sigma_{2n-1}$ порядка $2n-1$. Изменение локальной ориентации дает действие (тривиальное при $n=1$, свободное при $n>1$) группы $\mathbb Z_2$ на множестве $\Sigma_{2n-1}$. Получаем биекцию между множеством всевозможных циклических порядков для погружений и пространством орбит $\Sigma_{2n-1}/\mathbb Z_2$.
Так как фундаментальная группа проективной плоскости абелева, она не зависит от выбора базисной точки. Поэтому будем опускать базисную точку в обозначении фундаментальной группы.
\[[*индекс самопересечения*]{} замкнутой регулярной кривой на поверхности\]Пусть $\gamma:[0,1]\looparrowright M$ — замкнутая регулярная кривая (быть может, имеющая “излом” в точке $\gamma(0)=\gamma(1)$), и в точке $\gamma(0)$ фиксирована ориентация касательной плоскости $T_{\gamma(0)}M$, причем векторы $\dot\gamma(0)$ и $\dot\gamma(1)$ не являются противоположно направленными. Предположим, что все точки самопересечения этой кривой являются трансверсальными (этого можно добиться малой деформацией кривой). Пусть $(t_1,t_2)$ — точка самопересечения кривой $\gamma$, т.е. $\gamma(t_1)=\gamma(t_2)=:A$, причем $0\leq t_1<t_2\leq 1$ и $(t_1,t_2)\ne(0,1)$. Перенесем ориентацию вдоль кривой $\gamma|_{[0,t_1]}$ в точку $A$. Возьмем репер $(\dot\gamma(t_1),\dot\gamma(t_2))$ и сравним его ориентацию с перенесенной ориентацией. В зависимости от согласованности припишем точке самопересечения $(t_1,t_2)$ знак $+1$ или $-1$. Просуммировав эти числа по всем точкам самопересечения, получим [*индекс самопересечения*]{} кривой $\gamma$ относительно заданной локальной ориентации, обозначим его через $I(\gamma)\in\mathbb Z$. Заметим, что число $I(\gamma)\ mod \ 2\in\mathbb Z_2$ не зависит от выбора локальной ориентации.
Пусть $\gamma_1,\gamma_2$ — два погружения графа $G$ в проективную плоскость $\mathbb RP^2$. Эти погружения регулярно гомотопны (т.е. $\gamma_1 \overset{\text{\it reg}}{\sim} \gamma_2$) тогда и только тогда, когда $Inv(\gamma_1) = Inv(\gamma_2)$. Здесь функционал $Inv:\{\gamma:G\looparrowright \mathbb RP^2\}\to(\Sigma_{2n-1}/\mathbb Z_2)\times \{0,1\}^{2n}$ определяется формулами $$Inv_1(\gamma):=\Big( \mbox{циклический порядок полуребер для $\gamma$} \Big)\in\Sigma_{2n-1}/\mathbb Z_2,$$ $$Inv_2(\gamma):=\Big([\gamma|_{e_1}],\dots,[\gamma|_{e_n}]\Big) \in \Big(\pi_1 (\mathbb RP^2)\Big)^n \cong (\mathbb Z_2)^n=\{0,1\}^n,$$ $$Inv_3(\gamma):=\Big(I (\gamma|_{e_1}) \ mod \ 2,\dots,I (\gamma|_{e_n}) \ mod \ 2
\Big) \in (\mathbb Z_2)^n=\{0,1\}^n,$$ $$Inv(\gamma):=\Big(Inv_1(\gamma),Inv_2(\gamma),Inv_3(\gamma)\Big).$$
Теорема следует из лемм 1 и 3, приведенных ниже. Для каждой петли $e_i$, $i=1,\dots,n$, графа $G$ возможны два случая. Первый — когда петля $\gamma_1|_{e_i}$ стягиваема. Второй — когда эта петля нестягиваема.
![Движения Рейдемейстера типов 1, 2 и 3[]{data-label="fig:Reid"}](2a.jpg "fig:"){width="20.00000%"} ![Движения Рейдемейстера типов 1, 2 и 3[]{data-label="fig:Reid"}](3a.jpg "fig:"){width="30.00000%"} ![Движения Рейдемейстера типов 1, 2 и 3[]{data-label="fig:Reid"}](4a.jpg "fig:"){width="36.00000%"}
Замкнутую кривую $\gamma:S^1\to M$ назовем [*кривой общего положения*]{}, если она регулярна (и, возможно, имеет “излом” в вершине), все ее точки самопересечения (см. определение 4) являются трансверсальными, и образы $\gamma(t_1)=\gamma(t_2)$ ее точек самопересечения $(t_1,t_2)$ попарно различны. Согласно известной теореме Рейдемейстера \[3\], две замкнутые кривые общего положения $\gamma_0,\gamma_1:[0,1]\to M$ на поверхности $M$ гомотопны тогда и только тогда, когда существует гомотопия, разлагающаяся в последовательность движений Рейдемейстера 1-го, 2-го и 3-го типов (рис. \[fig:Reid\]).
Если $\gamma_1 \overset{\text{\it reg}}{\sim} \gamma_2$, то $Inv(\gamma_1)=Inv(\gamma_2)$.
Если погружения $\gamma_1$ и $\gamma_2$ регулярно гомотопны, то они локально регулярно гомотопны в некоторой окрестности вершины $v$. Значит, циклический порядок относительно ориентации меняться не будет.
Если погружения $\gamma_1$ и $\gamma_2$ регулярно гомотопны, то две петли $\gamma_1|_{e_i}$ и $\gamma_2|_{e_i}$ регулярно гомотопны, а потому гомотопны. Значит, они либо обе стягиваемы, либо обе нестягиваемы.
Для замкнутой регулярной кривой индекс самопересечения $mod \ 2$ при регулярной гомотопии меняться не будет. Это следует из “регулярного аналога” теоремы Рейдемейстера \[3\]: индекс самопересечения кривой сохраняется при движениях Рейдемейстера второго и третьего типов, а потому и при любой регулярной гомотопии.
Таким образом, если $\gamma_1$ и $\gamma_2$ регулярно гомотопны, то значения введенного функционала на них совпадают. То есть, функционал инвариантен относительно регулярной гомотопности погружений.
Теперь проведем доказательство в обратную сторону.
[Лемма 2.]{} Если замкнутые регулярные кривые $\gamma_0,\gamma_1:S^1\looparrowright M$ на поверхности $M$ гомотопны, то $\gamma_0 \overset{\text{\it reg}}{\sim} \tilde\gamma_1$ для некоторой регулярной кривой $\tilde\gamma_1$, получающейся из кривой $\gamma_1$ прибавлением некоторого количества маленьких петель (завитков). Если при этом кривые $\gamma_0,\gamma_1$ совпадают на некоторой дуге $\delta\subset S^1$, то указанная регулярная гомотопия может быть выбрана неподвижной на дуге $\delta$.
По теореме Рейдемейстера \[3\], ввиду гомотопности кривых, существует последовательность движений Рейдемейстера, приводящая кривую $\gamma_0$ к кривой $\gamma_1$. Выполним эту последовательность движений со следующей модификацией: каждый раз, когда предполагается совершить движение Рейдемейстера типа 1 (рождение или уничтожение завитка), мы вместо его выполнения будем оставлять на малом участке кривой маленький завиток (который будет сохраняться, жестко двигаясь по поверхности вместе с этим участком кривой, в течение всей последующей регулярной гомотопии).
![Взаимное уничтожение пары маленьких петель[]{data-label="fig:pair"}](9a.jpg){width="60.00000%"}
Более подробно: если движение Рейдемейстера типа 1 состоит в [*уничтожении*]{} завитка, то вместо его уничтожения мы сохраним его (в течение всей последующей регулярной гомотопии) в виде маленького завитка. А если движение Рейдемейстера типа 1 состоит в [*рождении*]{} завитка, то мы непосредственно перед выполнением этого движения породим (как на рис. \[fig:pair\]) регулярной гомотопией пару близлежащих завитков противоположных знаков, и один из них (нужного знака) используем в качестве рожденного завитка при движении Рейдемейстера, а другой завиток сохраним (в течение всей последующей регулярной гомотопии) в виде маленького завитка на данном участке кривой.
В результате мы получим регулярную гомотопию, преобразующую кривую $\gamma_0$ к некоторой кривой $\tilde\gamma_1$ искомого вида, т.е. получающейся из $\gamma_1$ добавлением нескольких завитков.
![“Протаскивание через бесконечность”[]{data-label="fig:protask"}](protask1.jpg "fig:"){width="45.00000%"} ![“Протаскивание через бесконечность”[]{data-label="fig:protask"}](protask2.jpg "fig:"){width="45.00000%"}
Для замкнутой регулярной кривой $\gamma$ на $\mathbb RP^2$ введем движение, которое будем называть *протаскиванием через бесконечность*. Для этого перейдем к накрытию проективной плоскости $\mathbb RP^2$ сферой $S^2$ и продеформируем некоторую простую дугу $\gamma|_{\delta}$, $\delta \not\ni v$, при помощи регулярной гомотопии, неподвижной вне $\delta$ и преобразующей дугу $\gamma|_\delta$ в дугу вида $\gamma|_\delta$ с двумя петлями одного знака, как показано на рис. \[fig:protask\]. Итак, с помощью данного движения можно менять индекс самопересечения любой замкнутой регулярной кривой $\gamma$ в $\mathbb RP^2$ на $+2$ или $-2$ при помощи регулярной гомотопии, неподвижной в некоторой окрестности концов кривой.
Если $Inv(\gamma_1)=Inv(\gamma_2)$, то $\gamma_1 \overset{\text{\it reg}}{\sim} \gamma_2$.
Совместим точки $\gamma_1(v)=v'_1$ и $\gamma_2(v)=v'_2$ так, чтобы совпали циклические порядки в вершине $v$ для погружений $\gamma_1$ и $\gamma_2$. Так как циклический порядок один и тот же, то сами эти погружения можно совместить в малой окрестности $U \subseteq G$ вершины $v$ путем регулярной гомотопии, неподвижной в вершине $v$.
Пусть $U'\subset M=\mathbb R P^2$ — малая круговая окрестность вершины $v'=v_1'=v_2'$ графа $G_1'=\gamma_1(G)$. Фиксируем локальную ориентацию касательной плоскости $T_{v'}\mathbb R P^2$. Фиксируем любое ребро $e_i$ графа $G$, $i=1,\dots,n$. Обозначим через $I_i$ индекс самопересечения $I(\gamma_1|_{e_i}) \ mod \ 2=I(\gamma_2|_{e_i}) \ mod \ 2\in\{0,1\}$, который будем рассматривать как целое число. Возможны два случая.
[*Случай 1:*]{} петля $\gamma_1|_{e_i}$ стягиваема. Из-за стягиваемости петли $\gamma_1|_{e_i}$ существует последовательность движений Рейдемейстера типов 2 и 3, деформирующих эту петлю к некоторой петле $\tilde{\gamma}_1|_{e_i} \subset U' \subset \mathbb RP^2$, причем эта деформация неподвижна в некоторой окрестности $V\subset U$ вершины $v$ графа $G$. Аналогично существует последовательность движений и для $\gamma_2$ такая, что $\tilde{\gamma}_2(e_i) \subset U'$. Петля $\tilde\gamma_1|_{e_i}$ по лемме 2 регулярно гомотопна в диске $U'$ положительно ориентированной окружности с некоторым числом завитков, причем соответствующая регулярная гомотопия неподвижна на $V\cap e_i$.
Любую пару соседних завитков, имеющих один знак, взаимно уничтожим регулярной гомотопией, обратной к “протаскиванию через бесконечность” (рис. \[fig:protask\]). А любую пару соседних завитков с противоположными знаками можно взаимно уничтожить регулярной гомотопией (неподвижной на $V\cap e_i$) в диске $U'$, изображенной на рис. \[fig:pair\]. Отметим, что все указанные гомотопии неподвижны на $V\cap e_i$. Таким образом, петлю $\tilde\gamma_1|_{e_i}$ можно преобразовать с помощью регулярной гомотопии, неподвижной на $V\cap e_i$, в положительно ориентированную окружность с 0 или 1 положительным завитком. Заметим, что количество завитков равно $I_i$ (так как индекс самопересечения по модулю 2 регулярной замкнутой кривой сохраняется при втором и третьем движениях Рейдемейстера, а потому и при любой регулярной гомотопии). Поэтому к этому же виду можно преобразовать и петлю $\tilde\gamma_2|_{e_i}$.
Отсюда следует, что петли $\gamma_1|_{e_i}$ и $\gamma_2|_{e_i}$ регулярно гомотопны с помощью регулярной гомотопии, неподвижной на $V\cap e_i$.
[*Случай 2:*]{} петля $\gamma_1|_{e_i}$ нестягиваема. В данном случае петля $\gamma_1|_{e_i}$ пересекает ленту Мебиуса $\mathbb R P^2\setminus U'$. Петля $\gamma_1|_{e_i}$ по лемме 2 регулярно гомотопна некоторой кривой, пересечение которой с указанным листом Мебиуса является простой дугой, а пересечение с диском $U'$ имеет вид простой дуги с некоторым количеством маленьких завитков. При этом соответствующая регулярная гомотопия неподвижна в малой окрестности $V\cap e_i$ вершины $v$.
Взаимно уничтожим пары соседних завитков как в случае 1 с помощью регулярной гомотопии, неподвижной на $V\cap e_i$. То есть, в данном случае получаем, что нестягиваемую петлю $\gamma_1|_{e_i}$ можно преобразовать с помощью регулярной гомотопии, неподвижной на $V\cap e_i$, в петлю, пересечение которой с листом Мебиуса $\mathbb R P^2\setminus U'$ является простой дугой, а пересечение с диском $U'$ имеет вид простой дуги с 0 или 1 положительным завитком. Как и в случае 1 получаем, что количество завитков равно $I_i$. Поэтому к этому же виду можно преобразовать и петлю $\gamma_2|_{e_i}$, что и завершает доказательство леммы.
Автор благодарен Е.А. Кудрявцевой за постановку задачи и полезные обсуждения.
[99]{} Graphs immersions to the projective plane // Moscow Univ. Math. Bull. 2017 (to appear). arXiv:1611.09634.
[Permyakov D. A.]{} Regular homotopy for immersions of graphs into surfaces // Sb. Math. 2016. [**207**]{}, N 6. 854–872.
Elementare Begründung der Knotentheorie // Abh. Math. Sem. Univ. Hamburg. 1926. [**5**]{}. 24–32.
Maxim A. Ivashkovskii
Moscow State University (Moscow, Russia)
[email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The interactions of the neutrini and antineutrini amongst themselves, as well as the interactions of these particles with the electrons and the positrons, are of interest in simulations of the early Universe and in studies of the processes involving compact stars. The objective in this paper is to create a reliable source of information regarding the differential and total cross sections of these interactions; expressions for these observables will be obtained using standard methodology. A number of relevant discrepancies in the literature will be addressed.\
[*PACS:*]{} 13.15.+g, 14.60.Lm, 14.60.St, 98.80.Cq
author:
- 'E. Matsinos'
title: 'Interactions relevant to the decoupling of the neutrini/antineutrini in the early Universe'
---
neutrino, differential cross section, total cross section, decoupling, early Universe
E-mail: evangelos\[dot\]matsinos\[at\]sunrise\[dot\]ch
\[sec:Introduction\]Introduction
================================
Despite the fact that the expressions for the cross sections, corresponding to the interactions of the neutrini and antineutrini (henceforth, (anti)neutrini for short) with the ingredients of the plasma of the early Universe, are straightforward to obtain, the information retrieved from several sources in the literature, be they books, scientific publications, or presentations in Conferences, is frequently incorrect. To the best of my knowledge, there is no place in the literature where the correct formulae, pertaining to these scattering processes, are listed in a manner which is not subject to misinterpretation and misunderstanding.
Such interactions are interesting not only in the context of simulations of the early Universe, but also in terms of the physical processes taking place in the collapsing cores of supernovae. In fact, the early papers on this subject had been stimulated by studies of the physics of supernova explosions. The neutrino-antineutrino annihilation to electron-positron pairs has been investigated for one additional reason, namely in terms of the possibility that it provides a driving mechanism for the creation of gamma-ray bursts in compact stars, i.e., in white dwarves, neutron stars, and black holes.
The goal in this paper is to create a reliable source of information with respect to these interactions, as well as to discuss a number of relevant discrepancies in the literature. One additional reason for writing this paper is that only convenient high-energy approximations are quoted in most cases for the interactions of the (anti)neutrini with the electrons and the positrons of the plasma, i.e., results obtained from calculations ignoring the rest mass of these particles; both the exact formulae and their approximated expressions at high energy are given in this work.
The structure of this paper is as follows. Sections \[sec:Conventions\]-\[sec:Preliminary\] provide the tools required for the evaluation of the cross sections. The cross sections are derived for all relevant processes in Section \[sec:Derivation\] and are presented in tabular and schematic forms in Section \[sec:Summary\]. Section \[sec:Discrepancies\] provides a discussion of those of the discrepancies in the literature which I am aware of. Section \[sec:Conclusions\] briefly summarises the findings of this paper.
\[sec:Conventions\]Notation and conventions
===========================================
In the present work, all physical variables and observables refer to the ‘centre of momentum’ (CM) frame of reference (unless, of course, they are Lorentz-invariant quantities, in which case there is no need to specify a frame of reference). Only processes with two particles in both the initial and final states are considered (i.e., $2 \to 2$ scattering). Used are the following notations and conventions.
- The speed of light in vacuum $c$ is equal to $1$.
- Einstein’s summation convention is used.
- $I_n$ denotes the $n \times n$ identity matrix.
- $A^\dagger$ denotes the Hermitian (conjugate transpose) of a square matrix $A$.
- $g_{\mu \nu}$ denotes the Minkowski metric with signature ‘$+ \, - \, - \, -$’.
- $\gamma_{\mu}$ are the standard Dirac $4 \times 4$ matrices, satisfying the relation $\{ \gamma_{\mu}, \gamma_{\nu} \} = 2 g_{\mu \nu} I_4$ for $\mu, \nu \in \{ 0,1,2,3 \}$. In addition, $\gamma_0^\dagger = \gamma_0$ and $\gamma_\mu^\dagger = \gamma_0 \gamma_\mu \gamma_0 = -\gamma_\mu$ for $\mu \in \{1,2,3\}$.
- The matrix $\gamma_5 \coloneqq i \gamma_0 \gamma_1 \gamma_2 \gamma_3$ enters the two projection operators to fermion states of definite chirality, i.e., the operators projecting a Dirac field onto its left- and right-handed components. The matrix $\gamma_5$ satisfies the relations: $\gamma_5^\dagger=\gamma_5$, $\gamma_5^2 = I_4$, and $\{ \gamma_{\mu}, \gamma_5 \} = 0$ for $\mu \in \{0,1,2,3\}$.
- The quantity $\epsilon_{i_0 i_1 \dots i_n}$ is the Levi-Civita symbol, defined as follows: $$\epsilon_{i_0 i_1 \dots i_n} \coloneqq \left\{
\begin{array}{rl}
+1 & \text{if $(i_0, i_1, \dots, i_n)$ is an even permutation of $(0, 1, \dots, n)$}\\
-1 & \text{if $(i_0, i_1, \dots, i_n)$ is an odd permutation of $(0, 1, \dots, n)$}\\
0 & \text{otherwise}
\end{array} \right.$$
- $s$, $u$, and $t$ are the standard Mandelstam variables. The quantity $s$ is equal to the square of the CM energy; $t$ is equal to the square of the $4$-momentum transfer; $u$ is equal to the square of the $4$-momentum transfer for interchanged final-state particles.
- For a $4$-vector $a$, ${\ensuremath{\mathrlap{\!\not{\phantom{a}}}a}} \coloneqq \gamma_{\mu} a^{\mu}$.
- $k$ and $p$ are the $4$-momenta of the incident particles (projectile and target, respectively).
- $k^\prime$ and $p^\prime$ are the $4$-momenta of the scattered particles.
- $q \coloneqq k-k^\prime=p^\prime-p$ is the $4$-momentum transfer.
- $\theta$ and $\Omega$ denote the scattering angle and the solid angle, respectively.
- $m_e$ stands for the electron (and positron) rest mass.
\[sec:FD\]Feynman graphs relevant to the decoupling of the (anti)neutrini
=========================================================================
It is generally believed that, between the temperatures of one hundred billion K ($T_1=10^{11}$ K), corresponding to a mere one-hundredth of a second after the Big Bang, and about $T_2=2 \cdot 10^{10}$ K, corresponding to the cosmological time of about $250$ ms, when the (anti)neutrini decoupled from the other ingredients of the primordial plasma [^1], our Universe was a superdense mixture of neutrini, antineutrini, electrons, positrons, and photons. In contrast to the huge densities of these particles, a few baryons (protons and neutrons) were also present, about six such particles for every ten billion of photons. Of interest in the context of this work are the interactions of the (anti)neutrini with other (anti)neutrini, as well as with the electrons and the positrons, during the temporal interval corresponding to the temperatures $T_1$ and $T_2$.
The permissible tree-level Feynman graphs (henceforth, simply graphs) for the interaction of an electron neutrino (projectile) with the leptons of the plasma (targets), resulting in no more than an electron-positron ($e^+ e^-$) pair in the final state, are shown in Figs. \[fig:NuNu\]-\[fig:NuElPos\]. All these graphs involve the exchange of either of the two intermediate vector bosons (IVBs) which are associated with the weak interaction, namely the $Z^0$ boson, which is associated with the neutral current (NC), and the $W^\pm$ boson, which is associated with the charged current (CC). Given that it yields a cross section several orders of magnitude smaller [@dr], the elastic scattering between the (anti)neutrini and the photons may be safely omitted.
![\[fig:NuNu\]The permissible tree-level Feynman graphs for the interaction of an electron neutrino (projectile) with the neutrini of the electron, muon, and $\tau$-lepton generations of matter (targets). Graphs (a) and (b) relate to the scattering off the electron neutrino. Graph (c) relates to the scattering processes $\nu_e \nu_{\mu,\tau} \to \nu_e \nu_{\mu,\tau}$. In all cases, these $t$-channel exchanges involve the weak neutral current. Similar graphs are applicable in case of the muon and $\tau$-lepton neutrini as projectiles.](NuNu.eps){width="15cm"}
![\[fig:NuNuBar\]The permissible tree-level Feynman graphs for the elastic scattering of an electron neutrino (projectile) off the antineutrini of the electron, muon, and $\tau$-lepton generations of matter (targets). Graphs (a) and (b) relate to the $t$-channel scattering off the electron antineutrino, graph (a) via the exchange of the weak neutral current, graph (b) via the exchange of the weak charged current. Graphs (c)-(e) relate to the $s$-channel scattering off the electron antineutrino (annihilation graphs); in all such cases, only the weak neutral current is involved. Graph (f) relates to the $t$-channel scattering processes $\nu_e \bar{\nu}_{\mu,\tau} \to \nu_e \bar{\nu}_{\mu,\tau}$, where only the weak neutral current may be exchanged. Similar graphs are applicable in case of the muon and $\tau$-lepton neutrini as projectiles, save for graph (b) which cannot contribute in the energy range explored in this paper.](NuNuBar.eps){width="15cm"}
![\[fig:NuElPos\]The permissible tree-level Feynman graphs for the elastic scattering of an electron neutrino (projectile) off an electron and a positron (targets). Graphs (a) and (b) relate to the scattering off an electron, graph (a) via the $t$-channel exchange of the weak neutral current, graph (b) via the $u$-channel exchange of the weak charged current. Graphs (c) and (d) relate to the scattering off a positron, graph (c) via the $t$-channel exchange of the weak neutral current, graph (d) via the $s$-channel creation of a $W^+$ boson. In the case of the muon and $\tau$-lepton neutrini as projectiles, only graphs (a) and (c) contribute: graph (b) cannot contribute in the energy range explored in this paper, whereas graph (d) does not contribute on the basis of first principles (conservation of the leptonic number within each generation of matter).](NuElPos.eps){width="15cm"}
Ignoring the (presently unknown) (anti)neutrini masses, the interactions (a) of the (anti)neutrini of the electron generation of matter and (b) of the (anti)neutrini of the muon and of the $\tau$-lepton generations of matter are different: three of the graphs in Figs. \[fig:NuNu\]-\[fig:NuElPos\] do not contribute to the collision rates of the muon and $\tau$-lepton (anti)neutrini. As the interactions of the muon and $\tau$-lepton (anti)neutrini are identical in the plasma, the study of the interactions induced only by $\nu_e$ and $\nu_\mu$ suffices for the purposes of this work.
In all cases, only Dirac neutrini are considered herein: therefore, each neutrino and its corresponding antineutrino are assumed to be distinguishable particles. The range of the available energy, corresponding to the temperatures $T_1$ and $T_2$, is such that no massive state may be produced ‘on shell’, save for electrons and positrons. In addition, $-q^2$ does not exceed $300$ MeV$^2$ in this temperature range, a value which is at least seven orders of magnitude smaller than the square of the masses of the two IVBs. This property results in the simplification of the expressions for the various weak-interaction scattering amplitudes, and the subsequent dependence of the observables on only two physical constants, namely $G_F$ and $\xi \coloneqq \sin^2 \theta_W$, known as Fermi coupling constant and (square of the sine of the) weak-mixing angle, respectively. According to the most recent compilation of the physical constants by the Particle-Data Group [@pdg], $G_F=1.1663787(6) \cdot 10^{-5}$ GeV$^{-2}$ and $\xi=0.23129(5)$. In this work, the differential cross sections (DCSs) and the total cross sections (TCSs) will be expressed as multiples of the representative weak-interaction cross section $\sigma_0 \coloneqq G^2_F s/\pi$. The $\sigma_0$ value, corresponding to $T=10^{11}$ K, is about $5 \cdot 10^{-42}$ cm$^2$ or $5 \cdot 10^{-18}$ b.
\[sec:Preliminary\]The weak-interaction scattering amplitude
============================================================
The literature on the weak interaction is vast. For the sake of example, a thorough introduction to the subject may be obtained from Refs. [@ah; @gk]. I generally follow the formalism of Ref. [@ah] herein. In this section, I only summarise those elements which are of interest in the narrow scope of this paper.
Inspection of Figs. \[fig:NuNu\]-\[fig:NuElPos\] reveals that, in order to derive the various weak-interaction scattering amplitudes, one needs to obtain the leptonic currents applicable in the cases of the NC (vertices involving $Z^0$) and of the CC (vertices involving $W^\pm$). The latter are purely $V-A$ in character (‘purely’ implies equal - and opposite in sign - contributions of the vector and axial-vector components, ensuring the left-handedness of the neutrino and the right-handedness of the antineutrino), given for the $l \to \nu_l$ transition by the expression $$\bra{u(\nu_l)} J_\mu \ket{u(l)} = \frac{g}{\sqrt{2}} \bar{u}(\nu_l) \gamma_\mu \frac{1-\gamma_5}{2} u(l) \, \, \, ,$$ where the constant $g$ determines the overall strength of the CC processes. Of course, $u(l)$ is the spinor associated with the lepton $l$ (where $l$ denotes an electron, a muon, or a $\tau$-lepton) and $u(\nu_l)$ is the spinor associated with the corresponding neutrino; in this work, only electrons and electron neutrini are of relevance. The dependence of the spinors on the spin and on the $4$-momentum of the fermions is assumed (but not explicitly given). The normalisation condition of the spinors, associated with positive-energy solutions, is $u^\dagger(l) u(l)=2 E$, $E$ being the particle’s total energy (where $l$ now stands for the neutrini as well). Assuming this normalisation condition, one can prove that $\bar{u}(l) u(l)=2 m$, $m$ being the particle’s rest mass and $\bar{u}(l) \coloneqq u^\dagger(l) \gamma_0$.
On the other hand, the leptonic NC follows the generic form $$\bra{u(l)} J_\mu \ket{u(l)} = \frac{g_N}{2} \bar{u}(l) \gamma_\mu (g_V - g_A \gamma_5) u(l) \, \, \, ,$$ where the constant $g_N$ determines the overall strength of the NC processes, and $g_V$ and $g_A$ are known as vector and axial-vector couplings, respectively. These couplings are different for the neutrini and for the other leptons: the neutrino NC remains purely $V-A$ in character, i.e., $$g_V^\nu = 1/2 \, \, \, , g_A^\nu = 1/2 \, \, \, ,$$ whereas for all other leptons $$g_V^l = -1/2 + 2\xi \, \, \, , g_A^l = -1/2 \, \, \, .$$
According to the Standard Model of the Electroweak Interactions, developed in the 1960s by Sheldon Lee Glashow, Abdus Salam, and Steven Weinberg, the couplings $g$ and $g_N$ are related via the expression: $g=g_N \xi$. A similar relation holds for the masses of the IVBs: $M_W = M_Z \xi$. As a result, $g/M_W=g_N/M_Z$. The Fermi coupling constant, the sole regulator of the strength of the weak-interaction processes in Fermi’s current-current framework (pointlike approximation), is related to the aforementioned quantities via the expression $$\frac{G_F}{\sqrt{2}} = \frac{g^2}{8 M_W^2} = \frac{g_N^2}{8 M_Z^2} \, \, \, .$$
The last ingredient, which is necessary in order to advance to the derivation of the various weak-interaction scattering amplitudes, is the propagator associated with the virtual state; it is of the form $$\Pi^{\mu \nu} (q) = i \frac{-g^{\mu \nu} + q^\mu q^\nu / M^2}{q^2 - M^2} \, \, \, ,$$ where $M$ stands for the mass of the exchanged IVB.
Let me now combine these elements and obtain the weak-interaction scattering amplitude of one simple process, e.g., of the one involving the scattering of two neutrini of different flavour: $\nu_a \nu_b \to \nu_a \nu_b$. Such a process, featuring an electron neutrino as projectile, is schematically shown in Fig. \[fig:NuNu\](c). The current-propagator-current form of the scattering amplitude reads as $$\begin{aligned}
\mathscr{T} & \coloneqq - i \bra{u(\nu_a)} J_\mu \ket{u(\nu_a)} \Pi^{\mu \nu} \bra{u(\nu_b)} J_\nu \ket{u(\nu_b)}\\
&= -i \frac{g_N}{2} \bar{u}(\nu_a) \gamma_\mu \frac{1 - \gamma_5}{2} u(\nu_a) i \frac{-g^{\mu \nu} + q^\mu q^\nu / M_Z^2}{q^2 - M_Z^2} \frac{g_N}{2} \bar{u}(\nu_b) \gamma_\nu \frac{1 - \gamma_5}{2} u(\nu_b)\\
&= \frac{g^2_N}{16} \bar{u}(\nu_a) \gamma_\mu (1 - \gamma_5) u(\nu_a) \frac{-g^{\mu \nu} + q^\mu q^\nu / M_Z^2}{q^2 - M_Z^2} \bar{u}(\nu_b) \gamma_\nu (1 - \gamma_5) u(\nu_b) \, \, \, ,\end{aligned}$$ which, for $q^2 \ll M_Z^2$, results in $$\begin{aligned}
\label{eq:EQ100}
\mathscr{T} &= \frac{g^2_N}{16 M_Z^2} \bar{u}(\nu_a) \gamma^\mu (1 - \gamma_5) u(\nu_a) \bar{u}(\nu_b) \gamma_\mu (1 - \gamma_5) u(\nu_b)\nonumber\\
&= \frac{G_F}{2\sqrt{2}} \bar{u}(\nu_a) \gamma^\mu (1 - \gamma_5) u(\nu_a) \bar{u}(\nu_b) \gamma_\mu (1 - \gamma_5) u(\nu_b) \, \, \, .\end{aligned}$$
The square of the scattering amplitude $\lvert \mathscr{T} \rvert^2$, summed over the final states and properly averaged (for unpolarised cross sections) over the spin orientations in the initial state, is the backbone of the evaluation of the DCS of a process, complemented by the flux of the incident beam and the permissible Lorentz invariant phase space.
For massless projectiles, the DCS is obtained via the expression $$\frac{d\sigma}{d\tau} = \frac{\lvert \mathscr{T} \rvert^2}{2(s-m_t^2)} \, \, \, ,$$ where $d\tau$ denotes the infinitesimal Lorentz invariant phase space and $m_t$ stands for the rest mass of the target, i.e., $0$ for neutrini (Sections \[sec:D1\]-\[sec:D5\] and \[sec:D8\]), and $m_e$ for electrons and positrons (Sections \[sec:D6\] and \[sec:D7\]). The infinitesimal Lorentz invariant phase space may be obtained via the relation $$d\tau = \frac{p_f d\Omega}{(4 \pi)^2 \sqrt{s}} \, \, \, ,$$ where $p_f$ denotes the modulus of the CM $3$-momentum in the final state. (Regarding the details relating to the Lorentz invariant phase space, see Ref. [@ah].) In Sections \[sec:D1\]-\[sec:D5\], $p_f=\sqrt{s}/2$ and $$\label{eq:EQ105}
\frac{d\sigma}{d\Omega} = \frac{\lvert \mathscr{T} \rvert^2}{(8 \pi)^2 s} \, \, \, .$$ In Sections \[sec:D6\] and \[sec:D7\], $p_f=(s-m_e^2)/(2\sqrt{s})$ and Eq. (\[eq:EQ105\]) also holds. In Section \[sec:D8\], $p_f=\sqrt{s/4-m_e^2}$ and $$\frac{d\sigma}{d\Omega} = \frac{\lvert \mathscr{T} \rvert^2}{(8 \pi)^2 s} \sqrt{1-\frac{4 m_e^2}{s}} \, \, \, .$$
I will finally list a number of helpful relations obeyed by the matrix operator Tr, which returns the trace of a square matrix, i.e., the sum of its diagonal elements. Put forth by Richard Feynman, the trace technique comprises a set of operations enabling the speedy derivation of the sum of the contributions to the scattering amplitude from all spin orientations pertaining to the scattering process in question. $$\begin{aligned}
{\rm Tr} [A+B] &= {\rm Tr} [A] + {\rm Tr} [B]\\
{\rm Tr} [AB] &= {\rm Tr} [BA]\\
{\rm Tr} [ABC] &= {\rm Tr} [CAB] = {\rm Tr} [BCA]\\
{\rm Tr} [g_{ab}] &= 4 g_{ab}\\
{\rm Tr} [\gamma_a \gamma_b] &= {\rm Tr} [2 g_{a b} - \gamma_b \gamma_a] = {\rm Tr} [2 g_{a b} - \gamma_a \gamma_b] \Rightarrow {\rm Tr} [\gamma_a \gamma_b] = 4 g_{a b}\\
{\rm Tr} [\gamma_a \gamma_b \gamma_c \gamma_d] &= 4 \left( g_{ab} g_{cd} - g_{ac} g_{bd} + g_{ad} g_{bc} \right)\\
{\rm Tr} [\gamma_5 \gamma_a \gamma_b \gamma_c \gamma_d] &= 4 i \epsilon_{abcd}\\\end{aligned}$$ $A$, $B$, and $C$ denote square matrices of the same dimension. The traces of products of odd numbers of $\gamma$ matrices, as well as those of odd numbers of $\gamma$ matrices with the $\gamma_5$ matrix, vanish.
\[sec:Derivation\]Derivation of the various weak-interaction differential cross sections
========================================================================================
In this section, the scattering amplitudes, as well as the DCSs and TCSs derived thereof, of the various scattering processes of the neutrini with the (anti)neutrini, as well as with electrons and positrons, will be extracted. Sections \[sec:D1\]-\[sec:D5\] deal with the channels in which only (anti)neutrini appear in the initial and final states. Sections \[sec:D6\] and \[sec:D7\] deal with the scattering of neutrini off electrons and positrons, respectively. Finally, Section \[sec:D8\] pertains to the neutrino-antineutrino annihilation to an $e^+ e^-$ pair.
\[sec:D1\]Scattering of neutrini of different flavour
-----------------------------------------------------
The scattering amplitude $\mathscr{T}$ for the interaction of neutrini of different flavour was obtained in Section \[sec:Preliminary\], see Eq. (\[eq:EQ100\]). For an electron neutrino as projectile, the graph of Fig. \[fig:NuNu\](c) is relevant. As the neutrini are solely left-handed, the averaging over the spin orientations in the initial state is not applicable (the average of one value is the value itself). The square of the scattering amplitude $\lvert \mathscr{T} \rvert^2$ is obtained after multiplying $\mathscr{T}$ with its complex conjugate. However, given that $\mathscr{T}$ is one (complex) element (i.e., not a matrix), this operation is equivalent to using the Hermitian of $\mathscr{T}$, namely $\mathscr{T}^\dagger$, in the product. $$\lvert \mathscr{T} \rvert^2 \coloneqq \mathscr{T} \mathscr{T}^\dagger = \frac{G_F^2}{8} \, {\rm Tr} [{\ensuremath{\mathrlap{\!\not{\phantom{k}}}k}}^\prime \gamma^\mu (1-\gamma_5) {\ensuremath{\mathrlap{\!\not{\phantom{k}}}k}} \gamma^\nu (1-\gamma_5)] \, {\rm Tr} [{\ensuremath{\mathrlap{\!\not{\phantom{p}}}p}}^\prime \gamma_\mu (1-\gamma_5) {\ensuremath{\mathrlap{\!\not{\phantom{p}}}p}} \gamma_\nu (1-\gamma_5)]$$ The application of the first trace operator yields the projectile-related tensor $$\label{eq:EQ110}
K^{\mu \nu} = 8 \left( k^{\prime \, \mu} k^\nu + k^{\prime \, \nu} k^\mu + \frac{q^2}{2} g^{\mu\nu} - i \epsilon^{\mu \nu a b} k_a k^\prime_b \right) \, \, \, .$$ The target-related tensor, resulting from the application of the second trace operator, is of similar structure $$P_{\mu \nu} = 8 \left( p^\prime_\mu p_\nu + p^\prime_\nu p_\mu + \frac{q^2}{2} g_{\mu\nu} - i \epsilon_{\mu \nu c d} p^c p^{\prime \, d} \right) \, \, \, .$$
It can be easily shown that $q_\mu K^{\mu \nu} = q_\nu K^{\mu \nu} = 0$. This property enables the simplification of the target-related tensor, via the introduction of its effective form $$(P_{\mu \nu})_{\rm eff} = 8 \left( 2 p_\mu p_\nu + \frac{q^2}{2} g_{\mu\nu} - i \epsilon_{\mu \nu c d} p^c q^d \right) \, \, \, .$$
The contraction of the tensors $K^{\mu \nu}$ and $(P_{\mu \nu})_{\rm eff}$ finally yields the value of $64 s^2$. Necessary in the extraction of this result are the relations: $s \coloneqq (k+p)^2 = 2 k \cdot p$, $u \coloneqq (k-p^\prime)^2 = -2 k \cdot p^\prime$, and $t \coloneqq (k-k^\prime)^2 = -2k \cdot k^\prime$, which are the expressions of the Mandelstam variables for massless initial and final states. One additional relation is needed, namely $$\epsilon_{\mu \nu c d} \epsilon^{\mu \nu a b} = - 2 \left( \delta^a_c \delta^b_d - \delta^b_c \delta^a_d \right) \, \, \, ,$$ where $\delta$ denotes the standard Kronecker delta, equal to $1$ for identical indices and $0$ for different ones. Evidently, $$\label{eq:EQ120}
\lvert \mathscr{T} \rvert^2 = 8 G_F^2 s^2 \, \, \, .$$
Using this result, one obtains the DCS $$\label{eq:EQ130}
\frac{d\sigma}{d\Omega} = \frac{G^2_F s}{8 \pi^2} = \frac{\sigma_0}{8 \pi}$$ and, integrating over the solid angle, the TCS $$\sigma_{\rm tot} = \frac{\sigma_0}{2} \, \, \, .$$
\[sec:D2\]Scattering of a neutrino off an antineutrino of different flavour
---------------------------------------------------------------------------
For an electron neutrino as projectile, the graph of Fig. \[fig:NuNuBar\](f) is relevant. Feynman’s interpretation of ‘the negative-energy particle solutions propagating backward in time’ as ‘the positive-energy antiparticle solutions propagating forward in time’ provides a speedy solution to the scattering of this section from the result of the previous one: the current for the incident/outgoing antineutrino of Fig. \[fig:NuNuBar\](f) moving forward in time is identical to the one for an outgoing/incident neutrino moving backward in time. This implies that the result, obtained for the graph Fig. \[fig:NuNu\](c) (i.e., the result of Section \[sec:D1\]), may be used, along with the obvious transformations $p \to - p^\prime$ and $p^\prime \to -p$, and enable the nearly effortless extraction of the DCS of this section. Evidently, one simply needs to interchange the Mandelstam variables $s$ and $u$ in the $\lvert \mathscr{T} \rvert^2$ result of Section \[sec:D1\]. According to Eq. (\[eq:EQ120\]), only the Mandelstam variable $s$ enters $\lvert \mathscr{T} \rvert^2$, hence the result for the process of this section is $$\lvert \mathscr{T} \rvert^2 = 8 G_F^2 u^2 \, \, \, ,$$ leading to the DCS in the form $$\frac{d\sigma}{d\Omega} = \frac{G^2_F u^2}{8 \pi^2 s} = \frac{\sigma_0 u^2}{8 \pi s^2} \, \, \, ,$$ which, integrated over the solid angle, results in $$\sigma_{\rm tot} = \frac{\sigma_0}{6} \, \, \, .$$
\[sec:D3\]Scattering of neutrini of the same flavour
----------------------------------------------------
The peculiarity of this case is that the particles in the final state are indistinguishable. For the scattering of electron neutrini, the graphs of Fig. \[fig:NuNu\](a) and (b) are relevant.
Starting from Eq. (\[eq:EQ100\]), one may write the scattering amplitude as $$\begin{aligned}
\label{eq:EQ140}
\mathscr{T} = \frac{G_F}{2\sqrt{2}} &\big( \bar{u}(\nu_a) \gamma^\mu (1 - \gamma_5) u(\nu_a) \bar{u}(\nu_b) \gamma_\mu (1 - \gamma_5) u(\nu_b)\nonumber\\
&+ \bar{u}(\nu_b) \gamma^\mu (1 - \gamma_5) u(\nu_a) \bar{u}(\nu_a) \gamma_\mu (1 - \gamma_5) u(\nu_b) \big) \, \, \, .\end{aligned}$$ Let me rewrite the second term within the brackets in the form $$( \bar{\chi} \gamma^\mu (1 - \gamma_5) \psi ) \, ( \bar{\psi} \gamma_\mu (1 - \gamma_5) \chi ) \, \, \, ,$$ where $\chi=u(\nu_b)$ (hence $\bar{\chi}=\bar{u}(\nu_b)$) and $\psi=u(\nu_a)$ (hence $\bar{\psi}=\bar{u}(\nu_a)$). By invoking two of the Fierz identities, namely $$\begin{aligned}
\label{eq:EQ150}
(\bar{\chi}\gamma^\mu \psi)(\bar{\psi} \gamma_\mu \chi)&=(\bar{\chi}\chi)(\bar{\psi}\psi)-\frac{1}{2}(\bar{\chi}\gamma^\mu \chi)(\bar{\psi}\gamma_\mu \psi)\nonumber\\
&-\frac{1}{2}(\bar{\chi}\gamma^\mu \gamma_5 \chi)(\bar{\psi}\gamma_\mu \gamma_5 \psi)-(\bar{\chi}\gamma_5 \chi)(\bar{\psi}\gamma_5 \psi) \, \, \, ,\nonumber\\
(\bar{\chi}\gamma^\mu \gamma_5 \psi)(\bar{\psi} \gamma_\mu \gamma_5 \chi)&=-(\bar{\chi}\chi)(\bar{\psi}\psi)-\frac{1}{2}(\bar{\chi}\gamma^\mu \chi)(\bar{\psi}\gamma_\mu \psi)\nonumber\\
&-\frac{1}{2}(\bar{\chi}\gamma^\mu \gamma_5 \chi)(\bar{\psi}\gamma_\mu \gamma_5 \psi)+(\bar{\chi}\gamma_5 \chi)(\bar{\psi}\gamma_5 \psi) \, \, \, ,\end{aligned}$$ one obtains the result $$(\bar{\chi}\gamma^\mu (1-\gamma_5) \psi)(\bar{\psi} \gamma_\mu (1-\gamma_5) \chi) = -(\bar{\chi}\gamma^\mu (1-\gamma_5) \chi)(\bar{\psi}\gamma_\mu (1-\gamma_5) \psi)$$ and, after reverting to the original spinors, $$(\bar{u}(\nu_b)\gamma^\mu (1-\gamma_5) u(\nu_a))(\bar{u}(\nu_a) \gamma_\mu (1-\gamma_5) u(\nu_b)) = -(\bar{u}(\nu_b)\gamma^\mu (1-\gamma_5) u(\nu_b))(\bar{u}(\nu_a)\gamma_\mu (1-\gamma_5) u(\nu_a)) \, \, \, .$$ Of course, the wavefunction of fermions is antisymmetric under particle exchange, hence $$-(\bar{u}(\nu_b)\gamma^\mu (1-\gamma_5) u(\nu_b))(\bar{u}(\nu_a)\gamma_\mu (1-\gamma_5) u(\nu_a)) = + (\bar{u}(\nu_a)\gamma^\mu (1-\gamma_5) u(\nu_a)) (\bar{u}(\nu_b)\gamma_\mu (1-\gamma_5) u(\nu_b)) \, \, \, .$$
Inserting this relation into Eq. (\[eq:EQ140\]), one obtains $$\begin{aligned}
\mathscr{T} = \frac{G_F}{2\sqrt{2}} &\big( \bar{u}(\nu_a) \gamma^\mu (1 - \gamma_5) u(\nu_a) \bar{u}(\nu_b) \gamma_\mu (1 - \gamma_5) u(\nu_b)\\
&+ \bar{u}(\nu_a)\gamma^\mu (1-\gamma_5) u(\nu_a) \bar{u}(\nu_b)\gamma_\mu (1-\gamma_5) u(\nu_b) \big)\\
= \frac{G_F}{\sqrt{2}} &\bar{u}(\nu_a) \gamma^\mu (1 - \gamma_5) u(\nu_a) \bar{u}(\nu_b) \gamma_\mu (1 - \gamma_5) u(\nu_b) \, \, \, .\end{aligned}$$ Therefore, the amplitude for the scattering of neutrini of the same flavour is twice the result of Eq. (\[eq:EQ100\]), i.e., twice the scattering amplitude for the scattering of neutrini of different flavour. Consequently, there is no need to repeat the calculation of Section \[sec:D1\] in order to obtain the result of this section; all one needs to do is multiply the DCS of Eq. (\[eq:EQ130\]) by $4$. Finally, $$\label{eq:EQ160}
\frac{d\sigma}{d\Omega} = \frac{4 G^2_F s}{8 \pi^2} = \frac{\sigma_0}{2 \pi} \, \, \, .$$
Attention is needed when deriving the TCS for the scattering of neutrini of the same flavour from the result of Eq. (\[eq:EQ160\]). Owing to the fact that the final state comprises indistinguishable particles, the integration of the DCS between the $\theta$ limits of $0$ and $\pi$ yields an erroneous result! The outgoing particles are indistinguishable and it is fallacious to attach labels to them [^2], e.g., by identifying particle (1) as the one scattered at angle $\theta$ and particle (2) as the one scattered at $\pi-\theta$. The indistinguishability of the particles in the final state requires that the integration be performed over the solid angle of $2 \pi$, i.e., over a hemisphere, thus yielding the TCS result $$\sigma_{\rm tot} = \sigma_0 \, \, \, .$$
To be able to compare the DCS of this section with those obtained for the other neutrino-induced processes, one may proceed in either of two ways:
- by making use of the DCS of Eq. (\[eq:EQ160\]), also bearing in mind that the corresponding TCS must involve an integration over a hemisphere,
- by halving the DCS of Eq. (\[eq:EQ160\]) and integrating over $4 \pi$.
I will follow the latter option. Of course, all expressions, obtained from the $\lvert \mathscr{T} \rvert^2$ result of this section (e.g., on the basis of interchanges of $4$-momenta) for other processes, a) must involve the original $\lvert \mathscr{T} \rvert^2$ result and b) must correspond to TCSs involving the integration of the corresponding DCSs over $4 \pi$. The restricted solid-angle domain pertains exclusively to processes yielding indistinguishable particles in the final state, which (for Dirac neutrini) is the case only in this section.
\[sec:D4\]Elastic scattering of a neutrino off an antineutrino of the same flavour
----------------------------------------------------------------------------------
For the elastic scattering of an electron neutrino off an electron antineutrino, the graphs of Fig. \[fig:NuNuBar\](a) and (c) are relevant; the latter graph represents the annihilation to a neutrino-antineutrino pair of the same flavour. The substitutions $p \to - p^\prime$ and $p^\prime \to -p$ enable the speedy extraction of the DCS of this section from the $\lvert \mathscr{T} \rvert^2$ result of Section \[sec:D3\]. Evidently, $$\label{eq:EQ180}
\frac{d\sigma}{d\Omega} = \frac{G^2_F u^2}{2 \pi^2 s} = \frac{\sigma_0 u^2}{2 \pi s^2} \, \, \, ,$$ which, integrated over the solid angle, results in $$\sigma_{\rm tot} = \frac{2 \sigma_0}{3} \, \, \, .$$
\[sec:D5\]Neutrino-antineutrino annihilation to a neutrino-antineutrino pair of different flavour
-------------------------------------------------------------------------------------------------
For an electron neutrino-antineutrino pair, the graph of Fig. \[fig:NuNuBar\](d) is relevant. Straightforward considerations, again invoking Feynman’s interpretation of the negative-energy particle solutions, lead to a result identical to the one obtained in Section \[sec:D2\]. The DCS is of the form $$\frac{d\sigma}{d\Omega} = \frac{\sigma_0 u^2}{8 \pi s^2} \, \, \, ,$$ which, integrated over the solid angle, results in $$\sigma_{\rm tot} = \frac{\sigma_0}{6} \, \, \, .$$
\[sec:D6\]Scattering of neutrini off an electron
------------------------------------------------
For the scattering of an electron neutrino off an electron, the graphs of Fig. \[fig:NuElPos\](a) and (b) are relevant. The scattering amplitude is of the form $$\begin{aligned}
\label{eq:EQ200}
\mathscr{T} &= \left( \frac{g_N}{2 M_Z} \right)^2 \bar{u}(\nu_e) \gamma^\mu \frac{1 - \gamma_5}{2} u(\nu_e) \bar{u}(e^-) \gamma_\mu (g_V^l - g_A^l \gamma_5) u(e^-)\nonumber\\
&+ \left( \frac{g}{\sqrt{2} M_W} \right)^2 \bar{u}(\nu_e) \gamma^\mu \frac{1 - \gamma_5}{2} u(e^-) \bar{u}(e^-) \gamma_\mu \frac{1 - \gamma_5}{2} u(\nu_e)\nonumber\\
&= \frac{G_F}{\sqrt{2}} \big( \bar{u}(\nu_e) \gamma^\mu (1 - \gamma_5) u(\nu_e) \bar{u}(e^-) \gamma_\mu (g_V^l - g_A^l \gamma_5) u(e^-)\nonumber\\
&+ \bar{u}(e^-) \gamma^\mu (1 - \gamma_5) u(\nu_e) \bar{u}(\nu_e) \gamma_\mu (1 - \gamma_5) u(e^-) \big) \, \, \, .\end{aligned}$$ After employing the two Fierz identities of Eqs. (\[eq:EQ150\]), the second term within the brackets may be rewritten as $\bar{u}(\nu_e) \gamma^\mu (1 - \gamma_5) u(\nu_e) \bar{u}(e^-) \gamma_\mu (1 - \gamma_5) u(e^-)$, resulting in $$\begin{aligned}
\label{eq:EQ220}
\mathscr{T} &= \frac{G_F}{\sqrt{2}} \bar{u}(\nu_e) \gamma^\mu (1 - \gamma_5) u(\nu_e) \bar{u}(e^-) \gamma_\mu \left( g_V^l + 1 - (g_A^l +1) \gamma_5 \right) u(e^-)\nonumber\\
&= \frac{G_F}{\sqrt{2}} \bar{u}(\nu_e) \gamma^\mu (1 - \gamma_5) u(\nu_e) \bar{u}(e^-) \gamma_\mu (C_V - C_A \gamma_5) u(e^-) \, \, \, .\end{aligned}$$ where $C_V \coloneqq g_V^l + 1$ and $C_A \coloneqq g_A^l +1$ for the scattering of an electron neutrino off an electron. Equation (\[eq:EQ220\]) provides an explanation for the differences in the interactions between electron and muon/$\tau$-lepton neutrini with the electron. As, below the temperature $T_1$, the CM energy is not sufficient for the creation of muons and $\tau$-leptons in the final state, the CC graph contributes to the scattering amplitude only in case of an incident electron neutrino. To be able to apply the results of the calculation also in the case of incident muon and $\tau$-lepton neutrini, I will retain the constants $C_V$ and $C_A$ (applicable for the scattering of an electron neutrino off an electron) and bear in mind that, in the case of an incident $\nu_{\mu,\tau}$, they must be replaced by $\widetilde{C}_V = C_V-1 = g_V^l$ and $\widetilde{C}_A = C_A-1 = g_A^l$. Evidently, $$\begin{aligned}
\lvert \mathscr{T} \rvert^2 = &\frac{G_F^2}{2} \, {\rm Tr} [{\ensuremath{\mathrlap{\!\not{\phantom{k}}}k}}^\prime \gamma^\mu (1-\gamma_5) {\ensuremath{\mathrlap{\!\not{\phantom{k}}}k}} \gamma^\nu (1-\gamma_5)]\\
&\frac{1}{2} \, {\rm Tr} [({\ensuremath{\mathrlap{\!\not{\phantom{p}}}p}}^\prime+m_e) \gamma_\mu (C_V-C_A \gamma_5) ({\ensuremath{\mathrlap{\!\not{\phantom{p}}}p}}+m_e) \gamma_\nu (C_V-C_A\gamma_5)] \, \, \,\end{aligned}$$ where the factor $1/2$ in front of the second trace takes account of the averaging of the spin orientations in the initial state (target electron). The application of the first trace operator yields the projectile-related tensor of Eq. (\[eq:EQ110\]). The application of the second trace yields the target-related tensor, which (owing to the fact that the electron NC is not purely $V-A$ in character) is now of a more complex form. $$P_{\mu \nu} = 4 \left( c_1 \left( p^\prime_\mu p_\nu + p^\prime_\nu p_\mu + \left( \frac{q^2}{2} - m_e^2 \right) g_{\mu\nu} \right) - i c_2 \epsilon_{\mu \nu c d} p^c p^{\prime \, d} + c_3 m_e^2 g_{\mu\nu} \right) \, \, \, ,$$ where $c_1=C_V^2+C_A^2$, $c_2=2 C_V C_A$, and $c_3=C_V^2-C_A^2$.
As in Section \[sec:D1\], an effective target-related tensor may be constructed on the basis of the properties: $q_\mu K^{\mu \nu} = q_\nu K^{\mu \nu} = 0$. $$(P_{\mu \nu})_{\rm eff} = 4 \left( c_1 \left( 2 p_\mu p_\nu + \left( \frac{q^2}{2} - m_e^2 \right) g_{\mu\nu} \right) - i c_2 \epsilon_{\mu \nu c d} p^c q^d + c_3 m_e^2 g_{\mu\nu} \right) \, \, \, .$$ The contraction of $K^{\mu\nu}$ of Eq. (\[eq:EQ110\]) and $(P_{\mu \nu})_{\rm eff}$ of the previous equation results in $$\label{eq:EQ240}
\lvert \mathscr{T} \rvert^2 = 4 G_F^2 \left( C_1 (s-m^2_e)^2 + C_2 (u - m^2_e)^2 + 2 C_3 m^2_e t \right) \, \, \, ,$$ where $C_1=c_1+c_2=(C_V+C_A)^2$, $C_2=c_1-c_2=(C_V-C_A)^2$, and $C_3=c_3=(C^2_V-C^2_A)$. As aforementioned, this result is valid for an electron neutrino as projectile. The same formula holds for muon and $\tau$-lepton neutrini as projectiles, but the constants must be redefined, following the replacements $C_V \to \widetilde{C}_V$ and $C_A \to \widetilde{C}_A$. (The constant $C_2$ is not affected by these substitutions.)
The DCS for the scattering of neutrini off an electron is of the form $$\label{eq:EQ225}
\frac{d\sigma}{d\Omega} = \frac{\sigma_0}{16 \pi s^2} \left( C_1 (s-m^2_e)^2 + C_2 (u - m^2_e)^2 + 2 C_3 m^2_e t \right) \, \, \, ,$$ where (as already explained) the constants $C_1$ and $C_3$ depend on the neutrino flavour. After integrating the DCS over the solid angle, one obtains $$\label{eq:EQ230}
\sigma_{\rm tot}=\frac{\sigma_0}{4} \left( 1 - \frac{m^2_e}{s} \right)^2 \left( C_1 + C_2 \frac{s^2 + m^4_e + s m^2_e}{3 s^2} - C_3 \frac{m^2_e}{s} \right) \, \, \, ,$$ which is applicable for $s \geq m_e^2$. These expressions are in agreement with those derived in the pioneering paper of Herrera and Hacyan [@herr].
\[sec:D7\]Scattering of neutrini off a positron
-----------------------------------------------
For the scattering of an electron neutrino off a positron, the graphs of Fig. \[fig:NuElPos\](c) and (d) are relevant. Following the line of argumentation of the previous section, the graph of Fig. \[fig:NuElPos\](d) does not contribute to the scattering amplitude in case of incident muon and $\tau$-lepton neutrini. The scattering amplitude may be obtained from Eq. (\[eq:EQ240\]) after the interchange $s \leftrightarrow u$, which is equivalent to the interchange $C_1 \leftrightarrow C_2$ in Eqs. (\[eq:EQ225\]) and (\[eq:EQ230\]). Therefore, the DCS for the scattering of neutrini off a positron reads as $$\frac{d\sigma}{d\Omega} = \frac{\sigma_0}{16 \pi s^2} \left( C_1 (u - m^2_e)^2 + C_2 (s-m^2_e)^2 + 2 C_3 m^2_e t \right) \, \, \, ,$$ where the constants $C_1$ and $C_3$ depend on the neutrino flavour as explained in the previous section. After integrating the DCS over the solid angle, one obtains $$\sigma_{\rm tot}=\frac{\sigma_0}{4} \left( 1 - \frac{m^2_e}{s} \right)^2 \left( C_1 \frac{s^2 + m^4_e + s m^2_e}{3 s^2} + C_2 - C_3 \frac{m^2_e}{s} \right) \, \, \, ,$$ which is applicable for $s \geq m_e^2$. As expected, these expressions are in agreement with those given in Ref. [@herr].
\[sec:D8\]Neutrino-antineutrino annihilation to an $e^+ e^-$ pair
-----------------------------------------------------------------
For the annihilation of an electron neutrino-antineutrino pair to an $e^+ e^-$ pair, the graphs of Fig. \[fig:NuNuBar\](b) and (e) are relevant. The scattering amplitude, corresponding to the neutrino-antineutrino annihilation to an $e^+ e^-$ pair, may be obtained from the $\lvert \mathscr{T} \rvert^2$ result of Section \[sec:D6\] after the replacements $k^\prime \to -p$, $p \to -p^\prime$, and $p^\prime \to k^\prime$. These replacements are equivalent to the substitutions $s \to u$, $u \to t$, and $t \to s$. One additional issue requires attention. Equation (\[eq:EQ240\]) for $\lvert \mathscr{T} \rvert^2$ was obtained after averaging over the spin orientations of the target electron; no averaging is to be performed in the case of the neutrino-antineutrino annihilation, hence the amplitude, obtained after the aforementioned substitutions are performed, must be multiplied by $2$. The DCS thus becomes $$\frac{d\sigma}{d\Omega} = \frac{\sigma_0}{8 \pi s^2} \sqrt{1-\frac{4 m^2_e}{s}} \left( C_1 (u - m^2_e)^2 + C_2 (t - m^2_e)^2 + 2 C_3 m^2_e s \right) \, \, \, ,$$ where the constants have been defined in Section \[sec:D6\]. The integration of the DCS over the solid angle yields $$\sigma_{\rm tot}=\sigma_0 \sqrt{1-\frac{4 m^2_e}{s}} \left( \frac{C_1+C_2}{6} \left( 1 - \frac{m^2_e}{s} \right) + C_3 \frac{m^2_e}{s} \right) \, \, \, ,$$ which is applicable for $s \geq 4 m_e^2$. As Kuznetsov and Savin [@ks] noticed (and proposed this property as a simple criterion for judging the correctness of relevant calculations [^3]), the TCS for the neutrino-antineutrino annihilation to an $e^+ e^-$ pair does not depend on the constant $C_A$ in the vicinity of $s=4m_e^2$; it comes out equal to $C_V^2 \sqrt{\epsilon} \sigma_0 / 2$, where $s = 4 m_e^2 (1 + \epsilon)$ and $\epsilon$ is positive and small (compared to $1$).
\[sec:Summary\]Summary of the results
=====================================
Figures \[fig:ElectronNeutrino\] and \[fig:MuonTauNeutrino\] provide plots of the angular distributions of the cross section, detailed in Sections \[sec:D1\]-\[sec:D8\], separately for electron and muon/$\tau$-lepton neutrini as projectiles. Visual inspection of these figures leaves no doubt that the interactions of the (anti)neutrini amongst themselves are more important than those involving electron and positron targets [^4].
![\[fig:ElectronNeutrino\]Plot of the angular distributions of the cross section, detailed in Sections \[sec:D1\]-\[sec:D8\], for an electron neutrino as projectile. The quantity $\sigma_0$ has been defined at the end of Section \[sec:FD\]. The differential cross section for antineutrino targets does not vanish at $\cos\theta=-1$ because it includes the annihilation channel to an $e^+ e^-$ pair. The square of the CM energy for the incident particles corresponds to a temperature of $3 \cdot 10^{10}$ K.](ElectronNeutrino.eps){width="15cm"}
![\[fig:MuonTauNeutrino\]Plot of the angular distributions of the cross section, detailed in Sections \[sec:D1\]-\[sec:D8\], for muon/$\tau$-lepton neutrini as projectiles. The quantity $\sigma_0$ has been defined at the end of Section \[sec:FD\]. The differential cross section for antineutrino targets does not vanish at $\cos\theta=-1$ because it includes the annihilation channel to an $e^+ e^-$ pair. The square of the CM energy for the incident particles corresponds to a temperature of $3 \cdot 10^{10}$ K.](MuonTauNeutrino.eps){width="15cm"}
All the TCSs, derived in Section \[sec:Derivation\], are listed in Table \[tab:TCSs\]. In the three last cases (corresponding to the results obtained in Sections \[sec:D6\]-\[sec:D8\]), expressions derived under the high-energy approximation ($s \gg m^2_e$) are quoted in the table.
****
Total cross sections (TCSs) for the interactions of the (anti)neutrini with the various ingredients of the plasma in the early Universe. The TCSs are expressed as multiples of the representative weak-interaction cross section $\sigma_0$, which has been defined at the end of Section \[sec:FD\]. In the three last cases, the formulae are high-energy approximations; the exact expressions are given in Section \[sec:Derivation\]. The TCS for the neutrino-antineutrino annihilation ($\nu_a \bar{\nu}_a$) contains all the final-state channels, save for the annihilation to an $e^+ e^-$ pair, which is shown separately in last row of the table.
Target $\downarrow$, Projectile $\to$ $\nu_e$ $\nu_\mu$ $\nu_\tau$
--------------------------------------- ------------------------------------------------------ ------------------------------------------------------ ------------------------------------------------------
$\nu_e$ $1$ $1/2$ $1/2$
$\nu_\mu$ $1/2$ $1$ $1/2$
$\nu_\tau$ $1/2$ $1/2$ $1$
$\bar{\nu}_e$ $1$ $1/6$ $1/6$
$\bar{\nu}_\mu$ $1/6$ $1$ $1/6$
$\bar{\nu}_\tau$ $1/6$ $1/6$ $1$
$e^-$ $\frac{1}{4}+\xi+\frac{4 \xi^2}{3}$ $\frac{1}{4}-\xi+\frac{4 \xi^2}{3}$ $\frac{1}{4}-\xi+\frac{4 \xi^2}{3}$
$e^+$ $\frac{1}{3} \left( \frac{1}{4}+\xi+4 \xi^2 \right)$ $\frac{1}{3} \left( \frac{1}{4}-\xi+4 \xi^2 \right)$ $\frac{1}{3} \left( \frac{1}{4}-\xi+4 \xi^2 \right)$
$e^+ e^-$ creation $\frac{1}{6} \left( 1+4 \xi+8 \xi^2 \right)$ $\frac{1}{6} \left( 1-4 \xi+8 \xi^2 \right)$ $\frac{1}{6} \left( 1-4 \xi+8 \xi^2 \right)$
: \[tab:TCSs\]
\[sec:Discrepancies\]Discrepancies in the literature
====================================================
Comparing the expressions of this paper with those appearing in a number of published works (peer-reviewed articles and books) reveals a number of discrepancies. I will next address in chronological order those of the discrepancies which I am aware of.
- Flowers and Sutherland [@fs] calculated the $\nu_e \nu_e$ DCS and TCS correctly, including the appropriate integration of the DCS for indistinguishable particles in the final state. However, their expression for the $\nu_e \bar{\nu}_e$ DCS, see their Eq. (8), is fourfold the value obtained from Eq. (\[eq:EQ180\]) of this work.
- Hannestad and Madsen [@hm] presented in tabular form the scattering amplitudes corresponding to the various neutrino-induced processes relevant to the decoupling. The entry for the scattering of neutrini of the same flavour in their tables yields a DCS which is half the result of this work. The same problem appeared in Tables 5.1 and 5.2 of Ref. [@hh], as well as in Table 1 of Ref. [@gg]. Dolgov also commented on these mismatches in Ref. [@dolg] (p. 356).
- Xing and Zhou [@xz] gave the expressions for the $\nu_a \nu_b$ and $\bar{\nu}_a \bar{\nu}_b$ TCSs for $a=b$ and $a \neq b$ (pp. 39-40). Those expressions are in agreement with the results of Table \[tab:TCSs\]. They then advanced to treat the TCS for the $\nu_a \bar{\nu}_a$ processes and came up with a value which is twice as large as the $\nu_a \nu_a$ TCS. Unfortunately, they obtained that result after employing the wrong DCS; evidently, they did not replace the Mandelstam variable $s$ with $u$ in $\lvert \mathscr{T} \rvert^2$ when extracting the $\nu_a \bar{\nu}_a$ scattering amplitude from the one they had obtained for the $\nu_a \nu_a$ process.
- Lesgourgues, Mangano, Miele, and Pastor [@lmmp] presented in tabular form the TCSs corresponding to the various neutrino-induced processes relevant to the decoupling (see their Tables 1.7 and 1.8). To the best of my knowledge, this is the only place in the literature where an observable (e.g., the TCS) is presented in tabular form for all the processes studied in this work. Nevertheless, their tables are particularly worrying. To start with, all the TCSs off (anti)neutrino targets are half the corresponding entries of Table \[tab:TCSs\] of this work. In addition, their TCSs for the neutrino-antineutrino annihilation to an $e^+ e^-$ pair are twice the corresponding entries of Table \[tab:TCSs\] of this work. (Interestingly, their TCSs for the scattering of neutrini off electrons or positrons agree with the results of this work!) These discrepancies are puzzling.
Additional discrepancies (relating to the DCSs and TCSs for the neutrino-antineutrino annihilation to an $e^+ e^-$ pair) were reported by Kuznetsov and Savin in Ref. [@ks]. As my overview in the domain of Neutrino Physics is limited, the chances are that the aforementioned list is anything but exhaustive. I would be indebted to the colleagues who could communicate further discrepancies to me; I will acknowledge such contributions in future versions of this paper.
\[sec:Conclusions\]Conclusions
==============================
The present study dealt with the neutrino-induced processes relevant to the physics of the early Universe, namely with the interactions of the neutrini and the antineutrini of the three generations of matter amongst themselves, as well as with the electrons and the positrons of the plasma. These processes are of interest in other domains too, namely in the physics of compact stars. All the differential cross sections of these processes were derived, hopefully in a didactical and elucidating manner, following the standard methodology of Ref. [@ah]. To facilitate the overview and the crosscheck of the results, the corresponding total cross sections are shown in tabular form (Table \[tab:TCSs\]). The angular distributions of the sums of the DCSs, detailed in Sections \[sec:D1\]-\[sec:D8\], are shown in Fig. \[fig:ElectronNeutrino\] and \[fig:MuonTauNeutrino\], separately for electron and muon/$\tau$-lepton neutrini as projectiles.
Discrepancies were reported with some scientific publications and books comprising part of the literature on this subject. By no means should the list of discrepancies, addressed in this work, be considered as exhaustive.
The present study deals only with Dirac neutrini; its extension to also cover Majorana neutrini may be worth pursuing.
I am grateful to Alexander V. Kuznetsov for having clarified a question regarding the results of Ref. [@herr] and to Joseph A. Formaggio for drawing my attention to Ref. [@gk]. It would be an unpardonable omission to fail acknowledging the exchange of constructive electronic mail with Carlo Giunti on a number of interesting subjects relating to this paper.
All Feynman graphs in this work were created with the software package JaxoDraw [@jd], available from http://jaxodraw.sourceforge.net/.
[99]{} D. A. Dicus, W. W. Repko, ‘Photon-neutrino interactions’, Phys. Rev. Lett. 79 (1997) 569–571. C. Patrignani (Particle Data Group), ‘The review of Particle Physics’, Chin. Phys. C 40 (2016) 100001. I. J. Aitchison, A. J. Hey, ‘Gauge Theories in Particle Physics’, Vol. 2, 3rd edn., IoP Publishing, 2004. C. Giunti, C. W. Kim, ‘Fundamentals of Neutrino Physics and Astrophysics’, Oxford University Press, 2007. M. A. Herrera, S. Hacyan, ‘Relaxation time of neutrinos in the early Universe’, Astrophys. J. 336 (1989) 539–543. A. V. Kuznetsov, V. N. Savin, ‘The cross-section of the neutrino-antineutrino pair conversion into the electron-positron pair and its use in astrophysical calculations: correction of mistakes’, Conference on ‘Physics of Fundamental Interactions’, Moscow, MEPhI, November 17-21, 2014; available online from http://www.icssnp.mephi.ru/content/file/2014/section8/8\_16\_Kuznetsov\_MEPhI\_14e.pdf. E. G. Flowers, P. G. Sutherland, ‘Neutrino-neutrino scattering and supernovae’, Astrophys. J. 208 (1976) L19–L21. S. Hannestad, J. Madsen, ‘Neutrino decoupling in the early Universe’, Phys. Rev. D 52 (1995) 1764–1769. S. Hannestad, ‘Aspects of Neutrino Physics in the Early Universe’, Ph.D. dissertation, University of Aarhus, 1997. N. Y. Gnedin, O. Y. Gnedin, ‘Cosmological neutrino background revisited’, Astrophys. J. 509 (1998) 11–15. A. D. Dolgov, ‘Neutrinos in Cosmology’, Phys. Rep. 370 (2002) 333–535. Zhizhong Xing, Shun Zhou, ‘Neutrinos in Particle Physics, Astronomy and Cosmology’, Springer, 2011. J. Lesgourgues, G. Mangano, G. Miele, S. Pastor, ‘Neutrino Cosmology’, Cambridge University Press, 2013. D. Binosi, L. Theußl, ‘JaxoDraw: A graphical user interface for drawing Feynman diagrams’, Comput. Phys. Commun. 161 (2004) 76–86.
[^1]: So long as their collision rates exceeded the expansion rate of the Universe (identified as the Hubble parameter, which is naturally dependent on the cosmological time), the (anti)neutrini remained in thermal equilibrium with the ingredients of the plasma. At the time when their collision rates dropped below the expansion rate of the Universe, the (anti)neutrini escaped the plasma. The detachment of the (anti)neutrini from their former engagement as an active part of the plasma is known as ‘decoupling’.
[^2]: I am indebted to Carlo Guinti for drawing my attention to this subtle point and for bringing forth the argument on the integration limits of $\theta$.
[^3]: Of course, the fulfilment of this condition may be a necessary, but is not a sufficient condition for the correctness of the calculations.
[^4]: Due to the spin degeneracy of the electrons and the positrons, the contributions of the graphs involving electron and positron targets to the mean-free path of the neutrini in simulations of the early Universe are twice as large as one obtains from Figs. \[fig:ElectronNeutrino\] and \[fig:MuonTauNeutrino\] (as well as from Table \[tab:TCSs\]); due to the two possible spin orientations in case of electrons and positrons, twice as many of these particles may be ‘packed together’ in a given volume, at a given temperature, as (anti)neutrini of each of the three flavours.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We demonstrate efficient second-harmonic generation at moderate input power for thin film Ni$_{81}$Fe$_{19}$ undergoing ferromagnetic resonance (FMR). Powers of the generated second-harmonic are shown to be quadratic in input power, with an upconversion ratio three orders of magnitude higher than that demonstrated in ferrites [@AyresJAP1956], defined as $\Delta P^{2\omega}/\Delta P^{\omega} \sim 4 \times 10^{-5} /W \cdot P^{\omega}$, where $\Delta P$ is the change in the transmitted rf power and $P$ is the input rf power. The second harmonic signal generated exhibits a significantly lower linewidth than that predicted by low-power Gilbert damping, and is excited without threshold. Results are in good agreement with an analytic, approximate expansion of the Landau-Lifshitz-Gilbert (LLG) equation.'
author:
- Cheng Cheng
- 'William E. Bailey'
title: 'High-efficiency GHz frequency doubling without power threshold in thin-film Ni$_{81}$Fe$_{19}$'
---
Nonlinear effects in magnetization dynamics, apart from being of fundamental interest[@WangPR1954; @AyresJAP1956; @SuhlJPCS1957; @BierleinPRB1970], have provided important tools for microwave signal processing, especially in terms of frequency doubling and mixing[@RidrigueJAP1969; @HarrisIEEE2012]. Extensive experimental work exists on ferrites[@AyresJAP1956; @BierleinPRB1970; @HarrisIEEE2012], traditionally used in low-loss devices due to their insulating nature and narrow ferromagnetic resonance (FMR) linewidth. Metallic thin-film ferromagnets are of interest for use in these and related devices due to their high moments, integrability with CMOS processes, and potential for enhanced functionality from spin transport; low FMR linewidth has been demonstrated recently in metals through compensation by the spin Hall effect[@DemidovNMat2012]. While some recent work has addressed nonlinear effects[@BerteaudJAP1966; @GerritsPRL2007; @OlsonJAP2007] and harmonic generation[@BaoAPL2008; @KhivintsevAPL2011; @MarshAPL2012] in metallic ferromagnets and related devices[@YanaJMMM2008; @DemidovAPL2011; @BiAPL2011], these studies have generally used very high power or rf fields, and have not distinguished between effects above and below the Suhl instability threshold. In this manuscript, we demonstrate frequency doubling below threshold in a metallic system (Ni$_{81}$Fe$_{19}$) which is three orders of magnitude more efficient than that demonstrated previously in ferrite materials[@AyresJAP1956]. The results are in good quantitative agreement with an analytical expansion of the Landau-Lifshitz-Gilbert (LLG) equation.\
For all measurements shown, we used a metallic ferromagnetic thin film structure, Ta(5 nm)/Cu(5 nm)/Ni$_{81}$Fe$_{19}$ (30 nm)/Cu(3 nm)/Al(3 nm). The film was deposited on an oxidized silicon substrate using magnetron sputtering at a base pressure of 2.0$\times$10$^{-7}$ Torr. The bottom Ta(5 nm)/Cu(5 nm) layer is a seed layer to improve adhesion and homogeneity of the film and the top Cu(3 nm)/Al(3 nm) layer protects the Ni$_{81}$Fe$_{19}$ layer from oxidation. A diagram of the measurement configuration, adapted from a basic broadband FMR setup, is shown in Fig.1. The microwave signal is conveyed to and from the sample through a coplanar waveguide (CPW) with a 400 $\mu$m wide center conductor and 50 $\Omega$ characteristic impedance, which gives an estimated rf field of 2.25 Oe rms with the input power of +30 dBm. We examined the second harmonic generation with fundamental frequencies at 6.1 GHz and 2.0 GHz. The cw signal from the rf source is first amplified by a solid state amplifier, then the signal power is tuned to the desirable level by an adjustable attenuator. Harmonics of the designated input frequency are attenuated by the bandpass filter to less than the noise floor of the spectrum analyzer (SA). The isolator limits back-reflection of the filtered signal from the sample into the rf source. From our analysis detailed in a later section of this manuscript, we found the second harmonic magnitude to be proportional to the product of the longitudinal and transverse rf field strengths, and thus place the center conductor of CPW at 45$^{\circ}$ from H$_B$ to maximize the $H^{rf}_y H^{rf}_z$ product. The rf signal finally reaches the SA for measurements of the power of both the fundamental frequency and its second harmonic.\
![\[Fig.1\] Experimental setup and the coordinate system, $\theta = 45^{\circ}$; see text for details. EM: electromagnet; SA: spectrum analyzer. Arrows indicate the transmission of rf signal.](fig1.eps)
![\[Fig.2\] Second harmonic generation with $\omega/2\pi = $ 6.1 GHz. a) *left panel*: 6.1 GHz input power +17.3 dBm; *right panel*: 6.1 GHz input power +8.35 dBm. b) amplitudes of the $\omega$ (FMR) and generated 2$\omega$ peaks as a function of input power $P^{\omega}$; right and top axes represent the data set in log-log plot (green), extracting the power index; c) ratio of the peak amplitudes of FMR and second harmonic generation as a function of the input 6.1 GHz power; green: log scale.](fig2.eps)
Fig.2(a) demonstrates representative field-swept FMR absorption and the second harmonic emission spectra measured by the SA as 6.1 GHz and 12.2 GHz peak intensities as a function of the bias field H$_B$. We vary the input rf power over a moderate range of +4 - +18 dBm, and fit the peaks with a Lorentzian function to extract the amplitude and the linewidth of the absorbed ($\Delta P^{\omega}$) and generated ($\Delta P^{2\omega}$) power. Noticeably, the second harmonic emission peaks have a much smaller linewidth, $\Delta H_{1/2}\sim$ 10 Oe over the whole power range, than those of the FMR peaks, with $\Delta H_{1/2}\sim$ 21 Oe. Plots of the absorption and emission peak amplitudes as a function of the input 6.1 GHz power, shown in Fig.2(b), clearly indicate a linear dependence of the FMR absorption and a quadratic dependence of the second harmonic generation on the input rf power. Taking the ratio of the radiated second harmonic power to the absorbed power, we have a conversion rate of 3.7$\times$10$^{-5}$/W, as shown in Fig.2(c).\
Since the phenomenon summarized in Fig.2 is clearly not a threshold effect, we look into the second-harmonic analysis of the LLG equation with small rf fields, which is readily described in Gurevich and Melkov’s text for circular precession relevant in the past for low-M$_s$ ferrites. For metallic thin films, we treat the elliptical case as follows. As illustrated in Fig.1, the thin film is magnetized in the *yz* plane along $\widehat{\textbf{z}}$ by the bias field H$_B$, with film-normal direction along $\widehat{\textbf{x}}$. The CPW exerts both a longitudinal rf field h$_z^{rf}$ and a transverse rf field h$_y^{rf}$ of equal strength. First consider only the transverse field h$_y^{rf}$. In this well established case\cite{}, the LLG equation $\dot{\textbf{m}} = -\gamma \textbf{m}\times \textbf{H}_{eff} + \alpha \textbf{m}\times \dot{\textbf{m}}$ is linearized and takes the form $$\begin{bmatrix}
\dot{\widetilde{m_x}} \\
\dot{\widetilde{m_y}}
\end{bmatrix}
=\begin{bmatrix}
-\alpha(\omega_H+\omega_M) & -\omega_H \\
\omega_H+\omega_M & -\alpha\omega_H
\end{bmatrix}
\begin{bmatrix}
\widetilde{m_x} \\
\widetilde{m_y}
\end{bmatrix}
+\begin{bmatrix}
\gamma \widetilde{h_y^{rf}} \\
0
\end{bmatrix}$$ , where $\gamma$ is the gyromagnetic ratio, $\alpha$ is the Gilbert damping parameter, $\omega_M \equiv \gamma 4\pi M_s$, and $\omega_H \equiv \gamma H_z$. Introducing first order perturbation to $m_{x,y}$ under additional longitudinal h$_z^{rf}$ and neglecting the second order terms, we have $$\begin{bmatrix}
\dot{\widetilde{m_x}} +\dot{\widetilde{\Delta m_x}}\\
\dot{\widetilde{m_y}} +\dot{\widetilde{\Delta m_y}}
\end{bmatrix}
=\begin{bmatrix}
-\alpha(\omega_H+\omega_M) & -\omega_H \\
\omega_H+\omega_M & -\alpha\omega_H
\end{bmatrix}
\begin{bmatrix}
\widetilde{m_x} + \widetilde{\Delta m_x} \\
\widetilde{m_y} + \widetilde{\Delta m_y}
\end{bmatrix}
+\begin{bmatrix}
\gamma \widetilde{h_z^{rf}} \widetilde{m_y}\\
-\gamma \widetilde{h_z^{rf}} \widetilde{m_x}
\end{bmatrix}
+\begin{bmatrix}
\gamma \widetilde{h_y^{rf}} \\
0
\end{bmatrix}$$ Subtracting (1) from (2) and taking $\widetilde{h^{rf}}_{y,z} = H_{y,z}^{rf}e^{-i\omega t}$, $\widetilde{m}_{x,y} = (H_y^{rf}/M_s)e^{-i\omega t}\widetilde{\chi}_{\perp,\parallel}(\omega)$, the equation for the perturbation terms is $$\begin{bmatrix}
\dot{\widetilde{\Delta m_x}}\\
\dot{\widetilde{\Delta m_y}}
\end{bmatrix}
=\begin{bmatrix}
-\alpha(\omega_H+\omega_M) & -\omega_H \\
\omega_H+\omega_M & -\alpha\omega_H
\end{bmatrix}
\begin{bmatrix}
\widetilde{\Delta m_x} \\
\widetilde{\Delta m_y}
\end{bmatrix}
+H_z^{rf}\frac{H_y^{rf}}{M_s}e^{-i2\omega t}\begin{bmatrix}
\gamma \widetilde{\chi_{\parallel}}(\omega)\\
-\gamma \widetilde{\chi_{\perp}}(\omega)
\end{bmatrix}$$ Since $\chi_{\perp}$ is one order of magnitude smaller than $\chi_{\parallel}$, we neglect the term $-\gamma \widetilde{\chi_{\perp}}(\omega)$. In complete analogy to equation (1), the driving term could be viewed as an effective transverse field of $H_z^{rf}(H_y^{rf}/M_s)\widetilde{\chi_{\parallel}}(\omega)e^{-i2\omega t}$, and the solutions to equation (3) would be $
\widetilde{\Delta m_x} = (H_z^{rf}H_y^{rf}/M_s^2)\widetilde{\chi_{\parallel}}(\omega)\widetilde{\chi_{\perp}}(2\omega)e^{-i2\omega t} $, $
\widetilde{\Delta m_y} = (H_z^{rf}H_y^{rf}/M_s^2)\widetilde{\chi_{\parallel}}(\omega)\widetilde{\chi_{\parallel}}(2\omega)e^{-i2\omega t}
$. We can compare the power at frequency $f$ and $2f$ now that we have the expressions for both the fundamental and second harmonic components of the precessing **M**. The time-averaged power per unit volume could be calculated as $
\langle P\rangle = [\int_{0}^{\frac{2\pi}{\omega}} P(t) dt]/(2\pi/\omega), \ \ P(t) = -\partial U/\partial t = 2\textbf{M}\partial \textbf{H}/\partial t
$ where only the transverse components of **M** and **H** contribute to P(t). Using the expression for $ \langle P\rangle$, **M** and **H**, we have $
P^{\omega} = \omega H_{y,rf}^2\chi(\omega)''_{\parallel}
$ and $
P^{2\omega} = 2\omega H_{z,rf}^2(H_y^{rf}/M_s)^2|\tilde{\chi}(\omega)_{\parallel}|^2\chi(2\omega)''_{\parallel}
$, from which we conclude that under H$_B$ for FMR at frequency $f=\omega/(2\pi)$, we should see a power ratio $$P^{2\omega}/P^{\omega} = 2(H_z^{rf}/M_s)^2\chi(\omega)''_{\parallel}\chi(2\omega)''_{\parallel}$$ With $M_s =$ 844 Oe, $\alpha =$ 0.007 as measured by FMR for our Ni$_{81}$Fe$_{19}$ 30 nm sample and 2.25 Oe rf field amplitude at input power of 1 W for the CPW, we have a calculated $2f/f$ power ratio of 1.72$\times$10$^{-5}$/W, which is in reasonable agreement with the experimental data 3.70$\times$10$^{-5}$/W as shown in Fig.2(c). To compare this result with the ferrite experiment in ref.\[1\], we further add the factor representing the ratio of FMR absorption to the input rf power, which is $3.9 \times 10^{-2}$ in our setup. This leads to an experimental upconversion ratio of $1.44 \times 10^{-6}$/W in ref.\[1\]’s definition ($\Delta P^{2\omega} / {P^{\omega}_{in}}^2 $), compared with $7.1 \times 10^{-10}$/W observed in Mg$_{70}$Mn$_8$Fe$_{22}$O (Ferramic R-1 ferrite).\
Examining Eq.(4), we notice that there should be two peaks in the field-swept $2f$ emission spectrum: the first coincides with the FMR but with a narrower linewidth due to the term $|\tilde{\chi}(\omega)_{\parallel}|^2$, and the second positioned at the H$_B$ for the FMR with a $2f$ input signal due to the term $\chi(2\omega)''_{\parallel}$. The second peak should have a much smaller amplitude. Due to the field limit of our electromagnet, we could not reach the bias field required for FMR at 12.2 GHz under this particular configuration and continued to verify Eq.(4) at a lower frequency of 2.0 GHz. We carried out an identical experiment and analysis and observed an upconversion efficiency of 0.39$\times$10$^{-3}$/W for the 4.0 GHz signal generation at 2.0 GHz input, again in reasonable agreement with the theoretical prediction 1.17$\times$10$^{-3}$/W. Fig.3 demonstrates the typical line shape of the 4 GHz spectrum, in which the input 2 GHz power being +18.9 dBm. A second peak at the H$_B$ for 4 GHz FMR is clearly visible with a much smaller amplitude and larger linewidth than the first peak, qualitatively consistent with Eq.(4). A theoretical line (dashed green) from equation (4) with fixed damping parameter $\alpha = 0.007$ is drawn to compare with the experimental data. The observed second peak at the $2f$ resonance H$_B$ shows a much lower amplitude than expected. We contribute this difference to the possible $2f$ component in the rf source which causes the $2f$ FMR absorption. The blue line shows the adjusted theoretical line with consideration of this input signal impurity.\
![\[Fig.3\] 4 GHz generation with input signal at 2 GHz, +18.9 dBm. A second peak at the bias field for 4 GHz FMR is clearly present; red dots: experimental data; dashed green: theoretical; blue: adjusted theoretical with input rf impurity. See text for details.](fig3.eps)
*Summary*: We have demonstrated a highly efficient frequency doubling effect in thin-film Ni$_{81}$Fe$_{19}$ for input powers well below the Suhl instability threshold. An analysis of the intrinsically nonlinear LLG equation interprets the observed phenomena quantitatively. The results explore new opportunities in the field of rf signal manipulation with CMOS compatible thin film structures.\
We acknowledge Stephane Auffret for the Ni$_{81}$Fe$_{19}$ sample. We acknowledge support from the US Department of Energy grant DE-EE0002892 and National Science Foundation ECCS-0925829.
[99]{}
W. P. Ayres, P. H. Vartanian, and J. L. Melchor, J. Appl. Phys. **27**, 188 (1956)
N. Bloembergen and S. Wang, Phys. Rev. **93**, 72 (1954)
H. Suhl, J. Phys. Chem. Solids. **1**, 209 (1957)
J. D. Bierlein and P. M. Richards, Phys. Rev. B **1**, 4342 (1970)
G. P. Ridrigue, J. Appl. Phys. **40**, 929 (1969)
V. G. Harris, IEEE Trans. Magn. **48**, 1075 (2012)
V. E. Demidov, S. Urazhdin, H. Ulrichs, V. Tiberkevich, A. Slavin, D. Baither, G. Schmitz and S. O. Demokritov, Nature Mat. **11**, 1028 (2012)
A. Berteaud and H. Pascard, J. Appl. Phys. **37**, 2035 (1966)
T. Gerrits, P. Krivosik, M. L. Schneider, C. E. Patton, and T. J. Silva, Phys. Rev. Lett. **98**, 207602 (2007)
H. M. Olson, P. Krivosik, K. Srinivasan, and C. E. Patton, J. Appl. Phys. **102**, 023904 (2007)
M. Bao, A. Khitun, Y. Wu, J. Lee, K. L. Wang, and A. P. Jacob, Appl. Phys. Lett. **93**, 072509 (2008)
Y. Khivintsev, J. Marsh, V. Zagorodnii, I. Harward, J. Lovejoy, P. Krivosik, R. E. Camley, and Z. Celinski, Appl. Phys. Lett. **98**, 042505 (2011)
J. Marsh, V. Zagorodnii, Z. Celinski, and R. E. Camley, Appl. Phys. Lett. **100**, 102404 (2012)
M. Yana, P. Vavassori, G. Leaf, F.Y. Fradin, and M. Grimsditch, J. Magn. Magn. Mater **320**, 1909 (2008)
V. E. Demidov, H. Ulrichs, S. Urazhdin, S. O. Demokritov, V. Bessonov, R. Gieniusz, and A. Maziewski, Appl. Phys. Lett. **99**, 012505 (2011)
C. Bi, X. Fan, L. Pan, X. Kou, J. Wu, Q. Yang, H. Zhang, and J. Q. Xiao, Appl. Phys. Lett. **99**, 232506 (2011)
S. E. Bushnell, W. B. Nowak, S. A. Oliver, and C. Vittoria, Rev. Sci. Instrum. **63**, 2021 (1992)
A. G. Gurevich and G. A. Melkov, Magnetization Oscillation and Waves (CRC, Boca Raton, 1996)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: '[Deterministic evolutionary theory robustly predicts that populations displaying altruistic behaviors will be driven to extinction by mutant “cheats” that absorb common benefits but do not themselves contribute. Here we show that when demographic stochasticity is accounted for, selection can in fact act in the reverse direction to that predicted deterministically, instead favoring cooperative behaviors that appreciably increase the carrying capacity of the population. Populations that exist in larger numbers experience a selective advantage by being more stochastically robust to invasions than smaller populations, and this advantage can persist even in the presence of reproductive costs. We investigate this general effect in the specific context of public goods production and find conditions for stochastic selection reversal leading to the success of public good producers. This insight, developed here analytically, is missed by both the deterministic analysis as well as standard game theoretic models that enforce a fixed population size. The effect is found to be amplified by space; in this scenario we find that selection reversal occurs within biologically reasonable parameter regimes for microbial populations. Beyond the public good problem, we formulate a general mathematical framework for models that may exhibit stochastic selection reversal. In this context, we describe a stochastic analogue to $r-K$ theory, by which small populations can evolve to higher densities in the absence of disturbance.]{}'
author:
- 'George W. A. Constable'
- Tim Rogers
- 'Alan J. McKane'
- 'Corina E. Tarnita'
nocite: '[@*]'
title: Demographic noise can reverse the direction of deterministic selection
---
,
Over the past century, mathematical biology has provided a framework with which to begin to understand the complexities of evolution. Historically, development has focused on deterministic models [@hofbauer_1998]. However, when it comes to questions of invasion and migration in ecological systems, it is widely acknowledged that stochastic effects may be paramount, since the incoming number of individuals is typically small. The importance of demographic (intrinsic) noise has long been argued in population genetics; it is the driver of genetic drift and can undermine the effect of selection in small populations [@fisher_1930; @wright_1931]. This concept has also found favor in game theoretic models of evolution which seek to understand how apparently altruistic traits can invade and establish in populations [@nowak_2006]. However, the last decade has seen an increase in the awareness of some of the more exotic and counter-intuitive aspects of demographic noise: it has the capacity to induce cycling of species [@mckane_2005], pattern formation [@butler_2009; @hallatschek_2007], speciation [@rossberg_2013] and spontaneous organization in systems that do not display such behavior deterministically.
Here we explore the impact of demographic noise on the direction of selection in interactions between multiple phenotypes or species. Historically, a key obstacle to progress in this area has been the analytical intractability of multidimensional stochastic models. This is particularly apparent when trying to investigate problems related to invasion, where systems are typically far from equilibrium. A promising avenue of analysis has recently become apparent however through stochastic fast-variable elimination [@parsons_quince_2007_1; @doering_2012]. If a system consists of processes that act over very different timescales, it is often possible to eliminate fast-modes, assumed to equilibrate quickly in the multidimensional model, and obtain a reduced dimensional description that is amenable to analysis [@gunawardena_2014]. This approach has been employed multiple times over the last decade to study a stochastic formulation of the classical Lotka-Volterra competition model for two competing phenotypes/species. In [@parsons_quince_2007_1; @parsons_quince_2007_2; @doering_2012; @nelson_2015; @constable_2015], such models were analyzed under the assumption that the dynamics regulating the total population size (birth, death and competition) occurred on a much faster timescale than the change in population composition. In particular [@parsons_quince_2007_1; @parsons_quince_2007_2; @doering_2012; @nelson_2015] have shown that it is possible for systems that appear neutral in a deterministic setting to become non-neutral once stochasticity is included. If the two phenotypes have equal deterministic fitness, but one is subject to a larger amount of demographic noise than the other, then the effect of this noise alone can induce a selective drift in favor of the phenotype experiencing less noise. This stems from the fact that it is easier to invade a noisy population than a stable one; furthermore, the direction of this induced selection can vary with the system’s state [@parsons_quince_2010]. The idea has been further generalized mathematically in [@kogan_2014].
Here we will show more generally that not only can stochasticity break deterministic neutrality, but that it has the capacity to reverse the direction of selection predicted deterministically. Thus while in a deterministic setting a certain phenotype will always reach fixation (and is resistant to invasions), in a stochastic setting its counterpart can in fact be more likely to invade and fixate (and less susceptible to invasions). These results generalize recent work on modified Moran and Wright-Fisher type models [@houch_2012; @houch_2014] to a large class of models consisting of two phenotypes interacting with their environment. We begin with the analysis of a prototypical public good model, which is used to illustrate our analysis. We find that stochastic selection reversal can alleviate the public good production dilemma. We further show how space can amplify this phenomenon, allowing the reversal of selection to emerge over a greater parameter range. Finally, we extend the ideas to a more general model framework, and explore the types of system in which we expect this behavior to be relevant. In particular we discuss the similarities with $r-K$ selection theory [@reznick_2002].
Public Good model
=================
It is generally accepted that random events play a strong role in the evolution of cooperative behavior, which is deterministically selected against [@nowak_2006]. The standard formulation of evolutionary game theory involves setting the problem in terms of a modified Moran model [@nowak_2004; @rice_evo_theory]. The Moran model is a population genetic model first developed as an abstract illustration of the effect of genetic drift in a haploid population of two phenotypes; an individual is picked to reproduce with a probability proportional to their fitness, whilst simultaneously a second individual is chosen randomly to die [@crow_kimura_into]. Coupling birth and death events keeps the population size fixed, which increases the tractability of the system.
The specification of fixed population size is however restrictive and can be problematic. Most prominently, a phenotype with increased fitness can be no more abundant in isolation than its ailing counterpart. Additional difficulties are encountered if one attempts to use simple game-theoretic models to quantitatively understand more complex experimental data. While, for example, assuming some arbitrary non-linearity in the model’s game payoff matrix may enable experimental findings to be elegantly recapitulated, it is more difficult to justify the origin of these assumptions on a mechanistic level [@gore_2009]. In light of such issues, it has been suggested that a more ecologically grounded take on the dynamics of cooperation might be preferable [@hauert_2006; @huang_2015], one in which the population size is not fixed and that is sufficiently detailed that mechanistic (rather than phenomenological) parameters can be inferred experimentally. In the following, we take such an approach. We begin by considering a prototypical model of public good production and consumption.
In our model, we consider a phenotype $X$ having the ability to produce a public good $Q$ that catalyzes its growth. We wish to capture the stochastic dynamics of the system. To this end we assume that the system is described by a set of probability transition rates, which describe the probability per unit time of each reaction occurring: $$\begin{aligned}
& X \overset{b_{x}}{\underset{\kappa/R^{2}}{ {\mathrel{\substack{\xrightarrow{\rule{0.4cm}{0cm}} \\[-.9ex] \xleftarrow{\rule{0.4cm}{0cm}}}}} } } X + X \,, \quad & X + Q \xrightarrow{ r / R^{2} } X + X + Q \,, \nonumber \\
& X\xrightarrow{ p_{x} } X + Q \,,\quad & Q\xrightarrow{\delta}\varnothing\,. \label{eq_reactions_1}\end{aligned}$$ In the absence of the public good, the producer phenotype $X$ reproduces at a baseline birthrate $b_{x}$. The phenotypes encounter each other and the public good at a rate $R^{-2}$; the quantity $R^{2}$ can be interpreted as a measure of the area (or volume) to which the system is confined. Death of the phenotype occurs solely due to crowding effects at rate $\kappa$, multiplied by the encounter rate. Phenotypes encounter and utilize the public good at a rate $r/R^{2}$. We study the case where this reaction is catalytic (i.e. the public good is conserved) and leads to a phenotype reproduction. Examples of catalytic (reusable) public goods are the enzyme invertase produced by the yeast *Saccharomyces cerevisiae* [@koschwanez_2011] or the siderophore pyoverdine produced by the bacterium *Pseudomonas aeruginosa* [@kummerli_2010]. The total rate at which the phenotype reproduces is thus increased in the presence of the public good. The public good itself is produced by the producer phenotype at a rate $p_{x}$ and decays at a rate $\delta$. Note that as well as controlling the spatial scale of the well-mixed system, the magnitude of $R$ will also control the typical number of individuals in the system, since larger $R$ (more space) allows the population to grow to greater numbers. We next introduce a mutant phenotype $Y$ that does not produce the public good; (i.e. $p_{y}=0$) consequently, it has a different baseline birth rate $b_{y}$ which we expect to be at least as high as that of the producer, due to the non-producers’ reduced metabolic expenditure. Its interactions with the public good are otherwise similar to those of $X$ (see [Eq.(\[eq\_reactions\_1\])]{}).
The state of the system is specified by the discrete variables $n_{x}$, $n_{y}$ and $n_{q}$, the number of each phenotype and public good respectively. For the system described, we wish to know the probability of being in any given state at any given time. To answer this, we set up an infinite set of partial difference equations (one for each unique state $(n_x,n_y,n_q)$) that measures the flow of probability between neighboring states (controlled by the transitions [Eq.(\[eq\_reactions\_1\])]{}). These equations govern the time-evolution of a probability density function $P(n_{x},n_{y},n_{q},t)$ (see [Eq.(\[eq:supmat\_meqn\_general\])]{}). Such a model is sometimes termed a microscopic description [@gardiner_2009], since it takes account of the dynamics of discrete interactions between the system variables.
Although the probabilistic model is straightforward to formalize, it is difficult to solve in its entirety. We apply an approximation that makes the model more tractable, while maintaining the system’s probabilistic nature. Such approximations, which assume that the system under consideration has a large but finite number of individuals, are well practiced and understood [@gardiner_2009] and are analogous to the diffusion approximation [@crow_kimura_into] of population genetics. Assuming that $R$ is large, but finite, (which implies a large number of individuals in the system), we transform the system into the approximately continuous variables $(x,y,q)=(n_{x},n_{y},n_{q})/R^{2}$ and expand the partial difference equations in $1/R^{2}$. This allows us to to express the infinite set of partial difference equations as a single partial differential equation in four continuous variables, $(x,y,q,t)$. However, since the PDE results from a Taylor expansion, it has infinite order. Truncating the expression after the first term (at order $R^{-2}$), one obtains a deterministic approximation of the dynamics (valid for $R \rightarrow \infty$, or equivalently for infinite population sizes). Since we aim to make the system tractable but still retain some stochastic element to the dynamics, we truncate the expansion after the second term (at order $R^{-4}$, see [Eq.(\[eq:supmat\_eq\_FPE\_general\])]{}). The resulting model can be conveniently expressed as a set of Itō stochastic differential equations (SDEs): $$\begin{aligned}
\dot x &=& x \left[ b_{x} + r q -\kappa (x+y) \right] + R^{-1} \eta_{x}(t)\,, \nonumber \\
\dot y &=& y \left[ b_{y} + r q -\kappa (x+y) \right] + R^{-1} \eta_{y}(t)\,, \label{eq_PG_SDE} \\
\dot q &=& p_{x} x - \delta q + R^{-1} \eta_{q}(t) \,. \nonumber\end{aligned}$$ The $\eta_{i}(t)$ represent Gaussian white noise terms whose correlations depend on the state of the system (the noise is multiplicative). Importantly, because [Eq.(\[eq\_PG\_SDE\])]{} has been developed as a rigorous approximation of the underlying stochastic model, [Eq.(\[eq\_reactions\_1\])]{}, the precise functional form of the noise can be determined explicitly, rather than posited on an ad-hoc basis (see [Appendix\[sec:supmat\_sec\_obtain\_SDE\_system\]]{}). Setting $ R \rightarrow \infty $, the population size increases with the interaction scale and one recovers the deterministic limit. Since [Eq.(\[eq\_PG\_SDE\])]{} is a course-grained approximation of the underlying microscopic model but retains an inherent stochasticity, it is often referred to as the mesoscopic limit [@mckane_TREE].
First we analyze the dynamics of [Eq.(\[eq\_PG\_SDE\])]{} in the deterministic, $ R \rightarrow \infty $ limit. There exist three fixed points, or equilibria. The first, at the origin, is always unstable. The remaining fixed points occur when the system only contains a single phenotype: the producer fixed point, $(x,y,q)=( K_{x} , 0 , p_{x} K_{x}/\delta )$ and the non-producer fixed point, $(x,y,q)=( 0 , K_{y} , 0 )$. Thus $K_{x}$ and $K_{y}$ are measures of the phenotypes’ frequency (carrying capacity) in isolation, with precise forms $$K_{x}= \frac{ b_{x} \delta }{ \kappa \delta - p_{x} r }\,, \qquad K_{y}= \frac{ b_{y} }{ \kappa }\,. \label{eq_carrying_capacity}$$ If $b_y>b_x$ then the non-producer fixed point is always stable while the producer fixed point is always unstable. However, the non-producer fixed point is only globally attracting if $ \kappa \delta > r p_{x} $. If this condition is not met then there exist initial conditions for which the producers produce and process the public good faster than they die and faster than the public good degrades, resulting in unbounded exponential growth of the system. This biologically unrealistic behavior comes from the fact that we have assumed for simplicity that the public good uptake does not saturate. Since this behavior is unrealistic, we will work in the regime $ \kappa \delta > r p_{x} $ for the remainder of the paper. Finally, we are interested in systems where the size of the producer population in isolation is larger than that of the non-producer, $K_x>K_y$; this is true if the condition $b_{x}>b_{y}(1-r p_x/\delta\kappa)$ holds. Thus, deterministically, a non-producing mutant will always take over a producer population and, due to the absence of the public good, it will yield a smaller population at equilibrium.
This deterministic analysis predicts, unsurprisingly, that a population composed entirely of non-producers is the only stable state. We next explore the behavior of the system in [Eq.(\[eq\_reactions\_1\])]{} when demographic stochasticity is considered.
Mesoscopic selection reversal
-----------------------------
Due to noise, a stochastic system will not be positioned precisely on deterministic fixed points, but rather it will fluctuate around them. In the above system, these fluctuations will occur along the $y$-axis for the non-producer fixed point while in the absence of non-producers they will occur in the $(x,q)$ plane for the producer fixed point. We can define $N_{x} = R^{2} K_{x}$ and $N_{y} = R^{2} K_{y}$ to be the mean number of the phenotypes $X$ and $Y$ in isolation in the respective stationary states. We assume that the non-producing phenotype has a greater per-capita birth rate than the producer phenotype, i.e. $b_{y} > b_{x}$, and we introduce a single non-producing mutant into a producer population. While the deterministic theory predicts that the non-producer should sweep through the population until it reaches fixation, in the stochastic setting fixation of the non-producer is by no means guaranteed: there is a high probability that the single mutant might be lost due to demographic noise. However, since the non-producer is deterministically selected for, we might expect the probability of a non-producer mutant invading and fixating in a resident producer population to be greater than the probability of a producer mutant invading and fixating in a resident non-producer population. We will explore this question below.
In order to make analytic predictions about the stochastic model, we need to reduce the complexity of the system. This can be done if we employ methods based on the elimination of fast variables [@parsons_rogers_2015] to obtain an effective one-dimensional description of the system dynamics. To this end, we begin by assuming that the public good production and decay, $p_{x}$ and $\delta$, and the phenotypes’ reproduction and death, $b_{x}$, $b_{y}$, and $\kappa$, occur on a much faster timescale than the rate of change of population composition, which is governed by the difference in birth rates, $b_x - b_y$. Essentially this assumes that the cost of public good production is marginal. In the case of *S. cerevisiae*, this assumption is supported by empirical work (see [Table\[tab:supmat\_fig\_parameter\_list\_1\]]{}). In order to mathematically investigate this timescale-separation we define $$b_{x} = b(1-\varepsilon)\,,\quad b_{y} = b \,, \label{eq_def_b}$$ where the parameter $\varepsilon$ represents the metabolic cost that $X$ pays for producing the public good. The parameter $\varepsilon$ now controls the rate of change of population composition, and if $1>>\varepsilon$, we have our desired timescale separation in the deterministic system. Because the parameters $K_{x}$, $K_{y}$, $N_{x}$ and $N_{y}$ depend on $\varepsilon$, we will find it convenient to define their values when $\varepsilon=0$ as $K_{x}^{(0)}$, $K_{y}^{(0)}$, $N_{x}^{(0)}$ and $N_{y}^{(0)}$ respectively. In order to maintain our assumption that the composition of the phenotype population changes slowly in the stochastic system, we additionally require that the noise is small. However this assumption has already been implicitly made in the derivation of [Eq.(\[eq\_PG\_SDE\])]{}, where it was assumed that $R$ is large, and thus $R^{-1}$, the prefactor for the noise terms, is small. In order to formalize this, we will find it convenient to assume $R^{-2}\approx \mathcal{O}\left( \varepsilon \right)$.
Under the above assumptions, the system features a separation of timescales. Next, we take advantage of this to reduce the complexity of the system. Deterministically, the existence of a set of fast timescales suggests the existence of a lower-dimensional subspace, the slow manifold (SM), shown in [Fig.\[fig\_streamplots\]]{}(a), to which the system quickly relaxes, and along which it slowly moves, until it reaches the system’s stable fixed point. This behavior can be exploited if we assume that the system reaches the SM instantaneously. We can then describe the dynamics of the entire system in this lower dimensional space, and thus reduce the number of variables in our description of the deterministic system. However, we are interested in the stochastic dynamics.
![System dynamics in the phenotype plane. Deterministic trajectories shown as gray arrows. Panel (a): Trajectories rapidly collapse to a SM (black dashed line), before slowly moving to the non-producing $Y$ fixed-point. Stochastic trajectories (histogram overlaid in orange) remain in the region of the SM but may fluctuate away from it. Panel (b): Illustration of the origin of noise-induced selection. The orange ellipse depicts the standard deviation of Gaussian fluctuations originating at its center. Fluctuations (black dashed arrows) to points $\alpha$ are equally likely, however when projected back to the CM (black dashed line) to points $\beta$, a bias for producing $X$ phenotype is observed. Parameters used are $p_{x}=9.5\times10^{-4}$, $\varepsilon=0.08$ in panel (a), $\varepsilon=0$ in panel (b) and the remaining parameters are given in [Table\[tab:supmat\_fig\_parameter\_list\_1\]]{}. []{data-label="fig_streamplots"}](fig1.pdf){width="50.00000%"}
The stochastic trajectories initially collapse to the region around the SM, about which they are confined, but along which they can move freely until one of the phenotypes fixates (see [Fig.\[fig\_streamplots\]]{}(a)). Fluctuations that take the system off the SM are quickly quashed back to another point on the SM; however the average position on the SM to which a fluctuation returns is not necessarily the same as that from which the fluctuation originated. A crucial element of the dynamics in this stochastic setting is that the form of the noise, combined with that of the trajectories back to the SM, can induce a bias in the dynamics along the SM (see [Fig.\[fig\_streamplots\]]{}(b)). This is the origin of the stochastic selection reversal that we will explore. In order to capture this behavior while simultaneously removing the fast timescales in the stochastic system, we map all fluctuations off the SM along deterministic trajectories back to the SM [@parsons_rogers_2015]. This essentially assumes that any noisy event that takes the system off the SM is instantaneously projected back to another point on the SM.
For clarity, we briefly describe the dynamics when $\varepsilon=0$. In this case the birth rates of phenotypes $X$ and $Y$ are identical. Instead of the two non-zero fixed points, $K_{x}$ and $K_{y}$, found above, the deterministic system now has a line of fixed points, referred to as a center manifold (CM) [@arnold_2003]. The CM is identical to the SM in the limit $\varepsilon\rightarrow0$. It is given by $$y = \frac{ K_{y}^{(0)} }{ K_{x}^{(0)} } \left( K_{x}^{(0)} - x \right) \,, \quad q = \frac{ p_{x} }{ \delta } x \,, \label{eq_CM}$$ and shown graphically in [Fig.\[fig\_streamplots\]]{}(b). The separation of timescales in the system is now at its most pronounced, since there are strictly no deterministic dynamics along the CM following the fast transient to the CM. However the stochastic system still features dynamics along the CM. Applying the procedure outlined in [@parsons_rogers_2015] we arrive at a description of the stochastic dynamics in a single variable, the frequency of producers along the CM; $$\dot{x} = \frac{ b }{ R^{2} } x \left( 1 - \frac{ x }{ K_{x}^{(0)} } \right) \,\mathcal{F}(x) + \frac{1}{R} \zeta(t) \,, \label{eq_reduced_PG_SDE_neutral}$$ where $$\mathcal{F}(x) = 2 \left( \frac{ K_{x}^{(0)} - K_{y}^{(0)} }{ (K_{x}^{(0)}K_{y}^{(0)})^{2} } \right) \left[ K_{x}^{(0)} K_{y}^{(0)} + \left( K_{x}^{(0)} - K_{y}^{(0)} \right) x \right] \,. \nonumber$$ Here $\zeta(t)$ is a Gaussian white noise term with a correlation structure given in [Eq.(\[eq:supmat\_eq\_PG\_clean\_Bbar\])]{}. Together with [Eq.(\[eq\_CM\])]{}, [Eq.(\[eq\_reduced\_PG\_SDE\_neutral\])]{} approximates the dynamics of the entire system. Note that while [Eq.(\[eq\_reduced\_PG\_SDE\_neutral\])]{} predicts a noise-induced directional drift along the CM (controlled by $\mathcal{F}(x)$), a deterministic analysis predicts no dynamics, since the CM is by definition a line of fixed points. This directional drift along the CM results from the projection bias illustrated in [Fig.\[fig\_streamplots\]]{}(b). If $p_{x}>0$, then $K_{x}^{(0)}>K_{y}^{(0)}$, and so $\mathcal{F}(x) > 0 $; thus the public good production by phenotype $X$ induces a selective pressure that selects for $X$ along the center manifold.
The origin of the term $\mathcal{F}(x)$ in [Eq.(\[eq\_reduced\_PG\_SDE\_neutral\])]{} can be understood more fully by exploring its implications for the invasion probabilities of $X$ and $Y$, denoted $\phi_{x}$ and $\phi_{y}$. These can be straightforwardly calculated since the system is one-dimensional (see [Appendix\[sec:supmat\_sec\_fixprob\]]{}). We find $$\phi_{x} = \frac{1}{N_{y}} \,, \qquad \textrm{and} \qquad \phi_{y} = \frac{1}{N_{x}} \,, \label{eq_phi_quasineutral}$$ where $\phi_{x}>\phi_{y}$ so long as $p_{x} > 0$ (see [Eq.(\[eq\_carrying\_capacity\])]{}). The term $\mathcal{F}(x)$ can thus be interpreted as resulting from the stochastic advantage the producers have at the population level from reaching higher carrying capacities in isolation, which makes them more stochastically robust to invasion attempts. This result is independent of the spatial scale $R$ (and therefore population size) so long as $R$ is finite.
If $\varepsilon \neq 0 $, the system does not collapse to the CM, but rather to the SM. At leading order in $\varepsilon$, the equation for the SM is given by [Eq.(\[eq\_CM\])]{}. Upon removing the fast dynamics, the effective dynamics of $x$ can now be shown to take the form (see [Eq.(\[eq:supmat\_eq\_PG\_maintext\_effective\])]{}) $$\dot{x} = b x \left( 1 - \frac{x}{K_{x}^{(0)}} \right) \left( \frac{ 1 }{ R^{2} } \,\mathcal{F}(x) - \varepsilon \right) + \frac{1}{R} \zeta(t) \,, \label{eq_reduced_PG_SDE}$$ where $\zeta(t)$ and $\mathcal{F}(x)$ are the same as in [Eq.(\[eq\_reduced\_PG\_SDE\_neutral\])]{}. The SDE now consists of two components. The deterministic contribution, governed by $\varepsilon$, exerts a selective pressure against phenotype $X$, due to its reduced birth rate. The stochastic term, $\mathcal{F}(x)$ exerts a pressure in favor of phenotype $X$, resulting, as in the case $\varepsilon = 0$ discussed above, from the producers’ stochastic robustness to invasions.
![Stochasticity can render non-producers more susceptible to invasion by producers than vice versa. Plots of the difference in invasion probabilities between producers $X$ and non-producers $Y$ as a function of the cost to birth for production, $\varepsilon$, and good production rate $p_{x}$. The remaining parameters are taken from [Table\[tab:supmat\_fig\_parameter\_list\_1\]]{}. Left: analytic results for a single small patch (see [Eq.(\[eq\_eps\_condition\])]{}). The critical cost $\varepsilon$ for selection reversal, [Eq.(\[eq\_eps\_condition\])]{}, is shown here as the black dashed line. Right; results from Gillespie simulations [@gillespie_1976] of the stochastic process [Eq.(\[eq\_reactions\_1\])]{}, averaged over $2000$ runs.[]{data-label="fig_trade_off"}](fig2.pdf){width="50.00000%"}
Thus, when $\varepsilon > 0$, a trade-off emerges in the stochastic system between the stochastic advantage to public good production (due to increased population sizes) and the deterministic cost producers pay (in terms of birth rates). If the birth costs are not too high, producers will be selected for, which constitutes a reversal in the direction of selection from the deterministic prediction. Specifically, we can calculate the condition on the metabolic cost that ensures that the producers are fitter than the non-producers (i.e. $\phi_{x}>\phi_{y}$): $$\varepsilon < \frac{\kappa}{ b R^{2} } \log \left[ \frac{\delta \kappa }{ \delta \kappa - p_{x} r } \right] \,. \label{eq_eps_condition}$$ Whereas for no metabolic cost producers consistently have a stochastic advantage regardless of typical population size (see [Eq.(\[eq\_phi\_quasineutral\])]{}), for non-zero production costs, the population must be sufficiently small that stochastic effects, governed by $R^{-2}$, are dominant. [Fig.\[fig\_trade\_off\]]{} shows that the theory predicts well the trade-off in the underlying stochastic model (\[eq\_reactions\_1\]).
We have shown that stochastic selection reversal is more prevalent when $R$ is not large. Meanwhile our analytic results results have been obtained under the assumption that $R$ is large, which allowed us to utilize the diffusion approximation leading to [Eq.(\[eq\_PG\_SDE\])]{} and aided the timescale elimination procedure that yielded [Eq.(\[eq\_reduced\_PG\_SDE\])]{}. We therefore expect that although stochastic selection reversal will become more prominent as $R$ is reduced, the quality of our analytic predictions may suffer. Despite this caveat, it is the small $R$ regime that is interesting to us. Small values of $R$ are associated with small population sizes. While it is conceivable that populations of macro-organisms may consist of a small number of individuals, this limit is not so pertinent to the study of micro-organisms. In the next section however, we will show that by incorporating space, the constraint of small population size can be relaxed.
Spatial amplification
=====================
In this section we consider a metapopulation on a grid: each subpopulation (patch) has a small size so that demographic noise continues to be relevant locally, but the number of subpopulations is large so that the overall population in the system is large. This method of incorporating demographic stochasticity into spatial systems has proved to be successful in the modeling of microbial populations [@hallatschek_2007]. We consider a grid of $C$ patches. The dynamics within each patch are given by the transitions in [Eq.(\[eq\_reactions\_1\])]{}, and coupled to the surrounding patches by the movement of the phenotypes and public good. A patch will produce migrants at a rate proportional to its density. Producers $X$ and non-producers $Y$ disperse with a probability rate $m$ to a surrounding region, while the public good diffuses into neighboring regions at a rate $D$. Once again the diffusion approximation can be applied to obtain a set of SDEs approximating the system dynamics; $$\begin{aligned}
\frac{\mathrm{d}x_{ij}}{\mathrm{d}\tau} &=& x_{ij} \left( b_{x} + r q_{ij} - \kappa ( x_{ij} + y_{ij} ) \right) + m \left( L \bm{x} \right)_{ij} + \frac{ \eta_{xij}(t) }{ R } \, , \nonumber \\
\frac{\mathrm{d}y_{ij}}{\mathrm{d}\tau} &=& y_{ij} \left( b_{y} + r q_{ij} - \kappa ( x_{ij} + y_{ij} ) \right) + m \left( L \bm{y} \right)_{ij} + \frac{ \eta_{yij}(t) }{ R }\, , \nonumber \\
\frac{\mathrm{d}q_{ij}}{\mathrm{d}\tau} &=& p_{x} x_{ij} - \delta q + D \left( L \bm{q} \right)_{ij} + \frac{ \eta_{qij}(t) }{ R } \,,
\label{eq_spatial_ODEs} \end{aligned}$$ where $ij$ is the patch on row $i$ and column $j$. The operator $L$ is the discrete Laplacian operator $(L\bm{x})_{ij} = - 4 x_{ij} + x_{(i-1)j} + x_{(i+1)j} + x_{i(j-1)} + x_{i(j+1)} $. If $b_{y}>b_{x}$, the deterministic dynamics predict that the producers will always go extinct.
![Left panel: analytic results show that space amplifies stochastic selection reversal in the low dispersal, zero diffusion limit. The critical maximum cost $\varepsilon$ for selection reversal (see [Eq.(\[eq\_eps\_crit\_space\])]{}) is plotted as black dashed line. Right panel: simulation results are shown for varied $m$ and $D$, averaged over $2000$ runs. Panel (a) gives results which in the low dispersal, zero diffusion limit ($m=3.7\times10^{-8}$, $D=0$), which match our theoretical predictions. Panel (b) gives the result with a set of biologically plausible parameters ($m=3.7\times10^{-7}$ and $D=2.2\times10^{-5}$ derived in [Table\[tab:supmat\_fig\_parameter\_list\_1\]]{}). Panel (c) gives results in a system with high dispersal ($m=3.7\times10^{-5}$, $D=2.2\times10^{-5}$), while panel (d) shows the case of a system with high diffusion ($m=3.7\times10^{-7}$ and $D=2.2\times10^{-3}$). The number of patches is given by $C=16$ and the remaining parameters are listed in [Table\[tab:supmat\_fig\_parameter\_list\_1\]]{}.[]{data-label="fig_trade_off_16_islands"}](fig3.pdf){width="50.00000%"}
First we will discuss some important limit case behavior for this system. In the limit of large dispersal rate $m$ and diffusion rate $D$, the stochastic system behaves like a well-mixed population with a spatial scale $cR^{2}$ (i.e. the spatial structure is lost). In this case, as the size of the spatial system is increased, the effective population size also increases, and as a consequence selection reversal for producing phenotypes becomes less likely (see [Eq.(\[eq\_eps\_condition\])]{}).
We next consider the low-dispersal, zero diffusion limit. For sufficiently low dispersal, any incoming mutant will first either fixate or go to extinction locally before any further dispersal event occurs. Since each dispersal/invasion/extinction event resolves quickly, at the population level, the system behaves like a Moran process on a graph [@nowak_2006], with each node representing a patch. The ‘fitness’ of a patch is the probability that it produces a migrant, and that that migrant successfully invades a homogeneous patch of the opposite type, following the approach used in [@houch_2012]. Denoting the ‘fitness’ of producing and non-producing patches by $W_{x}$ and $W_{y}$ respectively, we have $$W_{x} = m N_{x} \phi_{x} \,, \qquad W_{y} = m N_{y} \phi_{y} \,, \label{eq_patch_fitness}$$ where $N_{i}$ ($i = x,y$) is the mean carrying capacity of phenotype $i$ in a homogeneous patch, and $\phi_{i}$ are the invasion probabilities of a type $i$ mutant in a type $j\neq i$ patch. The fixation probabilities of a homogeneous patch in a population of the opposite phenotype can now be calculated using standard results [@nowak_2006]. Let $\rho_{i}$ ($i = x,y$) denote the probability that type $i$ takes over the metapopulation when starting from one patch of type $i$ in a population otherwise comprised entirely of patches of the opposite phenotype. Then $$\rho_{i} = \frac{1 - r_{i}^{-1} }{ 1 - r_{i}^{-C} } \,, \quad \textrm{for} \quad i = x,y \quad \textrm{and} \quad r_{x} = \frac{W_{x}}{W_{y}} \,, \quad r_{y} = \frac{W_{y}}{W_{x}} \,. \label{eq_regular_graph_fix}$$ If we start from a single invading mutant, the probability that it takes over the entire population (i.e. invasion probability) is the product between the probability that it takes over its home patch $\phi_{i}$, and the probability that the newly invaded home patch fixates into the metapopulation, $\rho_{i}$: $$\begin{aligned}
\Pi_{x} = \phi_{x} \rho_{x} \,, \quad \Pi_{y} = \phi_{y} \rho_{y} \,. \label{eq_pi}\end{aligned}$$ In the infinite patch limit ($C\rightarrow\infty$), $\rho_{x}$ and $\rho_{y}$ depend on $r_{x}$, the patch fitness ratio defined in [Eq.(\[eq\_regular\_graph\_fix\])]{}. If $r_{x}>1$, $\rho_{x}\rightarrow1-r_{x}^{-1}$ and $\rho_{y}\rightarrow0$, whereas if $r_{x}<1$ the converse is true. This means that, in the infinite patch, low dispersal, zero diffusion limit, the condition for the stochastic reversal of selection is weakened from $\phi_{x} > \phi_{y}$ to $$\begin{aligned}
N_{x} \phi_{x} > N_{y} \phi_{y} \,. \label{eq_weakening_threshold}\end{aligned}$$ Spatial structure therefore has the ability to enhance the stochastic reversal observed in the small well-mixed system. An approximate analytic form for the above condition can be obtained in terms of the original parameters; $$\varepsilon < 2 \frac{ \kappa }{ b R^{2} } \log \left[ \frac{\delta \kappa }{ \delta \kappa - p_{x} r } \right] \,. \label{eq_eps_crit_space}$$
Once again, our analytical results are well supported by simulations (see [Fig.\[fig\_trade\_off\_16\_islands\]]{}). The critical production rate for the invasion probability of producers to exceed that of non-producers has been decreased, as predicted by Eqs. (\[eq\_eps\_condition\]) and (\[eq\_eps\_crit\_space\]). Producers can therefore withstand higher production costs in spatially structured environments.
It is important to note that while [Eq.(\[eq\_weakening\_threshold\])]{} depends on the mean number of producers and non-producers on a homogeneous patch ($N_x$ and $N_y$), it is independent of the number of individuals in the entire metapopulation in the large $C$ limit. The interaction between these two spatial scales leads to results that can appear counter-intuitive. Demographic noise, as we have discussed, leads to producing patches being ‘more fit’ at the patch level (see [Eq.(\[eq\_patch\_fitness\])]{}). However, when a large number of patches is considered, the demographic noise at the metapopulation level is reduced. This leads to the system following trajectories that appear deterministic at the level of the metapopulation, even though the path they follow is entirely the result of demographic stochasticity at the within-patch level (see [Fig.\[fig\_10000\_islands\]]{}). The movie S1 (see [Appendix\[sec:supmat\_sec\_movies\]]{}) displays the individual dynamics of the patches that comprise the trajectory illustrated in [Fig.\[fig\_10000\_islands\]]{}.
![Demographic stochasticity at the local ‘patch’ scale profoundly alters the system dynamics at the population level. Results are obtained from stochastic and deterministic ($R\rightarrow \infty $) simulations of [Eq.(\[eq\_spatial\_ODEs\])]{} with a grid of $C=100\times100$ patches, $p_{x}=1\times10^{-4}$, $\varepsilon=0.02$, $m=3.7\times10^{-5}$ and the remaining parameters taken from [Table\[tab:supmat\_fig\_parameter\_list\_1\]]{}. Initial conditions are a single producer and non-producer on each patch. The initial (fast) transient collapse to the SM occurs occurs in the shaded gray region. Following this, the deterministic system slowly moves along the slow manifold until the non-producers fixate, whereas in the stochastic system, the producers experience a selective pressure in their favor. For dynamics at the patch level, see Supplementary Information movie S1.[]{data-label="fig_10000_islands"}](fig4.pdf){width="50.00000%"}
Away from the small dispersal, zero diffusion limit, the dramatic selection reversal predicted by the analytical results is clearly weakened (see [Fig.\[fig\_trade\_off\_16\_islands\]]{}). Though selection reversal is still found across a range of $m$ and $D$ values, if either dispersal or diffusion are too high, the selection reversal breaks down. It is therefore important to understand what order of magnitude estimates for the values of $m$ and $D$ may be biologically reasonable.
Insights from
--------------
In the following section, we will attempt to contextualize our model with reference to a *S. cerevisiae* yeast system, which has been previously identified as a biological example of a population that features public good producers and non-producers. The model we have presented is general and therefore it could not capture the full biological detail of this particular system. For instance, it has been noted that some degree of privatization of the public good occurs in even the well-mixed experimental system [@gore_2009], a behavior we do not consider in our model. However, setting our model in this context can provide some insight into the scenarios in which we might expect stochastic selection reversal to be a biologically relevant phenomenon.
An *S.cerevisiae* yeast cell metabolizes simple sugars, such as glucose, in order to function. However, when simple sugars are scarce, the yeast can produce invertase, an enzyme that breaks down complex sugars, such as sucrose, to release glucose [@maclean_2010]. Invertase is produced at a metabolic cost and, since digestion of sucrose occurs extracellularly, most of the benefits of its production are shared by the population. Specifically in the case of S. cerevisiae, $SUC2$, the wild-type strain, produces invertase, while the lab cultured mutant $suc2$ does not [@maclean_2008]. In terms of our model parameters, the baseline birth rates, $b_{x}$ and $b_{y}$, represent respectively $SUC2$ and $suc2$ reproduction in the absence of invertase. This could be understood as arising from yeast directly metabolizing sucrose (a less energetically beneficial metabolic route [@maclean_2010]) or as the result of some extrinsically imposed low glucose concentration in the system. The rate $r$ would then represent the additional birth rate in the presence of invertase. The form of our specified reactions (see [Eq.(\[eq\_reactions\_1\])]{}), assumes that the presence of invertase leads directly to a yeast reproduction event. In reality invertase must break down the sucrose into glucose, and then slowly absorb the glucose. We are therefore essentially assuming that the sucrose is abundant, its breakdown by invertase instantaneous, and the glucose absorption rapid and occurring in discrete packets, with each packet absorbed leading to a reproduction event.
In the well-mixed system, our analytic predictions indicate that stochastic selection reversal can occur only if the population is very small. Since this is an unrealistic assumption in the case of yeast cultures, we would predict that non-producers should come to dominate a well-mixed population. In a spatially structured population however, this constraint is relaxed since it only requires small interaction regions. For *S.cerevisiae*, we can obtain order of magnitude estimates for the majority of parameters in our model, including public good diffusion (see [Appendix\[sec:supmat\_sec\_params\]]{}). Using these estimates together with our analytic results for the spatial public goods system, we find that stochastic selection reversal could feasibly be an important phenomenon for promoting the evolution of microbial public goods production in spatial settings (see [Fig.\[fig\_trade\_off\_16\_islands\]]{}, Panel b). Given this, we now consider the results of a spatial experiment on *S. cerevisiae*, and ask how its results might be interpreted in light of the insights developed with our simple model.
![Plots of the pairwise invariability scenarios possible for non-producing, producing and hyper-producing phenotypes. Arrows point away from the dominant phenotype in a pair, which is defined as that with a larger invasion probability (Fig. S3). Non-transitive dynamics are not possible. It is possible however for an optimal intermediate good production rate to emerge (cyan and blue regions), if $a_{p}<a_{b}$. In this scenario the hyper-producer receives diminishing good production as a function of cost to birth rate compared to the producer. Left panels, $a_{b}=1.3$ and $a_{p}=1.5$. Right panels, $a_{b}=3$ and $a_{p}=1.5$. Remaining parameters given in [Table\[tab:supmat\_fig\_parameter\_list\_1\]]{}. []{data-label="fig_pairwise_invade"}](fig5.pdf){width="50.00000%"}
In [@maclean_2008], $SUC2$ and $suc2$ were experimentally competed on an agar plate. It was found that non-producing $suc2$ could not invade from rare ($1\%$ of initial yeast population), and in fact decreased in frequency, becoming undetectable at long times (around 800 generations). This suggests that in a spatial setting, invertase producing $SUC2$ yeast are robust to invasions, which is in qualitative agreement with our theoretical predictions. The experiments yielded an additional result, the appearance of a hyper-producing mutant. This hyper-producing phenotype produced invertase at approximately $1.5$ times the rate of standard producers and existed at higher densities. The hyper-producer appeared to evolve naturally and establish robust colonies during the competition experiments between non-producers and producers. However, when separate competition experiments were conducted between the hyper-producers and the producers, the hyper-producers failed to demonstrate any appreciable fitness advantage over the producers. This potentially suggests an optimal invertase production rate, whereby the hyper-producers managed to establish and grow during the $SUC2$-$suc2$ competition experiments by exploiting non-producing regions due to a relative fitness advantage, but could not invade regions of space occupied by producers. Interestingly, our model also predicts that an intermediate optimal production rate may exist, depending on how the cost of production scales with the production rate. Suppose a hyper-producer, $U$, produces at a rate $p_{u}=a_{p}p_{x}$, paying a metabolic cost $a_{b}\varepsilon$ to its birth rate, such that $b_{u}=b(1-a_{b}\varepsilon)$. The pairwise invasion probabilities of each phenotype can then be calculated (see Supplementary Information,
SectionS.4). We define the fitter phenotype in a pair as that with the larger invasion probability. The potential fitness rankings are investigated in [Fig.\[fig\_pairwise\_invade\]]{} as a function of $p_{x}$ and $\varepsilon$, (which we recall also alter $p_{u}$ and $b_{u}$). We draw particular attention to the right panels, in which $a_{b}>a_{p}$. In this scenario, the hyper-producers pay a disproportionate cost for their increased production rate compared to the producers. This can be interpreted as diminishing returns for production. In this case, there exist regions where the producer is the optimal phenotype (regions (a) and (b), in blue and cyan respectively). Specifically, scenario (a) displays a similar behavior to that observed in [@maclean_2008], in which producers win out over both non-producers and hyper-producers, but hyper-producers are more likely to invade non-producing populations.
Generality of results
=====================
We have shown that demographic stochasticity can reverse the direction of selection in a public good model. In this section we will show that the mechanism responsible for this phenomenon is by no means particular to this model. We consider a general scenario, with a phenotype $X_{1}$, which is at the focus of our study, and a number of discrete ecosystem constituents, $E_{i}$. In the public good model for instance, we would label the public good itself as an ecosystem constituent, however more generally this could be a food source, a predator or anything else that interacts with the phenotypes. The state of the ecosystem influences the birth and death of the phenotype and in turn the presence of the phenotype influences the state of the ecosystem, altering the abundances of the constituents. We assume that the system lies at a unique, stable stationary state, precluding the possibility of periodic behavior. Suppose that a new phenotype, $X_{2}$, arises. We assume that the second phenotype is only slightly better at exploiting the ecosystem than $X_{1}$, though its influence on the ecosystem may be very different. For instance, in the public goods model, non-producers have a small birth-rate advantage over producers, but do not produce the public good. Which phenotype is more likely to invade and fixate in a resident population of the opposite type?
The stochastic model for this system can be constructed in a similar manner to the public good model; the dynamics are described by a set of probability transition rates (analogous to [Eq.(\[eq\_reactions\_1\])]{}). We restrict the transitions by specifying that although the two phenotypes compete, there is no reaction that instantaneously changes both of their numbers in the population. This final condition simply means that they should not, for instance, be able to mutate from one type to another during their lifetime, or to prey on each other. A parameter $R$ is introduced, to once again govern the typical scale of the system. The model is analyzed in the mesoscopic limit, by introducing $(x_{1},x_{2},\bm{e}) = (n_{x1},n_{x2},\bm{n}_{e})/R^{2}$ and applying the diffusion approximation. For large but finite $R$, the mesoscopic description takes the form $$\begin{aligned}
\dot{ x_{1} } &=& x_{1} F^{(0)}(\bm{x},\bm{e}) - \varepsilon x_{1} F^{(\varepsilon)}(\bm{x},\bm{e}) + R^{-1}\eta_{1}(t) \, , \nonumber \\
\dot{ x_{2} } &=& x_{2} F^{(0)}(\bm{x},\bm{e}) + R^{-1} \eta_{2}(t) \, , \label{eq_general_SDEs} \\
\dot{ e_{i} } &=& F_{i}(\bm{x},\bm{e}) + R^{-1} \beta_{i}(t) \,, \quad \forall \, i = 3 , \ldots J \,, \nonumber\end{aligned}$$ where $\varepsilon$ is small and governs selective pressure against $X_{1}$. The assumption that there is no reaction that instantaneously changes the number of both phenotypes ensures that the correlation structure of the noise terms takes the form $$\begin{aligned}
& \langle \eta_{1}(t) \eta_{1}(t') \rangle = \delta(t-t') x_{1} H^{(0)}(\bm{x},\bm{e}) \,,& \nonumber \\
& \langle \eta_{2}(t) \eta_{2}(t') \rangle = \delta(t-t') x_{2} H^{(0)}(\bm{x},\bm{e}) \,,& \quad \langle \eta_{1}(t) \eta_{2}(t') \rangle = 0 \nonumber \,,\end{aligned}$$ with $\varepsilon$ taken to be of order $R^{-2}$. This assumption, made here to isolate the effect of varying carrying capacity from any other intraspecies dynamics, means that while the magnitude of fluctuations in the number of both phenotypes is dependent on the state of the system, $(\bm{x},\bm{e})$, the fluctuations themselves are not correlated with each other. Restrictions on the microscopic model that yield the above SDE description are addressed more thoroughly in [Appendix\[sec:supmat\_sec\_general\_results\]]{}. The form of [Eq.(\[eq\_general\_SDEs\])]{} makes the nature of the system we describe more clear; it consists of two competing phenotypes, which reproduce according to replicator dynamics [@hofbauer_1998] with equal fitness at leading order in $\varepsilon$.
In the special case $\varepsilon=0$, both phenotypes are equally fit, regardless of their influence on the ecosystem variables $e_{i}$. The degeneracy of the dynamics in $x_{1}$ and $x_{2}$ ensures the existence of a deterministic CM. We assume that the structure of $F^{(0)}(\bm{x})$ and $F_{i}(\bm{x})$ is such that the CM is one-dimensional (there are no further degenerate ecosystem variables) and that it is the only stable state in the interior region $x_{i}>0$. A separation of timescales is present if the system collapses to the CM much faster than the stochastic dynamics. In practical terms, the timescale of collapse can be inferred as the inverse of the non-zero eigenvalues of the system, linearised about the CM [@constable_2013], while the timescale of fluctuations will be of order $R^{-2}$ [@constable_2014_phys]. When $\varepsilon>0$, the timescale elimination procedure can still be applied if $\varepsilon\approx\mathcal{O}\left( R^{-2} \right)$ . The effective one-dimensional description of the system now takes the form $$\dot{x_{1}} = - \varepsilon \, \mathcal{D}(x_{1}) + R^{-2} \,\mathcal{S}(x_{1}) + R^{-1} \zeta(t) \,, \label{eq_general_reduced_PG_SDE}$$ where the term $ \mathcal{D}(x_{1})$ is the deterministic contribution to the effective dynamics and $\mathcal{S}(x_{1})$ is the stochastic contribution, while $\zeta(t)$ is an effective noise term. The form these functions take is dependent on $F^{(0)}(\bm{x},\bm{e})$, $F^{(\varepsilon)}(\bm{x},\bm{e})$ and $F_{i}(\bm{x},\bm{e})$, as well as the noise correlation structure, $H^{(0)}(\bm{x},\bm{e})$; however it is independent of the structure of the demographic noise acting on the ecosystem variables (see Eqs.(\[eq:supmat\_eq\_general\_Abar\_det\]), (\[eq:supmat\_eq\_Abar\_stoch\_2\]) and (\[eq:supmat\_eq\_general\_Bbar\])).
The core assumption we have made to derive [Eq.(\[eq\_general\_reduced\_PG\_SDE\])]{} is essentially that the system’s ecological processes act on a faster timescale than its evolutionary processes. Even in this general setting, insights about the system’s stochastic dynamics can still be drawn (see [Appendix\[sec:supmat\_sec\_general\_results\]]{}). If $\varepsilon=0$, the fixation probability of phenotype $X_{1}$ is independent of the initial conditions of the ecosystem variables $\bm{e}$. In fact it is equal to the initial fraction of $X_{1}$ in the population, $ n_{10} / (n_{10} + n_{20} )$. The invasion probability of mutant $X_{1}$ phenotype fixating in a resident $X_{2}$ population however depends on the stationary state of the $X_{2}$ population; this defines the initial invasion conditions (the denominator for the fixation probability of $X_{1}$). Denoting by $N_{1}$ and $N_{2}$ the average numbers of phenotypes $X_{1}$ and $X_{2}$ in their respective stationary states, we find $\phi_{1} = 1/N_{2}$ and $\phi_{2} = 1/N_{1}$, generalizing [Eq.(\[eq\_phi\_quasineutral\])]{}. Therefore, for $\varepsilon=0$ the phenotype that exists at higher densities is more likely to invade and fixate than its competitor, a consequence of its robustness to invasions. This result holds for any choice of finite $R$. In an ensemble of disconnected populations subject to repeated invasions, we would observe the emergence of high density phenotypes if this phenotype does not carry a cost. While this seems like a reasonable and indeed natural conclusion, it is one entirely absent from the deterministic analysis.
If $\varepsilon>0$, general results for the phenotype fixation probabilities cannot be obtained. However, if $N_{1}>N_{2}$, in the limit $\varepsilon\rightarrow0$ we have shown that $\phi_{1}>\phi_{2}$. From this, it can be inferred that the term $\mathcal{S}(x_{1})$ is positive on average along the slow manifold (see [Eq.(\[eq:supmat\_integral\_S\])]{}). Therefore, if phenotype $X_{1}$ exists at higher densities in isolation than phenotype $X_{2}$, there will exist a stochastically-induced pressure favoring the invasion of phenotype $X_{1}$. Meanwhile, by construction we expect the form of $\mathcal{D}(x_{1})$ to be positive, since phenotype $X_{1}$ exploits the ecosystem environment less effectively than phenotype $X_{2}$. There is therefore a trade-off for competing phenotypes between increasing their phenotype population density and increasing their per capita growth rate. Note that the noise-induced selection function $\mathcal{S}(x_{1})$ need not be strictly positive; indeed it may become negative along regions of the SM. This potentially allows for stochastically induced ‘fixed points’ along the SM, around which the system might remain for unusually large periods of time. This may provide a theoretical understanding of the coexistence behavior observed in [@behar_2015].
The term $\mathcal{S}(x_{1})$ is moderated by factor $R^{-2}$ (see [Eq.(\[eq\_reduced\_PG\_SDE\])]{}), or more physically, the typical size of the population. The stochastically induced selection for the high-density phenotype therefore becomes weaker as typical system sizes increase. The trade-off will be most crucial in small populations, or as illustrated in the public good model, systems with a spatial component. If the phenotypes and ecosystem variables move sufficiently slowly in space, the results of Eqs. (\[eq\_pi\]) and (\[eq\_weakening\_threshold\]) can be imported, with the understanding that $\phi_{1}$ and $\phi_{2}$ must be calculated for the new model under consideration.
It is worth noting that the precise functional form of $\phi_{1}$ and $\phi_{2}$ identified in the deterministically neutral case ($\varepsilon=0$) is dependent on the assumption that phenotype noise fluctuations are uncorrelated. While correlated fluctuations (for instance resulting from mutual predation of the phenotypes) can still be addressed with similar methods to those employed here, there is then the potential for the emergence of further noise-induced selection terms (see [Appendix\[sec:supmat\_sec\_comp\_models\]]{}). Careful specification of the phenotype interaction terms is therefore needed to determine to what degree these additional processes might amplify or dampen the induced selection we have identified.
Discussion
==========
In this paper, we have shown that stochastic effects can profoundly alter the dynamics of systems of phenotypes that change the carrying capacity of the total population. Most strikingly, selection can act in the opposite direction from that of the deterministic prediction if the phenotype that is deterministically selected for also reduces the carrying capacity of the population. The methods used to analyze the models outlined in the paper are based on the removal of fast degrees of freedom [@parsons_rogers_2015]. The conclusions drawn are therefore expected to remain valid as long as the rate of change of the phenotype population composition occurs on a shorter timescale than the remaining ecological processes.
By illustrating this phenomenon in the context of public good production, we have revealed a mechanism by which the dilemma of cooperation can be averted in a very natural way: by removing the unrealistic assumptions of fixed population size inherent in Moran-type game theoretic models. The potential for such behavior has been previously illustrated with the aid of a modified Moran model [@houch_2012] and a single variable Wright-Fisher type model [@houch_2014] that assumes discrete generations. However we have shown that the mechanism can manifest more generally in multivariate continuous time systems. Our analysis may also provide a mathematical insight into the related phenomenon of fluctuation-induced coexistence that has been observed in simulations of a similar public good model featuring exogenous additive noise [@behar_2015]: such coexistence may rely on a similar conflict between noise-induced selection for producing phenotypes and deterministic selection against them.
For biologically reasonable public good production costs, selection reversal is only observed in systems that consist of a very small number of individuals. However, by building a metapopulation analogue of the model to account for spatial structure, the range of parameters over which selection reversal is observed can be dramatically increased, so long as public good diffusion and phenotype dispersal between populations are not large. Two distinct mechanisms are responsible for these results. First, including spatial structure allows for small, local effective population sizes, even as the total size of the population increases. This facilitates the stochastic effects that lead to selection reversal. Second, since producer populations tend to exist at greater numbers (or higher local densities) they produce more migrants. The stochastic advantage received by producers is thus amplified, as not only are they more robust stochastically to invasions, but also more likely to produce invaders. Away from the low-dispersal, zero public good diffusion limit, the effect of selection reversal is diminished, but is still present across a range of biologically reasonable parameters. The analytical framework we have outlined may prove insightful for understanding the simulation results observed in [@behar_2014], where a similar metapopulation public good model was considered. In addition to fixation of producers (in the low dispersal-diffusion limit) and fixation of non-producers (in the high dispersal-diffusion limit), [@behar_2014] observed an intermediate parameter range in which noise induced coexistence was possible. Though our model does not feature such a regime, extending our mathematical analysis to their model would be an interesting area for future investigations. However it must be noted that coexistence in a stochastic setting is inherently difficult to quantify analytically, as for infinite times some phenotype will always go extinct.
That space can aid the maintenance of cooperation is well known [@tarnita_review_2009; @wakano_2011]. Generally, however, this is a result of spatial correlations between related phenotypes, so that cooperators are likely to be born neighboring other cooperators (and share the benefits of cooperation) while defectors can only extract benefits at the perimeter of a cooperating cluster. This is not what occurs in the model presented in this paper. Indeed, while we have assumed in our analytic derivation of the invasion probability that dispersal is small enough that each patch essentially contains a single phenotype, we find that the phenomenon of selection reversal manifests outside this limit (see [Appendix\[sec:supmat\_sec\_movies\]]{}, movie S2 in which a majority of patches contain a mix of producers and non-producers). Instead, producing phenotypes have a selective advantage due to the correlation between the fraction of producers on a patch and the total number of individuals on a patch, which provides both resistance to invasions and an increased dispersal rate.
Most commonly in spatial game theoretic models of cooperation-defection, individuals are placed at discrete locations on a graph [@nowak_1992; @allen_2013]. In contrast, by using a metapopulation modelling framework we have been able to capture the effect of local variations in phenotype densities across space, which is the driver of selection amplification in our model. Nevertheless, the question that remains is which modelling methodology is more biologically reasonable. This clearly depends on the biological situation. However, in terms of test-ability, our model makes certain distinct predictions. In [@allen_2013], producers and non-producers were modeled as residing on nodes of a spatial network, with a public good diffusing between them. The investigation concludes that both lower public good diffusion and lower spatial dimensions (e.g. systems on a surface rather than in a volume) should encourage public good production, essentially by limiting the ‘surface area’ of producing clusters. While our investigation certainly predicts that lower public good diffusion is preferable, stochastic selection reversal does not require that the spatial dimension of the system is low. In fact the result utilized in [Eq.(\[eq\_regular\_graph\_fix\])]{} holds for patches arranged on any regular graph (where each vertex has the same number of neighbors), and thus could be used to describe patches arranged on a cubic, or even hexagonal, lattice.
In our final investigation, we have shown that stochastic selection reversal is not an artifact of a specific model choice, but may be expected across a wide range of models. These models consist of two phenotypes, competing under weak deterministic selection strength, reproducing according to replicator dynamics and interacting with their environment. Thus the phenomenon of selection reversal is very general; however, it depends strongly on how one specifies a selective gradient. We take one phenotype to have a stochastic selective advantage over the other if a single mutant is more likely to invade a resident population of the opposite type. Such a definition is also used in standard stochastic game theoretic models [@nowak_2006]. A key difference here however (where the population size is not fixed) is that the invasion probability is not specified by a unique initial condition; we must also specify the size of the resident population. We have assumed that the invading mutant encounters a resident population in its stationary state. This is by no means an unusual assumption; it is the natural analogue of the initial conditions in a fixed population size model. Essentially it assumes a very large time between invasion or mutation events, an approach often taken in adaptive dynamics [@waxman_2005].
If instead we assumed a well-mixed system far from the steady state, our results would differ. For instance, suppose the system initially contains equal numbers of the two phenotypes. For the case when the two phenotypes have equal reproductive rates ($\varepsilon=0$), the phenotypes have equal fixation probability. For $\varepsilon>0$, the phenotype with the higher birth rate has the larger fixation probability, regardless of its influence on the system’s carrying capacity. This apparent contradiction with the results we developed in the body of the paper echos the observations of $r-K$ selection theory [@pianka_1970]: selection for higher birth rates ($r$-selection) acts on frequently disturbed systems that lie far from equilibrium, while selection for improved competitive interactions or carrying capacities ($K$-selection) acts on rarely disturbed systems. In addition, $r-K$ selection theory suggests that $K$-selected species are typically larger in size and, as a consequence, consist of a lower number of individuals [@reznick_2002]. This indicates a further parallel with our stochastic model framework, since selection for higher carrying capacities requires that the typical number of individuals (of both the low and high carrying capacity phenotypes) is small. Though the mechanism that leads us to these conclusions is distinct, our stochastic analysis provides a complementary view of $r-K$-selection theory, which may be applicable to simple microorganisms. In exploring this analogous behavior further, future investigations may also benefit from considering the results of [@parsons_quince_2010], where it was shown that stochastically induced selection can change direction near carrying capacity.
Although we have implicitly developed our results in the low mutation limit, including mutation explicitly in the modeling framework is possible. This would be an interesting extension to the framework. In the well-mixed scenario, it is likely that the inclusion of mutation will complicate the intuition developed here: while larger populations are more robust to invasions, they are also more prone to mutations, by virtue of their size. While this may be offset by the additional benefits garnered in the spatial analogue of the model, a complex set of timescale-dependent behaviors is likely to emerge.
Finally, we propose a rigorous analytical investigation of existing models that conform to the framework we have outlined; an example is the work conducted in [@behar_2014; @behar_2015], which we believe to be mathematically explainable within our formalism. In the context of induced selection, whereby deterministically neutral systems become non-neutral in the stochastic setting, similar ideas have already been extended to disease dynamics [@kogan_2014] and the evolution of dispersal [@lin_mig_1; @lin_mig_2]. The extension of selection reversal to such novel ecological models may provide further insight. Furthermore, this general scheme may be of relevance to many other systems in ecological and biological modeling, such as cancer, for which the evolution of phenotypes that profoundly alter cell carrying capacity can be of primary importance.
TR acknowledges funding from the Royal Society of London.
Obtaining the SDE system from the microscopic individual based model {#sec:supmat_sec_obtain_SDE_system}
====================================================================
We begin with a model consisting of a discrete number of entities, two phenotypes of a species, $X$ and $Y$ and a public good $Q$. They interact according to the transitions $$\begin{aligned}
& X \overset{b_{x}}{\underset{\kappa/R^{2}}{ {\mathrel{\substack{\xrightarrow{\rule{0.4cm}{0cm}} \\[-.9ex] \xleftarrow{\rule{0.4cm}{0cm}}}}} } } X+X\,,\quad Y + X \xrightarrow{\kappa/R^{2}} X \,, & \nonumber \\
& Y \overset{b_{y}}{\underset{\kappa/R^{2}}{ {\mathrel{\substack{\xrightarrow{\rule{0.4cm}{0cm}} \\[-.9ex] \xleftarrow{\rule{0.4cm}{0cm}}}}} } } Y+Y\,,\quad Y + X \xrightarrow{\kappa/R^{2}} Y \,, & \,, \nonumber \\
& X+Q\xrightarrow{r/R^{2} } X+X+Q\,,\quad Y+Q\xrightarrow{r/R^{2} } Y+Y+Q&\,,\nonumber \\
& X\xrightarrow{p_{x}}X+Q\,, \quad Q\xrightarrow{\delta}\varnothing \,. &\label{supmat_eq_reactions}\end{aligned}$$ The term $R^{-2}$ occurs in all terms involving two reactants. It thus controls the interaction probability between instances of the phenotypes and the public good. Taking larger $R$ decreases the interaction probability of phenotypes $X$ and $Y$ and the public good and allowing the populations to grow to greater numerical abundances. The parameter $R$ can thus be understood as a measure of the spatial scale of the system; when $R$ is increases, the probability of interactions in the well-mixed system is decreased while the number of individuals the system can contain is increased.
Let us denote $\bm{n}=(n_{x}$, $n_{y}$, $n_{q})$ the numbers of $X$, $Y$ and $Q$ respectively. Then the dynamics of this system can be described by the set of partial difference equations $$\frac{d P(\bm{n},t)}{d t} = \sum_{\bm{n}'\neq \bm{n}} \left[ T(\bm{n}|\bm{n}')P(\bm{n}',t) - T(\bm{n}'|\bm{n})P(\bm{n},t) \right] \, ,\label{eq:supmat_meqn_general}$$ where $P(\bm{n},t)$ is the probability of the state being in state $\bm{n}$ at time $t$, and $T(\bm{n}'|\bm{n})$, the probability transition rate, is the probability per unit time of transitioning from state $\bm{n}$ to $\bm{n}'$. Formally this is known as the master equation [@van_kampen_2007]. Given the reactions [Eq.(\[supmat\_eq\_reactions\])]{} the probability transition rates can be expressed as $$\begin{aligned}
&T_1(n_x + 1,n_y,n_{q}|n_x,n_y,n_{q}) = b_{x}n_{x} + \dfrac{ r }{ R^{2} } n_{x} n_{q} \, ,& \nonumber \\
&T_2(n_x,n_y + 1,n_{q}|n_x,n_y,n_{q}) = b_{y}n_{y} + \frac{ r }{ R^{2} } n_{y} n_{q} \, ,& \nonumber \\
&T_3(n_x - 1,n_y,n_{q}|n_x,n_y,n_{q}) = \dfrac{ \kappa }{ R^{2} } n_{x} \left( n_{x} + n_{y} \right) \,, & \nonumber \\
&T_4(n_x,n_y - 1,n_{q}|n_x,n_y,n_{q}) = \frac{ \kappa }{ R^{2} } n_{y} \left( n_{x} + n_{y} \right) \,, & \nonumber \\
&T_5(n_x,n_y,n_q + 1|n_x,n_y,n_{q}) = p_{x} n_{x} \,, & \nonumber \\
&T_6(n_x,n_y,n_{q}-1|n_x,n_y,n_{q}) = \delta n_{q} \,. \label{supmat_eq_trans}\end{aligned}$$
Let us now make a change of variables into the scaled expressions $\bm{x}=(x,y,q)=(n_{x},n_{y},n_{q})/R^{2}$. Substituting the probability transition rates into [Eq.(\[eq:supmat\_meqn\_general\])]{}, we find recurrent factors of $1/R^{2}$ appearing in the resulting expression. These terms are associated with the local transitions from state $\bm{n}$ to the surrounding states. If $R^{2}$ is sufficiently large, the population grows larger (as the crowding terms in [Eq.(\[supmat\_eq\_reactions\])]{} grow small). We may then Taylor expand [Eq.(\[eq:supmat\_meqn\_general\])]{} in $R^{-1}$, assuming that the variables $(x,y,q)$ are approximately continuous [@gardiner_2009]. Truncating at second order in $R^{-4}$, we arrive at a partial differential equation for $p(x,y,q,t)$ of the form $$\begin{aligned}
\frac{\partial p(\bm{x},t)}{\partial t} = &-& \frac{1}{R^{2}} \sum_{i} \frac{\partial}{\partial x_{i}} \left[A_{i}(\bm{x})p(\bm{x},t)\right] \nonumber \\ &+& \frac{1}{ 2 R^{4} }\sum_{i,j} \frac{\partial^{2}}{\partial x_{i} \partial x_{j}} \left[B_{ij}(\bm{x})p(\bm{x},t)\right] \, , \label{eq:supmat_eq_FPE_general} \\
\bm{x} &=& (x_{1},x_{2},x_{3}) \equiv (x,y,q) \,. \nonumber\end{aligned}$$ This is a diffusion approximation in a population genetics context [@crow_kimura_into], but more generally is akin to the Kramers-Moyal expansion [@gardiner_2009] or a nonlinear analogue of the van Kampen expansion [@van_kampen_2007]. The forms of $\bm{A}(\bm{x})$ and $B(\bm{x})$, given transition rates [Eq.(\[supmat\_eq\_trans\])]{} are found to be $$\begin{aligned}
A_{x}(\bm{x}) &=& x \left( b_{x} + r q - \kappa ( x + y ) \right) \,, \nonumber \\
A_{y}(\bm{x}) &=& y \left( b_{y} + r q - \kappa ( x + y ) \right) \,, \nonumber \\
A_{q}(\bm{x}) &=& p_{x} x - \delta q \,, \label{supmat_eq_A_supMat}\end{aligned}$$ and $$\begin{aligned}
&B_{xx}(\bm{x}) = x \left( b_{x} + r q + \kappa x + \kappa y \right) \,, &\nonumber \\
&B_{yy}(\bm{x}) = y \left( b_{y} + r q + \kappa x + \kappa y \right) \,, &\nonumber \\
&B_{qq}(\bm{x}) = p_{x} x + \delta q \,, &\nonumber \\
&B_{ij} = 0 \quad \forall \quad i \neq j \,. & \label{supmat_eq_B_supMat}\end{aligned}$$ Further, it can be shown that the above PDE is equivalent to the set of Itō SDEs [@risken_1989] $$\frac{ \mathrm{d} \bm{x} }{ \mathrm{d} \tau} = \bm{A}(\bm{x}) + \frac{1}{R} \bm{\eta}(\tau) \,, \label{supmat_eq_SDE_very_general}$$ where $ \tau = t R^{2}$ and $\bm{\eta}(t)$ are Gaussian white noise terms with zero mean and correlations $$\langle \eta_{i}(\tau) \eta_{j}(\tau') \rangle = \delta( \tau - \tau' ) B_{ij}(\bm{x}) \,. \label{supmat_eq_correlations}$$ Notice that the correlations are multiplicative and thus dependent on the state of the system.
Obtaining one-dimensional effective public good model {#supmat_sec_public_good_reduction}
=====================================================
In this section we seek to identify and remove the fast-modes of the SDE system [Eq.(\[supmat\_eq\_SDE\_very\_general\])]{}, and thus obtain an effective one-dimensional description of the dynamics. We make use of methods of fast-mode elimination described in [@parsons_rogers_2015]. Firstly we note that the deterministic nullcline for $q$ is given by $$q = \frac{ p_{x} x }{ \delta } \equiv Z_{q}(x,y) \,. \label{supmat_eq_q_cm}$$ Therefore, if the production and decay of public good occur much faster than the processes associated with the phenotypes, we would expect the public good to quickly attain this value, after which its dynamics would be slaved to those of $x$ and $y$. Notice that deterministically, substituting [Eq.(\[supmat\_eq\_q\_cm\])]{} into [Eq.(\[supmat\_eq\_A\_supMat\])]{} recovers a Lotka-Volterra competition model for two competing species.
To make further analytic progress, we begin by considering the quasi-neutral limit in which $b_{x}=b_{y} \equiv b$. Under these conditions, the deterministic system exhibits a center manifold (CM) given by [Eq.(\[supmat\_eq\_q\_cm\])]{} and $$y = \frac{ \left[ b \delta - (\delta \kappa - r p_{x} ) x \right] }{ \delta \kappa} \equiv Z_{y}(x) \,. \label{supmat_eq_y_cm}$$ The CM is stable for $ \kappa \delta > r p_{x}$, and we assume that this condition holds throughout the paper. Calculating the intersection of the center manifold at the boundaries $y=0$ and $x=0$ allows us to determine the mean population size in the quasi-neutral ($\varepsilon=0$) limit when it consists of only producers and non-producers respectively; $$\begin{aligned}
N_{x}^{(0)} = R^{2} K_{x}^{(0)} \,, \quad K_{x}^{(0)} = \left( \frac{ b \delta }{ \delta \kappa - r p_{x} } \right) \,, \\
N_{y}^{(0)} = R^{2} K_{y}^{(0)} \,, \quad K_{y}^{(0)} = \left( \frac{ b \delta }{ \delta \kappa - r p_{y} } \right) \,. \label{eq:supmat_eq_NX_NY_0}\end{aligned}$$ These parameters will be useful in the following analysis.
Deterministically, the system comes to rest on a point along the CM (defined by Eqs. (\[supmat\_eq\_q\_cm\]) and(\[supmat\_eq\_y\_cm\])), which depends on the system’s initial conditions. When stochasticity is included, the CM ceases to exist in any true sense. However, when the noise is small (already assumed in the derivation of SDEs (\[supmat\_eq\_SDE\_very\_general\])) we can say that far from the CM, we expect the dynamics to be dominated by the deterministic collapse to the CM, while in the vicinity of the CM, we expect noise to play a more important role, driving the slow change in population composition until one or other of the phenotypes fixates. We wish to exploit this timescale separation, and obtain an effective description of the dynamics in terms of a single variable.
To begin, we note that the stochastic dynamics along the CM has two components. First, noise can move the system neutrally *along* the CM. Second, noise can take the system *off* the CM, at which point we expect the deterministic component of the dynamics to become more prevalent, driving the system back to the CM. In order to capture the effect of both of these processes on the effective dynamics along the CM, we implement a non-linear projection of the stochastic system to the CM. Essentially this assumes that fluctuations which take the system away from the manifold are instantaneously mapped along deterministic trajectories back to the CM. In order to formalize this, the mapping $z=f(x,y,q)$ is introduced, where $f(x,Z_{y}(x),Z_{q}(x))=x$; that is $z$ gives the position on the CM, parameterized by $x$, which intersects a deterministic trajectory beginning at $(x,y,q)$. The mapping can be determined analytically from the observation that the quantity $x/y$ in [Eq.(\[supmat\_eq\_SDE\_very\_general\])]{} is invariant in this quasi-neutral ($b_{x}=b_{y}$) scenario. Therefore $$\frac{z}{Z_{y}(z)} = \frac{x}{y} \,, \quad z = \frac{ b \delta x }{ (\delta \kappa - p r ) x + \delta \kappa y } \,. \label{supmat_eq_PG_define_z}$$
The effective dynamics for $z$ can now be straightforwardly calculated by differentiating [Eq.(\[supmat\_eq\_PG\_define\_z\])]{} with respect to $t$. One must note however that since the original SDE system is defined in the Itō sense, the normal rules of calculus no longer apply. Applying Itō’s rules of calculus appropriately [@van_kampen_2007; @parsons_rogers_2015], we find that the effective dynamics along the CM take the following form $$\dot{z} = \frac{1}{R^{2}}\mathcal{S}( z ) + \frac{1}{R} \zeta (t) \,, \label{supmat_eq_quasineutral_effective_SDE}$$ where $$\begin{aligned}
\mathcal{S}( z ) &=& \frac{1}{2} \left( \frac{\partial^{2} z }{\partial x^{2} } B_{xx}(\bm{x}) + \frac{\partial^{2} z }{\partial y^{2} } B_{yy}(\bm{x}) \right)|_{x=z,y=Z_{y}(z),q=Z_{q}(z)} \,, \label{supmat_eq_PG_general_Abar} \\
&=& \frac{ 2 p_{x} r }{ \delta } z \left\{ 1 + \frac{1}{ b^{2} \delta^{2} } z \left[ b \delta \left( 2 p_{x} r - \delta \kappa \right) + p_{x} r \left( p_{x} r - \delta \kappa \right) z \right] \right\} \,, \nonumber \\
&=& 2 b \left( \frac{ K_{x}^{(0)} - K_{y}^{(0)} }{ (K_{x}^{(0)})^{3} (K_{y}^{(0)})^{2} } \right) z \left( K_{x}^{(0)} - z \right) \left[ K_{x}^{(0)} K_{y}^{(0)} \right. \nonumber \\ &+& \left. \left( K_{x}^{(0)} - K_{y}^{(0)} \right) z \right] \,, \label{supmat_eq_PG_clean_Abar}\end{aligned}$$ and $$\begin{aligned}
\langle \zeta(t) \rangle &=& 0 \,, \quad \langle \zeta(t) \zeta(t') \rangle = \delta( t - t' ) \mathcal{B}(z) \,, \nonumber \end{aligned}$$ with $$\begin{aligned}
\mathcal{B}(z) &=&\left(\left[ \frac{\partial z }{\partial x } \right]^{2}B_{xx}(\bm{x}) + \left[ \frac{\partial z }{ \partial y }\right]^{2}B_{yy}(\bm{x}) \right)|_{x=z,y=Z_{y}(z),q=Z_{q}(z)} \,, \label{supmat_eq_PG_general_Bbar}\\
&=& 2 z \left\{ b + \frac{1}{ b^{2} \delta^{3} } z \left[ b^2 \delta^2 \left( 3 p_{x} r - \delta \kappa \right)
\right. \right. \nonumber \\ &+& \left. \left. b p_{x} r \delta \left( 3 p_{x} r - 2 \delta \kappa \right) z
p_{x}^2 r^2 \left( p_{x} r - \delta \kappa \right)z^{2} \right] \phantom{\frac{1}{2}} \right\} \,, \nonumber \\
&=& 2 b \left( \frac{ 1 }{ (K_{x}^{(0)})^{3} (K_{y}^{(0)})^{2} } \right) z \left( K_{x}^{(0)} - z \right) \left[ K_{x}^{(0)} K_{y}^{(0)} \right. \nonumber \\ &+& \left. \left( K_{x}^{(0)} - K_{y}^{(0)} \right) z \right]^{2} \,. \label{eq:supmat_eq_PG_clean_Bbar}\end{aligned}$$ Notice that since the mapping [Eq.(\[supmat\_eq\_PG\_define\_z\])]{} is independent of $q$, both [Eq.(\[supmat\_eq\_PG\_general\_Abar\])]{} and [Eq.(\[supmat\_eq\_PG\_general\_Bbar\])]{} do not depend on the noise correlations in $q$.
![Figure illustrating the origin of stochastically induced drift along the center manifold (CM). The gray dashed line shows the form of the deterministic center manifold, which intersects the $x$ axis at a higher value than the $y$ axis (phenotype $X$ has a higher carrying capacity due to the production of the public good). The red shaded circle illustrates the form of the Gaussian noise centered on the point $\bm{x}^{(0)}$ on the CM. Fluctuations in the population are equally likely to increase or decrease the frequency of the $Y$ phenotype to the points $\bm{x}^{(1)}$. Away from the CM, the deterministic pressure to the CM becomes prominent, forcing the system along quasi-deterministic trajectories back to the CM, at the points $\bm{x}^{(2)}$. The resulting distribution of $\bm{x}^{(2)}$ does not have a mean centered on $\bm{x}^{(0)}$. Rather, the distribution is shifted, inducing a drift in favor of the producing $X$ phenotype. []{data-label="fig:supmat_fig_projection_diagram"}](figS1.pdf){width="45.00000%"}
While the deterministic system features no dynamics along the CM, the effective SDE (\[supmat\_eq\_quasineutral\_effective\_SDE\]) does feature a drift in the mean state, embodied by $\mathcal{S}( z )$. Understanding the origin of this induced drift term requires considering the following. We envisage fluctuations arising from a single point on the CM, $\bm{x}^{(0)}$, which take to the system to a point off the CM, $\bm{x}^{(1)}$ (see [Fig.\[fig:supmat\_fig\_projection\_diagram\]]{}). The point $\bm{x}^{(1)}$ is clearly stochastic, but its distribution is approximately Gaussian, with a variance defined by $B(\bm{x}^{(0)})$. The fluctuation is now mapped back along a deterministic trajectory to a point $\bm{x}^{(2)}$ on the CM. The location $\bm{x}^{(2)}$ is also stochastic (dependent as it is on $\bm{x}^{(1)}$), and has its own distribution. The presence of the term $\mathcal{S}(z)$ in [Eq.(\[supmat\_eq\_quasineutral\_effective\_SDE\])]{} is indicative of the fact that the mean of the distribution of $\bm{x}^{(2)}$ is not $\bm{x}^{(0)}$; fluctuation events on average are mapped back to the CM with a preferred direction, inducing drift along the CM. Note that $\mathcal{S}(z)$ is positive along the length of the CM, which is defined on the interval $[0,K^{(0)}_{x}]$.
We now turn our attention to the case when $\varepsilon > 0 $. So long as $\varepsilon$ is small, a separation of timescales is still present, though now no center manifold exists. Instead there is a slow manifold (SM), to which the deterministic system quickly relaxes, before slowly moving along it until phenotype $Y$ fixates. The equations for the population size at the boundaries of the SM are formally given by $$\begin{aligned}
N_{x} = R^{2} K_{x} \,, \quad K_{x} = \left( \frac{ b_{x} \delta }{ \delta \kappa - r p_{x} } \right) \equiv K^{(0)}_{x} + \mathcal{O}(\varepsilon) \,, \nonumber \\
N_{y} = R^{2} K_{y} \,, \quad K_{y} = \left( \frac{ b_{y} \delta }{ \delta \kappa - r p_{y} } \right) \equiv K^{(0)}_{x} + \mathcal{O}(\varepsilon) \,. \nonumber \\ \label{eq:supmat_eq_NX_NY}\end{aligned}$$ In order to proceed with the stochastic calculation, we assume $\varepsilon \approx \mathcal{O}(R^{-2})$, and work order by order in $R^{-1}$. At leading order, the equation for the SM is identical to that of the CM, Eqs. (\[supmat\_eq\_q\_cm\]) and (\[supmat\_eq\_y\_cm\]). The mapping to the SM is also unchanged at leading order from the quasi-neutral case (see [Eq.(\[supmat\_eq\_PG\_define\_z\])]{}). We proceed as before to obtain an effective description of the system dynamics in terms of $z$ [@parsons_rogers_2015], now obtaining the dynamics, $$\dot{z} = - \varepsilon \mathcal{D}( z ) + \frac{1}{R^{2}}\mathcal{S}( z ) + \frac{1}{R} \zeta (t) \,. \label{supmat_eq_nonneutral_effective_SDE}$$ where $$\begin{aligned}
\mathcal{D}( z ) &=& - \left( \frac{ d z }{ d x } A_{x}(\bm{x}) + \frac{ d z }{ d y } A_{y}(\bm{x}) \right)|_{x=z,y=Z_{y}(z),q=Z_{q}(z)} \,, \nonumber \\
&=& b z \left[ 1 - \left(\frac{\delta \kappa - p_{x} r }{ b \delta }\right) z \right]\,, \nonumber \\
&=& \frac{b}{K_{x}^{(0)}} z \left( K_{x}^{(0)} - z \right) \,, \label{supmat_eq_PG_clean_AbarDet}\end{aligned}$$ and $\mathcal{S}(t)$ and $\zeta(t)$ retain their form from the quasi-neutral case, Eqs. (\[supmat\_eq\_PG\_clean\_Abar\]) and (\[eq:supmat\_eq\_PG\_clean\_Bbar\]). The function $\mathcal{D}(z)$ is the deterministic contribution to the dynamics along the SM. This expression is that which would be obtained using standard fast variable elimination techniques on the deterministic system. From [Eq.(\[supmat\_eq\_PG\_clean\_AbarDet\])]{}, we can see that $\mathcal{D}(z)$ is positive along the length of the SM and therefore acts (as we would expect) to increase the selective advantage of the non-producers, phenotype $Y$. There is therefore a conflict between the two components of the drift in the system. The term $\mathcal{D}(z)$ works against producers along the length of the SM, while $\mathcal{S}(z)$ creates a selective pressure in favor of producers. Ultimately, which term is more prevalent is dependent on the parameters $\varepsilon$ and $R$ (see [Eq.(\[supmat\_eq\_nonneutral\_effective\_SDE\])]{}); small $R$ leads to a small population size in which stochastic effects are stronger, and so producers are more likely to be selected for. In contrast, when the deterministic cost for good production is increased, the non-producers have an increased advantage over producers.
Adopting the notation used in the main text, in which we set $z=x$ (which is valid on the CM and SM at leading order), the expression for the SDE (\[supmat\_eq\_nonneutral\_effective\_SDE\]) can alternatively be written $$\dot{x} = \frac{b}{K_{x}^{(0)}} x \left( K_{x}^{(0)} - x \right) \left( \frac{ 1 }{ R^{2} } \,\mathcal{F}(x) - \varepsilon \right) + \frac{1}{R} \zeta(t) \,, \label{eq:supmat_eq_PG_maintext_effective}$$ where $$\mathcal{F}(x) = 2 \left( \frac{ K_{x}^{(0)} - K_{y}^{(0)} }{ (K_{x}^{(0)})^{2} (K_{y}^{(0)})^{2} } \right) \left[ K_{x}^{(0)} K_{y}^{(0)} + \left( K_{x}^{(0)} - K_{y}^{(0)} \right) x \right] \,.\label{eq:supmat_eq_Fx}$$
Probability of fixation for the reduced public good model {#sec:supmat_sec_fixprob}
===========================================================
The fixation probability for a phenotype in a single variable system can be calculated using standard methods [@gardiner_2009]. In order to conduct the calculation, we need expressions for the absorbing boundaries of the problem. For the reduced system given in [Eq.(\[supmat\_eq\_nonneutral\_effective\_SDE\])]{}, these lie at $z=0$ and $z=K^{(0)}_{x}$. The fact that the boundary for the problem exists at $z=K^{(0)}_{x}$, rather than $z=K_{x}$, is a consequence of the order to which we are working in $\varepsilon$. At this order the SM is approximated by the expression for the CM, which intersects the absorbing boundaries $x=0$ and $y=0$ at $z=0$ and $z=K^{(0)}_{x}$ respectively. Denoting $Q(z_{0})$ the fixation probability of producing phenotype $X$ given an initial frequency $z_{0}$ on the CM/SM, the fixation probability can be conveniently be expressed $$\begin{aligned}
Q(z_{0}) &=& \frac{ \int_{z=0}^{z_{0}} \psi(z) dz }{ \int_{ z=0 }^{ K_{x} } \psi(z) dz } \,, \nonumber \\
\psi(z) &=& \exp \left[ \int_{0}^{z} \frac{2 ( - \varepsilon R \mathcal{D}( z' ) + \mathcal{S}( z' )) }{ \mathcal{B}(z') } dz' \right] \,.\end{aligned}$$ Substituting for $\mathcal{D}(z)$, $\mathcal{S}(z)$ and $\mathcal{B}(z)$ from Eqs. (\[supmat\_eq\_PG\_clean\_AbarDet\]), (\[supmat\_eq\_PG\_clean\_Abar\]) and (\[eq:supmat\_eq\_PG\_clean\_Bbar\]), we find $$\begin{aligned}
Q(z_{0}) &=& \frac{1 - G(z_{0})}{ 1 - G(K_{x}) } \, , \nonumber \\
G(z_{0}) &=& \exp \left[ \frac{ \left( \varepsilon N_{y}^{(0)} K_{x}^{(0)} z_{0} \right) }{ \left( K_{x}^{(0)} K_{y}^{(0)} + (K_{x}^{(0)} - K_{y}^{(0)}) z \right) } \right] \,. \label{supmat_eq_PG_fixprob_z}\end{aligned}$$
The nature of these expressions can be understood more intuitively if we move from considering the initial frequency of $X$ on the CM, $z_{0}=n_{x0}/R^{2}$, to considering the initial fraction of phenotype $X$ on the CM, $f_{z0}$. The fraction and number of phenotype $X$ on the CM are related by $$\begin{aligned}
f_{z0} &=& \frac{ z }{ z + Z_{y}(z) } \,, \nonumber \\
z &=& \frac{ K_{x}^{(0)} K_{y}^{(0)} f_{z0} }{ K_{x}^{(0)} - ( K_{x}^{(0)} - K_{y}^{(0)} ) f_{z0} } \,. \end{aligned}$$ Substituting this into [Eq.(\[supmat\_eq\_PG\_fixprob\_z\])]{}, we find $$Q(f_{z0}) = \frac{1 - \exp \left[ \varepsilon N_{y}^{(0)} f_{z0} \right] }{ 1 - \exp \left[ \varepsilon N_{y}^{(0)} \right] } \, , \qquad Q(f_{z0})|_{\varepsilon = 0} = f_{z0} \,. \label{supmat_eq_PG_fixprob_f}$$ On first appraisal, the fixation probabilities [Eq.(\[supmat\_eq\_PG\_fixprob\_f\])]{} appear to share the form of the well-mixed Moran model with weak selection. There is however one crucial distinction; the relation between $f_{z0}$ and $(x_{0},y_{0},q_{0})$ is dependent on the form of the CM/SM, and is not necessarily symmetric under the interchange of $X$ and $Y$. For instance, let us consider the quasi-neutral case ($\varepsilon=0$) with the population initially consisting of a mutant $X$ in a population of the $Y$ phenotype in its stationary state. Then $f_{z0} = 1/N_{y}$. In contrast, if the mutant is of phenotype $Y$, and the resident population consists of phenotype $X$ in the stationary state, $f_{z0} = 1 - 1/N_{x}$. Since $N_{x}$ and $N_{y}$ are distinct, these frequencies are not the same, and [Eq.(\[supmat\_eq\_PG\_fixprob\_f\])]{} is not symmetric under the interchange of phenotypes, undermining its apparent similarities with the Moran model.
In this section a crucial aspect of the selection reversal has been elucidated. The selection reversal along the SM is a result of the differing densities at which the populations of $X$ and $Y$ phenotypes reside in isolation. In a deterministic system, we would define the fitter phenotype as the one which fixates at long times. In stochastic Moran-type model, the fitter phenotype is defined as that with the greater invasion probability. Since Moran-type models feature a constant population size, $N$, the invasion probability of a mutant phenotype is defined by a unique initial condition; a single mutant, and $N-1$ residents. In systems such as the public good model discussed in this paper, the invasion probability is no longer defined uniquely by the specification of a single invading mutant; we must also define the size of the resident phenotype population and the public good density. If the system has been allowed to relax to a stationary state before the mutant is introduced, then selection reversal along the CM may be present, and it is possible for the producing phenotype to have a larger fixation probability than the non-producing phenotype. Thus the producing phenotype may be fitter.
Pairwise invasibility for non-producers, producers and hyper-producers {#sec:supMat_sec_pairwise}
========================================================================
In this section we explore the pairwise invasibility of three separate phenotypes, non-producers, producers and hyper-producers. We begin by noting that, under the assumption that the birth rates differ by only a small amount from phenotype to phenotype, the invasion probability of phenotype $i$ in a resident population $j$, $\phi_{i|j}$, can be expressed $$\begin{aligned}
\phi_{i|j} = \frac{1 - \exp\left[ (b_{i} - b_{j} ) / ( \kappa N_{j}^{(0)} ) \right] }{ 1 - \exp\left[ (b_{i} - b_{j} ) R^{2} / \kappa \right] } \,.\end{aligned}$$ We therefore define phenotype $i$ as fitter than phenotype $j$ if $\phi_{i|j}>\phi_{j|i}$. Let us now explicitly express the birth rates of each of the phenotypes as $$\begin{aligned}
\mathrm{Non-producer}: \, b_{y} &=& b \,, \nonumber \\
\mathrm{Producer}: \, b_{x} &=& b(1-\varepsilon) \,, \nonumber \\
\mathrm{Hyper-producer}: \, b_{u} &=& b( 1 - a_{b} \varepsilon) \,. \nonumber \\ \label{supmat_eq_birth_pairwise}\end{aligned}$$
We now wish to obtain an expression for the critical costs to birth rate $\varepsilon$ at which producers are fitter than non-producers, hyper-producers are fitter than non-producers and hyper-producers are fitter than producers. To do this we must solve $\phi_{i|j}=\phi_{j|i}$ for $\varepsilon$ for each pair of phenotypes. An analytic solution is available if we set $\varepsilon = \tilde{\varepsilon}R^{-2}$ with $\tilde{\varepsilon}$ of order one, and expand Taylor expand in $R^{-2}$. Truncating at first order, we find that the critical cost for species $i$ to be fitter than species $j$, $\varepsilon_{i|j}$ is given by $$\begin{aligned}
\varepsilon_{i|j} = \frac{\kappa \log \left[ ( p_{i} r - \delta \kappa ) / ( p_{j} r - \delta \kappa )\right] }{ \left[ (b_{i} - b_{j})/\varepsilon \right] R^{2} } \,.\end{aligned}$$ We note that this provides eight different possible scenarios of fitness ranking, described in [Fig.\[supmat\_fig\_invasion\_scenarios\]]{}. Substituting in our equations for the birth rates, [Eq.(\[supmat\_eq\_birth\_pairwise\])]{}, these expressions become $$\begin{aligned}
\varepsilon_{x|y} &=& \frac{ \kappa }{ b R^{2} } \log \left[ -\frac{ \delta \kappa }{ p_{x} r - \delta \kappa } \right] \,, \\
\varepsilon_{u|y} &=& \frac{ \kappa }{ a b R^{2} } \log \left[ - \frac{ \delta \kappa }{ p_{u} r - \delta \kappa }\right] \,,\\
\varepsilon_{u|x} &=& \frac{ \kappa }{ (a-1) b R^{2} } \log \left[ \frac{ p_{x} r - \delta \kappa }{ p_{u} r - \delta \kappa }\right] \,.\end{aligned}$$
![ Eight different fitness rankings are possible based on the pairwise invasibility probabilities of non-producers, producers and hyper-producers. (A) Producers have a larger invasion probability than both hyperproducers and nonproducers, while hyperproducers have a larger invasion probability than nonproducers. (B) Producers have a larger invasion probability than both hyperproducers and nonproducers, while nonproducers have a larger invasion probability than hyperproducers. (C) Hyperproducers have a larger invasion probability than both producers and nonproducers, while nonproducers have a larger invasion probability than producers. (D) Hyperproducers have a larger invasion probability than both producers and nonproducers, while producers have a larger invasion probability than producers. (E) Nonproducers have a larger invasion probability than both producers and hyperproducers, while hyperproducers have a larger invasion probability than producers. (F) Nonproducers have a larger invasion probability than both producers and hyperproducers, while producers have a larger invasion probability than hyperproducers. (G) Producers have a larger invasion probability than hyperproducers. Hyperproducers have a larger invasion probability than nonproducers. Nonproducers have a larger invasion probability than producers. (H) Producers have a larger invasion probability than nonproducers. Nonproducers have a larger invasion probability than hyperproducers. Hyperproducers have a larger invasion probability than producers. The nontransitive dynamics of G and H are not seen in the public good model. []{data-label="supmat_fig_invasion_scenarios"}](figS2.pdf){width="45.00000%"}
Clearly the exact scenarios which emerge for a given set of parameters depends on the relationship between $p_{x}$ and $p_{u}$. We make the assumption $$\begin{aligned}
\quad p_{u} = a_{p} p_{x} \,.\end{aligned}$$ For $a_{p}>a_{b}$, the hyper-producer pays a discounted cost to its birth rate for its additional good production. In this situation, only scenarios (c-f) are possible in [Fig.\[supmat\_fig\_invasion\_scenarios\]]{}. It is always better to be a hyper-producer or a non-producer, depending on the production rate $p_{x}$ and $\epsilon$. This ‘all or nothing’ result makes intuitive sense; if the hyper-producer produces much more than the producer, but pays only fractionally more to its birth rate, any region in which production is favored will be disproportionately advantageous to the hyper-producers. In contrast, if $a_{p}<a_{b}$, the hyper-producer receives decreasing production returns as a function of the cost it pays to birth in comparison with the producer. In this case, scenarios (a-b) and (e-f) are possible. Either producers or non-producers are favored, and hyper-producers are never favored.
Generality of results {#sec:supmat_sec_general_results}
=====================
We begin by specifying in a very general way the dynamics of an arbitrary IBM with $m$ distinct types of constituent, fully described by a set of $u$ reaction rates. The model can be expressed in chemical reaction notation as $$\begin{aligned}
\label{supmat_generalReaction}
\sum_{i=1}^{m} a_{\mu i}X_{i} \xrightarrow{ r_{\mu} } \sum_{i=1}^{m}b_{ \mu i}X_{i}, \quad \forall \mu=1, \dots u,\end{aligned}$$ where $a_{\mu i}$ and $b_{ \mu i}$ respectively specify the reactants and products of the $\mu^{th}$ reaction, and $r_{\mu}$ are the reaction rate constants (see, for example, [Eq.(\[supmat\_eq\_reactions\])]{}). The stoichiometric matrix is defined by $\nu_{i \mu } = b_{\mu i}-a_{\mu i}$, whose elements give the change in number of the $i^{\mathrm{th}}$ species due to the $\mu^{\mathrm{th}}$ reaction. Together with the rate constants $r_{\mu}$, the stoichiometric matrix allows us to express the transition rates $$\begin{aligned}
T_{\mu}( \bm{n} + \bm{\nu}_{\mu} |\bm{n}) = r_{\mu} \prod_{i=1}^{m} a_{\mu i} \frac{n_{i}}{R^{2}} \,, \label{supmat_relate_reactants_transitions}\end{aligned}$$ where $R^{2}$ once again is a controlled measure of how often constituents interact (see [Eq.(\[supmat\_eq\_trans\])]{}). In the well-mixed model, it therefore directly controls the typical area of the system. Together with the master equation (\[eq:supmat\_meqn\_general\]), the full stochastic dynamics are specified.
With a general notation now in hand, we now begin to define the specific type of system we will analyze. We consider a system consisting of two phenotypes, $X_{1}$ and $X_{2}$, who interact with a set of discrete ecosystem variables $X_{i}$, for $i = 3, \ldots , N$. The state of the system at any time is given by the number of each phenotype and ecosystem constituent $\bm{n}=(n_{1},n_{2},n_{3},\ldots,n_{N})$. The situation we envisage is as follows; while the interplay between the phenotypes and the ecosystem is relevant for the dynamics, we are primarily interested in the evolutionary dynamics and outcome of competition between the two phenotypes. We make the following assumptions on their dynamics;
1. Each phenotype birth and death event is proportional to the number of that phenotype; $$\begin{aligned}
\mathrm{if} \quad \nu_{1\mu} &\neq& 0 \quad \mathrm{then} \quad a_{\mu1}>0 \,, \qquad \mathrm{and} \nonumber \\ \mathrm{if} \quad \nu_{2\mu} &\neq& 0 \quad \mathrm{then} \quad a_{\mu2}>0 \,.\end{aligned}$$
2. The phenotypes are very similar in their utilization of the ecosystem. For each $\mu^{th}$ reaction that changes the frequency of $X_{1}$, there therefore exists a similar reaction $\mu'$ that changes the frequency of $X_{2}$ such that; $$\nu_{1 \mu} r_{\mu} = \nu_{2 \mu' } ( r_{\mu'} + \mathcal{O}(\varepsilon) ) \,.$$
3. There is no reaction which simultaneously changes the frequencies of the phenotypes (i.e. no cannibalization or simultaneous killing); $$\nu_{1 \mu} \nu_{2 \mu } = 0 \qquad \forall \, \mu \,.$$
The phenotypes may however differ significantly in their effect on the ecosystem, so that one phenotype may deplete or increase ecosystem constituents in an entirely distinct way to the other (for instance, the production of a public good by phenotype $X$ in [Eq.(\[supmat\_eq\_reactions\])]{}).
As $R$ is increases so too does the number of each phenotype and ecosystem constituent. If $R$ is sufficiently large, once again a system-size expansion of the master equation can be conducted. Making the change of variables $x_{1} = n_{1}/R^{2}$, $x_{2} = n_{2} /R^{2} $ and $e_{i}=n_{i-2}/R^{2}$, we obtain the set of Itō SDEs $$\begin{aligned}
\frac{\mathrm{d}x_{1}}{\mathrm{d}t} &=& x_{1} \left[ F^{(0)}(\bm{x},\bm{e}) - \varepsilon F^{(1)}(\bm{x},\bm{e}) \right] + \frac{1}{R}\eta_{1}(t) \, , \nonumber \\
\frac{\mathrm{d}x_{2}}{\mathrm{d}t} &=& x_{2} F^{(0)}(\bm{x},\bm{e}) + \frac{1}{R}\eta_{2}(t) \, , \nonumber \\
\frac{\mathrm{d}e_{i}}{\mathrm{d}t} &=& h_{i}(\bm{x},\bm{e}) + \frac{1}{R}\beta_{i}(t) \,, \quad \forall \, i = 1 , \ldots N \,. \label{supmat_eq_general_SDEs}\end{aligned}$$ The deterministic contribution to the SDEs can be determined from the transitions via $$\begin{aligned}
x_{1} \left[ F^{(0)}(\bm{x},\bm{e}) \right. &-& \left. \varepsilon F^{(1)}(\bm{x},\bm{e}) \right] = \nonumber \\ && \sum_{\mu=1}^{u} \nu_{1 \mu}T_{\mu}\left[ R^{2} (\bm{x},\bm{e})^{T} + \bm{\nu}_{\mu} | (\bm{x},\bm{e})^{T} \right] \,, \label{supmat_eq_general_drift_1} \\
x_{2} F^{(0)}(\bm{x},\bm{e}) &=& \sum_{\mu=1}^{u} \nu_{2 \mu}T_{\mu}\left[ R^{2} (\bm{x},\bm{e})^{T} + \bm{\nu}_{\mu} | (\bm{x},\bm{e})^{T} \right] \,, \label{supmat_eq_general_drift_2} \\
h_{i}(\bm{x},\bm{e}) &=& \sum_{\mu=1}^{u} \nu_{(i+2) \mu}T_{\mu}\left[ R^{2} (\bm{x},\bm{e})^{T} + \bm{\nu}_{\mu} | (\bm{x},\bm{e})^{T} \right] \,. \nonumber\end{aligned}$$ Notice that the relationship between Eqs. (\[supmat\_eq\_general\_drift\_1\]) and (\[supmat\_eq\_general\_drift\_2\]) is controlled by assumption 2. The correlations in the noise meanwhile are given by $$\begin{aligned}
&& \langle \eta_{1}(t) \eta_{1}(t') \rangle = \nonumber \\ && \delta(t-t') \lim_{\varepsilon \to 0 } \sum_{\mu=1}^{u} \nu_{1 \mu}^{2} T_{\mu}\left[ R (\bm{x},\bm{e}) + \bm{\nu}_{\mu} | (\bm{x},\bm{e}) \right] \,, \label{supmat_eq_general_phenotype_correlation_1}\\
&& \langle \eta_{2}(t) \eta_{2}(t') \rangle = \nonumber \\ && \delta(t-t') \lim_{\varepsilon \to 0 } \sum_{\mu=1}^{u} \nu_{2 \mu}^{2} T_{\mu}\left[ R (\bm{x},\bm{e}) + \bm{\nu}_{\mu} | (\bm{x},\bm{e}) \right] \,, \label{supmat_eq_general_phenotype_correlation_2}\\
&& \langle \eta_{1}(t) \eta_{2}(t') \rangle = 0 \,, \label{supmat_eq_general_phenotype_correlation_3}\\
&& \langle \eta_{i}(t) \beta_{j}(t') \rangle = \nonumber \\ && \delta(t-t')\lim_{\varepsilon \to 0 } \sum_{\mu=1}^{u} \nu_{i \mu}\nu_{(j+2) \mu} T_{\mu}\left[ R (\bm{x},\bm{e}) + \bm{\nu}_{\mu} | (\bm{x},\bm{e}) \right] \,, \\
&&\langle \beta_{i}(t) \beta_{j}(t') \rangle = \nonumber \\ && \delta(t-t') \lim_{\varepsilon \to 0 } \sum_{\mu=1}^{u} \nu_{(i+2) \mu}\nu_{(j+2) \mu} T_{\mu}\left[ R (\bm{x},\bm{e}) + \bm{\nu}_{\mu} | (\bm{x},\bm{e}) \right] \,, \nonumber \\\end{aligned}$$ at leading order in $\varepsilon$. The lack of noise correlation between the phenotypes, [Eq.(\[supmat\_eq\_general\_phenotype\_correlation\_3\])]{}, is a consequence of assumption 3. Assumption 2 allows us to rewrite Eqs. (\[supmat\_eq\_general\_phenotype\_correlation\_1\]) and (\[supmat\_eq\_general\_phenotype\_correlation\_2\]) as $$\begin{aligned}
\langle \eta_{1}(t) \eta_{1}(t') \rangle &=& \delta(t-t') x_{1} H(\bm{x},\bm{e}) \,, \nonumber\\
\langle \eta_{2}(t) \eta_{2}(t') \rangle &=& \delta(t-t') x_{2} H(\bm{x},\bm{e}) \,. \label{supmat_eq_phenotype_correlation_rule}\end{aligned}$$ An example of a system where this condition is not enforced is explored in [Section\[sec:supmat\_sec\_comp\_models\]]{}.
To begin our analysis of the SDEs, a quasi-neutral limit is considered in which $\varepsilon=0$. Then the deterministic ODEs for the system (the SDEs in the limit $R\rightarrow \infty$) lead to a manifold of fixed points associated with the focus phenotypes. We now make two additional assumptions;
5. There exits a single stable, well behaved, manifold
6. This manifold is one-dimensional, and so can be paramaterized by a single variable
We then choose to parameterize the manifold in terms of $x_{1}$, which for clarity we label $z$ on the CM. The CM is then defined by the set of equations $$x_{1} = z \,, \qquad x_{2} = Z_{2}(z) \, ,\qquad e_{i} = Z_{ei}(z) \quad \forall \, i=2, \ldots N \,. \label{supmat_eq_general_CM}$$ The system dynamics are now entirely analogous to that of the public good model in the quasi-neutral limit. Deterministically, the system comes to rest at a point on the CM (which depends on the system’s initial conditions) at which it stays indefinitely, and when stochasticity is included the system moves along the CM until one of the phenotypes fixates. A timescale separation is present so long as the composition of the population changes on a slower timescale to that of the collapse to the CM. In practice, the timescale of the collapse to the CM can be inferred from the eigenvalues of [Eq.(\[supmat\_eq\_general\_SDEs\])]{} linearised about the CM. The magnitude of the smallest non-zero eigenvalue is indicative of the slowest component of collapse to the CM [@constable_2013]. This should be much larger than the timescale at which the system moves along the CM, which is of order $R^{-1}$ [@constable_2014_phys].
In order to implement the timescale separation, a non-linear projection is applied to the system which maps fluctuations back to the CM. This can be seen to be equivalent to transforming into the deterministically invariant variable whose existence is guaranteed by the existence of the CM [@arnold_2003], setting the dynamics in all other variables equal to zero, and evaluating the variables themselves on the CM. What form does this mapping take, in the quasi-neutral limit, for [Eq.(\[supmat\_eq\_general\_SDEs\])]{}? Since the dynamical equations for the phenotypes take on the form of degenerate replicator equations in the limit $\varepsilon \rightarrow 0$, the ratio $x_{1}/x_{2}$ is deterministically invariant, regardless of the other parameters. Therefore the non-linear mapping may be obtained by solving the following equation for $z$; $$\frac{ z }{ Z_{2}( z ) } = \frac{ x_{1} }{ x_{2} } \, , \quad \rightarrow \quad z = Y(x_{1},x_{2}) \,. \label{supmat_eq_general_projection}$$ The resulting effective description for the quasi-neutral system on the CM can be denoted $$\dot{z} = \frac{1}{R}\mathcal{S}( z ) + \frac{1}{\sqrt{R}} \zeta(t) \,.$$ Note that while the deterministic system evaluated on the CM had no drift dynamics, the reduced system may. Mathematically, this is a consequence of the fact that the equations are defined strictly in the Itō sense (from the underlying IBM) and therefore the normal rules of calculus do not apply. Instead, any nonlinear transformation induces a drift, in general given by $$\begin{aligned}
\mathcal{S}( z ) &=& \frac{1}{2} \left. \left[ \sum_{ij}^{2} \left( \frac{\partial z }{\partial x_{i} \partial x_{j} } B_{ij} \right) \right. \right. \nonumber \\ &+& \left. \left. \sum_{ij}^{N} \left( \frac{\partial z }{\partial e_{i} \partial e_{j} } B_{eij} \right) \right] \right|_{x_{1}=z,x_{2}=Z_{2}(z),e_{i}=Z_{ei}(z)} . \label{supmat_eq_Abar_stoch_1}\end{aligned}$$ However, since the mapping $z$ is independent of the ecosystem variables $\bm{e}$ (see [Eq.(\[supmat\_eq\_general\_projection\])]{}), [Eq.(\[supmat\_eq\_Abar\_stoch\_1\])]{} can be simplified to $$\mathcal{S}( z ) = \frac{1}{2} \left.\sum_{ij}^{2} \left( \frac{\partial z }{\partial x_{i} \partial x_{j} } B_{ij}\right) \right|_{x_{1}=z,x_{2}=Z_{2}(z),e_{i}=Z_{ei}(z)} \,. \label{eq:supmat_eq_Abar_stoch_2}$$ The form of the correlations in $\zeta(t)$ are now given by $$\mathcal{B}(z) = \left. \sum_{ij}^{2} \left(\left[ \frac{\partial z }{\partial x_{i} } \right]_{i} \left[ \frac{\partial z }{ \partial x_{j} }\right]_{j} B_{ij}(\bm{x}) \right) \right|_{x=z,x_{2}=Z_{2}(z),e_{i}=Z_{ei}(z)} \,, \label{eq:supmat_eq_general_Bbar}$$ where once again we have taken advantage of the property $(dz/de_{i})=0$ for all $i$.
In this very general scenario, what inferences can we make about $\mathcal{S}( z )$? To answer this, it is convenient to return to our original SDEs, [Eq.(\[supmat\_eq\_general\_SDEs\])]{}, and implement the timescale separation in a different fashion. We begin by transforming into variables measuring the total size of the $x_{1}$ and $x_{2}$ population and the fraction of type $x_{1}$; $$\begin{aligned}
N_{T} = x_{1} + x_{2} \,, \quad f_{1} = \frac{ x_{1} }{ x_{1} + x_{2} } \,, \nonumber \\ \rightarrow \quad x_{1} = f_{1} N_{T} \,, \quad x_{2} = N_{T}(1 - f_{1}) \,.\end{aligned}$$ Applying this transformation, taking care to account for the impact of Itō calculus, we arrive at the following SDEs for the system; $$\begin{aligned}
\frac{\mathrm{d}f_{1}}{\mathrm{d}t} &=& \frac{1}{2R^{2}}\sum_{i,j=1}^{2}\frac{\partial^{2} f_{1} }{\partial x_{i} \partial x_{j} } B_{ij} + \frac{1}{R}\tilde{\eta}_{1}(t) \, , \nonumber \\
\frac{\mathrm{d}N_{T}}{\mathrm{d}t} &=& N_{T} F^{(0)}(\bm{x},\bm{e}) + \frac{1}{R}\tilde{\eta}_{2}(t) \, , \nonumber \\
\frac{\mathrm{d}e_{i}}{\mathrm{d}t} &=& h_{i}(\bm{x},\bm{e}) + \frac{1}{R} \tilde{\beta_{i}}(t) \,, \quad \forall \, i = 1 , \ldots N \,.\end{aligned}$$ By conducting the transformation, we immediately notice a few things. Most trivially, the forms of the noise correlations are now altered in all variables. Second, since the transformation into the variable $N_{T}$ was linear, its governing SDE contains no noise-induced elements. Finally, the non-linear transformation into $f_{1}$ has resulted in a noise induced drift term. This drift term however is only dependent on the noise correlation structure between $x_{1}$ and $x_{2}$. Evaluating the dynamics for $N_{T}$ and $\bm{e}$ on the CM and substituting in the remaining expressions from Eqs. (\[supmat\_eq\_general\_phenotype\_correlation\_3\]) and (\[supmat\_eq\_phenotype\_correlation\_rule\]), we obtain the following one-dimensional SDE for $f_{1}$; $$\frac{ \mathrm{d} f_{1} }{ \mathrm{d} t } = \frac{1}{R}\tilde{\eta}_{1}(t) \,, \label{eq_f1_neutral}$$ where $\tilde{\eta}_{1}(t)$ is evaluated on the CM. There are no deterministic dynamics in our reduced dimension description of $f_{1}$. This is a consequence of assumptions 2 and 3. The equation for the fixation probability of phenotype $X_{1}$ given an initial *fraction* $f_{10}$ on the CM, $Q(f_{10})$, is then, regardless of the noise form, $$\begin{aligned}
Q(f_{10}) = f_{10} \,.\end{aligned}$$ Crucially however, $f_{1}$ is evaluated on the CM, which may vary depending on the constitution of the population; $$\begin{aligned}
f_{10} = \frac{ x_{10} }{ x_{10} + Z_{2}(x_{10}) } \,.\end{aligned}$$ If $\left[ d Z_{2}(x_{10}) /dx_{10} \right] < 1 $, then the total phenotype population decreases with increasing $x_{20}$, and phenotype $X_{1}$ has a larger invasion probability than $X_{2}$. From this we can infer that $\mathcal{S}(z)$ will be positive on average along the length of the CM; $$\begin{aligned}
\int_{z=0}^{N_{1}/R^{2}} \mathcal{S}(z) dz > 0 \,. \label{eq:supmat_integral_S}\end{aligned}$$ Therefore, the phenotype with the higher carrying capacity will be stochastically selected for in this quasi-neutral case, regardless of their interaction with the environment. We note once again that this result is in general dependent on assumption 2. If assumption 2 does not hold then there will be correlations between the fluctuations $\eta_{1}(t)$ and $\eta_{2}(t)$ and, rather than the equation for the time evolution of $f_{1}$ featuring no mean drift (as in Eq. (\[eq\_f1\_neutral\]) there will be a noise induced drift term favoring one or other of the phenotypes. The exact form of this term will be highly dependent on the exact form of the interactions between the phenotypes, a full treatment of which lies outside the scope of this paper.
Now suppose that $\varepsilon>0$, so that the system is non-neutral. Now there exists no CM. There is no line of deterministic fixed points, and therefore no invariant variable to project our variables on to and reduce the problem. However, under the assumption that $\varepsilon$ is small there is still a separation of timescales. If $\varepsilon$ is sufficiently small, the slow manifold (and the projection to it) can be approximated by the results from the quasi-neutral case (see Eqs. (\[supmat\_eq\_general\_CM\]) and (\[supmat\_eq\_general\_projection\])), plus an $\varepsilon$ correction. A perturbative analysis can thus be conducted, and, under the assumption the $\varepsilon \approx \mathcal{O}(R^{-2})$, at leading order we have $$\begin{aligned}
\dot{z} = - \varepsilon \mathcal{D}(z) + \frac{1}{R^{2}}\mathcal{S}( z ) + \frac{1}{R} \bar{\eta}(t) \,. \label{supmat_eq_Abar_stoch_det_SDE}\end{aligned}$$ The form of $\mathcal{S}( z )$ is unchanged from [Eq.(\[supmat\_eq\_Abar\_stoch\_1\])]{}, while the new deterministic contribution to the drift takes the form $$\mathcal{D}(z) = - \left. \sum_{i=1}^{N}\left( \frac{ d z }{ d x_{i} } \frac{ d x_{i} }{ d t } \right) \right|_{x_{1}=z,x_{2}=Z_{2}(z),\bm{e}=\bm{Z_{e}}(z) } \,.$$ Once again however, the projection is simply a function of $x_{1}$ and $x_{2}$, and so $$\begin{aligned}
\mathcal{D}(z) &=& - \left. \left( x_{1} F^{(0)}(\bm{x}) \frac{ d z }{ d x_{1} } + x_{2} F^{(0)}(\bm{x}) \frac{ d z }{ d x_{2} }
\right. \right. \nonumber \\ &-& \left. \left. \varepsilon x_{1} F^{(1)} \frac{ d z }{ d x_{1} } \right) \right|_{x_{1}=z,x_{2}=Z_{2}(z),\bm{e}=\bm{Z_{e}}(z) } \,.\end{aligned}$$ Finally, we also know that in the limit $ \varepsilon \rightarrow 0$ this deterministic contribution to the dynamics on the CM, $\mathcal{D}(z)$, should disappear. Therefore the first two terms in the above equation must cancel, leaving us with $$\mathcal{D}(z) = \left. \varepsilon z \left(F^{(1)}(\bm{x}) \frac{ d z }{ d x_{1} } \right) \right|_{x_{1}=z,x_{i}=Z_{i}(z)} \,. \label{eq:supmat_eq_general_Abar_det}$$ We now have a much simpler system to deal with. Say that $F^{(1)}( \bm{x} )$ is strictly positive. Then this will be a term which consistently decreases the value of $x_{1}$. Based on physical arguments, we would expect that, regardless of the form of $\zeta$, $\mathcal{D}(z)$ must be positive. We still require the exact form of $z$ (see [Eq.(\[supmat\_eq\_general\_projection\])]{}) to make analytic progress and specific predictions. Generally however, we have shown that $\mathcal{S}( z )$ will be positive so long as species $X_{1}$ has a larger carrying capacity (subject to the above conditions). A consideration of [Eq.(\[supmat\_eq\_Abar\_stoch\_det\_SDE\])]{} shows that even when the system is non-neutral, for sufficiently weak selection/small $R$, there will be a tradeoff between stochastic ‘strength in numbers’ and deterministic costs for high-density behavior.
Illustrating generality with reference to a complimentary systems: The stochastic Lotka-Volterra system {#sec:supmat_sec_comp_models}
========================================================================================================
In [Section\[supmat\_sec\_public\_good\_reduction\]]{} it was noted that deterministically the public good model reduces to a competitive Lotka-Volterra model under the elimination of the fast public good dynamics. However, it is important to note that though they may be deterministically equivalent at long times, due to alterations in the demographic noise structure the two systems have distinct behaviors. Despite this, the qualitative picture remains the same; for the quasi-neutral system, the fixation probability of each type is simply proportional to its initial fraction in the population, while when selection is introduced, there is playoff between stochastic and deterministic effects. To illustrate this, we investigate the stochastic Lotka-Volterra competition model (SLVC), derived from first principles.
In this section we analyze a stochastic Lotka-Volterra competition model using the methods developed in [Section\[sec:supmat\_sec\_general\_results\]]{}. We assume a population composed of two phenotypes, $X_{1}$ and $X_{2}$, whose numbers in the system are measured by $\bm{n}=(n_{1},n_{2})$. The phenotypes are born, die and compete with each other. In particular, we define the system to be governed by the probability transition rates $$\begin{aligned}
\label{supmat_eq_SLVC_trans_rates}
&T_1(n_1 + 1,n_2|n_1,n_2) = b_1 n_1 \, , &\nonumber \\
&T_2(n_1 - 1,n_2|n_1,n_2) = d_1 n_1 + \frac{ c_{1} }{ R^{2} } n_1^{2} + \frac{ c_{2} }{ R^{2} } n_{1} n_{2} \, , & \nonumber \\
&T_3(n_1,n_2 + 1|n_1,n_2) = b_2 n_2\,, &\nonumber \\
&T_4(n_1,n_2 - 1|n_1,n_2) = d_2 n_2 + \frac{ c_{1} }{ R^{2} } n_1 n_2 + \frac{ c_{2} }{ R^{2} } n_{2}^{2} \, . &\nonumber\end{aligned}$$ Together with [Eq.(\[eq:supmat\_meqn\_general\])]{}, this fully specifies the stochastic dynamics. Taking the limit of large $R$, we can once again obtain a mesoscopic description of the system; $$\begin{aligned}
\frac{\mathrm{d}x_{1}}{\mathrm{d} t } &=& x_{1}\left( (b_{1} - d_{1} ) - c_{1} x_{1} - c_{2} x_{2} \right) + \frac{1}{ R } \eta_{1}( t ), \nonumber \\
\frac{\mathrm{d}x_{2}}{\mathrm{d} t } &=& x_{2}\left( (b_{2} - d_{2} ) - c_{1} x_{1} - c_{2} x_{2} \right) + \frac{1}{ R } \eta_{2}( t ), \nonumber \\
\label{supmat_eq_SLVC_SDE_orig}\end{aligned}$$ where $\eta_{i}(t)$ have correlation structure [Eq.(\[supmat\_eq\_correlations\])]{} with $B_{ij}(\bm{x})$ term given by $$\begin{aligned}
B_{11}(\bm{x}) &=& x_{1}\left( ( b_{1} + d_{1} ) + c_{1} x_{1} + c_{2} x_{2} \right) \,, \nonumber \\
B_{22}(\bm{x}) &=& x_{2}\left( ( b_{2} + d_{2} ) + c_{1} x_{1} + c_{2} x_{2} \right) \,, \nonumber \\
B_{12}(\bm{x}) &\equiv& B_{21}(\bm{x}) = 0 \,.\end{aligned}$$ Note that the noise structure is *not* the same as that in [Eq.(\[supmat\_eq\_phenotype\_correlation\_rule\])]{}; two phenotypes with an equal effective reproduction rate $b_{1}-d_{1}=b_{2}-d_{2}$ have the same deterministic fitness, but distinct multiplicative noise. Phenotypes which are reproducing and dying more quickly are subject to greater noise as they have a larger rate of population turnover. We will however proceed to consider this more general scenario in order to illustrate what can happen when this assumption is not enforced. Finally, we impose a separation of timescales by setting $$\begin{aligned}
b_{1} - d_{1} = \tilde{b}( 1 - \varepsilon ) \,, \quad b_{2} - d_{2} = \tilde{b} \,.\end{aligned}$$ A CM thus exists if $\varepsilon = 0$, and an SM while $\varepsilon$ is small. The parameter $\tilde{b}$ is an effective birth rate encompassing birth and death, while $\varepsilon$ is a fitness cost paid by phenotype $X_{1}$ either in terms of a decreased birth rate, or increased death rate, relative to phenotype $X_{2}$.
In the case $\varepsilon = 0$, the system is quasi-neutral, and so a CM exists. The equation for the CM $x_{2}=Z_{2}(x_{1})$ (see [Eq.(\[supmat\_eq\_general\_CM\])]{}) and its intersection with the boundaries $x_{2}=0$ and $x_{1}=0$, $K_{1}^{(0)}$ and $K_{2}^{(0)}$ respectively, are $$\begin{aligned}
Z_{2}(x_{1}) = \frac{1}{c_{2}} \left( \tilde{b} - c_{1} x_{1} \right) \,, \nonumber \\ \quad K_{1}^{(0)} = \frac{ \tilde{b} }{ c_{1} } \,, \quad K_{2}^{(0)} = \frac{ \tilde{b} }{ c_{2} } \,.\end{aligned}$$ The parameters $K_{1}^{(0)}$ and $K_{2}^{(0)}$ give the frequency of $X_{1}$ and $X_{2}$ phenotypes in isolation. We assume that $c_{2}>c_{1}$ and thus that phenotype $X_{1}$ exists at higher densities than phenotype $X_{2}$. Finally, the mapping from any point $(x_{1},x_{2})$ to a coordinate $z=x_{1}$ on the CM is determined from [Eq.(\[supmat\_eq\_general\_projection\])]{}; $$\begin{aligned}
z = \frac{ \tilde{b} x_{1} }{ c_{1} x_{1} + c_{2} x_{2} } \,.\end{aligned}$$ We can now obtain expressions for $\mathcal{D}( z )$, $\mathcal{S}( z )$ and $\mathcal{B}(z)$ directly from Eqs. (\[eq:supmat\_eq\_general\_Abar\_det\]), (\[supmat\_eq\_Abar\_stoch\_1\]) and (\[eq:supmat\_eq\_general\_Bbar\]); $$\begin{aligned}
\mathcal{D}( z ) &=& - z \left( \tilde{b} - c_{1} z\right) \,,\\
\mathcal{S}( z ) &=& \frac{2}{ \tilde{b}^{2} } z \left( \tilde{b} - c_{1} z\right) \left( c_{2} (\tilde{b} + d_{2} ) - c_{1} (\tilde{b} + d_{1} ) \right) \,, \\
\mathcal{B}(z) &=& \frac{2}{ \tilde{b}^{2} } z \left( \tilde{b} - c_{1} z\right) \left[ z \left( c_{2} (\tilde{b} + d_{2} ) \right. \right. \nonumber \\ &-& \left. \left. c_{1} (\tilde{b} + d_{1} ) \right) + \tilde{b} ( d_{1} + \beta)\right] \,.\end{aligned}$$ The equation can now be solved to calculate the fixation probability of phenotype $X_{1}$ along the CM. In terms of the initial fraction of $X_{1}$, $f_{1}$, we find $$\begin{aligned}
Q(f_{1}) &=& \frac{1 - \chi( f_{1} ) }{ 1 - \chi( 1 ) } \,, \nonumber \\
\chi(f_{1}) &=& \left[ \left( \frac{ K_{1}^{(0)} }{ d_{1} + \tilde{b} } \right) \left( \frac{ d_{1}( 1 - f_{1} ) + d_{2} f_{1} + \tilde{b} }{ K_{1}^{(0)}(1-f_{1}) + f_{1} K_{2}^{(0)} } \right) \right] ^{ - \theta }\,,\end{aligned}$$ where $\theta$ is a parameter given by $$\begin{aligned}
\theta = \left( 1 + \frac{ K_{1}^{(0)} K_{2}^{(0)} R^{2} \tilde{b} \varepsilon }{ K_{2}^{(0)} (d_{1} + \tilde{b} ) - K_{1}^{(0)} ( d_{2} + \tilde{b} ) } \right) \,.\end{aligned}$$
Let us consider the special case $\varepsilon = 0$. The fixation probability then becomes $$\begin{aligned}
Q(f_{1})|_{\varepsilon=0} = \frac{ f_{1} ( d_{2} + \tilde{b} ) }{ d_{1}( 1 - f ) + d_{2} f + \tilde{b} } \,.\end{aligned}$$ The species with the lower death rate (and death rate, since $\tilde{b}$ is fixed), has a greater probability of fixation than the species with the higher birth rate/death rate. This insight, made in [@parsons_quince_2007_1; @doering_2012], is a result of the higher levels of noise experienced by the phenotype with the high birth and death rates. This makes it easier for the longer lived phenotype, (lower birth/death rates), to invade and fixate. For the purposes of this paper, we ignore such effects in order to focus on systems in which the carrying capacity of the phenotypes alone is responsible for the differences in noise experienced by the phenotypes on the CM/SM.
To this end, we now focus on the case $b_{1}=b_{2}\equiv b$, $d_{1}=d_{2}\equiv d$. In this case, $Q(f_{1})|_{\varepsilon=0} = f_{1}$, and $Q(f_{1})$ in general becomes $$\begin{aligned}
Q(f_{1}) &=& \frac{1 - \chi( f_{1} ) }{ 1 - \chi( 1 ) } \,, \nonumber \\
\chi(f_{1}) &=& \left( \frac{ K_{1}^{(0)} } { K_{1} (1 -f_{1} ) +f_{1}K_{2}^{(0)} } \right)^{-\theta} \,,\end{aligned}$$ where $\theta$ is now given by $$\begin{aligned}
\theta = \left( 1 + \frac{ K_{1}^{(0)} K_{2}^{(0)} R^{2} ( b - d ) \varepsilon }{ ( K_{2}^{(0)} - K_{1}^{(0)} ) b } \right) \,.\end{aligned}$$
The invasion probabilities $\phi_{1}$ and $\phi_{2}$ meanwhile are given by $$\begin{aligned}
\phi_{1} = Q(N_{2}^{-1}) \,,\quad \phi_{2} = 1 - Q( 1 - N_{1}^{-1} ) \,.\end{aligned}$$ We can use the above expressions to obtain an approximate value for the maximum cost to birth rate that can be paid in order that selection reversal is observed. Assuming $N_{1}^{-1}$ and $N_{2}^{-1}$ are of order $\varepsilon$ and Taylor expanding in $\varepsilon$, we find the cost to birth must obey $$\frac{1}{N_{2}} \left( \frac{ b }{ b - d } \right) \left( 1 - \frac{ N_{2} }{ N_{1} } \right) > \varepsilon$$ for the direction of selection to be reversed. This is analogous to Eq. ($8$) in the main text.
Order of magnitude parameter estimates {#sec:supmat_sec_params}
======================================
In this section we seek an illustrative set of parameters for use in the model in order emphasis that the insights developed are biologically reasonable. We wish to obtain order of magnitude estimates for the set of parameters, $b$, $p_{x}$, $p_{u}$, $r$, $\delta$, $\kappa$, $R$, $m$ and $D$. We choose the yeast *Saccharomyces cerevisiae* as our model organism. While our model is more physically realistic than many mathematical public good models, we note that there are still choices that must be made in relating this physical system to our general framework.
Our model is constructed such that the uptake of one constituent of the public good, $Q$, by a phenotype, results in a reproduction event. In the context of *S. cerevisiae*, the type $Q$ is thus shorthand for the amount of invertase that must be present in the system to break down sucrose into sufficient glucose for a reproduction event of the yeast. Let us define $\sigma$ to be the scaling between $n_{q}$ and the total number of invertase molecules, such that the number of invertase molecules is $\sigma n_{q}$. In order to understand the relationship between our model parameters and physically measurable parameters, we begin by considering a simplified ODE system of our model. $$\begin{aligned}
\frac{ d x }{ d t } &=& x( b + r q - \kappa x ) \,,\\
\frac{ d q }{ d t } &=& p x - \delta q \,. \end{aligned}$$ While the total number of discrete invertase constituents is $n_{q}\approx R^{2} q$, the total number of invertase molecules is $R^{2}\sigma q$. Let $\theta$ be a measure of the number of invertase molecules, such that $\theta=\sigma q$. The ODEs in this more natural variable read $$\begin{aligned}
\frac{ d x }{ d t } &=& x( b + \frac{r}{\sigma} \theta - \kappa x ) \,,\\
\frac{ d \theta }{ d t } &=& \sigma p x - \delta \theta \,. \end{aligned}$$ The decay rate $\delta$ is independent of the number of molecules which make up an invertase constituent $Q$, so we can take experimental measurements of the invertase molecular decay rate as values for $\delta$. Meanwhile the molecular invertase production rate and reproduction rate due to invertase take on scaled forms of the parameters in our original ODEs; $$\begin{aligned}
r_{mol} &=& \frac{r}{\sigma} \,, \\
p_{mol} &=& \sigma p \,.\label{eq:eq_pmol}\end{aligned}$$
While measurements of $p_{mol}$ are obtainable in the literature (see Table \[tab:supmat\_fig\_parameter\_exp\]), our estimation of $r_{mol}$ is complicated by the fact that it is an effective parameter. It must capture the increase in the reproductive rate due to invertase, which in reality is coupled to both the reaction rate of invertase and sucrose into glucose, as well as the uptake rate of glucose by yeast and the energy conversion to reproduction. We do however know the typical range of yeast reproduction rates. Let us define $\lambda_{exp}$ as the yeast reproduction rate as measured experimentally. In turn, let $\lambda_{\text{eff}}$ be the effective per capita reproduction rate of yeast in the model; $$\begin{aligned}
\lambda_{\text{eff}} = b + r q \,.\end{aligned}$$ The yeast reproduction rate clearly depends on the amount of public good in the system, typically varying from $$\begin{aligned}
\lambda_{\text{eff}} &=& b \qquad\mathrm{(all\,non-producers)} \nonumber \\
\mathrm{to} \qquad \lambda_{\text{eff}} &=& b \frac{ \delta \kappa }{ \delta \kappa - p_{x} r } \qquad \mathrm{(all\,producers)} \,.\end{aligned}$$ In reality, the reproduction rate of yeast in a system without any invertase is effectively zero; we have assumed some baseline birth rate for convenience in the model, which could be physically interpreted as being associated with an exogenous glucose concentration in the system. We assume that this is typically low, such that $b$ is small, while the yeast approaches its maximum reproductive rate when it consists entirely of producers.
The parameter $\kappa$ controls death due to crowding. For simplicity this is the only form of death in the model. This choice leads, perhaps unnaturally, to the non-producers (who exist at typically lower densities) having a much smaller death rate than producers. For the parameters chosen however, we obtain per- capita death rates on the order of an hour for producers, and ten hours for non-producers. The parameter $R$ meanwhile measures the assumed spatial interaction scale. It determines the typical number of individuals on each patch. We can use this value to infer the size of a patch. Denoting the diameter of a yeast cell as $L_{c}$, and assuming that the hyper-producing cells in the stationary state can be packed on a grid, the size of each patch, $L_{p}$, can be approximated by $$\begin{aligned}
L_{p} &=& L_{c} \sqrt{ N_{u} } \,,\\
&=& L_{c} R \sqrt{K_{u}} \,. \label{eq:eq_Lp}\end{aligned}$$ The parameters $m$ and $D$ are effective migration and diffusion rates in our model. To map these physical parameters these must be in turn scaled by the patch length. The public good diffusion rate must also be scaled by $\sigma$, which maps the discrete amount of invertase constituents $Q$ to the number of invertase molecules. Denoting $m_{exp}$ and $D_{exp}$ the physical migration and public good diffusion rate rates, it can be shown that [@gardiner_2009] $$\begin{aligned}
D_{exp} = \sigma L_{p}^{2} D \,,\qquad m_{exp} = L_{p}^{2} m \,. \label{eq:eq_D_m_exp}\end{aligned}$$ The parameter choices which follow from these calculations are summarized in [Table\[tab:supmat\_fig\_parameter\_list\_1\]]{}.
![Figure illustrating the larger range of values for the parameter $\varepsilon$ over which the approximation [Eq.(\[supmat\_eq\_PG\_fixprob\_f\])]{} is accurate. Parameters are given in [Table\[tab:supmat\_fig\_parameter\_list\_1\]]{}, with the exception of $p_{x}$ and $\varepsilon$ which are varied. Note that this figure is similar to Fig. 2, plotted in the main text, but plotted over a grater range of $\varepsilon$ and $\phi_{x}-\phi_{y}$. The parameter region plotted in black is that for which $ \kappa \delta > r p_{x} $.[]{data-label="supmat_fig_trade_off_larger_eps"}](figS3.pdf){width="56.00000%"}
When considering the parameter choices summarized in [Table\[tab:supmat\_fig\_parameter\_list\_1\]]{}, it is important to make a final point. While both the approximations we have employed, the system size expansion in [Section\[sec:supmat\_sec\_obtain\_SDE\_system\]]{} and the fast-variable elimination in [Section\[supmat\_sec\_public\_good\_reduction\]]{}, rely formally on $R^{2}$ being large and $\varepsilon$ being small ($\mathcal{O}(R^{-2}) > \varepsilon$), in practical terms the procedures are relatively robust to this restriction. Indeed, throughout the body of the main text, $R=4$, while $\varepsilon$ is varied on the interval $[0,0.1]$. In fact we find that the approximate analytic expression we obtain for the invasion probabilities of the phenotypes, [Eq.(\[supmat\_eq\_PG\_fixprob\_f\])]{}, describes the results obtained from simulation well, up to $\varepsilon=R^{-2}$, as illustrated in [Fig.\[supmat\_fig\_trade\_off\_larger\_eps\]]{}.
In terms of the system size expansion, this robustness can in part be explained by the fact that the typical population sizes ($N_{x}$, $N_{y}$ and $N_{q}$) are proportional to $R$. For populations of fixed size $N$, SDEs for the system can be obtained by means of a Taylor expansion of the master equation (for example, [Eq.(\[eq:supmat\_meqn\_general\])]{}) as a series in $1/N$. A crucial feature of the system we are concerned with here however, is that population sizes may vary, and so this technique is unavailable. Instead we conduct an expansion in the interaction scale $R$, which is proportional to the mean population size. Though $R$ may not be a large number itself, increasing $R$ leads to an associated increase in population size (see [Table\[tab:supmat\_fig\_parameter\_list\_1\]]{}). In turn, this leads to terms of higher order in the Taylor expansion of the master equation becoming subdominant [@van_kampen_2007], justifying the truncation which leads to [Eq.(\[eq:supmat\_eq\_FPE\_general\])]{}. In contrast, the resilience of the fast-variable elimination approximation to such large values of $\varepsilon$ is surprising.
[ |l|l|l| ]{}\
Experimental Parameter & Value & Description\
$p_{mol}$ & $0.46 \, \mathrm{mol}\,s^{-1}$ & Production rate of a molecule of invertase per\
& & producing yeast cell [@koschwanez_2011].\
$\delta$ & $2 \times 10^{-3} \, \mathrm{mol}\,s^{-1}$ & Estimated efficacy decay rate of invertase (see [@gomez_2008], Fig 5.)\
$\lambda_{exp}$ & $0.31-0.5 \, hr^{-1}$ & Yeast reproduction rate in producing population [@sanchez_2013; @snoep_2009]\
$\varepsilon_{exp}$ & $0.06$ & Cost of public good production to yeast reproduction rate [@sanchez_2013]\
$D_{exp}$ & $100\, \mu m^{2} s^{-1}$ & Diffusion rate of invertase molecules estimated in [@allen_2013]\
$L_{c}$ & $3\,\mu m$ & Cell length physical approximation [@allen_2013].\
[ ]{}
[ |l|l|l| ]{}\
Parameter & Value & Justification\
$\sigma$ & $4000$ & Assumed parameter. Presence of $4000$ invertase\
& & molecules required for yeast reproduction.\
$p_{y}$ & 0 & True non-producer does not produce invertase.\
$p_{x}$ & $1.14\times10^{-4}\,s^{-1} $, & Experimental value of molecular invertase production\
& ($0.41\,hr^{-1}$) & rate (see Table \[tab:supmat\_fig\_parameter\_exp\]) scaled by $\sigma$ (see [Eq.(\[eq:eq\_pmol\])]{}).\
$p_{u}$ & $ 1.2 \times 10^{-4} \,s^{-1} $ & Leads to factor $1.7$ increase in the steady state invertase\
& ($0.43\,hr^{-1}$) & from producing to hyper-producing population, consistent with [@maclean_2008].\
$b$ & $6.94\times10^{-6}\,s^{-1}$ & Small baseline yeast birthrate assumed.\
& ($0.025\,hr^{-1}$) &\
$r$ & $ 1.58 \times 10^{-5}\, s^{-1} $ & Chosen so as to give per-capita yeast reproduction rate\
& ($0.057\,hr^{-1}$) & $( b + r q) \approx \lambda_{exp}$ when system entirely producers (see Table \[tab:supmat\_fig\_parameter\_exp\]).\
$\delta$ & $0.002\,s^{-1}$ & Taken from experimentally measured values. (see Table \[tab:supmat\_fig\_parameter\_exp\])\
$\kappa$ & $1 \times 10^{-6}\,s^{-1}$ & Suggested parameter for illustrating effects in paper;\
& & restricted by $\delta \kappa > p_{i} r $, $i=x,y,u$.\
$R$ & $ 2 $ & Suggested parameter for illustrating effects in paper.\
$\varepsilon$ & $0.06$ & Taken from experiments (see Table \[tab:supmat\_fig\_parameter\_exp\])\
$N_{y}$ & $ 28 $ & See [Eq.(\[eq:supmat\_eq\_NX\_NY\])]{}.\
$N_{x}$ & $ 302 $ & See [Eq.(\[eq:supmat\_eq\_NX\_NY\])]{}.\
$N_{u}$ & $ 499 $ & See [Eq.(\[eq:supmat\_eq\_NX\_NY\])]{} for $N_{x}$ and substitute $p_{u}$ for $p_{x}$.\
$L_{p}$ & $67 \,\mu m$ & See [Eq.(\[eq:eq\_Lp\])]{}.\
$m$ & $3.4\times10^{-7} \, s^{-1}$ & Yields a migration to birth-rate ratio between $m/b=4.9\times10^{-2}$\
& & (all non-producers) and $m/(b+r q)=4.5\times10^{-3}$ (all producers).\
$D$ & $2.22 \times 10^{-5} \, s^{-1}$ & Obtained using experimental value $D_{exp}$ from Table \[tab:supmat\_fig\_parameter\_exp\] and\
& & [Eq.(\[eq:eq\_D\_m\_exp\])]{}.\
Movies {#sec:supmat_sec_movies}
======
![Still from Movie S1: Movie of a simulation of the metapopulation public good model on a $100\times100$ grid ($C=10^{4}$). Colors indicate the number of producers on each patch; patches with a small number of producers are colored red while patches with a large number of producers are colored blue. Parameters used are $p_{x}=1\times10^{-4}$, $\varepsilon=0.02$, $m=3.7\times10^{-5}$ and the remaining parameters taken from [Table\[tab:supmat\_fig\_parameter\_list\_1\]]{}. With these parameters, $N_{y}\approx28$ and $N_{x}\approx129$. Initial conditions are a single producer and non-producer on each patch. Large numbers of producers on a patch are correlated with low numbers of non-producers on the same patch. The space-averaged dynamics of this simulation are given in the main text, Fig.4. Counter to the deterministic prediction, the number of producers increases with time, while the number of non-producers decreases.[]{data-label="supmat_fig_video_S1"}](movieS1_still.pdf){width="45.00000%"}
![Still from Movie S2: Movie showing the distribution of homogeneous non-producing patches (red), homogeneous producing patches (blue) and heterogeneous mixed patches (gray-green) in the simulation of the metapopulation public good model given in Video S1. For the majority of the observation time, every patch contains a heterogeneous mix of producers and non-producers. Homogeneous producer patches only begin to emerge as producers approach fixation in the system.[]{data-label="supmat_fig_video_S2"}](movieS2_still.pdf){width="45.00000%"}
[10]{}
Hofbauer J, Sigmund K (1998) *[E]{}volutionary [G]{}ames and [P]{}opulation [D]{}ynamics* (Cambridge University Press, Cambridge).
Fisher RA (1930) *The Genetical Theory of Natural Selection* (Clarendon Press, Oxford).
Wright S (1931) Evolution in [M]{}endelian populations. *Genetics* 16:97–159.
Nowak MA (2006) *[E]{}volutionary [D]{}ynamics: [E]{}xploring the [E]{}quations of [L]{}ife* (Harvard University Press, Cambridge, Massachusetts).
McKane AJ, Newman TJ (2005) Predator-prey cycles from resonant amplification of demographic stochasticity. *Phys. Rev. Lett.* 94:218102.
Butler T, Goldenfeld N (2009) Robust ecological pattern formation induced by demographic noise. *Phys. Rev. E* 80:030902.
Hallatschek O, Hersen P, Ramanathan S, Nelson DR (2007) Genetic drift at expanding frontiers promotes gene segregation. *PNAS* 104:19926–19930.
Rossberg AG, Rogers T, McKane AJ (2013) Are there species smaller than 1 mm? *Proc. Roy. Soc. B* 280:20131248.
Parsons T, Quince C (2007) Fixation in haploid populations exhibiting density dependence [I]{}: The non-neutral case. *Theor. Pop. Biol.* 72:121–135.
Lin YT, Kim H, Doering CR (2012) Features of fast living: On the weak selection for longevity in degenerate birth-death processes. *J. Stat. Phys.* 148:646–662.
Gunawardena J (2014) Time-scale separation - [M]{}ichaelis and [M]{}enten’s old idea, still bearing fruit. *FEBS Journal* 281:473–488.
Parsons T, Quince C (2007) Fixation in haploid populations exhibiting density dependence [II]{}: The quasi-neutral case. *Theor. Pop. Biol.* 72:468–479.
Chotibut T, Nelson DR (2015) Evolutionary dynamics with fluctuating population sizes and strong mutualism. *Phys. Rev. E* 92:022718.
Constable GWA, McKane AJ (2015) Models of genetic drift as limiting forms of the [L]{}otka-[V]{}olterra competition model. *Phys. Rev. Lett.* 114:038101.
Parsons TL, Quince C, Plotkin JB (2010) Some consequences of demographic stochasticity in population genetics. *Genetics* 185:1345–1354.
Kogan O, Khasin M, Meerson B, Schneider D, Myers CR (2014) Two-strain competition in quasi-neutral stochastic disease dynamics. *Phys. Rev. E* 90:042149.
Houchmandzadeh B, Vallade M (2012) Selection for altruism through random drift in variable size populations. *BMC Evol. Biol.* 12:61.
Houchmandzadeh B (2014) Fluctuation driven fixation of cooperative behavior. *Biosystems* 127:60–66.
Reznick D, Bryant MJ, Bashey F (2002) r- and k-selection revisited: The role of population regulation in life-history evolution. *Ecology* 83:1509–1520.
Nowak MA, Sasaki A, Taylor C, Fudenberg D (2004) Emergence of cooperation and evolutionary stability in finite populations. *Nature* 428:646–650.
Rice SH (2004) *[E]{}volutionary [T]{}heory* (Sinauer Associates, Sunderland, Massachusetts).
Crow JF, Kimura M (1970) *[A]{}n [I]{}ntroduction to [P]{}opulation [G]{}enetics [T]{}heory* (The Blackburn Press, New Jersey).
Gore J, Youk H, van Oudenaarden A (2009) Snowdrift game dynamics and facultative cheating in yeast. *Nature* 459:253–256.
Hauert C, Holmes M, Doebeli M (2006) Evolutionary games and population dynamics: maintenance and cooperation in public goods games. *Proc. R. Soc. B* 273:2565–2570.
Huang W, Hauert C, Traulsen A (2015) Stochastic game dynamics under demographic fluctuations. *Proc. Natl. Acad. Sci.* 112:9064–9069.
Koschwanez JH, Foster KR, Murray AW (2011) Sucrose utilization in budding yeast as a model for the origin of undifferentiated multicellularity. *[PLoS Biol. ]{}* 9:e1001122.
Kümmerli R, Brown SP (2010) Molecular and regulatory properties of a public good shape the evolution of cooperation. *Proc. Natl. Acad. Sci.* 107:18921–18926.
Gardiner CW (2009) *[H]{}andbook of [S]{}tochastic [M]{}ethods* (Springer, Berlin).
Black AJ, McKane AJ (2012) Stochastic formulation of ecological models and their applications. *Trends Ecol. Evol.* 27:337–345.
Parsons TL, Rogers T (2015) Dimension reduction via timescale separation in stochastic dynamical systems. *ar[X]{}iv* p 1510.07031.
Arnold L (2003) *[R]{}andom [D]{}ynamical [S]{}ystems*, [S]{}pringer [M]{}onographs in [M]{}athematics (Springer, Berlin).
Maclean RC, Fuentes-Hernandez A, Greig D, Hurst LD, Gudelj I (2010) A mixture of “cheats” and “co-operators” can enable maximal group benefit. *[PLoS Biol.]{}* 21:e1000486.
Maclean RC, Brandon C (2008) Stable public goods cooperation and dynamic social interactions in yeast. *J. Evol. Biol.* 21:1836–1843.
Constable GWA, McKane AJ, Rogers T (2013) Stochastic dynamics on slow manifolds. *J. Phys. A: Math. Theor.* 46:295002.
Constable GWA, McKane AJ (2014) Fast-mode elimination in stochastic metapopulation models. *Phys. Rev. E* 89:032141.
Behar H, Brenner N, Ariel G, Louzoun Y (2016) Fluctuations-induced coexistence in public good dynamics.
Behar H, Brenner N, Louzoun Y (2014) Coexistance of productive and non-productive populations by fluctuation-driven spatio-temporal patterns. *Theor. Popul. Biol.* 96:20–29.
Nowak MA, Tarnita CE, Antal T (2009) Evolutionary dynamics in structured populations. *Philosophical Transactions of the Royal Society of London B: Biological Sciences* 365:19–30.
Wakano JY, Hauert C (2011) Pattern formation and chaos in spatial ecological public goods games. *J. Theor. Biol.* 268:30–38.
Nowak MA, May RM (1992) Evolutionary games and spatial chaos. *Nature* 359:826–829.
Allen B, Gore J, Nowak MA (2013) Spatial dilemmas of diffusible public goods. *[eLife]{}* 2:01169.
Waxman D, Gavrilets S (2005) 20 questions on adaptive dynamics. *Journal of Evolutionary Biology* 18:1139–1154.
Pianka ER (1970) On r- and k-selection. *The American Naturalist* 104:592–597.
Lin YT, Kim H, Doering CR (2015) Demographic stochasticity and evolution of dispersion [I]{}. spatially homogeneous environments. *J. Math. Biol.* 70:647.
Lin YT, Kim H, Doering CR (2015) Demographic stochasticity and evolution of dispersion [II]{}. spatially inhomogeneous environments. *J. Math. Biol.* 70:679.
Gillespie DT (1976) A general method for numerically simulating the stochastic time evolution of coupled chemical reactions. *J. Comput. Phys.* 22:403–434.
van Kampen NG (2007) *Stochastic [P]{}rocesses in [P]{}hysics and [C]{}hemistry* (Elsevier, Amsterdam).
Risken H (1989) *The [F]{}okker-[P]{}lanck [E]{}quation* (Springer, Berlin).
Gomez L, Ramirez HL, Cabrera G, Simpson BK, Villalonga R (2008) Immobilization of invertase–chitosan conjugate on hyaluronic-acid-modified chitin. *J. Food Biochem.* 32:264 – 277.
Sanchez A, Gore J (2013) Feedback between population and evolutionary dynamics determines the fate of social microbial populations. *[PLoS Biol. ]{}* 11:e1001547.
Snoep JL, Mrwebi M, Schuurmans JM, M. RJ, de Mattos M. J. T (2009) Control of specific growth rate in [S]{}accharomyces cerevisiae. *Microbiology* 155:1699 – 1707.
|
{
"pile_set_name": "ArXiv"
}
|
6.0in 9.0in 0.0in 0.5in 0.5in 0.08in
epsf
\#1
[**\#1**]{}
\#1
[**\#1**]{}
1[$U(1)$]{} 5[$SU(5)$]{} 10[$SO(10)$]{} 22[$SU(4)\otimes SU(2)_L \otimes SU(2)_R$]{} . \#1[(\[\#1\])]{} \_U[ ]{} \_P[ ]{} +N
1Charges[I]{}
hep-ph/0009287
\
[S. F. King and M.Oliveira]{}\
\
[*[Department of Physics and Astronomy, University of Southampton\
Southampton, SO17 1BJ, U.K]{}*]{}\
[**Abstract**]{}
[We analyse a supersymmetric string-inspired model of all fermion masses and mixing angles based on the Pati-Salam $SU(4)\times SU(2)_L \times SU(2)_R$ gauge group supplemented by a $U(1)_X$ flavour symmetry. The model involves third family Yukawa unification and predicts the top mass and the ratio of the vacuum expectation values $\tan \beta$. The model also provides a successful description of the CKM matrix and predicts the masses of the down and strange quarks. However our main focus is on the neutrino masses and MNS mixing angles, and we show how the recent atmospheric neutrino mixing observed by Super-Kamiokande, and the MSW solution to the solar neutrino problem lead to important information about the flavour structure of the model near the string scale. We show how single right-handed neutrino dominance may be implemented by the use of “Clebsch zeros”, leading to the LMA MSW solution, corresponding to bi-maximal mixing. The LOW MSW and SMA MSW solutions are also discussed.]{}
The problem of understanding the quark and lepton masses and mixing angles represents one of the major unsolved questions of the standard model. Recently additional information on the fermion mass spectrum has come from the measurement of the atmospheric neutrino masses and mixing angles by Super-Kamiokande [@SKamiokandeColl]. The most recent data disfavours mixing involving a sterile neutrino, and finds a good fit for $\nu_{\mu} \rightarrow \nu_{\tau}$ mixing with $\sin^22\theta_{23}>0.88$ and a mass square splitting $\Delta m^2_{23}$ in the $1.5-5\times 10^{-3} {\rm\ eV}^2$ range at 90% CL [@HSobel]. Super-Kamiokande has also provided additional support for solar neutrino mixing. The most recent Super-Kamiokande data does not show a significant day-night asymmetry and shows an energy independent neutrino spectrum, thus it also disfavours the sterile neutrino mixing hypothesis, the just-so vacuum oscillation hypothesis, and the small mixing angle (SMA) MSW [@MSWMechanism] solution [@YSuzuki]. The preferred solution at the present time seems to be the large mixing angle (LMA) MSW solution, although a similar solution with a low mass splitting (the LOW) solution is also possible. A typical point in the LMA MSW region is $\sin^22\theta_{12}\approx 0.75$, and $\Delta m^2_{12}\approx 2.5\times
10^{-5} {\rm\ eV}^2$ [@BaKrSm].
If one accepts the recent data as evidence for neutrino masses and mixing angles, then the obvious question is how these can be accommodated in the standard model, or one of its supersymmetric extensions. The simplest possibility to account for the smallness of the neutrino masses is the see-saw mechanism [@seesaw] in which one introduces right-handed neutrinos which acquire very large Majorana masses at a super-heavy mass scale. When one integrates out the right-handed neutrinos the “normal sized” Dirac Yukawa couplings, which connect the left-handed to the right-handed neutrinos, are transformed into very small couplings which generate very light effective left-handed physical Majorana neutrino masses. Given the see-saw mechanism, it is natural to expect that the spectrum of the neutrino masses will be hierarchical, since the Dirac Yukawa couplings in the charged fermion sector are observed to be hierarchical, and if they are related to the Dirac neutrino Yukawa couplings then they should also be hierarchical, leading to hierarchical light Majorana masses. [^1]
Having assumed the see-saw mechanism and a hierarchical neutrino mass spectrum, the next question is how such large (almost maximal) lepton mixing angles such as $\theta_{23}$ could emerge? There are several possibilities that have been suggested in the literature. One possibility is that it happens as a result of the off-diagonal 23 entries in the left-handed Majorana matrix being large, and the determinant of the 23 sub-matrix being accidentally small, leading to a neutrino mass hierarchy with large neutrino mixing angles [@ElLeLoNa]. Another possibility is that the neutrino mixing angles start out small at some high energy scale, then get magnified by renormalization group (RG) running down to low energies [@BaLePa:MTanimoto]. A third possibility is that the off-diagonal elements of the left-handed neutrino Majorana matrix are large, but the 23 sub-determinant of the matrix is small for a physical reason, as would be the case if a single right-handed neutrino were providing the dominant contribution to the 23 sub-matrix [@KingSRND; @SFKing1; @SFKing2]. We shall refer to these three approaches as the accidental, the magnification and the single right-handed neutrino dominance (SRHND) mechanisms, respectively. As we shall see, in the model under consideration, only the SRHND mechanism provides a successful description of the atmospheric neutrino data, and the results in this paper will rely on this mechanism.
A promising approach to understanding the fermion mass spectrum is within the framework of supersymmetric (SUSY) unified theories. Within the framework of such theories the quark and lepton masses and mixing angles become related to each other, and it begins to be possible to understand the spectrum. The simplest grand unified theory (GUT) is $SU(5)$ but this theory in its minimal version does not contain any right-handed neutrinos. Nevertheless three right-handed neutrinos may be added, and in this theory it is possible to have a large 23 element [^2] on the Dirac neutrino Yukawa matrix without introducing a large 23 element into any of the charged fermion Yukawa matrices. The problem of maintaining a 23 neutrino mass hierarchy in these models may be solved for example by assuming SRHND [@AlFe2]. Another possibility within the framework of $SU(5)$ is to maintain all the off-diagonal elements to be small, but require the 22 and 32 elements of the Dirac neutrino Yukawa matrix to be equal and the second right-handed neutrino to be dominant, in which case SRHND again leads to a large 23 neutrino mixing angle with hierarchical neutrino masses [@AlFeMa]. However the drawback of $SU(5)$ is that it does not predict any right-handed neutrinos, which must be added as an afterthought.
From the point of view of neutrino masses, the most natural GUTs are those like $SO(10)$ that naturally predict right-handed neutrinos. However within the framework of $SO(10)$ the quark masses and mixing angles are related to the lepton masses and mixing angles, and the existence of large neutrino mixing angles is not expected in the minimal versions of the theory in which the Higgs doublets are in one (or two) ${\bf 10}'s$ (ten dimensional representations of $SO(10)$) and each matter family is in a ${\bf 16}$. Nevertheless various possibilities have been proposed in $SO(10)$ in order to account for the large neutrino mixing angles. Within the framework of minimal $SO(10)$ with third family Yukawa unification, it has been suggested that if two operators with different Clebsch coefficients contribute with similar strength then, with suitable choice of phases, in the case of the lepton Yukawa matrices one may have large numerical 23 elements, which add up to give a large lepton mixing angle, while for the quarks the 23 elements can be small due to approximate cancellation of the two contributing operators [@BaPaWi]. This is an example of the accidental mechanism mentioned above, where in addition one requires the quark mixing angles to be small by accident, although it remains to be seen if the LMA MSW solution could be understood in this framework. Moving away from minimal $SO(10)$, one may invoke a non-minimal Higgs sector in which one Higgs doublet arises from a ${\bf 10}$ and one from a ${\bf 16}$, and in this framework it is possible to understand atmospheric neutrino mixing [@AlBa]. Alternatively, one may invoke a non-minimal matter sector in which parts of a quark and lepton family arise from a ${\bf 16}$ and other parts from a ${\bf 10}$, and in these models one may account for atmospheric and solar neutrinos via an inverted mass hierarchy mechanism [@ShTa].
In the present paper we shall discuss neutrino masses and mixing angles in a particular string-inspired [*minimal*]{} model based on the Pati-Salam $SU(4)\times SU(2)_L \times SU(2)_R$ (422) group [@PaSa]. As in $SO(10)$ the presence of the gauged $SU(2)_R$ predicts the existence of three right-handed neutrinos. However, unlike $SO(10)$, there is no Higgs doublet-triplet splitting problem since in the minimal model both Higgs doublets are contained in a $(1,2,2)$ representation. Moreover, since the left-handed quarks and leptons are in the $(4,2,1)$ and the right-handed quarks and leptons in the $(4,1,2)$ representations, the model also leads to third family Yukawa unification as in minimal $SO(10)$. Although the Pati-Salam gauge group is not unified at the field theory level, it readily emerges from string constructions either in the perturbative fermionic constructions [@AnLe], or in the more recent type I string constructions [@ShTy], unlike $SO(10)$ which typically requires large Higgs representations which do not arise from the simplest string constructions. The question of fermion masses and mixing angles in the string-inspired Pati-Salam model has already been discussed for the case of charged fermions [@SFking2; @AlKi3], and later for the case of neutrinos [@AlKi4]. For the neutrino study [@AlKi4] it was assumed that the heavy Majorana neutrino mass matrix was proportional to the unit matrix, and only small neutrino mixing angles were considered. Later on a $U(1)_X$ family symmetry was added to the model, in order to understand the horizontal hierarchies, although in this case the neutrino spectrum was not analysed at all [@AlKiLeLo1].
The purpose of the present paper is to discuss neutrino masses and mixing angles in the string-inspired Pati-Salam model supplemented by a $U(1)_X$ flavour symmetry. The model involves third family Yukawa unification and predicts the top mass and the ratio of the vacuum expectation values $\tan \beta$, as we recently discussed in Ref. [@KiOl2]. It is already known that the model can provide a successful description of the CKM matrix and predicts the down and strange quark masses, although our present analysis differs from that presented previously [@AlKiLeLo1] partly due to the recent refinements in third family Yukawa unification [@KiOl2], but mainly as a result of the recent Super-Kamiokande data which has important implications for the flavour structure of the model. In fact our main focus here is on the neutrino masses and mixing angles which were not previously discussed at all in this framework. We assume a minimal version of the model, and avoid the use of the accidental cancellation mechanism, which in any case has difficulties in accounting for bi-maximal neutrino mixing. We also show that the mixing angle magnification mechanism can only provide limited increases in the mixing angles, due to the fact that the unified third family Yukawa coupling is only approximately equal to 0.7 [@KiOl2] and is therefore too small to have a dramatic effect. Instead, we rely on the SRHND mechanism, and we show how this mechanism may be implemented in the 422 model by appropriate use of operators with “Clebsch zeros” resulting in a natural explanation for atmospheric neutrinos via a hierarchical mass spectrum. We specifically focus on the LMA MSW solution in the text, with the LOW and SMA MSW solutions relegated to Appendices.
The layout of the remainder of the paper is as follows. In section II we briefly review the see-saw mechanism in the Minimal Supersymmetric Standard Model (MSSM) [@MSSM] with right-handed neutrinos. In section III we review some useful analytic results for SRHND, for the case of an approximately diagonal right-handed Majorana mass matrix. In section IV we introduce the string-inspired Pati-Salam model, and in section V we introduce an Abelian anomalous gauge $U(1)_X$ family symmetry into the model, and show how horizontal Yukawa hierarchies may be generated. In section VI we describe our operator approach to fermion masses, including the heavy Majorana neutrino masses. Section VII contains the main results of the paper. In this section we show how a particular choice of $U(1)_X$ family charges, and operators with certain Clebsch coefficients can lead to a successful description of quark and lepton masses and mixing angles, and in particular describe atmospheric and solar neutrinos via SRHND. Although the neutrino masses and mixing angles correspond to the usual LMA MSW solution, in Appendix C we show how a modification of the heavy Majorana mass matrix can lead to a large mixing angle MSW solution with a LOW mass splitting. In Appendix D we present a different choice of $U(1)_X$ charges and operators which can lead to the SMA MSW solution.
The superpotential of the MSSM with right-handed neutrinos is given by : $$\begin{aligned}
{\cal W} &=& {\cal W}_{MSSM}+{\cal W}_{\nu^c} \label{SRNDSuperPot} \\
{\cal W}_{MSSM} &=&
q_A (\lambda_u)_{AB} u^c_B h_u
- q_A (\lambda_d)_{AB} d^c_B h_d
- l_A (\lambda_e)_{AB} e^c_B h_d
+ \mu h_u h_d \\
{\cal W}_{\nu^c} &=&
l_A (\lambda_\nu)_{AB} \nu^c_B h_u
+ {\textstyle {1 \over 2}} \nu^c_A (M_{RR})_{AB} \nu^c_B \end{aligned}$$ where $A,B=1,..,3$ are family indices, $u^c$, $d^c$, $e^c$ and $\nu^c$ are the right-handed $SU(2)_L$ singlet superfields, $q=(u,d)$ and $l=(\nu,e)$ are the $SU(2)_L$ quark and lepton doublets, and $h_u$ ($h_d$) is the up (down) Higgs boson doublet. The Dirac neutrino coupling and the heavy Majorana mass for the right-handed neutrinos are denoted by $\lambda_\nu$ and $M_{RR}$ respectively. When the neutral components of the two MSSM Higgs bosons $h^0_{u,d}$ acquire their vacuum expectation values (VEV)s $v_{2,1}$ ($\tan\beta=v_2/v_1 \sim 40-50$) the superpotential in Eq. generates the following sum of mass terms : $$\begin{aligned}
{\cal L}_{U,D,E} &=&-U (\lambda_u v_2) {U^c}
-D (\lambda_d v_1) {D^c}
-E (\lambda_e v_1) {E^c} + {\rm h.c.} \label{SRNDLagUDE} \\
{\cal L}_{N} &=&-N (\lambda_\nu v_2) {N^c}
-{\textstyle {1\over 2}} {N^c} M_{RR} {N^c} + {\rm h.c.}
\label{SRNDLagN}\end{aligned}$$ where the upper case letters now denote the fermionic components of the superfields in ${\cal W}$, for example $u$ contains $(U,\tilde u)\equiv (U_L,\tilde u_L)$ and $u^c$ contains $(U^c, \tilde u^c)\equiv(U^*_R,\tilde u^*_R)$. The Yukawa matrices in Eq. can be diagonalized by bi-unitary transformations $S$ and $T$ defined by : $$T^{u*} \lambda_u S^{uT} = \lambda'_u \qquad
T^{d*} \lambda_d S^{dT} = \lambda'_d \qquad
T^{e*} \lambda_e S^{eT} = \lambda'_e$$ Thus the physical (primed) states $U'_{R,L}$ are related to the gauge eigenstates $U_{R,L}$ by $U'_R = S^u U_R$ and $U'_L = T^u U_L$, [*etc*]{}.. In this model, the left-handed neutrino masses are generated via see-saw mechanism [@seesaw] by the terms in Eq. which can be re-arranged into a two-by-two block matrix in the following way : $${\cal L}_N = - {\textstyle {1\over 2}}
( N \, N^c )
\left( \matrix{ 0 & m_{LR} \cr
m^T_{LR} & M_{RR} }\right)
\left( \matrix{ N \cr N^c} \right) + {\rm h.c.}$$ where $m_{LR} = \lambda_\nu v_2$. Thus, after the heavy $N^c$ fields are integrated out, the light left-handed neutrinos $N$ effectively acquire a small mass given by : $$m_{LL} = m_{LR} M^{-1}_{RR}\, m_{LR}^T
\label{SRNDSeeSaw}$$ Finally, the diagonalization of $m_{LL}$ : $$T^{N*} m_{LL} T^{N\dagger}_{L} =
{\rm diag}(m_{\nu_1},m_{\nu_2},m_{\nu_3})$$ allows the determination of the masses of the physical neutrinos $m_{\nu_A}$ and enables the physical neutrino states $N' = (\nu_1,\nu_2,\nu_3)$ to be related to the neutrino gauge fields $N = (\nu_e, \nu_\mu, \nu_\tau)$ by $N' = T^N N$.
Taking into account the above conventions, we now proceed to give expressions for the Cabbibo-Kobayashi-Maskawa (CKM) matrix [@CKM] ($V^{CKM}$) and the corresponding lepton analogue, the Maki-Nakawaga-Sakata (MNS) matrix [@MaNaSa] ($V^{MNS}$). Their definitions derive from the charged current interactions : [^3] $$- {g \over \sqrt{2}} W^+_\mu \bar\Psi_U \gamma^\mu P_L \Psi_D \to
- {g \over \sqrt{2}} W^+_\mu \Psi_{U'} \gamma^\mu P_L V^{CKM} \Psi_{D'}$$ $$- {g \over \sqrt{2}} W^-_\mu \bar\Psi_E \gamma^\mu P_L \Psi_{N} \to
- {g \over \sqrt{2}} W^-_\mu \bar\Psi_{E'} \gamma^\mu P_L V^{MNS} \Psi_{N'}$$ that imply : $$V^{CKM} = T^u T^{d\dagger} \qquad
V^{MNS} = T^e T^{N\dagger}
\label{SRNDCKMMNS}$$ In what follows we will assume that the matrices in Eq. are real. [^4] Thus, we will write $V^{MNS}$ in terms of three rotation matrices : $$V^{MNS} = R_{23} R_{13} R_{12}
\label{MNSRot}$$ given by : $$R_{23} = \left( \matrix{ 1 & 0 & 0 \cr
0 & \phantom{-}c_{23} & s_{23} \cr
0 & -s_{23} & c_{23} } \right) \quad
R_{13} = \left( \matrix{ \phantom{-}c_{13} & 0 & s_{13} \cr
0 & 1 & 0 \cr
-s_{13} & 0 & c_{13} } \right) \quad
R_{12} = \left( \matrix{ \phantom{-}c_{12} & s_{12} & 0 \cr
-s_{12} & c_{12} & 0 \cr
0 & 0 & 1 } \right)
\label{RotMtr}$$ where $s_{AB} = \sin\theta_{AB}$, $c_{AB} = \cos\theta_{AB}$ refer to the lepton mixing angles between the $A$ and $B$ generation. Using Eq. into Eq. gives : $$V^{MNS} = \left( \matrix{
c_{12} c_{13} &
s_{12} c_{13} &
s_{13} \cr
-s_{12} c_{23}-c_{12} s_{23} s_{13} &
\phantom{-} c_{12} c_{23}-s_{12} s_{23} s_{13} &
s_{23} c_{13} \cr
\phantom{-} s_{12} s_{23}-c_{12} c_{23} s_{13} &
-c_{12} s_{23}-s_{12} c_{23} s_{13} &
c_{23} c_{13} } \right)
\label{VMNS}$$ It is also practical to have expressions for the $\theta_{AB}$ angles in terms of the $V^{MNS}$ entries. Inverting Eq. we find that : [^5] $$\sin\theta_{13} = V_{e3} \qquad
\sin\theta_{23} = {V_{\mu 3} \over \sqrt{1-V_{e3}^2}} \qquad
\sin\theta_{12} = {V_{e2} \over \sqrt{1-V_{e3}^2}}$$
Finally we note that while the above expressions were derived in the context of three neutrino species, the analysis of the experimental results assumed only two, thus a direct comparison of mixing angles is not exactly valid.
Third family single right-handed neutrino dominance (SRHND) [@KingSRND; @SFKing1; @SFKing2] is a mechanism that can explain the large atmospheric ($\theta_{23}$) and the solar LMA MSW ($\theta_{12}$) neutrino mixing angles and a small $\theta_{13}$. SRHND relies on the possibility that the neutrino mass matrix ($m_{LL}$) is dominated by the contributions coming solely from a single right-handed neutrino (for example $\nu^c_\tau$.) In this scheme a maximal $\theta_{23}$ angle arises when the tau right-handed neutrino $\nu^c_\tau$ couples to the left-handed muon $\nu_\mu$ and tau neutrino $\nu_\tau$ with equal strength. Similarly, if $\nu^c_\mu$ couples to $\nu_e$ and to $\nu_\mu$ with comparable strength then $\theta_{12}$ is large. The role of the (sub-dominant) muon neutrino is also important since it provides small perturbations to the $m_{LL}$ matrix (which otherwise has one heavy and two massless eigenstates), thus leading to a neutrino mass splitting $\Delta m^2_{12}=|m^2_{\nu_2}-m^2_{\nu_1}|$ compatible with experiment. In this section we summarize the theory behind SRHND and review the analytic results presented in Ref. [@SFKing2] for the case of the diagonal dominated right-handed neutrino mass matrix $M_{RR}$.
The see-saw formula for the left-handed neutrino matrix in Eq. depends explicitly on $M_{RR}$. Although $M_{RR}$ might have a non-trivial structure we find instructive to start our analysis by considering the very simple case of $M_{RR}$ given by : $$M^{-1}_{RR} \sim {\rm diag}(M^{-1}_{\nu_1},M^{-1}_{\nu_2},M^{-1}_{\nu_3}) \sim
{\rm diag}(0,0,M^{-1}_{\nu_3})
\label{SRNDmRR}$$ which effectively corresponds to taking $M_{\nu_1},M_{\nu_2} \gg M_{\nu_3}$. Replacing Eq. into Eq. we find that : [^6] $$m_{LL} =
v_2^2 \lambda_\nu M^{-1}_{RR} \lambda^T_\nu
\sim
\lambda M^{-1}_{RR} \lambda^T
\sim
M^{-1}_{\nu_3}
\left(\matrix{
\lambda_{13}^2 & \lambda_{13} \lambda_{23} & \lambda_{13} \lambda_{33} \cr
\lambda_{13} \lambda_{23} & \lambda_{22}^2 & \lambda_{23} \lambda_{33} \cr
\lambda_{13} \lambda_{33} & \lambda_{23} \lambda_{33} & \lambda_{33}^2 \cr}
\right)
\label{SRNDmLLMTR1}$$ The $m_{LL}$ matrix above is easily diagonalized by the matrices $R_{23}$, $R_{13}$, $R_{12}$ [^7] in Eq. with rotation angles given by : $$\matrix{
\displaystyle
s_{23} = {\lambda_{23} \over A} &
\displaystyle
c_{23} = {\lambda_{33} \over A} & &
A^2 = \lambda_{33}^2+\lambda_{23}^2 \hfill \cr
& & \qquad & \cr
\displaystyle
s_{13} = {\lambda_{13} \over B} &
\displaystyle
c_{13} = { A \over B} & &
\displaystyle
B^2 = \lambda_{33}^2+\lambda_{23}^2+\lambda_{13}^2
}
\label{SRNDRotAngles}$$ that successively act on $m_{LL}$ as follows : $$m_{LL}^{\prime\prime\prime} =
R_{12}^\dagger R_{13}^\dagger R_{23}^\dagger m_{LL} R_{23} R_{13} R_{12} =
{\rm diag}(m_{\nu_1},m_{\nu_2},m_{\nu_3})$$ It is also convenient to define the following primed matrices : $$m_{LL}^{\prime} = R_{23}^\dagger m_{LL} R_{23} \quad\quad
m_{LL}^{\prime\prime} = R_{13}^\dagger m_{LL}^{\prime} R_{13} \quad\quad
m_{LL}^{\prime\prime\prime} = R_{12}^\dagger m_{LL}^{\prime\prime} R_{12}
\label{SRNDmLLprime}$$ which, for $m_{LL}$ as in Eq. , are explicitly given by : $$m_{LL}^\prime \sim M^{-1}_{\nu_3} \left(\matrix{
\lambda_{13}^2 & 0 & \lambda_{13} A \cr
0 & 1 & 0 \cr
\lambda_{13} A & 0 & A^2}\right)\qquad
m_{LL}^{\prime\prime} \equiv m_{LL}^{\prime\prime\prime}
\sim M^{-1}_{\nu_3} \left(\matrix{
0 & 0 & 0 \cr 0 & 0 & 0 \cr 0 & 0 & B^2}\right)\quad
\label{SRNDmLLprimeMTR}$$
We can see from Eq. that if $\lambda_{23} =
\lambda_{33}$ then a maximal $\theta_{23}=45^0$ angle results. Moreover, if $\lambda_{13} \ll \lambda_{23},\lambda_{33}$ then $\theta_{13}$ is small. Although SRHND, in the limiting case of Eq. , is successful in predicting a maximal atmospheric neutrino angle, it fails to account for a viable neutrino spectrum. Indeed, from Eq. , we see that the two lightest neutrinos are massless $m_{\nu_1} = m_{\nu_2} = 0$. Moreover the solar neutrino angle $\theta_{12}$ is undetermined. These two problems can be solved by allowing the right-handed muon neutrino $\nu^c_\mu$ to play a sub-dominant/perturbative role in the structure of $m_{LL}$ in Eq. .
We now turn to the more realist model in which $M_{RR}$ can be approximated by : [^8] $$M^{-1}_{RR} \sim
{\rm diag}(M^{-1}_{\nu_1},M^{-1}_{\nu_2},M^{-1}_{\nu_3}) \sim
{\rm diag}(0,M^{-1}_{\nu_2},M^{-1}_{\nu_3})
\label{SRNDmLLMTR2}$$ Using Eq. into Eq. we find that : $$m_{LL} \sim
M^{-1}_{\nu_3}
\left(\matrix{
\lambda_{13}^2 &
\lambda_{13}\lambda_{23} &
\lambda_{13}\lambda_{33} \cr
\lambda_{13}\lambda_{23} &
\lambda_{23}^2 &
\lambda_{23}\lambda_{33} \cr
\lambda_{13}\lambda_{33} &
\lambda_{23}\lambda_{33} &
\lambda_{33}^2
}\right)+
M^{-1}_{\nu_2}
\left(\matrix{
\lambda_{12}^2 &
\lambda_{12}\lambda_{22} &
\lambda_{12}\lambda_{32} \cr
\lambda_{12}\lambda_{22} &
\lambda_{22}^2 &
\lambda_{22}\lambda_{32} \cr
\lambda_{12}\lambda_{32} &
\lambda_{22}\lambda_{32} &
\lambda_{32}^2
}\right)
\label{SRNDmLLM2M3}$$
Given that we assumed SRHND by the $\nu^c_\tau$ neutrino, it follows that the contributions to the 23 block of $m_{LL}$ in Eq. arising from the terms proportional to $M^{-1}_{\nu_3}$ dominate over the ones proportional to $M^{-1}_{\nu_2}$. [^9] Clearly, the rotations $R_{12}$, $R_{13}$ parameterised by the angles in Eq. diagonalize $m_{LL}$ in Eq. up to terms of order ${\cal O}(M^{-1}_{\nu_2})$. Thus the new primed matrices $m^{\prime}_{LL}$ and $m^{\prime\prime}_{LL}$ are given by : $$\begin{aligned}
m_{LL}^{\prime} &\sim&
M^{-1}_{\nu_3}
\left(\matrix{
\lambda_{13}^2 &
0 &
\lambda_{13} A \cr
0 &
0 &
0 \cr
\lambda_{13} A &
0 &
A^2
}\right)+
M^{-1}_{\nu_2}
\left(\matrix{
\lambda_{12}^2 &
\lambda_{12} {C^2 \over A} &
\lambda_{12} {D^2 \over A} \cr
\lambda_{12} {C^2 \over A} &
{C^4 \over A^2} &
{C^2 D^2 \over A^2} \cr
\lambda_{12} {D^2 \over A} &
{C^2 D^2 \over A^2} &
{D^4 \over A^2}
}\right) \\
& &
\nonumber \\
m_{LL}^{\prime\prime} &\sim&
M^{-1}_{\nu_3}
\left(\matrix{
0 & 0 & 0 \cr
0 & 0 & 0 \cr
0 & 0 & B^2
}\right)+
M^{-1}_{\nu_2}
\left(\matrix{
{E^6 \over A^2 B^2} &
{C^2 E^3 \over A^2 B} &
{F^2 E^3 \over A B^2} \cr
{C^2 E^3 \over A^2 B} &
{C^4 \over A^2} &
{C^2 F^2 \over AB} \cr
{F^2 E^3 \over A B^2} &
{C^2 F^2 \over AB} &
{F^4 \over B^2}
}\right)
\label{SRNDmLLpp}\end{aligned}$$ where $$\begin{aligned}
C^2 &=& \lambda_{22}\lambda_{33}-\lambda_{32}\lambda_{23} \\
D^2 &=& \lambda_{33}\lambda_{32}+\lambda_{22}\lambda_{23} \\
E^3 &=& \lambda_{12} (\lambda_{33}^2+\lambda_{23}^2) -
\lambda_{13} (\lambda_{33}\lambda_{32}+\lambda_{22}\lambda_{23}) \\
F^2 &=& \lambda_{33}\lambda_{32}+\lambda_{22}\lambda_{23}+\lambda_{12}\lambda_{13}\end{aligned}$$ The diagonalization of the 12 block of $m^{\prime\prime}_{LL}$ in Eq. is achieved by a $R_{12}$ matrix parameterised by the following $\theta_{12}$ rotation angle : $$s_{12} ={E^3 \over \sqrt{E^6+B^2 C^4}} \qquad
c_{12} ={B C^2 \over \sqrt{E^6+B^2 C^4}}$$ Thus we find : $$m^{\prime\prime\prime}_{LL} \sim
M^{-1}_{\nu_3}
\left(\matrix{
0 & 0 & 0 \cr
0 & 0 & 0 \cr
0 & 0 & B^2
}\right)+
M^{-1}_{\nu_2}
\left(\matrix{
0 & 0 & 0 \cr
0 & {E^6+B^2 C^4 \over A^2 B^2} & {F^2 \sqrt{E^6+B^2 C^4} \over A B^2} \cr
0 & {F^2 \sqrt{E^6+B^2 C^4} \over A B^2} & {F^4 \over B^2}
\label{SRNDmLLppp}
}\right)$$ It is interesting to note that the $R_{12}$ rotation has not only diagonalized the 12 block of $m^{\prime\prime}_{LL}$ but also put zeros in the 13,31 entries of $m^{\prime\prime\prime}_{LL}$. The reason is because $m^{\prime\prime}_{LL}$ displays a special structure. Indeed, their elements obey : $$t_{12} =
{s_{12} \over c_{12}} =
{(m^{\prime\prime}_{LL})_{12} \over (m^{\prime\prime}_{LL})_{22}} =
{(m^{\prime\prime}_{LL})_{11} \over (m^{\prime\prime}_{LL})_{12}} =
{(m^{\prime\prime}_{LL})_{13} \over (m^{\prime\prime}_{LL})_{23}} =
{E^3 \over B C^2}$$ Explicitly $t_{12} = \tan\theta_{12}$ is given by : $$t_{12} =
{\lambda_{12}(\lambda_{33}^2+\lambda_{23}^2)-
\lambda_{13}(\lambda_{33}\lambda_{32}+\lambda_{22}\lambda_{23}) \over
(\lambda_{22}\lambda_{33}-\lambda_{32}\lambda_{23})
\sqrt{\lambda_{33}^2+\lambda_{23}^2+\lambda_{13}^2}}
\sim {\lambda_{12} \over \lambda_{22}}
\label{SRNDt12}$$ From Eq. we see that, although $t_{12}$ generally depends on the second and third family neutrino Yukawa couplings, if $\lambda_{33}$ is much bigger than the other Yukawa couplings then $t_{12} \sim \lambda_{12} / \lambda_{22}$. This means that the $\theta_{12}$ angle is set not by the dominant neutrino couplings, but by the sub-dominant $\nu^c_\mu$ neutrino couplings to the $\nu_e$ and $\nu_\mu$ neutrinos. Thus while a large atmospheric neutrino mixing angle $\theta_{23}$ can be achieved by requiring $\lambda_{23}\sim\lambda_{33}$, a large MSW solar neutrino angle $\theta_{12}$ results from $\lambda_{12}\sim\lambda_{22}$. Moreover, Eq. and Eq. show that bi-maximal $\theta_{23}$, $\theta_{12}$ mixing can be achieved with a small $\theta_{13}$ angle as long as $\lambda_{13} \ll \lambda_{23},\lambda_{33}$. The neutrino mass spectrum can be read from $m^{\prime\prime\prime}_{LL}$ in Eq. . We find a massless neutrino state $m_{\nu_1} = 0$, plus a light state with mass $m_{\nu_2} \sim \lambda^2_{22} / M_{\nu_2}$ and a heavy neutrino with mass $m_{\nu_3} \sim \lambda^2_{33} / M_{\nu_3}$.
Here we briefly summarize the parts of the Pati-Salam model [@PaSa] that are relevant for our analysis. For a more complete discussion see Ref. [@AnLe]. The SM fermions, together with the right-handed neutrinos, are conveniently accommodated in the following $F=(4,2,1)$ and $F^c=(\bar 4,1,\bar 2)$ representations : $$\qquad
F_A = \left(\matrix{
u & u & u & \nu \cr
d & d & d & e \cr}
\right)_A
\qquad
F^c_B = \left(\matrix{
d^c & d^c & d^c & e^c \cr
u^c & u^c & u^c & \nu^c \cr}
\right)_B
\label{FcFields}$$ The MSSM Higgs bosons fields are contained in $h=(1,\bar 2,2)$ : $$h = \left(\matrix{h_d^- & h_u^0 \cr
h_d^0 & h_u^+ \cr} \right)
\label{SMHiggs}$$ whereas the heavy Higgs bosons $\bar H=(\bar 4,1,\bar 2)$ and $H=(4,1,2)$ are denoted by : $$\bar H = \left(\matrix{ \bar H_d & \bar H_d & \bar H_d & \bar H_e \cr
\bar H_u & \bar H_u & \bar H_u & \bar H_\nu
\cr}\right)
\qquad
H = \left(\matrix{ H_d & H_d & H_d & H_e \cr
H_u & H_u & H_u & H_\nu \cr}\right).
\label{RHiggs}$$ In addition to the Higgs fields in Eqs. , the model also involves an $SU(4)$ sextet field $D=(6,1,1)=(D_3,D^c_3)$.
The superpotential of the minimal 422 model is : $$\begin{aligned}
{\cal W} &=& F \lambda F^c h + \lambda_h S h h + \nonumber \\
& & \lambda_S S (\bar H H - M^2_H) +
\lambda_H H H D + \lambda_{\bar H} {\bar H} {\bar H} D +
F^c \lambda' F^c \,{H H \over M_V}
\label{W3}\end{aligned}$$ where $S$ denotes a gauge singlet superfield, the $\lambda$’s are real dimensionless parameters and $M_H \sim M_X \sim 10^{16} {\rm\ GeV}$. Additionally, $M_V > M_X$ denotes the mass of extra exotic matter that has been integrated out from the model at high energy. As a result of the superpotential terms involving the singlet $S$, the Higgs fields develop VEVs $\langle H \> \rangle = \langle H_\nu \rangle \sim M_X$ and $\langle \bar H \rangle = \langle \bar H_\nu \rangle \sim M_X$ which lead to the symmetry breaking : $$SU(4) \otimes SU(2)_L \otimes SU(2)_R \to
SU(3)_c \otimes SU(2)_L \otimes U(1)_Y.
\label{422:321}$$ The singlet $S$ itself also naturally develops a small VEV of the order of the SUSY breaking scale [@KiSh] so that the $\lambda_h S$ term in Eq. gives an effective $\mu$ parameter of the correct order of magnitude. Under Eq. the Higgs field $h$ in Eq. splits into the familiar MSSM doublets $h_u$ and $h_d$ whose neutral components subsequently develop weak scale VEVs $\langle h_u^0 \rangle = v_2$ and $\langle h_d^0 \rangle = v_1$ with $\tanb = v_2/v_1$. The neutrino fields $\nu^c$ acquires a large mass $M_{RR} \sim \lambda' \langle HH \rangle / M_V$ through the non-renormalizable term in ${\cal W}$ which, together with the Dirac $\nu^c$ – $\nu$ interaction (proportional to $\lambda \langle h_u^0 \rangle$), gives rise to a 2 $\times$ 2 matrix that generates, via see-saw mechanism [@seesaw], a suppressed mass for the left-handed neutrino states. The $D$ field does not develop a VEV but the terms $HHD$ and $\bar H \bar H D$ combine the colour triplet parts of $H$, $\bar H$ and $D$ into acceptable GUT scale mass terms [@AnLe].
The pattern of fermion masses and mixing angles is one of the fundamental problems is particle physics that has not yet been understood. The importance of this unsolved puzzle is demonstrated by the numerous works published in the literature over the past years (see Refs. [@MatrixModels1]-[@BaDo] for a “short” list.) In the standard model (SM) the quark/lepton masses and the CKM matrix are input parameters fixed by laboratory experiments. Surprisingly, however, their values, though unconstrained and [*a priori*]{} arbitrary, do display a certain degree of organization. The fermion masses are highly hierarchical and the CKM matrix can be described in terms of the small Wolfenstein expansion parameter $\lambda \sim |V_{12}| \sim 0.22$ [@LWolfenstein]. These results suggest that a broken flavour symmetry might be playing an important role in the setting of the structure of the Yukawa matrices.
In this work we will assume that the “vertical” gauge group is supplemented by an additional $U(1)_X$ “horizontal” flavour symmetry that constraints the nature of the couplings of quarks and leptons to SM singlet fields $\theta$ and $\bar\theta$. The family symmetry, however, is broken at some high energy scale $M_\theta > M_X$ by the VEVs of the $\theta$, $\bar\theta$ fields which under the $U(1)_X$ group have charges $X_\theta = -1$ and $X_{\bar\theta}= +1$. As a consequence of the $U(1)_X$ symmetry breaking, the low energy effective theory includes Dirac interactions between the $F$ and $F^c$ fields of the following form : $$F_A F^c_B h \left( {\theta}
\over M_V \right)^{p_{AB}} \to
F_A F^c_B h \left( \langle\theta\rangle
\over M_V \right)^{p_{AB}} \sim
F_A F^c_B h \epsilon^{p_{AB}}
\label{NonRenTheta}$$ $$F_A F^c_B h \left( {\bar\theta}
\over M_V \right)^{p_{AB}} \to
F_A F^c_B h \left( \langle\bar\theta\rangle
\over M_V \right)^{p_{AB}} \sim
F_A F^c_B h \epsilon^{p_{AB}}
\label{NonRenThetaBar}$$ where $p_{AB}$ is the modulos of the sum of the $U(1)_X$ charges of the $F_A$, $F^c_B$ and $h$ fields, [*i.e.*]{} $p_{AB} = |X_{AB}| = |X_{F_A}+X_{F^c_B}+X_h|$. Thus Eq. holds if $X_{AB} > 0$ whereas Eq. holds if $X_{AB} < 0$. The non-renormalizable terms in Eqs. , might originate from interactions between the $F$ and $\theta$ fields with additional exotic vector matter with mass $M_V > M_X$ that lead to “spaghetti” diagrams as discussed in Ref. [@AlKiSpaghetti]. In summary, the equations above show that, in the context of a $U(1)_X$ symmetry, the observed hierarchy in the fermion masses and mixing angles might be the result of the flavour charges carried by the fields of the 422 model which act to suppress the Yukawa couplings by some $\epsilon$-power.
The introduction of the $U(1)_X$ symmetry provides a way to relate the various flavour parameters of the model thus making it more predictive. However, one should be careful. Generally the $U(1)_X$ group is potentially dangerous since it can introduce, through triangle diagrams, mixed anomalies with the SM gauge group. [^10] In the last part of this section we review the constraints imposed on $X$ charges of the fields of our model enforced by the requirement of anomaly cancellation [@JaSh].
The mixed anomalies that we shall consider are : [^11] $$\begin{aligned}
\!\!\!\!\!\!\!\!
SU(3)^2 U(1)_X & : &
A_3 = \sum_{A=1}^3 (2 X_{q_A}+ X_{u^c_A}+ X_{d^c_A}) \label{AnomA3} \\
\!\!\!\!\!\!\!\!
SU(2)^2 U(1)_X & : &
A_2 = \sum_{A=1}^3 (3 X_{q_A}+ X_{l_A})+ X_{h_u}+ X_{h_d} \label{AnomA2} \\
\!\!\!\!\!\!\!\!
U(1)^2_Y U(1)_X & : &
A_1 = \sum_{A=1}^3 ({\textstyle {1\over 3}} X_{q_A}+
{\textstyle {8\over 3}} X_{u^c_A}+
{\textstyle {2\over 3}} X_{d^c_A}+
X_{l_A}+ 2 X_{e^c_A})+ X_{h_u}+ X_{h_d} \label{AnomA1} \\
\!\!\!\!\!\!\!\!
U(1)_Y U(1)_X^2 & : &
A'_1 = \sum_{A=1}^3 ( X_{q_A}^2- 2X_{u^c_A}^2+X_{d^c_A}^2
-X_{l_A}^2+ X_{e^c_A}^2) + X_{h_u}^2+ X_{h_d}^2 \label{AnomA1p}\end{aligned}$$ For example, $A_3$ corresponds to the anomalous term generated by the Feynman diagram that has two $SU(3)$ gluons and one $U(1)_X$ gauge boson attached to the triangle vertices. We note that the first three anomalies $A_3$, $A_2$ and $A_1$ are linear in the trace of the charges, [*i.e.*]{} $X_f = \sum_{A=1}^3 X_{f_A}$, where $f$ is any of the $q,u^c,d^c,l,e^c$ fields, thus they constraint only the family independent (FI) part of the $U(1)_X$ charges. On the other hand, $A'_1$ is quadratic in the $X$ charges, thus it generally constraints the FI and family dependent (FD) part of the $U(1)_X$ charges.
In this paper we will assume that the cancellation of anomalies results from the Green-Schwartz (GS) mechanism [@GrSc]. This is possible if the $A_3$, $A_2$ and $A_1$ anomalies are in the ratio $A_3:A_2:A_1=k_3:k_2:k_1$ where the $k_i$ are the Kac-Moody levels of the $SU(3)$, $SU(2)$ and $U(1)_Y$ gauge groups that determine the boundary conditions for the gauge couplings at the string scale $g_3^2 k_3 : g_2^2 k_2 : g_1^2 k_1$. Hence, using the canonical GUT normalization for the gauge couplings (that successfully predicts $\sin^2(\theta_W)=3/8$ [@Ibanez]), anomalies can be cancelled if we require that : $$A_3=A_2={\textstyle {5 \over 3}}A_1
\label{AnomCanc}$$ As a consequence of the two constraints implicit in Eq. , the set of solutions for the $X$ charges appearing in Eqs. (\[AnomA3\])-(\[AnomA1\]) is given by [@JaSh] : $$\matrix{
\displaystyle
X_{e^c} = \sum_{A=1}^3 X_{e^c_A} = x \hfill &
\displaystyle
X_{l^c} = \sum_{A=1}^3 X_{l_A} = y \hfill &
\displaystyle
X_{h_u} = -z \hfill \cr
\displaystyle
X_{q\phantom{^c}} = \sum_{A=1}^3 X_{q_A} = x+u \hfill &
\displaystyle
X_{d^c} = \sum_{A=1}^3 X_{d^c_A} = y+v \hfill &
\displaystyle
X_{h_d} = +z+(u+v) \hfill \cr
\displaystyle
X_{u^c} = \sum_{A=1}^3 X_{u^c_A} = x+2u \hfill &
& }
\label{AnomSol}$$ where $x,y,z,u,v$ are free parameters. However not all the solutions in Eqs. are valid after $A'_1=0$ is enforced. In fact, as we said before, generally $A'_1$ constrains both the FI and FD charges of $U(1)_X$. By this we mean that, if we conveniently write the charge of the $f_A$ field $X_{f_A}$ as a sum of a FI part $X_f$ plus a FD part $X'_{f_A}$, [*i.e.*]{} $X_{f_A}= {1 \over 3} X_f+X'_{f_A}$, then $A'_1=0$ is a complicated equation on all $X_f$, $X'_{f_A}$ and $X_{h_u}$, $X_{h_d}$ charges. However, it is easy to see that, if all the left-handed fields and if all the right-handed fields have the same FD charges, [*i.e.*]{} $X'_{q_A}=X'_{l_A}$ and $X'_{u^c_A}=X'_{d^c_A}=X'_{e^c_A}$, as is the case of the 422 model, then $A'_1=0$ is an equation on the FI charges only : $$A'_1 = {\textstyle {2 \over 3}}
(X^2_q- 2 X^2_{u^c}+ X^2_{d^c}- X^2_{l}+ X^2_{e^c}+ X^2_{h_u}-X^2_{h_d})=0$$ Thus, a simple solution to all the anomaly constraints is given by Eq. with $u=v=0$. Finally, we must add that since the Pati-Salam model unifies all the left/right-handed quark and lepton fields in the $F$/$F^c$ multiplets, and the MSSM Higgs fields $h_u$, $h_d$ in the $h$ Higgs bi-doublet we must also have $x=y$ and $z=0$. Thus, anomaly cancellation in the 422 model via GS mechanism is possible if the traces of the $U(1)_X$ charges of the $F$ and $F^c$ fields are equal, [*ie*]{} $X_{F} = \sum_{A=1}^3 X_{F_A} \equiv \sum_{A=1}^3 X_{F^c_A} = X_{F^c}$.
In the simplest formulation of the 422 model extended by a $U(1)_X$ horizontal symmetry all the Yukawa couplings originate from a single matrix. The Abelian $U(1)_X$ symmetry introduced in the previous section mainly serves one purpose, it establishes an hierarchy between the flavour dependent couplings. Thus, it provides no precise/predictive information about the relationships between the different Yukawa coupling matrices. As a result, all the SM fermions of a given family have identical Yukawa couplings at the unification scale. Naturally, when the fermion masses are run from the $M_X$ to the $M_Z$ scale they lead to quark and lepton masses that are incompatible with the experimental data.
The idea of Yukawa unification, though unsuccessful in its most simpler form, is not, however, a complete failure. As a matter of fact, it turns out that third family Yukawa unification works rather well. It is well known that the GUT boundary condition for the Yukawa couplings : $$\lambda_t(M_X)=\lambda_b(M_X)=\lambda_\tau(M_X)=\lambda_{\nu_\tau}(M_X)$$ leads to a large pole top mass prediction $M_t \sim 175$ GeV and $\tan\beta\sim m_t/m_b$. On the other hand, the first and second family fermion masses can be predicted if special relations between the “vertical” intra-generation Yukawa couplings at $M_X$ hold. For example, the Georgi-Jarlskog (GJ) [@GeJa] relation between the muon and strange Yukawa couplings $\lambda_\mu \sim 3 \lambda_s$ successfully reproduces the low energy experimental $m_s/m_\mu \sim 1$ mass ratio. In the context of GUT theories the appearance of numerical factors relating the couplings of the up-down-lepton Yukawa matrices might originate from non-renormalizable operators involving the interaction between the fermions and the heavy Higgs that break the GUT symmetry [@FrNi; @AnRaDiHaSt].
In the Pati-Salam model, we will have in mind operators of the following form [@AlKiLeLo1] : $$F_A F^c_B h \left( H \bar H \over M_V^2 \right)^{n}
\left(\theta \over M_V \right)^{p_{AB}}
\quad{\rm\ and}\quad
F_A F^c_B h \left( H \bar H \over M_V^2 \right)^{n}
\left(\bar\theta \over M_V \right)^{p_{AB}}
\label{SRNDopRL}$$ The idea is that when the $H$ and $\theta$ fields develop their VEVs such operators reduce to effective Yukawa couplings with small coefficients. For example, if $F_2$, $F^c_2$ and $h$ carry a charge $X_{F_2} = 0$, $X_{F^c_2} = 2$ and $X_h = 0$ under the $U(1)_X$ symmetry then Eq. (with $n=1$) generates the following terms : $$(x_u u_2 u^c_2 h^0_u+x_d d_2 d^c_2 h^0_d+
x_e e_2 e^c_2 h^0_d+x_\nu \nu_2 \nu^c_2 h^0_u)\delta\epsilon^2$$ where $\delta=\langle H \rangle\langle\bar H \rangle/M_V^2$ and $\epsilon=\langle\theta\rangle / M_V$ are small dimensionless parameters, $u_2$, $d_2$, $e_2$, $\nu_2$ are the charm, strange, muon, muon neutrino superfields, and $x_f$ ($f=u,d,e,\nu$) are Clebsch factors that depend on the group theoretical contractions between the fields in Eq. [@SFking2; @AlKi3]. In Table (Appendix A) we present a complete list of all $x_f$ values that result from $n=1$ operators in the 422 model [@AlKiLeLo1] normalized by : $$x_u^2+x_d^2+x_e^2+x_\nu^2 = 4$$ It is interesting to point out that different operators imply zero Clebsches for different $x_f$’s. For example, CLASS-I operators are rather special since of all $x_f$’s only one is non-zero (and significantly large). The CLASS-II operators have $x_u=x_\nu=0$ while CLASS-III have $x_d=x_e=0$. Additionally CLASS-IV operators have $x_u=x_d=0$ and CLASS-V have $x_e=x_\nu=0$. Finally CLASS-VI operators have all $x_f$’s different from zero. The variety of the operator Clebsches is to be welcome since, as we will see, they open the possibility of avoiding the disastrous fermion mass predictions characteristic of the minimal 422 model with a unified renormalizable interaction.
Finally we shall mention the origin of the heavy Majorana neutrino mass matrix. Generally $M_{RR}$ results from non-renormalizable operators of the form : $$F^c_A F^c_B \left( H H \over M_V^2 \right)
\left( H \bar H \over M_V^2 \right)^n
\left(\theta \over M_V \right)^{q_{AB}}
\to
\nu^c_A \nu^c_B \delta^{n+1}\epsilon^{q_{AB}}
\label{SRNDopRR}$$ where $q_{AB}=|X_{F^c_A}+X_{F^c_B}+\sigma|$ and $\sigma = 2 X_H$. Three important differences distinguish Eq. from Eq. . Firstly we note that while Eq. allows for renormalizable operators, $M_{RR}$ as given by Eq. is always the result of non-renormalizable operators. Secondly we note that the combination of the $HH$ fields in Eq. introduces an additional free parameter $\sigma$ that may be fixed at our convenience. Thirdly we observe that while Eq. is able to generate precise relationships between the up-down-lepton Yukawa couplings (via Clebsch factors), Eq. is an expression that constrains only the hierarchy of $M_{RR}$ (via the $U(1)_X$ symmetry), as a result it is less predictive.
In this section we show how the $U(1)_X$ horizontal family symmetry of section V can be combined with the operator approach of section VI to give predictions for the fermion masses and mixing angles in the 422 model. In particular we are interested in the predictions for the neutrino masses and mixing angles for the LMA MSW solution to the solar neutrino problem. The LOW and the SMA MSW solutions are discussed in Appendices C and D. We start by listing the quark and charged lepton experimental data used in our analysis : [^12] $$\begin{aligned}
m_u({\rm 1\ GeV}) &=& 4.7 {\rm\ MeV} \qquad
\hfill
(1.35-6.75) {\rm\ MeV} \label{SRNDdatamu} \\
m_c(M_c) &=& 1.21 {\rm\ GeV} \qquad
\hfill
(1.15-1.35) {\rm\ GeV} \\
M_t & \sim & 175 {\rm\ GeV} \qquad
(170-180) {\rm\ GeV}\end{aligned}$$ $$\begin{aligned}
m_d({\rm 1\ GeV}) & \sim & 6.0 {\rm\ MeV} \qquad
\hfill
(4-12) {\rm\ MeV} \\
m_s(M_s) & \sim & 160 {\rm\ MeV} \qquad
\hfill
(100-230) {\rm\ MeV} \\
m_b(M_b) &=& 4.15 {\rm\ GeV} \qquad
\hfill
(4.0-4.4) {\rm\ GeV}\end{aligned}$$ $$\begin{aligned}
M_e &=& 0.511 {\rm\ MeV} \qquad
m_e(M_e) = 0.496 {\rm\ MeV} \\
M_\mu &=& 105.7 {\rm\ MeV} \qquad
m_\mu(M_\mu) = 104.6 {\rm\ MeV} \\
M_\tau &=& 1777.0 {\rm\ MeV} \qquad
m_\tau(M_\tau) = 1772.8 {\rm\ MeV} \label{SRNDdatamtau}\end{aligned}$$ where $m_u({\rm 1\ GeV})$, $m_d({\rm 1\ GeV})$ denote the running masses of the up and down quarks at $Q=1{\rm\ GeV}$; [^13] $m_c(M_c)$, $m_s(M_c)$, $m_b(M_b)$ the running masses of the charm, strange and bottom quarks at their pole masses ($M_c = 1.6 {\rm\ GeV}$, $M_b = 4.8 {\rm\ GeV}$); $M_t$ the top pole mass; and $M_{e,\mu,\tau}$ ($m_{e,\mu,\tau}$) the well known pole (running) charged lepton masses. We converted the above pole masses to running masses using the expressions in Ref. [@ArCaKeMiPiRaWr] with : $$\alpha_s(M_Z) = 0.120 \qquad\qquad
\alpha_e^{-1}(M_Z) = 127.8$$ Finally the CKM matrix at $Q=M_Z$ was fixed by : [^14] $$|V_{12}| = 0.2215 \qquad |V_{23}|= 0.040 \qquad |V_{13}| = 0.0035
\label{SRNDdataCKM}$$ It is important to note that, in fact, not all the parameters above were taken as an input. Indeed, $M_t \sim 175 {\rm\ GeV}$ is a prediction that results from third family Yukawa unification $\lambda_t = \lambda_b = \lambda_\tau$ at the GUT scale. Moreover, as we will see, our model is also able to predict the masses of the down and charm quarks, thus their values listed above should be taken merely as a guide and/or convenient initial estimates.
We now turn to the neutrino experimental data. The results from the Super-Kamiokande collaboration [@SKamiokandeColl; @HSobel] indicate that the atmospheric neutrino anomaly can be understood in terms of $\nu_\mu\leftrightarrow\nu_\tau$ oscillations with : $$\sin^2(2\theta_{23}) > 0.88 \qquad\qquad
1.5 \times 10^{-3}{\rm\ eV}^2 < \Delta m^2_{23} < 5\times 10^{-3} {\rm\ eV}^2
\label{SRNDSKdata}$$ at 90 % confidence level. On the other hand, the large mixing angle (LMA) MSW solution to the solar neutrino deficit suggests that [@BaKrSm] : $$\sin^2(2\theta_{12}) \sim 0.75 \qquad\qquad
\Delta m^2_{12} \sim 2.5 \times 10^{-5} {\rm\ eV}^2
\label{SRNDLMAMSWdata}$$ Assuming that the neutrino spectrum is hierarchical, [*i.e.*]{} $\Delta m_{23}^2 = |m^2_{\nu_3}-m^2_{\nu_2}| \sim m^2_{\nu_3}$ and $\Delta m_{12}^2 = |m^2_{\nu_2}-m^2_{\nu_1}| \sim m^2_{\nu_2}$ the values in Eqs. , give : $$\sin(\theta_{23}) > 0.57 \quad m_{\nu_3} \sim 0.05 {\rm\ eV}
\qquad\qquad
\sin(\theta_{12}) \sim 0.50 \quad m_{\nu_2} \sim 0.005 {\rm\ eV}
\label{SRNDneutrinoDATA}$$ The latest results from the CHOOZ experiment also show that, over the interesting $\Delta m^2_{23}$ range suggested by the Super-Kamiokande data, $\sin^2(2\theta_{13})<0.10$ at 90% CL [@CHOOZ].
The experimental data in Eqs. - constrains the parameters of our model at low energy. However, the GUT symmetry is broken at an energy $M_X \sim 10^{16} {\rm\ GeV}$. Thus, before we start our analysis we should correct the fermion masses and mixing angles for the radiative corrections that result from the running of the RGEs between the $Q=M_Z$ and $Q=M_X$ scales. The implementation of the RGEs, decoupling of SUSY particles and boundary conditions is a complicated subject whose detailed description is beyond the scope of this work. Here we will only mention that we used 2-loop RGEs in the gauge and Yukawa couplings and refer the interested reader to Ref. [@KiOl2] where the issue of Yukawa unification in the 422 model is discussed. As a result of the RGEs running, subjected to third family Yukawa unification at $M_X$, the low energy input values for the fermion masses in Eqs. - effectively constraint the eigenvalues of the Yukawa couplings at $Q=M_X$ to be : $$\lambda_u(M_X) = 4.738\times 10^{-6} \qquad
\lambda_c(M_X) = 1.529\times 10^{-3} \qquad
\lambda_t(M_X) = 0.677
\label{SRNDdataMXUP}$$ $$\lambda_d(M_X) \sim 3.208\times 10^{-4} \qquad
\lambda_s(M_X) \sim 9.612\times 10^{-3} \qquad
\lambda_b(M_X) = 0.677
\label{SRNDdataMXDOWN}$$ $$\lambda_e(M_X) = 1.490\times 10^{-4} \qquad
\lambda_\mu(M_X) = 3.154\times 10^{-2} \qquad
\lambda_\tau(M_X) = 0.677
\label{SRNDdataMXLEP}$$ and the CKM matrix at $Q=M_X$ to be : $$|V_{12}(M_X)| = 0.2215 \qquad |V_{23}(M_X)|= 0.032 \qquad |V_{13}(M_X)| = 0.0028
\label{SRNDdataMXCKM}$$ At this point it is convenient to re-write Eqs. - in terms of the Wolfenstein [@LWolfenstein] expansion parameter $\lambda=0.22\sim|V_{12}|$. We find : $$\lambda_u(M_X) = \lambda^{8.097} \qquad
\lambda_c(M_X) = \lambda^{4.282} \qquad
\lambda_t(M_X) = \lambda^{0.257}
\label{SRNDUPdata}$$ $$\lambda_d(M_X) \sim \lambda^{5.313} \qquad
\lambda_s(M_X) \sim \lambda^{3.068} \qquad
\lambda_b(M_X) = \lambda^{0.257}$$ $$\lambda_e(M_X) = \lambda^{5.820} \qquad
\lambda_\mu(M_X) = \lambda^{2.283} \qquad
\lambda_\tau(M_X) = \lambda^{0.257}$$ $$|V_{12}(M_X)| = \lambda^{0.996} \qquad
|V_{23}(M_X)| = \lambda^{2.273} \qquad
|V_{13}(M_X)| = \lambda^{3.882}
\label{SRNDCKMdata}$$ Equations - neatly summarize the hierarchy of the quark and charged lepton sectors at $M_X$ that we aim to reproduce/predict.
It is now time to specify the structure of the LMA model in more detail. We start by indicating the nature of the (non-)renormalizable operators responsible for the structure of the Dirac and neutrino Majorana matrices : $$\begin{aligned}
\lambda_{AB} &:&
F_3 F^c_3 h +
F_A F^c_B h {H \bar H \over M^2_V}
\left[
1+
\left(H \bar H \over M^2_V \right)+
\left(H \bar H \over M^2_V \right)^2+\ldots
\right]
\left(\theta \over M_V\right)^{p_{AB}} \label{SRNDMRL} \\
(M_{RR})_{AB} &:&
\left\{
F^c_3 F^c_3+
F^c_A F^c_B
\left[
{H \bar H \over M^2_V}+\ldots
\right]
\right\}
{H H \over M^2_V}
\left(\theta \over M_V\right)^{q_{AB}} \label{SRNDMRR}\end{aligned}$$
The first term in Eq. is renormalizable, thus it implies third family Yukawa unification at $M_X$. The second term, which we shall assume to be present for $AB\ne 33$, on the other hand, is a sum of non-renormalizable operators. For the sake of simplicity we will consider that the $H\bar H/M_V^2$ part of $\lambda_{AB}$ that lies outside the square brackets in Eq. has non-trivial gauge contractions with the $F_A F^c_B h$ fields next to it, thereby generating the Clebsch factors in Table (Appendix A). On the other hand, the $(H\bar H/M_V^2)^{1,2}$ factors inside the square brackets will form gauge singlet terms that will be responsible for the appearance of higher $\delta$ powers in the entries of $\lambda_{AB}$. The $M_{RR}$ matrix, as given by Eq. , depends only on non-renormalizable operators because gauge invariance demands that every combination of $F^c F^c$ fields must be paired with at least a couple of $HH$ fields. However, we will assume that the only $n=1$ operator in $M_{RR}$ is placed on the 33 entry. All other entries of $M_{RR}$ result from $n=2$ operators. [^15]
We can see from Eqs. , that the structure of the Yukawa and Majorana matrices can be decomposed into a “vertical” $\delta$-component and a “horizontal” $\epsilon$-component. Thus we write : $$(\lambda_f)_{AB} \sim (\lambda^\delta)_{AB} (\lambda^\epsilon)_{AB}
\qquad\qquad
(M_{RR})_{AB} \sim (M_{RR}^\delta)_{AB} (M_{RR}^\epsilon)_{AB}
\label{YukMrrLMA}$$ The hierarchies of $\lambda^\epsilon$ and $M_{RR}^\epsilon$ are fixed by the choice of the $U(1)_X$ charges. Using the results of Ref. [@SFKing1], we can write the most general form of the unified $(\lambda^\epsilon)_{AB}$ matrix in the 422 model, constrained by the absence of anomalies, in terms of only four independent parameters $\bar X_{F_1}$, $\bar X_{F_2}$, $\bar X_{F^c_1}$ and $\bar X_{F^c_2}$ : [^16] $$\lambda^\epsilon =
\left(\matrix{
\epsilon^{|\bar X_{F_1}+\bar X_{F^c_1}|} &
\epsilon^{|\bar X_{F_1}+\bar X_{F^c_2}|} &
\epsilon^{|\bar X_{F_1}|} \cr
\epsilon^{|\bar X_{F_2}+\bar X_{F^c_1}|} &
\epsilon^{|\bar X_{F_2}+\bar X_{F^c_2}|} &
\epsilon^{|\bar X_{F_2}|} \cr
\epsilon^{|\bar X_{F^c_1}|} &
\epsilon^{|\bar X_{F^c_2}|} &
1 \cr}\right)
\label{lambdaepsilonLMA}$$ From the equation above it is easy to see that the values of $\bar X_{F_2}$, $\bar X_{F^c_2}$, $\bar X_{F_1}$ and $\bar X_{F^c_1}$ are closely related with the large neutrino $\theta_{23}$ angle, the second family Yukawa couplings, the $V_{12}$ CKM angle, and the masses of the lightest fermions respectively. In the first row of Table we list our choices for the $\bar X$’s parameters which we will, from now on, refer to as $U(1)_{\bar X}$ charges. In the second row we indicate the values of the physical (anomaly free) $U(1)_X=U(1)_{FD}+U(1)_{FI}$ charges of the fields of our model. In the third and forth row we list the values of the family dependent (traceless) and family independent (unphysical) charges that sum up to give $U(1)_X$.
We note that the $U(1)_{\bar X}$ and $U(1)_X$ charges are “equivalent” in the sense that they determine equal family structures for the Yukawa and neutrino Majorana matrices. [^17]
The charges in Table fix the $\epsilon$-structure of $\lambda^\epsilon$ and $M_{RR}^\epsilon$ to be : $$\lambda^\epsilon\sim\left(\matrix{
\epsilon^5 & \epsilon^3 & \epsilon \cr
\epsilon^4 & \epsilon^2 & 1 \cr
\epsilon^4 & \epsilon^2 & 1 \cr }\right) \qquad\qquad
M_{RR}^\epsilon\sim\left(\matrix{
\epsilon^8 & \epsilon^6 & \epsilon^4 \cr
\epsilon^6 & \epsilon^4 & \epsilon^2 \cr
\epsilon^4 & \epsilon^2 & 1\cr}\right)
\label{SRNDmRLmRReps}$$ Comparing $\lambda^\epsilon$ above with the hierarchy of the Yukawa couplings listed in Eqs. - we see that, the $U(1)_X$ symmetry (by itself) can not explain the pattern of all fermion masses and mixing angles. For example, although the symmetry allows a large 23 entry suitable for generating large 23 mixing from the neutrino Yukawa matrix, it also allows similarly large 23 entries in the charged lepton and quark Yukawa matrices which are not welcome. In order to overcome this we shall assume that although a renormalizable operator in the 23 position is allowed by the $U(1)_X$ symmetry, it is forbidden by some unspecified string symmetry which however allows a 23 operator containing one factor of $(H\bar{H})$. We shall further select a 23 operator which will involve a Clebsch factor of zero for the charged lepton and quark entries, with only its neutrino component having a non-zero contribution, thereby generating a large 23 mixing from the neutrino sector, with only small 23 mixing in the charged lepton and quark sectors arising from operators containing higher powers of $(H\bar{H})^n$, with $n>1$. The existence of such operators with “Clebsch zeros” is clearly crucial for the success of our approach.
In general, we shall show that by a suitable choice of non-renormalizable operators, which determine the $\lambda^\delta$ “vertical” structure of $\lambda_f$, we can obtain a successful description of all quark and lepton masses and mixing angles. For example, let us consider $\lambda^\delta$ and $M^\delta_{RR}$ given by the following operator matrices : $$\lambda^\delta \sim \left(\matrix{
{\cal O}^R+{\cal O}^{\prime\prime V} &
{\cal O}^J+{\cal O}^{\prime Q} &
{\cal O}^g+{\cal O}^{\prime f} \cr
{\cal O}^G+{\cal O}^{\prime\prime K} &
{\cal O}^W+{\cal O}^{\prime H} &
{\cal O}^I+{\cal O}^{\prime W} \cr
{\cal O}^R+{\cal O}^{\prime\prime V} &
{\cal O}^M+{\cal O}^{\prime K} &
1 }\right)
\qquad
M_{RR}^\delta \sim
\left(\matrix{
{\cal O} & {\cal O} & {\cal O} \cr
{\cal O} & {\cal O} & {\cal O} \cr
{\cal O} & {\cal O} & 1
}\right)
\label{SRNDmRLopMtr}$$ where ${\cal O}$, ${\cal O}^{\prime}$ and ${\cal O}^{\prime\prime}$ are $n=1$, $n=2$ and (very small) $n=3$ operators respectively [^18] where $n$ is defined in Eq. and refers to the powers of $(H\bar{H})^n$. Using Eqs. , into Eq. gives : $$\begin{aligned}
\lambda_f(M_X)
&=&
\left(\matrix{
\hfil x^R_f a_{11} \delta\epsilon^5 &
\hfil x^J_f a_{12} \delta\epsilon^3 &
\hfil x^g_f a_{13} \delta\epsilon^2 \cr
\hfil x^G_f a_{21} \delta\epsilon^4 &
\hfil x^W_f a_{22} \delta\epsilon^2 &
\hfil x^I_f a_{23} \delta\phantom{\epsilon^2} \cr
\hfil x^R_f a_{31} \delta\epsilon^4 &
\hfil x^M_f a_{32} \delta\epsilon^2 &
\hfil a_{33}\phantom{\epsilon^2} \cr
}\right)+ \nonumber \\
& & \cr
& &
\left(\matrix{
0 &
x^Q_f a^\prime_{12} \delta^2\epsilon^3 &
x^f_f a^\prime_{13} \delta^2\epsilon^2 \cr
0 &
x^H_f a^\prime_{22} \delta^2\epsilon^2 &
\hfil x^W_f a^\prime_{23} \delta^2\phantom{\epsilon^2} \cr
0 &
x^K_f a^\prime_{32} \delta^2\epsilon^2 &
0
}\right)+ \nonumber \\
& & \cr
& &
\left(\matrix{
x^V_f a^{\prime\prime}_{11} \delta^3\epsilon^5 &
0 &
0 \cr
x^K_f a^{\prime\prime}_{21} \delta^3\epsilon^4 &
0 &
0 \cr
x^V_f a^{\prime\prime}_{23} \delta^3\epsilon^4 &
0 &
0
}\right) \label{lfMXLMAan} \\
& & \cr
{\displaystyle {M_{RR}(M_X)\phantom{_{33}} \over M_{RR}(M_X)_{33}}}
&=& \left(\matrix{
A_{11} \delta\epsilon^8 & A_{12} \delta\epsilon^6 & A_{13} \delta\epsilon^4 \cr
A_{21} \delta\epsilon^6 & A_{22} \delta\epsilon^4 & A_{23} \delta\epsilon^2 \cr
A_{31} \delta\epsilon^4 & A_{32} \delta\epsilon^2 & A_{33}
\phantom{\delta\epsilon^2} \cr
}\right) \label{MrrMXLMAan}\end{aligned}$$ where the subscript $f$ stands for any of the $u,d,e,\nu$ indices, $x^{\cal O}_f$ is the Clebsch of the ${\cal O}$ operator of the $f$-type fermion, and the $a$’s ($A$’s) are order-one $f$-independent Yukawa (Majorana) parameters that parameterise $\lambda_f$ ($M_{RR}$). The first matrix on the right-hand side of Eq.\[lfMXLMAan\] contains the leading $n=1$ operators giving contributions of order $\delta$, while the second and third matrices contain the $n=2$ and $n=3$ operators which give contributions of order $\delta^2$ and $\delta^3$ and provide the leading contributions in the cases where the $n=1$ operators involve Clebsch zeros.
The effective matrices resulting from Eqs. , are approximately given in Table . This table shows an interesting structure for the Yukawa matrices. We find that $\lambda_u\sim\lambda^8$, $\lambda_c\sim\lambda^4$ and $\lambda_d\sim\lambda_e\sim\lambda^6$, $\lambda_s\sim\lambda_\mu\sim\lambda^3$. Furthermore the CKM matrix has $|V_{12}|\sim\lambda$, $|V_{23}|\sim\lambda^2$, $|V_{13}|\sim\lambda^3$. Comparing these approximate results with the data in Eqs. - we see that only the $\lambda_d$, $\lambda_\mu$ couplings need substantial (one $\lambda$-power) corrections. On the other hand, the neutrino sector described by $\lambda_\nu$ and $M_{RR}$ in Table is clearly dominated by the right-handed tau neutrino and predicts $(\lambda_{\nu})_{12}\sim(\lambda_\nu)_{22}$ which according to Eq. successfully generates a large $\theta_{12}$ solar neutrino angle. However, the subdominant perturbation to $m_{LL}$ in Eq. resulting from $\lambda_\nu$ and $M_{RR}$ in Table is too small to correctly predict the neutrino mass ratio $m_{\nu_2}/m_{\nu_3}\sim\lambda^{1.5}$ required by Eq. . These approximate predictions can be further improved because Table does not include the numerical effects of the operator Clebsches and of the order-one $a$,$A$ factors.
The success of our model (in the SM sector) depends on the ability to find suitable solutions for the $a$’s in Eq. which simultaneously can account for all the hierarchies in Eqs. -. Generally we will require that $0.5 < |a_{AB}|, |a^\prime_{AB}|, |a^{\prime\prime}_{AB}| < 2.0$ for all $A,B=1,2,3$. At first, it looks that such solution is trivial since Eq. depends on 16 parameters, [^19] while Eqs. - is a set of 9 constraints (on the first and second family Yukawa couplings and CKM entries.) However, we should not forget that ${\cal O} \gg {\cal O}^\prime \gg {\cal O}^{\prime\prime}$ and that the CKM matrix constrains only on the 12,13 and 23 entries of $\lambda_f$. As a consequence, we find that the parameters in Eqs. - are mainly sensitive to $a_{22}\leftrightarrow\lambda_{s,\mu}$, $a_{11}\leftrightarrow\lambda_{d,e}$, an independent combination of $(a^{\prime\prime}_{11},a^{\prime\prime}_{21},
a^{\prime\prime}_{31})\leftrightarrow\lambda_u$, $a'_{22}\leftrightarrow\lambda_c$ and $a'_{12}\leftrightarrow V_{12}$, $a'_{23}\leftrightarrow V_{23}$, $a'_{13}\leftrightarrow V_{13}$, which allows two predictions to be made — $\lambda_d$ and $\lambda_s$. Thus we fitted the $a_{22}$,$a_{11}$,$a^{\prime\prime}_{11}$, $a'_{22}$,$a'_{12}$,$a'_{23}$,$a'_{13}$ dependence of $\lambda_f$ [^20] to the $\lambda_\mu$, $\lambda_e$, $\lambda_u$, $\lambda_c$ and $V_{12}$, $V_{23}$, $V_{13}$ experimental constraints in Eqs. -. The results are shown in Table . [^21]
Thus, using Eq. with the $a$’s of Table and the Clebsch factors of Table (Appendix A) we get the numerical results for $\lambda_f(M_X)$ shown in Table . In Table we present the results of Table expanded in powers of $\delta=\epsilon=\lambda=0.22$.
We can analyse the effect of the operator Clebsches by comparing Table against Table . We see that the ${\cal O}^W$ operator [^22] in the 22 entry of $\lambda_f$ split $(\lambda_d)_{22}=(\lambda_e)_{22}\sim\lambda^3$ in Table into $|(\lambda_d)_{22}|=\lambda^{3.124}$ and $(\lambda_e)_{22}=\lambda^{2.313}$ in Table , thus allowing for a proper GJ ratio $\lambda_\mu / \lambda_s\sim 3$ at $M_X$. Similarly, the operators in the 12 block have allowed for a more appropriate $\lambda_d / \lambda_e$ Yukawa ratio. Numerically we have the following predictions for the lightest eigenvalues of the down-Yukawa matrix at $M_X$ : $\lambda_d=\lambda^{5.469}$ and $\lambda_s=\lambda^{3.087}$.
The effect of the Clebsch factors also modified the neutrino Yukawa matrix $\lambda_\nu(M_X)$ in Table . Due to the ${\cal O}^I$ operator in the 23 position of $\lambda_f$, which has a large $x^I_\nu=2$ Clebsch, we are now able to predict a large $\theta_{23}$ atmospheric neutrino mixing angle. Indeed using Eq. we can roughly estimate that $\tan\theta_{23}\sim 0.44/0.68=0.65$ implying $\sin^2(2\theta_{23}) \sim 0.83$.
It is interesting to check that $\lambda_\nu$ and $M_{RR}$ in Table do lead to a $m_{LL}$ matrix dominated by the right-handed tau neutrino. As a result of the small mixing angles of $M_{RR}$ it is convenient to work in a basis where $M_{RR}$ is diagonal. Furthermore it is practical to scale $\lambda_\nu$ [^23] such that the 23 and 33 entries are $(\lambda_\nu)_{23}\sim(\lambda_\nu)_{33}\sim 1$, and approximate the normalized entries to (semi-)integer $\lambda$-powers. Thus using : $$M_{RR}\sim\left(\matrix{
\lambda^{9.0} & 0 & 0 \cr
0 & \lambda^{5.0} & 0 \cr
0 & 0 & 1 }\right)
\qquad
m_{RL}\sim\left(\matrix{
\lambda^{7.5} & \lambda^{3.5} & \lambda^{1.5} \cr
\lambda^{6.5} & \lambda^{3.5} & 1 \cr
\lambda^{6.5} & \lambda^{3.5} & 1 }\right)$$ into Eq. gives : $$m_{LL} \sim \left(\matrix{
\lambda^{3.0}+\lambda^{2.0}+\lambda^{6.0} &
\lambda^{1.5}+\lambda^{2.0}+\lambda^{5.0} &
\lambda^{1.5}+\lambda^{2.0}+\lambda^{5.0} \cr
\lambda^{1.5}+\lambda^{2.0}+\lambda^{5.0} &
\hfill 1+\lambda^{2.0}+\lambda^{4.0} &
\hfill 1+\lambda^{2.0}+\lambda^{4.0} \cr
\lambda^{1.5}+\lambda^{2.0}+\lambda^{5.0} &
\hfill 1+\lambda^{2.0}+\lambda^{4.0} &
\hfill 1+\lambda^{2.0}+\lambda^{4.0} }\right)
\label{mLLGUTapr}$$ where the first (second and third) term in each entry corresponds to the third (second and first) family neutrino contribution $\nu^c_\tau$ ($\nu^c_\mu$ and $\nu^c_e$) coming from $M_{RR}$. Clearly Eq. shows that even though, in this case, $\nu^c_\tau$ is the heaviest right-handed neutrino, it nevertheless dominates the 23 block, and that the sub-dominant contribution from $\nu^c_\mu$ induces $\lambda^2$ perturbations in $m_{LL}$ that are compatible with the $m_{\nu_2} / m_{\nu_3}$ mass ratio.
Using the MSSM RGEs adapted and properly extended to take into account the presence and successive decoupling of the right-handed neutrinos between the $Q=M_X$ and $Q=M_Z$ scales (see Appendix B) we find that the Yukawa and neutrino Majorana matrices at low energy are given by Table .
Thus, inserting the results of Table into Eq. and Eq. we get the mass matrix for the left-handed neutrinos $m_{LL}$ and the $V^{MNS}$ mixing matrix shown in Table . The predictions for the neutrino masses and squared mass splittings are shown in Table . In Table we examine how the neutrino mixing angles evolve between the unification $Q=M_X\sim 3\times 10^{16} {\rm\ GeV}$, the right-handed tau neutrino mass $Q=M_{\nu_3}\sim 3\times 10^{14} {\rm\ GeV}$ and the $M_Z$ scale. We see that the effect of the radiative corrections has increased the magnitude of $\sin\theta_{12}$, $\sin\theta_{23}$ and $\sin\theta_{13}$ by 2.5%, 6.4% and 2.4% respectively. These corrections agree with the results found in Ref. [@KiSi]. Finally, we present in Table the predictions for the down and strange quark masses.
We would like to conclude this section by noting that the predictions for the neutrino parameters, in particular for the neutrino $\Delta m_{12}^2$ squared mass splitting, should be taken carefully. Generally we expect at least 20% (theoretical) errors in the quoted values which, for example, arise from our inability to fix order-one factors in the entries of $M_{RR}(M_X)$.
We have discussed a theory of all fermion masses and mixing angles based on a particular string-inspired [*minimal*]{} model based on the Pati-Salam group $SU(4)\times SU(2)_L \times SU(2)_R$ [@PaSa] supplemented by a gauged $U(1)_X$ family symmetry. We argued that this gauge group preserves the attractive features of $SO(10)$ such as predicting three right-handed neutrinos, and Yukawa unification, while avoiding the doublet-triplet splitting problem. Although it is not a unified gauge group at the field theory level, it naturally arises from string constructions and so in principle may be fully unified with gravity.
Earlier work in collaboration with one of us [@AlKiLeLo1] had already shown that the model can provide a successful description of the charged fermion masses and the CKM matrix. The use of the $U(1)_X$ family symmetry to provide the horizontal mass splittings combined with the Clebsch factors arising from the $(H\bar{H})^n$ insertion in the operators has already been shown to provide a powerful approach to the fermion mass spectrum in this model [@AlKiLeLo1]. The present analysis differs from that presented previously partly due to the recent refinements in third family Yukawa unification [@KiOl2], but mainly due to the recent data from Super-Kamiokande which implies that the 23 operator should be allowed by the $U(1)_X$ family symmetry. We have therefore extended our previous analysis to the atmospheric and solar neutrino masses and mixing angles, and showed that all three MSW solutions to the solar neutrino data may be accommodated, namely the LMA MSW region discussed in the main text as well as the LOW MSW and the SMA MSW regions discussed in the Appendices.
The approach to neutrino masses and mixing angles followed here makes use of the SRHND mechanism [@KingSRND; @SFKing1; @SFKing2] in which one of the right-handed neutrinos (the $\nu_{\tau}^c$) gives the dominant contribution to the 23 block of the light effective Majorana matrix. This mechanism avoids reliance on accidental cancellations, and does not rely on excessive magnification of mixing angles, although a mild enhancement was observed in the numerical results in agreement with that observed in [@KiSi]. Crucial to the implementation of SRHND in this model is the assumption that the renormalizable 23 operator is forbidden by unspecified string selection rules, and the leading 23 operator contains $(H\bar{H})$ and involves “Clebsch zeros”, which give a zero contribution to the charged lepton and quark Yukawa matrices, but a non-zero contribution to the neutrino Yukawa matrix, thereby allowing small $V_{cb}$ but large 23 mixing in the lepton sector.
The analysis in this paper is essentially “bottom-up”. A particular choice of $U(1)_X$ family symmetry charges was used to give the horizontal mass splittings, and the vertical mass splittings were achieved by particular choices of operators corresponding to different Clebsch factors in the leading contributions to each entry of the Yukawa matrix. It would be very nice to understand these choices from the point of view of a “top-down” string construction, such as the Type I string construction which has recently led to the Pati-Salam gauge group with three chiral families [@ShTy]. We believe that only by a combination of top-down and bottom-up approaches (such as that presented here) will a completely successful string theory of fermion masses and mixing angles emerge. We have shown that the recent discovery of neutrino mass by Super-Kamiokande provides precious information about the flavour structure of such a future string theory.
The work of M.O. was supported by JNICT under contract grant : PRAXIS XXI/BD/ 5536/95.
-20pt
In this appendix we briefly review some technical issues related to the presence of the right-handed neutrinos. Firstly we show how the decoupling of the neutrinos affects the one-loop RGEs for the Yukawa couplings in the MSSM+$\nu^c$ model : $$\begin{aligned}
{1 \over 16\pi^2}
{d\lambda_u\over dt} &=&
[ 3{\rm Tr}U+{\rm Tr}N+3U+D-G^u] \lambda_u \label{SRNDrgeU} \\
{1 \over 16\pi^2}
{d\lambda_d\over dt} &=&
[ 3{\rm Tr}D+{\rm Tr}E+3D+U-G^d] \lambda_d \\
{1 \over 16\pi^2}
{d\lambda_e\over dt} &=&
[ 3{\rm Tr}D+{\rm Tr}E+3E+N-G^e] \lambda_e \\
{1 \over 16\pi^2}
{d\lambda_\nu\over dt} &=&
[ 3{\rm Tr}U+{\rm Tr}N+3N+E-G^\nu] \lambda_\nu \label{SRNDrgeN}\end{aligned}$$ where $t=\ln(Q)$, $$\matrix{
U = \lambda_u\lambda_u^\dagger \hfill & \phantom{space}
G^u = {26 \over 30} g^2_1+3g^2_2+{16\over 3}g^2_3 \hfill \cr
D = \lambda_d\lambda_d^\dagger \hfill & \phantom{space}
G^d = {14 \over 30} g^2_1+3g^2_2+{16 \over 3}g^2_3 \hfill \cr
E = \lambda_e\lambda_e^\dagger \hfill & \phantom{space}
G^e = {18 \over 10} g^2_1+3g^2_2 \hfill \cr
N = \lambda_\nu\lambda_\nu^\dagger \hfill & \phantom{space}
G^\nu = {6 \over 10} g^2_1+3g^2_2 \hfill } \label{UDEN}$$ and ${\rm\ Tr}U= U_{11}+U_{22}+U_{33}$ [*etc..*]{} The general idea behind the process of decoupling the right-handed neutrinos (in the “step” approximation) is that a Feynman diagram that includes a specific flavour of a right-handed neutrino $\nu^c_A$, with mass $M_{\nu_A}$, in an internal line only makes a contribution to the RGEs in Eqs. (\[SRNDrgeU\])-(\[SRNDrgeN\]) for energies $Q$ bigger than $M_{\nu_A}$. Thus, the procedure depends on properly adapting the $N$ parameter in Eq. . We shall now make this statement more precise. Let us assume that the neutrino Majorana matrix $M_{RR}$ is diagonalized by the following transformation : $$S^{\nu^c\dagger} M_{RR} S^{\nu^c} = M'_{RR}
= {\rm\ diag}(M_{\nu_1},M_{\nu_2},M_{\nu_3})$$ Then, the decoupling of the right-handed neutrinos in Eqs. (\[SRNDrgeU\])-(\[SRNDrgeN\]) can be accounted by replacing $N$ in Eq. by $N_\theta$ given by : $$N=\lambda_\nu \lambda^\dagger_\nu =
\lambda_\nu S^{\nu^c} S^{\nu^c\dagger} \lambda^\dagger_\nu \to
\lambda_\nu S^{\nu^c} \Theta S^{\nu^c\dagger} \lambda^\dagger_\nu = N_\theta$$ where $\Theta(Q)$ is a energy dependent diagonal matrix defined by : $$\Theta(Q) = {\rm\ diag}(
\theta(Q-M_{\nu_1}),
\theta(Q-M_{\nu_2}),
\theta(Q-M_{\nu_3}))$$ with $\theta(x)=0$ for $x<0$ and $\theta(x)=1$ for $x>0$.
The second issue that we would like to address concerns the effect of a large $(\lambda_\nu)_{23}$ coupling on third family Yukawa unification, and as a consequence, for example, on the prediction for the top mass. We claim that the effect is small. To see why let us assume that the only large Yukawa couplings in Eqs. (\[SRNDrgeU\])-(\[SRNDrgeN\]) are $\lambda_t = (\lambda_u)_{33}$, $\lambda_b = (\lambda_d)_{33}$, $\lambda_\tau = (\lambda_e)_{33}$ and $\lambda_{\nu_\tau} = (\lambda_\nu)_{33}$, $\lambda_{23} = (\lambda_\nu)_{23}$. In this limit, the RGEs simplify to : $$\begin{aligned}
{1 \over 16\pi^2}
{d\lambda_t \over dt} &=&
\lambda_t ( 6 \lambda_t^2 + \lambda_b^2 + \lambda_{\nu_\tau}^2 +
\lambda_{23}^2-G^u) \label{SRNDrget} \\
{1 \over 16\pi^2}
{d\lambda_b \over dt} &=&
\lambda_b ( 6 \lambda_b^2 + \lambda_t^2 + \lambda_{\tau}^2 - G^d)
\label{SRNDrgeb} \\
{1 \over 16\pi^2}
{d\lambda_\tau \over dt} &=&
\lambda_\tau ( 4 \lambda_\tau^2 + 3 \lambda_b^2 + \lambda_{\nu_\tau}^2 - G^e )
\label{SRNDrgee}\\
{1 \over 16\pi^2}
{d\lambda_{\nu_\tau} \over dt} &=&
\lambda_{\nu_\tau} ( 4 \lambda_{\nu_\tau}^2 + 4 \lambda_{23}^2 +
3 \lambda_t^2 + \lambda_\tau^2-G^\nu) \label{SRNDrgen}\end{aligned}$$ From Eqs. , we see that the presence of the $\lambda_{23}$ coupling does not affect the RGEs of $\lambda_{b,\tau}$. Moreover the effect of $\lambda_{23}$ on the RGE of $\lambda_t$ is small ($1/8 \sim 12 \%$.) The only RGE that is significantly affected by $\lambda_{23}$ is the RGE of $\lambda_{\nu_\tau}$. However, since the correct prediction for the heaviest left-handed neutrino $m_{\nu_3}\sim 0.05 {\rm\ eV}$ requires that $M_{\nu_3} > 10^{13} {\rm\ GeV}$, the $\lambda_{\nu_\tau}^2$ and $\lambda_{23}^2$ terms in Eqs. , are only present in a rather short energy range, [*i.e.*]{} between $10^{13} {\rm\ GeV} < Q < M_X \sim 10^{16} {\rm\ GeV}$. As a consequence the presence/absence of the neutrino Yukawa couplings, as far as third family Yukawa unification is concerned, is not important. [^24]
Finally we find interesting to comment on the radiative corrections to the neutrino atmospheric mixing angle $\theta_{23}$ between the GUT and the $M_Z$ scale. It is well known [@BaLePa:MTanimoto] that the running of $\sin^2(2\theta_{23})$ can be understood from the following evolution equation : $${1 \over 16\pi^2}
{1\over \sin^2(2\theta_{23})}
{d \sin^2(2\theta_{23}) \over dt} =
-2(\lambda_\tau^2-\lambda^2_\mu)
{ (m_{LL})^2_{33} - (m_{LL})^2_{22} \over
[(m_{LL})_{33}-(m_{LL})_{22}]^2+4(m_{LL})^2_{23}}
\label{SRNDsin2t}$$ which displays a resonance peak at $(m_{LL})_{33} \sim (m_{LL})_{22}$ when $(m_{LL})_{23}$ is small. Generally, it is possible that $(m_{LL})_{33}$ starts at $Q=M_X$ bigger than $(m_{LL})_{22}$ but, due to the third family Yukawa radiative effects, to be driven to smaller values faster than $(m_{LL})_{22}$. As a result, even if the initial values of $(m_{LL})_{33}$ and $(m_{LL})_{22}$ at $M_X$ are different, they may at some point become comparable. If this is the case then a large $\theta_{23}$ angle can be generated radiatively from a small tree level $\theta_{23}$ at $M_X$. This mechanism, of amplifying $\theta_{23}$ radiatively, as been studied for example in Refs. [@ElLeLoNa; @BaLePa:MTanimoto]. However, in these works, and as can be seen from Eq. , the amplification is only efficient if at least the $\lambda_\tau$ Yukawa coupling is large (about 2 or 3.) In our model, since we demanded top-bottom-tau Yukawa unification, the value of the third family Yukawa coupling is rather small ($\sim 0.7$), thus the $\sin^2(2\theta_{23})$ is stable under radiative corrections.
In this appendix we show that is easy to convert the results of the LMA MSW solution found in the main body of this paper into results for the LOW solution which is also characterized by maximal $\nu_e\to\nu_\mu$ oscillations but smaller $\Delta m_{12}^2$ [@GoPe; @JEllis] : $$\hskip -2cm {\rm LOW : } \quad\quad
\sin^2(2\theta_{12}) \sim 1 \qquad
\Delta m^2_{12} \sim 10^{-7} {\rm\ eV}^2$$ The reason why we can adapt the LMA results is because, as we showed in section III, in the SRHND approach, the $\theta_{12}$ and $\theta_{23}$ neutrino mixing angles come “solely” from the neutrino Yukawa matrix. On the other hand, the neutrino mass spectrum depends on the hierarchies of $\lambda_\nu$ and $M_{RR}$. Thus, as long as we keep within the SRHND scenario, we can change $M_{RR}$ to fit the LOW $\Delta m^2_{12}$ solution without that implying a significant change in $\theta_{12}$ and $\theta_{23}$.
Let as consider a LOW model with the same $U(1)_X$ flavour charges and the same operator matrix for $\lambda_f$ as in the LMA model, given by Table and Eq. , but with a $M_{RR}$ matrix with the following structure : $${\displaystyle {M_{RR}(M_X)\phantom{_{33}} \over M_{RR}(M_X)_{33}}}
= \left(\matrix{
A_{11} \epsilon^8 & A_{12} \epsilon^6 & A_{13} \epsilon^4 \cr
A_{21} \epsilon^6 & B_{22} \epsilon^4 & A_{23} \epsilon^2 \cr
A_{31} \epsilon^4 & A_{32} \epsilon^2 & A_{33}
\phantom{\epsilon^2} \cr
}\right) \label{MrrMXLOWan}$$ Comparing Eq. with Eq. we see that these two equations differ only by their “vertical” $\delta$-component (we assumed that Eq. has $M_{RR}^\delta\sim{\bf 1}$) and by the numerical factor $B_{22}=1.821 \ne 1.072=A_{22}$. We note that the removal the $\delta$-factor in the 22 entry of $M_{RR}$ and the increase of the $B_{22} > A_{22}$ coefficient act to decrease $\Delta m^2_{12}$.
The Majorana matrix $M_{RR}(M_Z)$ and the neutrino Yukawa matrix $\lambda_\nu(M_Z)$ in the LOW model resulting from $M_{RR}(M_X)$ in Eq. and the Yukawa matrices $\lambda_f(M_X)$ given by Table (recall that we take $\lambda_f^{LOW}(M_X) = \lambda_f^{LMA}(M_X)$) are shown in Table . In Table we present the predicted values for the left-handed neutrino matrix and the MNS matrix in the LOW model. The results for the neutrino masses in the LOW model are given in Table . Finally in Table we show the values of the neutrino mixing angles.
In this appendix we briefly present a model that explores the possibility of a SMA MSW solution to the solar neutrino anomaly. Although the SMA region is disfavoured by the latest results from the Super-Kamiokande experiment, the SMA solution is not completely ruled out. [^25] The SMA solution data indicates [@GoPe]: $${\rm SMA : } \qquad
\sin^2(2\theta_{12}) \sim 1.6\times 10^{-3} \qquad
\Delta m^2_{12} \sim 5\times 10^{-6} {\rm\ eV}^2$$
In analogy with the LMA model we start by recalling that the Yukawa and the neutrino Majorana matrices in the SMA model can be decomposed into a “vertical” $\delta$-component and a “horizontal” $\epsilon$-component given by : $$(\lambda_f)_{AB} \sim (\lambda^\delta)_{AB} (\lambda^\epsilon)_{AB}
\qquad\qquad
(M_{RR})_{AB} \sim (M_{RR}^\delta)_{AB} (M_{RR}^\epsilon)_{AB}
\label{YukMrrSMA}$$
The $U(1)_{\bar X}$ charges of the SMA given in Table fix the “horizontal” structure of $\lambda^\epsilon$ and $M_{RR}^\epsilon$ in the SMA model to be : $$\lambda^\epsilon
\sim
\left(\matrix{
\epsilon^5 & \epsilon^4 & \epsilon^2 \cr
\epsilon^3 & \epsilon^2 & 1 \cr
\epsilon^3 & \epsilon^2 & 1
}\right)\qquad
M_{RR}^\epsilon
\sim
\left(\matrix{
\epsilon^6 & \epsilon^5 & \epsilon^3 \cr
\epsilon^5 & \epsilon^4 & \epsilon^2 \cr
\epsilon^3 & \epsilon^2 & 1 \cr
}\right)
\label{SRNDmRLmRRepsSMA}$$ On the other hand, the “vertical” structure of $\lambda^\delta$ and $M_{RR}^\delta$ is given by the following operator matrices : $$\lambda^\delta \sim \left(\matrix{
{\cal O}^G+{\cal O}^{\prime\prime K} &
{\cal O}^R+{\cal O}^{\prime O} &
{\cal O}^H+{\cal O}^{\prime a} \cr
{\cal O}^G+{\cal O}^{\prime\prime V} &
{\cal O}^W+{\cal O}^{\prime H} &
{\cal O}^I+{\cal O}^{\prime W} \cr
{\cal O}^M+{\cal O}^{\prime\prime V} &
{\cal O}^g+{\cal O}^{\prime T} &
1 }\right)\qquad
M_{RR}^\delta \sim {\bf 1}
\label{SRNDmRLopMtrSMA}$$ As a result of Eqs. , the Yukawa and the neutrino Majorana matrices in Eq. can be written as : $$\begin{aligned}
\lambda_f(M_X)
&=&
\left(\matrix{
\hfil x^G_f c_{11} \delta\epsilon^5 &
\hfil x^R_f c_{12} \delta\epsilon^4 &
\hfil x^H_f c_{13} \delta\epsilon^2 \cr
\hfil x^G_f c_{21} \delta\epsilon^3 &
\hfil x^W_f c_{22} \delta\epsilon^2 &
\hfil x^I_f c_{23} \delta\phantom{\epsilon^2} \cr
\hfil x^M_f c_{31} \delta\epsilon^3 &
\hfil x^g_f c_{32} \delta\epsilon^2 &
\hfil c_{33}\phantom{\epsilon^2} \cr
}\right)+ \nonumber \\
& & \cr
& &
\left(\matrix{
0 &
x^O_f c^\prime_{12} \delta^2\epsilon^4 &
x^a_f c^\prime_{13} \delta^2\epsilon^2 \cr
0 &
x^H_f c^\prime_{22} \delta^2\epsilon^2 &
\hfil x^W_f c^\prime_{23} \delta^2\phantom{\epsilon^2} \cr
0 &
x^T_f c^\prime_{32} \delta^2\epsilon^2 &
0
}\right)+ \nonumber \\
& & \cr
& &
\left(\matrix{
x^K_f c^{\prime\prime}_{11} \delta^3\epsilon^5 & 0 & 0 \cr
x^V_f c^{\prime\prime}_{21} \delta^3\epsilon^3 & 0 & 0 \cr
x^V_f c^{\prime\prime}_{31} \delta^3\epsilon^3 & 0 & 0
}\right) \label{lfMXSMAan} \\
& & \cr
{\displaystyle {M_{RR}(M_X)\phantom{_{33}} \over M_{RR}(M_X)_{33}}}
&=& \left(\matrix{
C_{11} \epsilon^6 & C_{12} \epsilon^5 & C_{13} \epsilon^3 \cr
C_{21} \epsilon^5 & C_{22} \epsilon^4 & C_{23} \epsilon^2 \cr
C_{31} \epsilon^3 & C_{32} \epsilon^2 & C_{33} \phantom{\epsilon^2} \cr
}\right) \label{MrrMXSMAan}\end{aligned}$$
In the rest of this appendix we apply the same systematic approach used in the main part of the paper for the LMA solution to the SMA model. The approximate structure of the effective matrices resulting from Eqs. , is given in Table . In Table we give the values of the $c$,$C$ parameters appearing in Eqs. ,. In Table we present the exact numerical values of the Yukawa and Majorana matrices at the unification scale and in Table the values of the same matrices at the $M_Z$ scale. In Table we present the predicted values for the left-handed neutrino mass and for the MNS mixing matrices. The predictions for masses of the physical neutrinos in the SMA model is listed in Table and in Table we give the predictions for the neutrino mixing angles at several energy scales. Finally, in Table we show the predictions for the masses of the down and strange quarks in the SMA model.
[99]{}
Y. Fukuda , Super-Kamiokande Collaboration, Phys. Lett. [**B433**]{}, 9 (1998); Phys. Lett. [**B436**]{}, 33 (1998); Phys. Rev. Lett. [**81**]{}, 1562 (1998).
H. Sobel, talk presented at the XIX International Conference on Neutrino Physics and Astrophysics, Sudbury, Canada, June 16-21, 2000.
L. Wolfenstein, Phys. Rev. [**D17**]{}, 2369 (1978); Phys. Rev. [**D20**]{}, 2634 (1979); S. P. Mikheyev, A. Y. Smirnov, Yad. Fiz. [**42**]{}, 1441 (1985) \[Sov. J. Nucl. Phys. [**42**]{}, 913 (1985)\]; Nuovo Cimento [**9C**]{}, 17 (1986).
Y. Suzuki, talk presented at the XIX International Conference on Neutrino Physics and Astrophysics, Sudbury, Canada, June 16-21, 2000.
J. N. Bahcall, P. I. Krastev, A. Y. Smirnov, Phys. Rev. [**D60**]{} 093001 (1999) [hep-ph/9905220]{}.
M. Gell-Mann, P. Ramond, R. Slansky, in Sanibel Talk, CALT-68-709, Feb. 1979 and in [*Supergravity*]{} (North Holland, Amsterdam 1979); T. Yanagida in [*Proc. of the Workshop on Unified Theory and Baryon Number of the Universe*]{}, KEK, Japan (1979); R. N. Mohapatra, G. Senjanovic, Phys. Rev. Lett. [**44**]{}, 912 (1980).
S. F. King, N. N. Singh, [hep-ph/0007243]{}.
J. Ellis, G. K. Leontaris, S. Lola, D. V. Nanopoulos, Eur. Phys. J. [**C9**]{}, 389 (1999);
K. S. Babu, C. N. Leung, J. Pantaleone, Phys. Lett. [**B319**]{}, 191 (1993); M. Tanimoto, Phys. Lett. [**B360**]{}, 41 (1995).
S. F. King, Phys. Lett. [**B439**]{}, 350 (1998).
S. F. King, Nucl. Phys. [**B562**]{}, 57 (1999).
S. F. King, Nucl. Phys. [**B576**]{}, 85 (2000).
G. Altarelli, F. Feruglio, Phys. Lett. [**B451**]{}, 388 (1999).
G. Altarelli, F. Feruglio, I. Masina, Phys. Lett. [**B472**]{}, 382 (2000).
K. S. Babu, J. C. Pati, F. Wilczek, Nucl. Phys. [**B566**]{}, 33 (2000).
C. H. Albright, S. M. Barr, Phys. Rev. Lett. [**85**]{}, 244 (2000).
Q. Shafi, Z. Tavartkiladze, Phys. Lett. [**B487**]{}, 145 (2000).
J. C. Pati, A. Salam, Phys. Rev. [**D10**]{}, 275 (1974).
I. Antoniadis and G. K. Leontaris, Phys. Lett. [**B216**]{}, 333 (1989); I. Antoniadis and G. K. Leontaris, and J. Rizos, Phys. Lett. [**B245**]{}, 161 (1990).
G. Shiu, S. H. H. Tye, Phys. Rev. [**D58**]{}, 106007 (1998).
S. F. King, Phys. Lett. [**B325**]{}, 129 (1994).
B. C. Allanach, S.F. King, Nucl. Phys. [**B456**]{}, 57 (1995).
B. C. Allanach, S. F. King, Nucl. Phys. [**B459**]{}, 75 (1996).
B. C. Allanach, S. F. King, G. K. Leontaris, S. Lola, Phys. Rev. [**D56**]{}, 2632 (1997).
S. F. King, M. Oliveira, [hep-ph/0008183]{}.
H. P. Nilles, Phys. Rep. [**110**]{}, 1 (1984) ; H. E. Haber and G. L. Kane, Phys. Rep. [**117**]{}, 75 (1985) ; A. B. Lahanas and D. V. Nanopoulos, Phys. Rep [**145**]{}, 1 (1987); H. E. Haber, Presented at Theoretical Advanced Study Institute (TASI 92): From Black Holes and Strings to Particles, Boulder, CO, 3-28 Jun 1992. Published in Boulder TASI 92:0589-688, [hep-ph/9305248]{}.
N. Cabibbo, Phys. Rev. Lett. [**10**]{}, 531 (1963); M. Kobayashi, T. Maskawa, Prog. Theor. Phys. [**49**]{}, 652 (1973).
Z. Maki, M. Nakagawa, S. Sakata, Prog. Theor. Phys. [**28**]{}, 870 (1962).
S. F. King and Q. Shafi, Phys. Lett. [**B422**]{}, 135 (1998).
P. Ramond, R. G. Roberts, G. G. Ross, Nucl. Phys. [**B406**]{}, 19 (1993); Y. Grossman, Y. Nir, Nucl. Phys. [**B448**]{}, 30 (1995); M. Leurer, Y. Nir, N. Seiberg, Nucl. Phys. [**B398**]{}, 319 (1993); , Nucl. Phys. [**B420**]{}, 468 (1994).
L. Ibáñez, G. G. Ross, Phys. Lett. [**B332**]{}, 100 (1994).
V. Jain, R. Shrock, Phys. Lett. [**B352**]{}, 83 (1995).
P. Binetruy, P. Ramond, Phys. Lett. [**B350**]{}, 49 (1995); E. Dudas, S. Pokorski, C.A. Savoy, Phys. Lett. [**B356**]{}, 45 (1995); P. Binétruy, S. Lavignac, P. Ramond, Nucl. Phys. [**B477**]{}, 353 (1996); E. J. Chun, A. Lukas, Phys. Lett. [**B387**]{}, 99 (1996). H. Dreiner, G. K. Leontaris, S. Lola, G. G. Ross, C. Scheich, Nucl. Phys. [**B436**]{}, 461 (1995); G. K. Leontaris, S. Lola ,C. Scheich, J. D. Vergados, Phys. Rev. [**D53**]{}, 6381 (1996); S. Lola, G. G. Ross, Nucl. Phys. [**B553**]{}, 81 (1999); M. Carena, J. Ellis, S. Lola, C. E. M. Wagner, Eur. Phys. J. [**C12**]{}, 507 (2000).
R. Barbieri, L. J. Hall, D. Smith, A. Strumia, N. Weiner, JHEP 9812, 17 (1998); Y. Grossman, Y. Nir, Y. Shadmi, JHEP [**9810**]{}, 007 (1998); G. Altarelli, F. Feruglio, Phys. Lett. [**B439**]{}, 112 (1998); S. M. Barr, I. Dorsner, [hep-ph/0003058]{}.
L. Wolfenstein, Phys. Rev. Lett. [**51**]{}, 1945 (1983).
B. C. Allanach, S. F. King, Nucl. Phys. [**B507**]{}, 91 (1997).
M. Green, J. Schwarz, Phys. Lett. [**149**]{}, 117 (1984).
L. E. Ibáñez, Phys. Lett. [**B303**]{}, 55 (1993).
H. Georgi, C. Jarlskog, Phys. Lett. [**B86**]{} 297 (1979); S. Dimopoulos, L. J. Hall, S. Raby, Phys. Rev. [**D45**]{}, 4192 (1992).
C. D. Froggatt, H. B. Nielsen, Nucl. Phys. [**B147**]{}, 277 (1979).
G. Anderson, S. Raby, S. Dimopoulos, L.J. Hall, G.D. Starkman, Phys. Rev. [**D49**]{}, 3660 (1994).
D. E. Groom, (Particle Data Group), Eur. Phys. Jour. [**C15**]{}, 1 (2000).
H. Arason, D. J. Castaño, B. Kesthelyi, S. Mikaelian, E. J. Piard, P. Ramond and B. D. Wright, Phys. Rev. [**D46**]{}, 3945 (1992).
M. Apollonio (CHOOZ collaboration), Phys. Lett. [**B466**]{}, 415, (1999); Phys. Lett. [**B420**]{}, 397 (1998).
S .F. King and N. N. Singh, [hep-ph/0006229]{}, Nucl. Phys. B (to appear).
M. C. Gonzalez-Garcia and C. Peña-Garay, [hep-ph/0009041]{}.
J. Ellis, Talk given at the 19TH International Conference on Neutrino Physics and Astrophysics - Neutrino 2000, Sudbury, Ontario, Canada, 16-21 Jun 2000. [hep-ph/0008334]{}
[^1]: However this is not guaranteed due to the unknown structure of the heavy Majorana matrix, and for example an inverted neutrino mass hierarchy could result although this relies on some non-hierarchical couplings in the Dirac Yukawa matrix [@KiSi2].
[^2]: We use Left-Right (LR) convention for Yukawa matrices in this paper.
[^3]: The four component fermion fields $\Psi$ are given by $\Psi_F = (F , -i\sigma^2 F^{c*})$ for $F=U,D,E$ and $\Psi_{N} = (N, -i\sigma^2 N^{*})$ for the neutrinos.
[^4]: We shall not address the question of CP violation in this paper.
[^5]: $V_{e2} = V^{MNS}_{12}$, $V_{e3} = V^{MNS}_{13}$ and $V_{\mu 3} = V^{MNS}_{23}$.
[^6]: In this section we will use the following simplified notation $m_{LR} = \lambda_\nu v_2 \equiv \lambda v_2 \sim \lambda$.
[^7]: Note that $R_{12}$ for $m_{LL}$ in Eq. is undetermined.
[^8]: Note that although Eq. still looks very simple it can, in many cases, provide a good qualitative description of the physics involving the heaviest neutrinos. Indeed, if $M_{RR}$ is diagonal dominated and if $m_{RL}$ is highly hierarchical then the limiting case of Eq. applies.
[^9]: Note that this does not necessarily imply that $M^{-1}_{\nu_3}$ is larger than $M^{-1}_{\nu_2}$ since the Yukawa couplings must also be taken into account.
[^10]: The cancellation of anomalies requires the vanishing of the trace ${\rm Tr}(T^a \{T^b,T^c\}) = 0$ where $T^{a,b,c}$ are any of the group generators which stand at the three gauge boson vertices of the triangle diagrams.
[^11]: We will not include the analysis of the $U(1)_X^3$ or of the gravitational anomaly because they depend exclusively on SM singlet fields.
[^12]: The numbers inside the curly brackets indicate the experimental ranges according to Ref. [@pdg].
[^13]: All running masses are given in the $\overline{\rm MS}$ scheme.
[^14]: The experimental ranges are $|V_{12}| = 0.219 {\rm\ to\ } 0.226$, $|V_{23}|= 0.037 {\rm\ to\ } 0.043$ and $|V_{13}| = 0.002 {\rm\ to\ } 0.005$ [@pdg].
[^15]: We note that these assumptions about the nature of the Majorana matrix are unique to the LMA MSW solution. The SMA MSW and LOW solutions discussed in Appendices C and D are characterized by a Majorana matrix filled with $n=1$ operators only.
[^16]: In [@SFKing1] these charges were called $\alpha$, $\beta$, $\gamma$, $\delta$. Roughly, this corresponds to choosing a basis of charges, that has $\bar X_{F_3}=\bar X_{F^c_3}=\bar X_{h}=0$.
[^17]: However, only the $U(1)_X$ symmetry is anomaly free.
[^18]: The $n=3$ operators can, to a very good approximation, be neglected. Their inclusion here serves only to fill the 11, 21, 31 entries of the $\lambda_{u,\nu}$ Yukawa matrices, thereby ensuring (for example) that the up quark is given a very small mass.
[^19]: We note that $a_{33}$ is fixed by quadruple Yukawa unification at $M_X$, [*i.e.*]{} $\lambda_t=\lambda_b=\lambda_\tau=\lambda_{\nu_\tau}$.
[^20]: Fixing all other $a$’s to be one.
[^21]: The values of the $A$ parameters in Eq. are not constrained by the experimental data, thus we chose them to be “arbitrary” numbers in the $0.5 < A_{AB} < 2.0$ range.
[^22]: That has Clebsches $x^W_e = -3 x^W_d$
[^23]: $(\lambda_\nu)_{AB} \to (\lambda_\nu)_{AB} / k$ with $k = {1\over 2} [ (\lambda_\nu)_{23}+(\lambda_\nu)_{33}] = \lambda^{0.384}$.
[^24]: Numerically we found that when $(\lambda_\nu)_{23}$ is allowed to take values comparable with $(\lambda_\nu)_{33}$ the prediction for the top mass roughly decreased by 1 GeV, the value of $\tan\beta$ decrease by 0.5 and the value of the unified third family Yukawa coupling at the unification scale decreased by 0.015.
[^25]: Statistically, the SMA solution can still describe the neutrino data with a probability of 34 %. [@GoPe]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Algorithms for embedding certain types of nilpotent subalgebras in maximal subalgebras of the same type are developed, using methods of real algebraic groups. These algorithms are applied to determine non-conjugate subalgebras of the symmetry algebra of the wave equation, which in turn are used to determine a large class of invariant solutions of the wave equation. The algorithms are also illustrated for the symmetry algebra of a classical system of differential equations considered by Cartan in the context of contact geometry.'
address:
- 'Department of Basic Sciences, School of Electrical Engineering and Computer Science, National University of Sciences and Technology, Islamabad 44000, Pakistan'
- 'Department of Mathematics and Statistics, King Fahd University, Saudi Arabia'
- 'School of Mathematics, Tata Institute of Fundamental Research, Homi Bhabha Road, Mumbai 400005, India'
- 'Virginia Commonwealth University in Qatar, Education City Doha, Qatar'
- 'Department of Mathematics, Statistics and Physics, Qatar University, Doha, 2713, State of Qatar'
author:
- Sajid Ali
- Hassan Azad
- Indranil Biswas
- Ryad Ghanam
- 'M. T. Mustafa'
title: Embedding algorithms and applications to differential equations
---
Introduction
============
One of the main applications of Lie algebras is to find solutions of differential equations, by reduction of order, or by using conjugacy classes of its subalgebras to find invariant solutions. The method of invariant solutions goes back to [@Lie Ch. X]. This method is also explained in detail in the books of Ibragimov [@ib Ch 9], Ibragimov [@ibsw], Bluman [@bluman1; @bluman2] and Olver [@Ol Ch 3].
The Lie theoretic input in this method is a list of conjugacy classes of subalgebras of dimension depending on the order of the equation. A detailed structure of the symmetry algebra is also useful in finding linearizing coordinates for linearizable equations.
It is our experience, based on [@ADGM], that if the algebras are not chosen appropriately, they are practically useless, because a preliminary step is to find their invariants and there is no algorithmic procedure to do that. However, if the subalgebras are constructed from the geometry of the space on which one is studying a given equation, for example by embedding translations or scalings in maximal subalgebras, the characteristics of the subalgebras obtained are manageable.
The principal aim of this note is to give algorithms for embedding given abelian and solvable algebras of certain types in maximal subalgebras of the same type, using standard commands of Maple.
The precise types of the subalgebras are given in the algorithms constructed below.
Maple is able to find the Cartan decomposition as well as root space decompositions for semisimple algebras of fairly high dimensions. The algorithms it uses are based on the fundamental papers of Rand, Winternitz and Zassenhaus [@RWZ], of de Graff [@Dg], and of Dietrich, Faccin and de Graaf [@DFG] and Ian Anderson [@An]. The recent book of Šnobl-Winternitz [@SW] gives a detailed account of some of these algorithms.
Derksen, Jeandel and Koiran [@DJK] have also developed algorithms for computing the Zariski closures of linear solvable Lie groups and the algorithms in this paper reduce the computation of Zariski closures of linear groups to those of abelian subgroups. The algorithms given in this paper are based on results of Mostow [@Mo] on real algebraic groups; see a recent account of the subject in [@AB]. All the algebras considered in this paper are assumed to be real algebraic Lie algebras. We recall their definition and some basic facts about them.
1. Let ${\mathfrak g}\, \subset\, {\rm gl}(n, {\mathbb R})$ be a Lie algebra, with $G$ the corresponding Lie group. The algebra ${\mathfrak g}$ is called an algebraic Lie algebra if the group $G^{\mathbb C}\,\subset\, {\rm GL}(n, {\mathbb C})$ with Lie algebra ${\mathfrak g}^{\mathbb C}\,=\, {\mathfrak g} \oplus \sqrt{-1}{\mathfrak g}$ is an algebraic group in as defined in [@Bo].
2. The group $G^{\mathbb C}$ is generated by complex $1$-parameter subgroups $$\{\exp (zX) \,\mid\, z\,\in\, {\mathbb C}\},\ \ X \,\in\, {\mathfrak g}$$ and the connected component of the real points of $G^{\mathbb C}$ is the group $G$.
All real semisimple Lie algebras, all real linear Lie algebras generated by nilpotents, all abelian real linear algebras of semisimple elements defined by integral equations as well as all subalgebras generated by the types of algebras already listed are examples of algebraic Lie algebras.
If an abstract Lie algebra is given by its commutator table, then one is tacitly working in its adjoint representation and all the concepts regarding semisimplicity, nilpotency etc. are with reference to this representation.
As regards the reduced root space decompositions, Maple needs a maximal split abelian algebra of semisimple elements. If that is not specified, the commands will give, in general, roots taking complex eigen-values.
The algorithms constructed in this paper give both the relative and absolute root system of the Lie algebra more or less automatically.
We will illustrate the algorithms by working out in detail an embedding of a subalgebra, which is clearly abelian, of the symmetry algebra of the wave equation on flat 4 dimensional space, in a maximal solvable subalgebra: this gives at the same time detailed structure of the symmetry algebra and several non-conjugate subalgebras.
These algebras are non-conjugate by construction. A solvable linear algebraic algebra is itself a sum of a torus and its maximal nilpotent ideal. A linear algebraic torus has either all real eigen-values or all purely imaginary eigen-values or it is itself a sum of two such tori [@Bo], [@AB Proposition 1]. Thus if given subalgebras of a semisimple algebra are abstractly isomorphic and solvable one looks at their semisimple and nilpotent parts in the adjoint representation of any semisimple subalgebra containing them to decide if they are non-conjugate. The uniqueness of the Jordan decomposition [@Bo] ensures that it is immaterial which ambient subalgebra one chooses as long as it is semisimple.
If the subalgebras are semisimple and abstractly isomorphic, one looks at their centralizers or normalizers to decide if they are non-conjugate.
It may happen that their centralizers or normalizers are the same. In that case one has to use Bruhat decomposition or its variants [@Kn] or methods similar to [@GW], [@PWZ] to decide non-conjugacy under the adjoint group. This type of complete symmetry analysis has not been done in this paper and the examples were chosen to illustrate that while doing the algorithms interactively and computing normalizers or centralizers one gets several subalgebras for which one could check their non-conjugacy without using Bruhat or Iwasawa decomposition.
We do the same for the symmetry algebra of a classical nonlinear ODE, whose symmetry algebra was determined and identified as the exceptional algebra $G_2$ by Cartan — see [@Agricola] for further references, and also [@AKO] and [@Ke]. The group $G_2$ also has a very interesting relation to mechanics. A recent account is in Bor and Montgomery [@BM]. One of the main points in [@BM] is to identify a maximal compact subalgebra of $G_2$ and give its explicit decomposition. The identification and decomposition of the maximal compact subalgebra of $G_2$ follows from algorithm 1 in a straight forward way.
The structural information obtained in Section 4 gives several three and four dimensional subalgebras that are non-conjugate in the adjoint representation. This gives a more extensive and useable list of subalgebras than that given in [@ADGM] and these subalgebras are used in the last section to give solutions of the wave equation in flat 4-d space.
In a follow up of this paper, a similar analysis of the wave equation on all static spherically symmetric spaces times will be given to obtain a more extensive list of solutions than that given in [@ADGM] of the wave equation on certain spherically symmetric four dimensional spaces. In particular, such invariant solutions will be given for all types of $3$ dimensional subalgebras that can arise as subalgebras of the symmetry algebra of the equation.
The reader is referred to [@ibsw] for very general results on the wave equation on Riemannian manifolds and several other equations of relevance to physics.
As far as Lie algebras are concerned, we need the following results:
1. If $X$ is an element of a Lie algebra $L$ which is the Lie algebra of a real algebraic subgroup of ${\rm GL}(n,{\mathbb R})$, and $X\,=\, X_s+X_n$ is the Jordan decomposition of $X$, then both $X_s$ and $X_n$ are in $L$ [@Bo p. 14, Section 3.7, Proposition 1, Lemma 1], [@AB]. In fact, they are in the center of the centralizer of $X$ in $L$.
2. If $S$ is an abelian subalgebra consisting of semisimple elements of a real algebraic Lie algebra, then its centralizer $Z(S)$ has the Levi decomposition $$Z(S) \,=\, [Z(S),\, Z(S)] \oplus Z(Z(S))\, ,$$ where $Z(Z(S))$ is the center of $Z(S)$ [@BT] (a proof of this is also given below in Proposition \[propL\]).
3. The derived algebra of any solvable algebra is ad-nilpotent. In particular, if $H$ is any subalgebra of $L$ then the derived algebras of the radical of centralizer of $H$ and of the normalizer of $H$ are ad-nilpotent [@HN p. 105, Theorem 5.4.7], [@AB Proposition 1].
4. If $H$ is a semisimple subalgebra of $L$ and $X$ is an ad-semisimple or ad-nilpotent element of $H$ in the adjoint representation of $H$ on itself, then $X$ is also ad-semisimple or ad-nilpotent in the adjoint representation of $L$ [@Bo p. 14, 3.7].
These facts are very useful in verifying that a certain element is semisimple or nilpotent by reducing the computations to subalgebras of small dimensions.
Roots {#sR}
=====
Roots of a semisimple algebra
-----------------------------
A few words regarding the section on roots are in order. Maple will give – for any Cartan algebra – an array of complex numbers. In the following section we explain how to extract a simple system of roots and the corresponding Dynkin diagram directly from such a list.
Let $L$ be a semisimple Lie subalgebra of $\mathfrak{g}\mathfrak{l}(n,
{\mathbb R})$ and $C$ a Cartan subalgebra of $L$. The algebra $C$ is, by definition, a maximal abelian subalgebra of diagonalizable elements in the complexification of $L$. A nonzero vector $v$ in ${L}\oplus\sqrt{-1}L$ such that $$[h\, ,v] \,=\, \lambda(h)\cdot v$$ for all $h\, \in\, C$ is called a *root vector* and the corresponding linear functional $\lambda$ is called a root of the Cartan algebra $C$.
In general, the roots will be complex valued, so one needs to define what it means for a complex valued root to be positive – based only on the list of roots provided by the program. This is sufficient to describe the Dynkin diagram algorithmically, as detailed below.
A complex number $z\,=\, a+\sqrt{-1}b$, where $a\, , b\, \in\, \mathbb R$, is positive if either its real part $a$ is positive or $a\,=\, 0$ but $b\, >\,0$.
Fix a basis $h_1\, , \cdots\, , h_r$ of $C$. A non-zero root $\lambda$ is positive if the first nonzero number $\lambda(h_i)$ is a complex positive number. Otherwise, it is called a negative root.
Positive roots which are not a sum of two positive roots are called simple roots.
For sake of convenience, henceforth a root will mean a non-zero root.
Restricted roots {#secrr}
----------------
An abelian subalgebra of $L$ consisting of semisimple elements in the adjoint representation on $L$, is, by definition, a [*torus*]{}. If, moreover, all its elements in the adjoint representation of $L$ have real eigen-values, then it is a [*real*]{} torus; if all eigen-values are purely imaginary, it is called a [*compact*]{} torus.
Any real algebraic torus is a sum of a real and a compact torus and the dimensions its real and compact parts are invariants of the torus [@AB Proposition 1].
Moreover, all maximal solvable subalgebras $B$ of a real semisimple algebra with real eigen-values in the adjoint representation are conjugate [@Mo], [@AB]. In the context of the Iwasawa decomposition [@Kn], [@HN], this is the algebra $A\oplus N$.
If $A$ is a maximal torus of B then the full algebra is a sum of $A$-invariant subspaces — of dimension possibly greater than one — and the roots $A$ in $B$ which are not a sum of two roots in $B$ are simple roots of a root system — in the sense of [@HN], [@Kn]. In case that the real semisimple algebra has a maximal torus with all real eigen-values we can define positive roots without going to the complexification of the algebra. In this case the positive root spaces together with the torus give a maximal solvable algebra whose eigen-values are all real and all such algebras are conjugate. Therefore, the restricted root system and the absolute root system coincide in this case.
For each positive simple root $\alpha$ we can find a standard set of generators $X_\alpha,\,
Y_\alpha,\, H_\alpha$ with $X_\alpha,\, Y_\alpha$ eigen-vectors of ${\rm ad}(H_\alpha)$ with opposite and nonzero eigen-values. This three dimensional subalgebra is therefore isomorphic to ${\rm sl}(2, {\mathbb R})$. If, for each simple root $\alpha$ we fix an isomorphism $$\varphi_\alpha\, :\, {\rm sl}(2, {\mathbb R}) \,\longrightarrow\,
\langle X_\alpha,\, Y_\alpha,\, H_\alpha\rangle\, ,$$ where $\langle X_\alpha,\, Y_\alpha,\, H_\alpha\rangle$ is the Lie subalgebra generated by $X_\alpha,\, Y_\alpha,\, H_\alpha$, then the elements $$\varphi_\alpha\begin{pmatrix}0 &1\\ -1 & 0\end{pmatrix}$$ generate, as a Lie subalgebra, a maximal compact subalgebra of the given Lie algebra — only in the case that the Lie algebra has a maximal torus whose eigen-values are all real [@St p. 100, Lemma 43].
Procedure for constructing Dynkin diagram
-----------------------------------------
Positive roots which are not a sum of two positive roots are called simple roots. Thus to obtain simple roots from a given set of positive roots, one adds pairs of positive roots and marks those that are sums of positive roots; at the end, one strikes out those roots that are sums of two positive roots and the remaining ones will be simple roots.
Let $a$, $b$ be simple roots. The positive roots among the integral combinations of them determine the bond between $a$ and $b$. The simple roots $a$, $b$ are not joined if $a+b$ is not a root. They are joined by a single bond if $a$, $b$ and $a+b$ are the only positive roots among the integral combinations of $a$ and $b$. They are joined by a double bond with arrow pointing from $a$ to $b$ if $a$, $b$, $a+b$ and $a+2b$ are the only positive roots among the integral combinations of $a$ and $b$. They are joined by a triple bond with arrow pointing from $a$ to $b$ if $a$, $b$, $a+b$, $a+2b$, $a+3b$ and $2a+3b$ are the only positive roots among the integral combinations of $a$ and $b$.
The diagram then identifies the complexification of the Lie algebra $L$.
Algorithms {#aL}
==========
We need the following result on centralizers of semisimple elements to implement the algorithms.
\[propL\] Let $S$ be a commuting algebra of diagonalizable elements in a real semisimple algebra $L$. Then the centralizer of $S$ has the Levi decomposition $$Z_L(S)\,=\, [Z_L(S),\, Z_L(S)]\oplus Z(Z_L(S))\, ,$$ where $Z_L(S)$ is the centralizer of $S$ in $L$, and $Z(Z_L(S))\, \subset\, Z_L(S)$ is its center.
Recall that we defined a nonzero complex number to be positive if its real part is positive or if its real part is zero and the imaginary part is positive.
Include $S$ in a maximal torus $T$. The complexification of $T$ is a maximal torus of the complexification $L^{\mathbb C}$ of $L$ [@AB Corollary 7], and the centralizer $Z_L(S)$ of $S$ in $L$ is the same as the real points of the centralizer of $S^{\mathbb C}$ in $L^{\mathbb C}$. Now $Z_{L^\mathbb{C}}(S^{\mathbb C})$ is generated by $T^{\mathbb C}$ and the root vectors $X_\alpha$ such that $\alpha(S)\,=\, 0$. This is a closed set of roots and the Lie algebra $Z_{L^\mathbb{C}}(S^{\mathbb C})$ contains the root vector $X_{-\alpha}$ for every $\alpha$ as above. By extending a basis of $S^{\mathbb C}$ say $\{s_i\}_{i=1}^m$ to a basis of $T^{\mathbb C}$, say $\{s_i\}_{i=1}^n$ and declaring a root $r$ to be positive if the first nonzero number $r(s_i)$ is positive, we see that the indecomposable positive roots of $Z_{L^\mathbb{C}}(S^{\mathbb C})$ are simple roots of $T^{\mathbb C}$. Thus these generate a semisimple subalgebra $L_1$ of $Z_{L^\mathbb{C}}(S^{\mathbb C})$ and $Z_{L^\mathbb{C}}(S^{\mathbb C}) \,=\, \langle L_1,\, T^{\mathbb C}\rangle$. Since $T^{\mathbb C}$ normalizes $L_1$ the commutator is $L_1$. Hence $Z_{L^\mathbb{C}}(S^{\mathbb C})/L_1$, being an image of $T^{\mathbb C}$, contains no nilpotents.
Therefore, the Levi decomposition of $Z_{L^\mathbb{C}}(S^{\mathbb C})$ is $$Z_{L^\mathbb{C}}(S^{\mathbb C})\,=\, L_1+ R\, ,$$ and $R$, being solvable with no nilpotents, is a torus. As $[L_1,\, R]$ is contained in both $L_1$ and $R$ it must be $0$. Hence $R$ is a central torus and it is equal to $Z(Z_{L^\mathbb{C}}(S^{\mathbb C}))$ — the center of $Z_{L^\mathbb{C}}(S^{\mathbb C})$. Therefore, the Levi decomposition of $Z_{L^\mathbb{C}}(S^{\mathbb C})$ is $[Z_{L^\mathbb{C}}(S^{\mathbb C}),\, Z_{L^\mathbb{C}}(S^{\mathbb C})]
\oplus Z(Z_{L^\mathbb{C}}(S^{\mathbb C}))$. Taking real points gives the Levi decomposition $Z_L(S)\,=\, [Z_L(S),\, Z_L(S)]\oplus Z(Z_L(S))$.
The algorithms given below are similar to each other. For the convenience of the user we have written down complete details-at the expense of repetition-of the most frequent types of algebras encountered in practice.
Algorithm for embedding a given abelian subalgebra of semisimple elements with real eigen-values
------------------------------------------------------------------------------------------------
Here we give an algorithm for embedding a given abelian subalgebra of semisimple elements with real eigen-values in a maximal algebra of such elements and in a maximally real Cartan algebra: in this algorithm, the ambient algebra is assumed to be semisimple.
For a subalgebra $H$ of $L$, let $N_L(H)$, $Z_L(H)$, $Z(H)$ and $H'$ denote its normalizer in $L$, its centralizer in $L$, its center and its derived algebra respectively. For notational convenience, we will also write $N(H)$ for $N_L(H)$.
Let $A$ be real torus (defined in Section \[secrr\]).
**Step 1:** Compute $Z_L(A)$, the centralizer of $A$ in $L$, the derived algebra $Z_L(A)'$ of $Z_L(A)$ and the center $Z(Z_L(A))$ of $Z_L(A)$. Then one has the direct sum decomposition $$Z_L(A)\,=\, Z_L(A)'\oplus Z(Z_L(A))\, .$$
**Step 2:** Compute the Killing form of $Z_L(A)'$. If it is negative definite, then the real part of the subalgebra $Z(Z_L(A))$ is a maximal real torus.
**Step 3:** If the Killing form of $Z_L(A)'$ is indefinite, compute the Cartan decomposition of $Z_L(A)'$, and pick any nonzero element from the radial part of the decomposition and adjoin it to $A$.
Repeat Step 1 and Step 2 till an abelian algebra, which we again denote by $A$, is obtained which has all real eigen-values in the adjoint representation – and in the decomposition $Z_L(A)\,=\, Z_L(A)'\oplus Z(Z_L(A))$, the Killing form of $Z_L(A)'$ is negative definite.
At this stage, a maximal real torus containing the given algebra has been obtained. Denote it again by $A$.
The compact part of $Z(Z_L(A))$ together with a maximal torus of $Z_L(A)'$ is a compact torus. Adjoining it to $A$ gives a maximally real Cartan algebra.
By an entirely similar procedure, a compact torus can be embedded in a maximally compact Cartan subalgebra.
Algorithm for embedding a commutative subalgebra of ad-nilpotent elements to a maximal commutative subalgebra of such elements
------------------------------------------------------------------------------------------------------------------------------
Let $U$ be an abelian algebra of ad–nilpotent elements. Let $$Z_L(U)\,=\, S\oplus R$$ be the Levi decomposition of $Z_L(U)$. Compute the derived subalgebra $R'\, \subset\, R$. If $\dim (R'+U)\, >\, \dim U$, adjoin any element of $R'$ complementary to $U$ to obtain a commutative subalgebra of ad–nilpotent elements. Repeat this procedure until an abelian algebra of ad–nilpotent elements is obtained – which we denote again by $U$ – so that in the Levi decomposition on $Z_L(U)$, the algebra $R'$ is contained in $U$.
At this stage if $R$ contains $U$ as a proper subalgebra, consider an element $x$ of $R$ complementary to $U$. Then the nilpotent and semisimple parts of $x$ belong to $R$. If $x$ has a nonzero nilpotent part $x_n$, then the subalgebra generated by $U$ and $x_n$ is commutative consisting of ad–nilpotent elements.
Repeating the above procedure, we may assume that $U$ is a commutative subalgebra of ad–nilpotent elements such that in the Levi decomposition $$Z_L(U) \,=\, S\oplus R\, ,$$ $R'\, \subset\, U$, and every element in a basis of $R$ complementary to $U$ consists of semisimple elements.
If the Killing form of $S$ is not negative definite, then $S$ will have a nontrivial Cartan decomposition. Take an element $\alpha$ in the radial part of the Cartan decomposition of $S$. As $S$ has no center, the endomorphism ${\rm ad}(\alpha)$ of $S$ has a nonzero real eigen-value. In fact any element all of whose eigen-values are real will do. If $u$ is a nonzero eigen-vector of ${\rm ad}(\alpha)$ for such a nonzero real eigen-value, then ${\rm ad}(u)$ is nilpotent on $S$. The reason is that $S$ is a direct sum of eigen-spaces for ${\rm ad}(\alpha)$ and if $S_a$,$S_b$ are two such eigen-spaces, then $[S_a,\, S_b]$ is contained in $S_{a+b}$; therefore a sufficiently high power of ${\rm ad}(u)$ annihilates $S$. This implies that ${\rm ad}(u)$ is also nilpotent on $L$, by uniqueness of Jordan decomposition. Adjoin $u$ to $U$ to obtain a higher dimensional commutative algebra of nilpotents.
Thus repeating the above procedure we ultimately have a commutative subalgebra consisting of ad–nilpotent elements, which we again denote by $U$, such that $$Z_L(U)\,=\, S+R\, ,$$ $R'\, \subset\, U$, every element in a basis of $R$ complementary to $U$ consists of semisimple elements and $S$ has a negative definite Killing form.
At this stage $U$ is a maximal abelian subalgebra of ad–nilpotent elements.
A very similar argument (-detailed below-) gives the embedding of a given ad–nilpotent subalgebra in a maximal ad–nilpotent subalgebra. Here one is assured of conjugacy of these subalgebras [@AB]. Also, its normalizer will pick up a torus whose eigen-values are all real and which is maximal with these properties. Embedding this in a maximal torus – using Algorithm 3.1 – one will obtain a maximally split Cartan subalgebra of $L$.
Algorithm for embedding a subalgebra $U$ of ad–nilpotent elements to a maximal subalgebra of such elements
----------------------------------------------------------------------------------------------------------
**Step 1:** Find a normalizer of $U$ and compute its Levi decomposition $$N(U)\,=\, S\oplus R\, ,$$ where $R$ is the radical.
**Step 2:** Compute the derived algebra $R'$ of $R$. If $\dim (R'+U)\, >\, \dim
U$, then this is again an ad–nilpotent algebra.
Repeat Step 1 and Step 2 so that ultimately $R'+U\,=\, U$. At this stage $R/U$ is abelian.
**Step 3:** We want to enlarge $U$ further so that $R/U$ consists entirely of semisimple elements. To do this, take a basis of $U$, say $u_1\, , \cdots\, ,u_k$, and enlarge it to a basis of $R$ by adjoining $v_1\, , \cdots\, ,v_\ell$. Find the Jordan decomposition of all $v_1\, , \cdots\, ,v_\ell$. Adjoin to $U$ the nilpotent parts of all the $v_1\, , \cdots\, ,v_\ell$. Denote by $\widetilde U$ the algebra obtained this way.
Now repeat Step 1, Step 2 and Step 3 till we have the Levi decomposition $$N(\widetilde U) \,=\, S+R$$ such that $R'\, \subset\, \widetilde U$ and $R/\widetilde U$ consists only of semisimple elements.
**Step 4:** If the Killing form of $S$ is not negative definite, then it will have a nontrivial Cartan decomposition. Take an element $\alpha$ in the radial part of the Cartan decomposition of $S$. As the center of $S$ is trivial, the endomorphism ${\rm
ad}(\alpha)$ of $S$ has a nonzero real eigen-value. Let $u$ be a nonzero eigen-vector for such an eigen-value. Then ${\rm ad}(u)$ is nilpotent on $S$ and therefore on the Lie algebra $L$ – by uniqueness of the Jordan decomposition. Adjoin $u$ to $U$ to obtain a higher dimensional algebra of nilpotents.
Iterating this procedure, we finally reach the situation that we have an ad–nilpotent subalgebra, which we denote again by $U$, that contains the original ad–nilpotent subalgebra, such that the Levi decomposition of $N(U)$ is $$N(U)\,=\, S\oplus R\, ,$$ where $S$ has a negative definite Killing form, $R'\, \subset\, U$ and $R/U$ consists only of semisimple elements in the sense that if we extend a basis of $U$ to a basis of $R$ and find the Jordan decomposition of the basis elements outside $U$, then the nilpotent parts all belong to $U$ [@AB].
At this stage, $U$ is a maximal ad–nilpotent algebra containing the given ad–nilpotent algebra.
Finally, the abelian algebra representing $R/U$ is a torus and its real part is a maximal abelian algebra consisting of real semisimple elements [@AB]. Denote this algebra by $A$. Then $A$ can be enlarged to a maximally split Cartan algebra of the whole algebra, using Algorithm 3.1. This algebra permutes the common eigen-spaces of $A$.
Applications to structure of symmetry algebras related to certain equations of physics {#Se4}
======================================================================================
Before giving applications to symmetry algebras of higher dimensions, we illustrate the algorithms to obtain structural information for the algebras ${\mathfrak
s}{\mathfrak o}(4)$, ${\mathfrak s}{\mathfrak o}(1,3)$ and ${\mathfrak s}{\mathfrak
o}(2,2)$.
Recall that if $A$ is an $n\times n$ diagonal matrix with diagonal entries $1$, $-1$, then the Lie algebra of the corresponding orthogonal group has generators $e_{ij}-e_{ji}$ if $a_{ii}a_{jj}
\,=\, 1$, and $e_{ij}+e_{ji}$ if $a_{ii}a_{jj} \,=\, -1$. Moreover, if $A$ has the first $p$ diagonal entries $1$, and the next $q$ diagonal entries $-1$, where $p+q\,=\, n$, then the Lie algebra of the corresponding orthogonal group is generated by $$e_{1,2}-e_{2,1}\, , \cdots\, , e_{p-1,p}-e_{p,p-1}\, , e_{p,p+1}+e_{p+1,p}\, ,
e_{p+1,p+2}-e_{p+2, p+1}\, , \cdots\, , e_{n-1,n}-e_{n,n-1}\, .$$
${\mathfrak s}{\mathfrak o}(4)$ {#sec4.1l}
-------------------------------
We now derive the factorization of ${\mathfrak s}{\mathfrak o}(4)$ using roots of its maximal torus.
A basis of $V\,=\,{\mathfrak s}{\mathfrak o}(4)$ is $$e_1\,=\, e_{12}-e_{21}\, ,e_2\,=\, e_{13}-e_{31}\, ,e_3\,=\, e_{14}-e_{41}\, ,
e_4\,=\, e_{23}-e_{32}\, ,e_5\,=\, e_{24}-e_{42}\, , e_6\,=\, e_{34}-e_{43}\, .$$ Using Algorithm 3.1, a Cartan subalgebra $C$ is generated by $\{e_1\,, e_6\}$. The roots of $C$ are $$a\,:=\, (\sqrt{-1}\, , \sqrt{-1})\, ,\ b\,:=\, (\sqrt{-1}\, ,
-\sqrt{-1})\, ,\ -a\, ,\ -b\, .$$ As $a+b$ is not a root, this root system is of type $A_1\times A_1$.
Also, conjugation maps a root to its negative. Thus the subalgebras generated by the eigen-spaces $V_r\, , V_{-r}$, $r\,=\, a\, , b$, contain a real form of $\text{sl}(2, {\mathbb C})$, which must be isomorphic to ${\mathfrak s}{\mathfrak o}(3)$.
In more detail, we need to compute only the eigen-vectors for the positive roots. Their real and imaginary parts will give the decomposition of the compact algebra $V$. Now, $$V_a\,=\, \langle e_2+\sqrt{-1}e_3+\sqrt{-1}e_4-e_5\rangle\, ,\
V_b\,=\, \langle e_2+\sqrt{-1}e_3-\sqrt{-1}e_4+e_5\rangle\, .$$ The real and imaginary parts of the basis elements in $V_a$ and $V_b$ and generate ${\mathfrak s}{\mathfrak o}(4)$.
Let $$u_1\,=\, e_2-e_5\, , \ v_1\,=\, e_3+e_4\, ,\ u_2\,=\, e_2+e_5\, ,\ v_2\,=\, e_3-e_4\, .$$ Then $u_1\, , v_1\, , [u_1\, , v_1]\,=\, -2(e_1+e_6)$ generate a copy of ${\mathfrak s}{\mathfrak o}(3)$. Note that $u_2\, , v_2\, , [u_2\, , v_2]\,=\,
2(e_1-e_6)$ also generate a copy of ${\mathfrak s}{\mathfrak o}(3)$. These two copies of ${\mathfrak s}{\mathfrak o}(3)$ commute because the root system is of type $A_1\times A_1$. This gives the well known fact that ${\mathfrak s}{\mathfrak
o}(4)\,=\,{\mathfrak s}{\mathfrak o}(3)\oplus {\mathfrak s}{\mathfrak o}(3)$.
${\mathfrak s}{\mathfrak o}(1,3)$
---------------------------------
A basis of $V\,=\, {\mathfrak s}{\mathfrak o}(1,3)$ is $$e_1\,=\, e_{12}+e_{21}\,
,e_2\,=\, e_{13}+e_{31}\, ,e_3\,=\, e_{14}+e_{41}\, , e_4\,=\, e_{23}-e_{32}\,
,e_5\,=\, e_{24}-e_{42}\, , e_6\,=\, e_{34}-e_{43}\, .$$ Using Algorithm 3.1 a Cartan subalgebra $C\,=\, \langle e_1\, , e_6\rangle$ is obtained. Note that there is no real split or compact Cartan subalgebra.
The roots of $C$ are $$a\,:=\, (1\, , -\sqrt{-1})\, ,\ b\,:=\, (1\, , \sqrt{-1})\, , \-a\, ,\
-b\, .$$ The root system is of type $A_1\times A_1$, with positive roots $a\, , b$ and conjugation maps $a$ to $b$.
The real rank is one, and the eigen-values of $\text{ad}(e_1)$ are $-1,\,-1,\,0,\,0,\,1,\,1$. The corresponding root spaces are $$V_1\,=\, \langle e_3+e_5\, , e_2+e_4\rangle\, ,
V_{-1}\,=\, \langle -e_3+e_5\, , -e_2+e_4\rangle\, .$$ Also, $[e_3+e_5\, , -e_3+e_5]\,=\,2e_1$; the subalgebra generated by $e_3+e_5\, , -e_3+e_5$ is $\text{sl}(2,{\mathbb R})$, while the subalgebra generated by $e_4\, , e_5\, , e_6$ is ${\mathfrak s}{\mathfrak o}(3)$.
Thus a maximal solvable subalgebra consisting of elements with real eigen-values in the adjoint representation is $$\langle e_1\, , e_3+e_5\, , e_2+e_4\rangle\, ,$$ and a maximal solvable subalgebra is $$\langle e_6\, ,\, e_1\, ,\, e_3+e_5\, ,\, e_2+e_4\rangle\, .$$
${\mathfrak s}{\mathfrak o}(2,2)$
---------------------------------
A basis of $V\,=\, {\mathfrak s}{\mathfrak o}(2,2)$ is $$e_1\,=\, e_{12}-e_{21}\,
,e_2\,=\, e_{13}+e_{31}\, ,e_3\,=\, e_{14}+e_{41}\, , e_4\,=\, e_{23}+e_{32}\,
,e_5\,=\, e_{24}+e_{42}\, , e_6\,=\, e_{34}-e_{43}\, .$$ Using Algorithm 3.1, a real split Cartan subalgebra is $C\,=\,
\langle e_2\, , e_5\rangle$, while a compact Cartan subalgebra is $\langle e_1\, , e_6\rangle$.
The roots of $C$ are $$a\,:=\, (1\, , 1)\, , \ b\,:=\, (1\, , -1)\, ,\ -a\, ,\ -b\, .$$ The root spaces are $$V_a\,=\, \langle e_1-e_3+e_4-e_6\rangle \, ,\ V_b\,=\, \langle e_1+e_3+e_4+e_6\rangle$$ $$V_{-a}\,=\, \langle e_1+e_3-e_4-e_6\rangle\, ,\
V_{-b}\,=\, \langle e_1-e_3-e_4+e_6\rangle\, .$$ Conjugation fixes the roots. Consequently, the subalgebra generated by a root spaces of a root and its negative is isomorphic to $\text{sl}(2,{\mathbb R})$.
Therefore, denoting the subalgebra generated by $V_r\, , V_{-r}$ by $\langle
V_r\, , V_{-r}\rangle$ the decomposition $\langle
V_a\, , V_{-a}\rangle\oplus \langle V_b\, , V_{-b}\rangle$ gives an isomorphism of $\text{sl}(2,{\mathbb R})\oplus \text{sl}(2,{\mathbb R})$ with ${\mathfrak s}{\mathfrak o}(2,2)$.
Lie symmetries of wave equations {#se4.4}
--------------------------------
The algebra of Lie point symmetries of the wave equation in a flat 4-d space is sixteen dimensional and determined by the vector fields (following the same order as given in [@ADGM] and using the notation in which $X$ is represented with $e$):
$$\begin{aligned}
&e_{1} = yt \partial_{t}+xy\partial_{x}+\frac{(y^2+t^2-x^2-z^2)}{2}\partial_{y}+yz\partial_{z}-uy\partial_{u}, \\
&e_{2} = y\partial_{t} + t\partial_{y}, \\
&e_{3} = xt \partial_{t}+\frac{(x^2+t^2-y^2-z^2)}{2}\partial_{x}+xy\partial_{y}+xz\partial_{z}-ux\partial_{u}, \\
&e_{4} = x\partial_{t} + t\partial_{x}, \\
&e_{5} = zt \partial_{t}+zx\partial_{x}+yz\partial_{y}+\frac{(z^2+t^2-y^2-x^2)}{2}\partial_{z}-uz\partial_{u},\\
&e_{6} = z\partial_{t} + t\partial_{z}, \\
&e_{7} = t\partial_{t} + x\partial_{x}+y\partial_{y} + z\partial_{z}, \\
&e_{8} = \partial_{t}, \\
&e_{9} = (t^2+x^2+y^2+z^2) \partial_{t} + 2tx \partial_{x}+ 2ty \partial_{y}+ 2tz \partial_{z}- 2ut\partial_{u}, \\
&e_{10} = \partial_{y}, \\
&e_{11} = \partial_{x}, \\
&e_{12} = \partial_{z}, \\
&e_{13} = z\partial_{y}-y\partial_{z}, \\
&e_{14} = z\partial_{x}-x\partial_{z}, \\
&e_{15} = y\partial_{x}-x\partial_{y}, \\
&e_{16} = u\partial_{u}.\end{aligned}$$
The commutator algebra of the finite dimensional part of the symmetry algebra of the wave equation on Minkowski space-time is 15 dimensional. By computing its radical or the Killing form one sees that this 15-dimensional algebra is semisimple.
Its basis is $e_1,\ldots,e_6,\,e_7-e_{16},e_{8},\ldots, e_{15}.$ For notational convenience we will denote-only in this section- $e_7-e_{16}$ by $e_7$.
The commutator table is reproduced in Appendix 1. The translations parallel to the coordinate axes are $$\partial_x\,=\, e_{11}\, ,\partial_y\,=\, e_{10}\, ,\partial_z\,=\, e_{12}\, ,
\partial_t\,=\, e_8$$ and they form an ad–nilpotent subalgebra.
Let $$U\, =\, \langle e_8\, , e_{10}\, , e_{11}\, ,e_{12}\rangle\, .$$ We will use standard Maple commands and Algorithm 3.3 to embed $U$ in a maximal ad-nilpotent subalgebra $\widetilde{U}$ and also compute the normalizer of this subalgebra. It may be pointed out that the implementation of Algorithm 3.3 will give rise to the identification of several lower dimensional non-conjugate subalgebras. Since the standard symmetry reduction from translations yield trivial solutions, therefore our algorithm will provide translations embedded into those subalgebras that will provide non-trivial solutions of the wave equation. Consequently this approach provides a direct use of translational subalgebras.
As explained in Algorithm 3.3, in general, if $$N(\widetilde{U})\, =\, S+R$$ is the Levi decomposition of $N(\widetilde{U})$ then $S$ has a negative definite Killing form, $R'$ is contained in $U$ and $R/U$ consists only of semisimple elements in the sense that if we extend a basis of $U$ to a basis of $R$ and find the Jordan decomposition of the basis elements outside $U$, then the nilpotent parts all belong to $U$.
As the abelian algebra representing $R/U$ is a torus and its real part $A$ is a maximal abelian algebra consisting of real semisimple elements, $A$ can be enlarged to a maximally split Cartan algebra of the whole algebra, using Algorithm 3.1. This algebra permutes the common eigen-spaces of $A$.
Following Algorithm 3.3, we first compute $N(U)$ and its Levi decomposition. We have $$N(U)\,=\, R\oplus S\, ,$$ where $R\,=\, \langle U\, , e_7\rangle$ is the radical, and $$S\,=\,
\langle e_{15}\, , e_{14}\, , e_{13}\, ,e_{6}\, , e_{5}\, , e_{4} \rangle$$ is semisimple with Cartan decomposition $$S\,=\, \langle e_{15}\, , e_{14}\, , e_{13}\rangle\oplus
\langle e_{6}\, , e_{5}\, , e_{4} \rangle$$ with compact part $K\, =\, \langle e_{15}\, , e_{14}\, ,
e_{13}\rangle$ and radial part $P\,=\, \langle e_{6}\, , e_{5}\, ,
e_{4} \rangle$.
The element $e_7$ representing $R/U$ is real semisimple and $e_6$ is maximal abelian in $P$. The compact subalgebra $K$ is the subalgebra of spatial rotations.
The eigen-values of $\text{ad}(e_6)$ in $S$, counting multiplicities, are $1,1,-1,-1,0,0$ and eigen-vectors for eigen-value 1 are $-e_{15}+e_4$, $-e_{13}+e_6$. Therefore, as the eigen-vectors for positive eigen-values of a real semisimple element of $S$ form an ad-nilpotent subalgebra, following Algorithm 3.3, we adjoin $-e_{15}+e_4$, $-e_{13}+e_6$ to $U$ to get an ad-nilpotent algebra $\tilde{U}$ and compute its normalizer. We find that $$N(\widetilde{U})\,=\, \langle \widetilde{U}\, , e_2\, ,e_7\, ,e_{14}\rangle\, .$$ The subalgebra $\langle e_2\, ,e_7\, ,e_{14}\rangle$ is abelian, and is a torus, whose real part is $\langle e_2\, ,e_7\rangle$ and compact part is $\langle e_{14}\rangle$.
Thus $N(\widetilde{U})$ is self–normalizing and solvable. Therefore, by Algorithm 3.3, $\widetilde{U}$ is a maximal ad-nilpotent subalgebra containing $U$. Using Algorithm 3.1, we find that $\langle e_2\, ,e_7\, ,e_{14}\rangle$ is a Cartan subalgebra and $A\,=\,\langle e_2\, ,e_7\rangle$ is a maximal abelian subalgebra of real semisimple elements.
The roots of $A$ on $N(\widetilde{U})$ are $$(-1\, ,0)\, , (-1\, ,-1)\, , (-1\, ,1)\, , (0\, ,1)\, ;$$ here, to say that $(r\, ,s)$ is a root means that there is a common eigen-vector $X$ for $A$ which is not centralized by $A$ and $$[e_7\, , X]\, =\, rX\, , [e_2\, , X]\, =\, sX\, .$$ Let $$a\,=\, (-1\, ,0)\, , b\,=\, (-1\, ,-1)\, , c\,=\, (-1\, ,1)\, , d\,=\, (0\, ,1)\, .$$ This is a positive system of roots for $A$ determined by $N(\widetilde{U})$. The only positive roots which are sums of positive roots are $a+d\,=\, c$ and $b+d\,=\, a$. Therefore, the simple roots are $b\, , d$ and the roots as nonnegative integral combinations of the simple roots are $b\, , d\, , b+d\, ,b+2d$.
Therefore, the Dynkin diagram of the reduced root system is of type $B_2$ with $b$ a long root.
Let $\omega_7\, , \omega_2$ be linear functions on $A$ dual to the ordered basis $e_7\, , e_2$. With this notation, the roots are $$-\omega_7\, ,-\omega_7-\omega_2\, ,-\omega_7+\omega_2\, , \omega_2\,
.$$ Let $L\,=\,N(\widetilde{U})$. The corresponding eigen-spaces in $L$ are $$L_{-\omega_7}\,=\, \langle e_{12}\, ,e_{11}\rangle\,
,L_{-\omega_7-\omega_2}\,=\,\langle e_8+e_{10} \rangle\, ,
L_{-\omega_7+\omega_2}\,=\, \langle e_8-e_{10}\rangle\, ,
L_{\omega_2}\,=\, \langle -e_{13}+e_6\, ,-e_{15}+e_4\rangle\, .$$ Finally $L_0\,=\, \langle A\, ,e_{14}\rangle$ and $e_{14}$ operates on these eigen-spaces, as rotations on $L_{-\omega_7}$ and $L_{\omega_2}$, while it commutes with $L_{-w_7-w_2}$ and $L_{-\omega_7+\omega_2}$.
The absolute root system is determined by common eigen-vectors for the Cartan algebra $$C\,=\, \langle e_7\, ,e_2\, ,e_{14}\rangle\, .$$ The positive roots are $$a\,=\, (0,1,-\sqrt{-1})\, , b\,=\,(0,1,\sqrt{-1})\, , c\,=\, (1,0,\sqrt{-1})$$ $$d\,=\, (1,1,0)\, , e\,=\, (1,-1,0)\, ,f\,=\, (1,0,-1)\, .$$ Forming sums of pairs of positive roots and removing those roots that are sums of positive roots, we find that the simple roots are $a\, ,e\, ,b$ with Dynkin diagram of type $A_3$ with $e$ the simple middle root.
Conjugation maps $a$ to $b$ and fixes $e$. Thus the algebra is a real form of $\text{sl}(4,{\mathbb C})$.
To find a maximally compact subalgebra – if any – we follow a procedure analogous to Algorithm 3.1 – starting with a compact element – namely one which generates a compact subgroup. For example, as $\langle e_{15}\, ,e_{14}\, ,e_{13}\rangle$ generate ${\mathfrak s}{\mathfrak o}(3)$, because $e_{15}\,=\,
x\partial_y-y\partial_x$ and $e_{14}\,=\, x\partial_z-z\partial_x$, we can with begin with $e_{15}$, compute its centralizer, the center of its centralizer and its derived algebra. If the derived algebra is trivial, then the centralizer $e_{15}$ of would be a maximal torus and its compact part will be a maximal compact subalgebra containing $e_{15}$. If the derived algebra is nontrivial, it must have a compact element, say $t$. Adjoining it to $e_{15}$ and computing the centralizer of $\langle e_{15}\, ,t\rangle$ and its derived algebra and repeating the process, we will ultimately obtain a maximal compact subalgebra containing $e_{15}$.
In this case we find that $C_k\,=\, \langle t_1\, ,t_2\, ,e_{15}\rangle$ is a maximally compact Cartan subalgebra, where $$t_1\,=\, 2e_{12}+e_5\, ,t_2\,=\, e_9+4e_8\, .$$ The positive roots are $$a\,=\, (\sqrt{-1}, -4\sqrt{-1},0)\, , b\,=\,(0,4\sqrt{-1},\sqrt{-1})\, , c\,=\,
(\sqrt{-1},0,-\sqrt{-1})\, ,$$ $$d\,=\, (0,4\sqrt{-1},-\sqrt{-1})\, , e\,=\, (\sqrt{-1},4\sqrt{-1},0)\, ,f\,=\,
(\sqrt{-1},0,\sqrt{-1})\, .$$ The simple positive roots are $d\, ,a\, ,b$ with Dynkin diagram of type $A_3$ with $a$ the middle simple root.
Conjugation maps every root to its negative. The root algebras generated by the real and imaginary parts of the root vectors are copies of $\text{sl}(2,{\mathbb R})$ except for roots $d+a\, ,a+b$, where they generate copies of ${\mathfrak s}{\mathfrak o}(3)$. Specifically, the subalgebras generated by the real and imaginary parts of root vectors for $d+a$ and $a+b$ are $$\langle e_1+2e_{10}-2e_{14},\, e_3+2e_{11}+2e_{13},\, -4e_5- 8e_{12}+8e_{15}\rangle\, ,$$ $$\langle e_1+2e_{10}+2e_{14},\,-e_3-2e_{11}+2e_{13},\, -4e_5- 8e_{12}-8e_{15}\rangle\, .$$ Both are isomorphic to ${\mathfrak s}{\mathfrak o}(3)$. Denoting these subalgebras by $k_1$ and $k_2$ respectively, we find that the centralizer of $k_1$ is $$k_2\oplus\langle 4e_8+e_9\rangle\, .$$ Moreover as the centralizer of the copy of ${\mathfrak s}{\mathfrak o}(3)$ given by $k_0\,=\,\langle e_{15}\,
,e_{14}\, ,e_{13}\rangle$ is $\langle e_7\,
,e_8\, ,e_9\rangle\,=\,\text{sl}(2,{\mathbb R})$, the subalgebra $k_1$ is not conjugate to $k_0$.
Finally, as the Killing form of the full algebra has seven negative eigen-values, a maximal compact subalgebra is $$k_1\oplus k_2\oplus \langle 4e_8 +e_9\rangle$$ because $4e_8 +e_9$ generates the maximal compact subalgebra of $\langle e_7\, ,e_8\, ,e_9\rangle\,=\,\text{sl}(2,{\mathbb R})$.
Lie Symmetries of $f_{xx}=\frac{4}{3}f_{yy}^3, f_{xy}=f^2_{yy}$ and $v'= (u'')^2$ {#sec5}
=================================================================================
These equations were considered by Cartan in the context of symmetries of a certain system of equations defined by differential forms [@Ca]; he showed that their symmetry algebra was the 14 dimensional simple group $G_{2}$. Maple is able to compute both the algebras by using commands for contact symmetries and for generalized symmetries as well as its root space decomposition and its maximal compact subalgebra. The latter equation was also considered by Anderson, Kamran and Olver, [@AKO], in the context of generalized symmetries. To illustrate the algorithms of this paper, we will use the table given in [@AKO] – reproduced in Appendix 2 to identify the algebra and determine several interesting subalgebras. To streamline the calculations we will use repeatedly the following facts-already mentioned in the Introduction.
If $H$ is a semisimple subalgebra of a semisimple algebra $G$ then and element $X$ of $H$ is real semisimple, compact or nilpotent in the adjoint representation $H$ of on itself,if and only if it is, respectively, real semisimple, compact or nilpotent in the adjoint representation of $H$ in $G$. Moreover, the derived algebras of the radical of normalizer or centralizer of any subalgebra are nilpotent.
The symmetry algebra has basis $X_1, X_2, \cdots , X_{14}$ and is given by:
$$\begin{aligned}
&X_{1} = (\frac{2}{3}u'^2-uu'')\partial_{x}+(\frac{1}{2}uv+\frac{4}{9}u'^3-uu'u'')\partial_{u}+(\frac{1}{2}v^2-\frac{1}{3}uu''^3)\partial_{v}, \\
&X_{2}=(\frac{4}{3}x^2u'-2xu-\frac{1}{3}x^3u'')\partial_{x}+(\frac{1}{6}x^3v+\frac{2}{3}x^2u'^2-2u^2-\frac{1}{3}x^3u'u'')\partial_{u}+(2xu'v-2uv-\frac{1}{9}x^3u''^3-\frac{8}{9}u'^3)\partial_{v}, \\
&X_{3} = (\frac{8}{3}xu'-2u-x^2u'')\partial_{x}+(\frac{1}{2}x^2v+\frac{4}{3}xu'^2-x^2u'u'')\partial_{u}+(2vu'-\frac{1}{3}x^2u''^3)\partial_{v}, \\
&X_{4} = (\frac{8}{3}u'-2xu'')\partial_{x}+(xv+\frac{4}{3}u'^2-2xu'u'')\partial_{u}-\frac{2}{3}xu''^3\partial_{v}, \\
\\
&X_{5} = -2u''\partial_{x}+(v-2u'u'')\partial_{u}-\frac{2}{3}u''^3\partial_{v}, \\
&X_{6} = \frac{1}{2}u\partial_{u} + v\partial_{v}, \\
&X_{7} = -\frac{1}{2}x^2\partial_{x} -\frac{3}{2}xu\partial_{u}-2u'^2\partial_{v}, \\
&X_{8} = -x\partial_{x}-\frac{3}{2}u\partial_{u}, \\
&X_{9} = -\partial_{x} , \\
&X_{10} = \frac{1}{6}x^3\partial_{u}+2(xu'-u)\partial_{v}, \\
&X_{11} = \frac{1}{2}x^2\partial_{u}+2u'\partial_{v}, \\
&X_{12} = x\partial_{u}, \\
&X_{13} = \partial_{u}, \\
&X_{14} = \partial_{v}.\end{aligned}$$
Using the commutator table in Appendix 2, and computing the determinant of the Killing form by Maple, we find that it is non-zero. Thus, the algebra is semisimple. The translations $$X_{14}\,=\, \partial_v\, ~~~ X_{13}\,=\, \partial_u$$ clearly commute and $X_{12}\,=\, x\partial_u$ commutes with both. One can check that they are nilpotent: this also follows from computing the derived algebra of the radical of normalizer of U. We want to embed $$U\,=\, \langle X_{14}\, , X_{13}\, , X_{12}\rangle$$ in a maximal subalgebra whose elements are all nilpotent. Following Algorithm 3.3 we compute the normalizer $N(U)$ and find its Levi decomposition, using standard Maple commands: $$N(U)\,=\, \langle X_9\, , X_8\, , X_6\, , X_5\, , X_{14}\, , X_{13}\, , X_{12}
\, , X_{11}\, , X_{10}\rangle$$ and its Levi decomposition is $R(N(U))\oplus S$, where the radical $$R(N(U))\,=\, \langle X_9\, , X_8 - 3X_6\, , X_{14}\, , X_{13}\, , X_{12}\, , X_{11}\rangle$$ and the semisimple part $S\,=\, \langle X_8 + X_6\, , X_5\, , X_{10}\rangle$. The commutators for the semisimple part are $$[X_8 + X_6\, , X_5] \,=\, 2X_5,\, [X_8 + X_6\, , X_{10} ]\,=\,
-2X_{10},\, [X_5 ,\, X_{10}]\,= -2(X_8 + X_6)\, .$$ This means that $X_5$ and $X_{10}$ are nilpotent in the full algebra, $X_8 + X_6$ is real semisimple in the full algebra and $X_5 + X_{10}$ is a compact element. Following Algorithm 3.3, we compute the derived algebra of the radical $R(N(U))$. It is $${\widetilde U}\,=\,\langle X_9\, , X_{14}\, , X_{13}\, , X_{12}\, , X_{11}\rangle\, .$$ The quotient $R(N(U)) /{\widetilde U}$ is represented by $X_8 - 3X_6$, which is a real semisimple element. (This also follows by computing the centralizer of $X_5 + X_{10}$ and its derived algebra, which turns out to be $\langle X_8 -3X_6\, , X_{12}\, , X_3\rangle$. This is a standard ${\rm sl}(2, {\mathbb R})$ with $X_8 -3X_6$ as real semisimple element.)
Following Algorithm 3.3 we compute again $N(\widetilde{U})$ and its Levi decomposition. It turns out to be identical to the Levi decomposition of $N({U})$. We therefore adjoint a nilpotent element coming from the semisimple part of the decomposition, say $X_5$. Let $$\widetilde{\widetilde{U}}\,=\, \langle \widetilde{U}, X_5\rangle\, .$$ Its normalizer is $\langle \widetilde{\widetilde{U}}\, , X_8\, , X_6\rangle$ and it is solvable, with commutator $\widetilde{\widetilde{U}}$ and the quotient is represented by the real torus $\langle X_6\, , X_8\rangle$. This also follows from noticing that $X_8 + X_6$ is also real semisimple and commutes with $X_8 - 3X_6$.
Thus a maximal nilpotent subalgebra containing $$U\,=\, \langle X_{14}\, , X_{13}\, , X_{12}\rangle$$ is $\widetilde{\widetilde{U}}\,=\,\langle X_5\, , X_{14}\, , X_{13}\, , X_{12}\, ,
X_{11}\, , X_9\rangle$. Finally, $C \,=\,\langle X_6\, , X_8\rangle$ is self-centralizing and it is a real split Cartan subalgebra of the full 14 dimensional algebra $L$.
Maple gives the following roots for $C$ in $\widetilde{\widetilde{U}}$: in fact, the basis vectors for $\widetilde{\widetilde{U}}$ listed above are common eigen-vectors for $C$ with eigen-values $$a\,=\,(\frac{1}{2}, \frac{3}{2}),\, b\,=\, (-1,0),\, c\,=\, (-\frac{1}{2}, \frac{3}{2}),\,
d\,=\, (-\frac{1}{2}, \frac{1}{2}),\, e\,=\, (-\frac{1}{2}, -\frac{1}{2}),\, f\,=\,
(0,1)\, .$$ As explained in Section \[sR\], this is a positive system of roots and a simple system of roots is given by adding pairs of positive roots and removing those that are a sum of positive roots. We have $$a + b\,=\, c,\, a + e\,=\, f,\, d + e\,=\, b,\, e + f\,=\, d\, .$$ Thus the simple roots are $a, e$ and the positive roots written in terms of these roots are $$a,\, e,\, a + e \,=\, f,\, a + 2e \,= \, d,\, a + 3e \,=\, b,\, 2a + 3e \,=\, c\, .$$ Therefore, the algebra $L$ is of type $G_2$ with a real split Cartan subalgebra. Any semisimple split real Lie algebra is generated by copies of ${\rm sl}(2,{\mathbb R})$ corresponding to the simple roots, with relations $$[X\, , Y] \,=\, H,\, [H\, , X] \,=\, 2X,\, [H\, , Y] \,=\, -2Y$$ and its maximal compact subalgebra is generated by copies of the compact element $X-Y$, which generates a circle, in these generating root ${\rm sl}(2,{\mathbb R})$ copies; see [@St pp. 99–100] for a global version of these results.
Here, the root vectors corresponding to $a, -a$ are $X_5 ,X_{10}$; the root vectors corresponding to $e, -e$ are $X_{11} , X_4$ and a maximal compact compact subalgebra $K$ is thus generated by $$J_1 \,=\, X_5 + X_{10} ,\, J_2 \,=\, X_4 - X_{11}\, .$$ The algebra is spanned by $$J_1 ,\, J_2 ,\, J_3 \,=\, X_1 + \frac{3}{8}X_{14} ,\, J_4 \,=\, X_2 - \frac{3}{4}X_{13} ,\,
J_5 \,=\, X_3 +\frac{3}{4}X_{12} ,\, J_6 \,=\, X_7 - \frac{3}{2}X_9\, .$$ To identify the structure of $K$, we must choose a Cartan subalgebra of $K$ and compute its roots in the complexification of $K$ — exactly as for ${\mathfrak s}{\mathfrak o}(4)$ in Section \[sec4.1l\]. Now the centralizer of $J_1$ is $\langle J_1\, , J_5\rangle$, and it is therefore a Cartan subalgebra of $K$. Its positive roots are $$(\sqrt{-2}\, , -\frac{1}{\sqrt{-2}}),\, (\sqrt{-2}\, , \frac{3}{\sqrt{-2}})\, .$$ Therefore the system is of type $A_1\times A_1$. The real and imaginary parts for the root vectors of $(\sqrt{-2}\, , -\frac{1}{\sqrt{-2}})$ are $$\label{g1}
J_3 - \frac{1}{6} J_6 ,\, \frac{\sqrt{2}}{4}J_4- \frac{\sqrt{2}}{8}J_2$$ and for the root $(\sqrt{-2}\, , \frac{3}{\sqrt{-2}})$ they are $$\label{g2}
J_3 + \frac{1}{2} J_6 ,\, \frac{-\sqrt{2}}{4}J_4+
\frac{3\sqrt{2}}{8}J_2\, .$$ The vectors in generate a copy of ${\mathfrak s}{\mathfrak o}(3)$ and in also a copy of ${\mathfrak s}{\mathfrak o}(3)$ and these subalgebras commute.
This gives an explicit decomposition of a maximal compact subalgebra of $L$ as a sum of copies of ${\mathfrak s}{\mathfrak o}(3)$.
Solutions of the wave equation
==============================
Section \[Se4\] gives several non-conjugate subalgebras of the symmetry algebra of the wave equation. The reason that they are non-conjugate is that the structure of low dimensional Lie algebras is well documented in literature [@PBNL], [@SW]. In the case of three dimensional algebras, this is the Lie-Bianchi classification [@Lie1 p. 479–562], [@Bian] In this case, this can be described very briefly. If $L$ is a 3 dimensional algebra and its commutator $L'\,=\,[L,\, L]$ is 1–dimensional, then $L$ is completely determined by the dimension of the centralizer of $L'$ in $L$; if $L'$ is 2–dimensional, then $L'$ is abelian and the structure of $L$ is completely determined by the eigen-values of $L/L'$ in $L'$ and their multiplicities; in case $L'$ is of dimension 3, the eigen-values of a single element suffice to determine the structure of $L$ [@ADMM Corollary 2.2 and Section 4.3]. The algebras given below were of all possible Lie–Bianchi types. Their identification is facilitated by determining the reduced root system using the algorithms of Section \[aL\], by enlarging a given subalgebra of commuting ad-nilpotent elements to a maximal solvable subalgebra that contains up to conjugacy all solvable subalgebras with real eigen-values. If the vector fields are in polynomial form and contain translations with respect to the independent variables, then these translations are ad-nilpotent. In case the Cartan algebra so obtained has a compact part, it must operate on the positive root spaces of its real part and this way one may obtain all 3 dimensional subalgebras of solvable subalgebras of all Lie–Bianchi types.
In the specific example of the 15 dimensional algebra considered in Section \[se4.4\], we denoted — for simplicity of notation the element $e_7 -e_{16}$ by $e_7$. Taking this into account, the Cartan algebras obtained in Section \[se4.4\] were $\langle e_2,\, e_7 -e_{16},\, e_{14}\rangle$ and $\langle 2e_{12} + e_5,
\,e_9+ 4e_8,\, e_{15} \rangle$. This first Cartan algebra is maximally real with its real part $A\,=\, \langle e_2,\, e_7 -e_{16}\rangle$. For this reason, the relative root system is different from the absolute root system. The roots of $A$ were determined in Section \[se4.4\] as $$b\,=\, -\omega_7-\omega_2\, , \ d\,=\, \omega_2\, , \ b+d\,=\, -\omega_7\, ,
\ b+2d\,=\, \omega_2- \omega_7\, ,$$ where $\omega_2,\, \omega_7$ are dual to the ordered basis $e_2,\, e_7 -e_{16}$ of $A$.
The root spaces of $A$ in the maximal solvable algebra $L$ determined in Section \[se4.4\] were
- $L_b\,=\, L_{-\omega_7-\omega_2}\,=\, \langle e_8,\, e_{10}\rangle$,
- $L_d\,=\, L_{\omega_2}\,=\, \langle e_6-e_{13},\, e_{4}-e_{15}\rangle$,
- $L_{b+d}\,=\, L_{-\omega_7}\,=\, \langle e_{12},\, e_{11}\rangle$,
- $L_{b+2d}\,=\, L_{\omega_2-\omega_7}\,=\, \langle e_{8}- e_{10}\rangle$.
Thus, this displays the common eigen-vectors of $A$ and their multiplicities. Moreover as $[L_r,\, L_s]\, \subset\, L_{r+s}$, and $2r$ is not a root, the root spaces given above contain commuting eigen-vectors of $e_2$ with different eigen-values and of $e_7 -e_{16}$ with repeated eigen-values. As any element centralizing $A$ operates on each root space of $A$, applying this to the compact part of the Cartan algebra $\langle A,\, e_{14}\rangle$ gives all possible solvable 3–dimensional solvable Lie–Bianchi types. Finally, using the compact Cartan subalgebra, we found a maximal compact subalgebra, whose derived algebra was ${\rm so}(4)\,=\, {\rm so}(3)\oplus {\rm so}(3)$. The centralizer for the spatial rotations was ${\rm sl}(2,{\mathbb R})$. For this reason, the 3–dimensional simple algebras given below are non-conjugate. We now proceed to find the corresponding reductions and invariant solutions.
A preliminary step is to find the invariants of a given algebra of vector fields. The number of functionally independent invariants can be found from the row reduced echelon form of the operators. In the row reduced echelon form, the resulting operators always commute [@ABGM].
The equation is $$\label{eq-w}
{\frac {\partial ^{2}}{\partial {x}^{2}}}u \left( t,x,y,z
\right) +{\frac {\partial ^{2}}{\partial {y}^{2}}}u \left( t,x
,y,z \right) +{\frac {\partial ^{2}}{\partial {z}^{2}}}u \left(
t,x,y,z \right) -{\frac {\partial ^{2}}{\partial {t}^{2}}}u
\left( t,x,y,z \right) =0$$
**(I)** [$\dim G' =0$]{}
**(I.a)** ${\mathcal L}_{1,0}
\,=\, \langle e_{8},e_{10},e_{11}\rangle$
The joint invariants of ${\mathcal L}_{1,0}$ are $$\nonumber
z, u \\$$ so that the corresponding similarity transformations $$p=z, w(p)=u \\$$ transform wave equation to $${\frac {{\rm d}^{2}}{{\rm d}{p}^{2}}}w \left( p \right) =0$$ which has the solution $$w \left( p \right) ={\it C_1}\,p+{\it C_2}.$$ This leads to solution $$u \left( t,x,y,z \right) ={\it C_1}\,z+{\it C_2}$$ of wave equation .\
**(I.b)** ${\mathcal L}_{2,0} \,=\,\langle e_{2},e_7 -
e_{16},e_{14}\rangle$
The joint invariants of ${\mathcal L}_{2,0}$ are $$\nonumber
-{\frac {{x}^{2}+{z}^{2}}{{y}^{2}-{{\it t}}^{2}}}, u_{{}}\sqrt {-{y}^{2}+{{\it t}}^{2}} \\$$ so that the corresponding similarity transformations $$p=-{\frac {{x}^{2}+{z}^{2}}{{y}^{2}-{{\it t}}^{2}}}, w(p)=u_{{}}\sqrt {-{y}^{2}+{{\it t}}^{2}} \\$$ transform wave equation to Jacobi ODE $$4\, \left( {\frac {{\rm d}^{2}}{{\rm d}{p}^{2}}}w \left( p \right)
\right) {p}^{2}-4\, \left( {\frac {{\rm d}^{2}}{{\rm d}{p}^{2}}}w
\left( p \right) \right) p+8\, \left( {\frac {\rm d}{{\rm d}p}}w
\left( p \right) \right) p-4\,{\frac {\rm d}{{\rm d}p}}w \left( p
\right) +w \left( p \right) =0$$ which has the solution $$w \left( p \right) ={\it C_1}\,{\it EllipticK} \left( \sqrt {p}
\right) +{\it C_2}\,{\it EllipticCK} \left( \sqrt {p} \right)$$ in terms of complete and complementary complete elliptic integrals of the first kind\
(ref: http://www.maplesoft.com/support/help/Maple/view.aspx?path=EllipticF). This leads to solution $$u \left( {\it t},x,y,z \right) ={\frac {1}{\sqrt {-{y}^{2}+{{\it t}}
^{2}}} \left( {\it C_1}\,{\it EllipticK} \left( \sqrt {{\frac
{-{x}^{ 2}-{z}^{2}}{{y}^{2}-{{\it t}}^{2}}}} \right) +{\it
C_2}\,{\it EllipticCK} \left( \sqrt {{\frac
{-{x}^{2}-{z}^{2}}{{y}^{2}-{{\it t}} ^{2}}}} \right)\right) }$$ of wave equation .\
**(I.c)** ${\mathcal L}_{3,0} \,=\, \langle e_{12} +
\frac{1}{2}e_5, e_9 + 4 e_{8},e_{15}\rangle$
The joint invariants of ${\mathcal L}_{3,0}$ are $$\nonumber
{\frac {-{{\it t}}^{4}+ \left( 2\,{x}^{2}+2\,{y}^{2}+2\,{z}^{2}-8
\right) {{\it t}}^{2}-{x}^{4}+ \left( -2\,{y}^{2}-2\,{z}^{2}
\right) {x}^{2}-{y}^{4}-2\,{y}^{2}{z}^{2}- \left( {z}^{2}+4 \right) ^
{2}}{4\,{x}^{2}+4\,{y}^{2}}} , u \sqrt {{x}^{2}+{y}^{2}}$$ so that the corresponding similarity transformations $$\label{p-3-0}
p= {\frac {-{{\it t}}^{4}+ \left( 2\,{x}^{2}+2\,{y}^{2}+2\,{z}^{2}-8
\right) {{\it t}}^{2}-{x}^{4}+ \left( -2\,{y}^{2}-2\,{z}^{2}
\right) {x}^{2}-{y}^{4}-2\,{y}^{2}{z}^{2}- \left( {z}^{2}+4 \right) ^
{2}}{4\,{x}^{2}+4\,{y}^{2}}},$$ $$w(p)=u \sqrt {{x}^{2}+{y}^{2}}$$ transform wave equation to $$4\, \left( {\frac {{\rm d}^{2}}{{\rm d}{p}^{2}}}w \left( p \right)
\right) {p}^{2}+8\, \left( {\frac {\rm d}{{\rm d}p}}w \left( p
\right) \right) p-16\,{\frac {{\rm d}^{2}}{{\rm d}{p}^{2}}}w \left(
p \right) +w \left( p \right) =0$$ which has the solution $$w \left( p \right) ={\it C_1}\,{\it LegendreP} \left( -1/2,p/2
\right) +{\it C_2}\,{\it LegendreQ} \left( -1/2,p/2 \right)$$ in terms of Legendre functions of the first and second kind\
(ref: http://www.maplesoft.com/support/help/Maple/view.aspx?path=Legendre). This leads to solution $$u \left( {\it t},x,y,z \right) = \frac{1}{\sqrt {{x}^{2}+{y}^{2}}}
\left( {\it C_1}\,{\it LegendreP} \left( -1/2,p/2
\right) +{\it C_2}\,{\it LegendreQ} \left( -1/2,p/2 \right)
\right)$$ of wave equation where $p$ is given by .\
**(II)** [$\dim G' =1$]{}
**(II.a)** ${\mathcal L}_{1,1}
\,=\,\langle e_{2},e_{7}-e_{16},e_8 + e_{10}\rangle$
The joint invariants of ${\mathcal L}_{1,1}$ are $$\nonumber
{\frac {z}{x}}, u_{{}}x \\$$ so that the corresponding similarity transformations $$p={\frac {z}{x}}, w(p)=u_{{}}x \\$$ transform wave equation to $$\left( {\frac {{\rm d}^{2}}{{\rm d}{p}^{2}}}w \left( p \right)
\right) {p}^{2}+4\, \left( {\frac {\rm d}{{\rm d}p}}w \left( p
\right) \right) p+2\,w \left( p \right) +{\frac {{\rm d}^{2}}{
{\rm d}{p}^{2}}}w \left( p \right) =0$$ which has the solution $$w \left( p \right) ={\frac {{\it C_1}\,p+{\it C_2}}{{p}^{2}+1}}.$$ This leads to solution $$u \left( {\it t},x,y,z \right) ={\frac {{\it C_1}\,z+{\it C_2}\,x}{
{x}^{2}+{z}^{2}}}$$ of wave equation .\
**(II.b)** ${\mathcal L}_{2,1}
\,=\,\langle e_{12},-e_{6}+e_{13},-e_8 + e_{10}\rangle$
The joint invariants of ${\mathcal L}_{2,1}$ are $$\nonumber
x, t+y, u$$ which gives the similarity transformation $$p\,=\,x,\, q\,=\,t+y,\, w(p,q)\,=\,u_{{}}{ } \\$$ that transforms the wave equation into $$w_{pp}\,=\,0\, ,$$ which gives the solution $$u(t,x,y,z) \,=\, xF_{1} (y+t) + F_{2} (y+t)\, .$$
**(III)** [$\dim G' =2$]{}
**(III.a)** ${\mathcal L}_{1,2}
\,=\,\langle e_{7}-e_{16},e_{11}, e_{12}\rangle$
The joint invariants of ${\mathcal L}_{1,2}$ are $$\nonumber
{\frac {y}{{\it t}}}, u_{{}}{\it t} \\$$ so that the corresponding similarity transformations $$p={\frac {y}{{\it t}}}, w(p)=u_{{}}{\it t} \\$$ transform wave equation to $$\left( {\frac {{\rm d}^{2}}{{\rm d}{p}^{2}}}w \left( p \right)
\right) {p}^{2}+4\, \left( {\frac {\rm d}{{\rm d}p}}w \left( p
\right) \right) p+2\,w \left( p \right) -{\frac {{\rm d}^{2}}{
{\rm d}{p}^{2}}}w \left( p \right) =0$$ which has the solution $$w \left( p \right) ={\frac {{\it C_1}\,p+{\it C_2}}{{p}^{2}-1}}.$$ This leads to solution $$u \left( {\it t},x,y,z \right) ={\frac {{\it C_1}\,y+{\it C_2}\,{
\it t}}{{y}^{2}-{{\it t}}^{2}}}$$ of wave equation .\
**(III.b)** ${\mathcal L}_{2,2}
\,=\,\langle e_{2},\, e_8+e_{10}, e_8-e_{10}\rangle$ $$\,=\, \langle y\frac{\partial}{\partial t}+
t\frac{\partial}{\partial y},\, \frac{\partial}{\partial t}+
\frac{\partial}{\partial y},\, \frac{\partial}{\partial t}-
\frac{\partial}{\partial y}\rangle\, .$$ Clearly, joint invariants the same as the invariants of $\langle \frac{\partial}{\partial y},\,
\frac{\partial}{\partial t}\rangle$, therefore the basic invariants are $x,\, z,\, u$. Substituting $u\,=\, u(x,\, z)$ in the wave equation shows that $u$ must be a harmonic function, so they are real parts of holomorphic functions in the variable $x+\sqrt{-1}z$.\
**(III.c)** ${\mathcal L}_{3,2} \,=\,\langle e_{14},e_{11},
e_{12}\rangle$.
The joint invariants of ${\mathcal L}_{3,2}$ are $$\nonumber
t,\, y, \, u \\$$ so that the corresponding similarity transformations $$p\,=\,t,\, q\,=\,y, \, w(p,q)\,=\,u \\$$ transform wave equation to $${\frac {\partial^{2}}{\partial {q}^{2}}}w \left( p,q \right) -{\frac
{\partial ^{2}}{\partial {p}^{2}}}w \left( p,q \right) =0$$ which has the solution $$w \left( p,q \right) ={\it F_1} \left( q+p \right) +{\it F_2}
\left( q-p \right).$$ This leads to solution $$u \left( {\it t},x,y,z \right) ={\it F_1} \left( y+{\it t} \right)
+{\it F_2} \left( y-{\it t} \right)$$ of wave equation .\
**(III.d)** ${\mathcal L}_{4,2} \,=\,\langle e_{14},- e_6+
e_{13}, -e_4 + e_{15}\rangle$
The joint invariants of ${\mathcal L}_{4,2}$ are $$\nonumber
y+{\it t}, {x}^{2}-2\,{\it t}\,y+{z}^{2}-2\,{{\it t}}^{2}, u \\$$ so that the corresponding similarity transformations $$p\,=\,y+{\it t},\ q\,=\, {x}^{2}-2\,{\it t}\,\ y+{z}^{2}-2\,{{\it t}}^{2},\ w(p,q)\,=\,u \\$$ transform wave equation to $$\left( -{p}^{2}+q \right) {\frac {\partial ^{2}}{\partial {q}^{2}}}w
\left( p,q \right) + \left( {\frac {\partial ^{2}}{\partial q
\partial p}}w \left( p,q \right)\right) p+2\,{\frac {\partial }{
\partial q}}w \left( p,q \right) =0$$ which has the solution $$w \left( p,q \right) ={\frac {1}{p} \left( {\it F_2} \left( p
\right) p+{\it F_1} \left( {\frac {{p}^{2}+q}{p}} \right) \right) }.$$ This leads to solution $$u \left( {\it t},x,y,z \right) ={\it F_2} \left( y+{\it t} \right)
+{\frac {1}{y+{\it t}}{\it F_1} \left( {\frac {{x}^{2}+{y}^{2}+{z}^{
2}-{{\it t}}^{2}}{y+{\it t}}} \right) }$$ of wave equation .\
**(IV)** [$\dim G' =3$]{}
**(IV.a)** ${\mathcal L}_{1,3} \,=\,\langle e_{15},e_{14},
e_{13}\rangle$
The joint invariants of ${\mathcal L}_{1,3}$ are $$\nonumber
t, \, {x}^{2}+{y}^{2}+{z}^{2},\, u \\$$ so that the corresponding similarity transformations $$p\,=\,t, \, q\,=\, {x}^{2}+{y}^{2}+{z}^{2}, \, w(p,q)\,=\,u \\$$ transform wave equation to $$4\, \left( {\frac {\partial ^{2}}{\partial {q}^{2}}}w \left( p,q
\right) \right) q+6\,{\frac {\partial }{\partial q}}w \left( p,q
\right) -{\frac {\partial ^{2}}{\partial {p}^{2}}}w \left( p,q
\right) =0$$ which has the solution $$w \left( p,q \right) ={\frac {{\it F_1} \left( \sqrt {q}+p \right)
+{ \it F_2} \left( -\sqrt {q}+p \right) }{\sqrt {q}}}.$$ This leads to solution $$u \left( {\it t},x,y,z \right) ={\frac {{\it F_1} \left( \sqrt {{x}^
{2}+{y}^{2}+{z}^{2}}+{\it t} \right) +{\it F_2} \left( -\sqrt {{x}^{
2}+{y}^{2}+{z}^{2}}+{\it t} \right) }{\sqrt
{{x}^{2}+{y}^{2}+{z}^{2}} }}$$ of wave equation .\
**(IV.b)** ${\mathcal L}_{2,3} \,=\,\langle e_7 -e_{16},e_{8},
e_{9}\rangle$
The joint invariants of ${\mathcal L}_{2,3}$ are $$\nonumber
{\frac {y}{x}},\, {\frac {z}{x}}, \, u x \\$$ so that the corresponding similarity transformations $$p\,=\,{\frac {y}{x}},\, q\,=\, {\frac {z}{x}},\, w(p,q)\,=\,u x\\$$ transform wave equation to $$\left( {\frac {\partial ^{2}}{\partial {p}^{2}}}w
\right) {p}^{2}+2\,pq{\frac {\partial ^{2}}{\partial q\partial p}}w
+ \left( {\frac {\partial ^{2}}{\partial {q}^{2}}}
w \right) {q}^{2}+4\,p{\frac {\partial }{\partial p}}w +4\,
\left( {\frac {\partial }{\partial q}}w
\right) q+{\frac {\partial ^{2}}{\partial {p}^{2}
}}w +2\,w +{\frac {\partial ^{2} }{\partial {q}^{2}}}w =0$$ which has the solution $$w \left( p,q \right) ={\frac {{\it C_1}\,p+{\it C_2}}{{p}^{2}+1}}+{
\frac {{\it C_3}\,q+{\it C_4}}{{q}^{2}+1}}.$$ This leads to solution $$u \left( {\it t},x,y,z \right) ={\frac {1}{x} \left( {\frac { \left(
{\it C_1}\,y+{\it C_2}\,x \right) x}{{x}^{2}+{y}^{2}}}+{\frac {
\left( {\it C_3}\,z+{\it C_4}\,x \right) x}{{x}^{2}+{z}^{2}}}
\right) }$$ of wave equation .\
**(IV.c)** ${\mathcal L}_{3,3}
\,=\,\langle e_1 + 2e_{10} + 2 e_{14}, -e_{3}-2 e_{11} + 2 e_{13}, -4 e_5 - 8 e_{12} - 8 e_{15}\rangle$\
The joint invariants of ${\mathcal L}_{3,3}$ are $$\nonumber
{\frac {{x}^{2}+{y}^{2}+{z}^{2}-{{\it t}}^{2}+4}{{\it t}}} ,\, u t \\$$ so that the corresponding similarity transformations $$p\,=\,{\frac {{x}^{2}+{y}^{2}+{z}^{2}-{{\it t}}^{2}+4}{{\it t}}}, \, w(p)\,=\, u t \\$$ transform wave equation to $$\left( {\frac {{\rm d}^{2}}{{\rm d}{p}^{2}}}w \left( p \right)
\right) {p}^{2}+4\, \left( {\frac {\rm d}{{\rm d}p}}w \left( p
\right) \right) p+2\,w \left( p \right) +16\,{\frac {{\rm d}^{2}}{
{\rm d}{p}^{2}}}w \left( p \right) =0$$ which has the solution $$w \left( p \right) ={\frac {{\it C_1}\,p+{\it C_2}}{{p}^{2}+16}}$$ This leads to solution $$u \left( {\it t},x,y,z \right) = {\frac {-{\it C_1}\,{{\it t}}^{2}+{
\it C_2}\,{\it t}+{\it C_1}\, \left( {x}^{2}+{y}^{2}+{z}^{2}+4
\right) }{{{\it t}}^{4}+ \left( -2\,{x}^{2}-2\,{y}^{2}-2\,{z}^{2}+8
\right) {{\it t}}^{2}+ \left( {x}^{2}+{y}^{2}+{z}^{2}+4 \right) ^{2}
}}$$ of wave equation .\
Appendices {#appendices .unnumbered}
==========
**Appendix 1**\
\[T32\]
where $[X_{i},X_{j}]\,=\,-[X_{j},X_{i}]$.
**Appendix 2**\
\[T32b\]
where $[X_{i},X_{j}]\,=\,-[X_{j},X_{i}]$.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank the two referees for helpful comments. One of us (HA) thanks Karl-Hermann Neeb for a very helpful correspondence. The third-author is supported by a J. C. Bose Fellowship.
[ZZZZZ]{}
I. Agricola, Old and new on the exceptional group $G_2$, [*Notices Amer. Math. Soc.*]{} [**55**]{} (2008), 922–929.
I. M. Anderson, N. Kamran and P. J. Olver, Internal, external, and generalized symmetries, [*Adv. Math.*]{} [**100**]{} (1993), 53–100.
I. Anderson, http://digitalcommons.usu.edu/dg/.
H. Azad, A. Y. Al-Dweik, R. Ghanam and M. T. Mustafa, Symmetry analysis of wave equation on static spherically symmetric spacetimes with higher symmetries, [*Jour. Math. Phys.*]{} [**54**]{}, 063509 (2013).
H. Azad, A. Y. Al-Dweik, F. M. Mahomed and M. T. Mustafa, A point symmetry based method for transforming ODEs with three–dimensional symmetry algebras to their canonical forms, [*Appl. Math. Comp.*]{} [**289**]{} (2016), 444–463.
H. Azad and I. Biswas, A note on real algebraic groups, [*Forum Math.*]{} [**28**]{} (2016), 539–543.
H. Azad, I. Biswas, R. Ghanam and M. T. Mustafa, On computing joint invariants of vector fields, [*Jour. Geom. Phys.*]{} [**97**]{} (2015), 69–76.
L. Bianchi, Sugli spazii a tre dimensioni che ammettono un gruppo continuo di movimenti, [*Soc. Ital. Sci. Mem. di Mat.*]{} [**11**]{} (1898), 267–352.
G. Bluman and S. Anco, *Symmetry and Integration Methods for Differential Equations*, Applied Mathematical Sciences, 154, Springer-Verlag, New York, 2002.
G. Bluman, A. Cheviakov and S. Anco, *Applications of Symmetry Methods to Partial Differential Equations*, Applied Mathematical Sciences, 168, Springer-Verlag, New York, 2010.
G. Bor and R. Montgomery, $G_2$ and the “Rolling Distribution”, [*L’Enseign. Math.*]{} [**55**]{} (2009), 157–196.
A. Borel, Lie groups and linear algebraic groups I. Complex and real groups, 1–49, in: [*Lie Groups and Automorphic Forms*]{}, AMS/IP Stud. Adv. Math. 37, American Mathematical Society, Providence (2006). http://hkumath.hku.hk/$\sim$imr/records0001/borel.pdf.
A. Borel and J. Tits, Groupes réductifs, [*Inst. Hautes Études Sci. Publ. Math.*]{} [**27**]{} (1965), 55–150.
E. Cartan, Les systèmes de Pfaff, à cinq variables et les équations aux dérivées partielles du second ordre, [*Ann. Sci. École Normale Sup*]{} [**27**]{} (1910), 109–192.
H. Derksen, E. Jeandel and P. Koiran, Quantum automata and algebraic groups, [*Jour. Symbolic Comput.*]{} [**39**]{} (2005), 357–371.
H. Dietrich, P. Faccin and W. A. de Graaf, Computing with real Lie algebras: Real forms, Cartan decompositions, and Cartan subalgebras, [*J. Symbolic Comput.* ]{} [**56**]{} (2013), 27–45.
W. A. de Graaf, [*Lie Algebras: Theory and Algorithms*]{}, North-Holland Mathematical Library, 36, Amsterdam: Elsevier Science, 2000.
L. Gagnon and P. Winternitz, Lie symmetries of a generalised nonlinear Schrödinger equation. I. The symmetry group and its subgroups, [*Jour. Phys.*]{} A [**21**]{} (1988), 1493–1511.
J. Hilgert and K.-H. Neeb, *Structure and Geometry of Lie Groups*, Springer-Verlag, New York 2011.
N. H. Ibragimov, [*Elementary Lie group analysis and ordinary differential equations*]{}, Wiley series in Mathematical Methods in Practice, 4, John Wiley & Sons, Ltd. Chichester, 1999.
N. H. Ibragimov, Selected works, vol I and vol II, ALGA Publications Blekinge Institute of Technology Karlskrona, Sweden, 2006.
P. H. M. Kersten, The general symmetry algebra structure of the underdetermined equation $u_x= (v_{xx})^2$, [*J. Math. Phys.*]{} [**32**]{} (1991), 2043–2050.
A. W. Knapp, *Lie groups beyond an introduction*, second edition, Progress in Mathematics, 140, Birkhäuser Boston, Inc., Boston, MA, 2002.
S. Lie, Vorlesungen über Differentialgleichungen mit bekannten infinitesimalen Transformationen, B.G. Teubner, 1891.
S. Lie, *Gesammelte Abhandlungen, vol. 4*, B.G.Teubner, Leipzig, (1929), 320–384.
G. D. Mostow, On maximal subgroups of real Lie groups, *Ann. of Math.* **73** (1961), 20–48.
P. J. Olver, *Applications of Lie Groups to Differential Equations*, Springer-Verlag, New York, 1993.
D. Rand, P. Winternitz and H. Zassenhaus, On the identification of a Lie algebra given by its structure constants. I. Direct decompositions, Levi decompositions, and nilradicals, [*Linear Algebra Appl.*]{} [**109**]{} (1988), 197–246.
J. Patera, P. Winternitz and H. Zassenhaus, Continuous subgroups of the fundamental groups of physics. I. General method and the Poincaré group, [*J. Math. Phys.*]{} [**16**]{} (1975), 1597–1614.
R, O. Popovych, V. M. Boyko, M. O. Nesterenko and M. W. Lutfullin, Realizations of real low-dimensional Lie algebras, [*Jour. Phys. A*]{} [**36**]{} (2003), 7337–7360.
L. Šnobl and P. Winternitz, *Classification and Identification of Lie Algebras*, American Mathematical Society: Providence, RI, USA, 2014.
R. Steinberg, [*Lectures on Chevalley groups*]{}. Notes prepared by John Faulkner and Robert Wilson, Yale University, New Haven, Conn., 1968.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The adversarial training procedure proposed by [@madry2017towards] is one of the most effective methods to defend against adversarial examples in deep neural networks (DNNs). In our paper, we shed some lights on the practicality and the hardness of adversarial training by showing that the effectiveness (robustness on test set) of adversarial training has a strong correlation with the distance between a test point and the manifold of training data embedded by the network. Test examples that are relatively far away from this manifold are more likely to be vulnerable to adversarial attacks. Consequentially, an adversarial training based defense is susceptible to a new class of attacks, the *“blind-spot attack”*, where the input images reside in “blind-spots” (low density regions) of the empirical distribution of training data but is still on the ground-truth data manifold. For MNIST, we found that these blind-spots can be easily found by simply scaling and shifting image pixel values. Most importantly, for large datasets with high dimensional and complex data manifold (CIFAR, ImageNet, etc), the existence of blind-spots in adversarial training makes defending on any valid test examples difficult due to the curse of dimensionality and the scarcity of training data. Additionally, we find that blind-spots also exist on provable defenses including [@kolter2017provable] and [@sinha2018certifying] because these trainable robustness certificates can only be practically optimized on a limited set of training data.'
author:
- |
Huan Zhang^1^$^{*}$, Hongge Chen^2^$^{*}$, Zhao Song^3^, Duane Boning^2^, Inderjit Dhillon^3^,\
**Cho-Jui Hsieh^1^**\
^1^UCLA, Los Angeles, CA 90095\
^2^MIT, Cambridge, MA 02139\
^3^UT Austin, Austin, TX 78712\
`[email protected], [email protected], [email protected]`\
`[email protected], [email protected], [email protected]`\
$^{*}$Huan Zhang and Hongge Chen contributed equally to this work.
bibliography:
- 'iclr2019\_conference.bib'
title: 'The Limitations of Adversarial Training and the Blind-Spot Attack'
---
Introduction
============
Related Works
=============
Methodology
===========
Experiments
===========
Conclusion
==========
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Many physical problems can be described using the integral equations known as Volterra equations. There exists quite a number of analytical methods that can handle these equations for linear cases. For non-linear cases, a numerical scheme is needed to obtain an approximation. Recently a new concept of fractional differentiation was introduced using the Mittag-Leffler function as kernel, and the associate fractional integral was also presented. Up to this point there is no numerical approximation of this new integral in the literature. Therefore, to accommodate researchers working in the field of numerical analysis, we propose in this paper a new numerical scheme for the new fractional integral. To test the accuracy of the new numerical scheme, we first revisit the groundwater model within a leaky aquifer by reverting the time classical derivative with the Atangana-Baleanu fractional derivative in Caputo sense. The new model is solved numerically by using this new scheme.'
address:
- 'African Institute for Mathematical Sciences (AIMS), P.O. Box 608, Limbe Crystal Gardens, South West Region, Cameroon. '
- 'Departamento de Matemática Aplicada II, E.E. Aeronáutica e do Espazo, Universidade de Vigo, Campus As Lagoas s/n, 32004 Ourense, Spain.'
- 'Institute for Groundwater Studies, Faculty of Natural and Agricultural Sciences, University of the Free State, 9301, Bloemfontein, South Africa.'
author:
- 'J.D. Djida'
- 'I. Area'
- 'A. Atangana'
title: 'New numerical scheme of Atangana-Baleanu fractional integral: an application to groundwater flow within leaky aquifer'
---
[^1]
Introduction
============
The concept of fractional differentiation with non-singular and non-local kernel has been suggested recently and is becoming a hot topic in the field of fractional calculus. The concept was tested in many fields including chaotic behaviour, epidemiology, thermal science, hydrology and mechanical engineering [@Abdon2; @Ivan2015; @Abdon_Nieto]. The numerical approximation of this differentiation was also proposed in [@Abdon_Nieto]. In the recent decade, the integral equations were revealed to be great mathematical tools to model many real world problems in several fields of science, Technology and Engineering. In many research papers under some conditions (see e.g. [@Meerschaert; @Zhuang; @Hao] and references therein), it was proven that there is equivalence between a given differential equation and its integral equation associate.
Recently [@Zhuang] proposed a method based on a semi-discrete finite difference approximation in time and Galerkin finite element method in space. In this work we propose a new numerical approximation of Atangana-Baleanu integral which is the summation of the average of the given function and its fractional integral in Riemann-Liouville sense. This numerical scheme will be validate by solving the partial differential equation describing the subsurface water flowing within a confined aquifer model with the new derivative with fractional order in time component.
This paper is organized as follows: In Section \[Sec:2\], for the convenience of the reader, we recall some definitions and properties of fractional Calculus within the scope of Atangana-Baleanu. In Section \[Sec:3\] a numerical approach of Atangana-Baleanu derivative with fractional order is introduced. In Section \[Sec:4\], as application of the new numerical approximation of Atangana-Baleanu fractional integral. We study analytically and numerically the model of groundwater flow within a leaky aquifer based upon the Mittag-Leffler function. Finally, Section \[Sec:5\] is dedicated to our perspectives and conclusions.
Somes definitions of the Atangana-Baleanu fractional derivative and integral {#Sec:2}
============================================================================
We recall the definitions of the new derivative with non singular kernel and integral introduced by Atangana and Baleanu in the senses of Caputo and Riemann-Liouville derivatives [@Abdon2; @Abdon3].
Let $(a,b) \subset {\mathbb{R}}$ and let $u$ be a function of the Hilbert space $L^{2}(a,b)$. We define by $u'$ the derivative of $u$ as distribution on $(a,b)$.
The Sobolev space of order $1$ in $(a,b)$ is defined as $$H^{1}(a,b)=\{ u \in L^{2}(a,b) \,\vert\, \,\, u' \in L^{2}(a,b) \}.$$
\[Atangana-Baleanu\_Caputo\] Let $\alpha \in (0,1)$ and a function $u \in H^{1}(a,b)$, $b > a$. The Atangana-Baleanu fractional derivative in Caputo sense of order $\alpha$ of $u$ with a based point $a$ is defined as $$\label{Atangana-Baleanu_Caputo:definition}
^{\textsc{ABC}}{\mathcal{D}}_t^{\alpha}u\,(t)= \frac{B(\alpha)}{1 - \alpha}\int_{a}^{t} u'(s)E_{\alpha}\bigg[-\frac{\alpha}{1-\alpha}\big(t - s\big)^{\alpha} \bigg]ds,$$ where $B(\alpha)$ has the same properties as in Caputo and Fabrizio case, and is defined as $$B(\alpha)=1-\alpha +\frac{\alpha }{\Gamma (\alpha )},$$ $E_{\alpha, \beta}(\lambda z^{\alpha})$ is the Mittag-Leffler function, defined in terms of a series as the following entire function $$\label{def:mittag}
E_{\alpha,\beta}(z)=\sum_{k=0}^{\infty} \frac{\big(\lambda z^{\alpha}\big)^{k}}{\Gamma(\alpha k + \beta)}, \quad \alpha > 0, \quad \lambda <
\infty,~~\text{and}~~
\quad \beta>0,$$ $\lambda = -\alpha(1-\alpha)^{-1}$.
The Mittag-Leffler functions appear in the solution of linear and nonlinear fractional differential equations [@MR1658022]. The above definition is very helpful to discuss real world problems and will also have a great advantage when using the Laplace transform with initial condition. Now let us recall the definition of the Atangana-Baleanu fractional derivative in the Riemann-Liouville sense.
\[Atangana-Baleanu\_Riemann\] Let $\alpha \in (0,1)$ and a function $u \in H^{1}(a,b), b > a$. The Atangana-Baleanu fractional derivative in the Riemann-Liouville sense of order $\alpha$ of $u$ is defined as $$\label{Atangana-Baleanu_Riemann:definition}
^{\textsc{ABR}}{\mathcal{D}}_t^{\alpha}u\,(t)= \frac{B(\alpha)}{1 - \alpha}\frac{d}{dt}\int_{a}^{t}u(s) E_{\alpha}\bigg[-\frac{\alpha}{1-\alpha}\big(t -
s\big)^{\alpha} \bigg]ds.$$ Notice that, when the function $u$ is constant, we get zero.
\[Atangana-Baleanu\_Integral:definition\] The Atangana-Baleanu fractional integral of order $\alpha$ with base point $a$ is defined as $$\label{Atangana-Baleanu_Integral:equation}
^{\textsc{AB}}{\mathcal{I}}_t^{\alpha} u(t) = \frac{1-\alpha}{B(\alpha)}u(t) + \frac{\alpha}{B(\alpha) \Gamma(\alpha)} \int_{a}^{t} u(s) (t-s)^{\alpha-1} ds.$$
Numerical approach of the Atangana-Baleanu integral with fractional order {#Sec:3}
=========================================================================
In this section, we will start by given the discretization for Riemmann-Liouville fractional integral [@Hao; @Khalil]. Next, following the same idea as [@Zhuang; @Gallegos], we introduce a numerical scheme to discretize the temporal fractional integral, and give the corresponding error analysis which will be used to give the numerical solution of the Atangana-Baleanu fractional integral and also to obtain the solution of the modified groundwater flow within a leaky aquifer.
Let $f \in \mathcal{C}^{2}(a,b)$. Then $\int_{a}^{b}f(x)dx$ could be discretized as follows as $$\begin{gathered}
\label{descrete1}
\int_{a_{0} = a}^{a_{n}=b}f(x)dx = \sum_{j = 0}^{n} \int_{t_{j}}^{t_{j+1}}\frac{f(t_{j+1}) + f(t_{j})}{2}dy
= \sum_{j = 0}^{n} \frac{f(t_{j+1}) + f(t_{j})}{2}\int_{t_{j}}^{t_{j+1}}dy\\
= \sum_{j = 0}^{n} \frac{f(t_{j+1}) + f(t_{j})}{2}\big(t_{j+1} - t_{j}\big).\end{gathered}$$ On the other hand $$\label{descrete2}
\int_{a}^{b}f(x)dx = \sum_{j=0}^{n}\int_{t_{j}}^{t_{j+1}} f(t_{j+1})dy = \sum_{j=0}^{n}f(t_{j+1})\big(t_{j+1}-t_{j}\big).$$ With the Riemann-Liouville fractional integral we have the following.
We choose $t\in [0,T]$, the fractional order is denoted by $\alpha \in (0,1)$, the step $\tau = \frac{T}{n}$, $(n\in \mathbb{N})$, and the grid points are $t_{k}$, $k\in \{0,1,2,\dots, n\}$, where $t_{k} = k\tau$. Thus for $k\in \{0,1,2,\dots, n\}$, we have $$\label{descrete_Riemman}
_{0}^{\textsc{RL}}{\mathcal{I}}_t^{\alpha}f(t_{k}) = \frac{1}{\Gamma(\alpha)} \sum_{i=0}^{k-1}\int_{t_{i}}^{t_{i+1}}f(y)\big(t_{k} - y \big)^{\alpha-1}dy.$$ If we consider the following discretization for the function $f$. $$\label{descrete3}
f(x) = \frac{f(x_{i+1}) + f(x_{i})}{2},$$ by replacing into we obtain $$\begin{gathered}
\label{descrete4}
_{0}^{}{\mathcal{I}}_t^{\alpha} = \frac{1}{\Gamma(\alpha)} \sum_{j=0}^{k-1}\int_{t_{j}}^{t_{j+1}}\frac{f(t_{j+1}) + f(t_{j})}{2}\big(t_{k} - s\big)^{\alpha-1}ds, \\
= \frac{\tau^{\alpha}}{\Gamma(\alpha+1)}\sum_{j=0}^{k-1}\frac{f(t_{j+1}) + f(t_{j})}{2}\bigg[(k - j)^{\alpha} - (k-j+1)^{\alpha} \bigg] + R_{k, \alpha}.\end{gathered}$$ Using the same approach as in [@Wang], the error is given as $$\begin{gathered}
\label{error}
R_{k,\alpha} = \frac{1}{\Gamma(\alpha)}\sum_{j=0}^{k-1}\int_{t_{j}}^{t_{j+1}}\frac{f(y) - f(t_{j+1}) - f(t_{j})}{(t_{k}-y)^{1-\alpha}}
= \frac{1}{\Gamma(\alpha)}\sum_{j=0}^{k-1}\frac{f(y) - \frac{f(t_{j+1}) - f(t_{j})(t_{j+1}-t_{j})}{(t_{j+1}-t_{j})}}{(t_{k}-y)^{1-\alpha}} \\
= \frac{1}{\Gamma(\alpha)}\sum_{j=0}^{k-1}\frac{(t_{j+1}-t_{j})}{(t_{k}-y)^{1-\alpha}}\big( f(y) - f'(\xi) \big)dy, \quad t_{j} < \xi < t_{j+1} .\end{gathered}$$ From the Taylor series at the point $\xi$ $$f(y) = f(\xi) + yf'(\xi) + \cdots$$ we can approximate $$f(y) -f'(\xi) \approx f(\xi) + yf'(\xi) - f'(\xi) =f(\xi) - f'(\xi)(y-1).$$ Now taking the norm, we get $${\left\lVertf(y) - f'(\xi)\right\rVert} \approx {\left\lVertf(\xi) - f'(\xi)(y-1)\right\rVert} \leq M,$$ since the function $f$ is differentiable.Hence $${\left\lVertR_{k,\alpha}\right\rVert} = \frac{\tau}{\Gamma(\alpha+1)}M t_{k}^{\alpha}.$$
Let $f \in \mathcal{C}^{2}[0,T]$. The Atangana-Baleanu fractional integral is given by $$\begin{aligned}
\label{descrete5}
_{0}^{\textsc{AB}}{\mathcal{I}}_t^{\alpha}\big(f(t_{k})\big) = \frac{1-\alpha}{B(\alpha)}f(t_{k}) + \frac{\alpha \tau^{\alpha}}{B(\alpha)\Gamma(\alpha + 1)}
\sum_{j=0}^{k-1}b_{j}^{\alpha}\frac{f(t_{k-j}) + f(t_{k-j+1})}{2} + R_{k,\alpha}, \end{aligned}$$ where $\vert R_{k,\alpha} \vert \leq K t_{k}^{\alpha}\tau$, $k = 1,2,3,\dots, n$, and $b_{j}^{\alpha} = (j+1)^{\alpha} - j^{\alpha}$, $j = 0,1,2,\dots, n$.
Using the second approach of the discretization , we obtain the following approximation $$\label{descrete6}
_{0}^{\textsc{AB}}{\mathcal{I}}_t^{\alpha}\big(f(t_{k})\big) = \frac{1-\alpha}{B(\alpha)}f(t_{k}) + \frac{\alpha \tau^{\alpha}}{B(\alpha)\Gamma(\alpha + 1)}
\sum_{j=0}^{k-1}y(t_{j+1})\big[(k-j)^{\alpha} - (k-j-1)^{\alpha} \big] + \tilde{R}_{k,\alpha},$$ where $$\tilde{R}_{k,\alpha} = \sum_{j = 0}^{k-1}\int_{t_{j}}^{t_{j+1}}\frac{f(y) + f(t_{j+1})}{(t_{k}-y)^{1-\alpha}}dy,$$ and $$\vert \tilde{R}_{k,\alpha} \vert \leq \frac{\tau}{\Gamma(\alpha)}t_{k}^{\alpha}\max_{0 \leq t \leq t_{k}} \vert f'(t) \vert.$$ If the function $f$ is differentiable such that its Atangana-Baleanu fractional integral exists then $$\frac{d}{dt} ~ _{0}^{\textsc{AB}}{\mathcal{I}}_t^{\alpha}\big(f(t)\big) = \frac{1-\alpha}{B(\alpha)}f'(t) + \frac{\alpha}{B(\alpha)\Gamma(\alpha)}\frac{d}{dt}
\int_{0}^{t}f(y)(t-y)^{\alpha-1}dy.$$ Then for $k=0,1,2,3,\dots, n$ the Atangana-Baleanu integral can be decomposed as follows $$\begin{gathered}
_{0}^{\textsc{AB}}{\mathcal{I}}_t^{\alpha}\big(f(t)\big) \\ = \frac{1-\alpha}{B(\alpha)} \big(f(t_{k} - f(t_{k-1}) \big) + \Psi(\alpha) \int_{0}^{t_{k}}
\frac{f(s)}{(t_{k} - y)^{1-\alpha}}ds
- \Psi(\alpha)\int_{0}^{t_{k-1}}\frac{f(y)}{(t_{k-1}-y)^{\alpha-1}}dy \\
= \frac{1-\alpha}{B(\alpha)} \big(f(t_{k} - f(t_{k-1}) \big) + \Psi(\alpha) \int_{0}^{\tau}\frac{f(y)}{(t_{k} - y)^{1-\alpha}}dy
+ \Psi(\alpha)\int_{0}^{t_{k-1}}\frac{f(y+\tau) - f(y)}{(t_{k-1} - y)^{1-\alpha}}dy \\
= \Psi(\alpha) \int_{0}^{\tau}\frac{f(y)}{(t_{k} - y)^{1-\alpha}}dy + ~ _{0}^{\textsc{AB}}{\mathcal{I}}_t^{\alpha}\big(\beta(t_{k-1})\big), \end{gathered}$$ where $\Psi(\alpha)=\frac{\alpha}{B(\alpha)\Gamma(\alpha)}$. Nevertheless $$\begin{aligned}
\int_{0}^{\tau}\frac{f(y)}{(t_{k} - y)^{1-\alpha}}dy = \int_{0}^{\tau}\frac{f(y)}{(t_{k} - y)^{1-\alpha}}dy + \int_{0}^{\tau}\frac{f(y) -
f(\tau)}{(t_{k} - y)^{1-\alpha}}ds, \end{aligned}$$ where also $$\begin{aligned}
\int_{0}^{\tau}\frac{f(y) - f(\tau)}{(t_{k} - y)^{1-\alpha}}ds = \int_{0}^{\tau}\frac{f'(\xi)(y-\tau)}{(t_{k} - y)^{1-\alpha}}ds, \qquad y < \xi < \tau.\end{aligned}$$ Let us now assume that $f$ is two times differentiable on $[0,T]$. Then, we obtain $$\vert \Psi(\alpha)\int_{0}^{\tau}\frac{f(y) - f(\tau)}{(t_{k} - y)^{1-\alpha}}dy \vert \leq \frac{\tau^{1+\alpha}}{\frac{\alpha}{B(\alpha)
\Gamma(\alpha)}}b^{\alpha}_{k-1} \max_{0\leq t \leq \tau} \vert f'(t) \vert.$$ With the above relation in hand, we can conclude that $$\begin{aligned}
_{0}^{\textsc{AB}}{\mathcal{I}}_t^{\alpha}\big(\beta(t_{k-1})\big) &= \frac{1-\alpha}{B(\alpha)} \beta(t_{k-1}) + \frac{\tau^{\alpha}}{B(\alpha)\Gamma(\alpha)}
\sum_{j=1}^{k-1}b_{j-1}^{\alpha}\beta(t_{k-j}) + R'_{k,\alpha}.\end{aligned}$$ Here we have $$\vert R'_{k,\alpha} \vert \leq \frac{\tau^{\alpha}}{B(\alpha)\Gamma(\alpha)} t_{k-1}^{\alpha} \max_{0 \leq t \leq t_{k}} \vert f''(\xi) \vert, \qquad 0
\leq \xi < t_{k}.$$ Therefore, we have the following relationship $$_{0}^{\textsc{AB}}{\mathcal{I}}_\Delta t^{\alpha}~f(t_{k-1}) = \frac{1-\alpha}{B(\alpha)} \beta(t_{k-1}) + \frac{\tau^{\alpha}}{B(\alpha)\Gamma(\alpha)} \bigg\{
b_{k-1}^{\alpha}f(\tau)
+ \sum_{j=1}^{k-1}b_{j-1}^{\alpha}\big[f(t_{k-j + 1}) - f(t_{k-j}) \big] \bigg\}+ R^{2}_{k,\alpha},$$ where $\vert R^{2}_{k,\alpha} \vert < c_{\alpha}b_{k-1}^{\alpha}\tau^{\alpha+1} + c_{\alpha}\tau^{2} t^{\alpha}_{k-1}$
Finally the equation above can be reformulated as $$_{0}^{\textsc{AB}}{\mathcal{I}}_\Delta t^{\alpha}~f(t_{k-1}) = \frac{1-\alpha}{B(\alpha)}\frac{(f(t_{k} - t_{k-1})}{\tau} + \frac{\tau^{\alpha}}{B(\alpha)\Gamma(
\alpha)}\big\{ f(t_{k})
+ \sum_{j=1}^{k-1}b_{j}^{k}- \big(b_{j-1}^{\alpha} \big)f(t_{k-j}) \big\} + R^{2}_{k,\alpha},$$ where indeed $\vert R^{2}_{k,\alpha} \vert \leq c_{\alpha}b_{k-1}^{\alpha}\tau^{\alpha+1}$.
Application to groundwater flow within Leaky aquifer based upon Atangana-Baleanu fractional derivative {#Sec:4}
======================================================================================================
The concept of groundwater flow within the geological formation is a very complex physical problem and has attracted the attention of several scholars from different branches of sciences and technology. In particular the model portraying the movement of this subsurface water within the medium called leaky aquifer. In this section, we focus our attention on the model based on differentiation with non-local and non-singular kernel. To do this we consider the time derivative to be the time fractional derivative based on the Mittag-Leffler function. The new model will be analysed analytically and numerically. For analytical investigation, we shall focus on the analysis of existence and uniqueness of the solution of the new model. Then we apply the new numerical scheme to derive the numerical solution of the new model.
Analytical solution of the flow within the leaky aquifer based upon Atangana-Baleanu fractional derivative
----------------------------------------------------------------------------------------------------------
Let $\Omega = (a,b)$ be an open and bounded subset of $\mathbb{R}^{n}~(n \geq 1)$, with boundary $\partial \Omega$. For a given $\alpha \in (0,1)$, and a function $\varphi(r,t) \in H^{1}(\Omega) \times [0,T]$, which represents the head, we seek $\varphi$ such that the flow of water within the leaky aquifer is governed by $$\label{problem_Atangana-Baleanu}
\beta^{2}~{}^{ABC}{\mathcal{D}}_{t}^{\alpha}~\varphi = \partial_{rr}\varphi + \frac{1}{r} \partial_{r}\varphi - \frac{\varphi}{\varpi^{2}},$$ where $\varpi = KDc$ and $\beta^{2} = Sc\varpi^{-2}$, $S$ denotes the coefficient of storage, $K$ denotes the conductivity.
The problem of groundwater flow within the leaky aquifer will be approached analytically and solved numerically using the implicit scheme.
Existence and uniqueness of the solution of the problem
-------------------------------------------------------
In the following, we discuss the existence and uniqueness of solutions of the direct problem. $$\label{direct_Atangana_Baleanu_leaky}
\begin{cases}
~{}^{ABC}{\mathcal{D}}_{t}^{\alpha}~\varphi(t,r) = g(t,\varphi(t,r)),\\
\varphi(r_{c}, 0) = h(r) \qquad \qquad \text{on} \qquad [0,T]\times \partial \Omega,\\
\varphi(r_{c}, t) = \varphi_{c}\qquad \qquad \text{in}~~ \{0\}~\times
~\Omega,\\
\end{cases}$$ for the initial datum $\varphi_{c} \in L^{\infty}(\Omega), \varphi_{c} > 0$, and where $$\label{function_g}
g\big(r,t, \varphi(r,t)\big) = \beta^{-2} \bigg( \partial_{rr}\varphi(t,r) + \frac{1}{r} \partial_{r}\varphi(t,r) - \frac{\varphi(t,r)}{\varpi^{2}} \bigg).$$
Given $\varphi_{c} \in L^{\infty}(\Omega),~ \varphi_{c} > 0$, a solution of is a positive function $\varphi \in H^{1}(\Omega) \times [0,T]$ such that if we apply the fractional integral defined in to it yields $$\label{solution_1}
\varphi(t,r) = \varphi(0) + \frac{1-\alpha}{B(\alpha)}g(t,\varphi(t,r)) + \frac{\alpha}{B(\alpha) \Gamma(\alpha)} \int_{0}^{t} g(s,\varphi(s,r)) (t-s)^{\alpha-1} ds,$$ for all $t \in [0,T]$
We want to use the contraction mapping theorem, so for this purpose we need to build a closed set $\mathcal{\varepsilon}$ of $H^{1}(\Omega) \times [0,T]$ such that the nonlinear operator $g$ be a contraction which maps $\mathcal{\varepsilon}$ into itself.
So we first show that $g$ is a contraction mapping.
The nonlinear operator $g \in H^{1}(\Omega) \times [0,T]$ is locally Lipschitz.
We consider two bounded functions $\varphi$ and $\vartheta$ in $\in H^{1}(\Omega) \times [0,T]$. Then, $$\begin{aligned}
{\left\lVertg(\varphi) - g(\vartheta)\right\rVert}_{H^{1}} = \beta^{-2} {\left\lVert \partial_{rr}\varphi - \partial_{rr}\vartheta + r^{-1} \big( \partial_{r}\varphi- \partial_{r}\vartheta \big) - \varpi^{-2}\big( \varphi- \vartheta \big)\right\rVert}.\end{aligned}$$ We have by the triangular inequality $$\begin{aligned}
{\left\lVertg(\varphi) - g(\vartheta)\right\rVert}_{H^{1}} \leq \beta^{-2} \bigg \{ {\left\lVert \partial_{rr}\varphi - \partial_{rr}\vartheta \right\rVert} + {\left\lVert r^{-1} \big( \partial_{r}\varphi- \partial_{r}\vartheta \big)\right\rVert} - \varpi^{-2}{\left\lVert \varphi- \vartheta \big)\right\rVert} \bigg \}.\end{aligned}$$ As the derivative operator satisfy the Lipschitz conditions in $H^{1}$, hence there exist two positive constants $c_{1}$ and $c_{2}$ such that $$\begin{gathered}
{\left\lVertg(\varphi) - g(\vartheta)\right\rVert}_{H^{1}} \leq \beta^{-2}c_{1} {\left\lVert \varphi - \vartheta \right\rVert} + \beta^{-2}c_{2} {\left\lVertr^{-1}\right\rVert}{\left\lVert\varphi- \vartheta\right\rVert} - \beta^{-2}\varpi^{-2}{\left\lVert \varphi- \vartheta \right\rVert},\\
\leq C {\left\lVert\varphi- \vartheta \right\rVert},\end{gathered}$$ where $C = |\beta^{-2} (c_{1} + c_{2}{\left\lVertr^{-1}\right\rVert} - \varpi^{-2})| < 1$. Then $g$ is a contraction mapping.
\[theo1\] For a given initial datum $\varphi_{c} \in L^{\infty}(\Omega),~ \varphi_{c} > 0$, there exists an unique positive solution $\varphi$ of on $ H^{1}(\Omega) \times [0,T]$, for all $t < T < \infty$, and $$\lim_{t \to T} {\left\lVert\varphi(t,r)\right\rVert}_{L^{\infty}} \to \infty.$$
We shall follow the idea of [@Song]. Since the nonlinear operator $g$ is locally Lipschitz, for $\bar{\varphi_{c}} = {\left\lVert\varphi_{c}\right\rVert}_{H^{1}}$ there exists $C_{\bar{\varphi_{c}}}$ such that $0 < C_{\bar{\varphi_{c}}} < \infty$ and $${\left\lVertg(\varphi) - g(\vartheta)\right\rVert}_{H^{1}} \leq C_{\bar{\varphi_{c}}} {\left\lVert\varphi- \vartheta \right\rVert}_{H^{1}}.$$ Let $T_{1}>0$ be a constant such that $T_{1} < \frac{1}{C_{\bar{\varphi_{c}}}}$.
Set $$\varepsilon = \big\{ \varphi \in H^{1}(\Omega) \times [0,T];~~{\left\lVert\varphi\right\rVert}_{H^{1}} \leq \bar{\varphi_{c}}, \text{for all}~ t \in [0,T_{1}] \big\},$$ endowed with the norm $${\left\lVert\varphi\right\rVert}_{\varepsilon} = \sup_{0 \leq t \leq T_{1}}{\left\lVert\varphi\right\rVert}_{H^{1}},$$ then $\varepsilon$ is a closed convex subset of $H^{1}(\Omega) \times [0,T]$.
Now we consider the following associated problem of defined on $\varepsilon$, $$\label{solution_2}
\psi(\varphi) = \varphi_{c} + \frac{1-\alpha}{B(\alpha)}g(t,\varphi(t,r)) + \frac{\alpha}{B(\alpha) \Gamma(\alpha)} \int_{0}^{t} g(s,\varphi(s,r)) (t-s)^{\alpha-1} ds.$$ For all $\varphi \in \varepsilon$ and $t \geq 0$ we have $$\begin{aligned}
& {\left\lVert\psi(\varphi)\right\rVert}_{\varepsilon} = \sup_{0 \leq t \leq T_{1}} {\left\lVert \varphi_{c} + \frac{1-\alpha}{B(\alpha)}g(t,\varphi(t,r)) + \Psi(\alpha) \int_{0}^{t} g(s,\varphi(s,r)) (t-s)^{\alpha-1} ds\right\rVert}_{H^{1}}\\
&\leq {\left\lVert\varphi_{c}\right\rVert}_{H^{1}} + {\left\lVert \frac{1-\alpha}{B(\alpha)}g(t,\varphi(t,r))\right\rVert}_{H^{1}} + \frac{\alpha}{B(\alpha) \Gamma(\alpha)} \int_{0}^{t} {\left\lVert g(s,\varphi(s,r))\right\rVert}_{H^{1}}ds\\
&\leq {\left\lVert\varphi_{c}\right\rVert}_{H^{1}} + \int_{0}^{t} {\left\lVert g(s,\varphi(s,r))\right\rVert}_{H^{1}}ds.\end{aligned}$$ But since $g$ is locally Lipschitz for all $s \in [0,T_{1}]$, it follows that $$\begin{aligned}
& {\left\lVert\psi(\varphi)\right\rVert}_{\varepsilon} \leq {\left\lVert\varphi_{c}\right\rVert}_{H^{1}} + \int_{0}^{t} \big(C_{\bar{\varphi_{c}}} {\left\lVert\varphi(s,r)\right\rVert}_{H^{1}} + c_{3} \big)~ds.\\\end{aligned}$$ Thus for all $\varphi_{1}$, $\varphi_{2} \in \varepsilon $ $$\begin{aligned}
& {\left\lVert\psi(\varphi_{1}) - \psi(\varphi_{2})\right\rVert}_{\varepsilon} = \sup_{0 \leq t \leq T_{1}}T_{1}C_{\bar{\varphi_{c}}} {\left\lVert\varphi_{1} - \varphi_{2}\right\rVert}_{\varepsilon}.\\\end{aligned}$$
This shows that $\psi$ is a contraction mapping in $\varepsilon$. Thus $\psi$ has a fixed point which is a solution to .
Now let us show that the problem has an unique solution.
Let $\varphi_{1}$, $\varphi_{2} \in H^{1}$ be two solutions of and let $\varphi = \varphi_{1} - \varphi_{2}$. Then $$\varphi = \frac{1-\alpha}{B(\alpha)} \bigg( g(t,\varphi_{1}(t,r))- g(t,\varphi_{2}(t,r)) \bigg) \\ + \frac{\alpha}{B(\alpha) \Gamma(\alpha)} \int_{0}^{t} \bigg( g(s,\varphi_{1}(s,r))- g(s,\varphi_{2}(s,r)) \bigg)ds.$$ By thanking the norm on both sides, $$\begin{gathered}
{\left\lVert\varphi\right\rVert} \leq \frac{1-\alpha}{B(\alpha)} {\left\lVert g(t,\varphi_{1}(t,r))- g(t,\varphi_{2}(t,r)) \right\rVert} + \frac{\alpha}{B(\alpha) \Gamma(\alpha)} \int_{0}^{t} {\left\lVert g(s,\varphi_{1}(s,r))- g(s,\varphi_{2}(s,r)) \right\rVert}ds\\
\leq T_{1}C_{\bar{\varphi_{0}}} \int_{0}^{t} {\left\lVert\varphi_{1}(s,r)\right\rVert}_{H^{1}}~ds.\end{gathered}$$ By the Gronwall inequality [@Song], the result follows.
Exact solution of the flow within the leaky aquifer based upon Atangana-Baleanu fractional derivative
-----------------------------------------------------------------------------------------------------
In the model here discussed for water flow in the leaky aquifer, the head $\varphi(r,t)$, which appears in , is assumed to be governed by the one-dimensional time-fractional differential equation involving the Atangana-Baleanu fractional derivative.
Applying Laplace transform to , the fundamental solution $ \tilde{\varphi(r,p)}$, $p \in \mathbb{N}$ results to be: $$-\beta^{2} \mathcal{L}_{t} \bigg[ ~{}^{ABC}{\mathcal{D}}_{t}^{\alpha}~\varphi \bigg](p) + \mathcal{L}_{t} \bigg[ \partial_{rr}\varphi \bigg](p) +
\mathcal{L}_{t} \bigg[ \frac{1}{r} \partial_{r}\varphi\bigg](p) - \varpi^{-2} \mathcal{L}_{t} \big[\varphi \big](p) = 0,$$ where $\mathcal{L}_{t} := \tilde{\varphi}$ denotes the Laplace transform. Replacing each term by its value, we get $$-\frac{B(\alpha)\beta^{2}\big[p^{\alpha}\varphi - p^{\alpha-1}\varphi_{c} \big]}{(1-\alpha)p^{\alpha} + \alpha} +
\partial_{rr} \tilde{\varphi} + \frac{1}{r}\partial_{r}\tilde{\varphi} - \varpi^{-2} \tilde{\varphi}
= 0.$$ Hence the following differential equation in the form holds $$\label{diff_bessel_laplace}
r^{2} \partial_{rr} \tilde{\varphi} + r~\partial_{r}\tilde{\varphi} - qr^{2}\tilde{\varphi} = 0,$$ with $\varphi_{c} \approx 0$, where $$q = \frac{B(\alpha)\beta^{2}p^{\alpha}}{(1-\alpha)p^{\alpha} + \alpha} + \varpi^{-2}.$$ Since $q$ is positive, the exact solution of the differential equation is given in terms of Bessel function of the first kind, $J_0$ and modified kind, $K_0$ as $$\label{solution_laplace1}
\tilde{\varphi}(r,p) = AJ_{0}(r\sqrt{q}) + BK_0(r\sqrt{q}),$$ where $A$ and $B$ are the constants and $J_0$, $K_0$ respectively given as $$J_{\nu}(z) = \sum_{k=0}^{\infty}\frac{(-1)^k}{k!\Gamma(k+\nu+1)}\left( \frac{z}{2}\right)^{2k+\nu} \text{ and } K_{\nu}(z) =
\frac{\pi}{2}(-i)^{\nu}\left(\frac{-J_{\nu}(z)+J_{-\nu}(z)}{\sin(\nu \pi)}\right).$$
Using the boundary condition in , we obtain $B=0$, then the solution is reduced to $$\label{solution_laplace2}
\tilde{\varphi}(r,p) = AJ_{0}(r\sqrt{q}) = A\sum_{k=0}^{\infty}\frac{(-1)^k}{k!\Gamma(k+\nu+1)}\left( \frac{r\sqrt{q}}{2}\right)^{2k+\nu}.$$ Due to the difficulties to obtain the inverse Laplace transform of , we therefore propose to obtain the approximate solution of by using the proposed numerical approximation of the Atangana-Baleanu integral.
Numerical analysis of the ground water flow within the leaky aquifer based upon Atangana-Baleanu fractional derivative
----------------------------------------------------------------------------------------------------------------------
To achieve this, we revert the fractional differential equation to the fractional integral equation using the link between the Atangana-Baleanu derivative and the Atangana-Baleanu integral to obtain $$\varphi(r,t) - \varphi(r_{c},0) = ~~_{0}^{AB}{\mathcal{I}}_{t}^{\alpha}~g\big(r,t, \varphi(r,t)\big).$$ For some positive and large integers $M = N = 350$, the grid sizes in space and time is denoting respectively by $\xi = 1/M$ and $\tau = 1/N$. The grid points in the space interval $(0,1]$ are the numbers $r_{i} = i\xi$, $i = 1,2,\dots,M$ and the grid points in the time interval $[0,1]$ are the numbers $t_{k} = k \tau$, $k = 0,1,2,\dots,N$. The value of the function $\varphi$ at the grid points are denoted by $\varphi_{i}^{k} =
\varphi(r_{i},t_{k})$.
Using the implicit finite differences method, a discrete approximation of $g\big(r, t, \varphi(r,t) \big)$ given by can be obtained as follows $$\label{descrite_function_g}
g \big(r_{i}, t_{k}, \varphi(r_{i},t_{k})\big) = \frac{\beta^{-2}}{\xi^2} \bigg\{ \big(1 + \frac{1}{2i}\big)\varphi_{i+1}^{k} - \big(2 +
\frac{\xi^{2}}{\varpi^{2}} \big) \varphi_{i}^{k} + \big(1 - \frac{1}{2i}\big)\varphi_{i-1}^{k} \bigg\}.$$ In order to use the numerical approximation proposed in this work, discrete solution of is then given as follows$$\label{solutiona}
\varphi_{i}^{k+1} - \varphi_{c}^{0} = ~{}^{AB}{\mathcal{I}}_{t}^{\alpha}~g(r_{i},t_{k}, \varphi(r_{i},t_{k+1})).$$
$$\begin{gathered}
_{0}^{\textsc{AB}}{\mathcal{I}}_{t}^{\alpha}~g = \frac{1-\alpha}{B(\alpha)}g \big(r_{i}, t_{k+1}, \varphi(r_{i},t_{k+1})\big) \\
+ \Psi(\alpha)
\sum_{j=0}^{k}\frac{b_{j}^{\alpha}}{2} \bigg(
g\big(r_{i},t_{k-j}, \varphi_{i}(r_{i},t_{k-j})\big) + g\big(r_{i},t_{k-j+1}, \varphi(r_{i},t_{k-j+1}\big)\bigg).\end{gathered}$$
Therefore the numerical approximation can be given as follows $$\begin{gathered}
\label{solutionb}
\varphi_{i}^{k+1} - \varphi_{c}^{0} = \frac{(1-\alpha)}{B(\alpha)}\frac{\beta^{-2}}{\xi^2} \bigg\{ \big(1 + \frac{1}{2i}\big)\varphi_{i+1}^{k+1} -
\big(2 + \frac{\xi^{2}}{\varpi^{2}} \big) \varphi_{i}^{k+1} + \big(1 - \frac{1}{2i}\big)\varphi_{i-1}^{k+1} \bigg\} \\
+ \Psi(\alpha)\frac{\beta^{-2}}{\xi^2}
\sum_{j=0}^{k}\frac{b_{j}^{\alpha}}{2} \bigg\{
\big(1 + \frac{1}{2i}\big)\varphi_{i+1}^{k-j} - \big(2 + \frac{\xi^{2}}{\varpi^{2}} \big) \varphi_{i}^{k-j} + \big(1 -
\frac{1}{2i}\big)\varphi_{i-1}^{k-j} \\
+ \big(1 + \frac{1}{2i}\big)\varphi_{i+1}^{k-j+1} - \big(2 + \frac{\xi^{2}}{\varpi^{2}} \big) \varphi_{i}^{k-j+1} + \big(1 -
\frac{1}{2i}\big)\varphi_{i-1}^{k-j+1} \bigg\} .\end{gathered}$$
In order obtain the plots of the solution given by , we shall consider $350$ equidistant nodes in $[0,1]$. For the value $\alpha= 1/2$, the approximate solution found by using the numerical method of the Atangana-Baleanu integral proposed is obtained. Figure \[fig:fig\_a\] shows the approximate solution for the different time step $k = 0$, $k = 20$ and $k = 50$ respectively.
![Numerical solution of for the specific value of the parameters $\alpha=1/2$, and different time step $k = 0$, $k = 20$ and $k =
50$ respectively.[]{data-label="fig:fig_a"}](n-350-m-350-alpha-1-2-k0.eps "fig:"){width="30.00000%"} $\quad$ ![Numerical solution of for the specific value of the parameters $\alpha=1/2$, and different time step $k = 0$, $k = 20$ and $k =
50$ respectively.[]{data-label="fig:fig_a"}](n-350-m-350-alpha-1-2-k20.eps "fig:"){width="30.00000%"} $\quad$ ![Numerical solution of for the specific value of the parameters $\alpha=1/2$, and different time step $k = 0$, $k = 20$ and $k =
50$ respectively.[]{data-label="fig:fig_a"}](n-350-m-350-alpha-1-2-k50.eps "fig:"){width="30.00000%"}
Moreover, to have an overview of the variation of flow or the behaviour of the function $\varphi$ in a finite time in terms of the parameter $\alpha$, we consider two different time steps $k = 0$ and $k = 50$. For this purpose, Figure gives us the approximate result using the method proposed. We would like to notice that for larger number of nodes in $[0,1]$ the better approximated solution is obtained in the whole interval.
![Approximate solution of for initial time step $k=0$ and $\alpha = 1/2$—blue— and $\alpha = 9/10$ —orange— (graph in the left). Approximate solution of for time step $k=50$ and $\alpha = 1/2$—blue— and $\alpha = 9/10$ —orange— (graph in the right)[]{data-label="fig:fig_b"}](n-350-m-350-alpha-1-2-k0-k50.eps "fig:"){width="35.00000%"} $\quad$ ![Approximate solution of for initial time step $k=0$ and $\alpha = 1/2$—blue— and $\alpha = 9/10$ —orange— (graph in the left). Approximate solution of for time step $k=50$ and $\alpha = 1/2$—blue— and $\alpha = 9/10$ —orange— (graph in the right)[]{data-label="fig:fig_b"}](n-350-m-350-alpha-9-10-k0-k50.eps "fig:"){width="35.00000%"}
Next in Figure \[fig:fig\_c\] we show the approximate solution of by using the numerical approximation of the Atangana-Baleanu method proposed for different time steps $k = 0$ and $k = 50$ for $\alpha = 1/2$ and $\alpha = 9/10$ in $[0,1]$.
![Approximate solution of for $\alpha = 1/2$, $k = 0$ —blue— and $k = 50$—orange— (graph in the left). Approximate solution of for $\alpha = 9/10$, $k = 0$ —blue— and $k = 50$—orange— (graph in the right).[]{data-label="fig:fig_c"}](n-350-m-350-alpha-1-2-alpha-9-10-k0.eps "fig:"){width="35.00000%"} $\quad$ ![Approximate solution of for $\alpha = 1/2$, $k = 0$ —blue— and $k = 50$—orange— (graph in the left). Approximate solution of for $\alpha = 9/10$, $k = 0$ —blue— and $k = 50$—orange— (graph in the right).[]{data-label="fig:fig_c"}](n-350-m-350-alpha-1-2-alpha-9-10-k50.eps "fig:"){width="35.00000%"}
Conclusions {#Sec:5}
===========
The new scheme of the fractional integral in the sense Atangana-Baleanu has been proposed in this work. The error analysis of the novel scheme was successfully presented and the error obtained shows that the scheme is highly accurate. A new model of groundwater flowing within a leaky aquifer was suggested using the concept of fractional differentiation based on the generalised Mittag-Leffler function in order to fully introduce into mathematical formulation the complexities of the physical problem as the flow taking place in a very heterogeneous medium. The Mittag-Leffler operator provide more natural observed fact than the more used power law. The new model was analysed, as the uniqueness and the existence of the solution was investigated with care. To further access the accuracy of the proposed numerical scheme, we solved the new model numerically using this suggested scheme. Some simulations have been presented for different values of fractional order. We strongly believe that this numerical scheme will be applied in many fields of science, technology and engineering for those problems based on the new fractional calculus.
References {#references .unnumbered}
==========
[99]{} A. Atangana. *On the new fractional derivative and application to nonlinear Fisher’s reaction-diffusion equation.* Applied Mathematics and Computation, 1, 273, pp. 948–956, (2016).
I. Area, J. D. Djida, J. Losada, and Juan J. Nieto. *On Fractional Orthonormal Polynomials of a Discrete Variable*. Discrete Dynamics in Nature and Society, vol. 2015, Article ID 141325, 7 pages, 2015. doi:10.1155/2015/141325
A. Atangana, and J. J. Nieto. *Numerical solution for the model of RLC circuit via the fractional derivative without singular kernel*. Advances in Mechanical Engineering, vol.7, no.10, pp. 1–7, (2015).
A. Atangana and D. Baleanu. *New fractional derivatives with nonlocal and non-singular kernel: Theory and application to heat transfer model*. Thermal Science, 20(2), pp. 763–769, (2016).
M. Meerschaert. M and C. Tadjeran. *Finite difference approximations for fractional advection-dispersion flow equations*, Journal of Computational and Applied Mathematics 1, 172, pp. 65–77 (2004),
P. Zhuang, Liu, F, I. Turner and V. Anh. *Galerkin finite element method and error analysis for the fractional cable equation*, Numerical Algorithms 72(2), pp. 447–466, (2016).
Z. Hao, W. Cao, and G. Lin. *A second-order difference scheme for the time fractional substantial diffusion equation*, Journal of Computational and Applied Mathematics, 313, pp. 54–69, (2017).
I. Podlubny. Fractional differential equations. Academic Press, Inc., San Diego, CA, (1999).
R. Khalil, M. Al Horani, A. Yousef, and M. Sababheh. *A new definition of fractional derivative*, Journal of Computational and Applied Mathematics, 264, pp. 65–70 (2014).
P. Wang, C. Huang, and L. Zhao. *Point-wise error estimate of a conservative difference scheme for the fractional Schrödinger equation*, Journal of Computational and Applied Mathematics, 306, pp. 231–247, (2016).
J.A Gallegos, M.A Duarte-Mermoud. *Boundedness and convergence on fractional order systems*, Journal of Computational and Applied Mathematics, 296, pp. 815–826, (2016).
Song-Mu Zheng. *Nonlinear Evolution Equations*. Chapman and Hall/CRC monographs and surveys in pure and applied mathematics, 133, (2004).
O. I. Marichev. Handbook of Integral Transforms of Higher Transcendental Functions: Theory and Algorithmic Tables. Ellis Horwood & Halsted Press, New York, 1983.
[^1]: The first author is indebted to the AIMS-Cameroon 2015–2017 tutor fellowship. The work of I. Area has been partially supported by the Ministerio de Ciencia e Innovación of Spain under grant MTM2016–75140–P, co-financed by the European Community fund FEDER.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper deals with the routing protocols for distributed wireless sensor networks. The conventional protocols for WSNs like Low Energy adaptive Clustering Hierarchy (LEACH), Stable Election Protocol (SEP), Threshold Sensitive Energy Efficient Network (TEEN), Distributed Energy Efficient Clustering Protocol (DEEC) may not be optimal. We propose a scheme called Away Cluster Head (ACH) which effectively increases the efficiency of conventional clustering based protocols in terms of stability period and number of packets sent to base station (BS). We have implemented ACH scheme on LEACH, SEP, TEEN and DEEC. Simulation results show that LEACH-ACH, SEP-ACH, TEEN-ACH and DEEC-ACH performs better than LEACH, SEP, TEEN and DEEC respectively in terms of stability period and number of packets sent to BS. The stability period of the existing protocols prolongs by implementing ACH on them.'
author:
- |
N. Javaid$^{\ddag}$, M. Waseem$^{\ddag}$, Z. A. Khan$^{\$}$, U. Qasim$^{\pounds}$, K. Latif$^{\ddag}$, A. Javaid$^{\natural}$\
COMSATS Institute of IT, $^{\ddag}$Islamabad, $^{\natural}$Wah Cantt, Pakistan.\
$^{\$}$Faculty of Engineering, Dalhousie University, Halifax, Canada.\
$^{\pounds}$University of Alberta, Alberta, Canada.
title: 'ACH: Away Cluster Heads Scheme for Energy Efficient Clustering Protocols in WSNs'
---
Wireless sensor networks, Distributed networks, Clustering Protocol.
Background
==========
In Direct Transmission \[1\], each node in the sensor network communicates directly to BS. In the aforementioned protocol, farthest nodes die faster than the nearest nodes. In Minimum transmission energy \[2\] routing protocol each node transmits to its nearest node so the nearest nodes die at a faster rate because they receive data from the farther nodes. In the current body of research going in the field of WSNs clustering based protocols have attain significant attraction. In clustering based routing protocols the sensor nodes form clusters. In these clusters, one node is selected as CH. The nodes sense data and send to their respective CHs which aggregate and fuse the data, thus saving the energy as global communication is reduced due to local compression.
Once the CH receives data from its nodes it aggregates and fuses the data into a small set and sends to BS. Unbalanced energy consumption among the sensor nodes may cause network partition and node failures where transmission from some sensors to the sink node becomes blocked. Therefore, construction of a stable backbone is one of the challenges in sensor network applications.
LEACH \[3\] proposes a clustering based routing protocol for homogenous networks in which a node becomes CH by a probabilistic equation and forms a cluster of those nodes which receive strong signal to noise ratio from it. The nodes sense the environment and send data to CH where it is aggregated and finally send to BS. In LEACH there is a localized coordination amongst the nodes for cluster set up and locally compress the data to reduce global communication. CHs in LEACH are rotated randomly. Heterogeneous networks are more stable and beneficiary than homogenous networks. A number of protocols like SEP, DEEC and Threshold Distributed Energy-Efficient Clustering protocol (T-DEEC) have been proposed for WSNs. SEP \[4\] has two level of heterogeneity. In DEEC \[5\], CH selection is based on the ratio of residual energy and average energy of the network. The high energy nodes have more chances to become CH. In this way the energy is evenly distributed in the network. These routing protocols have some limitation due to their design and performance.
The ACH Scheme
==============
Optimal Number Of CHs
---------------------
The optimal probability of a node to take part in election for selection of CHs is a function of the spatial density when the nodes are uniformly distributed over the sensors’ field. When the total energy consumption is minimum and energy consumption is well distributed over all sensors, the clustering is then called optimal clustering. The energy model we use for our simulation effect the optimal number of CHs. We use similar energy model as proposed in LEACH, SEP and DEEC. We have been giving particular attention to distribution of CHs in network so as energy in the network. Once nodes are deployed in region of interest the nodes locally coordinate for cluster set up and operation. Each node decides whether to become a CH or not. The node generates a random number and compares it with the threshold value. If the number generated is less than or equal to the threshold value and the node has not been CH for the last $\frac{1}{p}$ round the node is marked as to be one of the CH. p is the probability of a node to become CH. In ACH scheme, once the CHs have been formed the CHs send a confirmation message to one another using CSMA-MAC protocol. The CH which receives a strong SNR from its adjacent CH will be marked as a normal node. For simplicity in our simulations we replace SNR by distance. We assume the CH will be made unmark and will become a normal node if its distance from the nearest CH is less than $12m$. The distance between CHs less than $12m$ is shown by “a” otherwise “b”. After confirmation of CHs, nodes receive an association message from CHs and respond according to the strength of SNR. The clusters are thus formed and the CHs are well distributed in the network. This makes clusters even in terms of number of nodes in each cluster. In this way energy of CH dissipated in each round is comparably equal. The distant CHs’ network is shown in fig. 2. We implement ACH scheme on LEACH, SEP, TEEN and DEEC. Simulation results show that ACH scheme performs better with all of 4 selected protocols. We initialize parameters for simulation, randomly deploy our nodes and start network’s operation. In network’s operation each node is checked whether eligible to become CH. We call this operation for epoch. If a node successfully pass through this test energy of the node is checked. Next comes the turn of CH formation. A node becomes CH if it has energy and is eligible to be CH which is shown in CH formation block. The CH then goes through neighbors’ association phase and data transmission phase as shown in fig. 1.
{height="7cm" width="9cm"}
{height="7cm" width="8cm"}
Simulations
===========
We have implemented our protocol on Matlab \[6\] to evaluate its performance with LEACH, SEP, TEEN and DEEC. We have proposed LEACH-ACH, SEP-ACH, TEEN-ACH and DEEC-ACH. Our goals in conducting the simulation are as follows:
- Compare the performance of LEACH, SEP, TEEN, DEEC and their ACH versions on the basis of longevity of the network.
- Compare throughput of LEACH, SEP, TEEN, DEEC and the respective ACH schemes.
We have performed our simulation on $100$ nodes and a fixed BS located in the center of the field. We randomly distributed $100$ nodes in a $100m×100m$ field. The most distant node from BS is at $70.7m$. The nodes have their horizontal and vertical coordinates located between $0$ and maximum value of the dimension which is $100$. All the nodes have different energies as the environment is heterogeneous. We simulate our protocol on the basis of initial energy as follows:
- The maximum energy of a node in the field is not more than 0.5J/bit.
The parameters used in our simulation are summarized in Table. 1.
Parameter Value
----------------------------------------- ------------
Simulation Area 100m 100m
Location of BS (50m, 50m)
Number Of Nodes 100
Initial Energy of Nodes (Maximum Value) 0.5 J/node
Packet Size 4000 bits
: Parameters used in our simulations
To analyze and compare the performance of our protocol with LEACH, SEP, TEEN and DEEC we have used two metrics. They are:
- Total number of dead nodes:
This metric show the overall lifetime of the network. It gives us an idea about the stability period and instability period. This metric is an indication of the number of dead nodes with time.
- Through put:
This metric is an indication of the rate of packets sent to BS.
Implementing ACH Scheme on LEACH, TEEN, DEEC, and SEP
-----------------------------------------------------
### LEACH-ACH
LEACH is a clustering based routing protocol for homogenous networks. In LEACH probability is same for all nodes to become CHs. The nodes compare the random value generated at each round with the threshold equation and become CH if the threshold value is less than the random number. The threshold formula is given by:
$
T(\emph{n})=\begin{array}{cc}
\{ &
\begin{array}{cc}
\frac{P}{1-P(rmod\frac{1}{P})} & n\in G \\
0 & {else}
\end{array}
\end{array}
$
We implement ACH scheme in LEACH and found its results better than LEACH as shown in fig. 3.
{height="6cm" width="8cm"}
### TEEN-ACH
TEEN is a routing protocol for reactive networks. In TEEN two thresholds: hard and soft have been introduced to reduce number of communications. After deployment of nodes CH set up phase starts in which CHs are formed. Once the CHs are confirmed the nodes sense environment and on their transmitter. When the the sense value reach hard threshold the transmitter is on and data is send to CH. This value is stored in an internal variable. Next time when the sense value reach the hard threshold, difference of stored value and sense value is obtained, if this value is greater or equal to soft threshold transmission is done otherwise transmitter are kept off. In TEEN-ACH we make CHs distant which makes energy dissipation even among CHs and thus very less energy is consumed in each round. Fig. 4 shows the behavior of TEEN against TEEN-ACH.
{height="6cm" width="8cm"}
### DEEC-ACH
DEEC-ACH is an extension of DEEC protocol which enhances stability period of DEEC. The CH selection criterion is based on DEEC protocol however we introduce ACH scheme which enhances the performance of DEEC.
DEEC is a heterogeneous routing protocol in which nodes have different initial energy as the network starts. DEEC uses initial and residual energy level of nodes to form CHs. Once sensor nodes are deployed in the region they locally coordinate for cluster set up and operation. Let $\emph{ni}$ denotes the number of rounds for which the node $\emph{Si}$ is CH often referred as the rotating epoch. $\emph{Popt}$ is our desired percentage of CHs and $\emph{ni}=1/\emph{popt}$ is the rotating epoch. By epoch we means that a node once becomes CH will not take part in CH formation for the next $1/\emph{popt}$ rounds. As in DEEC nodes have different energy levels the CH selection probability is different for each node and we call it average probability $\emph{pi}$. $\emph{pi}$ of nodes with more energy is greater. The average energy of network denoted by $\overline{E(r)}$ is given by eq.(1):
$$\overline{E}(r)=\frac{1}{N}\sum_{i=1}^{N}E_i(r)$$
The average probability of CHs per round per epoch is represented in eq.(2):
$$\emph{pi}=\emph{popt}\,\,[1-\frac{\overline{E}(r)-Ei(r)}{\overline{E}(r)}]$$
$$\sum_{i=1}^{N}pi=\sum_{i=1}^{N}popt\frac{Ei(r)}{{\overline{E}(r)}}=popt\sum_{i=1}^{N}\frac{Ei(r)}{{\overline{E}(r)}}=Npopt$$
Eq.(3) shows the optimal number of CHs we want to achieve. The probability of nodes in the network to become CHs is based on the ratio of their residual energy and average residual energy of the network. The probability equation for nodes to become CH is given by eq.(4):
$$pi=\frac{potp N (1+a) E_i(r)}{(N+\sum_{i=1}^{N}a_i)\overline{E(r)}}$$
Where $\emph{popt}$ is the desired percentage of CHs, $N$ is number of nodes, $\emph{Ei(r)}$ is residual energy of a node and $\overline{E}(r)$ is network’s average energy. $Pi$ is average probability of a node to become CH. Each node creates a random number for itself and compares it with threshold equation, if the number generated is less than or equal to the threshold value the node is selected as CH for that round. The threshold equation is given by:
$
T(\emph{Si})=\begin{array}{cc}
\{ &
\begin{array}{cc}
\frac{pi}{1-pi(rmod\frac{1}{pi})} & Si\in G \\
0 & {else}
\end{array}
\end{array}
$
Where $G$ represents the set of nodes eligible to take part in CH selection at round $r$. $\emph{Si}\,\,\,$$\epsilon$ $G$ consists of all those nodes which have not been CHs for the most recent $ni$ rounds. Once a node becomes CH it sends a confirmation message to all CHs. As soon as the CH is confirmed it sends an association message to all the nodes using CSMA-MAC protocol. The nodes respond according to strength of SNR received. The nodes associate themselves to that CH whose SNR is stronger. The CH then allocates TDMA slot to each node in the cluster. The nodes sense the environment and send data to their respective CHs in the TDMA slots allocated to them.
CH formation depends on random number generated, threshold value and energy of nodes. At some stages two or more very close (or intersecting) nodes become CHs and energy dissipation is even more and unbalance. We make CHs away in our protocol. Once a node becomes a CH it virtually take part in election of next CH. The CH decides area of next node taking part in election for CH. We assume that no node can become CH in (15m, 15m). A node which has become CH at (5m, 5m) will force the nodes to be normal nodes in the area ((5+10)m, (5+10)m) . In this way the CHs are made distant and we are able to achieve $Npopt$ CHs each round. Fig. 5 shows the comparison of DEEC with DEEC-ACH.
{height="6cm" width="8cm"}
### SEP-ACH
In this section we implement ACH scheme on SEP. In SEP we have two level heterogeneity. The normal nodes in SEP have $a$ times less energy than advance nodes. The probability of normal nodes in SEP differs from advance nodes as follows:
$$pnrm=\frac{p_{opt}}{1+a.m}$$
$$padv=\frac{p_{opt}(1+a)}{1+a.m}$$
where $pnrm$ in eq.(5) and $padv$ in eq.(6) is the probability equation for normal and advance nodes respectively. m is the fraction of advance nodes. Eq.(6) shows greater probability of advance nodes to become CHs. Each node in the network generates a random number for itself, compares itself with the threshold value and become CH if the number is less than or equal to threshold value. After the CH formation CH confirmation phase starts in which the CHs are made distant. We introduce ACH scheme in SEP which makes CHs formed in SEP distant. The energy in the nodes are thus conserved and the stability period of SEP is enhanced. Fig. 6 shows that SEP-ACH performs better than SEP. The $1st$ node in SEP dies at round $1130$ whereas in SEP-ACH the $1st$ node dies at round $2004$.
{height="6cm" width="8cm"}
Conclusion
==========
This paper deals with ACH scheme, a clustering technique for WSNs that enhances life time of LEACH, TEEN, DEEC, SEP and minimizing global energy consumption by distributing the load to all the nodes at different points in time. In DEEC-ACH high energy nodes are made CHs frequently than low energy nodes, making energy distribution evenly in the network. Also a CH take part in the selection of next CH thus the number of CHs are reduced and the CHs are made distant. We have forced those nodes which have not become CHs and are close to each other or intersecting. The energy of nodes are conserved in this way and stability period of the network is prolonged. DEEC-ACH outperforms LEACH as LEACH is not suited with heterogeneous environment. DEEC-ACH distribute the energy evenly in the network by giving high priority to high energy nodes in election for CHs and making CHs away from one another. DEEC-ACH also perform well than DEEC as the adjacent and very close nodes are made CHs in DEEC. This consume much energy of the nodes in the process of aggregation and fusion. The global communication is increased and the nodes die at a faster rate. The stability period and throughput of the network is decreased.
[1]{} I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. A survey on sensor networks. IEEE Communications Magazine, 40(8):102–114, August 2002.
T. J. Shepard. A channel access scheme for large dense packet radio networks. In Proccedings of ACM SIGCOMM, pages 219–230, September 1996.
W. Heinzelman, A. Chandrakasan, and H. Balakrishnan. “Energy-Ef?cient Communication Protocols for Wireless Microsensor Networks”. In Proceedings of Hawaiian In- ternational Conference on Systems Science, January 2000.
G. Smaragdakis, I. Matta, A. Bestavros, “SEP: A Stable Election Protocol for clustered heterogeneous wireless sensor networks”, in: Second International Workshop on Sensor and Actor Network Protocols and Applications (SANPA 2004), 2004.
L. Qing, Q. Zhu, M. Wang, ”Design of a distributed energy-efficient clustering algorithm for heterogeneous wireless sensor networks”. In ELSEVIER, Computer Communications 29 (2006) 22302237.
MATLAB 7.4.0(R2007a) www.mathworks.com
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'While the precise mechanism responsible for the L to T dwarf transition remains unclear, it is clearly caused by changing cloud characteristics. Here we briefly review data relevant to understanding the nature of the transition and argue that changing atmospheric dynamics produce the transition by opening holes through the global iron and silicate cloud decks. Other possibilities, such as a sudden vertical collapse in these cloud decks are also considered. Any acceptable model of the L to T transition must ultimately connect changing cloud properties to the underlying atmospheric dynamics.'
author:
- 'M.S. Marley, M.C. Cushing,'
- 'D. Saumon'
title: The L to T dwarf Transition
---
Introduction
============
As the atmospheres of brown dwarfs cool with time, their spectral signatures reflect a progression of changes in their atmospheric chemical equilibrium and condensate structure. In M dwarfs the elements O, C, and N are predominantly found in $\rm H_2$O, CO, and $\rm N_2$ and the atmosphere is too warm for condensation of solids (Allard & Hauschildt 1995; Lodders 1999). As the effective temperature ($T_{\rm eff}$) falls, a variety of condensates form in the atmosphere, most notably iron and silicates. These condensates are apparently not well-mixed through the atmosphere, but are found in discrete cloud layers overlying the condensation level (Ackerman & Marley 2001; Marley et al. 2002; Tsuji 2002; Woitke & Helling 2004).
By the time the $T_{\rm eff}$ falls to that of a late L dwarf the cloud layer is optically thick and affects either directly (as a major opacity source) or indirectly (by altering the atmospheric temperature/pressure profile) all spectral regions. The exact spectral signature of the cloud depends both on its vertical thickness and the particle size distribution of the condensates. In addition, as the atmosphere cools, chemical equilibrium begins to favor first $\rm CH_4$ over CO and then $\rm NH_3$ over $\rm N_2$ (Tsuji 1964; Fegley & Lodders 1996). Thus $\rm CH_4$ absorption in the $K$ band begins to replace CO and $\rm NH_3$ appears (Roellig et al. 2004) in the mid-infrared by the late L’s. By the early to mid T dwarfs the condensate cloud is forming quite deep in the atmosphere. In the relatively clear, cool atmosphere above the cloud, chemical equilibrium begins to strongly favor $\rm CH_4$ and $\rm NH_3$ and their spectral features, along with particularly strong bands of water, grow in prominence (Marley et al. 1996, Burrows et al. 1997, Allard et al. 2001, Burrows et al. 2003). Figures 1 through 3 illustrate these spectral trends in the $K$ and $L$ bands for L5 through T5 brown dwarfs.
Signatures of the L to T Transition
===================================
Below we summarize the characteristics exhibited by brown dwarfs at and near the L to T transition (approximately L8 to T5):
[*Turn to the blue in $J-K$:*]{} The colors of L dwarfs become progressively redder until they saturate at $J-K \sim 2$ at spectral type L8 (Knapp et al. 2004). This color then rapidly turns to the blue, reaching $J-K\sim -0.8$ by T8 or so.
[*Color change at near constant $T_{\rm eff}$:*]{} Recent estimates of the bolometric $T_{\rm eff}$ from Golimowski et al. (2004) have quantified the rapid rate of this color change, as shown in Figures 4 and 5. Most ($>80\%$) of the change is $J-K$ color is seen to occur over a very small $T_{\rm eff}$ range near 1300 K. This is a remarkable result as it implies that brown dwarfs are undergoing substantial spectral and color changes over a very small temperature range.
[*Brightening at J Band:*]{} The L to T transition also appears to be associated with a brightening at $J$ band from late L to early T (T4 or so) (Knapp et al. 2004). $H$, $K$, $L$, and $M$ bands show no sign of such brightening (Knapp et al. 2004; Golimowski et al. 2004), while there is some evidence of a brightening at $Z$. It should be noted that the bolometric luminosity, as would be expected, does [*not*]{} increase across the transition (Golimowski et al. 2004).
[*Resurgence of FeH:*]{} Burgasser et al. (2002) argue there is evidence that, after decaying away as FeH is presumably lost to Fe drops and grains, the $0.997\,\rm\mu m$ FeH band shows a resurgence in strength, coincident with the $J-K$ color change.
[*Model Spectral Fits:*]{} The comparison of models and data shown in Figures 1 through 3 provides additional information about the transition. In Figure 1 a cloudy model does a good job of reproducing the $K$-band spectra of an L5 dwarf. A model with no cloud opacity predicts too much methane absorption in both $K$ and $L$ bands as well as a too-deep water band. Comparing the cloudy and cloudless models for this objects makes clear why the $J-K$ color is such an important diagnostic for the cloud. In the $L$ band the model gets the depth of the $3.3\,\rm\mu m$ methane band correct, which suggests the thermal structure of the model and the associated equilibrium methane abundance are reasonable.
By T2 (Figure 2), however, a model using the same cloud model (Ackerman & Marley 2001) is apparently somewhat too warm, predicting a bit too much CO and too little $\rm CH_4$. At $K$ band the observed spectrum lies between this cloudy model and a cloudless model. The overall shape of the $L$ band spectrum, which probes higher in the atmosphere, seems to be best fit by a combination of the cloudy and cloudless models. Interestingly the amplitude of the methane feature at $3.3\,\rm \mu m$ is larger than either the cloudy or cloudless models predict, which may indicate that the temperature gradient in the photosphere above the cloud deck is steeper than either model predicts.
Finally by T5 (Figure 3) a model with no cloud opacity (but with condensation included in the equilibrium chemistry) fits very well both at $K$ and $L$ bands, implying that condensates play a very small role in controlling the thermal profile and emitted flux. The difference between the best fitting models for the T2 and the T5 dwarfs is only $100\,\rm K$!
The Transition Mechanism
========================
Any explanation of the L to T transition mechanism must be consistent with the evidence summarized above. The unmistakable gross explanation–that condensates have been lost from the atmosphere–belies the difficulty in explaining this loss is a self-consistent manner. That a sinking, finite-thickness cloud deck will eventually disappear from sight allowing the atmosphere above to cool has been apparent for some time (Marley 2000, Allard et al. 2001, Marley et al. 2002, Tsuji 2002). The difficulty lies in explaining the rapidity of the color change in light of the measured effective temperatures (Figures 4 and 5). For example while nicely accounting for the $J-K$ colors of the reddest L dwarfs, the model of Ackerman & Marley (2001) takes much too long to ultimately sink out of sight (Burgasser et al. 2002; Knapp et al. 2004).
In a series of papers Tsuji (Tsuji 2001, Tsuji & Nakajima 2003, Tsuji et al. 2004) proposed that a physically thin cloud, thinner than predicted by the Ackerman & Marley model, could self-consistently explain the rapid L to T transition. These ‘UCM’ models indeed exhibit a faster L- to T-like transition, but as Figure 4 demonstrates the UCM models are not consistent with the observed rapidity of the color change. Even accounting for a likely spread in gravities across the transition can not account for the observations. In addition the UCM models, like the cloudy models of Marley et al., do not brighten in $J$ band across the transition. Tsuji et al. had to invoke an exceptionally large spread in atmospheric $\log g$ of known sources across the transition in order to account for both the reddest and dimmest late L dwarfs and the brightest and bluest early T dwarfs. Finally the UCM models could not explain the resurgence in FeH that is observed across the transition.
To overcome the sort of difficulties faced by the Tsuji et al. models, Burgasser et al. (2002), following a suggestion from Ackerman & Marley (2001), hypothesized that at the L to T transition the global cloud deck rapidly breaks apart. Under this scenario holes in the cloud deck appear at $T_{\rm eff} \sim 1300$ to $1400\,\rm K$. In the molecular window regions, particularly $Z$ and $J$ bands, bright flux from deeply seated regions pours out of the holes left by the departure of the cloud deck. This outpouring of flux is then responsible for the rapid color change in $J-K$ (Figure 5), the brightening in $J$ (and apparently also $Z$) band, and the reappearance of FeH. The fact that the T2 dwarf (Figure 2) seems to be a composite of the cloud free and cloudy model spectra supports this interpretation.
Burgasser et al. suggest that the cloud holes appear when the combined iron and silicate clouds sink sufficiently deeply into the global convection zone. On Earth clouds tend to be more spatially uniform when they form in relatively shallow convective layers. Regions in which the convective layer is thick, such as near the equator, seem to be inhabited by towering cumulus clouds separated by cloud-free regions. The presence of some relatively cloud free regions in the atmospheres of Venus and Jupiter provides evidence that cloud layers are generally not globally uniform in planetary atmospheres and supports the plausibility of the mechanism.
Finally Knapp et al. (2004) proposed a third alternative in which the sedimentation efficiency of the cloud substantially increases at the transition. In this case the cloud remains homogeneous across the disk, but particle growth becomes much more efficient. Efficient growth leads to larger particles which more rapidly fall out of the atmosphere. This leads to optically thinner clouds. In the language of Ackerman & Marley (2001) this is described as $f_{\rm sed}\rightarrow \infty$. As discussed elsewhere in these proceedings, Tsuji and collaborators now favor a sudden collapse of the global cloud deck ($T_{\rm crit} \rightarrow T_{\rm cond}$) at the transition. This is similar to the Knapp et al. (2004) suggestion with the exception that Tsuji et al. do not address the particle size.
Regardless of whether the L to T transition is explained by the appearance of holes in the global cloud deck or a sudden increase in the efficiency of condensate sedimentation, the root cause must lie with the atmospheric dynamics. What aspect of atmospheric circulation or dynamics would favor the appearance of holes or the sudden collapse of the cloud deck? Perhaps the behavior of condensates change when the cloud reaches a certain depth in the atmospheric convection zone or perhaps the second, detached convection zone found in brown dwarf atmosphere models (Marley et al. 1996, Burrows et al. 1997, Allard et al. 2001, Tsuji 2002) plays a role. Another possibility is that there is a change in the global atmospheric circulation that affects the behavior of the cloud decks. Schubert & Zhang (2000) found that brown dwarfs likely exhibit one of two styles of global atmospheric circulation: dominated by rotation, like Jupiter, or fairly independent of rotation, like the sun. Since the luminosity falls with age, the Rayleigh and Eckman numbers of brown dwarfs, which influence the regime in which the atmospheric dynamics falls, likewise vary with time. Perhaps the L to T transition is associated with a change between the two regimes. Until such possible mechanisms have been quantitatively addressed the nature of the L to T transition will remain the domain of plausible, if ad hoc, modeling.
MSM acknowledges support from NASA grants NAG2-6007 and NAG5-8919 and NSF grant AST 00-86288. MC acknowledges support from the Spitzer Fellow Program. This work was supported in part by the U.S. Department of Energy under contract W-7405-ENG-36.
Ackerman, A. S. & Marley, M. S. 2001, Ap. J., 556, 872
Allard, F. & Hauschildt, P. H. 1995, Ap. J., 445, 433
Allard, F. Hauschildt, P. H., Alexander, D. R., Tamanai, A., & Schweitzer, A. 2001, Ap. J., 556, 357
Burgasser, A. J. et al. 2002, Ap. J., 564, 421
Burrows, A., et al. 1997, Ap. J., 491, 856
\[ Fegley, B. J. & Lodders, K. 1996, Ap. J.l, 472, L37
Golimowski, D. et al. 2004, A. J. 127, 3516
Knapp, G. K., et al., 2004, A.J. 127, 3553
Lodders, K. 1999, Ap. J., 519, 793
Marley, M. S., Saumon, D., Guillot, T., Freedman, R. S., Hubbard, W. B., Burrows, A., & Lunine, J. I. 1996, Science, 272, 1919
Marley, M. 2000, ASP Conf. Ser. 212: From Giant Planets to Cool Stars, 152
\[ Marley, M. S., Seager, S., Saumon, D., Lodders, K., Ackerman, A. S., Freedman, R. S., & Fan X. 2002, Ap. J., 568, 335
Roellig, T. et al. 2004, Ap. J. Sup., 154, 418
Schubert, G. & Zhang, K. 2000, ASP Conf. Ser. 212: From Giant Planets to Cool Stars, 210
Tsuji, T. 1964, Annals Tokyo Astron. Obs., 9, 1
Tsuji, T. 2002, Ap. J., 575, 264
Tsuji, T. & Nakajima, T. 2003, ApJ, 585, L151
Tsuji, T., Nakajima, T., & Yanagisawa, K. 2004, ApJ, 607, 511
Woitke, P. & Helling, C. 2004, A&A, 414, 335
|
{
"pile_set_name": "ArXiv"
}
|
---
address: 'Fachbereich Mathematik, Technische Universität Darmstadt, 64289 Darmstadt, Germany.'
author:
- 'Andrew R. Linshaw'
title: Invariant theory and the Heisenberg vertex algebra
---
[The invariant subalgebra $\cH^+$ of the Heisenberg vertex algebra $\cH$ under its automorphism group $\mathbb{Z}/2\mathbb{Z}$ was shown by Dong-Nagatomo to be a $\cW$-algebra of type $\cW(2,4)$. Similarly, the rank $n$ Heisenberg vertex algebra $\cH(n)$ has the orthogonal group $O(n)$ as its automorphism group, and we conjecture that $\cH(n)^{O(n)}$ is a $\cW$-algebra of type $\cW(2,4,6,\dots,n^2+3n)$. We prove our conjecture for $n=2$ and $n=3$, and we show that this conjecture implies that $\cH(n)^G$ is strongly finitely generated for any reductive group $G\subset O(n)$. ]{}
Introduction
============
We call a vertex algebra $\cV$ [*strongly finitely generated*]{} if there exists a finite set of generators such that the collection of iterated Wick products of the generators and their derivatives spans $\cV$. This property has several important consequences, and in particular implies that both Zhu’s associative algebra $A(\cV)$, and Zhu’s commutative algebra $\cV / C_2(\cV)$, are finitely generated. By an [*invariant vertex algebra*]{}, we mean a subalgebra $\cV^G\subset \cV$, where $G$ is a group of automorphisms of $\cV$. It is our belief that if $\cV$ is a simple, strongly finitely generated vertex algebra, and $G$ is reductive, $\cV^G$ will be strongly finitely generated under fairly general circumstances. Isolated examples of this phenomenon have been known for some years (see for example [@BFH][@FKRW][@EFH][@DNI][@KWY]), although the first general results of this kind were obtained in [@LII], in the case where $\cV$ is the $\beta\gamma$-system $\cS(V)$, $bc$-system $\cE(V)$, or $bc\beta\gamma$-system $\cE(V)\otimes \cS(V)$, associated to $V = \mathbb{C}^n$. The strong finite generation property is a subtle and essentially quantum" phenomenon, and is generally destroyed by passing to the classical limit before taking invariants. Often, $\cV$ admits a $G$-invariant filtration for which $gr(\cV)$ is a commutative algebra with a derivation (i.e., an abelian vertex algebra), and the classical limit $gr(\cV^G)$ is isomorphic to $(gr(\cV))^G$ as a commutative algebra. Unlike $\cV^G$, $gr(\cV^G)$ is generally not finitely generated as a vertex algebra, and a presentation will require both infinitely many generators and infinitely many relations.
One of the most basic examples of an invariant vertex algebra was studied by Dong-Nagatomo in [@DNI]. Let $\cH$ denote the Heisenberg vertex algebra, which is generated by a field $\alpha$ satisfying the operator product expansion (OPE) relation $\alpha(z) \alpha(w)\sim (z-w)^{-2}$. Clearly the automorphism group $Aut(\cH)$ is isomorphic to $\mathbb{Z}/2\mathbb{Z}$, and is generated by the involution $\theta$ sending $\alpha\mapsto -\alpha$. In [@DNI] it was shown that the invariant subalgebra $\cH^+$ under $\theta$ is a $\cW$-algebra of type $\cW(2,4)$, and in particular is strongly generated by the Virasoro element $L$ and an element $J$ of weight four. Using this result, the authors described the Zhu algebra of $\cH^+$, which is a commutative algebra on two generators, and they classified the irreducible modules of $\cH^+$. In [@DNII], Dong-Nagatomo considered a higher-rank analogue $\cH(n)^+$ of $\cH^+$. In this notation, $\cH(n)$ is the rank $n$ Heisenberg algebra, which is just the tensor product of $n$ copies of $\cH$, and $\cH(n)^+$ is the invariant subalgebra under the $-1$ involution. Unlike the rank $1$ case, the Zhu algebra of $\cH(n)^+$ is nonabelian, and it is difficult to describe it completely by generators and relations. However, the authors obtained enough information about it to classify the irreducible modules of $\cH(n)^+$. This result is important for understanding the structure and representation theory of vertex algebras of the form $V_L^+$. Here $V_L$ is the lattice vertex algebra attached to some lattice $L$ of rank $n$, and $V_L^+$ is the invariant subalgebra under the $-1$ involution.
In this paper, we study general invariant vertex algebras $\cH(n)^G$, where $G$ is an arbitrary reductive group of automorphisms of $\cH(n)$. In the case where the action of $G$ extends to $V_L$, an understanding of $\cH(n)^G$ is a necessary first step in studying $V_L^G$. The [*full*]{} automorphism group of $\cH(n)$ preserving a natural conformal structure is the orthogonal group $O(n)$. We begin by studying $\cH(n)^{O(n)}$, which coincides with $\cH^+$ in the case $n=1$. For an arbitrary reductive group $G\subset O(n)$, $\cH(n)^G$ is completely reducible as a module over $\cH(n)^{O(n)}$, and this module structure is an essential ingredient in our description. Our approach in this paper is quite parallel to our earlier study of invariant subalgebras of the $\beta\gamma$-system $\cS(V)$. The automorphism group of $\cS(V)$ preserving its conformal structure is $GL_n$, and $\cS(V)^{GL_n}$ is isomorphic to $\cW_{1+\infty,-n}$ by a theorem of Kac-Radul [@KR]. In [@LI], we studied $\cW_{1+\infty,-n}$ via classical invariant theory, and we use a similar method in Section \[secortho\] of this paper to study $\cH(n)^{O(n)}$. There are many parallels between these two vertex algebras; for example, they both have abelian Zhu algebras, which implies that their irreducible, admissible modules are all highest-weight modules. This observation is crucial in our description in Section \[secgeneral\] of $\cH(n)^G$ as a module over $\cH(n)^{O(n)}$, which uses essentially the same ideas as [@LII].
First of all, there is an $O(n)$-invariant filtration on $\cH(n)$ such that the associated graded object $gr(\cH(n))$ is isomorphic to $S=Sym \bigoplus_{j\geq 0} V_j$ as a commutative ring, where $V_j\cong \mathbb{C}^n$ as $O(n)$-modules. In fact, $\cH(n)\cong S$ as vector spaces, and we view $\cH(n)$, equipped with the Wick product, as a deformation of $S$. Using Weyl’s first and second fundamental theorems of invariant theory for the standard representation of $O(n)$, we obtain a natural (infinite) strong generating set $\{\omega_{a,b}|~0\leq a\leq b\}$ for $\cH(n)^{O(n)}$, as well as an infinite set of relations among these generators. A linear change of variables produces a slightly more economical set of strong generators $\{j^{2m}|~m\geq 0\}$, where $j^{2m}$ has weight $2m+2$. In fact, $\cH(n)^{O(n)}$ is generated as a vertex algebra by $\{j^0, j^2\}$ for all $n\geq 1$, although this is only a strong generating set in the case $n=1$. The relation of minimal weight among $\{j^{2m}|~m\geq 0\}$ and their derivatives occurs at weight $n^2+3n+2$, and we conjecture that it gives rise to a decoupling relation $$j^{n^2+3n} = P(j^0,j^2,\dots,j^{n^2+3n-2}).$$ Here $P$ is a normally ordered polynomial in $j^0, j^2,\dots,j^{n^2+3n-2}$, and their derivatives. An easy consequence of our conjecture is that higher decoupling relations of the form $j^{2r} = Q_{2r}(j^0,j^2,\dots,j^{n^2+3n-2})$ exist for all $r\geq\frac{1}{2} (n^2+3n)$. Hence $\cH(n)^{O(n)}$ has a minimal strong generating set $\{j^0, j^2, \dots, j^{n^2+3n-2}\}$, and in particular is a $\cW$-algebra of type $\cW(2,4,6,\dots, n^2+3n)$. By computer calculation, we prove our conjecture for $n=2$ and $n=3$, but we are unable to prove it in general.
By a fundamental result of Dong-Li-Mason [@DLM], $\cH(n)$ has a decomposition of the form $$\cH(n) \cong \bigoplus_{\nu\in H} L(\nu)\otimes M^{\nu},$$ where $H$ indexes the irreducible, finite-dimensional representations $L(\nu)$ of $O(n)$, and the $M^{\nu}$’s are inequivalent, irreducible, admissible $\cH(n)^{O(n)}$-modules. Since the Zhu algebra of $\cH(n)^{O(n)}$ is abelian, each $M^{\nu}$ above is a highest-weight module. For any reductive group $G\subset O(n)$, $\cH(n)^G$ is also a direct sum of irreducible, highest-weight $\cH(n)^{O(n)}$-modules. Using a classical theorem of Weyl, we show that there is a finite set of irreducible $\cH(n)^{O(n)}$-submodules of $\cH(n)^G$ whose direct sum contains an (infinite) strong generating for $\cH(n)^G$. Since $\cH(n)^{O(n)}$ is finitely generated, this shows that $\cH(n)^G$ is finitely generated as a vertex algebra. Finally, assuming our conjecture that $\cH(n)^{O(n)}$ is strongly finitely generated, we show that $\cH(n)^G$ is strongly finitely generated as well. Since our conjecture holds for $n=2$ and $n=3$, the strong finite generation of $\cH(2)^G$ and $\cH(3)^G$ for an arbitrary reductive $G$ is an immediate consequence.
There is an application of these results to invariant subalgebras of affine vertex algebras which we develop in a separate paper [@LIII]. Let $\gg$ be a simple, finite-dimensional Lie algebra, and let $V_k(\gg)$ denote the universal affine vertex algebra at level $k$ associated to $\gg$. It is freely generated by vertex operators $X^{\xi}$, which are linear in $\xi\in\gg$ and satisfy the OPE relations $$X^{\xi}(z) X^{\eta}(w) \sim k\bra \xi,\eta\ket (z-w)^{-2} + X^{[\xi,\eta]}(w)(z-w)^{-1},$$ where $\bra ,\ket$ denotes the normalized Killing form $\frac{1}{2 h^{\vee}} \bra ,\ket_K$. Let $G$ be a reductive group of automorphisms of $V_k(\gg)$ for all $k\in \mathbb{C}$. In particular, $G$ acts on the weight-one subspace $V_k(\gg)[1]\cong \gg$, and $G$ preserves both the bracket and the bilinear form on $\gg$. Therefore $G$ lies in the orthogonal group $O(n)$ for $n= \text{dim}(\gg)$, so $G$ also acts on the Heisenberg algebra $\cH(n)$. As vector spaces, we have $V_k(\gg)^G \cong (Sym \oplus_{j\geq 0} V_j)^G\cong \cH(n)^G$, where $V_j\cong \mathbb{C}^n$ for all $j\geq 0$, and we regard $\cH(n)^G$ as a partial abelianization" of $V_k(\gg)^G$. Invariant subalgebras of $V_k(\gg)$ are much more complicated and difficult to study than invariant subalgebras of $\cH(n)$, but for generic values of $k$, a strong generating set for $\cH(n)^G$ gives rise to a strong generating set for $V_k(\gg)^G$. Therefore the conjectured strong finite generation of $\cH(n)^{O(n)}$ has a far-reaching consequence; it implies that $V_k(\gg)^G$ is strongly finitely generated for generic values of $k$. Finally, since our conjecture holds for $n=3$, this statement is true in the case $\gg = \gs\gl_2$.
Vertex algebras
===============
In this section, we define vertex algebras, which have been discussed from various different points of view in the literature [@B][@FBZ][@FHL][@FLM][@K][@LiI][@LZ]. We will follow the formalism developed in [@LZ] and partly in [@LiI]. Let $V=V_0\oplus V_1$ be a super vector space over $\mathbb{C}$, and let $z,w$ be formal variables. By $QO(V)$, we mean the space of all linear maps $$V\ra V((z)):=\{\sum_{n\in\mathbb{Z}} v(n) z^{-n-1}|
v(n)\in V,\ v(n)=0\ \text{for} \ n>>0 \}.$$ Each element $a\in QO(V)$ can be uniquely represented as a power series $$a=a(z):=\sum_{n\in\mathbb{Z}}a(n)z^{-n-1}\in End(V)[[z,z^{-1}]].$$ We refer to $a(n)$ as the $n$th Fourier mode of $a(z)$. Each $a\in
QO(V)$ is of the shape $a=a_0+a_1$ where $a_i:V_j\ra V_{i+j}((z))$ for $i,j\in\mathbb{Z}/2\mathbb{Z}$, and we write $|a_i| = i$.
On $QO(V)$ there is a set of nonassociative bilinear operations $\circ_n$, indexed by $n\in\mathbb{Z}$, which we call the $n$th circle products. For homogeneous $a,b\in QO(V)$, they are defined by $$a(w)\circ_n b(w)=Res_z a(z)b(w)~\iota_{|z|>|w|}(z-w)^n-
(-1)^{|a||b|}Res_z b(w)a(z)~\iota_{|w|>|z|}(z-w)^n.$$ Here $\iota_{|z|>|w|}f(z,w)\in\mathbb{C}[[z,z^{-1},w,w^{-1}]]$ denotes the power series expansion of a rational function $f$ in the region $|z|>|w|$. We usually omit the symbol $\iota_{|z|>|w|}$ and just write $(z-w)^{-1}$ to mean the expansion in the region $|z|>|w|$, and write $-(w-z)^{-1}$ to mean the expansion in $|w|>|z|$. It is easy to check that $a(w)\circ_n b(w)$ above is a well-defined element of $QO(V)$.
The nonnegative circle products are connected through the [*operator product expansion*]{} (OPE) formula. For $a,b\in QO(V)$, we have $$\label{opeform} a(z)b(w)=\sum_{n\geq 0}a(w)\circ_n
b(w)~(z-w)^{-n-1}+:a(z)b(w):\ ,$$ which is often written as $a(z)b(w)\sim\sum_{n\geq 0}a(w)\circ_n b(w)~(z-w)^{-n-1}$, where $\sim$ means equal modulo the term $$:a(z)b(w):\ =a(z)_-b(w)\ +\ (-1)^{|a||b|} b(w)a(z)_+.$$ Here $a(z)_-=\sum_{n<0}a(n)z^{-n-1}$ and $a(z)_+=\sum_{n\geq
0}a(n)z^{-n-1}$. Note that $:a(w)b(w):$ is a well-defined element of $QO(V)$. It is called the [*Wick product*]{} of $a$ and $b$, and it coincides with $a\circ_{-1}b$. The other negative circle products are related to this by $$n!~a(z)\circ_{-n-1}b(z)=\ :(\partial^n a(z))b(z):\ ,$$ where $\partial$ denotes the formal differentiation operator $\frac{d}{dz}$. For $a_1(z),\dots ,a_k(z)\in QO(V)$, the $k$-fold iterated Wick product is defined to be $$\label{iteratedwick} :a_1(z)a_2(z)\cdots a_k(z):\ =\ :a_1(z)b(z):~,$$ where $b(z)=\ :a_2(z)\cdots a_k(z):\ $. We often omit the formal variable $z$ when no confusion can arise.
The set $QO(V)$ is a nonassociative algebra with the operations $\circ_n$, which satisfy $1\circ_n a=\delta_{n,-1}a$ for all $n$, and $a\circ_n 1=\delta_{n,-1}a$ for $n\geq -1$. In particular, $1$ behaves as a unit with respect to $\circ_{-1}$. A linear subspace $\cA\subset QO(V)$ containing $1$ which is closed under the circle products will be called a [*quantum operator algebra*]{} (QOA). Note that $\cA$ is closed under $\partial$ since $\partial a=a\circ_{-2}1$. Many formal algebraic notions are immediately clear: a homomorphism is just a linear map that sends $1$ to $1$ and preserves all circle products; a module over $\cA$ is a vector space $M$ equipped with a homomorphism $\cA\rightarrow
QO(M)$, etc. A subset $S=\{a_i|\ i\in I\}$ of $\cA$ is said to generate $\cA$ if every element $a\in\cA$ can be written as a linear combination of nonassociative words in the letters $a_i$, $\circ_n$, for $i\in I$ and $n\in\mathbb{Z}$. We say that $S$ [*strongly generates*]{} $\cA$ if every $a\in\cA$ can be written as a linear combination of words in the letters $a_i$, $\circ_n$ for $n<0$. Equivalently, $\cA$ is spanned by the collection $\{ :\partial^{k_1} a_{i_1}(z)\cdots \partial^{k_m} a_{i_m}(z):| ~i_1,\dots,i_m \in I,~ k_1,\dots,k_m \geq 0\}$.
We say that $a,b\in QO(V)$ [*quantum commute*]{} if $(z-w)^N
[a(z),b(w)]=0$ for some $N\geq 0$. Here $[,]$ denotes the super bracket. This condition implies that $a\circ_n b = 0$ for $n\geq N$, so (\[opeform\]) becomes a finite sum. A [*commutative quantum operator algebra*]{} (CQOA) is a QOA whose elements pairwise quantum commute. Finally, the notion of a CQOA is equivalent to the notion of a vertex algebra. Every CQOA $\cA$ is itself a faithful $\cA$-module, called the [*left regular module*]{}. Define $$\rho:\cA\rightarrow QO(\cA),\ \ \ \ a\mapsto\hat a,\ \ \ \ \hat
a(\zeta)b=\sum_{n\in\mathbb{Z}} (a\circ_n b)~\zeta^{-n-1}.$$ Then $\rho$ is an injective QOA homomorphism, and the quadruple of structures $(\cA,\rho,1,\partial)$ is a vertex algebra in the sense of [@FLM]. Conversely, if $(V,Y,{\bf 1},D)$ is a vertex algebra, the collection $Y(V)\subset QO(V)$ is a CQOA. [*We will refer to a CQOA simply as a vertex algebra throughout the rest of this paper*]{}.
The following are useful identities that measure the nonassociativity and noncommutativity of the Wick product, and the failure of the positive circle products to be derivations of the Wick product. Let $a,b,c$ be vertex operators in some vertex algebra $\cA$, and let $n > 0$. Then $$\label{vaidi} :(:ab:)c:-:abc:=\sum_{k\geq0}{1\over(k+1)!}\left(:(\partial^{k+1}a)(b\circ_k
c): +(-1)^{|a||b|}:(\partial^{k+1}b)(a\circ_k c):\right),$$ $$\label{vaidii} :ab:-(-1)^{|a||b|}:ba:=\sum_{k\geq0}{(-1)^k\over(k+1)!}\partial^{k+1}(a\circ_kb),$$ $$\label{vaidiii} a\circ_n(:bc:)-:(a\circ_nb)c:-(-1)^{|a||b|}:b(a\circ_nc):= \sum_{k=1}^n\left(\begin{matrix} n\cr k\end{matrix} \right)(a\circ_{n-k}b)\circ_{k-1}c,$$ $$\label{vaidiv} (:ab:)\circ_n
c=\sum_{k\geq0}{1\over k!}:(\partial^ka)(b\circ_{n+k}c):
+(-1)^{|a||b|}\sum_{k\geq0}b\circ_{n-k-1}(a\circ_k c) .$$
Category $\cR$
==============
In [@LL] we considered a certain category $\cR$ of vertex algebras, together with a functor from $\cR$ to the category of supercommutative rings.
Let $\cR$ be the category of vertex algebras $\cA$ equipped with a $\mathbb{Z}_{\geq 0}$-filtration $$\label{goodi} \cA_{(0)}\subset\cA_{(1)}\subset\cA_{(2)}\subset \cdots,\ \ \ \cA = \bigcup_{k\geq 0}
\cA_{(k)}$$ such that $\cA_{(0)} = \mathbb{C}$, and for all $a\in \cA_{(k)}$, $b\in\cA_{(l)}$, we have $$\label{goodii} a\circ_n b\in\cA_{(k+l)},\ \ \ \text{for}\
n<0,$$ $$\label{goodiii} a\circ_n b\in\cA_{(k+l-1)},\ \ \ \text{for}\
n\geq 0.$$ Elements $a(z)\in\cA_{(d)}\setminus \cA_{(d-1)}$ are said to have degree $d$, and morphisms in $\cR$ are vertex algebra homomorphisms that preserve the filtration.
Filtrations on vertex algebras satisfying (\[goodii\])-(\[goodiii\]) were introduced in [@LiII] and are known as [*good increasing filtrations*]{}. Setting $\cA_{(-1)} = \{0\}$, the associated graded object $gr(\cA) = \bigoplus_{k\geq 0}\cA_{(k)}/\cA_{(k-1)}$ is a $\mathbb{Z}_{\geq 0}$-graded associative, supercommutative algebra with a unit $1$ under a product induced by the Wick product on $\cA$. In general, there is no natural linear map $\cA\ra gr (\cA)$, but for each $r\geq 1$ we have the projection $$\label{proj} \phi_r: \cA_{(r)} \ra \cA_{(r)}/\cA_{(r-1)}\subset gr(\cA).$$ Moreover, $gr(\cA)$ has a derivation $\partial$ of degree zero (induced by the operator $\partial = \frac{d}{dz}$ on $\cA$), and for each $a\in\cA_{(d)}$ and $n\geq 0$, the operator $a\circ_n$ on $\cA$ induces a derivation of degree $d-k$ on $gr(\cA)$, which we also denote by $a\circ_n$. Here $$k = sup \{ j\geq 1|~ \cA_{(r)}\circ_n \cA_{(s)}\subset \cA_{(r+s-j)}~\forall r,s,n\geq 0\},$$ as in [@LL]. Finally, these derivations give $gr(\cA)$ the structure of a vertex Poisson algebra.
The assignment $\cA\mapsto gr(\cA)$ is a functor from $\cR$ to the category of $\mathbb{Z}_{\geq 0}$-graded supercommutative rings with a differential $\partial$ of degree 0, which we will call $\partial$-rings. A $\partial$-ring is the same thing as an [*abelian*]{} vertex algebra, that is, a vertex algebra $\cV$ in which $[a(z),b(w)] = 0$ for all $a,b\in\cV$. A $\partial$-ring $A$ is said to be generated by a subset $\{a_i|~i\in I\}$ if $\{\partial^k a_i|~i\in I, k\geq 0\}$ generates $A$ as a graded ring. The key feature of $\cR$ is the following reconstruction property [@LL]:
\[recon\] Let $\cA$ be a vertex algebra in $\cR$ and let $\{a_i|~i\in I\}$ be a set of generators for $gr(\cA)$ as a $\partial$-ring, where $a_i$ is homogeneous of degree $d_i$. If $a_i(z)\in\cA_{(d_i)}$ are vertex operators such that $\phi_{d_i}(a_i(z)) = a_i$, then $\cA$ is strongly generated as a vertex algebra by $\{a_i(z)|~i\in I\}$.
As shown in [@LI], there is a similar reconstruction property for kernels of surjective morphisms in $\cR$. Let $f:\cA\ra \cB$ be a morphism in $\cR$ with kernel $\cJ$, such that $f$ maps $\cA_{(k)}$ onto $\cB_{(k)}$ for all $k\geq 0$. The kernel $J$ of the induced map $gr(f): gr(\cA)\ra gr(\cB)$ is a homogeneous $\partial$-ideal (i.e., $\partial J \subset J$). A set $\{a_i|~i\in I\}$ such that $a_i$ is homogeneous of degree $d_i$ is said to generate $J$ as a $\partial$-ideal if $\{\partial^k a_i|~i\in I,~k\geq 0\}$ generates $J$ as an ideal.
\[idealrecon\] Let $\{a_i| i\in I\}$ be a generating set for $J$ as a $\partial$-ideal, where $a_i$ is homogeneous of degree $d_i$. Then there exist vertex operators $a_i(z)\in \cA_{(d_i)}$ with $\phi_{d_i}(a_i(z)) = a_i$, such that $\{a_i(z)|~i\in I\}$ generates $\cJ$ as a vertex algebra ideal.
The structure of $\cH(n)^{O(n)}$ {#secortho}
================================
The ring of Laurent polynomials $\mathbb{C}[t,t^{-1}]$ may be regarded as an abelian Lie algebra. It has a central extension $\gh = \mathbb{C}[t,t^{-1}]\oplus \mathbb{C}\kappa$ with bracket $[t^n,t^m] = n \delta_{n+m,0} \kappa$, and $\mathbb{Z}$-gradation $deg(t^n) = n$, $deg(\kappa) = 0$. Let $\gh_{\geq 0} = \oplus_{n\geq 0} \gh_n$, and let $C$ be the one-dimensional $\gh_{\geq 0}$-module on which $t^n$ acts trivially for $n\geq 0$, and $\kappa$ acts by the identity. Define $V = U(\gh)\otimes_{U(\gh_{\geq 0})} C$, and let $\alpha(n)\in End(V)$ be the linear operator representing $t^n$ on $V$. Define $\alpha(z) = \sum_{n\in\mathbb{Z}} \alpha(n) z^{-n-1}$, which is easily seen to lie in $QO(V)$ and satisfy the OPE relation $$\alpha(z)\alpha(w)\sim (z-w)^{-2}.$$ The vertex algebra $\cH$ generated by $\alpha$ is known as the [*Heisenberg vertex algebra*]{}. The rank $n$ Heisenberg algebra $\cH(n)$ is just the tensor product of $n$ copies of $\cH$, with generators $\alpha^1,\dots,\alpha^n$. There is a natural conformal structure of central charge $n$ on $\cH(n)$, with Virasoro element $$\label{virasoro} L(z) = \frac{1}{2} \sum_{i=1}^n :\alpha^i(z) \alpha^i(z):,$$ under which each $\alpha^i$ is primary of weight one. The full automorphism group of $\cH(n)$ preserving $L(z)$ is easily seen to be the orthogonal group $O(n)$, which acts linearly on the vector space $U$ spanned by $\{\alpha^1,\dots,\alpha^n\}$. First, any conformal automorphism $\phi$ of $\cH(n)$ must lie in $GL(U)$ by weight considerations. Moreover, since $\alpha^i \circ_1 \alpha^j = \phi(\alpha^i)\circ_1 \phi(\alpha^j) = \delta_{i,j}$, $\phi$ must preserve the pairing $\bra,\ket$ on $U$ defined by $\bra \alpha^i,\alpha^j\ket = \delta_{i,j}$.
We define a good increasing filtration on $\cH(n)$ as follows: $\cH(n)_{(r)}$ is spanned by the set $$\label{goodsv} \{:\partial^{k_1} \alpha^{i_1} \cdots \partial^{k_s}\alpha^{i_s} :| ~ k_j\geq 0,~ s \leq r\}.$$ Then $\cH(n)\cong gr(\cH(n))$ as linear spaces, and as commutative algebras we have $$\label{structureofgrs} gr(\cH(n))\cong Sym \bigoplus_{k\geq 0} V_k.$$ In this notation, $V_k$ is the linear span of $\{\alpha^{i}_k |~ i=1,\dots,n\}$, where $\alpha^i_k$ is the image of $\partial^k \alpha^i(z)$ in $gr(\cH(n))$ under the projection $\phi_1: \cH(n)_{(1)}\ra \cH(n)_{(1)}/\cH(n)_{(0)}\subset gr(\cH(n))$. The action of $O(n)$ on $\cH(n)$ preserves this filtration, and induces an action of $O(n)$ on $gr(\cH(n))$ by algebra automorphisms. For all $k\geq 0$ we have isomorphisms of $O(n)$-modules $V_k\cong \mathbb{C}^n$. Finally, for any reductive subgroup $G\subset O(n)$, $\cH(n)^G\cong gr(\cH(n)^G)$ as linear spaces, and $$\label{structureofgrsinv} gr(\cH(n)^G )\cong (gr(\cH(n))^G \cong (Sym \bigoplus_{k\geq 0} V_k)^G$$ as commutative algebras. In the case $G=O(n)$, the following classical theorem of Weyl [@W] describes the generators and relations of the ring $(Sym \bigoplus_{k\geq 0} V_k)^{O(n)}$:
\[weylfft\] (Weyl) For $k\geq 0$, let $V_k$ be the copy of the standard $O(n)$-module $\mathbb{C}^n$ with orthonormal basis $\{x_{i,k}| ~i=1,\dots,n\}$. The invariant ring $(Sym \bigoplus_{k\geq 0} V_k )^{O(n)}$ is generated by the quadratics $$\label{weylgenerators} q_{a,b} = \sum_{i=1}^n x_{i,a} x_{i,b},\ \ \ \ \ \ 0\leq a\leq b.$$ For $a>b$, define $q_{a,b} = q_{b,a}$, and let $\{Q_{a,b}|\ a,b\geq 0\}$ be commuting indeterminates satisfying $Q_{a,b} = Q_{b,a}$ and no other algebraic relations. The kernel $I_n$ of the homomorphism $\mathbb{C}[Q_{a,b}]\ra (Sym \bigoplus_{k\geq 0} V_k)^{O(n)}$ sending $Q_{a,b}\mapsto q_{a,b}$ is generated by the $(n+1)\times (n+1)$ determinants $$\label{weylrel} d_{I,J} = \det \left[\begin{matrix} Q_{i_0,j_0} & \cdots & Q_{i_0,j_n} \cr \vdots & & \vdots \cr Q_{i_n,j_0} & \cdots & Q_{i_n,j_n} \end{matrix} \right].$$ In this notation, $I=(i_0,\dots, i_{n})$ and $J = (j_0,\dots, j_{n})$ are lists of integers satisfying $$\label{ijineq} 0\leq i_0<\cdots <i_n,\ \ \ \ \ \ 0\leq j_0<\cdots <j_n.$$ Since $Q_{a,b} = Q_{b,a}$, it is clear that $d_{I,J} = d_{J,I}$.
Under the projection $$\phi_2: (\cH(n)^{O(n)})_{(2)}\ra (\cH(n)^{O(n)})_{(2)}/(\cH(n)^{O(n)})_{(1)}\subset gr(\cH(n)^{O(n)}) \cong (Sym \bigoplus_{k\geq 0} V_k)^{O(n)},$$ the generators $q_{a,b}$ of $(Sym \bigoplus_{k\geq 0} V_k)^{O(n)}$ correspond to vertex operators $\omega_{a,b}$ given by $$\label{omegagen} \omega_{a,b} = \sum_{i=1}^n :\partial^a \alpha^i \partial^b \alpha^i:,\ \ \ \ \ \ 0\leq a\leq b.$$ By Lemma \[recon\], the set $\{\omega_{a,b}|~0\leq a \leq b\}$ is a strong generating set for $\cH(n)^{O(n)}$. Note that $\omega_{0,0} = 2L$, where $L$ is the Virasoro element (\[virasoro\]). The subspace $(\cH(n)^{O(n)})_{(2)}$ of degree at most 2 has a basis $\{1\} \cup \{\omega_{a,b}\}$, and for all $n\geq 0$, the operators $\omega_{a,b}\circ_n$ preserve this vector space. It follows that every term in the OPE formula for $\omega_{a,b}(z) \omega_{c,d}(w)$ is a linear combination of these generators, so they form a Lie conformal algebra. We calculate that for $a,b,c\geq 0$ and $0\leq m \leq a+b+c+1$, $$\label{bcalc} \omega_{a,b} \circ_m \partial^c \alpha^i = \lambda_{a,b,c,m} \partial^{a+b+c+1-m} \alpha^i$$ where $$\lambda_{a,b,c,m} = (-1)^b \frac{(b+c+1)!}{(b+c+1-m)!} + (-1)^a \frac{(a+c+1)!}{(a+c+1-m)!}.$$ It follows that for $m\leq a+b+c+1$ we have $$\label{opeformula} \omega_{a,b}\circ_m \omega_{c,d} = \lambda_{a,b,c,m} \omega_{a+b+c+1-m,d} + \lambda_{a,b,d,m}\omega_{c,a+b+d+1-m}.$$ In fact, there is a somewhat more economical set of strong generators for $\cH(n)^{O(n)}$. For each $m\geq 0$, let $A_m$ denote the vector space spanned by $\{\omega_{a,b}|~ a+b = m\}$, which is homogeneous of weight $m+2$. Clearly $\text{dim}(A_{2m}) = m+1 = \text{dim}(A_{2m+1})$ for $m\geq 0$. Moreover, $\partial(A_m)\subset A_{m+1}$, and we have $$\label{deca} \text{dim} \big(A_{2m} / \partial(A_{2m-1})\big) = 1,\ \ \ \ \ \ \text{dim} \big(A_{2m+1} / \partial(A_{2m})\big) = 0.$$ For $m\geq 0$, define $$\label{defofj} j^{2m} = \omega_{0,2m},$$ which is clearly not a total derivative. Hence $A_{2m}$ has a decomposition $$\label{decompofa} A_{2m} = \partial (A_{2m-1})\oplus \bra j^{2m}\ket = \partial^2 (A_{2m-2})\oplus \bra j^{2m}\ket ,$$ where $\bra j^{2m}\ket$ is the linear span of $j^{2m}$. Similarly, $A_{2m+1}$ has a decomposition $$\label{decompofai} A_{2m+1} = \partial^2(A_{2m-1})\oplus \bra \partial j^{2m}\ket = \partial^3 (A_{2m-2})\oplus \bra \partial j^{2m}\ket.$$ It is easy to see that $\{\partial^{2i} j^{2m-2i}|~ 0\leq i\leq m\}$ and $\{\partial^{2i+1} j^{2m-2i}|\ 0\leq i\leq m\}$ are bases of $A_{2m}$ and $A_{2m+1}$, respectively. Hence each $\omega_{a,b}\in A_{2m}$ and $\omega_{c,d}\in A_{2m+1}$ can be expressed uniquely in the form $$\label{lincomb} \omega_{a,b} =\sum_{i=0}^m \lambda_i \partial^{2i}j^{2m-2i},\ \ \ \ \ \ \omega_{c,d} =\sum_{i=0}^m \mu_i \partial^{2i+1}j^{2m-2i}$$ for constants $\lambda_i,\mu_i$, $i=0,\dots,m$. Hence $\{j^{2m}|\ m\geq 0\}$ is an alternative strong generating set for $\cH(n)^{O(n)}$, and it will be convenient to pass back and forth between the sets $\{j^{2m}|\ m\geq 0\}$ and $\{\omega_{a,b}|\ 0\leq a\leq b\}$.
\[ordfingen\] $\cH(n)^{O(n)}$ is generated as a vertex algebra by $j^0$ and $j^2$.
Let $\cJ$ denote the vertex subalgebra of $\cH(n)^{O(n)}$ generated by $j^0$ and $j^2$. We need to show that $j^{2m}\in\cJ$ for all $m\geq 2$. Specializing (\[opeformula\]) shows that $$\label{specialope} j^2\circ_1 j^{2k} = 4 \omega_{2,2k} + (4+4k) j^{2k+2}.$$ Moreover, it is easy to check that $\omega_{2,2k} \equiv - j^{2k+2}$ modulo second derivatives. It follows that $j^2\circ_1j^{2k} \equiv (3+4k)j^{2k+2}$ modulo a linear combination of elements of the form $\partial^{2i} j^{2k+2-2i}$ for $1\leq i \leq k+1$. The claim then follows by induction on $k$.
Consider the category of all vertex algebras with generators $\{J^{2m}|~m\geq 0\}$, which satisfy the same OPE relations as the generators $\{j^{2m}|~m\geq 0\}$ of $\cH(n)^{O(n)}$. Since the vector space with basis $\{1\}\cup \{\partial^lj^{2m}|~l,m\geq 0\}$ is closed under all nonnegative circle products, it forms a Lie conformal algebra. By Theorem 7.12 of [@BK], this category possesses a universal object $\cV_n$, which is [*freely*]{} generated by $\{J^{2m}|\ m\geq 0\}$. In other words, there are no nontrivial normally ordered polynomial relations among the generators and their derivatives in $\cV_n$. Then $\cH(n)^{O(n)}$ is a quotient of $\cV_n$ by an ideal $\cI_n$, and since $\cH(n)^{O(n)}$ is a simple vertex algebra, $\cI_n$ is a maximal ideal. Let $\pi_n: \cV_n\ra \cH(n)^{O(n)}$ denote the quotient map, which sends $J^{2m}\mapsto j^{2m}$. Using the formula (\[lincomb\]), which holds in $\cH(n)^{O(n)}$ for all $n$, we can define an alternative strong generating set $\{\Omega_{a,b}| ~0\leq a\leq b\}$ for $\cV_n$ by the same formula: for $a+b = 2m$ and $c+d = 2m+1$, $$\Omega_{a,b} =\sum_{i=0}^m \lambda_i \partial^{2i}J^{2m-2i},\ \ \ \ \ \ \Omega_{c,d} =\sum_{i=0}^m \mu_i \partial^{2i+1}J^{2m-2i}.$$ Clearly $\pi_n(\Omega_{a,b}) = \omega_{a,b}$. We will use the same notation $A_m$ to denote the linear span of $\{\Omega_{a,b}|~ a+b = m\}$, when no confusion can arise. Finally, $\cV_{n}$ has a good increasing filtration in which $(\cV_n)_{(2k)}$ is spanned by iterated Wick products of the generators $J^{2m}$ and their derivatives, of length at most $k$, and $(\cV_n)_{(2k+1)} = (\cV_n)_{(2k)}$. Equipped with this filtration, $\cV_n$ lies in the category $\cR$, and $\pi_n$ is a morphism in $\cR$.
Since the vertex operators $J^{2m}$ satisfy the same OPE relations as the $j^{2m}$, $\cV_n$ is also generated as a vertex algebra by $J^0$ and $J^2$.
Recall the variables $Q_{a,b}$ and $q_{a,b}$ appearing in Theorem \[weylfft\]. Since $\cV_n$ is freely generated by $\{J^{2m}|\ m\geq 0\}$, and $\{\Omega_{a,b}|\ 0\leq a\leq b\}$ and $\{\partial^k J^{2m}|\ k,m\geq 0\}$ form bases for the same space, we may identify $gr(\cV_n)$ with $\mathbb{C}[Q_{a,b}]$, and we identify $gr(\cH(n)^{O(n)})$ with $\mathbb{C}[q_{a,b}]/I_n$. Under this identification, $gr(\pi_{n}): gr(\cV_n) \ra gr(\cH(n)^{O(n)})$ is just the quotient map sending $Q_{a,b}\mapsto q_{a,b}$. Clearly the projection $\pi_{n}: \cV_n\ra \cH(n)^{O(n)}$ maps each filtered piece $(\cV_n)_{(k)}$ onto $(\cH(n)^{O(n)})_{(k)}$, so the hypotheses of Lemma \[idealrecon\] are satisfied. Since $I_{n} = Ker (gr(\pi_{n}))$ is generated by the determinants $d_{I,J}$, we can apply Lemma \[idealrecon\] to find vertex operators $D_{I,J}\in (\cV_{n})_{(2n+2)}$ satisfying $\phi_{2n+2}(D_{I,J}) = d_{I,J}$, such that $\{D_{I,J}\}$ generates $\cI_{n}$. Since $\Omega_{a,b}$ has weight $a+b+2$, $$\label{wtod} wt(D_{I,J}) = |I| + |J| +2n+2,\ \ \ \ \ \ |I| =\sum_{a=0}^{n} i_a,\ \ \ \ \ \ |J| =\sum_{a=0}^{n} j_a.$$
In general, the vertex operators $a_i(z)$ furnished by Lemma \[idealrecon\] satisfying $\phi_{d_i}(a_i(z)) = a_i$ which generate $\cI$ are not unique. However, in our case, $D_{I,J}$ is uniquely determined by the conditions $$\label{uniquedij} \phi_{2n+2}(D_{I,J}) = d_{I,J},\ \ \ \ \ \ \pi_{n}(D_{I,J}) = 0.$$ If $D'_{I,J}$ is another vertex operator satisfying (\[uniquedij\]), $D_{I,J} - D'_{I,J}$ lies in $(\cV_n)_{(2n)} \cap \cI_{n}$, and since there are no relations in $\cH(n)^{O(n)}$ of degree less than $2n+2$, we have $D_{I,J} - D'_{I,J}=0$. There is a distinguished element $D_0 = D_{(0,\dots,n),(0,\dots,n)}$ which is the unique element of $\cI_n$ of minimal weight $n^2+3n+2$. It is annihilated by all operators $J^{2m}(k)$ for $k>2m+1$, which lower the weight by $k-2m-1$.
Given a homogeneous polynomial $p\in gr(\cV_{n})\cong \mathbb{C}[Q_{a,b}|~0\leq a\leq b]$ of degree $k$ in the variables $Q_{a,b}$, a [*normal ordering*]{} of $p$ will be a choice of normally ordered polynomial $P\in (\cV_{n})_{(2k)}$, obtained by replacing $Q_{a,b}$ by $\Omega_{a,b}$, and replacing ordinary products with iterated Wick products of the form (\[iteratedwick\]). Of course $P$ is not unique, but for any choice of $P$ we have $\phi_{2k}(P) = p$, where $\phi_{2k}: (\cV_{n})_{(2k)} \ra (\cV_{n})_{(2k)} /(\cV_{n})_{(2k-1)} \subset gr(\cV_{n})$ is the usual projection. For the rest of this section, $D^{2k}$, $E^{2k}$, $F^{2k}$, etc., will denote elements of $(\cV_{n})_{(2k)}$ which are homogeneous, normally ordered polynomials of degree $k$ in the vertex operators $\Omega_{a,b}$.
Let $D_{I,J}^{2n+2}\in (\cV_{n})_{(2n+2)}$ be some normal ordering of $d_{I,J}$. Then $$\pi_{n}(D_{I,J}^{2n+2}) \in (\cH(n)^{O(n)})_{(2n)},$$ and $\phi_{2n}(\pi_n(D_{I,J}^{2n+2})) \in gr(\mathcal{H}(n)^{O(n)})$ can be expressed uniquely as a polynomial of degree $n$ in the variables $q_{a,b}$. Choose some normal ordering of the corresponding polynomial in the variables $\Omega_{a,b}$, and call this vertex operator $-D^{2n}_{I,J}$. Then $D^{2n+2}_{I,J} + D^{2n}_{I,J}$ has the property that $\pi_{n}(D^{2n+2}_{I,J} + D^{2n}_{I,J})\in (\cH(n)^{O(n)})_{(2n-2)}.$ Continuing this process, we arrive at a vertex operator $\sum_{k=1}^{n+1} D^{2k}_{I,J}$ in the kernel of $\pi_{n}$. We must have $$\label{decompofd} D_{I,J} = \sum_{k=1}^{n+1}D^{2k}_{I,J},$$ since $D_{I,J}$ is uniquely characterized by (\[uniquedij\]).
In this decomposition, the term $D^2_{I,J}$ lies in the space $A_m$ spanned by $\{\Omega_{a,b}|~a+b=m\}$, for $m = |I| +|J|+2n$. Recall that for $m$ even, $A_m = \partial^2 A_{m-2} \oplus \bra J^{m}\ket$, and for $m$ odd, $A_m = \partial^3 A_{m-3} \oplus \bra \partial J^{m-1}\ket$. For $m$ even (respectively odd), define $pr_m: A_m\ra \bra J^m\ket$ (respectively $pr_m: A_m\ra \bra \partial J^{m-1}\ket$) to be the projection onto the second term. Define the [*remainder*]{} $$\label{defofrij} R_{I,J} = pr_m(D^2_{I,J}).$$
\[uniquenessofr\] Given $D_{I,J}\in\cI_n$ as above, suppose that $D_{I,J} = \sum_{k=1}^{n+1} D^{2k}_{I,J}$ and $D_{I,J} = \sum_{k=1}^{n+1} \tilde{D}^{2k}_{I,J}$ are two different decompositions of $D_{I,J}$ of the form (\[decompofd\]). Then $$D^2_{I,J} - \tilde{D}^2_{I,J} \in \partial^2 (A_{m-2}),$$ where $m = |I| +|J| +2 n$. In particular, $R_{I,J}$ is independent of the choice of decomposition of $D_{I,J}$.
This is analogous to Corollary 4.8 of [@LI], and the proof is almost the same. First, we claim that for all $j,k,l,m\geq 0$, $\Omega_{j,k}\circ_0\Omega_{l,m}$ is a total derivative. In view of the decomposition (\[deca\]) which holds in $\cV_n$ as well as $\cH(n)^{O(n)}$, it suffices to show that $J^{2k}\circ_0 J^{2l}$ is a total derivative for all $k,l$. This is clear because $J^{2k}\circ_0 J^{2l}$ lies in $A_m$ for $m=2k+2l+1$, and $A_m = \partial(A_{m-1})$.
Next, let $\mu = ~:a_1 \cdots a_m:$ be a normally ordered monomial in $(\cV_{n})_{(2m)}$, where each $a_i$ is one of the generators $\Omega_{a,b}$. Let $\tilde{\mu} = :a_{i_1} \cdots a_{i_m}:$, where $(i_1,\dots, i_m)$ is some permutation of $(1,\dots,m)$. We claim that for any decomposition $\mu - \tilde{\mu} = \sum_{k=1}^{m-1} E^{2k}$ of the difference $\mu - \tilde{\mu}\in (\cV_{n})_{(2m-2)}$, the term $E^2$ is a second derivative. To prove this statement, we proceed by induction on $m$. For $m=1$, there is nothing to prove since $\mu - \tilde{\mu} =0$. For $m=2$, and $\mu = ~:\Omega_{a,b}\Omega_{c,d}:$, we have $$\label{reari} \mu - \tilde{\mu} = ~:\Omega_{a,b}\Omega_{c,d}: ~- ~: \Omega_{c,d}\Omega_{a,b}:~ = \sum_{i\geq 0} \frac{(-1)^i}{(i+1)!}\partial^{i+1}(\Omega_{a,b}\circ_{i} \Omega_{c,d}),$$ by (\[vaidii\]). Since $\Omega_{a,b}\circ_{0} \Omega_{c,d}$ is already a total derivative, it follows that $\mu -\tilde{\mu}$ is a second derivative, as claimed. Next, we assume the result for $r\leq m-1$. Since the permutation group on $m$ letters is generated by the transpositions $(i,i+1)$ for $i=1,\dots,m-1$, we may assume without loss of generality that $$\tilde{\mu} = ~:a_1 \cdots a_{i-1} a_{i+1} a_i a_{i+2} \cdots a_m:.$$ If $i>1$, we have $\mu - \tilde{\mu} = ~:a_1 \cdots a_{i-1} f:$, where $f =~ :a_i \cdots a_m:~ -~: a_{i+1}a_i a_{i+2} \cdots a_m$, which lies in $(\cM_{-n})_{(2m-2i+2)}$. Since each term of $f$ has degree at least 2, it follows that $\mu - \tilde{\mu}$ can be expressed in the form $\sum_{k=i}^{m-1}E^{2k}$. Since $i>1$, there is no term of degree $2$. Given any rearrangement $\mu - \tilde{\mu}=\sum_{k=1}^{m-1}F^{2k}$, it follows from our inductive hypothesis that the term $F^2$ is a second derivative.
Suppose next that $i=1$, so that $\tilde{\mu} = ~:a_2 a_1 a_3 \cdots a_m:$. Define $$\nu = \ :(:a_1 a_2:) a_3 \cdots a_m:,\ \ \ \ \ \ \tilde{\nu} = \ :(:a_2 a_1:) a_3 \cdots a_m:,$$ and note that $\nu - \tilde{\nu} = \ : (:a_1 a_2: -: a_2 a_1: )f:$, where $f = \ :a_3 \cdots a_m:$. By (\[reari\]), $:a_1 a_2: - : a_2 a_1:$ is homogeneous of degree 2, so $\nu - \tilde{\nu}$ is a linear combination of monomials of degree $2m-2$. By inductive assumption, any rearrangement $\nu - \tilde{\nu} =\sum_{k=1}^{m-1}F^{2k}$ has the property that $F^2$ is a second derivative.
Next, by (\[vaidi\]), we have $$\label{diffi} \mu - \nu = - \sum_{k\geq 0} \frac{1}{(k+1)!} \bigg( :(\partial^{k+1} a_1) (a_2 \circ_k f): + : (\partial^{k+1} a_2)(a_1\circ_k f):\bigg).$$ Since the operators $\circ_k$ for $k\geq 0$ are homogeneous of degree $-2$, each term appearing in (\[diffi\]) has degree at most $2m-2$. Moreover, $$deg \big(:(\partial^{k+1} a_1) (a_2\circ_k f):\big) = 2+ deg(a_2 \circ_k f),\ \ \ deg \big(:(\partial^{k+1} a_2) (a_1\circ_k f):\big) = 2+ deg(a_1 \circ_k f),$$ so the only way to obtain terms of degree $2$ is for $a_2\circ_k f$ or $a_1\circ_k f$ to be a scalar. This can only happen if $k>0$, in which case we obtain either $\partial^{k+1} a_1$ or $\partial^{k+1} a_2$, which are second derivatives. By inductive assumption, any rearrangement of $\mu - \nu$ can contain only second derivatives in degree 2. Similarly, $\tilde{\mu} - \tilde{\nu}$ has degree at most $2m-2$, and any rearrangement of $\tilde{\mu} - \tilde{\nu}$ can only contain second derivatives in degree 2. Since $\mu - \tilde{\mu} = (\mu - \nu) + (\nu - \tilde{\nu}) + (\tilde{\nu} - \tilde{\mu})$, the claim follows.
An immediate consequence is the following statement. Let $E\in (\cV_{n})_{(2m)}$ be a vertex operator of degree $2m$, and choose a decomposition $$\label{dcnew} E = \sum_{k=1}^m E^{2k},$$ where $E^{2k}$ is a homogeneous, normally ordered polynomial of degree $k$ in the variables $\Omega_{a,b}$. If $E = \sum_{k=1}^m F^{2k}$ is any rearrangement of (\[dcnew\]), i.e., another decomposition of $E$ of the same form, then $E^{2}-F^2$ is a second derivative. Finally, specializing this to the case $E=D_{I,J}$ proves the lemma.
Let $R_0$ denote the remainder of the element $D_0$. The condition $R_0 \neq 0$ is equivalent to the existence of a decoupling relation of the form $j^{n^2+3n} = P(j^0,j^2,\dots, j^{n^2+3n-2})$ in $\cH(n)^{O(n)}$.
Let $D_0 = \sum_{k=1}^{n+1}D^{2k}_0$ be a decomposition of $D_0$ of the form (\[decompofd\]). If $R_0\neq 0$, we have $D^2_0 = \lambda J^{n^2+3n} + \partial^2 \omega$ for some $\lambda \neq 0$ and some $\omega\in A_{n^2+3n-2}$. Applying the projection $\pi_{n}:\cV_{n}\ra \cH(n)^{O(n)}$, since $\pi_{n}(D_0)=0$ we obtain $$j^{n^2+3n}= -\frac{1}{\lambda}\big( \partial^2 \pi_{n}(\omega) + \sum_{k=2}^{n+1} \pi_{n}(D^{2k}_0) \big),$$ which is a decoupling relation of the desired form. The converse follows from the fact that $D_0$ is the unique element of the ideal $\cI_{n}$ of weight $n^2+3n+2$, up to scalar multiples.
\[mainconj\] For all $n\geq 1$, the remainder $R_0$ is nonzero.
For $n=1$, it is easy to check that $R_0 = -\frac{5}{4} J^4$, and in the Appendix, we write down computer calculations that prove this conjecture in the cases $n=2$ and $n=3$. For $n=2$, we have $R_0 = \frac{149}{600} J^{10}$, and for $n=3$ we have $R_0 = -\frac{2419}{705600} J^{18}$. However, for $n>3$ it is difficult to calculate $R_0$ directly. Unlike the case of $\cW_{1+\infty,-n}$ where a similar remainder was shown to be nonzero in [@LI], there seems to be no nice recursive structure that allows us to proceed by induction on $n$. Even in the case $\cW_{1+\infty,-n}$, we still lack a conceptual explanation for this phenomenon, and we expect that such an explanation will be necessary to prove our conjecture for $\cH(n)^{O(n)}$.
The next theorem shows that the strong finite generation of $\cH(n)^{O(n)}$ is an easy consequence of our conjecture.
\[stronggen\] Suppose that Conjecture \[mainconj\] holds. Then for all $r\geq \frac{1}{2}(n^2+3n)$, there exists a decoupling relation $$\label{maindecoup} j^{2r} = Q_{2r}(j^0,j^2,\dots,j^{n^2+3n-2}),$$ where $Q_{2r}$ is some normally ordered polynomial in $j^0,j^2,\dots,j^{n^2+3n-2}$, and their derivatives. It follows that $\{j^0,j^2,\dots,j^{n^2+3n-2}\}$ is a minimal strong generating set for $\cH(n)^{O(n)}$.
The decoupling relation $j^{n^2+3n} = P(j^0,j^2,\dots,j^{n^2+3n-2})$ given by Conjecture \[mainconj\] corresponds to an element $J^{n^2+3n} - P(J^0,J^2,\dots,J^{n^2+3n-2})\in \cI_{n}$. We need to show that for all $r\geq \frac{1}{2}(n^2+3n)$, there exists an element $J^{2r} - Q_{2r}(J^0,J^2,\dots,J^{n^2+3n-2})\in \cI_{n}$, so we assume inductively that $Q_{2r-2}$ exists. Choose a decomposition $$Q_{2r-2} = \sum_{k=1}^{d} Q^{2k}_{2r-2},$$ where $Q_{2r-2}^{2k}$ is a homogeneous normally ordered polynomial of degree $k$ in the vertex operators $J^0,\dots,J^{n^2+3n-2}$ and their derivatives. In particular, $$Q^2_{2r-2} = \sum_{i=0}^{\frac{1}{2}(n^2+3n-2)} c_i \partial^{2r-2i-2} J^{2i},$$ for constants $c_i$. We apply the operator $J^2 \circ_1$, which raises the weight by two. By (\[specialope\]), we have $J^2 \circ_1 J^{2r-2} \equiv (4r-1)J^{2r}$ modulo second derivatives. Moreover, using (\[vaidiii\]) and (\[specialope\]), we see that $J^2 \circ_1 \big(\sum_{k=1}^{d} Q^{2k}_{2r-2}\big)$ can be expressed in the form $\sum_{k=1}^{d} E^{2k}$ where each $E^{2k}$ is a normally ordered polynomial in $J^0,\dots,J^{n^2+3n}$ and their derivatives. If $J^{n^2+3n}$ or its derivatives appear in $E^{2k}$, we can use the element $J^{n^2+3n} -P(J^0,\dots, J^{n^2+3n-2})$ in $\cI_{n}$ to eliminate the variable $J^{n^2+3n}$ and any of its derivatives, modulo $\cI_{n}$. Hence $J^2 \circ_1 \big(\sum_{k=1}^{d} Q^{2k}_{2r-2}\big)$ can be expressed modulo $\cI_{n}$ in the form $\sum_{k=1}^{d'} F^{2k}$, where $d'\geq d$, and $F^{2k}$ is a normally ordered polynomial in $J^0,\dots, J^{n^2+3n-2}$ and their derivatives. It follows that $$\frac{1}{4r-1} J^2 \circ_1 \big(J^{2r-2} - Q_{2r-2}(J^0,\dots, J^{n^2+3n-2})\big)$$ can be expressed as an element of $\cI_{n}$ of the desired form.
Representation theory of $\cH(n)^{O(n)}$
========================================
The basic tool in studying the representation theory of vertex algebras is the [*Zhu functor*]{}, which was introduced by Zhu in [@Z]. Given a vertex algebra $\cW$ with weight grading $\cW = \bigoplus_{n\in\mathbb{Z}} \cW_n$, this functor attaches to $\cW$ an associative algebra $A(\cW)$, together with a surjective linear map $\pi_{Zh}:\cW\ra A(\cW)$. For $a\in \cW_{m}$ and $b\in\cW$, define $$\label{defzhu} a*b = Res_z \bigg (a(z) \frac{(z+1)^{m}}{z}b\bigg),$$ and extend $*$ by linearity to a bilinear operation $\cW\otimes \cW\ra \cW$. Let $O(\cW)$ denote the subspace of $\cW$ spanned by elements of the form $$\label{zhuideal} a\circ b = Res_z \bigg (a(z) \frac{(z+1)^{m}}{z^2}b\bigg)$$ where $a\in \cW_m$, and let $A(\cW)$ be the quotient $\cW/O(\cW)$, with projection $\pi_{Zh}:\cW\ra A(\cW)$. Then $O(\cW)$ is a two-sided ideal in $\cW$ under the product $*$, and $(A(\cW),*)$ is a unital, associative algebra. The assignment $\cW\mapsto A(\cW)$ is functorial, and if $\cI$ is a vertex algebra ideal of $\cW$, we have $A(\cW/\cI)\cong A(\cW)/ I$, where $I = \pi_{Zh}(\cI)$. A well-known formula asserts that for all $a\in \cW_m$ and $b\in \cW$, $$\label{zhucomm} a*b - b*a \equiv Res_z (1+z)^{m -1} a(z) b \ \ \ \ \text{mod} \ O(\cW).$$
A $\mathbb{Z}_{\geq 0}$-graded module $M = \bigoplus_{n\geq 0} M_n$ over $\cW$ is called [*admissible*]{} if for every $a\in\cW_m$, $a(n) M_k \subset M_{m+k -n-1}$, for all $n\in\mathbb{Z}$. Given $a\in\cW_m$, the Fourier mode $a(m-1)$ acts on each $M_k$. The subspace $M_0$ is then a module over $A(\cW)$ with action $[a]\mapsto a(m-1) \in End(M_0)$. In fact, $M\mapsto M_0$ provides a one-to-one correspondence between irreducible, admissible $\cW$-modules and irreducible $A(\cW)$-modules. If $A(\cW)$ is a commutative algebra, all its irreducible modules are one-dimensional, and the corresponding $\cW$-modules $M = \bigoplus_{n\geq 0} M_n$ are cyclic and generated by any nonzero $v\in M_0$. Accordingly, we call such a module a [*highest-weight module*]{} for $\cW$, and we call $v$ a [*highest-weight vector*]{}.
Let $\cW$ be a vertex algebra which is strongly generated by a set of weight-homogeneous elements $\alpha_i$ of weights $w_i$, for $i$ in some index set $I$. Then $A(\cW)$ is generated by $\{ a_i = \pi_{Zh}(\alpha_i(z))|~i\in I\}$. Moreover, $A(\cW)$ inherits a filtration (but not a grading) by weight.
\[abelianzhu\] For all $n\geq 1$, $A(\cH(n)^{O(n)})$ is a commutative algebra.
In the case $n=1$, the Zhu algebra $A(\cH^+)$ is clearly abelian since it has two generators, one of which is central since it corresponds to the Virasoro element [@DNI]. For all $n\geq 1$, the circle products $j^{2l}\circ_k j^{2m}$ among the generators of $\cH(n)^{O(n)}$ are independent of $n$ for $0\leq k<2l+2m+3$. For $k = 2l+2m+3$, $j^{2l}\circ_k j^{2m}$ is a constant (which depends on $n$), and $j^{2l}\circ_k j^{2m} = 0$ for $k> 2l+2m+3$. Let $a^{2m}= \pi_{Zh}(j^{2m})$. It is clear from (\[zhucomm\]) that the commutator $[a^{2l},a^{2m}]$ depends only on $j^{2l}\circ_k j^{2m}$ for $0\leq k\leq 2l+2$, and therefore is not affected by the value of this constant. So this commutator is the same for all $n$, and in particular must vanish since it vanishes for $n=1$. Note that this argument is independent of Conjecture \[mainconj\].
All irreducible, admissible $\cH(n)^{O(n)}$-modules are highest-weight modules.
Since $\cV_n$ has the same generators and OPE relations as $\cH(n)^{O(n)}$, the same argument shows that the Zhu algebra of $\cV_n$ is abelian. Since $\cV_n$ is freely generated by $J^0,J^2,\dots$, it follows that $A(\cV_n)$ is the polynomial algebra $\mathbb{C}[A^0,A^2,\cdots]$, where $A^{2m} = \pi_{Zh}(J^{2m})$. Moreover, $A(\cH(n)^{O(n)}) \cong \mathbb{C}[a^0,a^2,\dots] / I_n$, where $I_n = \pi_{Zh}(\cI_n)$, and we have a commutative diagram
$$\label{commdiag} \begin{array}[c]{ccc}
\cV_n &\stackrel{\pi_n}{\rightarrow}& \cH(n)^{O(n)} \\
\downarrow\scriptstyle{\pi_{Zh}}&&\downarrow\scriptstyle{\pi_{Zh}}\\
A(\cV_n) &\stackrel{A(\pi_n)}{\rightarrow}& A(\cH(n)^{O(n)})
\end{array} .$$
For $n=1$, it was shown by Dong-Nagatomo [@DNI] that $A(\cH^+) \cong \mathbb{C}[x,y] / I$, where $I$ is the ideal generated by the polynomials $$P=(y+x-4x^2)(70y+908x^2-515x+27),\ \ \ \ \ \ Q = (x-1)(x-\frac{1}{16})(x-\frac{9}{16})(y+x-4x^2).$$It follows that the irreducible, admissible $\cH^+$-modules are parametrized by the points on the variety $V(I)\subset \mathbb{C}^2$. Suppose now that Conjecture \[mainconj\] holds. Then $A(\cH(n)^{O(n)})$ is generated by $\{a^0, a^2,\dots,a^{n^2+3n-2}\}$, and it follows that $$A(\cH(n)^{O(n)}) \cong \mathbb{C}[a^0,a^2,\dots, a^{n^2+3n-2}]/ I_{n},$$ where $I_{n}$ is now regarded as an ideal inside $\mathbb{C}[a^0,a^2\dots, a^{n^2+3n-2}]$. The corresponding variety $V(I_{n})\subset \mathbb{C}^{\frac{1}{2}(n^2+3n)}$ then parametrizes the irreducible, admissible modules over $\cH(n)^{O(n)}$. The problem of classifying these modules is equivalent to giving a description of the ideal $I_n$.
Invariant subalgebras of $\cH(n)$ under arbitrary reductive groups {#secgeneral}
==================================================================
In this section, we study $\cH(n)^G$ for a general reductive group $G\subset O(n)$. Our approach is similar to our study of invariant subalgebras of ghost systems in [@LII]. By a fundamental result of Dong-Li-Mason [@DLM], $\cH(n)$ has a decomposition of the form $$\label{dlmdecomp} \cH(n) \cong \bigoplus_{\nu\in H} L(\nu)\otimes M^{\nu},$$ where $H$ indexes the irreducible, finite-dimensional representations $L(\nu)$ of $O(n)$, and the $M^{\nu}$’s are inequivalent, irreducible, highest-weight $\cH(n)^{O(n)}$-modules. The modules $M^{\nu}$ appearing in (\[dlmdecomp\]) have an integrality property; the eigenvalues of $\{j^{2m}(2m+1)|~m\geq 0\}$ on the highest-weight vectors $f_{\nu}$ are all integers. These modules therefore correspond to certain rational points on the variety $V(I_n)$, if it exists.
Using the decomposition (\[dlmdecomp\]), together with a classical theorem of Weyl, we show that $\cH(n)^G$ is finitely generated as a vertex algebra. This statement is analogous to Lemma 2 of [@LII], and is independent of Conjecture \[mainconj\]. We then prove some combinatorial properties of the modules $M^{\nu}$ appearing in (\[dlmdecomp\]), which are also independent of Conjecture \[mainconj\]. Next, we show that Conjecture \[mainconj\] implies that each module $M^{\nu}$ appearing in (\[dlmdecomp\]) possesses a certain finiteness property. Together with the finite generation of $\cH(n)^G$, this is enough to prove that $\cH(n)^G$ is strongly finitely generated. Since Conjecture \[mainconj\] holds for $n=2$ and $n=3$, the strong finite generation of $\cH(2)^G$ and $H(3)^G$ is an immediate consequence.
\[ordfg\] For any reductive $G\subset O(n)$, $\cH(n)^G$ is finitely generated as a vertex algebra.
Recall that $\cH(n)\cong gr(\cH(n))$ as linear spaces, and $$gr(\cH(n)^G )\cong (gr(\cH(n))^G \cong (Sym \bigoplus_{k\geq 0} V_k)^G = R$$ as commutative algebras, where $V_k\cong \mathbb{C}^n$ as $O(n)$-modules. For all $p\geq 0$, there is an action of $GL_p$ on $\bigoplus_{k =0}^{p-1} V_k $ which commutes with the action of $G$. The natural inclusions $GL_p\hookrightarrow GL_q$ for $p<q$ sending $$M \ra \bigg[ \begin{matrix} M & 0 \cr 0 & I_{q-p} \end{matrix} \bigg]$$ induce an action of $GL_{\infty} = \lim_{p\ra \infty} GL_p$ on $\bigoplus_{k\geq 0} V_k$. We obtain an action of $GL_{\infty}$ on $Sym \bigoplus_{k\geq 0} V_k$ by algebra automorphisms, which commutes with the action of $G$. Hence $GL_{\infty}$ acts on $R$ as well. By a basic theorem of Weyl, $R$ is generated by the set of translates under $GL_{\infty}$ of any set of generators for $(Sym \bigoplus_{k = 0} ^{n-1} V_k)^G$ [@W]. Since $G$ is reductive, $(Sym \bigoplus_{k = 0} ^{n-1} V_k)^G$ is finitely generated. Hence there exists a finite set of homogeneous elements $\{f_1,\dots, f_k\}\subset R$ such that $\{ \sigma f_i|~ i=1,\dots,k,~ \sigma\in GL_{\infty}\}$ generates $R$. It follows from Lemma \[recon\] that the set of vertex operators $$\{(\sigma f_i)(z)\in \cH(n)^G|~i=1,\dots,k,~ \sigma\in GL_{\infty}\}$$ which correspond to $\sigma f_i$ under the linear isomorphism $\cH(n)^G\cong gr(\cH(n)^G) \cong R$, is a set of strong generators for $\cH(n)^G$.
In the decomposition (\[dlmdecomp\]) of $\cH(n)$ as a bimodule over $O(n)$ and $\cH(n)^{O(n)}$, the $O(n)$-isotypic component of $\cH(n)$ of type $L(\nu)$ is isomorphic to $L(\nu)\otimes M^{\nu}$. Each $L(\nu)$ is a module over $G\subset O(n)$, and since $G$ is reductive, it has a decomposition $L(\nu) =\oplus_{\mu\in H^{\nu}} L(\nu)_{\mu}$. Here $\mu$ runs over a finite set $H^{\nu}$ of irreducible, finite-dimensional representations $L(\nu)_{\mu}$ of $G$, possibly with multiplicity. We thus obtain a refinement of (\[dlmdecomp\]): $$\label{decompref} \cH(n) \cong \bigoplus_{\nu\in H} \bigoplus_{\mu\in H^{\nu}} L(\nu)_{\mu} \otimes M^{\nu}.$$ Let $f_1(z),\dots,f_k(z)\in \cH(n)^G$ be the vertex operators corresponding to the polynomials $f_1, \dots,f_k$ under the linear isomorphism $\cH(n)^G\cong gr(\cH(n)^G) \cong R$. Clearly $f_1(z),\dots, f_k(z)$ must live in a finite direct sum $$\label{newsummation} \bigoplus_{j=1}^r L(\nu_j)\otimes M^{\nu_j}$$ of the modules appearing in (\[dlmdecomp\]). By enlarging the collection $f_1(z),\dots,f_k(z)$ if necessary, we may assume without loss of generality that each $f_i(z)$ lives in a single representation of the form $L(\nu_j)\otimes M^{\nu_j}$. Moreover, we may assume that $f_i(z)$ lives in a trivial $G$-submodule $L(\nu_j)_{\mu_0} \otimes M^{\nu_j}$, where $\mu_0$ denotes the trivial, one-dimensional $G$-module. (In particular, $L(\nu_j)_{\mu_0}$ is one-dimensional). Since the actions of $GL_{\infty}$ and $O(n)$ on $\cH(n)$ commute, we may assume that $(\sigma f_i)(z)\in L(\nu_j)_{\mu_0}\otimes M^{\nu_j}$ for all $\sigma\in GL_{\infty}$. Since $\cH(n)^G$ is strongly generated by the set $\{ (\sigma f_i)(z)|~i=1,\dots,k,~ \sigma\in GL_{\infty}\}$, and each $M^{\nu_j}$ is an irreducible $\cH(n)^{O(n)}$-module, $\cH(n)^G$ is generated as an algebra over $\cH(n)^{O(n)}$ by $f_1(z),\dots,f_k(z)$. Finally, since $\cH(n)^{O(n)}$ is itself a finitely generated vertex algebra by Lemma \[ordfingen\], we conclude that $\cH(n)^G$ is finitely generated.
Next, we need a fact about representations of associative algebras which can be found in [@LII]. Let $A$ be an associative $\mathbb{C}$-algebra (not necessarily unital), and let $W$ be a linear representation of $A$, via an algebra homomorphism $\rho: A\ra End(W)$. Regarding $A$ as a Lie algebra with commutator as bracket, let $\rho_{Lie}:A\ra End(W)$ denote the map $\rho$, regarded now as a Lie algebra homomorphism. There is an induced algebra homomorphism $U(A)\ra End(W)$, where $U(A)$ denotes the universal enveloping algebra of $A$. Given elements $a,b\in A$, we denote the product in $U(A)$ by $a*b$ to distinguish it from $ab\in A$. Given a monomial $\mu = a_1* \cdots * a_r\in U(A)$, let $\tilde{\mu} = a_1\cdots a_r$ be the corresponding element of $A$. Let $U(A)_+$ denote the augmentation ideal (i. e., the ideal generated by $A$), regarded as an associative algebra with no unit. The map $U(A)_+ \ra A$ sending $\mu\mapsto \tilde{\mu}$ is then an algebra homomorphism which makes the diagram $$\label{commutativediag} \begin{matrix} U(A)_+ & & \cr \downarrow & \searrow \cr A & \ra & End(W) \cr \end{matrix}$$ commute. Let $Sym(W)$ denote the symmetric algebra of $W$, whose $d$th graded component is denoted by $Sym^d(W)$. Clearly $\rho_{Lie}$ (but not $\rho$) can be extended to a Lie algebra homomorphism $\hat{\rho}_{Lie}: A\ra End(Sym(W))$, where $\hat{\rho}_{Lie}(a)$ acts by derivation on each $Sym^d(W)$: $$\hat{\rho}_{Lie}(a)( w_1\cdots w_d) = \sum_{i=1}^d w_1 \cdots \hat{\rho}_{Lie}(a)(w_i) \cdots w_d.$$ This extends to an algebra homomorphism $U(A)\ra End(Sym(W))$ which we also denote by $\hat{\rho}_{Lie}$, but there is no commutative diagram like (\[commutativediag\]) because the map $A\ra End(Sym(W))$ is not a map of associative algebras. In particular, the restrictions of $\hat{\rho}_{Lie}(\mu)$ and $\hat{\rho}_{Lie}(\tilde{\mu})$ to $Sym^d(W)$ are generally not the same for $d>1$. The following result appears in [@LII].
\[first\] Given $\mu \in U(A)$ and $d\geq 1$, define a linear map $\Phi^d_{\mu} \in End(Sym^d(W))$ by $$\label{mapmu} \Phi^d_{\mu} = \hat{\rho}_{Lie}(\mu) \big|_{Sym^d(W)}.$$ Let $E$ denote the subspace of $End(Sym^d(W))$ spanned by $\{\Phi^d_{\mu}|~\mu\in U(A)\}$. Note that $E$ has a filtration $$E_1\subset E_2\subset \cdots,\ \ \ \ \ \ E = \bigcup_{r\geq 1} E_r,$$ where $E_r$ is spanned by $\{\Phi^d_{\mu}|~ \mu \in U(A),~ deg(\mu) \leq r\}$. Then $E = E_d$.
Given a monomial $\mu = a_1* \cdots * a_r \in U(A)$ of arbitrary degree $r>d$, we need to show that $\Phi^d_{\mu}$ can be expressed as a linear combination of elements of the form $\Phi^d_{\nu}$ where $\nu \in U(A)$ and $deg (\nu) \leq d$. Fix $p\leq d$, and let $Part^r_{p}$ denote the set of partitions $\phi$ of $\{1,\dots,r\}$ into $p$ disjoint, non-empty subsets $S^{\phi}_1,\dots, S^{\phi}_p$ whose union is $\{1,\dots,r\}$. Each subset $S^{\phi}_i$ is of the form $$S^{\phi}_i = \{i_1, \dots, i_{k_i} \},\ \ \ \ \ \ i_1<\cdots < i_{k_i}.$$ For $i=1,\dots, p$, let $m_i\in U(A)$ be the corresponding monomial $m_i = a_{i_1} *\cdots *a_{i_{k_i}}$. Let $J = (j_1,\dots,j_p)$ be an (ordered) subset of $\{1,\dots,d\}$. Define a linear map $g_{\phi}\in End(Sym^d(W))$ by $$\label{fphi} g_{\phi}(w_1 \cdots w_d) = \sum_J g_{\phi}^1(w_1) \cdots g_{\phi}^d (w_d),\ \ \ \ \ \ g_{\phi}^k (w_k) = \bigg\{ \begin{matrix} \hat{\rho}_{Lie}(m_i) (w_{j_i}) & k= j_i \cr & \cr w_k & k\neq j_i \end{matrix},$$ where the sum runs over all (ordered) $p$-element subsets $J$ as above. Note that we could replace $m_i\in U(A)$ with $\tilde{m_i}\in A$ in (\[fphi\]), since $\hat{\rho}_{Lie}(m_i)(w_{j_i}) = \hat{\rho}_{Lie}(\tilde{m}_i)(w_{j_i})$.
We claim that for each $\phi\in Part^r_p$, $g_{\phi} \in E_d$. We proceed by induction on $p$. The case $p=1$ is trivial because $g_{\phi} = \hat{\rho}_{Lie}(a)$ as derivations on $Sym^d(W)$, where $a = a_1\cdots a_r$. Next, assume the result for all partitions $\psi\in Part^s_{q}$, for $q<p$ and $s\leq r$. Let $m_1,\dots,m_p\in U(A)$ be the monomials corresponding to $\phi$ as above, and define $m_{\phi} = \tilde{m}_1*\cdots * \tilde{m}_p \in U(A)$. By definition, $\Phi^d_{m_{\phi}} \in E_p\subset E_d$, and the leading term of $\Phi^d_{m_{\phi}}$ is $g_{\phi}$. The lower order terms are of the form $g_{\psi}$, where $\psi\in Part^p_q$ is a partition of $\{1,\dots, p\}$ into $q$ subsets, which each corresponds to a monomial in the variables $\tilde{m}_1,\dots,\tilde{m}_p$. By induction, each of these terms lies in $E_q$, and since $g_{\phi} \equiv \Phi^d_{m_{\phi}}$ modulo $E_q$, the claim is proved.
Finally, using the derivation property of $A$ acting on $Sym^d(W)$, one checks easily that $$\label{triplesum} \Phi^d_{\mu} = \sum_{p=1}^d \sum_{\phi\in Part^r_p} g_{\phi}.$$ Since each $g_{\phi}$ lies in $E_d$ by the above claim, this completes the proof of the lemma.
\[firstcor\] Let $f\in Sym^d(W)$, and let $M\subset Sym^d(W)$ be the cyclic $U(A)$-module generated by $f$. Then $\{\hat{\rho}_{Lie}(\mu)(f)|~ \mu\in U(A),~ deg(\mu)\leq d\}$ spans $M$.
Let $\cL$ denote the Lie algebra generated by the Fourier modes $\{j^{2m}(k)|~k\in\mathbb{Z},~m\geq 0\}$ of the generators of $\cH(n)^{O(n)}$, and let $\cP\subset \cL$ be the subalgebra generated by the annihilation modes $\{j^{2m}(k)|~k\geq 0\}$. Note that $\cP$ has a decomposition $$\cP = \cP_- \oplus \cP_0\oplus \cP_+,$$ where $\cP_-$, $\cP_0$, and $\cP_+$ are the Lie algebras spanned by $\{j^{2m}(k)|~0\leq k< 2m+1\}$, $\{j^{2m}(2m+1)\}$, and $\{j^{2m}(k)|~k>2m+1\}$, respectively. Clearly $\cP$ preserves the filtration on $\cH(n)$, so each element of $\cP$ acts by a derivation of degree zero on $gr(\cH(n))$.
Let $\cM$ be an irreducible, highest-weight $\cH(n)^{O(n)}$-submodule of $\cH(n)$ with generator $f(z)$, and let $\cM'$ denote the $\cP$-submodule of $\cM$ generated by $f(z)$. Since $f(z)$ has minimal weight among elements of $\cM$ and $\cP_+$ lowers weight, $f(z)$ is annihilated by $\cP_+$. Moreover, $\cP_0$ acts diagonalizably on $f(z)$, so $f(z)$ generates a one-dimensional $\cP_0\oplus \cP_+$-module. By the Poincare-Birkhoff-Witt theorem, $\cM'$ is a quotient of $$U(\cP)\otimes_{U(\cP_0\oplus \cP_+)} \mathbb{C} f(z),$$ and in particular is a cyclic $\cP_-$-module with generator $f(z)$. Suppose that $f(z)$ has degree $d$, that is, $f(z)\in \cH(n)_{(d)} \setminus \cH(n)_{(d-1)}$. Since each element of $\cP$ preserves the filtration on $\cH(n)$, and $\cM$ is irreducible, it is easy to see that the nonzero elements of $\cM'$ lie in $\cH(n)_{(d)} \setminus \cH(n)_{(d-1)}$. Therefore, the projection $\cH(n)_{(d)}\ra \cH(n)_{(d)}/\cH(n)_{(d-1)} \subset gr(\cH(n))$ restricts to an isomorphism of $\cP$-modules $$\label{isopmod} \cM'\cong gr(\cM')\subset gr(\cS(V)).$$ By Lemma \[firstcor\], we conclude that $\cM'$ is spanned by elements of the form $$\{j^{2l_1}(k_1)\cdots j^{2l_r}(k_r) f(z) |~ j^{2l_i}(k_i)\in \cP_-,~ r\leq d\}.$$
Next, we need a basic fact from linear algebra. Let $A = (A_{i,j})$, $i,j = 1,\dots,n$ be an $n\times n$-matrix, whose entries $A_{i,j}$ are all positive real numbers. We call $A$ [*totally increasing*]{} if $A_{i,j} \leq A_{i,j+1}$ and $A_{i,j}\leq A_{i+1,j}$ for all $i,j$, and for each $i,j = 1,\dots,n-1$, the $2\times 2$ matrix $$\bigg[ \begin{matrix} A_{i,j} & A_{i,j+1} \cr A_{i+1,j} & A_{i+1,j+1}\end{matrix} \bigg]$$ has positive determinant.
\[totinc\] Any totally increasing matrix is nonsingular.
Let $A$ be such a matrix. First, for $i=1,\dots,n$ we rescale the $i$th row by a factor of $1/A_{i,1}$. Next, for $j=2,\dots,n$, we rescale the $j$th column by a factor of $A_{1,1}/A_{1,j}$ to obtain $$\left[ \begin{matrix}1 & 1 & 1 & \cdot & \cdot & \cdot & 1 \cr 1 & \frac{A_{1,1} A_{2,2}}{A_{2,1}A_{1,2}} & \frac{A_{1,1}A_{2,3}}{A_{2,1}A_{1,3}} & \cdot & \cdot & \cdot & \frac{A_{1,1}A_{2,n}}{A_{2,1}A_{1,n}} \cr \vdots & \vdots & \vdots & & & & \vdots \cr 1 & \frac{A_{1,1}A_{n,2}}{A_{n,1}A_{1,2}} & \frac{A_{1,1} A_{n,3}}{A_{n,1}A_{1,3}} & \cdot & \cdot & \cdot & \frac{A_{1,1} A_{n,n}}{A_{n,1} A_{1,n}} \end{matrix}\right],$$ which is easily seen to be totally increasing. Finally, we subtract the first row from the $i$th row, for $i=2,\cdots,n$, obtaining $$\left[ \begin{matrix}1 & 1 & 1 & \cdot & \cdot & \cdot & 1 \cr 0 & \frac{A_{1,1} A_{2,2}}{A_{2,1}A_{1,2}} -1 & \frac{A_{1,1}A_{2,3}}{A_{2,1}A_{1,3}} -1 & \cdot & \cdot & \cdot & \frac{A_{1,1}A_{2,n}}{A_{2,1}A_{1,n}} -1 \cr \vdots & \vdots & \vdots & & & & \vdots \cr 0 & \frac{A_{1,1}A_{n,2}}{A_{n,1}A_{1,2}} -1 & \frac{A_{1,1} A_{n,3}}{A_{n,1}A_{1,3}} -1 & \cdot & \cdot & \cdot & \frac{A_{1,1} A_{n,n}}{A_{n,1} A_{1,n}} -1 \end{matrix} \right].$$ It is easy to check that the $(n-1)\times (n-1)$ matrix $$\left[ \begin{matrix} \frac{A_{1,1} A_{2,2}}{A_{2,1}A_{1,2}} -1 & \frac{A_{1,1}A_{2,3}}{A_{2,1}A_{1,3}} -1 & \cdot & \cdot & \cdot & \frac{A_{1,1}A_{2,n}}{A_{2,1}A_{1,n}} -1 \cr \vdots & \vdots & & & & \vdots \cr \frac{A_{1,1}A_{n,2}}{A_{n,1}A_{1,2}} -1 & \frac{A_{1,1} A_{n,3}}{A_{n,1}A_{1,3}} -1 & \cdot & \cdot & \cdot & \frac{A_{1,1} A_{n,n}}{A_{n,1} A_{1,n}} -1 \end{matrix} \right]$$ is totally increasing, so the claim follows by induction on $n$.
This lemma allows us to prove a certain useful property of $\cH(n)$ as a module over $\cP$. For simplicity of notation, we take $n=1$, but the result we are going to prove holds for any $n$. In this case, $\cH = \cH(1)$ is generated by $\alpha(z)$. Recall from (\[structureofgrs\]) that $\alpha_j$ denotes the image of $\partial^j\alpha$ in $gr(\cH)$. Let $W\subset gr(\cH)$ be the vector space with basis $\{\alpha_j|~ j\geq 0\}$, and for each $m\geq 0$, let $W_m$ be the subspace with basis $\{\alpha_j|~ 0\leq j\leq m\}$. Let $\phi:W\ra W$ be a linear map of weight $w\geq 1$, such that $$\label{arbmap} \phi(\alpha_j) = c_j \alpha_{j+w},$$ for constants $c_j \in \mathbb{C}$. For example, the restriction $j^{2k}(2k-w+1)\big|_{W}$ of any $j^{2k}(2k-w+1)\in \cP$, is such a map.
\[third\] Fix $w\geq 1$ and $m\geq 0$, and let $\phi$ be a linear map satisfying (\[arbmap\]). Then the restriction $\phi \big|_{W_m}$ can be expressed uniquely as a linear combination of the operators $j^{2k}(2k-w+1)\big|_{W_m}$ for $0\leq 2k+1-w \leq 2m+1$.
Suppose first that $w$ is odd, and let $k_j = j+ \frac{1}{2}(w-1)$, for $j=0,\dots,m$. In this notation, we need to show that $\phi \big|_{W_m}$ can be expressed uniquely as a linear combination of the operators $j^{2k_j}(2j)\big|_{W_m}$ for $j=0,\dots,m$. Using (\[bcalc\]), we calculate $$\label{actionofp} j^{2k_j}(2j)(\alpha_i) = \lambda_{0,2k_j,i,2j} (\alpha_{i+w}) = \bigg(\frac{(2k_j+i+1)!}{(2k_j+i+1-2j)!} + \frac{(i+1)!}{(i+1-2j)!} \bigg)\alpha_{i+w}.$$
Let $M^w$ be the $(m+1)\times(m+1)$ matrix with entries $M^w_{i,j} = \lambda_{0,2k_j,i,2j}$, for $i,j = 0,\dots,m$. Let ${\bf c}$ be the column vector in $\mathbb{C}^{m+1}$ whose transpose is given by $(c_0,\dots,c_m)$. Given an arbitrary linear combination $$\psi = t_0 j^{2k_0}(0) + t_1 j^{2k_1}(2) + \cdots + t_{m} j^{2k_m}(2m)$$ of the operators $j^{2k_j}(2j)$ for $0\leq j\leq m$, let ${\bf t}$ be the column vector whose transpose is $(t_0,\dots, t_{m})$. Note that $\phi \big|_{W_m} = \psi \big|_{W_m}$ precisely when $M^w {\bf t} = {\bf c}$, so in order to prove the claim, it suffices to show that $M^w$ is invertible. It is easy to check using (\[actionofp\]) that $M^w$ is totally increasing, so this is immediate from Lemma \[totinc\]. Finally, if $w$ is even, the same argument shows that for $k_j = j+\frac{w}{2}$, $j=0,\dots,m$, $\phi$ can be expressed uniquely as a linear combination of the operators $j^{2k_j}(2j+1)$ for $j=0,\dots,m$.
Since (\[bcalc\]) holds for any $n\geq 1$, it follows that the statement of Lemma \[third\] holds for any $n$. More precisely, let $W\subset gr(\cH(n))$ be the vector space with basis $\{\alpha^i_j|~i=1,\dots,n,~j\geq 0\}$, and let $W_m\subset W$ be the subspace with basis $\{\alpha^i_j|~i=1,\dots, n,~0\leq j\leq m\}$. Let $\phi:W\ra W$ be a linear map of weight $w\geq1$ taking $$\label{actiongencase} \alpha^i_j\mapsto c_j \alpha^i_{j+w},\ \ \ \ \ \ i=1,\dots,n,$$ where the constants $c_j$ are independent of $i$. For example, each $\phi = j^{2k}(2k-w+1)\big|_W$ satisfies (\[actiongencase\]). Then $\phi\big|_{W_m}$ can be expressed uniquely as a linear combination of $j^{2k}(2k-w+1)\big|_{W_m}$ for $0\leq 2k+1-w \leq 2m+1$.
The next result is analogous to Lemma 7 of [@LII], and the proof is almost identical.
\[fourth\] Let $\cM$ be an irreducible, highest-weight $\cH(n)^{O(n)}$-submodule of $\cH(n)$ with highest-weight vector $f(z)$ of degree $d$. Let $\cM'$ be the corresponding $\cP$-module generated by $f(z)$, and let $f$ be the image of $f(z)$ in $gr(\cH(n))$, which generates $M = gr(\cM')$ as a $\cP$-module. Fix $m$ so that $f\in Sym^d(W_m)$. Then $\cM'$ is spanned by $$\{j^{2l_1}(k_1) \cdots j^{2l_r}(k_r) f(z)|~j^{2l_i}(k_i)\in \cP_-,\ \ r\leq d,\ \ 0\leq k_i \leq 2m+1\}.$$
We may work with $M = gr(\cM')$ rather than $\cM'$, and for notational convenience, we do not distinguish between elements of $U(\cP_-)$ and their images in $End(Sym^d(W))$. As in the proof of Lemma \[first\], let $E$ denote the subspace of $End(Sym^d (W))$ spanned by $U(\cP_-)$, and let $E_r$ be the subspace spanned by elements of $U(\cP_-)$ of degree at most $r$. Let $\tilde{E}_r$ be the subspace of $E_r$ spanned by elements of $U(\cP_-)$ which only depend on $j^{2l}(k)$ for $k\leq 2m+1$.
It is not true that $E_d = \tilde{E}_d$ as subspaces of $End(Sym^d(W))$, but it suffices to show that these spaces of endomorphisms coincide when restricted to $Sym^d(W_m)$. Since $E = E_d$, and hence is spanned by monomials $\mu = a_1*\cdots * a_r\in U(\cP_-)$ of degree $r\leq d$, we have $$\label{triplesumii} \mu = \sum_{p=1}^r \sum_{\phi\in Part^r_p} g_{\phi},$$ where each partition $\phi\in Part^r_p$ corresponds to a set of monomials $m_1,\dots,m_p$, and $g_{\phi}$ is given by (\[fphi\]). For $p=r$, there is only one partition $\phi_0$ of $\{1,\dots,r\}$ into disjoint, non-empty subsets, and $g_{\phi_0}$ is defined on monomials $w_1\cdots w_d\in Sym^d(W)$ by $$\label{fphii} g_{\phi_0}(w_1\cdots w_d) = \sum_J g_{\phi_0}^1(w_1) \cdots g_{\phi_0}^d (w_d),\ \ \ \ \ \ g_{\phi_0}^k (w_k) = \bigg\{ \begin{matrix} a_i (w_{j_i}) & k= j_i \cr & \cr w_k & k\neq j_i\end{matrix},$$ where the sum runs over all (ordered) $r$-element subsets $J \subset \{1,\dots,d\}$. By Lemma \[third\], the restriction of $a_i$ to $W_m$ coincides with a linear combination $S_i$ of the elements $j^{2l}(k)\big|_{W_m}$ for $k\leq 2m+1$. Replace each of the factors $a_i (w_{j_i})$ appearing in (\[fphii\]) with $S_i (w_{j_i})$, and let $Q = \prod_{i=1}^r S_i$, which lies in $U(\cP_-)$, and depends only on $j^{2l}(k)$ for $k\leq 2m+1$. Clearly the restriction of $Q$ to $Sym^d(W_m)$ agrees with the restriction of $\mu$ to $Sym^d(W_m)$, modulo terms lying in $E_{r-1}$. The lemma then follows by induction on $r$.
As in [@LII], we may order the elements $j^{2l}(k)\in \cP_-$ as follows: $j^{2l_1}(k_1) > j^{2l_2}(k_2)$ if $l_1>l_2$, or $l_1=l_2$ and $k_1<k_2$. Then Lemma \[fourth\] can be strengthened as follows: $\cM'$ is spanned by elements of the form $j^{2l_1}(k_1)\cdots j^{2l_r}(k_r) f(z)$ with $$\label{shapealpha} j^{2l_i}(k_i)\in \cP_-,\ \ \ \ r\leq d,\ \ \ \ 0\leq k_i\leq 2m+1,\ \ \ \ j^{2l_1}(k_1)\geq \cdots \geq j^{2l_r}(k_r).$$
Up to this point, everything we have proven in this section is independent of Conjecture \[mainconj\]. Then next lemma is where this assumption will enter. We use the notation $\cH(n)^{O(n)}[k]$, $\cM[k]$, and $\cM'[k]$ to denote the homogeneous components of these spaces of conformal weight $k$. As in [@LII], we define the [*Wick ideal*]{} $\cM_{Wick}\subset \cM$ to be the subspace spanned by elements of the form $$:a(z) b(z):,\ \ \ \ \ a(z)\in \bigoplus_{k>0} \cH(n)^{O(n)}[k],\ \ \ \ \ b(z)\in \cM.$$ Despite the choice of terminology, $\cM_{Wick}$ is not a vertex algebra ideal. It is properly contained in the space $C_1(\cM)$ defined in [@LiIII], and in particular it does not contain all elements of the form $L\circ_0 b = \partial b$ for $b\in\cM$.
\[fifth\] Let $\cM$ be an irreducible, highest-weight $\cH(n)^{O(n)}$-submodule of $\cH(n)$ with highest-weight vector $f(z)$. If Conjecture \[mainconj\] holds, any homogeneous element of $\cM$ of sufficiently high weight lies in the Wick ideal. In particular, $\cM / \cM_{Wick}$ is finite-dimensional.
It suffices to show that $\cM'[k]$ lies in the Wick ideal for $k>>0$, where $\cM'$ is the $\cP$-module generated by $f(z)$. As usual, let $d$ be the degree of $f(z)$, and fix $m$ so that $f \in Sym^d(W_m)$. Recall that $\cM'$ is spanned by elements of the form $j^{2l_1}(k_1)\cdots j^{2l_r}(k_r) f(z)$ satisfying (\[shapealpha\]). Fix an element $\alpha(z)$ of this form of weight $K>>0$. Since each operator $j^{2l_i}(k_i)$ has weight $2l_i+1-k_i$, $k_i\leq 2m+1$, and $K>>0$, we may assume that $l_1>>\frac{1}{2}(n^2+3n)$. Then the decoupling relation (\[maindecoup\]) allows us to express $j^{2l_1}(z)$ as a normally ordered polynomial $Q_{l_1}(z)$ in the generators $$\label{genera} \partial^t j^{2l}(z),\ \ \ \ \ 0\leq l\leq \frac{1}{2}(n^2+3n-2),\ \ \ \ \ t\geq 0.$$ We claim that for any weight-homogeneous, normally ordered polynomial $Q(z)$ in the generators (\[genera\]) of sufficiently high weight, any element $c(z)\in \cM$, and any $k$ satisfying $0\leq k\leq 2m+1$, $Q(z)\circ_k c(z)$ lies in $\cM_{Wick}$. Specializing this to the case $Q(z) = Q_{l_1}(z)$, $c(z) = j^{2l_2}(k_2)\cdots j^{2l_r}(k_r) f(z)$, and $k=k_1$, proves the lemma.
We may assume without loss of generality that $Q(z)=:a(z)b(z):$ where $a(z) = \partial^t j^{2l}(z)$ for some $0\leq l\leq \frac{1}{2}(n^2 +3n-2)$. Then using (\[vaidiv\]), and suppressing the formal variable $z$, we have $$\label{appvaid} Q\circ_k c = \big(:ab:\big) \circ_{k} c = \sum_{r\geq0}{1\over r!}:(\partial^r a)(b\circ_{k+r}c):
+\sum_{r\geq 0}b\circ_{k-r-1}(a\circ_r c) .$$ Suppose first that $b = \lambda 1$ for some constant $\lambda$. Then $Q= \lambda \partial^t j^{2l}$, and since $wt(Q) >>0$, we have $t>>0$. Hence $Q\circ_k= \lambda(\partial^t j^{2l})\circ_k = 0$ as an operator (since this operator vanishes whenever $t>k$). So we may assume without loss of generality that $b$ is not a constant.
We proceed by induction on $k$. For $k=0$, each term appearing in (\[appvaid\]) lies in $\cM_{Wick}$, so there is nothing to prove. For $k>0$, the only terms appearing in (\[appvaid\]) that need not lie in $\cM_{Wick}$ a priori, are those of the form $\sum_{r=0}^{k-1} b\circ_{k-r-1}(a\circ_r c)$. However, each of these terms is weight-homogeneous, and the weight of $a\circ_r c = \partial^t j^{2l} \circ_r c$ is bounded above by $wt(c) + n^2+3n+1$, since $\partial^t j^{2l} \circ_r c=0$ for $t>r$. So we may still assume that $wt(b)>>0$. By our inductive assumption, all these terms then lie in $\cM_{Wick}$.
\[sixth\] Let $\cM$ be an irreducible, highest-weight $\cH(n)^{O(n)}$-submodule of $\cH(n)$. Given a subset $S\subset \cM$, let $\cM_S\subset \cM$ denote the subspace spanned by elements of the form $$:\omega_1(z)\cdots \omega_t(z) \alpha(z):,\ \ \ \ \ \omega_j(z)\in \cH(n)^{O(n)},\ \ \ \ \ \alpha(z)\in S.$$ If Conjecture \[mainconj\] holds, there exists a finite set $S\subset \cM$ such that $\cM = \cM_S$.
Now we are ready to prove our main result.
\[sfg\] Suppose that Conjecture \[mainconj\] holds. Then for any reductive group $G$ of automorphisms of $\cH(n)$ preserving the conformal structure (\[virasoro\]), $\cH(n)^G$ is strongly finitely generated.
By Theorem \[ordfg\], we can find vertex operators $f_1(z),\dots, f_k(z)$ such that the corresponding polynomials $f_1,\dots, f_k\in gr(\cH(n))^G$, together with all $GL_{\infty}$ translates of $f_1,\dots, f_k$, generate the invariant ring $gr(\cH(n))^G$. As in the proof of Lemma \[ordfg\], we may assume that each $f_i(z)$ lies in an irreducible, highest-weight $\cH(n)^{O(n)}$-module $\cM_i$ of the form $L(\nu)_{\mu_0}\otimes M^{\nu}$, where $L(\nu)_{\mu_0}$ is a trivial, one-dimensional $G$-module. Furthermore, we may assume without loss of generality that $f_1(z),\dots, f_k(z)$ are highest-weight vectors for the action of $\cH(n)^{O(n)}$. Otherwise, we can replace these with the highest-weight vectors in the corresponding modules.
For each $\cM_i$, choose a finite set $S_i \subset \cM_i$ such that $\cM_i = (\cM_i)_{S_i}$, using Corollary \[sixth\]. Define $$S=\{j^0(z),j^2(z),\dots, j^{n^2+3n-2}(z) \} \cup \big(\bigcup_{i=1}^k S_i \big).$$ Since $\{j^0(z),j^2(z),\dots, j^{n^2+3n-2}(z)\}$ strongly generates $\cH(n)^{O(n)}$ (assuming Conjecture \[mainconj\]), and the set $\bigcup_{i=1}^k \cM_i$ strongly generates $\cH(n)^G$, it is immediate that $S$ is a strong, finite generating set for $\cH(n)^G$.
We include one more secondary result in this section, which is independent of Conjecture \[mainconj\]. Recall that $\cH(n)^{O(n)}$ is the quotient of $\cV_n$ by the ideal $\cI_n$, which is generated by the set $\{D_{I,J}\}$, where $I,J$ satisfy (\[ijineq\]). We will show that $\cI_n$ is finitely generated as a vertex algebra ideal. Define $$U_n = (\cV_{n})_{(2n+2)}\cap \cI_{n},$$ which is just the vector space spanned by $\{D_{I,J}\}$, where $I,J$ satisfy (\[ijineq\]). We have $D_{I,J} = D_{J,I}$, but there are no other linear relations among these elements. It is easy to see that $U_n$ is a module over the Lie algebra $\cP$ generated by $\{j^{2m}(k) = j^{2m}\circ_k |~k,m\geq 0\}$, since the action $\cP$ preserves both the filtration degree and the ideal $\cI_n$. Note that $\cP$ has an alternative generating set $\{\Omega_{a,b}\circ_{a+b+1-w}|~ 0\leq a\leq b,~ a+b+1-w\geq 0\}$, where $\Omega_{a,b}\circ_{a+b+1-w}$ is homogeneous of weight $w$. Recall that $gr(\cV_n)$ is the polynomial algebra generated by $\alpha^i_k$ for $i=1,\dots,n$ and $k\geq 0$. The action of $\cP$ by derivations of degree zero on $gr(\cV_{n})$ coming from the vertex Poisson algebra structure is independent of $n$, and is specified by (\[bcalc\]). Using this formula, it is easy to see that the action of $\cP$ on $U_n$ is by weighted derivation" in the following sense. Fix $I = (i_0,\dots,i_n)$ and $J = (j_0,\dots,j_n)$, and let $D_{I,J}$ be the corresponding element of $U_n$. Given $p = \Omega_{a,b}\circ_{a+b+1-w}\in \cP$, we have $$\label{paraction} p(D_{I,J}) = \sum_{r=0}^n c_r D_{I^r,J} + \sum_{r=0}^n d_r D_{I,J^r},$$ for lists $I^r = (i_0,\dots, i_{r-1}, i_r + w,i_{r+1},\dots, i_n)$ and $J^r = (j_0,\dots, j_{r-1}, j_r+ w,j_{r+1},\dots, j_n)$, and constants $c_r,d_r$. If $i_r + w$ appears elsewhere on the list $I^r$, $c_r = 0$, and if $j_r + w$ appears elsewhere on the list $J^r$, $d_r = 0$. Otherwise, $$\label{actioni} c_r = \pm \lambda_{a,b,i_r,t},\ \ \ \ \ \ d_r = \pm \lambda_{a,b,j_r,t},$$ where $t= a+b+1-w$, and the signs $\pm$ are the signs of the permutations transforming $I^r$ and $J^r$ into lists in increasing order, as in (\[ijineq\]).
$\cI_n$ is generated as a vertex algebra ideal by the set of elements $D_{I,J} \in U_n$ for which $|I| + |J| \leq 2n^2+3n$.
Let $\cI'_n$ denote the ideal in $\cV_n$ generated by $\{D_{I,J}|~ |I| + |J| \leq 2n^2+3n\}$. Since $U_n$ generates $\cI_n$, $\cI'_n$ is properly contained in $\cI_n$ if and only if there exists some $D_{I,J}\in \cI_n \setminus \cI'_n$. Suppose that $\cI_n \setminus \cI'_n$ is nonempty, and let $D_{I,J}$ be an element of this form of minimal weight $d$, lying in $\cI_n \setminus \cI'_n$.
Note that elements $D_{I,J}$ satisfying $|I|+|J| \leq 2n^2+3n$ have weight at most $2n^2+5n+2$. This choice guarantees that all elements $D_{I,J}$ for which the union of $I$ and $J$ is $\{0,1,\dots,n,n+1,n+2,\dots,2n+1\}$, must lie in $\cI'_n$. We say that $D_{I,J}$ has a [*hole*]{} at some integer $k\geq 0$ if $k$ does not appear in either $I$ or $J$. Since $d = wt(D_{I,J}) >2n^2+5n+2$, it follows that $D_{I,J}$ has a hole for some $k$ satisfying $0\leq k\leq 2n+1$, and also that there is some $l>2n+1$ such $D_{I,J}$ does [*not*]{} have a hole at $l$. Without loss of generality, we may assume that $l$ appears in $I$. Let $w = l-k$, and fix an integer $m$ greater than all entries of both $I$ and $J$.
By Lemma \[third\], we can choose an element $p\in \cP$ which is a linear combination of the operators $\Omega_{a,b} \circ_{a+b+1-w}$, for which $p(\alpha^i_k) = \alpha^i_l$ and $p(\alpha^i_r) = 0$ for all $i=1,\dots, n$ and all $r\neq k$ satisfying $0\leq r\leq m$. Let $I'$ be the list obtained from $I$ by replacing $l$ by $k$, and let $D_{I',J}$ be the corresponding element of $U_n$. Since $D_{I',J}$ has weight $d-w$, it lies in $\cI'_n$ by inductive assumption, and we clearly have $p(D_{I',J}) = D_{I,J}$. This shows that $D_{I,J}$ lies in $\cI'_n$ as well.
Appendix
========
In this Appendix, we present the results of computer calculations that prove Conjecture \[mainconj\] in the cases $n=2$ and $n=3$. These calculations were done using Kris Thielemann’s OPE package for Mathematica [@T]. For $n=2$, define $$D^6_0 =:\Omega_{0,0} \Omega_{1,1} \Omega_{2,2}: - :\Omega_{0,2} \Omega_{0,2} \Omega_{1,1}: + 2 :\Omega_{0,1} \Omega_{0,2} \Omega_{1,2}: - :\Omega_{0,0} \Omega_{1,2}\Omega_{1,2}: - :\Omega_{0,1} \Omega_{0,1} \Omega_{2,2}:,$$
$$D^4_0 = - 1/30 :\Omega_{0,0} \Omega_{1,7}: + 3/4 :\Omega_{0,0} \Omega_{2,6}: - 5/6: \Omega_{0,1} \Omega_{1,6}: - 13/15 :\Omega_{0,1} \Omega_{2,5}:$$ $$+ 61/30 : \Omega_{0,2} \Omega_{1,5}: - 1/6 : \Omega_{0,2} \Omega_{2,4}: + 5/12 :\Omega_{1,1} \Omega_{0,6}: - 1/4 :\Omega_{1,1} \Omega_{2,4}:$$ $$- 1/10 :\Omega_{1,2} \Omega_{0,5}: + :\Omega_{1,2} \Omega_{1,4}: + :\Omega_{1,2} \Omega_{2,3}: + 1/12 :\Omega_{0,4} \Omega_{2,2}: - 7/6 :\Omega_{1,3} \Omega_{2,2}: ,$$
$$D^2_0 = \frac{149}{600} \Omega_{0,10} +\partial^2 \bigg(-
\frac{1697}{10800} \Omega_{0,8} + \frac{1637}{18900} \Omega_{1,7} - \frac{33241}{37800} \Omega_{2,6} +
\frac{16117}{9450} \Omega_{3,5}- \frac{4223}{3780} \Omega_{4,4}\bigg) .$$
A computer calculation shows that the corresponding normally ordered polynomial in the variables $\omega_{a,b}$ is identically zero, so $D^6_0 + D^4_0 + D^2_0$ lies in $Ker(\pi_2)$, and hence must coincide with $D_0$. It follows that $R_0 = \frac{149}{600} J^{10}$, and in particular is nonzero.
Similarly, in the case $n=3$, define $$D^8_0 = :\Omega_{0,3} \Omega_{0,3} \Omega_{1,2} \Omega_{1,2}: - 2 :\Omega_{0,2}\Omega_{0,3} \Omega_{1,2} \Omega_{1,3} : +
:\Omega_{0,2}\Omega_{0,2} \Omega_{1,3} \Omega_{1,3}:$$ $$- :\Omega_{0,3} \Omega_{0,3} \Omega_{1,1} \Omega_{2,2}: +
2 :\Omega_{0,1} \Omega_{0,3} \Omega_{1,3} \Omega_{2,2}: - :\Omega_{0,0}\Omega_{1,3} \Omega_{1,3} \Omega_{2,2}:$$ $$+
2 :\Omega_{0,2}\Omega_{0,3}\Omega_{1,1}\Omega_{2,3}: - 2 :\Omega_{0,1}\Omega_{0,3}\Omega_{1,2}\Omega_{2,3}: -
2 :\Omega_{0,1}\Omega_{0,2}\Omega_{1,3}\Omega_{2,3}:$$ $$+ 2 :\Omega_{0,0}\Omega_{1,2}\Omega_{1,3}\Omega_{2,3}: +
:\Omega_{0,1}\Omega_{0,1}\Omega_{2,3}\Omega_{2,3}: - :\Omega_{0,0}\Omega_{1,1}\Omega_{2,3}\Omega_{2,3}:$$ $$-
:\Omega_{0,2}\Omega_{0,2}\Omega_{1,1}\Omega_{3,3}: + 2 :\Omega_{0,1}\Omega_{0,2}\Omega_{1,2}\Omega_{3,3}: -
:\Omega_{0,0}\Omega_{1,2}\Omega_{1,2}\Omega_{3,3}:$$ $$- :\Omega_{0,1}\Omega_{0,1}\Omega_{2,2}\Omega_{3,3}: +
:\Omega_{0,0}\Omega_{1,1}\Omega_{2,2}\Omega_{3,3}:,$$
$$D^6_0 = 1/56 :\Omega_{2,10}\Omega_{1,1}\Omega_{0,0}: - 23/42 :\Omega_{3,9}\Omega_{1,1}\Omega_{0,0}: -
1/56 :\Omega_{2,10}\Omega_{0,1}\Omega_{0,1}: + 23/42 :\Omega_{3,9}\Omega_{0,1}\Omega_{0,1}:$$ $$+
7/12 :\Omega_{2,9}\Omega_{1,2}\Omega_{0,0}: + 62/105 :\Omega_{3,8}\Omega_{1,2}\Omega_{0,0}:-
7/12 :\Omega_{2,9}\Omega_{0,2}\Omega_{0,1}:- 62/105 :\Omega_{3,8}\Omega_{0,2}\Omega_{0,1}:$$ $$-
139/105 :\Omega_{2,8}\Omega_{1,3}\Omega_{0,0}:+ 1/15 :\Omega_{37}\Omega_{13}\Omega_{00}: -
7/24 :\Omega_{1,9}\Omega_{2,2}\Omega_{0,0}: + 1/4 :\Omega_{3,7}\Omega_{2,2}\Omega_{0,0}:$$ $$+
139/105 :\Omega_{2,8}\Omega_{0,3}\Omega_{0,1}: - 1/15 :\Omega_{3,7}\Omega_{0,3}\Omega_{0,1}: +
3/20 :\Omega_{2,8}\Omega_{1,2}\Omega_{0,1}: - 81/70 :\Omega_{3,7}\Omega_{1,2}\Omega_{0,1}:$$ $$+
7/24 :\Omega_{1,9}\Omega_{0,2}\Omega_{0,2}: - 1/4 :\Omega_{3,7}\Omega_{0,2}\Omega_{0,2}: -
3/20 :\Omega_{2,8}\Omega_{0,2}\Omega_{1,1}: + 81/70 :\Omega_{3,7}\Omega_{0,2}\Omega_{1,1}:$$ $$+
1/21 :\Omega_{1,8}\Omega_{2,3}\Omega_{0,0}: - 2/3 :\Omega_{2,7}\Omega_{2,3}\Omega_{0,0}: -
7/10 :\Omega_{3,6}\Omega_{2,3}\Omega_{0,0}: + 9/35 :\Omega_{2,7}\Omega_{1,3}\Omega_{0,1}:$$ $$+
5/6 :\Omega_{3,6}\Omega_{1,3}\Omega_{0,1}: - 3/20 :\Omega_{1,8}\Omega_{2,2}\Omega_{0,1}: -
1/5 :\Omega_{3,6}\Omega_{2,2}\Omega_{0,1}: - 1/21 :\Omega_{1,8}\Omega_{0,3}\Omega_{0,2}:$$ $$+
2/3 :\Omega_{2,7}\Omega_{0,3}\Omega_{0,2}: + 7/10 :\Omega_{3,6}\Omega_{0,3}\Omega_{0,2}:+
3/20 :\Omega_{1,8}\Omega_{1,2}\Omega_{0,2}: + 1/5 :\Omega_{3,6}\Omega_{1,2}\Omega_{0,2}:$$ $$-
9/35 :\Omega_{2,7}\Omega_{0,3}\Omega_{1,1}: - 5/6 :\Omega_{3,6}\Omega_{0,3}\Omega_{1,1}: -
1/30 :\Omega_{1,7}\Omega_{3,3}\Omega_{0,0}: + 3/4 :\Omega_{2,6}\Omega_{3,3}\Omega_{0,0}:$$ $$+
9/10 :\Omega_{1,7}\Omega_{2,3}\Omega_{0,1}: + 1/5 :\Omega_{2,6}\Omega_{2,3}\Omega_{0,1}: +
13/15 :\Omega_{3,5}\Omega_{2,3}\Omega_{0,1}:- 81/70 :\Omega_{1,7}\Omega_{1,3}\Omega_{0,2}:$$ $$+
2/5 :\Omega_{2,6}\Omega_{1,3}\Omega_{0,2}: - 61/30 :\Omega_{3,5}\Omega_{1,3}\Omega_{0,2}:+
3/40 :\Omega_{0,8}\Omega_{2,2}\Omega_{1,1}: - 1/2 :\Omega_{3,5}\Omega_{2,2}\Omega_{1,1}:$$ $$+
1/30 :\Omega_{1,7}\Omega_{0,3}\Omega_{0,3}: - 3/4 :\Omega_{2,6}\Omega_{0,3}\Omega_{0,3}: +
9/35 :\Omega_{1,7}\Omega_{1,2}\Omega_{0,3}: - 3/5 :\Omega_{2,6}\Omega_{1,2}\Omega_{0,3}:$$ $$+
7/6 :\Omega_{3,5}\Omega_{1,2}\Omega_{0,3}: - 3/40 :\Omega_{0,8}\Omega_{1,2}\Omega_{1,2}:+
1/2 :\Omega_{3,5}\Omega_{1,2}\Omega_{1,2}: - 5/6 :\Omega_{1,6}\Omega_{3,3}\Omega_{0,1}:$$ $$-
13/15 :\Omega_{2,5}\Omega_{3,3}\Omega_{0,1}: - 3/5 :\Omega_{1,6}\Omega_{2,3}\Omega_{0,2}: +
1/6 :\Omega_{3,4}\Omega_{2,3}\Omega_{0,2}: - 24/35 :\Omega_{0,7}\Omega_{2,3}\Omega_{1,1}:$$ $$+
4/5 :\Omega_{2,5}\Omega_{2,3}\Omega_{1,1}: - 1/2 :\Omega_{3,4}\Omega_{2,3}\Omega_{1,1}: +
5/6 :\Omega_{1,6}\Omega_{1,3}\Omega_{0,3}: + 13/15 :\Omega_{2,5}\Omega_{1,3}\Omega_{0,3}:$$ $$+
3/5 :\Omega_{1,6}\Omega_{2,2}\Omega_{0,3}: - 1/6 :\Omega_{3,4}\Omega_{2,2}\Omega_{0,3}: +
24/35 :\Omega_{0,7}\Omega_{1,3}\Omega_{1,2}: - 4/5 :\Omega_{2,5}\Omega_{1,3}\Omega_{1,2}:$$ $$+
1/2 :\Omega_{3,4}\Omega_{1,3}\Omega_{1,2}: + 61/30 :\Omega_{1,5}\Omega_{3,3}\Omega_{0,2}: -
1/6 :\Omega_{2,4}\Omega_{3,3}\Omega_{0,2}: + 5/12 :\Omega_{3,3}\Omega_{0,6}\Omega_{1,1}:$$ $$-
1/4 :\Omega_{2,4}\Omega_{3,3}\Omega_{1,1}: - 61/30 :\Omega_{1,5}\Omega_{2,3}\Omega_{0,3}: +
1/6 :\Omega_{2,4}\Omega_{2,3}\Omega_{0,3}:- 1/10 :\Omega_{3,3}\Omega_{0,5}\Omega_{1,2}:$$ $$+ :\Omega_{3,3}\Omega_{1,4}\Omega_{1,2}: + 1/15 :\Omega_{0,6}\Omega_{2,3}\Omega_{1,2}: - 4/5 :\Omega_{1,5}\Omega_{2,3}\Omega_{1,2}:+ 1/12 :\Omega_{3,3}\Omega_{2,2}\Omega_{0,4}:$$ $$- 5/12 :\Omega_{0,6}\Omega_{1,3}\Omega_{1,3}:+ 1/4 :\Omega_{2,4}\Omega_{1,3}\Omega_{1,3}: - 1/15 :\Omega_{0,6}\Omega_{2,2}\Omega_{1,3}: + 4/5 :\Omega_{1,5}\Omega_{2,2}\Omega_{1,3}:$$ $$- 1/6 :\Omega_{3,3}\Omega_{2,2}\Omega_{1,3}: -
1/12 :\Omega_{2,3}\Omega_{2,3}\Omega_{0,4}: + 1/10 :\Omega_{2,3}\Omega_{0,5}\Omega_{1,3}:-
:\Omega_{1,4}\Omega_{2,3}\Omega_{1,3}:$$ $$+ 1/6 :\Omega_{2,3}\Omega_{2,3}\Omega_{1,3}:,$$
$$D^4_0 = \frac{451}{114660} :\Omega_{0,0}\Omega_{1,15}: + \frac{6251}{102960} :\Omega_{0,0}\Omega_{2,14}: +
\frac{10261}{69300} :\Omega_{0,0}\Omega_{3,13}: - \frac{1}{140} :\Omega_{0,0}\Omega_{6,10}:$$ $$+
\frac{137}{840} :\Omega_{0,0}\Omega_{7,9}: + \frac{4849}{22050} :\Omega_{0,0}\Omega_{8,8}: -
\frac{451}{114660} :\Omega_{0,1}\Omega_{0,15}: + \frac{96}{2695} :\Omega_{0,1}\Omega_{1,14}:$$ $$-
\frac{1857}{28600} :\Omega_{0,1}\Omega_{2,13}: - \frac{467}{2475} :\Omega_{0,1}\Omega_{3,12}:-
\frac{1}{560} :\Omega_{0,1}\Omega_{5,10}: - \frac{347}{1680} :\Omega_{0,1}\Omega_{6,9}:$$ $$-
\frac{643}{9800} :\Omega_{0,1}\Omega_{7,8}: + \frac{719}{15015} :\Omega_{0,2}\Omega_{0,14}: +
\frac{24929}{1201200} :\Omega_{0,2}\Omega_{1,13}: + \frac{155}{1584} :\Omega_{0,2}\Omega_{2,12}:$$ $$- \frac{1427}{9900} :\Omega_{0,2}\Omega_{3,11}:+ \frac{203}{720} :\Omega_{0,2}\Omega_{5,9}: +
\frac{209}{8400} :\Omega_{0,2}\Omega_{6,8}: - \frac{1}{10} :\Omega_{0,2}\Omega_{7,7}:$$ $$+ \frac{59}{1470} :\Omega_{1,1}\Omega_{0,14}: + \frac{811}{6300} :\Omega_{1,1}\Omega_{2,12}:+ \frac{191}{1260} :\Omega_{1,1}\Omega_{3,11}:+ \frac{1}{112} :\Omega_{1,1}\Omega_{4,10}:$$ $$- \frac{23}{105} :\Omega_{1,1}\Omega_{5,9}: + \frac{5}{48} :\Omega_{1,1}\Omega_{6,8}: + \frac{1809}{9800} :\Omega_{1,1}\Omega_{7,7}: - \frac{4271}{69300} :\Omega_{0,3}\Omega_{0,13}: +$$ $$\frac{118}{4725} :\Omega_{0,3}\Omega_{1,12}: + \frac{22669}{118800} :\Omega_{0,3}\Omega_{2,11}: +
\frac{1279}{10800} :\Omega_{0,3}\Omega_{3,10}: + \frac{257}{6300} :\Omega_{0,3}\Omega_{5,8}:$$ $$+
\frac{1537}{4200} :\Omega_{0,3}\Omega_{6,7}: + \frac{2467}{46200} :\Omega_{1,2}\Omega_{0,13}: +
\frac{2431}{25200} :\Omega_{1,2}\Omega_{1,12}:+ \frac{591}{2200} :\Omega_{1,2}\Omega_{2,11}:$$ $$+
\frac{3523}{6300} :\Omega_{1,2}\Omega_{3,10}: + \frac{5}{8} :\Omega_{1,2}\Omega_{4,9}: +
\frac{2383}{8400} :\Omega_{1,2}\Omega_{5,8}: + \frac{3}{700} :\Omega_{1,2}\Omega_{6,7}:$$ $$+ \frac{7}{96} :\Omega_{0,4}\Omega_{2,10}: -
\frac{367}{504} :\Omega_{0,4}\Omega_{3,9}: + \frac{1019}{18900} :\Omega_{1,3}\Omega_{0,12}: -
\frac{11}{1260} :\Omega_{1,3}\Omega_{1,11}:$$ $$- \frac{4399}{3600} :\Omega_{1,3}\Omega_{2,10}: - \frac{5851}{7560} :\Omega_{1,3}\Omega_{3,9}: - \frac{139}{210} :\Omega_{1,3}\Omega_{4,8}: -
\frac{2447}{4200} :\Omega_{1,3}\Omega_{5,7}:$$ $$+ \frac{1}{6} :\Omega_{1,3}\Omega_{6,6}: - \frac{155}{2376} :\Omega_{2,2}\Omega_{0,12}: -
\frac{283}{1100} :\Omega_{2,2}\Omega_{1,11}: + \frac{271}{1080} :\Omega_{2,2}\Omega_{3,9}:$$ $$+ \frac{1}{48} :\Omega_{2,2}\Omega_{4,8}: +
\frac{1}{10} :\Omega_{2,2}\Omega_{5,7}:+ \frac{7}{150} :\Omega_{2,2}\Omega_{6,6}: - \frac{17}{560} :\Omega_{0,5}\Omega_{1,10}:$$ $$+
\frac{83}{240} :\Omega_{0,5}\Omega_{2,9}: + \frac{799}{2100} :\Omega_{0,5}\Omega_{3,8}: + \frac{17}{56} :\Omega_{1,4}\Omega_{1,10}: +
\frac{17}{16} :\Omega_{1,4}\Omega_{2,9}:$$ $$+ \frac{27}{35} :\Omega_{1,4}\Omega_{3,8}:- \frac{169}{5400} :\Omega_{2,3}\Omega_{0,11}: +
\frac{1021}{3150} :\Omega_{2,3}\Omega_{1,10}: - \frac{163}{540} :\Omega_{2,3}\Omega_{2,9}:$$ $$-
\frac{2539}{2520} :\Omega_{2,3}\Omega_{3,8}:- \frac{1}{4} :\Omega_{2,3}\Omega_{4,7}: - \frac{59}{100} :\Omega_{2,3}\Omega_{5,6}: +
\frac{4691}{10080} :\Omega_{0,6}\Omega_{1,9}:$$ $$- \frac{39899}{25200} :\Omega_{0,6}\Omega_{2,8}: -
\frac{1889}{5040} :\Omega_{0,6}\Omega_{3,7}: + \frac{1037}{1680} :\Omega_{1,5}\Omega_{0,10}: -
\frac{22}{105} :\Omega_{1,5}\Omega_{1,9}:$$ $$- \frac{1}{15} :\Omega_{1,5}\Omega_{2,8}: - \frac{1369}{4200} :\Omega_{1,5}\Omega_{3,7}: -
\frac{17}{336} :\Omega_{2,4}\Omega_{0,10}: - \frac{37}{96} :\Omega_{2,4}\Omega_{1,9}:$$ $$- \frac{13}{240} :\Omega_{2,4}\Omega_{2,8}: +
\frac{419}{1680} :\Omega_{2,4}\Omega_{3,7}: + \frac{7}{400} :\Omega_{3,3}\Omega_{0,10}: +
\frac{137}{1260} :\Omega_{3,3}\Omega_{1,9}:$$ $$+ \frac{3397}{5040} :\Omega_{3,3}\Omega_{2,8}: + \frac{7}{90} :\Omega_{3,3}\Omega_{3,7}: +
\frac{19}{72} :\Omega_{3,3}\Omega_{4,6}: + \frac{937}{1800} :\Omega_{3,3}\Omega_{5,5}:$$ $$+
\frac{9697}{29400} :\Omega_{0,7}\Omega_{1,8}: + \frac{37}{60} :\Omega_{0,7}\Omega_{2,7}: +
\frac{301}{450} :\Omega_{0,7}\Omega_{3,6}: - \frac{967}{5040} :\Omega_{1,6}\Omega_{0,9}:$$ $$- \frac{47}{240} :\Omega_{1,6}\Omega_{1,8}: +
\frac{123}{700} :\Omega_{1,6}\Omega_{2,7}: + \frac{4}{5} :\Omega_{1,6}\Omega_{3,6}: - \frac{13}{360} :\Omega_{2,5}\Omega_{0,9}:$$ $$-
\frac{4733}{8400} :\Omega_{2,5}\Omega_{1,8}: - \frac{4}{5} :\Omega_{2,5}\Omega_{2,7}: + \frac{17}{225} :\Omega_{2,5}\Omega_{3,6}: +
\frac{11}{252} :\Omega_{3,4}\Omega_{0,9}:$$ $$+ \frac{27}{140} :\Omega_{3,4}\Omega_{1,8}: + \frac{163}{420} :\Omega_{3,4}\Omega_{2,7}: +
\frac{59}{180} :\Omega_{3,4}\Omega_{3,6}: - \frac{1199}{2352} :\Omega_{0,8}\Omega_{1,7}:$$ $$+
\frac{11969}{16800} :\Omega_{0,8}\Omega_{2,6}: - \frac{4399}{6300} :\Omega_{0,8}\Omega_{3,5}: -
\frac{45}{196} :\Omega_{1,7}\Omega_{1,7}: - \frac{97}{175} :\Omega_{1,7}\Omega_{2,6}:$$ $$- \frac{857}{4200} :\Omega_{1,7}\Omega_{3,5}: +
\frac{1}{6} :\Omega_{2,6}\Omega_{2,6}: - \frac{103}{360} :\Omega_{2,6}\Omega_{3,5}: - \frac{127}{300} :\Omega_{3,5}\Omega_{3,5}:,$$
$$D^2_0 = -\frac{2419}{705600} \Omega_{0,18} + \frac{2356854113}{13722508800} \partial^2 \Omega_{0,16} -
\frac{3876250811}{34306272000} \partial^2 \Omega_{1,15} + \frac{710040893}{4678128000} \partial^2 \Omega_{2,14}$$ $$-
\frac{3598850419}{44108064000} \partial^2 \Omega_{3,13} + \frac{50867963}{212058000} \partial^2 \Omega_{4,12} +
\frac{5617559}{75398400} \partial^2 \Omega_{5,11} - \frac{47629609}{62832000}\partial^2 \Omega_{6,10}$$ $$+
\frac{12154709537}{7916832000} \partial^2 \Omega_{7,9} - \frac{13317559687}{15833664000} \partial^2 \Omega_{8,8}.$$
A computer calculation shows that $D^8_0 + D^6_0 + D^4_0 + D^2_0$ lies in $Ker(\pi_3)$, so it must coincide with $D_0$. In particular, $R_0 = -\frac{2419}{705600} J^{18}$.
[ABKS]{}
B. Bakalov and V. Kac, *Field algebras*, Internat. Math. Res. Notices No. 3, 123-159 (2003). J. de Boer, L. Feher, and A. Honecker, *A class of $\cW$-algebras with infinitely generated classical limit*, Nucl. Phys. B420 (1994), 409-445. R. Borcherds, *Vertex operator algebras, Kac-Moody algebras and the monster*, Proc. Nat. Acad. Sci. USA 83 (1986) 3068-3071. C. Dong, H. Li, and G. Mason, *Compact automorphism groups of vertex operator algebras*, Internat. Math. Res. Notices 1996, no. 18, 913–921. C. Dong and K. Nagatomo, *Classification of irreducible modules for the vertex operator algebra $M(1)^+$*, J. Algebra 216 (1999) no. 1, 384-404. C. Dong and K. Nagatomo, *Classification of irreducible modules for the vertex operator algebra $M(1)^+$ II. Higher rank*, J. Algebra 240 (2001) no. 1, 289-325. W. Eholzer, L. Feher, and A. Honecker, *Ghost systems: a vertex algebra point of view*, Nuclear Phys. B 518 (1998), no. 3, 669–688. E. Frenkel and D. Ben-Zvi, *Vertex Algebras and Algebraic Curves*, Math. Surveys and Monographs, Vol. 88, American Math. Soc., 2001. I.B. Frenkel, Y.Z. Huang, and J. Lepowsky, *On axiomatic approaches to vertex operator algebras and modules*, Mem. Amer. Math. Soc. 104 (1993), no. 494, viii+64. E. Frenkel, V. Kac, A. Radul, and W. Wang, *$\cW_{1+\infty}$ and $\cW(\gg\gl_N)$ with central charge $N$*, Commun. Math. Phys. 170 (1995), 337-357. I.B. Frenkel, J. Lepowsky, and A. Meurman, *Vertex Operator Algebras and the Monster*, Academic Press, New York, 1988. V. Kac, *Vertex Algebras for Beginners*, University Lecture Series, Vol. 10. American Math. Soc., 1998. V. Kac and A. Radul, *Representation theory of the vertex algebra $\cW_{1+\infty}$*, Transformation Groups, Vol. 1 (1996) 41-70. V. Kac, W. Wang, and C. Yan, *Quasifinite representations of classical Lie subalgebras of $\cW_{1+\infty}$*, Advances in Mathematics, vol. 139 (1), (1998) 59–140. H. Li, *Local systems of vertex operators, vertex superalgebras and modules*, J. Pure Appl. Algebra 109 (1996), no. 2, 143–195. H. Li, *Vertex algebras and vertex Poisson algebras*, Commun. Contemp. Math. 6 (2004) 61-110. H. Li, Some finiteness properties of regular vertex operator algebras, J. Algebra 212, 495-514 (1999). B. Lian and A. Linshaw, *Howe pairs in the theory of vertex algebras*, J. Algebra 317, 111-152 (2007). B. Lian and G.J. Zuckerman, *Commutative quantum operator algebras*, J. Pure Appl. Algebra 100 (1995) no. 1-3, 117-139. A. Linshaw, *Invariant theory and the $\cW_{1+\infty}$ algebra with negative integral central charge*, J. Eur. Math. Soc., to appear. A. Linshaw, *A Hilbert theorem for vertex algebras*, Transformation Groups, Vol. 15, No. 2 (2010), 427-448. A. Linshaw, *Invariant subalgebras of affine vertex algebras*, arXiv:1011.2281. K. Thielemans, A Mathematica package for computing operator product expansions, Int. Jour. Mod. Phys. C2 (1991) p.787. H. Weyl, *The Classical Groups: Their Invariants and Representations*, Princeton University Press, 1946. Y. Zhu, Modular invariants of characters of vertex operators, J. Amer. Math. Soc. 9 (1996) 237-302.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Kanan K. Datta$^{1}$ [^1], Martina M. Friedrich$^{1}$, Garrelt Mellema$^{1}$, Ilian T. Iliev$^{2}$, Paul R. Shapiro$^{3}$\
$^1$Department of Astronomy & Oskar Klein Centre, AlbaNova, Stockholm University, SE-106 91 Stockholm, Sweden\
$^2$Astronomy Centre, Department of Physics & Astronomy, Pevensey II Building, University of Sussex, Falmer, Brighton BN1 9QH\
$^3$Department of Astronomy and Texas Cosmology Center, University of Texas, Austin, TX 78712, USA
title: Prospects of observing a quasar HII region during the Epoch of Reionization with redshifted 21cm
---
\[firstpage\]
methods: numerical – radiative transfer – galaxies:intergalactic medium – H II regions
Acknowledgments {#acknowledgments .unnumbered}
===============
This study was supported in part by the Swedish Research Council grant 2009-4088. The authors acknowledge the Swedish National Infrastructure for Computing (SNIC) resources at HPC2N (Umeå, Sweden) and PDC (Stockholm, Sweden) and TeraGrid and the Texas Advanced Computing Center (TACC) at The University of Texas at Austin ( URL: http://www.tacc.utexas.edu) for providing HPC resources. The authors also acknowledge the Royal Society International Join Project grant. The authors would like to thank Michiel Brentjens for providing us with the LOFAR station coordinates and Kyungjin Ahn for providing the recipe for including sub-resolution sources in the 504 Mpc simulation volume. KKD is grateful for financial support from Swedish Research Council (VR) through the Oscar Klein Centre (grant 2007-8709). ITI was supported by The Southeast Physics Network (SEPNet) and the Science and Technology Facilities Council grants ST/F002858/1 and ST/I000976/1. PRS was supported by NSF grants AST-0708176 and AST-1009799, NASA grants NNX07AH09G, NNG04G177G and NNX11AE09G, and Chandra grant SAO TM8-9009X.
\[lastpage\]
[^1]: e–mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'An unexpected distribution of temperatures in the CMB could be a sign of new physics. In particular, the existence of cosmic defects could be indicated by temperature discontinuities via the Kaiser-Stebbins effect. In this paper, we show how performing finite differences on a CMB map, with the noise regularized in harmonic space, may expose such discontinuities, and we report the results of this process on the 7-year Wilkinson Microwave Anisotropy Probe data.'
author:
- 'Jude Bowyer$^{*}$'
- 'Andrew H. Jaffe$^{\dag}$'
title: Improved Method for Detecting Local Discontinuities in CMB data by Finite Differencing
---
The Kaiser-Stebbins effect\[1\], a cosmological manifestation of a discontinuous temperature gradient, is a well-known result of the presence of defects in the CMB such as cosmic strings\[2\]. The magnitude of the gradient relates to the string tension $G\mu$ by $$\frac{\delta T}{T}=8\pi G\mu\alpha_{s}$$ where $\alpha_{s}$ contains the dynamical and observational information of the string such as the Lorentz factor and the orientation of the string segment; we will assume $\alpha_{s}=1$. Current limits are of order $G\mu\lesssim10^{-6}$\[3\].\
One method of detecting such a discontinuity is by looking for characteristic signals in the derivatives of the CMB temperature field. Real-space derivative operators produce significant small-scale noise, limiting their utility for cosmology. This is particularly relevant when searching for highly localized sources such as local defects. The noise properties of harmonic and other transform-based methods, such as wavelets, tend to be well behaved up to the Nyquist limit but these transforms act to smooth out discontinuities. However, it is the different error characteristics of these two types of approaches (with respect to each other) in the presence of discontinuous signals that suggest one might calculate the difference of the derivative maps created by each method in order to isolate discontinuous signals.\
When calculating derivatives over a pixelated sampling grid, a finite difference scheme \[4\] is both popular and effective. It can be shown that a given finite difference scheme is related to an underlying interpolating polynomial \[5\]; this being the case, the scheme will perform poorly when in the presence of functions which are not well modeled by such a polynomial, due to Runge’s phenomenon \[6\]. This is true for discontinuous signals, with the error in the resultant derivative scaling with both the order of the differencing scheme and the severity (the step size) of the discontinuity. These errors can then be used as a crude detector of discontinuities.\
In the following text, a method to detect discontinuities and the results of an application to the Wilkinson Microwave Anisotropy Probe (WMAP) 7-year internal linear combination (ILC) CMB temperature anisotropy map are reported. The ILC map is known to suffer from systematic noise at multipoles $l\gtrsim100$, which we further confirm with the detection of a relic of the galaxy boundary. This is performed using the MASQU software package \[5\], which was initially designed for clean masked CMB polarization mode separation \[7\] on the popular HEALPIX \[8\] spherical grid, utilizing real-space differential operators which act on the basis-dependent orthogonal Stokes’ linear polarization parameters $Q$ and $U$ to produce the scalar and pseudoscalar $e$ and $b$ fields, real-space analogues of the $E$ and $B$ modes. In the past, the signal-to-noise ratio in the Stokes’ parameters has been too low for such a method; projects such as the Planck Surveyor \[9\] will go some way toward rectifying that. In the present case, derivative operations will be performed on the temperature anisotropy map instead. The derivatives of a field with respect to coordinate basis $\hat{n}$ at each point on a pixelated grid can be computed by the finite difference approximation \[4\]
$$\partial_{\hat{n}}F_{i}\approx\sum_{j}w_{ij}^{(\hat{n})}F_{j}$$
where the $w_{ij}$ are numerical weights attached to a sample of surrounding pixels (the ‘‘pixel stencil’’). The weights are calculated by inverting a matrix $V$ at each pixel, whose elements depend on the positions of the stencil pixels. In the following, $n$th-order calculations $O_{n}$ refer to calculations using a square $(n+1)^{2}$ pixel stencil. Figure 1 shows how the results vary with both step size and stencil size, for the $E/B$ mode-separation. Note that the error increases both in magnitude and in range for larger stencils in the presence of a discontinuity, whereas it would decrease for a function well-modeled by a polynomial.
{height="50mm"}
The ILC map\[10\] is an $N_{side}=512$ resolution temperature aniostropy map, created by the WMAP team by taking weighted combinations of the WMAP band-limited data (bands $K$,$Ka$,$Q$,$V$ and $W$), with special attention paid to the masked region corresponding to the galactic plane. For our analysis of the ILC map, it is useful to calculate the Laplacian of the field, which corresponds to a simple power multiplier $$\nabla^{2}\rightleftharpoons-l(l+1)$$ when applied to the harmonic space reconstruction of a spherical function $$F(\theta,\phi)=\sum_{l=0}^{\infty}\sum_{m=-l}^{l}a^{f}_{lm}Y_{lm}(\theta,\phi),$$ calculated explicitly as $$\nabla^{2}F(\theta,\phi)=\sum_{l=0}^{\infty}\sum_{m=-l}^{l}-l(l+1)a^{f}_{lm}Y_{lm}(\theta,\phi)$$ where the $Y_{lm}$ are the ordinary spherical harmonics.\
From here on in, we refer to the method of taking a HEALPIX map, generating the $a_{lm}$s via the iterative HEALPIX $map2alm$ technique\[11\], multiplying these coefficients by the factor in Eq. (3), and creating a Laplacian map by the HEALPIX $alm2map$ technique as merely ‘‘the spectral method.’’ The spectral method will also expose any conspicuous boundaries, and suffers from Gibbs’ phenomenon\[12,13\], similar to that of Runge. Further, the iterative HEALPIX method used to recreate the $a_{lm}$s is of reasonable but nevertheless limited accuracy. These errors will also propagate through to the Laplacian map. Since the interpolating polynomial and the spherical harmonics are not identical, a discontinuity should be enhanced by calculating the map of differences between the finite difference Laplacian and the spectral Laplacian. This is then particularly useful when the discontinuity is subtle. As a toy model for the calculation, we show the results for a CMB simulation injected with a diamond-shaped discontinuity (i.e. by linearly adding $\delta T/T=100\mu$K to the map at the discontinuity region, see Fig. 2).
{height="50mm"}
To estimate this calculation, it is noted that the differencing error for derivatives of a function $f(\theta,\phi)$ at pixel $i$ is approximately $$\delta f_{\mathrm{diff},i}(\theta,\phi)\approx\frac{f^{(n+1,p+1)}(\theta,\phi)}{(n+1)!(m+1)!}\mathrm{det}(V)$$ where $(n,p)$ are the orders of the derivatives in coordinate bases $(\theta,\phi)$, yielding for the full Laplacian $$\begin{split}\delta(\nabla^{2}f)_{\mathrm{diff},i}\approx\left(\csc^{2}\theta \frac{f^{(1,3)}(\theta,\phi)}{3!}+\cot\theta\frac{f^{(2,1)}(\theta,\phi)}{2!}\right.
\\ \hspace{1in}+\left.\frac{f^{(3,1)}(\theta,\phi)}{3!}\right)\mathrm{det}(V).\end{split}$$
Meanwhile, the function reconstruction error from the Gibbs’ phenomena receives contributions from two places; First, the Fourier sum in the limit of $l\rightarrow\infty$ will reconstruct the signal everywhere except at the discontinuity. Secondly, truncation creates overshoot (ringing) in the vicinity of the discontinuity. These errors will be of the form $$\delta f_{\mathrm{recon},i}=\left|f(\theta,\phi)-\sum_{lm}^{\infty}a^{f}_{lm}Y_{lm}(\theta,\phi)\right|$$ and $$\delta f_{\mathrm{trun},i}=\left|\sum_{lm}(a^{f}_{lm}-a^{g}_{lm}) Y_{lm}(\theta,\phi))\right|$$ where $g$ is a truncated approximation to the original function $f$, cut off at $l_{max}$ $$g(\theta,\phi)=\sum_{lm}^{l_{max}}a^{f}_{lm}Y_{lm}(\theta,\phi)$$ An analysis of this type on the sphere has been carried out for longitudinal ringing via the Gegenbauer polynomials in \[14\]. Finally, there is an error contribution $\delta f_{\mathrm{hpix}}$ which accounts for any small remaining errors accumulated from the HEALPIX routines. In calculating the Laplacian of the field, the full harmonic function errors $\delta f_{\mathrm{spec}}$ will have additional scaling as $\sim l^{2}$. Our method thus calculates $$\delta(\nabla^{2}f)_{i}\approx\delta(\nabla^{2}f)_{\mathrm{diff},i}-\nabla^{2}(\delta f_{\mathrm{spec},i}).$$ at each pixel position $i$. The noise characteristics (Fig. 3) of the absolute difference map accounts for why a discontinuity can be seen by eye in the toy model absolute difference map; the $\sim l^{2}$ scaling in either Laplacian map is approximately canceled, leaving white noise on small scales.
![Thick line: power spectrum for the $O_{2}$ absolute difference map calculated over a map synthesized from a Gaussian random sample of $a_{lm}$s provided by the ($\sigma=1\mu$K) HEALPIX random number generator facility. The dashed and dotted lines are the spectra for the spectrally-calculated and $O_{2}$ finite-differenced Laplacian maps respectively. The discrepancy between the dashed and dotted lines decreases with an increased number of sampling points in the finite-differencing calculations. The power spectrum $C_{l}^{|\nabla^{2}T_{\mathrm{spec}}-\nabla^{2}T_{\mathrm{F.D.}}|}$ of the map differences has a white spectrum on small scales due to the strong correlation of noise between the maps and a boost on large scales due to the HEALPIX geometry in the polar cap (see \[5\]).](gauss_noise.eps){height="60mm" width="85mm"}
The results for the software performance on the ILC data are shown in Fig. 4. Taking the absolute difference between the finite-differenced ILC map and spectrally calculated ILC map reveals the galaxy mask boundary very clearly, regardless of the smoothing used in the ILC map construction\[10\].
{height="50mm"}
The ILC map is not directly useful for cosmological parameter estimation; instead, it is used as an approximately noise-free CMB map in order to isolate galactic foregrounds and generate a zeroth-order noise template \[15\] which is then supplemented by a maximal entropy method. In this sense the impact of removing the boundary discontinuity region for analyses is unlikely to seriously effect the cosmological implications of the WMAP results. Instead of implementing the complicated WMAP-type analysis to reconstruct the angular power spectrum, the boundary is isolated in a separate mask. One can then estimate the contribution of the masked region to the ILC power spectrum by using the relation between the pseudo-$C_l$ spectrum $\tilde{C}_l$ and the theoretical power spectrum given below:
$$\tilde{C}_{l}=\sum_{ll'}K_{ll'}C_{l'}$$
via a mode-coupling matrix $K_{ll'}$. More typically, this technique is used in MASTER-type calculations\[16\] (a Monte Carlo-type power spectrum estimator, such as SPICE\[17\]) since the equation above properly recovers only the Monte Carlo average of the spectral signal. This is performed for a range of boundary thicknesses (2-,4- and 6-pixel mask thicknesses), since the smoothing kernel corresponds to 1.5$\deg$\[18\] which is not rigorously followed in the code used to create the boundary mask.
{height="50mm"}
It can be seen in Fig. 6 that the impact of the boundary region is small; however, as cosmological experiments probe the microwave sky with higher precision and theory discrimination becomes more sensitive to subtle differences in predicted power spectra, it may be necessary to ensure that excess power from the mask boundary does not bias results. Secondly, given recent claims of detections of non-Gaussianity\[19,20\] it would certainly be necessary to account for the boundary region in any calculation of the nonlinear term $f_{NL}$\[21\]. The effect of the boundary region on a number of the so-called anomalies\[22\] has also been calculated, with no significant divergence from their manifestation when the boundary region is included (Fig. 5).
{height="60mm"}
One can continue the search for anomalies by excluding the galaxy region precalculation and generating a new Laplacian-difference map. For firm statistical analysis (eschewing any by-eye method), the needlet approach\[23\] is particularly suitable to decomposing a discontinuity into power due to its localization properties, while the Canny algorithm\[24\] is suited to edge detection. Further, it should be noted that there may be correlations at the discontinuity boundary with the Stokes’ $Q$ and $U$ parameters, since cosmic strings have a vector $B$-mode signal\[25\] in real space analogous to the Kaiser-Stebbins effect in the temperature case. These issues will be discussed in a future paper.\
In conclusion, the method proposed finds no immediate evidence for anomalies such as strings, with the caveat that only lower-order calculations have been performed, coupled with by-eye detection. This finding may change with higher-order calculations and a more rigorous examination using the needlet algorithm. Also, since the signal in a pixel is an average over signals within the pixel area, it is possible that the results of the Planck Surveyor with an improved resolution of $N_{side}=2048$, or more local balloon surveys such as the E and B Experiment \[26\] and their successors, will test these results. The method did encounter a success: the detection of the ‘‘smoothed’’ galaxy boundary in the map. It was found that the effect of excluding the boundary from the map was negligible, but that its mere detection might warrant added care in future surveys when performing map synthesis.
Thanks go to H. V. Peiris for promoting the needlet approach. This work was supported by the STFC.
[99]{}
$^{*}$Electronic address: [email protected]\
$^{\dag}$Electronic address: [email protected]\
$[$1$]$ Kaiser, N. & Stebbins, A. \[1984\], Nature [**310**]{}, 391.\
$[$2$]$ Kibble, T.W.B. \[1976\], J. Phys. [**A9**]{}, 1387\
$[$3$]$ Battye, R. & Moss, A. \[2010\], Phys. Rev. D [**82**]{}, 023521\
$[$4$]$ Ames, W. \[1977\], [*Numerical Methods for Partial Differential Equations*]{}, Academic Press, New York\
$[$5$]$ Bowyer, J., Jaffe, A.H. & Novikov, D.I., ArXiv:1101.0520\
$[$6$]$ Runge, C. \[1901\], Z. Math. Phys. [**46**]{} 224\
$[$7$]$ Zaldarriaga, M. & Seljak, U. \[1997\], Phys. Rev. D. [**55**]{} 1830\
$[$8$]$ Gorski, K.M. [*et al*]{} \[2005\], Astrophys. J. [**622**]{} 759\
$[$9$]$ The Planck Collaboration, ArXiv:astro-ph/0604069\
$[$10$]$ Gold, B. [*et al*]{}, ArXiv:1001.4555\
$[$11$]$ Gorski, K.M. [*et al*]{}, ArXiv:astro-ph/9905275\
$[$12$]$ Weyl, H. & Chandrasekharan, K. \[1968\], [*Gesammelte Abhandlungen*]{}, Springer-Verlag, Berlin\
$[$13$]$ Gibbs, J. W. \[1899\], Nature [**59**]{} 606\
$[$14$]$ Gelb, A. \[1997\], Math. Comput. [**66**]{} 699\
$[$15$]$ Bennett, C. [*et al*]{} \[2003\] Astrophys. J. Supp. Ser. [**148**]{} 97\
$[$16$]$ Hivon, E., Gorski, K.M. & Netterfield, C.B. \[2002\], Astrophys. J. [**567**]{} 2\
$[$17$]$ Szapudi, I [*et al*]{} \[2001\], Astrophys. J. [**548**]{} L115\
$[$18$]$ Hinshaw, G. [*et al*]{} \[2007\]. Astrophys. J. Supp. Ser. [**170**]{}, 288\
$[$19$]$ McEwen, J. [*et al*]{} \[2008\], Mon. Not. R. Astron. Soc. [**388**]{} 659\
$[$20$]$ Yadav, A. & Wandelt, B. \[2008\] Phys. Rev. Lett. [**100**]{} 181301\
$[$21$]$ Komatsu, E. & Spergel, D. N. \[2001\], in [*Proceedings of the 9th Marcel Grossmann Meeting*]{}, edited by Gurzadyan, V. G., Jantzen, R., and Ruffini, R., World Scientific, Singapore\
$[$22$]$ Bennett, C. [*et al*]{}, ArXiv:1001.4758\
$[$23$]$ Marinucci, D. [*et al*]{} \[2008\], Mon. Not. R. Astron. Soc. [**383**]{} 539\
$[$24$]$ Danos, R. & Brandenberger, R. \[2010\], Int. J. Mod. Phys. D [**19**]{} 183\
$[$25$]$ Benabed, K. & Bernardeau, F. \[2000\], Phys. Rev. [**D61**]{}, 123510.\
$[$26$]$ Oxley, P. [*et al*]{} \[2004\], Proc. SPIE Int. Soc. Opt. Eng. [**5543**]{} 320
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In 1996, Jackson and Martin [@JM1] proved that a strong ideal ramp scheme is equivalent to an orthogonal array. However, there was no good characterization of ideal ramp schemes that are not strong. Here we show the equivalence of ideal ramp schemes to a new variant of orthogonal arrays that we term *augmented orthogonal arrays*. We give some constructions for these new kinds of arrays, and, as a consequence, we also provide parameter situations where ideal ramp schemes exist but strong ideal ramp schemes do not exist.'
author:
- |
Douglas R. Stinson[^1]\
David R. Cheriton School of Computer Science\
University of Waterloo\
Waterloo, Ontario N2L 3G1, Canada
title: Optimal Ramp Schemes and Related Combinatorial Objects
---
Introduction {#intro.sec}
============
Informally, a *$(t,n)$ threshold scheme* ([@Blakley; @Shamir]) is a method of distributing secret information (called *shares*) to $n$ players, in such a way that any $t$ of the $n$ players can compute a predetermined *secret*, but no subset of $t-1$ players can determine the secret. The integer $t$ is called the *threshold*; we assume that $1 \leq t \leq n$. It is well-known that the number of possible shares in a threshold scheme must be greater than or equal to the number of possible secrets. If the number of possible secrets in a threshold scheme equals the number of possible shares, the scheme is termed *ideal*.
*An $(s,t,n)$ ramp scheme* ([@BM]) is a generalization of a threshold scheme in which there are two thresholds. The value $s$ is the *lower threshold* and $t$ is the *upper threshold*. In a ramp scheme, any $t$ of the $n$ players can compute the secret (exactly as in a $(t,n)$ threshold scheme). It is also required that no subset of $s$ players can determine the secret. We note that a $(t-1,t,n)$ ramp scheme is exactly the same thing as a $(t,n)$ threshold scheme. The parameters of a ramp scheme satisfy the conditions $0 \leq s < t \leq n$.
A ramp scheme with $s < t-1$ possibly permits a larger number of possible secrets for a given number of shares than is the case in a threshold scheme. If there are $v$ possible shares in an $(s,t,n)$ ramp scheme, then the number of possible secrets is bounded above by $v^{t-s}$. Of course, for $s < t-1$, it holds that $v^{t-s} > v$. When $s = t-1$, the bound is equal to $v$, agreeing with the above-mentioned bound for threshold schemes. If an $(s,t,n)$ ramp scheme can be constructed with $v^{t-s}$ possible secrets (given $v$ possible shares), then we say that the ramp scheme is *ideal*. Thus, an ideal $(t-1,t,n)$ ramp scheme is the same thing as an ideal $(t,n)$ threshold scheme.
One of the very first constructions for threshold schemes, the Shamir threshold scheme [@Shamir], yields ideal schemes. It is also well-known that ideal threshold schemes are equivalent to certain well-studied combinatorial structures, namely, orthogonal arrays and maximum distance separable (MDS) codes [@Mar; @DMR; @BK].
There is less work on combinatorial characterizations of optimal ramp schemes. The main result in this direction is due to Jackson and Martin [@JM1 Theorem 9], who show that an ideal $(s,t,n)$ ramp scheme that satisfies certain additional conditions (they call such a scheme a *strong* ramp scheme) is equivalent to an ideal $(t,n+t-s-1)$ threshold scheme. This result is perhaps not completely satisfying because the additional conditions used to define strong ramp schemes are rather restrictive. In [@JM1], the authors ask if it is possible to construct ideal ramp schemes that are not strong. This is one of the open questions that we answer in this paper.
Our approach is to define a new type of combinatorial structure that we term an *augmented orthogonal array*, or *AOA*. We prove that any optimal ramp scheme is equivalent to a certain augmented orthogonal array. This equivalence can be proven in a straightforward manner, analogous to the proof that an ideal threshold scheme is equivalent to an orthogonal array. We then investigate some methods of constructing augmented orthogonal arrays. There is a natural way to construct augmented orthogonal arrays from orthogonal arrays. (Roughly speaking, the resulting augmented orthogonal arrays correspond to the strong ramp schemes considered in [@JM1; @Yamamoto].) However, we observe that there are also constructions of augmented orthogonal arrays which yield ideal ramp schemes that are not strong. Moreover, we show there are parameter situations for which there exist ideal ramp schemes, but where there do not exist strong ideal ramp schemes. These results provide answers to the questions that were first posed in [@JM1].
For future reference, Figure \[fig1\] shows the relationships between the ramp schemes and combinatorial structures we discuss in this paper.
--------------------------------------------------------------- ----------------------- -----------------
strong $(s,t,n)$ ramp scheme defined over a set of $v$ shares $\Longleftrightarrow$ OA$(t,n+t-s,v)$
$\Downarrow$ $\Downarrow$
$(s,t,n)$ ramp scheme defined over a set of $v$ shares $\Longleftrightarrow$ AOA$(s,t,n,v)$
--------------------------------------------------------------- ----------------------- -----------------
The rest of this paper is organized as follows. In Section \[defns.sec\], we give formal definitions of ramp and threshold schemes, based on the “distribution rules” for the scheme. Section \[equiv.sec\] reviews combinatorial structures equivalent to ideal threshold schemes (orthogonal arrays, MDS codes, etc.) in both the linear and general cases (in this context, the term “linear” means that the object in question can be viewed as a subspace of a vector space over a finite field). Section \[AOA.sec\] introduces the new notion of an augmented orthogonal array (AOA). We then discuss the connection between AOAs and orthogonal arrays. We also provide some constructions for AOAs in situation where “associated” orthogonal arrays do not exist. Section \[ramp-equiv.sec\] gives the proof that an ideal ramp scheme is equivalent to an AOA. We also provide examples of ideal ramp schemes that are not strong in this section. Finally, we conclude with some discussion and comments in Section \[summary.sec\].
Formal Definitions of Ramp and Threshold Schemes {#defns.sec}
================================================
In this section, we provide formal definitions of ramp schemes in two flavours, namely, “weak” and “perfect”. Our definitions are phrased in terms of “distribution rules”, which is one of the standard ways of defining these types of schemes. (For a discussion of this model in relation to other models, we refer the reader to [@JM2].)
Suppose there is an $(s,t,n)$ ramp scheme defined over a set of $v$ secrets. We will assume without loss of generality that the set of possible shares for any player is $\mathcal{X} = \{1, \dots, v\}$, and we denote the set of possible secrets by $\mathcal{K}$.
We now present a formal mathematical model for a ramp scheme. Denote the set of $n$ *players* by $\mathcal{P} = \{P_1, \dots , P_n\}$. A *distribution rule* $d$ represents a possible distribution of shares to the $n$ players. So we can view $d$ as a function, i.e., $d : \mathcal{P} \rightarrow \mathcal{X}$. The share given to $P_i$ is $d(P_i)$, $1 \leq i \leq n$. We may also represent $d$ as an $n$-tuple $(d_1, \dots , d_n)$, where $d_i = d(P_i)$, for $1 \leq i \leq n$. Finally, for a distribution rule $d$ and a subset of players $\mathcal{P}_0 \subseteq \mathcal{P}$, we define the *projection of $d$ to $\mathcal{P}_0$*, denoted $d|_{\mathcal{P}_0}$, to be the restriction of $d$ to the subdomain $\mathcal{P}_0$. A projection can be represented as a tuple of length $|\mathcal{P}_0|$.
For every possible secret $K \in \mathcal{K}$, we have a collection of distribution rules denoted by $\mathcal{D}_K$. The collection $\mathcal{D}_K$ is the subset of distribution rules for which $K$ is the value of the secret. The entire set of distribution rules is denoted by $\mathcal{D} = \cup_K \mathcal{D}_K$. We assume without loss of generality that the distribution rules in $\mathcal{D}$ are all distinct. When the *dealer* wishes to share a secret $K \in \mathcal{K}$, they first choose a distribution rule $d \in \mathcal{D}_K$ and then they use $d$ to distribute shares to the $n$ players. The choice of the secret $K$ and the distribution rule $d$ will be determined by appropriate probability distributions. The only property that we will require moving forward is that every possible distribution rule is used with positive probability (which implies that every possible secret occurs with positive probability).
\[weak.rs\] A set of distribution rules $\mathcal{D}$ is a *weak $(s,t,n)$ ramp scheme* if the following two properties are satisfied:
(1)
: Suppose $K,L \in \mathcal{K}$, $|\mathcal{P}_0| \geq t$, $d \in \mathcal{D}_K$, $e \in \mathcal{D}_L$ and $d|_{\mathcal{P}_0} = e|_{\mathcal{P}_0}$. Then $K = L$. (This property is saying that $t$ or more shares determine a unique secret.)
(2)
: Suppose $K \in \mathcal{K}$, $|\mathcal{P}_0| \leq s$ and $d \in \mathcal{D}_K$. Then, for every $L \in \mathcal{K}$, there is at least one distribution rule $e \in \mathcal{D}_L $ such that $d|_{\mathcal{P}_0} = e|_{\mathcal{P}_0}$.
Property [**(2)**]{} is saying that $s$ or fewer shares do not rule out any possible value of the secret. An alternative definition is to require that the probability distribution on the set of possible secrets is unchanged even when $s$ shares are known. Suppose we consider $\mathbf{K}$ to be the random variable defined by the probability distribution on the set of secrets $\mathcal{K}$. For any subset of players $\mathcal{P}_0$, define $\mathbf{X}(\mathcal{P}_0)$ to be the random variable determined by the probability distribution induced on the possible lists (i.e., tuples) of shares given to the players in $\mathcal{P}_0$.
A set of distribution rules $\mathcal{D}$ is a *perfect $(s,t,n)$ ramp scheme* if the following two properties are satisfied:
(1)
: Suppose $K,L \in \mathcal{K}$, $|\mathcal{P}_0| \geq t$, $d \in \mathcal{D}_K$, $e \in \mathcal{D}_L$ and $d|_{\mathcal{P}_0} = e|_{\mathcal{P}_0}$. Then $K = L$.
(2\*)
: Suppose $|\mathcal{P}_0| \leq s$ and let $d$ be any distribution rule. Then, for every $L \in \mathcal{K}$, it holds that $$\mathbf{Pr}[ \mathbf{K} = L \mid \mathbf{X}(\mathcal{P}_0) = d|_{\mathcal{P}_0}]
= \mathbf{Pr}[\mathbf{K} = L].$$
Of course, any perfect ramp scheme is also a weak ramp scheme.
The following is a standard method to produce ramp schemes. We use the presentation of the construction from [@OK] (see also [@Stinson §14.2.1]).
\[Shamir-ramp\] Let $q \geq n+1$ be a prime power, define $t_0 = t-s$, let $\mathcal{K} = ({\ensuremath{\mathbb F}}_q)^{t_0}$, and let $\mathcal{X} = {\ensuremath{\mathbb F}}_q$. Define $x_1,x_2, \dots ,x_n$ to be $n$ distinct non-zero elements of ${\ensuremath{\mathbb F}}_q$. The value $x_i$ is associated with $P_i$, for all $i$, $1 \leq i \leq n$. There are $q^{t}$ distribution rules in the scheme. For any $t$-tuple $\mathbf{a} = (a_0, \dots , a_{t-1}) \in ({\ensuremath{\mathbb F}}_q)^{t}$ we have an associated distribution rule $d_{\mathbf{a}} \in \mathcal{D}_{(a_0, \dots , a_{t_0-1})}$ defined by the equation $$d_{\mathbf{a}}(P_j) = \sum_{i=0}^{t-1} a_i (x_j)^i,$$ for $1 \leq i \leq n$. The result is an ideal $(s,t,n)$ ramp scheme with shares from ${\ensuremath{\mathbb F}}_q$.
\[shamir.rem\] In the construction above, there is a distribution rule corresponding to every possible polynomial $a(x) \in {\ensuremath{\mathbb F}}_q[x]$ of degree at most $t-1$. The shares are evaluations of the polynomial $a(x)$, analogous to a Reed-Solomon code. The secret consists of the first $t-s$ coefficients of the polynomial $a(x)$. In the case $t-s = 1$, we have the Shamir threshold scheme [@Shamir] and the secret is the constant term of the the polynomial $a(x)$.
The following result is well-known. We provide the proof, which uses a simple counting argument, for completeness.
\[idealbound\] Suppose that $\mathcal{K}$ is the set of possible secrets and $\mathcal{X}$ is the set of possible shares for any $(s,t,n)$ ramp scheme. Then $|\mathcal{K}| \leq |\mathcal{X}|^{t-s}$.
Let $\mathcal{P}_0 \subseteq \mathcal{P}$ and $\mathcal{P}_1 \subseteq \mathcal{P}$, where $|\mathcal{P}_0| = s$, $|\mathcal{P}_1| = t-s$, and $\mathcal{P}_0 \cap \mathcal{P}_1 = \emptyset$. Fix any distribution rule $d$. Property [**(2)**]{} states that, for every $L \in \mathcal{K}$, there exists a distribution rule $e_L \in \mathcal{D}_L $ such that $d|_{\mathcal{P}_0} = e_L|_{\mathcal{P}_0}$. Consider the set $$\{ e_L|_{\mathcal{P}_1} : L \in \mathcal{K}\} .$$ This set consists of $|\mathcal{K}|$ different $(t-s)$-tuples of shares. Therefore, $|\mathcal{K}| \leq |\mathcal{X}|^{t-s}$.
If equality is achieved in Lemma \[idealbound\], then the ramp scheme is termed *ideal*.
\[idealTS\] Suppose that $\mathcal{K}$ is the set of possible secrets and $\mathcal{X}$ is the set of possible shares for any $(t,n)$ threshold scheme. Then $|\mathcal{K}| \leq |\mathcal{X}|$.
Set $s = t-1$ in Lemma \[idealbound\].
Combinatorial Equivalences Involving Threshold Schemes {#equiv.sec}
======================================================
In this section, we review known combinatorial equivalences involving threshold schemes.
An *orthogonal array*, denoted $\mathrm{OA}(t,k,v)$, is a $v^t$ by $k$ array $A$ defined on an alphabet $\mathcal{X}$ of cardinality $v$, such that any $t$ of the $k$ columns of $A$ contain all possible $k$-tuples from $\mathcal{X}^t$ exactly once.
A *maximum distance separable code*, or *MDS code*, of length $k$ and size $v^t$ over an alphabet $\mathcal{X}$ of size $v$, is a set of $v^t$ vectors (called *codewords*) in $\mathcal{X}^k$, having the property that the hamming distance between any two distinct codewords is at least $k-t+1$.
\[equiv1\] The following are equivalent:
1. an ideal $(t,k-1)$-threshold scheme with shares from an alphabet of size $v$
2. an $\mathrm{OA}(t,k,v)$
3. an MDS code of length $k$ and size $v^t$ over an alphabet of size $v$.
Given an ideal $(t,k-1)$-threshold scheme, if we write out all the possible distribution rules in the form of an array, then this array is the stated orthogonal array. The rows of the orthogonal array form the codewords in the stated code.
The equivalence of 2. and 3. is a “classical” result that has been known at least since the work of Delsarte [@Delsarte]. The equivalence of 1. and 2. was proven by Keith Martin in 1991 in his PhD thesis [@Mar], and independently by Dawson, Mahmoodian and Rahilly in 1993 [@DMR] and by Blakley and Kabatianski in 1995 [@BK]. Much earlier, in 1983, it was stated in a paper by Karnin, Greene and Hellman [@KGH] that an ideal threshold scheme implies the existence of the corresponding MDS code, but their proof is incomplete.
In this paper, we use the term “linear” to refer to combinatorial structures that can be viewed as subspaces of a vector space over a finite field[^2]. The following theorem is the analogue of Theorem \[equiv1\] restricted to the setting of linear threshold schemes. First, however, we define some relevant concepts.
Let $q$ be a prime power and let $t\geq 2$. The *desarguesian projective geometry PG$(t-1,q)$* is based on the vector space $\mathcal{V} = ({\ensuremath{\mathbb F}}_q)^t$. The *points* in PG$(t-1,q)$ are the one-dimensional subspaces of $\mathcal{V}$; the *lines* are the two-dimensional subspaces of $\mathcal{V}$, etc. A *hyperplane* in the geometry is a $(t-1)$-dimensional subspace of $\mathcal{V}$.
A *$k$-arc* in the projective geometry PG$(t-1,q)$ is a set of $k$ points in PG$(t-1,q)$ such that no $t$ of them are on a hyperplane.
\[equiv2\] The following are equivalent:
1. an ideal linear $(t,k-1)$-threshold scheme with shares from ${\ensuremath{\mathbb F}}_q$
2. a linear OA$(t,k,q)$ defined over ${\ensuremath{\mathbb F}}_q$
3. a linear MDS code of length $k$ and dimension $t$ over ${\ensuremath{\mathbb F}}_q$
4. a $k$-arc in PG$(t-1,q)$
5. a $t \times k$ matrix $M$ over ${\ensuremath{\mathbb F}}_q$ such that any $t$ columns of $M$ are linearly independent.
The equivalence of 1., 2. and 3. in Theorem \[equiv2\] is the same as in Theorem \[equiv1\], since it is clear that the relevant transformations “preserve” linearity.
We note that the matrix defined in 5. is just a basis for the code or the orthogonal array. The points in the $k$-arc are the one-dimensional subspaces of $({\ensuremath{\mathbb F}}_q)^t$ generated by the columns of this matrix (recall that the “points” in a projective geometry of dimension $d$ are one-dimensional subspaces of a $(d+1)$-dimensional vector space).
The main existence results for these structures come from Reed-Solomon codes (RS-codes), which are linear MDS codes. The following result, stated in terms of the equivalent orthogonal arrays, is well-known (see, e.g., [@MS Ch. 11, §5]).
\[RS.thm\] Suppose $q$ is a prime power and $t \geq 2$. Then there is a linear OA$(t,q+1,q)$.
Let $\alpha_1, \dots, \alpha_{q-1}$ be the nonzero elements of ${\ensuremath{\mathbb F}}_q$. Consider the following $t$ by $q+1$ matrix $M_0$: $$M_0 =
\left(
\begin{array}{cccccc}
1 & 1 & 1 & \cdots & 1 & 0\\
0 & \alpha_1 & \alpha_2 & \cdots & \alpha_{q-1} & 0\\
0 & {\alpha_1}^2 & {\alpha_2}^2 & \cdots & {\alpha_{q-1}}^2 & 0\\
\vdots & \vdots & \vdots & \ddots & \vdots & \vdots\\
0 & {\alpha_1}^{t-1} & {\alpha_2}^{t-1} & \cdots & {\alpha_{q-1}}^{t-1} & 1
\end{array}
\right) .$$ Any $t$ columns of $M_0$ are linearly independent, so it generates a linear OA$(t,q+1,q)$
It has been conjectured (e.g., by Hedayat, Sloane and Stufken, [@HSS p. 96]) that if an MDS code of length $k$ and size $q^t$ over ${\ensuremath{\mathbb F}}_q$ exists, then there is a [*linear*]{} MDS code with the same parameters.
In the “general” case, the first necessary conditions derive from the classical Bush bounds for orthogonal arrays, which were proven in 1952 (see [@CD]).
\[bush.thm\] If there is an OA$(t,k,v)$, then $$k \leq
\begin{cases}
v+t-1 & \text{if $t=2$, or if $v$ is even and $3 \leq t \leq v$}\\
v+t-2 & \text{if $v$ is odd and $3 \leq t \leq v$}\\
t+1 & \text{if $t \geq v$.}
\end{cases}$$
There have been some relatively minor improvements to these general bounds over the years. On the other hand, the linear case has received considerably more attention and much more is known in this case.
The following is known as the [**Main Conjecture**]{} for linear MDS codes. It is attributed to Segre (1955).
Suppose $q$ is a prime power. Let $M(t,q)$ denote the maximum value of $k$ such that there exists a linear MDS code of length $k$ and dimension $t$ over ${\ensuremath{\mathbb F}}_q$. If $2 \leq t < q$, then $$M(t,q) =
\begin{cases}
q+2 & \text{if $q$ is a power of $2$ and $t \in \{3,q-1\}$}\\
q+1 & \text{otherwise.}
\end{cases}$$ If $t \geq q$, then $M(t,q) = t+1$.
The [**Main Conjecture**]{} has been shown to be true in many parameter situations, including all the cases where $q$ is prime. This is a famous result of Simeon Ball [@Ball] proven in 2012.
The following theorem summarizes some of the known results. These and other related results are surveyed in [@Huntemann].
\[mainconj.thm\] Suppose that $q = p^j$ where $p$ is prime, and suppose $2 \leq t < q$. Then the [**Main Conjecture**]{} is true in the following cases:
1. $q$ is prime (for all relevant $t$)
2. $q \leq 27$ (for all relevant $t$)
3. $t \leq 5$ or $t \geq q-3$
4. $t \leq p$.
Augmented Orthogonal Arrays {#AOA.sec}
===========================
Our objective is to generalize the results of the previous section to ideal ramp schemes. It turns out that a certain (apparently new) type of combinatorial array is required in order to state the resulting equivalences. We define these arrays now, and then we prove some basic results about them and give some constructions.
An *augmented orthogonal array*, denoted $\mathrm{AOA}(s,t,k,v)$, is a $v^t$ by $k+1$ array $A$ that satisfies the following properties:
1. the first $k$ columns of $A$ form an orthogonal array $\mathrm{OA}(t,k,v)$ on a symbol set $\mathcal{X}$ of size $v$
2. the last column of $A$ contains symbols from a set $\mathcal{Y}$ of size $v^{t-s}$
3. any $s$ of the first $k$ columns of $A$, together with the last column of $A$, contain all possible $(s+1)$-tuples from $\mathcal{X}^s \times Y$ exactly once.
There are generalizations of orthogonal arrays, known as *mixed orthogonal arrays*, that may contain different symbol sets in different columns. See, for example, [@HSS Chapter 9]. The AOAs that we have defined above are not mixed orthogonal arrays, however.
\[1333.ex\] We give an example of an $\mathrm{AOA}(1,3,3,3)$. We take $\mathcal{X} = {\ensuremath{\mathbb F}}_3$ and $\mathcal{Y} = {\ensuremath{\mathbb F}}_3 \times {\ensuremath{\mathbb F}}_3$. The AOA has the following $27$ rows: $$\begin{array}{|c|c|c|c|}
\hline
\alpha & \beta & \gamma & (\alpha + \beta, \alpha + \gamma) \\ \hline \end{array}$$ where $\alpha , \beta , \gamma \in {\ensuremath{\mathbb F}}_3$.
The following theorem is immediate from the definitions.
\[OA-AOA\] An $\mathrm{AOA}(t-1,t,k,v)$ is equivalent to an $\mathrm{OA}(t,k,v)$.
The next theorem shows an obvious way to construct AOAs from OAs.
\[OAtoAOA\] If there exists an $\mathrm{OA}(t,k+t-s,v)$, then there exists an $\mathrm{AOA}(s,t,k,v)$.
Merge the last $t-s$ columns of an $\mathrm{OA}(t,k+t-s,v)$ to form a single column whose entries are $(t-s)$-tuples of symbols.
We note that the converse of Theorem \[OAtoAOA\] is not always true. Example \[1333.ex\], which gives a construction of an $\mathrm{AOA}(1,3,3,3)$, provides an illustration of what can go wrong.
The natural way to attempt to construct an $\mathrm{OA}(3,5,3)$ from the $\mathrm{AOA}(1,3,3,3)$ presented in Example \[1333.ex\] would be to split the last column into two columns of elements from ${\ensuremath{\mathbb F}}_3$. We would get the array having the following $27$ rows: $$\begin{array}{|c|c|c|c|c|}
\hline
\alpha & \beta & \gamma & \alpha + \beta & \alpha + \gamma \\ \hline \end{array}$$ where $\alpha , \beta , \gamma \in {\ensuremath{\mathbb F}}_3$. It is easy to see that the fourth column is the sum of the first two columns, so these three columns cannot contain all possible $3$-tuples.
In fact, there does not exist any $\mathrm{OA}(3,5,3)$, because the parameters violate the Bush bound (Theorem \[bush.thm\]).
Later in this section, we will construct some additional examples of $\mathrm{AOA}(s,t,k,v)$ in situations where $\mathrm{OA}(t,k+t-s,v)$ do not exist.
Next, we present an obvious but useful method to construct linear AOAs.
\[linearAOA\] Suppose that $q$ is a prime power. Suppose there is a $t$ by $k+t-s$ matrix $M$, having entries from ${\ensuremath{\mathbb F}}_q$, which satisfies the following two properties:
1. any $t$ of the first $k$ columns of $M$ are linearly independent, and
2. any $s$ of the first $k$ columns of $M$, along with with the last $t-s$ columns of $M$, are linearly independent.
Then there exists a linear AOA$(s,t,k,q)$.
The following application of Construction \[linearAOA\] is a slight generalization of the Shamir ramp schemes that we already presented in Example \[Shamir-ramp\].
\[Shamir.const\] Suppose $q$ is a prime power and $1 \leq s < t \leq k \leq q$. Then there exists a linear AOA$(s,t,k,q)$.
We obtain an AOA$(s,t,k,q)$ using Construction \[linearAOA\], by defining a suitable $t$ by $k+t-s$ matrix $M$ having entries from ${\ensuremath{\mathbb F}}_q$.
The first $k$ columns of $M$ are obtained by deleting the first column and the last $q-k$ columns from the matrix $M_0$ defined in the proof of Theorem \[RS.thm\]. Cal this matrix $M_1$. The last $t-s$ columns of $M$ consist a $t-s$ by $t-s$ identity matrix and $s$ rows of zeroes: $$M_2 = \left(
\begin{array}{cccc}
1 & 0 & \cdots & 0\\
0 & 1 & \cdots & 0\\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \cdots & 1\\
0 & 0 & \cdots & 0\\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \cdots & 0
\end{array}
\right) = \left(
\begin{array}{c}
I_{t-s} \\ \hline \mathbf{0}
\end{array}
\right).$$ Then, construct the matrix $M = \left( \begin{array}{c|c} M_1 & M_2 \end{array} \right)$. The matrix $M$ satisfies the conditions of Theorem \[linearAOA\] and therefore it yields an AOA having the stated parameters.
We are interested in identifying parameters for which there exists an AOA$(s,t,k,q)$ but there does not exist an OA$(t,k+t-s,q)$. Suppose $q$ is odd and $3 \leq t \leq q$. If an OA$(t,k+t-s,q)$ exists, then Theorem \[bush.thm\] asserts that $k+t-s \leq q+t-2$, so $k \leq q+s-2$. Therefore an OA$(t,k+t-s,q)$ does not exist if $k = q$ and $s=1$. So we have the following.
\[nonexist1\] Suppose $q$ is an odd prime power and $3 \leq t \leq q$. Then there exists an AOA$(1,t,q,q)$ but there does not exist an OA$(t,q+t-1,q)$.
Here is another application of Construction \[linearAOA\].
\[second.thm\] Suppose there is a linear OA$(t-s,t,q)$. Then there exists an AOA$(s,t,t,q)$.
Let $N$ be the $t-s$ by $t$ matrix whose rows form a basis for a linear OA$(t-s,t,q)$. Then the matrix $M = \left( \begin{array}{c|c} I_t & N^T \end{array} \right)$ satisfies the conditions of Construction \[linearAOA\] and hence it yields an AOA$(s,t,t,q)$.
We can apply Theorem \[second.thm\] with $t=q+1$. From Theorem \[RS.thm\], there is a linear OA$(q+1-s,q+1,q)$ for any $s \leq q-1$ whenever $q$ is a prime power. Theorem \[second.thm\] then yields an AOA$(s,q+1,q+1,q)$. However, an OA$(q+1,2(q+1)-s,q)$ does not exist because any OA$(q+1,k,q)$ has $k \leq q+2$ by the Bush bound (Theorem \[bush.thm\]) and $2(q+1)-s \geq q+3$ for $s \leq q-1$.
We have shown the following.
\[nonexist2\] Suppose $q$ is a prime power and $s \leq q-1$. Then there exists an AOA$(s,q+1,q+1,q)$ but there does not exist an OA$(q+1,2(q+1)-s,q)$.
We can take $q = 3$, $s = 2$ in Theorem \[nonexist2\]. Here we could let $$N = \left(
\begin{array}{cccc}
1 & 1 & 1 & 0 \\
0 & 1 & 2 & 1
\end{array}
\right).$$ Then we obtain an AOA$(2,4,4,3)$: $$M = \left(
\begin{array}{cccccc}
1 & 0 & 0 & 0 & 1 & 0\\
0 & 1 & 0 & 0 & 1 & 1\\
0 & 0 & 1 & 0 & 1 & 2\\
0 & 0 & 0 & 1 & 0 & 1
\end{array}
\right).$$ However, there is no OA$(4,6,3)$.
Equivalence of Ideal Ramp Schemes and AOAs {#ramp-equiv.sec}
==========================================
We already mentioned that Jackson and Martin [@JM1] proved that strong ideal ramp schemes are equivalent to ideal threshold schemes. For future reference, we state their result in terms of orthogonal arrays.
[@JM1] \[equivalence-JM\] There exists a strong ideal $(s,t,n)$ ramp scheme with $v$ possible shares if and only if there exists an OA$(t,n+t-s,v)$.
Our main goal in this section is to prove the equivalence of ideal ramp schemes and AOAs.
\[L1\] Suppose $\mathcal{D}$ is the set of distribution rules of an ideal $(s,t,n)$ ramp scheme having shares from $\mathcal{X}$ and secrets from $\mathcal{K}$, where $|\mathcal{X}| = v$ and $|\mathcal{K}| = v^{t-s}$. Suppose $\mathcal{P}_0 \subseteq \mathcal{P}$ and $\mathcal{P}_1 \subseteq \mathcal{P}$, where $|\mathcal{P}_0| = s$, $|\mathcal{P}_1| = t-s$ and $\mathcal{P}_0 \cap \mathcal{P}_1 = \emptyset$. Let $d^* \in \mathcal{D}$ and define $$\mathcal{E} = \{ d \in \mathcal{D}: d|_{\mathcal{P}_0} = d^*|_{\mathcal{P}_0} \}.$$ Then, for every $K \in \mathcal{X}$, there is a distribution rule $d \in \mathcal{D}_K \cap \mathcal{E}$. Further, there exists a bijection $\pi : \mathcal{K} \rightarrow \mathcal{X}^{t-s}$ such that, if $d \in \mathcal{D}_K \cap \mathcal{E}$, then $d|_{\mathcal{P}_0} = \pi(K)$.
Define $$\mathcal{Z} =
\{ (d|_{\mathcal{P}_1}, K ) : d \in \mathcal{D}_K \cap \mathcal{E}, K \in \mathcal{K}\}
.$$
- First, we observe, for every $K \in \mathcal{K}$, that there exists $\mathbf{x} \in \mathcal{X}^{t-s}$ such that $(\mathbf{x}, K) \in \mathcal{Z}$ (this follows from property [**(2)**]{} of Definition \[weak.rs\]).
- Next, we note that, if $(\mathbf{x}, K) \in \mathcal{Z}$ and $(\mathbf{x}, L) \in \mathcal{Z}$, then $K = L$ (this follows from property [**(1)**]{} of Definition \[weak.rs\]).
Recall that $|\mathcal{K}| = |\mathcal{X}|^{t-s}$. Hence (i) and (ii) imply that, for any $K \in \mathcal{K}$, there is a *unique* $\mathbf{x} \in \mathcal{X}^{t-s}$ such that $(\mathbf{x}, K) \in \mathcal{Z}$. This allows us to define $$\pi(K) = \mathbf{x} \Leftrightarrow (\mathbf{x}, K) \in \mathcal{Z},$$ and the function $\pi$ will be a bijection.
\[main.thm\] If there exists an ideal $(s,t,n)$ ramp scheme defined over a set of $v$ shares, then there exists an $\mathrm{AOA}(s,t,n,v)$.
Suppose $\mathcal{D}$ is the set of distribution rules of an ideal $(s,t,n)$ ramp scheme having $v$ possible shares. We now describe how to construct an $(s,t,n,v)$-AOA. For every $K \in \mathcal{X}$ and every distribution rule $d \in \mathcal{D}_K$, construct the $(n+1)$-tuple $r_d = (d_1, \dots , d_n, K)$. We will show that the array $\mathcal{A}$ whose rows are all the $(n+1)$-tuples $r_d$ ($d \in \mathcal{D}$) is the desired augmented orthogonal array.
1. We first show that the restriction of $\mathcal{A}$ to any $t$ of the first $n$ columns of $\mathcal{A}$ consists of all $v^t$ possible $t$-tuples, each occurring exactly once. Without loss of generality, choose the first $t$ columns of $\mathcal{A}$ and let $\mathcal{A}'$ denote the subarray of $\mathcal{A}$ consisting of the first $t$ columns.
Suppose $r$ is a row of $\mathcal{A}'$. This means that there is a distribution rule $d$ such that $d |_{\mathcal{P}_0} = r$. Suppose that $r'$ is any $t$-tuple that differs from $r$ in a single co-ordinate, say $r_i \neq r'_i$. Choose any $s$ of the $t-1$ co-ordinates that exclude co-ordinate $i$. It then follows immediately from Lemma \[L1\] that there is a distribution rule $d'$ such that $d' |_{\mathcal{P}_0} = r'$. Therefore, all $t$-tuples that are hamming distance $1$ from $r$ are rows in $\mathcal{A}'$.
This argument can be repeated as often as desired, to establish that $\mathcal{A}'$ contains all the $v^t$ possible $t$-tuples as rows. (More formally, we can induct on the hamming distance between $r$ and $r'$.)
Next, we show that no $t$-tuple $r$ occurs more than once as a row of $\mathcal{A}'$. Suppose that $d |_{\mathcal{P}_1} = d' |_{\mathcal{P}_1} = r$, say, where $|\mathcal{P}_1| = t$. From property [**(1)**]{} of a ramp scheme, we have that $d, d' \in \mathcal{D}_K$, for some $K$. The two distribution rules $d$, $d'$ are not identical, so there is a $P_i \not\in \mathcal{P}_1$ such that $d(P_i) \neq d'(P_i)$. Choose any subset $\mathcal{P}_0 \subseteq \mathcal{P}_1$ such that $|\mathcal{P}_0| = s$. Let $P_j \in \mathcal{P}_1 \setminus \mathcal{P}_0$ and define $$\mathcal{P}_2 = \mathcal{P}_1 \cup \{ P_i \} \setminus \{P_j\}.$$ We have $d |_{\mathcal{P}_0} = d' |_{\mathcal{P}_0}$, $d, d' \in \mathcal{D}_K$, and $d |_{\mathcal{P}_0 \setminus \mathcal{P}_2} \neq d' |_{\mathcal{P}_0 \setminus \mathcal{P}_2}$, which contradicts Lemma \[L1\].
2. So far, we have shown that the first $n$ columns of $\mathcal{A}$ form a $(t,n,v)$-orthogonal array. We now have to consider $s$ of the first $n$ columns, together with the last column. Without loss of generality, take the first $s$ columns of $\mathcal{A}$ and denote them by $\mathcal{A}''$. We already know that there are exactly $v^t$ rows in $\mathcal{A}$. Given any $s$-tuple $r$, there are $v^{t-s}$ occurrences of $r$ in rows of $\mathcal{A}''$. By Lemma \[L1\], these $v^{t-s}$ rows of $\mathcal{A}''$ correspond to all the $v^{t-s}$ different possible values of the secret.
1. and 2. provide the proof of the desired result.
The proof of Theorem \[main.thm\] applies to both weak and perfect ramp schemes. We now observe that the AOA yields a weak or perfect ramp scheme, depending on the probability distributions that are defined on the set of distribution rules. If we define *any* probability distribution on the set of distribution rules, we get (at least) a weak ramp scheme (provided that every distribution rule is used with positive probability). Further, we can ensure that the ramp scheme is perfect by defining probability distributions as follows:
1. Define an arbitrary probability distribution on $\mathcal{K}$ (ensuring that $\mathbf{Pr}[ \mathbf{K} = L ] > 0$ for all $L \in \mathcal{K}$).
2. For any $L \in \mathcal{K}$ and any distribution rule $d \in \mathcal{D}_L$, define $\mathbf{Pr}[ d ] = \mathbf{Pr}[ \mathbf{K} = L ] / v^s$.
It is straightforward to verify that the resulting ramp scheme is perfect.
Summarizing, we have the following
\[equivalence\]
1. If there exists a weak ideal $(t,n)$ ramp scheme defined over a set of $v$ shares, then there exists an $\mathrm{AOA}(s,t,n,v)$.
2. If there exists an $\mathrm{AOA}(s,t,n,v)$, then there exists a perfect ideal $(t,n)$ ramp scheme defined over a set of $v$ shares.
We can now identify some parameter situations in which ideal ramp schemes exist, but strong ideal ramp schemes do not exist.
Suppose $q$ is an odd prime power and $3 \leq t \leq q$. Then there exists an ideal $(1,t,q,q)$ ramp scheme but there does not exist a strong ideal $(1,t,q,q)$ ramp scheme.
This is an immediate consequence of Theorems \[nonexist1\], \[equivalence-JM\] and \[equivalence\].
Suppose $q$ is a prime power and $s \leq q-1$. Then there exists an ideal $(s,q+1,q+1,q)$ ramp scheme but there does not exist a strong ideal $(s,q+1,q+1,q)$ ramp scheme.
This follows from Theorems \[nonexist2\], \[equivalence-JM\] and \[equivalence\].
Summary and Conclusion {#summary.sec}
======================
We showed various parameter situations where ideal ramp schemes exist but strong ideal ramp schemes do not exist. Our approach was to construct linear AOAs and show that the “corresponding” OAs (linear or not) do not exist. It would be easy to find additional parameter sets for which linear AOAs exist but *linear* OAs do not exist, by making use of Theorem \[mainconj.thm\] and other related results in the literature.
All the ramp schemes we constructed in this paper are linear, in the sense that they are subspaces of vector spaces over a finite field. Constructions of ideal ramp schemes over alphabets of non-prime power order would also be of interest.
[XX]{}
S. Ball. On sets of vectors of a finite vector space in which every subset of basis size is a basis. [*Journal of the European Mathematical Society*]{} [**14**]{} (2012), 733–748.
G.R. Blakley. Safeguarding cryptographic keys. [*Proceedings AFIPS 1979 National Computer Conference*]{}, pp. 313–317.
G.R. Blakley and G.A Kabatianski. Ideal perfect threshold schemes and MDS codes. [*Proceedings of the 1995 IEEE International Symposium on Information Theory*]{}, p. 488.
G.R. Blakley and C. Meadows. Security of ramp schemes. [*Lecture Notes in Computer Science*]{} [**196**]{} (1985), 242–268 (Advances in Cryptology: Proceedings of CRYPTO ’84).
C.J. Colbourn and J.H. Dinitz, eds. [*The CRC Handbook of Combinatorial Designs, Second Edition*]{}, CRC Press, 2006.
E. Dawson, E.S. Mahmoodian and A. Rahilly. Orthogonal arrays and ordered threshold schemes. [*Australasian Journal of Combinatorics*]{} [**8**]{} (1993), 27–44.
P. Delsarte. *An Algebraic Approach to the Association Schemes of Coding Theory*. Philips Research Reports Supplement no. 10, 1973.
A.S. Hedayat, N.J.A. Sloane and J. Stufken. [*Orthogonal Arrays: Theory and Applications.*]{} Springer, 1999.
S. Huntemann. [*The Upper Bound of General Maximum Distance Separable Codes.*]{} Honours project, University of New Brunswick, Saint John, May 2012.
W.A. Jackson and K.M. Martin. A combinatorial interpretation of ramp schemes. *Australasian Journal of Combinatorics* [**14**]{} (1996), 51–60.
W.A. Jackson and K.M. Martin. Combinatorial models for perfect secret sharing schemes. *Journal of Combinatorial Mathematics and Combinatorial Computing* [**28**]{} (1998), 249–265.
E.E. Karnin, J.A. Greene and M.E. Hellman. On secret sharing systems. *IEEE Transactions on Information Theory* [**IT-29**]{} (1983), 35–41.
F.J. MacWilliams and N.J.A. Sloane. [*The Theory of Error-Correcting Codes.*]{} North-Holland, 1977.
K.M. Martin. *Discrete Structures in the Theory of Secret Sharing*. PhD Thesis, University of London, 1991.
W. Ogata and K. Kurosawa. Some basic properties of general nonperfect secret sharing schemes. [*Journal of Universal Computer Science*]{} [**4**]{} (1998), 690–704.
A. Shamir. How to share a secret. [*Communications of the ACM*]{} [**22**]{} (1979), 612–613.
D.R. Stinson. Cryptography: Theory and Practice, Third Edition. Chapman & Hall/CRC, 2006.
H. Yamamoto. Secret sharing schemes using $(k,L,n)$ threshold schemes. *Electronics and Communications in Japan* [**69**]{} (1985), 46–54.
[^1]: Research supported by NSERC discovery grant RGPIN-03882
[^2]: However, note that the term “linear ramp scheme” has a completely different meaning in [@JM1].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The quenched dynamics of an ultracold homogeneous atomic two-dimensional Bose gas subjected to periodic quenches across the Berezinskii-Kosterlitz-Thouless (BKT) phase transition are discussed. Specifically, we address the effect of periodic cycling of the effective atomic interaction strength between a thermal disordered state above, and a highly ordered state below the critical BKT interaction strength, by means of numerical simulations of the stochastic projected Gross-Pitaevskii equation. Probing the emerging dynamics as a function of the frequency of sinusoidal driving from low to high frequencies reveals diverse dynamical features, including phase-lagged quasi-adiabatic reversible condensate formation, resonant excitation consistent with an intrinsic system relaxation timescale, and gradual establishment of dynamically-recurring or time-averaged non-equilibrium states with enhanced coherence which are neither condensed, nor thermal. Our study paves the way for experimental observation of such driven non-equilibrium ultracold superfluid states.'
author:
- 'K. Brown'
- 'T. Bland'
- 'P. Comaron'
- 'N. P. Proukakis'
bibliography:
- 'biblio.bib'
title: 'Periodic quenches across the Berezinskii-Kosterlitz-Thouless phase transition'
---
= 1
Introduction {#sec:intro}
============
The quench dynamics of a quantum system across a phase transition are an exciting subject of active ongoing research [@bray_theory_1994; @dziarmaga_review_2010; @polkovnikov_review_2011; @del_campo_universality_2014; @proukakis_snoke_littlewood_2017]. Controlled studies have been performed in a plethora of diverse systems, including spin systems, superconductors, superfluids, ultracold atoms and exciton-polariton systems [@KZnum-a; @KZnum-b; @KZnum-c; @KZnum-d; @KZexp-a; @KZexp-b; @KZexp-c; @KZexp-d; @KZexp-e; @KZexp-f; @KZexp-g; @KZexp-gg; @weiler_08; @erne_universal_2018]. Most studies to date have focussed on a simplified scenario, whereby the system is driven once across the phase transition by a time-dependent external control parameter such as chemical potential, temperature, or pumping laser [@weiler_08; @zurek_09; @jelic2011quench; @damski_10; @das_12; @liu_18; @zamora2020kibble; @comaron_19; @ComaronPRLpolaritons]. The particular case of linear quenching is known to lead to the established Kibble-Zurek scaling law, which quantifies the dependence of correlation functions and spontaneous emergence of defects on the quench rate. Such a model, first proposed in the cosmological context [@kibble1976topology] and subsequently carried through to the condensed matter realm [@zurek1985cosmological], has since been extensively studied numerically [@del_campo_universality_2014; @KZnum-a; @KZnum-b; @KZnum-c; @KZnum-d; @weiler_08; @damski_10; @das_12; @Su13; @Matuszewski2014; @McDonald2015; @liu_18; @liu_20; @bland_marolleau_20; @liu_20; @zamora2020kibble], and in a broad range of experimental settings [@KZexp-a; @KZexp-b; @KZexp-c; @KZexp-d; @KZexp-e; @KZexp-f; @KZexp-g], with experimental studies in ultracold atoms addressing the effect across different dimensionalities and geometries [@KZexp-gg; @weiler_08; @das_12; @lamporesi2013spontaneous; @corman2014quench; @navon2015critical; @chomaz2015emergence; @braun2015emergence; @donadello2016creation; @liu_18; @ko2019kibble; @keesling2019quantum]. The related question of how the presence of a sinusoidally-modulated driving across a phase transition may affect the system dynamics arises naturally. Such a periodic cycling was experimentally investigated in the context of three-dimensional (3D) harmonically trapped ultracold atoms, where a time-dependent dimple microtrap was used to controllably induce a periodic phase space modulation, leading to a reversible cycling across the thermal and the Bose-Einstein condensation phase [@stamperkurn1998reversible]. Condensate formation was found to lag behind the applied constant-frequency modulation of the laser power. Such findings were subsequently reproduced qualitatively by means of numerical simulations based on the stochastic Gross-Pitaevskii equation [@stoof2001dynamics]. Motivated by the above pioneering works [@stamperkurn1998reversible; @stoof2001dynamics], and by our recent studies of instantaneously quenched two-dimensional quantum gases [@comaron_19; @groszek2020crossover], here we address the corresponding periodically driven phase transition crossing in the context of (quasi) two-dimensional homogeneous ultracold atomic Bose gases. Two-dimensional systems are interesting in their own right, due to the different nature of the underlying Berezinskii-Kosterlitz-Thouless (BKT) phase transition [@berezinskii1972destruction; @kosterlitz1973ordering], associated with binding-unbinding of vortex-antivortex pairs, previously observed in diverse physical contexts [@resnick1981kosterlitz; @PhysRevLett.40.1727; @safonov1998observation; @nitsche2014algebraic]. Such properties have also been studied in ultracold atomic systems both experimentally [@stock2005; @hadzibabic2006berezinskii; @kruger2007p; @clade2009observation; @PhysRevLett.105.230408; @hung2011observation; @yefsah_11; @LiChung2013; @Fletcher2015; @chomaz2015emergence], and theoretically [@prokofev2001critical; @Prokofev2002; @davis_01; @simula_06; @Giorgetti2007; @Holzmann2008; @Simula2008; @Bisset2009; @foster2010vortex; @holzmann_10; @mathey2010light; @Cockburn_2012; @mathey_17; @Karl2017; @gawryluk_19; @comaron_19] with the transversal (typically harmonic) confinement in the other direction offering a way to control the effective two-dimensional interaction experienced between the atoms.
In this work, we perform a detailed quantitative analysis of the role of driving frequency on the cyclic phase transition crossing, focussing on the particular case of a periodically-driven 2D homogeneous ultracold atomic Bose gas. This is facilitated by a periodically-modulated-in-time interaction strength, between an initial incoherent state close to, but above, the BKT phase transition, and a state with lower interaction strength below the BKT phase transition exhibiting a high degree of coherence. Our numerical study, performed at fixed system temperature, focusses in parallel on the effect of external driving on density, vortex number, momentum spectrum and coherence. Our parameter choice is based on accessible experimental regimes, building on our earlier work based on an [*instantaneous*]{} interaction quench from above to below the BKT threshold, which focussed on equilibrium properties and late-time phase-ordering dynamics [@comaron_19].
Deep in the superfluid regime, where the system is highly condensed, modulations of the interaction strength have been studied both experimentally and theoretically, in different contexts. Periodic modulations of the interaction strength between two values in the superfluid regime – conducted within the context of the Gross-Pitaevskii equation, for which the system is assumed to be coherent, and often termed ‘Feshbach Resonance Management’ [@Kevrekidis_FRM_03] – have demonstrated the emergence of Faraday patterns [@staliunas_02] and have been used to study aspects of condensate stability and soliton dynamics [@saito_FRM_03; @abdullaev_FRM_03]. Condensate experiments based on a sudden interaction modulation found an interesting analogy with Sakharov oscillations in the early Universe [@Hung_2013], while periodic interaction strength modulations led to matter-wave jet emission (Bose fireworks) [@chin_prl_18; @chin_nature_17] and other interesting patterns [@chin_natphys_20].
The novel feature of our present study is that our periodic interaction quenches are neither conducted between two states deep in the superfluid regime, nor do they start from a well-formed condensate – but instead the interaction strength is modulated across the phase transition multiple times, in direct analogy to the work of [@stamperkurn1998reversible]. For this reason, our analysis is based on the stochastic (projected) Gross-Pitaevskii equation [@Bradley2008; @Blakie2008; @Proukakis2008; @Proukakis13].
Beyond the expected regimes of extremely slow pumping (which allows the system to proceed adiabatically through instantaneous equilibrium states) and extremely rapid pumping which only mildly perturbs the initial state, we find the periodic driving to be intrinsically resonant with a characteristic relaxation time, at which the periodic modulation of the scattering length causes strongly-non-equilibrium features to emerge. This resonance separates two interesting distinct and experimentally relevant physical regimes: Driving frequencies lower than the resonant value lead to reproducible dynamics which are independent of the quench cycle, and – while not fully equilibrated – resemble certain features of the corresponding-parameter equilibrium states. This regime bears close analogy to the experimental 3D findings of Ref. [@stamperkurn1998reversible] (upon excluding observed atom losses). In the opposite regime of frequencies exceeding the resonant value, the system grows gradually over multiple quench cycles, and can – for frequencies few times the resonant frequency – accommodate significant coherence.
This paper is structured as follows: After presenting a brief overview of the 3D experiment of Ref. [@stamperkurn1998reversible], we outline the numerical scheme and quench protocol to be used in our work (Sec. II). We start by discussing the dependence of the emerging density and vortex dynamics on driving frequency, and present a detailed quantitative analysis of their respective dynamical time-delay with respect to the external periodic driving (Sec. III): this allows us to identify an intrinsic system frequency, and thus to focus our subsequent analysis on the physically interesting and experimentally relevant regime of driving frequencies which are up to 20 times slower, or faster, than the identified resonant frequency. Within this experimentally interesting range where most novel dynamical features emerge, we analyse the dependence of the momentum spectrum on the driving frequency and identify parameter regimes of high coherence in spite of fast driving (Sec. IV). We then study the maximum attained coherence as a function of quench frequency, and the emerging non-equilibrium steady-state under rapid pumping (Sec. V), before concluding.
Quench Protocol and Modelling Scheme {#sec:quench}
====================================
Experiment of Stamper-Kurn [*et al.*]{} [@stamperkurn1998reversible]
--------------------------------------------------------------------
In a ground-breaking paper, Stamper-Kurn *et al.* [@stamperkurn1998reversible] achieved reversible condensate formation in a three-dimensional gas of sodium atoms. Initially, this gas was confined to a broad harmonic trap and cooled to just above the critical temperature, $T_c$, at which condensation onsets. Subsequently, a thin well (dimple trap) was introduced at the centre of the trap, using an infrared laser. This granted the sodium atoms access to a new, lower-energy state. The well was made so steep that it could only contain a single new energy level. As well depth was increased, by increasing the strength of the laser, the energy of this single state decreased to the extent that a condensate formed in this well. The condensate fraction grew with well depth, and hence with laser power. The power of the laser beam was then sinusoidally modulated between 0 and 7mW at a frequency of 1Hz, which resulted in the periodic cyclic growth and decay of the condensate fraction, $N_0/N$, from 0 to 6$\%$. A non-zero condensate fraction was observed to recur for 15 oscillations, even though the peak condensate fraction decayed slightly with each oscillation. This decay was shown to arise from atom loss, rather than as a consequence of this periodic crossing of the phase transition. The latter was confirmed by the fact that the same decay in the peak values was observed even when the power of the laser beam was held constant. Importantly, a $\sim70$ms time delay was observed between the times of maximum laser power and corresponding time of peak condensate fraction, giving some insight into the condensate formation time under this protocol.
To interpret the observed findings, Stoof and Bijlsma [@stoof2001dynamics] pioneered the use of a stochastic Gross-Pitaevskii equation as a model to study reversible condensate formation. Introducing a steep well in the potential, as done experimentally, is equivalent to altering the effective chemical potential, $\mu_{\textrm{eff}}$, of the system. Hence, in this numerical investigation, the reversible formation of a one-dimensional condensate was simulated by modulating the value of $\mu_{\textrm{eff}}$ as $\mu_{\textrm{eff}}(t) = \mu\textrm{sin}(\omega_D t)$, where $\mu$ is the chemical potential of the system in the absence of a well and $\omega_D$ is the frequency of the imposed periodic modulation in $\mu_{\textrm{eff}}(t)$. Stoof and Bijlsma analysed the evolution of the condensate density at the trap centre and qualitatively reproduced the delay in condensate growth observed by Stamper-Kurn *et. al* [@stamperkurn1998reversible] (but observing a slightly longer time of $\sim$100ms, which could be attributed to the different probed dimensionality and potential uncertainties of exact experimental parameters). That study provided the first numerical evidence of a repeated phase transition crossing in ultracold atomic gases. Since that work, the stochastic Gross-Pitaevskii model introduced by Stoof [@stoof2001dynamics], and its closely-related stochastic projected Gross-Pitaevskii equation [@Bradley2008; @Blakie2008], have been used to study a broad range of dynamical ultracold phenomena [@Blakie2008; @Proukakis2008; @Proukakis13; @berloff_brachet_14], including the study of quenched phase transitions across different geometries, dimensionalities and mixtures [@Proukakis03; @Proukakis_Schmiedmayer_2006; @weiler_08; @Proukakis09; @zurek_09; @damski_10; @das_12; @Su13; @De_2014; @McDonald2015; @Liu_2016; @gallucci2016engineering; @Kobayashi_16a; @Kobayashi_16b; @Eckel_2018; @liu_18; @Ota_2018; @comaron_19; @bland_marolleau_20; @liu_20], with direct successful modelling of phase transition experiments [@weiler_08; @liu_18; @bland_marolleau_20].
Numerical Model and Parameter Regime
------------------------------------
In this work we analyse the quenched system dynamics in terms of the Stochastic Projected Gross-Pitaevskii Equation (SPGPE), which describes the ‘classical’ field, $\Phi({\mathbf r},t)$, of all highly-populated modes up to a fixed energy cut-off. Its dynamics are governed by [@Blakie2008]: $$\begin{aligned}
i\hbar\frac{\partial\Phi}{\partial t} = \hat{\mathcal{P}}\Big\{ \Big(1- i\gamma\Big)\Big[-\frac{\hbar^2}{2m} \nabla^2 + g|\Phi|^2 - \mu \Big]\Phi + \eta\Big\}\,.
\label{eqn:spgpe}\end{aligned}$$ Here $-i \gamma$ corresponds to a dissipative/growth term, arising from the coupling of the ‘classical field’ modes $\Phi$ to the high-lying modes, which are treated as a heat bath. Consistent with other treatments [@comaron_19; @Ota_2018], we set $\gamma = 0.01$ in our current calculations, noting that typical physical evolution timescales are set by the scaled time $\gamma t$, controlling system dynamics and growth. The long-term evolution/steady-state is determined by the balancing of the kinetic energy contribution and the nonlinear interaction term $g |\Phi({\mathbf r}, t)|^2$ with the bath chemical potential $\mu$. The presence of the dynamical noise term $\eta$, associated with the collisional randomness of growth/decay processes differentiates each numerical run in a manner analogous to experimental shot-to-shot fluctuations. The noise correlations are given by $\langle \eta^{*}(\textbf{r},t)\eta(\textbf{r}',t') \rangle = 2\hbar\gamma k_{\textrm{B}}T\delta(t-t')\delta(\textbf{r}-\textbf{r}')$. Finally, $\hat{\mathcal{P}}$ is a projector which constrains the dynamics of the system within a finite number of macroscopically occupied modes, up to an ultraviolet energy cutoff $\epsilon_\mathrm{cut} (\mu,T) =k_\mathrm{B}T\log(2)+\mu$, with the mean occupation of the last included mode set to be $\sim 1$. This relation sets our numerical grid spacing to $\Delta x \le \pi/\sqrt{8m\epsilon_\text{cut}}$ [@Blakie2008].
In this work, we consider a two-dimensional, weakly interacting, homogeneous Bose gas, confined to a square box of sides $L_x = L_y = 100 \mu\textrm{m}$. Considering experimentally accessible regimes, our presented analysis is based on a gas of ultracold ^39^K bosons with mass ${m = 6.47 \times 10^{-26}}$ kg, a chemical potential $\mu = 1.92 k_B$nK and a temperature, $T = 50$nK [@footnote1]. The corresponding 2D interaction strength is given in terms of the tight transversal harmonic confinement of frequency $\omega_{\perp}$ by ${g_{2D}= g_{3D} / \sqrt{2\pi}\ell_\perp}$, where $g_{3D} = 4 \pi \hbar^2 a_{s}/m$, $a_s=7.36$nm is the background s-wave scattering length [@derrico2007feshbach], and ${\ell_\perp=\sqrt{\hbar/m\omega_\perp}}$ the transverse harmonic length. This yields a dimensionless coupling constant $g = m g_{2D} / \hbar^2 = \sqrt{8\pi}{a_s}/{\ell_\perp}$.
The location of the BKT phase transition is fixed by the relation [@prokofev2001critical] $$\label{eqn:g_crit}
\frac{\mu}{k_\mathrm{B}T} \bigg\rvert_\mathrm{BKT} \approx \frac{g_c}{\pi} \ln\left(\frac{C}{g_c}\right),$$ where $g_c$ denotes the critical value of the dimensionless strength at the BKT transition, and Monte-Carlo analysis gives the constant $C\sim 13.2$ [@Prokofev2002]. This relation sets the value of $g_c=1.83 \times 10^{-2}$ for our parameters. This coupling constant can be varied by changing the transverse confinement, and the parameters considered in this work relate to the range $1.83 \times 10^{-3} < g < 3.48 \times 10^{-2}$. Implicit in our description is the modulation of the scattering length through Feshbach resonance [@chin2010feshbach], or the transverse confining harmonic oscillator frequency $\omega_\perp$. All presented results correspond to an average over $\mathcal{N} = 50$ stochastic realizations, which is large enough such that the error bars on all figures are of the same size order as the marker itself.
Quench Protocol
---------------
Stimulated by the work of Refs. [@stamperkurn1998reversible; @stoof2001dynamics], we impose here a symmetric interaction quench, as follows: Firstly, we allow the system to dynamically equilibrate close to, and on the disordered side of, the BKT critical region. Specifically we choose to equilibrate initially to a value $g=1.9 g_c$ (which corresponds to an initial number of $N=3\times10^4$ atoms in the classical field). We verify that this state is in equilibrium via calculation of the first order correlation function, which exhibits the correct exponential decay law, and ensuring the vortex number reaches a steady state value. We then initiate a periodic quench in the interaction strength which is symmetric about the critical point, via $$\begin{aligned}
g(t) = g_c(1+\alpha\textrm{cos}(\omega_D t))\,,
\label{eqn:gt}\end{aligned}$$ where, $\omega_D$ is the driving frequency of the quench and $\alpha$ is the constant periodic amplitude. Based on our chosen initial state, we fix the modulation amplitude to $\alpha = 0.9$, implying that the quench proceeds between an initial state with $g=1.9 g_c$ (corresponding to $T>T_\text{BKT}$), and a final highly-ordered state with $g=0.1 g_c$ ($T<T_\text{BKT}$) [@footnote2].
Throughout this work, we examine how the system behaviour varies for different driving frequencies in the range $\omega_D = 2 \pi \times [1/40 ,\, 500]$Hz, which cover the entire behaviour from quasi-adiabatic ($ \omega_D = 2 \pi \times (1/40)$Hz) to extremely rapid ($ \omega_D = 2 \pi \times 500$Hz), which fully encompass the entire range of potential parameters for experimental investigation.
In presenting our results, we introduce the time-dependent deviation from criticality $$\begin{aligned}
\Delta g(t) = \frac{g(t)-g_c}{g_c},
\label{eqn:delg}\end{aligned}$$ which is constrained to oscillate between $+\alpha$ (disordered) and $-\alpha$ (ordered).
![Repeated quenches through the BKT phase transition, with time normalised to the quench period through the time scale $2\pi/\omega_D$. (a) Quench protocol \[Eq. \]. Dotted line through $\Delta g = 0$ separates the incoherent (thermal) ($\Delta g >0$) and (partly) superfluid ($\Delta g <0$) states. (b) Evolution of the c-field density averaged over space and stochastic realisations, compared to its initial value $\langle n(t=0) \rangle \equiv \langle n _{\textrm{bg}} \rangle$ for various driving frequencies $\omega_{D}$. Also shown are the equilibrium densities for the initial thermal state ($\text{eq}_\text{th}$, lower purple horizontal line) and the final-parameter part-superfluid system $\text{eq}_\text{c}$ (upper blue horizontal line) respectively, found by fixing $g(t)=g_\text{max}$ ($\Delta g=0.9$) and $g(t)=g_\text{min}$ ($\Delta g=-0.9$), and the equilibrium density $n_0(t)=\mu/g(t)$. The slowest quench and condensate equilibrium data are reduced by a factor of 4 for visual aid. For a given driving frequency, the phase lag between the $i_{\textrm{th}}$ trough of $\Delta g$ and the $i_{\textrm{th}}$ peak of the average density is denoted $\phi_{\textrm{lag}}^{n_i}$. The vertical grey dashed line throughout all three panels highlights the time of the first trough in $\Delta g$. (c) Evolution of the average number of vortices, $\langle N_{v} \rangle$, for the same quench frequencies and equilibria as above.[]{data-label="fig:quenchprot"}](figure_1.pdf){width="0.9\columnwidth"}
This quench protocol can be easily visualized in Fig. \[fig:quenchprot\](a). Positive values of $\Delta g$ indicate that $g > g_{\textrm{c}}$, with the system in the disordered (thermal) phase, while $\Delta g < 0$ indicate that $g < g_{\textrm{c}}$, which initiates the (quasi)-condensate formation process. For better understanding, in what follows we also briefly compare our dynamical periodic quench results to the corresponding equilibrium regimes in the two limiting cases of our quench cycle.
Dynamical Results {#sec:dyn}
=================
The system dynamics under periodic quenching can be characterized in terms of its density, vortex number, occupation spectrum and coherence, to which we next turn our attention.
Density and Vortex Delayed Oscillations
---------------------------------------
We start our analysis by considering the impact of the periodic interaction strength modulation shown in Fig. \[fig:quenchprot\] on density and vortex number. Specifically, Fig. \[fig:quenchprot\](b) shows the evolution of the average modulation $\langle n \rangle - \langle n _{\textrm{bg}} \rangle$ of the c-field density $\langle n \rangle$, compared to its initial value $\langle n _{\textrm{bg}} \rangle$, averaged over space and stochastic realisations, for a range of driving frequencies, while Fig. \[fig:quenchprot\](c) shows the corresponding average evolution of the vortex number in the system. It is well-known that above the BKT phase transition ($\Delta g>0$) the phase of the system is random, corresponding to an exponentially-decaying phase correlation function: this incoherent state can be thought of as a system with a large number of ‘free’ vortices, whose exact number is fixed by the system size, grid resolution, atomic mass, temperature and interaction strength (the number itself is not important). As we scan the system across different equilibrium configurations, we find that the number of vortices at equilibrium decreases abruptly across the BKT critical region, reaching a very low, or even zero value as the system approaches its pure superfluid limit ($g \ll g_c$) [@foster2010vortex; @gawryluk_19; @comaron_19].
Let us first consider the limiting (idealized) cases of extremely slow and rapid driving: For extremely slow driving, we expect the system to evolve adiabatically through corresponding equilibrium states, oscillating between the initial purely incoherent state with $\langle n \rangle = \langle n _{\textrm{bg}} \rangle$ which is filled by vortices (here $\langle N_{v} \rangle \sim 160$), and a higher value ($\langle n \rangle - \langle n _{\textrm{bg}} \rangle \approx 80\mu$m$^{-2}$ ), which corresponds to the equilibrium density value at $g=g_\text{min}=0.1g_c$, and $T=50$nK, with such state having $\langle N_{v} \rangle = 0$. The timescale for such periodic change would be set by the slow drive frequency $2 \pi / \omega_D$, in the limit $\omega_D \rightarrow 0$. In fact, the equilibrium density is simply given by the time-dependent homogeneous solution $n_0(t)=\mu/g(t)$, which the idealized case $\omega_D\to0$ follows in Fig.\[fig:quenchprot\](b). The corresponding averaged densities and vortices in the two limiting cases are shown Fig.\[fig:quenchprot\](b)-(c) by the horizontal purple (initial incoherent state with $g=g_{\rm max}$) and blue (part-superfluid state with $g=g_{\rm min}$) lines. In the opposite extreme limit of very rapid driving ($\omega_D \rightarrow \infty$), the system has no time to adjust to the driving parameters, and remains in an effective incoherent steady-state close to the initial state, i.e. its density and vortex number closely mimic the flat behaviour of the purple horizontal lines, as shown here for $\omega_D/2 \pi = 500$Hz.
Neither of those regimes is intrinsically interesting to study, so our numerical analysis focuses on a broad range of intermediate driving frequencies, from the quasi-adiabatic to the dynamically-driven regimes identified below.
The slowest quench analyzed in detail in this work is shown by the red curves in Fig. \[fig:quenchprot\](b)-(c), and corresponds to a drive frequency $\omega_D = 2 \pi \times (1/40)$ Hz. This has been selected such that both the density and vortex number closely resemble the adiabatic case (with $\langle N_{v} \rangle \approx 0$), but strictly speaking the frequency of the drive is such that the state reached at $g_{\rm min}$ is not yet fully equilibrated, consistent with the state likely to arise under experimental driving conditions: we label such state here as quasi-adiabatic. We have explicitly verified that driving the system with such $\omega_D$ for half a period (i.e. up to $t = \pi / \omega_D$), and subsequently removing the drive and waiting sufficient time for the system to relax indeed recovers the corresponding equilibrium state at $g_{\rm min}$.
For such slow quenches, the maximum system density is achieved almost exactly at each local minimum $\Delta g(t)$, while density minima temporally coincide with maxima in $\Delta g(t)$. In other words, the density modulation oscillations are in phase with the underlying drive. Moreover, the vortex number rapidly decreases acquiring its minimum value at the local $\Delta g$ minimum when the density modulation is also maximised.
As the oscillation frequency increases beyond the adiabatic limit, the overall dynamical evolution becomes faster. To best characterize the underlying dynamics, we consider the dynamical density and vortex number modulation as a function of time scaled to the drive timescale $2 \pi / \omega_D$: this enables evolution graphs to be compared against the same $\Delta g(t)$ graph of Fig. \[fig:quenchprot\](a). In such scaled time units, the evolution is comparatively slowed down with increased driving frequency. As a result of the faster driving, the system no longer has sufficient time to adjust to the same density maximum within each cycle, with the oscillating density modulation going out of phase (lagging behind) the external periodic driving, by an amount $\phi_{\rm lag}^{n_i}$, corresponding to the phase lag of the $i^{\text{th}}$ trough in $\Delta g(t)$. The local density maximum is thus reached at a time after $\Delta g(t)$ has reached its minimum, when the interaction strength is $g_\text{min}<g(t)<g_c$.
This effect becomes pronounced with faster driving, which leads both to a lower local density maximum being reached, and a longer time delay between driving and density growth. For intermediate external driving frequencies, the first density maximum reached is lower than the subsequent ones, with the system evidently requiring multiple driving cycles for the temporally-local density maximum to plateau to its overall maximum attainable value under periodic driving. Moreover, the rapid evolution implies that the overall observed density modulation amplitude per cycle is decreasing with increasing $\omega_D$, so that – as the density maximum is gradually increasing – the corresponding density modulation minimum is no longer constrained to coincide with the initial density value: such features are most evident when comparing $\omega_D/2\pi=1/2$Hz (blue) and $\omega_D/2\pi=5/3$Hz (brown) curves in Fig. \[fig:quenchprot\](b). This effect becomes more pronounced with higher values of $\omega_D$, until the driving becomes so rapid that the system has practically no time to adjust.
Such dynamical behaviour can be better understood by looking at the corresponding evolution of mean vortex number $\langle N_{v} \rangle$ in the system. As the driving frequency increases, the rate of decrease of vortices in a given driving cycle becomes significantly lower, with the local vortex minimum being significantly delayed from the corresponding $\Delta g$ minimum. The delay of the first local vortex minimum $\phi_\text{lag}^{V_1}$ as measured from the $\Delta g$ minimum at $t=\pi/\omega_D$ is in fact significantly pronounced, and occurs clearly after the corresponding density modulation maximum (which was itself found to phase-lag behind the external driving). This can be understood in terms of the additional phase-ordering time required for the annihilation of a vortex-antivortex pair, which is a competing effect to the external drive. For faster quenches the gradual vortex number decrease occurs over multiple quench cycles and does not necessarily reach $\langle N_{v} \rangle \sim 0$ as evident from the brown quench in Fig. \[fig:quenchprot\](c). (In the idealized limiting case of extremely rapid quenches (pink), the system has insufficient time to react, exhibiting only a marginal periodic decrease in vortex number).
The above discussion has identified an evident difference in the evolutions of density and vortex number during the driven quench, identifying two rather different behaviours with respect to the driving frequency. This points to the existence of a critical (resonant) frequency which separates these two regimes. We can identify a characteristic relaxation, or ‘resonant’ frequency through the identification of a characteristic timescale in the system. A relevant timescale in the system is [@McDonald2015] $$\begin{aligned}
\tau_0=\frac{\hbar}{\gamma|\mu|}\approx 0.4\,\text{s}\,
\label{eqn:relax1}\end{aligned}$$ This timescale corresponds to a relaxation time towards the symmetric equilibrium above the critical point, while below the critical point it marks a timescale for dynamical instability, through exponential growth of small fluctuations.
This timescale in turn implies a characteristic system frequency $\omega_0$, defined as $$\begin{aligned}
\omega_0=\frac{1}{\tau_0}\approx2\pi\times0.4\,\text{Hz}\,.
\label{eqn:relax}\end{aligned}$$ This frequency is marked across Fig. \[fig:taus\] (and relevant subsequent figures) as a vertical red dashed line, and clearly marks the boundary between quasi-adiabatic driving ($\omega_D < \omega_0$) and dynamically-driven systems ($\omega_D > \omega_0$). The role of this intrinsic ‘resonant’ frequency on momentum spectrum and coherence is further highlighted below.
To characterize the above behaviour quantitatively, we next consider the mean phase lag of both density and vortex evolution as a function of the driving frequency $\omega_D$, as shown in Fig. \[fig:taus\](a). The average vortex number phase lag is greater than the density phase lag for all driving frequencies considered, in keeping with observations from Fig. \[fig:quenchprot\]. A phase lag of $\phi_\text{lag}/2\pi=0.25$ corresponds to the external driving parameter crossing of the BKT threshold, which the density phase lag never exceeds, implying the maximum density per cycle is achieved during the coherent phase. However, the minimum in $\langle N_{v} \rangle$ occurs later whilst in the thermal phase, suggesting that the vortex number is not in equilibrium before the next cycle occurs.
Given that the vortex number decay exhibits oscillatory behaviour, and that for a large range of probed driving frequencies it actually requires multiple such oscillations before the vortex number decreases to its steady-state (potentially non-zero) values, it is also interesting to characterize the time taken for the vortex to decay to a near-zero final vortex number. Similarly, the average density also requires several cycles to saturate to its maximum value. To quantify these effects, we have here chosen arbitrarily to define a timescale, $n_\text{cycle}$, as the number of quench cycles at which the measured quantity reaches 90% saturation of its steady-state values. This corresponds to the time the vortex number first reaches 10% of its initial value, or the density reaches 90% of its final maximum value.
![Quench dependent time shifts probed here in the physically interesting range $0.05 \lesssim (\omega_D / \omega_0) \lesssim 10$. (a) The mean phase shift for the density ($\phi^n$) and vortex number ($\phi^V$) between the $i^{\textrm{th}}$ peaks of $\Delta g$ and the peak average density, as a function of $\omega_D$. Error bars are smaller than the marker size. Points corresponding to the quench frequencies considered previously are coloured according to the legend of Fig. \[fig:quenchprot\]. (b) Average number of oscillation cycles, $n_\text{cycle}$, between the first quench and density (vortex number) reaching 90% of its maximum value (10% of its initial value). The relaxation frequency $\omega_0$ (Eq. ) is marked by a vertical dashed red line, with the grey-shaded area corresponding to the ‘quasi-adiabatic’ regime $\omega_D < \omega_0$. Horizontal black lines in (a) and (b) respectively denote the values of $\phi_{\rm lag} / 2 \pi = 1/4$, marking the transition from the observed phase lag occurring within a quarter of the driving cycle (i.e. while the system is within the ‘coherent’ phase), and $n_{\rm cycle} =1$, marking the transition between vortex number counts that are independent and dependent on cycle number. For $\omega_D / \omega_0 \gtrsim 100$ (not shown here) the extremely modest response of the system to the drive leads to practically no change in density, or phase. []{data-label="fig:taus"}](figure_2.pdf){width="1\columnwidth"}
{width="95.00000%"}
Apart from the very slow quenches, Fig. \[fig:taus\](b) shows a near-linear increase in the number of cycles required for the system to reach the steady-state as a function of driving frequency. Moreover, even though it may take multiple cycles to reach that value, the linear-like curve implies a near constant duration of time to saturation ($2\pi n_\text{cycle}/\omega_D$) for $\omega_D \gtrsim \omega_0 $. This figure thus confirms the two distinct regimes either side of $\omega_D \sim \omega_0$ probed here in the interesting range $0.05 \lesssim (\omega_D / \omega_0) \lesssim 10$. For relatively slow driving $\omega_D < \omega_0$, the vortex number and average density saturate within a single quench cycle, suggesting maximum coherence has been reached within one cycle and the vortex number and density evolution are only dependent on the time during a cycle when they are measured, and not the total number of cycles completed. However, when driving faster than $\omega_0$, the system saturates after multiple quench cycles, with extreme limits of $(\omega_D / \omega_0) \gg 10$ (not shown here) leading to only very small variations from the initial (incoherent) state.
We now turn our attention to the evolution of the momentum spectrum as a function of different driving frequencies.
Momentum Evolution and Condensate Mode Dynamics
-----------------------------------------------
The density analysis considered above focussed on the entire classical field region, which encompasses both the condensate–which requires macroscopic occupation of the lowest lying momentum mode–and those higher-lying macroscopically occupied modes affected by its presence. To quantify the emerging (quasi)condensate dynamics, we thus consider the evolution of the momentum distribution.
Firstly, we define the fractional momentum occupation of mode $k$ via $$\begin{aligned}
\hat{n}_k(t) = \frac{ \langle \tilde{n}_k(t) \rangle_{\mathcal{N}} }{ \sum\limits_{k} \langle \tilde{n}_k(t) \rangle_{\mathcal{N}} }\,,
\label{eqn:nknormal}\end{aligned}$$ where $\langle\cdot\rangle_\mathcal{N}$ denotes averaging over stochastic realisations. The evolution of the normalized momentum density $\hat{n}_k$ is plotted in Fig. \[fig:momentumspace\](a) for different driving frequencies from slow (left, $\omega_D/\omega_0 \sim 0.05$) through to fast (right, $\omega_D/\omega_0 \sim 10$) quenches. For a truly adiabatic evolution, proceeding slowly enough that the system passes through [*fully equilibrated*]{} states during its entire evolution, one would expect to see the periodic emergence of a ‘condensate’ mode, in the sense of a highly populated $k = 0$ mode Indeed, this emerges for the slowest quasi-adiabatic periodic quench \[panel (a)(i)\], whose modes up to $k \sim 0.1 \mu$m are maximally populated at times $t = (2\pi/\omega_D )(2m+1)/2$ (where $m$ is an integer), when $\Delta g$ has a local minimum.
This regime clearly shows a periodic reversible condensate formation process, qualitatively similar to the behaviour observed in harmonic dimple microtraps cycled across the phase transition [@stamperkurn1998reversible; @stoof2001dynamics]. As the quench frequency increases beyond $\omega_0 \sim 2 \pi \times 0.4$ Hz we expect coherence to form over multiple quench cycles. Indeed for $\omega_D \sim 1.25 \omega_0$ \[panel (a)(ii)\], we find the initial momentum peak after a single quench cycle to be significantly broadened around $k \sim 0$, consistent with the existence of multiple highly-occupied modes (and the gradual emergence of a quasi-condensate). Further increasing the driving frequency leads to coherence only appearing after multiple cycles, consistent with earlier findings, due to the more gradual decrease of vortex numbers which–for faster quenches–becomes significant only after multiple driving cycles through the critical point.
Example slices of $\hat{n}_k$ vs. $k$ (for $k>0$), at extreme limits of $\Delta g$ are shown in Fig. \[fig:momentumspace\](b) for both a relatively slow (red; corresponding to a(i)) and a relatively fast (green; corresponding to a(iv)) drive. As a guide for the eye, solid grey lines show the following limiting behaviour: a thermal field is expected to follow an $e^{-\xi k}$ scaling, where $\xi$ is the coherence length from the first order correlation function $g^{(1)}(r)$ for a thermal cloud [@footnote_g1], and a fully equilibrated 2D Bose gas is expected to exhibit a $k^{-2}$ scaling [@pitaevskii2003bose]. Our findings show that the driven states oscillate between these limiting cases. As a comparison, we show the extracted momentum distribution for the thermal field at $t=0$ (purple circles), and of the Bose gas in equilibrium showing the expected $k^{-2}$ behaviour (blue dash dotted), and find excellent agreement to the expected scaling.
Considering first the slow driving case \[case (a)(i)\], we note that the system exhibits a monotonically-decaying $\hat{n}_k$ \[Fig. \[fig:momentumspace\](b)\] with more than 95% condensate fraction \[Fig. \[fig:momentumspace\](c)\] both at the first (solid red line) and subsequent $\Delta g$ minima (dotted red line). This is in contrast to the corresponding behaviour at $\Delta g$ maxima (dashed red line) which shows decreasing occupation as $k \rightarrow 0$.
Considering now a much faster quench \[case a(iv)\], the first $\Delta g$ minimum \[solid green line\] is clearly not condensed, as evident from the low $k=0$ occupation in Fig. \[fig:momentumspace\](c)(i) \[green open circle\], and still resembles the initial thermal field distribution at $t=0$. After many driving cycles, such a rapidly-driven system develops clear evidence of non-equilibrium condensation, with increasing occupation into the lower $k$ modes \[dotted and dashed green lines\] and $\sim50\%$ occupation into the $k=0$ mode (Fig. \[fig:momentumspace\](c)(i)), distinct from the expected thermal distribution and with little variation between the maximum (dotted) and minimum (dashed) cases of $\Delta g$. Remarkably, for fast quenches the system maintains a coherent fraction regardless of the drive.
In all cases, it’s worth noting that the spectrum at large $k$ acquires a power-law behaviour with an exponent resembling $\sim-1$ [@footnote_scaling], clearly distinct from both expected thermal and Bose gas scalings. We attribute such numerically observed ‘anomalous’ scaling to the highly non-equilibrium nature of the driven system, a feature that could be tested in experiments.
The maximum of the condensate fraction as a function of driving frequency is shown in Fig. \[fig:momentumspace\](c)(i), and the total atom number in (c)(ii). The open circles show the condensate fraction measured at the first peak in the total density, whereas the open squares show the condensate fraction measured at an arbitrary peak at sufficiently late times. For faster driving frequencies the condensate fraction [*increases*]{} over time, as a result of the repeated crossing of the phase transition, thus leading to a non-equilibrium steady-state with [*irreversible*]{} growth. This can be attributed to the total time spent above the phase transition, $\pi/\omega_D$, being smaller than the relaxation time $\tau_0$, and hence a small fraction of the atoms remain Bose condensed after each repeated crossing of the phase transition. This effect disappears both for small driving frequencies, where the condensate fraction follows the equilibrium value expected from an instantaneous evaluation of the dissipative GPE, and for very rapid frequencies ($\omega_D / \omega_0 \gtrsim 100$ – beyond the scale of this graph) which do not allow for any noticeable coherence to form even after multiple quench cycles. In this limit the condensate feels a time averaged drive, $\langle g(t)\rangle_t=g_c$, and the c-field remains incoherent.
Interestingly, the smallest condensate fraction occurs around $\omega_D \approx \omega_0$, when the competing driving and relaxation are in [*resonance*]{} with one another, leading to a highly non-equilibrium partly-superfluid state, an effect we explore in more detail in Sec. \[sec:resonance\].
Evolution of Phase Coherence {#sec:g1}
----------------------------
We now discuss the effect of driving on the spatial coherence of the system.
For a homogeneous 2D system, the spatial correlation function, or normalised first order correlation function, defined by [@pitaevskii2003bose]: $$\begin{aligned}
g^{(1)}(r)=\frac{\langle\Phi^*(r_0)\Phi(r_0+r)\rangle}{\sqrt{\langle|\Phi(r_0)|^2\rangle\langle|\Phi(r_0+r)|^2\rangle}}\,,
\label{eqn:g1}\end{aligned}$$ is known (at equilibrium) to exhibit a transition from exponential (above the critical region) to algebraic-order decay (below the critical region), a characteristic of the BKT phase transition [@Nazarenko2014; @prokofev_two-dimensional_2002; @foster2010vortex; @gawryluk_19; @comaron_19; @hadzibabic2006berezinskii]. We have indeed verified such behaviour in the steady-state profiles for the two limiting equilibrium cases $g=g_\text{max}$ and $g=g_\text{min}$. For a thermal cloud $g^{(1)}(r)\sim e^{-r/\xi}$, where $\xi$ is the coherence length \[see also Fig. \[fig:momentumspace\](b)\]. Here we are instead interested in investigating the effect of the periodic driving on the emerging maximum spatial phase coherence. With this in mind, we perform a detailed analysis of the system coherence when $\Delta g =-0.9$. To account for the fact that the system does not necessarily relax after a single driving cycle (for faster driving), we consider the spatial phase coherence at two times. These correspond to: (a) the end of the first driving half-period ($t_{p_{1}}$) when the system reaches its minimum $\Delta g$ value after a single phase transition crossing – this is shown in Fig. \[fig:g1\](a). This corresponds to the same point in the driving cycle irrespective of $\omega_D$, thus amounting to different physical times. (b) We also measure $g^{(1)}$ at a fixed absolute late time for all $\omega_D$, corresponding to a sufficiently long time such that all considered cases have already acquired their maximum steady-state values. Specifically: we investigate $g^{(1)}(r)$ at the $\Delta g$ trough at, or just after the time $\gamma t_* \approx 0.12$s: such a time corresponds to the first trough in the limiting case of quasi-adiabatic driving ($\omega_D = 2 \pi (1/40)$ Hz $=0.05 \omega_0$) and the 60$^\text{th}$ trough, at which a non-equilibrium steady-state has been reached for the ultrafast quench $\omega_D = 2 \pi \times5\text{Hz} = 10 \omega_0$ \]. This is shown in Fig. \[fig:g1\](b).
![Growth of coherence through repeated quenches. (a) First order correlation function, $g^{(1)}(r)$, evaluated at $t=t_{p_1}$, the time of the first peak in the average density in Fig. \[fig:quenchprot\]. Each curve is fitted with an algebraic function, $g^{(1)}(r) \sim r^{-\alpha}$, plotted using a dashed line. (b) Correlation function evaluated at the density peak closest to $\gamma t_* = 0.12$s. (c) Correlation function evaluated at $r_\star=40 \mu$m, \[indicated by the dashed vertical grey lines in panels (a)-(b)\], as a function of time. Error bars are located where $g^{(1)}$ is measured at a peak in the average density. Data for $\omega_D/2\pi=1/40$ (red) and $5/33$ (grey) Hz extend far beyond $\gamma t=0.12$s, with corresponding red/grey bands denoting error bars over the total range covered. Plotted error bars indicate one standard deviation of $g^{(1)}(r)$. []{data-label="fig:g1"}](figure_4.pdf){width="1\columnwidth"}
Fig. \[fig:g1\](a) shows that for the quasi-adiabatic driving \[$\omega_D = 2 \pi \times (1/40)$Hz\], the system exhibits very strong spatial coherence at the first $\Delta g$ minimum, which is perfectly fit with an algebraic-decay function $g^{(1)}(r) \sim r^{-\alpha}$, where $\alpha = 6.8\times10^{-3}$. This correlation is actually identical in all subsequent peaks with the same driving. Although this is highly coherent, consistent with the complete absence of vortices in most numerical realizations, we nonetheless note here once again that this system is not yet fully equilibrated. This is because the emerging state is found to exhibit a momentum spectrum distinct from the expected $k^{-2}$ equilibrium spectrum, whereas a $k^{-2}$ spectrum does emerge upon allowing this same system to relax (without further driving) for a significantly longer timescale at $\Delta g = -0.9$.
Increasing the driving frequency leads to the establishment of less coherence in the system after a single cycle, with coherence nonetheless building up gradually after multiple cycles. This can be seen by comparing the computed $g^{(1)}(r)$ between Figs. \[fig:g1\](a) and (b). Although very fast quenches lead to very small coherence after a single drive half-cycle, probing the correlation function instantaneously at $g=g_\text{min} \ll g_{c}$ nonetheless still facilitates a very good algebraic-decay fit (but with a higher value of $\alpha$, implying a highly non-equilibrium state). Thus, in order to compare with the maximal coherence imparted at steady-state, we analyze all plots here in terms of an algebraic-decay fit.
The increase of the (algebraic) decay rate of $g^{(1)}(r)$ with increasing $\omega_D$ seen in Fig. \[fig:g1\](a) should come as no surprise, since the build-up of significant coherence requires the decay of all vortices in the system, whereas all probed quenches except the quasi-adiabatic (red) one still have at least some vortices present at the first half-cycle \[see Fig. \[fig:quenchprot\](c)\]. Remarkably, even the presence of a few vortices is enough to significantly destroy the system’s overall phase coherence, as evident from the blue curves \[$\omega_D = 2 \pi \times (1/2)$ Hz\] in Fig. \[fig:g1\](a) and Fig. \[fig:quenchprot\](c).
An interesting and rather distinctive feature arises when looking at corresponding late-time peaks shown in Fig. \[fig:g1\](b), after allowing the periodically-driven system to reach its non-equilibrium steady-state, following a sufficiently large number of drive cycles, and repeated crossings of the phase transition in both directions. We have already shown earlier that for sufficiently fast driving, the vortex number curve decreases periodically in an oscillating manner, but with a rapidly decaying envelope, requiring multiple cycles for a significant decrease of its mean vortex number $\langle N_{v}\rangle$. As a result, coherence builds up gradually in such systems, and this was to be expected. The interesting emerging feature here is that at steady-state, there is an optimum driving frequency, above which the final acquired system coherence (at one of the $\Delta g$ peaks) at steady-state grows over longer distances again.
This is easily seen by comparing Fig. \[fig:g1\](a)-(b). Focussing at the late time $\gamma t_* \sim 0.12$s, we see that increasing the frequency (in Hz) from $\omega_D / 2 \pi = 1/40$ (red) through $5/33$ (brown) to $1/2 \sim O(\omega_0 / 2 \pi)$ (blue), $g^{(1)}(r)$ decays faster spatially with increasing drive frequency. However, driving the system even faster than $\omega_0$ periodically across a sufficiently large number of phase-transition cycles leads to an [*increase*]{} in the system phase coherence, with the spatial coherence at the maximum distance (half the box length) of the $\omega_D / 2 \pi = 5$ Hz (green) being comparable to that of $\omega_D / 2 \pi = 5/33$ Hz (grey), which is driven 33 times slower (and whose coherence does not change between consecutive peaks, having already saturated at its maximal coherence at the end of the first half-cycle).
Such behaviour becomes more evident in Fig. \[fig:g1\](c), when examining the temporal evolution of the value of the phase correlation function near the edge of the system, at $g^{(1)}(r_\star=40\mu$m). Slow quenches ($\omega_D/2\pi<0.5$ Hz) allow maximal coherence to be effectively established already after a single quench half-cycle: this is indicated by the red/grey horizontal lines/bands in Fig. \[fig:g1\](c). For faster driving than that, the system initially reacts rather slowly, with the coherence growth rate gradually picking up around $\gamma t \sim 0.01$s, eventually saturating at a higher value (well before the $\gamma t \sim 0.12$s used for the comparison in Fig. \[fig:g1\](b)). Interestingly the extra growth of coherence after multiple cycles is very small around the frequency $\omega_D / 2 \pi \sim 1/2$ Hz (blue curve), but increases noticeably for quenches few times faster than that, as evident from the enhanced values shown by the yellow, brown and green curves.
Resonances with intrinsic timescales {#sec:resonance}
------------------------------------
The behaviour of the spatial correlation function is further characterized in comparison to the relaxation frequency $\omega_\text{0}=2\pi\times0.4$ Hz in Fig. \[fig:analysis\]. Here we plot (a) the steady-state value ($t=t_*$) of $g^{(1)}$ near the box edge ($r=r_\star$), (b) the fractional increase of its value at $t_*$, compared to that at its first peak, (c) the value of the power-law decay exponent $\alpha$, and (d) the density phase lag $\gamma \tau^{n}_\text{delay}= 2\pi\phi^n_\text{lag}/\omega_D$ (a) (corresponding to Fig. \[fig:taus\](a), but now plotted in terms of average delay time which offers a different perspective). This figure clearly highlights the importance of the critical driving frequency, corresponding to ‘resonant’ driving distinguishing the slow and fast driving regimes. This resonant driving frequency is well explained in terms of the frequency $\omega_{0}$. Looking at the various subplots, we easily infer that maximal coherence is achieved in the quasi-adiabatic regime (leftmost, red points in all subplots), with $g^{(1)}(r_\star=40\mu m,\, \gamma t_*=0.12s)$ $\sim 0.96$ and $\alpha = 6.8\times 10^{-3}$, which is, as expected for such a low-temperature state with $T/T_{BKT} \sim 0.15$, much less than the value $\alpha = 0.25$ occurring at equilibrium for $g=g_c$ [@Nelson1977].
As the system is driven faster, the late-time coherence at the system edge rapidly decreases (down to $\sim 0.11$) \[Fig. \[fig:analysis\](a)\], with such value being already reached after a single quench half-cycle: the latter can be inferred by looking at the fractional change in the value of the correlation function at $r_\star$ from the first peak up to a late converged peak at $t_*$. Correspondingly, faster driving leads to an increasing delay (phase lag) of density growth, as shown in Fig. \[fig:analysis\](d). Interestingly, an algebraic power-law fit $g^{(1)}(r,\,t) \sim r^{-\alpha}$ of the correlation function at $\gamma t_* \sim 0.12$s – when the system clearly possesses a non-zero occupation of the $k=0$ mode – leads to a significant increase in the value of $\alpha$, much exceeding the equilibrium value of (1/4). This latter observation provides strong evidence of the non-equilibrium nature of the achieved steady-state, and points to the strong interplay between the system’s external driving across the phase transition (quantified in $\omega_D$), and its ability to adjust (quantified in $\omega_0$) to an averaged behaviour across two vastly different equilibrium states located well above and well below the critical region.
![Resonances of observables with the relaxation time. (a) Correlation function evaluated at $r_{\star} = 40\mu$m, $\gamma t_* \approx 0.12$s, see Fig \[fig:g1\]. (b) Percentage change in $g^{(1)}(r_\star,t)$ from $t_1$ to $t_*$ (early to late times). (c) Exponent from algebraic fits of Fig. \[fig:g1\](b). (d) Average time delay in the average peak density.[]{data-label="fig:analysis"}](figure_5.pdf){width="1\columnwidth"}
When driving the system even faster than that, the system loses its ability to adjust rapidly, as it starts experiencing an averaged quench effect. The overall steady-state coherence at the trap edge starts increasing again with increasing frequency \[Fig. \[fig:analysis\](a)\], with an increase of as much as 50% over its initial (but rather low) acquired value at the first quench half-cycle \[Fig. \[fig:analysis\](b)\]. As the system has less and less time to react to the external driving, it effectively only exhibits a time-averaged effect, and so becomes less non-equilibrium, consistent with a decreasing value of $\alpha$ with faster driving $\omega_D > \omega_0$ \[Fig. \[fig:analysis\](c)\]. As a result of the rapid driving, the overall density evolution time-delay now also decreases significantly (in ‘absolute’ time $\gamma t$ \[Fig. \[fig:analysis\](d)\]), which is testament to the very tiny increase of the system density facilitated.
Nonetheless, it is important to note that this delay time is actually a significant fraction of the drive time, as evident from Fig. \[fig:taus\](a), revealing that the density delay time for the fastest quench considered here is 20% of the quench period, while the vortex delay time \[not shown here, see Fig. \[fig:taus\](a)\] exceeds 60% of the period.
Much faster driving than that ($\omega_D / \omega_0 \gtrsim 100$) \[not shown here\] leads to a practically monotonic time-averaged evolution and a minor decrease of about 30% in vortex number, with an associated increase of less than 10%in the population of the $k=0$ mode, indicating a vortex-filled, low-coherence, non-equilibrium state. Moreover, driving at an idealized ‘infinite’ frequency $\omega_D / \omega_0 \sim 1000$ was found to lead to practically no change to the system’s initial configuration.
Conclusions {#sec:conclusion}
===========
We have considered the dynamical response of a homogeneous two-dimensional ultracold Bose gas under periodic quenches of its interaction strength through the Berezinskii-Kosterlitz-Thouless phase transition at a driving frequency $\omega_D$. We have identified an intrinsic system response frequency $\omega_0$ and demonstrated that resonant driving leads to a highly non-equilibrium state exhibiting only limited coherence growth, compared to when driving the system on either side of this resonance. Focusing on the most interesting regime of driving at a frequency within one order of magnitude from the resonant value, we characterized the system response in terms of the driving frequency, by analysing the evolution of densities, vortices, condensate fractions, spectra and coherence.
Specifically, we identified two distinct driving regimes of experimental relevance: Driving at a frequency $\omega_D$ much smaller than $\omega_0$ gives rise to quasi-adiabatic dynamics, with maximum coherence achieved after a single quench half-cycle, even if the system has not had sufficient time to fully equilibrate in momentum space (with the limiting case of extremely slow drive, $\omega_D \rightarrow 0$, corresponding to the adiabatic regime). Under (quasi)-adiabatic conditions it is possible to analytically describe the evolution of the average density.
In the opposite regime, driving the system faster than the system response frequency leads to a dynamically-driven non-equilibrium system, whose coherence grows gradually through multiple quench cycles. Interestingly, such driving can lead to a significant condensate fraction and enhancement in coherence: for example, even for large driving frequencies $\omega_D/\omega_0 \sim 10$, the coherence can grow up to 50% over the range of the box, despite only achieving $<1\%$ coherence after the first quench. As expected, quenching much faster than that (e.g. $\omega_D / \omega_0 \gtrsim 100$) leads to practically no generation of coherence, with the system entering an incoherent, vortex-filled steady-state.
Our study extends earlier works on cyclic phase transition crossing, and may help guide future non-equilibrium driven quench experiments and lead to the generation of interesting non-equilibrium steady-states of ultracold atoms with partial superfluidity.
Data supporting this publication is openly available under a Creative Commons CC-BY-4.0 License on the data.ncl.ac.uk site [@data].
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Gerasimos Rigopoulos for discussions. KB acknowledges support from a Newcastle University summer studentship. TB thanks EPSRC Doctoral Prize Fellowship Grant No. EP/R51309X/1. We acknowledge financial support from the Quantera ERA-NET cofund project NAQUAS through the Engineering and Physical Science Research Council, Grant No. EP/R043434/1.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present the first measurements of charge-dependent correlations on angular difference variables $\eta_1 - \eta_2$ (pseudorapidity) and $\phi_1 - \phi_2$ (azimuth) for primary charged hadrons with transverse momentum $0.15 \leq p_t \leq 2$ GeV/$c$ and $|\eta| \leq 1.3$ from Au-Au collisions at $\sqrt{s_{NN}} = 130$ GeV. We observe correlation structures not predicted by theory but consistent with evolution of hadron emission geometry with increasing centrality from one-dimensional fragmentation of color strings along the beam direction to an at least two-dimensional hadronization geometry along the beam and azimuth directions of a hadron-opaque bulk medium.'
author:
- 'J. Adams'
- 'M.M. Aggarwal'
- 'Z. Ahammed'
- 'J. Amonett'
- 'B.D. Anderson'
- 'D. Arkhipkin'
- 'G.S. Averichev'
- 'S.K. Badyal'
- 'Y. Bai'
- 'J. Balewski'
- 'O. Barannikova'
- 'L.S. Barnby'
- 'J. Baudot'
- 'S. Bekele'
- 'V.V. Belaga'
- 'A. Bellingeri-Laurikainen'
- 'R. Bellwied'
- 'J. Berger'
- 'B.I. Bezverkhny'
- 'S. Bharadwaj'
- 'A. Bhasin'
- 'A.K. Bhati'
- 'V.S. Bhatia'
- 'H. Bichsel'
- 'J. Bielcik'
- 'J. Bielcikova'
- 'A. Billmeier'
- 'L.C. Bland'
- 'C.O. Blyth'
- 'S-L. Blyth'
- 'B.E. Bonner'
- 'M. Botje'
- 'A. Boucham'
- 'J. Bouchet'
- 'A.V. Brandin'
- 'A. Bravar'
- 'M. Bystersky'
- 'R.V. Cadman'
- 'X.Z. Cai'
- 'H. Caines'
- 'M. Calderón de la Barca Sánchez'
- 'J. Castillo'
- 'O. Catu'
- 'D. Cebra'
- 'Z. Chajecki'
- 'P. Chaloupka'
- 'S. Chattopadhyay'
- 'H.F. Chen'
- 'J.H. Chen'
- 'Y. Chen'
- 'J. Cheng'
- 'M. Cherney'
- 'A. Chikanian'
- 'H.A. Choi'
- 'W. Christie'
- 'J.P. Coffin'
- 'T.M. Cormier'
- 'M.R. Cosentino'
- 'J.G. Cramer'
- 'H.J. Crawford'
- 'D. Das'
- 'S. Das'
- 'M. Daugherity'
- 'M.M. de Moura'
- 'T.G. Dedovich'
- 'M. DePhillips'
- 'A.A. Derevschikov'
- 'L. Didenko'
- 'T. Dietel'
- 'S.M. Dogra'
- 'W.J. Dong'
- 'X. Dong'
- 'J.E. Draper'
- 'F. Du'
- 'A.K. Dubey'
- 'V.B. Dunin'
- 'J.C. Dunlop'
- 'M.R. Dutta Mazumdar'
- 'V. Eckardt'
- 'W.R. Edwards'
- 'L.G. Efimov'
- 'V. Emelianov'
- 'J. Engelage'
- 'G. Eppley'
- 'B. Erazmus'
- 'M. Estienne'
- 'P. Fachini'
- 'J. Faivre'
- 'R. Fatemi'
- 'J. Fedorisin'
- 'K. Filimonov'
- 'P. Filip'
- 'E. Finch'
- 'V. Fine'
- 'Y. Fisyak'
- 'K.S.F. Fornazier'
- 'J. Fu'
- 'C.A. Gagliardi'
- 'L. Gaillard'
- 'J. Gans'
- 'M.S. Ganti'
- 'F. Geurts'
- 'V. Ghazikhanian'
- 'P. Ghosh'
- 'J.E. Gonzalez'
- 'H. Gos'
- 'O. Grachov'
- 'O. Grebenyuk'
- 'D. Grosnick'
- 'S.M. Guertin'
- 'Y. Guo'
- 'A. Gupta'
- 'N. Gupta'
- 'T.D. Gutierrez'
- 'T.J. Hallman'
- 'A. Hamed'
- 'D. Hardtke'
- 'J.W. Harris'
- 'M. Heinz'
- 'T.W. Henry'
- 'S. Hepplemann'
- 'B. Hippolyte'
- 'A. Hirsch'
- 'E. Hjort'
- 'G.W. Hoffmann'
- 'M.J. Horner'
- 'H.Z. Huang'
- 'S.L. Huang'
- 'E.W. Hughes'
- 'T.J. Humanic'
- 'G. Igo'
- 'A. Ishihara'
- 'P. Jacobs'
- 'W.W. Jacobs'
- M Jedynak
- 'H. Jiang'
- 'P.G. Jones'
- 'E.G. Judd'
- 'S. Kabana'
- 'K. Kang'
- 'M. Kaplan'
- 'D. Keane'
- 'A. Kechechyan'
- 'V.Yu. Khodyrev'
- 'B.C. Kim'
- 'J. Kiryluk'
- 'A. Kisiel'
- 'E.M. Kislov'
- 'J. Klay'
- 'S.R. Klein'
- 'D.D. Koetke'
- 'T. Kollegger'
- 'M. Kopytine'
- 'L. Kotchenda'
- 'K.L. Kowalik'
- 'M. Kramer'
- 'P. Kravtsov'
- 'V.I. Kravtsov'
- 'K. Krueger'
- 'C. Kuhn'
- 'A.I. Kulikov'
- 'A. Kumar'
- 'R.Kh. Kutuev'
- 'A.A. Kuznetsov'
- 'M.A.C. Lamont'
- 'J.M. Landgraf'
- 'S. Lange'
- 'F. Laue'
- 'J. Lauret'
- 'A. Lebedev'
- 'R. Lednicky'
- 'C-H. Lee'
- 'S. Lehocka'
- 'M.J. LeVine'
- 'C. Li'
- 'Q. Li'
- 'Y. Li'
- 'G. Lin'
- 'S.J. Lindenbaum'
- 'M.A. Lisa'
- 'F. Liu'
- 'H. Liu'
- 'J. Liu'
- 'L. Liu'
- 'Q.J. Liu'
- 'Z. Liu'
- 'T. Ljubicic'
- 'W.J. Llope'
- 'H. Long'
- 'R.S. Longacre'
- 'M. Lopez-Noriega'
- 'W.A. Love'
- 'Y. Lu'
- 'T. Ludlam'
- 'D. Lynn'
- 'G.L. Ma'
- 'J.G. Ma'
- 'Y.G. Ma'
- 'D. Magestro'
- 'S. Mahajan'
- 'D.P. Mahapatra'
- 'R. Majka'
- 'L.K. Mangotra'
- 'R. Manweiler'
- 'S. Margetis'
- 'C. Markert'
- 'L. Martin'
- 'J.N. Marx'
- 'H.S. Matis'
- 'Yu.A. Matulenko'
- 'C.J. McClain'
- 'T.S. McShane'
- 'F. Meissner'
- 'Yu. Melnick'
- 'A. Meschanin'
- 'M.L. Miller'
- 'N.G. Minaev'
- 'C. Mironov'
- 'A. Mischke'
- 'D.K. Mishra'
- 'J. Mitchell'
- 'B. Mohanty'
- 'L. Molnar'
- 'C.F. Moore'
- 'D.A. Morozov'
- 'M.G. Munhoz'
- 'B.K. Nandi'
- 'S.K. Nayak'
- 'T.K. Nayak'
- 'J.M. Nelson'
- 'P.K. Netrakanti'
- 'V.A. Nikitin'
- 'L.V. Nogach'
- 'S.B. Nurushev'
- 'G. Odyniec'
- 'A. Ogawa'
- 'V. Okorokov'
- 'M. Oldenburg'
- 'D. Olson'
- 'S.K. Pal'
- 'Y. Panebratsev'
- 'S.Y. Panitkin'
- 'A.I. Pavlinov'
- 'T. Pawlak'
- 'T. Peitzmann'
- 'V. Perevoztchikov'
- 'C. Perkins'
- 'W. Peryt'
- 'V.A. Petrov'
- 'S.C. Phatak'
- 'R. Picha'
- 'M. Planinic'
- 'J. Pluta'
- 'N. Porile'
- 'J. Porter'
- 'A.M. Poskanzer'
- 'M. Potekhin'
- 'E. Potrebenikova'
- 'B.V.K.S. Potukuchi'
- 'D. Prindle'
- 'C. Pruneau'
- 'J. Putschke'
- 'G. Rakness'
- 'R. Raniwala'
- 'S. Raniwala'
- 'O. Ravel'
- 'R.L. Ray'
- 'S.V. Razin'
- 'D. Reichhold'
- 'J.G. Reid'
- 'J. Reinnarth'
- 'G. Renault'
- 'F. Retiere'
- 'A. Ridiger'
- 'H.G. Ritter'
- 'J.B. Roberts'
- 'O.V. Rogachevskiy'
- 'J.L. Romero'
- 'A. Rose'
- 'C. Roy'
- 'L. Ruan'
- 'M.J. Russcher'
- 'R. Sahoo'
- 'I. Sakrejda'
- 'S. Salur'
- 'J. Sandweiss'
- 'M. Sarsour'
- 'I. Savin'
- 'P.S. Sazhin'
- 'J. Schambach'
- 'R.P. Scharenberg'
- 'N. Schmitz'
- 'K. Schweda'
- 'J. Seger'
- 'P. Seyboth'
- 'E. Shahaliev'
- 'M. Shao'
- 'W. Shao'
- 'M. Sharma'
- 'W.Q. Shen'
- 'K.E. Shestermanov'
- 'S.S. Shimanskiy'
- E Sichtermann
- 'F. Simon'
- 'R.N. Singaraju'
- 'N. Smirnov'
- 'R. Snellings'
- 'G. Sood'
- 'P. Sorensen'
- 'J. Sowinski'
- 'J. Speltz'
- 'H.M. Spinka'
- 'B. Srivastava'
- 'A. Stadnik'
- 'T.D.S. Stanislaus'
- 'R. Stock'
- 'A. Stolpovsky'
- 'M. Strikhanov'
- 'B. Stringfellow'
- 'A.A.P. Suaide'
- 'E. Sugarbaker'
- 'M. Sumbera'
- 'B. Surrow'
- 'M. Swanger'
- 'T.J.M. Symons'
- 'A. Szanto de Toledo'
- 'A. Tai'
- 'J. Takahashi'
- 'A.H. Tang'
- 'T. Tarnowsky'
- 'D. Thein'
- 'J.H. Thomas'
- 'A.R. Timmins'
- 'S. Timoshenko'
- 'M. Tokarev'
- 'T.A. Trainor'
- 'S. Trentalange'
- 'R.E. Tribble'
- 'O.D. Tsai'
- 'J. Ulery'
- 'T. Ullrich'
- 'D.G. Underwood'
- 'G. Van Buren'
- 'N. van der Kolk'
- 'M. van Leeuwen'
- 'A.M. Vander Molen'
- 'R. Varma'
- 'I.M. Vasilevski'
- 'A.N. Vasiliev'
- 'R. Vernet'
- 'S.E. Vigdor'
- 'Y.P. Viyogi'
- 'S. Vokal'
- 'S.A. Voloshin'
- 'W.T. Waggoner'
- 'F. Wang'
- 'G. Wang'
- 'G. Wang'
- 'X.L. Wang'
- 'Y. Wang'
- 'Y. Wang'
- 'Z.M. Wang'
- 'H. Ward'
- 'J.W. Watson'
- 'J.C. Webb'
- 'G.D. Westfall'
- 'A. Wetzler'
- 'C. Whitten Jr.'
- 'H. Wieman'
- 'S.W. Wissink'
- 'R. Witt'
- 'J. Wood'
- 'J. Wu'
- 'N. Xu'
- 'Z. Xu'
- 'Z.Z. Xu'
- 'E. Yamamoto'
- 'P. Yepes'
- 'I-K. Yoo'
- 'V.I. Yurevich'
- 'I. Zborovsky'
- 'H. Zhang'
- 'W.M. Zhang'
- 'Y. Zhang'
- 'Z.P. Zhang'
- 'C. Zhong'
- 'R. Zoulkarneev'
- 'Y. Zoulkarneeva'
- 'A.N. Zubarev'
- 'J.X. Zuo'
title: 'Hadronization geometry from net-charge angular correlations on momentum subspace $(\eta,\phi)$ in Au-Au collisions at $\sqrt{s_{NN}}$ = 130 GeV'
---
Introduction
============
Analysis of correlations and fluctuations plays an important role in studies of the colored medium produced in ultrarelativistic heavy-ion collisions [@stock; @poly; @dcc]. [*In-medium modification*]{} of parton scattering and fragmentation of energetic partons by the bulk medium produced in heavy-ion collisions may significantly alter large-momentum-scale two-particle correlations relative to those observed in p-p collisions. Large-momentum-scale correlations may result from initial-state multiple scattering [@iss; @jetquench], in-medium dissipation of scattered energetic partons [@newref] and hadronization of the colored medium to final-state hadrons (fragmentation of color strings in p-p, hadronization of the bulk medium in A-A). The local geometry of hadronization, which can be accessed by net-charge correlations, is the subject of this paper.
String fragmentation models [@lund] describe two-particle correlations on pseudorapidity and azimuth $(\eta,\phi)$ in high-energy p-p collisions in terms of local conservation of transverse momentum and net charge leading to canonical suppression of event-wise net-momentum and net-charge fluctuations. The nature of the corresponding process in A-A collisions remains an open question. Some change should be expected in the correlation structure as the medium evolves from that produced in very peripheral collisions (approximating minimum-bias proton-proton collisions) to that in central heavy-ion collisions. Predictions have been made of dramatic suppression of net-charge [*fluctuations*]{} in central A-A collisions as a signal of quark-gluon plasma formation [@jeon]. The question arises what detailed net-charge [*correlation structure*]{} would correspond to such predictions, and what structure is actually present in heavy-ion collisions.
In this Letter we report the first measurements in heavy-ion collisions of the centrality dependence of two-particle [*charge-dependent*]{} (net-charge) correlations on angular subspace $(\eta,\phi)$, where charge-dependent here refers to the difference between correlations for like-charge-sign pairs and unlike-sign pairs. This analysis is based on Au-Au collisions at $\sqrt{s_{NN}} = 130$ GeV obtained with the STAR detector at the Relativistic Heavy Ion Collider (RHIC). The observed correlation structure suggests that local charge conservation at hadronization combined with increasing system density and spatial extent results in evolution with Au-Au centrality from one-dimensional (1D) [*charge-ordering*]{} (locally alternating charge signs) on configuration space $z$ (the collision axis), coupled to $p_z$ (or pseudorapidity $\eta$) by longitudinal Bjorken expansion, to two-dimensional (2D) charge ordering on beam and azimuth directions $(z,\phi)$. Those results have not been anticipated by theoretical models [@jetquench; @rqmd].
Analysis Method
===============
We wish to access the complete [*charge-dependent*]{} (CD) structure of two-particle density $\rho(\vec{p_1},\vec{p_2})$ with minimal distortion and without imposition of a correlation model. In this analysis of net-charge [*angular*]{} correlations we project the two-particle momentum space onto angular subspace $(\eta_1,\eta_2,\phi_1,\phi_2)$ by integrating over a specific transverse momentum interval. The structure of net-charge correlations on [transverse momentum]{} with specific angular constraints will be considered in a future analysis.
Correlations are obtained with a [*differential*]{} analysis which compares object and reference pair density distributions. The object distribution is comprised of particle pairs formed from single events, referred to as [*sibling*]{} pairs, and the reference distribution consists of pairs combining particles from two different but similar events, referred to as [*mixed*]{} pairs. The corresponding pair densities are denoted by $\rho_{sib}(\vec{p}_1,\vec{p}_2)$ and $\rho_{mix}(\vec{p}_1,\vec{p}_2)$ respectively. The two-particle correlation function $C$ (as commonly defined in nuclear physics) and pair-number density ratio $r$ (as used in the study of quantum correlations or HBT [@starhbt]) are then defined and related by C(\_1,\_2) & = & \_[sib]{}(\_1,\_2) - \_[mix]{}(\_1,\_2)\
& = & \_[mix]{}(\_1,\_2) (r(\_1,\_2) - 1), \[Eq1\] with $r \equiv \rho_{sib}/ \rho_{mix}$. Difference $r - 1$ is the correlation measure we use. In order to visualize the CD correlation structure in the 4D angular subspace $(\eta_1,\eta_2,\phi_1,\phi_2)$ pair densities can be projected onto separate 2D subspaces ($\eta_1,\eta_2$) and ($\phi_1,\phi_2$). Those projections, discussed further below, discard a substantial amount of the information in the full two-particle space. However, they reveal that significant variation is restricted to [*difference variables*]{} $\eta_\Delta \equiv \eta_1 - \eta_2$ and $\phi_\Delta \equiv \phi_1 - \phi_2$ (the notation is explained in Sec. \[2part\]). For this analysis we therefore [*simultaneously*]{} project the 4D subspace onto those angular difference variables. The resulting 2D distribution is referred to as a [*joint autocorrelation*]{}. An autocorrelation is a projection [*by averaging*]{} [@average] from subspace $(x_1,x_2)$ onto difference variable $x_\Delta = x_1 - x_2$. A [*joint*]{} autocorrelation is a simultaneous projection onto two difference variables. The result of this projection technique is a [*nearly lossless*]{} (distortion free) projection from the initial 4D angular subspace onto a 2D autocorrelation space.
In this analysis, sibling and mixed pair-number densities $\rho(\vec{p}_1,\vec{p}_2)$ for four charge-pair combinations $(++,+-,-+,--)$ were projected onto $(\eta_1,\eta_2)$, $(\phi_1,\phi_2)$ and $(\eta_\Delta,\phi_\Delta)$. The projection was done by filling histograms of pair numbers $n_{ab} \simeq \epsilon_x \, \epsilon_y \,\rho(x_a,y_b)$, where subscripts $ab$ denote the 2D bin indices and $\epsilon_x, \epsilon_y$ are histogram bin widths on variables $x,y \in \{\eta_1, \eta_2, \phi_1, \phi_2, \eta_\Delta, \phi_\Delta\}$. Sibling and mixed pair-number histograms for each charge-pair combination were separately normalized to the total number of detected pairs in each centrality class: $\hat n_{ab,sib} = n_{ab,sib} / \sum_{ab} n_{ab,sib}$ and $\hat n_{ab,mix} = n_{ab,mix} / \sum_{ab} n_{ab,mix}$. Normalized pair-number ratios $\hat{r}_{ab} = \hat{n}_{ab,sib}/\hat{n}_{ab,mix}$ are the basis for this analysis.
To reduce systematic error, ratio histograms were obtained for subsets of events within a given centrality class which have similar multiplicities (differences $\leq 50$) and primary collision vertex locations within the detector (within 7.5 cm along the beam axis). Ratios $\hat r_{ab}$ for each centrality class were defined as weighted (by total number of sibling pairs) averages over all subsets in that centrality class. Ratios were further combined to form like-sign (LS: $++,--$), unlike-sign (US: $+-,-+$), and charge-dependent (CD = LS $-$ US) ratios. In this analysis we adopt a CD sign convention compatible with standard particle physics isospin convention and net-charge fluctuation measures [@meanptprl].
Data
====
Data for this analysis were obtained with the STAR detector [@star] using a 0.25 T uniform magnetic field parallel to the beam axis. A minimum-bias event sample required coincidence of two Zero-Degree Calorimeters (ZDC); a 0-15% of total cross section event sample was defined by a threshold on the Central Trigger Barrel (CTB), with ZDC coincidence. Event triggering and charged-particle measurements with the Time Projection Chamber (TPC) are described in [@star]. Tracking efficiencies, event and track quality cuts and primary-particle definition are described in [@meanptprl; @spectra]. Charged particles were accepted in $|\eta| \leq 1.3$, full azimuth and transverse momentum ($p_t$) range $0.15 \leq p_t \leq 2$ GeV/$c$. Particle identification was not implemented but charge sign was determined. Corrections were made to ratio $\hat r$ for two-track inefficiencies due to overlapping space points in the TPC (merging) and intersecting trajectories reconstructed as $>2$ particles (splitting) by applying two-track proximity cuts in the TPC to both $\rho_{sib}$ and $\rho_{mix}$ similar to that done in HBT analyses.
Small-momentum-scale correlation structures due to quantum, Coulomb and strong-interaction correlations [@starhbt] were suppressed by eliminating sibling [*and*]{} mixed track pairs ($\sim$22% of total) with $|\eta_{\Delta}| < 1.0$, $|\phi_{\Delta}| < 1.0$ and $|p_{t1} - p_{t2}| < 0.2$ GeV/$c$ if $p_t < 0.8$ GeV/$c$ for either particle. Those cuts do not significantly affect the correlation structures shown here. Four centrality classes for 300k events labeled (a) - (d) for central to peripheral were defined by cuts on TPC track multiplicity $N$ within the acceptance defined here relative to minimum-bias event multiplicity frequency distribution upper half-maximum end-point $N_0$, which corresponds to the maximum participant number [@meanptprl; @nu]. Four centrality classes were defined by (d) $0.03 < N/N_0 \leq 0.21$, (c) $0.21 < N/N_0 \leq 0.56$, (b) $0.56 < N/N_0 \leq 0.79$ and (a) $N/N_0 > 0.79$.
![\[Figure1\] Normalized LS pair-number ratios $\hat{r}$ for collisions in centrality class (a) (most central) for $(\eta_1,\eta_2)$ (left panel) and $(\phi_1,\phi_2)$ (right panel). ](mtxmtlsus3 "fig:"){width="1.65in"} ![\[Figure1\] Normalized LS pair-number ratios $\hat{r}$ for collisions in centrality class (a) (most central) for $(\eta_1,\eta_2)$ (left panel) and $(\phi_1,\phi_2)$ (right panel). ](mtxmtlsus1 "fig:"){width="1.65in"}
Two-particle Distributions {#2part}
==========================
Fig. \[Figure1\] shows ratio histograms $\hat r_{ab}$ for the LS charge combination on $(\eta_1,\eta_2)$ and $(\phi_1,\phi_2)$ for the most central event class, denoted (a). Deviations from unity ($\hat r - 1$) of this [*per-pair*]{} correlation measure contain a [*dilution factor*]{} [@dilution] $1/\bar N$ ($\bar N$ is defined as the mean multiplicity in the detector acceptance) and are therefore numerically a few [*permil*]{} for central Au-Au collisions. However, the correlation structure is large compared to statistical errors ([*cf.*]{} Figs. \[Figure2\]-\[Figure4\]). A sinusoid associated with elliptic flow (consistent with conventional reaction-plane measurements) dominates the $(\phi_1,\phi_2)$ correlations in the right panel. The [*anti*]{}correlated LS distribution on $(\eta_1,\eta_2)$ in the left panel (anticorrelated: depression along the $\eta_1 = \eta_2$ diagonal) suggests charge ordering from longitudinal string fragmentation as in p-p collisions [@lund; @isrpp]. However, these correlations projected separately onto $(\eta_1,\eta_2)$ and $(\phi_1,\phi_2)$ are incomplete, and quite misleading for A-A collisions. A more complete picture is obtained from 2D [joint]{} autocorrelations on difference variables $(\eta_\Delta,\phi_\Delta)$ as shown in Fig. \[Figure2\].
![\[Figure2\] Perspective views of two-particle CD joint autocorrelations $\bar N (\hat r-1)$ on $(\eta_{\Delta},\phi_{\Delta})$ for central (a) to peripheral (d) collisions. Center bins at $\phi_{\Delta} = \eta_{\Delta} = 0$, containing photon-conversion electron pairs, were omitted from model fits.](cdfitaya1 "fig:"){width="1.65in"} ![\[Figure2\] Perspective views of two-particle CD joint autocorrelations $\bar N (\hat r-1)$ on $(\eta_{\Delta},\phi_{\Delta})$ for central (a) to peripheral (d) collisions. Center bins at $\phi_{\Delta} = \eta_{\Delta} = 0$, containing photon-conversion electron pairs, were omitted from model fits.](cdfitaya2 "fig:"){width="1.65in"} ![\[Figure2\] Perspective views of two-particle CD joint autocorrelations $\bar N (\hat r-1)$ on $(\eta_{\Delta},\phi_{\Delta})$ for central (a) to peripheral (d) collisions. Center bins at $\phi_{\Delta} = \eta_{\Delta} = 0$, containing photon-conversion electron pairs, were omitted from model fits.](cdfitaya3 "fig:"){width="1.65in"} ![\[Figure2\] Perspective views of two-particle CD joint autocorrelations $\bar N (\hat r-1)$ on $(\eta_{\Delta},\phi_{\Delta})$ for central (a) to peripheral (d) collisions. Center bins at $\phi_{\Delta} = \eta_{\Delta} = 0$, containing photon-conversion electron pairs, were omitted from model fits.](cdfitaya4 "fig:"){width="1.65in"}
Because of the symmetry of these distributions on the angular spaces $(x_1,x_2)$ their description is more natural on [*diagonal*]{} sum and difference variables $x_\Sigma$ and $x_\Delta$ (reserving conventional difference notation $\Delta x$ for displacement on a 1D space $x$). The invariance of correlation structure on sum variables $\eta_\Sigma \equiv \eta_1 + \eta_2$ and $\phi_\Sigma \equiv \phi_1 + \phi_2$ in Fig. \[Figure1\] ([*i.e.,*]{} parallel to the $\eta_1 = \eta_2$ or $\phi_1 = \phi_2$ diagonals) implies that each distribution can be projected onto its difference variable $\phi_{\Delta} \equiv \phi_1 - \phi_2$ and $\eta_{\Delta} \equiv \eta_1 - \eta_2$ to form an autocorrelation [*without loss of information*]{}. The projection is done by averaging bin contents along each diagonal in Fig. \[Figure1\] parallel to the sum axis ([*e.g.*]{} the $\eta_1 = \eta_2$ diagonal) to obtain the bin contents of a 1D autocorrelation on $\eta_{\Delta}$ or $\phi_{\Delta}$ (the difference axes). Autocorrelation details are described in [@inverse; @mitmeths]. If projections are made simultaneously onto both difference variables of Fig. \[Figure1\] the resulting 2D joint autocorrelation on $(\eta_\Delta,\phi_\Delta)$ compactly represents [*all*]{} significant correlation structure on 4D angular subspace $(\eta_1,\eta_2,\phi_1,\phi_2)$.
In Fig. \[Figure2\] perspective views are shown of CD joint autocorrelations for four centrality classes of Au-Au collisions at $\sqrt{s_{NN}} =$ 130 GeV. Quantity $\bar N(\hat r - 1)$[ [@nbar]]{} represents [*per-particle*]{} correlations ([*i.e.,*]{} distribution of average numbers of correlated pairs per final-state particle) and is $O(1)$ for all centralities. Distributions in Fig. \[Figure2\] are dominated by a 2D negative peak which is broader and elliptical for peripheral collisions (d) with major axis along $\phi_\Delta$, transitioning smoothly to a narrower and deeper peak symmetric on $(\eta_\Delta,\phi_\Delta)$ for central collisions (a). The negative peak means that unlike-sign charge pairs are more probable than like-sign pairs for small angular separations on pseudorapidity and azimuth, consistent with local charge conservation (suppression of net-charge fluctuations). The vertical axis limits common to all panels were chosen to enhance the visibility of structure at large angular separations as opposed to showing the full depth of the negative peak at $\phi_\Delta = \eta_\Delta = 0$. Note that no CD (charge-dependent) component of elliptic flow is observed at the sensitivity level of these data. 1D projections of Fig. \[Figure2\] distributions and their 2D model fits (discussed below) onto individual difference variables $\phi_{\Delta}$ and $\eta_{\Delta}$ are shown in Fig. \[Figure3\]. Solid dots and curves (open triangles and dashed curves) correspond to $\eta_{\Delta}$ $(\phi_{\Delta})$ projections. The projections are over the pair acceptances apparent in Fig. \[Figure2\].
![\[Figure3\] Projections of 2D CD autocorrelations $\bar N(\hat{r} - 1 )$ in Fig. \[Figure2\] onto individual difference variables $\eta_{\Delta}$ (solid dots) and $\phi_{\Delta}$ (open triangles) for central (a) to peripheral (d) collisions. Solid (dashed) curves represent projections of 2D analytical model fits to data on $\eta_{\Delta}$ ($\phi_{\Delta}$). The 2D negative peaks are substantially reduced in amplitude after projecting onto 1D. ](cdprojnew1 "fig:"){width="1.65in" height="1.65in"} ![\[Figure3\] Projections of 2D CD autocorrelations $\bar N(\hat{r} - 1 )$ in Fig. \[Figure2\] onto individual difference variables $\eta_{\Delta}$ (solid dots) and $\phi_{\Delta}$ (open triangles) for central (a) to peripheral (d) collisions. Solid (dashed) curves represent projections of 2D analytical model fits to data on $\eta_{\Delta}$ ($\phi_{\Delta}$). The 2D negative peaks are substantially reduced in amplitude after projecting onto 1D. ](cdprojnew2 "fig:"){width="1.65in" height="1.65in"} ![\[Figure3\] Projections of 2D CD autocorrelations $\bar N(\hat{r} - 1 )$ in Fig. \[Figure2\] onto individual difference variables $\eta_{\Delta}$ (solid dots) and $\phi_{\Delta}$ (open triangles) for central (a) to peripheral (d) collisions. Solid (dashed) curves represent projections of 2D analytical model fits to data on $\eta_{\Delta}$ ($\phi_{\Delta}$). The 2D negative peaks are substantially reduced in amplitude after projecting onto 1D. ](cdprojnew3 "fig:"){width="1.65in" height="1.65in"} ![\[Figure3\] Projections of 2D CD autocorrelations $\bar N(\hat{r} - 1 )$ in Fig. \[Figure2\] onto individual difference variables $\eta_{\Delta}$ (solid dots) and $\phi_{\Delta}$ (open triangles) for central (a) to peripheral (d) collisions. Solid (dashed) curves represent projections of 2D analytical model fits to data on $\eta_{\Delta}$ ($\phi_{\Delta}$). The 2D negative peaks are substantially reduced in amplitude after projecting onto 1D. ](cdprojnew4 "fig:"){width="1.65in" height="1.65in"}
Errors
======
Statistical errors for $\hat r$ in Fig. \[Figure1\] (central collisions) are $\pm$0.00015 for all bins. Statistical errors for 1D autocorrelations are uniform on $\phi_{\Delta}$ (since $\phi$ is a periodic variable) but approximately double as $|\eta_{\Delta}|$ increases from 0 to 2 (due to finite $\eta$ acceptance). Statistical errors at $\eta_{\Delta} \sim 0$ vary from $\pm$0.00015 for central collisions to $\pm$0.0007 for peripheral collisions, again reflecting the $1/\bar N$ dilution factor. In contrast, statistical errors for $\bar N(\hat r - 1)$ in Fig. \[Figure2\] are approximately $\pm 0.2$ (one tick) for $\eta_\Delta \sim 0$ and are independent of centrality. Statistical errors for projections in Fig. \[Figure3\] are shown explicitly in that figure by error bars. Systematic errors were estimated as in [@meanptprl]. Systematic uncertainties associated with two-track inefficiency corrections and small momentum scale correlation cuts are negligible for this analysis. Systematic error due to non-primary backgrounds (dominant source) [@spectra], whose correlation with true primary particles is unknown, is estimated to be at most $\pm$7%, assumed uniform for all $(\eta_{\Delta},\phi_{\Delta})$ in the STAR acceptance. Contributions from resonance $(\rho^0 , \omega )$ decays are estimated to be at most about 10% of the negative peaks at $\phi_\Delta = \eta_\Delta = 0$ in Fig. \[Figure2\] in the range $|\eta_{\Delta}| < 0.5$, $|\phi_{\Delta}| < 2$ [@mevsim].
Model Fits
==========
The distributions in Fig. \[Figure2\] and their counterpart for p-p collisions [@jeffpp] reveal two asymptotic forms at the centrality limits: a 1D gaussian on $\eta_\Delta$ ([*uniform*]{} on $\phi_\Delta$) for p-p collisions and a 2D exponential on $(\eta_\Delta,\phi_\Delta)$ for central Au-Au collisions. The two forms may be limiting cases of a single evolving structure, or they may correspond to two independent correlation mechanisms with complementary centrality trends. A preliminary fitting exercise indicated that these 130 GeV Au-Au data do not have sufficient statistical power or centrality range to explore the possibility of a single evolving peak structure. We therefore used the simpler superposition model.
The distributions in Fig. \[Figure2\] were fitted with a five-parameter model function consisting of a 2D exponential function peaked on both $\eta_{\Delta}$ and $\phi_{\Delta}$ and a 1D gaussian on $\eta_{\Delta}$, constant on $\phi_\Delta$ (the latter motivated by the p-p limiting case [@isrpp; @jeffpp]) plus a constant offset, all defined relative to quantity $\hat r - 1$ as \[Eq3\] F & = & A\_0 + A\_1 { -\^ }\
& + & A\_2 {-( )\^2 } . $F$ interpolates between the 1D gaussian peak observed in p-p and the 2D exponential peak observed in central Au-Au collisions. Correlations between amplitudes $A_1$ and $A_2$ were negligible because of the distinct one- and two-dimensional peak shapes. Parameters $\sigma_{\phi_\Delta}$ and $\sigma_{\eta_\Delta}$ are the r.m.s. widths of the 2D exponential peak when projected onto the respective difference variables.
Best-fit values for varied parameters and $\chi^2/$DoF for the four centralities are listed in Table \[TableI\]. The width of the 1D gaussian, most evident near $|\phi_\Delta| \sim \pi$ in Fig. 2(d), was best determined by those peripheral data to be $1.5\pm0.25$ and was held fixed at that value for the other centralities to obtain the amplitude estimates. The observed peripheral Au-Au $\phi_\Delta$ width is definitely larger than the corresponding width for p-p collisions. Also included is tracking efficiency-correction factor ${\tilde S}$ [@extrap]. Total systematic error for efficiency-corrected amplitudes in Table \[TableI\] was 11% (errors added in quadrature). The model fits indicate that with increasing centrality the 2D exponential peak exhibits 1) strong amplitude increase, 2) significant width reduction and 3) approach to approximately equal widths on $\phi_{\Delta}$ and $\eta_{\Delta}$ for central collisions ([*cf.*]{} Fig. \[Figure3\]; [*e.g.,*]{} at mid-rapidity $\sigma_{\eta_\Delta} = 0.6$ corresponds to polar angle difference $0.57$, which is directly comparable to $\sigma_{\phi_\Delta}$).
centrality (d) (c) (b) (a) error[^1](%)
-------------------------- ------------------- ------------------- ------------------- ------------------- ----------------
${\tilde S}$ [@extrap] 1.19 1.22 1.25 1.27 8 (syst.)
$\bar{N}$ 115.5 424.9 789.3 983.0
${\tilde S} \bar{N} A_0$ 0.98 0.80 0.91 0.79 11-12
${\tilde S} \bar{N} A_1$ -4.1 -6.8 -7.7 -7.7 6-4
$\sigma_{\phi_{\Delta}}$ 0.94 0.75 0.72 0.72 11-5
$\sigma_{\eta_{\Delta}}$ 0.66 0.59 0.58 0.58 10-5
${\tilde S} \bar{N} A_2$ -0.51 -0.11 -0.15 -0.021 0.17-0.19 [^2]
$\chi^2$/DoF $\frac{380}{315}$ $\frac{315}{315}$ $\frac{314}{315}$ $\frac{329}{315}$
: \[TableI\] Parameters and fitting errors (only) for model fits \[Eq. (\[Eq3\])\] to joint autocorrelation data in Fig. \[Figure2\] for centrality bins (a) - (d) (central - peripheral). Total systematic error for tracking efficiency-corrected amplitudes is 11% [@extrap].
Discussion
==========
This analysis demonstrates for the first time that charge-dependent angular correlations for central Au-Au collisions differ dramatically from those for p-p collisions. CD angular correlations for p-p collisions are dominated by a 1D negative gaussian peak on $\eta_{\Delta}$ with $\sigma_{\eta_\Delta} \simeq 1$ [@isrpp; @jeffpp], conventionally associated with longitudinal charge ordering on $z$ during string fragmentation [@lund], plus a 2D gaussian peak associated with quantum correlations. For the most peripheral Au-Au centrality (d) in this analysis we observe CD correlation structure intermediate between p-p and central Au-Au collisions, consistent with the fact that collision events in centrality class (d) for these 130 GeV data are not very peripheral: they contain about 100 particles in the STAR acceptance (see Table \[TableI\]). In central Au-Au collisions the 1D gaussian peak is no longer detectable. Instead, a large-amplitude 2D negative exponential peak dominates the correlation structure, with similar widths on $\eta_\Delta$ and $\phi_\Delta$ much reduced from those measured in p-p collisions.
![\[Figure4\] Left panel: Efficiency corrected correlation amplitudes for 2D exponential (dots) and 1D gaussian (triangles) components from Table \[TableI\] for negative peaks in Fig. \[Figure2\] are plotted on mean path length $\nu$ [@nu]. Right panel: Fitted widths $\sigma_{\eta_\Delta}$ (dots) and $\tan^{-1}\, \sigma_{\phi_\Delta}$ (triangles) are plotted on $\nu$. Plotting variable $\tan^{-1}$ permits the divergent p-p $\sigma_{\phi_\Delta}$ value to be included. Hatched regions and $\nu = 1$ data points summarize p-p limiting values. Curves guide the eye.](cdcentrality1.eps){width="3.3in"}
Variations of peak amplitudes and widths with Au-Au centrality are shown in Fig. \[Figure4\], along with p-p limiting cases (cross-hatched bands) from STAR p-p data at 200 GeV [@jeffpp], consistent with ISR p-p data at 52.5 GeV [@isrpp]. The p-p data points in Fig. \[Figure4\] (values at $\nu = 1$) indicate the amplitude and r.m.s. width of the 1D gaussian on $\eta_\Delta$, the uniformity of that correlation on $\phi_\Delta$ ($\sigma_{\phi_\Delta} \gg 1$) and the absence of a 2D exponential on $(\eta_\Delta,\phi_\Delta)$ in the fit residuals, represented by the solid dot in the left panel at $\nu = 1$. Comparison of the low-$p_t$ ($0.15 \leq p_t \leq 0.5$ GeV/c) p-p results with the present Au-Au results is qualitative but reasonable given the similarity in shape of the Au-Au CD correlations for $0.15 \leq p_t \leq 0.5$ (discussed below) to those in Fig. \[Figure2\].
The collision centrality is represented by mean participant path length $\nu$ [@nu], defined as the average number of nucleons encountered by a participant nucleon. That centrality measure is desirable because it permits comparisons with p-A collisions, initial-state scattering should follow a trend linear in $\nu$ and $\nu$ also provides an estimate (proportionality) of final-state pathlength.
We adopt the strategy of plotting $\tan^{-1}(\sigma_{\phi_\Delta})$ rather than $\sigma_{\phi_\Delta}$ so as to include the p-p ‘infinite azimuth width’ on the same plot, since that distribution is approximately uniform on $\phi$. Interpolations among the measured Au-Au points are sketched by the solid and dash-dot curves. [*Extrapolations*]{} to corresponding p-p values are sketched by the dashed and dotted curves. The extrapolations contain [*substantial uncertainties*]{} in relating p-p to mid-peripheral Au-Au results. *Efficiency-corrected*
per-particle correlation amplitudes $\tilde S\bar NA$ for central Au-Au collisions exceed in magnitude those for p-p collisions [*by a factor 10*]{}. The dramatic shape and amplitude changes strongly contradict a p-p linear superposition hypothesis [@nbar] for all but the most peripheral Au-Au collisions.
These results for net-charge angular correlations suggest that CD correlations in Au-Au collisions, as in p-p collisions, derive from configuration-space charge ordering as a consequence of local charge conservation during hadronization, but the hadronization geometry changes from 1D ($\eta$) in p-p collisions to [*at least*]{} 2D ($\eta,\phi$) in central Au-Au collisions, leading to an approach to angular symmetry on $(\eta_\Delta,\phi_\Delta)$. Transverse charge ordering (on $p_t$) is also possible but is studied in a separate analysis. Hadronic rescattering in A-A collisions could reduce the CD correlation amplitude at large $\phi_\Delta$ but would also reduce the width on $\eta_\Delta$ and therefore cannot be solely responsible for the nearly symmetric peak shape in central Au-Au collisions. In Fig. \[Figure4\] the contribution from 1D charge ordering (gaussian peak on $\eta_\Delta$) is already substantially reduced for centrality (d) ($\nu \sim 2.5$) in favor of the symmetric component (exponential peak).
A hadron-opaque medium in more central collisions may contribute to the newly-observed [*exponential*]{} peak shape. An exponential distribution on pair opening angle \[radius on $(\eta,\phi)$\] is consistent with: 1) correlations detected only if both members of a correlated pair are not significantly scattered, 2) scattering probability determined by a mean free path, 3) mean path length in the medium increasing monotonically with pair opening angle. That rescattering picture assumes that CD correlations do not result from hadronization outside the medium. Contributions from charge ordering in jet fragmentation were studied by splitting central Au-Au data at $p_t = 0.5$ GeV/$c$, below which jet fragments should be negligible. Negative peak structures as in Fig. \[Figure2\] were observed to dominate both subsamples, although the amplitudes were not identical.
[hijing]{} [@jetquench] and [rqmd]{} [@rqmd; @starhbt] charge-dependent angular correlations qualitatively disagree with data. [hijing]{} charge-dependent correlations are determined by the Lund model [@lund] [*via*]{} [pythia]{} [@pythia], and are consequently consistent with p-p 1D string fragmentation for all A-A centralities: a 1D gaussian on $\eta_\Delta$ with amplitude about 10% of the exponential peak in Fig. \[Figure2\] (a). RQMD, representing mainly resonance decays and hadronic rescattering, exhibits a broad 2D gaussian on ($\eta_\Delta,\phi_\Delta $), with amplitude also about 10% of the exponential peak in the data for central collisions. Large-scale correlations as in Fig. \[Figure1\] observed for US [*and*]{} LS pairs in data are consistent with local charge ordering but [*inconsistent*]{} with CD correlations from decays of hadronic resonances such as the $\rho^0$, which affect only the US pair type. That observation further argues against a resonance-gas scenario. Measurements of net-charge fluctuations have been advocated as a probe of heavy-ion collisions. Predictions of dramatic suppression of net-charge fluctuations in the case of QGP formation based on entropy arguments [@jeon] refer by implication to an [*integral*]{} of net-charge angular correlations over a detector acceptance. Phenix observed net-charge fluctuations in Au-Au at 130 GeV [@phen-net] slightly reduced from ‘stochastic behavior’ and independent of collision centrality. The data were consistent with RQMD representing a resonance gas. STAR observed net-charge fluctuations in Au-Au at 200 GeV [@star-net] intermediate between what is expected from canonical suppression in a partial acceptance and a resonance gas, again with little or no centrality dependence. Those conclusions are in sharp contrast to what we observe in the present analysis.
It is important to note that net-charge fluctuations within a given detector acceptance integrate CD joint autocorrelations such as those presented in this paper (within a constant offset) over that acceptance, as described in [@inverse]. As integral quantities, fluctuation measurements are insensitive to the [*differential structure*]{} of angular correlations. In the present analysis we observe dramatic changes in differential structure (10-fold amplitude increase, nearly two-fold width reduction) while corresponding peak integrals exhibit only modest change with collision centrality (integrals of observed CD peaks using peak parameters in Table \[TableI\] increase linearly in magnitude on $\nu$ by about 20%). We suggest that the theoretical connection between net-charge fluctuation suppression and QGP formation, currently based only on large-scale integral measures, should be re-examined in the more differential context of CD autocorrelation structure.
Summary
=======
In summary, we have measured charge-dependent angular correlations on pseudorapidity and azimuth difference variables $(\eta_1 - \eta_2)$ and $(\phi_1 - \phi_2)$ for Au-Au collisions at $\sqrt{s_{NN}}$ = 130 GeV. The data are consistent with [*local*]{} charge conservation or canonical suppression of net-charge fluctuations, evolving from 1D (along $\eta$) color-string fragmentation in p-p collisions to exponentially-attenuated (on opening angle) 2D charge-ordered emission from a hadron-opaque medium in central Au-Au collisions. The transition from 1D to 2D correlation structure occurs rapidly with increasing collision centrality. These results are qualitatively inconsistent with predictions from standard Monte Carlo collision models typically applied to single-particle differential distributions and integrated yields from relativistic heavy-ion collisions. Charge-dependent angular autocorrelations provide unique [*differential*]{} access to the changing geometry of hadronization and hadronic rescattering as the energy density and spatial extent of A-A collisions vary with centrality. We thank the RHIC Operations Group and RCF at BNL, and the NERSC Center at LBNL for their support. This work was supported in part by the HENP Divisions of the Office of Science of the U.S. DOE; the U.S. NSF; the BMBF of Germany; IN2P3, RA, RPL, and EMN of France; EPSRC of the United Kingdom; FAPESP of Brazil; the Russian Ministry of Science and Technology; the Ministry of Education and the NNSFC of China; IRP and GA of the Czech Republic, FOM of the Netherlands, DAE, DST, and CSIR of the Government of India; Swiss NSF; the Polish State Committee for Scientific Research; STAA of Slovakia, and the Korea Sci. & Eng. Foundation.
[9]{}
R. Stock, Nucl. Phys. [**A661**]{}, 282c (1999); H. Heiselberg, Phys. Rep. [**351**]{}, 161 (2001).
A. Dumitru, R. Pisarski, Phys. Lett. [**B504**]{}, 282 (2001). L.M. Bettencourt, K. Rajagopal and J. V. Steele, Nucl. Phys. [**A693**]{}, 825 (2001). M. Gaździcki, A. Leonidov, G. Roland, Eur. Phys. J. [**C6**]{}, 365 (1999).
X.-N. Wang, M. Gyulassy, Phys. Rev. D [**44**]{}, 3501 (1991).
Q.J. Liu and T.A Trainor, Phys. Lett. [**B567**]{}, 184 (2003).
B. Andersson, G. Gustafson, G. Ingelman and T. Sjöstrand, Phys. Rep. [**97**]{}, 31 (1983).
M. Asakawa, U. Heinz, B. Müller, Phys. Rev. Lett. [**85**]{}, 2072 (2000); S. Jeon, V. Koch, Phys. Rev. Lett. [**85**]{}, 2076 (2000).
H. Sorge, H. Stöcker, W. Greiner, Nucl. Phys. [**A498**]{}, 567c (1989); Ann. Phys. (N.Y.) [**192**]{}, 266 (1989). C. Adler [*et al.*]{} Phys. Rev. Lett. [**87**]{}, 082301 (2001).
Averaging rather than simple projection is an essential feature of autocorrelations required to properly account for acceptance effects in two-particle spaces.
J. Adams [*et al.*]{} (STAR Collaboration), Phys. Rev. C [**71**]{}, 064906 (2005).
K. H. Ackermann [*et al.*]{}, Nucl. Instrum. Meth. A [**499**]{}, 624 (2003); see other STAR papers in volume A[**499**]{}.
C. Adler [*et al.*]{}, Phys. Rev. Lett. [**87**]{}, 112303 (2001); [*ibid*]{}. [**89**]{}, 202301 (2002).
Centrality measure $\nu$ estimates the mean participant path length as a number of encountered nucleons. For this analysis $\nu \equiv 5.5 \, (N/N_0)^{1/3} \simeq 5.5\, (N_{part}/N_{part,max})^{1/3} \simeq 2N_{bin}/N_{part}$, based on Glauber-model simulations. $N_{part}$ is the number of participants, $N_{bin}$ the number of binary collisions, and $N_0$ is the upper half-maximum endpoint of the minimum-bias data distribution plotted as $d\sigma/dN^{1/4}_{}$.
Under composition of independent (uncorrelated) subsystems ([*e.g.,*]{} hypothetical independent N-N collisions within A-A collisions) correlations [*per particle*]{} remain constant, whereas correlations [*per pair*]{} go as $1/n_{ch}$. An exception to that behavior is quantum correlations for bosons (pions) where [*all pairs*]{} with momentum difference $q$ below some value are correlated, in which case $r -1 \rightarrow 1$ as $q \rightarrow 0$, independent of system size.
D. Drijard [*et al.*]{} (ACCDHW Collaboration), Nucl. Phys. B [**166**]{}, 233 (1980). T. A. Trainor, R. J. Porter and D. J. Prindle, J. Phys. G [**31**]{}, 809 (2005) (hep-ph/0410182).
D. J. Prindle and T. A. Trainor, hep-ph/0506173. To appear in: Proceedings of the MIT Workshop on Correlations and Fluctuations in Relativistic Nuclear Collisions, Cambridge, Massachusetts, 21-23 April, 2005.
$\bar{N} (\hat r - 1 )$, measuring correlations per final-state particle (typically $O(1)$ for all centralities), is [*invariant*]{} with centrality if A-A collisions are linear superpositions of p-p collisions.
R. Ray and R. Longacre, nucl-ex/0008009.
R.J. Porter and T.A. Trainor (STAR Collaboration), hep-ph/0406330.
Extrapolation factors ${\tilde S}$ for $\bar{N}(\hat r - 1 )$ provide corrections to amplitudes $A_0$, $A_1$ and $A_2$ for background contamination and tracking inefficiency [@spectra]. Systematic error in ${\tilde S}$ was estimated to be $\pm$8%.
T. Sjöstrand, Comput. Phys. Commun. [**82**]{}, 74 (1994); T. Sjöstrand, L. Lönnblad, S. Mrenna and P. Skands, hep-ph/0308153.
K. Adcox [*et al.*]{} (PHENIX Collaboration), Phys. Rev. Lett. [**89**]{}, 082301 (2002) (nucl-ex/0203014).
J. Adams [*et al.*]{} (STAR Collaboration), Phys. Rev. C [**68**]{}, 044905 (2003) (nucl-ex/0307007).
[^1]: Range of fitting errors in percent, from peripheral to central.
[^2]: Magnitude of fitting errors.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We provide a streamlined proof and improved estimates for the weak multivariate Gnedenko law of large numbers on concentration of random polytopes within the space of convex bodies (in a fixed or a high dimensional setting), as well as a corresponding strong law of large numbers.'
author:
- 'Daniel J. Fresen[^1] and Richard A. Vitale [^2]'
title: Concentration of random polytopes around the expected convex hull
---
Introduction
============
Let $d\in \mathbb{N}$ and let $\mu $ be a probability measure on $\mathbb{R}^{d}$ with a log-concave density $f=d\mu /dx$, i.e. $-\log f$ is a convex extended real valued function. Let $n\geq d+1$ and let $(X_{i})_{1}^{n}$ denote an i.i.d. sequence of random vectors with common distribution $\mu $. The convex hull$$P_{n}=\mathrm{conv}\{X_{i}\}_{1}^{n} \label{random poly def}$$is a random polytope and, as such, is a random element w.p.1 of the space $\mathcal{K}_{d}$ of all convex bodies in $\mathbb{R}^{d}$ (compact convex sets with non-empty interior). There are various metrics and metric-like functions on $\mathcal{K}_{d}$, such as the Hausdorff distance $d_{\mathcal{H}}$ and the Banach-Mazur distance $\delta ^{BM}$ (for origin symmetric bodies). We refer the reader to [@Schn] for general background on convex bodies, and to [@Gruber] specifically for metric, and other, structures on $\mathcal{K}^{d}$.
It was shown in [@Fr2013] that if $n\geq c\exp (\exp (5d))$, then with probability at least $1-3^{d+3}(\log n)^{-1000}$, there exists $x\in \mathbb{R}^{n}$ and$$\lambda \leq 1+c^{\prime }d^{2}\frac{\log \log n}{\log n}$$such that$$\lambda ^{-1}(F_{1/n}-x)+x\subseteq P_{n}\subseteq \lambda (F_{1/n}-x)+x
\label{sandwhiching}$$where $c,c^{\prime }>0$ are universal constants and $F_{1/n}$ is the floating body defined by$$F_{\delta }=\cap \{\mathfrak{H}:\mu (\mathfrak{H})\geq 1-\delta \}
\label{floatin}$$where the intersection runs through the collection of all closed half-spaces $\mathfrak{H}$ of $\mu $-mass at least $1-\delta $ ($\delta <e^{-1}$). The body $F_{1/n}$ was originally defined by Schütt and Werner [@ScWe] in the case of Lebesgue measure on a convex body and has often been used to model random polytopes, see for example [@Bar07; @BaLa; @Vu].
Being log-concave, the density $f$ decays at least as quickly as an exponential function. Any bound on the decay rate of $f$ translates to a bound on the Hausdorff distance $d_{\mathcal{H}}(P_{n},F_{1/n})$. For example if the tails of $\mu $ are sub-Gaussian (with universally bounded constants), then $\mathrm{diam}(F_{1/n})\leq c(\log n)^{1/2}$ and ([sandwhiching]{}) translates to$$d_{\mathcal{H}}(P_{n},F_{1/n})\leq c^{\prime }d^{2}\frac{\log \log n}{\sqrt{\log n}}$$where $c,c^{\prime }>0$ are universal constants. This is an embodiment of the concentration of measure phenomenon: the polytope $P_{n}$, as a random element of the metric space $(\mathcal{K}_{d},d_{\mathcal{H}})$, is concentrated around $F_{1/n}$.
In the case $d=1$, $P_{n}$ reduces to the interval$$\lbrack \min \{X_{i}\}_{1}^{n},\max \{X_{i}\}_{1}^{n}]$$and we see that the above mentioned result generalizes a theorem of Gnedenko [@Gn] on concentration of the maximum and minimum of a large i.i.d. sample (under rapid decay of the tails of $\mu $). Other multivariate analogs of Gnedenko’s law of large numbers are included in [@Geff] for the multivariate normal distribution, [@Good] for Gaussian measures on infinite dimensional spaces, [@DMR; @Fish; @1966; @Fish; @1969] for regularly varying distributions, and [@KiRe; @MR94] for more general distributions.
The proof of (\[sandwhiching\]) was complicated by the fact that there is no convenient expression for the support function of the floating body,$$h_{F_{1/n}}(\theta )=\max_{x\in F_{1/n}}\left\langle \theta ,x\right\rangle$$In this paper we study concentration of $P_{n}$ around the expected convex hull$$\mathbb{E}P_{n}=\{x\in \mathbb{R}^{n}:\forall \theta \in
S^{d-1},\left\langle \theta ,x\right\rangle \leq \mathbb{E}\max_{1\leq i\leq
n}\left\langle \theta ,X_{i}\right\rangle \} \label{def expected hull}$$which is easily seen to be a convex body with support function$$h_{\mathbb{E}P_{n}}(\theta )=\mathbb{E}\max_{1\leq i\leq n}\left\langle
\theta ,X_{i}\right\rangle \label{support expected}$$Using the expected convex hull leads to a streamlined proof of ([sandwhiching]{}). The notion of the expectation of a random convex body follows the theory of integrals of set valued functions, see for example [@Art; @Au; @Deb; @Kudo] and the references therein. It was used in [ArtVit]{} for the purpose of a Kolmogorov strong law of large numbers and has appeared as an approximant to floating bodies in bounded domains [@BaVi], as well as in other contexts e.g. [@GW92; @GW12; @Vitale87; @Vitale; @90; @Vitale; @91; @Vitale; @94; @Weil95].
In the original paper [@Fr2013] we were mainly interested in a quantitative dependence on $n$. Although our bounds included dependence on dimension, the required sample size was very large. Theorem \[polytope and expected val\] includes improved bounds on the required sample size and is more in the spirit of the high dimensional theory. The quantitative dependence that we achieve is essentially the same as that in Dvoretzky’s theorem, see for example [@Sch]. This result should also be compared to the main result in [@DGT].
To make the present exposition brief, we refer the reader to [@Fr2013] for a more detailed discussion.
Main results
============
\[polytope and expected val\]Let $d\in \mathbb{N}$ and let $\mu $ be a log-concave probability measure on $\mathbb{R}^{d}$ with center of mass at the origin and non-singular covariance matrix. Consider any $\varepsilon \in
(0,1/2)$ and let $n\geq \exp (7d\varepsilon ^{-1}\log \varepsilon ^{-1})$. Let $(X_{i})_{1}^{n}$ be an i.i.d. sample from $\mu $, $P_{n}=\mathrm{conv}\{X_{i}\}_{1}^{n}$, and let $\mathbb{E}P_{n}$ denote the expected convex hull as defined by (\[support expected\]). With probability at least $1-3n^{-\varepsilon /4}$,$$(1-\varepsilon )\mathbb{E}P_{n}\subseteq P_{n}\subseteq (1+\varepsilon )\mathbb{E}P_{n}$$
Using the bound $d_{\mathcal{H}}(A,B)\leq \mathrm{diam}(B)\inf \{\lambda
\geq 1:\lambda ^{-1}A\subseteq B\subseteq \lambda A\}$, Theorem [polytope and expected val]{} may be transferred to a bound on $d_{\mathcal{H}}(P_{n},\mathbb{E}P_{n})$. The following Corollary, which is similar to the main result in [@BaVi], is a consequence of Lemma \[2 sided 1D concentration\].
\[expected val and floating body\]Let $d\in \mathbb{N}$ and let $\mu $ be a log-concave probability measure on $\mathbb{R}^{d}$ with center of mass at the origin and non-singular covariance matrix. Let $\mathbb{E}P_{n}$ denote the expected convex hull as defined by (\[support expected\]), and let $F_{1/n}$ denote the floating body defined by (\[floatin\]). Then provided $n\geq 12$,$$(1-3/\log n)\mathbb{E}P_{n}\subseteq F_{1/n}\subseteq (1+1/\log n)\mathbb{E}P_{n}$$
\[strong LLN\]Let $d\in \mathbb{N}$ and let $\mu $ be a log-concave probability measure on $\mathbb{R}^{d}$ with center of mass at the origin and non-singular covariance matrix. Let $(X_{i})_{1}^{\infty }$ be an i.i.d. sample from $\mu $, and let $(P_{n})_{n=d+1}^{\infty }$ and $(\mathbb{E}P_{n})_{3}^{\infty }$ be the random polytopes and expected convex hulls defined by (\[random poly def\]) and (\[support expected\]) respectively. Then with probability $1$, there exists $N\in \mathbb{N}$ such that for all $n\geq N$,$$\left( 1-\frac{3\log \log n}{\log n}\right) \mathbb{E}P_{n}\subseteq
P_{n}\subseteq \left( 1+\frac{8\log \log n}{\log n}\right) \mathbb{E}P_{n}
\label{strong bound}$$
Notation
========
If $J$ is the cumulative distribution function associated to a probability measure $\mu $ on $\mathbb{R}$, then the generalized inverse $J^{-1}:(0,1)\rightarrow \mathbb{R}$ is defined as$$J^{-1}(t)=\sup \{x\in \mathbb{R}:J(x)<t\}=\inf \{x\in \mathbb{R}:J(x)\geq
t\}$$If $\mu $ has a log-concave density function then $J(J^{-1}(t))=t$ for all $t\in (0,1)$ and $J^{-1}(J(x))=x$ for all $x$ in the support of $\mu $. If $(Y_{i})_{1}^{n}$ is an i.i.d. sample from $\mu $, then $Y_{(n)}=\max_{1\leq
i\leq n}Y_{i}$ denotes the $n^{th}$ order statistic.
If $K\subset \mathbb{R}^{d}$ is a convex body then the function$$h_{K}(x)=\max_{y\in K}\left\langle x,y\right\rangle$$is known as the support function of $K$. If $0\in \mathrm{int}(K)$ then the Minkowski functional is defined as $$\left\Vert x\right\Vert _{K}=\min \{\lambda \geq 1:x\in \lambda K\}$$and the support function is the Minkowski functional of the polar body$$K^{\circ }=\{y\in \mathbb{R}^{d}:\forall x\in K,\left\langle
x,y\right\rangle \leq 1\}$$i.e. $h_{K}(\cdot )=\left\Vert \cdot \right\Vert _{K^{\circ }}$. In the case when $K$ is centrally symmetric, i.e. $K=-K$, then $h_{K}(\cdot )$ and $\left\Vert \cdot \right\Vert _{K}$ are norms.
Proofs
======
The following lemma is a natural extension of Lemma 7 in [@Fr2013].
\[2 sided 1D concentration\]Let $\mu $ be a probability measure on $\mathbb{R}$ with mean $0$ and log-concave density $f=d\mu /dx$. Let $n\geq
12 $ and let $(Y_{i})_{1}^{n}$ be an i.i.d. sample from $\mu $. Then for all $t>0$,$$\begin{aligned}
\mathbb{P}\{Y_{(n)} &\leq &(1+t)\mathbb{E}Y_{(n)}\}\geq 1-n^{-t/2}
\label{right tail} \\
\mathbb{P}\{Y_{(n)} &\geq &(1-t)\mathbb{E}Y_{(n)}\}\geq 1-\exp (-n^{t/2}/3)
\label{left tail}\end{aligned}$$
Let $J$ be the common distribution function of each $Y_{i}$. Let $f_{n}$ and $J_{n}$ denote the density and distribution function of $Y_{(n)}$,$$\begin{aligned}
J_{n}(t) &=&J(t)^{n} \\
f_{n}(t) &=&\frac{d}{dt}J_{n}(t)=nJ(t)^{n-1}f(t)\end{aligned}$$Since $f$ is log-concave, so is $J$ (see for example Theorem 5.1 in [LoVe]{} or Lemma 5 in [@Fr2013]). The product of log-concave functions is certainly log-concave, and therefore so is $f_{n}$. By a standard result, see for example Lemma 5.4 in [@LoVe], $J_{n}^{-1}(e^{-1})\leq \mathbb{E}Y_{(n)}\leq J_{n}^{-1}(1-e^{-1})$. Just as the left tail $J$ is log-concave, so is the right tail $1-J$, and the function $u(t)=-\log (1-J(t))$ is convex. This implies that,$$\frac{u(J^{-1}(1-n^{-t/2}/n))-u(J^{-1}(1-1/n))}{J^{-1}(1-n^{-t/2}/n)-J^{-1}(1-1/n)}\geq \frac{u(J^{-1}(1-1/n))-u(0))}{J^{-1}(1-1/n)}$$which translates to$$\frac{J^{-1}(1-n^{-t/2}/n)-J^{-1}(1-1/n)}{J^{-1}(1-1/n)}\leq \frac{t\log n}{2(\log n-1)}\leq t$$Now,$$\mathbb{P}\{Y_{(n)}\leq J^{-1}(1-n^{-t/2}/n)\}=(1-n^{-t/2}/n)^{n}\geq
1-n^{-t/2}$$By definition of $J_{n}$, $J_{n}(J^{-1}(1-1/n))=(1-1/n)^{n}<e^{-1}$, so $\mathbb{E}Y_{(n)}\geq J_{n}^{-1}(e^{-1})\geq J^{-1}(1-1/n)$ and (\[right tail\]) follows. Again by convexity of $u$, $$\frac{u(J^{-1}(1-9/(20n)))-u(J^{-1}(1-9n^{t/2-1}/20))}{J^{-1}(1-9/(20n))-J^{-1}(1-9n^{t/2-1}/20)}\geq \frac{u(J^{-1}(1-9/(20n)))-u(0))}{J^{-1}(1-9/(20n))}$$which translates to$$\frac{J^{-1}(1-9/(20n))-J^{-1}(1-9n^{t/2-1}/20)}{J^{-1}(1-9/(20n))}\leq
\frac{(t/2)\log n}{\log n-1+\log (20/9)}\leq t$$Now,$$\mathbb{P}\{Y_{(n)}\leq J^{-1}(1-9n^{t/2-1}/20)\}=(1-9n^{t/2-1}/20)^{n}\leq
\exp (-9n^{t/2}/20)$$As before, $J_{n}(J^{-1}(1-9/(20n)))=(1-9/(20n))^{n}>1-e^{-1}$, so $\mathbb{E}Y_{(n)}\leq J_{n}^{-1}(1-e^{-1})<J^{-1}(1-9/(20n))$ and (\[left tail\]) follows.
Since $J^{-1}(1-1/n)=J_{n}^{-1}((1-1/n)^{n})$, where $J_{n}(x)=\mathbb{P}\{Y_{(n)}\leq x\}$, $\mathbb{P}\{Y_{(n)}\leq J^{-1}(1-1/n)\}\geq 1/3$ and by inequality (\[left tail\]) of Lemma \[2 sided 1D concentration\], this can only be true if $J^{-1}(1-1/n)\geq (1-(\log 18)/\log n)\mathbb{E}Y_{(n)}$. By similar reasoning, $\mathbb{P}\{Y_{(n)}>J^{-1}(1-1/n)\}\geq 1-e^{-1}$, which by inequality (\[right tail\]) of Lemma \[2 sided 1D concentration\] implies that $J^{-1}(1-1/n)\leq (1+1/\log n)\mathbb{E}\gamma
_{(n)}$. The result now follows from the definitions of $F_{1/n}$ and $\mathbb{E}P_{n}$, see (\[floatin\]) and (\[support expected\]).
The following lemma appears as Lemmas 4.10 and 4.11 in [@Pi] under the assumption that $K$ is centrally symmetric. We sketch the proof to show that it can also be used in the non-symmetric case.
\[net lemma\]Let $K\subset \mathbb{R}^{d}$ be any convex body with $0\in
\mathrm{int}(K)$ and $0<\varepsilon <1/2$. Then there exists a set $\mathcal{N}\subset \partial K$ with $|\mathcal{N}|\leq (3/\varepsilon )^{d}$ such that for all $\theta \in \partial K$ there exist sequences $(\omega
_{i})_{0}^{\infty }\subseteq \mathcal{N}$ and $(\varepsilon
_{i})_{1}^{\infty }\subseteq \lbrack 0,\infty )$ such that $0\leq
\varepsilon _{i}\leq \varepsilon ^{i}$ for all $i$ and$$\theta =\omega _{0}+\sum_{i=1}^{\infty }\varepsilon _{i}\omega _{i}$$
Consider a subset $\mathcal{N}\subset \partial K$, minimal with respect to set inclusion, with the following property: for all $z\in \partial K$ there exists $\omega \in \mathcal{N}$ such that $\left\Vert z-\omega \right\Vert
_{K}\leq \varepsilon $. Such a set can easily be constructed recursively, and we shall refer to $\mathcal{N}$ as an $\varepsilon $-net. Note that since $K$ may be non-symmetric, we may have $\left\Vert z-\omega \right\Vert
_{K}\neq \left\Vert \omega -z\right\Vert _{K}$ and order becomes important. By the standard volumetric argument $|\mathcal{N}|\leq (3/\varepsilon )^{d}$. By the defining property of $\mathcal{N}$, for all $x\in \mathbb{R}^{d}$ there exists $\omega \in \mathcal{N}$ such that$$\left\Vert x-\left\Vert x\right\Vert _{K}\omega \right\Vert _{K}\leq
\varepsilon \left\Vert x\right\Vert _{K} \label{net approxi}$$Now consider $\theta \in \partial K$. By (\[net approxi\]) there exists $\omega _{0}\in \mathcal{N}$ such that $\left\Vert \theta -\omega
_{0}\right\Vert _{K}\leq \varepsilon $. By applying (\[net approxi\]) again, there exists $\omega _{1}\in \mathcal{N}$ such that $\left\Vert
\theta -\omega _{0}-\left\Vert \theta -\omega _{0}\right\Vert _{K}\omega
_{1}\right\Vert _{K}\leq \varepsilon \left\Vert \theta -\omega
_{0}\right\Vert _{K}\leq \varepsilon ^{2}$. Iterating this procedure defines a sequence $(\omega _{i})_{0}^{\infty }$ such that for all $N\in \mathbb{N}$,$$\left\Vert \theta -\omega _{0}-\sum_{i=1}^{N}\varepsilon _{i}\omega
_{i}\right\Vert _{K}\leq \varepsilon ^{N+1}$$where $\varepsilon _{i}=\left\Vert \theta -\omega
_{0}-\sum_{i=1}^{i-1}\varepsilon _{i}\omega _{i}\right\Vert _{K}\leq
\varepsilon ^{i}$.
Set $\delta =3n^{-\varepsilon /(4d)}$ and let $\mathcal{N\subset \partial ((}\mathbb{E}P_{n})^{\circ })$ be a $\delta $-net as in Lemma \[net lemma\]. By the bounds imposed on $n$, $\delta \leq \varepsilon /5<1/10$. From the union bound and Lemma \[2 sided 1D concentration\], the following event occurs with probability at least $1-(3/\delta )^{d}3n^{-\varepsilon /2}\geq
1-3n^{-\varepsilon /4}$: for all $\omega \in \mathcal{N}$,$$(1-\varepsilon /2)\left\Vert \omega \right\Vert _{(\mathbb{E}P_{n})^{\circ
}}\leq \left\Vert \omega \right\Vert _{P_{n}{}^{\circ }}\leq (1+\varepsilon
/2)\left\Vert \omega \right\Vert _{(\mathbb{E}P_{n})^{\circ }}
\label{first bound}$$For any $\theta \in \mathcal{\partial ((}\mathbb{E}P_{n})^{\circ })$, write $\theta =\omega _{0}+\sum_{1}^{\infty }\delta _{i}\omega _{i}$, with $\omega
_{i}\in \mathcal{N}$ and $0\leq \delta _{i}\leq \delta ^{i}$ for all $i$. By the triangle inequality and (\[first bound\]),$$\left\Vert \theta \right\Vert _{P_{n}{}^{\circ }}\leq (1+\varepsilon
/2)\sum_{i=0}^{\infty }\delta ^{i}\leq (1+2\delta )(1+\varepsilon /2)\leq
1+\varepsilon$$and$$\left\Vert \theta \right\Vert _{P_{n}{}^{\circ }}\geq \left\Vert \omega
_{0}\right\Vert _{P_{n}{}^{\circ }}-\sum_{1}^{\infty }\delta ^{i}\left\Vert
\omega _{i}\right\Vert _{P_{n}{}^{\circ }}\geq 1-\varepsilon
/2-(1+\varepsilon /2)\delta (1-\delta )^{-1}\geq 1-\varepsilon$$and the result follows.
Here $d$ and $\mu $ are fixed, and we treat $n\rightarrow \infty $ as a variable. From comparing successive terms in the binomial theorem and using the fact that $n^{-k}{n\choose k}$ is a decreasing function of $k$, for all $\delta \in (0,1/2)$$$(1-2\delta /n)^{n}=(1-\delta )-\delta +{n\choose 2}\left( \frac{2\delta }{n}\right) ^{2}+\sum_{k=3}^{n}(-1)^{k}{n\choose k}\left( \frac{2\delta }{n}\right) ^{k}\leq 1-\delta$$Since $1-3n^{-\varepsilon /4}\leq \mathbb{P}\{P_{n}\subseteq (1+\varepsilon )\mathbb{E}P_{n}\}=(\mathbb{P}\{X_{1}\in (1+\varepsilon )\mathbb{E}P_{n}\})^{n}$, it follows that $\mu ((1+\varepsilon )\mathbb{E}P_{n})\geq
(1-3n^{-\varepsilon /4})^{1/n}\geq 1-6n^{-1-\varepsilon /4}$ (provided $3n^{-\varepsilon /4}<1/2$). Setting $\varepsilon =8(\log \log n)/\log n$ yields$$\sum_{n=12}^{\infty }\mathbb{P}\{X_{n}\notin (1+\varepsilon )\mathbb{E}P_{n}\}\leq 2\sum_{n=12}^{\infty }n^{-1-\varepsilon /4}=2\sum_{n=12}^{\infty
}\frac{1}{n(\log n)^{2}}<\infty$$Therefore, by the Borel-Cantelli lemma, with probability 1 there exists $N^{(1)}\in \mathbb{N}$ such that for all $n\geq N^{(1)}$,$$P_{n}\subseteq (1+8(\log \log n)/\log n)\mathbb{E}P_{n}
\label{outer sandwich}$$For each $n\in \mathbb{N}$, let $E_{n}$ be the event that (\[outer sandwich\]) holds. Consider any sufficiently large (deterministic) $n\in
\mathbb{N}$. Set $\varepsilon =3(\log \log n)/\log n$ and $\delta =3\exp
(-n^{-\varepsilon /2}/(6d))$. Let $\mathcal{N\subset \partial ((}\mathbb{E}P_{n})^{\circ })$ be a $\delta $-net as in Lemma \[net lemma\]. As before, $\delta \leq \varepsilon /10\leq 1/20$. By the union bound and Lemma \[2 sided 1D concentration\], the following event, to be denoted $F_{n}$, occurs with probability at least $1-(3/\delta )^{d}\exp (-n^{\varepsilon /2}/3)\geq
1-\exp (-n^{\varepsilon /2}/6)\geq 1-n^{-2}$: for all $\omega \in \mathcal{N}
$,$$(1-\varepsilon /2)\left\Vert \omega \right\Vert _{(\mathbb{E}P_{n})^{\circ
}}\leq \left\Vert \omega \right\Vert _{P_{n}{}^{\circ }}$$The Borel-Cantelli lemma again implies that with probability $1$ there exists $N^{(2)}\in \mathbb{N}$ such $F_{n}$ occurs for all $n\geq N^{(2)}$. For all $n\geq \max \{N^{(1)},N^{(2)}\}$, $E_{n}\cap F_{n}$ occurs, and expressing an arbitrary $\theta \in \mathcal{\partial ((}\mathbb{E}P_{n})^{\circ })$ as $\theta =\omega _{0}+\sum_{1}^{\infty }\delta
_{i}\omega _{i}$ as in Lemma \[net lemma\] and using the triangle inequality,$$\left\Vert \theta \right\Vert _{P_{n}{}^{\circ }}\geq \left\Vert \omega
_{0}\right\Vert _{P_{n}{}^{\circ }}-\sum_{1}^{\infty }\delta ^{i}\left\Vert
\omega _{i}\right\Vert _{P_{n}{}^{\circ }}\geq 1-\varepsilon /2-2\delta
(1-\delta )^{-1}\geq 1-\varepsilon$$which implies (\[strong bound\]).
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors would like to thank Mokshay Madiman for comments related to the paper.
[99]{} Artstein, Z.: On the calculus of closed set-valued functions. Indiana Univ. Math. J. 24, 433-441(1974)
Artstein, Z., Vitale, R. A.: A strong law of large numbers for random compact sets. Ann. Probab. 3 (5), 879-882 (1975)
Aumann, R. J.: Integrals of set-valued functions. J. Math. Anal. Appl. 12, 1-12 (1965)
Bárány, I.: Random polytopes, convex bodies, and approximation. Stochastic geometry, 77-118, Lecture Notes in Math., 1892, Springer, Berlin, (2007)
Bárány, I., Larman, D. G.: Convex bodies, economic cap coverings, random polytopes. Mathematika 35, 274-291 (1988)
Bárány, I., Vitale, R. A.: Random convex hulls: floating bodies and expectations. J. Approx. Theory 75 (2), 130-135 (1993)
Dafnis, N., Giannopoulos, A., Tsolomitis, A.: Asymptotic shape of a random polytope in a convex body. J. Funct. Anal. 257 (9), 2820-2839 (2009)
Davis, R., Mulrow, E., Resnick, S.: Almost sure limit sets of random samples in $\mathbb{R}^{d}$. Adv. in Appl. Probab. 20 (3), 573-599 (1988)
Debreu, G.: Integration of correspondences. Proc. Fifth Berkeley Symp. Math. Statist. and Probability (Berkeley, Calif., 1965/66), Vol. II: Contributions to Probability Theory, Part 1, 351-372 Univ. California Press, Berkeley, Calif. (1967)
Fisher, L. D.: The convex hull of a sample. Bull. Amer. Math. Soc. 72, 555-558 (1966)
Fisher, L. D.: Limiting sets and convex hulls of samples from product measures. Ann. Math. Statist. 40, 1824-1832 (1969)
Fresen, D.: A multivariate Gnedenko law of large numbers. Ann. Probab. 41 (5), 3051-3080, (2013)
Geffroy, J.: Localisation asymptotique du polyèdre d’appui d’une échantillon Laplacien à $k$ dimensions. Publ. Inst. Statist. Univ. Paris 10, 213-228 (1961)
Gnedenko, B.: Sur la distribution limite du terme maximum d’une série aléatoire. Ann. of Math. (2) 44, 423-453, (1943)
Goodey, P., Weil, W.: The determination of convex bodies from the mean of random sections. Math. Proc. Cambridge Philos. Soc. 112 (2), 419-430 (1992)
Goodey, P., Weil, W.: A uniqueness result for mean section bodies. Adv. Math. 229 (1), 596-601 (2012)
Goodman, V.: Characteristics of normal samples. Ann. Probab. 16, 1281-1290 (1988)
Gruber, P. M.: The space of convex bodies. Handbook of Convex Geometry, Vol. A, B, 301-318 (1993)
Kinoshita, K., Resnick, S.: Convergence of scaled random samples in $\mathbb{R}^{d}$. Ann. Probab. 19 (4), 1640-1663 (1991)
Kudo, H.: Dependent experiments and sufficient statistics. Nat. Sci. Rept. Ochanomizu Univ., Tokyo 4, 151-163 (1954)
Lovász, L., Vempala, S.: The geometry of logconcave functions and sampling algorithms. Random Structures Algorithms 30, 307-358 (2007)
McBeth, D., Resnick, S.: Stability of random sets generated by multivariate samples. Comm. Statist. Stochastic Models 10 (3), 549-574 (1994)
Pisier, G.: The Volume of Convex Bodies and Banach Space Geometry. Cambridge Tracts in Mathematics 94, Cambridge University Press, (1989)
Schechtman, G.: Two observations regarding embedding subsets of Euclidean space in normed spaces. Adv. Math. 200 (1), 125-135 (2006)
Schneider, R.: Convex bodies: the Brunn-Minkowski theory. 2nd ed. Encyclopedia of Mathematics and its Applications, 151. Cambridge University Press, (2014)
Schütt, C., Werner, E.: The convex floating body. Math. Scand. 66, 275-290, (1990)
Vitale, R. A.: Expected convex hulls, order statistics, and Banach space probabilities. Acta Appl. Math. 9 (1-2), 97-102, (1987)
Vitale, R. A.: The Brunn-Minkowski inequality for random sets. J. Multivariate Anal. 33 (2), 286-293, (1990)
Vitale, R. A.: Expected absolute random determinants and zonoids. Ann. Appl. Probab. 1 (2), 293-300 (1991)
Vitale, R. A.: Stochastic smoothing of convex bodies: two examples. First International Conference on Stochastic Geometry, Convex Bodies and Empirical Measures (Palermo, 1993). Rend. Circ. Mat. Palermo (2) Suppl. 35, 315-322 (1994)
Vu, V.: Sharp concentration of random polytopes. Geom. Funct. Anal. 15 (6), 1284-1318, (2005)
Weil, W.: The estimation of mean shape and mean particle number in overlapping particle systems in the plane. Adv. in Appl. Probab. 27 (1), 102-119 (1995)
[^1]: Yale University, Department of Mathematics, [email protected]
[^2]: The University of Connecticut, Department of Statistics, [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We report on a two-loop supersymmetric contribution to the magnetic moment $(g-2)_\mu$ of the muon [which is enhanced by two powers of $\tan\beta$. This contribution arises from a shift]{} in the relation between the muon mass and Yukawa coupling and can [increase]{} the supersymmetric contribution to $(g-2)_\mu$ [sizably]{}. As a result, if the currently observed $3\sigma$ deviation between the experimental and SM theory value of $(g-2)_\mu$ [is analyzed within the Minimal Supersymmetric Standard Model (MSSM), the derived constraints on the parameter space are modified significantly: If $(g-2)_\mu$ is used to determine $\tan \beta$ as a function of the other MSSM parameters, our corrections decrease $\tan \beta$ by roughly 10% for $\tan \beta =50$.]{}'
author:
- Schedar Marchetti$^1$
- Susanne Mertens$^1$
- Ulrich Nierste$^1$
- Dominik Stöckinger$^2$
title: '${\mbox{Tan}\,}\beta$-enhanced supersymmetric corrections to the anomalous magnetic moment of the muon'
---
The anomalous magnetic moment ${a_\mu}=(g-2)_\mu/2$ of the muon is one of the most precisely measured and calculated quantities in particle physics — and recently it has developed into one of the observables with the [most significant]{} deviations between the experimental value and the corresponding Standard Model (SM) theory prediction. The review [@MillerRR] obtains$$\begin{aligned}
{a_\mu}^{\rm exp}-{a_\mu}^{\rm SM} &= 29.5(8.8)\times10^{-10},
\label{deviation}\end{aligned}$$ a $3.4\sigma$ deviation between ${a_\mu}^{\rm exp}$ and ${a_\mu}^{\rm SM}$, the experimental [@BNL6] and SM theory value, respectively.
Eq. (\[deviation\]) represents dramatic progress. It has been made possible by better determinations of the hadronic $e^+e^-$ cross section by SND, CMD-II, KLOE and BaBar [@experiments]. These are crucial ingredients for all recent evaluations of the hadronic vacuum polarization contribution to ${a_\mu}^{\rm SM}$ [@Jegerlehner07; @Davier07; @HMNT06]. [In Ref. [@Benayoun], further progress on the “$\tau$-puzzle” has been achieved, confirming the $e^+e^-$-based result (\[deviation\]).]{} Now all evaluations of the SM theory prediction have a smaller error than ever before and agree very well. [With this]{} progress the case for physics beyond the SM in ${a_\mu}$ has become stronger. [Generically,]{} contributions from [new]{} physics with characteristic mass scale $M_{\rm BSM}$ are suppressed as $(M_W/M_{\rm BSM})^2$ compared to the SM electroweak contribution [of ${a_\mu}^{\rm
weak}=15.4(0.2)\times10^{-10}$]{}, which is only half as large as the observed deviation. [Thus some parametric enhancement of the new contribution is required.]{}
Supersymmetry (SUSY), implemented in the Minimal Supersymmetric Standard Model (MSSM), [can naturally explain the]{} observed deviation [for two reasons:]{} First, the masses of smuons and charginos, the most relevant SUSY particles, can be as small as $M_{\rm
SUSY}\sim{\cal O}($100 GeV$)$ without contradicting current experimental data, [allowing a rather mild]{} suppression factor $(M_W/M_{\rm SUSY})^2$. Second, the SUSY contributions to ${a_\mu}$ are enhanced by the parameter $$\begin{aligned}
\tan\beta&=\frac{v_2}{v_1},\end{aligned}$$ the ratio of the vacuum expectation values (vevs) of the two Higgs doublets $H_{1,2}$ in the MSSM, [which governs the size of the down-type Yukawa couplings. We normalize the vevs as $v\equiv\sqrt{v_1^2+v_2^2}=174\; {\mbox{GeV}}$. Since $a_{\mu}$ involves a chirality-flip, it is proportional to the muon Yukawa coupling $y_{\mu}$.]{} Large values $\tan\beta\sim50$ lead to similar top and bottom Yukawa couplings and are therefore preferred [in scenarios with]{} Yukawa unification. [Remarkably, naive multi-Higgs doublet models fail to explain the deviation in Eq. (\[deviation\]), because the corresponding loop diagrams involve at least three powers of the small coupling $y_{\mu}$.]{}
The SUSY contributions to ${a_\mu}$ are approximately given by $$\begin{aligned}
{a_\mu}^{\rm SUSY}\approx
13 \times10^{-10}\left(\frac{100\,\rm GeV}{M_{\rm SUSY}}\right) ^2
\ \tan\beta\ \mbox{sign}(\mu),
\label{approx}\end{aligned}$$ if the SUSY parameters for the smuon, gaugino and Higgsino masses have a common scale $M_{\rm SUSY}$, see [@review] and references therein. The sign of the contributions is given by the sign of the Higgsino mass parameter $\mu$ [(choosing the gaugino mass parameters $M_1$, $M_2$ positive)]{}. [We restrict our analysis to the case of real $\mu$, $M_1$ and $M_2$, because sizable CP-violating phases of these parameters are in conflict with the bounds on electric dipole moments, if $M_{\rm SUSY}$ is in the range needed to accommodate ${a_\mu}^{\rm exp}$.]{} The SUSY contributions explain the entire deviation of $29.5\times10^{-10}$ if [$\tan\beta$ is given by approximately $ 2.3 ({M_{\rm SUSY}}/{100\,\rm GeV}) ^2$.]{}
[Clearly, ${a_\mu}$ plays an eminent role in studies of the MSSM parameter space, see e.g. [@scanrefs]. In Ref. [@Ellis] a preference of a constrained version of the MSSM over the SM is found from a global fit to collider and electroweak precision data. [This conclusion is primarily driven by ${a_\mu}$. In Refs. [@Hertzog:2007hz; @PlehnRauch] a future $\tan\beta$-determination using LHC-data combined with ${a_\mu}$ is outlined. Owing to its importance, ${a_\mu}$ therefore deserves a theoretical precision analysis including radiative corrections.]{}]{}
So far, all SUSY one-loop contributions are known [@oneloop]. At the two-loop level two kinds of [relevant]{} SUSY contributions [have been identified:]{} QED-logarithms $\log(M_{\rm SUSY}/m_\mu)$ arising from SUSY one-loop diagrams with additional photon exchange have been evaluated in [@DG98] and amount to $-7$% to $-9$% of the one-loop contributions. Two-loop diagrams involving closed loops of either sfermions (stops, sbottoms, etc) or charginos/neutralinos have been evaluated in [@HSW0304]. They amount to about $2\%$ of the one-loop contributions if all SUSY masses are degenerate but can be much larger, if e.g. smuon masses are very heavy but stops and/or charginos and Higgs bosons are light.
All these known SUSY contributions to ${a_\mu}$ share the feature of [Eq. ]{}(\[approx\]): For large $\tan\beta$ they are linear in $\tan\beta$, $$\begin{aligned}
\label{tborder}
{a_\mu}^{\rm SUSY,\ known} \propto \alpha^l
\left(\frac{m_\mu}{M_{\rm SUSY}}\right)^2\tan\beta,\end{aligned}$$ where $l=1,2$ denotes the loop order. In this paper we identify and discuss a SUSY contribution ${a_\mu}^{\rm
SUSY,\ \Delta_\mu}$ which is quadratic in $\tan\beta$, i.e. of the order $$\begin{aligned}
\label{tbsqorder}
{a_\mu}^{\rm SUSY,\ \Delta_\mu} \propto \alpha^2
\left(\frac{m_\mu}{M_{\rm SUSY}}\right)^2{\tan^2\beta},\end{aligned}$$ and can therefore be a significant correction in the large-$\tan\beta$ region.
The physical origin of these ${\tan^2\beta}$-corrections is a shift in the muon Yukawa coupling $y_\mu$ due to $\tan\beta$-enhanced one-loop effects. In the computation of ${a_\mu}^{\rm SUSY}$ beyond the one-loop level this shift appears in the muon mass renormalization constant $\delta m_\mu$, defined in the on-shell scheme: $$\begin{aligned}
m_\mu+\delta m_\mu &=
\frac{m_\mu}{1+\Delta_\mu}+\mbox{non-}\tan\beta\mbox{-enhanced terms},
\nonumber\\
\label{shift}
y_\mu &= \frac{m_\mu}{v \cos \beta (1+\Delta_\mu)}
\left( 1 + {\cal O} (\cot \beta) \right),\end{aligned}$$ where $m_\mu$ is the physical, pole-mass of the muon and where the [shift $\Delta_\mu\propto\alpha\tan\beta$ will be given below]{}. [This type of $\tan\beta$-enhanced corrections has been studied intensely in the down-quark sector [@tanbe; @CGNWlong]. In the standard approach one employs the limit $M_{\rm SUSY}\gg v_2$ and derives an effective loop-induced coupling of $H_2$ to down-type fermions, which results in relations between masses and Yukawa couplings of the type in Eq. (\[shift\]) [@tanbe]. For ${a_\mu}$, however, this procedure fails, because ${a_\mu}^{\rm SUSY}$ vanishes in the limit $M_{\rm SUSY}\gg v_2$, so that the important corrections associated with $\Delta_\mu$ were overlooked so far. In the case of ${a_\mu}$ one must resort to the method of Ref. [@CGNWlong], which explicitly identifies $\tan\beta$-enhanced loop diagrams and resums them to all orders in perturbation theory for the case of interest $M_{\rm SUSY}\sim v_2$. Eq. (\[shift\]) contains the desired effect to all orders $\alpha^l \tan^l\beta$, $l=1,2,\ldots$. For the phenomenology of ${a_\mu}$ only the term with $l=1$, contributing to ${a_\mu}^{\rm SUSY}$ at the two-loop level, is relevant.]{}
In the following we will show that the shift [in Eq. ]{}(\[shift\]) is the only source of the [$\tan\beta$-enhanced radiative corrections of type $\alpha^l \tan^l\beta$ and that there are no enhancement factors with even more powers of $\tan\beta$.]{} The proof relies on an analysis of mass singularities similar to the analysis presented in [@CGNWlong]:
The one-loop diagram proportional to $y_\mu$ gives one power of $m_\mu \tan\beta$. (The second factor of $m_\mu$ in Eqs. (\[tborder\]) and (\[tbsqorder\]) stems from the definition of ${a_\mu}$.) A genuine $l$-loop diagram (i.e. without counterterms) may involve $n$ powers of $\tan \beta$ stemming from the muon Yukawa coupling $y_\mu$ or any other Yukawa coupling $y_f\propto
(m_f/M_W) \tan \beta$. It will result in a desired $\tan\beta$-enhanced correction if $n\geq l$ and the loop diagram diverges as $1/(m_f^{n-1})$ for $m_f\to 0$ to compensate for the factor of $m_f^n$ in $y_f^n$.
Such mass singularities can be analyzed by passing from the MSSM to an effective field theory in which all heavy particles are integrated out and only particles with mass $m_f$ or less are retained. Heavy loops are represented by point-like interactions in the effective theory and the infrared structure of any MSSM loop diagram and its counterpart in the effective theory are the same. A novel feature compared to the analysis of Yukawa interactions in Ref. [@CGNWlong] is the appearance of one dimension-5 coupling, the magnetic interaction term ${\overline}\mu_L \sigma_{\nu\rho} \mu_R
F^{\nu\rho}$. On dimensional grounds any loop corrections involving this term can only depend logarithmically on light fermion masses $m_f$. Potentially dangerous loops involve effective couplings of dimension 4 or less, since they might come with one or more inverse power of $m_f$. However, the only such couplings induced by heavy loops are those which are already present in low-energy QED and QCD and the effect of the heavy particles in the underlying theory can be completely absorbed into the renormalization of masses and couplings in the effective theory [@ac].
In conclusion the only effective diagrams with inverse powers of $m_f$ are the known QED diagrams proportional to $1/m_\mu$ and they are unaffected by our MSSM short-distance structure. These findings only hold, if a decoupling scheme is adopted for the renormalization [@ac]; for our case it is important that $m_\mu$ is renormalized in the on-shell scheme. Next we inspect other diagrams involving counterterms: The counterterm for the Yukawa coupling $y_\mu$ is $\delta y_\mu= y_\mu \delta m_\mu/m_\mu
\propto m_\mu \tan^2 \beta$ and gives rise to the enhanced corrections in Eq. (\[shift\]) [@CGNWlong]. The counterterms for gauge couplings and the muon and photon fields cannot be $\tan\beta$-enhanced, because unlike $\delta y_\mu/y_\mu$ they do not involve any factor of $1/m_f$. Finally the renormalization of the Higgsino mass parameter $\mu$, the soft SUSY breaking terms and the parameter $\tan\beta$ can be chosen at will, and our statement about the absence of $\tan\beta$-enhanced corrections beyond those in Eq. (\[shift\]) is valid
for all renormalization schemes in which [these]{} renormalization constants are [*not*]{} $\tan\beta$-enhanced. This includes common schemes such as the $\overline{\mbox{DR}}$-scheme for all SUSY parameters, or a mixed scheme where $\tan\beta$ and the $A$-parameter are defined in the $\overline{\mbox{DR}}$-scheme but the [smuon, chargino and neutralino masses are renormalized on-shell.]{}
The shift $\Delta_\mu$ is given by the $\tan\beta$-enhanced terms of the muon self energy. In terms of the loop function $$\begin{aligned}
I(a,b,c)&=\frac{
a^2b^2\log{\frac{a^2}{b^2}}
+b^2c^2\log{\frac{b^2}{c^2}}
+c^2a^2\log{\frac{c^2}{a^2}}}{(a^2-b^2)(b^2-c^2)(a^2-c^2)}
,\end{aligned}$$ which satisfies $I(a,a,a)=1/(2a^2)$, it can be written as $$\begin{aligned}
\Delta_\mu =
&-{\mu \ \tan\beta\ }
\frac{g_2^2\ M_2}{16\pi^2}\ I(m_1,m_2,m_{{\tilde{\nu}_\mu}})\nonumber\\
&
-{\mu \ \tan\beta\ }
\frac{g_2^2\ M_2}{16\pi^2}\ \frac{1}{2} I(m_1,m_2,m_{\tilde{\mu}_L})
\nonumber\\
& -{\mu \ \tan\beta\ } \frac{g_1^2\ M_1}{16\pi^2}\
\Big[
I(\mu,M_1,m_{\tilde{\mu}_R})\nonumber\\
& -\frac{1}{2}
I(\mu,M_1,m_{\tilde{\mu}_L})
-I(M_1,m_{\tilde{\mu}_L},m_{\tilde{\mu}_R})\Big].
\label{DeltamuResult}\end{aligned}$$ The appearing gaugino, Higgsino, and smuon mass parameters [and the Standard Model parameters $g_{1,2}$, $s_W$ are defined as usual, see e.g. [@review]]{}, and we have defined $$\begin{aligned}
m_{1,2}^2 &=\frac{1}{2}\Big[
(M_2^2 + \mu^2 + 2M_W^2)\nonumber\\
&\qquad\mp\sqrt{(M_2^2 + \mu^2 + 2M_W^2)^2-4M_2^2\mu^2}\Big], \nonumber \\
m^2_{\tilde{\nu}_\mu}&
=m_{L,\tilde{\mu}}^2- \frac{M_Z^2}{2},\quad
m^2_{\tilde{\mu}_L} = m_{L,\tilde{\mu}}^2-M_Z^2 (s_W^2-\frac12),
\nonumber \\
m^2_{\tilde{\mu}_R}&=m_{R,\tilde{\mu}}^2+M_Z^2 s_W^2.\end{aligned}$$ While the chargino contributions are exact in the large-$\tan\beta$ limit, the neutralino contributions in (\[DeltamuResult\]) have been simplified using the approximation $M_Z\ll \mu, M_1, M_2$. [The]{} deviation of $\Delta_\mu$ as given in (\[DeltamuResult\]) from the exact result [satisfies]{} $|\Delta_\mu-\Delta_\mu^{\rm exact}|<0.01$ over the entire parameter range (all supersymmetry masses [are varied]{} independently between $100$ GeV and $2$ TeV, $\tan\beta\le100$) for which $|{a_\mu}^{\rm SUSY}|<10^{-8}$.
Similar shifts exist for all down-type fermions, and in particular the shift of the bottom-quark Yukawa coupling $\Delta_b$ has been analyzed in detail in the literature [@tanbe; @CGNWlong], and the results can be readily applied to the muon case.
[The contribution ${a_\mu}^{\rm SUSY,\Delta_\mu}$ of the new ${\tan\beta}$-enhanced contributions to [${a_\mu}=-2m_\mu F_M(0)$ can be easily obtained by noting that the magnetic form factor $F_M(0)$]{} is proportional to $y_\mu$, apart from numerically irrelevant terms with three or more powers of $y_\mu$. Now $y_\mu$ enters $F_M$ in two ways: First it appears explicitly in the higgsino-muon couplings or in the Higgs-smuon coupling triggering the left-right mixing in the smuon mass matrix. Second it appears implicitly through $m_\mu \propto y_\mu \cos\beta$, which arises from the application of the Dirac equation ${{\makebox[.2ex][l]{$p$}/}}\mu_L=m_\mu \mu_R$. The second contribution is suppressed by a factor of $\cot\beta$ compared to the first. The ${\tan\beta}$-enhanced corrections to the first contribution are obtained by using the expression in Eq. (\[shift\]) for $y_\mu$. Therefore]{} $$\begin{aligned}
{a_\mu}^{\rm SUSY,1L}+{a_\mu}^{\rm SUSY,\Delta_\mu} &=
{a_\mu}^{\rm SUSY,1L}\left(\frac{1}{1+\Delta_\mu}\right).
\label{ResultamuDeltamu}\end{aligned}$$ Note that [this formula is]{} only correct for the [enhanced terms of order $\alpha^l \tan^l\beta$]{}, but this is sufficient for our purposes.
Equation (\[ResultamuDeltamu\]) is the main result of this paper. We are now in the position to write down the most accurate prediction for ${a_\mu}^{\rm
SUSY}$, replacing the result given in [@review] [by]{} [^1] [^2] $$\begin{aligned}
{a_\mu^{\rm SUSY}}&=
{a_\mu^{\rm SUSY,1L}}\left(1-\frac{4\alpha}{\pi}\log\frac{M_{\rm
SUSY}}{m_\mu}\right)\left(\frac{1}{1+\Delta_\mu}\right)\nonumber\\
&+
{a_\mu}^{(\chi\gamma H)}+{a_\mu}^{(\tilde{f}\gamma H)}
+{a_\mu}^{(\chi\{W,Z\} H)}+{a_\mu}^{(\tilde{f}\{W,Z\}H)}\nonumber\\
&
+{a_\mu^{\rm SUSY,ferm,2L}}+ {a_\mu^{\rm SUSY,bos,2L}}+\ldots.
\label{amuSUSYknown}\end{aligned}$$ The first line contains the one-loop result, corrected by large QED-logarithms [@DG98] and by the [new ${\tan\beta}$-enhanced terms]{} discussed here. The second and third lines contain further known two-loop contributions [@HSW0304]. The terms ${a_\mu}^{(\tilde{p}VS)}$ denote contributions from diagrams where a vector boson $V$ and scalar $S$ couple to the muon line and which involve a closed $\tilde{p}$-loop; ${a_\mu^{\rm SUSY,ferm,2L}}$ and ${a_\mu^{\rm SUSY,bos,2L}}$ denote the difference of diagrams without SUSY particles between the MSSM and the SM, arising from the different Higgs sectors. The dots denote known but negligible terms computed in [@HSW0304], the contributions computed partially in [@FengLM06], and the remaining, unknown contributions. For analytical results see the original references and [@review].
In order to discuss the phenomenological impact of [our new]{} contributions, we start by noting that $$\begin{aligned}
\Delta_\mu &= - 0.0018 \tan\beta\;\mbox{sign}\, \mu
\label{DeltamuMSUSY}\end{aligned}$$ in the case where all SUSY masses are equal [and much larger than $M_W$. Hence in the interesting region with $\tan\beta\sim 50$ the value of $\tan\beta$ extracted from ${a_\mu}^{\rm exp}$ will be off by roughly $10$%, if ${a_\mu}^{\rm SUSY,\Delta_\mu}$ is omitted in Eq. (\[ResultamuDeltamu\]). Fig. \[fig\] shows the impact of the new contribution on the dependence of ${a_\mu}$ on ${\tan\beta}$.]{}
[Importantly, as a dimensionless quantity $\Delta_\mu$ does not decouple for arbitrarily large SUSY masses. For slight mass splittings, $\Delta_\mu$ can be even larger than in Eq. (\[DeltamuMSUSY\]). For example, for $\tan\beta=50$ and $m_{L,\tilde{\mu}}=300$, $m_{R,\tilde{\mu}}=500$, $M_2=650$, $\mu=800$ GeV and $M_1=M_2/2$ one obtains a correction of $+14\%$ for ${a_\mu}^{\rm SUSY}$.]{}
[Among the SPS SUSY benchmark parameter points [@SPSDef] large effects are obtained at SPS 4 with $\tan\beta=50$ ($+8\%$), and at SPS 1b with $\tan\beta=30$ ($+6\%$). In particular, for SPS 4, which is already experimentally disfavoured by ${a_\mu}$, the contribution rises from ${a_\mu}^{\rm SUSY}(\mbox{SPS 4})
= 49 \times 10^{-10}$ to $53\times10^{-10}$ by including the new $\tan\beta$-enhanced correction. This corresponds to a rise of the deviation from the experimental value from $2.2\sigma$ to $2.6\sigma$. ]{}
![${a_\mu}$ as a function of ${\tan\beta}$ for four different values of degenerate SUSY masses. Solid (red) lines: correct ${a_\mu}$ as in Eq. (\[amuSUSYknown\]). Dashed (black) lines: ${a_\mu}$ without ${a_\mu}^{\rm SUSY,\Delta_\mu}$. Gray band: $1\sigma$ range of Eq. (\[deviation\]). \[fig\]](AmuPlotSmall.eps){width="\columnwidth"}
In conclusion we have identified a new ${\tan\beta}$-enhanced contribution to ${a_\mu}$ which first enters at the two-loop level. In scenarios with large values of ${\tan\beta}$ the new term ${a_\mu}^{\rm SUSY,\Delta_\mu}$ alters the MSSM phenomenology by typically 10%, but can have an even larger impact in certain regions of the parameter space. Our contribution is typically larger than the previously known supersymmetric two-loop corrections and should be included in global fits of electroweak precision data to the MSSM.
This work is supported by the DFG–SFB/TR9 *Computergestützte Theoretische Teilchenphysik*, by BMBF grant 05 HT6VKB and by the EU Contract No. MRTN-CT-2006-035482, FLAVIAnet”.
[999]{} J. P. Miller, E. de Rafael and B. L. Roberts, Rept. Prog. Phys. [**70**]{} (2007) 795. G. W. Bennett \[Muon Collaboration\], Phys. Rev. D [**73**]{} (2006) 072003. M. N. Achasov [*et al.*]{}, J. Exp. Theor. Phys. [**103**]{} (2006) 380 \[Zh. Eksp. Teor. Fiz. [**130**]{} (2006) 437\]. V. M. Aulchenko [*et al.*]{} \[CMD-2 Collaboration\], JETP Lett. [**82**]{} (2005) 743; R. R. Akhmetshin [*et al.*]{}, JETP Lett. [**84**]{} (2006) 413; Phys. Lett. B [**648**]{} (2007) 28. A. Aloisio [*et al.*]{} \[KLOE Collaboration\], Phys. Lett. B [**606**]{} (2005) 12. B. Aubert [*et al.*]{} \[BABAR Collaboration\], Phys. Rev. D [**70**]{} (2004) 072004; Phys. Rev. D [**71**]{}, 052001 (2005); Phys. Rev. D [**74**]{}, 091103 (2006); Phys. Rev. D [**73**]{}, 052003 (2006). F. Jegerlehner, Acta Phys. Polon. B [**38**]{} (2007) 3021. M. Davier, Nucl. Phys. Proc. Suppl. [**169**]{} (2007) 288. K. Hagiwara, A. D. Martin, D. Nomura and T. Teubner, Phys. Lett. B [**649**]{} (2007) 173. M. Benayoun, P. David, L. DelBuono, O. Leitner and H. B. O’Connell, Eur. Phys. J. C [**55**]{} (2008) 199. D. Stöckinger, J. Phys. G [**34**]{} (2007) R45. F. Feroz, B. C. Allanach, M. Hobson, S. S. AbdusSalam, R. Trotta and A. M. Weber, arXiv:0807.4512; L. Roszkowski, R. Ruiz de Austri and R. Trotta, JHEP [**0707**]{}, 075 (2007); S. Heinemeyer, X. Miao, S. Su and G. Weiglein, arXiv:0805.2359. J. R. Ellis, S. Heinemeyer, K. A. Olive, A. M. Weber and G. Weiglein, JHEP [**0708**]{}, 083 (2007). D. W. Hertzog, J. P. Miller, E. de Rafael, B. Lee Roberts and D. Stöckinger, arXiv:0705.4617. M. Alexander, S. Kreiss, R. Lafaye, T. Plehn, M. Rauch, and D. Zerwas, Chapter 9 in M. M. Nojiri [*et al.*]{}, arXiv:0802.3672. J. L. Lopez, D. V. Nanopoulos and X. Wang, Phys. Rev. D [**49**]{}, 366 (1994); U. Chattopadhyay and P. Nath, Phys. Rev. D [**53**]{}, 1648 (1996); T. Moroi, [Phys. Rev. D]{} [**53**]{} (1996) 6565 \[Erratum-ibid. [**56**]{} (1997) 4424\]. G. Degrassi and G. F. Giudice, Phys. Rev. D [**58**]{} (1998) 053007. S. Heinemeyer, D. Stöckinger and G. Weiglein, Nucl. Phys. B [**690**]{} (2004) 62; Nucl. Phys. B [**699**]{} (2004) 103. M. Carena, M. Olechowski, S. Pokorski and C.E.M. Wagner, Nucl. Phys. [**B426**]{} (1994) 269; L.J. Hall, R. Rattazzi and U. Sarid, Phys. Rev. [**D50**]{} (1994) 7048; C. Hamzaoui, M. Pospelov and M. Toharia, Phys. Rev. D [**59**]{} (1999) 095005; K. S. Babu and C. F. Kolda, Phys. Rev. Lett. [**84**]{} (2000) 228; G. Isidori and A. Retico, JHEP [**0111**]{} (2001) 001; A. J. Buras, P. H. Chankowski, J. Rosiek and L. Slawianowska, Nucl. Phys. B [**659**]{} (2003) 3; T. Blazek, S. Raby and S. Pokorski, Phys. Rev. D [**52**]{} (1995) 4151. M. S. Carena, D. Garcia, U. Nierste and C. E. M. Wagner, Nucl. Phys. B [**577**]{} (2000) 88 \[arXiv:hep-ph/9912516\]. T. Appelquist and J. Carazzone, Phys. Rev. D [**11**]{} (1975) 2856. T. F. Feng, X. Q. Li, L. Lin, J. Maalampi and H. S. Song, Phys. Rev. D [**73**]{} (2006) 116001 \[arXiv:hep-ph/0604171\].
B. C. Allanach [*et al.*]{}, in [*Proc. of the APS/DPF/DPB Summer Study on the Future of Particle Physics (Snowmass 2001)* ]{} ed. N. Graf, Eur. Phys. J. C [**25**]{} (2002) 113.
[^1]: [Note that there is no double-counting between the non-$\tan^2\beta$-enhanced terms implicitly contained in (\[ResultamuDeltamu\]) and the terms in the second and third line of (\[amuSUSYknown\]).]{}
[^2]: In our numerical analysis we parametrize the one-loop result in terms of the muon decay constant $G_\mu$, i.e. we replace $\pi\alpha/s_W^2\to \sqrt2 G_\mu M_W^2$, in order to absorb further universal two-loop corrections. This gives rise to slight numerical differences compared to [@review].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'By studying the optical conductivity of Bi$_2$Sr$_{2-x}$La$_x$CuO$_6$ and Y$_{0.97}$Ca$_{0.03}$Ba$_{2}$Cu$_{3}$O$_{6}$, we show that the metal-to-insulator transition (MIT) in these hole-doped cuprates is driven by the opening of a small gap at low $T$ in the far infrared. Its width is consistent with the observations of Angle-Resolved Photoemission Spectroscopy in other cuprates, along the nodal line of the $k$-space. The gap forms as the Drude term turns into a far-infrared absorption, whose peak frequency can be approximately predicted on the basis of a Mott-like transition. Another band in the mid infrared softens with doping but is less sensitive to the MIT.'
author:
- 'S. Lupi$^{1}$, D. Nicoletti$^{1}$, O. Limaj$^{1}$, L. Baldassarre$^{1}$, M. Ortolani$^{1}$, S. Ono$^{2}$, Yoichi Ando$^{3}$, and P. Calvani$^{1}$'
title: 'Far-infrared absorption and the metal-to-insulator transition in hole-doped cuprates'
---
epsf
The parent compounds of high-$T_c$ cuprates are half-filled antiferromagnetic (AF) insulators where the Coulomb repulsion opens a wide charge-transfer (CT) gap ($\agt$ 1.5 eV) in the excitation spectrum. By adding a few percent holes ($p$) or electrons ($n$) per Cu ion, the CT gap is filled by a broad infrared absorption, while Angle-Resolved Photoemission Spectroscopy (ARPES) reveals fully gapped single particle excitations at low temperature [@ShenARPES]. At $p \agt$ 0.05, like in La$_{2-x}$Sr$_x$CuO$_4$ (LSCO), or $n \agt$ 0.12 like in Nd$_{2-x}$Ce$_x$CuO$_4$ (NCCO), the insulator eventually turns at low $T$ into a superconductor, above $T_c$ into a “strange metal” characterized by a pseudogap in the density of states [@Timusk99]. Much effort has been devoted in the last two decades to understand these three phases and their excitation spectrum. However, less attention has been paid to the mechanism of the transition from the AF insulator to the strange metal and vice versa, which is far from being clear. For example, the onset of a metallic state is observed at $p$ or $n$ values much higher than those which destroy the AF long range order [@Iye92]. Therefore, magnetism should not play a dominant role in the metal-to-insulator transition (MIT). The mechanisms responsible for the localization of the carriers then may be disorder due to dopant ions [@Atkinson], local electron-spin interactions [@Defilippis], or the electron-phonon coupling [@Cimento].
In the present paper we investigate the MIT by studying the behavior of the optical conductivity $\sigma_{1}(\omega)$ across the transition, in the single Cu-O layer cuprate Bi$_2$Sr$_{2-x}$La$_x$CuO$_6$ (BSLCO). Therein, hole doping can be accurately controlled like in LSCO by replacing Sr by La. In BSLCO, $p$ decreases [@Ono00] for increasing $x$ according to a relation which is not linear, due to compensation effects, but well known [@Ando00; @Ono03]. Finally, $T_c$ is low enough ($\simeq$ 30 K at optimum doping) to allow for a study of the normal state of the Cu-O planes at low $T$. Here we have measured four single crystals [@Ono03] with $x$ = 1.0, 0.9, 0.8, and 0.7, corresponding to $p \simeq$ 0.03, 0.07, 0.10, and 0.12, respectively. As at $x=1.0$ one reaches the lowest $p\simeq 0.03$ allowed by the miscibility of La in the Bi$_2$Sr$_{2}$CuO$_6$ matrix. Therefore, we have also measured a single crystal of Y$_{1-x}$Ca$_{x}$Ba$_{2}$Cu$_{3}$O$_{6}$ (YCBCO) with $x$ = 0.03 and $p \simeq$ 0.015. As in YCBCO the CuO chains are empty and only the Cu-O plane contributes to the $ab$ optical response [@Erb], we could thus probe five Cu-O planes with decreasing hole doping, from an underdoped metallic state ($p \simeq$ 0.12) to an AF insulating phase ($p \simeq$ 0.015).
=8.6cm
The $ab$-plane resistivity $\rho_{ab}(T)$ of the four BSLCO crystals is reported in the insets of Fig. \[sigma\] and dc conductivity values are shown, for selected temperatures, on the vertical axis of the same Figure. In $a$, for $p\simeq$ 0.12, $\rho_{ab}(T)$ shows a metallic behavior above the superconducting transition at $T_c =$ 13 K. The crystal with $p \simeq$ 0.10 in Fig. \[sigma\]-$b$ displays a metallic behavior above 50 K, and a broad minimum above the transition to a superconducting state [@Ono03] at $T_c \simeq$ 1.4 K. At $p\simeq 0.07$, $\rho_{ab}(T)$ in $c$ is nearly constant down to $\sim$ 50 K. Below, it diverges for $T \to$ 0 according to a variable range hopping regime [@Ono03]. A clearly semiconducting behavior at any $T$ ($d\rho_{ab}/dT<0$) is instead shown by the compound with $p\simeq$ 0.03 in Fig. \[sigma\]-$d$. Therefore the MIT can be placed between $p\simeq 0.10$ and $p\simeq0.07$. This finding is consistent with the Mott-Ioffe-Regel limit, which fixes the metal-to-insulator crossover and, for the Cu-O planes, can be written as [@Ando08] $k_F \cdot l = (h c_0 / e^2 \rho_{ab}) \sim 1$. Here, $k_F$ is the Fermi wavevector, $l$ the carrier mean free path, and $c_0$ is the $c$-axis lattice spacing. Indeed, following Ref. , from the $\rho_{ab}$ (10 K) in Fig. \[sigma\] one obtains $k_F \cdot l $ = 3.4 at $p$ = 0.10, $k_F \cdot l $ = 0.05 at $p$ = 0.07. Magnetic fields on the order of 60 T displace the MIT [@Ono00] to $p \simeq$ 1/8.
The $ab$-plane reflectivity $R(\omega)$ of the five samples was measured at near-normal incidence from 40 or 50 to 22000 cm$^{-1}$ at different $T > T_c$, shortly after cleaving the sample. The real part of the optical conductivity $\sigma_{1}(\omega)$, as obtained from $R(\omega)$ via Kramers-Kronig (KK) transformations, is shown in Fig. \[sigma\]. The extrapolations to high frequency were based on the data of Ref. , those to zero frequency on Drude-Lorentz fits, which provided deviations from the measured $\sigma_{dc}(T)$ of a few percent. Afterwards, the extrapolations were adjusted to $\sigma_{dc}(T)$ (including that at 20 K in $a$, not shown). In insulating YCBCO, a check value of $\sigma_{dc}$ was measured at 200 K [@private] and reported in Fig. \[sigma\]-$e$. Figures \[sigma\]-$a$ ($p \simeq$ 0.12) and -$b$ ($p \simeq$ 0.10) exhibit a Drude term which partially shields the phonon peaks. These are instead well evident in panels $c$ ($p \simeq$ 0.07), $d$ ($p \simeq$ 0.03), and $e$ ($p \simeq$ 0.015). Their frequencies are in good agreement with those previously measured [@TajimaPhon] on Bi$_2$SrLaCuO$_y$ and YBa$_{2}$Cu$_{3}$O$_{y}$. Around 10000 cm$^{-1}$, $\sigma_{1}(\omega)$ increases steeply due to the CT transition between Cu $3d$ and O $2p$ orbitals, as in the other cuprates.
Let us now focus on the broad far- and mid-infrared contributions in Fig. \[sigma\]. To better understand their behavior with doping and temperature, in Fig. \[subtract\] $\sigma_1(\omega)$ is reported at 300 K (dashed lines) and at the lowest $T$ (solid lines), after the phonon lines have been subtracted by accurate Lorentzian fits (those at the lowest T are shown in Fig. \[sigma\]). Further fits - not reported in Fig. 2 as they could be hardly distinguished from the data - were then made on the subtracted spectra, by using a Drude term and two broad bands at $\omega_{FIR}$ and $\omega_{MIR}$ (open symbols). In the metallic phase (Fig. \[subtract\]-$a$ and -$b$) at all temperatures, data are fit by a simple Drude term plus a $T$-independent mid-infrared (MIR) band peaked at $\omega_{MIR} \sim$ 2000 cm$^{-1}$. As $p$ decreases below the critical hole content $p_{MIT}$ (Fig. \[subtract\]-$c$), the Drude term turns at low $T$ into a far-infrared (FIR) band peaked at $\omega_{FIR} \sim$ 200 cm$^{-1}$. Correspondingly, $\sigma_{dc}$ becomes vanishingly small. At $p\simeq$ 0.03 (Fig. \[subtract\]-$d$), a gap $\sim$ 100 cm$^{-1}$ opens in $\sigma_1(\omega)$ at $T$ = 10 K. Meanwhile, $\omega_{FIR}$ shifts to $\sim$ 400 cm$^{-1}$. Finally, at $p\simeq$ 0.015 in Fig. \[subtract\]-$e$, the insulating gap extends along the whole far-infrared region, and the FIR peak is displaced to $\sim$ 1000 cm$^{-1}$. At 300 K, both in $c$ and $d$ the FIR band includes a background - suggestive of incoherent charge transport - which extends to $\omega$ = 0 and accounts for the residual dc conductivity of these samples at room temperature.
=8.6cm
The results of Fig. \[subtract\] can be compared with those of ARPES experiments. Figure \[ARPES-IR\] reproduces that in Ref. , which shows the gap $\Delta$ at low doping and low $T$ in the density of states of different cuprates, measured at the leading edge midpoint along the nodal line of the Brillouin zone. In BSLCO with $p \simeq$ 0.05 and $p \simeq$ 0.07, a low-$T$ suppression of the spectral-weight was also observed along that direction in the $k$-space, but the gap width could not be measured for lack of resolving power [@HashimotoARPES]. We have also plotted in Fig. \[ARPES-IR\] both the present determinations of the gap in BSLCO, YCBCO, and those extracted from previous measurements in LSCO [@Lucarelli03] and in NCCO [@Lupi99]. In all cases, the infrared gap 2$\Delta$ is obtained after phonon subtraction, by a linear extrapolation to zero [@Katsufuji] of the optical conductivity along the gap edge (see an example in the inset), and then divided by two (by assuming full hole-electron symmetry). If also the IR gap were taken at mid-height, the full symbols would displace toward high frequencies by about a factor of 2. In Fig. \[ARPES-IR\], independently of the procedure employed, both the infrared and ARPES observations consistently indicate that a small gap - just a few meV wide - opens in the density of states at the MIT. In LSCO, however, for reasons which deserve further investigations, the ARPES gap is not observed the infrared [@Lucarelli03; @Padilla]. In turn, the band at $\omega_{MIR}$ can be associated with the incoherent background observed at $\sim 0.5$ eV below the quasi-particle peak in the antinodal ARPES spectra of cuprates [@ShenARPES; @HashimotoARPES; @MengCondMat].
The mechanism of the MIT in the Cu-O plane appears clearly in Fig. \[subtract\]: as $p$ decreases below $p_{MIT}$, the carriers increasingly localize at low $T$ into states having an optical ionization energy $E_0$ measured by $\omega_{FIR}$. $E_0$ increases for decreasing $p$ and reaches about 0.12 eV in the limit of high dilution (Fig. \[sigma\]-$e$). This value is in excellent agreement with previous observations in electron-doped NCCO at the lowest doping [@Calvani94].
A simple calculation can predict $E_0$. Indeed, as usually done in semiconductors [@Mott], one may model the charge injected in the Cu-O plane as a hole orbiting with radius $R$ either around a defect or an impurity (in case of disorder) or within an attractive potential well due to lattice distortion (in a polaronic model). $E_0$ is related to $R$ through the hydrogen-like equation
$$E_{0}=\left ( \frac{3}{2}\frac{a_{0}}{R} \right )^{2}Ry
\label{E0}$$
where $a_0$ = 0.0529 nm is the Bohr radius and $Ry$ = 109737 cm$^{-1}$ is the Rydberg constant. The Mott transition occurs at a 2-D critical density of holes per Cu ions $p_{MIT}$ such that [@nota] the orbits of adjacent charges overlap in the Cu-O planes. This condition implies that the hole density $\rho_{MIT} \sim \frac{1}{\pi R^2}$ and the Cu density $\rho_{Cu} =1/a^2$ satisfy the relation
$$p_{MIT} = \rho_{MIT}/ \rho_{Cu} \sim \frac{a^2}{\pi R^2}
\label{pmit}$$
Here, we use $a$ = 0.386 nm, averaged between the Cu-O lattice parameters of [@TajimaPhon] Bi$_2$SrLaCuO$_y$ and YBa$_2$Cu$_3$O$_y$ (0.383 nm and 0.389 nm, respectively). Assuming from Fig. 2 $p_{MIT} \simeq 0.08$, and an average $\rho_{Cu}$ = 6.7 nm$^{-2}$, one obtains, from Eq. 2, $R \simeq 0.77$ nm and, from Eq. 1, $E_0 \simeq 1150$ cm$^{-1}$. This value is in very good agreement with the $\omega_{FIR} \simeq 1000$ cm$^{-1}$ observed here at high hole dilution (in the YCBCO with $p\simeq 0.015$ of Fig. \[subtract\]-$e$).
Therefore, one should find the origin of the FIR band in order to understand the MIT mechanism. A first candidate is disorder. In cuprates added with Zn impurities, or irradiated by high-energy particles, the metallic phase can be destroyed and the Drude spectral weight strongly reduced, due to a poor screening of the impurities and to the resulting fluctuating potentials in the Cu-O planes [@Basov98; @Basov94]. Recent calculations [@Atkinson] show that such disorder effects are amplified in a $d$-wave electronic symmetry and that, for increasing impurity content, the Drude term turns into a FIR peak at a finite frequency. This framework is consistent with the behavior of BSLCO, where the insulating phase is reached by adding La impurities, not with that of LSCO which is opposite. It seems then that the parameter which governs the metal-to-insulator transition is $p$, rather than $x$.
A different scenario can then be invoked to explain the MIT. Indeed, in BSLCO the FIR peak behaves with $p$ like a FIR band of NCCO vs. $n$ [@Lupi99]. This absorption was attributed to large polarons [@Cimento] and its softening was explained in terms of polaron-polaron interactions which increase with $n$ [@Devreese; @Lorenzana]. At room temperature, where $k_BT \approx \Delta$, incoherent polaron hopping takes place: this may explain the flat background observed in Fig. \[subtract\]-$c$ and -$d$ and the resulting, non-vanishing $\sigma_{dc}$ at $p < p_{MIT}$. The above scenario is also consistent with recent calculations of the optical conductivity in hole-doped cuprates. They are based on a $t$-$J$-Holstein approach, where the FIR band has a dominant electron-phonon character, while the MIR band is attributed mainly to electron-spin interactions [@Mishchenko08]. The observation of the MIR band at $p > p_{MIT}$ therefore indicates that local antiferromagnetic fluctuations survive in the underdoped metallic state, as also reported previously for YBCO [@Basov]. As $p$ decreases, in Fig. \[subtract\] $\omega_{MIR}$ shifts steadily to higher energies, to reach $\sim$ 4500 cm$^{-1}$ at $p \simeq$ 0.015. This value is consistent with the determinations of the MIR peak in the other cuprates ($\sim$ 0.5 eV) [@TimuskBasov]. Despite that shift, in Fig. \[subtract\] the MIR absorption does not seem to play a major role in the MIT.
=8.6cm
In conclusion, we have measured both the dc and infrared conductivity of BSLCO with different doping, and of a lightly doped sample of YCBCO. We have thus monitored the mechanism of the metal-to-insulator transition in the hole-doped Cu-O plane. For any $p$ from 0.12 down to 0.015 we have detected a MIR band which hardens for decreasing $p$ but is poorly sensitive to the MIT. In the far infrared, on the contrary, the Drude term of the metallic phase collapses at the MIT into a FIR band at finite frequency. As $p$ is further decreased, this band shifts to higher energies, leaving behind a gap at low $T$, a flat absorption tail at high $T$. The FIR gap here observed is in quantitative agreement with that reported by ARPES in other cuprates with similar doping at low $T$, and the FIR peak frequency is correctly predicted by a simple Mott-transition model. Both bands here identified are consistent with a recent model, where the carrier localization is interpreted in terms of charge and spin polarons. According to this interpretation, the FIR band is mainly due to electron-phonon interaction, the MIR band to electron-spin interaction.
We are gratefully indebted to Andreas Erb for providing the YCBCO crystal. This work has been partially funded by PRIN 2005022492. S.O. was supported by KAKENHI No. 20740213.
K. M. Shen *et al.*, Phys. Rev. B **69**, 054503 (2004). For a review, see T. Timusk and B. Statt, Rep. Progr. Phys. **62**, 61 (1999). Y. Iye, in *Properties of High Temperature Superconductors*, edited by D. M. Ginsberg (World Scientific, Singapore, 1992), Vol. 3, pp. 285-361. W.A. Atkinson and P.J. Hirschfeld, Phys. Rev. Lett. **88**, 187003 (2002). G. De Filippis, V. Cataudella, A. S. Mishchenko, and N. Nagaosa, Phys. Rev. Lett. **99**, 146405 (2007). For a review, see P. Calvani, Riv. Nuovo Cimento **24**, 1 (2001). S. Ono *et al.*, Phys. Rev. Lett. [**85**]{}, 638 (2000). Yoichi Ando *et al.*, Phys. Rev. B **61**, R14956 (2000); Phys. Rev. B **63**, 069902 (2001). S. Ono and Yoichi Ando, Phys. Rev. B **67**, 104512 (2003). A. Jánossy, T. Fehér, and A. Erb, Phys. Rev. Lett. **91**, 177001 (2003). Yoichi Ando, J. Phys. Chem. Solids **69**, 3195 (2008). I. Terasaki *et al.*, Phys. Rev. B **41**, 865 (1990). A. Jánossy, private communication. S. Tajima *et al.*, Phys. Rev. B **43**, 10496 (1991). M. Hashimoto *et al.*, Phys. Rev. B **77**, 094516 (2008). A. Lucarelli *et al.*, Phys. Rev. Lett. **90**, 037002 (2003). S. Lupi *et al.*, Phys. Rev. Lett. **83**, 4852 (1999). T. Katsufuji et al., Phys. Rev. Lett. **75**, 3497 (1995). W. J. Padilla *et al.*, Phys. Rev. B **72**, 060511(R) (2005). J. Meng *et al.*, arXiv:0808.0806v2 \[cond-mat.supr-con\] (2008). P. Calvani *et al.*, Solid State Commun. **91**, 113 (1994). N. Mott, in *Metal-Insulator Transitions* (Taylor & Francis, London, 1990). The three dimensional Mott criterion can be extended to a charge moving in the Cu-O plane with small corrections, as shown by Kastner *et al.*, \[Rev. Mod. Phys. **70**, 897 (1998)\]. In the present general approach, we adopted the simple geometrical factor $1/\sqrt{\pi}$ instead of the usual, phenomenological valuer 0.25 which leads, for example, to a large mismatch with the $p_{MIT}$ of Si:P \[C. Kittel, Introduction to Solid State Physics, 7th ed., John Wiley and Sons, New York 1996, p. 285\]. D. N. Basov, B. Dabrowski, and T. Timusk, Phys. Rev. Lett. **81**, 2132 (1998). D. N. Basov, *et al.*, Phys. Rev. B **49**, 12165 (1994). J. Tempere and J. T. Devreese, Phys. Rev. B **64**, 104504 (2001); Eur. Phys. J. B **20**, 27 (2001). J. Lorenzana, Europhys. Lett. **53**, 532 (2001). A. S. Mishchenko *et al.*, Phys. Rev. Lett. **100**, 166401 (2008). Y. S. Lee *et al.*, Phys. Rev. B **72**, 054529 (2005). D. N. Basov and T. Timusk, Rev. Mod. Phys. **77**, 721 (2005).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We give a detailed analysis of the particle spectrum and the perturbative unitarity of the recently introduced Weyl-invariant version of the new massive gravity in 2+1 dimensions. By computing the action up to second order in the fluctuations of the metric, the gauge and the scalar fields around the anti-de Sitter (AdS) and flat vacua, we find that the theory describes unitary (tachyon and ghost-free) massive spin-2, massive (or massless) spin-1 and massless spin-0 excitations for certain ranges of the dimensionless parameters. The theory is not unitary in de Sitter space. Scale invariance is either broken spontaneously (in AdS background) or radiatively (in flat background) and hence the masses of the particles are generated either spontaneously or at the second loop order.'
author:
- 'M. Reza Tanhayi'
- Suat Dengiz
- Bayram Tekin
title: 'Unitarity of Weyl-Invariant New Massive Gravity and Generation of Graviton Mass via Symmetry Breaking'
---
Introduction
============
Einstein’s general relativity (GR) is expected to be modified at both large (astrophysical) (IR) and small (UV) regions. There are ample theoretical (in the case of UV) and experimental (in the case of IR) reasons to conclude that GR can only be an effective theory that works perfectly in the intermediate regions, such as the solar system and *etc*. Apriori, the nature of UV and IR modifications is quite different. For UV modifications, experience from quantum field theory dictates that if one is to define a perturbatively well-behaved (that is renormalizable and unitary) gravity theory, then one must introduce higher powers of curvature that modify both the tree-level propagator structure and the interactions. Unfortunately, it is well-known that such a theory simply does not exist in four dimensions [@stelle]. On the other extreme, IR modifications consist of introducing a cosmological constant and/or mass to the graviton. Even though, theoretically, cosmological Einstein theory is the easiest extension of GR, the problems with the cosmological constant are well-known (such as the difficulty of keeping it small in the quantum theory). Graviton mass on the other hand is a very subtle issue. Given a massless free spin-2 field about a maximally symmetric background, one can introduce the Fierz-Pauli term respecting the background symmetries to get a massive spin-2 field. But such a theory does not seem to arise from a diffeomorphism invariant interacting gravity theory save the unique case of the 2+1 dimensions.
For $ D= 2+1 $ dimensions, new massive gravity (NMG) introduced in [@BHT] provides a non-linear extension of the Fierz-Pauli massive spin-2 theory. For the mostly plus signature the action reads[^1] $$I_{NMG}=\frac{1}{\kappa^2} \int d^3 x \sqrt{-g} \Big [\sigma R-2 \lambda m^2 + \frac{1}{m^2} \Big ( R^2_{\mu\nu}-\frac{3}{8}R^2 \Big )\Big],
\label{nmg}$$ which attracted a lot of attention: Detailed works on it appeared in [@BHT; @GulluTekin; @deser; @nakasone; @liusun; @canonical; @cubic; @Aliev] regarding its unitarity, solutions and *etc*. Remarkably, higher curvature terms in this theory provide in some sense both the viable UV and IR modifications that one is interested in. Unfortunately, this state of happy affairs do not extend to four dimensions. But in any case, 2+1 dimensional gravity is a valuable theoretical lab for ideas in quantum gravity. \[In fact, according to the proposal of Horava for which spacetime’s spectral dimension reduces at high energies, 3D gravity becomes much more relevant and NMG appears as part of the non-covariant 3+1 dimensional action [@horava].\]
Having understood that NMG describes a consistent parity-invariant[^2] massive spin-2 theory in 2+1 dimensions, the next natural question is to ask if graviton mass can be generated from breaking a symmetry (not diffeomorphism invariance) in this theory in analogy with the Higgs mechanism in the Standard Model. This question was answered in the affirmative recently in [@DengizTekin] by finding a local scale-invariant (Weyl-invariant) version of NMG and showing that in the case of (A)dS space, the vacuum breaks the conformal symmetry spontaneously and for flat space conformal symmetry is broken at the two loop level [@tantekin] via the Coleman-Weinberg mechanism [@coleman]. Referring to [@DengizTekin] for the details of how the Weyl-invariant extension of NMG (and other higher curvature models, such as the Born-Infeld NMG [@gullusismantekin]) was introduced and how symmetry gets broken, here we just quote the final expression $$\begin{aligned}
S_{WNMG}&= \int d^3 x \sqrt{-g} \bigg \{\sigma \Phi^2 \Big(R-4\nabla \cdot A -2A^2 \Big) \\
& \qquad \quad \qquad \quad +\Phi^{-2} \bigg [R^2_{\mu\nu}-\frac{3}{8} R^2-2 R^{\mu\nu}\nabla_\mu A_\nu+ 2R^{\mu\nu}A_\mu A_\nu \\
& \qquad \quad \qquad \quad\qquad \quad +R\, \nabla \cdot A-\frac{1}{2}R A^2+2 F_{\mu\nu}^2 +(\nabla_\mu A_\nu)^2 \\
& \qquad \quad\qquad \quad \qquad \quad -2 A_\mu A_\nu \nabla^\mu
A^\nu-(\nabla \cdot A)^2+\frac{1}{2}A^4 \bigg ] \bigg
\}+S_\Phi+S_{A_\mu},
\label{winmg}
\end{aligned}$$ where $S_\Phi$ and $S_{A_\mu}$ are the Weyl-invariant scalar and gauge field actions, which are given by $$\begin{aligned}
& S_{\Phi}=- \frac{1}{2}\int d^3 x \sqrt{-g}\bigg \{\Big (
\partial_\mu \Phi-\frac{1}{2} A_\mu\Phi\Big)^2+\nu\Phi^6\bigg\},\\
&S_{A_\mu}= \beta \int d^3 x \sqrt{-g}\,\, \Phi^{-2}
F^2_{\mu \nu}.
\end{aligned}$$ The action (\[winmg\]) is invariant under the simultaneous transformation of the metric and the fields as $$\label{gaugetran}
g_{\mu\nu} \rightarrow g^{'}_{\mu\nu}=e^{2 \zeta(x)} g_{\mu\nu}, \hskip 1 cm \Phi \rightarrow \Phi^{'} =e^{-\frac{(n-2)}{2}\zeta(x)}
\Phi, \hskip 1 cm A_\mu \rightarrow A^{'}_\mu = A_\mu - \partial_\mu \zeta(x).$$ It is important to note that there are no dimensionful parameters in the theory, on the other hand, local scale invariance does not fix the relative numerical coefficients of various parts which are independently scale-invariant. Generically, up to a numerical scaling of the total action, there are 4 dimensionless parameters that one can introduce. By scaling the total action, we set the coefficient of the kinetic part of the scalar action to its canonical non-ghost form and to keep contact with the NMG, we take the numerical coefficient of the quadratic part of the action to be 1 (this choice can easily be relaxed). Therefore, we have 3 dimensionless parameters $\sigma, \nu, \beta$.
The Weyl-invariant theory (\[winmg\]) obviously is much larger than NMG (\[nmg\]) in the sense that when one sets $
\Phi=\sqrt{m} $, $ \nu=2 \lambda $ and $ A_\mu= 0 $, at the level of the action, one recovers NMG (\[nmg\]), with a fixed gravitational coupling $ \kappa=m^{-1/2} $ and a fixed cosmological. \[In fact all the dimensionful scales are determined by the symmetry breaking order parameter $ <\Phi>
=\Phi=\sqrt{m} $\]. Let us consider the infinite dimensional field space $ {\cal M}=[g_{\mu \nu}, A_\mu, \Phi]$ to be the space of all fields satisfying the field equations derived from the action (\[winmg\]). It was shown in [@DengizTekin] that “NMG-point”, that is \[$ g_{\mu \nu}, 0, \sqrt{m} $\], is in ${\cal
M}$. Moreover, if one freezes the scalar and the gauge fields to these NMG-point values and consider fluctuations just in the metric directions, one exactly gets the same spectrum as NMG around its AdS or flat vacua. This was shown in [@DengizTekin] but what was left out in that work and which will be remedied here, is a complete study of the second order fluctuations of all the fields around the NMG point (or the vacuum of the theory). Namely, apriori, the stability and unitarity of (\[winmg\]) is not clear for all allowed fluctuations in the metric, gauge and scalar field directions on ${\cal M}$. The main task of this paper is to show that NMG-point is stable by proving that there are no ghosts and tachyons in the particle spectrum of the Weyl-invariant action (\[winmg\]). Therefore, mass of the graviton and the mass of the gauge field is consistently generated by the symmetry breaking mechanism of the conformal symmetry.
The layout of the paper is as follows: In section II, we first find the expansion of the action up to second order in the fields around the (A)dS or flat vacua. This section also discusses issues about Jordan versus Einstein frame and the Weyl-invariant gauge-fixing in the gauge sector. In section III, we decouple the fields and identify the masses and also the unitarity regions of the dimensionless parameters. We collect some useful computations in the appendices.
Quadratic fluctuations about the vacuum
=======================================
In [@DengizTekin], the field equations coming from the action (\[winmg\]) and its vacuum solution were given. Here, we do not depict the field equations, since they are rather lengthy, instead we note that the vacuum solution (let us first take a dS or AdS vacuum, as the flat vacuum will follow these) is given as $$\Phi_{vac}= \sqrt{m}, \hskip 1 cm A^\mu_{vac}=0, \hskip 1 cm g_{\mu \nu}=\bar{g}_{\mu \nu},$$ here $ \bar{R}_{\mu \nu}=2 \Lambda \bar{g}_{\mu \nu} $. And the cosmological constant satisfies[^3] $$\Lambda^2+4 \sigma m^2 -\nu m^4 =0.
\label{vacuum}$$ Given $m^2$, generically, there are two vacua $$\Lambda_{\pm}= m^2 \Big[ - 2 \sigma \pm \sqrt{4\sigma^2 + \nu} \Big ].
\label{lambda_exp}$$ Our task now is to study the stability of these vacua and also study the particle spectrum of the model. This can be achieved by considering the second order fluctuations about the vacuum following from $$\Phi=\sqrt{m}+\tau \Phi_{L}, \hskip 1 cm A_\mu=\tau A^{L}_\mu, \hskip 1 cm g_{\mu \nu}=\bar{g}_{\mu \nu}+\tau h_{\mu \nu},
\label{agg}$$ where we have introduced $ \tau$, a small dimensionless parameter to keep the track of the expansion orders. In what follows, we will use the conventions given in [@ubinmg]. The expansions of various curvature terms are needed in the computations, so we collect them in Appendix A.
Since the action (\[winmg\]) is highly complicated with fields coupled to each other, it is a non-trivial task to find the basic oscillators (free particles) of the theory. There are a couple of paths one can take. For example, one can linearize field equations and try to decouple the fields. Or, one can transform the action to the Einstein frame and then find the field equations and do the linearization. These two paths do not give an efficient way for the study of the spectrum. \[See Appendix B for the Einstein frame version of the Weyl-invariant quadratic theory.\] As a third way, one can directly compute the action up to quadratic order in the fluctuations about its vacua, which we shall adopt here. This will lead to coupled fields at the quadratic level. Then we will find a way to decouple the basic free fields in the theory. This procedure is quite lengthy but there seems to be no way of avoiding it and it is still easier than the above mentioned procedures. The action (\[winmg\]), after making use of the field fluctuations (\[agg\]) and the relevant formulas in the Appendix A, can be written as $$S_{WNMG}=\bar{S}_{WNMG}+\tau S^{(1)}_{WNMG}+\tau^2 S^{(2)}_{WNMG}+{\cal
O}(\tau^3),$$ where $ \bar{S}_{WNMG} $ is the value of action evaluated in the background which is irrelevant for our purposes. On the other hand $ S^{(1)}_{WNMG} $ vanishes in the vacuum, which also gives us the vacuum equations without going into the details of finding the full field equations [@DengizTekin]. Finally the quadratic part $ S^{(2)}_{WNMG} $, after making use of the vacuum equations and dropping the boundary terms, reads as $$\begin{aligned}
{S}_{WNMG}^{(2)}=\int d^3x\sqrt{-\bar{g}}\bigg \{& -\frac{1}{2}(\partial_\mu \Phi^L)^2 + \Big (6 \sigma \Lambda -\frac{9 \Lambda^2}{2m^2}-
\frac{15 \nu m^2}{2} \Big ) \Phi^2_L \\
& +\frac{2\beta +5 }{2m}(F^L_{\mu\nu})^2- \Big ( 2 \sigma m +\frac{\Lambda}{m}+\frac{m}{8} \Big ) A^2_L-\frac{1}{m}(\bar{\nabla}\cdot
A^L)^2 \\
&+\frac{1}{m}({\cal G}^L_{\mu\nu})^2-\Big(\frac{\sigma m}{2}-\frac{\Lambda}{4m} \Big )h^{\mu \nu}
{\cal G}^{L}_{\mu \nu}-\frac{1}{8m}R^2_L \\
&+ \Big (2 \sigma \sqrt{m}+\frac{\Lambda}{m \sqrt{m}} \Big ) \Phi^L R^L - \Big ( 8 \sigma \sqrt{m}+\frac{4 \Lambda}{m
\sqrt{m}}+\frac{\sqrt{m}}{2} \Big ) \Phi^L \bar{\nabla}\cdot A^L\bigg \}.
\label{linearform}
\end{aligned}$$ In deriving the above expansion, in addition to the formulas in the Appendix, the following relations have also been used $$\Phi^2= m \Big (1+2 \tau\frac{\Phi_{L}}{\sqrt{m}}+\tau^2\frac{\Phi^2_{L}}{m}
+{\cal O}(\tau^3) \Big ), \hskip 1 cm \Big (\nabla_\mu A_\nu \Big)=\tau \bar{\nabla}_\mu A^{L}_\nu-\tau^2
\Big(\Gamma^{\gamma}_{\mu \nu} \Big)_{L} A^{L}_\gamma+{\cal
O}(\tau^3).$$ The first thing to observe is that to have a non-ghost and canonically normalized (that is $-\frac{1}{4}$) kinetic term for the Maxwell field, we should set $\beta =- \frac{11}{4}$, which we do from now on. As it stands, the fields are still coupled and one should find a way to decouple them. Such a coupling between the scalar field and the curvature is expected, since we are dealing with a non-minimally (in fact conformally) coupled scalar field to gravity. The scalar field also couples to the gauge field as demanded by conformal invariance. To understand how one could decouple these fields in (\[linearform\]), let us study a simpler model (scalar-tensor theory) first, and then come back to our problem.
Quadratic fluctuations and the spectrum of the conformally coupled scalar-tensor theory
---------------------------------------------------------------------------------------
We choose the 2+1 dimensional conformally coupled scalar-tensor action which is given by $$S_{S-T}=\int d^3 x \sqrt{-g} \Big (\Phi^2 R + 8 \partial_\mu \Phi \partial^\mu \Phi-\frac{\nu}{2}\Phi^6 \Big),
\label{scalartensor}$$ and ask what the particle spectrum is around its (A)dS vacuum. By just inspecting the action, one mistakingly namely think that the scalar field is a ghost since it comes with a negative kinetic energy part. But this is actually a red-herring, since the action is in the Jordan frame one cannot draw such a conclusion from the full non-linear theory. One must either go to the Einstein frame where the fundamental degrees of freedom are more transparent or, in the Jordan frame, study the quadratic fluctuations of the fields around the vacuum. We will do both below.
Under the conformal rescaling $g_{\mu \nu}(x)=\Omega^{-2}(x)g^E_{\mu \nu}(x)$, with $\Omega\equiv(\frac{\Phi}{\Phi_0})^{2}$, the action (\[scalartensor\]) transforms into the Einstein frame, as $$\label{steski}
^E S_{S-T}=\int d^3 x\sqrt{-g^E}
\Phi_0^{2}\Big(R^E-\frac{\nu}{2}\Phi_0^{4}\Big),$$ in which $\Phi_0$ is a constant and introduced in order to keep $\Omega$ dimensionless. Therefore, the conformally-coupled scalar field simply disappears and one is left with pure cosmological Einstein theory which has a massless spin-2 particle in its spectrum. How does one see this result in the Jordan frame (which we need for our main problem). We take (\[scalartensor\]) and expand up to quadratic order in the scalar and tensor fields about the (A)dS vacuum to get $$\begin{aligned}
S_{S-T}=\int d^3x\sqrt{-\bar{g}}\bigg \{
&6m\Lambda-\frac{\nu}{2}m^3+\tau\Big[(3m\Lambda-\frac{\nu}{4})h+(12\sqrt{m}-3\nu
m^{5/2})\Phi^L+m\,R^L\Big]\\
&+\tau^2\Big[(-\frac{1}{2}m\Lambda+\frac{\nu}{8}m^3)h_{\mu\nu}^2-\frac{1}{2}m
h^{\mu\nu}{\cal
G}_{\mu\nu}^L+(\frac{1}{4}m\Lambda-\frac{\nu}{16}m^3)h^2 \\
&+2\sqrt{m}R^L\,\Phi^L
+(6\Lambda-\frac{15}{2}\nu m^2)\Phi_L^2+8(\partial_\mu
\Phi_L)(\partial^\mu\Phi_L)\Big]\bigg \}.
\end{aligned}$$ Again $ {\cal O}(\tau^0) $ part is not relevant. $ {\cal O}(\tau^1)$ part gives the vacuum of the theory, inserting the value of $R_L$ (Appendix) in the linear part and dropping the boundary terms, one obtains $$\Lambda=\frac{\nu m^2}{4}.$$ Using this value in the quadratic part results in $$S^{(2)}_{S-T}= \int d^3x\sqrt{-\bar{g}}\bigg \{-\frac{1}{2}m h^{\mu\nu}{\cal G}^L_{\mu\nu}+2\sqrt{m}R_L\Phi_L-24\Lambda\Phi_L^2
+8(\partial_\mu\Phi_L)^2 \bigg \}.
\label{quadratic}$$ By the redefinition of tensor field as follows $$h_{\mu\nu} \equiv\widetilde{h}_{\mu\nu}-\frac{4}{\sqrt{m}}
\bar{g}_{\mu\nu}\Phi_L,
\label{rede}$$ (\[quadratic\]) reduces to the linearized version of the cosmological Einstein theory $$S^{(2)}_{S-T}= -\frac{1}{2}m \int d^3x\sqrt{-\bar{g}}\,\widetilde{h}^{\mu\nu}\widetilde{{\cal
G}}^L_{\mu\nu}.
\label{ste}$$ As it is clear, just like in the Einstein frame, here at the quadratic level of the Jordan frame, the conformally-coupled scalar field with the wrong-sign kinetic energy disappears in (\[ste\]). We will use a similar field redefinition in (\[linearform\]).
Weyl-invariant gauge-fixing condition
-------------------------------------
Before we can identify the fundamental degrees of freedom, there is one more issue that we must discuss: The gauge field in its locally Lorentz invariant form, has spurious (non-propagating) degrees of freedom, which we must eliminate. This can be done with a Weyl-invariant gauge-fixing. Such a gauge condition can be found as follows: Let the gauge-covariant derivative act on the gauge field as [@DengizTekin] $$\mathcal{D}_\mu A_\nu\equiv\nabla_\mu A_\nu+A_\mu A_\nu.$$ Under the transformations (\[gaugetran\]), in $n$ dimensions, it is easy to show that the divergence transforms as $$(\mathcal{D}_\mu
A^\mu)'=e^{-2\zeta}\Big(\mathcal{D}_\mu
A^\mu-\mathcal{D}_\mu
\partial^\mu\zeta+(n-3)(A^\alpha\partial_\alpha\zeta-\partial_\alpha\zeta\partial^\alpha \zeta)\Big),$$ so in 3 dimensions by setting $\mathcal{D}_\mu
\partial^\mu\zeta=0$, we have $$(\mathcal{D}_\mu A^\mu)'=e^{-2\zeta}(\mathcal{D}_\mu A^\mu).$$ Therefore, we can choose a Lorenz-like condition $$\mathcal{D}_\mu A^\mu=\nabla\cdot A+A^2=0, \label{gaugefixing}$$ as a Weyl-invariant gauge-fixing condition. It is important to note that $\mathcal{D}_\mu \partial^\mu\zeta=0$ is also Weyl-invariant. \[This is a Weyl-invariant generalization of the leftover gauge-invariance, $ \partial^2 \zeta=0 $, after the usual Lorenz gauge $\partial_\mu A^\mu =0$ is chosen.\] At the linear level, (\[gaugefixing\]) reduces to the background covariant Lorenz condition: $\bar{\nabla}\cdot A_L=0$.
These tools are sufficient to decouple the fundamental degrees of freedom in (\[linearform\]) which we do in the next section.
Particle spectrum and their masses
==================================
The Weyl-invariant gauge-fixing term (\[gaugefixing\]) at the linear level eliminates the cross term between the gauge field and scalar field in (\[linearform\]). On the other hand, redefinition (\[rede\]) works well in decoupling scalar and tensor fields, so at the end (\[linearform\]) becomes $$\begin{aligned}
\widetilde{S}_{WNMG}=\int d^3x\sqrt{-\bar{g}}\bigg
\{&-\frac{1}{2}\Big(16\sigma+\frac{8\Lambda}{m^2}+1\Big)(\partial_\mu
\Phi^L)^2 \\
&-\frac{1}{4m}(F^L_{\mu\nu})^2 -\Big(2\sigma
m+\frac{\Lambda}{m}+\frac{m}{8}\Big)(A^L_\mu)^2 \\
&-\Big(\frac{\sigma m}{2}-\frac{\Lambda}{4m} \Big
)\widetilde{h}^{\mu \nu}
{\cal \widetilde{G}}^{L}_{\mu \nu}
+\frac{1}{m}({\cal\widetilde{G}}^L_{\mu\nu})^2-\frac{1}{8m}
\widetilde{R}^2_L
\bigg \},
\label{WeylNMGde}
\end{aligned}$$ where we have used the following relations that arise after the field redefinition of (\[rede\]) $$\begin{aligned}
(R_{\mu\nu})_L=&(\widetilde{R}_{\mu\nu})_L+\frac{2}{\sqrt{m}}(\bar{\nabla}_\mu\partial_\nu\Phi_L+\bar{g}_{\mu\nu}\bar{\Box}\Phi_L),
\, \, \, \, \, \, \, R_L=\widetilde{R}_L+\frac{8}{\sqrt{m}}(\bar{\Box}\Phi_L+3\Lambda\Phi_L),\\
{\cal G}_{\mu\nu}^L=&\widetilde{{\cal
G}}^L_{\mu\nu}+\frac{2}{\sqrt{m}}\Big(\bar{\nabla}_\mu\partial_\nu\Phi_L-\bar{g}_{\mu\nu}\bar{\Box}\Phi_L-2\Lambda
\bar{g}_{\mu\nu}\Phi_L\Big),\\
h^{\mu\nu}{\cal G}^L_{\mu\nu}=&\widetilde{h}^{\mu\nu}{\cal
\widetilde{G}}^L_{\mu\nu}+\frac{4}{\sqrt{m}}\widetilde{R}_L\Phi_L+\frac{16}{m}\Phi_L\bar{\Box}\Phi_L+\frac{48}{m}\Lambda\Phi_L^2,\\
({\cal
G}_{\mu\nu}^L)^2=&({\cal{\widetilde{G}}}^L_{\mu\nu})^2+\frac{8}{m}(\bar{\Box}\Phi_L)^2+\frac{40}{m}\Lambda\Phi_L\bar{\Box}\Phi_L
+\frac{48}{m}\Lambda^2\Phi_L^2+\frac{2}{\sqrt{m}}\widetilde{R}_L\bar{\Box}\Phi_L+\frac{4}{\sqrt{m}}\Lambda\widetilde{R}_L\Phi_L.
\end{aligned}$$ The expression (\[WeylNMGde\]) is what we were looking for to identify the fundamental excitations and their masses. The first line shows that we have a unitary massless scalar field as long as we have a non-ghost kinetic term which is guaranteed by $$16\sigma+\frac{8\Lambda}{m^2}+1 \ge 0.
\label{constraint_scalar}$$ In fact, when the bound is saturated the scalar field ceases to be dynamical. The second line is the action for a massive spin-1 field (a Proca field) which propagates 2 unitary degrees of freedom in 2+1 dimensions with mass-square $$M_A^2= (4\sigma +\frac{1}{4})m^2+2\Lambda \ge 0,$$ which is exactly equal to the constraint on the kinetic energy of the scalar field (\[constraint\_scalar\]). The third line needs a little more explanation, since the fundamental degrees of freedom are not transparent by a cursory look. But, the action is exactly what one gets from the linearization of the NMG (\[nmg\]) around its (A)dS vacuum \[albeit with fixed ratios of dimensionful parameters\]. There are two ways to find that it describes a massive spin-2 field. The first way is to show with the help of an auxiliary field that the action reduces to the massive Fierz-Pauli spin-2 theory [@BHT]. The second way is to explicitly decompose the $h_{\mu \nu}$ into its irreducible components and at the end express all the fundamental degrees of freedom as scalar fields [@canonical]. These two pictures yield obviously the same result showing that the third line of (\[WeylNMGde\]) describes a massive spin-2 field with mass-square $$M_g^2 = - \sigma m^2 + \frac{\Lambda}{2}.
\label{bf_bound}$$ Unitarity of the massive spin-2 theory depends whether one is dealing with an AdS ($\Lambda < 0$ ) or a dS ($\Lambda > 0$) background. For the AdS background, Breintenlohner-Freedman (BF) bound [@bf; @waldron] $ M_g^2 \ge \Lambda$ must be satisfied, on the other hand for the dS background, Higuchi [@higuchi] bound $ M_g^2 \ge \Lambda >0$ must be satisfied. What we have not yet shown is that all the unitarity conditions on the spin-0, spin-1 and the spin-2 fields are compatible with each other and with the condition that the theory has a maximally symmetric vacuum (\[lambda\_exp\]). First of all one should notice that $\Lambda_{+}$ corresponds to the dS and $\Lambda_{-}$ corresponds to the AdS spaces. It is easy to see that (\[bf\_bound\]) is not compatible with the dS branch, therefore the theory is not unitary in dS. On the other hand, considering all the conditions together, one finds that the theory is unitary in AdS with a massless spin-0, a massive spin-1 and a massive spin-2 field as long as[^4] $$\begin{aligned}
-&\frac{1}{16} < \sigma \le 0, \hskip 1.5 cm 0< \nu \le \frac{1}{64} ( 1- 256 \sigma^2) , \\
&0 < \sigma \le \frac{1}{16} , \hskip 1.5 cm 0 \le \nu \le \frac{1}{64} ( 1- 256 \sigma^2).
\end{aligned}$$ On the other hand, for AdS the theory has a massless spin-1 field, a massive spin-2 field (with $ M_g^2 = - \frac{ 3
m^2}{16}$) (no scalar field) for $$\sigma=\frac{1}{16}, \hskip 1 cm \nu=0, \hskip 1 cm \Lambda_{-}= -\frac{m^2}{4}.$$ For flat vacuum the theory becomes unitary when $$-\frac{1}{16} \le \sigma \le 0, \hskip 1 cm \nu=0,$$ for which generically the theory has a massless spin-0, massive spin-1 and massive spin-2 fields. There are two special points: for $ \sigma = -\frac{1}{16}$, there is no scalar field, there is a massless gauge field and a massive spin-2 field with mass $M_g = \frac{m}{4}$. For $\sigma = 0$, there is a massless spin-0, a massless spin-2 and a massive spin-1 field with $M_A = \frac{m}{2}$.
Conclusions {#conclusions .unnumbered}
===========
By computing the action up to second order in all directions in the space of gauge and scalar fields and the metric, we have shown that the Weyl-invariant extension of NMG is a unitary theory generically describing a massive spin-2, a massive (or massless) spin-1 and a massless spin-0 field around its AdS and flat vacua. The mere existence of an AdS background spontaneously breaks the conformal symmetry and provides mass to the spin-1 and spin-2 fields in analogy with the Higgs-mechanism. Breaking of the conformal symmetry also fixes all the relevant couplings between the fields. In flat space, dimensionful parameter (that is the expectation value of the scalar field) comes from dimensional transmutation in the quantum theory and the conformal symmetry is broken at the two loop level [@tantekin] via the Coleman-Weinberg mechanism. Weyl-invariant version of NMG seems to be the only known toy model where a graviton mass is generated by the breaking of a symmetry in such a way that the resultant mass has a non-linear, fully covariant, local extension in terms of quadratic curvature terms. In a separate work [@tanhayi], we study the particle spectrum of similar Weyl-invariant quadratic theories in generic $n$-dimensions whose actions were introduced in [@DengizTekin]. It would also be interesting to add conformally coupled spin fields to these models.
\[ackno\] Acknowledgments
=========================
The work of B.T. is supported by the TUBITAK Grant No. 110T339. S.D. is supported by TUBITAK Grant No. 109T748.
Appendix A: Quadratic expressions of the curvature terms {#appendix-a-quadratic-expressions-of-the-curvature-terms .unnumbered}
========================================================
This part compiles all the relevant tensors expanded up to second order around a background ($\bar{g}_{\mu \nu}$). We take the expressions directly from [@ubinmg] for generic $n$-dimensions. The metric perturbation $ h_{\mu \nu} $ is defined as $$g_{\mu \nu} \equiv \bar{g}_{\mu \nu}+\tau h_{\mu \nu}, \hskip 1 cm g^{\mu \nu}=\bar{g}^{\mu \nu}-\tau h^{\mu \nu}+\tau^2 h^{\mu \rho} h^\nu_\rho+ {\cal O}(\tau^3).$$ The Christoffel connection can be expanded as $$\Gamma^\rho_{\mu \nu}=\bar{\Gamma}{^\rho_{\mu \nu}}+ \tau \Big (\Gamma^\rho_{\mu \nu} \Big)_{L}
-\tau^2 h^\rho_\beta \Big (\Gamma^\beta_{\mu \nu} \Big)_{L} +{\cal
O}(\tau^3),$$ from which follows the expansions of the Riemann and all the related tensors. We just need the following expressions: $$\Big (\Gamma^\rho_{\mu \nu} \Big)_{L}=\frac{1}{2}\bar{g}^{\rho \lambda} \Big (\bar{\nabla}_\mu h_{\nu \lambda}+ \bar{\nabla}_\nu
h_{\mu \lambda}-\bar{\nabla}_\lambda h_{\mu \nu} \Big ),$$ $$\sqrt{-g} =\sqrt{-\bar{g}} \bigg[1+\frac{\tau}{2}h+\frac{\tau^2}{8} \Big (h^2-2 h^2_{\mu \nu} \Big)+{\cal O}(\tau^3) \bigg ],$$ where $ h=\bar{g}^{\mu\nu} h_{\mu \nu} $. $$\begin{aligned}
R{^\mu}{_{\nu \rho \sigma}}=& \bar{R}{^\mu}{_{\nu \rho \sigma}}+ \tau \Big (R{^\mu}{_{\nu \rho \sigma}} \Big)_{L}
-\tau^2 h^\mu_\beta \Big ( R{^\beta}{_{\nu \rho \sigma}} \Big )_{L} \\
&-\tau^2 \bar{g}^{\mu \alpha} \bar{g}_{\beta \gamma} \bigg [\Big
(\Gamma^\gamma_{\rho \alpha} \Big)_{L} \Big(\Gamma^\beta_{\sigma
\nu} \Big)_{L}- \Big (\Gamma^\gamma_{\sigma \alpha} \Big)_{L}
\Big(\Gamma^\beta_{\rho \nu} \Big)_{L} \bigg ]+{\cal O}(\tau^3),
\end{aligned}$$ where the linearized Riemann tensor is defined as follows $$\Big ( R{^\mu}{_{\nu \rho \sigma}} \Big )_{L}= \frac{1}{2} \Big
(\bar{\nabla}_\rho \bar{\nabla}_\sigma h^\mu_\nu+\bar{\nabla}_\rho
\bar{\nabla}_\nu h^\mu_\sigma-\bar{\nabla}_\rho \bar{\nabla}^\mu
h_{\sigma \nu}-\bar{\nabla}_\sigma \bar{\nabla}_\rho
h^\mu_\nu-\bar{\nabla}_\sigma \bar{\nabla}_\nu
h^\mu_\rho+\bar{\nabla}_\sigma \bar{\nabla}^\mu h_{\rho \nu} \Big
). \label{riem}$$ The quadratic expansion of the Ricci tensor follows as $$\begin{aligned}
R_{\nu \sigma}=& \bar{R}_{\nu \sigma}+\tau \Big (R_{\nu \sigma} \Big)_{L}-\tau^2 h^\mu_\beta \Big(R^\beta{_{\nu \mu \sigma}} \Big)_{L} \\
&- \tau^2 \bar{g}^{\mu \alpha} \bar{g}_{\beta \gamma} \bigg [\Big
(\Gamma^\gamma_{\mu \alpha} \Big)_{L} \Big(\Gamma^\beta_{\sigma
\nu} \Big)_{L}- \Big (\Gamma^\gamma_{\sigma \alpha} \Big)_{L}
\Big(\Gamma^\beta_{\mu \nu} \Big)_{L} \bigg ]+{\cal O}(\tau^3),
\label{ricc}
\end{aligned}$$ where the linearized Ricci tensor is $$R^{L}_{\nu \sigma}=\frac{1}{2} \Big (\bar{\nabla}_\mu \bar{\nabla}_\sigma h^\mu_\nu+\bar{\nabla}_\mu \bar{\nabla}_\nu
h^\mu_\sigma- \bar{\Box}h_{\sigma \nu}-\bar{\nabla}_\sigma \bar{\nabla}_\nu h \Big).$$ The quadratic expansion of the curvature scalar is $$\begin{aligned}
R=\bar{R}+\tau R_{L}+\tau^2 \bigg \{& \bar{R}^{\rho \lambda}h_{\alpha \rho}h^\alpha_\lambda-h^{\nu \sigma}
\Big(R_{\nu \sigma} \Big)_{L}-\bar{g}^{\nu \sigma} h^\mu_\beta \Big(R^\beta{_{\nu \mu \sigma}} \Big)_{L} \\
& -\bar{g}^{\nu \sigma} \bar{g}^{\mu \alpha} \bar{g}_{\beta
\gamma} \bigg [\Big (\Gamma^\gamma_{\mu \alpha} \Big)_{L}
\Big(\Gamma^\beta_{\sigma \nu} \Big)_{L}- \Big
(\Gamma^\gamma_{\sigma \alpha} \Big)_{L} \Big(\Gamma^\beta_{\mu
\nu} \Big)_{L} \bigg ] \bigg \}+{\cal O}(\tau^3),
\end{aligned}$$ where $$\label{rl}
R_{L}=\bar{g}^{\alpha \beta} R^{L}_{\alpha \beta}-\bar{R}^{\alpha \beta} h_{\alpha
\beta}.$$ The linear form of the Einstein tensor that we frequently used in the text is $${\cal
G}_{\mu\nu}^L=(R_{\mu\nu})^L-\frac{1}{2}\bar{g}_{\mu\nu}R^L-\frac{2\Lambda}{n-2}
h_{\mu\nu}.$$
Appendix B: Weyl-invariant action in the Einstein frame {#appendix-b-weyl-invariant-action-in-the-einstein-frame .unnumbered}
=======================================================
Here we will transform Weyl-invariant new massive gravity (\[winmg\]), which is necessarily in the Jordan frame to the Einstein frame. In what follows we will keep some of the computations in $n$-dimensions for the sake of generality and set $n=3$ later. $g_{\mu \nu}(x)$ denotes the Jordan frame metric and $g^E_{\mu \nu}(x)$ denotes the Einstein frame metric which are related as $$g_{\mu \nu}(x)=\Omega^{-2}(x)g^E_{\mu \nu}(x), \hskip 1.5 cm \sqrt{-g}=\Omega^{-n}\sqrt{-g^E}.$$ The Riemann and the Ricci tensors and the curvature scalars in the two frames are related to each other, respectively as follows $$\begin{aligned}
{R}^\mu{_{\nu\rho\sigma}} [g]&= (R^\mu{_{\nu\rho\sigma}})^E-2
\delta^\mu{_{[\sigma}} \nabla_{\rho]} \partial_\nu\ln\Omega
-2 g_{\nu[\rho}\nabla_{\sigma]} \partial^\mu\ln\Omega \\
& \quad -2 \partial_{[\sigma}\ln\Omega \,\, \delta^\mu_{\rho]}
\partial_\nu\ln\Omega +2 g_{\nu[\sigma} \partial_{\rho]}\ln\Omega \,\,\partial^\mu\ln\Omega +2 g_{\nu[\rho}
\delta^\mu_{\sigma]} (\partial_\lambda\ln\Omega)^2.
\end{aligned}$$ $$\begin{aligned}
{R}_{\nu\sigma}[g]&=(R_{\nu\sigma})^E+(n-2)\Big [\nabla_\sigma
\partial_\nu\ln\Omega + \partial_\nu\ln\Omega\,\,
\partial_\sigma\ln\Omega -g^E_{\nu\sigma}(\partial_\lambda\ln\Omega)^2 \Big ] \\
& \quad +g^E_{\nu\sigma}\Box\ln\Omega.
\end{aligned}$$ $${R}[g]=\Omega^2\Big(R^E+2(n-1)\Box\ln\Omega-(n-1)(n-2)
(\partial_\lambda\ln\Omega)^2\Big).$$ From these we can get the invariant relevant for quadratic gravity as : $$\begin{aligned}
(R_{\mu\nu\rho\sigma})^2 [g]&= \Omega^4\bigg
\{(R^E_{\mu\nu\rho\sigma})^2+8(R^E_{\mu\nu})\nabla^\mu\partial^\nu\ln\Omega+8(R^E_{\mu\nu})\partial^\mu\ln\Omega\,\,\partial^\nu\ln\Omega-
4R^E(\partial_\mu\ln\Omega)^2\\
&+4(n-2)(\nabla_\mu\partial_\nu\ln\Omega)^2+4(\Box\ln\Omega)^2-8(n-2)\Box\ln\Omega(\partial_\mu\ln\Omega)^2\\
&+8(n-2)(\partial_\mu\ln\Omega)(\partial_\nu\ln\Omega)\nabla^\mu\partial^\nu\ln\Omega
+2(n-1)(n-2)(\partial_\mu\ln\Omega)^4\bigg \},
\end{aligned}$$ $$\begin{aligned}
(R_{\mu\nu})^2[g]&=\\
& \Omega^4\bigg
\{(R^E_{\mu\nu})^2+2(n-2)(R^E_{\mu\nu})\nabla^\nu\partial^\mu\ln\Omega+2R^E\Box\ln\Omega+2(n-2)R^E_{\mu\nu}\partial^\mu\ln\Omega\,\partial^\nu\ln\Omega\\
&-2(n-2)R^E(\partial_\mu\ln\Omega)^2+(n-2)^2(\nabla_\nu\partial_\mu\ln\Omega)^2+(3n-4)(\Box\ln\Omega)^2\\
& +2(n-2)^2(\partial_\mu\ln\Omega)(\partial_\nu\ln\Omega)\nabla^\mu\partial^\nu\ln\Omega-(4n-6)(n-2)\Box\ln\Omega(\partial_\mu\ln\Omega)^2\\
&+
(n-2)^2(n-1)(\partial_\mu\ln\Omega)^4\bigg
\},
\end{aligned}$$ $$\begin{aligned}
R^2[g]&=\\
& \Omega^4\bigg
\{(R^E)^2+4(n-1)R^E\Box\ln\Omega-2(n-1)(n-2)R^E(\partial_\mu\ln\Omega)^2+4(n-1)^2(\Box\ln\Omega)^2\\
&-4(n-1)^2(n-2)\Box\ln\Omega(\partial_\mu\ln\Omega)^2+(n-1)^2(n-2)^2(\partial_\mu\ln\Omega)^4\bigg \}.
\end{aligned}$$ As noted in the text above equation (\[steski\]) one should choose $\Omega\equiv(\frac{\Phi}{\Phi_0})^{2}$, where $\Phi_0$ is a dimensionful constant which keeps the Einstein-frame metric dimensionless. Note that the mere requirement that there is a transformation between the Jordan and the Einstein frame introduces a dimensionful constant and breaks the scaling symmetry. This symmetry breaking is not the spontaneous symmetry breaking in the vacuum that we discussed in the bulk of the paper. With the help of the above formulas, we can now write the Einstein frame version of the Weyl-invariant quadratic theory (\[winmg\]) (not to clutter the notation, below we will drop the superscript $E$ ) $$\begin{aligned}
\tilde{S}_{NMG}=\int d^3 x \sqrt{-g} \bigg \{&\sigma\Phi_0^2 \Big[R-8(\partial_\mu \ln \Phi)^2+8 A^\mu \partial_\mu \ln \Phi-2 A^2 \Big] \\
&+\Phi_0^{-2}\Big[R^2_{\mu \nu}-\frac{3}{8}R^2+4R_{\mu \nu}\nabla^\mu \partial^\nu \ln \Phi+8R_{\mu \nu}\partial^\mu \ln \Phi \partial^\nu \ln \Phi-2 R \Box \ln \Phi \\
&-2R (\nabla_\alpha \ln \Phi)^2+4(\nabla_\mu \partial_\nu \ln \Phi)^2+8(\nabla_\alpha \ln \Phi)^4-4(\Box \ln \Phi)^2 \\
&+16(\nabla_\mu \partial_\nu \ln \Phi)\partial^\mu \ln \Phi \partial^\nu \ln \Phi -8R_{\mu \nu}A^\mu \partial^\nu \ln \Phi \\
&+2R A^\mu \partial_\mu \ln \Phi-\frac{1}{2} R A^2+2R_{\mu \nu}A^\mu A^\nu+4 \Box \ln \Phi \nabla\cdot A \\
&-8 \partial_\mu \ln \Phi \partial_\nu \ln \Phi \nabla^\mu A^\nu-16 A^\mu \partial_\mu \ln \Phi (\partial_\nu \ln \Phi)^2 \\
&-4\nabla_\mu \partial_\nu \ln \Phi \Big(\nabla^\mu A^\nu+ 4A^\mu \partial^\nu \ln \Phi \Big)+8(A^\mu \partial_\mu \ln \Phi)^2 \\
&+4 (\nabla_\mu \partial_\nu \ln \Phi)A^\mu A^\nu +4A^2 (\partial_\mu \ln \Phi)^2+(2+\beta) F^2_{\mu \nu}+(\nabla_\mu A_\nu)^2-(\nabla.A)^2 \\
&+4\nabla_\mu A_\nu \Big(A^\mu \partial^\nu \ln \Phi+ A^\nu \partial^\mu \ln \Phi \Big )-2A^\mu A^\nu \nabla_\mu A_\nu \\
&-4A^2 A^\mu \partial_\mu \ln \Phi+\frac{1}{2}A^4\Big]
-\frac{1}{2}\Phi_0^2\Big[(\partial_\mu\ln\Phi)^2+\frac{1}{4}A^2
-A^\mu\partial_\mu\ln\Phi+\nu\Phi_0^4\Big] \bigg \}.
\end{aligned}
\label{ewnmg}$$ To simplify[^5] this action let us define $\Phi = \Phi_0 e^\varphi $ and $D_\mu \varphi \equiv \partial_\mu \varphi - \frac{1}{2} A_\mu$ which is actually gauge invariant since under the gauge transformations (\[gaugetran\]) $\varphi \rightarrow \varphi - \frac{1}{2} \zeta(x) $. Then (\[ewnmg\]) becomes
$$\begin{aligned}
\tilde{S}_{NMG}=&\int d^3 x \sqrt{-g} \bigg \{\Phi_0^2 \Big[ \sigma R-(8\sigma + \frac{1}{2})(D_\mu \varphi)^2 - \frac{1}{2}\nu\Phi_0^4
+(\frac{5}{2}+\beta)\Phi_0^{- 4} F^2_{\mu \nu} \Big] \\
&+\Phi_0^{-2}\Big[R^2_{\mu \nu}-\frac{3}{8}R^2+4R_{\mu \nu}D^\mu \varphi D^\nu \varphi
-2R (D_\mu \varphi)^2+8(D_\alpha \varphi)^4 \\
&+16(\nabla_\mu D_\nu \varphi)\partial^\mu \varphi \partial^\nu \varphi +8 D_\mu \varphi \partial^\mu \varphi \nabla \cdot A
-2 A^2 \nabla_\mu D_\mu \varphi \Big] \bigg \}.
\end{aligned}$$
From this action, one can find the vacuum and study the excitations about it but this route, as we we noted in the text, is rather tedious compared to the Jordan frame action that we worked with. More importantly, the Einstein-frame action is not scale invariant and hence the idea, put forward above and in [@DengizTekin], that graviton becomes massive after the scale symmetry gets broken spontaneously (or radiatively) needs to be reinterpreted.
[0]{}
K. S. Stelle, Phys. Rev. D **16**, 953 (1977); Gen. Rel. Grav. **9**, 353 (1978).
E. A. Bergshoeff, O. Hohm and P. K. Townsend, Phys. Rev. Lett. **102**, 201301 (2009); Phys. Rev. D [**79**]{}, 124042 (2009).
I. Gullu and B. Tekin, Phys. Rev. D **80**, 064033 (2009).
S. Deser, Phys. Rev. Lett. **103**, 101302 (2009).
M. Nakasone and I. Oda, Prog. Theor. Phys. **121**, 1389 (2009).
Y. Liu and Y. W. Sun, Phys. Rev. D **79**, 126001 (2009).
I. Gullu, T. C. Sisman and B. Tekin, Phys. Rev. D **81**, 104017 (2010).
I. Gullu, T. C. Sisman and B. Tekin, Phys. Rev. D **83**, 024033 (2011).
H. Ahmedov and A. N. Aliev, Phys. Rev. Lett. **106**, 021301 (2011).
P. Horava, Phys. Rev. D **79**, 084008 (2009).
S. Deser, R. Jackiw and S. Templeton, Phys. Rev. Lett. **48**, 975 (1982); Annals Phys. **140**, 372 (1982).
S. Dengiz and B. Tekin, Phys. Rev. D **84**, 024033 (2011).
P. N. Tan, B. Tekin and Y. Hosotani, Phys. Lett. B **388**, 611 (1996); Nucl. Phys. B **502**, 483 (1997).
S. R. Coleman and E. J. Weinberg, Phys. Rev. D **7**, 1888 (1973).
I. Gullu, T. C. Sisman and B. Tekin, Class. Quant. Grav. **27**, 162001 (2010).
I. Gullu, T. C. Sisman and B. Tekin, Phys. Rev. D **82**, 124023 (2010).
P. Breitenlohner and D. Z. Freedman, Phys. Lett. B [**115**]{}, 197 (1982).
A. R. Gover, A. Shaukat and A. Waldron, Nucl. Phys. B [**812**]{}, 424 (2009).
A. Higuchi, Nucl. Phys. B [**282**]{}, 397 (1987).
M. R. Tanhayi, S. Dengiz and B. Tekin, “Weyl-Invariant Higher Curvature Gravity Theories in $ n$ Dimensions,” arXiv:1201.5068 \[hep-th\].
[^1]: To have a maximally symmetric vacuum one must have $\lambda$ > -1 and one can normalize $\sigma^2=1$ in NMG. On the other hand, $\lambda =0$ and $\sigma$ should be free in the Weyl-invariant version, since the numerical values of various couplings play a key role in the unitarity analysis.
[^2]: Parity non-invariant massive spin-2 theory with a single helicity degree of freedom, that is the Topologically Massive Gravity, was found in 1982 [@DJT].
[^3]: Here we take the point of view that $\Phi_{vac}$ is given and $\Lambda$ is determined, one can also take a different point of view that $\Lambda$ is given and $\Phi_{vac}$ is determined. See [@DengizTekin] for a discussion on this.
[^4]: Note that negative $\nu$ is not allowed for the scalar field to have a viable potential with a lower bound. Moreover, for $\nu =0$, the conformal symmetry cannot be broken since Coleman-Weinberg potential at any loop would vanish in the case of flat space, but in what follows below, we include this point to explore the full parameter range.
[^5]: We thank a very conscientious referee who not only suggested this simplification but also actully computed a different version of it and offered several useful remarks on the rest of the paper.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recent exam of large samples of [$\omega$ Cen]{} giants shows that it shares with mono-metallic globular clusters the presence of the sodium versus oxygen anticorrelation, within each subset of stars with iron content in the range –1.9$\simlt$\[Fe/H\]$\simlt$–1.3. These findings suggest that, while the second generation formation history in [$\omega$ Cen]{} is more complex than that of mono-metallic clusters, it shares some key steps with those simpler cluster. In addition, the giants in the range –1.3$<$\[Fe/H\]$\simlt$–0.7 show a [*direct*]{} O–Na correlation, at moderately low O, but Na up to 20 times solar. These peculiar Na abundances are not shared by stars in other environments often assumed to undergo a similar chemical evolution, such as in the field of the Sagittarius dwarf galaxy. These O and Na abundances match well the yields of the massive asymptotic giant branch stars (AGB) in the same range of metallicity, suggesting that the stars at \[Fe/H\]$>$–1.3 in [$\omega$ Cen]{} are likely to have formed directly [*from the pure ejecta of massive AGBs of the same metallicities*]{}. This is possible if the massive AGBs of \[Fe/H\]$>$–1.3 in the progenitor system evolve when all the pristine gas surrounding the cluster has been exhausted by the previous star formation events, or the proto–cluster interaction with the Galaxy caused the loss of a significant fraction of its mass, or of its dark matter halo, and the supernova ejecta have been able to clear the gas out of the system. The absence of dilution in the metal richer populations lends further support to a scenario of the formation of second generation stars in cooling flows from massive AGB progenitors. We suggest that the entire formation of [$\omega$ Cen]{} took place in a few 10$^8$yr, and discuss the problem of a prompt formation of s–process elements.'
author:
- 'F. D’Antona'
- 'A. D’Ercole'
- 'A.F. Marino'
- 'A. P. Milone'
- 'P. Ventura'
- 'E. Vesperini'
title: 'The oxygen vs. sodium (anti)correlation(s) in [$\omega$ Cen]{}'
---
Introduction {#sec:intro}
============
A number of surprising observational findings of the latest dozen years suggest a completely new paradigm for the formation of Globular Cluster (GC) stars. The abundances of stars in most Galactic GCs observed so far show particular patterns, such as the oxygen vs. sodium and magnesium vs. aluminum anticorrelations, not shown by the halo stars [@gratton-ar], both in giants and in scarcely evolved stars [e.g. @gratton2001; @ramirezcohen2002; @gratton-ar; @carretta2009a; @carretta2009b]. The observed amount of chemically anomalous stars and the relative number of normal and anomalous stars suggest a sequence of events starting with the birth of a first stellar generation (FG) in a “protocluster" much more massive than todays’ cluster, followed by the birth of a second generation (SG) from gas having the same iron content, but including hot–CNO cycled matter ejected by the FG stars evolving during the first phases of the cluster life [for a summary, see, e.g. @dercole2008]. Theoretical studies have identified the matter constituting the SG stars as the gaseous ejecta either by massive asymptotic giant branch (AGB) stars or by fast rotating massive stars . These studies investigated whether the observed physical properties of the Galactic GCs are consistently accounted for in the models. Problems are present in either scenario [for a critical discussion, see @dercole2010], although some chemical signatures of the SG, such as the Lithium abundances [@dorazimarino2010; @shen2010] or the magnesium depletion [@vd2009; @vcd2011] favour the AGB scenario, in which the hot CNO processing occurs by “hot bottom burning" (HBB) in the convective envelopes, that are recycled to the intra–cluster medium by wind mass loss. The AGB ejecta are likely to be converted into new stars, thanks both to their low ejection velocities and to the fact that they evolve after the end of the energetic Supernovae II epoch.
Most GCs show none or scarce iron spread among their stars [e.g. @carretta2009ferro]. A few clusters, however, must be born in a more complex “protocluster" environment. It has been often suggested that [$\omega$ Cen]{}, the most massive galactic GC, formed in a now dispersed dwarf galaxy [e.g. @norris1997; @rey2004; @bekki-norris2006] and, possibly, has an evolutionary history similar to that of M54, whose host galaxy, the Sagittarius dwarf, is now merging into the Galaxy. Both [$\omega$ Cen]{} and M 54 show a large spread in iron. [$\omega$ Cen]{} contains stars with iron contents spanning a range from \[Fe/H\]$\sim$–2 to \[Fe/H\]$\sim$–0.6 [@norris1996; @suntzeff1996; @lee1999; @pancino2000; @sollima2005a; @johnson2008; @johnson2010; @marino2011].
The complexity of [$\omega$ Cen]{} star formation episodes is also shown from its multiple red giant and sub giant branches [@lee1999; @pancino2000; @sollima2005b]. [@bedin2004] found that the main sequence (MS) contains a well separated blue branch, probably then hiding a large overabundance of He [@norris2004], as its stars are more iron-rich than stars on the redder side [@piotto2005]. A third, less populated MS (MSa) has been discovered by [@bedin2004] on the red side of the red MS and it has been associated to the most metal rich stars [e.g. @villanova2007]. As well as blue MS stars, the colors and the metallicity of the MSa suggest that it is populated by He rich stars [@bellini2010].
The presence of large variations in the light elements abundances and s–process elements is well documented [see, e.g. @paltoglou1989; @ndc1995; @smith2000]. [@ndc1995] show that [$\omega$ Cen]{} giants also display huge aluminum variations, anticorrelated with oxygen and correlated with sodium, another typical pattern present in mono–metallic GCs. More recently, the spectral analysis of a sample of 840 giants by [@johnson2010] and 300 giants by [@marino2011] shows with more clarity a very interesting peculiarity. Although the presence of a wide range of iron contents points towards a chemical evolution similar to that of galactic environments, the stars in each bin of iron content for \[Fe/H\]$<$–1.3 [*also show the typical O–Na anticorrelation of GCs having a homogeneous iron content*]{}. On the contrary, at \[Fe/H\]$>$–1.3 the O and Na abundances show a direct correlation, and the \[Na/Fe\] values are always much larger than the solar ratio.
The attempts to model the chemical evolution of [$\omega$ Cen]{}, in the framework of a dwarf galaxy environment, have not been fully successful [e.g. @romano2007; @choi2008] especially for the difficulty in modeling the origin of the very helium-rich population(s).
The most recent modelling by [@romano2010] does not yet include the spectroscopic results concerning the O–Na anticorrelation [*for each subset of iron content*]{} we just discussed. In fact, a linear chemical evolution path for each element can not describe this complex situation. Their work has to make a series of complex hypotheses concerning winds that must preferentially eject some products of nuclear evolution, in order to reproduce some characteristics of the cluster stars. In particular, a timescale of a few Gyrs is needed to form the stellar generations with different iron content in this cluster, and this rules out the massive AGB scenario for the formation of the super–helium rich population(s); the formation of this population, according to [@romano2010] should then be attributed to the yields of FRMS, but see, e.g., [@renzini2008] against this possibility. Since our work is focused on the AGB scenario also for the formation of the very helium rich population(s), we will discuss later on (Sect.\[sprocess\]) why this fails in the approach by [@romano2010].
We point out another ingredient of the [$\omega$ Cen]{} puzzle: several studies have shown that the blue MS stars are more strongly concentrated in the cluster core than the red MS stars, and that this higher concentration holds for the most metal rich populations too [@suntzeff1996; @norris1996; @rey2004; @bellini2009; @johnson2010]. These observed differences in the spatial distribution of first- and second-generation stars are consistent with the predictions of the simulations presented in @dercole2008. Specifically, @dercole2008 show that the SG formation in “simple", mono–metallic GCs can occur in a cooling flow that gathers the massive AGB ejecta in the cluster core [see also @bekki2010 for 3D simulations confirming the results of D’Ercole et al. 2008].
[@dercole2010] extended the dynamical model to predict the abundances in the SG and, in particular, how the O–Na and Mg–Al anticorrelations are built up.
The remarkable O–Na–Fe abundance distribution described above, introduces a new complex ingredient in the evolution of [$\omega$ Cen]{} and suggests that a new approach is needed to model the chemical evolution of this cluster. We suggest that, despite the complexity of the abundance patterns in [$\omega$ Cen]{} stars, the formation history of this cluster may be explained according to the key steps of the [@dercole2008] model, that also provides the key to understand how the stars with \[Fe/H\]$>$–1.3 may be directly born from the pure ejecta of massive AGB stars [as already recognized by @johnson2009]. We also show that the cluster formation outlined in this work is in agreement with the abundances of other light elements (Mg and Al). It is also consistent with the spatial distribution of [$\omega$ Cen]{} populations. The critical problem of the formation timescale(s) of the multiple populations constituting the GC as we see it now is finally discussed in Sect.\[sprocess\], where we remark that the abundances of s–process elements as a function of metallicity may be uncorrectly biasing our views.
The p–capture data in [$\omega$ Cen]{} {#dilut}
======================================
While in this work we will focus mainly on the Na and O abundances in [$\omega$ Cen]{}, and will use the data from [@marino2011] for the sake of homogeneity, here we provide some more information on results for Mg and Al, whose abundances constitute another powerful indicator of p–capture processing in the gas from which the SG stars formed. Fig. \[f1ab\] shows the O, Na, Mg and Al abundances as a function of \[Fe/H\] from different data, and the lines corresponding to theoretical models. These latter will be discussed in Sect. \[mgal\]. Mg data are taken from [@ndc1995], and Al data are from [@johnson2008; @johnson2009; @johnson2010]. The figure shows that Na and O data from [@johnson2010] are qualitatively compatible with the [@marino2011] data, although a trend to have lower Na and O values is evident, especially at \[Fe/H\]$>$–1.3. [@marino2011] show in their Fig. 6 that their own oxygen determinations can be larger than the [@johnson2008; @johnson2009] and [@johnson2010] determinations, and the difference may reach 0.3 dex or more at low oxygen values. Similar systematic (although smaller) differences appear also in the sodium determinations. For this reason we decided not to combine the different datasets, and we mainly use our own data to put constraints on the O–Na (anti)correlation(s).
We show the O and Na abundances from the data by [@marino2011] in Fig. \[f2\]. We consider 6 bins of metallicity, keeping in mind that the errors on the abundance determination may easily shift some points from one bin to the adjacent one(s) [@marino2011]. We plot \[O/H\] and \[Na/H\] instead of the values normalized to the iron abundance, scaling all the values to the lowest metallicity \[Fe/H\]=–2. In order to increase the relatively few data at \[Fe/H\]$>$–1, we add the values by [@johnson2010] for these metallicities. These data increase the scatter, but, in this kind of plot, they do not change the qualitative behaviour of the metal rich sub–populations. We point out here some interesting patterns emerging from this figure (see also the right panel of Fig. \[f3\]): 1) the lowest metallicity bin shows a scarce, if any, anticorrelation; 2) [*each*]{} intermediate metallicity bin in the range –1.9 $<$–1.3 shows [*independent*]{} extended O–Na anticorrelations; 3) the highest metallicity bins (apart from a few stars in the range –1.3 $<$ –1.1 located in the O–rich, low–Na region) show very clearly a radically different behaviour, a [*direct*]{} O–Na correlation, with oxygen abundances slightly reduced and very large sodium values, up to $\sim$20 times the solar ratio.
The huge sodium values of the most iron rich bins is a very atypical feature; we are not aware of any other galactic environment in which such a feature is found. In particular, the field stars of the Sagittarius dwarf galaxy having comparable iron contents, have solar, or lower than solar, \[Na/Fe\] [@sbordone2007; @carretta2010sag]. Thus the field of the Sagittarius formed in a very different context than the metal–rich population in [$\omega$ Cen]{}, although other elemental abundances seem to suggest a similar chemical evolution [@carretta2010m54wcen].
In Fig. \[f2\] we also plot the yields of the massive AGB models (see Fig. \[f3\] for the yields down to 3[$\, {M}_\odot$]{}), scaled in the same way as the data. We limit the masses to the range 4.5$\leq$M/[$\, {M}_\odot$]{}$\leq$6.5. Both O and Na increase when the stellar mass decreases (see Fig. \[f3\] for the super–AGB behaviour). In fact, smaller masses have lower temperature at the HBB boundary of convection ([T$_{\rm bce}$]{}), and only allow a milder ON and Na cycling. Increasing the metallicity, the models define parallel and partially superimposed diagonal lines, that shift to larger O and Na values for larger metallicity. This is due to two concomitant reasons.
The first reason is that, by increasing the metal content, the average content of each element increases. The sodium yield is a result of four different processes occurring in the HBB envelopes: 1) second dredge up of the sodium produced in the stellar interior in the previous phases of evolution; 2) burning to sodium of the $^{22}$Ne achieved during the same process of second dredge up; 3) and burning during the HBB phase along the Ne–Na cycle. 4) In addition, in the relatively small masses (M$\le$4.5[$\, {M}_\odot$]{}), sodium is also produced by HBB of the $^{22}$Ne dredged in the envelope at each third dredge up process. As sodium is produced from $^{20}$Ne and $^{22}$Ne, and this latter is ultimately produced from $^{14}$N, the average production of $^{23}$Na, [*for similar HBB conditions*]{} must increase with stellar metallicity. For the same reason, the equilibrium value of O in the full CNO processing is larger at larger C, N and O abundances.
The second reason is that the HBB efficiency decreases with increasing metallicity, as the higher opacities affect the envelope structure and allow smaller [T$_{\rm bce}$]{}’s, lower ON processing and Na depletion [@vd2009]. Notice that the computation of the sodium yields is a delicate issue, as they depend on uncertain cross sections, and also on the initial abundance of $^{20}$Ne. This problem is examined extensively by [@dercole2010], who adopt Na abundances larger by 0.2–0.3dex with respect to those plotted in Fig. \[f3\]. Yields from other modelers [e.g. @karakas2007] attain smaller oxygen depletion and larger sodium abundances for a fixed mass, due to the different assumptions made for the convective efficiency and mass loss. This issue is largely discussed in [@vd2005a; @vd2005b].
The theoretical yields seem to define a border limit to the abundances of the data, and the yields of the massive AGBs of the highest available metallicity Z=$4\times 10^{-3}$ go through the data in the two bins of largest iron content. As the model yields reproduce the observed O–Na correlation for the high-metallicity population, we are led to suggest that the stars at \[Fe/H\]$>$–1.3 [*form from the pure ejecta of massive AGBs of the same metallicity*]{}, and this may represent a key ingredient to understand the key lines of chemical evolution in [$\omega$ Cen]{}. If we had adopted [@johnson2010] data, the same conclusion would have been reached, but only for \[Fe/H\]$>$–1, as their oxygen values for giants having –1.3$<$\[Fe/H\]$<$–1 can reach much lower abundances, likely due to systematics both in \[Fe/H\] and \[O/Fe\] between the two data-sets [see Figs. 4–6 in @marino2011]. We now try to use this finding as a constraint for the formation model of the cluster.
The O–Na plane: data points, theoretical yields and predictions of D’Ercole et al. model {#theory}
========================================================================================
Before we discuss a possible modelling of the O–Na data in [$\omega$ Cen]{}, it is propaedeutic to summarize the status of art of modelling the O–Na anticorrelation in single–metallicity GCs. The left panel of Fig. \[f3\] shows the data by [@carretta2009a] for giants in these clusters. Superimposed, we plot the whole oxygen and sodium yields of AGB stars from [@vd2009]. As discussed in Sect. \[dilut\], masses increase from top right to bottom left. For the Z=10$^{-3}$ models, we also plot the results of the yield computation by @vd2010b for the masses 6.5$\leq$M/[$\, {M}_\odot$]{}$\leq$8, that, after having ignited Carbon in conditions of semi–degeneracy, and formed an O–Ne degenerate core, follow the super–AGB phase. We see that these more massive progenitors do not change the location of the Na–O model correlation, but simply reverse it, and now the O and Na content increase for increasing mass. This is due to the huge mass loss rates of super–AGBs. So by increasing the initial mass from 3[$\, {M}_\odot$]{} to the maximum mass that does not explode as supernova, the Na–O line first goes down, reaches a minimum value in the yields (at masses $\sim$6.5[$\, {M}_\odot$]{}), and then reverts back to larger O and Na.
Three main features are evident from the comparison between yields and single-metallicity globular cluster data:
1. the AGB models of different masses provide a [*direct*]{} correlation between sodium and oxygen abundance in their ejecta. The presence of the O–Na anticorrelation in the GC data is generally attributed to dilution of the hot–CNO processed gas of the stellar ejecta with “pristine" gas [@bekki2007; @dv2007; @dercole2008; @dercole2010]. Dilution is also needed when dealing with FRMS ejecta [@prantzos2006; @decressin2007; @lind2011]. In all clusters dilution seems to be present. Detailed hydrodynamical models able to explain the dynamics and the presence of pristine gas during the SG formation of all clusters are still lacking. This issue is discussed in detail by [@dercole2011].
2. the yields for masses M$\simlt$4.5–5[$\, {M}_\odot$]{}lie outside the plane region occupied by the O–Na observational values. This represents a powerful constraint of the models, showing that the winds of massive AGB, for some reason, can not contribute to form SG stars at the epoch in which these –or smaller– masses begin to evolve ($>$130Myr). A powerful similar constraint is imposed also by the observed constancy of the sum of CNO abundances among the stars of many GCs (e.g. M92 [@pilachowski1988], NGC 288 and NGC 362 [@dickens1991], M3 and M13 [@smith1996], M4 [@ivans1999]) The occurrence of numerous third dredge up events during the AGB evolution of masses M$<$4.5[$\, {M}_\odot$]{} [@vd2009; @dercole2010] increases the total CNO.
3. the yields do not represent well the [*lowest*]{} oxygen values that can be present in the data. [@dercole2010] assumed semi empirical yields down to \[O/Fe\]$=-1$ in the super–AGB masses, by adopting a ‘deep–mixing" suggestion. [@dv2007] attribute the very small oxygen abundance present in the giant stars of the most extreme SG to deep extra–mixing acting in the progenitors having a very high helium content. In these stars, the small molecular weight discontinuity during the red giant branch evolution may not be able to preclude deep mixing and lowering of oxygen in the envelope. Stars having the largest helium abundances would then populate the blue MS and show the strongest chemical anomalies (see later). The presence of \[O/Fe\] abundances smaller than those predicted by AGB models only among the giants, and not among the turnoff or subgiant stars so far examined [@carretta2006], and in stars belonging to clusters also showing the signature of a very helium rich MS, such as NGC 2808 and [$\omega$ Cen]{}, are a possible argument in favour of this interpretation. Obviously, the determination of a very low oxygen content —similar to the one in the most anomalous red giants— in the atmosphere of blue MS stars in these clusters would falsify this suggestion [@bragaglia2010 were not able to measure oxygen in their spectrum of the blue MS in NGC 2808.].
Let us discuss how the O–Na anticorrelation is introduced in [@dercole2008; @dercole2010]. In order to explain the data of NGC 2808, they propose a model in which the “extreme" SG, corresponding to the most oxygen-poor stars and to the stars populating the very helium rich blue MS, are born directly from the pure ejecta of the super–AGBs of the FG. In fact, after the SN II epoch, that has cleared the cluster from the residual gas, a cooling flow sets in and brings to the core the low velocity winds of the super–AGBs, so that stars can form directly from their very helium rich and hot–CNO processed material. Then a phase of mixing of the massive AGB ejecta with pristine gas (reaccreted in the cluster core) follows, giving origin to the “intermediate" SG stars, with milder oxygen depletion. The SG formation stops due to the onset of the Type Ia SN epoch before the pristine gas is exhausted. In the right panel of Fig. \[f3\], the slanted V–shaped line shows an example of the O–Na patterns predicted by the [@dercole2010] model, based on the whole set of yields for masses from 3 to 8[$\, {M}_\odot$]{} shown in the left panel, and extended for 200Myr after the FG formation (D’Ercole et al. 2011, in preparation). Time along the line proceeds from the left side to down right, and then climbs up towards the upper right side. The “extreme" SG is the left side of the V–shaped line, up to the point at $t$=80 Myr. The “intermediate" SG stars follow the line from 80 to 100 Myr. If star formation proceeds further, the subsequent SG would follow the right side of the V, from $t$=100 Myr to the end, when stars form directly from the ejecta of the evolving AGBs, so that the simulation follows the [*direct*]{} correlation O–Na shown in the left panel, but this side of the curve has no observational counterpart in “normal" GCs. The simulation stops at 2$\times 10^8$yr, when the evolving AGB mass is $\sim$3.7[$\, {M}_\odot$]{}. The model is plotted on the \[O/Fe\], \[Na/Fe\] data for [$\omega$ Cen]{} from [@marino2011], and we see that data points are present also on the upper side of the simulation curve, but they have larger metallicity than stars showing the anticorrelation(s). This suggestive comparison requires some deeper understanding.
Interpretation of the [$\omega$ Cen]{} data in terms of D’Ercole et al. model
=============================================================================
We discussed in the previous Sect. \[theory\] how the O–Na anticorrelation in the GC data is attributed to dilution of the hot–CNO processed gas of the stellar ejecta with “pristine" gas. Where does this pristine gas come from? [@dercole2008] suggest that it may survive in a torus that collapses back on the cluster after the SN II epoch. If the system is much more massive than a typical proto–cluster as in the case of [$\omega$ Cen]{} (possibly including a dark matter halo) the SN II ejecta (and type Ia SN ejecta as well) can not be lost outside the potential well of the protocluster. The 3D hydro simulations by [@marcolini2006] show in fact that the collapse back includes the matter enriched by the SN II ejecta. Their model was first developed to be compared with Draco’s dwarf galaxy data. Afterwards, they attempted to model directly the case of [$\omega$ Cen]{} [@marcolini2007]. Star formation in the cluster region stops when SN II explode, and a successive burst occurs when the interstellar medium (enriched by SN ejecta) flows back down the cluster potential well. Thus the gas content in the cluster center has an oscillating temporal profile whose period is given by the time-interval between two successive starbursts. Each starburst provides star formation with increasing metallicity. If we add to this model the possible role of the winds of the AGBs evolving in the progenitor systems, these may account for star formation of the very helium rich stars when the cluster core is still devoid of matter (before each collapse back of the ISM), and for the formation of stars from AGB matter diluted with the ISM enriched by the SN II ejecta, similar to the events in the one–burst only model by [@dercole2010]. Suppose now that, after a last burst of star formation, there is no longer gas that can collapse back from the outskirts of the cluster. Nevertheless, there can be further episodes of star formation thanks to the cooling flows of the AGB winds. Whatever is the detailed mechanism by which the O–Na anticorrelation(s) are built in [$\omega$ Cen]{} for \[Fe/H\]$<$–1.3, the O–Na correlation we see at larger metallicities is telling us that the ISM in the outskirt of the cluster must be so reduced that only the AGB winds take part in the cooling flow and that these winds belong to the massive AGBs of the more metal rich populations. This interpretation only requires that these AGBs are delayed until there is no longer any pristine gas available to collapse back in the core of [$\omega$ Cen]{}. A possible hypothesis is that the gravitational interaction of the [$\omega$ Cen]{} progenitor with the Galaxy [@dinescu1999] caused the loss of a significant fraction of its mass, or of the dark matter halo, so that the last SN II burst, or the type Ia SN explosions could clean completely the central region and the whole proto–cluster from its residual gas.
Following this scenario, we now can re-interpret the path of the simulation shown in Fig. \[f3\] and explain why it resembles the shape of the [$\omega$ Cen]{} data in the Na–O plane. In a mono–metallic cluster, the stars on the upper branch would be those formed by [*smaller mass AGB stars ejecta*]{} (M$\simlt$4.5[$\, {M}_\odot$]{}), evolving at a later time, when the diluting gas is over. In [$\omega$ Cen]{}, the high Na abundances belong to the stars in the most metal rich bins, and correspond to the yields of (again) more massive AGB stars for those metallicities (Fig. \[f2\]). The only requirements to explain why the most iron-rich giants in [$\omega$ Cen]{} have large sodium and moderately low oxygen are: 1) these stars are formed in a cooling flow [*directly*]{} from the ejecta of the massive AGBs having the same metallicities; 2) the AGBs that provide the ejecta evolve at the time when all the gas surrounding the cluster has been already exhausted.
Consistency of the scenario: the magnesium and aluminum abundances {#mgal}
------------------------------------------------------------------
We have so far focused our attention on the Na–O (anti)correlation(s) to outline a possible scenario for the formation of the multiple populations in [$\omega$ Cen]{} because of the wealth of data available for the abundances of these two elements.
In this section, we show that existing data on other light elements lend further support to the proposed scenario. In Fig. \[f1ab\] the top panels show the Mg and Al data as a function of \[Fe/H\]. We plot on the data the [*minimum*]{} Mg yield and [*maximum*]{} Al yields from the [@vd2009] models. These values then represent the maximum degree of Mg burning and Al production in massive AGBs of each metallicity. The Mg data by [@ndc1995] show a remarkable constancy of magnesium for all metal contents, but the few low Mg points are at iron content for which models predict Mg depletion. Quantitatively, the models are a factor $\sim$2 off the data, but see [@vcd2011] for a discussion of the dependence of the Mg depletion in AGB and super–AGBs on the input physics of the models. It is important to point out that [*no Mg depletion*]{} is predicted at the highest metallicities, for which, according to our model, the SG formation occurred directly from the pure AGB ejecta. The upper boundary of Al data is in good agreement with the model yields. In particular, a smaller Al enhancement is predicted at high metallicity than at the lower ones, as observed. The Al data are also consistent with the hypothesis that some dilution of AGB ejecta with pristine gas occurs up to \[Fe/H\]$\simlt$–1.3, while either a dichotomy in the Al values is found at –1.3$\simlt$\[Fe/H\]$\simlt$–1, or values compatible with pure ejecta at \[Fe/H\]$>$–1.
The timescale of formation of the [$\omega$ Cen]{} populations {#sprocess}
==============================================================
We do not attach a fixed timescale to the model presented to account for the high sodium of the metal rich populations in [$\omega$ Cen]{}. Its main requirement is that the time delay for the formation of the massive AGB progenitors in these iron bins must be of the order of the time required for the preceding formation events to exhaust the galactic gas in the neighborhood of the cluster, so that this gas does not collapse back to the cluster core after the last SN II epoch. We think that, in order to satisfy these requirements, the timescale for the formation of all the populations in [$\omega$ Cen]{} must not exceed a few 10$^8$yr.
Notice that the upper boundary for \[O/Fe\] (see the right panel of Fig. \[f3\]) increases with increasing metallicity, in spite of the presence of the many O–poor giants defining the O–Na anticorrelation(s). Standard galactic chemical evolution shows that \[O/Fe\] remains $>0$ until the iron is produced by SN II, and shifts to the solar ratio only when Type Ia SN, producing much more iron and no oxygen, become the main source of metals [@matteuccigreggio1986]. Thus the iron chemical evolution has been dominated by Type II SN, suggesting that the initial chemical evolution inside the progenitor system occurred on a [*fast*]{} timescale. As also the other $\alpha$–elements seem to go towards the solar ratio only for the most metal rich population, corresponding to the MSa and RGBa [@pancino2002; @origlia2003], the timescale for the formation of the last, metal richest, population in [$\omega$ Cen]{} must be in fact very close to the timescale for the onset of SN Ia explosions. Although this timescale is uncertain, its value is observationally estimated to be in the range 40–300 Myr [@mare01; @mapa06; @raskin2009], and certainly not longer than 500Myr [e.g. @galyam2004; @sullivan2006]. Simulations adopting the double–degenerate model for this event show that the peak of the distribution is reached after only 250Myr [@tornambe1989]. It is difficult to see how this epoch can be delayed by more than a few 10$^8$ yr in [$\omega$ Cen]{}. In the [@marcolini2006] and in other similar models, the presence of a dark matter halo prevents the clearing of gas by type Ia SN, so some other process must have precluded further star formation in the proto–[$\omega$ Cen]{} (and in spheroidal dwarf galaxies as well), e.g. the gravitational interaction with the Galaxy.
As the iron production is delayed in time and massive AGBs of different metallicity evolve at different epochs, also less massive (M$<$4.5[$\, {M}_\odot$]{}) AGBs of the lower metallicity bins would contribute to the cooling flow. These may be, e.g., some of the stars in the bin –1.3$\leq$\[Fe/H\]$\leq$–1). To understand the role of smaller AGB masses of lower metallicities, notice that the metal content of the massive AGB of high metallicity dominates over the anomalies of these latter stars (see the discussion of the O–Na yields shown in Fig. \[f2\]). As a net result, the \[O/Fe\] and \[Na/Fe\] of the gas may be not affected too much by the yields of the less massive stars and the latter contribution to the average abundances remains a small perturbation. Of course, a full model is needed to quantify the extent of this problem.
HR diagram dating results in age differences from less than 2 Gyr [e.g. @lee2005; @sollima2005a] to 5 Gyr or more [@villanova2007] in recent literature. The influence of different helium content and/or of the total C+N+O content on the isochrones turnoff level may easily change these age determinations [@dantona2009; @pietrinferni2009] and bring them to much closer values. Joo, S. & Lee, Y.-W. (2011, in preparation) have indeed shown that it is possible that all the populations are coeval, by taking into account the combined effect of helium and metallicity. The result is valid within the errors of $\pm$0.3 Gyr that affect the determination of relative ages from the turnoffs’ location. This requires that the MS-a is indeed helium rich, as suggested also by [@bellini2010].
We address now the most important clue, that we consider a red herring, for the timescale of formation of this peculiar cluster, namely the evidence of a very fast and large growth, by more than a decade, in the abundances of Lanthanum and Barium [@johnson2010; @marino2011] in the very small range of metallicity between \[Fe/H\]$\sim$–1.9 and \[Fe/H\]$\sim$–1.5. This very dramatic s-process increase with iron, at such a low metallicity, is not found in other galactic environments [@tolstoy2009], as shown in Figure 17 of [@johnson2010]. [@marino2011] suggest a rapid production of these elements. The r-process is excluded, as it would also affect Europium as well, while the \[Eu/Fe\] ratio does not increase in the same metallicity interval [@smith2000; @johnson2010]. Another plausible production site are the cores of massive stars [weak s–process, e.g. @raiteri1993 and references therein], but this chain should mainly increase the lighter nuclei in the s–chain, like $^{58}$Fe, $^{63}$Cu, or $^{65}$Cu [see, e.g. @pumo2010], while Cu in the metal rich stars of [$\omega$ Cen]{} remains at low abundances [@cunha2002; @pancino2002]. Further non–standard models for the production of these s–process elements may occur in rotating massive stars [@pignatari2008].
Nevertheless, the best acknowledged production site of these s–process elements are the lower mass AGB stars (M$<$3[$\, {M}_\odot$]{}) [@busso1999] that evolve on timescales of several hundreds of million years or on gigayears timescales, and [*these stars have been always considered the obvious responsible for this s–process production*]{}. Of course, if the –say– 3[$\, {M}_\odot$]{} are responsible for the s–process production, this happens at a time at which all the possible producers of helium and of the O–Na anticorrelations (the stars of M$\simgt$5[$\, {M}_\odot$]{}!) have already evolved. We remark again that the 3[$\, {M}_\odot$]{} stars would produce ingent amounts of sodium, but do not deplete oxygen nor produce helium. It may well be that these masses contribute to increase the s–process abundances when the \[s/Fe\] value flattens (at \[Fe/H\]$\simgt -1.5$), but it is difficult to envision how they can contribute to the extreme s–process increase that occurs at –1.5$>$\[Fe/H\]$>$–1.9, a range of metallicities that show independent, and even extreme, O–Na anticorrelations (see the blue squares and the red triangles of Fig. \[f2\]) and where there is no correlation of the s–process abundances with, e.g., the oxygen depletion. So the problem is that the timescale to build up the O–Na anticorrelation(s) is much shorter than the timescale needed for standard s–process production sites. We do not see an easy way out of this problem, and in fact this is the reason why [@romano2010] had to exclude the AGB contribution for the formation of the helium rich population, and resort to the contribution of massive stars.
We suggest that, in the same way as the O–Na anticorrelation occurs only in GCs, and is due to the peculiar chemical evolution provoked by the collection of winds from a restricted range of initial stellar masses, we may be facing a process that has not yet been explored in stellar models. A not well discussed site of s–nucleosynthesis occurs in the carbon burning shells of the tail of lower mass progenitors of SNII [e.g. @the2007], and their contribution may peculiarly become apparent mainly in the evolution of the progenitor systems of clusters like [$\omega$ Cen]{}, that are unable to get rid of the SN II ejecta, but still give birth to a cluster in which cooling flows play a dominant role. We remark that the only similar trend of increase in the s–process abundance, in the same range of metallicities, is found in M 22 [@marino2009; @dacostamarino2010], although this cluster completely lacks the most metal rich populations (\[Fe/H\]$>$–1.5) present in [$\omega$ Cen]{}. Even the smaller cluster NGC 1851 shows an s–process increase by a factor 2–3 [@yong2008], and this may be correlated to a possible, still discussed, small increase in the iron content [@carretta2010.1851; @villanova1851.2010]. It would be interesting to know whether the cluster NGC 2419, that harbors two populations with a small variation in Calcium content [@cohen2010], and a population with very high helium [@dicriscienzo2011] also shows a similar s–process dichotomy. It looks very difficult that these very different clusters managed to evolve in such a similar way, if there is not an unifying model for the s–process production, linked mainly to stars that evolve into SN II. Further work is required to clarify this issue.
Conclusions
===========
We have examined the patterns of the O–Na–Fe data by [@marino2011] for the giants in [$\omega$ Cen]{}. Most stars at the largest iron contents (\[Fe/H\]$>$–1.3) are all very sodium-rich and moderately oxygen depleted, a peculiarity not predicted in standard galactic evolution, and not found in other environments apparently similar, like in the metal rich field stars of the Sagittarius dwarf galaxy. Furthermore, the metal-rich stars populating the MS–a might be helium-rich [@bellini2010]. We have shown that these peculiar abundances (including high helium content) match those of the ejecta of the massive AGB stars of comparable metallicities [@vd2009].
We attribute the presence of a direct O–Na correlation for the stars at \[Fe/H\]$>$–1.3 to formation, in a cooling flow, from the ejecta of massive AGB stars of the same metallicity, that evolve in the cluster when the gas in the surroundings has been all exhausted by previous episodes of star formation. These previous events have formed the populations with \[Fe/H\]$<$–1.3, from the ejecta of AGBs, diluted with the recollapsed gas, metal enriched by the SN II preceding burst. This process forms the typical Na–O anticorrelation patterns. The reason why massive, metal rich AGBs are available for this phase of star formation is that, with respect to the AGBs at smaller iron content, they evolve with the time delay that has been necessary to form the metal-rich components in the progenitor system. Consequently, [$\omega$ Cen]{} differs from mono-metallic clusters because its initial mass must have been even larger (possibly including a dark matter halo) than that of a typical cluster.
Consequently, [$\omega$ Cen]{} differs from mono–metallic clusters because its initial mass must have been even larger (possibly including a dark matter halo) than the initial mass of a typical cluster. This led to an evolution in which SN II ejecta could not be expelled from the cluster, but were incorporated in successive bursts of star formation, as described by [@marcolini2006]. The existence of helium-rich populations and the concentrated spatial distribution of the helium-rich and of the most metal-rich populations [@suntzeff1996; @norris1996; @rey2004; @bellini2009; @johnson2010], however, shows that also in this cluster the gas from which second generation stars form collects in a cooling flow in the cluster central regions as described in the simulations by D’Ercole et al. (2008) for standard GCs. A further reason, to think that these latter stars were formed in cooling flows, is the O–Na composition of their stars, that follows the direct O–Na correlation predicted by massive AGB stellar models. We showed that the formation of the [$\omega$ Cen]{} multiple populations seems to follow the basic steps of the [@dercole2010] model for mono–metallic GCs. The O–Na patterns, and the Mg and Al abundances in [$\omega$ Cen]{} then contribute to validate the AGB pollution model for the formation of multiple populations in GCs.
An important and still unsolved issue in the formation of [$\omega$ Cen]{} multiple populations is which are the stars responsible for the fast increase in the s–process abundances, that occurs at the epoch in which the metallicity raises from $\sim$–1.9 to $\sim$–1.5. We discussed the common view of s–elements production in low mass AGB stars, and concluded that it is in contrast with the other strong indicators of formation of the GC populations in AGB–dominated cooling flows. We suggest that the s–process patterns are a red herring, and propose that their evolution is due to a process not yet identified or well explored in stellar models, such as, e.g., the nucleosynthesis in the carbon burning shells of the tail of lower mass progenitors of SNII. This (postulated) nucleosynthesis site becomes dominant in the chemistry of clusters showing a spread in metallicity, in the same way as the O–Na anticorrelation, due to the peculiar chemical evolution provoked by the collection of winds from a restricted range of initial stellar masses, dominates the chemistry of GC stars. Further work is required to clarify this suggestion. For the same reasons, we suggest that the whole time of formation of [$\omega$ Cen]{} is at most a few 10$^8$yr.
F.D., A.D. and P.V. have been supported by the PRIN-INAF 2009 grant “Formation and Early Evolution of Massive Star Cluster". E.V. was supported in part by the grant NASA-NNX10AD86G. We thank Y.W. Lee for discussion on the problem of the age spread in [$\omega$ Cen]{}, and A. Chieffi, M.L. Pumo and R.G. Gratton for discussion on the stellar sites for the s–process nucleosynthesis. We thank the referee for a constructive report. We are grateful to A. Renzini for a critical reading of a first version of this manuscript. This work is dedicated to F.D.’s grandson Filippo and his battle against leukemia.
[99]{}
Bedin, L. R., Piotto, G., Anderson, J., Cassisi, S., King, I. R., Momany, Y., & Carraro, G. 2004, ApJ Letters, 605, L125
Bekki, K., & Norris, J. E. 2006, ApJ Letters, 637, L109
Bekki, K., Campbell, S. W., Lattanzio, J. C., & Norris, J. E. 2007, , 377, 335
Bekki, K. 2010, arXiv:1011.5956
Bellini, A., Piotto, G., Bedin, L. R., King, I. R., Anderson, J., Milone, A. P., & Momany, Y. 2009, , 507, 1393
Bellini, A., Bedin, L. R., Piotto, G., Milone, A. P., Marino, A. F., & Villanova, S. 2010, , 140, 631
Bragaglia, A., et al. 2010, , 720, L41
Busso, M., Gallino, R., & Wasserburg, G. J. 1999, , 37, 239
Carretta, E., Bragaglia, A., Gratton, R. G., Leone, F., Recio-Blanco, A., & Lucatello, S. 2006, , 450, 523
Carretta, E., et al. 2009a, A&A, 505, 117
Carretta, E., Bragaglia, A., Gratton, R., & Lucatello, S. 2009b, A&A, 505, 139
Carretta, E., Bragaglia, A., Gratton, R., D’Orazi, V., & Lucatello, S. 2009c, , 508, 695
Carretta, E., et al. 2010a, , 520, A95
Carretta, E., et al. 2010b, , 714, L7
Carretta, E., et al. 2010c, , 722, L1
Choi, E., & Yi, S. K. 2008, , 386, 1332
Cohen, J. G., Kirby, E. N., Simon, J. D., & Geha, M. 2010, , 725, 288
Cunha, K., Smith, V. V., Suntzeff, N. B., Norris, J. E., Da Costa, G. S., & Plez, B. 2002, , 124, 379
Da Costa, G. S., & Marino, A. F. 2010, arXiv:1009.1955
D’Antona, F., & Caloi, V. 2004, , 611, 871
D’Antona, F., & Ventura, P. 2007, , 379, 1431
D’Antona, F., Stetson, P. B., Ventura, P., Milone, A. P., Piotto, G., & Caloi, V. 2009, , 399, L151
Decressin, T., Meynet, G., Charbonnel, C., Prantzos, N., & Ekstr[ö]{}m, S. 2007, A&A, 464, 1029
D’Ercole, A., Vesperini, E., D’Antona, F., McMillan, S. L. W., & Recchi, S. 2008, , 391, 825
D’Ercole, A., D’Antona, F., Ventura, P., Vesperini, E., & McMillan, S. L. W. 2010, , 407, 854
D’Ercole, A., D’Antona, F. & Vesperini, E. 2011, , in press, arXiv:1103.4715
Dickens, R. J., Croke, B. F. W., Cannon, R. D., & Bell, R. A. 1991, , 351, 212
Di Criscienzo, M., et al. 2011, arXiv:1103.0867
Dinescu, D. I., Girard, T. M., & van Altena, W. F. 1999, , 117, 1792
D’Orazi, V., & Marino, A. F. 2010, , 716, L166
Gal-Yam, A., & Maoz, D. 2004, , 347, 942
Gratton, R. G., Bonifacio, P., Bragaglia, A., et al. 2001, A&A, 369, 87
Gratton, R., Sneden, C., & Carretta, E. 2004, ARA&A, 42, 385
Karakas, A. I., Fenner, Y., Sills, A., Campbell, S. W., & Lattanzio, J. C. 2006, , 652, 1240
Ivans, I. I., Sneden, C., Kraft, R. P., Suntzeff, N. B., Smith, V. V., Langer, G. E., & Fulbright, J. P. 1999, , 118, 1273
Johnson, C. I., Pilachowski, C. A., Simmerer, J., & Schwenk, D. 2008, , 681, 1505
Johnson, C. I., Pilachowski, C. A., Michael Rich, R., & Fulbright, J. P. 2009, , 698, 2048
Johnson, C. & I., Pilachowski, C. A. 2010, arXiv:1008.2332
Karakas, A., & Lattanzio, J. C. 2007, Publications of the Astronomical Society of Australia, 24, 103 Lee, Y.-W., Joo, J.-M., Sohn, Y.-J., Rey, S.-C., Lee, H.-C., & Walker, A. R. 1999, , 402, 55
Lee, Y.-W., et al. 2005, , 621, L57
Lind, K., Charbonnel, C., Decressin, T., Primas, F., Grundahl, F., & Asplund, M. 2011, , 527, A148
Mannucci, F., Della Valle, M., & Panagia, N. 2006, , 370, 773
Marcolini, A., D’Ercole, A., Brighenti, F., & Recchi, S. 2006, , 371, 643
Marcolini, A., Sollima, A., D’Ercole, A., Gibson, B. K., & Ferraro, F. R. 2007, , 382, 443
Marino, A. F., Milone, A. P., Piotto, G., Villanova, S., Bedin, L. R., Bellini, A., & Renzini, A. 2009, , 505, 1099
Marino, A. F., et al. 2011, , 731, 64
Matteucci, F., & Greggio, L. 1986, , 154, 279
Matteucci, F., & Recchi, S. 2001, , 558, 351
Meynet, G., Ekstr[ö]{}m, S., & Maeder, A. 2006, A&A, 447, 623
Norris, J. E., & Da Costa, G. S. 1995, , 447, 680
Norris, J. E., Freeman, K. C., & Mighell, K. J. 1996, , 462, 241
Norris, J. E., Freeman, K. C., Mayor, M., & Seitzer, P. 1997, , 487, L187
Norris, J. E. 2004, ApJ Letters, 612, L25
Origlia, L., Ferraro, F. R., Bellazzini, M., & Pancino, E. 2003, , 591, 916
Paltoglou, G., & Norris, J. E. 1989, , 336, 185
Pancino, E., Ferraro, F. R., Bellazzini, M., Piotto, G., & Zoccali, M. 2000, , 534, L83
Pancino, E., Pasquini, L., Hill, V., Ferraro, F. R., & Bellazzini, M. 2002, , 568, L101
Pignatari, M., Gallino, R., Meynet, G., Hirschi, R., Herwig, F., & Wiescher, M. 2008, , 687, L95
Pietrinferni, A., Cassisi, S., Salaris, M., Percival, S., & Ferguson, J. W. 2009, , 697, 275
Pilachowski, C. A. 1988, , 326, L57
Piotto, G., et al. 2005, ApJ, 621, 777
Prantzos, N., & Charbonnel, C. 2006, , 458, 135
Pumo, M. L., Contino, G., Bonanno, A., & Zappal[à]{}, R. A. 2010, , 524, A45
Raiteri, C. M., Gallino, R., Busso, M., Neuberger, D.,& Kaeppeler, F. 1993, , 419, 207
Ram[í]{}rez, S. V., & Cohen, J. G. 2002, , 123, 3277
Raskin, C., Scannapieco, E., Rhoads, J., & Della Valle, M. 2009, , 707, 74
Renzini, A. 2008, , 391, 354
Rey, S.-C., Lee, Y.-W., Ree, C. H., Joo, J.-M., Sohn, Y.-J., & Walker, A. R. 2004, , 127, 958
Romano, D., Matteucci, F., Tosi, M., Pancino, E., Bellazzini, M., Ferraro, F. R., Limongi, M., & Sollima, A. 2007, , 376, 405
Romano, D., Tosi, M., Cignoni, M., Matteucci, F., Pancino, E., & Bellazzini, M. 2010, , 401, 2490
Sbordone, L., Bonifacio, P., Buonanno, R., Marconi, G., Monaco, L., & Zaggia, S. 2007, , 465, 815
Shen, Z.-X., Bonifacio, P., Pasquini, L., & Zaggia, S. 2010, , 524, L2
Smith, G. H., Shetrone, M. D., Bell, R. A., Churchill, C. W., & Briley, M. M. 1996, , 112, 1511
Smith, V. V., Suntzeff, N. B., Cunha, K., Gallino, R., Busso, M., Lambert, D. L., & Straniero, O. 2000, , 119, 1239
Sollima, A., Pancino, E., Ferraro, F. R., Bellazzini, M., Straniero, O., & Pasquini, L. 2005a, , 634, 332
Sollima, A., Ferraro, F. R., Pancino, E., & Bellazzini, M. 2005b, , 357, 265
Sullivan, M., et al. 2006, , 648, 868
Suntzeff, N. B., & Kraft, R. P. 1996, , 111, 1913
The, L.-S., El Eid, M. F., & Meyer, B. S. 2007, , 655, 1058
Tolstoy, E., Hill, V., & Tosi, M. 2009, , 47, 371
Tornambè, A. 1989, , 239, 771
Ventura, P., D’Antona, F., Mazzitelli, I., & Gratton, R. 2001, ApJ Letters, 550, L65
Ventura, P., & D’Antona, F. 2005a, A&A, 431, 279
Ventura, P. & D’Antona, F. 2005b, , 439, 1075
Ventura P., D’Antona F., 2009, A&A, 499, 835
Ventura, P., & D’Antona, F. 2011, , 410, 2760
Ventura, P., Carini, R. & D’Antona, F. 2011, ,
Villanova, S., et al. 2007, , 663, 296
Villanova, S., Geisler, D., & Piotto, G. 2010, , 722, L18
Yong, D., & Grundahl, F. 2008, , 672, L29
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We discuss the geometry of Laplacian eigenfunctions $-\Delta \phi = \lambda \phi$ on compact manifolds $(M,g)$ and combinatorial graphs $G=(V,E)$. The ’dual’ geometry of Laplacian eigenfunctions is well understood on $\mathbb{T}^d$ (identified with $\mathbb{Z}^d$) and $\mathbb{R}^n$ (which is self-dual). The dual geometry is of tremendous role in various fields of pure and applied mathematics. The purpose of our paper is to point out a notion of similarity between eigenfunctions that allows to reconstruct that geometry. Our measure of ’similarity’ $ \alpha(\phi_{\lambda}, \phi_{\mu})$ between eigenfunctions $\phi_{\lambda}$ and $\phi_{\mu}$ is given by a global average of local correlations $$\alpha(\phi_{\lambda}, \phi_{\mu})^2 = \| \phi_{\lambda} \phi_{\mu} \|_{L^2}^{-2}\int_{M}{ \left( \int_{M}{ p(t,x,y)( \phi_{\lambda}(y) - \phi_{\lambda}(x))( \phi_{\mu}(y) - \phi_{\mu}(x)) dy} \right)^2 dx},$$ where $p(t,x,y)$ is the classical heat kernel and $e^{-t \lambda} + e^{-t \mu} = 1$. This notion recovers all classical notions of duality but is equally applicable to other (rough) geometries and graphs; many numerical examples in different continuous and discrete settings illustrate the result.'
address:
- 'Department of Mathematics, University of California, San Diego, CA 92093, USA'
- 'Department of Mathematics, Yale University, New Haven, CT 06511, USA'
author:
- Alexander Cloninger
- Stefan Steinerberger
title: |
On the Dual Geometry of\
Laplacian Eigenfunctions
---
Introduction
============
Introduction.
-------------
The Laplacian eigenfunctions on $\mathbb{T}^2$ are easily determined and given by $\phi_k = e^{i\left\langle k, x\right\rangle}$, where $k \in \mathbb{Z}^2$ is a lattice point. They are orthogonal in $L^2$ and allow for a representation of a function $f \in L^2(\mathbb{T}^2)$. However, it becomes very quickly apparent that the geometry of $\mathbb{Z}^2$ is not only a convenient enumeration but plays a fairly fundamental role itself. Examples include
1. a beautiful inequality of Zygmund [@zyg] stating that for any $r > 0$ $$\sum_{ \|k\| = r}{ |\widehat{f}(k)|^2} \leq 5^{1/2} \|f\|^2_{L^{4/3}(\mathbb{T}^2)}$$ and, more generally, restriction phenomena in Harmonic Analysis.
2. the analysis of the nonlinear Schrödinger equation (see e.g. [@tao]) $$i u_t + \Delta u = |u|^{p-1}u$$ as well as other general nonlinear dispersive equations,
3. the structure of pseudo-differential operators [@coifman],
4. the operations of wavelets and spectral filters on images [@coif2]
5. or, a personal example [@stein], an inequality for $f \in C^1(\mathbb{T}^2)$ with mean value 0 $$\| \nabla f\|_{L^2(\mathbb{T}^2)} \left\| \partial_x f + \sqrt{2} \partial_y f\right\|_{L^2(\mathbb{T}^2)} \geq c \| \nabla f\|^2_{L^2(\mathbb{T}^2)}.$$
It is clear that the torus $\mathbb{T}^d$ is a special case. The same is true for $\mathbb{R}^d$ whose eigenfunctions (now understood in a distributional sense) satisfy the same additive relationship $$e^{i\left\langle m, x\right\rangle} e^{i\left\langle n, x\right\rangle} = e^{i\left\langle m + n, x\right\rangle}.$$ This is one of the reasons why so much mathematical analysis is possible on $\mathbb{T}^d$ and $\mathbb{R}^d$ that cannot be easily generalized. A somewhat philosophical question, intentionally put vaguely, is
> whether or not, for a given manifold $(M,g)$ or Graph $G=(V,E)$, there is additional structure in the eigenfunctions beyond orthogonality and the eigenvalue.
Graph Signal Processing
-----------------------
This question has turned out to be of increasing importance in modern problems of data science for rather obvious reasons: if we are interested in either selecting or filtering out certain types of substructures and if those substructures tend to be connected to certain types of eigenvectors or eigenfunctions, then one needs to understand that interplay. Conversely, ’proximity’ of eigenvectors should be indicative of capturing the same kind of phenomenon. This is easily observed on $\mathbb{T}^d$ and $\mathbb{R}^d$ where eigenvectors $e^{i\left\langle m, x\right\rangle}, e^{i\left\langle n, x\right\rangle}$ are ’close’ if $\|m - n\|$ is small and this corresponds to having oscillations point in roughly the same direction. The importance of this elementary observation is difficult to overstate; take, for example, $f \in L^2(\mathbb{R}^2)$ and define $g \in L^2(\mathbb{R}^2)$ via restricting the Fourier transform $$\widehat{g}(\xi_1, \xi_2) = \chi_{\xi_1 + \xi_2 \geq 0}(\xi_1, \xi_2) \widehat{f}(\xi_1, \xi_2),$$ then this corresponds to a filter with an obvious geometric interpretation. More precisely, the very notion of a Fourier multiplier (and, correspondingly entire branches of mathematics) is intimately entangled with this underlying geometry of the eigenfunctions of the Laplacian.
> **Challenge** (Graph Signal Processing)**.** Given a Graph $G=(V,E)$, define, if possible, an analogous geometry on its eigenvectors.
This is an absolutely fundamental problem, we refer to [@ankenmann; @ank2; @bron; @co1; @co2; @co3; @co4; @co5; @sarah; @ham; @irion; @per; @saito; @shu; @shu2; @shu3] for recent examples. Moreover, it is not expected that this is always (or even generically) possible – even in Euclidean space, one would expect that eigenfunctions on generic domains do not have any distinguishing features except for their eigenvalue; this vague statement is made precise in different ways in the study of quantum chaos [@haa; @nonne].
Our contribution.
-----------------
The purpose of this paper is to present a somewhat curious definition of a notion of affinity or similarity between two eigenfunctions (or eigenvectors). This notion is identical on manifolds and graphs and results in a number in $[0,1]$ with 1 indicating strong similarity and 0 denoting weak similarity. This then allows us to take a finite set of eigenfunctions $\left\{ \phi_1, \dots, \phi_n \right\}$ and compute an $n \times n$ matrix $A \in [0,1]^{n \times n}$ where $A_{ij}$ denotes the similarity of $\phi_i$ and $\phi_j$. This is then interpreted as the weighted affinity matrix of a complete weighted Graph on $\left\{ \phi_1, \dots, \phi_n \right\}$ that we hope encodes the underlying geometry of the eigenfunctions. We then use a fairly standard visualization technique that embeds complete weighted graphs into Euclidean space in a geometry-preserving manner; there is some flexibility in this step and one could use any number of visualization techniques. The main purpose of this paper is to discuss this algorithm in detail and show that it recovers the geometry of the eigenfunctions in all classical cases where such a geometry exists. This has substantial theoretical and practical applications.
1. Practically, this allows us to group eigenvectors of a Laplacian into various natural geometric substructures beyond just ordering by their eigenvalue. This is of obvious significance in Graph Signal Processing, as well as in the choice of eigenvectors used to visualize a number of non-linear dimensionality reduction embeddings.
2. Theoretically, it raises a curious connection between the geometry of eigenfunctions, how the pointwise product $\phi_{i} \phi_{j}$ spreads over the spectrum (i.e. what $\left\langle \phi_i \phi_j, \phi_k\right\rangle$ looks like as a sequence in $k$, see also [@stein2]) and, through its definition as a local corellation, the local structure. This gives rise to a number of fascinating theoretical prolems.
A Notion of Similarity
======================
Similarity.
-----------
We define a quantitative notion of similarity $0 \leq \alpha(\phi_{\lambda}, \phi_{\mu}) \leq 1$ between two eigenfunctions. We first define it on compact manifolds $(M,g)$ without boundary (this assumption can be dropped but simplifies exposition). We denote the solution of the heat equation $$\begin{aligned}
(\partial_t - \Delta) u(t,x) &= 0\\
u(0,x) &= f(x)\end{aligned}$$ by $e^{t\Delta}f$. This induces a heat kernel $p(t,x,y)$ satisfying $$\left[ e^{t\Delta}f \right] (x) = \int_{M}{ p(t,x,y)f(y) dy}.$$ We have $p(t,x,y) = p(t,y,x)$ and conservation of the $L^1-$mass implies that $p(t,x,\cdot)$ is a probability distribution. Moreover, Varadhan’s short-time asymptotic implies that for $t$ small, $p(t,x,\cdot)$ is essentially a Gaussian centered at $x$ and scale $\sim \sqrt{t}$. We introduce the following measure of similarity between two eigenfunctions $$\alpha(\phi_{\lambda}, \phi_{\mu})^2 = \| \phi_{\lambda} \phi_{\mu} \|_{L^2}^{-2}\int_{M}{ \left( \int_{M}{ p(t,x,y)( \phi_{\lambda}(y) - \phi_{\lambda}(x))( \phi_{\mu}(y) - \phi_{\mu}(x)) dy} \right)^2 dx},$$ where $t$ is the unique solution of $$e^{-t \lambda} + e^{-t \mu} = 1.$$ This is *not* a metric. The main motivation for this quantity is that
1. it can be interpreted as an average over local correlations: eigenfunctions that behave similar should look locally similar in lots of places and
2. that it appeared somewhat naturally in studies on products of eigenfunctions [@stein2] where it was shown to satisfy the identity $$\alpha(\phi_{\lambda}, \phi_{\mu}) = \frac{\|e^{t\Delta} (\phi_{\lambda} \phi_{\mu}) \|}{\|\phi_{\lambda} \phi_{\mu} \|_{L^2}}$$ for exactly $e^{-t \lambda} + e^{-t \mu} = 1$.
The second property immediately shows $$0 \leq \alpha( \phi_{\mu}, \phi_{\lambda}) \leq 1.$$
Both local correlations as well as the diffusion under the heat equation (and thus, implicitly, the distribution in the spectrum) is measured by the same quantity $\alpha$. It being large implies strong local correlations and frequency transport to lower frequencies (meaning that the expansion of the product $\phi_{\lambda} \phi_{\mu}$ into eigenfunctions contains some low-frequency contributions) and, conversely, it being small implies frequency transport to higher frequencies (see [@stein2] for details). This motivated it as an interesting object worthy of further study and the investigations reported on in this paper.
Creating a Landscape.
---------------------
We observe that the definition $\alpha(\phi_{\mu}, \phi_{\lambda})$ only gives rise to a weighted graph, however, we would very much like to understand its intrinsic structure. This leads to an a very substantial problem that is currently receiving a great deal of interest: how to ’accurately’ map vertices of a weighted graph to $\mathbb{R}^d$ in such a way that vertices that are ’close to each other’ are also close in the embedding and, conversely, vertices that are far away are will be mapped to different regions in space (see Figure 1). We often want $d \in \left\{2,3\right\}$ for visualization purposes.
(0,0) circle (0.05cm); (1,0) circle (0.05cm); (0,1) circle (0.05cm); (1,1) circle (0.05cm); (0,0) to\[out=350, in = 200\] (1,0); (0,0) to\[out=50, in = 200\] (1,1); (0,0) to\[out=100, in = 260\] (0,1); (1,0) to\[out=80, in = 280\] (1,1); (1,0) to\[out=45, in = 300\] (1.3, 1.3) to\[out =120, in =45\] (0,1); (0,1) to\[out=10, in = 170\] (1,1); at (-0.3, -0.15) [$\phi_{\lambda}$]{}; at (1.3, -0.15) [$\phi_{\mu}$]{}; at (0.5, -0.25) [$\alpha(\phi_{\lambda},\phi_{\mu})$]{}; at (2, 0.5) [$\implies$]{}; (3,0) – (3,1); (3,0) – (4,0); (3.2,0.3) circle (0.03cm); at (3.2, 0.5) [$\phi_{\lambda}$]{}; (3.6,0.2) circle (0.03cm); at (3.7, 0.3) [$\phi_{\mu}$]{}; (3.8,0.8) circle (0.03cm); (3.8,0.85) circle (0.03cm);
There are a number of methods for creating a low dimensional embedding of points given some notion of distance or similarity between points [@belkin2; @lafon; @kruskal; @jolliffe; @pearson; @saul]. Ultimately, if the structure is well encoded in the mutual distances, then it does not matter very much which method is used. Throughout this paper, we only use one of the very simplest methods: let us enumerate the eigenfunctions under consideration by $\left\{1, \dots, n\right\}$ and let $$A_{ij} = \alpha(\phi_i, \phi_j)$$ be the $n \times n$-matrix containing all mutual affinities. We then map $$\phi_i \rightarrow\left(v_{1,i}, v_{2, i}, v_{3, i}\right),$$ where $v_1, v_2, v_3$ are the three eigenvectors corresponding to the three largest (in absolute value) eigenvalues of the matrix $A$. The subscript denotes the $i-$th entry of the vector; in particular, for every eigenfunction $\phi_i$ we use the $i-$th entry in the first, second and third largest eigenvectors as $(x,y,z)-$coordinates for a point in $\mathbb{R}^3$. Sometimes it may be advantageous to not use the first three eigenvectors but the best result is usually obtained by picking three eigenvectors associated to the lowest eigenvalues. We will also sometimes use $$A_{ij} = \alpha(\phi_i, \phi_j)^p \qquad \mbox{for some}~p \geq 1$$ which has the effect of making strong existing affinities even more pronounced by disproportionately weakening smaller affinities. We recall that $$\alpha(\phi_i, \phi_j) = \frac{\|e^{t\Delta} (\phi_{i} \phi_{j}) \|_{L^2}}{\|\phi_{i} \phi_{j} \|_{L^2}}$$ requires the computation of a suitable $t$ depending on the eigenvalues of $\phi_i$ and $\phi_j$. We emphasize that the value depends smoothly on $t$ and is not very sensitive. Sometimes we will even use a fixed value $t_0$, independently of the eigenvalues, for all computations to demonstrate robustness. Finally, at points in this section, we also deal with eigenvectors of the normalized graph Laplacian $\mathcal{L}$ defined on a finite collection of points $\left\{x_1, \dots, x_n \right\} \subset \mathbb{R}^d$. This would correspond to a numerical computation of eigenvectors of a discretization of a continuous manifold. We follow standard procedure and define neighbors of points via the matrix $$\begin{aligned}
K_{i,j} = \exp\left(-\|x_i-x_j\|_2^2/\sigma^2\right),\end{aligned}$$ where $\sigma > 0$ is a parameter that fixes a scale. The normalized graph Laplacian is then denoted $$\begin{aligned}
L = \mbox{Id}_{n \times n} - D^{-1/2} K D^{-1/2}, & \textnormal{ where } D_{i,j} = \begin{cases} \sum_{j=1}^{n} K_{i,j} , & \textnormal{ for } i=j\\ 0, & \textnormal{ for } i\neq j\end{cases}.\end{aligned}$$ This normalized Laplacian has an eigendecomposition which satisfies properties similar to those of the manifold Laplacian eigenfunctions; if the points are actually sampled from an underlying manifold, then this construction is known to converge to the continuous Laplacian [@belkin1]. We will use this notion of Graph Laplacian throughout the paper when constructing affinities of eigenvectors.
Numerical Examples
==================
The remainder of the paper is devoted to the study of dual landscapes of Laplacian eigenfunctions and eigenvectors obtained by the method outlined above. More precisely, we will consider
1. The one-dimensional Torus $\mathbb{T}$ (with endpoints identified) discretized to a cycle graph $C_n$, the eigenfunctions are given by discrete approximations of $\sin{(kx)}$ and $\cos{(kx)}$.
2. Spherical harmonics $\phi_{\ell}^m$ as eigenfunctions of $\mathbb{S}^2$; we recover both indices $\ell$ and $m$.
3. A standard rectangle $[0,4] \times [0,1] \subset \mathbb{R}^2$. Eigenfunctions of the Laplacian are grouped by specifying the number of oscillations in each direction; the method recovers this perfectly.
4. We then study more general cartesian products; the Laplacian eigenfunction of $A \times B$ is merely the product of eigenfunctions on $A$ and eigenfunctions on $B$. The underlying cartesian structure is perfectly recovered even if $A, B$ are rather structure-less objects.
5. On the other end of the spectrum, we consider Erdős-Renyi random graphs. Laplacian eigenfunctions on these objects should not exhibit any particular structure nor any distinguishing features. We recover an ordering with respect to the eigenvalue.
6. We conclude by demonstrating several surprising eigenfunctions landscapes, and discuss this method’s uses in exploratory spectral graph theory.
We emphasize that, while most examples were computed on the known eigenfunctions, equivalent landscapes are discovered for the empirical eigenvectors of the graph Laplacian on the domain.
The one-dimensional Torus.
--------------------------
We begin with the Laplacian eigenfunctions of the torus restricted to a uniform grid consisting of $n=100$ equispaced points (resulting in $n$ eigenvectors with at least low-frequency eigenvectors approximating the classical trigonometric functions). We emphasize that especially at high frequencies, we only have one sampling point for a wavelength (which is very little). The method nonetheless remains functional (but the effect can be seen in the affinity matrix). Figure \[fig:analyticFourier\] shows the distance matrix $A$, Figure \[fig:analyticFourier2\] shows the spectral embedding.
![Affinity Matrix of the Eigenfunctions on a discretized Torus.[]{data-label="fig:analyticFourier"}](code/AnalyticCircFourier/analyticFourierDistanceMatrix.png){height=".45\textwidth"}
The method clearly recovers the linear dual structure of the eigenfunctions. It also clearly states that eigenvectors approximating $\sin{kx}$ and $\cos{kx}$ are quite different functions but orders them in nearby points in the landscape because they behave similarly with respect to other eigenfunctions. Finally, the identity $$\exp{\left(i \left[ \frac{n}{2} - k \right] \frac{j}{n} \right)} \exp{\left(i \left[ \frac{n}{2} + k \right] \frac{j}{n} \right)} = e^{2\pi i j} = 1$$ is automatically discovered. The embedding of the fourth (pointwise) power of the affinity matrix (i.e. $p=4$ as described above) shows a linear structure underlying the eigenvectors.
![Embedding of the distance matrix $A^4$[]{data-label="fig:analyticFourier2"}](code/AnalyticCircFourier/analyticFourierVaryTime4power_nolabel.png){height=".5\textwidth"}
Spherical Harmonics
-------------------
We now describe the process on spherical harmonics, i.e. the eigenfunctions of the Laplacian on $\mathbb{S}^2$. They can be separated into levels $\phi_{\ell}^{m}$, where $-\ell\le m \le \ell$, with corresponding eigenvalue $\ell(\ell+1)$. We generate these eigenfunctions by taking linear spacing of $181$ points in both spherical coordinate angles $(\alpha, \beta)$, which makes the eigenfunctions satisfy the orthogonality relation $$\begin{aligned}
\int_0^\pi \int_0^{2\pi} \phi_{\ell}^{m}(\alpha,\beta) \phi_{\ell'}^{m'}(\alpha,\beta) \sin(\alpha) d\beta d\alpha = \delta_{\ell,\ell'} \delta_{m, m'}.\end{aligned}$$ Figure \[fig:analyticSphHar\] describes an embedding of the 256 lowest frequency spherical harmonic functions ($\ell \leq 14$).
![Embedding of $A^{(2)}$ for the low frequency spherical harmonics. For the spectral embedding, each point corresponds to a specific spherical harmonic. Lines are drawn to connect all $\phi_{\ell}^m$ for a fixed $\ell$, and the color of the node corresponds to $m/\ell$. $\phi_{\ell}^m$ and $\phi_{\ell}^{-m}$ are effectively on top of one another in the embedding, $\phi_{\ell}^m$ for $m<0$ are represented by dots and $\phi_{\ell}^m$ for $m\ge 0$ are circles.[]{data-label="fig:analyticSphHar"}](code/AnalyticSphHar/analyticSphHarVaryTime2power_full_nolabel_large2.png){height=".45\textwidth"}
The method clearly recovers the level $\ell$ of the spherical harmonics. It also recovers the relative ordering of the degree $m$ of the spherical harmonic, with $\phi_{\ell}^m$ and $\phi_\ell^{-m}$ being found to be quite different functions. Despite this, $\phi_{\ell}^m$ and $\phi_\ell^{-m}$ are ordered into nearby points in the landscape as they behave similarly with respect to the rest of the spherical harmonics.
Cartesian Product Structure
---------------------------
We consider the rectangle $[0,4]\times [0,1] \subset \mathbb{R}^2$. The eigenfunctions of the Laplacian with Dirichlet boundary conditions are given by $$\phi_{mn} = \sin\left(\frac{mx}{4}\right) \sin\left(\frac{nx}{1}\right) \quad \mbox{ with corresponding eigenvalue} \quad \lambda_{mn} = \frac{m^2}{16} + \frac{n^2}{1^2}.$$
![Low frequency eigenfunctions of the Laplacian on a rectangular grid, with images of each eigenvector displayed in their respective positions. []{data-label="fig:analyticRectangleImages"}](code/AnalyticRectangle/analyticRectangleVaryTime5x10EigImagesRectangle_3.png){width="90.00000%"}
The precise geometric structure, how many times an eigenfunction oscillates in the $x-$direction vs. how many times it oscillates in the $y-$direction cannot be understood by eigenvalue alone (especially in the high-frequency limit, these points are equally spaced on an ellipse). Our method clearly recovers the underlying oscillation structure and orders eigenfunctions accordingly.
![Embedding of eigenfunctions on $[0,4] \times [0,1]$. []{data-label="fig:analyticRectangle"}](code/AnalyticRectangle/analyticRectangleVaryTime2power_nolabel_connected.png){height=".45\textwidth"}
We restricted the exact eigenfunctions to $n=400$ grid points arranged as a $40 \times 10$ rectangle. The spectral embedding reflects the separable relationship between eigenfunctions oscillating in differing directions, with the $10$ lines of $40$ eigenfunctions perfectly grouping $\phi_{mn}$. Figure \[fig:analyticRectangle\] displays the full landscape of the $400$ eigenfunctions. Figure \[fig:analyticRectangleImages\] also displays the the embedding for $1\le m\le 5$ and $1\le n \le 10$ with each point being the image of the respective $\phi_{mn}$, in order to demonstrate that the spectral embedding does in fact organize the eigenfunctions correctly.
More General Cartesian Products
-------------------------------
The reconstruction of tensor product geometry and separable eigenfunctions holds at a much greater level of generality.
![Embedding of eigenvectors recovers $X\times Y$ with $x_i\sim \mathcal{N}(0,\sigma^2)$.[]{data-label="fig:empiricalTensorProduct"}](code/EmpiricalTensorProduct/empiricalTensorProductVaryTime4power.png){height=".5\textwidth"}
We sample a set $X$ of 100 points in $\mathbb{R}^2$ from a Gaussian distribution $X\sim \mathcal{N}(0,\sigma^2 I_d)$, where $\sigma^2 =0.01$, then take the set $Y$ of 10 equispaced points from $[0,1]$ and consider the set $X \times Y$. The Cartesian Product Structure is perfectly recovered, the result is shown in Figure \[fig:empiricalTensorProduct\]. The method clearly recovers the frequency of oscillation in the $Y$ direction, and separates the eigenfunctions into different groups that constructively interfere with one another. In particular, the point at the tail of the $k^{th}$ line corresponds to the eigenfunction that is constant on $X$ and varies $k$ times in the $Y$ direction.
Objects without Structure
-------------------------
We now consider an Erdős-Renyi Graph $G(1000,0.2)$. This is a random object on $n=1000$ vertices with likelihood of two vertices being connected being $p=0.2$. We expect this object to be completely random and eigenfunctions to all behave in a fairly uniform manner. Results of this flavor have been of great interest recently. We refer, for example, to a rather general result of Rudelson & Vershynin [@rudi] who prove that eigenvectors of graphs with i.i.d. random entries are uniformly flat in the sense of not having any entries substantially larger than $\sim n^{-1/2}$ (up to logarithmic factors).
![Embedding of eigenvectors of the normalized graph Laplacian on Erdős-Renyi graph.[]{data-label="fig:empiricalRandomGraph"}](code/EmpiricalRandomGraph/empiricalRandomGraphVaryTime1power_normalized.png){height=".45\textwidth"}
The result of the embedding is shown in Figure \[fig:empiricalRandomGraph\]. We clearly observe that the first eigenvector (which does not change sign) is separated from the rest but the remaining eigenvectors are fairly structureless and clearly ordered with respect to their eigenvalue.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Embedding of eigenvectors recovers $X\times Y$ for cartesian product graph. (Left) first three coordinates of embedding, (Right) second, third, and forth coordinates of embedding.[]{data-label="fig:empiricalRandomCrossGrid"}](code/EmpiricalRandomCrossGrid/empiricalRandomCrossGridVaryTime2power_repeatGraph.png "fig:"){width=".5\textwidth"} ![Embedding of eigenvectors recovers $X\times Y$ for cartesian product graph. (Left) first three coordinates of embedding, (Right) second, third, and forth coordinates of embedding.[]{data-label="fig:empiricalRandomCrossGrid"}](code/EmpiricalRandomCrossGrid/empiricalRandomCrossGridVaryTime2power_repeatGraph_234.png "fig:"){width=".5\textwidth"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Exploratory Spectral Graph Theory
---------------------------------
This technique can be used to discover interesting structures in the eigenspace that are either not obvious or even previously unknown. We return to the example of separable eigenfunctions: take $K_1$ to be the adjacency matrix of an Erdős-Renyi random graph $G(100,0.2)$, and $$K_2(x_i, x_j) = e^{-\|x_i - x_j\|^2/\sigma^2}$$ for uniformly sampled points $x_i\in [0,1]$. We take the final kernel $K$ to be the Kronecker product $K = K_2\otimes K_1$, which corresponds to the Cartesian product of Erdős-Renyi grap across $10$ points on a uniformly spaced grid, and build the normalized Laplacian from $K$. Spectral Theory of cartesian product graphs implies that the eigenfunctions are separable across each dimension [@seary]. Figure \[fig:empiricalRandomCrossGrid\] shows the landscape of the eigenfunctions as displayed by the three largest eigenvectors of the eigenfunctions affinity matrix, as well as for the second, third, and forth largest eigenvectors. The method clearly recovers a very interesting structure to the eigenfunctions, with the lowest frequency eigenfunctions exhibiting a different structure from the majority that are organized in a two-dimensional grid. Moreover, this landscape can be used to find interesting connections between eigenfunctions that are otherwise non-obvious.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![$\phi_{10}$ and $\phi_{11}$ eigenvectors of the cartesian product graph. In the landscape, the nearest neighbor of $\phi_{10}$ is $\phi_{11}$. $\phi_9$ is significantly further away in the landscape despite being equally spaced in the spectrum. []{data-label="fig:empiricalTensorHadamardExample"}](code/EmpiricalRandomCrossGrid/eig10ForProduct.png "fig:"){width=".5\textwidth"} ![$\phi_{10}$ and $\phi_{11}$ eigenvectors of the cartesian product graph. In the landscape, the nearest neighbor of $\phi_{10}$ is $\phi_{11}$. $\phi_9$ is significantly further away in the landscape despite being equally spaced in the spectrum. []{data-label="fig:empiricalTensorHadamardExample"}](code/EmpiricalRandomCrossGrid/eig11ForProduct.png "fig:"){width=".5\textwidth"}
![$\phi_{10}$ and $\phi_{11}$ eigenvectors of the cartesian product graph. In the landscape, the nearest neighbor of $\phi_{10}$ is $\phi_{11}$. $\phi_9$ is significantly further away in the landscape despite being equally spaced in the spectrum. []{data-label="fig:empiricalTensorHadamardExample"}](code/EmpiricalRandomCrossGrid/randomoscillationProductAndSmoothed_eig9.png "fig:"){width=".5\textwidth"} ![$\phi_{10}$ and $\phi_{11}$ eigenvectors of the cartesian product graph. In the landscape, the nearest neighbor of $\phi_{10}$ is $\phi_{11}$. $\phi_9$ is significantly further away in the landscape despite being equally spaced in the spectrum. []{data-label="fig:empiricalTensorHadamardExample"}](code/EmpiricalRandomCrossGrid/2oscillationProductAndSmoothed.png "fig:"){width=".5\textwidth"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
To demonstrate this, we look at the nearest neighbors of $\phi_{10}$ in the landscape, and examine their Hadamard product with $\phi_{10}$. $\phi_{11}$ is the nearest neighbor of $\phi_{10}$ in the landscape, and is 75 times closer to $\phi_{10}$ than $\phi_9$ is to $\phi_{10}$ despite the fact that they are about equidistant in the spectrum. Figure \[fig:empiricalTensorHadamardExample\] shows these eigenfunctions and their Hadamard product: we discover that despite $\phi_{10}$ and $\phi_{11}$ being chaotic, their product perfect cuts $Y$ in half.\
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Landscape of the eigenfunctions of the unnormalized graph Laplacian on a random graph, and the $\ell^1$ norm of these ($\ell^2-$normalized) eigenfunctions.[]{data-label="fig:unnormalizedErdosLaplacian"}](code/EmpiricalRandomGraph/empiricalRandomGraphVaryTime1power_unnormalized.png "fig:"){width=".5\textwidth"} ![Landscape of the eigenfunctions of the unnormalized graph Laplacian on a random graph, and the $\ell^1$ norm of these ($\ell^2-$normalized) eigenfunctions.[]{data-label="fig:unnormalizedErdosLaplacian"}](code/EmpiricalRandomGraph/eigvecL1normLabeled.png "fig:"){width=".5\textwidth"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
We also use this technique to look at the differences between the normalized and unnormalized graph Laplacian for and Erdős-Renyi random graph $G(1000,0.2)$. Figure \[fig:unnormalizedErdosLaplacian\] shows the landscape of the eigenfunctions of the unnormalized graph Laplacian. Notice that, in contrast to the normalized Laplacian where only $\phi_0$ stands out, in the unnormalized Laplacian there are several low-freqency and high-frequency eigenfunctions that are clearly separated from the vast majority. This was of curiousity to the authors, and upon further investigation, it was discovered that this deviation corresponded to the fact that some eigenfunctions of the unnormalized Laplacian have slightly smaller $\ell^1$ norm than the vast majority. Indeed, as can be seen in Figure \[fig:unnormalizedErdosLaplacian\], we observe that the $\ell_1$ seems to have a nontrivial limiting distribution over the spectrum. We are not aware of any results in that direction.
[10]{}
J. Ankenmann, Geometry and Analysis of Dual Networks on Questionnaires, PhD thesis, Yale, 2014
J. Ankenmann and W. Leeb, Mixed Hölder matrix discovery via wavelet shrinkage and Calderon-Zygmund decompositions, to appear in Appl. Comp. Harm. Anal.
M. Belkin and P. Niyogi, Convergence of Laplacian eigenmaps. Advances in Neural Information Processing Systems. 2007.
M. Belkin and P. Niyogi, Laplacian eigenmaps for dimensionality reduction and data representation. Neural computation 15.6 (2003): 1373–1396.
M. Bronstein, J. Bruna, Y. LeCun, A. Szlam, P. Vandergheynst, Geometric deep learning: going beyond euclidean data, IEEE Signal Processing Magazine 34 (4), 18–42, (2017).
F. R. Chung, Spectral graph theory. CBMS Regional Conference Series in Mathematics, 92. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 1997.
R. Coifman and M. Gavish, Harmonic analysis of digital data bases. in:Wavelets and multiscale analysis, 161–197, Appl. Numer. Harmon. Anal., Birkhäuser/Springer, New York, 2011.
R. Coifman and M. Gavish, Sampling, denoising and compression of matrices by coherent matrix organization. Appl. Comput. Harmon. Anal. 33 (2012), no. 3, 354–369.
R. Coifman and M. Gavish, Harmonic analysis of databases and matrices. Excursions in harmonic analysis. Volume 1, 297–310, Appl. Numer. Harmon. Anal., Birkhäuser/Springer, New York, 2013.
R. Coifman and M. Gavish, Information integration, organization, and numerical harmonic analysis. Mathematical and computational modeling, 254–271, Pure Appl. Math. (Hoboken), Wiley, Hoboken, NJ, 2015.
R. Coifman and S. Lafon, Diffusion maps. Applied and computational harmonic analysis 21.1 (2006): 5–30.
R. Coifman and W. Leeb. Hölder-Lipschitz norms and their duals on spaces with semigroups, with applications to Earth Mover’s Distance. Journal of Fourier Analysis and Applications, 22(4) (2016) 910–953.
R. Coifman and Y. Meyer, Au dela des operateurs pseudo-differentiels. With an English summary. Asterisque, 57. Societe Mathematique de France, Paris, 1978.
Wavelets. Calderon-Zygmund and multilinear operators. Translated from the 1990 and 1991 French originals by David Salinger. Cambridge Studies in Advanced Mathematics, 48. Cambridge University Press, 1997.
S. Constantin, Diffusion Harmonics and Dual Geometry on Carnot Manifolds, PhD thesis, Yale, 2015.
F. Haake, Quantum signatures of chaos. With a foreword by H. Haken. Second edition. Springer Series in Synergetics. Springer-Verlag, Berlin, 2001
D. Hammond, P. Vandergheynst, R. Gribonval, Wavelets on graphs via spectral graph theory, Applied and Computational Harmonic Analysis 30 (2), 129–150
J. Irion, N. Saito, Hierarchical graph Laplacian eigen transforms. JSIAM Lett. 6 (2014), 21–24.
J. Kruskal, and M. Wish, Multidimensional scaling. Vol. 11. Sage, 1978.
I. Jolliffe, Principal component analysis. New York:Springer-Verlag, 1986.
S. Nonnenmacher, Anatomy of quantum chaotic eigenstates. (English summary) Chaos, 193–238, Prog. Math. Phys., 66, Birkhäuser/Springer, Basel, 2013.
K. Pearson, On lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science 2.11 (1901): 559–572.
N. Perraudin, P. Vandergheynst, Stationary signal processing on graphs, IEEE Transactions on Signal Processing 65 (13), 3462–3477, (2017).
M. Rudelson and R. Vershynin, Delocalization of eigenvectors of random matrices with independent entries. Duke Math. J. 164 (2015), no. 13, 2507–2538.
N. Saito, How can we naturally order and organize graph Laplacian eigenvectors?, arXiv:1801.06782
L. Saul and S. Roweis, Think globally, fit locally: unsupervised learning of low dimensional manifolds. Journal of Machine Learning Research 4, p. 119–155 (2003).
A. Seary, W. Richards, Spectral methods for analyzing and visualizing networks: an introduction. 2003.
D. Shuman, S. Narang, P. Frossard, A. Ortega, P. Vandergheynst, The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains, IEEE Signal Processing Magazine 30 (3), 83–98.
D. Shuman, B. Ricaud, P. Vandergheynst, A windowed graph Fourier transform, Statistical Signal Processing Workshop (SSP), 2012 IEEE, 133–136, (2012).
D. Shuman, M. Faraji, P. Vandergheynst, multiscale pyramid transform for graph signals, IEEE Transactions on Signal Processing 64 (8), 2119–2134, (2016).
S. Steinerberger, Directional Poincare Inequalities along Mixing Flows, Arkiv för Matematik 54 , 555–569, 2016
S. Steinerberger, On the Spectral Resolution of Products of Laplacian Eigenfunctions, arXiv:1711.09826
T. Tao, Nonlinear dispersive equations. Local and global analysis. CBMS Regional Conference Series in Mathematics, 106. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 2006.
A. Zygmund, On Fourier coefficients and transforms of functions of two variables. Studia Math. 50 (1974), 189–201.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We give an exact self-consistent operator description of the spin and orbital angular momenta, position, and spin-orbit interactions of nonparaxial light in free space. Both quantum-operator formalism and classical energy-flow approach are presented. We apply the general theory to symmetric and asymmetric Bessel beams exhibiting spin- and orbital-dependent intensity profiles. The exact wave solutions are clearly interpreted in terms of the Berry phases, quantization of caustics, and Hall effects of light, which can be readily observed experimentally.'
author:
- 'Konstantin Y. Bliokh'
- 'Miguel A. Alonso'
- 'Elena A. Ostrovskaya'
- Andrea Aiello
title: 'Angular Momenta and Spin-Orbit Interaction of Nonparaxial Light in Free Space'
---
Introduction
============
The problem of the identification of the spin and orbital parts of the angular momentum (AM) of an electromagnetic wave has a long history and has posed fundamental difficulties in both quantum electrodynamics and classical optics [@AB; @EN; @BA; @B].
It is known that the photon AM operator in the momentum (plane-wave) representation has the form [@AB]: $$\label{eqn:1}
\mathbf{\hat J} = - i\left( {\mathbf{k} \times \partial _\mathbf{k} } \right) + \mathbf{\hat S} \equiv \mathbf{\hat L} + \mathbf{\hat S}~.$$ Here the orbital part is $\mathbf{\hat L}=\mathbf{\hat r} \times \mathbf{\hat p}$ ($\mathbf{\hat p}={\bf k}$, $\mathbf{\hat r}=i{\partial_{\bf k}}$, ${\bf k}$ is the wave vector, and we use units $\hbar=c=1$), whereas $\mathbf{\hat S}$ is the spin-1 operator given by $3\times 3$ matrices $(\hat{S}_a)_{ij}=-i\epsilon_{aij}$ ($\epsilon_{aij}$ is the Levi-Civita symbol) that act on the Cartesian components of the wave electric field. Canonical orbital AM (OAM) and spin AM (SAM) operators, $\mathbf{\hat L}$ and $\mathbf{\hat S}$, satisfy $so(3)$ algebra and generate rotations in spatial and polarization degrees of freedom, respectively. However, “*the separation of the total AM into orbital and spin parts has restricted physical meaning. ... States with definite values of OAM and SAM do not satisfy the condition of transversality in the general case.*” [@AB]. In 1994, Van Enk and Nienhuis put forward an alternative, non-canonical AM separation, where the modified spin and orbital parts are measurable and consistent with the transversality of the wave, although they are not generators of rotations [@EN].
In classical optics, the two parts of Eq. (\[eqn:1\]) can be unambiguously associated with the OAM and SAM for *paraxial* light, where the eigenmodes of $\hat{L}_z=-i\partial_{\phi}$ ($\phi$ is the azimuthal angle in ${\bf k}$ space) and $\hat{S}_z$ are circularly polarized vortex beams with the corresponding quantum numbers $\ell = 0,\pm 1,\pm 2,...$ (topological charge of the vortex $e^{i\ell\phi}$) and $\sigma=\pm 1$ (helicity) [@OAM]. However, for *non-paraxial* fields the identification of OAM and SAM meets serious difficulties [@EN; @BA; @B]. Calculations based on the recently suggested division of the Poynting energy flow into spin and orbital parts [@BA; @Berry2009; @Li; @Beksh] show that the non-paraxial correction to the OAM is proportional to $\sigma$ rather than to $\ell$ [@BA; @Li]. This resulted in the conclusion that “*in the general non-paraxial case there is no simple separation into $\ell$-dependent orbital and $\sigma$-dependent spin component of AM*” [@BA].
In this paper we re-examine the problem and give an exact self-consistent solution in terms of both the fundamental photon operators and classical energy flows. The identification of the well-defined measurable OAM and SAM of light is shown to be closely related to the analogous problem for the *position* of localized photons [@Pryce; @Photon; @Hawton]. Our approach generalizes and unifies previously disjointed results: (i) non-canonical OAM and SAM operators obtained earlier for the second-quantized fields [@EN]; (ii) non-commutative photon position operator and Berry monopole field in momentum space [@Pryce; @Photon]; and (iii) separation of the spin and orbital parts of Poynting energy flows [@Berry2009; @Li; @Beksh]. We find that the $\sigma$-dependent non-paraxial part of the OAM arises from Berry-phase terms describing the *spin-orbit interaction* (SOI) of light. A similar effect occurs dynamically upon *spin-to-orbital AM conversion* in focusing and scattering of polarized light [@Beksh; @AMC1; @AMC2; @Oscar]. Other manifestations of the SOI are the *spin* [@SHE1; @SHE2; @Half; @Aiello] and *orbital* [@OHE1; @OHE2; @OHE3; @OHE4] *Hall effects of light* (i.e., $\ell$- and $\sigma$-dependent transverse shifts of the field center of gravity) that are described by our position operator and take place even in free space [@Aiello; @OHE4]. We apply the general theory to vector Bessel beams, for which the fundamental operators manifest themselves in immediately observable $\ell$- and $\sigma$-dependent intensity distributions. The exact wave results are also explained in terms of the underlying geometrical-optics rays and caustics.
Operator formalism
==================
We consider an electromagnetic field in free space, characterized by its plane-wave electric-field spectrum ${\bf \tilde E}\left( {\bf k} \right)$ without evanescent modes. The SOI of light originates from the transversality constraint, ${\bf k} \cdot {\bf \tilde E} = 0$, which couples polarization to the wave vector and reduces the full 3D vector space of the electric field components to the 2D subspace of the components tangential to a sphere of directions in ${\bf k}$-space. The operators ${\bf \hat L}$ and ${\bf \hat S}$ do not keep this subspace invariant, i.e., their action on a transverse mode results in a non-zero longitudinal component [@AB; @EN]. However, this subspace is invariant for the total AM operator ${\bf \hat J}$, and one can divide it into two parts consistent with the transversality condition: $$\label{eqn:2}
{\bf \hat J} = {\bf \hat L'} + {\bf \hat S'},~{\bf \hat L'} = {\bf \hat L} - {\bm \kappa } \times \left( {{\bm \kappa } \times {\bf \hat S}} \right),~{\bf \hat S'} = {\bm \kappa }\left( {{\bm \kappa } \cdot {\bf \hat S}} \right),$$ where ${\bm \kappa } = {\bf k}/k$ and the modified OAM and SAM operators ${\bf \hat L'}$ and ${\bf \hat S'}$ can be regarded as projections of the operators ${\bf \hat L}$ and ${\bf \hat S}$ onto the transversality subspace [@EN].
The modified SAM operator ${\bf \hat S'}$ is proportional to the helicity operator $\hat \sigma = {\bm \kappa } \cdot {\bf \hat S}$, whereas the OAM operator can be written as ${\bf \hat L'} = {\bf \hat r'} \times {\bf k}$ with $$\label{eqn:3}
{\bf \hat r'} = {\bf \hat r} + \frac{{\bf k} \times {\bf \hat S}}{k^2 } = i\partial _{\bf k} + \frac{{\bf k} \times {\bf \hat S}}{k^2 }~.$$ The modified position operator (\[eqn:3\]) has been considered in the context of photon localization and Berry phase [@Pryce; @Photon; @Hawton]. It describes the observable center of gravity of the field and brings about the space non-commutativity with the monopole term in ${\bf k}$-space: $$\label{eqn:4}
\left[ {\hat r'_i ,\hat r'_j } \right] = - i\epsilon _{ijl} \hat \sigma \frac{k_l }{k^3 }~.$$ The operators ${\bf \hat L'}$ and ${\bf \hat S'}$ do not satisfy the $so(3)$ AM algebra and have unusual commutation relations: $$\label{eqn:5}
\left[ {\hat S'_i ,\hat S'_j } \right] = 0,~\left[ {\hat L'_i ,\hat L'_j } \right] = i\epsilon _{ijl} (\hat L'_l -\hat S'_l),~
\left[ {\hat L'_i ,\hat S'_j } \right] = i\epsilon _{ijl} \hat S'_l.$$ At the same time, the modified operators transform as vectors under rotations: $\left[ {\hat J_i ,\hat O'_j } \right] = i\epsilon _{ijl} \hat O'_l$, ${\bf \hat O'} = {\bf \hat L'}$, ${\bf \hat S'}$, and ${\bf \hat r'}$. The commutation relations (\[eqn:5\]) unveil the similarity of operators ${\bf \hat L'}$ and ${\bf \hat S'}$ to those obtained for the second-quantized fields in [@EN]. Although they do not generate rotations, it is suggested that they do correspond to observable continuous values of the OAM and SAM of a non-paraxial transverse field [@EN].
Remarkably, in the helicity representation the matrix components of the operators (\[eqn:2\]) and (\[eqn:3\]) become diagonal. We introduce spherical coordinates $\left(\theta,\phi,k\right)$ with basic vectors $\left( {{\bf e}_\theta ,{\bf e}_\phi ,{\bm \kappa }} \right)$ in ${\bf k}$-space, so that the free electric field has only $\left( {{\bf e}_\theta ,{\bf e}_\phi} \right)$-components. The helicity basis of circular polarizations corresponds to the basic vectors ${\bf e}^ \pm = e^{\pm im\phi } \left({{\bf e}_{\theta} \pm i{\bf e}_{\phi} }\right)/{\sqrt 2 }$, where $e^{\pm im\phi }$ is an arbitrary gauge factor [@Hawton]. Transition from the global Cartesian field components $( {\tilde E_x ,\tilde E_y ,\tilde E_z } )^T$ to the helicity amplitudes $( {\tilde E^ + ,\tilde E^ - ,\tilde E_\parallel } )^T$ is realized via the local unitary transformation $\hat U\left( {\theta ,\phi } \right) = \hat R_z \left( { - \phi } \right)\hat R_y \left( { - \theta } \right)\hat R_z \left( {m\phi } \right)\hat V $, where $\hat R_a \left( \alpha \right) = e^{i\alpha \hat S_a }$ is the matrix of rotation by an angle $\alpha$ with respect to the $a$-axis, whereas $\hat V$ is the constant transformation from linear- to circular-polarization basis. Making the transformation of operators (\[eqn:2\]) and (\[eqn:3\]) to the helicity basis, ${\bf \hat O'} \to \hat U^\dag {\bf \hat O'}\hat U$, we obtain: $$\begin{aligned}
\label{eqn:6}
{\bf \hat S'} = {\bm \kappa }\hat \sigma~,~~{\bf \hat L'} =
- i{\bf k} \times \partial _{\bf k} - {\bf \hat A}_B \times {\bf k}~,\\
\label{eqn:7}
{\bf \hat r'} = i\partial _{\bf k} - {\bf \hat A}_B~,~~{\bf \hat p} = {\bf k}~,~~{\hat w}=\omega\end{aligned}$$ Here we included the momentum and energy operators, ${\bf \hat p}$ and ${\hat w}$ (which are unaffected by the transformations), $\omega$ is the frequency, the helicity is diagonal: $\hat \sigma = {\mathop{\rm diag}\nolimits} \left( {1, - 1,0} \right)$, and $$\label{eqn:8}
{\bf \hat A}_B = - \frac{{{\bf k} \times {\bf \hat S}}}{{k^2 }} - i\hat U^\dag \partial _{\bf k} \hat U = \frac{{m - \cos \theta }}{{k\sin \theta }}\hat \sigma {\bf e}_\phi$$ is the Berry gauge field (connection) which corresponds to the monopole curvature ${\bf \hat F}_B = \partial _{\bf k} \times {\bf \hat A}_B = \hat \sigma \,{\bf k}/k^3$ [@Photon; @Hawton]. Hereafter we choose the gauge $m=1$, which corresponds to the absence of the phase singularity (Dirac string) along the positive $z$-axis in Eq. (\[eqn:8\]) [@Hawton], allowing a smooth transition to the paraxial case, $\theta \to 0$.
It is worth noticing that the transformation to the helicity basis is associated with the transition to the local coordinate frame with the $z$-axis attached to the current ${\bf k}$-vector, which induces pure gauge Coriolis-type potential ${\bf \hat A}= - i\hat U^\dag \partial _{\bf k} \hat U$, i.e., $( {{\bf \hat A} } )_{ij} = - i{\bf e}_i^* \cdot \left( {\partial _{\bf k} } \right){\bf e}_j $, where $\mathbf{e}_{1,2,3} \equiv \left( {\mathbf{e}^ + ,\mathbf{e}^ - ,\bm{\kappa }} \right)$ [@Photon; @Coriolis]. At the same time, non-canonical operators and commutation relations (\[eqn:2\])–(\[eqn:5\]) essentially owe their origin to the projection onto the transversality subspace, which is equivalent to the diagonalization of the potential ${\bf \hat A}$ [@Photon; @Coriolis]: ${\bf \hat A}_B={\rm dg}{\bf \hat A}$, i.e., $$\label{eqn:9}
( {{\bf \hat A}_B } )_{ij} = - i{\bf e}_i^* \cdot \left( {\partial _{\bf k} } \right){\bf e}_j \delta _{ij}~.$$ While such diagonalization (which uncouples the two helicity components) is an adiabatic approximation for a nearly-transverse paraxial wave beam propagating in an inhomogeneous medium [@Coriolis; @Berry1987], it is exact for transverse plane waves in free space where the helicities are truly independent.
The measurable expectation (mean) values of the OAM, SAM, coordinate, momentum, and energy obtained from the diagonal operators (\[eqn:6\])–(\[eqn:8\]) can be written as $$\begin{aligned}
\label{eqn:10}
{\bf S} =
\left\langle {\tilde E^\sigma } \right| \sigma {\bm \kappa }\left| {\tilde E^\sigma } \right\rangle,~\\
\label{eqn:11}
{\bf L} =
\left\langle {\tilde E^\sigma } \right|{\bf \hat L}\left| {\tilde E^\sigma } \right\rangle - \left\langle {\tilde E^\sigma } \right| \sigma {\bf A}_B \times {\bf k} \left| {\tilde E^\sigma } \right\rangle,\\
\label{eqn:12}
{\bf R} =
\left\langle {\tilde E^\sigma } \right|i\partial _{\bf k} \left| {\tilde E^\sigma } \right\rangle - \left\langle {\tilde E^\sigma } \right| \sigma {\bf A}_B \left| {\tilde E^\sigma } \right\rangle,\\
\label{eqn:13}
{\bf P}= \left\langle {\tilde E^\sigma } \right|{\bf k}\left| {\tilde E^\sigma } \right\rangle,~
W= \left\langle {\tilde E^\sigma } \right|\omega\left| {\tilde E^\sigma } \right\rangle.\end{aligned}$$ Here ${\bf A}_B = {\bf e}_\phi k^{ - 1} \left( {1 - \cos \theta } \right)/\sin \theta$, convolution implies summation over $\sigma=\pm 1$ and integration in the ${\bf k}$-space, and we assume normalization $N=\left\langle {\tilde E^\sigma } \right|\left. {\tilde E^\sigma } \right\rangle=1$ (see Appendix for details). While the SAM is purely *intrinsic* (origin-independent), the OAM, in general, has both intrinsic and *extrinsic* contributions [@Extrinsic]: $$\label{eqn:14}
{\bf L}^{\rm ext}={\bf R}\times {\bf P}~,~~{\bf L}^{\rm int}={\bf L} - {\bf L}^{\rm ext}~.$$ Equations (\[eqn:10\])–(\[eqn:14\]) contain all the main observable results related to the AM and SOI of light. First, the $\sigma$-dependent non-paraxial Berry-phase term in ${\bf L}$ should be associated with the *spin-to-orbit AM conversion* [@Li; @Beksh; @AMC1; @AMC2; @Oscar]. Particular cases of this term have appeared in [@BA; @Li; @Beksh]. Second, the *orbital* [@OHE1; @OHE2; @OHE3; @OHE4] and *spin* [@SHE1; @SHE2; @Half; @Aiello] *Hall effects of light* are described by the two terms in the position of the center of gravity, Eq. (\[eqn:12\]). Indeed, for a symmetric vortex beam propagating along the $z$-axis, the transverse coordinates of the center of gravity vanish, $(X,Y)=0$, after integration over $\phi$, but any asymmetry of the field distribution along, say, the $x$-axis immediately causes an $\ell$- and $\sigma$-dependent shift along the orthogonal $y$-axis, $Y\neq 0$ together with tilt $P_x\neq 0$ (see the example in Section IV).
We emphasize that our results (\[eqn:6\])–(\[eqn:14\]) are *exact* and no approximations were made. They are equivalent to application of the canonical operators ${\bf \hat L}$, ${\bf \hat S}$, and ${\bf \hat r}$ to the laboratory-frame field components $( {\tilde E_x ,\tilde E_y ,\tilde E_z } )^T$ supplied with the transversality condition.
Energy flow approach
====================
Remarkably, the same results, Eq. (\[eqn:10\])–(\[eqn:14\]), can be derived from an approach based on the separation of the spin and orbital parts in the Poynting energy flow [@Berry2009; @Li]. Let us consider a monochromatic beam-like field propagating in the positive $z$-direction. Pecularities of the $(2+1)D$ formalism for such a problem are discussed in the Appendix, Eqs. (\[eqn:A6\])–(\[eqn:A9\]).
The transverse center of gravity (\[eqn:A9\]) obtained in the momentum representation from operator equation (\[eqn:12\]) can be equally derived from the traditional coordinate-representation definition $$\label{eqn:31}
\mathbf{R}_ \bot \left( z \right) = \frac{1}{g^2} \int {\mathbf{r}_ \bot \left| {\mathbf{E}\left( {\mathbf{r}_ \bot ,z} \right)} \right|^2 \,} d^2 \mathbf{r}_ \bot~,$$ where ${\bf r}_{\bot}=(x,y)$, $g=\sqrt{2\omega/\varepsilon_0}$, and we used normalization $\int { \left| {\mathbf{E}\left( {\mathbf{r}_ \bot ,z} \right)} \right|^2 \,} d^2 \mathbf{r}_ \bot = g^2$ corresponding to $N=1$ (see Appendix). Substituting here Fourier representation (\[eqn:A6\]) with the helicity-basis expansion (\[eqn:A2\]), and using $\int {\mathbf{r}_ \bot e^{i\mathbf{k}_ \bot \cdot \mathbf{r}_ \bot } d^2 \mathbf{r}_ \bot } = \left( {2\pi } \right)^2 \delta ^2 \left( {\mathbf{k}_ \bot } \right)\partial _{\mathbf{k}_ \bot }$ (${\bf k}_\bot =(k_x , k_y )$) together with expression (\[eqn:9\]) for the Berry connection, we arrive at Eq. (\[eqn:A9\]) which is equivalent to Eq. (\[eqn:12\]).
To derive the linear and angular momenta of the field, we use the Poynting vector which determines the momentum density (energy flow) [@Born]: $$\label{eqn:32}
{\bm{\pi}} = \frac{1}{g^2} \operatorname{Im} \left[ {{{\mathbf{E}}^*} \times \left( {\nabla \times {\mathbf{E}}} \right)} \right]~.$$ Substituting here the Fourier decomposition (\[eqn:A6\]), we obtain $$\label{eqn:33}
\bm{\pi} = {\mathop{\rm Re}\nolimits} \iint {e^{i\left( {{\bf k} - {\bf k'}} \right) \cdot {\bf r}} \tilde E^{\sigma\prime *} \tilde E^\sigma {\bf e}^{\sigma\prime *} \times \left( {{\bf k} \times {\bf e}^\sigma } \right)} \frac{d^2 {\bf k}^{\prime}_\bot}{2\pi} {\frac{d^2 {\bf k}_\bot}{2\pi} },$$ where $\tilde E^{\sigma\prime} = \tilde E^\sigma \left( {{\bf k'}} \right)$, ${\bf e}^{\sigma\prime} = {\bf e}^\sigma \left( {{\bf k'}} \right)$, and summation over $\sigma =\pm 1$ is implied hereinafter. After some calculations the total momentum density (\[eqn:33\]) can be decomposed into the orbital and spin parts as suggested in [@Berry2009; @Li], $\bm{\pi}=\bm{\pi}^{\rm o}+\bm{\pi}^{\rm s}$: $$\label{eqn:34}
{\bm \pi}^{\rm o} = \iint {e^{i {\bf K}^{-} \cdot {\bf r}} {\tilde E}^{\sigma} {\tilde E}^{\sigma\prime *} \left( {\bf e}^{\sigma} \cdot {\bf e}^{\sigma\prime *} \right) \frac{{\bf K}^+}{2}}
\frac{{d^2 {\mathbf{k}}^{\prime}_\bot }}
{{2\pi }} \frac{{d^2{\mathbf{k}}_\bot}}
{{2\pi }},$$ $$\label{eqn:35}
{\bm \pi}^{\rm s} = \iint {
e^{i {\bf K}^{-} \cdot {\bf r}}{\tilde E}^{\sigma} {\tilde E}^{\sigma\prime *} \left( {\bf e}^{\sigma} \times {\bf e}^{\sigma\prime *} \right) \times \frac{{\bf K}^-}{2}} \frac{d^2 {\bf k}^{\prime}_\bot }
{2\pi } \frac{d^2 {\bf k}_\bot }{2\pi },$$ where ${\bf K}^{\pm}={\mathbf{k}} \pm {\mathbf{k'}}$. Note that in the decomposition (\[eqn:33\])–(\[eqn:35\]) the two helicity components are exactly separated without an interference term [@Berry2009]. Also, the “electro-magnetic democracy” discussed by Berry [@Berry2009] is accomodated, because switching to magnetic-field plane-wave helicity amplitudes, ${{\tilde E}^\sigma } \rightarrow {{\tilde H}^\sigma } = -i\sigma {{\tilde E}^\sigma }$, keeps Eqs. (\[eqn:33\])–(\[eqn:35\]) invariant.
The linear momentum of the beam (per unit $z$-length) is given by the 2D space integration of the momentum densities (\[eqn:33\])–(\[eqn:35\]): $\mathbf{P} = \int {\bm{\pi }\,d^2 \mathbf{r}_ \bot }$. In doing so, we find that the spin momentum density makes no contribution to the linear momentum [@Li]: ${\bf P}^{\rm s} = \int {\bm{\pi }^{\rm s} \,d^2 \mathbf{r}_ \bot = 0}$, while the orbital contribution yields $$\label{eqn:36}
\mathbf{P}^{\rm o} = \mathbf{P} = \int {\bm{\pi }^{\rm o} d^2 \mathbf{r}_ \bot = \int {\mathbf{k}\left| {\tilde E^\sigma } \right|^2 } d^2 \mathbf{k}_ \bot }~,$$ which obviously coincides with Eq. (\[eqn:13\]).
The SAM and OAM of the beam (per unit $z$-length) can be obtained by the 2D space integration of their densities, i.e.: $$\label{eqn:37}
\mathbf{L} = \int {\mathbf{r} \times \bm{\pi }^{\rm o} d^2 \mathbf{r}_ \bot }~,~~
\mathbf{S} = \int {\mathbf{r} \times \bm{\pi }^{\rm s} d^2 \mathbf{r}_ \bot }~.$$ Substituting Eqs. (\[eqn:34\]) and (\[eqn:35\]) into Eqs. (\[eqn:37\]) and employing properties of Fourier integrals, we arrive at $$\label{eqn:38}
\mathbf{S} = i\int { ({\bf e}^{\sigma}\times {\bf e}^{\sigma *}) \left| {\tilde E^\sigma } \right|^2 d^2 \mathbf{k}_ \bot }
= \int {\sigma \bm{\kappa }\left| {\tilde E^\sigma } \right|^2 d^2 \mathbf{k}_ \bot }~,$$
$$\label{eqn:39}
\mathbf{L} = \int {\tilde E^{\sigma *} \mathbf{e}^{\sigma *} \cdot \left( { - i\mathbf{k} \times \partial _\mathbf{k} } \right)\tilde E^\sigma \mathbf{e}^\sigma d^2 \mathbf{k}_ \bot }
= \int {\tilde E^{\sigma *} \left( -i{\mathbf{k} \times \partial _\mathbf{k} } \right)\tilde E^\sigma d^2 \mathbf{k}_{\bot} } - \int { \sigma (\mathbf{A}_B \times \mathbf{k}) \left| {\tilde E^\sigma } \right|^2 d^2 \mathbf{k}_{\bot} },$$
where identity ${\bf e}^{\sigma *}\times {\bf e}^{\sigma}=i\sigma{\bm \kappa}$ and Eq. (\[eqn:9\]) were used. Clearly, the values of SAM and OAM, Eqs. (\[eqn:38\]) and (\[eqn:39\]), derived from the Poynting energy flows are in perfect agreement with our operator formalism, Eqs. (\[eqn:10\]) and (\[eqn:11\]) (see also Appendix).
![(Color online) (a) Bessel-beam distribution (\[eqn:11\]) on the sphere in ${\bf k}$-space with the azimuthal phase $2\pi\ell$. (b) Cylindrical caustic in the real space, an example of the closed orbit $\Gamma$ on it, and the corresponding GO rays tangent to the caustic. Scalar phases are color-coded for $\ell=-1$, $\theta_0=\pi/4$. Points $A$ and $A'$ on the caustic are connected by two paths: the straight line and the Poynting-flow helix. The phase matching yields the phase difference $2\pi\ell$ between the paths and quantization of the caustic radius. For circularly-polarized waves, the helical path brings about an additional Berry phase $\sigma\Phi_B$, Eq. (\[eqn:45\]).[]{data-label="fig1"}](Fig1_3jpeg.eps){width="8.6cm"}
Application to Bessel beams
===========================
Importantly, our theory has a number of directly observable consequences. As the simplest example we take non-paraxial vector Bessel-beam solutions which are eigenmodes of $\hat J_z$ constructed from plane waves with well-defined helicity $\sigma$ (cf. [@EN; @Li; @Hawton; @Bessel]). The angular spectrum of such beams is $$\label{eqn:41}
{\bf \tilde E}_\ell ^\sigma
%\left( {\theta ,\phi } \right)
= {\bf e}^\sigma \left( {\theta ,\phi } \right) \tilde E_\ell ^\sigma \left( {\theta ,\phi } \right),~{\tilde E}_\ell ^\sigma = A^\sigma \delta \left( {\theta - \theta _0 } \right)e^{i\ell \phi },$$ where $A^\sigma$ is a constant amplitude, $\theta_0$ is the polar angle of conical distribution of the ${\bf k}$-vectors, Fig. 1(a), and no summation over $\sigma$ is implied here.
For the $z$-components of OAM and SAM, Eqs. (\[eqn:10\]) and (\[eqn:11\]), or (\[eqn:38\]) and (\[eqn:39\]), of a superposition of $\sigma =\pm 1$ beams (\[eqn:41\]) we obtain [@R1]: $$\label{eqn:42}
{L_z } = \ell + \bar \sigma \frac{\Phi _B }{2\pi },~
{S_z } = \bar \sigma \left( {1 - \frac{\Phi _B }{2\pi }} \right),~
{J_z } = \ell + \bar \sigma.$$ Here $\bar \sigma = ( {\left| {A^ + } \right|^2 - \left| {A^ - } \right|^2 } )/( {\left| {A^ + } \right|^2 + \left| {A^ - } \right|^2 } )$ is the averaged helicity and $$\label{eqn:43}
\Phi _B = \oint\limits_{\rm C} {{\bf A}_B \cdot d{\bf k}} = 2\pi \left( {1 - \cos \theta _0 } \right)$$ is the Berry phase associated with the contour ${\rm C}=\{\theta=\theta_0, \phi\in(0,2\pi)\}$ formed by the ${\bf k}$-vectors distribution on the sphere of directions, Fig. 1(a) [@GP]. The Berry phase is equal to the flux of the monopole field $\mathbf{F}_B = \partial _\mathbf{k} \times \mathbf{A}_B = \mathbf{k}/k^3$ through the area of the ${\bf k}$-space sphere bounded by the contour ${\rm C}$. In this manner, the $\bar\sigma$-dependent term in $L_z$ represents a monopole-flux contribution to the OAM, cf. Eq. (87) in [@SUSY]. In the paraxial limit the Berry-phase terms vanish as $\Phi _B \simeq \pi \theta _0^2 \to 0$. The values (\[eqn:42\]) evidence an apparent partial conversion from SAM to OAM in non-paraxial light with the total AM being constant [@BA; @Li], akin to the spin-to-orbit AM conversion upon focusing of polarized light [@Li; @Beksh; @AMC1; @AMC2; @Oscar]. Indeed, in the Richards-Wolf approximation [@RW], the focusing represents a geometric conical redirection of partial plane waves with their helicity being conserved. It is described exactly by the same transformation operator $\hat U (\theta,\phi)$ that describes transition to the helicity basis [@Oscar].
![(Color online) (a) Intensity distributions, Eq. (\[eqn:45\]), marked by quantum numbers $(\ell,\sigma)$ for Bessel beams with $\theta_0=3\pi /8$. The spin-dependent profiles are shown for $\ell=4$ with $\sigma=-1$, 0 (scalar case), and 1. Dashed circles indicate the GO caustics (\[eqn:45\]). (b) Radial intensity profiles of the scalar ($\sigma=0$) or paraxial ($\theta_0\rightarrow0$) Bessel beams $I_{|\ell|}=J_{\ell}^2(\xi)$. (c) The SOI splitting of the profile of the polarized non-paraxial Bessel beam $I_{\pm4}^{\sigma}(\xi)$, Eq. (\[eqn:44\]), at different values of $\theta_0$; vertical lines indicate GO caustics (\[eqn:45\]), cf. (a). (d) The GO caustics (\[eqn:45\]) marked by ${\rm sgn}(\ell)(\ell,\sigma)$ as functions of $\theta_0$. (e) The SOI splitting of the maxima of intensity (\[eqn:44\]) \[cf. (c)\] as dependent on $\theta_0$, approaching the GO limit (\[eqn:45\]) at $|\ell| \gg 1$.[]{data-label="fig2"}](Fig2_6jpeg.eps){width="8.6cm"}
Simultaneously with a $\sigma$-dependent OAM, the non-paraxial fields exhibit $\sigma$-dependent intensity distributions related to the modified position operator. The real-space field of the circularly-polarized Bessel beam, calculated via the Fourier transformation (\[eqn:A6\]) of Eq. (\[eqn:41\]), is $$\mathbf{E}_\ell ^\sigma \propto A^\sigma \left( {\begin{array}{*{20}c}
{\frac{{1 + \sigma }}
{2}J_\ell \left( \xi \right) - \sigma b\,e^{i\left( {\sigma - 1} \right)\varphi } J_{\ell + \sigma - 1} \left( \xi \right)} \\
{\frac{{1 - \sigma }}
{2}J_\ell \left( \xi \right) + \sigma b\,e^{i\left( {\sigma + 1} \right)\varphi } J_{\ell + \sigma + 1} \left( \xi \right)} \\
{ - i\sigma \sqrt {2ab} e^{i\sigma \varphi } J_{\ell + \sigma } \left( \xi \right)} \\
\end{array} } \right)e^{ik_{\parallel} z + i\ell \varphi },$$ where $\left( {\rho ,\varphi ,z} \right)$ are the cylindrical coordinates in real space, $a = \cos ^2 \left( {\theta _0 /2} \right)$, $b = \sin ^2 \left( {\theta _0 /2} \right)$, $\xi=k_\bot\rho$, $k_ \bot = k\sin \theta _0$, $k_{\parallel} = k\cos \theta _0$, $J_n (\xi)$ are the Bessel functions of the first kind, and the field components are written in the basis $\left(\frac{{\bf e}_x+i{\bf e}_y}{\sqrt{2}},\frac{{\bf e}_x-i{\bf e}_y}{\sqrt{2}},{\bf e}_z\right)$. This field has a cylindrically symmetric intensity distribution, $I_\ell ^\sigma = \left| {\mathbf{E}_\ell ^\sigma } \right|^2$, given by $$\label{eqn:44}
I_\ell ^\sigma
\propto \left| {A^\sigma } \right|^2 \left[ {a^2 J_\ell^2 \left( \xi \right) + b^2 J_{\ell + 2\sigma }^2 \left( \xi \right) + 2abJ_{\ell + \sigma }^2 \left( \xi \right)} \right],$$ Below we show that the polarization-dependent intensity distributions (\[eqn:44\]) (see Fig. 2(a)) signify the SOI of light.
The $\ell$- and $\sigma$-dependence of the radial intensity profile (\[eqn:44\]) can be explained via a geometrical-optics (GO) ray picture and the *quantization of caustic* underlying the maximum of the intensity. The rays associated with a Bessel beam are those that form an angle $\theta_0$ with the $z$-axis and touch a cylindrical caustic of radius $\rho = R_\ell ^\sigma$ [@Berry2008], Fig. 1. The quantization condition for a closed orbit $\Gamma$ is $\oint\limits_\Gamma {\bf k} \cdot d{\bf r} = 2\pi \ell$. Using the underlying position (\[eqn:7\]), ${\bf r'}^\sigma = {\bf r} - \sigma {\bf A}_B$, we observe that the Berry phase changes the effective optical length of a closed orbit on the cylindrical surface, Fig 1(b). For the orbit $\Gamma = \left\{ {\rho = R_\ell ^\sigma ,\varphi \in \left( {0,2\pi } \right)} \right\}$ it becomes $k_ \bot \left[ {2\pi \, {\rm sgn}(\ell) R_\ell ^\sigma - \sigma {\kern 1pt} \Phi _B } \right]$, which yields $$\label{eqn:45}
k_ \bot R_\ell ^\sigma = \left| \ell + \sigma {\kern 1pt} \frac{\Phi _B }{2\pi } \right|~.$$ Similar Berry-phase effects appear in quantum quantization problems [@Quantization], e.g., the half-integer Hall effect in graphene [@Graphene]. Note also the exact correspondence between the GO caustic (\[eqn:45\]) and the wave OAM (\[eqn:42\]), $|L_z|=k_{\bot}R_\ell ^\sigma$, which reflects the OAM interpretation as ${\bf r}\times{\bf k}$ for the rays. Figure 2 shows $\ell$- and $\sigma$-dependent intensity distributions (\[eqn:44\]) of the Bessel beams vs. the GO caustics (\[eqn:45\]). Spin-dependent splitting of caustics and intensity maxima are the optical analogues of the fine spin-orbit splitting of levels in quantum systems. The $\sigma$-dependence in radial distributions of non-paraxial vortex fields can be observed experimentally by tightly focusing paraxial light with different polarizations, cf. [@Gorod].
![(Color online) Transverse intensity distributions at $z=0$ of the asymmetric Bessel beams with $\delta=\pi/2$ and $\theta_0 =3\pi/8$ marked by quantum numbers $(\ell,\sigma)$. Dashed lines indicate the $\ell$- and $\sigma$-dependent transverse shifts of the centers of gravity, Eq. (\[eqn:46\]), i.e., orbital and spin Hall effects of light in free space. We have verified that the centers of gravity calculated numerically from the intensity distributions and theoretically from Eq. (\[eqn:46\]) coincide.[]{data-label="fig3"}](Fig3_2jpeg.eps){width="8.6cm"}
Finally, we consider the Hall effects of light, which are described by the position (\[eqn:12\]). For this purpose we break the symmetry of the Bessel beams (\[eqn:41\]) along the $k_x$-axis and assume that the plane-wave components are distributed in the range $\phi\in (-\delta,\delta)$, $0<\delta <\pi$. (Such truncated azimuthal distributions can be generated via focusing by the corresponding sector of a lens [@Half].) Substitution of this distribution in Eqs. (\[eqn:12\]) and (\[eqn:13\]), or (\[eqn:36\]) and (\[eqn:A9\]), shows mutually orthogonal tilt and displacement of the beam: $$\label{eqn:46}
P_x =\gamma k_{\bot}~,~~k_\bot Y_{\ell}^{\sigma}= - \gamma \left( \ell + \bar \sigma \frac{{\Phi _B }}{{2\pi }}\right)~.$$ Here $\gamma = (\sin\delta)/\delta$, $X(z)=z P_x/P_z$ ($P_z=k_\parallel$), and the second expression (\[eqn:46\]) closely resembles Eqs. (\[eqn:42\]) and (\[eqn:45\]). The $\ell$- and $\sigma$-dependent parts of the transverse shift of the center of gravity of the beam, $Y_{\ell}^{\sigma}$, describe the orbital and spin Hall effects of light in free space, Fig. 3. A related spin-Hall effect has been observed upon focusing of light with a “half-lens” ($\ell=0$ for $\delta=\pi/2$) [@Half], whereas the orbital-Hall effect can be measured in a similar manner by focusing vortex beams with broken symmetry. The values of $L_z$ and $S_z$ for the asymmetric beam are given by the same Eq. (\[eqn:42\]), but in this case the OAM has an extrinsic contribution , $L_z^{\rm ext}=-P_x Y_{\ell}^{\sigma}$ (\[eqn:14\]): $$\label{eqn:47}
L_z^{\rm ext}=\gamma^2 L_z,~L_z^{\rm int}=(1-\gamma^2) L_z.$$ Hence, the Hall effects of light can be interpreted as an *intrinsic-to-extrinsic OAM conversion* [@SHE1; @OHE1] which is also accompanied by generation of a transverse OAM component $L_x^{\rm ext}=P_z Y_{\ell}^{\sigma}=- \gamma\cot\theta_0 L_z$ [@Aiello]. The total conversion is achieved at $\delta\rightarrow 0$, $\gamma\rightarrow 1$.
Conclusion
==========
To summarize, we have revisited the problem of the identification of the spin and orbital angular momenta of nonparaxial light in free space. It has been shown that this issue is closely related to the determination of the position of the center of gravity of a light beam or a wave packet. We have given an exact self-consistent solution to these problems in terms of quantum-operator formalism and using classical Poynting energy flows. In the helicity representation, taking into account the transverse nature of the electromagnetic fields, the operators of the OAM, SAM, and position become diagonal, but exhibit non-canonical commutation relations. We have shown that the unusual features of these operators originate from the Berry-phase terms and can be associated with manifestations of the spin-orbit interaction of light. Indeed, anomalous Berry terms in the OAM and position operators describe spin-dependent part of OAM (responsible for spin-to-orbital AM conversion) and spin-dependent shift of the center of gravity of light (i.e., the spin-Hall effect of light). We have applied the general theory to symmetric and asymmetric vector Bessel beams and found that our non-canonical operators indeed correspond to the observable quantities. The obtained Bessel-beam intensity distributions exhibit fine SOI splitting of caustics and Hall effects of light in perfect agreement with the derived OAM and position operators. These effects can be observed experimentally in tightly focused fields.
This work was supported by the European Commission (Marie Curie Action), Science Foundation Ireland (Grant No. 07/IN.1/I906), the Australian Research Council (ARC), and von Humboldt foundation. We are grateful to A. Y. Bekshaev and M. V. Berry for fruitful discussions.
Operator formalism for wave packets and beams
=============================================
One can separate two basic situations, for which the operator formalism of Section II can be adopted in a slightly different way. The first one is evolution of a wave-packet-like field localized in 3D space. Obviously, such field is nonmonochromatic and time-dependent. The plane-wave Fourier decomposition of the complex electric field can be written as $$\label{eqn:A1}
{\bf E}\left( {\bf r},t \right) = \frac{g}{(2\pi)^{3/2} }\int {{\bf \tilde E}\left( {\bf k} \right)e^{i{\bf k} \cdot {\bf r}-i\omega({\bf k})t} } d^3 {\bf k}~,$$ where $d^3{\bf k}=dk_x dk_y dk_z = k^2\sin\theta dk d\theta d\phi$, $\omega({\bf k})=k$ is the dispersion relation, factor $g=\sqrt{2\omega/\varepsilon_0}$ ($\varepsilon_0$ is the vacuum permittivity) is introduced for proper normalization of energy below, and the real wave electric field is given by ${\bm {\mathcal E}}\left( {\bf r},t \right) = {\rm Re}\, {\bf E}\left( {\bf r},t \right)$. In the helicity basis one has $$\label{eqn:A2}
{\bf \tilde E}\left( {\bf k} \right) = \tilde E^{+} \left( {\bf k} \right){\bf e}^{+} \left( {\bf k} \right) +
\tilde E^{-} \left( {\bf k} \right){\bf e}^{-} \left( {\bf k} \right)~.$$ The energy of the wave-packet field is given by the 3D space integral of the intensity (we omit inessential constant factors) and can be written as: $$\begin{aligned}
\label{eqn:A3}
\nonumber
W = \frac{1}{2}\int {\left(\varepsilon_0{\left| \bm{\mathcal E} \right|}^2+ \mu_0{\left| \bm{\mathcal H} \right|}^2\right)} {d^3}{\mathbf{r}} \\
\nonumber
= \frac{1}{4}\int {\left(\varepsilon_0{{\left| \bf{E} \right|}^2}+ \mu_0{{\left| \bf{H} \right|}^2}\right)} {d^3}{\mathbf{r}} \\
= \int {\omega{{\tilde E}^{\sigma *}}{{\tilde E}^\sigma }{d^3}{\mathbf{k}}} \equiv \left\langle {{{\tilde E}^\sigma }} \right|\omega \left| {{{\tilde E}^\sigma }} \right\rangle.\end{aligned}$$ Here ${\bm {\mathcal H}}\left( {\bf r},t \right) = {\rm Re}\, {\bf H}\left( {\bf r},t \right)$ is the magnetic field, $\mu_0$ is the vacuum permeability, $d^3{\bf r} = dx dy dz$, summation over $\sigma=\pm 1$ is implied hereinafter, and we performed some standard calculations with Maxwell equations and the Fourier transform (\[eqn:A1\]). Thus, the convolution implies 3D integration of the field spectral amplitudes over the ${\bf k}$-space. At the same time, to determine properly the state vector $\left| {\tilde E^\sigma } \right\rangle$, one has to take into account the temporal dependence of the field, namely: $$\label{eqn:A4}
\left| {\tilde E^\sigma } \right\rangle = \tilde E^\sigma \left( \mathbf{k} \right)e^{-i\omega \left( \mathbf{k} \right)t}~.$$ We assume normalization which has the meaning of the unit number of photons in the wave packet: $N= \left\langle {{{\tilde E}^\sigma }} \right| \left. {{{\tilde E}^\sigma }} \right\rangle =1$. Substituting the state vector (\[eqn:A4\]) with the definition of convolution (\[eqn:A3\]) into Eq. (\[eqn:12\]), we obtain the time-dependent position of the center of gravity of the wave packet moving in space: $$\label{eqn:A5}
\mathbf{R}\left( t \right) = - \operatorname{Im} \int {\tilde E^{\sigma *} \partial _\mathbf{k} \tilde E^\sigma } d^3 \mathbf{k} - \int {\sigma \mathbf{A}_B \left| {\tilde E^\sigma } \right|^2 } d^3 \mathbf{k} + \mathbf{V}t,$$ where the velocity of the wave-packet motion is given by $$\mathbf{V} = \int {(\partial _\mathbf{k} \omega) \left| {\tilde E^\sigma } \right|^2 } d^3 \mathbf{k} = \int {\bm{\kappa }\left| {\tilde E^\sigma } \right|^2 } d^3 \mathbf{k}.$$ Note that the same expression for the wave-packet center can be obtained by convolution of the canonical coordinate operator $\mathbf{\hat r} = i\partial _\mathbf{k}$ with the vector state $\left| {\tilde {\bf E}^\sigma } \right\rangle = \tilde {\bf E}^\sigma \left( \mathbf{k} \right)e^{-i\omega \left( \mathbf{k} \right)t}$. The Berry-connection term arises in this case from the $\partial _\mathbf{k}$ derivatives of the helicity basic vectors ${\bf e}^{\sigma}$, Eq. (\[eqn:9\]). The linear and angular momenta of the field, Eqs. (\[eqn:10\]), (\[eqn:11\]), and (\[eqn:13\]), can be calculated in a manner similar to Eqs. (\[eqn:A3\])–(\[eqn:A5\]).
The second typical problem that arises in optics deals with a beam-like monochromatic field ($\omega=k=const$) propagating in the positive $z$-direction and localized only in the transverse $(x,y)$-dimensions. In this case, it is natural to use the (2+1)D version of quantum-like formalism, where $z$ instead of time plays the role of the independent variable, whereas ${\bf r}_{\perp}=(x,y)$ is the effective 2D space allowing normalization of the transverse field distributions [@Aiello2]. Because of the monochromaticity, only two components of the ${\bf k}$-vector are independent, and the $z$-component can be expressed as $k_z = k_z \left( {\mathbf{k}_ \bot } \right) = \sqrt {\omega ^2 - k_ \bot ^2 }$, ${\bf k}_\bot=(k_x,k_y)$. This determines the following 2D plane-wave Fourier decomposition of the complex time-independent electric field [@Aiello2]: $$\label{eqn:A6}
{\bf E}\left( {\bf r}_{\perp},z \right) = \frac{g}{{2\pi }}\int {{\bf \tilde E}\left( {\bf k}_{\perp} \right)e^{i{\bf k}_{\perp} \cdot {\bf r}_{\perp} +ik_z({\bf k}_{\perp})z} } d^2 {\bf k}_{\perp}~,$$ where the real wave electric field is given by ${\bm{\mathcal E}}\left( {\bf r},t \right) = {\rm Re} \left[ {\bf E}\left( {\bf r}\right)e^{-i\omega t} \right]$ and the element of the 2D area of integration is ${d^2}{\mathbf{k}}_{\perp} = d{k_x}d{k_y}$. Alternatively, one can use $\mathbf{\tilde E} = \mathbf{\tilde E}\left( {\theta ,\phi } \right)$ and ${d^2}{\mathbf{k}}_{\perp} = {k^2}\cos \theta \sin \theta d\theta d\phi$ in spherical coordinates with two independent dimensions $(\theta,\phi)$. The characteristic energy of the wave beam is, in fact, the energy per unit $z$-length which is obtained by the 2D integration of the time-averaged intesity over $d^2{\bf r}_{\bot} = dx dy$: $$\begin{aligned}
\label{eqn:A7}
\nonumber
W = \frac{1}{2}\int {\left(\varepsilon_0\overline{{\left| \bm{\mathcal E} \right|}^2}+ \mu_0\overline{{\left| \bm{\mathcal H} \right|}^2}\right)} {d^2}{\mathbf{r}}_{\bot} \\
\nonumber
= \frac{1}{4}\int {\left(\varepsilon_0{{\left| \bf{E} \right|}^2}+ \mu_0{{\left| \bf{H} \right|}^2}\right)} {d^2}{\mathbf{r}}_{\bot} \\
= \int {\omega{{\tilde E}^{\sigma *}}{{\tilde E}^\sigma }{d^2}{\mathbf{k}}_{\bot}} \equiv \left\langle {{{\tilde E}^\sigma }} \right|\omega \left| {{{\tilde E}^\sigma }} \right\rangle=\omega~.\end{aligned}$$ Here ${\bm{\mathcal H}}\left( {\bf r},t \right) = {\rm Re} \left[ {\bf H}\left( {\bf r}\right)e^{-i\omega t} \right]$, the overline stands for the time averaging, and we assumed the unit number of photons per unit $z$-length in the beam: $N= \left\langle {{{\tilde E}^\sigma }} \right| \left. {{{\tilde E}^\sigma }} \right\rangle =1$. Thus, the convolution for beam-like fields implies 2D integration over $(k_x,k_y)$ or $(\theta,\phi)$ in the ${\bf k}$-space (these are equivalent unless we consider evanescent modes). To determine properly the state vector $\left| {\tilde E^\sigma } \right\rangle$, one has to take into account the $z$-dependence of the field, cf. Eq. (\[eqn:A4\]): $$\label{eqn:A8}
\left| {\tilde E^\sigma } \right\rangle = \tilde E^\sigma \left( \mathbf{k}_{\bot} \right)e^{ik_z \left( \mathbf{k}_{\bot} \right)z}~.$$ Substituting definitions (\[eqn:A7\]) and (\[eqn:A8\]) into Eq. (\[eqn:12\]), we obtain the $z$-dependent transverse position of the center of gravity of the propagating wave beam [@Aiello2]: $$\begin{aligned}
\label{eqn:A9}
\nonumber
\mathbf{R}_ \bot \left( z \right) = - \operatorname{Im} \int {\tilde E^{\sigma *} \partial _{\mathbf{k}_ \bot } \tilde E^\sigma } d^2 \mathbf{k}_ \bot \\
- \int {\sigma \mathbf{A}_B \left| {\tilde E^\sigma } \right|^2 } d^2 \mathbf{k}_ \bot
+ \mathbf{V}z~,\end{aligned}$$ where the ‘velocity’ of the motion along $z$ is given by $$\mathbf{V} = - \int { (\partial _{\mathbf{k}_ \bot } k_z) \left| {\tilde E^\sigma } \right|^2 } d^2 \mathbf{k}_ \bot = \int {\frac{\bm{\kappa }}
{{k_z }}\left| {\tilde E^\sigma } \right|^2 } d^2 \mathbf{k}_ \bot.$$ The linear and angular momenta (more precisely, their values per unit $z$-length) are calculated from Eqs. (\[eqn:10\]), (\[eqn:11\]) and (\[eqn:13\]) in a similar manner (see also Sections III and IV). Note that despite the above $(2+1)D$ quantum-like formalism, they are vectors in 3D space. This does not cause any difficulties if one uses $\hat{z}=i\partial_{k_z} = z$ which yields $Z=z$ in the 3D calculations, cf. [@Aiello].
It should be emphasized that, despite our using the same letters for the unifying formalism, quantities ${\bf E}$, $\tilde{\bf E}$, $W$, ${\bf R}$, etc. have different meanings for the 3D-localized wave-packet polychromatic fields, Eqs. (\[eqn:A1\])–(\[eqn:A5\]), and 2D-localized monochromatic beams, Eqs. (\[eqn:A6\])–(\[eqn:A9\]).
[99]{} A. I. Akhiezer and V. B. Berestetskii, *Quantum electrodynamics* (Interscience Publishers, New York, 1965). S. J. van Enk and G. Nienhuis, Europhys. Lett. **25**, 497 (1994); J. Mod. Opt. **41**, 963 (1994). S. M. Barnett and L. Allen, Opt. Commun. **110**, 670 (1994). S. M. Barnett, J. Opt. B: Quantum and Semiclass. Opt. **4**, S7 (2002); S. M. Barnett, J. Mod. Opt. **57**, 1339 (2010). L. Allen *et al.*, Phys. Rev. A **45**, 8185 (1992); L. Allen, M. J. Padgett, M. Babiker, Prog. Opt. **39**, 291 (1999). M. V. Berry, J. Opt. A: Pure Appl. Opt. **11**, 094001 (2009); A. Y. Bekshaev and M. S. Soskin, Opt. Commun. **271**, 332 (2007). C.-F. Li, Phys. Rev. A **80**, 063814 (2009). A. Y. Bekshaev, Cent. Eur. J. Phys. **8**, 947 (2010). M. H. L. Pryce, Proc. Roy. Soc. London, Ser. A **195**, 62 (1948). I. Bialynicki-Birula and Z. Bialynicka-Birula, Phys. Rev. D **35**, 2383 (1987); B.-S. K. Skagerstam, arXiv:hep-th/9210054; A. Bérard and H. Mohrbach, Phys. Lett. A **352**, 190 (2006). M. Hawton and W. E. Baylis, Phys. Rev. A **64**, 012101 (2001); Phys. Rev. A **71**, 033816 (2005). Z. Bomzon, M. Gu, and J. Shamir, Appl. Phys. Lett. **89**, 241104 (2006). Y. Zhao *et al.* Phys. Rev. Lett. **99**, 073901 (2007). C. Schwartz and A. Dogariu, Opt. Express **14**, 8425 (2006); D. Haefner, S. Sukhov, and A. Dogariu, Phys. Rev. Lett. **102**, 123903 (2009). O. G. Rodríguez-Herrera *et al*., Phys. Rev. Lett. **104**, 253601 (2010). V.S. Liberman and B.Y. Zel’dovich, Phys. Rev. A [**46**]{}, 5199 (1992); M. Onoda *et al.*, Phys. Rev. Lett. **93**, 083901 (2004); K. Y. Bliokh and Y. P. Bliokh, Phys. Lett. A **333**, 181 (2004); Phys. Rev. Lett. **96**, 073903 (2006). O. Hosten and P. Kwiat, Science **319**, 787 (2008); K. Y. Bliokh *et al.*, Nature Photon. **2**, 748 (2008). N. B. Baranova *et al.*, JETP Lett. **59**, 232 (1994); A. V. Volyar and T. A. Fadeeva, Tech. Phys. Lett. **26**, 740 (2000); K. Y. Bliokh *et al.*, Phys. Rev. Lett. **101**, 030404 (2008). A. Aiello *et al.*, Phys. Rev. Lett. **103**, 100401 (2009). V. G. Fedoseyev, Opt. Commun. **193**, 9 (2001); R. Dasgupta and P. K. Gupta, Opt. Commun. **257**, 91 (2006); K. Y. Bliokh *et al.*, Opt. Lett. **34**, 389 (2009); M. Merano *et al.*, Phys. Rev. A **82**, 023817 (2010). K. Y. Bliokh, Phys. Rev. Lett. **97**, 043901 (2006); K. Y. Bliokh and A. S. Desyatnikov, Phys. Rev. A **79**, 011807(R) (2009). V. Garbin [*et al.*]{}, New J. Phys. [**11**]{}, 013046 (2009). A. Y. Bekshaev, J. Opt. A: Pure Appl. Opt. [**11**]{}, 094003 (2009). S. G. Lipson, Opt. Lett. **15**, 154 (1990); K. Y. Bliokh, D. Y. Frolov, and Y. A. Kravtsov, Phys. Rev. A **75**, 053821 (2007); K. Y. Bliokh, J. Opt. A: Pure Appl. Opt. **11**, 094009 (2009). M. V. Berry, Nature **326**, 277 (1987). A. T. O’Neil [*et al.*]{}, Phys. Rev. Lett. [**88**]{}, 053601 (2002); A. Aiello and K. Y. Bliokh (in preparation). M. Born and E. Wolf, *Principles of Optics* (Oxford: Pergamon Press, 1970). R. Jáuregui and S. Hacyan, Phys. Rev. A **71**, 033411 (2005). As it is known, the Bessel beams cannot be normalized and we calculate the finite ratio of the quantities Eqs. (\[eqn:10\])–(\[eqn:14\]) to the infinite norm $N=\left\langle {\tilde E^\sigma } \right|\left. {\tilde E^\sigma } \right\rangle$. A. Shapere and F. Wilczek (ed) *Geometric Phases in Physics* (Singapore: World Scientific, 1989). J.-P. Ngome, P. A. Horváthy, and J. W. van Holten, J. Phys. A: Math. Theor. **43**, 285401 (2010). B. Richards and E. Wolf, Proc. R. Soc. London A [**253**]{}, 358 (1959). M. V. Berry and K. T. McDonald, J. Opt. A: Pure Appl. Opt. **10**, 035005 (2008). G. P. Mikitik and Y. V. Sharlai, Phys. Rev. Lett. **82**, 2147 (1999); P. Gosselin, H. Boumrar, and H. Mohrbach, Europhys. Lett. **84**, 50002 (2008). K. S. Novoselov *et al.*, Nature **438**, 197 (2005); Y. Zhang *et al.*, Nature **438**, 201 (2005). Z. Bomzon and M. Gu, Opt. Lett. **32**, 3017 (2007); Y. Gorodetski [*et al.*]{}, Phys. Rev. Lett. [**101**]{}, 043903 (2008). A. Aiello, M. Merano, and J. P. Woerdman, Phys. Rev. A **80**, 061801(R) (2009).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
C. J. Evans, N. Castro, O. A. Gonzalez, M. Garcia, N. Bastian, M.-R. L. Cioni, J. S. Clark,\
B. Davies, A. M. N. Ferguson, S. Kamann, D. J. Lennon, L. R. Patrick, J. S. Vink, D. R. Weisz
bibliography:
- '34145.bib'
date: 'Received 27 August 2018 / Accepted 21 December 2018'
title: First stellar spectroscopy in Leo P
---
Introduction {#intro}
============
Leo P is a relatively nearby, dwarf irregular galaxy. Its discovery and first studies of its physical properties were reported in a series of five papers by @g13, @r13, @s13, @m13, and @bc14. Initially discovered from H observations from the Arecibo Legacy Fast ALFA Survey (ALFALFA), further radio observations revealed its spatial extent and velocity structure [@g13]. Optical imaging from the 3.5m WIYN telescope gave a first view of its stellar population, with evidence of ongoing star formation in a luminous H region [@r13]. Spectroscopy of the H region from the 4m Mayall Telescope and the Large Binocular Telescope (LBT) was used to estimate an oxygen abundance, via analysis of the \[O\] 4363 line, of \[O/H\]$=$7.17$\pm$0.04 [just 3% of solar, @s13]. We lack an estimate of iron abundance at present, but even if Leo P were found to have a sub-solar $\alpha$/Fe ratio [as found for massive stars in other metal-poor dwarf irregulars, e.g. @v03; @tau07; @g14; @h14], such a low oxygen abundance suggests it is one of the most metal-poor star-forming galaxies known[^1].
Follow-up imaging with the LBT was used by @m13 to estimate a distance of 1.72$^{+0.14}_{-0.40}$Mpc, from the location of the tip of the red giant branch (TRGB) in the colour-magnitude diagram (CMD), with an improved measurement of 1.62$\pm$0.15Mpc from the luminosity of the horizontal branch and light curves of candidate RRLyrae stars [@m15]. The system is thought to have a stellar mass of 5.7$\times$10$^5$M$_\odot$ [@m13], with a neutral-hydrogen mass of 9.5$\times$10$^5$M$_\odot$ and a dynamical mass in excess of 2.6$\times$10$^7$M$_\odot$ [@bc14].
Following its discovery, @bofb13 suggested Leo P is a member of the NGC3109 association and is one of five ‘dwarfs walking in a row’ (together with NGC3109, Sextans A, Sextans B, and Antlia). They speculated this association could have originated from tidal interaction or, if arriving at the Local Group for the first time, that the structure could be a remnant of a cosmological filament.
![Observed MUSE field (in red) overlaid on the [*F475W*]{} [*HST*]{}-ACS image of the Leo P region [@m15]. The MUSE data encompass most of the visible extent of this low-luminosity dwarf system.[]{data-label="spatial_ACS"}](34145_fig1.pdf)
In this context, Leo P is of interest as the potential site of the most metal-poor massive stars in the local Universe. The populations of high-mass stars in metal-poor galaxies are a prime contender for the intense ultraviolet radiation which reionised the intergalactic medium by z$\sim$6 [e.g. @f12; @stark16; @wbk17]. However, while there are theoretical predictions for the properties of high-mass stars at these early epochs (including the putative ‘Population III’ stars), we lack observational tests of this very metal-poor regime.
In the Local Group (and nearby), the range of low-metallicity environments in which to study massive-star evolution is limited to the Magellanic Clouds [e.g. the VLT-FLAMES surveys, @esl05; @vfts] and other metal-poor galaxies such as IC1613, WLM, NGC3109 and Sextans A [e.g. @t11; @t14; @gh13; @c16]. To reach metallicities significantly below those in the Clouds for substantial stellar populations, we need to look much further afield to star-forming galaxies such as DDO 68 [$\sim$12.7Mpc, @c14; @s16] or I Zwicky 18 [18.2$\pm$1.5Mpc, @a07]. The presence of young blue populations in Leo A and the Sag DIG [@weisz14; @g18] and the discovery of Leo P, gives us a first chance to explore this regime. Tantalisingly, Leo P has a comparable oxygen abundance to both DDO 68 and IZw18 [@sk93; @p05] but is signficantly closer, potentially providing a unique opportunity to study high-mass stars in a pristine environment similar to those in the earliest stages of the Universe.
The cool-star population of Leo P is of also interest. The main sites of dust production in the early Universe are thought to be supernovae [e.g. @mmh10] and stars on the asymptotic giant branch [AGB, e.g. @v09; @zh13]; luminous blue variables might also play a role [@gn14]. Considerable effort has been undertaken to identify and characterise the AGB populations of local metal-poor galaxies to investigate this channel for dust production . Identification of AGB stars in Leo P would be a natural extension to such studies. @l16 used single-epoch, near-IR imaging of Leo P to identify 22 candidate AGB stars, but used selection criteria from @s12 from their observations of NGC6822 [where the cool stars are more metal-rich than Leo P, e.g. @p15]. Spectroscopic confirmation of their nature is a first step to planning longer-wavelength observations to investigate dust-production rates.
Here we present the first stellar spectroscopy in Leo P, obtained with the Multi-Unit Spectroscopic Explorer [MUSE; @b10] on the Very Large Telescope (VLT) at Paranal. The observations and data reduction are detailed in Section 2, followed by discussion of the stellar spectra (and detections of background galaxies) in Section 3. In Section 4 we discuss the potential impact of the low metallicity of Leo P on the magnitudes of massive stars, and in Section 5 we briefly investigate the morphology and dynamics of the nebular emission. Concluding remarks are given in Section 6.
Observations & data reduction
=============================
We observed Leo P using the extended wide-field mode of MUSE, without the blue filter to extend the wavelength coverage down to 4650Å (to include He 4686 if present in our targets). Eight observations, spanning 13 December 2015 to 8 March 2016, were executed from the service queue on UT4 of the VLT. Each comprised 2$\times$1505s exposures (so a total integration of 6.7hrs), centred on a position of $\alpha$$=$10$^{\rm h}$21$^{\rm
m}$443, $\delta$$=$18$^\circ$05$'$305 (J2000). As shown in Fig. \[spatial\_ACS\], the observed MUSE field spans most of the visible extent of the galaxy. All of the observations were obtained in dark time (in terms of lunar illumination and angular distance), except for one that was on the boundary of grey and bright time.
The data were reduced using the MUSE pipeline (v1.6.1), with steps including: bias correction, wavelength calibration, reconstruction of datacubes from the individual spectra on the detectors, correction of the spectra to the heliocentric frame, sky subtraction, and merging of the individual exposures to form a combined datacube. Our reductions were largely indistinguishable to the ‘MUSE-DEEP’ processing and stacking of the data undertaken as part of ESO’s archiving activities, and the combined cube is available from the ESO Phase 3 archive[^2]. The central parts of the cube are relatively crowded (cf. the typical seeing of the observations of $\sim$06). To extract our sources we used the [pampelmuse]{} software [@k13] that was developed to recover spectra from blended sources in MUSE data [see, e.g. @k16].
For the input catalogue for the [pampelmuse]{} extractions we used the [*Hubble Space Telscope (HST)*]{} imaging from @m15, who used the Advanced Camera for Surveys (ACS) to observe Leo P with the [*F475W*]{} and [*F814W*]{} filters (see Fig. \[spatial\_ACS\]). We created a source catalogue from the [*HST*]{} data with [dolphot]{}, a version of [hstphot]{} [@d00 that has been updated with a specific module for ACS], and using the parameter file and photometric criteria from @w14. The angular resolution of these data is a factor of ten better than the MUSE observations, and many of the MUSE sources are resolved into multiple components.
[pampelmuse]{} is designed to fit the point spread function (PSF) of each source throughout each wavelength slice of the combined MUSE datacube. In general, this requires a subset of bright, high signal-to-noise stars to define the PSF and coordinate transformations between the input catalogue and the MUSE observations prior to the extractions. Unfortunately, our Leo P field lacked sufficiently bright stars for adequate PSF modelling. This is in contrast to the recent analysis of the luminous stellar population of NGC300 which, although more distant at $\sim$1.9Mpc, has a much greater stellar density and a significant population of luminous stars [@r18].
We therefore used [pampelmuse]{} to fit all resolved sources simultaneously, which entailed computing both the Gaussian PSF profile and the coordinate transformations at each layer. After [ pampelmuse]{} had selected all the stars that were suitable for fitting and extraction, we performed a first run using a binning of 50 pixels in the spectral direction. This allowed us to smooth the wavelength dependencies to the fitted parameters by using a simple polynomial fit before performing the final extraction at each wavelength for 341 sources from the [*HST*]{} catalogue. In the cases of well isolated stars the resulting spectra were comparable to simple aperture extractions from the cube. However, the [pampelmuse]{} approach has the advantage over aperture extractions of not simply co-adding multiple components together in one spectrum (where objects are either blended or simply within the defined aperture).
The positions and photometry of the MUSE spectra that we were able to classify are given in Tables \[targets\] and \[cool\_targets\]. We also give the identifiers and photometry from the ground-based imaging from @m13. Those data were obtained with a seeing of 07, so are a good match to the angular resolution of our MUSE cube for relatively isolated stars (or where additional components detected in the [*HST*]{} images are sufficiently faint not to contribute more than a few percent of the flux). Given the small field of the ACS images, the right ascension and declination of our sources are given in the astrometric frame of @m13, where the [*HST*]{} and MUSE data were transformed using matched stars from their catalogue.
Spectral content
================
Our motivation for the MUSE observations was a first spectroscopic census of Leo P, primarily to identify candidate massive stars for quantitative analysis with higher-quality, follow-up spectroscopy. The spatial locations of our classified spectra are shown in Fig. \[spatial\], with their positions in the [*HST*]{} CMD shown in Fig. \[cmd\]. For reference, @m13 estimated the TRGB as $I$$=$22.11mag. As such, we expected to find a number of (modestly) massive stars in the MUSE spectroscopy in the ‘blue’ plume at ($F475W$$-$$F814W$)$\sim$0.0mag, and luminous evolved red stars above the TRGB. We have found examples of both hot and cool objects, as well as background galaxies in the MUSE field, as outlined in the following sections.
![Colour-magnitude diagram for the MUSE field using our photometry of the [*HST*]{} imaging from @m15. Symbols as follows: massive stars (closed cyan circles), candidate massive stars (open cyan circles), carbon stars (closed red circles), other sources previously classified as potential AGB stars (open red circles), background galaxies (green circles), foreground star (open green square).[]{data-label="cmd"}](34145_fig3.pdf)
![Blueward part of the MUSE spectrum of the extracted central source of the H[II]{} region in Leo P (LP26). The identified emission lines, from left-to-right by species, are: H$\beta$; He [I]{} 4713,4922,5016; \[O [III]{}\] 4959,5007. The He [II]{} 4686, 5411 absorption lines provide the first direct evidence for an O-type star in Leo P.[]{data-label="HIIregion"}](34145_fig4.pdf)
HII region
----------
There is a prominent H region in Leo P, discussed by @s13, and resolved into multiple sources and nebulosity by the [*HST*]{} images [@m15]. Skillman et al. mentioned an absence of absorption features, but the MUSE spectrum of the bright, central star (LP26) displays He 4686, 5411 absorption (see Fig. \[HIIregion\]), arguing for the presence of an O-type star. @m15 estimated the absolute magnitude of this source as $M_V$$=$$-$4.43, equivalent to what we would expect for a mid to late O-type dwarf at Galactic metallicity [e.g. @w72; @msh05]. In the absence of published calibrations for the absolute magnitudes of O-type stars at low metallicities, we discuss this aspect further in Section \[mags\].
LP\# MSB $\alpha$(J2000) $\delta$(J2000) $V$ $I$ [*F475W*]{} [*F814W*]{} S/N Comments
------ ----- ----------------- ----------------- ------- ------- ------------- ------------- ----- ------------------------------------
1 138 10 21 43.66 $+$18 05 28.50 24.40 23.90 24.459 24.087 ø4 candidate
2 207 10 21 43.71 $+$18 05 21.69 24.79 24.64 24.737 24.819 ø5 candidate
3 94 10 21 43.89 $+$18 05 19.35 23.98 23.68 24.156 23.880 ø7 candidate
4 16 10 21 43.92 $+$18 05 56.59 22.50 22.29 22.521 22.561 19 RV(H$\beta$)$=$240$\pm$19
5 202 10 21 43.97 $+$18 05 19.29 24.83 24.09 24.973 24.046 ø4 candidate
6 40 10 21 44.01 $+$18 05 23.81 23.45 23.08 23.707 23.348 ø8
7 98 10 21 44.37 $+$18 05 29.45 24.09 23.52 24.613 24.431 ø5 candidate
8 … 10 21 44.39 $+$18 05 25.31 … … 24.419 24.318 ø9
9 92 10 21 44.47 $+$18 05 25.62 23.95 23.66 24.111 23.961 ø6 candidate
10 … 10 21 44.55 $+$18 05 20.00 … … 23.958 24.013 ø9 candidate
11 … 10 21 44.63 $+$18 05 16.14 … … 23.834 23.908 ø4 candidate
12 … 10 21 44.69 $+$18 05 21.07 … … 24.363 24.488 ø7 candidate
13 … 10 21 44.69 $+$18 05 23.47 … … 24.090 24.155 ø9 candidate
14 25 10 21 44.76 $+$18 05 27.17 23.14 23.00 23.276 23.421 10 Be/Ae? (H$\alpha$ em.?)
15 … 10 21 44.79 $+$18 05 25.15 … … 24.251 24.540 ø7 candidate
16 15 10 21 44.81 $+$18 05 22.31 … … 22.790 23.021 21
17 … 10 21 44.89 $+$18 05 24.33 … … 24.634 24.850 ø8
18 12 10 21 44.90 $+$18 05 22.38 22.16 22.02 22.255 22.304 28 RV(H$\beta$)$=$274$\pm$19
19 41 10 21 44.95 $+$18 05 06.28 23.46 23.37 23.483 23.497 ø8
20 … 10 21 44.97 $+$18 05 24.72 … … 24.768 24.511 ø6 candidate
21 … 10 21 45.00 $+$18 05 25.59 … … 22.579 22.568 15 RV(H$\beta$)$=$260$\pm$22
22 24 10 21 45.01 $+$18 05 13.21 23.15 22.87 23.369 23.308 11 Be/Ae? (H$\alpha$ & H$\beta$ em.?)
23 … 10 21 45.03 $+$18 05 24.80 … … 22.473 22.585 12 RV(H$\beta$)$=$247$\pm$20
24 … 10 21 45.03 $+$18 05 25.02 … … 23.066 23.052 ø3 candidate
25 … 10 21 45.08 $+$18 05 16.35 … … 24.290 24.684 15 On SE edge of H [II]{} region
26 2 10 21 45.10 $+$18 05 16.62 … … 21.464 21.837 52 O-type star in H [II]{} region
27 82 10 21 45.12 $+$18 05 06.92 23.92 23.36 24.322 23.740 ø5 candidate
28 22 10 21 45.13 $+$18 05 13.76 23.01 23.16 22.999 23.349 13 candidate
29 33 10 21 45.14 $+$18 05 20.31 23.27 23.09 23.480 23.465 ø7
30 63 10 21 45.21 $+$18 05 09.09 23.73 23.41 24.435 24.488 ø8
31 114 10 21 45.26 $+$18 05 29.82 24.17 23.89 24.294 24.276 ø4 candidate
32 106 10 21 45.27 $+$18 05 07.84 24.11 23.90 24.238 24.080 ø5 candidate
Early-type stars
----------------
At the spectral resolution ($R$$\sim$2000 at 5500Å) and signal-to-noise (S/N) of the MUSE data, we would not expect to be able to identify many prominent spectral lines in early-type stars (which is further exacerabated by the lack of spectral coverage shortwards of $\sim$4650Å). Nonetheless, on the basis of H$\beta$ absorption, combined with indications of stellar features at H$\alpha$, we classify 14 sources as massive stars, with spectra of some of the brighter sources shown in Fig. \[blue\] (with a S/N ranging from 12 to 28 per pixel). We classified a further 17 spectra as possible massive stars, typically where the H$\beta$ absorption is less secure, but still supported by an increasingly blue flux distribution.
Given this data quality, combined with the spectral range, sky residuals, and resolution of MUSE, it is not surprising that we do not see other strong features. That said, weak absorption consistent with arising from He 5876, 6678 is seen in both LP16 and LP18 (see Fig. \[blue\]), but we also see evidence for oversubtraction of the \[O \] nebular lines (and similarly in the core of H$\alpha$) suggesting these could be related to limitations of the background subtraction. The only other feature of note is Paschen9 9229 absorption in the spectra of LP21 and LP23. From the steepening blue flux distributions of these spectra, combined with an absence of features that might be expected at cooler types, we expect that most of these are B-type objects (or potentially late O-type for the brightest few objects). The location of these sources in the CMDs (Fig. \[cmd\]) provides support for our classifications in most cases.
For four spectra with S/N$>$10 and well-defined H$\beta$ absorption (and without obvious evidence for nebular over-subtraction) we estimated radial velocities (RV) from Gaussian fits to the observed H$\beta$ absorption (included in last column of Table \[targets\]). The weighted-mean velocity ($\overline{\rm RV}$) from the four spectra is $\overline{\rm RV}$$=$255$\pm$13. The velocity of the He 4686 absorption in the spectrum of LP26 in the H region is consistent with these values (252$\pm$35). Given the relatively large uncertainties on the MUSE values, they are in good agreement with the systemic velocity of neutral hydrogen of $v_{\rm LSRK}$$=$260.8$\pm$2.5 [@bc14], which equates to a heliocentric velocity of approx. 265 in the direction of Leo P.
Cool stars
----------
Near-infrared imaging of Leo P was used by @l16 to identify 22 candidate AGB stars, six of which lie within the MUSE field. The spectra of five of these are shown in Fig. \[agb\_lee\], with their observational properties summarised in Table \[cool\_targets\]; the sixth (MSB\#56) is a background galaxy (see Table \[highz\_targets\]). The two brightest candidates are carbon stars, as revealed by the strong C$_2$ Swan bandheads at 5165, 5636. We compared the wavelengths of the bandheads in the MUSE spectra to those in an archival UVES spectrum from @uvespop of the Galactic carbon star, W Ori (HD32736), finding differential velocities that are consistent with their membership of Leo P (i.e. $\delta$RV$\sim$250).
The remaining three stars, suggested as oxygen-rich M-type AGB stars by @l16, are relatively featureless. However, there appears to be absorption in the near-IR Ca Triplet (CaT, with rest wavelengths of 8498, 8542, 8662) for MSB13. An expanded view of the CaT region for the candidate AGB sources is shown in Fig. \[CaT\], with absorption also present for 8542 in MSB18. Indeed, of the three CaT lines in the MUSE data, the central 8542 line is the most robust in terms of being less influenced by sky residuals. From a Gaussian fit to the 8542 line in MSB13 we estimated RV$=$262$\pm$8, in good agreement with the values for the hotter stars. To investigate the CaT region for the fainter stars in the RGB, the lower spectrum in Fig. \[CaT\] shows the co-added data of the 16 sources with [*F814W*]{}$\le$23.0. The stronger CaT components can again be seen, confirming the presence of cool, evolved stars as expected from the CMD (Fig. \[cmd\]).
Thus, we conclude that at least two of these candidate AGB stars appear to be cool, evolved stars with CaT absorption (and with $F814W$ magnitudes that are brighter than the TRGB). The corresponding lack of strong molecular bands (e.g. from TiO) in their spectra in Fig. \[agb\_lee\] is notable compared with what is usually seen in (M-type) AGB specta, but not unexpected given the low metallicity of Leo P (as traced by the published oxygen abundance). Indeed, the ratio of C-rich to M-type (O-rich) AGB stars is known to be a function of metallicity, with an increasing C/M ratio towards lower metallicities [e.g. @ir83; @bd05]. The relative dearth of M-type objects at low metallicities is attributed to the quicker dredge-up timescales to become a carbon star, combined with higher temperatures from the evolutionary models, which would act to reduce the TiO absorption, hence fewer M-type spectra [see, e.g., discussion by @k12]. Equally, it is plausible that simple metallicity effects give the impression of earlier spectral types, even if the temperatures were not that different. For example, a luminous, cool star with an effective temperature of 4000K would be classified as M-type in the LMC, but at the metallicity of Leo P, the TiO bands would weaken sufficiently that it would be classified as a K-type spectrum.
All five of the objects from @l16 are relatively isolated, point-like sources in the [*HST*]{} imaging, and we highlight that four are somewhat removed from the central part of the system. This suggests the intermediate-age population is either more extended or well mixed, with the latter similar to the findings for six of the nine dwarf galaxies studied by @m17. Following the arguments of @m17, we note that MSB18 is at a distance of 50$''$ (equivalent to 400pc) from the H region, potentially providing a source of chemical enrichment in the outer regions of Leo P.
![Region around the Ca [II]{} triplet (CaT) for the three stars previously identified as candidate AGB stars, together with a stacked spectrum of the 16 next brightest objects at the top of the red giant branch (i.e. excluding the three candidate AGB stars). CaT absorption is seen in the brightest two (MSB13 and 18) and in the stacked data.[]{data-label="CaT"}](34145_fig7.pdf)
Foreground stars
----------------
There are three seemingly foreground stars included in our extractions from the MUSE cube, as listed at the end of Table \[cool\_targets\]. One star (MSB1) has a somewhat larger RV than might be expected for a foreground Galactic object. Cross-correlation of the two M-type spectra (over the 4600-5565 range) yielded a differential RV of 168$\pm$20, in good agreement with absolute estimates from the CaT of 172$\,\pm$6.
Background galaxies {#highz_sources}
-------------------
From initial extractions of the MUSE sources using source detections and simple aperture extractions, we found 20 sources which are background galaxies. Indeed, a number of these are clearly visible as such in the [*HST*]{} imaging [see discussion of two examples by @l16], meaning that they are not included in the CMD in Fig. \[cmd\] as they were rejected as non-stellar sources in the photometric analyis. Observational details for the spectroscopically-confirmed galaxies are given in Table \[highz\_targets\], including estimates of their redshifts ($z$) using the diagnostic emission or absorption features as indicated. Of note are the two systems at $z$$=$2.5, located adjacent to each other on the sky. The [*HST*]{} imaging reveals that both MUSE sources are comprised of two components, as shown in Fig. \[z25\_hst\].
![Two $z$$\sim$2.5 sources in the [*HST*]{}-ACS [*F475W*]{} image from @m15. At the resolution of [*HST*]{}, the extracted MUSE sources are both resolved into two components.[]{data-label="z25_hst"}](34145_fig8.pdf)
MSB $\alpha$(J2000) $\delta$(J2000) $V_o$ $I_o$ $z$ Diagnostic
----- ----------------- ----------------- ------- ------- -------- ------------------------------
19 10 21 42.43 $+$18 05 17.38 23.41 21.70 0.5436 \[O [II]{}\] 3727
39 10 21 42.46 $+$18 05 42.80 23.88 22.33 0.7543 \[O [II]{}\] 3727
56 10 21 42.55 $+$18 05 47.16 24.62 22.53 0.5169 \[O [II]{}\] 3727
37 10 21 42.56 $+$18 05 17.56 24.19 22.24 0.7554 \[O [II]{}\] 3727
4 10 21 42.62 $+$18 05 14.59 21.95 20.26 0.3925 Ca $H+K$
62 10 21 42.87 $+$18 05 33.13 24.15 22.64 0.7539 \[O [II]{}\] 3727
48 10 21 42.97 $+$18 05 38.43 24.22 22.40 0.7549 \[O [II]{}\] 3727
30 10 21 42.99 $+$18 05 08.98 23.30 22.62 2.5007 Near-UV Fe [II]{}, Mg [II]{}
49 10 21 43.01 $+$18 05 59.13 23.98 22.50 0.8883 \[O [II]{}\] 3727
26 10 21 43.04 $+$18 05 08.34 23.29 22.36 2.5005 Near-UV Fe [II]{}, Mg [II]{}
7 10 21 43.10 $+$18 05 17.26 23.07 20.88 0.5440 \[O [II]{}\] 3727
31 10 21 43.39 $+$18 05 38.07 23.60 22.25 0.7462 \[O [II]{}\] 3727
17 10 21 44.15 $+$18 05 03.11 22.84 21.82 0.4763 \[O [II]{}\] 3727
108 10 21 44.15 $+$18 05 11.12 24.29 23.23 0.8320 \[O [II]{}\] 3727
– 10 21 44.44 $+$18 05 05.53 … … 0.4763 \[O [II]{}\] 3727
168 10 21 45.28 $+$18 05 34.77 24.57 24.17 0.3603 \[O [II]{}\] 3727
151 10 21 45.68 $+$18 05 38.30 25.00 23.38 0.6028 \[O [II]{}\] 3727
172 10 21 46.04 $+$18 05 35.46 24.78 23.67 0.8149 \[O [II]{}\] 3727
163 10 21 46.05 $+$18 05 49.46 24.86 23.52 0.9692 \[O [II]{}\] 3727
– 10 21 46.18 $+$18 05 59.60 … … 0.4137 Ca $H+K$
Expected absolute magnitudes at low $Z$ {#mags}
=======================================
We could potentially use the estimated absolute magnitude of LP26 to (coarsely) constrain its spectral type, but we are unfortunately limited to Galactic results [e.g. @w72; @w00; @msh05]. To investigate the potential impact of the very low metallicity in a system like Leo P on the expected stellar parameters we turned to evolutionary models calculated by @s15 for IZw18 (with $Z$$=$0.02$Z_{\odot}$). Fig. \[tracks\] shows the zero-age main sequence (ZAMS) as traced by their non-rotating models compared to results for the Galaxy, LMC and Small Magellanic Cloud (SMC) from @b11; for completeness the LMC tracks are supplemented by results for $M$$>$60$M_\odot$ from @k15. Also shown in Fig. \[tracks\] are the ZAMS positions for models with initial masses of 9, 10, 15, 20, and 30M$_\odot$, highlighting the temperature dependence vs. metallicity for stars of a given mass (at near constant luminosity, corresponding to smaller radii at lower metallicities); similar differences are also seen in the rotating models from the same grids. A shift in the same sense is also seen in the ‘Geneva’ evolutionary models, for example the comparison in Fig. 20 from @s15 of the ZAMS for the SMC-like models of @gee13 and the near-zero metallicity models of @mm02.
As an example case, consider the ZAMS temperatures in Fig. \[tracks\] for a 20M$_\odot$ star where T$_{\rm
eff}$(MW)$=$35.3kK cf. T$_{\rm
eff}$(IZw18)$=$41.2kK. The models at the ZAMS have (near) constant luminosities, but such stars will have different bolometric corrections (BCs) due to the increasing temperatures towards lower metallicity. To investigate the scale of this effect we highlight the BCs from the [tlusty]{} OSTAR2002 grids [@lh03]. A Galactic model for 35kK has BC$=$$-$3.24mag, cf. BC$\sim$$-$3.8mag for the higher temperature at $Z$$=$0.03$Z_{\odot}$ (interpolating between T$_{\rm
eff}$$=$40 and 42.5kK).
There is a similarly large difference for the 9M$_\odot$ models, ranging from 24.5kK in the Galaxy to 30kK for IZw18. We again turned to the [tlusty]{} results to assess the potential impact of this on the BC values, with BC$\sim$$-$2.95 for the hotter, metal-poor star. The cooler, Galactic temperature is beyond the OSTAR2002 grid, but an estimate of BC$\sim$$-$2.40 is available from the companion BSTAR2006 grid [@lh07]. We require similar calculations with the latest wind codes but these results illustrate that, for a given mass, the visual absolute magnitudes of high-mass stars at such low metallicities could be $\sim$0.5mag fainter than we would otherwise expect (together with an increase in the number of ionising photons). Alternatively, this also implies that fewer massive stars could potentially account for a given level of ionisation in unresolved systems at very low-$Z$.
![Zero-age main sequence in the Hertzsprung–Russell diagram for (non-rotating) evolutionary models with $Z$$=$0.02$Z_\odot$ [blue, @s15], and for the SMC (green), LMC (red), and Galaxy (black) from @b11; the LMC results are supplemented by higher-mass models (dotted red line) from @k15. Results for five initial masses (in units of M$_\odot$) are highlighted by the blue dots to illustrate the temperature dependence vs. metallicity (but with no 10M$_\odot$ model available for the SMC).[]{data-label="tracks"}](34145_fig9.pdf)
We are currently unable to link initial mass to spectral type as a function of metallicity, but the differences above echo the effect where metallicity is known to impact on the temperature scale of OB-type stars of a given spectral type [e.g. @m05; @m06; @m07; @t07]. At lower metallicities the cumulative opacity of the metal lines is diminshed, resulting in less ‘back warming’ by trapped radiation and hotter models are required to reproduce the observed ratios of spectral lines [also see discussion by @msh05].
Taking this into consideration for the calibrations from @msh05, the estimated absolute visual magnitude from @m15 (M$_V$$=$$-$4.43) constrains the ionising source in the H region in Leo P to a mid O-type star. Given the prevalence of binarity in massive stars, @m15 argued that this source could perhaps be two O7 or O8 stars, with stellar masses of $\sim$25M$_\odot$. Such a mass exceeds the maximum (of 2.5-3M$_\odot$) expected at the low star-formation rate of Leo P from the integrated galactic initial mass function (IGIMF) approach [@pwk07]. Aside from the specific physical properties of LP26, its spectroscopic confirmation as an O-type star supports the conclusions of @m15 in this regard, i.e. that the upper limit to the IMF – even if sparsely sampled – is not so significantly influenced by the star-formation rate. More recently, @jhk18 have argued that the inferred star-formation rate in the IGIMF theory for Leo P would be significantly larger than the value estimated from the current H$\alpha$ luminosity of the H [ II]{} region; this helps to reconcile the predictions with the detection of the O-type star.
Nebular emission
================
Inspecting the MUSE data at the wavelength of H$\alpha$ emission we discovered three new structures compared to the original discovery images, suggesting more than one site of (relatively) recent star formation, as shown in the contours in Fig.\[Ha\_overlay\] and the upper panels of Fig. \[emission\_maps\]. Part of the southern shell is also traced by H$\beta$ emission (middle panels), while \[O \] 5007 emission is only seen in the vicinity of the H region (lower panels). Intensity maps of \[O \] 4959, \[N \] 6584, 6548, and \[S \] 6717, 6731 were comparable to that for \[O \] 5007, in the sense that the large structures seen in H$\alpha$ are absent (perhaps simply linked to the low metallicity).
![H$\alpha$ emission in Leo P overlaid on the combined $F475W+F814W$ [*HST*]{} image (cyan, yellow and red contours are H$\alpha$-emission levels of 5, 10, and 15$\times$10$^{-20}$ergs$^{-1}$cm$^{-2}$, respectively). The known H [II]{} region is the compact blue source in the main southern feature. The MUSE data have also revealed a region of H$\alpha$ emission just to the north, and two large extended shells to the south and (far) north (see also Fig. \[emission\_maps\]).[]{data-label="Ha_overlay"}](34145_fig10.png){width="9.75cm"}
The northern H$\alpha$ shell has an apparent diameter of $\sim$15$''$. Adopting a distance of 1.62Mpc [@m15], this is equivalent to a projected diameter of $\sim$120pc. The northern and southern shells appear to be fairly well defined rings of emission, whereas a central shell appears somewhat more diffuse. The H region is located on northern edge of the southern ring.
{width="18.5cm"} {width="18.5cm"} {width="18.5cm"}
The positions of our spectroscopically-confirmed stars are overlaid in the left-hand panels of Fig. \[emission\_maps\]. The absence of massive stars in the northern shell suggests this is an older formation (e.g. supernova remnant) rather than a more recent wind-blown bubble. Only two of our sources lie within the southern ring, including LP22, one of the candidate Be/Ae stars, which is unlikely to be the driving source for such a ring, again suggesting it as a supernova remnant. Indeed, given the large number of high-mass x-ray binaries with Be-type components in the low-metallicity environment of the SMC [e.g. @c05], we speculate that LP22 warrants further study to investigate for the presence of a degenerate companion. Lastly, we also note the four stars along the southern edge of the southern ring, suggestive of a connection in their formation.
Using a similar profile-fitting approach to that used by @cce18 for MUSE observations of 30 Doradus, velocity maps for each emission line are shown in the centre panels of Fig. \[emission\_maps\] (in which the data are spatially smoothed with a Gaussian function with a spatial FWHM of 0.6$''$ prior to the fits). We also investigated maps of the standard deviation ($\sigma$) of the H$\alpha$ Gaussian fits to investigate the velocity dispersion of the gas, but these do not add much to the velocity maps, aside from hints of a slightly larger dispersion in the northern shell.
The systemic velocity of the gas ($\sim$265) is in good agreement with the H and stellar velocities. There appears to be a small velocity gradient of a few across the H region in the H$\beta$ map, with a similar trend in \[O \], but given the limitations of the MUSE data, we defer detailed study of the kinematics to future observations at higher spectral resolution [e.g. with the Keck Cosmic Web Imager, @kcwi].
Prompted by the discovery of these spatially-distinct structures we re-examined the [*HST*]{} photometry of all the stars in the four regions (i.e. including those without spectroscopic confirmation from MUSE). This led to relatively sparse samples and aside from reinforcing the relative dearth of luminous, blue stars in the northern and southern regions, we were unable to glean further insights into the histories of these regions.
Concluding remarks
==================
We have presented the first stellar spectroscopy in Leo P, a low luminosity, dwarf galaxy at a distance of 1.62$\pm$0.15Mpc. Our findings from the MUSE observations include:
- [Spectroscopic confirmation of an O-type star (LP26) in the H region via observations of He absorption. From consideration of its absolute magnitude (and assuming the published distance), this is probably a mid O-type star (Fig. \[HIIregion\]). ]{}
- [Fourteen sources with H$\beta$ absorption and rising blue flux distributions that confirms them as hot stars. Given the faintness of these targets (22$<$$V$$<$25) we were unable to comment more on their spectral classifications, but from their locations in the CMD we suggest these are B-type (or late O-type) objects (Fig. \[blue\]). We tentatively classify a further 17 sources as candidate hot stars, again via detection of H$\beta$ absorption.]{}
- [Confirmation of two candidate AGB stars from @l16 as carbon stars, and confirmation of two further candidates as luminous cool (presumably AGB) stars via detection of CaT absorption in their spectra (Figs \[agb\_lee\] and \[CaT\]).]{}
- [Confirmation of CaT absorption in the RGB population, via co-adding the spectra of the brightest members (Fig. \[CaT\]).]{}
- [Two 100pc-scale ring structures that are traced by H$\alpha$-emission from the gas, with the H region located on the northern edge of the southern ring (Fig. \[emission\_maps\]).]{}
In addition, to investigate the expected temperatures and magnitudes of massive stars in Leo P, we employed evolutionary tracks from @s15 for single stars at the (near identical) metallicity of the IZw18 galaxy. The shift of the ZAMS to higher temperatures with decreasing metallicity has long been known, but we argue that the significance of this effect at very low metallicites could impact on the apparent magnitudes of massive stars. For a given stellar luminosity, hotter temperatures give larger bolometric corrections, which would mean that stars of a comparable mass in Leo P or IZw18 will have fainter absolute visual magnitudes than in the Galaxy (by $\sim$0.5mag in the example cases of the 9 and 20M$_\odot$ models considered here).
While giving us our first tantalysing view of high-mass stars in a very metal-poor environment, the upper mass function is so sparsely populated that we were not able to test the predictions of @s15 regarding core-hydrogen-burning supergiants and blueward evolution of stars from the ZAMS (caused by chemically-homogeneous evolution). Nonetheless, the MUSE observations have given us a first census of the high-mass population of Leo P, providing confirmed targets for long-exposure spectroscopy to obtain higher S/N and better spectral resolution. In particular, we highlight future observations of the O-type star (LP26) to derive its physical properties and a more robust estimate of its mass (cf. predictions from the IGIMF). Moreover, to interpret such observations we will also require synthetic spectra at this low metallicity from the latest model-atmosphere codes.
The MUSE data have also provided a first look at the cool, evolved population of Leo P. Quantitative analysis to determine their physical parameters (in part to calibrate low-$Z$ evolutionary models) and detailed kinematic analysis to investigate the dynamical properties of the cool population will again require ambitious spectroscopic follow-up.
Looking further into the future, the Extremely Large Telescope (ELT) will have the combination of both angular resolution and sensitivity to probe the evolved population of Leo P in much greater depth, for example with the first-light HARMONI visible and near-IR, integral-field spectrograph [@harmoni]. Ultimately we also want ultraviolet spectroscopy of the population of massive stars in Leo P, to investigate their wind properties as well as their physical parameters in this important low-metallicity regime. This is unrealistic at present with [*HST*]{}, but would be well within the grasp of the proposed Large Ultraviolet Optical Infrared Surveyor (LUVOIR) concept currently under study [e.g. see Section 4.3 from @luvoir].
Based on observations at the European Southern Observatory Very Large Telescope in programme 094.D-0346. We thank the referee for their constructive and helpful suggestions, and are also grateful to Olivia Jones, Anna McLeod and Rubén Sánchez-Janssen for useful discussions in the course of this work. DRW is supported by a fellowship from the Alfred P. Sloan Foundation, and acknowledges support from the Alexander von Humboldt Foundation. This research made use of Astropy, a community-developed core Python package for Astronomy [@astropy], and APLpy, an open-source plotting package for Python [@rb12].
[^1]: Hence the moniker of ‘Leo P’ from @g13, where ‘P’ refers to its pristine nature.
[^2]: http://archive.eso.org/wdb/wdb/adp/phase3\_main/form
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In this paper we prove the existence of phase transitions at finite temperature for $O(n)$ classical ferromagnetic spin models on infrared finite graphs. Infrared finite graphs are infinite graphs with $\lim_{m\to 0^+}{\overline
Tr}(L+m)^{-1}<\infty$, where $L$ is the Laplacian operator of the graph. The ferromagnetic couplings are only requested to be uniformly bounded by two positive constants. The proof, inspired by the classical result of Fröhlich, Simon and Spencer on lattices, is given through a rigorous bound on the average magnetization. The result holds for $n\ge 1$ and it includes as a particular case the Ising model.
---
¶[[P]{}]{} Ø[[O]{}]{} §[[S]{}]{} ¶[[P]{}]{} Ø[[O]{}]{} [H]{} Ł[[L]{}]{} [v]{}
****
Magnetization bound for classical spin models on graphs\
, [Davide Cassi]{}[^1] [and Alessandro Vezzani]{}[^2]\
Istituto Nazionale di Fisica della Materia, Unità di Parma\
Dipartimento di Fisica, Università di Parma\
and INFN, Gruppo Collegato di Parma,\
Viale delle Scienze, 43100 Parma, Italy\
.125cm
Introduction
============
The study of statistical models on graphs requires the introduction of new techniques and concepts with respect to the well known case of lattices. The lack of translational invariance in general implies inhomogeneity in local variables and makes useless the introduction of such a powerful tool as Fourier transforms in space. In addition, due to the absence of a natural definition of dimension, the statement itself of general results and theorems is rather problematic. The key point is to find a connection between geometry and physical properties. This problem can be studied in the framework of algebraic graph theory. However, phase transitions only occur on infinite graphs and the algebraic theory of infinite graphs is a very recent field of research in mathematics [@gerl; @woess]. Up to now only few classical results on lattices have been extended to graphs.
In this paper we deal with the generalization of the proof of existence of spontaneous magnetization at finite temperature for ferromagnetic models on lattices in $d\ge 3$. This result was proven in 1976 by Fröhlich, Simon and Spencer [@FSS] by a general approach based on Gaussian domination of correlation functions in the infrared regime. Such approach, as well as the formulation of the result itself, is deeply related to the translation invariance of lattices. Therefore here we modify the mathematical techniques and we use an alternative definition of the order parameter. In particular the concept of dimension is replaced by the asymptotic behavior of the Laplacian spectral density at low eigenvalues, a choice which is meaningful also in many other contexts. The structure of the paper is as follows: in section 2 we recall the basic definitions and theorems of graph theory to be used in the proof; in section 3 we give the fundamental known results about statistical models on graphs and state our theorem on the magnetization bound; finally in section 4 we present the proof.
Some mathematical properties of graphs
======================================
A graph G consists of a countable set of vertices (points, sites) $V(G)=\{i,j,k,\ldots \}$ and of a set $E(G)$ of unordered pairs of vertices. The generic element $(i,j)$ of $E(G)$ is called a link and its vertices are said to be nearest neighbors (adjacent). If $V(G)$ is finite, $G$ is called a finite graph and we will denote by $N$ the number of vertices of $G$.
A path connecting two vertices $i$ and $j$ is an alternating sequence of vertices and links $[i, (i,k), k, (k,h), h \ldots, w, (w,j), j]$.
A graph $G$ is connected if, given any two vertices $i, j \in V(G)$, it exists a path between $i$ and $j$.
Here we will deal only with connected graphs. The topology of $G$ is algebraically described by its adjacency matrix.
The adjacency matrix of $G$, $A_{ij}$, is given by: $$A_{ij}=\left\{
\begin{array}{cl}
1 & {\rm if } \ (i,j)\ \in E(G)\cr
0 & {\rm otherwise}\cr
\end{array}
\right .
\label{defA}$$
Most physical models on $G$ can be defined in terms of the Laplacian matrix $L_{ij}$:
The Laplacian matrix of $G$, $L_{ij}$ is given b: $$L_{ij} = z_i \delta_{ij} - A_{ij}
\label{defL}$$ where $z_i\equiv\sum_{j\in V(G)} A_{ij}$ is called the coordination number of $i$.
In the following we will consider graphs with coordination numbers bounded from above: $z_i \le z_{max} \in {\bf N} ~~ \forall i \in V(G)$. If $G$ is infinite, i.e. $V(G)$ is infinite, $L_{ij}$ can be considered as the representation of the Laplacian operator $L$ acting on $\ell_2(V(G))$. Under the previous conditions for $G$, it can be shown that $L$ is symmetric, non negative and bounded [@woess]. A relevant property of $L$ we will exploit in the following is the Schwinger-Dyson identity [@hhw]. Let us consider a diagonal bounded operator $\eta_{ij} \equiv \eta_i \delta_{ij}$ with $ \eta_i \in $[**C**]{} such that $(L + \eta)^{-1}_{ij}$ exists. Then: $$\sum_{i \in V(G)} \eta_i (L + \eta)^{-1}_{ij}=1 .
\label{SD}$$ This identity follows from the property $\sum_i L_{ij}=0$.
The graph $G$ is naturally provided with an intrinsic metric defined by the chemical distance on $G$.
The chemical distance $r_{i,j} \in
{\bf N}$ between two vertices $i$ and $j$ is the number of links in the shortest path connecting $i$ and $j$.
The intrinsic metric is the fundamental tool to define the thermodynamical limit on a infinite graph $G$, allowing us to introduce the generalized Van Hove spheres.
The Van Hove sphere $S_{i,r}$ of center $i$ and radius $r$ is the subset of $V(G)$, $S_{i,r}= \{j \in V(G)) ~| r_{ij} \le r\}$. We will define $N_{i,r}$ as the number of vertices in $S_{i,r}$.
The behaviour of $N_{i,r}$ as a function of $r$ characterizes the growth of $G$. In particular if $$A_i r^c \leq N_{i,r} \leq B_i r^c \
{\rm where}~ c, \ A_i ~ {\rm and}\ B_i \ {\rm are \ positive \ constants}
~\forall i.
\label{defc}$$ $G$ is said to have polynomial growth and we define the connectivity exponent $c$ to be the inf of the set of all $c$ satisfying (\[defc\]). Here we will consider graphs with polynomial growth. Indeed this condition is sufficient to guarantee the convergence of the restriction $O_{i,r}$ to $S_{i,r}$ of a bounded operator $O$ defined on $G$ to $O$ itself for $r\to\infty$.
Let $G$ be an infinite graph satisfying all previous conditions. Given a complex function of the vertices $\phi: V(G) \rightarrow
{\bf C}$ we define the thermodynamic limit average $\overline
\phi$ of $\phi$ as: $$\overline{\phi}=\lim_{r\rightarrow\infty}
{1\over \displaystyle N_{i,r}} \sum_{j\in S_{i,r}} \phi_j .
\label{deftd}$$ $\overline \phi$ can be easily shown to be independent of $i$.
In this work we will use also on finite graphs the symbol $\overline{\phi}$ to indicate the average over all vertices. We will define the average on a subset $V_\lambda \subseteq V(G)$: $$\overline{\chi_{V_{\lambda}} \phi}=\lim_{r\rightarrow\infty}
{1\over \displaystyle N_{i,r}}
\sum_{i\in S_{i,r}}\chi_{V_{\lambda}}(i)\phi_{i} .
\label{deftd2}$$ where $\chi_{V_{\lambda}}(i)$ is the characteristic function of $V_{\lambda}$. We define the measure $\mu(V_{\lambda})$ of $V_{\lambda}$: $$\mu(V_{\lambda})= \overline {\chi_{V_\lambda}} .
\label{measure}$$ More generally, the thermodynamic limit average of a $k$ variable function $\Phi: V((G))^k \rightarrow $[**C**]{} is defined by: $$\overline \Phi =\lim_{r\rightarrow\infty}{1\over \displaystyle N_{i,r}}
\sum_{i_1,\ldots, i_k \in S_{i,r} } \Phi_{i_1,\ldots, i_k} .$$ We define the thermodynamic limit average trace of an infinite matrix $B_{ij}$ as: $$\overline{\Tr}B= \overline b
\label{dtr}$$ where $b_i \equiv B_{ii}$, and its restriction to a subset $V_\lambda \subseteq V(G)$: $$\overline{\Tr}_{V_{\lambda}} B \equiv \overline{\chi_{V_{\lambda}} b}$$
The large scale topology affecting the critical behavior of statistical models on graphs can be characterized by the properties of the trace of $(L+M)^{-1}$, where $M_{ij}= m ~\delta_{ij}$ is a real and positive diagonal matrix.
An infinite graph is infrared finite if $$\lim_{m \rightarrow 0^+} \overline{Tr}(L+M)^{-1} < \infty.
\label{limtr}$$ An infinite graph is infrared infinite if: $$\lim_{m \rightarrow 0^+} \overline{Tr}(L+M)^{-1} = \infty.
\label{ilimtr}$$
It can be shown that if we substitute the constant matrix $M$ with $M'_{ij} = m ~b_i ~\delta_{ij}$ where $0<\epsilon < b_i < K $ the infrared behavior for $m\to 0^+$ does not change [@debole].
In dealing with statistical models the adjacency matrix $A_{ij}$ is often generalized to the ferromagnetic interaction matrix $J_{ij}$:
$$J_{ij} = J_{ji} = \left\{
\begin{array}{cl}
J_{ij} & \ {\rm with\ } 0< \epsilon\leq J_{ij}\leq J\ {\rm if\ } A_{ij}=1 \cr
0 & \ {\rm if\ } A_{ij}=0 \cr
\end{array}
\right .
\label{defJ}$$
In this case the Laplacian generalizes to: $$L_{ij} = z_i \delta_{ij} - J_{ij}
\label{defgL}$$ where $ z_i =\sum_j J_{ij}$.
The generalized Laplacian has the same properties of the Laplacian associated to $A$: it is symmetric, positive and bounded, it satisfies the Schwinger-Dyson identity and finally it has the same infrared behaviour [@forte].
O(n) classical spin models on graphs
====================================
An important class of classical statistical spin models on graphs is defined by the Hamiltonian: $$H' = - {1\over 2}\sum_{i,j\in V(G)} J_{ij}
\vsigma_i \cdot \vsigma_j -\vh \cdot \sum_{i \in V(G)} \vsigma_i
\label{defH}$$ where $J_{ij}$ is a ferromagnetic interaction matrix defined on the graph $G$ satisfying conditions (\[defJ\]) and $\vsigma_j$ are n-dimensional real unit vectors $\vsigma_i \equiv (\sigma^1_i,\ldots,\sigma^n_i)$ defined on each vertex $i$ and satisfying the constraints: $$\vsigma_i^2=1 ~~ \forall i .
\label{const}$$ For $n=1$ $H'$ defines the Ising model which is invariant under the discrete symmetry group ${\bf Z_2}$, while for $n \geq 2$ $H'$ represents an $O(n)$ model with continuous symmetry. Finally $\vh \equiv (h,0,\ldots,0)$ is an external magnetic field coupled to $\vsigma_i$. In the following we will set $h>0$.
Due to (\[const\]) $H'$ is equivalent up to an additive constant to: $$H = {1\over 4}\sum_{ij} J_{ij} (\vsigma_i -\vsigma_j)^2
-\vh \cdot \sum_i \vsigma_i
= {1\over 2}\sum_{ij} L_{ij} \vsigma_i \cdot \vsigma_j
-\vh \cdot \sum_i \vsigma_i
\label{newH}$$ where the Laplacian operator defined in (\[defgL\]) has been introduced.
The Boltzmann measure $\mu_{\beta,h} (\vsigma)$ is given by: $$d \mu_{\beta,h}(\vsigma) \equiv \prod_i d\vsigma_i e^{-\beta H}~
\label{boltz}$$ where $\beta=1/k_B~T$, with $\vsigma_i$ satisfying the constraints (\[const\]). The order parameter of the model is the average magnetization in the $\vh$ direction: $$M (\beta,h) = {\displaystyle \int d \mu_{\beta,h}(\vsigma) ~ \overline
{\sigma^1} \over \displaystyle \int d \mu_{\beta,h}(\vsigma)} .
\label{defm}$$ On lattices, in the thermodynamical limit for $n\ge 2$ it is known that the limit for $h\rightarrow 0$ of the order parameter vanishes for $T > 0$ in the infrared infinite case (i.e. if the Euclidean dimension $d$ is 2 or 1) [@mw]. In the infrared finite case ($d \ge 3$) for low enough temperatures the limit is positive [@FSS], implying the existence of phase transitions. On graphs we have $\lim_{h\rightarrow 0}M(\beta,h)=0$ for every $\beta$, if for each positive measure subset $V_{\lambda} \subseteq V(G)$ we have [@mwg]: $$\lim_{b\rightarrow 0^+}\overline{\Tr}_{V_{\lambda}}(L+b)^{-1}= \infty~.
\label{mermin}$$ For Ising models on lattices the limit of the order parameter vanishes for all temperatures only in one dimension. There are no analogous results for generic graphs.
In this work we will show that, for the classical spin models (\[defH\]) on infrared finite graphs, below a certain temperature $M (\beta,h)$ satisfies the bound: $$\lim_{h\rightarrow 0}M (\beta,h) > k >0.
\label{bound}$$ This result for $O(n)$ models is not the exact inversion of [@mwg]. Indeed one should prove that if there is a positive measure subset $V_{\lambda}$, for which the limit in (\[mermin\]) is finite, then the bound (\[bound\]) is satisfied. Here the proof is given only for the case where the limit in (\[mermin\]) is finite for $V_{\lambda} \equiv G$. A complete inversion of [@mwg] will be given in a following paper [@invmwg]. We remark that this is the first result for the Ising model on a generic graph.
The magnetization bound
=======================
In this section we will prove the magnetization bound (\[bound\]) for the classical spin model (\[defH\]) according to the following strategy:
\(A) first of all we will prove some useful inequalities for the Laplacian on finite graphs;
\(B) then we will introduce for the constraint (\[const\]) an integral representation allowing us to perform a Gaussian integral with respect to the spin variables $\vsigma_i$;
\(C) we will take the thermodynamic limit of the inequalities (A) and in the expressions for $M$ and for the global constraint;
\(D) exploiting a saddle point technique for large $\beta$, we will apply the inequalities in (C) to $M$ getting the bound (\[bound\]), for infrared finite graphs.
0.3truecm As for the point (A):
The following inequalities hold for finite graphs, with $h_{ij} \equiv h ~\delta_{ij}$ and $\alpha \equiv \alpha_i ~\delta_{ij}$ with $\alpha_i \in {\bf R}~\forall i$:
$$1 \geq \Re\left( {h\over N} \sum_{ij}(L+h+i\alpha)_{ij}^{-1} \right)\geq
\Re\left( {h^2\over N} \sum_{ij}(L+h+i\alpha)_{ij}^{-2}\right) .
\label{ineq1}$$
$$0 \leq {1\over N}\sum_{ikj}\alpha_i(L+h-i\alpha)_{ik}^{-1}(L+h+i\alpha)_{kj}^{-1}\alpha_j
\leq\Re\left({1\over N h}\sum_{ij}\alpha_i(L+h+i\alpha)_{ij}^{-1}\alpha_j\right)
\label{ineq4}$$
$$0 \leq {1\over N}\Re\Tr_{V_{\lambda}}\left(L+h+i\alpha)^{-1}\right) \leq
{1\over N}\Re\Tr_{V_{\lambda}}\left((L+h)^{-1}\right) \leq
{1\over N}\Re\Tr\left((L+h)^{-1}\right)
\label{ineq2}$$
$${1\over N}| \Im \Tr_{V_{\lambda}} (L+h+i\alpha)^{-1}|
\leq {1\over 2N} \Tr _{V_{\lambda}}(L+h)^{-1} \leq {1\over2 N} \Tr (L+h)^{-1}
\label{ineq3}$$
where $V_{\lambda}$ is a generic subset $V(G)$ and $$\Tr_{V_{\lambda}} B \equiv \sum_i B_{ii}~\chi_{V_\lambda}(i)~,$$ $${1\over N} \sum_i \alpha_i^2 \leq f(\alpha) \left( {1\over N}\sum_{ikj}
\alpha_i(L+h-i\alpha)_{ik}^{-1}(L+h+i\alpha)_{kj}^{-1}\alpha_j\right)^{1/2}
\label{ineq5}$$ where $f(\alpha) \leq 1/N \sum_i (\alpha_i^2(l_{max}+h)^2+\alpha_i^4)$ and $l_{max}$ is the maximum eigenvalue of $L$.
The proof of these inequalities will be given at the end of this section. $\Box$
As for (B), we prove the following:
On a finite graph the expression of the average magnetization (\[defm\]) and of the constraint: $$1 = {\displaystyle \int \overline{\vsigma^2}
d\mu_{\beta,h}(\vsigma)
\over \displaystyle \int d\mu_{\beta,h}(\vsigma)}
\label{glconst}$$ can be written in the following way:
$$M(\beta,h) ={1\over Z}{\int d \mu_{\beta,h}(\alpha)
\ {h\over N} \sum_{kj}(L+h+i\alpha)^{-1}_{kj}}
\label{defm1}$$
$$1 = {1\over Z} \int d \mu_{\beta,h}(\alpha) \left(
{n \over \beta N} {\Tr}(L+h+i\alpha)^{-1} +
{h^2\over N} \sum_{ij}(L+h+i\alpha)^{-2}_{ij} \right)
\label{glconst1}$$
where: $$d \mu_{\beta,h}(\alpha)\equiv
\prod_{i\in G} d \alpha_i e^{i S_{\beta,h}}
\label{dmu}$$ $$i S_{\beta,h}\equiv
-{n\over 2} \Tr\left(\ln (L+h+i\alpha)\right)
+{\beta\over 2}
\left(i\sum_i\alpha_i+h^2\sum_{ij}(L+h+i\alpha)^{-1}_{ij}\right)$$ [and]{} $$Z\equiv \int d \mu_{\beta,h}(\alpha_i)$$
Let us write the constraints (\[const\]) using the complex integral representation of the delta function: $$\delta(x^2)=\int d\alpha \exp(-i\alpha x^2/2-\epsilon x^2/2)$$ where $\epsilon$ is a real arbitrary constant. We will put $\epsilon=h\beta$. Substituting the expression for $\delta$ in (\[defm\]) and (\[glconst\]) we obtain: $$M(\beta,h)= {\displaystyle \int d \vsigma \ d\alpha \ \hsigma
\exp\left(-\beta({1 \over 2} \sum_{ij}L_{ij}{\vsigma_i}\cdot{\vsigma_j}
-h \sum_i \sigma_i^1)
-\sum_i(i {\alpha_i \over 2} ({\vsigma_i}^2-1)
-{h \beta \over 2} {\vsigma_i^2})\right)
\over \displaystyle \int d\vsigma \ d\alpha
\exp\left(-\beta({1 \over 2} \sum_{ij}L_{ij}{\vsigma_i}\cdot{\vsigma_j}
-h \sum_i \sigma_i^1)
-\sum_i(i {\alpha_i \over 2} ({\vsigma_i}^2-1)
-{h \beta \over 2} {\vsigma_i^2})\right)}$$ and $$1 = {\displaystyle \int d \vsigma \ d\alpha \ \overline{\vsigma^2}
\exp\left(-\beta({1 \over 2} \sum_{ij}L_{ij}{\vsigma_i}\cdot{\vsigma_j}
-h \sum_i \sigma_i^1)
-\sum_i(i {\alpha_i \over 2} ({\vsigma_i}^2-1)-
{h \beta \over 2} {\vsigma_i^2})\right)
\over \displaystyle \int d\vsigma \ d\alpha
\exp\left(-\beta({1 \over 2} \sum_{ij}L_{ij}{\vsigma_i}\cdot{\vsigma_j}
-h \sum_i \sigma_i^1)
-\sum_i(i {\alpha_i \over 2} ({\vsigma_i}^2-1)-
{h \beta \over 2} {\vsigma_i^2})\right)}$$ where $d\alpha \equiv \prod_i d\alpha_i$ and $d\vsigma \equiv \prod_i d\vsigma_i$. Substituting in both integral $\alpha_i$ with $\beta\alpha_i$ and $\vsigma_i$ with $\vsigma_i/\sqrt{\beta}$, we can perform the Gaussian integral on the variables $\vsigma_i$, obtaining in this way (\[defm1\]) and (\[glconst1\]). $\Box$
Now we consider an infinite graph $G$ and using the thermodynamic limit, we will extend to $G$ the Lemmas (1) and (2). This is point (C) in our plan.
For an infinite graph the inequalities (\[ineq1\]), (\[ineq4\]), (\[ineq2\]), (\[ineq3\]), (\[ineq5\]) become:
$$1 \geq \Re\left( h \overline{(L+h+i\alpha)^{-1}} \right)\geq
\Re\left( h^2 \overline {(L+h+i\alpha)^{-2}}\right) .
\label{ineq1'}$$
$$0 \leq \overline{\alpha(L+h-i\alpha)^{-1}(L+h+i\alpha)^{-1}\alpha}
\leq\Re\left(\overline{\alpha(L+h+i\alpha)^{-1}\alpha}\right)
\label{ineq4'}$$
$$0 \leq \Re\left(\overline{\Tr_{V_{\lambda}}}(L+h+i\alpha)^{-1}\right) \leq
\Re\left(\overline{\Tr_{V_{\lambda}}}(L+h)^{-1}\right)\leq
\Re\left(\overline{\Tr}(L+h)^{-1}\right)
\label{ineq2'}$$
$$| \Im \overline{\Tr_{V_{\lambda}}} (L+h+i\alpha)^{-1}| \leq
{1\over 2}~ \overline{\Tr_{V_{\lambda}}} (L+h)^{-1} \leq
{1\over 2}~ \overline{\Tr} (L+h)^{-1}
\label{ineq3'}$$
here $V_{\lambda}$ is a generic positive measure subset of G. $$\overline{\alpha^2} \leq f(\alpha) \left( \overline{
\alpha(L+h-i\alpha)^{-1}(L+h+i\alpha)^{-1}\alpha}\right)^{1/2}
\label{ineq5'}$$ where $f(\alpha) \leq \overline{(\alpha^2(l_{max}+h)^2+\alpha^4)}$.
These inequalities are easily obtained by applying the inequalities (\[ineq1’\]-\[ineq5’\]) to the restriction to the Van Hove spheres $S_{i,r}$ of $L$, $\alpha$ and $h$ and taking the thermodynamical limit by letting $r\to\infty$. $\Box$
Now we will prove the main theorem showing the existence of the bound for the magnetization.
In an infinite and infrared finite graph $G$, $\forall \epsilon >0 \ \exists
\bar{\beta} >0$ such that $\forall \beta \geq \bar{\beta}$: $$\lim_{h\rightarrow 0}M(\beta,h)\geq 1 - \epsilon
\label{bound1}$$
We consider, in the thermodynamic limit, the expressions (\[defm1\]) and (\[glconst1\]) for $M(\beta,h)$ and for the global constraint, with the measure given by (\[dmu\]). At the end of this section we will prove the following lemma:
Consider a positive measure subset $V_{\lambda}$ of $V(G)$. If $G$ is infrared finite, the quantity $$\overline {\chi_{V_{\lambda}} S'_{\beta,h}}
\label{sadle1}$$ where $S'_{\beta,h}(i)= {\partial\over{\partial \alpha_i}} S_{\beta,h}$ is bounded for all $V_{\lambda}$ when $\beta\rightarrow\infty$ only if the $\alpha_i$ satisfy the condition: $$\overline{\alpha^2}=0 .
\label{sadle2}$$ This condition is true also in the limit $h\rightarrow 0$. $\Box$
We will call the set of values $\alpha_i \equiv \{\alpha_1, \alpha_2, \ldots\}$ satisfying (\[sadle2\]) the stationary point of the statistical weight $d \mu_{\beta,h}(\alpha)$. Now we separate the space of integration variables $\alpha_i$ into two subsets $\Gamma_{\beta,h}(\alpha)$ and its complement $\overline{\Gamma_{\beta,h}(\alpha)}$ such that $\Gamma_{\beta,h}(\alpha)$ is the region around the stationary point where the real part of the measure $d \mu_{\beta,h}(\alpha)$ is positive. We define: $$\begin{aligned}
\epsilon(\beta,h) & \equiv & {1\over Z} \int_{\overline{\Gamma_{\beta,h}}}
d\Re(\mu_{\beta,h}(\alpha)) \overline{(L+h+i\alpha)^{-1}}
+ {1\over Z} \int d\Im(\mu_{\beta,h}(\alpha))
\overline{(L+h+i\alpha)^{-1}} \nonumber\\
\epsilon'(\beta,h) & \equiv & {1\over Z} \int_{\overline{\Gamma_{\beta,h}}}
d\Re(\mu_{\beta,h}(\alpha)) \left( {n \over \beta}\overline{\Tr}(L+h+i\alpha)^{-1}
+ h^2 \overline{(L+h+i\alpha)^{-2}} \right) \label{epsilon}\\
& \ & + {1\over Z} \int d\Im(\mu_{\beta,h}(\alpha))
\left({n \over \beta}\overline{\Tr}(L+h+i\alpha)^{-1} +
h^2 \overline{(L+h+i\alpha)^{-2}} \right) \nonumber\end{aligned}$$ Lemma 4 and the saddle-point theorem imply that: $$\forall \ \epsilon >0 \ \exists\beta'\in {\bf R}~\ {\rm such ~that}\
\forall \ \beta \geq \beta':
\lim_{h\rightarrow 0}|\epsilon(\beta,h)|\leq\epsilon/3 \ {\rm and}
\ \lim_{h\rightarrow 0}|\epsilon'(\beta,h)|\leq\epsilon/3 .
\label{beta'}$$ Now for $M(\beta,h)$ we have: $$\begin{aligned}
M(\beta,h) & = & \epsilon(\beta,h) + {1\over Z} \int_{\Gamma_{\beta,h}(\alpha)}
d\Re(\mu_{\beta,h}(\alpha))
\overline{(L+h+i\alpha)^{-1}} \label{M1}\\
& = & \epsilon(\beta,h) + {1\over Z} \int_{\Gamma_{\beta,h}(\alpha)}
d\Re(\mu_{\beta,h}(\alpha))
\ \Re\left( h \overline{(L+h+i\alpha)^{-1}} \right)
\label{M2}\end{aligned}$$ The imaginary part of the second term in (\[M1\]) does not provide any contribution to the integral because $M(\beta,h)$ is a real quantity. In analogous way for the constraint (\[glconst1\]) we get: $$1= \epsilon'(\beta,h) + {1\over Z} \int_{\Gamma_{\beta,h}(\alpha)}
d\Re(\mu_{\beta,h}(\alpha))
\Re \left( {n \over \beta} \overline{\Tr}(L+h+i\alpha)^{-1}
+ h^2 \overline{(L+h+i\alpha)^{-2}} \right)
\label{gl1}$$ Now in (\[M2\]) we deal with a real quantity averaged with respect to a positive measure and we can use the inequalities of Lemma 3. In particular applying (\[ineq1’\]) in (\[M2\]) we get: $$M (\beta,h) \geq \epsilon(\beta,h) + {1\over Z} \int_{\Gamma_{\beta,h}
(\alpha)}
d\Re(\mu_{\beta,h}(\alpha))
\ \Re\left( h^2 \overline{(L+h+i\alpha)^{-2}} \right)$$ With this inequality and the expression for the global constraint (\[gl1\]) we get: $$M (\beta,h)\geq 1 - \epsilon'(\beta,h) + \epsilon(\beta,h)-
\int_{\Gamma_{\beta,h}(\alpha)} d\alpha \ \Re(\mu_{\beta,h}(\alpha))
{n \over \beta} \Re\left(\overline{\Tr}(L+h+i\alpha)^{-1}\right).
\label{M3}$$ Now we take the limit $h\rightarrow 0$ and we study the terms of this inequalities for large $\beta$. For $\epsilon(\beta,h)$ and $\epsilon'(\beta,h)$ the behaviour is given by (\[beta’\]). For the integral term in (\[M3\]) we can exploit the inequality (\[ineq2’\]) getting: $$0 \leq
\int_{\Gamma_{\beta,h}(\alpha)} d\alpha \ \Re(\mu_{\beta,h}(\alpha))
{n \over \beta} \Re\left(\overline{\Tr}(L+h+i\alpha)^{-1}\right) \leq
{n \over \beta}\overline{\Tr}(L+h)^{-1}
\int_{\Gamma_{\beta,h}(\alpha)} d\alpha \ \Re(\mu_{\beta,h}(\alpha)) .$$ The saddle-point theorem implies that: $$\exists \beta''\in{\bf R} ~\ {\rm such ~that}\ \forall \beta \geq \beta'':
\lim_{h\rightarrow 0} {1 \over Z}
\int_{\Gamma_{\beta,h}(\alpha)} d\Re(\mu_{\beta,h}(\alpha))
\geq 1/2.
\label{beta''}$$ If for each $\epsilon$ we choose $$\beta \geq \beta''~~{\rm and}~~\beta \geq {3 \over \epsilon}
2n \lim_{h\rightarrow 0}\overline{\Tr}(L+h)^{-1}=\beta'''
\label{beta'''}$$ we get: $$\lim_{h\rightarrow 0}{n\over \beta Z}
\int_{\Gamma_{\beta,h}(\alpha)} d\Re(\mu_{\beta,h}(\alpha))
\Re\left(\overline{\Tr}(L+h+i\alpha)^{-1}\right)\leq \epsilon/3.
\label{epsilon'}$$ Finally fixed $0\leq\epsilon\leq1$, if we choose $\bar{\beta}=\max(\beta',\beta'',\beta''')$, we have that $\forall\beta>\bar{\beta}$ $M(\beta,h)$ satisfies condition (\[bound1\]). Note that from (\[beta”’\]) we have that $\beta'''$ and then $\bar{\beta}$ are bounded only for infrared finite graphs. $\Box$
Inequalities
------------
Now we will prove the inequalities (\[ineq1\]), (\[ineq4\]), (\[ineq2\]), (\[ineq3\]) and (\[ineq5\]) for finite graphs. On a finite graph the set of all real functions of the vertices $\phi_i$ has a vector space structure. Let us define $\langle \phi| = (\phi_1,\ldots,\phi_N)$ and $|\phi \rangle = (\phi_1,\ldots,\phi_N)^t$. The scalar product $\langle
\phi | \psi \rangle$ is defined by: $$\langle \phi | \psi \rangle = \sum_i \phi_i \psi_i.$$ The matrices $L$, $A$, $h$ and $\alpha$ are operators on this space. The matrix $L$ is diagonalizable by a real transformation, and its eigenvalues $l$ satisfy $0\leq l\leq l_{max}$. Let us prove the following:
On a finite graph, it is possible to introduce a real operator $B$ defined by: $$B^t(L+h)B=I\ \ \ \ B^t\alpha B=c
\label{B}$$ where $c$ is a real diagonal operator and $I$ is the identity operator. Furthermore $B$ satisfies the properties: $$(L+h+i\alpha)^{-1}=B(1+ic)^{-1}B^t\ \ \ \ \ (L+h)^{-1}=BB^t
\label{propB}$$
$$\|B^tB\|={1\over h}
\label{normB}$$
where $$\|B^tB\|=
\sup_{\phi}{\langle\phi |B^tB |\phi\rangle \over \langle\phi| \phi\rangle}$$
The existence of $B$ is proved in [@gant] where it is also shown that is real. Properties (\[propB\]) can be immediately obtained by (\[B\]). Therefore it is easy to get the exact expression for $B$: $$B=TAT'$$ where $T$ is the orthogonal transformation that diagonalize $L$; $A$ is the matrix $A_{km}=(1/\sqrt{l_k+h}) \delta_{km}$, where $l_k$ is the eigenvalue of $L$ relative to the eigenvector $k$; finally $T'$ is the orthogonal operator that diagonalize the symmetric matrix $AT^t\alpha TA$. $B$ is not an orthogonal transformation but we compute its norm: $$\|B^tB\|=
\sup_{\phi}{\langle\phi| B^tB |\phi\rangle \over \langle\phi |\phi\rangle}=
\sup_{\phi}{\langle\phi |A^2 |\phi\rangle \over \langle\phi |\phi\rangle}
={1\over h}$$ This proves property (\[normB\]). $\Box$
[*Proof of Lemma 1*]{} Let $\langle\phi|$ be a generic vector of the space of the functions of vertices. We have: $$\begin{aligned}
\Re \langle \phi|h^2 (L+h+i\alpha)^{-2}|\phi\rangle
& = & \Re h^2 \langle \phi|B(1+ic)^{-1}B^tB(1+ic)^{-1}B^t|\phi\rangle\nonumber\\
& = & h^2 \langle \phi|B(\Re(1+ic)^{-1})B^tB(\Re(1+ic)^{-1})B^t|\phi\rangle\nonumber\\
& \ & - h^2 \langle \phi|B(\Im(1+ic)^{-1})B^tB(\Im(1+ic)^{-1})B^t|\phi\rangle\nonumber\\
& \leq & h^2 \langle \phi|B(\Re (1+ic)^{-1})B^tB(\Re(1+ic)^{-1})B^t|\phi\rangle\nonumber\\
& \leq & h \langle \phi|B(\Re (1+ic)^{-1})(\Re(1+ic)^{-1})B^t|\phi\rangle\nonumber\\
& \leq & h \langle \phi|B(1+c^2)^{-2}B^t|\phi\rangle\nonumber\\
& \leq & h \langle \phi|B(1+c^2)^{-1}B^t|\phi\rangle\nonumber\end{aligned}$$ where we used properties (\[propB\]) and (\[normB\]). Then we have: $$\begin{aligned}
\Re \langle \phi|h (L+h+i\alpha)^{-1}|\phi\rangle
& = & h \langle \phi|B(\Re (1+ic)^{-1})B^t|\phi\rangle\nonumber\\
& = & h \langle \phi|B(1+c^2)^{-1}B^t|\phi\rangle\nonumber\\
& \leq & h \langle \phi|BB^t|\phi\rangle \leq
\langle \phi|\phi\rangle\nonumber \end{aligned}$$ so we get: $$\Re \langle \phi|h^2 (L+h+i\alpha)^{-2}|\phi\rangle\leq
\Re \langle \phi|h (L+h+i\alpha)^{-1}|\phi\rangle \leq \langle \phi|\phi\rangle .
\label{ap2a}$$ Now choosing in (\[ap2a\]) $\langle\phi|$ to be the constant vector $\alpha_i=1$, we prove (\[ineq1\]).
To prove the inequality (\[ineq4\]) we consider the following expression: $$\begin{aligned}
0 \leq \langle \phi|(L+h-i\alpha)^{-1}(L+h+i\alpha)^{-1}|\phi\rangle
& = & \langle \phi|B(1-ic)^{-1}B^tB(1+ic)^{-1}B^t|\phi\rangle\nonumber\\
& \leq & {1\over h} \langle \phi|B(1-ic)^{-1}(1+ic)^{-1}B^t|\phi\rangle\nonumber\\
& \leq & {1\over h} \langle \phi|B(1+c^2)^{-1}B^t|\phi\rangle\nonumber\\
& \leq & {1\over h} \Re \langle \phi|B(1+ic)^{-1}B^t|\phi\rangle\nonumber\\
& \leq & {1\over h} \Re \langle \phi|(L+h+i\alpha)^{-1}|\phi\rangle .
\label{ap2b}\end{aligned}$$ where we used (\[propB\]) and (\[normB\]). Putting $\phi_i =\alpha_i$ in (\[ap2b\]), we get (\[ineq4\]).
To get inequalities (\[ineq2\]) and (\[ineq3\]) we introduce the base $\langle k|$, $k=1,\ldots,N$ in the vector space of the functions of the sites given by: $$\langle k|=\left\{
\begin{array}{cl}
k_i=1 & {\rm if} \ i=k \\
0 & {\rm in\ all\ others\ vertices.}\\
\end{array}
\right .$$ We have: $$(L+h+\alpha)^{-1}_{kk} = \langle k|(L+h+\alpha)^{-1}|k\rangle=
\langle k|B(1+ic)^{-1}B^t |k\rangle
\label{lreim}$$ For the real part of (\[lreim\]), with properties (\[propB\]) we get: $$0\leq \Re (L+h+\alpha)^{-1}_{kk} = \langle k|B(1+c^2)^{-1}B^t |k\rangle\leq
\langle k|BB^t |k\rangle=(L+h)^{-1}_{kk}$$ If we sum over all $k\in V_{\lambda}$ we get: $$0\leq\Re \Tr_{V_{\lambda}} (L+h+\alpha)^{-1}\leq
\Tr_{V_{\lambda}}(L+h)^{-1} \leq \Tr (L+h)^{-1}$$ This proves (\[ineq2\]). Now for the imaginary part of (\[lreim\]) we have: $$|\Im (L+h+\alpha)^{-1}_{kk}| = |\langle k|Bc(1+c^2)^{-1}B^t |k\rangle|\leq
\langle k|B|c|(1+c^2)^{-1}B^t |k\rangle\leq {1\over 2}
\langle k|BB^t |k\rangle={1\over 2} (L+h)^{-1}_{kk}$$ If we sum over all $k\in V_{\lambda}$ we get: $$|\Im \Tr_{V_{\lambda}} (L+h+\alpha)^{-1}|\leq
\sum_{i\in V_{\lambda}} |\Im (L+h+\alpha)^{-1}_{kk}|~
{1\over 2} ~\Tr_{V_{\lambda}}(L+h)^{-1} \leq {1\over 2} \Tr (L+h)^{-1}$$ In this way we proved inequality (\[ineq3\]).
For (\[ineq5\]) we have: $$\begin{aligned}
\sum_i \alpha_i^2 & = & \langle\alpha|\alpha\rangle
= \langle\alpha|(L+h+i\alpha)(L+h+i\alpha)^{-1}|\alpha\rangle \leq \nonumber\\
& \leq & \langle\alpha|(L+h+i\alpha)(L+h-i\alpha)|\alpha\rangle^{1/2}
\langle\alpha|(L+h-i\alpha)^{-1}(L+h+i\alpha)^{-1}|\alpha\rangle^{1/2} \nonumber\\
& \leq & f(\alpha)
\left(\sum_{ikj}
\alpha_i(L+h-i\alpha)_{ik}^{-1}(L+h+i\alpha)_{kj}^{-1}\alpha_j\right)^{1/2}=0
\nonumber\end{aligned}$$ where $f(\alpha)=\langle\alpha|(L+h)^2+\alpha^2|\alpha\rangle^{1/2}
\leq [\sum_i (\alpha_i^2(l_{max}+h^2+\alpha_i^4)]^{1/2}$ and this completes the proof of Lemma 1. $\Box$
The saddle-point condition
--------------------------
[*Proof of Lemma 4.*]{} Let us compute the quantity (\[sadle1\]). We have: $$\overline {\chi_{V_{\lambda}} S'_{\beta,h}}
= -{n\over 2}~~ \overline{\Tr}_{V_{\lambda}} (L+h+i\alpha)^{-1}
+{ \beta\over 2} \left(\mu(V_{\lambda}) -
h^2 \overline{(L+h+i\alpha)^{-1}\chi(V_\lambda)
(L+h+i\alpha)^{-1}}\right)$$ where $\chi(V_\lambda)_{ij}=\chi_{V_\lambda}~\delta_{ij}$. The first term in this expression is always bounded, also when $h\rightarrow 0$ for infrared finite graphs; inequalities (\[ineq2\]) and (\[ineq3\]) can be used for the proof. Then for $\beta \rightarrow \infty$ (\[sadle1\]) is bounded only when the condition: $$\mu(V_{\lambda}) =
h^2 \overline{(L+h+i\alpha)^{-1}\chi(V_\lambda)
(L+h+i\alpha)^{-1}}
\label{sadle'}$$ is satisfied. Let us consider the particular case where we put in (\[sadle’\]) $V_{\lambda}=G$. We get $$1=
h^2 \overline{(L+h+i\alpha)^{-2}}.
\label{sadle3}$$ Let us show that (\[sadle3\]) is satisfied only when (\[sadle2\]) holds. First of all we have that (\[sadle3\]) implies: $$\Re\left(h \overline{(L+h+i\alpha)^{-1}}\right)=1 .
\label{sadle4}$$ Indeed with the inequality (\[ineq1’\]) we get: $$1=\Re\left( h^2 \overline {(L+h+i\alpha)^{-2}}\right)
\leq \Re\left(h \overline{(L+h+i\alpha)^{-1}}\right)
\leq 1 .$$ Now using the Schwinger-Dyson identity (\[SD\]) in (\[sadle4\]) we have: $$1 = \Re\left(h \overline{(L+h+i\alpha)^{-1}}\right)
= 1 + \Re\left(\overline{(L+h+i\alpha)^{-1}(-i\alpha)}\right)$$ and then: $$\begin{aligned}
0 & = &
\Re\left(\overline {(L+h+i\alpha)^{-1}(-i\alpha)}\right)\nonumber\\
& = & \Re\left(-{i \overline{\alpha}\over h} +
{1\over h} \overline{(-i\alpha)
(L+h+i\alpha)^{-1}(-i\alpha)}\right)\nonumber\\
& = & \Re\left({1\over h}
\overline{\alpha(L+h+i\alpha)^{-1}\alpha}\right) .
\label{sadle5} \end{aligned}$$ where in the first step we used again the Schwinger-Dyson identity. Now applying the inequality (\[ineq4’\]) to (\[sadle5\]) we get: $$0 \leq \overline{\alpha(L+h-i\alpha)^{-1}(L+h+i\alpha)^{-1}\alpha}
\leq\Re\left({1\over h}\overline{\alpha_i(L+h+i\alpha)^{-1}\alpha}\right)$$ $$\overline{\alpha(L+h-i\alpha)^{-1}(L+h+i\alpha)^{-1}\alpha}=0
\label{sadle''}$$ Let us consider the inequality (\[ineq5’\]). We have: $$\overline{\alpha^2} \leq f(\alpha)
\left(\overline{
\alpha(L+h-i\alpha)^{-1}(L+h+i\alpha)^{-1}\alpha}\right)^{1/2}=0~.$$ Notice that $f(\alpha)$ is bounded if $\overline{\alpha^4}$ is bounded. So we proved that the condition (\[sadle2\]) for the variables $\alpha_i$ must be satisfied, if the quantity (\[sadle1\]) is bounded when $\beta \rightarrow\infty$. This proof also holds when $h\rightarrow 0$. Indeed in this case the stationary condition (\[sadle3\]) can be expressed by (\[sadle”\]) and the inequality (\[ineq5’\]) holds.
Now in order to complete our proof we must verify that also for a generic $V_{\lambda}$ (\[sadle1\]) is bounded at the stationary point when $\beta\rightarrow\infty$. In this way we will show that (\[sadle2\]) is not only a necessary condition but also a sufficient one. If we evaluate (\[sadle1\]) with the constraint (\[sadle2\]) we get: $$\begin{aligned}
& \ & i \left(-{n\over 2} \overline{\Tr}_{V_{\lambda}} (L+h)^{-1}
+{ \beta\over 2} \left(\mu(V_{\lambda}) -
h^2 \overline{(L+h)^{-1} \chi(V_\lambda)
(L+h)^{-1}}\right)\right)\nonumber\\
& = & i \left(-{n\over 2} \overline{\Tr}_{V_{\lambda}} (L+h)^{-1}
+{ \beta\over 2} \left(\mu(V_{\lambda}) -
\overline{\chi_{V_\lambda}} \right)\right)\nonumber\\
& = & -i~{n\over 2} \overline{\Tr}_{V_{\lambda}} (L+h)^{-1}
\nonumber\end{aligned}$$ where we used again the Schwinger-Dyson identity. For infrared finite graphs the last expression is always bounded (also when $h\rightarrow 0$), see (\[limtr\]). This completes our proof. $\Box$.
[99]{}
Burioni, R. and Cassi, D.: Geometrical universality in vibrational dynamics. Mod. Phys. Lett. B [**11**]{}, 1095-1101, (1997) Burioni, R. and Cassi, D.: Universal properties of spectral dimension. Phys. Rev. Lett. [**76**]{} 1091-1093, (1996) Burioni, R., Cassi, D. and Vezzani, A.: in preparation Cassi, D.: Local vs. average behavior on inhomogeneous structures: recurrence on the average and a further extension of Mermin-Wagner theorem on graphs. Phys. Rev. Lett. [**76**]{} 2941-2944, (1996) Fröhlich, J., Simon, B. and Spencer, T.: Infrared bound, phase transitions and continuous symmetry breaking. Commun. Math. Phys., [**50**]{}, 79-85 (1976) Gantmacher, F.R.: The Theory of Matrices. New York: Chelsea Publishing Company 1959 Hattori, K., Hattori, T. and Watanabe, H.: Gaussian Field Theories on General Networks and the Spectral Dimension. Progr. of Theor. Phys. Suppl. [**92**]{}, 108-143 (1987) Mermin, N.D.: Absence of ordering in certain classical systems. Journ. of Math. Phys. [**8**]{}, 1061-1064 (1967) Mohar, B.: The spectrum of an infinite graph. Linear Algebra Appl. [**48**]{}, 245-256 (1982) Mohar, B. and Woess, W.: A survey on spectra of infinite graphs. Bull. London Math. Soc. [**21**]{}, 209-234 (1989)
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider the exactly solvable spin-1/2 $XX$ chain with the three-spin interactions of the $XZX+YZY$ and $XZY-YZX$ types in an external (transverse) magnetic field. We calculate the entropy and examine the magnetocaloric effect for the quantum spin system. We discuss a relation between the cooling/heating efficiency and the ground-state phase diagram of the quantum spin model. We also compare ability to cool/heat in the vicinity of the quantum critical and triple points. Moreover, we examine the magnetocaloric effect for the spin-1/2 $XX$ chain with three-spin interactions in a random (Lorentzian) transverse magnetic field.'
author:
- Myroslava Topilko
- Taras Krokhmalskii
- Oleg Derzhko
- Vadim Ohanyan
title: 'Magnetocaloric effect in spin-1/2 $XX$ chains with three-spin interactions'
---
Introduction {#sec1}
============
In general, the magnetocaloric effect (MCE) refers to any change of the temperature of the magnetic material under the variation of external magnetic field. The revival of interest toward the various aspects of the physics of MCE which has been observed recently is mainly connected with the potential room-temperature cooling applications (see Refs. for recent reviews). Another important application of the MCE is the possibility to map out the $H$-$T$ phase diagram by detecting the magnetocaloric anomalies at a magnetic phase transition at high (pulsed) fields. For some materials there is no alternative way to do that. Since the first successful experiment of adiabatic demagnetization[@MacDougall], MCE is the standard technique of achieving the extremely low temperatures[@Strehlow].
Another important issue of the MCE is its intimate relation with the quantum critical points (QCPs)[@sachdev]. The MCE can be quantified by the adiabatic cooling rate $$\begin{aligned}
T\Gamma_H
&=&\left( \frac{\partial T}{\partial H}\right)_S
\nonumber\\
&=&-\frac{T}{C_H}\left( \frac{\partial S}{\partial H}\right)_T
=-\frac{T}{C_H}\left( \frac{\partial M}{\partial T}\right)_H,
\label{1.01}\end{aligned}$$ where $C_H$ is the heat capacity at the constant magnetic field, and $M$ is the magnetization. The dependence of the cooling rate on the magnetic field is an important characteristic of a specific magnetic material. The cooling rate $T\Gamma_H$ is related to the so-called generalized Grüneisen ratio, $$\begin{aligned}
\Gamma_r
=-\frac{1}{T}
\frac{\left({\partial S}/{\partial r}\right)_T}{\left({\partial S}/{\partial T}\right)_r}
=\frac{1}{T}\left(\frac{\partial T}{\partial r}\right)_S,
\label{1.02}\end{aligned}$$ the important quantity characterizing the QCP. It is known that the generalized Grüneisen ratio changes its sign when the parameter $r$ governing the zero-temperature quantum phase transitions crosses its critical value $r_c$, i.e., in the QCP[@QPT1; @QPT2]. In the case of MCE $r$ in Eq. (\[1.02\]) is the external magnetic field $H$ and QCP corresponds to the critical value $H_c$ at which the system undergoes the transition between different magnetic structures at zero temperature[@Zh_Hon; @trippe; @PNAS]. As the sign of the cooling rate depends on the way magnetic field affects the entropy at isothermal conditions, the system can undergo adiabatic cooling as well as adiabatic heating under the increasing (or under the decreasing) of the external magnetic field magnitude. Thus, the magnetic materials with complicated structure of zero-temperature (ground-state) phase diagram display non-trivial MCE with a sequence of cooling and heating.
Very recently, exact as well as numerical descriptions of the MCE in various one-dimensional interacting spin systems have been attracted much attention[@Zh_Hon; @trippe; @PNAS; @JLTP; @strecka; @derzhko2006; @derzhko2007; @pereira; @strecka2009; @hon_wes; @ribeiro; @ri_mce; @Jafari]. Some two-dimensional systems have been also investigated, mainly, in the context of effect of frustration on the MCE[@hon2d; @shannon2d]. The main features of MCE which have been revealed during the investigation of various models are: (i) essential enhancement of MCE in the vicinity of QCP, (ii) enhancement of MCE by frustration, (iii) appearance of the sequence of cooling and heating stages during adiabatic (de)magnetization for the systems demonstrating several magnetically ordered ground states, and (iv) potential application of MCE data for the investigation of critical properties of the system at hand.
In this paper we continue the investigation of the MCE in one-dimensional quantum spin systems admitting the exact solution in the form of free spinless fermions via the Jordan-Wigner transformation (see, for instance, Ref. ). Though, the cases of spin-1/2 $XX$ (isotropic) and $XY$ (anisotropic) models have been considered in Ref. , there is a series of spin chains with multiple spin interactions introduced by Suzuki in 70s[@suzuki1; @suzuki2] which can be solved by the standard Jordan-Wigner transformation. We consider the simplest model of the Suzuki series, the spin-1/2 $XX$ chain with three-spin interactions both of $XZX+YZY$ and $XZY-YZX$ type[@DRD; @Gottlieb; @Tit_Jap; @Lou; @Z; @krokhmalskii; @dksv; @G]. As has been shown in previous investigations, inclusion of the three-spin interactions leads to appearance of new phases in the ground state and, thus, to more reach behavior in the vicinity of quantum phase transitions. We study two different types of three-spin interactions, namely, the three-spin interaction of the $XZX+YZY$ type[@DRD; @Tit_Jap] and of the $XZY-YZX$ type[@Gottlieb; @Lou]. Although these types of three-spin interactions are connected to each other by a unitary transformation[@krokhmalskii], in the pure form they represent the systems with different symmetries and have different ground-state phase diagrams. In particular, in the former case the ground-state phase diagram contains a point, where three different ground states merge \[quantum triple point (QTP)\]. Appearance of additional parameters in the system, the three-spin coupling constants in our case, makes possible manipulation of the physical features of MCE, namely, the position of the QCP and the values of the maximal and minimal temperatures during the adiabatic (de)magnetization. The knowledge about manipulation of the MCE physical parameters can be very useful for the future quest for the novel magnetic materials and their applications in various aspects. On the other hand, the appearance of the points where several magnetically ordered ground states merge on the ground-state phase diagram, caused by the inclusion of additional three-spin interactions into the Hamiltonian, can lead to essential enhancement of MCE due to large entropy accumulation in such points. Finally, in real-life materials randomness is always present. It can be modeled assuming that on-site fields or intersite interactions acquire random values. The considered quantum spin chains admit an exact analytical solution for thermodynamics in the case when the transverse magnetic field is a random variable with the Lorentzian probability distribution[@ddr]. As a result, with such a model it is possible to discuss the MCE in the presence of randomness.
The paper is organized as follows. At first, we present a general consideration based on the Jordan-Wigner fermionization (Sec. \[sec2\]). Next, we consider separately the case of the of $XZX+YZY$ interaction and the case of the $XZY-YZX$ interaction (Secs. \[sec3\] and \[sec4\]). After that, we consider a random-field spin-1/2 $XX$ chain with three-spin interactions (Sec. \[sec5\]). We discuss the MCE in all these cases. Finally, we draw some conclusions (Sec. \[sec6\]).
Jordan-Wigner fermionization and thermodynamic quantities {#sec2}
=========================================================
Let us define the model under consideration. We consider $N\to\infty$ spins 1/2 placed on a simple chain. The Hamiltonian of the model looks as follows: $$\begin{aligned}
\label{2.01}
{\mathcal{H}} &=& \sum_{n=1}^N
\left[-h s^z_n + J(s^x_ns^x_{n+1} + s^y_ns^y_{n+1})
\right.
\nonumber\\
&+&
K(s^x_ns^z_{n+1}s^x_{n+2} + s^y_ns^z_{n+1}s^y_{n+2})
\nonumber\\
&+&
\left.
E(s^x_ns^z_{n+1}s^y_{n+2} - s^y_ns^z_{n+1}s^x_{n+2})
\right],\end{aligned}$$ where $h$ is the external (transverse) magnetic field, $J$ is the isotropic $XY$ (i.e., $XX$) exchange interaction constant (in what follows we will set $J=1$ to fix the units), and $K$ and $E$ are the constants of the two types of three-spin exchange interactions. We imply periodic boundary conditions in Eq. (\[2.01\]) for convenience.
The Hamiltonian (\[2.01\]) can be brought to the diagonal Fermi-form after applying at first the Jordan-Wigner transformation to spinless fermions, $$\begin{aligned}
\label{2.02}
s^+_n &=& s^x_n + is^y_n = P_{n-1}c^\dag_n,
\;
s^-_n = s^x_n - is^y_n = P_{n-1}c_n,
\nonumber\\
c^\dag_n &=& P_{n-1}s^+_n,
\;
c_n = P_{n-1}s^-_n,
\nonumber\\
P_m &=& \prod^m_{j=1}(1-2c^\dag_jc_j) = \prod^m_{j=1}(-2s_j^z),\end{aligned}$$ and performing further the Fourier transformation, $$\begin{aligned}
\label{2.03}
c^\dag_n &=& \frac{1}{\sqrt{N}} \sum_k e^{ikn}c_k^\dag,
\;
c_n = \frac{1}{\sqrt{N}} \sum_k e^{-ikn}c_k,
\nonumber\\
c^\dag_k &=& \frac{1}{\sqrt{N}} \sum_{n=1}^N e^{-ikn}c_n^\dag,
\;
c_k = \frac{1}{\sqrt{N}} \sum_{n=1}^N e^{ikn}c_n,\end{aligned}$$ $k= 2\pi m/N$, $m = -N/2,\ldots,N/2-1$ (we assume that $N$ is even without loss of generality). As a result, $$\begin{aligned}
\label{2.04}
{\cal{H}} &=& \sum_k\varepsilon_k\left(c_k^\dag c_k-\frac{1}{2}\right),
\nonumber\\
\varepsilon_k &=& -h+J\cos k-\frac{K}{2}\cos(2k)-\frac{E}{2}\sin(2k).\end{aligned}$$
Using Eq. (\[2.04\]) we can easily calculate the partition function for the spin model (\[2.01\]) $$\begin{aligned}
\label{2.05}
Z(T,h,N)={\rm{Tr}}e^{-{\cal{H}}/T}=\prod_k2{\rm{ch}}\frac{\varepsilon_k}{2T}\end{aligned}$$ (we set $k_{\rm{B}}=1$). Various thermodynamic quantities, such as the Helmholtz free energy, the entropy, and the specific heat (per site) immediately follow from Eq. (\[2.05\]): $$\begin{aligned}
\label{2.06}
f(T,h) &=& -\lim_{N\rightarrow\infty}\frac{T\ln Z(T,h,N)}N
\nonumber\\
&=&\frac1{2\pi}\int_{-\pi}^\pi dk \left(\frac{\varepsilon_k}2+T\ln n_k\right),
\nonumber\\
s(T,h)&=&-\frac{\partial f(T,h)}{\partial T}
\nonumber\\
&=& -\frac1{2\pi}\int_{-\pi}^\pi dk\left(\ln{n_k}+\frac{\varepsilon_k}Te^{\varepsilon_k/T}n_k\right),
\nonumber\\
c(T,h)&=&T\frac{\partial s(T,h)}{\partial T}
\nonumber\\
&=&\frac1{2\pi T^2}\int_{-\pi}^\pi dk\varepsilon_k^2n_k(1-n_k);\end{aligned}$$ here $n_k=1/(e^{\varepsilon_k/T}+1)$ are the occupation numbers of spinless fermions. Furthermore, we get $$\begin{aligned}
\label{2.07}
m(T,h)&=&\lim_{N\to\infty}\frac{1}{N}\sum_{n=1}^N\langle s_n^z\rangle
=-\frac{\partial f(T,h)}{\partial h}
\nonumber\\
&=&\frac{1}{2\pi}\int_{-\pi}^{\pi}dk \left(n_k-\frac{1}{2}\right),
\nonumber\\
\frac{\partial m(T,h)}{\partial T}
&=&\frac{1}{2\pi T^2}\int_{-\pi}^{\pi}dk \varepsilon_k n_k (1-n_k),
\nonumber\\
\Gamma_h
&=&-\frac{1}{c(T,h)}\frac{\partial m(T,h)}{\partial T}
\nonumber\\
&=&-\frac{\int_{-\pi}^{\pi}dk \varepsilon_k n_k(1-n_k)}{\int_{-\pi}^{\pi}dk \varepsilon_k^2 n_k (1-n_k)}.\end{aligned}$$
It may be useful to rewrite the formulas for thermodynamic quantities (\[2.06\]), (\[2.07\]) in terms of the density of states $$\begin{aligned}
\rho(\omega)=\lim_{N\to\infty}\frac{1}{N}\sum_k\delta(\omega-\varepsilon_k)
=\frac{1}{2\pi}\int_{-\pi}^{\pi}dk\delta(\omega-\varepsilon_k).
\label{2.08}\end{aligned}$$ We have $$\begin{aligned}
\label{2.09}
f(T,h)&=-T&\int_{-\infty}^{\infty} d\omega \rho(\omega)
\ln\left( 2{\rm{ch}}\frac{\omega}{2T}\right),
\nonumber\\
s(T,h)&=&\int_{-\infty}^{\infty} d\omega \rho(\omega)
\left[ \ln\left(2{\rm{ch}}\frac{\omega}{2T}\right)
-
\frac{\omega}{2T}{\rm{th}}\frac{\omega}{2T} \right],
\nonumber\\
c(T,h)&=&\frac{1}{4T^2}\int_{-\infty}^{\infty} d\omega \rho(\omega)
\frac{\omega^2}{{\rm{ch}}^2[{\omega}/(2T)]},
\nonumber\\
m(T,h)&=&-\frac{1}{2}\int_{-\infty}^{\infty} d\omega \rho(\omega){\rm{th}}\frac{\omega}{2T},
\nonumber\\
\frac{\partial m(T,h)}{\partial T}
&=&\frac{1}{4T^2}\int_{-\infty}^{\infty} d\omega \rho(\omega)\frac{\omega}{{\rm{ch}}^2[{\omega}/(2T)]},
\nonumber\\
\Gamma_h
&=&
-\frac{\int_{-\infty}^{\infty} d\omega \rho(\omega){\omega}/{\{{\rm{ch}}^2[{\omega}/(2T)]}\}}
{\int_{-\infty}^{\infty} d\omega \rho(\omega){\omega^2}/{\{{\rm{ch}}^2[{\omega}/(2T)]}\}}.\end{aligned}$$ Formulas (\[2.09\]) are extremely useful for consideration of the random spin-1/2 $XX$ chains within the Green’s functions approach[@gf; @ddr], since that method permits to calculate the random-averaged density of states (\[2.08\]), see Ref. and Sec. \[sec5\].
Although the presented above formulas give a comprehensive description of the quantum spin system (\[2.01\]) (and, in particular, the MCE), the thermodynamic behavior is somehow hidden behind one-fold integrals in Eqs. (\[2.06\]), (\[2.07\]). More explicit dependencies of thermodynamic quantities on temperature and field can be derived, e.g., in the low-temperature limit. Let us briefly discuss what happens with Eqs. (\[2.06\]), (\[2.07\]) when $T\to 0$, see also Sec. \[sec3\]. We note that $n_k(1-n_k)=1/\{4{\rm{ch}}^2[\varepsilon_k/(2T)]\}$ and therefore as $T\to 0$ only a small region where $\varepsilon_k\approx 0$ is relevant in the integrals yielding $c(T,h)$ or $\partial m(T,h)/\partial T$ in the low-temperature limit. Clearly, if the energy spectrum of spinless fermions is gapped we immediately get that $c(T,h)$ and $\partial m(T,h)/\partial T$ vanishes as $T\to 0$. We turn to the case of a gapless energy spectrum of spinless fermions. Assume that we have $\varepsilon_k=\varepsilon_i^{(z)}(k-k_i)^z/z!+\ldots$ around $k_i$ satisfying $\varepsilon_{k_{i}}=0$. Then we immediately find that $c(T,h)\propto T^{1/z}$. Note also that $s(T,h)\propto T^{1/z}$ and in a “flat-band-like” limit $z\to\infty$ the entropy becomes independent on temperature (for a discussion of true flat-band spin systems see Refs. ). While estimating $\partial m(T,h)/\partial T$ for odd $z$ (e.g., for $z=1$) we have to take higher-order terms in the expansion of $\varepsilon_k$ around $k=k_i$. For even $z$ \[$z=2$ for the QCP and $z=4$ for the QTP, see Eq. (\[3.01\])\] we get $\partial m(T,h)/\partial T \propto T^{1/z-1}$ and therefore $\Gamma_h\propto T^{-1}$.
Alternatively the critical behavior can be derived using formulas (\[2.09\]). The factor $1/{\rm{ch}}^2[\omega/(2T)]$ in the integrands for $c(T,h)$ and $\partial m(T,h)/\partial T$ implies that in the limit $T\to 0$ only a small region where $\omega\approx 0$ is relevant. Around the QCP $\rho(\omega)\propto\omega^{-1/2}$ and therefore $c(T,h)\propto T^{1/2}$, $\partial m(T,h)/\partial T\propto T$, whereas around the QTP $\rho(\omega)\propto\omega^{-3/4}$ (see Ref. ) and as a result $c(T,h)\propto T^{1/4}$, $\partial m(T,h)/\partial T\propto T^{-3/4}$.
In the case of randomness considered in Sec. \[sec5\], van Hove peculiarities in the density of states are smeared out, $\overline{\rho(\omega)}$ has a finite nonzero value for any $\omega$ and, in particular, $\overline{\rho(\omega)}=\overline{\rho(0)}+\ldots$ around $\omega=0$. As a result, $\overline{c(T,h)}\propto T$ for a sufficiently low temperature $T\to 0$ \[an estimate of $\partial \overline{m(T,h)}/\partial T$ requires higher-order terms in the expansion of $\overline{\rho(\omega)}$ around $\omega=0$\]. The boundaries between different ground-state phases disappear since the quantum phase transition transforms into a crossover[@ddr].
Three-spin interactions of $XZX+YZY$ type {#sec3}
=========================================
![ The ground-state phase diagram in the $K-h$ plane of the model (\[2.01\]) with $J=1$ and $E=0$. The dark-gray regions correspond to the spin-liquid II phase, the light-gray region corresponds to the spin-liquid I phase, and the white regions correspond to the ferromagnetic phase. The lines $h^{\star}(K)$ which separate different regions correspond to quantum phase transitions between different ground-state phases.[]{data-label="fig1"}](fig1.eps){width="8cm"}
Now we consider the chain with two-spin interactions and three-spin interactions of $XZX + YZY$ type[@Tit_Jap; @DRD]. In this case we put in Eq. (\[2.01\]) $E=0$ and the energy spectrum of spinless fermions in Eq. (\[2.04\]) reads: $$\begin{aligned}
\label{3.01}
\varepsilon_k=-h+J\cos k-\frac{K}{2}\cos(2k).\end{aligned}$$ The third term in (\[3.01\]) may lead to a new ground-state phase, the so-called spin-liquid II phase[@Tit_Jap]. The phase diagram in the ground state is shown in Fig. \[fig1\]. While $|K|$ is less than 1/2 ($J=1$) there are only two ground-state phases: the spin-liquid I phase and the ferromagnetic phase. However, when $|K|>1/2$ there is in addition one more ground-state phase, the spin-liquid II phase. It is worthy noting that there are two special points ($K=1/2$, $h=3/4$ and $K=-1/2$, $h=-3/4$) on the ground-state phase diagram at which all three ground-state phases meet (QTPs). For further details see Refs. .
The entropy of the spin system is a function of the temperature $T$, the magnetic field $h$, and of the parameter $K$ \[see Eq. (\[2.06\]) in which we have to use $\varepsilon_k$ given in Eq. (\[3.01\]) ($J=1$)\].
Now we turn to a discussion of the MCE in its classical interpretation as an adiabatic change of the temperature of the considered model under field variation. Grayscale plots of the temperature at constant entropy $s(T,h)=0.05$ in the $K-h$ plane are shown in the upper panel of Fig. \[fig2\]. The lowest values of the temperature are around the QTP $K=1/2$, $h=3/4$. Two lower panels of Fig. \[fig2\] supplement the upper one. In two lower panels of Fig. \[fig2\] we show by thin broken lines the dependencies $T(h)$ at fixed values of $s = 0.05, 0.10,\ldots,0.60$ for spin model (\[2.01\]) with $J=1$, $K=0.5,\;1.5$, $E=0$.
![ Upper panel: Grayscale plots of the temperature as it follows from the condition $s(T,h)=0.05$ in the $K-h$ plane for model (\[2.01\]) with $J=1$ and $E=0$. Two lower panels: Isentropic dependence $T$ vs $h$ at $s=0.05,0.10,\ldots,0.60$ (from bottom to top in each panel) for model (\[2.01\]) with $J=1$, $K=0.5,\;1.5$, $E=0$. Thin broken lines correspond to the nonrandom model, thick solid lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig2"}](fig2a.eps "fig:"){width="8cm"}\
![ Upper panel: Grayscale plots of the temperature as it follows from the condition $s(T,h)=0.05$ in the $K-h$ plane for model (\[2.01\]) with $J=1$ and $E=0$. Two lower panels: Isentropic dependence $T$ vs $h$ at $s=0.05,0.10,\ldots,0.60$ (from bottom to top in each panel) for model (\[2.01\]) with $J=1$, $K=0.5,\;1.5$, $E=0$. Thin broken lines correspond to the nonrandom model, thick solid lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig2"}](fig2b.eps "fig:"){width="7cm"}\
![ Upper panel: Grayscale plots of the temperature as it follows from the condition $s(T,h)=0.05$ in the $K-h$ plane for model (\[2.01\]) with $J=1$ and $E=0$. Two lower panels: Isentropic dependence $T$ vs $h$ at $s=0.05,0.10,\ldots,0.60$ (from bottom to top in each panel) for model (\[2.01\]) with $J=1$, $K=0.5,\;1.5$, $E=0$. Thin broken lines correspond to the nonrandom model, thick solid lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig2"}](fig2c.eps "fig:"){width="7cm"}
According to Eq. (\[1.01\]), to discuss the MCE we may analyze alternatively an isothermal change of the entropy under field variation. In the upper panel of Fig. \[fig3\] we show grayscale plots of the entropy in the $K-h$ plane for constant temperature $T=0.05$. Again two lower panels of Fig. \[fig3\] supplement the upper one. In two lower panels of Fig. \[fig3\] we show by thin broken lines the dependencies $s(h)$ at $T=0.05,\;0.10,\ldots,0.60$ for spin model (\[2.01\]) with $J=1$, $K=0.5,\;1.5$, $E=0$.
![ Upper panel: Grayscale plots of the entropy $s(T,h)$ in the $K-h$ plane at $T=0.05$ for model (\[2.01\]) with $J=1$ and $E=0$. Two lower panels: Isothermal dependence $s$ vs $h$ at $T=0.05,0.10,\ldots,0.60$ (from bottom to top in each panel) for model (\[2.01\]) with $J=1$, $K=0.5,\;1.5$, $E=0$. Thin broken lines correspond to the nonrandom model, thick solid lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig3"}](fig3a.eps "fig:"){width="8cm"}\
![ Upper panel: Grayscale plots of the entropy $s(T,h)$ in the $K-h$ plane at $T=0.05$ for model (\[2.01\]) with $J=1$ and $E=0$. Two lower panels: Isothermal dependence $s$ vs $h$ at $T=0.05,0.10,\ldots,0.60$ (from bottom to top in each panel) for model (\[2.01\]) with $J=1$, $K=0.5,\;1.5$, $E=0$. Thin broken lines correspond to the nonrandom model, thick solid lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig3"}](fig3b.eps "fig:"){width="7cm"}\
![ Upper panel: Grayscale plots of the entropy $s(T,h)$ in the $K-h$ plane at $T=0.05$ for model (\[2.01\]) with $J=1$ and $E=0$. Two lower panels: Isothermal dependence $s$ vs $h$ at $T=0.05,0.10,\ldots,0.60$ (from bottom to top in each panel) for model (\[2.01\]) with $J=1$, $K=0.5,\;1.5$, $E=0$. Thin broken lines correspond to the nonrandom model, thick solid lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig3"}](fig3c.eps "fig:"){width="7cm"}
Comparing Fig. \[fig2\] and Fig. \[fig3\] with the ground-state phase diagram in Fig. \[fig1\] one can note that the MCE in the low-$s$ or low-$T$ regimes perfectly reproduces the ground-state phase transition lines.
Furthermore, consider, e.g., $K=1/2$ (thin broken lines in the middle panel in Fig. \[fig2\]). Clearly if we decrease adiabatically $h$ from 2 to 3/4 the temperature noticeably falls down (e.g., approximately from 0.3388 to 0.0002 at $s=0.05$ or from 0.4331 to 0.0025 at $s=0.10$). We turn to the results shown by thin broken lines in the middle panel in Fig. \[fig3\]. If we decrease isothermally $h$ from 2 to 3/4, the entropy of spin system noticeably increases (e.g., approximately from 0.00000 to 0.2195 at $T=0.05$ or from 0.00001 to 0.2670 at $T=0.10$) meaning that the spin system absorbs from the thermostat the heat $T\left[s(h=3/4)-s(h=2)\right]$ (that is, $\approx 0.0110$ at $T=0.05$ or $\approx 0.0267$ at $T=0.10$) per site.
![ $\Gamma_h$ vs $h$ for model (\[2.01\]) with $J=1$, $K=0$ (dotted), $K=0.5$ (dashed), $K=1.5$ (solid), and $E=0$ at $T=0.05$. Thin lines correspond to the nonrandom model, thick lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig4"}](fig4.eps){width="7cm"}
In Fig. \[fig4\] we plot by thin lines the dependence $\Gamma_h(h)$ at $T=0.05$ for a few values of $K=0$ (dotted), $K=0.5$ (dashed), and $K=1.5$ (solid). From the first glance it becomes clear that one peak (around $h\approx 0.75$) is higher than others. It might be interesting to compare the height of maxima in the dependence $\Gamma_h(h)$ for $T=0.05$, in particular the height of the high-field ones which correspond to cooling while $h$ decreases starting from high fields (see thin lines in Fig. \[fig4\] for $h=0.75\ldots 2$). For $K=0,\;0.5,\;1.5$ we have $\Gamma_h(h)\approx 11.00,\;14.22,\;10.80$ at $h\approx 1.02,\;0.77,\;0.93$, respectively. (We recall here that the quantum phase transition occurs at $h^\star=1,\;3/4,\;11/12$ for $K=0,\;0.5,\;1.5$, respectively, see Fig. \[fig1\] and Ref. .) Clearly, the value of the cooling rate around the QTP is about 130% of such a value around the QCP (e.g., for the studied earlier $K=0$ case[@Zh_Hon]). At lower temperatures the heights of maxima increase (e.g., $\Gamma_h(h)\approx 55.56,\;73.04,\;55.41$ at $h\approx 1.003,\;0.754,\;0.920$ for $T=0.01$ and $\Gamma_h(h)\approx 556.51,\;739.48,\;556.61$ at $h\approx 1.0003,\;0.7504,\;0.9170$ for $T=0.001$) but the relation between the heights changes only very slightly and approaches roughly 3:4:3. Thus, one gets about 33% larger change in temperature for the same adiabatic change of the field being performed around QTP in comparison with being performed around QCP.
This can be illustrated further while considering Eq. (\[2.07\]) in the limit $T\to 0$. We recall that according to Eq. (\[3.01\]) $\varepsilon_k=-h+J-K/2-(J/2-K)k^2
+(J-8K)k^4/4!+\ldots$ and therefore while approaching a high-field peculiar point $h^\star$ from above, i.e., by decreasing $h$, $h\to h^\star+0$ (a ferromagnetic–to–spin-liquid transition), we have $$\begin{aligned}
\varepsilon_k=-(h-h^\star)-\frac{1}{2}k^2+\frac{1}{4!}k^4+\ldots
\label{3.02}\end{aligned}$$ for $K=0$ (QCP) with $h^\star=1$ and $$\begin{aligned}
\varepsilon_k=-(h-h^\star)-\frac{1}{8}k^4+\frac{1}{48}k^6 -\frac{1}{640}k^8+\ldots
\label{3.03}\end{aligned}$$ for $K=1/2$ (QTP) with $h^\star=3/4$. Moreover, we may single out two regimes. In the first regime, we first take the limit $T\to 0$ and then $h-h^\star\to +0$ \[i.e., $(h-h^\star)/T\gg 1$\]. In the second regime, we first put $h-h^\star=0$ and then take the limit $T\to 0$ \[i.e., $(h-h^\star)/T\ll 1$\]. \[For the high-field peaks in Fig. \[fig4\] we have $(h-h^\star)/T\approx 0.3\ldots 0.4$.\]
If $T\to 0$ and $(h-h^\star)/T\gg 1$ we can write $n_k(1-n_k)\approx e^{-\vert\varepsilon_k\vert/T}$ and hence $$\begin{aligned}
\label{3.04}
n_k(1-n_k)\propto e^{-k^2/(2T)}\end{aligned}$$ for $K=0$ and $$\begin{aligned}
\label{3.05}
n_k(1-n_k)\propto e^{-k^4/(8T)}\end{aligned}$$ for $K=1/2$, see Eqs. (\[3.02\]) and (\[3.03\]). Furthermore, we can extend the limits of integration with respect to $k$ in two relevant integrals in the formula for $\Gamma_h$ (\[2.07\]) to $-\infty$ and $\infty$. Using Eqs. (\[3.02\]), (\[3.03\]), (\[3.04\]), and (\[3.05\]) we immediately find that $\Gamma_h\to 1/(h-h^{\star})$ as $h\to h^\star+0$ in the limit $T=0$ for both cases, $K=0$ and $K=1/2$. For small but finite $T$ we obtain different results for $K=0$ and $K=1/2$. Although all relevant integrals are doable with the help of the well-known formula for the gamma function $\Gamma(z)$ (see, e.g., Ref. ) $$\begin{aligned}
\label{3.06}
\int_0^{\infty}dxx^{\beta-1}e^{-\lambda x^{\alpha}}
=\frac{1}{\alpha}\lambda^{-\beta/\alpha}\Gamma\left(\frac{\beta}{\alpha}\right),\end{aligned}$$ $\Re\lambda>0$, $\alpha,\;\beta>0$, it is simpler to find required results by using MAPLE codes. Namely, for the case $K=0$ we have $$\begin{aligned}
\label{3.07}
\Gamma_h\approx
\frac{1}{h-h^\star}
\frac{1+\epsilon/2-\epsilon T/8}
{1+\epsilon -(h-4)\epsilon^2/4}\end{aligned}$$ with $\epsilon=T/(h-h^\star)$, $h^\star=1$, whereas for the case $K=1/2$ we have $$\begin{aligned}
\label{3.08}
\Gamma_h\approx
\frac{1}{h-h^\star}
\frac{1+\epsilon /4 - 0.119 \epsilon \sqrt{T} +\epsilon T/32}
{1+\epsilon/2 - 0.239 \epsilon \sqrt{T} + (h+17/4)\epsilon^2/16}\end{aligned}$$ with $\epsilon=T/(h-h^\star)$, $h^\star=3/4$. Approximate analytical formulas (\[3.07\]) and (\[3.08\]) yield $1:1.39$ for ratio of the heights of peaks in the dependence $\Gamma_h(h)$ around QCP and QTP at $T=0.05$ that is in a reasonable agreement with exact numerical calculation according to Eq. (\[2.07\]).
If we put at first $h=h^\star$ and then assume $T\to 0$ we can write $$\begin{aligned}
\label{3.09}
n_k(1-n_k)=\frac{1}{2+e^{\varepsilon_k/T}+e^{-\varepsilon_k/T}}
\nonumber\\
\approx
\frac{1}{4+(ak^z)^2/T^2}
\approx
\frac{1}{4}e^{-a^2k^{2z}/(4T^2)}\end{aligned}$$ with $\varepsilon_k=-ak^z<0$, $z=2$ for the QCP and $z=4$ for the QTP, see Eqs. (\[3.02\]) and (\[3.03\]). Again the limits of integration with respect to $k$ in the two relevant integrals in Eq. (\[2.07\]) can be extended to $-\infty$ and $\infty$. After simple calculations using (\[3.06\]) we find that $\Gamma_h$ at QTP relates to $\Gamma_h$ at QCP as $\Gamma(5/8)/\Gamma(9/8)$ to $\Gamma(3/4)/\Gamma(5/4)$, i.e., as 1.1267…:1. Interestingly, for arbitrary $z$ we have $\Gamma_h\to \{\Gamma[(z+1)/(2z)]/\Gamma[(2z+1)/(2z)]\}/(2T)\to\sqrt{\pi}/(2T)$ if $z\to\infty$. Thus in such a case ($z\to\infty$) $\Gamma_h$ is about 131% of $\Gamma_h$ at the QCP with $z=2$.
In summary, as can be seen from numerical calculations of $\Gamma_h$ (\[2.07\]) reported in Fig. \[fig4\] as well as from analytical considerations in specific limits[@footnote], the efficiency of cooling while decreasing $h$ starting from the high-field limit is higher around the QTP than around the QCP. We note here that enhancement of the MCE in the frustrated $J_1-J_2$ antiferromagnetic Heisenberg chain due to a cancellation of the leading $k^2$-term in the one-particle energy spectrum at the point $J_2=J_1/4$ was discussed in Ref. . For the case at hand (\[2.01\]), (\[3.01\]), the softening in the excitation spectrum occurs at the QTP which emerges owing to the three-spin interactions of $XZX+YZY$ type.
Three-spin interactions of $XZY-YZX$ type {#sec4}
=========================================
Now we turn to the spin-1/2 $XX$ chain with three-spin interactions of $XZY-YZX$ type[@Lou; @Gottlieb]. In this case we put $K=0$ in Eq. (\[2.01\]) and the energy spectrum of spinless fermions in Eq. (\[2.04\]) reads: $$\begin{aligned}
\label{4.01}
\varepsilon_k =-h+J\cos k-\frac{E}{2}\sin(2k).\end{aligned}$$
The ground-state phase diagram of the model is shown in Fig. \[fig5\]. This type of three-spin interactions leads to the spin-liquid II phase too. But for the model (\[4.01\]), in contrast to the model (\[3.01\]), we do not have a QTP and all phase transition lines separate two different phases only, compare Fig. \[fig5\] and Fig. \[fig1\]. For further details see Refs. .
![ The ground-state phase diagram in the $E-h$ plane of the model (\[2.01\]) with $J=1$ and $K=0$. The dark-gray regions correspond to the spin-liquid II phase, the light-gray region corresponds to the spin-liquid I phase, and the white regions correspond to the ferromagnetic phase. The lines $h^{\star}(E)$ which separate different regions correspond to quantum phase transitions between different ground-state phases.[]{data-label="fig5"}](fig5.eps){width="8cm"}
Figs. \[fig6\], \[fig7\], and \[fig8\] are similar to Figs. \[fig2\], \[fig3\], and \[fig4\]. Again the MCE in the low-$s$ and low-$T$ regimes (Figs. \[fig6\], \[fig7\], \[fig8\]) indicates the ground-state phase transition lines seen in Fig. \[fig5\]. Again we observe that cooling/heating is especially efficient around QCPs. Since the model with three-spin interactions of $XZY-YZX$ type does not have a QTP, in all cases $\Gamma_h$ behaves like it should around QCP, see thin lines in Fig. \[fig8\]. For example, the height of the high-field peaks in the dependence $\Gamma_h(h)$ at $T=0.05$ for various $E$ is the same (see thin lines for $h=1\ldots 2$ in Fig. \[fig8\]).
![ Upper panel: Grayscale plots of the temperature as it follows from the condition $s(T,h)=0.05$ in the $E-h$ plane for model (\[2.01\]) with $J=1$ and $K=0$. Two lower panels: Isentropic dependence $T$ vs $h$ at $s=0.05,0.10,\ldots,0.60$ (from bottom to top in each panel) for model (\[2.01\]) with $J=1$, $K=0$, $E=1,\;2$. Thin broken lines correspond to the nonrandom model, thick solid lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig6"}](fig6a.eps "fig:"){width="8cm"}\
![ Upper panel: Grayscale plots of the temperature as it follows from the condition $s(T,h)=0.05$ in the $E-h$ plane for model (\[2.01\]) with $J=1$ and $K=0$. Two lower panels: Isentropic dependence $T$ vs $h$ at $s=0.05,0.10,\ldots,0.60$ (from bottom to top in each panel) for model (\[2.01\]) with $J=1$, $K=0$, $E=1,\;2$. Thin broken lines correspond to the nonrandom model, thick solid lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig6"}](fig6b.eps "fig:"){width="7cm"}\
![ Upper panel: Grayscale plots of the temperature as it follows from the condition $s(T,h)=0.05$ in the $E-h$ plane for model (\[2.01\]) with $J=1$ and $K=0$. Two lower panels: Isentropic dependence $T$ vs $h$ at $s=0.05,0.10,\ldots,0.60$ (from bottom to top in each panel) for model (\[2.01\]) with $J=1$, $K=0$, $E=1,\;2$. Thin broken lines correspond to the nonrandom model, thick solid lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig6"}](fig6c.eps "fig:"){width="7cm"}
![ Upper panel: Grayscale plots of the entropy $s(T,h)$ in the $E-h$ plane at $T=0.05$ for model (\[2.01\]) with $J=1$ and $K=0$. Two lower panels: Isothermal dependencies $s$ vs $h$ at $T=0.05,0.10,\ldots ,0.60$ (from bottom to top in each panel) for model (\[2.01\]) with $J=1$, $K=0$, $E=1,\;2$. Thin broken lines correspond to the nonrandom model, thick solid lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig7"}](fig7a.eps "fig:"){width="8cm"}\
![ Upper panel: Grayscale plots of the entropy $s(T,h)$ in the $E-h$ plane at $T=0.05$ for model (\[2.01\]) with $J=1$ and $K=0$. Two lower panels: Isothermal dependencies $s$ vs $h$ at $T=0.05,0.10,\ldots ,0.60$ (from bottom to top in each panel) for model (\[2.01\]) with $J=1$, $K=0$, $E=1,\;2$. Thin broken lines correspond to the nonrandom model, thick solid lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig7"}](fig7b.eps "fig:"){width="7cm"}\
![ Upper panel: Grayscale plots of the entropy $s(T,h)$ in the $E-h$ plane at $T=0.05$ for model (\[2.01\]) with $J=1$ and $K=0$. Two lower panels: Isothermal dependencies $s$ vs $h$ at $T=0.05,0.10,\ldots ,0.60$ (from bottom to top in each panel) for model (\[2.01\]) with $J=1$, $K=0$, $E=1,\;2$. Thin broken lines correspond to the nonrandom model, thick solid lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig7"}](fig7c.eps "fig:"){width="7cm"}
![ $\Gamma_h$ vs $h$ for model (\[2.01\]) with $J=1$, $K=0$, $E=0$ (dotted), $E=1$ (dashed), and $E=2$ (solid) at $T=0.05$. Thin lines correspond to the nonrandom model, thick lines correspond to the random-field model (\[5.01\]) with $\Gamma=0.1$.[]{data-label="fig8"}](fig8.eps){width="7cm"}
Random (Lorentzian) transverse magnetic field {#sec5}
=============================================
In this section we use recent results on thermodynamics of the spin-1/2 $XX$ chain with three-spin interactions in a random transverse field[@ddr] to discuss the influence of randomness on the MCE. To be specific, we consider the Hamiltonian (\[2.01\]) and make the change $h\to h_n$, where $h_n$ is the random transverse magnetic field with the Lorentzian probability distribution $$\begin{aligned}
\label{5.01}
p(h_n)=\frac{1}{\pi}\frac{\Gamma}{(h_n-h)^2+\Gamma^2}.\end{aligned}$$ Now $h$ is the mean value of $h_n$ and the parameter $\Gamma$ controls the strength of the Lorentzian disorder. The nonrandom case can be reproduced if $\Gamma$ is sent to 0. All (random-averaged) thermodynamic quantities of the random quantum spin system (\[2.01\]), (\[5.01\]) can be expressed through the (random-averaged) density of states[@ddr]. For instance, in the formula for the entropy in Eq. (\[2.09\]) we now have to use the density of states $$\begin{aligned}
\label{5.02}
\overline{\rho(\omega)} &=& \mp\frac{1}{\pi}\Im \overline{G_{jj}^{\mp}(\omega)},
\nonumber\\
\overline{G_{jj}^{\mp}(\omega)}
&=& \frac{4}{K}
\left[\frac{z_1}{(z_1-z_2)(z_1-z_3)(z_1-z_4)}
\right.
\nonumber\\
&+&
\left.
\frac{z_2}{(z_2-z_1)(z_2-z_3)(z_2-z_4)} \right]\end{aligned}$$ and $\vert z_1\vert\le \vert z_2\vert\le \vert z_3\vert\le \vert z_4\vert$ are the solutions of a certain quartic equation, see Ref. . In the case $E=0$ the relevant quartic equation, $$\begin{aligned}
\label{5.03}
z^4-\frac{2J}{K}z^3+\frac{4}{K}(\omega+h\pm i\Gamma)z^2-\frac{2J}{K}z+1=0,\end{aligned}$$ can be reduced to the quadratic one that further simplifies calculations. In the case $K=0$ the relevant quartic equation, $$\begin{aligned}
\label{5.04}
z^4-\frac{2iJ}{E}z^3+\frac{4i}{E}(\omega+h\pm i\Gamma)z^2-\frac{2iJ}{E}z-1=0,\end{aligned}$$ can be easily solved numerically.
Our findings for the random-field models (\[2.01\]), (\[5.01\]) with $\Gamma=0.1$ are shown by thick lines in the two lower panels of Figs. \[fig2\], \[fig3\], \[fig6\], \[fig7\] and in Figs. \[fig4\], \[fig8\]. From the analysis of nonrandom models in Secs. \[sec3\] and \[sec4\] we know that essential enhancement of MCE occurs around QCPs and QTPs. From the reported results for the random-field chains (compare, e.g., thin and thick lines in Figs. \[fig4\] and \[fig8\]) we conclude that just around these special points the MCE is extremely sensitive to randomness. As can be seen in Figs. \[fig4\] and \[fig8\], even small randomness leads to a rounding of sparks in the dependence $\Gamma_h(h)$ and noticeable diminishing of maximal values of $\Gamma_{h}$.
Conclusions {#sec6}
===========
In this work we have studied the MCE for the spin-1/2 $XX$ chain with the three-spin interactions of the $XZX+YZY$ and $XZY-YZX$ types. The considered models have more parameters (in addition to the magnetic field we can also vary the strength of three-spin interactions), they contain several lines of QCPs and QTPs, and manipulation with MCE becomes possible. We have found that the quantum phase transition lines clearly manifest themselves for the MCE in the low-$s$ or low-$T$ regimes. The ground-state phase diagrams can be perfectly reproduced by measuring of $\Gamma_h$ in the limit $T\to 0$. The vicinity of QCPs or QTPs is very effective for cooling since low temperatures are achieved by only small decrease of the field. Particularly strong variation of temperature (at $s={\rm{const}}$) or of entropy (at $T={\rm{const}}$) with varying the magnetic field occurs in the vicinity of a QTP.
We have discussed the MCE in a random quantum spin chain. We have found that even small randomness can noticeably diminish an enhanced MCE in proximity to a QCP/QTP.
The considered models, thanks to their simplicity, have enabled the rigorous analysis of the thermodynamic quantities of interest. Although we do not know any particular compound which can be described by the studied models, our results might have a general merit being useful for understanding the effects of proximity to QTP and randomness on the MCE. On the other hand, with further progress in material sciences and synthesis of new magnetic chain compounds, the lack of experimental data and comparison between theory and experiment may be resolved in future.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors thank A. Honecker, A. Klümper J. Richter, J. Sirker, T. Vekua, and M. E. Zhitomirsky for useful comments. O. D. and V. O. acknowledge financial support of the organizers of the SFB602 workshop on Localized excitations in flat-band models (Göttingen, April 12-15, 2012) and of the 504. WE-Heraeus-Seminar on “Quantum Magnetism in Low Spatial Dimensions” (Bad Honnef, 16-18 April 2012) where the paper was finalized. V. O. expresses his gratitude to the Department of Theoretical Physics of the Georg-August University (Göttingen) and ICTP (Trieste) for warm hospitality during the work on this paper. He also acknowledges financial support from DFG (Grant No. HO 2325/8-1), ANSEF (Grant No. 2497-PS), Volkswagen Foundation (Grant No. I/84 496), SCS-BFBR 11RB-001 grant, and joint grant of CRDF-NFSAT and the State Committee of Science of Republic of Armenia (Grant No. ECSP-09-94-SASP).
[99]{}
K. A. Gschneidner, Jr., V. K. Pecharsky, and A. O. Tsokol, Rep. Prog. Phys. [**68**]{}, 1479 (2005) and references therein.
A. M. Tishin and Y. I. Spichkin, [*The Magnetocaloric Effect and Its Applications*]{} (Institute of Physics Publishing, Bristol, Philadelphia, 2003).
W. F. Giauque and D. P. MacDougall, Phys. Rev. [**43**]{}, 768 (1933).
P. Strehlow, H. Nuzha, and E. Bork, J. Low Temp. Phys. [**147**]{}, 81 (2007).
S. Sachdev, [*Quantum Phase Transitions*]{} (2nd ed., Cambridge University Press, Cambridge, UK, 2011).
L. Zhu, M. Garst, A. Rosch, and Q. Si, Phys. Rev. Lett. [**91**]{}, 066404 (2003).
M. Garst and A. Rosch, Phys. Rev. B [**72**]{}, 205129 (2005).
M. E. Zhitomirsky and A. Honecker, Journal of Statistical Mechanics: Theory and Experiment P07012 (2004).
C. Trippe, A. Honecker, A. Klümper, and V. Ohanyan, Phys. Rev. B [**81**]{}, 054402 (2010).
B. Wolf, Y. Tsui, D. Jaiswal-Nagar, U. Tutsch, A. Honecker, K. Removic-Langer, G. Hofmann, A. Prokofiev, W. Assmus, G. Donath, and M. Lang, PNAS [**108**]{}, 6862 (2011).
M. Lang, Y. Tsui, B. Wolf, D. Jaiswal-Nagar, U. Tutsch, A. Honecker, K. Removic-Langer, A. Prokofiev, W. Assmus, and G. Donath, J. Low Temp. Phys. [**159**]{}, 88 (2010).
L. Čanová, J. Strečka, and M. Jaščur, J. Phys.: Condens. Matter [**18**]{}, 4967 (2006).
O. Derzhko and J. Richter, Eur. Phys. J. B [**52**]{}, 23 (2006).
O. Derzhko, J. Richter, A. Honecker, and H.-J. Schmidt, Fizika Nizkikh Temperatur (Kharkiv) [**33**]{}, 982 (2007); Low Temperature Physics [**33**]{}, 745 (2007).
M. S. S. Pereira, F. A. B. F. de Moura, and M. L. Lyra, Phys. Rev. B [**79**]{}, 054427 (2009).
L. Čanová, J. Strečka, and T. Lučivjansky, Condensed Matter Physics (L’viv) [**12**]{}, 353 (2009).
A. Honecker and S. Wessel, Condensed Matter Physics (L’viv) [**12**]{}, 399 (2009).
G. A. P. Ribeiro, Journal of Statistical Mechanics: Theory and Experiment P12016 (2010).
J. Schnack, R. Schmidt, and J. Richter, Phys. Rev. B [**76**]{}, 054413 (2007).
R. Jafari, arXiv:1105.0809.
A. Honecker and S. Wessel, Physica B [**378-380**]{}, 1098 (2006).
B. Schmidt, P. Thalmeier, and N. Shannon, Phys. Rev. B [**76**]{}, 125113 (2007).
M. Takahashi, [*Thermodynamics of One-Dimensional Solvable Models*]{} (Cambridge University Press, Cambridge, 1999); J. B. Parkinson and D. J. J. Farnell, [*An Introduction to Quantum Spin Systems*]{}, Lecture Notes in Physics Vol. 816 (Springer-Verlag, Berlin, Heidelberg, 2010).
M. Suzuki, Phys. Lett. A [**34**]{}, 94 (1971).
M. Suzuki, Prog. Theor. Phys. [**46**]{}, 1337 (1971).
D. Gottlieb and J. Rössler, Phys. Rev. B [**60**]{}, 9232 (1999).
O. Derzhko, J. Richter, and V. Derzhko, Annalen der Physik (Leipzig) [**8**]{}, SI-49 (1999) \[arXiv:cond-mat/9908425\].
I. Titvinidze and G. I. Japaridze, Eur. Phys. J. B [**32**]{}, 383 (2003).
P. Lou, W.-C. Wu, and M.-C. Chang, Phys. Rev. B [**70**]{}, 064405 (2004).
A. A. Zvyagin, Phys. Rev. B [**72**]{}, 064419 (2005).
T. Krokhmalskii, O. Derzhko, J. Stolze, and T. Verkholyak, Phys. Rev. B [**77**]{}, 174404 (2008).
O. Derzhko, T. Krokhmalskii, J. Stolze, and T. Verkholyak, Phys. Rev. B [**79**]{}, 094410 (2009).
F. G. Ribeiro, J. P. de Lima, and L. L. Gonçalves, J. Magn. Magn. Mater. [**323**]{}, 39 (2011).
V. Derzhko, O. Derzhko, and J. Richter, Phys. Rev. B [**83**]{}, 174428 (2011).
D. N. Zubarev, [*Njeravnovjesnaja Statistitchjeskaja Tjermodinamika*]{} (Nauka, Moskva, 1971) (in Russian); D. N. Zubarev, [*Nonequilibrium Statistical Thermodynamics*]{} (Consultants Bureau, New York, 1974); G. Rickayzen, [*Green’s Functions and Condensed Matter*]{} (Academic Press, London, 1991); G. D. Mahan, [*Many-Particle Physics*]{} (Plenum Press, New York, 1993).
J. Schulenburg, A. Honecker, J. Schnack, J. Richter, and H.-J. Schmidt, Phys. Rev. Lett. [**88**]{}, 167207 (2002); J. Richter, J. Schulenburg, A. Honecker, J. Schnack, and H.-J. Schmidt, J. Phys.: Condens. Matter [**16**]{}, S779 (2004); O. Derzhko and J. Richter, Phys. Rev. B [**70**]{}, 104415 (2004).
M. V. Fedoryuk, [*Mjetod Pjerjevala*]{} (Nauka, Moskva, 1977) (in Russian).
We may also estimate $\Gamma_h(h)$ in the limit $T\to 0$ within a spin-liquid phase. For this we have to take into account that the equation $\varepsilon_k=0$ has two ($K<1/2$), three ($K=1/2$), or four ($K>1/2$) solutions $\{k_i\}$ and $\varepsilon_k=\varepsilon_i^{(1)}(k-k_i)+\varepsilon_i^{(2)}(k-k_i)^2/2!+\ldots$ around $k_i$. Further we have to approximate $n_k(1-n_k)$ in the spirit of Eq. (\[3.09\]) and after extending the limits of integration to evaluate relevant integrals in Eq. (\[2.07\]). The final result reads: $$\begin{aligned}
\Gamma_h
\to
-\frac{\sum_i\varepsilon_i^{(2)}/\left\vert\varepsilon_i^{(1)}\right\vert^3}{\sum_i 1/\left\vert\varepsilon_i^{(1)}\right\vert}.
\nonumber\end{aligned}$$ Alternatively, we may begin with $\Gamma_h$ given in Eq. (\[2.09\]). As was mentioned in Sec. \[sec2\], $1/{\rm{ch}}^2[\omega/(2T)]$ tends to $4T\delta(\omega)$ in the limit $T\to 0$ and hence only a small region around $\omega=0$ is relevant. Expanding $\rho(\omega)$ around $\omega=0$, $\rho(\omega)=\rho(0)+\rho^{(1)}(0)\omega +\ldots$, we arrive at $$\begin{aligned}
\Gamma_h
\to
- \frac{d\rho(\omega)/d\omega\vert_{\omega=0}}{\rho(\omega)\vert_{\omega=0}}.
\nonumber\end{aligned}$$ These two results for $\Gamma_h$ are equivalent.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we develop a thorough analysis of the boundedness properties of the maximal operator for the Bochner-Riesz means related to the Fourier-Bessel expansions. For this operator, we study weighted and unweighted inequalities in the spaces $L^p((0,1),x^{2\nu+1}\, dx)$. Moreover, weak and restricted weak type inequalities are obtained for the critical values of $p$. As a consequence, we deduce the almost everywhere pointwise convergence of these means.'
address: |
Departamento de Matemáticas y Computación\
Universidad de La Rioja\
26004 Logroño, Spain
author:
- Óscar Ciaurri and Luz Roncal
date: 'July 24, 2012'
title: 'The Bochner-Riesz means for Fourier-Bessel expansions: norm inequalities for the maximal operator and almost everywhere convergence'
---
[^1]
Introduction and main results
=============================
Let ${J_{\nu}}$ be the Bessel function of order $\nu$. For $\nu>-1$ we have that $$\int_0^1 {J_{\nu}}(s_jx){J_{\nu}}(s_kx)x\,dx= \frac12
(J_{\nu+1}(s_j))^2\delta_{j,k},\quad j,k=1,2,\dots$$ where $\{s_j\}_{j\ge 1}$ denotes the sequence of successive positive zeros of ${J_{\nu}}$. From the previous identity we can check that the system of functions $$\label{eq:FBesselSystemI}
\psi_j(x)=\frac{\sqrt{2}}{|J_{\nu+1}(s_j)|}x^{-\nu}{J_{\nu}}(s_jx),\quad
j=1,2,\dots$$ is orthonormal and complete in $L^2((0,1),d\mu_\nu)$, with $d\mu_\nu(x)=x^{2\nu+1}\, dx$ (for the completeness, see [@Hochstadt]). Given a function $f$ on $(0,1)$, its Fourier series associated with this system, named as Fourier-Bessel series, is defined by $$\label{SeriesCoeficientes}
f\sim \sum_{j=1}^\infty a_j(f)\psi_j,\qquad\text{with}\qquad
a_j(f)=\int_0^1 f(y)\psi_j(y)\,d\mu_{\nu}(y),$$ provided the integral exists. When $\nu=n/2-1$, for $n\in
\mathbb{N}$ and $n\ge 2$, the functions $\psi_j$ are the eigenfunctions of the radial Laplacian in the multidimensional ball $B^n$. The eigenvalues are the elements of the sequence $\{s_j^2\}_{j\ge 1}$. The Fourier-Bessel series corresponds with the radial case of the multidimensional Fourier-Bessel expansions analyzed in [@Bal-Cor].
For each $\delta>0$, we define the Bochner-Riesz means for Fourier-Bessel series as $$\mathcal{B}_R^{\delta}(f,x)=\sum_{j\ge 1}
\left(1-\frac{s_j^2}{R^2}\right)_+^{\delta}a_j(f)\psi_j(x),$$ where $R>0$ and $(1-s^2)_+=\max\{1-s^2,0\}$. Bochner-Riesz means are a regular summation method used oftenly in harmonic analysis. It is very common to analyze regular summation methods for Fourier series when the convergence of the partial sum fails. Cesàro means are other of the most usual summation methods. B. Muckenhoupt and D. W. Webb [@Mu-We] give inequalities for Cesàro means of Laguerre polynomial series and for the supremum of these means with certain parameters and $1<p\leq
\infty$. For $p=1$, they prove a weak type result. They also obtain similar estimates for Cesàro means of Hermite polynomial series and for the supremum of those means in [@Mu-We-Her]. An almost everywhere convergence result is obtained as a corollary in [@Mu-We] and [@Mu-We-Her]. The result about Laguerre polynomials is an extension of a previous result in [@Stem]. This kind of matters has been also studied by the first author and J. L. Varona in [@Ciau-Var] for the Cesàro means of generalized Hermite expansions. The Cesàro means for Jacobi polynomials were analyzed by S. Chanillo and B. Muckenhoupt in [@Ch-Muc]. The Bochner-Riesz means themselves have been analyzed for the Fourier transform and their boundedness properties in $L^p(\mathbb{R}^n)$ is an important unsolved problem for $n>2$ (the case $n=2$ is well understood, see [@Car-Sjo]).
The target of this paper is twofold. First we will analyze the almost everywhere (a. e.) convergence, for functions in $L^p((0,1),d\mu_\nu)$, of the Bochner-Riesz means for Fourier-Bessel expansions. By the general theory [@Duoa Ch. 2], to obtain this result we need to estimate the maximal operator $$\mathcal{B}^{\delta}(f,x)=\sup_{R>0}\left|\mathcal{B}_R^{\delta}(f,x)\right|,$$ in the $L^p((0,1),d\mu_\nu)$ spaces. A deep analysis of the boundedness properties of this operator will be the second goal of our paper. This part of our work is strongly inspired by the results given in [@Ch-Muc] for the Fourier-Jacobi expansions.
Before giving our results we introduce some notation. Being $p_0=\frac{4(\nu+1)}{2\nu+3+2\delta}$ and $p_1=\frac{4(\nu+1)}{2\nu+1-2\delta}$, we define $$\begin{aligned}
\label{eq:endpoints0} p_0(\delta)&=\begin{cases}
1,& \delta> \nu+1/2\text{ or } -1<\nu\le -1/2,\\
p_0,& \delta\le \nu+1/2\text{ and } \nu>-1/2,
\end{cases}\\
\notag p_1(\delta)&=\begin{cases}
\infty,& \delta> \nu+1/2\text{ or } -1<\nu\le -1/2,\\
p_1,& \delta\le \nu+1/2\text{ and } \nu>-1/2.
\end{cases}\end{aligned}$$
Concerning to the a. e. convergence of the Bochner-Riesz means, our result reads as follows
\[th:fin\] Let $\nu>-1$, $\delta>0$, and $1\le p<\infty$. Then, $$\mathcal{B}_R^\delta(f,x)\to f(x)\quad \text{a.
e., for $f\in L^p((0,1),d\mu_\nu)$}$$ if and only if $p_0(\delta)\le p$, where $p_0(\delta)$ is as in .
Proof of Theorem \[th:fin\] is contained in Section \[sec:Proof Th1\] and is based on the following arguments. On one hand, to prove the necessity part, we will show the existence of functions in $L^p((0,1),d\mu_{\nu})$ for $p<p_0(\delta)$ such that $\mathcal{B}^\delta_R$ diverges for them. In order to do this, we will use a reasoning similar to the one given by C. Meaney in [@Meaney] that we describe in Section \[sec:Proof Th1\]. On the other hand, for the sufficiency, observe that the convergence result follows from the study of the maximal operator $\mathcal{B}^\delta f$. Indeed, it is sufficient to get $(p_0(\delta),p_0(\delta))$-weak type estimates for this operator and this will be the content of Theorem \[th:AcDebilMaxRedonda\].
Regarding the boundedness properties of $\mathcal{B}^\delta f$ we have the following facts. First, a result containing the $(p,p)$-strong type inequality.
\[th:max\] Let $\nu>-1$, $\delta>0$, and $1< p\le\infty$. Then, $$\left\|\mathcal{B}^{\delta}f\right\|_{L^p((0,1),d\mu_{\nu})}\le C
\|f\|_{L^p((0,1),d\mu_{\nu})}$$ if and only if $$\begin{cases}
1<p\le \infty, &\text{for $-1<\nu\le -1/2$ or $\delta>\nu+1/2$},\\
p_0<p<p_1, &\text{for $\delta\le \nu+1/2$ and $\nu>-1/2$.}
\end{cases}$$
In the lower critical value of $p_0(\delta)$ we can prove a $(p_0(\delta),p_0(\delta))$-weak type estimate.
\[th:AcDebilMaxRedonda\] Let $\nu>-1$, $\delta>0$, and $p_0(\delta)$ be the number in . Then, $$\left\|\mathcal{B}^{\delta}f\right\|_{L^{p_0(\delta),\infty}((0,1),d\mu_{\nu})}\le
C \|f\|_{L^{p_0(\delta)}((0,1),d\mu_{\nu})},$$ with $C$ independent of $f$.
Finally, for the upper critical value, when $0<\delta<\nu+1/2$ and $\nu>-1/2$, it is possible to obtain a $(p_1,p_1)$-restricted weak type estimate.
\[th:AcDebilRestMaxRedonda\] Let $\nu>-1/2$ and $0<\delta<\nu+1/2$. Then, $$\left\|\mathcal{B}^{\delta}\chi_E\right\|_{L^{p_1,\infty}((0,1),d\mu_{\nu})}\le
C \|\chi_E\|_{L^{p_1}((0,1),d\mu_{\nu})},$$ for all measurable subsets $E$ of $(0,1)$ and $C$ independent of $E$.
The previous results about norm inequalities are summarized in Figure 1 (case $-1<\nu \le -1/2$) and Figure 2 (case $\nu>-1/2$).
(0.1,0.1) – (2.1,0.1) – (2.1,2.1) – (0.1,2.1) – cycle; (2.1,0.1) – (2.1,2.1); (0.1,2.1) – (0.1,0.1); (0.1,0.1) – (2.15,0.1); at (0.1,0) [$0$]{}; at (2.1,0) [$1$]{}; at (2.2,0.1) [$\frac{1}{p}$]{}; at (0.1,2.175) [$\delta$]{}; (0,1.4) node \[rotate=90\] ; (2.2,1.4) node \[rotate=90\] ; (1.1,-0.2) node [Figure 1: case $-1<\nu\le-\tfrac{1}{2}$.]{};
(2.7,0.75) – (3.35,0.1) – (4.05,0.1) – (4.7,0.75) – (4.7,2.1) – (2.7,2.1) – cycle; (4.7,0.75) – (4.7,2.1); (2.7,2.1) – (2.7,0.78); (2.7,0.75) circle (0.8pt); (2.72,0.73) – (3.35,0.1); (4.05,0.1) – (4.7,0.75); (2.7,0.1) – (4.75,0.1); (2.7,0.1) – (2.7,0.72); (4.7,0.1) – (4.7,0.13); at (2.7,0) [$0$]{}; at (4.7,0) [$1$]{}; at (4.8,0.1) [$\frac{1}{p}$]{}; at (2.7,2.175) [$\delta$]{}; (2.43,0.75) node ; (3.35,0) node ; (4.05,0) node ; (2.6,1.4) node \[rotate=90\] ; (3.1,0.45) node \[rotate=-45\] ; (4.3,0.45) node \[rotate=45\] ; (4.8,1.4) node \[rotate=90\] ; (3.7,-0.2) node [Figure 2: case $\nu>-\tfrac{1}{2}$.]{};
At this point, a comment is in order. Note that J. E. Gilbert [@Gi] also proves weak type norm inequalities for maximal operators associated with orthogonal expansions. The method used cannot be applied in our case, and the reason is the same as can be read in [@Ch-Muc], at the end of Sections 15 and 16 therein. Following the technique in [@Gi] we have to analyze some weak type inequalities for Hardy operator and its adjoint with weights and these inequalities do not hold for $p=p_0$ and $p=p_1$.
The proof of the sufficiency in Theorem \[th:max\] will be deduced from a more general result in which we analyze the boundedness of the operator $\mathcal{B}^\delta f$ with potential weights. Before stating it, we need a previous definition. We say that the parameters $(b,B,\nu,\delta)$ satisfy the $C_p$ conditions if $$\begin{aligned}
b& > \frac{-2(\nu+1)}{p} \,\,\, (\ge \text{ if }p=\infty), \label{ec:con1B}\\
B& < 2(\nu+1)\left(1-\frac1p\right) \,\,\, (\le \text{ if }
p=1), \label{ec:con2B}\\
b& > 2(\nu+1)\left(\frac12-\frac1p\right)-\delta-\frac12\,\,\,
(\ge \text{ if }p=\infty), \label{ec:con3B}\\
B& \le 2(\nu+1)\left(\frac12-\frac1p\right)+\delta+\frac12, \label{ec:con4B}\\
B &\le b \label{ec:con5B},\end{aligned}$$ and in at least one of each of the following pairs the inequality is strict: and , and , and and except for $p=\infty$. The result concerning inequalities with potential weights is the following.
\[th:AcFuerteMaxRedonda\] Let $\nu>-1$, $\delta>0$, and $1<
p\le\infty$. If $(b,B,\nu,\delta)$ satisfy the $C_p$ conditions, then $$\left\|x^b\mathcal{B}^{\delta}f\right\|_{L^p((0,1),d\mu_{\nu})}\le
C \|x^Bf\|_{L^p((0,1),d\mu_{\nu})},$$ with $C$ independent of $f$.
A result similar to Theorem \[th:AcFuerteMaxRedonda\] for the partial sum operator was proved in [@GuPeRuVa Theorem 1]. It followed from a weighted version of a general Gilbert’s maximal transference theorem, see [@Gi Theorem 1]. The weighted extension of Gilbert’s result given in [@GuPeRuVa] depended heavily on the $A_p$ theory and it can not be used in our case because it did not capture all the information relative to the weights. On the other hand, it is also remarkable the paper by K. Stempak [@Stem2] in which maximal inequalities for the partial sum operator of Fourier-Bessel expansions and divergence and convergence results are discussed.
The necessity in Theorem \[th:max\] will follow by showing that the operator $\mathcal{B}^\delta f$ is neither $(p_1,p_1)$-weak nor $(p_0,p_0)$-strong for $\nu>-1/2$ and $0<\delta\le \nu+1/2$. This is the content of the next theorems.
\[th:noweak\] Let $\nu>-1/2$. Then $$\sup_{\|f\|_{L^{p_1}((0,1),d\mu_\nu)}=1}
\|\mathcal{B}^{\delta}_{R}f\|_{L^{p_1,\infty}((0,1),d\mu_\nu)}\ge
C (\log R)^{1/p_0},$$ if $0<\delta<\nu+1/2$; and $$\sup_{\|f\|_{L^\infty((0,1),d\mu_\nu)}=1}
\|\mathcal{B}^{\delta}_{R}f\|_{L^{\infty}((0,1),d\mu_\nu)}\ge C
\log R,$$ if $\delta= \nu+1/2$.
\[th:nostrong\] Let $\nu>-1/2$. Then $$\sup_{E\subset
(0,1)}\frac{\|\mathcal{B}^{\delta}_{R}\chi_E\|_{L^{p_0}((0,1),d\mu_\nu)}}{\|\chi_E\|_{L^{p_0}((0,1),d\mu_\nu)}}\ge
C (\log R)^{1/p_0},$$ if $0<\delta<\nu+1/2$; and $$\sup_{\|f\|_{L^1((0,1),d\mu_\nu)}=1}
\|\mathcal{B}^{\delta}_{R}f\|_{L^{1}((0,1),d\mu_\nu)}\ge C \log R,$$ if $\delta= \nu+1/2$.
The paper is organized as follows. In the next section, we give the proof of Theorem \[th:fin\]. In Section \[sec:proofAcFuerte\] we first relate the Bochner-Riesz means $\mathcal{B}_R^{\delta}$ to the Bochner-Riesz means operator associated with the Fourier-Bessel system in the Lebesgue measure setting. Then, we prove weighted inequalities for the supremum of this new operator. With the connection between these means and the operator $\mathcal{B}_R^{\delta}$, we obtain Theorem \[th:AcFuerteMaxRedonda\] and, as a consequence, the sufficiency of Theorem \[th:max\]. Sections \[sec:ProofThAcDebilMaxRedonda\] and \[sec:acdelrest\] will be devoted to the proofs of Theorems \[th:AcDebilMaxRedonda\] and \[th:AcDebilRestMaxRedonda\], respectively. The proofs of Theorems \[th:noweak\] and \[th:nostrong\] are contained in Section \[sec:negativeths\]. One of the main ingredients in the proofs of Theorems \[th:noweak\] and \[th:nostrong\] will be Lemma \[lem:pol\], this lemma is rather technical and it will be proved in the Section \[sec:techlemma\].
Throughout the paper, we will use the following notation: for each $p\in[1,\infty]$, we will denote by $p'$ the conjugate of $p$, that is, $\tfrac{1}{p}+\tfrac{1}{p'}=1$. We shall write $X\simeq Y$ when simultaneously $X\le C Y$ and $Y \le C X$.
Proof of Theorem \[th:fin\] {#sec:Proof Th1}
===========================
The proof of the sufficiency follows from Theorem \[th:AcDebilMaxRedonda\] and standard arguments.
In order to prove the necessity, let us see that, for $0<\delta<\nu+1/2$ and $\nu>-1/2$, there exists a function $f\in
L^{p}((0,1),d\mu_{\nu})$, $p\in [1,p_0)$, for which $\mathcal{B}_R^{\delta}(f,x)$ diverges. We follow some ideas contained in [@Meaney] and [@Stem2].
First, we need a few more ingredients. Recall the well-known asymptotics for the Bessel functions (see [@Wat Chapter 7]) $$\label{zero}
J_\nu(z) = \frac{z^\nu}{2^\nu \Gamma(\nu+1)} + O(z^{\nu+2}),
\quad |z|<1,\quad |\arg(z)|\leq\pi,$$ and $$\label{infty} J_\nu(z)=\sqrt{\frac{2}{\pi z}}\left[
\cos\left(z-\frac{\nu\pi}2
- \frac\pi4 \right) + O(e^{{\mathop{\rm Im}}(z)}z^{-1}) \right], \quad |z|
\ge 1,\quad |\arg(z)|\leq\pi-\theta,$$ where $D_{\nu}=-(\nu\pi/2+\pi/4)$. It will also be useful the fact that (cf. [@OScon (2.6)]) $$\label{eq:zerosCons} s_{j}=O(j).$$ For our purposes, we need estimates for the $L^p$ norms of the functions $\psi_j$. These estimates are contained in the following lemma, whose proof can be read in [@Ci-RoWave Lemma 2.1].
\[Lem:NormaFunc\] Let $1\le p\le\infty$ and $\nu>-1$. Then, for $\nu>-1/2$, $$\|\psi_j\|_{L^p((0,1),d\mu_{\nu})}\simeq
\begin{cases} j^{(\nu+1/2)-\frac{2(\nu+1)}{p}}, &
\text{if $p>\frac{2(\nu+1)}{\nu+1/2}$},\\
(\log j)^{1/p}, & \text{if $p=\frac{2(\nu+1)}{\nu+1/2}$},\\
1, & \text{if $p<\frac{2(\nu+1)}{\nu+1/2}$},
\end{cases}$$ and, for $-1<\nu\le-1/2$, $$\|\psi_j\|_{L^p((0,1),d\mu_{\nu})}\simeq
\begin{cases}
1, & \text{if $p<\infty$},\\
j^{\nu+1/2},& \text{if $p=\infty$}.
\end{cases}$$
We will also use a slight modification of a result by G. H. Hardy and M. Riesz for the Riesz means of order $\delta$, that is contained in [@HaRi Theorem 21]. We present here this result, adapted to the Bochner-Riesz means. We denote by $S_R(f,x)$ the partial sum associated to the Fourier-Bessel expansion, namely $$S_R(f,x)=\sum_{0<s_j\le R} a_j(f)\psi_j(x).$$ The result reads as follows.
\[lem:HardyRiesz\] Suppose that $f$ can be expressed as a Fourier-Bessel expansion and for some $\delta>0$ and $x\in(0,1)$ its Bochner-Riesz means $\mathcal{B}_R^\delta(f,x)$ converges to $c$ as $R\rightarrow \infty$. Then, for $s_n\le R < s_{n+1}$, $$|S_R(f,x)-c|\le A_{\delta}n^{\delta}\sup_{0<t\le
s_{n+1}}|\mathcal{B}_{t}^{\delta}(f,x)|.$$
By using this lemma, we can write $$\label{ConsecuenciaLemaHardyRiesz}
|a_j(f)\psi_j(x)|=|(S_{s_{j}}(f,x)-c)-(S_{s_{j-1}}(f,x)-c)|\le
A_{\delta}j^\delta\sup_{0<t\le
s_{j+1}}|\mathcal{B}_{t}^{\delta}(f,x)|.$$ Let us proceed with the proof of the necessity. Let $1\le p<p_0$. Note that $p_0'=p_1$. Therefore, $p'>p_0'>\tfrac{2(\nu+1)}{\nu+1/2}$, and $\delta<\nu+1/2-\frac{2(\nu+1)}{p'}:=\lambda$. By Lemma \[Lem:NormaFunc\], $\|\psi_j\|_{L^{p'}((0,1),d\mu_{\nu})}\ge C
j^{\lambda}$. Then, we have that the mapping $f\mapsto a_j(f)$, where $a_j(f)$ was given in , is a bounded linear functional on $L^{p}((0,1),d\mu_{\nu})$ with norm bounded below by a constant multiple of $j^\lambda$. By uniform boundedness principle, for $p$ conjugate to $p'$ and each $0\le
\varepsilon<\lambda$, there is a function $f_0\in L^p((0,1),
d\mu_{\nu})$ so that $a_j(f_0)j^{-\varepsilon}\rightarrow \infty$ as $j\rightarrow \infty$. By taking $\varepsilon=\delta$, we have that $$\label{coeficienteInfinito}
a_j(f_0)j^{-\delta}\rightarrow \infty \quad \textrm{ as } \quad
j\rightarrow \infty.$$
Suppose now that $B_R^\delta(f_0,x)$ converges. Then, by Egoroff’s theorem, it converges on a subset $E$ of positive measure in $(0,1)$ and, clearly, we can think that $E\subset (\eta, 1)$ for some fixed $\eta>0$. For each $x\in E$, we can consider $j$ such that $s_j x\ge 1$ and, by , $$\begin{aligned}
|a_j(f_0)\psi_j(x)|&=\big|a_j(f_0)
\Big(\frac{\sqrt2}{|J_{\nu+1}(s_j)|}
x^{-\nu}J_{\nu}(s_jx)\\
&-\frac{\sqrt2}{|J_{\nu+1}(s_j)|}x^{-\nu}
\Big(\frac{2}{\pi s_jx}\Big)^{1/2}
\cos(s_jx+D_{\nu})\Big)\\
&+a_j(f_0) \frac{\sqrt2}{|J_{\nu+1}(s_j)|}x^{-\nu}
\Big(\frac{2}{\pi s_jx}\Big)^{1/2}
\cos(s_jx+D_{\nu})\big|\\
&=Cs_j^{-1/2}\frac{\sqrt 2}{|J_{\nu+1}(s_j)|}|a_j(f_0)x^{-\nu-1/2}
\big(O((s_jx)^{-1})+\cos(s_jx+D_{\nu})\big)|\\
&\simeq |a_j(f_0)x^{-\nu-1/2}(\cos(s_jx+D_{\nu})+O((s_jx)^{-1}))|.\end{aligned}$$ By on this set $E$, $$|a_j(f_0)x^{-\nu-1/2}(\cos(s_jx+D_{\nu})+O((j)^{-1}))|\le
A_{\delta}j^{\delta}\sup_{0<t\le
s_{j+1}}|\mathcal{B}_{t}^{\delta}(f_0,x)|\le K_Ej^{\delta},$$ uniformly on $x\in E$. We also used in the latter. The inequality above is equivalent to $$|a_j(f_0)(\cos(s_jx+D_{\nu})+O(j^{-1}))|\le K_E
x^{\nu+1/2}j^{\delta}\le K_{E}j^{\delta}.$$ Therefore, $$\label{ec:boundFj}
|a_j(f_0)j^{-\delta}(\cos(s_jx+D_{\nu})+O((j)^{-1}))|\le K_E.$$ Now, taking the functions $$F_j(x)=a_j(f_0)j^{-\delta}(\cos(s_jx+D_{\nu})+O(j^{-1})), \qquad
x\in E,$$ and using an argument based on the Cantor-Lebesgue and Riemann-Lebesgue theorems, see [@Meaney Section 1.5] and [@Zyg Section IX.1], we obtain that $$\int_E |F_j(x)|^2\, dx\ge C |a_j(f_0)j^{-\delta}|^2|E|,$$ where, as usual, $|E|$ denotes the Lebesgue measure of the set $E$. On the other hand, by , $$\int_E |F_j(x)|^2\, dx\le K_E^2 |E|.$$ Then, from the previous estimates, it follows that $|a_j(f_0)j^{-\delta}|\le C$, which contradicts .
Bochner-Riesz means for Fourier-Bessel expansions in the Lebesgue measure setting. Proof of Theorem \[th:AcFuerteMaxRedonda\] {#sec:proofAcFuerte}
=============================================================================================================================
For our convenience, we are going to introduce a new orthonormal system. We will take the functions $$\phi_j(x)=\frac{\sqrt{2x}{J_{\nu}}(s_jx)}{|J_{\nu+1}(s_j)|},\quad
j=1,2,\dots.$$ These functions are a slight modification of the functions ; in fact, $$\label{eq:Relation}
\phi_j(x)=x^{\nu+1/2}\psi_j(x).$$ The system $\{\phi_j(x)\}_{j\ge1}$ is a complete orthonormal basis of $L^2((0,1),dx)$.
In this case, the corresponding Fourier-Bessel expansion of a function $f$ is $$f\sim\sum_{j=1}^{\infty}b_j(f) \phi_j(x), \qquad \text{with}
\qquad b_j(f)=\left(\int_0^1 f(y)\phi_j(y)\, dy\right)$$ provided the integral exists, and for $\delta>0$ the Bochner-Riesz means of this expansion are $$B_R^{\delta}(f,x)=\sum_{j\ge 1}
\left(1-\frac{s_j^2}{R^2}\right)_+^{\delta}b_j(f)\phi_j(x),$$ where $R>0$ and $(1-s^2)_+=\max\{1-s^2,0\}$. It follows that $$B_R^{\delta}(f,x)=\int_0^1 f(y)K_R^\delta(x,y)\, dy$$ where $$\label{ec:kern} K_R^\delta(x,y)=\sum_{j\ge
1}\left(1-\frac{s_j^2}{R^2}\right)_+^{\delta}\phi_j(x)\phi_j(y).$$
Our next target is the proof of Theorem \[th:AcFuerteMaxRedonda\]. Taking into account that $$\mathcal{B}_R^\delta f(x)=\int_0^1 f(y)\mathcal{K}_R^\delta(x,y)
\, d\mu_\nu(y),$$ where $$\mathcal{K}_R^\delta(x,y)=\sum_{j\ge
1}\left(1-\frac{s_j^2}{R^2}\right)_+^{\delta}\psi_j(x)\psi_j(y),$$ it is clear, from , that $\mathcal{K}_R^\delta(x,y)=(xy)^{-(\nu+1/2)}K_R^{\delta}(x,y)$. Then, it is verified that the inequality $$\|x^b\mathcal{B}^{\delta}(f,x)\|_{L^p((0,1),d\mu_{\nu})}\le
C\|x^Bf(x)\|_{L^p((0,1),d\mu_{\nu})}$$ is equivalent to $$\|x^{b+(\nu+1/2)(2/p-1)}B^{\delta}(f,x)\|_{L^p((0,1),dx)}\le
C\|x^{B+(\nu+1/2)(2/p-1)}f(x)\|_{L^p((0,1),dx)},$$ that is, we can focus on the study of a weighted inequality for the operator $B_R^{\delta}(f,x)$. The first results about convergence of this operator can be found in [@Ci-Ro].
We are going to prove an inequality of the form $$\|x^{a}B^{\delta}(f,x)\|_{L^p((0,1),dx)}\le C
\|x^{A}f(x)\|_{L^p((0,1),dx)}$$ for $\delta>0$, $1< p\leq \infty$, under certain conditions for $a, A,\nu$ and $\delta$. Besides, a weighted weak type result for $\sup_{R>0}|B_R^{\delta}(f,x)|$ will be proved for $p=1$. The abovementioned conditions are the following. Let $\nu>-1$, $\delta>0$ and $1\leq p\leq \infty$; parameters $(a,A,\nu,\delta)$ will be said to satisfy the $c_p$ conditions provided $$\begin{aligned}
a & > -1/p-(\nu+1/2) \,\,\, (\ge \text{ if } p=\infty), \label{ec:con1}\\
A & < 1-1/p+(\nu+1/2)\,\,\, (\le \text{ if } p=1),\label{ec:con2}\\
a &> -\delta-1/p\,\,\, (\ge \text{ if }p=\infty),\label{ec:con3}\\
A &\le 1+\delta-1/p, \label{ec:con4}\\
A &\le a\label{ec:con5}\end{aligned}$$ and in at least one of each of the following pairs the inequality is strict: and , and , and and except for $p=\infty$.
The main results in this section are the following:
\[th:main1\] Let $\nu>-1$, $\delta>0$ and $1< p\le \infty$. If $(a, A, \nu, \delta)$ satisfy the $c_p$ conditions, then $$\|x^{a}B^{\delta}(f,x)\|_{L^p((0,1),dx)}\le C
\|x^{A}f(x)\|_{L^p((0,1),dx)},$$ with $C$ independent of $f$.
\[th:main2\] Let $\nu>-1$ and $\delta>0$. If $(a, A, \nu,
\delta)$ satisfy the $c_1$ conditions and $$E_{\lambda}=\left\{x\in (0,1)\colon x^{a}
\sup_{R>0}\left(|B_R^{\delta}(f,x)|\right)>\lambda \right\},$$ then $$|E_{\lambda}|\leq C \frac{\|x^{A}
f(x)\|_{L^1((0,1),dx)}}{\lambda},$$ with $C$ independent of $f$ and $\lambda$.
Note that, taking $a=b+(\nu+1/2)(2/p-1)$ and $A=B+(\nu+1/2)(2/p-1)$, Theorem \[th:AcFuerteMaxRedonda\] follows from Theorem \[th:main1\].
The proofs of Theorem \[th:main1\] and Theorem \[th:main2\] will be achieved by decomposing the square $(0,1)\times (0,1)$ into five regions and obtaining the estimates therein. The regions will be: $$\begin{aligned}
\label{regions}
\notag A_1&=\{(x,y):0 < x, y\leq 4/R\},\\
\notag A_2&=\{(x,y):4/R<\max\{x,y\}<1,\, |x-y|\le 2/R \},\\
A_3&=\{(x,y): 4/R \leq x < 1,\, 0 < y\leq x/2\},\\
\notag A_4&=\{(x,y):0 < x \leq y/2,\, 4/R \leq y< 1\}, \\
\notag A_5&=\{(x,y): 4/R < x < 1, \, x/2 < y< x- 2/R\}\\
\notag &\kern25pt\cup
\{(x,y): y/2 < x \leq y-2/R,\, 4/R \leq y<1\}.\end{aligned}$$
Theorem \[th:main1\] and Theorem \[th:main2\] will follow by showing that, if $1\leq p\leq \infty$, then $$\label{eq:des_1} \left\|\sup_{R> 0}\int_0^1
y^{-A}x^a|K_R^\delta(x,y)||f(y)|\chi_{A_j}\,dy\right\|_{L^p((0,1),dx)}
\leq C\|f(x)\|_{L^p((0,1),dx)}$$ holds for $j=1,3,4$ and that $$\label{eq:des_2} \int_0^1
y^{-A}x^a|K_R^\delta(x,y)||f(y)|\chi_{A_j}\,dy \leq C M (f,x),$$ for $j=2,5$, where $M$ is the Hardy-Littlewood maximal function of $f$, and $C$ is independent of $R, x$ and $f$. These results and the fact that $M$ is $(1,1)$-weak and $(p,p)$-strong if $1<p\leq
\infty$ complete the proofs.
To get and we will use a very precise pointwise estimate for the kernel $K_R^\delta(x,y)$, obtained in [@Ci-Ro]; there, it was shown that $$\label{ec:kernel}
|K_R^\delta(x,y)|\le C \begin{cases} (xy)^{\nu+1/2}R^{2(\nu+1)}, &
(x,y) \in A_1,\\ R, & (x,y) \in A_2\\
\frac{\Phi_\nu(Rx)\Phi_{\nu}(Ry)}{R^{\delta}|x-y|^{\delta+1}}, &
(x,y) \in A_3\cup A_4 \cup A_5, \end{cases}$$ with $$\label{ec:aux}
\Phi_\nu(t)=\begin{cases}t^{\nu+1/2}, & \text{ if $0<t<2$},\\
1,& \text{ if $t\ge 2$}.\end{cases}$$
The proof of follows from the given estimate for the kernel $K_R^\delta(x,y)$ and $y^{-A}x^a\simeq C$ in $A_2\cup
A_5$ because $A\le a$. In the case of $A_2$, from $|K_R^\delta(x,y)|\le C R$ we deduce easily the required inequality. For $A_5$ the result is a consequence of $\Phi_{\nu}(Rx)\Phi_\nu(Ry)\le C$ and of a decomposition of the region in strips such that $R|x-y|\simeq 2^{k}$, with $k=0,\dots,
[\log_2 R]-1$; this can be seen in [@Ci-Ro p. 109]
In this manner, to complete the proofs of Theorem \[th:main1\] and Theorem \[th:main2\] we only have to show for $j=1,3,4$ in the conditions $c_p$ for $1\le p \le \infty$, and this is the content of Corollary \[cor:corolario2\] in Subsection \[subsec:reg1\]. In its turn, Corollary \[cor:corolario2\] follows from Lemmas \[lem:lema7\] and \[lem:lema8\] in the same subsection. Previously, Subsection \[subsec:lemmas\] contains some technical lemmas that will be used in the proofs of Lemmas \[lem:lema7\] and \[lem:lema8\].
Technical Lemmas {#subsec:lemmas}
----------------
To prove for $j=1,3,4$ we will use an interpolation argument based on six lemmas. These are stated below. They are small modifications of the six lemmas contained in Section 3 of [@Mu-We] where a sketch of their proofs can be found.
\[lem:lema1\] Let $\xi_0>0$, if $r<-1$, $r+t\leq-1$ and $r+s+t\leq-1$, then for $p=1$ $$\left\|x^r\chi_{[1,\infty)}(x) \sup_{\xi_0\leq\xi\leq x}\xi^s
\int_\xi^x y^t|f(y)|\,dy\right\|_{L^p((0,\infty),dx)}\leq
C\|f(x)\|_{L^p((0,\infty),dx)}$$ with $C$ independent of $f$. If $r\leq0$, $r+t\leq-1$ and $r+s+t\leq-1$ with equality holding in at most one of the first two inequalities, then this holds for $p=\infty$.
\[lem:lema2\] Let $\xi_0>0$, if $t\leq0$, $r+t\leq-1$ and $r+s+t\leq-1$, with strict inequality in the last two in case of equality in the first, then for $p=1$ $$\left\|x^r\chi_{[1,\infty)}(x) \sup_{\xi_0\leq\xi\leq x}\xi^s
\int_x^\infty y^t|f(y)|\,dy\right\|_{L^p((0,\infty),dx)}\leq
C\|f(x)\|_{L^p((0,\infty),dx)}$$ with $C$ independent of $f$. If $t<-1$, $r+t\leq-1$ and $r+s+t\leq-1$, then this holds for $p=\infty$.
\[lem:lema3\] If $s<0$, $s+t\leq0$ and $r+s+t\leq-1$,with equality holding in at most one of the last two inequalities, then for $p=1$ $$\left\|x^r\chi_{[1,\infty)}(x) \sup_{\xi\geq x}\xi^s \int_x^\xi
y^t|f(y)|\,dy\right\|_{L^p((0,\infty),dx)}\leq C\|f(x)\|_{L^p((0,\infty),dx)}$$ with $C$ independent of $f$. If $s<0$, $s+t\leq-1$ and $r+s+t\leq-1$ this holds for $p=\infty$.
\[lem:lema4\] If $t\leq0$, $s+t\leq0$ and $r+s+t\leq-1$,with strict inequality holding in the first two in case the third is an equality, then for $p=1$ $$\left\|x^r\chi_{[1,\infty)}(x) \sup_{\xi\geq x}\xi^s
\int_\xi^\infty y^t|f(y)|\,dy\right\|_{L^p((0,\infty),dx)}\leq
C\|f(x)\|_{L^p((0,\infty),dx)}$$ with $C$ independent of $f$. If $t<-1$, $s+t\leq-1$ and $r+s+t\leq-1$ then this holds for $p=\infty$.
\[lem:lema5\] If $s<0$, $r+s<-1$ and $r+s+t\leq-1$, then for $p=1$ $$\left\|x^r\chi_{[1,\infty)}(x) \sup_{\xi\geq x}\xi^s \int_1^x
y^t|f(y)|\,dy\right\|_{L^p((0,\infty),dx)}\leq C\|f(x)\|_{L^p((0,\infty),dx)}$$ with $C$ independent of $f$. If $s<0$, $r+s\leq 0$ and $r+s+t\leq-1$, with equality holding in at most one of the last two inequalities, this holds for $p=\infty$.
\[lem:lema6\] If $r<-1$, $r+s<-1$ and $r+s+t\leq-1$, then for $p=1$ $$\left\|x^r\chi_{[1,\infty)}(x) \sup_{1\leq\xi\leq x}\xi^s
\int_1^{\xi} y^t|f(y)|\,dy\right\|_{L^p((0,\infty),dx)}\leq
C\|f(x)\|_{L^p((0,\infty),dx)}$$ with $C$ independent of $f$. If $r\leq0$, $r+s\leq0$ and $r+s+t\leq-1$, with equality in at most one of the last two inequalities, this holds for $p=\infty$.
Proofs of Theorem \[th:main1\] and Theorem \[th:main2\] for regions $A_1$, $A_3$ and $A_4$ {#subsec:reg1}
------------------------------------------------------------------------------------------
This section contains the proofs of the inequality for regions $A_1$, $A_3$ and $A_4$. The results we will prove are included in the following
\[lem:lema7\] If $\nu>-1$, $\delta>0$, $R>0$, $j=1, 3, 4$ and $(a, A, \nu, \delta)$ satisfy the $c_1$ conditions, then holds for $p=1$ with $C$ independent of $f$.
\[lem:lema8\] If $\nu>-1$, $\delta>0$, $R>0$, $j=1, 3, 4$ and $(a, A, \nu, \delta)$ satisfy the $c_{\infty}$ conditions, then holds for $p=\infty$ with $C$ independent of $f$.
\[cor:corolario2\] If $1\leq p\leq\infty$, $\nu>-1$, $\delta>0$, $R>0$, $(a, A, \nu, \delta)$ satisfy the $c_p$ conditions and $j=1,3,4$, then holds with $C$ independent of $f$.
**Proof of Corollary \[cor:corolario2\]**. It is enough to observe that if $1< p<\infty$ and $(a, A, \nu, \delta)$ satisfy the $c_p$ conditions, then $(a-1+1/p, A-1+1/p, \nu, \delta)$ satisfy the $c_1$ conditions. So, by Lemma \[lem:lema7\] $$\begin{gathered}
\left\|\sup_{R\geq 0}\int_0^1
y^{-A+1-1/p}x^{a-1+1/p}|K_R^{\delta}(x,y)|\chi_{A_j}(x,y)|f(y)|\,
dy \right\|_{L^1((0,1),dx)}\\\leq C\|f(x)\|_{L^1((0,1),dx)},\end{gathered}$$ and this is equivalent to $$\int_0^1 x^{a+1/p}\left(\sup_{R\geq 0}\int_0^1
|K_R^{\delta}(x,y)|\chi_{A_j}(x,y)|f(y)|\, dy\right)
\frac{dx}{x}\leq C\int_0^1 x^{A+1/p}|f(x)|\frac{dx}{x},$$ where $j=1,3,4$. Similarly, if $(a,A,\nu,\delta)$ verify the $c_p$ conditions, then $(a+1/p, A+1/p, \nu, \delta )$ satisfy the $c_{\infty}$ conditions. Hence, by Lemma \[lem:lema8\] $$\begin{gathered}
\left\|x^{a+1/p}\sup_{R\geq 0}\int_0^1
|K_R^{\delta}(x,y)|\chi_{A_j}(x,y)|f(y)|\, dy\right\|_{L^{\infty}((0,1),dx)}
\\\leq C\|x^{A+1/p}f(x)\|_{L^{\infty}((0,1),dx)}.\end{gathered}$$ Now, we can use the Marcinkiewicz interpolation theorem to obtain the inequality $$\begin{gathered}
\int_0^1 \left(x^{a+1/p} \left(\sup_{R\geq 0}\int_0^1
|K_R^{\delta}(x,y)|\chi_{A_j}(x,y)|f(y)|\, dy\right)\right)^p
\frac{dx}{x}\\
\leq
C\int_0^1 \left(x^{A+1/p}|f(x)|\right)^p\frac{dx}{x},\end{gathered}$$ for $1<p<\infty$ and the proof is finished.
Finally, we will prove Lemmas \[lem:lema7\] and \[lem:lema8\] for $A_j$, $j=1, 3$ and $4$, separately.
**Proof of Lemma \[lem:lema7\] and Lemma \[lem:lema8\] for $A_1$**. First of all, we have to note that $B_R^\delta
(f,x)=0$ when $0<R<s_1$, being $s_1$ the first positive zero of $J_\nu$. Using the estimate , the left side of in this case is bounded by $$C\left\|x^{a+\nu+1/2}\chi_{[0,1]}(x)\sup_{s_1<R\leq
4/x}R^{2(\nu+1)} \int_0^{4/R}y^{-A+\nu+1/2}|f(y)|\,
dy\right\|_{L^p((0,1),dx)}.$$ Making the change of variables $x=4/u$ and $y=4/v$, we have $$C\left\|u^{-a-\nu-\frac12-\frac2p}\chi_{[4,\infty)}(u)
\sup_{s_1\leq R \leq u}R^{2(\nu+1)}\int_R^\infty
v^{A-(\nu+\frac12)-2+\frac2p}g(v)\,dv\right\|_{L^p((0,\infty),du)},$$ where $\|\cdot\|_{L^p((0,\infty),du)}$ denotes the $L^p$ norm in the variable $u$, and $$g(v)=v^{-2/p}|f(4v^{-1})|.$$ Note that function $g(v)$ is supported in $(1,\infty)$ and $\|g\|_{L^p((0,\infty),du)}=\|f\|_{L^p((0,1),dx)}$. The function $g$ will be used through the subsection, but the value $4$ may be changed by another one, at some points, without comment. Now, splitting the inner integral at $u$, we obtain the sum of $$\label{ec:pa11}
C\left\|u^{-a-\nu-\frac12-\frac2p}\chi_{[4,\infty)}(u)
\sup_{s_1\leq R \leq u}R^{2(\nu+1)}\int_R^u
v^{A-(\nu+\frac12)-2+\frac2p}g(v)\,dv\right\|_{L^p((0,\infty),du)}$$ and $$\label{ec:pa12}
C\left\|u^{-a-\nu-\frac12-\frac2p}\chi_{[4,\infty)}(u)
\sup_{s_1\leq R \leq u}R^{2(\nu+1)}\int_u^\infty
v^{A-(\nu+\frac12)-2+\frac2p}g(v)\,dv\right\|_{L^p((0,\infty),du)}.$$ From Lemma \[lem:lema1\] we get the required estimate for , using conditions and ; Lemma \[lem:lema2\] is applied to inequality , there we need conditions and and the restriction on them. This completes the proof of Lemmas \[lem:lema7\] and \[lem:lema8\] for $j=1$.
**Proof of Lemma \[lem:lema7\] and Lemma \[lem:lema8\] for $A_3$**. Clearly, the left side of is bounded by $$C\left\|x^a \chi_{[4/R,1]}(x)\sup_{4/x\leq R}\int_0^{x/2}y^{-A}
|K_R^{\delta}(x,y)||f(y)|\, dy\right\|_{L^p((0,1),dx)}.$$ Splitting the inner integral at $2/R$, using the bound for the kernel given in and the definition of $\Phi_\nu$, we have this expression majorized by the sum of $$\label{ec:pa21} \left\|x^a \chi_{[0,1]}(x)\sup_{4/x\leq
R}\int_0^{2/R}|f(y)|\frac{(Ry)^{\nu+1/2}y^{-A}}{R^{\delta}|x-y|^{\delta+1}}\,
dy\right\|_{L^p((0,1),dx)}$$ and $$\label{ec:pa22} \left\|x^a \chi_{[0,1]}(x)\sup_{4/x\leq
R}\int_{2/R}^{x/2}\frac{|f(y)|y^{-A}}{R^{\delta}|x-y|^{\delta+1}}\, dy
\right\|_{L^p((0,1),dx)}.$$ For , taking into account that $|x-y|\simeq x$ in $A_3$, the changes of variables $x=4/u$, $y=2/v$ give us $$\left\|u^{-a+(\delta+1)-\frac 2p}\chi_{[4,\infty)}(u)\sup_{u\leq
R} R^{-\delta+(\nu+1/2)}\int_R^{\infty}v^{-(\nu+1/2)+A+\frac
2p-2}g(v)\, dv\right\|_{L^p((0,\infty),du)}.$$ Lemma \[lem:lema4\] can be used here. The required conditions for $p=1$ are , and with the restriction in the pairs therein. For $p=\infty$ the same inequalities are needed.
On the other hand, in , using again that $|x-y|\simeq
x$, by changing of variables $x=4/u$ and $y=2/v$ we have $$\begin{gathered}
C\left\|u^{-a+(\delta+1)-\frac
2p}\chi_{[4,\infty)}(u)\sup_{u\leq R}R^{-\delta}\int_{2u}^R
v^{A+\frac 2p-2}g(v)\, dv\right\|_{L^p((0,\infty),du)}\\
\leq C\left\|u^{-a+(\delta+1)-\frac
2p}\chi_{[4,\infty)}(u)\sup_{u\leq R}R^{-\delta}\int_u^R
v^{A+\frac 2p-2}g(v)\, dv\right\|_{L^p((0,\infty),du)}.\end{gathered}$$ Lemma \[lem:lema3\] can then be applied. For $p=1$, we need $\delta>0$, which is an hypothesis, and and with its corresponding restriction. For $p=\infty$ the inequalities are the same, with the requirement that is strict. This completes the proof of Lemmas \[lem:lema7\] and \[lem:lema8\] for $j=3$.
**Proof of Lemma \[lem:lema7\] and Lemma \[lem:lema8\] for $A_4$**. In this case, the left hand side of is estimated by $$C\left\|x^a \chi_{[0,1/2]}(x)\sup_{R>4}\int_{\max(4/R,2x)}^1
y^{-A}|K_R^{\delta}(x,y)||f(y)|\, dy\right\|_{L^p((0,1),dx)}.$$ To majorize this, we decompose the $R$-range in two regions: $4<R\leq 2/x$ and $R\geq 2/x$. In this manner, with the bound for the kernel given in and the definition of $\Phi_\nu$, the previous norm is controlled by the sum of $$C\left\|x^a \chi_{[0,1/2]}(x)\sup_{4<R\leq 2/x}
\int_{4/R}^1 |f(y)|\frac{(Rx)^{\nu+1/2}y^{-A}}{R^{\delta}|x-y|^{\delta+1}}\, dy
\right\|_{L^p((0,1),dx)}$$ and $$C\left\|x^a \chi_{[0,1/2]}(x)\sup_{R\geq 2/x}\int_{2x}^1
\frac{|f(y)|y^{-A}}{R^{\delta}|x-y|^{\delta+1}}\, dy\right\|_{L^p((0,1),dx)}.$$ Next, using that $|x-y|\simeq y$ in $A_4$, with the changes of variables $x=2/u$ and $y=1/v$ the previous norms are controlled by $$\label{ec:pa31} C\left\|u^{-a-\frac 2p-(\nu+\frac
12)}\chi_{[4,\infty)}(u)\sup_{4<R\leq u}R^{-\delta+(\nu +\frac
12)} \int_1^{R/4}v^{A+\frac 2p
-2+(\delta+1)}g(v)\,dv\right\|_{L^p((0,\infty),du)}$$ and $$\label{ec:pa32} C\left\|u^{-a-\frac
2p}\chi_{[4,\infty)}(u)\sup_{R\geq u}R^{-\delta}\int_1^{u/4}
v^{A+\frac 2p -2+(\delta+1)}g(v)\,dv\right\|_{L^p((0,\infty),du)}.$$ In , we use Lemma \[lem:lema6\]; for $p=1$, conditions , and are needed; we need the same for $p=\infty$. For , Lemma \[lem:lema5\] requires the hypothesis $\delta>0$ and conditions and for $p=1$ and the same for $p=\infty$ with the restrictions in the pairs therein. This proves Lemmas \[lem:lema7\] and \[lem:lema8\] for $j=4$.
Proof of Theorem \[th:AcDebilMaxRedonda\] {#sec:ProofThAcDebilMaxRedonda}
=========================================
Now we shall prove Theorem \[th:AcDebilMaxRedonda\]. First note that, by , we can write $$\mathcal{B}_R^{\delta}(f,x)
=\int_0^1f(y)\left(\frac{y}{x}\right)^{\nu+1/2}K_R^{\delta}(x,y)\,dy,$$ where $K_R^\delta$ is the kernel in . By taking $g(y)=f(y)y^{\nu+1/2}$, to prove the result it is enough to check that $$\int_{E}\,d\mu_\nu(x)\le \frac{C}{\lambda^{p}}
\int_0^1|g(x)|^{p}x^{(\nu+1/2)(2-p)}\,dx,$$ where $E=\left\{x\in(0,1):
\sup_{R>0}x^{-(\nu+1/2)}\int_0^1|g(y)||K_R^{\delta}(x,y)|\,dy>\lambda\right\}$ and $p=p_0(\delta)$. We decompose $E$ into four regions, such that $E=\bigcup_{i=1}^{4}J_i$, where $$J_i=\left\{x\in(0,1): \sup_{R>0}x^{-(\nu+1/2)}\int_0^{1}|g(y)|
\chi_{B_i}(x,y)|K_R^{\delta}(x,y)|\,dy>\lambda\right\}$$ for $i=1,\dots,4$, with $B_1=A_1$, $B_2=A_2\cup A_5$, $B_3=A_3$, and $B_4=A_4$ where the sets $A_i$ were defined in . Note also that $\int_{E}\,d\mu_\nu(x)\le\sum_{i=1}^4\int_{J_i}\,d\mu_\nu(x)$, then we need to prove that $$\label{ec:boundweak}
\int_{J_i}\,d\mu_\nu(x)\le \frac{C}{\lambda^{p}}
\int_0^1|g(x)|^{p}x^{(\nu+1/2)(2-p)}\,dx,$$ for $i=1,\dots,4$ and $p=p_0(\delta)$. At some points along the proof we will use the notation $$\label{eq:integral}
I_p:=\int_0^{1}|g(y)|^{p}y^{(\nu+1/2)(2-p)}\,dy.$$
In $J_1$, by applying and Hölder inequality with $p=p_0$, we have $$\begin{gathered}
x^{-(\nu+1/2)}\int_0^{1}|g(y)|\chi_{B_1}(x,y)
|K_R^{\delta}(x,y)|\,dy\\
\begin{aligned}
&\le Cx^{-(\nu+1/2)}\int_0^{4/R}
|g(y)|(xy)^{\nu+1/2}R^{2(\nu+1)}\,dy\\
&\le C R^{2(\nu+1)}\left(\int_0^{4/R}
|g(y)|^{p_0}y^{(\nu+1/2)(2-p_0)}\,dy\right)^{1/p_0}
\left(\int_0^{4/R}y^{(2\nu+1)}\,dy\right)^{1/p'_0}\\
&=C R^{\frac{2(\nu+1)}{p_0}}
\left(\int_0^{4/R}|g(y)|^{p_0}y^{(\nu+1/2)(2-p_0)}\,dy\right)^{1/p_0}
\le C R^{\frac{2(\nu+1)}{p_0}}I_{p_0}^{1/p_0}.
\end{aligned}\end{gathered}$$ Therefore, $$\begin{aligned}
\sup_{R>0}x^{-(\nu+1/2)}
\int_0^{1}|g(y)|\chi_{B_1}(x,y)|K_R^{\delta}(x,y)|\,dy
&\le C\sup_{R>0}\chi_{[0,4/R]}(x)R^{\frac{2(\nu+1)}{p_0}}
I_{p_0}^{1/p_0}\\&\le C
x^{-\frac{2(\nu+1)}{p_0}}I_{p_0}^{1/p_0}.\end{aligned}$$ In the case $p=1$, it is clear that $$x^{-(\nu+1/2)}\int_0^{1}|g(y)|\chi_{B_1}(x,y)|K_R^{\delta}(x,y)|\,dy\le
C R^{2(\nu+1)}I_1$$ and $$\sup_{R>0}x^{-(\nu+1/2)}
\int_0^{1}|g(y)|\chi_{B_1}(x,y)|K_R^{\delta}(x,y)|\,dy\le C
x^{-2(\nu+1)}I_1.$$ Hence, for $p=p_0(\delta)$, $$J_1\subseteq \{x\in(0,1): C x^{-\frac{2(\nu+1)}{p}}I_{p}^{1/p} >\lambda\},$$ and this gives for $i=1$.
In $J_3$, note first that $$\begin{gathered}
\sup_{R>0}x^{-(\nu+1/2)}
\int_0^{1}|g(y)|\chi_{B_3}(x,y)|K_R^{\delta}(x,y)|\,dy\\
=\sup_{R>0}x^{-(\nu+1/2)}\chi_{[4/R,1]}(x)
\left(\int_0^{2/R}|g(y)||K_R^{\delta}(x,y)|\,dy+
\int_{2/R}^{x/2}|g(y)||K_R^{\delta}(x,y)|\,dy\right)\\:=R_1+R_2.\end{gathered}$$ For $R_1$, using , the inequality $x/2<x-y$, which holds in $B_3$, and Hölder inequality with $p=p_0$, $$\begin{aligned}
R_1
&\le \sup_{R>0}x^{-(\nu+3/2+\delta)}\chi_{[4/R,1]}(x)
\int_0^{2/R}R^{\nu+1/2-\delta}y^{\nu+1/2}|g(y)|\,dy\\
&\le\sup_{R>0}x^{-(\nu+3/2+\delta)}\chi_{[4/R,1]}(x)
R^{\nu+1/2-\delta}R^{-\frac{2(\nu+1)}{p'_0}}I_{p_0}^{1/p_0}
\le C x^{-\frac{2(\nu+1)}{p_0}}I_{p_0}^{1/p_0},\end{aligned}$$ where $I_{p_0}$ is the same as in . In the case $p=1$, the estimate $R_1\le C x^{-2(\nu+1)}I_{1}$ can be obtained easily.
On the other hand, for $R_2$, by using and Hölder inequality with $p=p_0$ again, $$\begin{aligned}
R_2
&\le \sup_{R>0}x^{-(\nu+3/2+\delta)}
\chi_{[4/R,1]}(x)I_{p_0}^{1/p_0}R^{-\delta}
\left(\int_{2/R}^{x/2}y^{-(\nu+1/2)\frac{(2-p_0)p'_0}{p_0}}\,dy\right)^{1/p'_0}\\
&\le \sup_{R>0}x^{-(\nu+3/2+\delta)}
\chi_{[4/R,1]}(x)I_{p_0}^{1/p_0}R^{-\delta}
\left(\int_{2/R}^{x/2}y^{(\nu+1/2)\frac{2-p_0}{1-p_0}}\,dy\right)^{1/p'_0}.\end{aligned}$$ Using that $(\nu+1/2)\frac{2-p_0}{1-p_0}<-1$ and $4/R<x<1$, we have that $$\begin{aligned}
R^{-\delta}\left(\int_{2/R}^{x/2}
y^{(\nu+1/2)\frac{2-p_0}{1-p_0}}\,dy\right)^{1/p'_0}
\le C\left(R^{-(\nu+1/2)\frac{2-p_0}{1-p_0}-1}\right)^{1/p'_0}
R^{-\delta}= C\end{aligned}$$ and the last inequality is true because the exponent of $R$ is zero. Then $$R_2\le C x^{\frac{-2(\nu+1)}{p_0}}I_{p_0}^{1/p_0}.$$ In the case $p=1$ applying Hölder inequality, then $$R_2\le \sup_{R>0}x^{-(\nu+3/2+\delta)}\chi_{[4/R,1]}(x)I_1
\,R^{-\delta}\sup_{y\in[2/R,x/2]}y^{-(\nu+1/2)}.$$ Now, if $\nu+1/2>0$ and $\nu+1/2<\delta$, $$\begin{gathered}
\sup_{R>0}\chi_{[4/R,1]}(x)R^{-\delta}\sup_{y\in[2/R,x/2]}y^{-(\nu+1/2)}\\
=C\sup_{R>0}\chi_{[4/R,1]}(x)R^{\nu+1/2-\delta}\le
Cx^{-\nu-1/2+\delta};\end{gathered}$$ and if $\nu+1/2\le0$, $$\begin{gathered}
\sup_{R>0}\chi_{[4/R,1]}(x)R^{-\delta}\sup_{y\in[2/R,x/2]}y^{-(\nu+1/2)}\\
=C\sup_{R>0}\chi_{[4/R,1]}(x)R^{-\delta}x^{-(\nu+1/2)}\le
Cx^{-\nu-1/2+\delta}.\end{gathered}$$ In this manner $$R_2\le C x^{-2(\nu+1)}I_{1}.$$ Therefore, collecting the estimates for $R_1$ and $R_2$ for $p=p_0$ and $p=1$, we have shown that $$J_3\subseteq \{x\in(0,1): C x^{\frac{-2(\nu+1)}{p}}(x)I^{1/p}
>\lambda\},$$ hence we can deduce for $i=3$.
For the region $J_4$, we proceed as follows $$\begin{aligned}
\sup_{R>0}x^{-(\nu+1/2)}&
\int_{0}^1|g(y)|\chi_{B_4}(x,y)|K_R^\delta(x,y)|\,dy\\
&\le \sup_{R>0}x^{-(\nu+1/2)}\chi_{[0,2/R]}(x)
\int_{4/R}^1|g(y)||K_R^\delta(x,y)|\,dy\\
&\kern20pt+\sup_{R>0}x^{-(\nu+1/2)}\chi_{[2/R,1]}(x)
\int_{2x}^1|g(y)||K_R^\delta(x,y)|\,dy\\
&\le C\sup_{R>0}x^{-(\nu+1/2)}\chi_{[0,2/R]}(x)(Rx)^{\nu+1/2}
\int_{4/R}^1\frac{|g(y)|}{R^{\delta}|x-y|^{\delta+1}}\,dy\\
&\kern20pt+ C\sup_{R>0}x^{-(\nu+1/2)}\chi_{[2/R,1]}(x)
\int_{2x}^1\frac{|g(y)|}{R^{\delta}|x-y|^{\delta+1}}\,dy:=S_1+S_2.\end{aligned}$$ We first deal with $S_1$, we use that $y-x>y/2$, then $$\begin{aligned}
S_1\le &C\sup_{R>0}\chi_{[0,2/R]}(x) R^{\nu+1/2-\delta}
\int_{4/R}^1\frac{|g(y)|}{y^{\delta+1}}\,dy\\
&\le C\sup_{R>0}\chi_{[0,2/R]}(x) R^{\nu+1}\int_{4/R}^1\frac{|g(y)|}{\sqrt{y}}\,dy
\le C x^{-(\nu+1)}\int_x^1\frac{|g(y)|}{\sqrt{y}}\,dy.\end{aligned}$$ Now for $p=p_0$ or $p=1$, we have that $2\nu+1-p(\nu+1)>-1$ and Hardy’s inequality [@SteinWeiss Lemma 3.14, p. 196] is applied in the following estimate $$\begin{aligned}
\int_0^1|S_1(x)|^{p}x^{2\nu+1}\,dx& \le C
\int_0^1\left(\int_x^1\frac{|g(y)|}{\sqrt{y}}\,dy\right)^{p}
x^{2\nu+1-p(\nu+1)}\,dx\\
&\le C \int_0^1\left|\frac{g(y)}{\sqrt y}\right|^{p}y^{2\nu+1-p\nu}\,dy
=C\int_0^1|g(y)|^{p}y^{(\nu+1/2)(2-p)}\,dy.\end{aligned}$$ Concerning $S_2$, observe that $\sup_{R>0}\chi_{[2/R,1]}(x)R^{-\delta}\le Cx^{\delta}$, thus $$S_2\le C
x^{-\nu-1/2+\delta}\int_x^1\frac{|g(y)|}{y^{\delta+1}}\,dy.$$ Since for $p=p_0$ or $p=1$ we have that $2\nu+1-p(\nu+1/2-\delta)>-1$, we can use again Hardy’s inequality to complete the required estimate. Indeed, $$\begin{aligned}
\int_0^1|S_2(x)|^{p}x^{2\nu+1}\,dx&
\le C\int_0^1\left(\int_x^1\frac{|g(y)|}{y^{\delta+1}}\,dy\right)^{p}
x^{2\nu+1-p(\nu+1/2-\delta)}\,dx\\
&\le C\int_0^1\left|\frac{g(y)}{y^{\delta+1}}\right|^{p}
y^{2\nu+1-p(\nu+1/2-\delta)+p}\,dy\\&
=C\int_0^1|g(y)|^{p}y^{(\nu+1/2)(2-p)}\,dy.\end{aligned}$$ With the inequalities for $S_1$ and $S_2$, we can conclude for $i=4$.
To prove for $i=2$ we define, for $k$ a nonnegative integer, the intervals $$I_k=[2^{-k-1},2^{-k}], \qquad N_k=[2^{-k-3},2^{-k+2}]$$ and the function $g_k(y)=|g(y)|\chi_{I_k}(y)$. By using for $x/2<y<2x$, with $x\in (0,1)$, we have the bound $$|K_R^\delta (x,y)|\le \frac{C}{R^{\delta}(|x-y|+2/R)^{\delta+1}}.$$ Then $$J_{2}\subset \left\{x\in (0,1): \sup_{R>0}\sum_{k=0}^\infty
\int_{x/2}^{\min{\{2x,1\}}}
\frac{g_k(t)}{R^{\delta}(|x-y|+2/R)^{\delta+1}}\, dy> C \lambda
x^{\nu+1/2}\right\}.$$ Since at most three of these integrals are not zero for each $x\in
(0,1)$ $$\begin{aligned}
J_2&\subset \bigcup_{k=0}^\infty \left\{x\in (0,1):
3\sup_{R>0}\int_{x/2}^{\min{\{2x,1\}}}
\frac{g_k(t)}{R^{\delta}(|x-y|+2/R)^{\delta+1}}\, dy> C \lambda
x^{\nu+1/2}\right\}\\
&\subset \bigcup_{k=0}^\infty \left\{x\in N_k : M(g_k,x)> C
\lambda x^{\nu+1/2}\right\}\end{aligned}$$ where in the las step we have used that $$\sup_{R>0}\int_{x/2}^{\min{\{2x,1\}}}
\frac{g_k(t)}{R^{\delta}(|x-y|+2/R)^{\delta+1}}\, dy\le C
M(g_k,x).$$ By using the estimate $x\simeq 2^{-k}$ for $x\in N_k$, we can check easily that $$J_2\subset \bigcup_{k=1}^\infty \left\{x \in N_k : M(g_k,x)> C
\lambda 2^{-k(\nu+1/2)}\right\}.$$ Finally by using again that $x\simeq 2^{-k}$ for $x\in I_k, N_k$ and the weak type norm inequality for the Hardy-Littlewood maximal function we have $$\begin{aligned}
\int_{J_2}x^{2\nu+1}\, dx &\le C \sum_{k=0}^\infty 2^{-k(2\nu+1)}
\int_{\left\{x\in N_k : M(g_k,x)> C \lambda
2^{-k(\nu+1/2)}\right\}}\, dx\\&\le C \sum_{k=0}^\infty
\frac{2^{pk(\nu+1/2)-k(2\nu+1)}}{\lambda^p}\int_{I_k}|g(y)|^p\,
dy\\&\le \frac{C}{\lambda^p}\int_0^1 |g(y)|^p y^{(\nu+1/2)(2-p)}\,
dy\end{aligned}$$ and the proof is complete.
Proof of Theorem \[th:AcDebilRestMaxRedonda\] {#sec:acdelrest}
=============================================
To conclude the result we have to prove with $g(x)=\chi_E(x)$ and $p=p_1$. For $J_1$ and $J_2$ the result follows by using the steps given in the proof of Theorem \[th:AcDebilMaxRedonda\] for the same intervals. To analyze $J_3$ we proceed as we did for $J_4$ in the proof of Theorem \[th:AcDebilMaxRedonda\]. In this case we obtain that $$\begin{gathered}
\sup_{R>0}x^{-(\nu+1/2)}\int_0^1
|g(y)|\chi_{B_3}(x,y)|K_r^\delta(x,y)|\\\le
C\left(x^{-(\nu+1)}\int_0^x \frac{|g(y)|}{\sqrt{y}}\, dy+
x^{-(\nu+3/2+\delta)}\int_0^x |g(y)| y^\delta\, dy\right).\end{gathered}$$ Now taking into account that for $p=p_1$ we have $2\nu+1-p(\nu+1)<-1$ and $2\nu+1-p(\nu+3/2+\delta)<-1$ we can apply Hardy’s inequalities to obtain that $$\int_0^1\left(x^{-(\nu+1)}\int_0^x \frac{|g(y)|}{\sqrt{y}}\,
dy\right)^{p}x^{2\nu+1}\, dx\le C \int_0^1 |g(y)|^p
y^{(\nu+1/2)(2-p)}\, dy$$ and $$\int_0^1\left(x^{-(\nu+3/2+\delta)}\int_0^x |g(y)| y^\delta\,
dy\right)^{p}x^{2\nu+1}\, dx\le C \int_0^1 |g(y)|^p
y^{(\nu+1/2)(2-p)}\, dy,$$ with these two inequalities we can deduce that holds for $J_3$ with $p=p_1$ in this case.
The main difference with the previous proof appears in the analysis of $J_4$. To deal with this case, we have to use the following lemma [@Ch-Muc Lemma 16.5]
\[lem:Muck\] If $1<p<\infty$, $a>-1$, and $E\subset
[0,\infty)$, then $$\left(\int_{E}x^a\, dx\right)^p\le
2^p(a+1)^{1-p}\int_{E}x^{(a+1)p-1}\, dx.$$
In this case, it is enough to prove that $$\int_{\mathcal{J}}\, d\mu_\nu(x)\le \frac{C}{\lambda^p}\int_{0}^1
\chi_E(y)\,d\mu_\nu(y),$$ where $$\mathcal{J}=\left\{x\in(0,1):
\sup_{R>0}x^{-(\nu+1/2)}\int_0^{1}\chi_E(y)
\chi_{B_4}(x,y)y^{\nu+1/2}|K_R^{\delta}(x,y)|\,dy>\lambda\right\},$$ and this can be deduced immediately by using the inclusion $$\label{ec:final}
\mathcal{J}\subseteq [0,\min\{1,H\}]$$ with $$H^{2(\nu+1)}=\frac{C}{\lambda^p}\int_{0}^1 \chi_E(y)\,d\mu_\nu(y).$$ Let’s prove . By using and the estimate $y-x>y/2$, we have $$\begin{gathered}
\sup_{R>0}x^{-(\nu+1/2)}\int_0^{1}\chi_E(y)
\chi_{B_4}(x,y)y^{\nu+1/2}|K_R^{\delta}(x,y)|\,dy
\\
\le C\sup_{R>0}
R^{-\delta+\nu+1/2}\chi_{[0,2/R]}(x)\int_{4/R}^1\chi_E(y)
y^{-\delta+\nu-1/2}\, dy\\
+ C\sup_{R>0}
R^{-\delta}x^{-(\nu+1/2)}\chi_{[2/R,1]}(x)\int_{2x}^1 \chi_E(y)
y^{-\delta+\nu-1/2}\, dy.\end{gathered}$$ In the first summand we can use that $R^{-\delta+\nu+1/2}\le C
x^{\delta-\nu-1/2}$ and in the second one that $R^{-\delta}\le
x^{\delta}$. Moreover observing that with $p=p_1$ it holds $-\delta+\nu+1/2=2(\nu+1)/p$ we obtain that $$\begin{aligned}
\sup_{R>0}x^{-(\nu+1/2)}\int_0^{1}\chi_E(y)\chi_{B_4}y^{\nu+1/2}|K_R^{\delta}(x,y)|\,dy&\le
C x^{-2(\nu+1)/p}\int_{E}y^{-1+2(\nu+1)/p}\, dy\\
&\le C x^{-2(\nu+1)/p}\int_{E}\,d\mu_\nu(y),\end{aligned}$$ where in the last step we have used Lemma \[lem:Muck\], and this is enough to deduce the inclusion in .
Proofs of Theorem \[th:noweak\] and Theorem \[th:nostrong\] {#sec:negativeths}
===========================================================
This section will be devoted to the proofs of Theorem \[th:noweak\] and Theorem \[th:nostrong\]. To this end we need a suitable identity for the kernel and in order to do that we have to introduce some notation. $H_{\nu}^{(1)}$ will denote the Hankel function of the first kind, and it is defined as follows $$H_{\nu}^{(1)}(z)=J_{\nu}(z)+iY_{\nu}(z),$$ where $Y_{\nu}$ denotes the Weber’s function, given by $$Y_{\nu}(z)=\frac{{J_{\nu}}(z)\cos \nu \pi-J_{-\nu}(z)}{\sin \nu
\pi},\,\, \nu\notin \mathbb{Z}, \text{ and } Y_n(z)=\lim_{\nu\to
n}\frac{{J_{\nu}}(z)\cos \nu \pi-J_{-\nu}(z)}{\sin \nu \pi}.$$ From these definitions, we have $$H_{\nu}^{(1)}(z)=\frac{J_{-\nu}(z)-e^{-\nu \pi i}{J_{\nu}}(z)}{i\sin
\nu
\pi}, \,\, \nu\notin \mathbb{Z},\\
\text{ and } H_n^{(1)}(z)=\lim_{\nu\to n}\frac{J_{-\nu}(z)-e^{-\nu
\pi i}{J_{\nu}}(z)}{i\sin \nu \pi}.$$ For the function $H_\nu^{(1)}$, the asymptotic $$\label{inftyH} H_{\nu}^{(1)}(z)=\sqrt{\frac{2}{\pi
z}}e^{i(z-\nu\pi/2-\pi/4)}[A+O(z^{-1})], \quad |z|>1,\quad -\pi <
\arg(z)<2\pi,$$ holds for some constant $A$.
In [@Ci-Ro Lemma 1] the following lemma was proved
\[lem:expresnucleo\] For $R>0$ the following holds: $$K_R^\delta(x,y)=I_{R,1}^\delta(x,y)+I_{R,2}^\delta(x,y)$$ with $$I_{R,1}^{\delta}(x,y)=(xy)^{1/2}\int_0^{R}z{\left(1-\frac{z^2}{R^2}\right)^\delta}{J_{\nu}}(zx){J_{\nu}}(zy)\,
dz$$ and $$I_{R,2}^{\delta}(x,y)=\lim_{\varepsilon\to 0}
\frac{(xy)^{1/2}}{2}\int_{\mathbf{S_\varepsilon}}
{\left(1-\frac{z^2}{R^2}\right)^\delta}\frac{z H^{(1)}_{\nu}(z){J_{\nu}}(zx){J_{\nu}}(zy)}{{J_{\nu}}(z)}\,dz,$$ where, for each $\varepsilon>0$, $\mathbf{S_\varepsilon}$ is the path of integration given by the interval $R+i[\varepsilon,\infty)$ in the direction of increasing imaginary part and the interval $-R+i[\varepsilon,\infty)$ in the opposite direction.
Then, by Lemma \[lem:expresnucleo\] we have $$\mathcal{K}_R^\delta(x,y)=\mathcal{I}_{R,1}^\delta(x,y)+\mathcal{I}_{R,2}^\delta(x,y)$$ where $\mathcal{I}_{R,j}^\delta(x,y)=(xy)^{-(\nu+1/2)}I_{R,j}^{\delta}(x,y)$ for $j=1,2$. The main tool to deduce our negative results will be the following lemma
\[lem:zero\] For $\nu>-1/2$, $\delta>0$, and $R>0$ it is verified that $$\mathcal{K}_R^\delta(0,y)=\frac{2^{\delta-\nu}\Gamma(\delta+1)}{\Gamma(\nu+1)}R^{2(\nu+1)}
\frac{J_{\nu+\delta+1}(yR)}{(yR)^{\nu+\delta+1}}+\mathcal{I}_{R,2}^\delta(0,y),$$ where $$\label{ec:boundI} \left|\mathcal{I}_{R,2}^\delta(0,y)\right|\le
C\begin{cases} R^{2\nu-\delta+1}, &
yR\le 1,\\
R^{\nu-\delta+1/2}y^{-(\nu+1/2)}, & yR>1.
\end{cases}$$
From , it is clear that $$\mathcal{I}_{R,1}^\delta(0,y)=\frac{y^{-\nu}}{2^\nu\Gamma(\nu+1)}\int_0^R
z^{\nu+1}{\left(1-\frac{z^2}{R^2}\right)^\delta}{J_{\nu}}(zy)\, dz.$$ Now, by using Sonine’s identity [@Wat Ch. 12, 12.11, p. 373] $$\int_0^1 s^{\nu+1}\left(1-s^2\right)^\delta{J_{\nu}}(sy)\,
ds=2^{\delta}\Gamma(\delta+1)\frac{J_{\nu+\delta+1}(y)}{y^{\delta+1}},
\qquad \nu,\delta>-1,$$ we deduce the leading term of the expression for $\mathcal{K}_{R}^\delta(0,y)$.
To control the term $$\mathcal{I}_{R,2}^\delta(0,y)=\lim_{\varepsilon\to
0}\frac{y^{-(\nu+1/2)}}{2}\int_{\mathbf{S_\varepsilon}} {\left(1-\frac{z^2}{R^2}\right)^\delta}\frac{z^{\nu+1/2}
H^{(1)}_{\nu}(z)(zy)^{1/2}{J_{\nu}}(zy)}{{J_{\nu}}(z)}\,dz,$$ we start by using the asymptotic expansions given in and for $H_{\nu}^{(1)}(z)$ and ${J_{\nu}}(z)$. We see that on $\mathbf{S_\varepsilon}$, the path of integration described in Lemma \[lem:expresnucleo\], for $t={\mathop{\rm Im}}(z)$ the estimate $$\left|\frac{H_{\nu}(z)}{{J_{\nu}}(z)}\right|\leq C e^{-2t},$$ holds for $t>0$. Now, from and , it is clear that for $z=\pm R+it$ $$|\sqrt{zy}J_{\nu}(zy)|\le Ce^{yt}\Phi_\nu((R+t)y)$$ where $\Phi_\nu$ is the function in . Then $$|\mathcal{I}_{R,2}^\delta(0,y)|\le C
y^{-(\nu+1/2)}R^{-2\delta}\int_0^\infty
t^{\delta}(R+t)^{\nu+\delta+1/2}\Phi_\nu((R+t)y)e^{-(2-y)t}\, dt.$$ If $y>1/R$ we have the inequality $\Phi_\nu((R+t)y)\le C$, then $$\begin{aligned}
|\mathcal{I}_{R,2}^\delta(0,y)| &\le C y^{-(\nu+1/2)} R^{-2\delta}
\int_0^\infty t^{\delta}(R+t)^{\nu+\delta+1/2}
e^{-(2-y)t}\, dt\\
&\le C y^{-(\nu+1/2)}R^{-\delta}(R^{\nu+1/2}+R^{-\delta})\le C
R^{\nu-\delta+1/2}y^{-(\nu+1/2)}\end{aligned}$$ and follows in this case. If $y\le 1/R$ we obtain the bound in with the estimate $\Phi_\nu((R+t)y)\le C (\Phi_\nu(yR)+(yt)^{\nu+1/2})$. Indeed, $$\begin{gathered}
|\mathcal{I}_{R,2}^\delta(0,y)| \le C y^{-(\nu+1/2)} R^{-2\delta}
\Phi_\nu(yR) \int_0^\infty t^{\delta}(R+t)^{\nu+\delta+1/2}
e^{-(2-y)t}\, dt\\+ C R^{-2\delta} \int_0^\infty
t^{\nu+\delta+1/2}(R+t)^{\nu+\delta+1/2} e^{-(2-y)t}\, dt\\\le C
(R^{2\nu-\delta+1}+R^{\nu-2\delta+1/2}+R^{\nu-\delta+1/2}+R^{-2\delta})\le
R^{2\nu-\delta+1}.\end{gathered}$$
\[lem:cota0\] For $\nu>-1/2$ and $0<\delta\le \nu+1/2$, the estimate $$\|\mathcal{K}_R^\delta(0,y)\|_{L^{p_0}((0,1),d\mu_\nu)}\ge C
R^{\nu-\delta+1/2}(\log R)^{1/p_0}$$ holds.
We will use the decomposition in Lemma \[lem:zero\]. By using and as was done in [@Ci-RoWave Lemma 2.1] we obtain that $$\left\|R^{2(\nu+1)}
\frac{J_{\nu+\delta+1}(yR)}{(yR)^{\nu+\delta+1}}\right\|_{L^{p_0}((0,1),d\mu_\nu)}\ge
C R^{\nu-\delta+1/2}(\log R)^{1/p_0}.$$ With the bound it can be deduced that $$\left\|\mathcal{I}_{R,2}^\delta(0,y)\right\|_{L^{p_0}((0,1),d\mu_\nu)}\le
C R^{\nu-\delta+1/2}.$$ With the previous estimates the proof is completed.
Finally, the last element that we need to prove Theorems \[th:noweak\] and \[th:nostrong\] is the norm inequality for finite linear combinations of the functions $\{\psi_j\}_{j\ge 1}$ contained in the next lemma. Its proof is long and technical and it will be done in the last section.
\[lem:pol\] For $\nu>-1/2$, $R>0$, $1<p<\infty$ and $f$ a linear combination of the functions $\{\psi_j\}_{1\le j\le N(R)}$ with $N(R)$ a positive integer such that $N(R)\simeq R$, the inequality $$\|f\|_{L^\infty ((0,1),d\mu_\nu)}\le C
R^{2(\nu+1)/p}\|f\|_{L^{p,\infty} ((0,1),d\mu_\nu)}$$ holds.
With the bound in Lemma \[lem:cota0\] we have $$\begin{aligned}
(\log R)^{1/p_0}&\le C R^{-2(\nu+1)/p_1}
\left\|\mathcal{K}_{R}^\delta(0,y)\right\|_{L^{p_0}((0,1),d\mu_\nu)}\\
& = C R^{-2(\nu+1)/p_1}
\sup_{\|f\|_{L^{p_1}((0,1),d\mu_\nu)}=1}\left|\int_0^1
\mathcal{K}_{R}^\delta(0,y) f(y)\, d\mu_\nu\right|\\
& = C R^{-2(\nu+1)/p_1}
\sup_{\|f\|_{L^{p_1}((0,1),d\mu_\nu)}=1}\left|\mathcal{B}_{R}^\delta
f(0)\right|.\end{aligned}$$ From the previous estimate the result for $\delta=\nu+1/2$ follows. In the case $\delta<\nu+1/2$ it is obtained by using Lemma \[lem:pol\] because $$\begin{gathered}
R^{-2(\nu+1)/p_1}
\sup_{\|f\|_{L^{p_1}((0,1),d\mu_\nu)}=1}\left|\mathcal{B}_{R}^\delta
f(0)\right|\\\le C
\sup_{\|f\|_{L^{p_1}((0,1),d\mu_\nu)}=1}\left\|\mathcal{B}_{R}^\delta
f(x)\right\|_{L^{p_1,\infty}((0,1),d\mu_\nu)}\end{gathered}$$ since $\mathcal{B}_{R}^\delta f(x)$ is a linear combination of the functions $\{\psi_j\}_{1\le j\le N(R)}$ with $N(R)\simeq R$.
In the case $\delta <\nu+1/2$, the result follows from Theorem \[th:noweak\] by using a duality argument. Indeed, it is clear that $$\begin{aligned}
\sup_{E\subset
(0,1)}\frac{\|\mathcal{B}^{\delta}_{R}\chi_E\|_{L^{p_0}((0,1),d\mu_\nu)}}
{\|\chi_E\|_{L^{p_0}((0,1),d\mu_\nu)}} &=\sup_{E\subset (0,1)}
\sup_{\|f\|_{L^{p_1}((0,1),d\mu_\nu)}=1}
\frac{\left|\int_0^1f(y)\mathcal{B}^{\delta}_{R}\chi_E(y)\,
d\mu_\nu\right|}
{\|\chi_E\|_{L^{p_0}((0,1),d\mu_\nu)}}\notag\\
&= \sup_{\|f\|_{L^{p_1}((0,1),d\mu_\nu)}=1} \sup_{E\subset (0,1)}
\frac{\left|\int_0^1\chi_E(y)\mathcal{B}^{\delta}_{R}f(y)\,
d\mu_\nu\right|}
{\|\chi_E\|_{L^{p_0}((0,1),d\mu_\nu)}}\label{ec:lambdacero}.\end{aligned}$$ By Theorem \[th:noweak\] it is possible to choose a function $g$ such that $\|g\|_{L^{p_1}((0,1),d\mu_\nu)}=1$ and $$\|\mathcal{B}_{R}^\delta
g(x)\|_{L^{p_1,\infty}((0,1),d\mu_\nu)}\ge C (\log R)^{1/p_0}.$$ Then, with the notation $$\mu_\nu(E)=\int_{E}\, d\mu_\nu,$$ we have $$\label{ec:lambda} \lambda^{p_1}\mu_\nu(A)\ge C (\log R)^{p_1/p_0},$$ for some positive $\lambda$ and $A=\{x\in (0,1): |B_{R}^\delta
g(x)|>\lambda\}$. Now, we consider the subsets of $A$ $$A_1=\{x\in (0,1): B_{R}^\delta g(x)>\lambda\} \qquad\text{ and }
\qquad A_2=\{x\in (0,1): B_{R}^\delta g(x)<-\lambda\}$$ and we define $D=A_1$ if $\mu_\nu(A_1)\ge \mu_\nu(A)/2$ and $D=A_2$ otherwise. Then, by , we deduce that $$\label{ec:lambda2} \lambda \ge C \frac{(\log
R)^{1/p_0}}{\mu_\nu(D)^{1/p_1}}.$$ Taking $f=g$ and $E=D$ in and using , we see that $$\sup_{E\subset
(0,1)}\frac{\|\mathcal{B}^{\delta}_{R}\chi_E\|_{L^{p_0}((0,1),d\mu_\nu)}}
{\|\chi_E\|_{L^{p_0}((0,1),d\mu_\nu)}} \ge C
\lambda\frac{\mu_\nu(D)}{\|\chi_D\|_{L^{p_0}((0,1),d\mu_\nu)}} \ge
C (\log R)^{1/p_0}$$ and the proof is complete in this case. For $\delta=\nu+1/2$ the result follows from Theorem \[th:noweak\] with a standard duality argument.
Proof of Lemma \[lem:pol\] {#sec:techlemma}
==========================
To proceed with the proof of Lemma \[lem:pol\] we need some auxiliary results that are included in this section.
We start by defining a new operator. For each non-negative integer $r$, we consider the vector of coefficients $\alpha=(\alpha_1,\dots,\alpha_{r+1})$ and we define $$T_{r,R,\alpha}f(x)=\sum_{\ell=1}^{r+1}\alpha_\ell
\mathcal{B}_{\ell R}^{r}f(x).$$ This new operator is an analogous of the *generalized delayed means* considered in [@SteinDuke]. In [@SteinDuke] the operator is defined in terms of the Cesàro means instead of the Bochner-Riesz means. The properties of $T_{r,R,\alpha}$ that we need are summarized in the next lemma
\[lem:delay\] For each non-negative integer $r$ and $\nu\ge
-1/2$, the following statements hold
1. $T_{r,R,\alpha}f$ is a linear combination of the functions $\{\psi_j\}_{1\le j\le N((r+1)R)}$, where $N((r+1)R)$ is a non-negative integer such that $N((r+1)R)\simeq (r+1)R$;
2. there exists a vector of coefficients $\alpha$, verifying that $|\alpha_\ell|\le A$, for $\ell=1,\dots, r+1$, with $A$ independent of $R$ and such that $T_{r,R,\alpha}f(x)=f(x)$ for each linear combination of the functions $\{\psi_j\}_{1\le j\le
N(R)}$ where $N(R)$ is a positive integer. Moreover, in this case, for $r>\nu+1/2$, $$\|Tf_{r,R,\alpha}\|_{L^1 ((0,1),d\mu_\nu)}\le
C\|f\|_{L^1((0,1),d\mu_\nu)}$$ and $$\|T_{r,R,\alpha}f\|_{L^\infty ((0,1),d\mu_\nu)}\le C
\|f\|_{L^\infty((0,1),d\mu_\nu)},$$ with $C$ independent of $R$ and $f$.
Part a) is a consequence of the definition of $T_{r,R,\alpha}$ and the fact that the $m$-th zero of the Bessel function $J_\nu$, with $\nu \ge-1/2$, is contained in the interval $(m\pi+\nu\pi/2+\pi/2,m\pi+\nu\pi/2+3\pi/4)$.
To prove b) we consider $f(x)=\sum_{j=1}^{N(R)} a_j \psi_j(x)$. In order to obtain the vector of coefficients such that $T_{r,R,\alpha}f(x)=f(x)$ the equations $$\sum_{\ell=1}^{r+1}\alpha_\ell \left(1-\frac{s_{k}^2}{(\ell
R)^2}\right)^r=1,$$ for all $k=1,\dots,N(R)$, should be verified. After some elementary manipulations each one of the previous equations can be written as $$\sum_{j=0}^r
s_k^{2j}\binom{r}{j}\frac{(-1)^j}{R^{2j}}\sum_{\ell=1}^{r+1}
\frac{\alpha_\ell}{\ell^{2j}}=1$$ and this can be considered as a polynomial in $s_k^2$ which must be equal $1$, therefore we have the system of equations $$\sum_{\ell=1}^{r+1}\frac{\alpha_\ell}{\ell^{2j}}=\delta_{j,0},
\qquad j=0,\dots,r.$$ This system has an unique solution because the determinant of the matrix of coefficients is a Vandermonde’s one. Of course for each $\ell=1,\dots,r+1$, it is verified that $|\alpha_\ell|\le A$, with $A$ a constant depending on $r$ but not on $N(R)$.
The norm estimates are consequence of the uniform boundedness $$\|\mathcal{B}_R^\delta f\|_{L^p((0,1),d\mu_\nu)}\le C
\|f\|_{L^p((0,1),d\mu_\nu)},$$ for $p=1$ and $p=\infty$ when $\delta > \nu+1/2$ (see [@Ci-Ro]).
In the next lemma we will control the $L^\infty$-norm of a finite linear combination of the functions $\{\psi_j\}_{j\ge 1}$ by its $L^1$-norm.
\[lem:infty1\] If $\nu>-1/2$ and $f(x)$ is a linear combination of the functions $\{\psi_j\}_{1\le j\le N(R)}$ with $N(R)$ a positive integer such that $N(R)\simeq R$, the inequality $$\|f\|_{L^\infty ((0,1),d\mu_\nu)}\le C
R^{2(\nu+1)}\|f\|_{L^1((0,1),d\mu_\nu)}$$ holds.
It is clear that $$f(x)=\sum_{j=1}^{N(R)} \psi_j(x)\int_0^1 f(y) \psi_j(y)\,
d\mu_\nu(y).$$ Now, using Hölder inequality and Lemma \[Lem:NormaFunc\] we have $$\begin{aligned}
\|f\|_{L^\infty ((0,1),d\mu_\nu)}&\le C\sum_{j=1}^{N(R)}
\|\psi_j\|_{L^\infty
((0,1),d\mu_\nu)}^2\|f\|_{L^{1}((0,1),d\mu_\nu)}\\&\le C
\|f\|_{L^{1}((0,1),d\mu_\nu)} \sum_{j=1}^{N(R)}j^{2\nu+1}\le C
R^{2(\nu+1)}\|f\|_{L^1((0,1),d\mu_\nu)}.\end{aligned}$$
The following lemma is a version in the space $((0,1),d\mu_\nu)$ of Lemma 19.1 in [@Ch-Muc]. The proof can be done in the same way, with the appropriate changes, so we omit it.
\[lem:fuerdeb\] Let $\nu>-1$, $1<p<\infty$ and $T$ be a linear operator defined for functions in $L^1((0,1),d\mu_\nu)$ and such that $$\|Tf\|_{L^\infty ((0,1),d\mu_\nu)}\le A
\|f\|_{L^1((0,1),d\mu_\nu)} \,\text{ and }\,\|Tf\|_{L^\infty
((0,1),d\mu_\nu)}\le B \|f\|_{L^\infty((0,1),d\mu_\nu)},$$ then $$\|Tf\|_{L^\infty ((0,1),d\mu_\nu)}\le C
A^{1/p}B^{1/p'}\|f\|_{L^{p,\infty}((0,1),d\mu_\nu)}.$$
Now, we are prepared to conclude the proof of Lemma \[lem:pol\].
We consider the operator $T_{r,R,\alpha}f$ given in Lemma \[lem:delay\] b) with $r>\nu+1/2$. By Lemma \[lem:delay\] and Lemma \[lem:infty1\] we have $$\begin{aligned}
\|T_{r,R,\alpha}f\|_{L^\infty((0,1),d\mu_\nu)}&\le C
((r+1)R)^{2(\nu+1)}
\|T_{r,R,\alpha}f\|_{L^1((0,1),d\mu_\nu)}\\&\le C
R^{2(\nu+1)}\|f\|_{L^1((0,1),d\mu_\nu)}.\end{aligned}$$ From b) in Lemma \[lem:delay\] we obtain the estimate $$\|T_{r,R,\alpha}f\|_{L^\infty((0,1),d\mu_\nu)}\le C
\|f\|_{L^\infty((0,1),d\mu_\nu)}.$$ So, by using Lemma \[lem:fuerdeb\], we obtain the inequality $$\|T_{r,R,\alpha}f\|_{L^\infty((0,1),d\mu_\nu)}\le C
R^{2(\nu+1)/p}\|f\|_{L^{p,\infty}((0,1),d\mu_\nu)}$$ for any $f\in L^1((0,1),d\mu_{\nu})$. Now, since $T_{r,R,\alpha}f(x)=f(x)$ for a linear combination of the functions $\{\psi_j\}_{1\le j\le N(R)}$, the proof is complete.
[0]{} P. Balodis and A. Córdoba, The convergence of multidimensional Fourier-Bessel series, J. Anal. Math. **77** (1999), 269–286.
L. Carleson and P. Sjölin, Oscillatory integrals and a multiplier problem for the disc, Studia Math. **44** (1972), 287–299.
S. Chanillo and B. Muckenhoupt, Weak type estimates for Cesàro sums of Jacobi polynomial series, Mem. Amer. Math. Soc. **102** (1993).
Ó. Ciaurri and L. Roncal, The Bochner-Riesz means for Fourier-Bessel expansions, J. Funct. Anal. **228** (2005), 89–113.
Ó. Ciaurri and L. Roncal, The wave equation for the Bessel Laplacian, preprint 2012.
Ó. Ciaurri and K. Stempak, Conjugacy for Fourier-Bessel expansions, Studia Math. **176** (2006), 215–247.
Ó. Ciaurri and J. L. Varona, Two-weight norm inequalities for the Cesàro means of generalized Hermite expansions, *J. Comput. Appl. Math.* **178** (2005), 99–110.
J. Duoandikoetxea, *“Fourier Analysis,”* Graduate Studies in Mathematics, 29, American Mathematical Society, Providence, RI, 2001.
J. E. Gilbert, Maximal theorems for some orthogonal series. I., *Trans. Amer. Math. Soc.* **145** (1969), 495–515.
J. J. Guadalupe, M. Pérez, F. J. Ruiz and J. L. Varona, Two notes on convergence and divergence a. e. of Fourier series with respect to some orthogonal systems, *Proc. Amer. Math. Soc.* **116** (1992), 457–464.
G. H. Hardy and M. Riesz, *“A General Theory of Dirichlet Series,”* Cambridge Univ. Press, Cambridge, 1915.
H. Hochstadt, The mean convergence of Fourier-Bessel series, *SIAM Rev.* **9** (1967), 211–218.
C. Meaney, Divergent Cesàro and Riesz means of Jacobi and Laguerre expansions, *Proc. Amer. Math. Soc.* **131** (2003), 3123–3128.
B. Muckenhoupt and D. W. Webb, Two-weight norm inequalities for Cesàro means of Laguerre expansions, *Trans. Amer. Math. Soc.* **353** (2000), 1119–1149.
B. Muckenhoupt and D. W. Webb, Two-weight norm inequalities for the Cesàro means of Hermite expansions, *Trans. Amer. Math. Soc.* **354** (2002), 4525–4537.
E. Stein, Interpolation in polynomial classes and Markoff’s inequality, *Duke Math. J.* **24** (1957), 467–-476.
E. Stein and G. Weiss, *“Introduction to Fourier aAnalysis on Euclidean Spaces,”* Princeton Univ. Press, Princeton, N. J. , 1970.
K. Stempak, Almost everywhere summability of Laguerre series II, *Stud. Math.* **103** (1992), 317-327.
K. Stempak, On convergence and divergence of Fourier-Bessel series, *Elect. Trans. Num. Anal.* **14** (2002), 223-235.
G. N. Watson, *“A Treatise on the Theory of Bessel Functions,”* Cambridge Univ. Press, 1966.
A. Zygmund, *Trigonometric series. Vol. I, II,* Reprinting of the 1968 version, Cambridge Univ. Press, 1977.
[^1]: Research supported by the grant MTM2009-12740-C03-03 from Spanish Government.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A certain class of Frobenius algebras has been used to characterize orthonormal bases and observables on finite-dimensional Hilbert spaces. The presence of units in these algebras means that they can only be realized finite-dimensionally. We seek a suitable generalization, which will allow arbitrary bases, and therefore observables with discrete spectra, to be described within categorical axiomatizations of quantum mechanics. We develop a definition of H\*-algebra that can be interpreted in any symmetric monoidal dagger category, reduces to the classical notion from functional analysis in the category of (possibly infinite-dimensional) Hilbert spaces, and hence provides a categorical way to speak about orthonormal bases and quantum observables in arbitrary dimension. Moreover, these algebras reduce to the usual notion of Frobenius algebra in compact categories. We then investigate the relations between nonunital Frobenius algebras and H\*-algebras. We give a number of equivalent conditions to characterize when they coincide in the category of Hilbert spaces. We also show that they always coincide in categories of generalized relations and positive matrices.'
author:
- |
Samson Abramsky[^1] and Chris Heunen[^2]\
Oxford University Computing Laboratory
bibliography:
- 'frobcliff.bib'
title: |
-algebras and nonunital Frobenius algebras:\
first steps in infinite-dimensional\
categorical quantum mechanics
---
Introduction
============
The context for this paper comes from the ongoing work on *categorical quantum mechanics* [@abramskycoecke:categoricalsemantics; @abramskycoecke:categoricalquantummechanics]. This work has shown how large parts of quantum mechanics can be axiomatized in terms of monoidal dagger categories and structures definable within them. This axiomatization can be used to perform high-level reasoning and calculations relating to quantum information, using diagrammatic methods [@selinger:graphicallanguages]; and also as a basis for exploring foundational issues in quantum mechanics and quantum computation. In particular, a form of Frobenius algebras has been used to give an algebraic axiomatization of *orthonormal bases* and *observables* [@coeckepavlovic:classicalobjects; @coeckepavlovicvicary:bases].
The structures used so far ([*e.g.* ]{}compact closure, Frobenius algebras) have only finite-dimensional realizations in Hilbert spaces. This raises some interesting questions and challenges:
- Find a good general notion of Frobenius structure which works in the infinite-dimensional case in ${{\ensuremath{\mathbf{Hilb}}}}$.
- Use this to characterize general bases and therefore general observables with discrete spectra.
- Similarly extend the analysis for other categories.
- Clarify the mathematics, and relate it to the wider literature.
As we shall see, an intriguing problem remains open, but much of this program of work has been accomplished.
The further contents of the paper are as follows. Section \[sec:background\] recalls some background on monoidal dagger categories and Frobenius algebras, and poses the problem. Section \[sec:hstar\] introduces the key notion of [H\*]{}-algebra, in the general setting of symmetric monoidal dagger categories. In Section \[sec:hstarinhilb\], we prove our results relating to ${{\ensuremath{\mathbf{Hilb}}}}$, the category of Hilbert spaces (of unrestricted dimension). We show how [H\*]{}-algebras provide exactly the right algebraic notion to characterize orthonormal bases in arbitrary dimension. We give several equivalent characterizations of when [H\*]{}-algebras and nonunital Frobenius algebras coincide in the category of Hilbert spaces. Section \[sec:hstarinrel\] studies [H\*]{}-algebras in categories of generalized relations and positive matrices. We show that in these settings, where no phenomena of ‘destructive interference’ arise, [H\*]{}-algebras and nonunital Frobenius algebras always coincide. Finally, Section \[sec:outlook\] provides an outlook for future work.
Background {#sec:background}
==========
The basic setting is that of *dagger symmetric monoidal categories*. We briefly recall the definitions, referring to [@abramskycoecke:categoricalquantummechanics] for further details and motivation.
A *dagger category* is a category ${\ensuremath{\mathbf{D}}}$ equipped with an identity-on-objects, contravariant, strictly involutive functor. Concretely, for each arrow $f \colon A \to B$, there is an arrow $f^\dag \colon B \to A$, and this assignment satisfies $${\ensuremath{\mathrm{id}_{}}}^\dag = {\ensuremath{\mathrm{id}_{}}}, \qquad
(g \circ f)^\dag = f^\dag \circ g^\dag, \qquad
f^{\dag\dag} = f \, .$$ An arrow $f \colon A \to B$ is *dagger monic* when $f^\dag
\circ f = {\ensuremath{\mathrm{id}_{A}}}$, and a *dagger iso(morphism)* if both $f$ and $f^\dag$ are dagger monics.
A *symmetric monoidal dagger category* is a dagger category with a symmetric monoidal structure $({\ensuremath{\mathbf{D}}}, {\ensuremath{\otimes}}, I, \lambda,
\rho, \alpha, \sigma)$ such that $$(f \otimes g)^\dag = f^\dag \otimes g^\dag$$ and moreover the natural isomorphisms $\lambda$, $\rho$, $\alpha$, $\sigma$ are componentwise dagger isomorphisms.
### Examples {#ex:cats .unnumbered}
- The category ${{\ensuremath{\mathbf{Hilb}}}}$ of Hilbert spaces and continuous linear maps, and its (full) subcategory ${{\ensuremath{\mathbf{fHilb}}}}$ of finite-dimensional Hilbert spaces. Here the dagger is the adjoint, and the tensor product has its standard interpretation for Hilbert spaces. Dagger isomorphisms are *unitaries*, and dagger monics are *isometries*.
- The category ${{\ensuremath{\mathbf{Rel}}}}$ of sets and relations. Here the dagger is relational converse, while the monoidal structure is given by the cartesian product. This generalizes to relations valued in a commutative quantale [@rosenthal:quantales], and to the category of relations of any regular category [@carbonikasangianstreet:relations]. This has a full sub-category ${{\ensuremath{\mathbf{fRel}}}}$, of finite sets and relations.
- The category ${{\ensuremath{\mathbf{lbfRel}}}}$, of *locally bifinite relations*. This is the subcategory of ${{\ensuremath{\mathbf{Rel}}}}$ comprising those relations which are image-finite, meaning that each element in the domain is related to only finitely many elements in the codomain, and whose converses are also image-finite. This forms a monoidal dagger subcategory of ${{\ensuremath{\mathbf{Rel}}}}$. It serves as a kind of qualitative approximation of the passage from finite- to infinite-dimensional Hilbert spaces. For example, a set carries a compact structure in ${{\ensuremath{\mathbf{lbfRel}}}}$ if and only if it is finite.
- A common generalization of ${{\ensuremath{\mathbf{fHilb}}}}$ and ${{\ensuremath{\mathbf{fRel}}}}$ is obtained by forming the category ${{\ensuremath{\mathbf{Mat}}}}(S)$, where $S$ is a commutative semiring with a specified involution [@heunen:hilbcatsembedding]. Objects of ${{\ensuremath{\mathbf{Mat}}}}(S)$ are finite sets, and morphisms are maps , which we think of as ‘$X$ times $Y$ matrices’. Composition is by matrix multiplication, while the dagger is conjugate transpose, where conjugation of a matrix means elementwise application of the involution on $S$. The tensor product of $X$ and $Y$ is given by $X \times Y$, with the action on matrices given by componentwise multiplication, corresponding to the ‘Kronecker product’ of matrices. If we take $S = {\ensuremath{\mathbb{C}}}$, this yields a category equivalent to ${{\ensuremath{\mathbf{fHilb}}}}$, while taking $S$ to be the Boolean semiring $\{0, 1\}$, with trivial involution, gives ${{\ensuremath{\mathbf{fRel}}}}$.
- An infinitary generalization of ${{\ensuremath{\mathbf{Mat}}}}({\ensuremath{\mathbb{C}}})$ is given by ${{{\ensuremath{\mathbf{Mat}}}}_{\ell^2}({\ensuremath{\mathbb{C}}})}$. This category has arbitrary sets as objects, and its morphisms $X \to Y$ are matrices such that for each $x \in X$, the family $\{ M(x, y) \}_{y \in Y}$ is $\ell^2$[-]{}summable; and for each $y \in Y$, the family $\{ M(x, y)
\}_{x \in X}$ is $\ell^2$[-]{}summable. The category ${{\ensuremath{\mathbf{Hilb}}}}$ is equivalent to a (nonfull) subcategory of ${{{\ensuremath{\mathbf{Mat}}}}_{\ell^2}({\ensuremath{\mathbb{C}}})}$ [@blanketal:hilbert Theorem 3.1.7].
### Graphical Calculus {#graphical-calculus .unnumbered}
We briefly recall the graphical calculus for symmetric monoidal dagger categories [@selinger:graphicallanguages]. This can be seen as a two-dimensional version of *Dirac notation*, which allows equational reasoning to be performed graphically in a sound and complete fashion. A morphism $f \colon X \to Y$ is represented pictorially as ${
\begin{tikzpicture}[scale=0.5,font=\scriptsize] \PandocStartInclude{tikz/morphism.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$, the identity on $X$ simply becomes ${
\begin{tikzpicture}[thick,scale=0.5,font=\scriptsize] \PandocStartInclude{tikz/id.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$, and composition and tensor products appear as follows. $${
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/composition-lhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\;=\;
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/composition-rhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\qquad\qquad
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/tensor-lhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\;=\;
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/tensor-rhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ The symmetry isomorphism $\sigma$ is drawn as $\vcenter{\hbox{{
\begin{tikzpicture}[scale=0.25] \PandocStartInclude{tikz/swap.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}}}$. The dagger is represented graphically by a horizontal reflection.
Dagger Frobenius algebras
-------------------------
Frobenius algebras are a classic notion in mathematics [@nakayama:frobenius]. A particular form of such algebras was introduced in the general setting of monoidal dagger categories by Coecke and Pavlovi[ć]{} in [@coeckepavlovic:classicalobjects]. In their version, a *dagger Frobenius structure* on an object $A$ in a dagger monoidal category is a commutative comonoid $(\xymatrix@1{I & A \ar|-{\varepsilon}[l] \ar|-{\delta}[r] & A {\ensuremath{\otimes}}A})$ satisfying certain additional equations: $$\begin{aligned}
({\ensuremath{\mathrm{id}_{A}}} \otimes \delta) {\circ}\delta
& = (\delta \otimes {\ensuremath{\mathrm{id}_{A}}}) {\circ}\delta, \tag{A} \\
({\ensuremath{\mathrm{id}_{A}}} {\ensuremath{\otimes}}\varepsilon) {\circ}\delta
& = {\ensuremath{\mathrm{id}_{A}}}, \tag{U} \\
\sigma {\circ}\delta
& = \delta, \tag{C} \\
\delta^\dag {\circ}\delta
& = {\ensuremath{\mathrm{id}_{A}}}, \tag{M} \\
\delta{\circ}\delta^\dag
& = (\delta^\dag {\ensuremath{\otimes}}{\ensuremath{\mathrm{id}_{A}}}) {\circ}({\ensuremath{\mathrm{id}_{A}}} {\ensuremath{\otimes}}\delta). \tag{F}\end{aligned}$$ These equations become more perspicuous when represented diagrammatically, as below. Here, we draw the comultiplication $\delta$ as $\vcenter{\hbox{{
\begin{tikzpicture}[scale=0.2] \PandocStartInclude{tikz/comultiplication.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}}}$, and the counit $\varepsilon$ as $\vcenter{\hbox{{
\begin{tikzpicture}[scale=0.2] \PandocStartInclude{tikz/counit.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}}}$. $${
\begin{tikzpicture}[] \PandocStartInclude{tikz/A-lhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}} \; \stackrel{\text{(A)}}{=} \; {
\begin{tikzpicture}[] \PandocStartInclude{tikz/A-rhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\qquad\qquad\qquad
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/U-lhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}} \; \stackrel{\text{(U)}}{=} \; {
\begin{tikzpicture}[] \PandocStartInclude{tikz/U-rhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\qquad\qquad\qquad
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/C-lhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}} \; \stackrel{\text{(C)}}{=} \; {
\begin{tikzpicture}[] \PandocStartInclude{tikz/C-rhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ $${
\begin{tikzpicture}[] \PandocStartInclude{tikz/M-lhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}} \; \stackrel{\text{(M)}}{=} \; {
\begin{tikzpicture}[] \PandocStartInclude{tikz/M-rhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\qquad\qquad\qquad
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/F-lhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}} \; \stackrel{\text{(F)}}{=} \; {
\begin{tikzpicture}[] \PandocStartInclude{tikz/F-rhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ A ‘right-handed version’ of the Frobenius law (F) follows from (C); in the noncommutative case we should add this symmetric version (F’) to axiom (F).
Dagger Frobenius algebras in quantum mechanics
----------------------------------------------
Frobenius algebras provide a high-level algebraic way of talking about *orthonormal bases*, and hence can be seen as modeling quantum mechanical *observables*.
To put this in context, we recall the *no-cloning theorem* [@wootterszurek:nocloning], which says that there is no quantum evolution ([*i.e.* ]{}unitary operator) $f
\colon H \to H {\ensuremath{\otimes}}H$ such that, for any ${{|} \phi\rangle} \in H$, $$f{{|} \phi\rangle} = {{|} \phi\rangle} {\ensuremath{\otimes}}{{|} \phi\rangle}.$$ A general form of no-cloning holds for structural reasons in categorical quantum mechanics [@abramsky:nocloning]. In particular, there is no *natural*, [*i.e.* ]{}uniform or basis-independent, family of diagonal morphisms in a compact closed category, unless the category collapses, so that endomorphisms are scalar multiples of the identity.
However, if we drop naturality, we *can* define such maps in ${\ensuremath{\mathbf{Hilb}}}$ in a basis-dependent fashion. Moreover, it turns out that such maps can be used to *uniquely determine* bases. Firstly, consider *copying maps*, which can be defined in arbitrary dimension: for a given basis $\{{{|} i\rangle}\}_{i \in I}$ of $H$, define $\delta \colon H \to H {\ensuremath{\otimes}}H$ by (continuous linear extension of) ${{|} i\rangle} \mapsto {{|} ii\rangle}$.
For example, consider the map $\delta_{\text{std}} \colon {\ensuremath{\mathbb{C}}}^2 \to
{\ensuremath{\mathbb{C}}}^2 {\ensuremath{\otimes}}{\ensuremath{\mathbb{C}}}^2$ defined by $${{|} 0\rangle} \mapsto {{|} 00\rangle}, \qquad {{|} 1\rangle} \mapsto {{|} 11\rangle} .$$ By construction, this copies the elements of the computational basis — and *only* these, as in general $$\delta_{\text{std}}(\alpha{{|} 0\rangle}+\beta{{|} 1\rangle}) =
\alpha{{|} 00\rangle}+\beta{{|} 11\rangle} \neq (\alpha{{|} 0\rangle}+\beta{{|} 1\rangle})
{\ensuremath{\otimes}}(\alpha{{|} 0\rangle}+\beta{{|} 1\rangle}).$$ Next, consider *deleting maps* $\varepsilon \colon H \to
{\ensuremath{\mathbb{C}}}$ by linearly extending ${{|} e_i\rangle} \mapsto 1$. In contrast to copying, these can be defined in *finite dimension only*. It is straightforward to verify that these maps define a dagger Frobenius structure on $H$. Moreover, the following result provides a striking converse.
\[cpvth\] [@coeckepavlovicvicary:bases] Orthonormal bases of a finite-dimensional Hilbert space $H$ are in one-to-one correspondence with dagger Frobenius structures on $H$.
This result in fact follows easily from previous results in the literature on Frobenius algebras [@abrams:thesis]; we will give a short proof from the established literature in Section \[subsec:furtherconditions\].
Another result provides a counterpart—at first sight displaying very different looking behaviour—in the category ${{\ensuremath{\mathbf{Rel}}}}$.
[@pavlovic:frobeniusinrel] Dagger Frobenius structures in the category ${{\ensuremath{\mathbf{Rel}}}}$ correspond to disjoint unions of abelian groups.
We shall provide a different proof of this result in Section \[frobrelsec\], which makes no use of units, and hence generalizes to a wide range of other situations, such as locally bifinite and quantale-valued relations, and positive $\ell_2$-matrices.
The problem
-----------
The notion of Frobenius structure as defined above, which requires a unit, limits us to the *finite-dimensional case* in ${{\ensuremath{\mathbf{Hilb}}}}$, as the following lemma shows.
A Frobenius algebra in ${\ensuremath{\mathbf{Hilb}}}$ is unital if and only if it is finite-dimensional.
Sufficiency is shown in [@kock:frobenius 3.6.9]. Necessity follows from [@kaplansky:dualrings Corollary to Theorem 4].
In fact, a Frobenius structure on an object $A$ induces a *compact* (or *rigid*) structure on $A$, with $A$ as its own dual (see [@abramskycoecke:categoricalquantummechanics]). Indeed, put . In the category ${{\ensuremath{\mathbf{fHilb}}}}$, for example, $\eta
\colon {\ensuremath{\mathbb{C}}} \to {\ensuremath{\mathbb{C}}}^2 {\ensuremath{\otimes}}{\ensuremath{\mathbb{C}}}^2$ is an *entangled state preparation*: $$\eta_{\text{std}}
= \delta_{\text{std}}{\circ}\varepsilon^\dag_{\text{std}}
= (1 \mapsto \delta_{\text{std}}({{|} 0\rangle}+{{|} 1\rangle}))
= (1 \mapsto {{|} 00\rangle} + {{|} 11\rangle}).$$ In general it is easy to see that $\eta$ indeed provides a dagger compact structure on $A$, with $A^*=A$: $${
\begin{tikzpicture}[scale=0.33] \PandocStartInclude{tikz/compact-1.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(C)}}{=} \;
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/compact-2.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(F)}}{=} \;
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/compact-3.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(U)}}{=} \;
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/compact-4.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}.$$ As is well-known, a compact structure exists only for finite-dimensional spaces in ${\ensuremath{\mathbf{Hilb}}}$. Thus to obtain a notion capable of being extended beyond the finite-dimensional case, we need to drop the assumption of a unit.
[H\*]{}-algebras {#sec:hstar}
================
We begin our investigation of suitable axioms for a notion of algebra which can characterize orthonormal bases in arbitrary dimension by recalling the axioms for Frobenius structures. $$\begin{aligned}
({\ensuremath{\mathrm{id}_{A}}} \otimes \delta) {\circ}\delta
& = (\delta \otimes {\ensuremath{\mathrm{id}_{A}}}) {\circ}\delta \tag{A} \\
({\ensuremath{\mathrm{id}_{A}}} {\ensuremath{\otimes}}\varepsilon) {\circ}\delta
& = {\ensuremath{\mathrm{id}_{A}}} \tag{U} \\
\sigma {\circ}\delta
& = \delta \tag{C} \\
\delta^\dag {\circ}\delta
& = {\ensuremath{\mathrm{id}_{A}}} \tag{M} \\
\delta{\circ}\delta^\dag
& = (\delta^\dag {\ensuremath{\otimes}}{\ensuremath{\mathrm{id}_{A}}}) {\circ}({\ensuremath{\mathrm{id}_{A}}} {\ensuremath{\otimes}}\delta) & \tag{F}\end{aligned}$$ We note in passing that there is some redundancy in the definition of Frobenius structure.
In any dagger monoidal category, (M), (F) and (F’) imply (A).
$$\begin{aligned}
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/A-lhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
& \; \stackrel{\text{(M)}}{=} \;
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/MFimplyA-1.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(F)}}{=} \;
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/MFimplyA-2.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(F')}}{=} \;
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/MFimplyA-3.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}} \\
& \; \stackrel{\text{(F)}}{=} \;
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/MFimplyA-4.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(F')}}{=} \;
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/MFimplyA-5.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(M)}}{=} \;
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/A-rhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
{\tag*{$\Box$}}\end{aligned}$$
The axioms (U), (C), and (F) are independent:
- As we have seen, for an orthonormal basis $\{{{|} n\rangle} \mid n \in
{\ensuremath{\mathbb{N}}}\}$ of a separable (infinite-dimensional) Hilbert space, the map $\delta({{|} n\rangle}) = {{|} nn\rangle}$ satisfies everything except for (U).
- Group algebras of finite noncommutative groups [@ambrose:hstar Example 4] satisfy everything except for (C).
- Any nontrivial commutative (unital) Hopf algebra satisfies everything except for (F) by [@kock:frobenius Proposition 2.4.10].
It is worth noting that under additional assumptions, such as unitality and enrichment in abelian groups, (A) and (M) are known to imply (F) [@longoroberts:dimension Section 6].
We shall now *redefine* a Frobenius algebra[^3] in a dagger monoidal category to be an object $A$ equipped with a comultiplication $\delta : A \to A {\ensuremath{\otimes}}A$ satisfying (A), (C), (M) and (F). A Frobenius algebra which additionally has an arrow $\varepsilon \colon A \to I$ satisfying (U) will explicitly be called *unital*.
Regular representation as pointwise abstraction
-----------------------------------------------
As we have seen, unital Frobenius algebras allow us to define compact, and hence closed, structure. How much of this can we keep in key examples such as ${\ensuremath{\mathbf{Hilb}}}$?
The category ${\ensuremath{\mathbf{Hilb}}}$ has well-behaved duals, since $H \cong
H^{**}$, and indeed there is a conjugate-linear isomorphism $H \cong
H^{*}$. However, it is *not* the case that the tensor unit ${\ensuremath{\mathbb{C}}}$ is exponentiable in ${\ensuremath{\mathbf{Hilb}}}$, since if it was, we would have a bounded linear evaluation map $$H {\ensuremath{\otimes}}H^* \to {\ensuremath{\mathbb{C}}},$$ and hence its adjoint ${\ensuremath{\mathbb{C}}} \to H {\ensuremath{\otimes}}H^*$, and a compact structure.
We shall now present an axiom which captures what seems to be the best we can do in general in the way of a ‘transfer of variables’. It is, indeed, a general form, meaningful in any monoidal dagger category, of a salient structure in functional analysis.
Suppose we have a comultiplication $\delta \colon A
\to A {\ensuremath{\otimes}}A$, and hence a multiplication $\mu = \delta^\dag
\colon A {\ensuremath{\otimes}}A \to A$. We can *curry* the multiplication (this process is also called $\lambda$-abstraction [@barendregt:lambdacalculus]) for *points*—this is just the regular representation![^4] Thus we have a function $R \colon {\ensuremath{\mathbf{D}}}(I,A) \to {\ensuremath{\mathbf{D}}}(A,A)$ defined by $$R(a) = \mu {\circ}({\ensuremath{\mathrm{id}_{}}}{\ensuremath{\otimes}}a) = {
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/regular.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}.$$ If $\mu$ is associative, this is a semigroup homomorphism.
Axiom (H)
---------
An endomorphism homset ${\ensuremath{\mathbf{D}}}(A,A)$ in a dagger category ${\ensuremath{\mathbf{D}}}$ is not just a monoid, but a *monoid with involution*, because of the dagger. We say that $(A, \mu)$ *satisfies axiom (H)* if there is an operation $a
\mapsto a^*$ on ${\ensuremath{\mathbf{D}}}(I, A)$ such that $R$ becomes a homomorphism of involutive semigroups, [*i.e.* ]{} $$R(a^*) = R(a)^\dag$$ for every $a \colon I \to A$. This unfolds to $$\begin{aligned}
\mu {\circ}(a^* {\ensuremath{\otimes}}{\ensuremath{\mathrm{id}_{}}})
= (a^\dag {\ensuremath{\otimes}}{\ensuremath{\mathrm{id}_{}}}) {\circ}\mu^\dag; \tag{H}\end{aligned}$$ or diagrammatically: $${
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/H-lhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(H)}}{=} \;
{
\begin{tikzpicture}[scale=0.5] \PandocStartInclude{tikz/H-rhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}.$$ Thus $a \mapsto a^*$ is indeed a ‘transfer of variables’.
Relationships between axioms (F) and (H)
----------------------------------------
The rest of this section compares axioms (F) and (H) at the abstract level of monoidal dagger categories. The following observation by Coecke, Pavlovi[ć]{} and Vicary is the central idea in their proof of Theorem \[cpvth\].
\[lem:cpv\] In any dagger monoidal category, (F) and (U) imply (H).
Define $a^* = (a^\dag {\ensuremath{\otimes}}{\ensuremath{\mathrm{id}_{}}}) {\circ}\delta {\circ}\varepsilon^\dag$. $${
\begin{tikzpicture}[] \PandocStartInclude{tikz/FUimplyH-1.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}} \;=\; {
\begin{tikzpicture}[] \PandocStartInclude{tikz/FUimplyH-2.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ This indeed satisfies (H). $${
\begin{tikzpicture}[] \PandocStartInclude{tikz/FUimplyH-3.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; = \;
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/FUimplyH-4.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(F)}}{=} \;
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/FUimplyH-5.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(U)}}{=} \;
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/FUimplyH-6.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$
Recall that a category is *monoidally well-pointed* if the following holds: $$\begin{aligned}
f = g \colon A {\ensuremath{\otimes}}A' \to B {\ensuremath{\otimes}}B'
\;\;\; \Longleftrightarrow \;\;\;
\forall x \colon I \to A, y \colon I \to A' .\,\,f {\circ}(x {\ensuremath{\otimes}}y) = g {\circ}(x {\ensuremath{\otimes}}y).\end{aligned}$$ All the categories listed in our Examples are monoidally well-pointed in this sense.
\[lem:HAWPimplyF\] In a monoidally well-pointed dagger monoidal category, (H) and (A) imply (F).
For any $a \colon I \to A$ we have the following. $${
\begin{tikzpicture}[] \PandocStartInclude{tikz/HAimplyF-1.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(H)}}{=} \;
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/HAimplyF-2.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(A)}}{=} \;
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/HAimplyF-3.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(H)}}{=} \;
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/HAimplyF-4.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ Then (F) follows from monoidal well-pointedness.
Lemma \[lem:cpv\] is strengthened by the following proposition, which proves that compactness implies unitality.
\[prop:cptimpliesunital\] Any Frobenius algebra in a dagger compact category is unital.
[@carboni:matrices Remark (1) on page 503] Suppose that $\delta \colon A \to A {\ensuremath{\otimes}}A$ is a nonunital Frobenius algebra in a compact category. Define $\varepsilon \colon A
\to I$ as follows. $$\varepsilon \;=\; {
\begin{tikzpicture}[] \PandocStartInclude{tikz/counit.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}} \;=\; {
\begin{tikzpicture}[] \PandocStartInclude{tikz/cptcounit.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ Then the following holds, where we draw the unit and counit of compactness by caps and cups (without dots). $$\begin{aligned}
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/cptFimpliesH-1.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
& =
& {
\begin{tikzpicture}[] \PandocStartInclude{tikz/cptFimpliesH-2.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(F)}}{=} \;
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/cptFimpliesH-3.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(C)}}{=} \;
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/cptFimpliesH-4.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}} \\
& \stackrel{\text{(F)}}{=}
& {
\begin{tikzpicture}[] \PandocStartInclude{tikz/cptFimpliesH-5.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; = \;
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/cptFimpliesH-6.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(C)}}{=} \;
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/cptFimpliesH-7.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; \stackrel{\text{(M)}}{=} \;
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/cptFimpliesH-8.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\end{aligned}$$ That is, (U) holds.
Thus in the unital, monoidally well-pointed case, (F) and (H) are essentially equivalent. Our interest is, of course, in the nonunital case. To explain the provenance of the (H) axiom, and its implications for obtaining a correspondence with orthonormal bases in ${\ensuremath{\mathbf{Hilb}}}$ in arbitrary dimension, we shall now study the situation in the concrete setting of Hilbert spaces.
[H\*]{}-algebras in ${\ensuremath{\mathbf{Hilb}}}$ {#sec:hstarinhilb}
==================================================
We begin by revisiting Theorem \[cpvth\]. How should the correspondence between Frobenius algebras and orthonormal bases be expressed mathematically? In fact, the content of this result is really a *structure theorem* of a classic genre in algebra [@albert:structureofalgebras]. The following theorem, the *Wedderburn structure theorem*, is the prime example; it was subsequently generalized by Artin, and there have been many subsequent developments.
Every finite-dimensional semisimple algebra is isomorphic to a product of full matrix algebras. In the commutative case over the complex numbers, this has the form: the algebra is isomorphic to a product of one-dimensional complex algebras.
To see the connection between the Wedderburn structure theorem and Theorem \[cpvth\], consider the coalgebra $A$ determined by an orthonormal basis $\{ {{|} i\rangle} \}$ on a Hilbert space: $$\delta \colon {{|} i\rangle} \mapsto {{|} ii\rangle}.$$ This is isomorphic as a coalgebra to a direct sum of one-dimensional coalgebras $$\delta_{{\ensuremath{\mathbb{C}}}} \colon {\ensuremath{\mathbb{C}}} \to {\ensuremath{\mathbb{C}}} {\ensuremath{\otimes}}{\ensuremath{\mathbb{C}}},
\qquad 1 \mapsto 1 {\ensuremath{\otimes}}1.$$ To say that a Frobenius algebra corresponds to an orthonormal basis is exactly to say that it is isomorphic as a coalgebra to a Hilbert space direct sum of one-dimensional coalgebras: $$A \; \cong \; \bigoplus_I \, ({\ensuremath{\mathbb{C}}}, \delta_{{\ensuremath{\mathbb{C}}}}),$$ where the cardinality of $I$ is the dimension of $H$. Applying dagger, this is equivalent to $A$ being isomorphic as an *algebra* to the direct sum of one-dimensional *algebras* $$A \; \cong \; \bigoplus_I \, ({\ensuremath{\mathbb{C}}}, \mu_{{\ensuremath{\mathbb{C}}}}),
\qquad \mu_{{\ensuremath{\mathbb{C}}}} \colon 1 {\ensuremath{\otimes}}1 \mapsto 1.$$ In this case, we say that the Frobenius algebra *admits the structure theorem*, making the view of bases as (co)algebras precise.
[H\*]{}-algebras {#h-algebras}
----------------
There is a remarkable generalization of the Wedderburn structure theorem to an infinite-dimensional setting, in a classic paper from 1945 by Warren Ambrose on ‘[H\*]{}-algebras’ [@ambrose:hstar]. He defines an [H\*]{}-algebra[^5] as a (not necessarily unital) Banach algebra based on a Hilbert space $H$, such that for each $x \in H$ there is an $x^* \in H$ with $${\ensuremath{\langle xy\,|\,z \rangle}} = {\ensuremath{\langle y\,|\,x^*z \rangle}}$$ for all $y,z \in H$, and similarly for right multiplication. Note that $${\ensuremath{\langle xy\,|\,z \rangle}}
= (\mu {\circ}(x {\ensuremath{\otimes}}y))^\dag {\circ}z
= (x^\dag {\ensuremath{\otimes}}y^\dag) {\circ}\mu^\dag {\circ}z,$$ where we identify points $x \in H$ with morphisms $x \colon
{\ensuremath{\mathbb{C}}} \to H$, and similarly $${\ensuremath{\langle y\,|\,x^*z \rangle}}
= y^\dag {\circ}\mu {\circ}(x^* {\ensuremath{\otimes}}z).$$ Using the monoidal well-pointedness of ${\ensuremath{\mathbf{Hilb}}}$, it is easy to see that this is equivalent to the (H) condition![^6] The following two lemmas show that the assumptions (A), (C), (M) and (H) indeed result in an H\*-algebra.
A monoid in ${\ensuremath{\mathbf{Hilb}}}$ satisfying (M) is a Banach algebra.
The condition (M) implies that $P = \mu^\dag {\circ}\mu$ is a projector: $$P^2
= \mu^\dag {\circ}\mu {\circ}\mu^\dag {\circ}\mu
= \mu^\dag {\circ}\mu
= P$$ and clearly $P = P^\dag$. Hence a monoid in ${\ensuremath{\mathbf{Hilb}}}$ satisfying (M) is a Banach algebra: $$\begin{aligned}
\| xy \|^2
& = {\ensuremath{\langle xy\,|\,xy \rangle}} \\
& = (x^\dag {\ensuremath{\otimes}}y^\dag) {\circ}\mu^\dag {\circ}\mu {\circ}(x {\ensuremath{\otimes}}y) \\
& = {\ensuremath{\langle x {\ensuremath{\otimes}}y\,|\,P(x {\ensuremath{\otimes}}y) \rangle}} \\
& \leq {\ensuremath{\langle x {\ensuremath{\otimes}}y\,|\,x {\ensuremath{\otimes}}y \rangle}} \\
& = {\ensuremath{\langle x\,|\,x \rangle}}{\ensuremath{\langle y\,|\,y \rangle}} \\
& = \| x \|^2 \| y \|^2 .
{\tag*{$\Box$}}\end{aligned}$$
### Remark {#remark .unnumbered}
In fact, it can be shown that the multiplication of a semigroup in ${\ensuremath{\mathbf{Hilb}}}$ satisfying (H) is automatically continuous, so that after adjusting by a constant, the semigroup is a Banach algebra [@ingelstam:realhstar Corollary 2.2].[^7]
The following lemma establishes *properness*, which corresponds to $x^*$ being the unique vector with the property defining H\*-algebras. It follows that $R(x^*)$ is the adjoint of $R(x)$.
Suppose $\delta \colon A \to A {\ensuremath{\otimes}}A$ in ${{\ensuremath{\mathbf{Hilb}}}}$ satisfies (A) and (H). Then (M) implies properness, [*i.e.* ]{}$aA=0
\Rightarrow a=0$. Hence (M) holds if and only if the regular representation is monic.
By [@ambrose:hstar Theorem 2.2], $A$ is the direct sum of its trivial ideal $A'$ and a proper H\*-algebra $A''$. Here, the trivial ideal is $A'=\{a \in A \mid aA=0\}$. Since the direct sum of Hilbert spaces is a dagger biproduct, we can write $\delta$ as $\delta' \oplus \delta''
\colon A \to A {\ensuremath{\otimes}}A$, where $\delta' \colon A' \to A' {\ensuremath{\otimes}}A'$ and $\delta'' \colon A'' \to A'' {\ensuremath{\otimes}}A''$. The latter two morphisms are again dagger monic as a consequence of (M). So the multiplication $\delta'^\dag$ of $A'$ is epic, which forces $A'=0$.
The following proposition summarizes the preceding discussion.
\[prop:hstaralgebra\] Any structure $(A, \mu)$ in ${\ensuremath{\mathbf{Hilb}}}$ satisfying (A), (H) and (M) is an [H\*]{}-algebra (and also satisfies (F)); and conversely, an [H\*]{}-algebra satisfies (A), (H) and (M), and hence also (F).
Ambrose proved a complete structure theorem for H\*-algebras, of which we now state the commutative case.
\[thm:ambrose\] Any proper commutative [H\*]{}-algebra (of arbitrary dimension) is isomorphic to a Hilbert space direct sum of one-dimensional algebras.
This is equivalent to asserting isomorphism qua coalgebras. So it is exactly the result we are after! Rather than relying on Ambrose’s results, we now give a direct, conceptual proof, using a few notions from Gelfand duality for commutative Banach algebras.
Copyables and semisimplicity
----------------------------
A *copyable element* of a semigroup $\delta \colon A \to A {\ensuremath{\otimes}}A$ in a monoidal category is a semigroup homomorphism to it from the canonical semigroup on the monoidal unit. More precisely, a copyable element is a morphism $a \colon I \to A$ such that $(a {\ensuremath{\otimes}}a) {\circ}\delta =
\delta {\circ}a$. In a monoidally well-pointed category such as ${\ensuremath{\mathbf{Hilb}}}$, we can speak of a copyable element of $\delta$ as a point $a \in A$ with $\delta(a)=a {\ensuremath{\otimes}}a$.[^8]
\[lem:independent\] Assuming only (A), nonzero copyable elements are linearly independent.
[@hofmann:duality Theorem 10.18(ii)] Suppose that $\{a_0,\ldots,a_n\}$ is a minimal nonempty linearly dependent set of nonzero copyables. Then we can write $a_0$ as $\sum_{i=1}^n
\alpha_i a_i$ for a suitable choice of coefficients $\alpha_i \in {\ensuremath{\mathbb{C}}}$. So $$\begin{aligned}
\sum_{i=1}^n \alpha_i (a_i {\ensuremath{\otimes}}a_i)
& = \sum_{i=1}^n \alpha_i \delta(a_i) \\
& = \delta(a_0) \\
& = (\sum_{i=1}^n \alpha_i a_i) {\ensuremath{\otimes}}(\sum_{j=1}^n \alpha_j a_j) \\
& = \sum_{i,j=1}^n \alpha_i \alpha_j (a_i {\ensuremath{\otimes}}a_j).
\end{aligned}$$ By minimality, $\{a_1,\ldots,a_n\}$ is linearly independent. Hence $\alpha_i^2=\alpha_i$ for all $i$, and $\alpha_i \alpha_j=0$ for $i \neq j$. So $\alpha_i=0$ or $\alpha_i=1$ for all $i$. If $\alpha_j=1$, then $\alpha_i=0$ for all $i \neq j$, so $a_0=a_j$. By minimality, then $j=1$ and $\{a_0,a_j\} = \{a_0\}$, which is impossible. So we must have $\alpha_i=0$ for all $i$. But then $a_0=0$, which is likewise a contradiction.
\[lem:normal\] Assuming only (M), nonzero copyable elements have unit norm.
Let $a$ be a copyable element. Then $\delta(a) = a {\ensuremath{\otimes}}a$. Hence $$\|a\| = \|\delta(a)\| = \|a
{\ensuremath{\otimes}}a\| = \|a\|^2.$$ It follows that $\|a\|$ is either 0 or 1. Therefore, if $a$ is a nonzero, then $\|a\|=1$.
\[prop:orthogonal\] Assuming only (F), copyable elements are pairwise orthogonal.
[@coeckepavlovicvicary:bases Corollary 4.7] Let $a,b$ be copyables. Then: $$\begin{aligned}
{\ensuremath{\langle a\,|\,a \rangle}} \cdot {\ensuremath{\langle a\,|\,a \rangle}} \cdot {\ensuremath{\langle b\,|\,a \rangle}}
& = {\ensuremath{\langle a {\ensuremath{\otimes}}a {\ensuremath{\otimes}}b\,|\,a {\ensuremath{\otimes}}a {\ensuremath{\otimes}}a \rangle}} \\
& = {\ensuremath{\langle (\delta {\ensuremath{\otimes}}{\ensuremath{\mathrm{id}_{}}})(a {\ensuremath{\otimes}}b)\,|\,({\ensuremath{\mathrm{id}_{}}}{\ensuremath{\otimes}}\delta)(a {\ensuremath{\otimes}}a) \rangle}} \\
& = {\ensuremath{\langle a {\ensuremath{\otimes}}b\,|\,(\delta^\dag {\ensuremath{\otimes}}{\ensuremath{\mathrm{id}_{}}}) {\circ}({\ensuremath{\mathrm{id}_{}}}{\ensuremath{\otimes}}\delta)(a {\ensuremath{\otimes}}a) \rangle}} \\
& = {\ensuremath{\langle a {\ensuremath{\otimes}}b\,|\,({\ensuremath{\mathrm{id}_{}}}{\ensuremath{\otimes}}\delta^\dag) {\circ}(\delta {\ensuremath{\otimes}}{\ensuremath{\mathrm{id}_{}}})(a {\ensuremath{\otimes}}a) \rangle}} \\
& = {\ensuremath{\langle ({\ensuremath{\mathrm{id}_{}}}{\ensuremath{\otimes}}\delta)(a {\ensuremath{\otimes}}b)\,|\,(\delta {\ensuremath{\otimes}}{\ensuremath{\mathrm{id}_{}}})(a {\ensuremath{\otimes}}a) \rangle}} \\
& = {\ensuremath{\langle a {\ensuremath{\otimes}}b {\ensuremath{\otimes}}b\,|\,a {\ensuremath{\otimes}}a {\ensuremath{\otimes}}a \rangle}} \\
& = {\ensuremath{\langle a\,|\,a \rangle}} \cdot {\ensuremath{\langle b\,|\,a \rangle}} \cdot {\ensuremath{\langle b\,|\,a \rangle}}.
\end{aligned}$$ Analogously ${\ensuremath{\langle b\,|\,b \rangle}}{\ensuremath{\langle b\,|\,b \rangle}}{\ensuremath{\langle a\,|\,b \rangle}} =
{\ensuremath{\langle b\,|\,b \rangle}}{\ensuremath{\langle a\,|\,b \rangle}}{\ensuremath{\langle a\,|\,b \rangle}}$. Hence, if ${\ensuremath{\langle a\,|\,a \rangle}}$ and ${\ensuremath{\langle b\,|\,a \rangle}}$ are both nonzero, then ${\ensuremath{\langle a\,|\,a \rangle}}={\ensuremath{\langle b\,|\,a \rangle}}$ and ${\ensuremath{\langle b\,|\,b \rangle}}={\ensuremath{\langle a\,|\,b \rangle}}$. So ${\ensuremath{\langle a\,|\,a \rangle}}, {\ensuremath{\langle b\,|\,b \rangle}} \in {\ensuremath{\mathbb{R}}}$ and ${\ensuremath{\langle a\,|\,a \rangle}}=
{\ensuremath{\langle a\,|\,b \rangle}} = {\ensuremath{\langle b\,|\,a \rangle}} = {\ensuremath{\langle b\,|\,b \rangle}}$. Now suppose ${\ensuremath{\langle a\,|\,b \rangle}}\neq 0$. Then we can conclude ${\ensuremath{\langle a-b\,|\,a-b \rangle}} = {\ensuremath{\langle a\,|\,a \rangle}} - {\ensuremath{\langle a\,|\,b \rangle}} - {\ensuremath{\langle b\,|\,a \rangle}} +
{\ensuremath{\langle b\,|\,b \rangle}} = 0$. So $a-b=0$, [*i.e.* ]{}$a=b$. Hence the copyables are pairwise orthogonal.
Applying dagger, a copyable element of $A$ corresponds exactly to a comonoid homomorphism $({\ensuremath{\mathbb{C}}}, \delta_{{\ensuremath{\mathbb{C}}}})
\to (A, \delta)$: $$\xymatrix{
1 \ar@{|->}[d] \ar@{|->}[r] & 1 {\ensuremath{\otimes}}1 \ar@{|->}[d] \\
a \ar@{|->}[r] & a {\ensuremath{\otimes}}a.
}$$ We have already seen that copyable elements correspond exactly to algebra homomorphisms $$(A, \mu) \to ({\ensuremath{\mathbb{C}}}, \mu_{{\ensuremath{\mathbb{C}}}}),$$ [*i.e.* ]{}to *characters* of the algebra—the elements of the Gelfand spectrum of $A$ [@pedersen:analysisnow]. This leads to our first characterization of when a (nonunital) Frobenius algebra in ${\ensuremath{\mathbf{Hilb}}}$ corresponds to an orthonormal basis.
A Frobenius algebra in ${{\ensuremath{\mathbf{Hilb}}}}$ admits the structure theorem and hence corresponds to an orthonormal basis if and only if it is semisimple.
We first consider sufficiency. Form a direct sum of one-dimensional coalgebras indexed by the copyables of $(A,\delta)$. This will have an isometric embedding as a coalgebra into $(A,\delta)$: $$e \colon \bigoplus_{\{a \mid \delta(a)=a {\ensuremath{\otimes}}a\}} ({\ensuremath{\mathbb{C}}},
\delta_{{\ensuremath{\mathbb{C}}}}) \to (A,\delta).$$ The image $S$ of $e$ is a closed subspace of $A$, and has an orthonormal basis given by the images of the characters of $A$ qua copyables. The structure theorem holds if the image of $e$ spans $A$.
Given $a \in A$ and a character $c$, the evaluation $c(a)$ gives the Fourier coefficient of $a$ at the basis element of $S$ corresponding to $c$. Now $S$ will be the whole of $A$ if and only if distinct vectors have distinct projections on $S$, [*i.e.* ]{}if and only if distinct vectors have distinct Gelfand transforms $\hat{a} \colon c
\mapsto c(a)$. Hence the Ambrose structure theorem holds when the Gelfand representation is injective, which holds if and only if the algebra is semisimple.
Necessity is easy to see from the form of a direct sum of one-dimensional algebras, as the lattice of ideals is a complete atomic boolean algebra, where the atoms are the generators of the algebras.
We shall restate the previous theorem in terms of axiom (H), so that we have a characterization that lends itself to categories other than ${\ensuremath{\mathbf{Hilb}}}$.
\[prop:semisimple\] A Frobenius algebra in ${{\ensuremath{\mathbf{Hilb}}}}$ satisfies (H) if and only if it is semisimple, and hence admits the structure theorem.
Semisimplicity of proper [H\*]{}-algebras follows from results in [@ambrose:hstar]. Conversely, $\bigoplus_I ({\ensuremath{\mathbb{C}}}, \mu_{{\ensuremath{\mathbb{C}}}})$ is easily seen to satisfy (H); we can define $x^*$ by taking conjugate coefficients in the given basis.
Categorical formulation
-----------------------
We can recast these results into a categorical form. Recall that there is a functor $\ell^2 \colon {\ensuremath{\mathbf{PInj}}} \to
{\ensuremath{\mathbf{Hilb}}}$ on the category of sets and partial injections [@barr:algebraicallycompact; @heunen:thesis]. It sends a set $X$ to the Hilbert space $\ell^2(X) = \{ \varphi \colon X \to
{\ensuremath{\mathbb{C}}} \mid \sum_{x \in X} |\varphi(x)|^2 < \infty \}$, which is the free Hilbert space on $X$ that is equipped with an orthonormal basis, [*i.e.* ]{}an [H\*]{}-algebra, in a sense we will now make precise. First, we make Frobenius algebras and [H\*]{}-algebras into categories. While other choices of morphisms can fruitfully be made [@heunen:copyables], the following one suits our current purposes.
\[def:morphisms\] Let ${\ensuremath{\mathbf{D}}}$ be a symmetric monoidal dagger category. We denote by ${\ensuremath{\mathbf{HStar}}}({\ensuremath{\mathbf{D}}})$ the category whose objects are [H\*]{}-algebras in ${\ensuremath{\mathbf{D}}}$, and by ${\ensuremath{\mathbf{Frob}}}({\ensuremath{\mathbf{D}}})$ the category whose objects are Frobenius algebras in ${\ensuremath{\mathbf{D}}}$. A morphism $(A,\delta) \to (A',\delta')$ in both categories is a morphism $f \colon A \to A'$ in ${\ensuremath{\mathbf{D}}}$ satisfying $(f {\ensuremath{\otimes}}f) {\circ}\delta = \delta' {\circ}f$ and $f^\dag {\circ}f = {\ensuremath{\mathrm{id}_{}}}$.
Every object in ${\ensuremath{\mathbf{PInj}}}$ carries a unique H\*-algebra structure, namely $\delta(a) = (a,a)$.
Let $\delta = (\xymatrix@1{A & \;D\;
\ar@{ >->}|-{\delta_1}[l] \ar@{ >->}|-{\delta_2}[r] & A \times A})$ be an object of ${\ensuremath{\mathbf{HStar}}}({\ensuremath{\mathbf{PInj}}})$. Because of (M), we may assume that $\delta_1 = {\ensuremath{\mathrm{id}_{}}}$. By (C), we find that $\delta_2$ is a tuple of some $d \colon A \to A$ with itself. It follows from (A) that $d = d {\circ}d$. Finally, since ${\ensuremath{\mathbf{PInj}}}$ is monoidally well-pointed, $\delta$ satisfies (F) by Lemma \[lem:HAWPimplyF\]. Writing out what (F) means gives $$\{((d(b),b),(b,d(b))) \mid b \in A \}
= \{((c,d(c)),(d(c),c)) \mid c \in A \}.$$ Hence for all $b \in A$, there is $c \in A$ with $b=d(c)$ and $d(b)=c$. Taking $b=d(a)$ we find that $c=a$, so that for all $a \in
A$ we have $d {\circ}d(a)=a$. Therefore $d = d {\circ}d = {\ensuremath{\mathrm{id}_{}}}$. We conclude that $\delta$ is the diagonal function $a \mapsto (a,a)$.
As a corollary one finds that an object in ${\ensuremath{\mathbf{PInj}}}$ with its unique H\*-algebra structure is unital if and only if it is a singleton set, which is another good argument against demanding (U).
If we drop the condition $f^\dag {\circ}f = {\ensuremath{\mathrm{id}_{}}}$ on morphisms in Definition \[def:morphisms\], the previous proposition can also be read as saying that the categories ${\ensuremath{\mathbf{HStar}}}({\ensuremath{\mathbf{PInj}}})$, ${\ensuremath{\mathbf{Frob}}}({\ensuremath{\mathbf{PInj}}})$, and ${\ensuremath{\mathbf{PInj}}}$ are isomorphic.
Since the Hilbert space $\ell^2(X)$ comes with a chosen basis induced by $X$, the $\ell^2$ construction is in fact a functor $\ell^2 \colon
{\ensuremath{\mathbf{HStar}}}({\ensuremath{\mathbf{PInj}}}) \to {\ensuremath{\mathbf{HStar}}}({\ensuremath{\mathbf{Hilb}}})$. Conversely, there is a functor $U$ in the other direction taking an [H\*]{}-algebra to the set of its copyables; this is functorial by [@ambrose:hstar Example 3]. These two functors are adjoints: $$\xymatrix@C+5ex{
{\ensuremath{\mathbf{HStar}}}({\ensuremath{\mathbf{PInj}}})
\ar@{}|-{\perp}[r] \ar@<1ex>^-{\ell^2}[r]
& {\ensuremath{\mathbf{HStar}}}({\ensuremath{\mathbf{Hilb}}}). \ar@<1ex>^-{U}[l]
}$$ The Ambrose structure theorem, Theorem \[thm:ambrose\], can now be restated as saying that this adjunction is in fact an equivalence.
Similarly, there is an adjunction between ${\ensuremath{\mathbf{Frob}}}({\ensuremath{\mathbf{PInj}}})$ and ${\ensuremath{\mathbf{Frob}}}({\ensuremath{\mathbf{Hilb}}})$, but it is not yet clear if this is an equivalence, too, [*i.e.* ]{}if ${\ensuremath{\mathbf{Frob}}}({\ensuremath{\mathbf{Hilb}}})$ and ${\ensuremath{\mathbf{HStar}}}({\ensuremath{\mathbf{Hilb}}})$ are equivalent categories. In fact, this question is the central issue of the rest of this paper, and will lead to the main open question in Section \[subsec:mainquestion\] to follow. In the meantime, we shall use the categorical formulation to give different characterizations of when Frobenius algebras in ${\ensuremath{\mathbf{Hilb}}}$ admit the structure theorem.
Further conditions {#subsec:furtherconditions}
------------------
There are in fact a number of conditions on Frobenius algebras in ${{\ensuremath{\mathbf{Hilb}}}}$ which are equivalent to admitting the structure theorem. This section gives two more.
A Frobenius algebra in ${\ensuremath{\mathbf{Hilb}}}$ is an H\*-algebra, and hence corresponds to an orthonormal basis, if and only if it is a directed colimit (in ${\ensuremath{\mathbf{Frob}}}({\ensuremath{\mathbf{Hilb}}})$) of unital Frobenius algebras.
Given an orthonormal basis $\{{{|} i\rangle}\}_{i \in I}$ for $A$, define $\delta \colon A \to A {\ensuremath{\otimes}}A$ by (continuous linear extension of) $\delta {{|} i\rangle} = {{|} ii\rangle}$. For finite subsets $F$ of $I$, define $\delta_F \colon \ell^2(F) \to \ell^2(F) {\ensuremath{\otimes}}\ell^2(F)$ by $\delta_F{{|} i\rangle} = {{|} ii\rangle}$. These are well-defined objects of ${\ensuremath{\mathbf{Frob}}}({\ensuremath{\mathbf{Hilb}}})$ by Theorem \[cpvth\]. Since $F$ is finite, every $\delta_F$ is a unital Frobenius algebra in ${\ensuremath{\mathbf{Hilb}}}$. Together they form a (directed) diagram in ${\ensuremath{\mathbf{Frob}}}({\ensuremath{\mathbf{Hilb}}})$ by inclusions $i_{F \subseteq F'} \colon
\ell^2(F) \hookrightarrow \ell^2(F')$ if $F \subseteq F'$; the latter are well-defined morphisms since $\delta_{F'} {\circ}i_{F \subseteq F'} {{|} i\rangle} = {{|} ii\rangle} =
(i_{F \subseteq F'} {\ensuremath{\otimes}}i_{F \subseteq F'}) {\circ}\delta_F {{|} i\rangle}$. Finally, we verify that $\delta$ is the colimit of this diagram. The colimiting cocone is given by the inclusions $i_F \colon
\ell^2(F) \hookrightarrow A$; these are morphisms $i_F \colon
\delta_F \to \delta$ in ${\ensuremath{\mathbf{Frob}}}({\ensuremath{\mathbf{Hilb}}})$ since $\delta {\circ}i_F =
(i_F {\ensuremath{\otimes}}i_F) {\circ}\delta_F$, that are easily seen to form a cocone. Now, if $f_F \colon \delta_F \to (A',\delta')$ form another cocone, define $m \colon X \to X'$ by (continuous linear extension of) $m {{|} i\rangle} = f_{\ell^2(\{{{|} i\rangle}\})} {{|} i\rangle}$. Then $m {\circ}i_F {{|} i\rangle} = f_{\ell^2(\{{{|} i\rangle}\})} {{|} i\rangle} = f_F {{|} i\rangle}$ for $i
\in F$, so that indeed $m {\circ}i_F = f_F$. Moreover, $m$ is the unique such morphism. Thus $\delta$ is indeed a colimit of the $\delta_F$.
Conversely, suppose $(A,\delta)$ is a colimit of some diagram $d \colon {\ensuremath{\mathbf{I}}} \to {\ensuremath{\mathbf{Frob}}}({\ensuremath{\mathbf{Hilb}}})$. We will show that the nonzero copyables form an orthonormal basis for $A$. By Lemma \[lem:independent\] and Proposition \[prop:orthogonal\], it suffices to prove that they span a dense subspace of $A$. Let $a \in A$ be given. Since the colimiting cocone morphisms $c_i \colon A_i \to A$ are jointly epic, the union of their images is dense in $A$, and therefore $a$ can be written as a limit of $c_i(a_i)$ with $a_i \in A_i$ for some of the $i \in {\ensuremath{\mathbf{I}}}$. These $a_i$, in turn, can be written as linear combinations of elements of copyables of $A_i$ by Theorem \[cpvth\]. Now, $c_i$ maps copyables into copyables, and so we have written $a$ as a limit of linear combinations of copyables of $A$. Hence the copyables of $A$ spans a dense subspace of $A$, and therefore form an orthonormal basis.
Finally, we verify that these two constructions are mutually inverse. Starting with a $\delta$, one obtains $E=\{e \mid \delta(e)
= e {\ensuremath{\otimes}}e\}$, and then $\delta' \colon A \to A {\ensuremath{\otimes}}A$ by (continuous linear extension of) $\delta'(e)=e {\ensuremath{\otimes}}e$ for $e \in E$. The definition of $E$ then gives $\delta' = \delta$.
Conversely, starting with an orthonormal basis $\{{{|} i\rangle}\}_{i \in I}$, one obtains a map by (continuous linear extension of) $\delta {{|} i\rangle} = {{|} ii\rangle}$, and then it follows that $E = \{a \in
A \mid \delta(a) = a {\ensuremath{\otimes}}a\}$. It is trivial that $\{{{|} i\rangle} \mid i
\in I\} \subseteq E$. Moreover, we know that $E$ is linearly independent by \[lem:independent\]. Since it contains a basis, it must therefore be a basis itself. Hence indeed $E = \{ {{|} i\rangle}
\mid i \in I\}$.
For *separable* Hilbert spaces, there is also a characterization in terms of approximate units as follows.
\[thm:approximateunit\] A Frobenius algebra on a separable Hilbert space in ${\ensuremath{\mathbf{Hilb}}}$ is an [H\*]{}-algebra, and hence corresponds to an orthonormal basis, if and only if there is a sequence $e_n$ such that $e_n a$ converges to $a$ for all $a$, and $({\ensuremath{\mathrm{id}_{}}}{\ensuremath{\otimes}}a^\dag) {\circ}\delta(e_n)$ converges.
Writing $a^*_{n} = ({\ensuremath{\mathrm{id}_{}}}{\ensuremath{\otimes}}a^\dag) {\circ}\delta(e_n)$, by assumption $a^* = \lim_{n \to \infty} a^*_n$ is well-defined. Since morphisms in ${\ensuremath{\mathbf{Hilb}}}$ are continuous functions and composition preserves continuity, (H) holds by the following argument. $${
\begin{tikzpicture}[] \PandocStartInclude{tikz/H-lhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; = \;
\lim_{n \to \infty}\left( {
\begin{tikzpicture}[] \PandocStartInclude{tikz/approxunit-lhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}} \right)
\; \stackrel{\text{(F)}}{=} \;
\lim_{n \to \infty}\left( {
\begin{tikzpicture}[] \PandocStartInclude{tikz/approxunit-rhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}} \right)
\; = \;
{
\begin{tikzpicture}[] \PandocStartInclude{tikz/H-rhs.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ Hence approximate units imply (H). Conversely, using the Ambrose structure theorem, Theorem \[thm:ambrose\], we can always define $e_n$ to be the sum of the first $n$ copyables.
Summarizing, we have the following result.
\[thm:whenFimpliesH\] For a Frobenius algebra in ${\ensuremath{\mathbf{Hilb}}}$, the following are equivalent:
1. it is induced by an orthonormal basis;
2. it admits the structure theorem;
3. it is semisimple;
4. it satisfies axiom (H);
5. it is a directed colimit (with respect to isometric homomorphisms) of finite-dimensional unital Frobenius algebras.
Moreover, if the Hilbert space is separable, these are equivalent to:
1. it has a suitable form of approximate identity.
We see that the finite-dimensional result follows immediately from our general result and Lemma \[lem:cpv\], which shows that the algebra is C\* and hence semisimple. In fact, the influential thesis [@abrams:thesis] (see also [@abrams:euler; @kock:frobenius]) already observes explicitly (and in much wider generality) that:
- If (M) holds, a unital Frobenius algebra is semisimple [@abrams:thesis Theorem 2.3.3].
- A commutative semisimple unital Frobenius algebra is a direct sum of fields [@abrams:thesis Theorem 2.2.5].
Thus the only additional ingredient required to obtain Theorem \[cpvth\] is the elementary Proposition \[prop:orthogonal\].
The main question {#subsec:mainquestion}
-----------------
The main remaining question in our quest for a suitable notion of algebra to characterize orthonormal bases in arbitrary dimension is the following.
We can ask this question for the central case of ${{\ensuremath{\mathbf{Hilb}}}}$, and for monoidal dagger categories in general.
If the answer is positive, then nonunital Frobenius algebras give us the right notion of observable to use in categorical quantum mechanics. If it is negative, we may consider adopting (H) as the right axiom instead of (F).
At present, these questions remain open, both for ${{\ensuremath{\mathbf{Hilb}}}}$ and for the general case. However, we have been able to achieve positive results for a large family of categories; these will be described in the following section. We shall conclude this section by further narrowing down the question in the category ${\ensuremath{\mathbf{Hilb}}}$.
Recall that the *Jacobson radical* of a commutative ring is the intersection of all its maximal regular ideals, and that a ring is called *radical* when it equals its Jacobson radical.
Frobenius algebras $A$ in ${\ensuremath{\mathbf{Hilb}}}$ decompose as a direct sum $$A \cong S \oplus R$$ of (co)algebras, where $S$ is an [H\*]{}-algebra and $R$ is a radical algebra.
Let $a$ be a copyable element of a Frobenius algebra $A$ in ${\ensuremath{\mathbf{Hilb}}}$. Consider the embedding into $A$ of $a$ as a one-dimensional algebra. This embedding is a kernel, since it is isometric and its domain is finite-dimensional. Observe that this embedding is an algebra homomorphism as well as a coalgebra homomorphism, because copyables are idempotents by (M). Now it follows from [@heunen:copyables Lemma 19] that also the orthogonal complement of the embedding is both an algebra homomorphism and a coalgebra homomorphism. Finally, Frobenius algebra structure restricts along such embeddings by [@heunen:copyables Proposition 9].
We can apply this to the embedding of the closed span of all copyables of $A$, and conclude that $A$ decomposes (as a (co)algebra) into a direct sum of its copyables and the orthogonal subspace. By definition, the former summand is semisimple, and is hence a [H\*]{}-algebra by Proposition \[prop:semisimple\]. The latter summand by construction has no copyables and hence no characters, and is therefore radical.
This shows how the Jacobson radical of a Frobenius algebra sits inside it in a very simple way. Indeed, we are left with not just a nonsemisimple algebra, but a radical one, which is the opposite of a semisimple algebra—an algebra is semisimple precisely when its Jacobson radical is zero. Therefore, in the category ${\ensuremath{\mathbf{Hilb}}}$, our main remaining question above reduces to finding out whether $R$ must be zero, as follows.
Although there is an extensive literature about commutative radical Banach algebras, including a complete classification that in fact ties in with approximate units [@esterle:radical], this question seems to be rather difficult.
[H\*]{}-algebras in categories of relations and positive matrices {#sec:hstarinrel}
=================================================================
We have been able to give a complete analysis of nonunital Frobenius algebras in several (related) cases, including:
- categories of relations, and locally bifinite relations, valued in cancellative quantales;
- nonnegative matrices with $\ell^2$-summable rows and columns.
The common feature of these cases can be characterized as *the absence of destructive interference*.
The main result we obtain is as follows.
Nonunital Frobenius algebras in all these categories decompose as direct sums of abelian groups, and satisfy (H).
The remainder of this section is devoted to the proof of this theorem. Our plan is as follows. First, we shall prove the result for ${{\ensuremath{\mathbf{Rel}}}}$, the category of sets and relations. In this case, our main question is already answered directly by Proposition \[prop:cptimpliesunital\] and Theorem \[thm:whenFimpliesH\]. Moreover, the result in this case has appeared in [@pavlovic:frobeniusinrel]. However, our proof is quite different, and in particular makes no use of units. This means that it can be carried over to the other situations mentioned above.
Frobenius algebras in ${{\ensuremath{\mathbf{Rel}}}}$ and ${{\ensuremath{\mathbf{lbfRel}}}}$ {#frobrelsec}
--------------------------------------------------------------------------------------------
We assume given a set $A$, and a Frobenius algebra structure on it given by a relation $\Delta \subseteq A \times (A \times A)$. We shall write $\nabla$ for $\Delta^\dag$.
Define $x \sim y$ if and only if $(x,y) \nabla z$ for some z. By (M), the relation $\nabla$ is single-valued and surjective. Therefore, we may also use multiplicative notation $xy$ (suppressing the $\nabla$), and write $x \sim y$ to mean that $xy$ is defined.
\[lem:reflexivity\] The relation $\sim$ is reflexive.
Let $a \in A$. By (M), we have $a=a_1a_2$ for some $a_1,a_2 \in A$. Then $(a_2,a) ({\ensuremath{\mathrm{id}_{}}}{\ensuremath{\otimes}}\Delta) (a_2,a_1,a_2)$ and $(a_2,a_1,a_2)
(\nabla {\ensuremath{\otimes}}{\ensuremath{\mathrm{id}_{}}}) (a,a_2)$ by (C), so by (F) we have $(a_2,a) \Delta{\circ}\nabla (a,a_2)$, so that $aa_2$ is defined. Diagrammatically, we annotate the lines with elements to show they are related by that morphism. $${
\begin{tikzpicture}[font=\small] \PandocStartInclude{tikz/reflexive-1.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; = \;
{
\begin{tikzpicture}[font=\small] \PandocStartInclude{tikz/reflexive-2.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ Also $(a,a) (\Delta {\ensuremath{\otimes}}{\ensuremath{\mathrm{id}_{}}}) (a_1,a_2,a)$ and $(a_1,a_2,a) ({\ensuremath{\mathrm{id}_{}}}{\ensuremath{\otimes}}(\Delta {\circ}\nabla)) (a_1,a,a_2)$, so by (F) we have $(a,a) ({\ensuremath{\mathrm{id}_{}}}{\ensuremath{\otimes}}\Delta) {\circ}\Delta {\circ}\nabla (a_1,a,a_2)$, so that $a^2$ is defined. $${
\begin{tikzpicture}[font=\small] \PandocStartInclude{tikz/reflexive-3.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; = \;
{
\begin{tikzpicture}[font=\small] \PandocStartInclude{tikz/reflexive-4.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ That is, $a \sim a$.
\[lem:transitivity\] The relation $\sim$ is transitive.
Suppose that $a \sim b$ and $b \sim c$. Then $d=ab$ is defined. By Lemma \[lem:reflexivity\], then $(ab)d=d^2$ is defined. Hence by (A), also $a(bd)$ is defined. Applying (F) now yields $\bar{b}$ such that $a=d\bar{b}$. $${
\begin{tikzpicture}[font=\small] \PandocStartInclude{tikz/transitive-1.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; = \;
{
\begin{tikzpicture}[font=\small] \PandocStartInclude{tikz/transitive-2.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ It now follows from (C) that $a=d\bar{b}=a\bar{b}b$; in particular $a\bar{b}$ is defined. But then also $ac=(a\bar{b}b)c=(a\bar{b})(bc)$ is seen to be defined by the assumption $b \sim c$ and another application of (F). $${
\begin{tikzpicture}[font=\small] \PandocStartInclude{tikz/transitive-3.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; = \;
{
\begin{tikzpicture}[font=\small] \PandocStartInclude{tikz/transitive-4.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ Hence $a \sim c$.
\[prop:disjointunion\] The Frobenius algebra $A$ is a disjoint union of totally defined commutative semigroups, each satisfying (F).
By the previous two lemmas and (C), the relation $\sim$ is a equivalence relation. Hence $A$ is a disjoint union of the equivalence classes under $\sim$. By definition of $\sim$, the multiplication $\nabla$ is totally defined on these equivalence classes. Moreover, they inherit the properties (M), (C), (A) and (F) from $A$.
\[lem:huntington\] [@huntington:group] A semigroup $S$ is a group if and only if $aS=S=Sa$ for all $a \in S$.
The condition $aS=S$ means $\forall b\exists c[b=ac]$. If $S$ is a group, this is obviously fulfilled by $c=a^{-1}b$. For the converse, fix $a \in S$. Applying the condition with $b=a$ yields $c$ such that $a=ac$. Define $e=c$, and let $x \in S$. Then applying the condition with $b=x$ gives $c$ with $x=ac$. Hence $ex=eac=ac=x$. Thus $S$ is a monoid with (global) unit $e$. Applying the condition once more, with $a=x$ and $b=e$ yields $x^{-1}$ with $xx^{-1}=e$.
\[lem:group\] $A$ is a disjoint union of commutative groups.
Let $A'$ be one of the equivalence classes of $A$, and consider $a,b \in A'$. This means that $a \sim b$. As in the proof of Lemma \[lem:transitivity\], there is a $\bar{b}$ such that $a=\bar{b}ba$. Putting $c=\bar{b}a$ thus gives $\forall a, b \in A'
\exists c \in A'[a=cb]$. In other words, $aA'=A'$ and similarly $A'=A'a$ for all $a \in A$. Hence $A'$ is a (commutative) group by Lemma \[lem:huntington\].
The following theorem already follows from Proposition \[prop:cptimpliesunital\] and Theorem \[thm:whenFimpliesH\], but now we have a direct proof that also carries over to the theorem after it, which does not follow from the earlier results.
\[thm:characterisationFinRel\] In ${{\ensuremath{\mathbf{Rel}}}}$, Frobenius algebras satisfy (H), and the conditions (F) and (H) are equivalent in the presence of the other axioms for Frobenius algebras.
This follows directly from Lemma \[lem:group\], since we can define $a^*=a^{-1}$, where $a^{-1}$ is the inverse in the disjoint summand containing $a$. More precisely, a point of $A$ in ${{\ensuremath{\mathbf{Rel}}}}$ will be a subset of $A$, and we apply the definition $a^*=a^{-1}$ pointwise to this subset. This assignment is easily seen to satisfy (H).
In ${{\ensuremath{\mathbf{lbfRel}}}}$, Frobenius algebras are disjoint unions of abelian groups and hence satisfy (H), and the conditions (F) and (H) are equivalent in the presence of the other axioms for Frobenius algebras.
The proof above made no use of units, and is equally valid in ${{\ensuremath{\mathbf{lbfRel}}}}$.
Quantale-valued relations
-------------------------
We shall now consider categories of the form ${{\ensuremath{\mathbf{Rel}}}}(Q)$, where $Q$ is a commutative, cancellative quantale. Recall that a commutative quantale [@rosenthal:quantales] is a structure $(Q, {\cdot}, 1, {\leq})$, where $(Q, {\cdot}, 1)$ is a commutative monoid, and $(Q, {\leq})$ is a partial order which is a complete lattice, [*i.e.* ]{}it has suprema of arbitrary subsets. In particular, the supremum of the empty set is the least element of the poset, written $0$. The multiplication is required to distribute over arbitrary joins, [*i.e.* ]{} $$x \cdot (\bigvee_{i \in I} y_i)
= \bigvee_{i \in I} x \cdot y_i,
\qquad
(\bigvee_{i \in I} x_i) \cdot y
= \bigvee_{i \in I} x_i \cdot y.$$ The quantale is called *cancellative* if $$x \cdot y = x \cdot z
\; \Rightarrow \;
x = 0 \vee y = z.$$ An example is given by the extended nonnegative reals $[0, \infty]$ with the usual ordering, and multiplication as the monoid operation. Note that the only nontrivial example when the monoid operation is idempotent, [*i.e.* ]{}when the quantale is a locale, is the two-element boolean algebra ${\ensuremath{\mathbf{2}}} = \{ 0, 1 \}$, since in the idempotent case $x \cdot 1
= x \cdot x$ for all $x$. We write ${{\ensuremath{\mathbf{canQuant}}}}$ for the category of cancellative quantales which are nontrivial, [*i.e.* ]{}in which $0 \neq 1$.
\[qtermprop\] The two element boolean algebra is terminal in ${{\ensuremath{\mathbf{canQuant}}}}$.
The unique homomorphism $h \colon Q \to {\ensuremath{\mathbf{2}}}$ sends $0$ to itself, and everything else to 1. Preservation of sups holds trivially, and cancellativity implies that multiplication is preserved.
The category ${{\ensuremath{\mathbf{Rel}}}}(Q)$ has sets as objects; morphisms $R \colon X
{\ensuremath{\xymatrix@1@C-2ex{\ar|(.4)@{|}[r]&}}}Y$ are $Q$-valued matrices, [*i.e.* ]{}functions $X \times Y \to
Q$. Composition is relational composition evaluated in $Q$, [*i.e.* ]{}if $R
\colon X {\ensuremath{\xymatrix@1@C-2ex{\ar|(.4)@{|}[r]&}}}Y$ and $S \colon Y {\ensuremath{\xymatrix@1@C-2ex{\ar|(.4)@{|}[r]&}}}Z$, then $$S {\circ}R(x, z) = \bigvee_{y \in Y} R(x, y) \cdot S(y, z).$$ It is easily verified that this yields a category, with identities given by diagonal matrices; that it has a monoidal structure induced by cartesian product; and that is has a dagger given by matrix transpose, [*i.e.* ]{}relational converse. Thus ${{\ensuremath{\mathbf{Rel}}}}(Q)$ is a symmetric monoidal dagger category, and the notion of Frobenius algebra makes sense in it. Note that ${{\ensuremath{\mathbf{Rel}}}}({\ensuremath{\mathbf{2}}})$ is just ${{\ensuremath{\mathbf{Rel}}}}$. A homomorphism of quantales $h \colon Q \to R$ induces a (strong) monoidal dagger functor $h^* \colon {{\ensuremath{\mathbf{Rel}}}}(Q) \to {{\ensuremath{\mathbf{Rel}}}}(R)$, which transports Frobenius algebras in ${{\ensuremath{\mathbf{Rel}}}}(Q)$ to Frobenius algebras in ${{\ensuremath{\mathbf{Rel}}}}(R)$. In particular, by Proposition \[qtermprop\], a Frobenius algebra $\Delta \colon A {\ensuremath{\xymatrix@1@C-2ex{\ar|(.4)@{|}[r]&}}}A \times A$ in ${{\ensuremath{\mathbf{Rel}}}}(Q)$ has a reduct $h^*
\Delta \colon A {\ensuremath{\xymatrix@1@C-2ex{\ar|(.4)@{|}[r]&}}}A \times A$ in ${{\ensuremath{\mathbf{Rel}}}}$. Hence Theorems \[lem:group\] and \[thm:characterisationFinRel\] apply to this reduct. The remaining degree of freedom in the Frobenius algebra in ${{\ensuremath{\mathbf{Rel}}}}(Q)$ is which elements of $Q$ can be assigned to the elements of the matrix.
Suppose that we have a Frobenius algebra $\Delta \colon A {\ensuremath{\xymatrix@1@C-2ex{\ar|(.4)@{|}[r]&}}}A
\times A$ in ${{\ensuremath{\mathbf{Rel}}}}(Q)$. We write $M \colon (A \times A) \times A \to Q$ for the matrix function corresponding to the converse of $\Delta$, and we write $M(a, b, c)$ rather than $M((a, b), c)$.
Because the unique homomorphism $Q \to {\ensuremath{\mathbf{2}}}$ reflects 0, an entry $M(a, b, c)$ is nonzero if and only if the corresponding relation $(a, b) (h^* \nabla) c$ holds. Applying Theorem \[lem:group\], this immediately implies that for each $a, b \in A$, there is exactly one $c \in A$ such that $M(a, b, c) \neq 0$.
We can use this observation to apply similar diagrams to those used in our proofs for ${{\ensuremath{\mathbf{Rel}}}}$ to obtain constraints on the values taken by the matrix in $Q$.
\[quantvalprop\] With notation as above:
1. If $e$ is an identity element in one of the disjoint summands, then for all $a,b$ in that disjoint summand we have $M(a, e, a) = M(b, e, b)$. We write $q_e$ for this common value.
2. For all $a,b \in A$, we have $M(a, b, ab)^2 = 1$.
For (a), consider the diagram $${
\begin{tikzpicture}[font=\small] \PandocStartInclude{tikz/unitquantale-1.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; = \;
{
\begin{tikzpicture}[font=\small] \PandocStartInclude{tikz/unitquantale-2.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ This implies the equation $$M(a, b, ab) \cdot M(ab, e, ab)
= M(a, b, ab) \cdot M(b, e, b)$$ and hence, by cancellativity, $M(ab, e, ab) = M(b, e, b)$. Hence for any $c$, taking $b = a^{-1}c$, $M(c, e, c) = M(b, e, b)$.
For (b), consider the diagram $${
\begin{tikzpicture}[font=\small] \PandocStartInclude{tikz/unitquantale-3.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}
\; = \;
{
\begin{tikzpicture}[font=\small] \PandocStartInclude{tikz/unitquantale-4.tikz}\PandocEndInclude{input}{24}{23} \end{tikzpicture}}$$ This implies the equation $M(a, b, ab)^2 = q_e^2$. Now applying (M), for each $c$ we obtain that $$\bigvee \{ M(a, b, ab)^2 \mid ab = c \} = 1.$$ As all the terms in this supremum are the same, $M(a, b, ab)^2 = 1$.
Thus if $q^2 = 1$ implies $q = 1$ in $Q$, the matrix $M$ is in fact valued in ${\ensuremath{\mathbf{2}}}$. Otherwise, we can choose square roots of unity for the entries.
Let $Q$ be a cancellative quantale. Suppose that $q^2 = 1$ implies $q = 1$ in $Q$. Then every Frobenius algebra in ${{\ensuremath{\mathbf{Rel}}}}(Q)$ satisfies (H).
Positive $\ell^2$ matrices
--------------------------
We now consider the case of matrices in ${{{\ensuremath{\mathbf{Mat}}}}_{\ell^2}({\ensuremath{\mathbb{C}}})}$ valued in the non-negative reals. These form a monoidal dagger subcategory of ${{{\ensuremath{\mathbf{Mat}}}}_{\ell^2}({\ensuremath{\mathbb{C}}})}$, which we denote by ${{{{\ensuremath{\mathbf{Mat}}}}_{\ell^2}({{\ensuremath{\mathbb{R}}}^{+}})}}$. Note that the semiring $({{\ensuremath{\mathbb{R}}}^{+}}, +, 0, {\times}, 1)$ has a unique 0-reflecting semiring homomorphism to ${\ensuremath{\mathbf{2}}}$. Hence a Frobenius algebra in ${{{{\ensuremath{\mathbf{Mat}}}}_{\ell^2}({{\ensuremath{\mathbb{R}}}^{+}})}}$ has a reduct to one in ${{\ensuremath{\mathbf{Rel}}}}$ via this homomorphism. Just as before, we can apply Theorem \[lem:group\] to this reduct.
We have the following analogue to Proposition \[quantvalprop\], where $M \colon (A \times A) \times A \to {{\ensuremath{\mathbb{R}}}^{+}}$ is the matrix realizing the Frobenius algebra structure.
\[Mconstprop\] The function $M$ is constant on each disjoint summand of $A$.
We can use the same reasoning as in Proposition \[quantvalprop\](a) to show that, if $e$ is an identity element in one of the disjoint summands, then for all $a$, $b$ in that disjoint summand, $M(a, e, a) = M(b, e, b)$. We write $r_e$ for this common value.
Using the same reasoning as in Proposition \[quantvalprop\](b) one finds $M(a, b, ab)^2 = r_e^2$. Since we are in ${{\ensuremath{\mathbb{R}}}^{+}}$, this implies $M(a, b, ab) = r_e$, so that $M$ is constant on each disjoint summand.
Each disjoint summand is finite, and the common value of $M$ on that summand is $1/\sqrt{d}$, where $d$ is the cardinality of the summand.
Applying (M), for each $c$ in the summand we obtain that $$\sum_{ab = c} M(a, b, c)^2 = 1.$$ Since the summand is a group, for each $c$ and $a$ there is a unique $b$ such that $ab = c$. Moreover, by Proposition \[Mconstprop\], all the terms in this sum are equal. Thus the sum must be finite, with the number of terms $d$ the cardinality of the summand. We can therefore rewrite the equation as $d r_e^2 = 1$, and hence $r_e = 1/\sqrt{d}$.
Every Frobenius algebra in ${{{{\ensuremath{\mathbf{Mat}}}}_{\ell^2}({{\ensuremath{\mathbb{R}}}^{+}})}}$ satisfies (H).
We make the same pointwise assignment $x^* = x^{-1}$ on the elements of $A$ as in the proof of Theorem \[thm:characterisationFinRel\], with the weight $1/\sqrt{d}$ determined by the summand.
In the case when the matrix represents a bounded linear map in ${{\ensuremath{\mathbf{Hilb}}}}$, we can apply Theorem \[thm:whenFimpliesH\], and obtain the following.
If a Frobenius algebra in ${{\ensuremath{\mathbf{Hilb}}}}$ can be represented by a non-negative real matrix, then it corresponds to a direct sum of one-dimensional algebras, and hence to an orthonormal basis.
Conversely, if a Frobenius algebra in ${\ensuremath{\mathbf{Hilb}}}$ satisfies (H), it is induced by an orthonormal basis, and hence has a matrix representation with nonnegative entries. Therefore we have found another equivalent characterization of when (F) implies (H) in ${\ensuremath{\mathbf{Hilb}}}$ to add to our list in Theorem \[thm:whenFimpliesH\].
A Frobenius algebra $A$ in ${\ensuremath{\mathbf{Hilb}}}$ satisfies (H) if and only if there is a basis of $A$ such that the matrix of the comultiplication has nonnegative entries when represented on that basis.
Discussion
----------
How different is the situation with Frobenius algebras in these matrix categories from ${{\ensuremath{\mathbf{Hilb}}}}$? In fact, it is not as different as it might at first appear.
- The category ${{\ensuremath{\mathbf{Hilb}}}}$ is equivalent to a full subcategory of the dagger monoidal category of complex matrices with $\ell^2$-summable rows and columns. The ‘only’ assumption needed but not satisfied is positivity.
- The result ‘looks’ different, but beware. Consider *group rings* (or algebras) over the complex numbers, for finite abelian groups. They can easily be set up to fulfil all our axioms, including (U), so that Theorem \[cpvth\] applies, and they decompose as direct sums of one-dimensional algebras. But the isomorphism which gives this decomposition may be quite non-obvious.[^9] Note that copyable elements are idempotents, so the only group element which is copyable is the identity.
- This decomposition result indeed shows that group rings over the complex numbers are very weak invariants of the groups. The group rings of two finite abelian groups will be isomorphic if the groups have the same order [@strichartz:groupalgebras]!
- However, this is highly sensitive to which field we are over. Group algebras over the *rationals* are isomorphism invariants of groups [@ingelstam:realhstar].
Outlook {#sec:outlook}
=======
We are still investigating our main question, of whether (F) implies (H), in ${{\ensuremath{\mathbf{Hilb}}}}$ and elsewhere.
Beyond this, we see the following main lines for continuing a development of categorical quantum mechanics applicable to infinite-dimensional situations.
- We are now able to consider observables with infinite discrete spectra. Beyond this lie continuous observables and projection-valued measures; it remains to be seen how these can be analyzed in the setting of categorical quantum mechanics.
- Complementary observables should be studied in this setting. The bialgebra approach studied in [@coeckeduncan:observables] is based on axiomatizing *mutually unbiased bases*, and does not extend directly to the infinite-dimensional case. However, complementary observables are studied from a much more general perspective in works such as [@buschetal:operational], and this should provide a good basis for suitable categorical axiomatizations.
- This leads on to another point. There may be other means, within the setting of categorical quantum mechanics, of representing observables, measurements and complementarity, which may be more flexible than the Frobenius algebra approach, and in a sense more natural, since tensor product structure is not inherent in the basic notion of measurement. Methodologically, one should beware of concluding over-hastily that a particular approach is canonical, simply on the grounds that it captures the standard notion in finite-dimensional Hilbert spaces. There may be several ways of doing this, and some more definitive characterization would be desirable.
- A related investigation to the present one is the work on nuclear and traced ideals in [@abramskyetal:nuclear]. It seems likely that some combination of the ideas developed there, and those we have studied in this paper, will prove fruitful.
### Acknowledgement {#acknowledgement .unnumbered}
We thank Rick Blute for stimulating the early stages of the research that led to this article.
[^1]: Supported by an EPSRC Senior Fellowship and by ONR.
[^2]: Supported by the Netherlands Organisation for Scientific Research (NWO).
[^3]: In the literature the unital version is more specifically termed a special commutative dagger Frobenius algebra (sometimes also called a separable algebra, or a Q-system). As we will only be concerned with these kinds of Frobenius algebras, we prefer to keep terminology simple and dispense with the adjectives.
[^4]: As we are in a commutative context, there is no need to distinguish between left and right regular representations.
[^5]: The notion termed 2-[H\*]{}-algebra in [@baez:twohilbertspaces] was inspired by Ambrose’s notion of [H\*]{}-algebra. The former could be seen as a categorification of the latter; the two notions should not be confused.
[^6]: Notice that neither axiom (H) nor Ambrose’s definition of H\*-algebra requires the operation $a \mapsto a^*$ to be continuous. However, in the setting of ${\ensuremath{\mathbf{Hilb}}}$, continuity follows automatically from axiom (H) [@ambrose:hstar Theorem 2.3].
[^7]: Hence Proposition \[prop:hstaralgebra\] and Theorem \[thm:ambrose\] can be altered to show that a monoid in ${\ensuremath{\mathbf{Hilb}}}$ satisfying properness, (A), (C), and (H) (but not necessarily (M)!), corresponds to an *orthogonal* basis. This may have consequences for attempts to classify multipartite entanglement according to various Frobenius structures [@coeckekissinger:multipartite]. Compare also the second entry in the table on page 11 of [@coeckepavlovicvicary:bases]: in finite dimension, $\delta$ is monic by (U), but in infinite dimension, one has to explicitly postulate $\delta$ to be monic to prevent [*e.g.* ]{}the trivial algebra $\delta(a)=0$ and obtain a correspondence with orthogonal bases. \[footnote:orthogonal\]
[^8]: Copyable elements are also called *primitive* in the context of C\*-bigebras [@hofmann:duality], and *grouplike* in the study of Hopf algebras [@sweedler:hopfalgebra; @kassel:quantumgroups].
[^9]: It would be interesting, for example, to know the computational complexity of determining this isomorphism, given a presentation of the group. As far as we know, this question has not been studied.
|
{
"pile_set_name": "ArXiv"
}
|
Introduction
============
The advances made in adaptive optics (AO) over the last decade have enabled many ground-based optical astronomical projects that were never before possible[@beck]. Most of these studies have been performed with natural guide star AO systems (NGSAO), which require a bright star near to the desired target. While there are many examples of strikingly successful studies done with NGSAO, the fraction of the sky accessible to them is rather small. In order to access the whole sky, one requires the ability to sample the atmospheric distortion anywhere, not just toward the direction of a bright star. The telluric sodium layer is a layer of sodium atoms that is typically present at an altitude of $\sim 90$ km over an extent of $\sim 15$ km. By illuminating the layer with a laser tuned to the wavelength of the $D_2$ transition ($589$ nm), one may excite emission from these sodium atoms. The resulting bright spot acts as an artificial guide star, and enables the operation of laser guide star AO (LGSAO) systems.
The development of a LGSAO system is of particular importance to the Thirty-Meter telescope (TMT). In order to take full advantage of the potential spatial resolution of the large aperture, it is imperative for many of the TMT’s instruments to be fed by the proposed LGSAO system, the Narrow Field InfraRed Adaptive Optic System (NFIRAOS). The development of NFIRAOS presents many technical challenges. As discussed in O’Sullivan et al. (2000)[@osull], one of these challenges—the refocusing of the telescope on the moving laser guide star spot—is due to the variability of the sodium layer. Though the physical mechanism is not entirely understood, atmospheric gravity wave propagating upward in the mesopause region is thought to be the most likely explanation. A typical Shack-Hartmant wavefront sensor will determine the centroid of a illuminated spot on a CCD, and will therefore respond to the instantaneous mean height of the sodium layer. A change in the mean height, $\Delta h$, of the sodium layer produces a RMS wavefront error, averaged over the aperture, $$\sigma=\frac{1}{8\sqrt{3}}\left(\frac{D}{h}\right)^2\Delta h,$$ where $D$ is the diameter of the telescope aperture, and $h$ is the instantaneous density-weighted mean height of the sodium layer [@herr]. Note that $\sigma$ is proportional to the square of the diameter of the telescope. Thus, for the TMT, we expect the effect to be an order of magnitude larger than for the largest optical telescopes presently in operation. For the TMT, $\sigma = 8$ nm for every meter change of the mean height.
While an AO system can use laser and natural guide stars to correct for focus error caused by the atmosphere and variations in the height of the sodium layer, it can not do so instantaneously. Therefore, the correction applied by the AO system will be imperfect, and a residual wavefront error (RWFE) will remain. While the mean height of the sodium layer is known to be variable on the order of several kilometers over timescales of minutes, LGSAO systems sample the atmosphere many times per second. Therefore, it is necessary to know how the mean height of the sodium layer will vary on millisecond timescales in order to estimate the RWFE. In order to assess the variability of the sodium layer on short time scales we examined data taken by the Colorado State University (CSU) Light Detection and Ranging (lidar) experiment over 89 nights in 2003. We determined the typical RWFE expected by multiplying the mean height power spectrum of the sodium layer by a model rejection transfer function determined specifically for the TMT.
Technique
=========
Reduction and extraction
------------------------
The CSU lidar samples the sodium layer by projecting laser pulses upwards, cyclically tuned to three preselected frequencies within the Doppler-broadened bandwidth of the sodium D$_2$ transition at 589 nm. The intensity of the return signal in three channels (one for each frequency), binned as a function of distance from the telescope can be processed for measuring atmospheric temperatures and horizontal wind between 80 and 105 km[@she]. The experiment employs two telescopes that can be independently pointed, and can therefore measure the Na-layer in two directions simultaneously, typically zenith, due north at an altitude of $60^\circ$, or due east at an altitude of $60^\circ$. In order to increase the signal-to-background ratio, the signal is integrated for a duration of $120$ seconds. Due to readout time, measurements are typically spaced on $125$ second intervals, resulting in a Nyquist frequency of $4$ mHz.
The signal-to-background ratio was calculated in the following manner: The atmosphere was divided into three regimes. The signal regime extended from $75$ km to $105$ km. Although the sodium layer only extends over approximately half this altitude range at any one time, the large range was necessary to ensure that all the signal was contained in the range regardless of the mean height. The two background regimes extended from $60$ to $72$ km and $108$ to $120$ km respectively. By integrating over the aforementioned regimes a total signal, $S_{total}$, and total background, $B_{total}$, were calculated. The requirements for acceptable data were $S_{total} > 10^4 $ photons and $B_{total} < 10^3$ photons.
The raw files were processed to yield time, mean height (corrected for zenith angle of the laser, and the altitude of the CSU lidar experiment), and a flag indicating if the signal-to-background ratio of an individual measurement was acceptable. The time series for December $24^{th}$ is shown in figure \[ts.fig\]. Channel \#1 and channel \#2 are shown as solid and dotted lines respectively.
Transformation and Fitting
--------------------------
We computed the mean height power spectrum for each night individually. Over the observed range of frequencies, from $\mu$Hz to mHz, the transformed data were well fit by a power law. Many of the nights were split into several runs of acceptable data quality separated by sections of poor-quality data. The extracted power spectrum is a convolution of the Fourier transforms of both the data and the window function. A window function with a broad Fourier transform will yield a power-law fit with an artificially shallow slope, which will in turn predict an inflated value for the RWFE at high frequencies. In order to derive the correct slope of the power-law fit we used only the longest continuous series of measurements from a given night. The power spectrum for the night of December $24^{th}$ is shown in figure \[ps.fig\]. Channel \#1 and channel \#2 are shown as solid and dotted lines respectively. The slope of the power-law fit for the various channels are $\alpha$ and the offsets are $\beta$.
Determination of residual wavefront error
-----------------------------------------
Because the distance to the sodium layer is not know a priori, natural guide stars are required to keep the telescope focused on science targets. In order for a sufficient number of guide stars to be present to make a large fraction of the sky available, the NGSAO system must be designed to work with relatively faint stars. The low fluxes from these stars necessitate a slow read out rate. Therefore, NGSAO can only provide low-frequency information regarding focus changes. A laser guide star can be designed to emit relatively high flux, and can therefore be read out quickly, providing high-frequency focus information. Herriot et al.[@herr] combine the high-frequency focus information from the laser guide star and the low-frequency focus information from the natural guide star to obtain focus information over a wide range of frequencies. Due to limited bandwidth, the aforementioned correction is not perfect, but may be characterized by the rejection transfer function. The residual wavefront variance (RWFE)$^2$ is proportional to the integral over the appropriate range of frequencies ($10^{-4}$ to $10^4$ Hz) of the product of the mean height power spectrum and the square of the AO system rejection transfer function, $$RWFE = \frac{1}{8\sqrt{3}}\left(\frac{D}{h}\right)^2
\left[
%\int_{2\pi10^{-4}Hz}^{2\pi10^4Hz} h_{Na} \left| H_{rej}\right|^2df.
\int_{\nu_1}^{\nu_2} h_{Na} \left| H_{rej}\right|^2df
\right]^{1/2}.$$
The residual wavefront error was calculated for each night with a rejection transfer function designed for the TMT. The cumulative distribution for all nights with more than $7$ hours of continuous acceptable data is shown in figure \[cum7.fig\]. We found that nights with less than $7$ hours of continuous data had a shallower slope than the datasets with longer continuous streams. This is due to the aforementioned effect of the convolution of a narrow window function with the true power spectrum.
Results and Conclusions
=======================
The data for each night with at least one hour of acceptable quality (76 of 89) were analyzed. It was found that nights with sections of usable data that were less than $7$ hours (48 nights) had a shallower slope with respect to those nights with longer sections of usable data (28 nights). Using just the longer nights we find the best-fit power law has the parameters $\alpha=-1.79\pm0.02$ and $\beta=1.12\pm0.40$. We found that RWFE for those nights has a median value of $20$ nm and a $90^{th}$ percentile value of $34$ nm. This effect is an important contribution to the overall error budget of NFIRAOS, and more generally, will likely be for most, if not all, of the LGSAO systems designed for the next generation of large ground-based telescopes.
This analysis was limited by the relatively long integration times of the CSU lidar experiment, which in turn set a low Nyquist frequency. The variability of the sodium layer height in the frequency range observed is mainly due to the upward passing of atmospheric waves. Atmospheric turbulence at the sodium layer height can perturb the mean height of the sodium layer on time scales of seconds or shorter. However, changes caused by atmospheric turbulence are likely to be small (cm or less), unlike the kilometer-scale changes reported here. The assumption that the mean height power spectrum can be fit by a single unbroken power law from $\mu$Hz to kHz is perhaps naïve, but reasonable. The actual high-frequency behavior of the mean sodium height power spectrum, of course, can only be determined by further experimentation. The temporal resolution could be improved by designing a lidar experiment with either a larger collecting telescope, or by sacrificing spatial resolution.
Acknowledgments {#acknowledgments .unnumbered}
===============
This research was supported by the Natural Sciences and Engineering Research Council of Canada. The work at Colorado State University was in part supported by National Aeronautics and Space Administration, under grant NAG5-10076 and National Science Foundation, under grant ATM-00-03171. S. Davis’ email is [[email protected]]{}.
[99]{} Jacques M. Beckers, “Adaptive optics for astronomy – principles, performance, and applications”, Annual review of astronomy and astrophysics, [**31**]{}, 13 (1993). C. O’Sullivan, R.M. Redfern, N. Ageorges, H.-C. Holstenberg, W. Hackenberg, T. Ott, S. Rabien, R. Davies, A. Eckart, “Short timescale variability of the mesospheric sodium layer”, Experimental Astronomy, [**10**]{}, 147 (2000). G. Herriot, P. Hickson, B.E. Ellerbroek, C.Y. She, “Focus errors from tracking sodium layer altitude variations with laser guide star adaptive optics for the Thirty meter Telescope”, Soc. Photo-Optical Instrum. Eng., [**6272**]{}, in press (2006). C.Y. She, T. Li, R. C. Collins, T. Yuan, B. P. Williams, T. D. Kawahara, J. D. Vance, P. Acott, D. A. Krueger, H.-L. Liu, and M. E. Hagan, “Tidal perturbations and variability in the mesopause region over Fort Collins, CO (41N, 105W): Continuous multi-day temperature and wind lidar observations”, Geophys. Res. Lett., [**31**]{}, L24111 (2004).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We introduce an individual claims forecasting framework utilizing Bayesian mixture density networks that can be used for claims analytics tasks such as case reserving and triaging. The proposed approach enables incorporating claims information from both structured and unstructured data sources, producing multi-period cash flow forecasts, and generating different scenarios of future payment patterns. We implement and evaluate the modeling framework using publicly available data.'
author:
- |
Kevin Kuo\
Kasa AI\
`[email protected]`\
bibliography:
- 'bnn-claims.bib'
title: Individual Claims Forecasting with Bayesian Mixture Density Networks
---
Introduction
============
Individual claims reserving has garnered increasing interest in recent years. While the main benefit cited for performing reserving at the claim level over aggregate loss triangle approaches, such as chain ladder and Bornhuetter-Ferguson, is potential improvement in predictive accuracy, especially in environments with changing portfolio mix [@boumezouedIndividualClaims2017], there are additional practical advantages to forecasting individual claim behavior. These include being able to obtain updated views of portfolio risk as claims are reported and optimize adjuster resource allocation based on severity predictions. Although the benefits of individual claims modeling are promising, it has not yet achieved widespread adoption in practice. One contributing reason for the lack of adoption, we hypothesize, is the absence of a modeling framework with features important to practitioners. We suggest that an effective loss reserves modeling framework should be able to:
- Incorporate arbitrary claims information as predictors,
- Produce multi-period forecasts that are sufficiently stable over time, and
- Sample different realizations of future payment patterns that encompass both process uncertainty and model risk.
Companies are capturing increasingly diverse data, such as unstructured text from claims adjusters’ notes and photographs of damages, that can potentially be predictive. Timing of cash flows and being able to sample from different future states have both business decision making and regulatory applications. A desirable characteristic of the forecasts is that they are stable over time, as management is averse to volatility in loss reserves figures from one accounting period to the next. A subjective criterion not listed above, which may affect adoption of an approach, is that the models should be implementable without the need for extensive bespoke feature engineering or specification of complex assumptions for underlying stochastic processes.
To the best of our knowledge, no existing loss reserving framework implements all of the above features. In this paper, we propose an extensible individual claims forecasting framework towards satisfying many of these criteria, utilizing ideas from Bayesian neural networks (BNN) [@nealBayesianLearning2012] and mixture density networks (MDN) [@bishopMixtureDensity1994]. While we discuss these concepts in detail later in the paper, at a high level,
- BNNs are non-linear supervised learning models that capture complex interactions among inputs, with prior distributions on model parameters; and
- MDNs are mixture models for conditional densities, where the mixture model parameters are the outputs of the neural network.
Concretely, our contributions are:
- Development of an individual claims forecasting framework based on Bayesian Mixture Density Networks (BMDN).
- Implementation of the proposed framework using publicly available claims-level data, which provides a baseline for future work to compare against.
Related Work
============
Claims-level reserving is a fast-moving research area, and [@boumezouedIndividualClaims2017] and [@taylorLossReserving2019] provide recent surveys. Many of the current works in the area utilize machine learning (ML) techniques. Wüthrich [@wuthrichMachineLearning2018] introduces using machine learning algorithms to incorporate diverse claims characteristics inputs. It demonstrates a simple model where regression trees are used to predict the number of payments, and suggests extensions such as compound modeling to predict severity and bootstrap simulation to obtain prediction uncertainty. More recently, [@duvalIndividualLoss2019; @lopezTreeBasedAlgorithm2019; @baudryMachineLearning] also utilize tree-based techniques for individual claims reserving. Another direction of research for the individual claims reserving problem are the generative approaches of, for example, [@antonioMicrolevelStochastic2014; @pigeonIndividualLoss2013; @pigeonIndividualLoss2014]. In this latter category of methodologies, a set of distributional assumptions are posited for the different drivers of claims, such as time to the next payment and its amount. These distributions are fit to the data, then, with the obtained parameters, the modeler is able to perform simulations of future development paths by sampling from the fitted distributions. While this approach provides a natural way to obtain samples of future cash flow paths, the distributional assumptions may be too rigid in some cases. It is also difficult to incorporate individual claim characteristics; to differentiate among different characteristics, one would have to segment the claims and fit separate models to each group.
In formulating our framework, we also draw inspiration from machine learning approaches to aggregate triangle data, including [@gabrielliNeuralNetwork2019a; @gabrielliNeuralNetwork2019], which embed a classical parametric loss reserving models into neural networks, and the DeepTriangle [@kuo2019deeptriangle] framework, whose neural network architecture we adapt for individual claims data.
Preliminaries
=============
We begin with a description of the loss reserving problem then briefly discuss BNN and MDN, two ideas that we incorporate into our claims forecasting neural network. As neural networks have been utilized and discussed extensively in recent loss reserving literature [@gabrielliNeuralNetwork2019a; @gabrielliNeuralNetwork2019; @kuo2019deeptriangle; @wuthrichNeuralNetworks2018; @gabrielliIndividualClaims2018], we defer discussion of neural network fundamentals to those works and the standard reference [@goodfellowDeepLearning2016]. To aid the practicing actuary in consuming this paper, we expand on certain concepts when we introduce the proposed neural network architecture in Section \[sec:claims-forecasting-model\].
We remark that in sections \[section:bnn\] and \[section:mdn\], we discuss Bayesian inference and mixture density networks, respectively, in general, and provide details on our specific choices of distributions later on in the paper.
The Loss Reserving Problem
--------------------------
Figure \[fig:claimrunoff\] shows a diagram of the development of a typical claim. We first point out that there can be a time difference, known as the reporting lag, between an accident’s occurrence and its reporting. The accidents which have been reported to the insurer, but not yet settled, are known as reported but not settled (RBNS) or, equivalently, incurred but not enough reported (IBNER) claims, while the accidents which have occurred but are yet unknown to the insurer are known as incurred but not reported (IBNR) claims. The reserving actuary is interested in estimating the ultimate loss amounts associated with accidents that have already occurred. As demonstrated in Figure \[fig:claimrunoff\], it is possible for a closed claim to re-open, and it is also possible for a claim to be closed without any cash flows.
In this paper, we are concerned with RBNS/IBNER, but not IBNR, claims, due to limitations of available data, as for IBNR claims we do not have individual claim feature information available. Also, the claims we study encompass closed claims to allow for the possibility of claim re-opening.
![Development of a claim[]{data-label="fig:claimrunoff"}](images/claim_runoff.png){width="70.00000%"}
Bayesian Inference on Neural Networks {#section:bnn}
-------------------------------------
In this section, we briefly discuss Bayesian inference on neural networks, following the discourse in [@gravesPracticalVariational2011] and [@blundellWeightUncertainty2015].
Consider a neural network parameterized by weights $w$ that takes input $x$ and returns output $y$. We can view this general formulation as a probabilistic model $P(y|x, w)$; as an example, in the case of linear regression where $y \in
\mathbb{R}$, mean squared loss is specified, and constant variance is assumed, $P(y|x, w)$ corresponds to a Gaussian distribution.
Rather than treating $w$ as a fixed unknown parameter, we adapt a Bayesian perspective and treat $w$ as a random variable with some prior distribution $P(w)$. The task is then to compute the posterior distribution $P(w|x, y)$ given the training data. We can then calculate the posterior predictive distribution $P(y^*|x^*) = \mathbb{E}_{P(w|x, y)}[P(y^*|x^*, w)]$, where $x^*$ is a new data point to be scored and $Y^*$ the corresponding unknown response. However, determining $P(w|x, y)$ analytically is intractable, and convergence of Markov chain Monte Carlo (MCMC) to the actual posterior for nontrivial neural networks is too slow to be feasible. Variational inference provides a workaround to this problem by approximating the posterior with a more tractable distribution $q(w|\theta)$, which is often chosen to come from the mean-field family, i.e., $q(z) = \prod_i q(z_i)$, but can be more general [@blundellWeightUncertainty2015]. We then formulate an optimization problem to find the parameters $\theta$. Specifically, we minimize the Kullback-Leibler (KL) divergence from the true posterior distribution $P$ to the approximate distribution $q$. KL divergence is defined in general for probability distributions $P$ to $Q$ as
$$D_{KL}(Q \Vert P) = \int q(x) \log \frac{q(x)}{p(x)} dx.$$
Our optimization problem can then be stated as
$$\begin{aligned}
\theta^* &= \arg \min_\theta D_{KL}(q(w|\theta) \Vert P(w|\mathcal{D}))\\
&= \arg \min_\theta \int q(w|\theta) \log \frac{q(w|\theta)}{P(w|\mathcal{D})} dw\\
&= \arg \min_\theta \int q(w|\theta) \log \frac{q(w|\theta)}{P(w)P(\mathcal{D}|w)} dw \label{eq:2}\\
&= \arg \min_\theta D_{KL}(q(w|\theta) \Vert P(w)) - \mathbb{E}_{q(w|\theta)}[\log P(\mathcal{D}|w)] \label{eq:1}\end{aligned}$$
where $\mathcal{D} = (x_i, y_i)_i$ is the training dataset. We remark that, in Equation \[eq:2\], the $P(\mathcal{D})$ term resulting from applying Bayes’ theorem to $P(w|\mathcal{D})$ disappears because it is irrelevant to the optimization.
Equation \[eq:1\] then gives us the optimization objective[^1]
$$\mathcal{F}(\mathcal{D}, \theta) = D_{KL}(q(w|\theta) \Vert P(w)) - \mathbb{E}_{q(w|\theta)}[\log P(\mathcal{D}|w)].$$
In practice, during training, where $\mathcal{D}$ is randomly split into $M$ mini-batches $\mathcal{D}_1,\dots,\mathcal{D}_M$, we compute $\mathcal{F}(\mathcal{D}, \theta) = \sum_1^M \mathcal{F}_i(\mathcal{D}_i, \theta)$, where
$$\mathcal{F}_i(\mathcal{D}_i, \theta) = \frac{|\mathcal{D}_i|}{|\mathcal{D}|}D_{KL}(q(w|\theta) \Vert P(w)) - \mathbb{E}_{q(w|\theta)}[\log P(\mathcal{D}_i|w)],$$
which we approximate with
$$\mathcal{F}_i(\mathcal{D}_i, \theta) \approx \sum_{j=1}^{|\mathcal{D}_i|} \frac{1}{|\mathcal{D}|} ( \log q(w^{(j)}|\theta) - \log P(w^{(j)}) ) - \log P(\mathcal{D}_i|w^{(j)}),$$
where the $w^{(j)}$ are sampled independently from $q(w|\theta)$. In other words, we sample from the weights distribution just once for each training sample.
Mixture Density Networks {#section:mdn}
------------------------
In practical applications, the response variables we try to predict often have multimodal distributions and exhibit heteroscedastic errors. This is particular relevant in forecasting claims cash flows since, in a given time period, there could be a large payment or little or no payment. An MDN allows the output to follow a mixture of arbitrary distributions and estimate each of its parameters with the neural network. Recall that a mixture distribution has a distribution function of the form
$$F(z) = \sum_{i = 1}^n w_i P_i(z),$$
where each $w_i \geq 0$, $\sum w_i = 1$, and $n$ is the number of component distributions $P_i$. Letting $\mathcal{P}$ denote the union of the sets of parameters for the distributions $P_1, \dots, P_n$, the neural network must then output $n + |\mathcal{P}|$ values. The first $n$ values determine the categorical distribution of the mixing weights, and the $|\mathcal{P}|$ outputs parameterize the component distributions.
By obtaining distributions rather than single points as prediction outputs, we also gain a straightforward mechanism to quantify the uncertainty of individual cash flow forecasts.
We emphasize now the difference between the uncertainty captured by specifying a distribution as the neural network’s output, as discussed here, and the uncertainty in the weight distribution, as discussed in Section \[section:bnn\]. The former corresponds to the irreducible pure randomness, while the latter reflects uncertainty in parameter estimation.
Data and Model
==============
In this section, we describe the data used for our experiments and the proposed model. The dataset used and relevant code are available on GitHub.[^2] The experiments are implemented using the R programming language [@rdevelopmentcoreteamLanguageEnvironment2011] and TensorFlow [@tensorflow2015-whitepaper].
Data
----
We utilize the individual claims history simulator of [@gabrielliIndividualClaims2018] to generate data for our experiments. For each claim, we have the following static information: line of business, labor sector of the injured, accident year, accident quarter, age of the injured, part of body injured, and the reporting year. In addition, we also have 12 years of claim development information, in the form of cash flows and claims statuses (whether the claim is open or not). The generated claims history exhibit behaviors such as negative cash flows for recoveries and late reporting, which mimic realistic scenarios.
Since we only have claims data and not policy level and exposure data, we study only reported claims.
Experiment Setup
----------------
We first simulate approximately 500,000[^3] claims using the simulator, which provides us with the full development history of these claims. The claims cover accident years 1994 to 2005. Since we are concerned with reported claims, we remove claims with report date after 2005, which leaves us with N = 497,516 claims. In this paper, we assume that each claim is fully developed at development year 11 (note that in the dataset the first development year is denoted year 0). More formally, we can represent the dataset as the collection
$$\mathcal{D} = \{ (X^{(j)}, (C_i^{(j)})_{0 \leq i \leq 11}, (S_i^{(j)})_{0 \leq i \leq 11}): j\in {1,\dots, N}\},$$
where $X$, $(C_i)$, and $(S_i)$ denote the static claim features, incremental cash flow sequences, and claim status sequences, respectively, and $j$ indexes the claims.
To create the training and testing sets, we select year 2005 as the evaluation year cutoff. For the training set, any cash flow information available after 2005 is removed. In symbols, we have
$$\mathcal{D}_{\text{train}} = \{ (X^{(j)}, (C_i^{(j)}), (S_i^{(j)})): i + \text{AY}^{(j)} \leq 2005, j \in {1,\dots, N}\},$$
where $\text{AY}^{(j)}$ denotes the accident year associated with claim $j$.
For each claim in the training set, we create a training sample for each time period after development year 0. The response variable consists of cash flow information available as of the end of each time period until the evaluation year cutoff, and predictors are derived from information available before the time period. For a claim $j$, we have the following input-output pairs:
$$\label{eq:training-samples}
\{((X^{(j)}, (C_0^{(j)}, \dots, C_i^{(j)}), (S_0^{(j)}, \dots, S_i^{(j)})), (C_{i+1}^{(j)}, \dots, C_{k^{(j)}}^{(j)})): i = 0, \dots, k^{(j)} - 1\},$$
where $k^{(j)}$ denotes the latest development year for which data is available for claim $j$ in the training set. As an example, if a claim has an accident year of 2000, five training samples are created. The first training sample has cash flows from 2001 to 2005 for the response and one cash flow value from 2000 for the predictor, while the last training sample has only the cash flow in year 2005 for the response and cash flows from 2000 to 2004 for the predictor.
We note here that we do not predict future claim statuses. As we will discuss in Section \[section:outputdist\], our output distributions, which contain point masses at zero, can accommodate behaviors of both open and closed claims.
The training samples in Equation \[eq:training-samples\] undergo additional transformations before they are used for training our model. We discuss these transformations in detail in the next section.
Feature Engineering {#sec:fe}
-------------------
We exhibit the predictors used and associated transformations in Table \[table:vars\]. In the raw simulated data, each time period has a cash flow amount along with a claim open status indicator associated with it. We take the cash flow value and derive two variables from it: the paid loss and the recovery, corresponding to the positive part and negative part of the cash flow, respectively. In other words, for each claim and for each time step, at most one of paid loss and recovery can be nonzero. For predictors, both the paid losses and recoveries are centered and scaled with respect to all instances of their values in the training set. The response cash flow values are not normalized.
Claim status indicator, which is a scalar value of 0 or 1, is one-hot encoded (i.e., represented using dummy variables); for each training sample, a claim status value is available for each time step.
Development year is defined as the number of years since the accident occurred. It is then scaled to $[0, 1]$.
The static claim characteristic variables include age, line of business, occupation of the claimant, and the injured body part. Of these, age is numeric while the others are categorical. We center and scale the age variable and integer-index the others, which are fed into embedding layers [@guoEntityEmbeddings2016], discussed in the next section.
As noted in the previous section, the response variable and the sequence predictors can have different lengths (in terms of time steps) from one sample to the next. To facilitate computation, we pre-pad and post-pad the sequences with a predetermined masking value (we use 99999) so that all sequences have a fixed length of 11.
[lll]{} Variable & Type & Preprocessing\
Paid loss history & Numeric sequence & Centered and scaled\
Recovery history & Numeric sequence & Centered and scaled\
Claim status history & Categorical sequence & One-hot encoded\
Development year & Integer & Scaled to \[0, 1\]\
Age & Numeric & Centered and scaled\
Line of business & Categorical & Indexed\
Claim code (occupation)& Categorical & Indexed\
Injured part & Categorical & Indexed\
\
Claims Forecasting Model {#sec:claims-forecasting-model}
------------------------
We utilize an encoder-decoder architecture with sequence output similar to the model proposed by [@kuo2019deeptriangle]. The architecture is illustrated in Figure \[fig:architecture\]. We first provide a brief overview of the architecture, then provide details on specific components later in the section.
The output in our case is the future sequence of distributions of loss payments, and the input is the cash flow and claim status history along with static claim characteristics.
![Claims forecasting model architecture. Note that the weights for the variational dense layers are shared across time steps.[]{data-label="fig:architecture"}](images/architecture.png){width="50.00000%"}
Static categorical variable inputs, such as labor sector of the injured, are indexed, then connected to embedding layers [@guoEntityEmbeddings2016] and sequential data, including payments and claim statuses, are connected to long short-term memory (LSTM) layers [@hochreiterLongShortterm1997].
The encoded inputs are concatenated together, then repeated 11 times before being passed to a decoder LSTM layer which returns a sequence. The length of this output sequence is so chosen to match our requirement to forecast a maximum of 11 steps into the future. Each time step of this sequence is connected to two dense variational layers, each of which parameterizes an output distribution, corresponding to paid losses and recoveries, respectively. The weights of the dense variational layer (for paid loss and recovery) are each shared across the time steps. In other words, for each training sample, we output two 11-dimensional random variables, each of which can be considered as a collection of 11 independent but non-identically distributed random variables.
We use the same loss function for each output and weight them equally for model optimization. The loss function is the negative log-likelihood given the target data with an adjustment for variable output lengths which we discuss in detail in Section \[sec:loss-function\].
In the remainder of this section, we discuss in more detail embedding layers, LSTM, our choices of the variational and distribution layers, the loss function, and model training.
### Embedding Layer
An embedding layer maps each level of a categorical variable to a fixed-length vector in $n$-dimensional Euclidean space. The value of $n$ is a hyperparameter chosen by the modeler; in our case we select $n = 2$ for all embedding layers, which means we map each factor level to a point in $\mathbb{R}^2$. In contrast to data-preprocessing dimensionality techniques such as principal component analysis (PCA) or t-distributed stochastic neighbor embedding (t-SNE), the values of the embeddings are learned during training of the neural network.
### Long Short-Term Memory
LSTM is a type of recurrent neural network (RNN) architecture that is appropriate for sequential data, such as natural language or time series (our use case). Empirically, LSTM is more apt at handling data with longer term dependencies than simple RNN [@lecunDeepLearning2015]. In our model, for both of the sequence input encoders and the decoder, we utilize a single layer of LSTM with 3 hidden units. We found that these relatively small modules provided reasonable results, but did not perform comprehensive tuning to test deeper and larger architectures.
### Dense Variational Layer
We choose Gaussians for both the prior and surrogate posterior distributions over the weights of the dense layers. The prior distribution is constrained to have zero mean and unit variance while for the surrogate posterior both the mean and variance are trainable. Symbolically, if we let $w$ denote the weights associated with each dense layer, we have
$$P(w) = \mathcal{N}(0, I) \text{ and}$$
$$q(w|\theta) = \mathcal{N}(\mu(\theta), \sigma(\theta)),$$
where $\theta$ represents trainable parameters. The KL term associated with estimation, as described in Section \[section:bnn\], is added to the neural network loss during optimization. The dense layers each output four units to parameterize the output distributions which we discuss next.
### Output Distribution {#section:outputdist}
For each of the 11 time steps we forecast, we parameterize two distributions, one for paid losses and one for recoveries. Each of the distributions is chosen to be a mixture of a shifted log-normal distribution and a deterministic degenerate distribution localized at zero, representing cash flow and no cash flow, respectively. Denoting $\{v_1,\dots,v_4\}$ to be the output of the preceding dense variational layer, we have the following as the distribution function:
$$\label{eq:fx}
F(y) = w_1(v_1, v_2)P(y; v_3, (\alpha + \lambda\sigma(\beta v_4))^2) + w_2(v_1, v_2)F_0(y).$$
We now describe the components of Equation \[eq:fx\]. First,
$$w_i = \frac{e^{v_i}}{e^{v_1} + e^{v_2}}\quad (i = 1, 2)$$
are the normalized mixing weights. $P(y; \mu, \sigma^2)$ denotes the distribution function for a log-normal distribution with location and scale parameters $\mu$ and $\sigma$, respectively, shifted to the left by 0.001 to accommodate zeros in the data. In other words, if $Y \sim LN(\mu, \sigma^2)$, then $Y - 0.001 \sim P$. This modification is necessary because we have zero cash flows in the response sequences, and computing the likelihood becomes problematic since the support of the log-normal distribution does not contain zero.[^4]
$$F_{0}(y)=\left\{\begin{matrix} 1, & \mbox{if }y\ge 0 \\ 0, & \mbox{if }y<0 \end{matrix}\right.$$
is the distribution function for the deterministic distribution, and $\sigma$ is the sigmoid function defined as
$$\sigma(t) = \frac{1}{1 + e^{-t}}.$$
In Equation \[eq:fx\], the constants $\alpha, \beta > 0$ are included for numerical stability and can be tuned as hyperparameters, although we set them to $0.001$ and $0.01$, respectively, for our experiments. The purpose of $\lambda$ is also numerical stability, as the scale parameter of the log-normal distribution is hard to learn in practice, so we bound it above by a constant. In our experiments we fix it to be $0.7$.
The net loss prediction for each development year is then the difference of the paid loss and recovery amounts. We assume that the output distributions, conditional on their parameters, are independent across the time steps, which facilitates the derivation of the loss function in the next section. We remark that this is a weakness of our approach, since the independence assumption is not fulfilled in practice.
### Loss Function {#sec:loss-function}
We calculate the log-likelihood loss for each output by computing the log probability of the true labels with respect to the output distributions. As mentioned in Section \[sec:fe\], the output sequences may contain masking values that need to be adjusted. Specifically, we marginalize out the components with the masking values. Formally, for a given training sample, let $M$ denote the masking constant, $Y=(Y_1,\dots,Y_{11})$ denote a single output sequence, $Y_{\mathcal{T}_M} = (Y_i: Y_i = M)$, and, abusing notation, also let $\mathcal{T}_M = \{i: Y_i = M\}$. Then,
$$\begin{aligned}
f_{Y\setminus Y_{\mathcal{T}_M}}(Y) &= \int_{Y_{\mathcal{T}_M}} f(Y) dY_{\mathcal{T}_M}\\
&= \int_{Y_{\mathcal{T}_M}} f(Y\setminus Y_{\mathcal{T}_M}, Y_{\mathcal{T}_M}) dY_{\mathcal{T}_M}\\
&= f_{Y\setminus Y_{\mathcal{T}_M}}(Y\setminus Y_{\mathcal{T}_M}) \int_{Y_{\mathcal{T}_M}} f_{Y_{\mathcal{T}_M}}(Y_{\mathcal{T}_M}) dY_{\mathcal{T}_M} \quad \text{(by independence)}\\
&= f_{Y\setminus Y_{\mathcal{T}_M}}(Y\setminus Y_{\mathcal{T}_M})\\
&= \prod_{i \notin \mathcal{T}_M} f_{Y_i}(Y_i) \quad \text{(by independence)}.\end{aligned}$$
The adjusted log-likelihood for a single training sample then becomes
$$\begin{aligned}
\log\mathcal{L}(Y) &= \log f_{Y\setminus Y_{\mathcal{T}_M}}(Y)\\
&= \log \prod_{i \notin \mathcal{T}_M} f_{Y_i}(Y_i)\\
&= \sum_{i = 1}^{T}\log f_{Y_i}(Y_i),\end{aligned}$$
where $T$ is $\min \mathcal{T}_M - 1$ if $\mathcal{T}_M$ is nonempty and 11 otherwise.
Hence, the log-likelihood associated with a training sample is the sum of the log probabilities of the non-masked elements. The negative log-likelihood, summed with the KL term associated with the variational layers, discussed in Section \[section:bnn\], comprise the total network loss for optimization. The contribution of a single training sample to the loss is then
$$\sum_{i = 1}^{T}\log f_{Y_i}(Y_i) + \frac{1}{|\mathcal{D}|}\sum_{k=1}^2D_{KL}(q_k(w_k|\theta_k) \Vert P_k(w_k)),$$
where $k=1, 2$ correspond to variational layers for parameterizing the paid loss and recovery distributions, respectively.
### Training and Scoring
We use a random subset of the training set consisting of 5% of the records as the validation set for determining early stopping and scheduling the learning rate. For optimizing the neural network, we use stochastic gradient descent with an initial learning rate of 0.01 and a minibatch size of 100,000, and halve the learning rate when there is no improvement in the validation loss for five epochs. Training is stopped early when the validation loss does not improve for ten epochs; we cap the maximum number of epochs at 100. Here, an *epoch* refers to a complete iteration through our training dataset, and the *minibatch* size refers to how many training samples we randomly sample for each gradient descent step.
To forecast an individual claim, we construct a scoring data point by using all data available to us as of the evaluation date cutoff. In other words, the last elements of the predictor cash flow sequences correspond to the actual value from the cutoff year. Hence, in the output sequence, the first element corresponds to the prediction of the cash flow distribution of the year after the cutoff.
Recall that our model weights are a random variable, so each time we score a new data point, we are sampling from the distribution of model weights and expect to obtain different parameters for the distributions of cash flows. The output distributions themselves are also stochastic and can be sampled from. Following [@kendallWhatUncertainties2017], we refer to the former variability as *epistemic* uncertainty and the latter as *aleatoric* uncertainty. Epistemic uncertainty reflects uncertainty about our model which can be reduced given more training data, while aleatoric uncertainty reflects the irreducible noise in the observations. In actuarial science literature, these concepts are also known as parameter estimation uncertainty and process uncertainty, respectively [@wuthrichNonlifeInsurance2017].
Generating a distribution of future paths of cash flows amounts to repeating the procedure of drawing a model from its distribution, calculating the output distribution parameters using the drawn model, then drawing a sample of cash flows from the distribution. Since we decompose the cash flows into paid losses and recoveries, we can obtain net amounts by subtracting the generated recoveries from the generated paid losses.
If we were only interested in point estimates, we can avoid sampling the output distributions and simply compute their means, since given the model weights we have closed form expressions for the distributions given by Equation \[eq:fx\]. Note that we would still have to sample the model weights.
Results and Discussion
======================
We evaluate our modeling framework by qualitatively inspecting samples paths generated for an individual claim and also comparing the aggregate estimate of unpaid claims with a chain ladder estimate. We note that, to the best of our knowledge, prior to the current paper, there is not a benchmark for individual claims forecasts. We also discuss the extensibility of our approach to accommodate company specific data elements and expert knowledge.
Individual Claim Forecasts
--------------------------
Figure \[fig:prob\_cash\_flows\] shows various posterior densities, obtained by sampling the model weights 1,000 times, of parameters for the output distributions of a single claim. Our model assigns a higher probability of payment in the next year along with more variability around that probability. It can be seen that the expected probability of a payment decreases as we forecast further into the future. Given that a loss payment occurs, we see from the middle and bottom plots that both the expected value and the variability for both the mean and variance of the log-normal distributions increase with time. In other words, for this particular claim, loss payments become less likely as time goes on, but if they occur, they tend to be more severe and the model is less certain about the severity distribution.
![Posterior distributions for a single claim at development year 4. (Top) Payment probability. (Middle) Mean of loss payment distribution. (Bottom) Log variance of payment distribution.[]{data-label="fig:prob_cash_flows"}](images/posteriors.png){width="80.00000%"}
![1,000 samples of cumulative cash flow paths for a single claim. The training data includes development year 4 and the predictions begin in development year 5.[]{data-label="fig:claim_paths"}](images/claim_paths.png){width="80.00000%"}
In Figure \[fig:claim\_paths\] we show plausible future developments of the claim. Note that one thousand samples are drawn, so the handful of scenarios that present large payments represents a small portion of the paths, which is consistent with the distributions of parameters we see above.
Aggregate estimates
-------------------
To compute a point estimate for the total unpaid losses, we score each claim once (i.e., sample from the weights distribution once) to obtain a distribution for each time step for paid losses and recoveries. We then compute the means of the paid loss and recovery distributions and take the differences to obtain the net cash flow amount. The final unpaid loss estimate is then the sum of all the net amounts up to and including development year 11. Since there is randomness in the neural network weight initialization, we instantiate and train the model 10 times, and take the average of the predicted future paid losses across the 10-model ensemble. In practice, one could distribute the prediction procedure over a computing cluster and draw more samples for an even more stable estimate.
For the chain ladder development method benchmark, we aggregate the data into a report year triangle (to exclude IBNR), calculate ultimate losses using all-year weighted age-to-age factors, then subtract the already paid amounts to arrive at the unpaid estimate. The unpaid amounts from the two approaches are then compared to the actual unpaid amounts. The results are shown in Table \[table:results\].
[lll]{} Model & Forecast & Error\
Actual & 109,216,388 & –\
Chain Ladder & 108,040,572 & $-1.08\%$\
Ours (10-Model Ensemble) & 102,502,710 & $-6.15\%$\
\
While we are not improving on chain ladder estimates at the aggregate level for this particular simulated dataset, our approach is able to provide individual claims forecasts, which are interesting in their own right.
Extensibility
-------------
One of the primary advantages of neural networks is their flexibility. Considering the architecture described in Section \[sec:claims-forecasting-model\], we can include additional predictors by appending additional input layers. These inputs can be unstructured and we can leverage appropriate techniques to process them before combining with the existing features. For example, we may utilize convolutional layers to transform image inputs and recurrent layers to transform audio or text inputs.
The forms of the output distributions can also be customized. We choose a log-normal mixture for our dataset, but depending on the specific book of business, one may want to specify a different distribution, such as a Gamma or a Gaussian. As we have done in constraining the scale parameter of the log-normal distribution, the modeler also has the ability to bound or fix specific parameters as situations call for them.
Conclusion
==========
We have introduced a framework for individual claims forecasting that can be utilized in loss reserving. By leveraging Bayesian neural networks and stochastic output layers, our approach provides ways to learn uncertainty from the data. It is also able to produce cash flow estimates for multiple future time periods. Experiments confirm that the approach gives reasonable results and provides a viable candidate for future work on individual claims forecasting to benchmark against.
There are a few potential avenues for enhancements and extensions, including larger scale simulations to evaluate the quality of uncertainty estimates, evaluating against new datasets, and relaxing simplifying assumptions, such as the independence of time steps to obtain more internally consistent forecasts. One main limitation of the proposed approach and experiments is that we focus on reported claims only. With access to policy level data, a particularly interesting direction of research would be to extend the approach to also estimate IBNR claims.
### Acknowledgments {#acknowledgments .unnumbered}
We thank Sigrid Keydana, Daniel Falbel, Mario Wüthrich, and volunteers from the Casualty Actuarial Society (CAS) for helpful discussions. This work is supported by the CAS.
[^1]: We note that $-\mathcal{F}(\mathcal{D}, \theta)$ is often referred to as the evidence lower bound (ELBO) in the machine learning literature.
[^2]: <https://github.com/kasaai/bnn-claims>
[^3]: The number of claims generated is stochastic; in our case, we draw 500,904 claims.
[^4]: As of the writing of this paper, TensorFlow Probability attempts to evaluate all terms in the likelihood of the mixture.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A class of states of the electromagnetic field involving superpositions of all the excited states above a specified low energy eigenstate of the electromagnetic field is introduced. These states and the photon-added coherent states are shown to be the limiting cases of a generalized photon-added coherent states. This new class of states is nonclassical, non-Gaussian and has equal uncertainties in the field quadratures. For suitable choices of parameters, these uncertainties are very close to those of the coherent states. Nevertheless, these states exhibit sub-Poissonian photon number distribution, which is a nonclassical feature. Under suitable approximations, these states become the generalized Bernoulli states of the field. Nonclassicality of these states is quantified using their entanglement potential.'
author:
- |
S. Sivakumar\
Materials Physics Division\
Indira Gandhi Centre for Atomic Research\
Kalpakkam 603 102 INDIA\
Email: [email protected]\
title: 'Truncated coherent states and photon-addition'
---
PACS: 42.50. -p, 42.50.Lc, 03.67.Bg\
Keywords: Truncated Coherent states, photon-added coherent states, inverse boson operators, entanglement potential
Introduction
============
Quantum theory stipulates a lower bound on the product of the uncertainties of in a pair of non-commuting observables such as the position and momentum, the phase and particle number, etc. In the context of the position-momentum uncertainty relation $\Delta x\Delta p\ge \hbar/2$, there are states which satisfy $\Delta x\Delta p = \hbar/2$. Such states are called minimum uncertainty states(MUS), for example, the coherent states and the squeezed vacuum. Of all the MUS, the coherent states $\vert\alpha\ra$[@glauber2; @ecg], =(-)\_[n=0]{}\^n, defined for every complex number $\alpha$, are very special as their phase space distributions (P- and Wigner-distributions) are well-defined probability densities[@gerryknight]. For that reason, the coherent states are considered to be “classical” among the quantum states. Any other pure state of a harmonic oscillator is non-classical, in the sense that its phase space distribution (P-function or Wigner) fails to be non-negative. Interestingly, superpositions of two coherent states generate non-classical states; the even and odd coherent states[@yurkestoler; @eocs; @ourjou] are prime examples of such states. In the case of the electromagnetic field, photon-addition and state truncation are two other processes to create nonclassical states from coherent states[@gsatara; @nha]. Both these routes have been experimentally realized[@zavatta; @babichev; @grangier; @marek]. In the process of state truncation, a finite number of Fock states are retained in the coherent states. Truncated coherent states (TCS) are defined as[@kuang; @leonski; @tanas94; @leonski97] ,N;u=N\_u\_[n=0]{}\^Nn, where $N_u^{-2}$ is $\exp(\vert\alpha\vert^2)[1-\gamma(N+1,\vert\alpha\vert^2)/N!]$. The incomplete Gamma function $\gamma(N,x)$ [@grad]is defined as \[ingamma\] (N,x)=(N-1)!, $N$ being the order of the function. Since there is an upper limit on the number of Fock states involved in the superposition, this state is referred as upper-truncated coherent states (UTCS). These experimentally realizable states have been studied extensively for their nonclassical features[@leonski; @ppb; @kuang_pla; @jcp1; @jcp2]. In the limit $N\rightarrow\infty$, $\vert\alpha,N;u\ra$ becomes the coherent state $\vert\alpha\ra$. Another class of states associated with the UTCS is defined as follows: \[Ltcsdef\] ,N;l=N\_l\_[n=0]{}\^N+1+n. The normalization constant satisfies $N_l^{-2}=\exp(\vert\alpha\vert^2)\gamma(N+1,\vert\alpha\vert^2)/N!$. The lower limit on the Fock states implies that these states could be aptly termed as lower-truncated coherent states (LTCS). By construction, the UTCS and LTCS are orthogonal to each other. If $\alpha\rightarrow 0$, then $\vert\psi,N\ra_u$ and $\vert\psi,N\ra_l$ are $\vert 0\ra$ and $\vert N+1\ra$ respectively. The canonical coherent state $\vert\alpha\ra$ is expressed as =.
In this work, the properties of the LTCS are presented and compared with those of the UTCS. In Section II, the relation between the LTCS and another well known nonclassical state, namely, the photon-added coherent states (PACS) is established. Nonclassical features such as squeezing and sub-Poissonian statistics are discussed in Section III, followed by a summary of the results in Section IV.
LTCS and Photon-added coherent states
=====================================
Photon-addition is mathematically represented by the action of a suitable creation operator on a state of the electromagnetic field. The process leads to generation of nonclassical states from the coherent states[@gsatara]. The process and its various generalizations have been studied both experimentally and theoretically[@zavatta]; see [@kim_pacs] for a recent review.
Photon-added coherent states (PACS) are defined as[@gsatara] \[pacs\] ,m\^[m]{}=(-)\_[n=0]{}\^n+m. The parameter $m$ is the order of the PACS. These states have been experimentally realized in optical parametric down-conversion[@zavatta]. Both the LTCS $\vert\alpha,N'l\ra$ and the PACS $\vert\alpha,m\ra$ with $m=(N+1)$ are defined are superpositions of the the same set of Fock states $\{\vert n\ra\}$ $[n=m+1,~m+2,\cdots]$. To see the connection between the two states, consider the deformed annihilation operator $\hat{A}=(1+k\ano)\aao$, where $k$ is nonnegative, not exceeding unity. Associated with $\hat{A}$ is the “deformed creation operator” $\hat{B}^\dagger$ $=\aco(1+k\ano)^{-1}$, and these two operators satisfy $\left[\hat{A},\hat{B}^\dagger\right]=I$[@shanta]. It is important to note that $\hat{B}^\dagger$ is not the adjoint of $\hat{A}$. Starting with the coherent state $\vert\alpha\ra$, define ,m\_k=N\_k\^[m]{}=N\_k\_[n=0]{}\^n+m. In the limit of vanishing $k$, the state $\vert\alpha,m\ra_k$ becomes the $m$-PACS $\vert\alpha,m\ra$; with $k=1$, the state becomes the LTCS. Hence, the PACS and LTCS are the limiting cases of the deformed-PACS $\vert\alpha,m\ra_k$. If $k=1$, the creation operation $\hat{B}^\dagger=\aco(1+\ano)^{-1}$ is the right inverse of the annihilation operator $\aao$, that is, $\aao\hat{B}^\dagger=I$[@roymehta]. Thus, the LTCS are the obtained by the repeated application of the right inverse of $\aao$ on the coherent state $\vert\alpha\ra$. The notion of deformed photon-addition has been discussed in the context of Posch-Teller potential[@dpacs] too. The overlap between the PACS and LTCS of same amplitude $\alpha$ and order $N+1$ is ,N+1,N;l\_l\^2=, where $L_N(x)$ is the Laguerre polynomial of order $N$[@grad]. For large values of $N$, the overlap becomes negligible indicating that the two limiting cases are nearly orthogonal to each other.
It is known that the PACS of order $N+1$ provide a resolution of identity for the subspace spanned by the number states ${\vert N+1\ra,~ \vert N+2\ra,~\vert N+3\ra, \cdots}$[@penson]. A similar relation holds for the LTCS too. In this case, d\^2,N;l,N;l=I-\_[n=0]{}\^Nnn. The RHS of the above equation is the identity operator on the subspace spanned by $ \vert N+1\ra,\vert N+2\ra,{\vert N+3\ra,\cdots}$. From the definition of the incomplete Gamma function given in Eq. \[ingamma\], it is seen that $\gamma(N+1, x)$ is non-negative if $x \ge 0$. This means that the weight function in the resolution of identity by the LTCS is nonnegative.
The result that a LTCS of order $N+1$ is obtained by the action of the deformed creation operator $\hat{B}^{\dagger N}$ on the coherent state $\vert\alpha\ra$ leads to a plausible method of generating these states, in close parallel to the suggestion by Agarwal and Tara[@gsatara] to generate the PACS. Consider a two-level atom interacting with a single-mode of the electromagnetic field. The Hamiltonian in the interaction picture is taken to be $H_I=\hbar\lambda\left[\hat{B}^\dagger\sigma_-+\hat{B}\sigma_+\right]$, which is a generalized Jaynes-Cummings model of atom-field interaction[@buck; @welsch]. This interaction corresponds to an intensity-dependent atom-field coupling. Here $\sigma_+$ and $\sigma_-$ are respectively the raising and lowering operators for the atomic states $\vert e\ra$ and $\vert g\ra$ and $\lambda$ is coupling constant. The actions of the raising and lowering operators on the atomic states are given by $\sigma_\pm\vert e\ra=(\vert g\ra\mp\vert g\ra)/2$ and $\sigma_\pm\vert g\ra=(\vert e\ra\pm\vert e\ra)/2$. The initial state of the atom-field state is $\vert\alpha\ra\vert e\ra$; the atom in the excited state $\vert e\ra$ and the field in the coherent state $\vert\alpha\ra$. If the interaction duration is small ($\vert t\lambda\vert<<1$), then the evolution operator is approximated to (-iH\_It/)I-it. The time-evolved state of the atom-field system is the superposition $\vert\alpha\ra\vert e\ra-it \lambda\hat{B}^\dagger\vert\alpha\ra\vert g\ra$, an entangled state of the system. Subsequent to the interaction, if the atom is detected in its ground state $\vert g\ra$, the field is in the state $\hat{B}^\dagger\vert\alpha\ra$, which is the LTCS with $N=0$. If another two-level atom in its excited state interacts with this resultant field $\hat{B}^\dagger\vert\alpha\ra$, and the the atom is detected in its ground state after the interaction, the field changes to $\hat{B}^{\dagger 2}\vert\alpha\ra$. After interacting with a sequence of $N$ atoms which are subsequently detected in their respective ground states, the initial coherent state of the field is transformed to $\hat{B}^{\dagger N}\vert\alpha\ra$, the LTCS of order $N$. It is possible to make the atom-field coupling to be time-dependent and its form can be tailored to generate arbitrary superpositions[@eberly]. Another route to realize the type of the intensity-dependent coupling is to consider situations in which the rotating-wave approximation is not suitable and the resulting Hamiltonian is equivalent to including an intensity-dependent interaction of the form $(1-k\ano)\aao$, which approximates the operator $(1+k\ano)^{-1}\aao$ when $k\la\ano\ra<1$[@naderi], wherein the expectation value $\la\ano\ra$ is in the state $\vert\alpha,N;l\ra$. Further, it has been shown that the properties of a cavity containing nonlinear media can be tuned to provide different forms of intensity-dependent atom-field couplings[@teppo].
Another state that is relevant in the context of the commutation $\left[\hat{A},\hat{B}^\dagger\right]=I$ is $\exp(\alpha\hat{B}^\dagger-\alpha^*\hat{A})\vert 0\ra$, whose Fock basis expansion is (\^-)0\_[n=0]{}\^n. Since $\hat{B}^\dagger$ is not the adjoint of $A$, the operator $\exp(\alpha\hat{B}^\dagger-\alpha^*\hat{A})$ is not unitary. However, its action on the vacuum results in a state that is normalizable for all values of $\alpha$. In this work, these states are not discussed further.
Nonclassical features
=====================
Pure states of the electromagnetic field, except the coherent states, are nonclassical. Squeezing and sub-Poissonian statistics are two of the experimentally verifiable nonclassical features that a state of the electromagnetic field can exhibit. In this section, these two aspects of the LTCS are discussed. If the deforming parameter $k=1$, the expectation of the annihilation operator and creations operators are \_u&=&\^\*\_u=N\
\^2\_u&=&\^[2]{}\^\*\_u=N(N-1)\^2 ,\
\_l&=&\^\*\_l=,\
\^2\_l&=&\^[2]{}\^\*\_l=\^2.The subscripts $u$ and $l$ refer to the UTCS and LTCS respectively. An interesting feature is that the expectations values of the two operators in LTCS are independent of $N$. However, the expectation value of the number operator $\ano$ does not possess this feature, \_u&=&N\^2\
\_l&=&N\^2
Squeezing in X and P quadratures
--------------------------------
For the single mode electromagnetic field, the following quadratures are defined, = =. The corresponding uncertainties are $(\Delta X)^2=\la\hat{X}^2\ra-\la\hat{X}\ra^2$ and $(\Delta P)^2=\la\hat{P}^2\ra-\la\hat{P}\ra^2$. For the coherent states of the electromagnetic field, uncertainties in the two quadratures are equal to 1/2. A state is said to exhibit squeezing in a quadrature if the uncertainty in that quadrature is smaller than 1/2. The PACS, which corresponds to $k=0$ in $\vert\alpha,m\ra_k$, exhibits squeezing. In Fig. \[fig:FigI\], the uncertainty in $X$ quadrature is shown as a function of the parameter $k$, for three different values of $N$. The lowest values of $\Delta X$ are achieved for the PACS ($k=0$ case in the figure). However, squeezing occurs over a substantial range of $k$. If $N < \vert\alpha\vert^2$, the amount of squeezing in the $X$-quadrature is larger for smaller $N$. The range of $k$ over which squeezing occurs, however, decreases as $N$ increases. In this work, the focus is on the properties of the LTCS. Explicit expressions for the uncertainties in the LTCS are (X)\^2\_l=(P)\^2\_l=. In Fig. \[fig:FigIII\] , the variation of $\Delta X$ and $\Delta P$ are shown for different values of $N$ for the LTCS corresponding to $\alpha=\sqrt{10}$.
It is seen from the figures that the quantities $\Delta X$ and $\Delta P$ are nearly equal to those of the coherent state $\alpha$, for $N\le \vert\alpha\vert^2$, though the average number of photons is high. If the condition on $N$ is satisfied, the states $\vert\psi\ra_l$ are MUS with equal variances in the quadratures, whereas the squeezed vacuum is a MUS though with unequal distribution of fluctuations. Unlike in the case of the PACS, the uncertainties in the two quadratures are equal for the LTCS, which holds for the number states and coherent states too. This, in turn, means that LTCS cannot exhibit reduced fluctuation in any of the quadratures, whereas PACS exhibits squeezing.\
For comparison, uncertainty profiles for the UTCS are also given in Fig. \[fig:FigIII\]. In the state $\vert\psi,N\ra_u$, squeezing occurs if the magnitude of $\alpha$ is small. In that limit, UTCS is well approximated by a superposition of the vacuum state $\vert 0\ra$ and the single photon state $\vert 1\ra$. Such states are known as Bernoulli states and they exhibit squeezing[@bernoulli].
The overlap $\vert\la\alpha\vert\alpha,N\ra_l\vert^2$ between the coherent state $\vert\alpha\ra$ and the UTCS is $\exp(-\vert\alpha\vert^2)\sum_{n=0}^N\vert\alpha\vert^{2n}/n!$. If $N >> \vert\alpha\vert^2$, the summation is nearly $\exp(\vert\alpha\vert^2)$ and the overlap is almost equal to unity. In that case, the UTCS is almost the coherent state $\vert\alpha\ra$, which is a MUS. It is interesting to note that for the UTCS the maximum uncertainties in the quadratures occur when $N\approx\vert\alpha\vert^2/\sqrt{2}$.
Phase distribution
------------------
Phase distribution of a quantum state of the electromagnetic field provides information about uncertainty in the phase of the field. This, in turn, imposes a lower limit on the the fluctuations in the number of photons. An useful approach to define phase distribution is the Pegg-Barnett formulation [@qphase]. Phase distribution of an arbitrary state of the field is obtained by the overlap of the state with the phase state =\_[n=0]{}\^s e\^[in]{}n. and taking the limit $s\rightarrow\infty$. The respective phase distributions for the states $\vert\psi\ra_u$ and $\vert\psi\ra_l$ are P\_u()=\_[s]{}\_u\^2=N\_u\^2|\_[n=0]{}\^N|\^2, and P\_l()=\_[s]{}\_l\^2=N\_l\^2|\_[n=N+1]{}\^|\^2. The phase distributions for the UTCS and LTCS are shown in Figs. \[fig:FigIV\] and \[fig:FigV\] respectively, for three different values of $N$. As the cutoff $N$ increases, the phase-distribution of UTCS approaches that of a coherent state of amplitude $\alpha$. In the case of LTCS, the phase-distribution broadens as $N$ increases. This feature is shown in Fig. \[fig:FigV\]. If $N>>\vert\alpha\vert^2$, the LTCS is a superposition of the states $\vert N+1\ra$ and $\vert N+2\ra$, \[bernoulli\] ,N;l. Such a superposition has a fairly well defined number of quanta, the corresponding uncertainty in the number of photons is (n)\^2=-\^2=. If $N >> \vert\alpha\vert^2$, then $\Delta n<< 1$, that is, the photon number fluctuation is small. Consequently, the phase distribution is broad.
The photon-number distribution of the LTCS is not Poissonian. Whether it is sub-Poissonian or not is inferred from the Q-parameter introduced by Mandel[@mandelQ], Q=-1=-1 < 0. for the state given in Eq. \[bernoulli\], which is the large-$N$ limit of the LTCS. This parameter is used to quantify the deviation from the Poissonian distribution. Consequently, the Q-parameter assumes negative values signalling the sub-Poissonian nature of the LTCS. Though the above closed form expressions have been obtained taking $N$ to be large, numerical evaluation of $Q$ for the exact LTCS shows that the states are indeed sub-Poissonian. The $N$-dependence of the Q parameter is shown in Fig. \[fig:FigVA\]. It is seen that for all values of $N$ the Q-parameter is negative.
Entanglement Potential
----------------------
Nonclassicality of a quantum state is amenable to quantification and many measures of nonclassicality are known: nonclassical distance[@hillery; @wunsche], nonclassical depth[@ncdepth], phase-space volume of the negative regions of the Wigner distribution[@wigner_neg], etc. Of these, the criterion based on the Wigner function is measurable as the Wigner function of a quantum state can be constructed by measuring a suitably chosen set of observables[@risken; @meyer; @leonhardt]. Recently, Asboth [*et al*]{} proposed a measure of nonclassicality of any quantum state of a single-mode of the electromagnetic field[@asboth]. Beam-splitter is a device in which two input modes interact to generate two output modes. In principle, due to the interaction, the output modes can be entangled. A necessary condition for a beam-splitter to generate entangled output state is that at least one of the inputs should be a nonclassical[@kim; @wang]. This feature has been used to define a new measure of nonclasciality called entanglement potential. It is the amount of entanglement between the two output modes when the input ports are fed with states $\vert 0\ra\vert\psi\ra$, where $\vert 0\ra$ is the vacuum of one of the modes and the $\vert\psi\ra$ is the state of field input in the other mode. Though both the LTCS and UTCS are nonclassical states, it is important to compare the entanglement potentials of the two states. Each of the input ports of a beam-splitter is modelled as a mode. The unitary operator representing the action of a beam-splitter is $\exp[\Gamma\aco\bao-\Gamma^*\aao\bco]$, $\Gamma$ represents the complex transmittance of the beam splitter[@welsch]. Here $\aao$ and $\aco$ are the annihilation and creation operators for one of the input ports. For the other input port, the corresponding operators are $\bao$ and $\bco$. The operators $K_+=\aco\bao$, $K_-=\aao\bco$ and $K_0=(\ano-\bno)/2$ are generators of the SU(1,1) algebra. One form of disentangled representation of the unitary operator is (K\_+-\^\* K\_+)=(2K\_+)(-e\^[-i]{}K\_0)(e\^[i]{}K\_-), where $\Gamma=\vert\Gamma\vert\exp(i\theta)$[@disent]. If vacuum is the input in one of the input ports and $\vert\psi\ra=\sum_{n=0}^\infty c_n\vert n\ra$ in the other input port, the output state $\vert\chi\ra$ is =\_[n=0]{}\^c\_n(nB)\_[k=0]{}\^n A\^k n-kk. Then the reduced density operator for the $a$-mode is \[rden\] \_[a]{}=\_[n,m=0]{}\^c\_n c\^\*\_m(B(n+m))\_[k=0]{}\^[(n,m)]{}n-km-k, where $B=-e^{-i\theta}\tan\vert\Gamma\vert$ and $A=2\log\sec\vert\Gamma\vert$. For 50-50, symmetric beam splitter, $\Gamma=i\pi/4$. The coefficients $c_n$ are those of the LTCS or UTCS depending on which state of the field is used in the input port of the beam-splitter.
A measure of entanglement for two-mode, pure states is the mixedness of the reduced density operator of one of the subsystems, and linear entropy $1-Tr[\tilde\rho_a^2]$, quantifies the mixedness[@bennett]. Hence, the entanglement between the output field modes is measured by the linear entropy of one of the output modes. The amount of entanglement introduced by the beam-splitter is a measure of nonclassicality of the input field[@asboth]. In Fig. \[fig:FigVI\], the linear entropy of the reduced density operator for the output from the beam-splitter is shown. With the UTCS and vacuum as the input fields, the linear entropy decreases as the number of Fock states in the UTCS increases. Higher the cutoff, the UTCS approximates the coherent state $\vert\alpha\ra$; consequently, the output fields are less entangled. The entanglement potential of the LTCS increases as the cutoff $N$ increases, a feature shared by the PACS as well[@usha]. As $N$ increases, the LTCS is essentially the number state $\vert N\ra$, as evident from the large $N$ limit of Eq. \[bernoulli\] and the entanglement potential of the number state increases with $N$. Since the nonclassicality of the UTCS decreases with $N$ and that of the LTCS increases, there exists a value of the cutoff parameter at which the states have nearly the same entanglement potential. If amplitude $\alpha=\sqrt{10}$, the entanglement potential of the two states are nearly equal if the cut-off parameter $N=4$.
Recently, the entanglement potential of the PACS has been studied in detail[@pacsep]. Since the LTCS and PACS of equal order are defined on the same sector of the Hilbert space, it is of interest to compare the entanglement potential of these states. The expression for the reduced density operator given in Eq. \[rden\] is used with the coefficients $c_n$ being those of the PACS given in Eq. \[pacs\]. In Fig. \[fig:FigVII\], the entanglement potential of the LTCS of order $N$ is compared with the PACS of same order for $N=$ 1 2,3 and 4. It is seen that the LTCS has higher entanglement potential compared to the PACS. This, in turn, means that the LTCS are better suited for generating entangled output modes in a beam-splitter than the PACS. Currently, a comparison of the two states in the context of quantum teleportation is being investigated, the results of which will be presented elsewhere.
Summary
=======
The states obtained by removing a set of contiguous, low energy Fock states from the coherent states of the electromagnetic field are non-Gaussian and nonclassical. These truncated coherent states possess equal uncertainties in the two quadratures, and do not exhibit squeezing. Additionally, if the cutoff parameter $N$ is less than $\vert\alpha\vert^2$, the product of the uncertainties attains its minimum, like that of the coherent states. In this approximation, these states are minimum uncertainty states whose average photon number is large (because $\vert\alpha\vert$ is large), but the photon statistics is sub-Poissonian. These states provide a resolution of identity for the subspace spanned by the Fock states $\vert N+1\ra,~\vert N+2\ra,~\vert N+3\ra, \cdots$. The noncalssicality, quantified by the entanglement potential, of the truncated states increases as the cutoff $N$ increases. This is related to the fact that if $N>\vert\alpha\vert^2$, the states are well approximated by a superposition the successive Fock states $\vert N+1\ra$ and $\vert N+2\ra$, which is the generalized Bernoulli state. Further, the entanglement potential of the truncated coherent states is higher than that of the photon-added coherent states of same amplitude and order. The photon-added coherent states and the lower truncated coherent states are the limiting cases of a suitably deformed photon-added coherent state, wherein the deformation is effected by a function of the number operator. The photon-added coherent states exhibit squeezing which is absent in truncated coherent states. The states which interpolate between these two liming cases show squeezing. Further, the entanglement potential of truncated coherent states is always higher than that of the photon-added coherent states. The Jaynes-Cummings hamilotnian without rotating wave approximation is the type of interaction required to produce these interpolating states.
[40]{}
![Uncertainty in $X$-quadrature as a function of the parameter $k$ for states $\vert\alpha, N\ra_k$, with $N=2$ (continuous), $N=3$ (line with dots) and $N=5$ (dashed). In all cases $\alpha=\sqrt{10}$. The limiting cases corresponding to $k=0$ and 1 are the PACS and LTCS respectively. []{data-label="fig:FigI"}](Fig_IR.eps){height="8cm" width="9cm"}
![Uncertainty profiles of $X-$ and $P-$quadratures for the UTCS and LTCS as a function of the cutoff number $N$. For the UTCS, $\Delta X$ ($\times$) and $\Delta P$ (.)are shown; for LTCS $\Delta X$(\*) is shown, which is also the profile for $\Delta P$. In all cases $\alpha=\sqrt{10}$.[]{data-label="fig:FigIII"}](Fig_IIIR.eps){height="8cm" width="10cm"}
![Phase distribution $P_u(\theta)$ for the UTCS for two different values of the cutoff parameter, $N=2$ (continuous), $N=10$ (dashed) and $N=20$ (dot-dash). The amplitude $\alpha=\sqrt{10}$.[]{data-label="fig:FigIV"}](Fig_IVR.eps){height="8cm" width="10cm"}
![Phase distribution $P_l(\theta)$ for the LTCS for two different values of the cutoff parameter, $N=2$ (continuous), $N=10$ (dashed) and $N=20$(dot-dash). The amplitude $\alpha=\sqrt{10}$.[]{data-label="fig:FigV"}](Fig_VR.eps){height="8cm" width="10cm"}
![$Q$-parameter as a function of the cut-off parameter $N$ for the LTCS of amplitude $\alpha=\sqrt{10}$.[]{data-label="fig:FigVA"}](Fig_VAR.eps){height="8cm" width="10cm"}
![Entanglement potential as a function of the cut-off $N$ for the UTCS (dots) and LTCS (star) of amplitude $\alpha=\sqrt{10}$.[]{data-label="fig:FigVI"}](Fig_VIR.eps){height="8cm" width="10cm"}
![Entanglement potential as a function of $\vert\alpha\vert$ for the PACS(continuous) and LTCS (dash). Figs. (a)-(d) correspond to the lower cutoff values $N$=1,2,3 and 4 respectively. []{data-label="fig:FigVII"}](Fig_VIIR.eps){height="8cm" width="10cm"}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the non-equilibrium behavior of optically driven dissipative coupled resonator arrays. Assuming each resonator is coupled with a two-level system via a Jaynes-Cummings interaction, we calculate the many-body steady state behavior of the system under coherent pumping and dissipation. We propose and analyze the many-body phases using experimentally accessible quantities such as the total excitation number, the emitted photon spectra and photon coherence functions for different parameter regimes. In parallel, we also compare and contrast the expected behavior of this system assuming the local nonlinearity in the cavities is generated by a generic Kerr effect rather than a Jaynes-Cummings interaction. We find that the behavior of the experimentally accessible observables produced by the two models differs for realistic regimes of interactions even when the corresponding nonlinearities are of similar strength. We analyze in detail the extra features available in the Jaynes-Cummings-Hubbard (JCH) model originating from the mixed nature of the excitations and investigate the regimes where the Kerr approximation would faithfully match the JCH physics. We find that the latter is true for values of the light-matter coupling and losses beyond the reach of current technology. Throughout the study we operate in the weak pumping, fully quantum mechanical regime where approaches such as mean field theory fail, and instead use a combination of quantum trajectories and the time evolving block decimation algorithm to compute the relevant steady state observables. In our study we have assumed small to medium size arrays (from 3 up to 16 sites) and values of the ratio of coupling to dissipation rate $g/\gamma \sim 20$ which makes our results implementable with current designs in Circuit QED and with near future photonic crystal set ups.'
address:
- '$^1$ Clarendon Laboratory, University of Oxford, Parks Road, Oxford OX1 3PU, UK'
- '$^2$ Centre for Quantum Technologies, National University of Singapore, 2 Science Drive 3, Singapore 117542'
- '$^3$ Science Department, Technical University of Crete, Chania, Crete, Greece, 73100'
author:
- 'T Grujic$^1$, S R Clark$^{2,1}$, D Jaksch$^{1,2}$ and D G Angelakis$^{2,3}$'
bibliography:
- 'mybib\_cavities.bib'
title: 'Non-equilibrium many-body effects in driven nonlinear resonator arrays'
---
Introduction {#sec:introduction}
============
Coupled resonator arrays (CRAs) interacting with single two-level systems embedded in each resonator have recently emerged as an exciting new platform for the realization of novel quantum many-body effects. CRAs may offer several features complementary to those of the successful and well-established ‘toolbox’ of cold atoms in optical lattices [@fisher1989boson; @jaksch1998cold], such as single-site addressibility and intrinsic non-equilibrium physics. This has led to proposals to realize a variety of phenomena of great interest in condensed matter physics such as Mott transitions [@angelakis2007photon; @hartmann2006strongly; @greentree2006quantum], effective spin models [@kay2008reproducing] and fractional quantum Hall states [@cho2008fractional], among others [@hartmann2008quantum; @tomadin2010many]. Though promising efforts are underway to build the first CRA systems, a number of technical challenges originating from the existence of strong dissipation must be overcome before equilibrium physics can be explored. On the other hand, the natural open nature of CRAs and the inherent ability to address and observe single resonators make this system an ideal test-bed to study many-body quantum lattice models out of equilibrium beyond the canonical Bose-Hubbard (BH) model and its realizations in optical lattices.
In this work we look for signatures of underlying many-body phenomena in CRAs by analyzing observables measured in the system’s non-equilibrium steady state (NESS). These include optically accessible observables like photon spectra and correlation functions. We focus on small to medium size CRAs of a few sites implementable with current or near future experimental technologies. We study coupled single mode resonators each interacting with a two-level system, a setup known now as the Jaynes-Cummings-Hubbard (JCH) model [@angelakis2007photon]. This simple model, in contrast to proposals involving multi-level atomic systems and external fields [@hartmann2006strongly], may be realized in a variety of technologies ranging from quantum dots [@badolato2005deterministic; @hennessy2007quantum] embedded in coupled defects in photonic crystals [@busch2007periodic; @notomi2008large] to coupled superconducting transmission line resonators [@wallraff2004strong] interacting with superconducting qubits [@blais2004cavity; @astafiev2010resonance; @frunzio2005fabrication]. In our study we have assumed small to medium size arrays (from 3 up to 16 sites) and values of the ratio of coupling to dissipation rate $g/\gamma \sim 20$ which makes our results implementable with current designs in Circuit QED [@houck2012chip] and with near future set ups involving fiber coupled cavities [@lepert2011arrays] and photonic crystals [@notomi2010manipulating].
The JCH as implemented in CRAs (with only the photonic excitations allowed to hop between neighboring cavities) motivated parallels with the predictions of the BH model. The latter naturally emerges in CRAs when one assumes generic nonlinear resonator effects instead of a Jaynes-Cummings interaction [@carusotto2009fermionized; @ssn2010signatures; @hartmann2010polariton; @gerace2009quantum]. Indeed, the JCH and BH share many similarities, both describing bosons hopping coherently between nearest neighbor sites, with local nonlinearities. An equilibrium quantum phase transition between Mott-insulating and superfluid-like phases exists for both models. However, the BH Hamiltonian involves only a single species of bosons (photons in this case), while the excitations of the JCH model have both photonic and atomic components. Consequently, the equilibrium physics of the JCH model is expected to be richer as shown in [@koch2009superfluid; @rossini2007mott; @mering2009analytic; @quach2009band; @makin2008quantum]. Studying the JCH out of equilibrium, as naturally implemented in open driven CRAs, is likely to highlight even more interesting differences with novel features beyond the realm of the driven BH model. In addition, most existing work to date in out of equilbrium CRAs has used mean-field theory to treat the system [@tomadin2009non; @ssn2010signatures; @liu2011quantum; @diehl2008quantum; @diehl2010dynamical; @tomadin2011nonequilibrium]. Going beyond mean-field theory is crucial to provide physically accurate insights, especially in the experimentally realistic few-resonator regime. This requires a faithful representation of the full Liouville space of the resonator system which in most cases beyond two resonators becomes challenging. For similar reasons of complexity, existing works to date on non-equilibrium resonator arrays exploring correlations and anti-bunching effects were always limited to minimal systems of two resonators [@ferretti2010photon; @liew2010single; @bamba2011origin; @angelakis2009steady; @angelakis2010coherent]. More recent work includes a study of the fluorescence spectrum of again two coupled Jaynes-Cummings resonators [@knap2011emission; @nissen2012non]. The possibility of simulating gauge fields in driven dissipative Jaynes-Cummings arrays was investigated in [@nunnenkamp2011synthetic], and artificial gauge fields in multi-resonator arrays in Bragg reflector micro-cavities assuming a Kerr interaction have also been studied recently [@umucalılar2011artificial].
In the present work we simulate CRAs beyond the two resonator regime, always assuming the full JCH model, by exploiting a combination of the matrix product state time-evolving block decimation (TEBD) algorithm [@vidal2003efficient; @vidal2004efficient] and quantum trajectories. We propose and analyze the many-body phases using experimentally accessible quantities such as the total excitation number, the emitted photon spectra and photon coherence functions for different, experimentally feasible, parameter regimes. In parallel, we also compare and contrast the expected behavior of this system when assuming that the local nonlinearity in the cavities is generated by a Kerr effect rather than a Jaynes-Cummings interaction. We find that the physics predicted by the two models differs for realistic regimes of interactions even if one meaningfully matches the respective strengths of the model nonlinearities. We analyze in detail the extra features of the JCH model originating from the mixed nature of the excitations and investigate the regimes where the Kerr description would faithfully match the JCH predictions. We find that the latter is only possible for values of the light-matter coupling and losses beyond the reach of current technology.
The structure of this manuscript is as follows. In the first section we describe the system and the dissipation/driving mechanisms. In Section 2 we discuss the local eigenstates and outline a scheme for mapping the effective nonlinearities between the models. In Section 3 we investigate the steady state spectra of single driven dissipative resonators, and demonstrate that Kerr physics can be achieved as a limiting case of the Jaynes-Cummings model. Section 4 treats arrays of a few sites as could be implemented in near future experimental setups, and results for yet larger arrays are presented in Section 5. In Section \[sec:novel\_effects\] we investigate the behavior of the system in the strongly repulsive regime and look for signatures of photon fermionization and crystallization in the JCH model. We then conclude in Section \[sec:discussion\].
The system {#sec:models}
==========
\
We study $M$ coupled single-mode resonators (indexed by $j$) in a circular configuration, as shown schematically in Fig. \[fig:system\_schematics\] (a) for $M=3$. The resonator frequency $\omega_c$ is the same for each resonator $j$. Coherent photon hopping between nearest-neighbor oscillators at a rate $J$ arises as a result of an overlap between neighboring resonator modes $\langle j, j' \rangle $. A large lattice spacing, relative to optical lattice setups, means that external driving lasers with amplitude $\Omega_j$ and frequency $\omega_L$ can independently excite and probe each resonator $j$. Assuming each resonator is coherently interacting with a two-level system, the physics is well described by the JCH model as defined below. We will analyze the non-equilibrium system response and in parallel contrast the results with the case where a generic Kerr nonlinearity is assumed in place of the Jaynes-Cummings interaction as described by the well known BH model.
The local physics for both cases is captured by the two Hamiltonians [^1] $$\begin{aligned}
\hat{h}'_{JC} & = & \omega_c \hat{a}^\dag \hat{a} + \left ( \omega_c - \Delta \right ) \hat{\sigma}^+ \hat{\sigma}^- + g \left ( \hat{a}^\dag \hat{\sigma}^- + \hat{a} \hat{\sigma}^+ \right ),\\
\hat{h}'_{BH} & = & \omega_c \hat{a}^\dag \hat{a} + \frac{U}{2} \hat{a}^\dag \hat{a}^\dag \hat{a} \hat{a},\end{aligned}$$ where we have set $\hbar=1$. The difference between the resonator frequency, and the atomic transition frequency is denoted by $\Delta$ and illustrated in Fig. \[fig:system\_schematics\] (b). The operator $\hat{a}^\dag$ creates a photon, while the operators $\hat{\sigma}^\pm$ denote the usual raising and lowering operations between the ground and excited states of the two-level system.
The eigenstates of $\hat{h}'_{\rm JC}$ are known as ‘dressed’ states, or polaritons. They are mixed atom-photon excitations, which are also eigenstates of the total excitation number operator $\hat{\mathcal{N}} = \hat{\sigma}^+ \hat{\sigma}^- + \hat{a}^\dag \hat{a}$ with eigenvalue $n$. The ground state is $|0\rangle \equiv |g,0\rangle$, with energy $E_0 = 0$. For a given excitation number $n \ne 0$, there are two polaritonic modes, designated $|n, \pm \rangle$, with associated frequencies in the bare frame $\omega_n^\pm = n \omega_c - (\Delta / 2) \pm \sqrt{\left ( \Delta / 2 \right )^2 + n g^2}$. The eigenstates can be written in the bare resonator-atom basis as $|n,\pm\rangle = \alpha_n^\pm |e,n-1\rangle + \beta_n^\pm |g,n\rangle$, with $\alpha_n^\pm = (\chi_n \mp \Delta) / (\sqrt{2}\sqrt{\chi_n^2 \mp \Delta \chi_n})$, and $\beta_n^\pm = \pm 2 g \sqrt{n}/(\sqrt{2}\sqrt{\chi_n^2 \mp \Delta \chi_n})$ [@barnett2003methods]. Here $\chi_n = \sqrt{\Delta^2 + 4 n g^2}$. It can be seen from these coefficients that setting a larger positive $\Delta$ results in an enhanced atomic component of the ‘-’ polaritons, while negative $\Delta$ increases the photonic contribution, with the relative composition of polaritons reversed for the ‘+’ species. The increased splitting of the polaritonic frequencies from the bare resonator frequency with increasing $n$ gives rise to an effective interaction between incoming photons, leading to a well known phenomenon in Cavity QED, the photon blockade [@birnbaum2005photon]. Figure \[fig:system\_schematics\] (c) shows the eigen-structure schematically for the particular case $\Delta/g = 0$, illustrating the energy mismatch between adjacent polariton manifolds. The resonance condition for an $n$-photon excitation of the $|n, \pm \rangle$ polaritonic mode in a single resonator from the ground state $|0 \rangle$ is to set the driving laser detuning $\Delta_c = \omega_L - \omega_c = \omega_n^\pm / n - \omega_c$ at $$\left ( \Delta_c \right )_n^\pm = \frac{1}{2 n} \left ( \Delta \pm \chi_n \right ).
\label{eq:driving_condition}$$
The local Hamiltonian for the driven BH model, $\hat{h}_{\rm BH}'$, describes a single-mode resonator with a Kerr-type nonlinearity in the particle number of strength $U$. The eigen-frequencies are $\omega_n = n \omega_c + \frac{U}{2} n (n-1)$, with corresponding eigenstates being the Fock number states $|n \rangle$. The eigen-structure of the three lowest-lying levels is shown in Fig. \[fig:system\_schematics\] (d). An $n$-photon excitation of the $n^{\rm th}$ mode occurs for the laser detuning $\Delta_c = \omega_L - \omega_c = \omega_n / n - \omega_c$ $$(\Delta_c)_n = \frac{U}{2}(n-1).
\label{eq:driving_condition_BH}$$
After transforming to a frame rotating at the driving laser frequency $\omega_L$ [@armen2006low], the Hamiltonian for both model systems can be written generically as $$\hat{H}_X = \sum_{j=1}^M \hat{h}_X^{(j)} + \sum_{j=1}^M \Omega_j \left ( \hat{a}_j^\dag + \hat{a}_j \right ) - J \sum_{<j,j'>} \hat{a}_j^\dag \hat{a}_{j'}.$$ The first term describes the local physics in resonator $j$ in the rotating frame, with $X \in \{JCH, BH\}$. The second term describes the action of a coherent driving laser on each resonator $j$, and the last term describes the photon hopping. The local contributions $\hat{h}_{JCH}^{(j)},\hat{h}_{BH}^{(j)}$ are identical in form to $\hat{h}_{JC}',\hat{h}_{BH}'$, with the bare cavity frequency replaced by the detuning $\omega_c \rightarrow -\Delta_c$.
We now describe the manner in which we compare the two models. As the two Hamiltonians are patently different, care must be taken when comparing their behavior. We map the Jaynes-Cummings nonlinearity to an effective Kerr interaction via a frequency mismatch argument, as considered in both equilibrium and non-equilibrium contexts [@makin2008quantum; @hãžmmer2012non]. In the following we focus on the ‘-’ species of polaritons, as the driving laser frequency necessary to resonantly excite the $|n, - \rangle$ mode increases with $n$, qualitatively similar to a Kerr nonlinearity. We define the effective Kerr nonlinearity $U_{\rm eff} = U_{\rm eff}(g, \Delta)$ as the energy penalty incurred in forming a two-particle polaritonic excitation in a resonator (with energy $\omega_2^-$) from two one-particle polaritons in neighboring resonators (with total energy $2 \omega_1^-$) $$\frac{U_{\rm eff}}{g} = \frac{\omega_2^- - 2 \omega_1^-}{g} = \frac{\Delta}{2g} + 2 \sqrt{\left ( \frac{\Delta}{2g} \right )^2 + 1} - \sqrt{\left ( \frac{\Delta}{2g} \right )^2 + 2} .
\label{eq:eff_U}$$ Figure \[fig:system\_schematics\] (c) shows these low-lying polaritonic eigenstates and the definition of $U_{\rm eff}$ pictorially, while Fig. \[fig:system\_schematics\] (d) shows the level structure for a Kerr-nonlinear resonator. Interestingly, an analogous definition for the ‘+’ polaritonic branch yields an attractive Kerr-type interaction. We note that as the distribution of energy levels with particle number $n$ is markedly different between the models, it is only sensible to compare them in the very weakly excited regime where the above definition is meaningful.
While the above mapping ‘matches’ the nonlinearities of the two models by considering transitions between the one- and two-particle manifolds, we expect the single particle resonances of both systems to occur at different spectral locations. The dominant spectral features in weakly excited systems will therefore be observed for different driving laser parameter regimes. Additionally, the mapping involves a comparison only of diagonal elements of the two governing Hamiltonians. However the different natures of the excitations of the models are not fully accounted for in such a mapping, in particular the additional internal degree of freedom possessed by the JCH model. We therefore expect the corresponding distinct off-diagonal Hamiltonian terms to lead to different physical observables between the models even under this mapping.
We assume a finite photon loss rate $\gamma_p$ from each resonator for both models, and that the spontaneous emission rate from the excited level $|e \rangle$ of the two-level systems in the JCH system is negligible, $\gamma_a = 0$. Competition between coherent resonator driving, and photon loss leads to NESS conditions. We employ a master equation formalism [@carmichael1991open] to describe the evolution of the system’s density matrix $\rho(t)$. The NESS density matrix $\rho_{\rm ss}$ is given by the stationary point of the master equation $\dot{\rho} = \mathcal{L}[\rho] = 0$, where the action of the Liouville super-operator is defined through $$\mathcal{L}_X [\rho_{\rm ss}] = \frac{1}{i} [\hat{H}_X, \rho_{\rm ss}] + \sum_{j=1}^M \frac{\gamma_p}{2} \left ( 2 \hat{a}_j \rho_{\rm ss} \hat{a}^\dag_j - [\hat{a}^\dag_j \hat{a}_j, \rho_{\rm ss}]_+ \right ) = 0,
\label{eq:master_equation}$$ with $[\cdot,\cdot]_+$ denoting the anti-commutator operation. In general, solving Eq. (\[eq:master\_equation\]) is a formidable task, owing to the exponential growth in the necessary size of a system’s description with the number of cavities $M$. We exploit the permutational symmetry of a homogeneous minimally sized three-site cyclic resonator system to enable solution via an exact diagonalization scheme, and employ a more sophisticated TEBD based stochastic unraveling of the master equation for larger systems. Details of the latter approach are provided in the Appendix. \[sec:single\_spectra\] Before discussing the solutions of Eq. (\[eq:master\_equation\]) for an array of cavities, we first revise and compare the stationary behavior of a single driven resonator with both Jaynes-Cummings and Kerr-type nonlinearities to illustrate their intrinsic differences in the local physics regime.
Local resonator physics: Kerr versus Jaynes-Cummings nonlinearities {#sec:local}
===================================================================
In Figures \[fig:example\_spectra\_single\_cavity\] (a) and (b) we show NESS particle numbers for both types of resonator as a function of the driving laser detuning from the bare cavity resonance. The most striking difference between the spectra is the existence of two ‘wings’ for the Jaynes-Cummings nonlinearity, symmetrically distributed about the cavity frequency for this special case $\Delta / g = 0$. The driving laser frequency necessary to excite the ‘-’ polaritons increases with the excitation number $n$ (the converse is true for the ‘+’ polaritons). Figure \[fig:example\_spectra\_single\_cavity\] (b) also demonstrates that the Kerr resonator always exhibits a single particle response at the bare cavity frequency, whereas the spectral location of the two single particle modes in a Jaynes-Cummings type system strongly depend on the atom-resonator coupling parameters.
The nonlinear response of the resonators can also be probed by analyzing the second order correlation function of the emitted photon from a certain resonator site. For later use, we define the generalized coherence function between any two resonators $(j,k)$ as $g^{(2)}(j,k) = \langle a_j^\dag a_k^\dag a_j a_k \rangle / \langle a^\dag_j a_j \rangle \langle a^\dag_k a_k \rangle$. In the following we refer to the on-site coherence function as $g^{(2)} \equiv g^{(2)}(j,j)$. As expected, in the strong coupling regime this quantity exhibits a dip below the coherent driving laser value $g^{(2)} = 1$ leading to photon anti-bunching when the driving laser is tuned to the corresponding single-particle modes for both models. For the Jaynes-Cummings resonator, the strongest anti-bunching is expected for laser detunings $\Delta_c = \pm g$ from the bare cavity frequency (for $\Delta = 0$). In contrast, the Kerr resonator demonstrates strongest anti-bunching when the driving laser is on resonance with the bare cavity mode, coinciding with the single particle mode at $\Delta_c = 0$. The reasonable atom-cavity coupling and cavity loss rates we have chosen give $g/\gamma=20$ (shown in Fig. \[fig:example\_spectra\_single\_cavity\]), causing the correlation function to dip to a minimum $g^{(2)} \approx 0.25$ for the Jaynes-Cummings resonator, and to $g^{(2)} \approx 0.15$ for the Kerr resonator.
Dips in the coherence function also appear when the driving laser is resonant with higher underlying quantum resonances (at equally spaced laser frequency intervals $U/2$ for the Kerr resonator, and laser frequencies $\Delta_c = \pm \frac{g}{\sqrt{n}}$ for the Jaynes-Cummings system). The magnitude of the correlation function at these resonances is highly sensitive to the magnitude of the driving laser strength and cavity loss rate.
\
Figures \[fig:example\_spectra\_single\_cavity\] (a) and (b) share several features common to driven dissipative nonlinear quantum systems. We see that, for both models spectral peaks corresponding to higher excitations $n$ become successively narrower, because the monochromatic driving laser cannot be simultaneously resonant with all intermediate levels. This means all but the lowest (polaritonic or Fock) peaks involve off-resonant transitions. The peaks also become less bright with increasing $n$ because in addition to being more difficult to populate, modes with more photons are more susceptible to photon loss. The Jaynes-Cummings system involves both atomic and photonic species, leading to different peak intensities between the models. This follows because we are only driving the photonic degree of freedom which in turn is coupled to the atomic degree of freedom in the Jaynes-Cummings resonator.
Thus the spectral signatures of driven dissipative JCH and BH systems for the case of vanishing photon hopping (single resonators) are in general quite different. However, the two models do possess broad qualitative similarities in that they both describe the interplay of coherent bosonic hopping with an on-site nonlinearity. Indeed, the BH Hamiltonian has been widely used as an approximation to treat CRAs in several works assuming a generic nonlinearity [@carusotto2009fermionized; @ssn2010signatures; @hartmann2010polariton; @gerace2009quantum]. It is natural to ask if the two models are equivalent in some regime or even if the physics of the JCH Hamiltonian can be mapped on to an effective BH model. Such a mapping would enable the large body of existing knowledge about the BH model to be applied directly to CRAs and would additionally significantly simplify numerical simulation. As the nonlinearity acts locally inside each resonator, we address this question below by first investigating single resonators.
*The photonic limit of the Jaynes-Cummings model* ($\Delta \ll -g$) - Intuitively, we expect agreement between the models as the polaritonic JCH excitations are made more photonic in nature. For the ‘-’ polariton species, this means setting $- g / \Delta\ll 1$. The two-level system in each resonator is then barely excited and the dispersive interaction available is capable of inducing a significant effective photon repulsion only for very large values of the coupling $g / \gamma_p \gg 1$. In this limit a Taylor expansion of the polaritonic eigen-energies $\omega_n^\pm$ in the small quantity $|g / \Delta|$ leads to frequencies quadratic in the particle number $n \ne 0$ as $\omega_{n,-}(g,\Delta) \approx n \left ( \omega_c + g^2 / \Delta \right ) + g (g / |\Delta|)^3 n (n-1)$. We see that the off-resonant interaction induces an effective shift in the bare resonator frequency $\Delta_{\rm shift} = - g^2 / |\Delta|$, and that the energy spectrum can be written in the canonical Kerr nonlinear form with effective repulsion $U_{\rm BH}^{\rm approx} \equiv 2 g \left (g / |\Delta| \right )^3$.
Figure \[fig:example\_spectra\_single\_cavity\] (c) shows the spectrum for a far detuned ($\Delta / g = -10$) driven dissipative Jaynes-Cummings resonator for the same experimentally realistic atom-cavity coupling rate $g / \gamma_p$ as in Fig. \[fig:example\_spectra\_single\_cavity\] (a). We see that the response is essentially that of an empty (linear) driven dissipative resonator, albeit with a peak response at $\Delta_{\rm shift}$ from the bare resonator frequency. The atomic excitation is an order of magnitude smaller, as expected. The effective Kerr nonlinearity is much smaller than the line-width ($U_{\rm BH}^{\rm approx} / \gamma_p = \frac{1}{50}$), so that nonlinear effects barely show up, apart from a slight asymmetry in the response. To observe sizable nonlinear effects for realistically achievable resonator parameters, we see that it is instead necessary to operate near the resonance point $\Delta / g \approx 0$.
For comparison, the physically unrealistic ultra-strong atom-resonator coupling limit is shown in Fig. \[fig:example\_spectra\_single\_cavity\] (d). The response of the Kerr-nonlinear resonator shown in Fig. \[fig:example\_spectra\_single\_cavity\] (b) is reproduced, and compared with the spectrum of a Jaynes-Cummings resonator again operating in the photonic regime ($\Delta / g = -10$). The coupling strength $g / \gamma_p > 10^4$ is chosen by setting $U_{\rm BH}^{\rm approx} = U$ in the above definition of the effective Kerr nonlinearity. We see good agreement between the photon number spectra, which worsens for higher excitation peaks as higher order terms in the Taylor expansion of the polaritonic eigen-frequencies become important. So while deep in the regime $- g / \Delta \ll 1$ the Jaynes-Cummings model can, with reasonable accuracy, be mapped on to a Kerr system, this regime is inaccessible with current technology.
*The atomic limit of the Jaynes-Cummings model* ($\Delta \gg g$) - In the opposite limit of $\Delta / g \gg 1$, the atomic component of the ‘-’ polaritons is maximal. A Taylor expansion of the eigen-frequencies now yields: $\omega_{n,-}(g,\Delta) \approx n \left ( \omega_c - g^2 / \Delta \right ) + g (g / \Delta)^3 n (n-1) - \Delta$, for $n \ne 0$. The ground state $|g,0\rangle$ still lies at the zero of energy. Therefore we see that to first order in $g / \Delta$, the spectrum becomes equally spaced with the same shift in the resonator frequency as for the limit $\Delta \ll -g$, with the exception of the interval $\omega_c - \Delta$ between the ground state and the first polaritonic mode $|1,-\rangle$ . In this limit $|1,-\rangle \approx |e,0\rangle$, with a small photonic component allowing transitions by the external driving laser. This may be summarized by the following effective Hamiltonian describing the ‘vacuum shifted’ $|n,-\rangle$ ladder of states for a single resonator in the atomic limit: $$\hat{h}_{g / \Delta \ll 1} = (\omega_c - g^2 / \Delta) \hat{a}^\dag \hat{a} + g \left ( \frac{g}{\Delta} \right )^3 \hat{a}^\dag \hat{a}^\dag \hat{a} \hat{a} + \Delta |g,0 \rangle \langle g,0| - \Delta.$$ Thus, population of the two-particle state from the resonantly driven single-particle mode is strongly inhibited. Higher lying levels could, however, be near resonantly populated from the one-particle mode given an additional driving laser with frequency $\omega_L = \omega_c$.
When strictly working within the low-excitation regime, it is then appropriate to assign a large effective nonlinearity describing this energy penalty to reach the two-excitation manifold. However, care must be taken to distinguish between setups such as we consider, where only the lowest lying excitations are directly probed, and others which access the approximately harmonic ladder at larger $n>2$.
Many-body signatures in steady state observables {#sec:finite_specta}
================================================
Moving beyond the single resonator regime we will first analyze a minimal cyclic nonlinear CRA of $M=3$ cavities. This case, though not truly many-body is interesting as this is where the first experimental implementations are likely to begin [@houck2012chip; @lepert2011arrays; @notomi2010manipulating]. We consider a finite coherent photon tunneling $J \ne 0$ between adjacent cavities which splits the local polaritonic resonances into delocalized global modes. For the moment we drive our system homogeneously so that all external lasers are in-phase with $\Omega_j = \Omega, \forall j$.
\
We obtain the NESS by diagonalizing the super-operator $\mathcal{L}$ after exploiting the permutational symmetry of the system to significantly decrease the number of unique density matrix elements and additionally retaingin only basis states allowing a maximum of $P = 4$ excitations in the system. [^2]
Figures \[fig:example\_spectra\_3\_cavities\] (a) and (b) show generalizations of the single resonator spectra discussed above for the relatively small rate of photon hopping $J / \gamma_p = 1$. We expect the response of the systems to strongly resemble the ‘local’ physics in this regime, as the photon hopping is essentially a weak perturbation on top of the atom-cavity coupling.
The homogeneous nature of our chosen driving means that the total momentum of excitations in the NESS must be zero. The one-particle excitation in the BH system is then just the zero-momentum free-particle Bloch mode $|k \!\!= \!\!0 \rangle = \frac{1}{\sqrt{M}} \sum_{j=1}^M \hat{a}^\dag_j |0 \rangle$ excited at a laser detuning $(\Delta_c)_{|k=0>} = -2 J$, as the Kerr nonlinearity does not influence a single particle. In contrast a single excitation in the JCH system can be shared between atomic and photonic degrees of freedom leading to two delocalized generalizations of the $|1, \pm \rangle$ polaritons, denoted $|k \!= \!0, \pm \rangle = A^\pm |E\rangle + B^\pm |k \!\!= \!\!0 \rangle$. Here $|E \rangle = \frac{1}{\sqrt{M}}\sum_{j=1}^M \hat{\sigma}^+_j |0 \rangle$ is a delocalized atomic excitation. The coefficients $A^\pm$ and $B^\pm$ are identical in form to the coefficients $\alpha_1^\pm$, $\beta_1^\pm$ of the localized polaritons, after making the replacement $\Delta \rightarrow \Delta_{k=0} = \Delta - 2 J$. These are resonantly excited at laser detunings $(\Delta_c)_{|k=0, \pm \rangle} = -2 J - \Delta_{k=0}/2 \pm \chi_{k=0}^{(1)}$, where $\chi_{k=0}^{(1)} = \sqrt{g^2 + \Delta_{k=0}^2/4}$. As only the cavities are directly coupled there is now an additional asymmetry between photons and atoms. The resonance point for delocalized polaritons is shifted to $\Delta = 2J$ from $\Delta=0$ found in the single resonator case. This is equivalent to setting the transition frequency of the atoms equal to the frequency of the lowest lying Bloch mode $|k \!\!= \!\!0 \rangle$.
Relative to Figs. \[fig:example\_spectra\_single\_cavity\] (a) and (b), we see new multi-particle modes appear between the delocalized generalizations of the dressed state resonances. By an $N$-particle mode we mean an eigenstate of the total excitation number operator for the whole system $\hat{\mathcal{N}} = \sum_{j=1}^M \hat{\mathcal{N}}_j$ with eigenvalue $N$. For this particularly small photon hopping rate, equal to the cavity line-width, the delocalized single particle mode of the BH system is smeared into a broad hump along with delocalized two- and three- particle modes, while new features can also be discerned in the JCH spectra. Additionally, new modes appear in the JCH model at approximately the bare cavity frequency. These modes are symmetrized superpositions of ‘+’ and ‘-’ polaritons, and for our chosen driving strength are barely populated relative to the response of the BH model at the bare cavity frequency.
There is an asymmetry between the delocalized generalizations of the ‘-’ and ‘+’ wings of the JCH resonator array spectrum, despite setting $\Delta = 2J$. We note that for large photon hopping $J \gg g$ and $\Delta = 2J$, the resonator array spectrum is again symmetric about the Bloch mode frequency (not shown), though still with quantitative differences to the single resonator case. It is in the intermediate regime ($J \approx g$) that asymmetries appear, as hopping brings multiple particle global excitations into the same spectral region. A signature of this hopping-induced behavior that is particularly amenable to experimental verification is a measurement of the photon correlation at the two delocalized single-particle resonance frequencies. For this particular driving and cavity loss rate, the values are $g^{(2)} \approx 0.18, 0.35$ for the $|k=0, \pm \rangle$ modes respectively. Meanwhile $g^{(2)}$ reaches a minimum of $\approx 0.26$ at the underlying single-particle mode of the BH system (indicated by the vertical dash-dotted line in Fig. \[fig:example\_spectra\_3\_cavities\] (b)). We note here again that if one assumes losses smaller than $g/\gamma_p=20$ used here then the correlation minimum approaches zero and clear anti-bunching should be achieved, as expected.
The spectral response of far-detuned finite-size CRAs operating in the photonic limit ($-g / \Delta \ll 1$) is shown in Figs. \[fig:example\_spectra\_3\_cavities\] (c) and (d). These spectra are direct generalizations of those shown in Fig. \[fig:example\_spectra\_single\_cavity\]. Again, we see that for our assumed atom-resonator couplings (i.e. $g / \gamma_p \approx 20$), the main effect of the atomic degree of freedom is to shift the free-particle mode by $\Delta_{\rm shift}$, in this case from the lowest Bloch mode. Figs. \[fig:example\_spectra\_3\_cavities\] (d) shows that (unrealistically) larger couplings ‘matched’ to a specific Kerr nonlinearity can reproduce the spectral features of a driven dissipative finite CRA with reasonable accuracy.
Larger resonator arrays
=======================
So far we have investigated few sites JCH systems of a few sites operating either on-resonance (i.e. $\Delta / g = 0$), or in the photonic limit where the weak nonlinearity is captured by an effective BH interaction. We now demonstrate the effect of the changing nature of the polaritonic excitations and also investigate larger arrays of $M=16$ sites as this parameter is varied about the strong-interaction regime $\Delta / g \approx 0$. Rather than construct a spectrum by varying the driving laser frequency, we instead selectively drive a particular spectral feature and vary $\Delta / g$. Specifically, in Fig. \[fig:8\_site\_system\_NESS\_expvals\] we show trajectory results for NESS particle numbers for larger ($M=16$) JCH and BH arrays driven at their single particle resonances, corresponding to the driving laser detunings $\Delta_c = (\Delta_c)_{|k=0,-\rangle}$ and $\Delta_c = -2J$ respectively.
As expected, we see the changing nature of the system’s excitations reflected in a larger atomic (and smaller photonic) occupation for increasing $\Delta / g$. For strong positive detunings, the excitations are predominantly atomic, indirectly excited via the resonator field. As we have assumed lossless atoms, the steady state corresponds to oscillations between the ground and excited states, leading to an average half atomic occupancy. Under the mapping of Eq. (\[eq:eff\_U\]), this regime corresponds to a BH system with large Kerr coefficient. Under the stronger driving we have chosen for this calculation, the BH system approximately oscillates coherently between zero and one photons per resonator, with occupation of higher Fock levels suppressed. Thus, while the total excitation number for the JCH asymptotically agrees with the NESS photon number for the BH in this limit, the underlying physics is very different.
![NESS particle numbers per resonator for sixteen-site Bose-Hubbard and Jaynes-Cummings CRAs driven at their single particle resonances. Parameters: photonic driving $\Omega / \gamma_p = 2$, atom-resonator coupling strength $g / \gamma_p = 20$, photon hopping rate $J / \gamma_p = 1$. Trajectory calculations retain $p = 6$ photons per resonator for the BH simulation, and $p=5$ for the JCH. Each Jaynes-Cummings detuning $\Delta / g$ is mapped to an effective Kerr nonlinearity via Eq. (\[eq:eff\_U\]) to enable a comparison. []{data-label="fig:8_site_system_NESS_expvals"}](fig_4.pdf)
In the opposite limit $\Delta / g < 0$, the excitations of the JCH model become more photonic, and the effective Kerr nonlinearity decreases. We have already seen that agreement between the JCH and Kerr photon number is reached far into this limit. In the intermediate regime, $\Delta / g \approx 0$, the total particle number for the JCH exhibits significant departures from the BH results. This is due both to the fact that only one component of the polaritons is driven in the JCH system, and also that this stronger driving allows access to higher-lying states ($n>2$) outside the regime of validity for the effective Kerr strength definition. We therefore see non-trivial differences between the models both in spectra measured as a function of driving laser frequency, and as a function of the nonlinearity.
The strong nonlinearity limit: fermionization and crystallization of photons in CRAs {#sec:novel_effects}
====================================================================================
Having demonstrated that the NESS in resonator arrays governed by the JCH and BH Hamiltonians are in general different we now evaluate two non-equilibrium effects in the large nonlinearity regime recently studied assuming a BH description, with proposals to experimentally realise such effects in Jaynes-Cummings type CRA systems.
*Fermionization in coupled resonator arrays -* Recently an exploration of the ultra-strong nonlinearity regime in a three-site driven dissipative Bose-Hubbard model was undertaken [@carusotto2009fermionized]. For a closed coherent system it was found that as the Kerr strength $U$ approached infinity double occupancy of any resonator was completely suppressed allowing the system’s bosonic wave-functions to be mapped to those of an equivalent fermionic system via a Jordan-Wigner transformation. In this interpretation double occupancy is prevented by fermionic statistics rather than a hard core bosonic interaction. The authors found that a cyclic system’s $N$-particle eigenfunctions and corresponding energies can then be identified uniquely for a given total particle momentum. Subjecting the system to coherent driving and a finite particle loss rate from each resonator results in readily classifiable peaks in the resonator occupancy as a function of the driving laser frequency.
Figure. \[fig:comparison\_of\_BH\_JCH\_fermionised\] (a) reproduces such a spectrum for a homogeneously driven ($\Omega_j = \Omega, \forall j$) minimal $M=3$ resonator BH system. For the spectral range shown the only possible modes that can be observed in the hard core limit NESS are the zero momentum one-particle mode, which for the BH model coincides with the Bloch mode $|k \!\!= \!\!0 \rangle$ of an empty resonator system, and a mode formed from two particles with opposite momenta $k = \pm 2 \pi / 3$. We show spectra for a strong nonlinearity $U / \gamma_p \gg 1$ (allowing for any number of photons per resonator in the calculation) and for the true fermionized limit $U \rightarrow \infty$ (when strictly $p=1$ photon per resonator is retained in calculations) illustrating the convergence. Also shown in Fig. \[fig:comparison\_of\_BH\_JCH\_fermionised\] (a) are NESS photon numbers for a strongly detuned JCH system operating in the ‘photonic regime’, again computationally retaining multiple and single photons per resonator. As expected from the discussion in Sec. \[sec:local\] good agreement with the BH results is observed, albeit for unrealistically large values of the light-matter coupling $g$.
\
Moving beyond from the photonic limit we now explore the novel physics of ultra-strong nonlinearity in JCH systems operating in the regime of strongest interaction between resonators and atoms, i.e. $\Delta / g \approx 0$. In this regime correspondence between the models is less clear and the fully polaritonic nature of excitations must be taken into account. To recover physics which resembles hard-core photons we operate in the limit $g / \gamma_p \gg 1$, along with $g \ll \omega_c$ to satisfy the rotating wave approximation, and drive at $\omega_L \approx \omega_1^-$ near the $|1,-\rangle$ polaritonic resonance of isolated resonators. Under these conditions the only relevant states in each resonator are the ground state and $|1,-\rangle$ polaritonic state.
Figure \[fig:comparison\_of\_BH\_JCH\_fermionised\] (b) shows hard-core JCH model spectra for two different atom-resonator detunings $\Delta = 0$ and $\Delta = g$. Increasing the coupling $g$ while holding the ratio $\Delta/ g$ fixed leads to a qualitatively similar evolution of spectral features as in the BH model. A two-particle excitation splits from the single particle resonance and asymptotically approaches a splitting that only depends on $J$ and $\Delta / g$. For the coupling $g/\gamma_p \sim 800$ used good agreement is seen in Fig. \[fig:comparison\_of\_BH\_JCH\_fermionised\] (b) between the full JCH model and the spectrum obtained when retaining only the $p=1$ photons per resonator necessary to describe the hard-core polariton limit. We see that even in the ultra-strong non-linearity limit of the JCH model the internal structure of polaritons (governed by the parameter $\Delta / g$) still plays a crucial role in determining the location and amplitude of spectral features. This behavior is a result of the effective hopping, driving and loss rates for polaritons being different from the bare photon parameters. Indeed the effective polaritonic hopping rate is $J_{\rm pol} = |\beta_1^-|^2 J$, the effective driving rate is $\Omega_{\rm pol} = (\beta_1^-) \Omega$, while the loss rate is $\gamma_{\rm pol} = |\beta_1^-|^2 \gamma_p$, where $\beta_1^-$ is the coefficient controlling the photonic component of the $|1,-\rangle$ polaritons defined in Sec. \[sec:models\]. This reflects that the hopping, driving and loss processes involve only the photonic component of the polaritons. Thus, on resonance ($\Delta / g = 0 \Rightarrow \beta_1^- = \frac{1}{\sqrt{2}}$) we see that the peak separation is $J_{\rm pol} = J/2$ as seen in Fig. \[fig:comparison\_of\_BH\_JCH\_fermionised\] (b).
Focusing on experimentally measurable photonic quantities, in Fig. \[fig:comparison\_of\_BH\_JCH\_fermionised\] (c) we track for both models as a function of their nonlinearity the photon density-density correlations measured at the two-particle peak on a single site via $g^{(2)}(j,j)$ and between neighboring sites via $g^{(2)}(j,j+1)$. The spectral location of the zero-momentum two-particle modes for the JCH and BH systems, are found using the results in Refs. [@wong2011two] and [@javanainen2010dimer] respectively. As outlined in Ref. [@carusotto2009fermionized] at small nonlinearities the two-particle peak resides within the one-particle spectral feature and so correlations inherit Poissonian statistics from the driving laser. Larger nonlinearities split the two-particle resonance from the one-particle peak, as shown in Figs. \[fig:comparison\_of\_BH\_JCH\_fermionised\] (a) and (b), and lead to strong bunching. For very strong nonlinearities on-site anti-bunching is expected as the two excitations are distributed in such a way that the pair of photons are never in the same resonator resulting in the nearest neighbor correlations becoming large. Such considerations, being consequences of generic nonlinear behavior, lead to qualitatively similar correlation functions for both the driven dissipative JCH and BH models. Thus fermionized photons are a feature of the on-resonance JCH model as well, once the different spectral frequencies for correlation measurements are taken into account.
*Polariton crystallization -* Another intriguing phenomenon of interacting photons in resonator systems, photon crystallization, was recently predicted to occur in a one dimensional ring of optical cavities with Kerr-type nonlinearity [@hartmann2010polariton]. Driving lasers with a phase difference of $\pi / 2$ between each site $k$, i.e. $\Omega_k = \Omega \exp (i k \pi / 2)$, create a flow of bosons around the system. The contact interaction energy $U$ was found to result in a ‘crystallization’ of particles as they flowed around the ring, even in the presence of dissipation. The signature of this effect was identified in the particle density-density correlations $g^{(2)}(j,k)$ between cavities $j$ and $k$, measured in the system’s steady state.
On-site anti-bunching is accompanied by nearest-neighbor density-density correlations stronger than correlations between more distant cavities. The conclusion drawn in Ref. [@hartmann2010polariton] was that particles form ‘dimers’ of light, which flow around the system. We demonstrate here that this dimerization can also be seen in a JCH system outside the photonic limit. The additional degree of freedom in the atom-resonator detuning $\Delta$ allows the strength of this effect to be adjusted on demand.
We study a system of $M = 16$ coupled Jaynes-Cummings resonators under periodic boundary conditions. In addition to the relative phases, we must choose the frequency of our driving lasers. For a BH-type system, a laser detuning $\Delta_c = 0$ directly drives the single particle $k = \pi / 2$ mode. For the JCH model we choose to drive the delocalized generalization of the single-particle polaritonic mode $|1, - \rangle_{k=\pi/2}$ to produce an analogous effect. Details of the parameters used for this numerically demanding calculation are given in the Appendix.
\
We see in Fig. \[fig:polariton\_crystallization\] (a) that the signatures of crystallization show up in our polaritonic system – there is a larger probability of finding photons in neighboring cavities than in cavities further apart. We have additional control over the atom-resonator detuning $\Delta$, which allows us to tune the nature of the system’s excitations either more photonic $\Delta / g < 0$, or atomic $\Delta / g > 0$. We see that the crystallization effect becomes weaker for larger $\Delta$ as the atoms and resonators are tuned away from resonance, while on-site anti-bunching is enhanced. This is a consequence of resonator photons being strongly coupled to the two-level systems, which hold most of the excitation for large $\Delta$, but longer-range correlations tend to unity. Figure \[fig:polariton\_crystallization\] (b) shows the actual photonic and atomic expectation values per resonator, for the three detunings considered illustrating the changing make-up and associated enhanced photon anti-bunching. Also shown for comparison in Figure \[fig:polariton\_crystallization\] are results for the driven dissipative Kerr system like that considered in Ref. [@hartmann2010polariton]. We conclude that the ‘dimerization’ predicted there for the BH model persists in CRA calculations using the full atom-resonator Hamiltonian without approximations, and that some degree of tunability should be observable by bringing the cavities in and out of resonance while maintaining significant resonator populations.
Discussion {#sec:discussion}
==========
We have proposed and analyzed excitation number and photon coherence spectra for small to medium sized weakly driven dissipative coupled resonator systems described by the Jaynes-Cummings-Hubbard Hamiltonian. Simulations were at all times performed using the full JCH Hamiltonian without approximations, using model parameters realizable by current state-of-the-art technology. We have also presented analogous spectra for resonator arrays governed by the Bose Hubbard model, drawing attention to the differences in experimentally accessible observables between the two models. These differences arise primarily because of the composite nature of the elementary excitations of the JCH Hamiltonian, leading to readily identifiable unique spectral characteristics. Therefore we conclude that when simulating coupled resonator arrays the full JCH Hamiltonian must be used in calculations to properly account for the underlying physics.
Generalizations of two bosonic interaction-induced non-equilibrium phenomena were observed in the NESS of CRAs modeled retaining the richer JCH physics, with additional tunability. The fingerprints of the polaritonic equivalents of ‘fermionic’ photons and photon crystallization exhibit subtle departures from the pure bosonic case, detectable via experimentally accessible spectral quantities.
Finally, this work has shown that the combination of the quantum trajectory method with the TEBD algorithm is an especially powerful tool in finding the NESS of open driven dissipative many-body systems. Such a numerical approach can be readily applied to calculate more complex out of equilibrium properties of coupled resonator arrays where an analytic approach is unfeasible. This might include important open problems such as determining the transport properties of linear coupled resonator arrays in the presence of photon loss, or exploring novel one-dimensional quantum states of light that are robust to experimentally realistic decay processes.
Appendix {#app:MPS_traj .unnumbered}
========
*Matrix product state quantum trajectory calculations -* To solve for NESS expectation values of larger driven dissipative CRAs we employ a stochastic unraveling of the time dependent quantum master equation. This involves propagating independent random wave-function trajectories through time rather than the (much larger) full density matrix [@dalibard1992wave; @dum1992monte]. We evolve $R$ stochastic wave-functions $|\Psi_i (t) \rangle$ under the action of a non-Hermitian Hamiltonian $\hat{H}^{\rm eff}_X = \hat{H}_X - i \frac{\gamma_p}{2} \sum_{j=1}^M \hat{a}^\dag_j \hat{a}_j$. The latter terms induce a decay in the wave-function normalization. In practice, quantum jumps corresponding to photon loss from one of the resonators are applied at times $t_{\rm jump}$ when the norm falls below a randomly chosen number $r \in [0,1]$. The particular resonator $j$ is selected by sampling from the probability distribution $P_j = || \hat{a}_j \Psi_i(t_{\rm jump}) \rangle||^2 / \left ( \sum_k || \hat{a}_k \Psi_i(t_{\rm jump}) \rangle||^2 \right )$. After the jump is applied, the wave-function is re-normalized and evolution continues.
The NESS density matrix $\rho_{\rm ss}$ is calculated by evolving this ensemble of wave-functions to times $t > t_{\rm trans}$ larger than the timescale over which transient dynamics die out, then averaging over realizations as $\bar{\rho}_{\rm ss} \approx (1 / R) \sum_i |\bar{\Psi}_i \rangle \langle \bar{\Psi}_i |$. Here, the over bars denote an additional average over time steps $t_i > t_{\rm trans}$ in the simulation. This important trick is possible due to the ergodicity of the unraveling of the master equation which means that in the NESS the stochastic wave-functions at each time step must average to the true steady state density matrix $\rho_{\rm ss}$ [@mejia2007heat; @michel2008transport; @daley2009atomic]. Hence the trajectory method is far more efficient at simulating non-equilibrium steady states than the transient dynamics. A single trajectory can yield estimates of steady state expectation values and multiple trajectory realizations give an indication of the statistical errors in calculated quantities. Yet another advantage of using a trajectory method here is that NESS expectation values for homogeneous systems can be further averaged over each site in the system and larger simulations gain considerable accuracy from this fact.
Time evolving the stochastic wave-functions is itself a nontrivial task, as the dimension of Hilbert space is still prohibitively large for a representation in the bare basis. We employ a matrix product state (MPS) [@perez2007matrix] ansatz to compress our description of the wave-functions $|\Psi_i(t) \rangle$, and propagate the state in time to near exact accuracy within this representation using the TEBD algorithm [@vidal2003efficient; @vidal2004efficient].
We note that in addition to the drastic reduction in computation time and improvement in accuracy afforded by the time averaging procedure, a relatively modest matrix dimension in the MPS system description is sufficient for accurate results. This is a consequence of long range correlations in the system being constantly broken up by the local incoherent processes. A quantitative analysis of the relationship between correlations in individual trajectory wave-functions, and those in the steady state density matrix, will be presented elsewhere [@our_future_paper].
*Polariton crystallization calculations -* To obtain the steady states used in Fig. \[fig:polariton\_crystallization\], we began time averaging at $g (T_{\rm start}) = 500$. We found that $N_T = 5000$ time-steps $g (\Delta t) = 0.5$, with around 100 trajectories retaining around 40 states in the matrix product state representation, were sufficient to obtain acceptable statistical fluctuations ($\approx 0.1\%$) in the NESS expectation values we are interested in. We found that retaining three or four photons in the bare resonator basis was sufficient for systems in the ‘atomic’ regime ($\Delta \geq 0$), but that more were required to properly evaluate steady state quantities for $\Delta < 0$ (results not shown).
References {#references .unnumbered}
==========
[^1]: Assuming the spacing between neighboring modes within a resonator to be much larger than all other scales we only employ a single photon mode per resonator.
[^2]: For linear cavities this permits the numerical analysis of driving strengths up to $\Omega / \gamma_p \approx 0.3$ without significant truncation errors. With a resonator nonlinearity $g, U \ne 0$, occupation of higher-lying levels is suppressed enabling accurate exact diagonalization results for yet stronger driving, confirmed by extensive quantum trajectory calculations as detailed in the Appendix.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Imaging systems based on a narrow-band tunable filter are used to obtain Doppler velocity maps of solar features. These velocity maps are created by taking the difference between the blue- and red-wing intensity images of a chosen spectral line. This method has the inherent assumption that these two images are obtained under identical conditions. With the dynamical nature of the solar features as well as the Earth’s atmosphere, systematic errors can be introduced in such measurements. In this paper, a quantitative estimate of the errors introduced due to variable seeing conditions for ground-based observations is simulated and compared with real observational data for identifying their reliability. It is shown, under such conditions, that there is a strong cross-talk from the total intensity to the velocity estimates. These spurious velocities are larger in magnitude for the umbral regions compared to the penumbra or quiet-sun regions surrounding the sunspots. The variable seeing can induce spurious velocities up to about 1 km s$^{-1}$. It is also shown that adaptive optics, in general, helps in minimising this effect.'
author:
- 'Sreejith $^{1,2}$, R. $^{3}$, K. $^{1}$'
title: 'Seeing-Induced Errors in Solar Doppler Velocity Measurements[^1]'
---
Introduction {#S-intro}
============
Recent observations have generated new interest in understanding the sunspot fine structures like umbral dots, light bridges, and penumbral filaments [@rimmele2008; @rimmele2006a; @scharmer2002; @shussler2006]. Detailed spatial and temporal observations of these fine structures are crucial in understanding the physical mechanisms behind the formation of these structures. With the success of the solar adaptive optics (AO), it is now feasible to study structures close to the diffraction limit of modern telescopes [@rimmele2004a; @sankar2003; @sankar2007; @rimmele2008]. These small-scale structures harbour flows that are important for understanding the interaction between magnetic fields and plasma. Observational study of these flows would provide constraints to the theoretical models and to the magnetohydrodynamic simulations of these flows and hence would lead to a better understanding of the overall structure.
Doppler shifts of spectral lines are regularly used to study the line of sight (LOS) velocity of solar features. They are observed either with a spectrograph-based instrument or with an instrument based on tunable narrow-band filter. In instruments based on a tunable narrow-band filter, Doppler velocities are obtained using the red- and blue-wing intensity images of a chosen spectral line. Intensity at a fixed wavelength point in the blue and/or red wing of any spectral line varies depending on the Doppler shift. At a fixed wavelength point in the red wing of an absorption line, a red-(blue-) shift will reduce (increase) the intensity. Similarly, at a fixed point in the blue wing of an absorption line, a red-(blue-) shift will increase (decrease) the intensity. Hence, the difference between the red- and blue-wing intensities obtained at a fixed wavelength point is used to estimate the Doppler shift and hence the Doppler velocity. A magnetic field and its gradients may affect the magnetically sensitive spectral line profiles and hence the Doppler velocities estimated from them [@wachter2006; @rajaguru2007]. Therefore, magnetically insensitive lines are preferred for a “clean" velocity estimation. For example, 5434 Å or 5576 Å are typically used to estimate Doppler velocities at the photosphere. Systems based either on a tunable Fabry-Pérot etalon or a universal birefringent filter (UBF) (see , , , and references therein for details about these instruments) are used to obtain the required spectral bandwidth (*e.g.*, about 200mÅ for photospheric spectral lines). In both schemes, Doppler velocities are estimated from the difference between intensities obtained at the blue and red wing of a chosen spectral line by using the following relation:
$$V = C.\frac{I_{\rm r} - I_{\rm b}}{I_{\rm r} + I_{\rm b}} ,
\label{eq1}$$
where $I_{\rm r}$ and $I_{\rm b}$ are the red- and blue-wing intensities and $C$ is a calibration constant which depends on the chosen spectral line and the spectral resolution. $C$ can be obtained using a well-known procedure [@Rimmele2004] briefly explained in Section \[S-simul\] of this paper. With this definition, positive (negative) velocity correspond to flows towards (away from) the observer. This sign convention is followed throughout this paper.
In such observations, the blue- and red-wing images are NOT recorded simultaneously. The time difference between the two depends on the wavelength tuning time, required number of wavelength positions, and the detector read-out time. In most cases the detector read-out time, which is typically a few seconds, limits the cadence. Hence, any appreciable change in the observing conditions within this time interval can introduce systematic errors in the velocity as well as spurious velocity structures. If these spurious velocities and these structures are comparable to those of the intrinsic velocities of photospheric structures, then the physical interpretation of the structures will be ambiguous. The typical intrinsic velocities in umbral dots are of the order of a few hundred m s$^{-1}$, whereas the penumbral Evershed flows and quiet-sun granular velocities are of the order of a few thousand m s$^{-1}$ [@rimmele1995; @lokesh2007].
It is a well-known fact that ground-based observations are affected by the atmospheric turbulence which is often characterised by Fried’s parameter ($r_0$) for long-exposure images. For this paper, a variable seeing condition refers to the time variation of the parameter $r_0$. If the time variation is a few cm within the few seconds required for obtaining the red- and blue-wing images, then the variable seeing conditions can introduce spurious velocity signals. In ground-based observations, an adaptive optics system is used to minimise the seeing effect. However, the performance of an adaptive optics system is a function of the seeing conditions at the time of observations. have shown that the Sterhl ratio (one of the metrics for evaluating AO performance) of an AO corrected image is a function of Fried’s parameter ($r_0$).
![Images showing the effect of variable seeing on velocity map. Top row images show a sample good red- (R1), blue-wing intensity images (B1) and the velocity map (V1) derived from these wing intensity images. Bottom row shows the same but observed during variable seeing conditions. []{data-label="vel_image"}](sree_fig01.eps)
An example of the effect of variable seeing is shown in Figure \[vel\_image\]. The observations were obtained at the Dunn Solar Telescope (DST) in Sunspot, NM, USA using the UBF system. The images marked with R and B (R1, R2, B1, and B2) are the red- and blue-wing images whereas the images marked with V (V1 and V2) are the estimated velocities from the respective R and B. The top row images were obtained under stable seeing conditions whereas the bottom row during variable seeing conditions. The velocity map clearly shows spurious velocity signals in the umbral and penumbral regions under variable seeing conditions. This is also true with quiet solar granulation (not shown here) but not so obvious to the eye due to the high intensity-velocity correlation at these regions.
The aim of this paper is to quantitatively estimate the variable seeing-induced spurious velocity signals, through simulations and to compare them with observations. In Section \[S-simul\], the method used to simulate the variable seeing conditions is explained and the input data used for the simulation are also explained. In Section \[S-results\], results from the simulation using both space- and ground-based data are discussed. The comparison of the simulation results with the observed data and a summary are given in Section \[S-summary\]. We conclude with the result that there is a good correlation between seeing difference and spurious velocity signals, especially in the umbral region of the spots. Our simulation indicates that spurious velocities can be as large as 1 km s$^{-1}$ and it is alarming to note that such high values are seen in the observed data.
Simulations {#S-simul}
===========
To simulate the effect of variable seeing on the Doppler velocities, the red- and blue-wing images, initially unaffected by the atmospheric turbulence, were convolved with point spread functions (PSFs) produced using different $r_0$. The PSFs were generated using the software tool Adaptive Optics Performance Evaluator (AOPE), originally developed for performing simulations on the design needs of solar adaptive optics systems [@Sridharan2004].
AOPE {#SS-aope}
----
A detailed description of the effect of the atmospheric turbulence on the quality of the images obtained with ground-based telescopes can be found in . The instantaneous wavefront perturbations induced by the atmosphere can be represented as a two-dimensional phase screen. AOPE generates such phase screens following the Kolmogorov model of turbulence, for any given value of Fried’s parameter ($r_0$) and derives a long-exposure PSF from them for a chosen telescope diameter and observing wavelength. AOPE also simulates the effect of the adaptive optics correction by fitting a model phase screen with finite number of Zernike polynomials (which are generally used in characterising the aberrations in optical systems) to the originally generated phase screen and subtracting the best fit model phase screen from the original phase screen. The long-exposure PSFs after adaptive optics correction are then generated from a series of residual phase screens, again for a chosen telescope diameter and wavelength. Thus, Fried’s parameter, telescope diameter, number of equivalent Zernike modes corrected by the adaptive optics system and the observing wavelength are the input parameters to be selected by the users. Long-exposure PSFs with and without a finite number of Zernike-mode correction, and ideal PSF of the telescope are the output parameters. These output parameters are characterised with the Strehl ratio, normalised Strehl resolution and the strehl width which quantify the final image quality for a given set of input parameters. In our simulations, we used this tool to obtain the ideal PSF of the telescope, and the PSFs with and without the required Zernike correction. The corrected PSF depends on all the four input parameters, whereas uncorrected PSFs do not depend on $Z$, the order of Zernike correction. The ideal PSFs depend only on telescope diameter and wavelength $\lambda$.
Input Data and Calibration {#SS-data}
--------------------------
The input data used for the simulation are obtained using the Solar Optical Telescope (SOT) on-board **Hinode** - a satellite dedicated for solar observations. *Hinode* is a joint mission between the space agencies of Japan, United States, Europe, and United Kingdom [@hinodemain]. Being a space-based instrument, the data obtained from SOT are free from atmospheric turbulence. The narrow-band filter imager (NFI) on SOT is used to observe wing images of magnetically insensitive ($g=0$) line ($\lambda$=5576Å ) [@sotmain; @sotref1]. The observations were carried out on 2007 July 14, 11:34 UT of an active region NOAA 10963. A pair of images observed at the wings ($\pm$136mÅ away from the line core) of the line 5576.09Å are used as the input images for this simulation.
![Velocity versus normalised intensity difference curves estimated for the spectral lines 5434 Å (wing points at $\pm$ 60 mÅ ; represented by asterisks) and 5576 Å (wing points at $\pm$ 136 mÅ ; represented by pluses). Solid and dashed lines are straight line fits to the respective curves. []{data-label="ubfcal"}](sree_fig02.eps)
The input data for the simulation of ground-based images are obtained using UBF observation of a sunspot carried out at the Dunn Solar Telescope (DST), NSO, NM, USA, on 2005 December 28. The calibration constant $C$ ([*cf.*]{} Equation (1)), required for deriving the velocity from the observed normalised intensity difference $\delta I$ ( = $\frac{I_{\rm r} - I_{\rm b}}{I_{\rm r} + I_{\rm b}}$), is estimated using the spectral profile from the Liège atlas. First, the atlas profile of the observed line is convolved with a Gaussian filter profile of passband specific to the instrument used. In this paper, data from the NFI on-board *Hinode* as well as from the UBF at the Dunn Solar Telescope are used. The passband is estimated to be 70 m[Å]{} in the case of NFI and 142 m[Å]{} for the UBF. The Doppler shift of the spectral line $\Delta\lambda$ for a defined velocity $v$ is calculated by using $\frac{\Delta\lambda}{\lambda} = \frac{v}{c} $ where $c$ is the speed of light. The convolved spectral line is then shifted by an amount of $\Delta\lambda$ and the normalised intensity difference is calculated. This is repeated for a range of velocities. Figure \[ubfcal\] shows the relation between $\Delta v$ and $\delta I$ for both 5576 Å (plus) and 5434 Å (asterisks) spectral lines. The linear part of the curve is fitted with a straight line (solid and dashed lines respectively) and 1/slope provides the calibration constant value $C$. For large velocity values, either the line core or the continuum will cross one of the chosen wing wavelengths and hence the $\delta I$ curve will deviate from the straight line. This sets the limit for the velocity range that can be measured using this method. The velocity range and the slope value (or calibration constant $C$) depends on the spectral line and the chosen wing pair. This is clearly reflected in Figure \[ubfcal\] in which the dynamic range achieved for 5576 Å is smaller compared to that of 5434 Å .
Procedure {#SS-procedure}
---------
The simulation is carried out for a telescope diameter of 50 cm (commensurate with *Hinode*) and for different Fried’s parameter values (starting from $r_0$ = 4 cm to 15 cm). In typical ground-based solar observations, Fried’s parameter of 4 cm or below is considered as bad seeing and an $r_0$ of 12 cm or above is considered as an excellent seeing condition. For each $r_0$ values, PSFs with and without Zernike corrections (of a particular order) are generated. For the simulation, Zernike order is varied from 11 to 55 depicting different amount of AO corrections. The PSFs generated are then convolved with the input image to generate different observational conditions. In order to simulate the variable seeing effects, the blue- and red-wing images are convolved with different PSFs generated using different $r_0$ as well as different Zernike correction.
The velocities are derived using Equation (\[eq1\]) for different simulated conditions. The generated velocity images are used to quantify the effect of variable seeing conditions. The umbra, penumbra, and quiet Sun regions are analysed separately in order to quantify the effects in these different regions.
For quantifying such variable seeing conditions between blue- and red-wing images, we define the normalised seeing difference $(\delta r_0)$ by,
$$\delta r_0 = \frac{r_{\rm 0R} -r_{\rm 0B}}{r_{\rm 0R} +r_{\rm 0B}} ,
\label{eq2}$$
where $r_{\rm 0R}$ and $r_{\rm 0B}$ are the seeing during the red- and blue-wing observations respectively.
Results {#S-results}
=======
The top row of Figure \[hinode\_fig2\] shows a pair of red- and blue-wing images and the velocity image used in our simulation. The velocity image in the top row shows the typical granular flows in the quiet region, Evershed flows in the penumbral region and uniform or zero flows in the umbral region. The grey scale used for the velocity images are also marked as a colour bar on the right side of the velocity figures. The bottom row shows the best case scenario when the seeing is good, like with the case of $r_0$ = 15 cm and high order AO correction ($Z$ = 66). The estimated velocity image looks similar to the original. The middle row shows a seeing-affected blue-wing image simulated for $r_0$ = 4 cm and without AO correction (the worst case scenario), the original red-wing image (convolved with ideal PSF), and the resulting velocity image. Note the change in the velocity values and the appearance of velocity structures in the umbra. These small-scale structures resemble umbral dots in the intensity image and hence are considered as cross-talk from intensity. This clearly shows the effect of the seeing difference inducing spurious velocity structures as well as modifying the velocity amplitudes. These effects are most notable in the umbral region.
![*Hinode* SOT-NFI images showing the effect of seeing difference on velocity. This figure is similar to Figure \[vel\_image\] except that it is simulated. The top row shows the original red- (R1), blue-wing (B1) images and the corresponding velocity map (V1). The middle row shows the original red- (R2), convolved blue-wing (B2) images and the corresponding velocity (V2). B2 is created by convolving original blue-wing image with a PSF of $r_0$ = 4 cm and without AO correction. In the bottom row, B3 is created by convolving original blue-wing image with the PSF of $r_0$ = 15 cm and $Z$=66. The colour bars on the right-hand side shows the grey scale range for the corresponding velocity images. Dark and white contours in V1 are the regions selected for measuring $\sigma$ and $\mu$, respectively. []{data-label="hinode_fig2"}](sree_fig03.eps)
In order to quantitatively study the effect on the mean velocity ($\mu$), a core umbral region with minimal intensity structures (like umbral dots) is selected manually. This region is shown as a white contour in V1 of Figure \[hinode\_fig2\]. The spurious velocity structures (as seen in V2 of Figure \[hinode\_fig2\]) are quantified using the standard deviation of the velocity of an umbral region which includes the umbral structures (like umbral dots) and this is shown as a dark contour in V1 of Figure \[hinode\_fig2\]. The two parameters ($\mu$ and $\sigma$) are plotted with normalised seeing difference (cf. Equation (\[eq2\])) in Figure \[hinode1\]. The seeing difference is simulated by convolving one of the wing images (either red or blue) with a PSF of a fixed Fried’s parameter (called base seeing in this paper), and the other wing is convolved with PSFs corresponding to variable Fried’s parameter ($r_0$ = 4 to 15 cm). This is repeated by varying the base seeing values from 4 to 15 cm. The normalised seeing difference is calculated as per Equation (\[eq2\]), and the mean velocity is calculated at the selected core umbral region. The same is also repeated with AO corrected PSFs, with Zernike correction order varying from $Z$=11 to $Z$=55.
In Figure \[hinode1\], the normalised seeing difference versus umbral core mean velocity is plotted on the left column. The top plot is the case without AO correction, the middle plot is with a Zernike correction of $Z$=11 and the bottom one is for $Z$=55. All the plots have three types of symbols. Squares are for base seeing, where either $r_{\rm 0R}$ or $r_{\rm 0B} $ is equal to 15 cm and the other varying from 15 to 4 cm. In other words, one of the wing images is affected by the best seeing condition and the other varies from best to worst. Similarly, asterisks are for a base seeing of 4 cm and pluses are for 9 cm. Each curve has two parts, separated by zero $\delta r_0$, and positive abscissa ( positive $\delta r_0$) means $r_{\rm 0R}~>~r_{\rm 0B}$. All the symbols are just connected by lines of different line types. As seen in the right column of Figure \[hinode1\], the standard deviation ($\sigma$) varies with seeing difference and hence cannot be used as an error estimate for $\mu$. The standard deviation of the velocity estimated in the selected umbral region of the original velocity image is used as the error in the velocity estimates for all simulated conditions.
All the three plots show that spurious velocity values increase with normalised seeing difference. For a particular normalised seeing difference value, there is a range of velocity values depending on the base seeing condition. A poor base seeing, or the dashed line in the plot (connecting asterisks, where either $r_{\rm 0R}$ or $r_{\rm 0B}$ is equal to 4 cm), represents the maximum limit, and a good base seeing, solid line (connecting squares, where either $r_{\rm 0R}$ or $r_{\rm 0B} $ is equal to 15 cm) is the minimum limit of these induced velocities. Notice that the range comes to almost zero at zero normalised seeing difference, irrespective of the difference in the base seeing. As AO corrections are applied (middle and bottom plots), both the maximum and minimum limit of the induced velocity decreases.
![ Effect of seeing difference on mean velocity (left column) and on velocity structures (right column) in the umbral region. The top row is the case without AO correction, the middle one with a Zernike correction of 11 and the bottom one with a correction of 55. In all the plots, squares represent the best seeing case when $r_{\rm 0R}$ or $r_{\rm 0B}$ = 15 cm and the solid line connects the symbols, pluses are for $r_{\rm 0R}$ or $r_{\rm 0B}$ = 9 cm, asterisks are for $r_{\rm 0R}$ or $r_{\rm 0B}$ = 0 cm and the dotted lines and dashed lines connect the symbols, respectively. []{data-label="hinode1"}](sree_fig04.eps "fig:") ![ Effect of seeing difference on mean velocity (left column) and on velocity structures (right column) in the umbral region. The top row is the case without AO correction, the middle one with a Zernike correction of 11 and the bottom one with a correction of 55. In all the plots, squares represent the best seeing case when $r_{\rm 0R}$ or $r_{\rm 0B}$ = 15 cm and the solid line connects the symbols, pluses are for $r_{\rm 0R}$ or $r_{\rm 0B}$ = 9 cm, asterisks are for $r_{\rm 0R}$ or $r_{\rm 0B}$ = 0 cm and the dotted lines and dashed lines connect the symbols, respectively. []{data-label="hinode1"}](sree_fig05.eps "fig:") ![ Effect of seeing difference on mean velocity (left column) and on velocity structures (right column) in the umbral region. The top row is the case without AO correction, the middle one with a Zernike correction of 11 and the bottom one with a correction of 55. In all the plots, squares represent the best seeing case when $r_{\rm 0R}$ or $r_{\rm 0B}$ = 15 cm and the solid line connects the symbols, pluses are for $r_{\rm 0R}$ or $r_{\rm 0B}$ = 9 cm, asterisks are for $r_{\rm 0R}$ or $r_{\rm 0B}$ = 0 cm and the dotted lines and dashed lines connect the symbols, respectively. []{data-label="hinode1"}](sree_fig06.eps "fig:") ![ Effect of seeing difference on mean velocity (left column) and on velocity structures (right column) in the umbral region. The top row is the case without AO correction, the middle one with a Zernike correction of 11 and the bottom one with a correction of 55. In all the plots, squares represent the best seeing case when $r_{\rm 0R}$ or $r_{\rm 0B}$ = 15 cm and the solid line connects the symbols, pluses are for $r_{\rm 0R}$ or $r_{\rm 0B}$ = 9 cm, asterisks are for $r_{\rm 0R}$ or $r_{\rm 0B}$ = 0 cm and the dotted lines and dashed lines connect the symbols, respectively. []{data-label="hinode1"}](sree_fig07.eps "fig:") ![ Effect of seeing difference on mean velocity (left column) and on velocity structures (right column) in the umbral region. The top row is the case without AO correction, the middle one with a Zernike correction of 11 and the bottom one with a correction of 55. In all the plots, squares represent the best seeing case when $r_{\rm 0R}$ or $r_{\rm 0B}$ = 15 cm and the solid line connects the symbols, pluses are for $r_{\rm 0R}$ or $r_{\rm 0B}$ = 9 cm, asterisks are for $r_{\rm 0R}$ or $r_{\rm 0B}$ = 0 cm and the dotted lines and dashed lines connect the symbols, respectively. []{data-label="hinode1"}](sree_fig08.eps "fig:") ![ Effect of seeing difference on mean velocity (left column) and on velocity structures (right column) in the umbral region. The top row is the case without AO correction, the middle one with a Zernike correction of 11 and the bottom one with a correction of 55. In all the plots, squares represent the best seeing case when $r_{\rm 0R}$ or $r_{\rm 0B}$ = 15 cm and the solid line connects the symbols, pluses are for $r_{\rm 0R}$ or $r_{\rm 0B}$ = 9 cm, asterisks are for $r_{\rm 0R}$ or $r_{\rm 0B}$ = 0 cm and the dotted lines and dashed lines connect the symbols, respectively. []{data-label="hinode1"}](sree_fig09.eps "fig:")
In the right column of Figure \[hinode1\], the standard deviation of the selected umbral area is plotted with normalised seeing difference. The top plot is without AO correction, the middle one with a Zernike correction of $Z$=11 and the bottom one with $Z$=55. The horizontal line is the value of $\sigma$ of the original velocity image. It is clear from the curves that the seeing difference introduces spurious velocity structures in the selected umbral region, and AO with higher order correction helps in minimising these induced spurious velocity structures. When both wing images are affected by similar bad seeing, the fine scale intensity structures are smeared in both. Hence the intrinsic velocity contrast of these structures is reduced representing the points below the horizontal line.
A similar study is carried out for the penumbral regions. However, due to the presence of Evershed flows, the mean velocity ($\mu$) was calculated separately for positive and negative velocities (flows away from and towards the observer). The standard deviation of the selected penumbral area is a measure of penumbral structures. These two parameters are plotted in Figure \[hinode2\] for different Zernike corrections. It is clear from the figure that $\mu$ (left column) in a penumbra increases very little (200-300 m s$^{-1}$) with normalised seeing difference, whereas $\sigma$ (right column) increases significantly. Similar to the umbral structures, with higher order AO correction $\sigma$ falls off to original and also $\sigma$ does not fall to zero due to the existence of penumbral velocity structures in the input image. In the case of the penumbra, $\sigma$ of the input image cannot be taken as the error in the velocity estimate due to the presence of Evershed flow velocity structures, and hence the error bars are not displayed. Horizontal lines show the value of $\sigma$ estimated from the initial velocity image.
The quiet sun-study showed a similar behaviour (with weaker amplitudes) like the penumbra and hence is not discussed in this paper.
![Similar to Figure \[hinode1\] but for the penumbral region. The mean velocity is calculated separately for positive (towards the observer) and negative (away from the observer) cases. Squares (positive side) and triangles (negative side) represent the best base seeing condition ($r_{\rm 0R}$ or $r_{\rm 0B}$ = 15 cm). Similarly, asterisks (positive side) and diamonds (negative side) represent the bad base seeing condition (4 cm), and pluses represent the medium seeing condition (9 cm). Lines are just a connection of data points. []{data-label="hinode2"}](sree_fig10.eps "fig:") ![Similar to Figure \[hinode1\] but for the penumbral region. The mean velocity is calculated separately for positive (towards the observer) and negative (away from the observer) cases. Squares (positive side) and triangles (negative side) represent the best base seeing condition ($r_{\rm 0R}$ or $r_{\rm 0B}$ = 15 cm). Similarly, asterisks (positive side) and diamonds (negative side) represent the bad base seeing condition (4 cm), and pluses represent the medium seeing condition (9 cm). Lines are just a connection of data points. []{data-label="hinode2"}](sree_fig11.eps "fig:") ![Similar to Figure \[hinode1\] but for the penumbral region. The mean velocity is calculated separately for positive (towards the observer) and negative (away from the observer) cases. Squares (positive side) and triangles (negative side) represent the best base seeing condition ($r_{\rm 0R}$ or $r_{\rm 0B}$ = 15 cm). Similarly, asterisks (positive side) and diamonds (negative side) represent the bad base seeing condition (4 cm), and pluses represent the medium seeing condition (9 cm). Lines are just a connection of data points. []{data-label="hinode2"}](sree_fig12.eps "fig:") ![Similar to Figure \[hinode1\] but for the penumbral region. The mean velocity is calculated separately for positive (towards the observer) and negative (away from the observer) cases. Squares (positive side) and triangles (negative side) represent the best base seeing condition ($r_{\rm 0R}$ or $r_{\rm 0B}$ = 15 cm). Similarly, asterisks (positive side) and diamonds (negative side) represent the bad base seeing condition (4 cm), and pluses represent the medium seeing condition (9 cm). Lines are just a connection of data points. []{data-label="hinode2"}](sree_fig13.eps "fig:") ![Similar to Figure \[hinode1\] but for the penumbral region. The mean velocity is calculated separately for positive (towards the observer) and negative (away from the observer) cases. Squares (positive side) and triangles (negative side) represent the best base seeing condition ($r_{\rm 0R}$ or $r_{\rm 0B}$ = 15 cm). Similarly, asterisks (positive side) and diamonds (negative side) represent the bad base seeing condition (4 cm), and pluses represent the medium seeing condition (9 cm). Lines are just a connection of data points. []{data-label="hinode2"}](sree_fig14.eps "fig:") ![Similar to Figure \[hinode1\] but for the penumbral region. The mean velocity is calculated separately for positive (towards the observer) and negative (away from the observer) cases. Squares (positive side) and triangles (negative side) represent the best base seeing condition ($r_{\rm 0R}$ or $r_{\rm 0B}$ = 15 cm). Similarly, asterisks (positive side) and diamonds (negative side) represent the bad base seeing condition (4 cm), and pluses represent the medium seeing condition (9 cm). Lines are just a connection of data points. []{data-label="hinode2"}](sree_fig15.eps "fig:")
Ground-Based Observations {#SS-ground}
-------------------------
A similar study is carried out with images obtained from ground-based observation. The seeing was variable at the time of observation. About 3 s was required to tune UBF from one wing to the other wing. Due to the variable seeing condition, there were many wing-pair images affected by the seeing difference in the observed data (one such an example is shown in Figure \[vel\_image\]). From the observed data, a pair of best images were chosen as input for the simulations. Differential seeing was simulated as explained in the previous section.
The curves of $\mu$ and $\sigma$ are very similar to the curves seen in Figure \[hinode1\], and hence they are not reproduced here in this paper. The only notable difference was in the $\sigma$ curves and that can be attributed to the difference in the umbral structures for these two different sunspot regions.
Summary and Discussions {#S-summary}
=======================
In the velocity measurements from the ground using narrow-band filters, there is always a chance of systematic error in velocity estimates due to the seeing difference. If seeing changes within the time of wavelength tuning from one wing to the other, it will create spurious velocity values. We simulated such seeing difference conditions using the AOPE code, and we studied the induced velocity values in different areas of a solar active region (like umbra, penumbra, and quiet sun). It is concluded that the seeing difference affects the velocity estimates more in the umbra than in the penumbral region of the sunspot. It is also seen that the effect is minimal in the quiet sun. Under worst normalised seeing difference conditions, where one wing image is obtained under a good seeing condition ($r_0$ = 15 cm) and the other during worst condition ($r_0$ = 4 cm; $\delta r_0$=0.58), the induced velocity is as high as 1 km s$^{-1}$. Such a large variation in seeing may not occur often in real cases, but our sample observations (*e.g.* Figure \[vel\_image\]) using the UBF at the DST have shown velocities up to 600 m s$^{-1}$. It is also clear that these induced velocity structures are due to cross-talk from intensity. A correlation analysis carried out between intensity and velocity shows a correlation coefficients up to 50% in the worst case scenario. These simulations also show that adaptive optics corrections can significantly reduce such induced velocities, and velocity structures.
![ Scatter plot between normalised seeing difference and normalised contrast difference from the simulations using NSO data as the input.[]{data-label="dcontra"}](sree_fig16.eps)
![Graph showing umbral core velocity with normalised contrast difference. The filled squares are from the observation and red pluses are from the simulation. The error bars on the observed points are three times the standard deviation of the umbral core velocity distribution.[]{data-label="final_plot"}](sree_fig17.eps)
We used UBF observations carried out during variable seeing conditions to compare with the simulations. However, in order to compare with the real observations, contrast values are used as a measure of seeing, since there were no simultaneous measurements of Fried’s parameter. The contrast at the umbra-penumbra border was used as a measure of seeing rather than the quiet-sun contrast. This is because the lock point of AO was in the umbra-penumbra border. The AO correction away from the lock point was really variable due to the variable seeing condition and hence contrast obtained at these areas will not represent the seeing changes alone in the umbra. We have also taken care to choose the region which has a minimum velocity throughout the time sequence of roughly 2 hours, as velocity can also change the intensity contrast. We define contrast as the standard deviation of the selected region in the intensity image. The normalised contrast difference is defined as the ratio between the difference and total of the blue- and red-wing contrasts and hence is a proxy for the normalised seeing difference value. Figure \[dcontra\] shows the relation between normalised seeing difference and normalised contrast difference from simulations with NSO data as the input. It is clear from the plot that the spread in the curve will cause a degeneracy in the normalised contrast difference values for a particular normalised seeing difference. Also, the zero normalised contrast difference does not fall on the zero normalised seeing difference and vice versa. This shift can be associated to the seeing difference in the input wing images and to the convective blue-shift.
Figure \[final\_plot\] shows the scatter plot between the normalised contrast difference versus $\mu$ at the core umbra estimated for the NSO data. Squares are from observation and pluses are from simulation. The simulation was extended to include all combinations of seeing values varying from $r_0$ = 4 to 15 cm and a Zernike correction from $Z$ = 11 to 77 to mimic a real observation. The error bar of the original data points are three times the standard deviation ($\sigma$) of the velocity distribution in the selected umbral core region. The value of this $\sigma$ will be a combination of the actual error in the velocity estimation and the standard deviation due to the induced structures because of the seeing difference.
In general most of the data points (70%) fall within the simulated velocity distribution. However, there are points well away from the simulated curve which may be due to (1) the observed data being a time sequence spanning approximately 2 hours, and hence any evolution can change both the velocity value and the contrast, (2) simulation assuming noiseless PSF after perfect AO correction whereas in a real situation it will not be.
We conclude the folloeing. (1) A seeing difference introduces spurious velocity signals which are large in the umbra compared to the penumbra and the quiet sun due to the low intrinsic velocities inside the umbra. (2) The spurious velocity also depends on the base seeing condition. If the general seeing conditions are very bad, then even a small seeing difference can cause large spurious velocity. (3) The simulations have given a range of spurious velocity, for a particular seeing difference in very bad base seeing condition ($r_0$= 4 cm) and for very good base seeing condition ($r_0$ = 15 cm). For a normalised seeing difference condition of 0.5 this spurious velocity can range from 600 m s$^{-1}$ ( good base seeing) to 1200 m s$^{-1}$ (bad base seeing). A normalised seeing difference of 0.5 in the umbra occurs when the ratio of Fried’s parameter between the red- and blue-wing images is larger than 3. A normalised seeing difference of 0.5 corresponds to a normalised contrast difference of approximately 0.2 and such values are seen in the observed data. (4) With higher order AO corrections, the spurious velocities are reduced by a factor greater than 4 for a normalised seeing difference of 0.5 or less for a perfect AO system.
We wish to caution the reader that this simulation is just to approximate the range of errors induced due to a seeing difference in velocity measurements based on a narrow-band filter. An error in the range of 1 km s$^{-1}$ is important in areas like the umbra and penumbra at the photospheric level. These results cannot be used to remove the spurious velocities from the seeing-affected observed images. At the same time, we cannot ignore all seeing-affected data, especially in the case of observation of transient phenomena like penumbral formation or a flare. In real observational situations, PSFs can vary even when $r_0$ remains constant (especially for exposures not long enough to average out the seeing fluctuations). In our simulation, it is assumed that the changes in PSFs are only due to changes in $r_0$. If PSFs can be estimated using the AO data simultaneously along with the Doppler measurements, then the variable seeing effects could be minimised by using deconvolution techniques [@rimmele2006a]. In the future, new methods and technologies may also be developed to observe the blue- and red-wing images simultaneously to avoid such effects.
[15]{} \#1[ISBN \#1]{}\#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[*\#1*]{}\#1[**\#1**]{}\#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1 \#1\#2[\#2]{} \#1[[\#1](http://dx.doi.org/#1)]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[*\#1*]{} \#1[\#1]{} \#1[**\#1**]{} \#1[\#1]{} \#1[*\#1*]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1 \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{}
, , : , ,
, , : , , .
: , , .
, , , , , , : , , .
, , , : , , .
: , , .
, : , , .
: , , .
, : , , .
: , , .
: 2004, In: [Bonaccini Calia, D., Ellerbroek, B.L., Ragazzoni, R.]{} (eds.), *Advancements in Adaptive Optics, Proc. SPIE* **5490**, 34.
: , , .
, : , , .
, : , , .
, , , , : , , .
, : , , .
, : 2004, In: [Fineschi, S., Gummin, M.A.,]{} (eds.), *Telescopes and Instrumentation for Solar Astrophysics, Proc. SPIE* **5171**, 219.
: , , Springer Publications., .
, , , , , , : , , .
, , , , , , : , , .
, , : , , .
[^1]: The original publication is at <http://www.springerlink.com/content/n5v33354r06j2017/>
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper aims to investigate a full numerical approximation of non-autonomous semilnear parabolic partial differential equations (PDEs) with nonsmooth initial data. Our main interest is on such PDEs where the nonlinear part is stronger than the linear part, also called reactive dominated transport equations. For such equations, many classical numerical methods lose their stability properties. We perform the space and time discretizations respectively by the finite element method and an exponential integrator. We obtain a novel explicit, stable and efficient scheme for such problems called Magnus-Rosenbrock method. We prove the convergence of the fully discrete scheme toward the exact solution. The result shows how the convergence orders in both space and time depend on the regularity of the initial data. In particular, when the initial data belongs to the domain of the family of the linear operator, we achieve convergence orders $\mathcal{O}\left(h^{2}+\Delta t^{2-\epsilon}\right)$, for an arbitrarily small $\epsilon>0$. Numerical simulations to illustrate our theoretical result are provided.'
author:
- 'Antoine Tambue, Jean Daniel Mukam'
date: 'Received: date / Accepted: date'
title: 'Convergence analysis of the Magnus-Rosenbrock type method for the finite element discretization of semilinear non-autonomous parabolic PDEs with nonsmooth initial data'
---
[example.eps]{} gsave newpath 20 20 moveto 20 220 lineto 220 220 lineto 220 20 lineto closepath 2 setlinewidth gsave .4 setgray fill grestore stroke grestore
Introduction {#intro}
============
We consider the following abstract Cauchy problem with boundary conditions $$\begin{aligned}
\label{model}
u'(t)= A(t)u(t)+F(t,u(t)), \quad u(0)=u_0, \quad t\in(0,T], \quad T>0,\end{aligned}$$ on the Hilbert space $H=L^2(\Lambda)$, where $\Lambda$ is an open bounded subset of $\mathbb{R}^d$ $(d=1,2,3)$. The family of unbounded linear operators $A(t)$ is assumed to generate an analytic semigroup $S_s(t):=e^{A(s)t}$. Suitable assumptions on the nonlinear function $F$ and the linear operator $A(t)$ to ensure the existence of a unique mild solution of [[(\[model\])]{}]{} are given in the following section. Equation of type [[(\[model\])]{}]{} finds applications in many fields such as quantum fields theory, electromagnetism, nuclear physics, see e.g. [@Blanes]. Since analytic solutions of [[(\[model\])]{}]{} are usually not available, numerical algorithms are the only tools to provide good approximations. Numerical schemes for [[(\[model\])]{}]{} with constant linear operator $A(t)=A$ are widely investigated in the scientific literature, see e.g. [@Stig1; @Ostermann5; @Hochbruck2; @Ostermann3] and the references therein. If we turn our attention to the non-autonomous case, the list of references becomes remarkably short. In the linear case, [[(\[model\])]{}]{} has been investigated in [@Hochbruck1], where the authors examined the convergence analysis of the Magnus integrator to Schrödinger equation. The Magnus integrator was further investigated in [@Ostermann1] for PDE [[(\[model\])]{}]{} with $F$ independent of $u$, where the authors applied the mid-point rule to approximate the Magnus expansion in order to achieve a second order approximation in time. Numercal scheme for semilinear PDEs [[(\[model\])]{}]{} was investigated in [@Ostermann4] and the convergence in time has been proved. In [@Ostermann4], the authors used the backward Euler method. Although backward Euler method has good stability properties, it is computationally expensive as nonlinear systems need to be solved at each time step. Our goal here is to provide a novel efficient scheme to solve [[(\[model\])]{}]{} by upgrading the scheme for linear PDEs in [@Ostermann1] and providing a mathematical rigorous convergence proof in space and in time. A standard direction to upgrade the Magnus integrator [@Ostermann1] to semilinear PDEs consists to keep the linear structure of [[(\[model\])]{}]{} at each time step. However, when the linear part of [[(\[model\])]{}]{} is stronger than its nonlinear part, the PDE [[(\[model\])]{}]{} is driven by the linear part and the good stability properties of a scheme from such approach it is not guaranteed. Indeed when the nonlinear part of a PDE is stronger than its linear part, the PDE is driven by its nonlinear part. For such problems, keeping the linear structure of [[(\[model\])]{}]{} at each step yields schemes behaving like the unstable explicit Euler method.
In this paper, we propose a novel numerical scheme by applying the Rosenbrock-Type method [@Ostermann3; @finance; @Hochbruck2; @Hairer; @Antjd2] to the semi-discrete problem [[(\[semi1\])]{}]{} combining with the Magnus-integrator to the linearized problem. This combination yields an explicit efficient numerical method for such problems. The linearization technique weakens the nonlinear part such that the linearized semi-discrete problem is driven by its new linear part. In contrast to [@Ostermann4], the linearization technique is done at every time step. Note that the Rosenbrock method was investigated in the scientific literature only for autonomous problems, see e.g. [@Hochbruck2; @Antjd2] for deterministic problem and recently in [@Antjd1] for stochastic parabolic PDEs to the best of our knowledge. Moreover, the convergence analyses in [@Ostermann1; @Ostermann2; @Ostermann4] are only in time. Furthermore, we examine the space and time convergence with non smooth initial data where the space discretization is performed using the finite element method. Comparing with scheme in [@Antjd2], the analysis here is extremely complicated due to the complexity of $A(t)$ and its semigroup $S_s(t)=e^{A(s)t}$. This complexity is broken through novel rigorous mathematical results obtained in Section \[preliminaires\]. Furthermore, in contrast to the scheme in [@Antjd2; @Hochbruck], the new scheme is second order accuracy in time for non-autonomous PDEs [[(\[model\])]{}]{} with constant linear operator $A$ without the extra matrix exponential function $\varphi_2$. Our final convergence result shows how the convergence orders in both space and time depend on the regularity of the initial data. In particular, when the initial data belongs to the domain of the family of the linear operator, we achieve convergence orders $\mathcal{O}\left(h^{2}+\Delta t^{2-\epsilon}\right)$, for an arbitrarily small $\epsilon>0$. The paper is organized as follows. In Section \[nummethod\], results about the well posedness are provided along with the Magnus-Rosenbrock scheme (MAGROS) and the main result. The proof of the main result is presented in Section \[proof1\]. In Section \[numericalexperiment\], we present some numerical simulations to sustain our theoretical result.
Mathematical setting and numerical method {#nummethod}
=========================================
Notations, settings and well posedness
--------------------------------------
Let us start by presenting briefly notations, the main function spaces and norms that will be used in this paper. We denote by $\Vert \cdot \Vert$ the norm associated to the inner product $\langle\cdot ,\cdot \rangle_H$ of the Hilbert space $H=L^{2}(\Lambda)$. The norm in the Sobolev space $H^m(\Lambda)$, $ m \geq 0$ will be denoted by $\Vert. \Vert_m$. For a Hilbert space $U$ we denote by $\Vert\cdot\Vert_{U}$ the norm of $U$, $L(U, H)$ the set of bounded linear operators from $U$ to $H$. For ease of notation, we use $L(U,U)=:L(U)$. To guarantee the existence of a unique mild solution of [[(\[model\])]{}]{}, and for the purpose of the convergence analysis, we make the following assumptions.
\[assumption2\]
- As in [@Ostermann1; @Ostermann2; @Gonz], we assume that $\mathcal{D}\left(A(t)\right)=D$, $0\leq t\leq T$ and the family of linear operators $A(t) : D\subset H\longrightarrow H$ to be uniformly sectorial on $0\leq t\leq T$, i.e. there exist constants $c>0$ and $\theta\in\left(\frac{1}{2}\pi, \pi\right)$ such that $$\begin{aligned}
\left\Vert \left(\lambda\mathbf{I}-A(t)\right)^{-1}\right\Vert_{L(L^2(\Lambda))}\leq \frac{c}{\vert \lambda\vert},\quad \lambda\in S_{\theta},\end{aligned}$$ where $S_{\theta}:=\left\{\lambda\in\mathbb{C} : \lambda=\rho e^{i \phi}, \rho>0, 0\leq \vert \phi\vert\leq \theta\right\}$. As in [@Ostermann2], by a standard scaling argument, we assume $-A(t)$ to be invertible with bounded inverse.
- Similarly to [@Ostermann2; @Ostermann1; @Praha; @Gonz], we require the following Lipschitz conditions: there exists a positive constant $K_1$ such that [$$\begin{aligned}
\label{conditionB}
\left\Vert \left(A(t)-A(s)\right)(-A(0))^{-1}\right\Vert_{L(H)}&\leq& K_1\vert t-s\vert,\quad s,t\in[0, T],\\
\left\Vert (-A(0))^{-1}\left(A(t)-A(s)\right)\right\Vert_{L(D,H)}&\leq& K_1\vert t-s\vert,\quad s,t\in[0, T].\end{aligned}$$ ]{}
- Since we are dealing with non smooth data, we follow [@Praha] and assume that $$\begin{aligned}
\label{domaine}
\mathcal{D}\left(\left(-A(t)\right)^{\alpha}\right)=\mathcal{D}\left(\left(-A(0)\right)^{\alpha}\right),\quad 0\leq t\leq T,\quad 0\leq \alpha\leq 1\end{aligned}$$ and there exists a positive constant $K_2$ such that the following estimate holds uniformly for $t\in[0,T]$ [$$\begin{aligned}
\label{equivnorme1}
K_2^{-1}\left\Vert \left(-A(0)\right)^{\alpha}u\right\Vert\leq \left\Vert (-A(t))^{\alpha}u\right\Vert\leq K_2\left\Vert (-A(0))^{\alpha}u\right\Vert,\quad u\in \mathcal{D}\left(\left(-A(0)\right)^{\alpha}\right).\end{aligned}$$ ]{}
- Similarly to [@Ostermann2 (3.17)] and [@Gonz; @Ostermann4], we assume that the map $t \longmapsto A(t)$ is twice differentiable and for any $\alpha_1,\alpha_2\in[0,1]$ such that $\alpha_1+\alpha_2=1$, the following estimates are satisfied [$$\begin{aligned}
\Vert (-A(s))^{-\alpha_1}A''(t)A(s)^{-\alpha_2}\Vert_{L\left(\left(-A(0)\right)^{1-\alpha_2}, H\right)}&\leq& C_0,\quad s,t\in[0,T],\\
\Vert (-A(0))^{-\alpha_1}(A(t)-A(s))(-A(0))^{-\alpha_2}\Vert_{L\left(\left(-A(0)\right)^{1-\alpha_2}, H\right)}&\leq& C_0\vert t-s\vert,\quad s,t\in[0,T],\end{aligned}$$ ]{} where $C_0$ is a positive constant independent of $t_1$ and $t_2$.
\[remark1\] From [[Assumption \[assumption2\]]{}]{} (i) and (iii), it follows that for all $\alpha\geq 0$ and $\delta\in[0,1]$, there exists a constant $C_1>0$ such that the following estimates hold uniformly for $t\in[0,T]$ $$\begin{aligned}
\label{smooth}
\left\Vert (-A(t))^{\alpha}e^{sA(t)}\right\Vert_{L(H)}\leq C_1s^{-\alpha},\quad \quad s>0,\\
\label{smootha}
\left\Vert(-A(t))^{-\delta}\left(\mathbf{I}-e^{sA(t)}\right)\right\Vert_{L(H)}\leq C_1s^{\delta}, \quad\quad\quad s\geq 0,\end{aligned}$$ see e.g. [@Ostermann2 (2.1)].
\[remark2\] Let $\Delta(T):=\{(t,s) : 0\leq s\leq t\leq T\}$. It is well known that [@Pazy Theorem 6.1, Chapter 5] under [[Assumption \[assumption2\]]{}]{} there exists a unique evolution system [@Pazy Definition 5.3, Chapter 5] $U : \Delta(T)\longrightarrow L(H)$ such that
- There exists a positive constant $K_0$ such that $$\begin{aligned}
\Vert U(t,s)\Vert_{L(H)}\leq K_0,\quad 0\leq s\leq t\leq T.\end{aligned}$$
- $U(.,s)\in C^1(]s,T] ; L(H))$, $0\leq s\leq T$, $$\begin{aligned}
\frac{\partial U}{\partial t}(t,s)=-A(t)U(t,s), \quad 0\leq s\leq t\leq T, \\
\Vert A(t)U(t,s)\Vert_{L(H)}\leq \frac{K_0}{t-s},\quad 0\leq s<t\leq T.\end{aligned}$$
- $U(t,.)v\in C^1([0,t[ ; H)$, $0<t\leq T$, $v\in\mathcal{D}(A(0))$ and $$\begin{aligned}
\frac{\partial U}{\partial s}(t,s)v=-U(t,s)A(s)v,\quad 0\leq s\leq t\leq T,\\
\Vert A(t)U(t,s)A(s)^{-1}\Vert_{L(H)}\leq K_0, \quad 0\leq s\leq t\leq T.\end{aligned}$$
We equip $V_{\alpha}(t) : = \mathcal{D}\left(\left(-A(t)\right)^{\alpha/2}\right)$, $\alpha\in \mathbb{R}$ with the norm $\Vert u\Vert_{\alpha,t} := \Vert (-A(t))^{\alpha/2}u\Vert$. Due to [[(\[domaine\])]{}]{}-[[(\[equivnorme1\])]{}]{} and for the seek of ease notations, we simply write $V_{\alpha}$ and $\Vert .\Vert_{\alpha}$ instead of $V_{\alpha}(t)$ and $\Vert .\Vert_{\alpha,t}$ respectively.
\[assumption1\] The initial data $u_0 : \Lambda\longrightarrow H$ is assumed to satisfy $u_0\in \mathcal{D}\left(\left(-A(0)\right)^{\beta/2}\right)$, $0\leq \beta\leq 2$.
Similarly to [@Lunardi (8.1.1)], [@Ostermann4] and [@Stig2 (5.3)], we make the following assumption on the nonlinear function.
\[assumption3\] The function $F : [0,T]\times H\longrightarrow H$ is assumed to be twice differentiable with respect to the first and second variables and with bounded partial derivatives, i.e. there exists $K_3\geq 0$ such that for $k=\{1,2\}$ we have [$$\begin{aligned}
\left\Vert\frac{\partial^2F}{\partial t\partial u}(t,u)\right\Vert_{L(H)}&\leq& K_3,\quad\left\Vert\frac{\partial^kF}{\partial t^k} (t,u)\right\Vert\leq K_3(1+\Vert u\Vert),\quad t\in[0,T],\quad u\in H,\\
\left\Vert\frac{\partial F}{\partial u} (t,u)\right\Vert_{L(H)}&\leq& K_3,\quad \left\Vert\frac{\partial^2 F}{\partial u^2} (t,u)\right\Vert_{L(H\times H, H)}\leq K_3,\quad t\in[0,T],\quad u\in H.\end{aligned}$$ ]{}
Moreover, we assume assume $F'(t, u)$ to be coercive for $t\in[0, T]$ and $u\in H$, i.e. there exists $\kappa > -b_0$ such that $$\begin{aligned}
\label{prix}
-\left\langle F'(t,u)v, v\right\rangle_H\geq \kappa \Vert v\Vert^2,\quad t\in[0, T],\quad v, u\in H,\\
b_0= \underset{ t \geq 0} {\inf} \{ \text{Re} (\lambda (t)),\, \lambda (t) \in \sigma (A(t)) \,\,(\text{spectrum of } A(t))\}\,\end{aligned}$$ where $F'(t,u):=\frac{\partial F}{\partial u}(t,u)$. We also assume the nonlinear function $F$ to satisfy the Lipschitz condition, i.e. there exists a constant $K_4\geq 0$ such that $$\begin{aligned}
\label{Lipschitz}
\Vert F(t,u)-F(s,v)\Vert\leq K_4(\vert t-s\vert+\Vert u-v\Vert),\quad s,t\in[0,T],\quad u,v\in H.\end{aligned}$$
Indeed from the coercivity [[(\[ellip2\])]{}]{}, we can take $b_0= \lambda_0$.
The following theorem provides the well posedness of problem [[(\[model\])]{}]{}.
\[theorem1\] Let [[Assumption \[assumption1\]]{}]{}, [[Assumption \[assumption2\]]{}]{} and [[Assumption \[assumption3\]]{}]{} be fulfilled. Then the initial value problem [[(\[model\])]{}]{} has a unique mild solution $u(t)$ given by $$\begin{aligned}
\label{mild0}
u(t)=U(t,0)u_0+\int_0^tU(t,s)F(s,u(s))ds,\quad t\in(0,T],\end{aligned}$$ where $U(t,s)$ is the evolution system defined in [[Remark \[remark2\]]{}]{}. Moreover, the following space regularity holds $$\begin{aligned}
\label{spacereg1}
\Vert (-A(0))^{\beta/2}u(t)\Vert\leq C\left(1+\Vert (-A(0))^{\beta/2}u_0\Vert\right),\quad \beta\in[0,2),\quad t\in[0,T].\end{aligned}$$
[[Theorem \[theorem1\]]{}]{} is an extension of [@Pazy Chapter 5, Theorem 7.1] to the full semilinear problem. Its proof can be done using arguments based on a fixed point theorem and the Gronwall’s lemma as of [@Pazy Chpater 6, Theorem 1.2]. The proof of [[(\[spacereg1\])]{}]{} follows from the regularities estimates of the evolution parameter $U(t,s)$.
Finite element discretization {#numscheme}
-----------------------------
For the seek of simplicity, we assume the family of linear operators $A(t)$ to be of second order and has the following form $$\begin{aligned}
\label{family}
A(t)u=\sum_{i,j=1}^d\frac{\partial}{\partial x_i}\left(q_{ij}(x,t)\frac{\partial u}{\partial x_j}\right)-\sum_{j=1}^dq_j(x,t)\frac{\partial u}{\partial x_j}.\end{aligned}$$ We require the coefficients $q_{i,j}$ and $q_j$ to be smooth functions of the variable $x\in\overline{\Lambda}$ and Hölder-continuous with respect to $t\in[0,T]$. We further assume that there exists a positive constant $c$ such that the following ellipticity condition holds $$\begin{aligned}
\label{ellip}
\sum_{i,j=1}^dq_{ij}(x,t)\xi_i\xi_j\geq c\vert \xi\vert^2, \quad (x,t)\in\overline{\Lambda}\times [0,T].\end{aligned}$$ Under the above assumptions on $q_{ij}$ and $q_j$, it is well known that the family of linear operators defined by [[(\[family\])]{}]{} fulfills [[Assumption \[assumption2\]]{}]{} (i)-(ii) with $D=H^2(\Lambda)\cap H^1_0(\Lambda)$, see [@Pazy Section 7.6] or [@Tanabe Section 5.2]. The above assumptions on $q_{ij}$ and $q_j$ also imply that [[Assumption \[assumption2\]]{}]{} (iii) is fulfilled, see e.g. [@Praha Example 6.1] or [@Amann; @Seely].
As in [@Suzuki; @Antonio1], we introduce two spaces $\mathbb{H}$ and $V$, such that $\mathbb{H}\subset V$, depending on the boundary conditions for the domain of the operator $-A(t)$ and the corresponding bilinear form. For Dirichlet boundary conditions we take $$\begin{aligned}
V=\mathbb{H}=H^1_0(\Lambda)=\{v\in H^1(\Lambda) : v=0\quad \text{on}\quad \partial \Lambda\}.\end{aligned}$$ For Robin boundary condition and Neumann boundary condition, which is a special case of Robin boundary condition ($\alpha_0=0$), we take $V=H^1(\Lambda)$ and $$\begin{aligned}
\mathbb{H}=\{v\in H^2(\Lambda) : \partial v/\partial v_A+\alpha_0v=0,\quad \text{on}\quad \partial \Lambda\}, \quad \alpha_0\in\mathbb{R}.\end{aligned}$$ Using Green’s formula and the boundary conditions, we obtain the corresponding bilinear form associated to $-A(t)$ $$\begin{aligned}
a(t)(u,v)=\int_{\Lambda}\left(\sum_{i,j=1}^dq_{ij}(x,t)\dfrac{\partial u}{\partial x_i}\dfrac{\partial v}{\partial x_j}+\sum_{i=1}^dq_i(x,t)\dfrac{\partial u}{\partial x_i}v\right)dx, \quad u,v\in V,\end{aligned}$$ for Dirichlet boundary conditions and $$\begin{aligned}
a(t)(u,v)=\int_{\Lambda}\left(\sum_{i,j=1}^dq_{ij}(x,t)\dfrac{\partial u}{\partial x_i}\dfrac{\partial v}{\partial x_j}+\sum_{i=1}^dq_i(x,t)\dfrac{\partial u}{\partial x_i}v\right)dx+\int_{\partial\Lambda}\alpha_0uvdx.\end{aligned}$$ for Robin and Neumann boundary conditions. Using Gårding’s inequality, it holds that there exist two constants $\lambda_0$ and $c_0$ such that $$\begin{aligned}
a(t)(v,v)\geq \lambda_0\Vert v \Vert^2_{1}-c_0\Vert v\Vert^2, \quad v\in V,\quad t\in[0,T].\end{aligned}$$ By adding and subtracting $c_{0}u $ on the right hand side of (\[model\]), we obtain a new family of linear operators that we still denote by $A(t)$. Therefore the new corresponding bilinear form associated to $-A(t)$ still denoted by $a(t)$ satisfies the following coercivity property $$\begin{aligned}
\label{ellip2}
a(t)(v,v)\geq \; \lambda_0\Vert v\Vert_{1}^{2},\;\;\;\;\; v \in V,\quad t\in[0,T].\end{aligned}$$ Note that the expression of the nonlinear term $F$ has changed as we included the term $-c_{0}u$ in a new nonlinear term that we still denote by $F$.
The coercivity property (\[ellip2\]) implies that $A(t)$ is sectorial on $L^2(\Lambda)$, see e.g. [@Stig2]. Therefore $A(t)$ generates an analytic semigroup $S_t(s)=e^{s A(t)}$ on $L^{2}(\Lambda)$ such that [@Henry] $$\begin{aligned}
S_t(s)= e^{s A(t)}=\dfrac{1}{2 \pi i}\int_{\mathcal{C}} e^{ s\lambda}(\lambda I - A(t))^{-1}d \lambda,\;\;\;\;\;\;\;
\;s>0,\end{aligned}$$ where $\mathcal{C}$ denotes a path that surrounds the spectrum of $A(t)$. The coercivity property [[(\[ellip2\])]{}]{} also implies that $-A(t)$ is a positive operator and its fractional powers are well defined and for any $\alpha>0$ we have $$\label{fractional}
\left\{\begin{array}{rcl}
(-A(t))^{-\alpha} & =& \frac{1}{\Gamma(\alpha)}\displaystyle\int_0^\infty s^{\alpha-1}{\rm e}^{sA(t)}ds,\\
(-A(t))^{\alpha} & = & ((-A(t))^{-\alpha})^{-1},
\end{array}\right.$$ where $\Gamma(\alpha)$ is the Gamma function (see [@Henry]). The domain of $(-A(t))^{\alpha/2}$ are characterized in [@Suzuki; @Stig1; @Stig2] for $1\leq \alpha\leq 2$ with equivalence of norms as follows. $$\begin{aligned}
\mathcal{D}((-A(t))^{\alpha/2})&=&H^1_0(\Lambda)\cap H^{\alpha}(\Lambda)\hspace{1cm}
\text{(for Dirichlet boundary condition)}\nonumber\\
\mathcal{D}(-A(t))&=&\mathbb{H},\quad \mathcal{D}((-A(t))^{1/2})=H^1(\Lambda)\hspace{0.5cm} \text{(for Robin boundary condition)}\nonumber\\
\Vert v\Vert_{H^{\alpha}(\Lambda)}&\equiv& \Vert ((-A(t))^{\alpha/2}v\Vert:=\Vert v\Vert_{\alpha},\quad v\in \mathcal{D}((-A(t))^{\alpha/2}).\nonumber\end{aligned}$$ The characterization of $\mathcal{D}((-A(t))^{\alpha/2})$ for $0\leq \alpha<1$ can be found in [@Nambu Theorem 2.1 & Theorem 2.2].
Let us now move to the space approximation of problem [[(\[model\])]{}]{}. We start with the discretization of our domain $\Lambda$ by a finite triangulation. Let $\mathcal{T}_h$ be a triangulation with maximal length $h$. Let $V_h \subset V$ denotes the space of continuous and piecewise linear functions over the triangulation $\mathcal{T}_h$. As in [@Luskin (1.6)], we assume that $$\begin{aligned}
\label{ritz0}
\inf_{\phi_h\in V_h}\Vert v-\phi_h\Vert_j\leq Ch^{r-j}\Vert v\Vert_r,\quad v\in V\cap H^r(\Lambda),\quad r\in\{1,2\},\end{aligned}$$ for all $j\in\{0,1\}$. Moreover, we assume that $$\begin{aligned}
\label{ritz0a}
\inf_{\phi_h\in V_h}\Vert v-\phi_h\Vert_2\leq C\Vert v\Vert_2,\quad v\in V\cap H^2(\Lambda).\end{aligned}$$ We consider the projection $P_h$ defined from $H=L^2(\Lambda)$ to $V_h$ by $$\begin{aligned}
\label{discrete1}
(P_hu,\chi)=(u,\chi), \quad \chi\in V_h,\, u\in H.\end{aligned}$$ For all $t\in[0, T]$, the discrete operator $A_h(t) : V_h\longrightarrow V_h$ is defined by $$\begin{aligned}
\label{discrete2}
(A_h(t)\phi,\chi)=(A(t)\phi,\chi)=-a(t)(\phi,\chi),\quad \phi,\chi\in V_h.\end{aligned}$$ The coercivity property [[(\[ellip2\])]{}]{} implies that there exist two constants $C_2>0$ and $\theta\in(\frac{1}{2}\pi,\pi)$ such that (see e.g. [@Stig2 (2.9)] or [@Suzuki; @Henry]) $$\begin{aligned}
\label{sectorial1}
\Vert (\lambda\mathbf{I}-A_h(t))^{-1}\Vert_{L(H)}\leq \frac{C_2}{\vert \lambda\vert},\quad \lambda \in S_{\theta}
\end{aligned}$$ holds uniformly for $h>0$ and $t\in[0,T]$. The coercivity condition [[(\[ellip2\])]{}]{} implies that for any $t\in[0,T]$, $A_h(t)$ generates an analytic semigroup $S^h_t(s):=e^{sA_h(t)}$, $s\in[0,T]$. The coercivity property [[(\[ellip2\])]{}]{} also implies that the smooth properties [[(\[smooth\])]{}]{} and [[(\[smootha\])]{}]{} hold for $A_h$ uniformly for $h>0$ and $t\in[0,T]$, i.e. for all $\alpha\geq 0$ and $\delta\in[0,1]$, there exists a positive constant $C_3$ such that the following estimates hold uniformly for $h>0$ and $t\in[0,T]$, see e.g. [@Suzuki; @Henry] $$\begin{aligned}
\label{smooth2}
\left\Vert(-A_h(t))^{\alpha}e^{sA_h(t)}\right\Vert_{L(H)}&\leq& C_3s^{-\alpha}, \quad s>0, \\
\label{smooth1}
\left\Vert (-A_h(t))^{-\delta}\left(\mathbf{I}-e^{sA_h(t)}\right)\right\Vert_{L(H)}&\leq& C_3s^{\delta}, \quad s\geq 0.
\end{aligned}$$ The semi-discrete in space of problem [[(\[model\])]{}]{} consists of finding $u^h(t)\in V_h$ such that $$\begin{aligned}
\label{semi1}
\dfrac{du^h(t)}{dt}=A_h(t)u^h(t)+P_hF(t,u^h(t)), \quad u^h(0)=P_hu_0,\quad t\in(0,T].\end{aligned}$$
Fully discrete scheme and main result
-------------------------------------
Throughout this paper, without loss of generality, we use a fixed time step $\Delta t=T/M$, $M\in\mathbb{N}$ and we set $t_m=m\Delta t$, $m\in \mathbb{N}$. The time discretization consists of computing the numerical approximation $u^h_m$ of $u^h(t_m)$ at discrete times $t_m=m\Delta t \in (0,T]$, $\Delta t >0$, $m=0,\cdots,M$. Let us build an explicit scheme, efficient to solve [[(\[model\])]{}]{}. The method is based on the following linearisation of [[(\[semi1\])]{}]{} at each time step, aiming to weaken the nonlinear part $$\begin{aligned}
\label{semi}
\dfrac{du^h(t)}{dt}=\left[A_h(t)+J_m^h\right]u^h(t)+a^h_mt+G^h_m(t,u^h(t)), \quad t_m\leq t\leq t_{m+1},\end{aligned}$$ for $m=0,\cdots, M-1$, where the derivatives $J_m^h$ and $a^h_m$ are respectively the partial derivatives of $F$ at $\left(t_m+\frac{\Delta t}{2}, u^h_m\right)$ with respect to $u$ and $t$, given by $$\begin{aligned}
\label{remainder1}
J^h_m :=P_h\frac{\partial F}{\partial u}\left(t_m+\frac{\Delta t}{2},u^h_m\right)\quad \text{and}\quad a^h_m:=P_h\frac{\partial F}{\partial t}\left(t_m+\frac{\Delta t}{2}, u^h_m\right)\end{aligned}$$ and the remainder $G^h_m$ is given by $$\begin{aligned}
\label{remainder2}
G^h_m(t,u^h(t)) :=P_hF(t,u^h(t))-J_m^hu^h(t)-a^h_mt.\end{aligned}$$ Note that using [[Assumption \[assumption3\]]{}]{} the following estimate holds $$\begin{aligned}
\label{remainder3}
\Vert J^h_mu-J^h_mv\Vert_{L(H)}\leq K_3\Vert u-v\Vert,\quad u, v\in H, \quad h>0,\quad m=0,\cdots,M.\end{aligned}$$ It follows therefore from [[(\[remainder3\])]{}]{}, [[(\[Lipschitz\])]{}]{} and [[(\[remainder2\])]{}]{} that the remainder $G^h_m$ satisfies the following Lipschitz estimate $$\begin{aligned}
\label{remainder4}
\Vert G^h_m(t,u)-G^h_m(t,v)\Vert \leq (K_3+K_4)\Vert u-v\Vert,\quad u,v \in H,\quad t\in[0,T].\end{aligned}$$ Applying the exponential-like Euler and Midpoint integrators [@LEM] to [[(\[semi\])]{}]{} gives the following numerical scheme, called Magnus-Rosenbrock method (MAGROS) [$$\begin{aligned}
\label{erem}
u^h_{m+1}&=& e^{\Delta t\left(A_{h,m}+J^h_m\right)}u^h_m+\Delta t\varphi_1\left(\Delta t(A_{h,m}+J^h_m)\right)a^h_m\left(t_m+\frac{\Delta t}{2}\right)\nonumber\\
&+&\Delta t\varphi_1\left(\Delta t(A_{h,m}+J^h_m)\right)G^h_m\left(t_m+\frac{\Delta t}{2},u^h_m\right),\quad m=0,\cdots,M-1,\end{aligned}$$ ]{} where the linear operator $A_{h,m}$ is given by $$\begin{aligned}
\label{defA}
A_{h,m}:=A_h\left(t_m+\frac{\Delta t}{2}\right)\end{aligned}$$ and the linear function $\varphi_1$ is given by $$\begin{aligned}
\label{phi1}
\quad\varphi_1\left(\Delta t\left(A_{h,m}+J^h_m\right)\right)&:=&\frac{1}{\Delta t}\int_0^{\Delta t}e^{\left(A_{h,m}+J^h_m\right)(\Delta t-s)}ds. \end{aligned}$$ Note that the numerical scheme [[(\[erem\])]{}]{} can be written in the following form, efficient for simulation $$\begin{aligned}
\label{erem1}
u^h_{m+1}=u^h_m+\Delta t\varphi_1\left(\Delta t(A_{h,m}+J^h_m)\right)\left[A_{h,m}u^h_m+P_hF\left(t_m+\frac{\Delta t}{2},u^h_m\right)\right].\end{aligned}$$ The numerical scheme [[(\[erem\])]{}]{} can also be written in the following integral form, useful for the error analysis $$\begin{aligned}
\label{semi2}
u^h_{m+1}&=&e^{\Delta t\left(A_{h,m}+J^h_m\right)}u^h_m+\int_0^{\Delta t}e^{\left(A_{h,m}+J^h_m\right)(\Delta t-s)}a^h_m\left(t_m+\frac{\Delta t}{2}\right)ds\nonumber\\
&+&\int_0^{\Delta t}e^{\left(A_{h,m}+J^h_m\right)(\Delta t-s)}G^h_m\left(t_m+\frac{\Delta t}{2},u^h_m\right)ds.\end{aligned}$$ We will need the following further assumption on the nonlinearity, useful to achieve full convergence order $2$ in space without any logarithmic perturbation when $u_0\in \mathcal{D}(-A(0))$. This assumption was also used in [@Antonio1 Remark 2.9].
\[assumption4\] We assume that $ F:[0,T]\times H\longrightarrow H$ satisfies the following estimate $$\begin{aligned}
\Vert (-A(s))^{\gamma}F(t, u(r))\Vert \leq C(\gamma)\left(1+\Vert (-A(s))^{\gamma}u(r)\Vert\right),\quad s,r,t\in[0,T],\end{aligned}$$ for any $\gamma>0$ small enough.
We can now state our convergence result, which is in fact the main result of this paper.
**\[Main result\]** \[mainresult1\] Let [[Assumption \[assumption2\]]{}]{}, [[Assumption \[assumption1\]]{}]{} and [[Assumption \[assumption3\]]{}]{} be fulfilled.
- If $0<\beta<2$, then the following error estimate holds $$\begin{aligned}
\Vert u(t_m)-u^h_m\Vert\leq C\left(h^{\beta}+\Delta t^{1+\beta/2-\epsilon}\right),\end{aligned}$$ where $\epsilon>0$ is a positive constant small enough.
- If $\beta=2$, then the following error estimate holds $$\begin{aligned}
\Vert u(t_m)-u^h_m\Vert\leq C\left(h^{2}\left(1+\max\left(0, \ln(t_m/h)\right)\right)+\Delta t^{2-\epsilon}\right).\end{aligned}$$
- If $\beta=2$ and moreover if [[Assumption \[assumption4\]]{}]{} is fulfilled then the following error estimate holds $$\begin{aligned}
\Vert u(t_m)-u^h_m\Vert\leq C\left(h^{2}+\Delta t^{2-\epsilon}\right).\end{aligned}$$
[[Theorem \[mainresult1\]]{}]{} extends the result in [@Ostermann1] to a fully semilinear problem with nonsmooth initial data. Note that the linearisation technique allows to achieve convergence order almost $2$ when $u_0\in\mathcal{D}(-A(0))$.
Proof of the main result {#proof1}
========================
Preliminaries results {#preliminaires}
---------------------
The following lemma will be useful in our convergence proof.
\[lemma0\] Let [[Assumption \[assumption2\]]{}]{} be fulfilled. Then for any $\gamma\in[0,1]$ the following estimates hold [$$\begin{aligned}
\label{equidiscrete1}
K^{-1}\Vert (-(A_h(0))^{-\gamma}v\Vert&\leq& \Vert ((-A_h(t))^{-\gamma}v\Vert\leq K\Vert ((-A_h(0))^{-\gamma}v\Vert,\quad v\in V_h,\\
\label{equidiscrete2}
K^{-1}\Vert (-(A_h(0))^{\gamma}v\Vert&\leq& \Vert ((-A_h(t))^{\gamma}v\Vert\leq K\Vert ((A_h(0))^{\gamma}v\Vert,\quad v\in V_h,\end{aligned}$$ ]{} uniformly in $h>0$ and $t\in[0,T]$, where $K$ is a positive constant independent of $t$ and $h$.
We only prove [[(\[equidiscrete1\])]{}]{} since the proof of [[(\[equidiscrete2\])]{}]{} is similar to [@Antjd1 Lemma 1] by using [[Assumption \[assumption2\]]{}]{} (iii). For relatively smooth coefficients ($q_j\in C^1(\Lambda)$), the formal adjoint of $A(t)$ denoted by $A^*(t)$ is given by (see e.g. [@Evans Section 6.2.3]) [ $$\begin{aligned}
\label{adjoint1}
A^*(t)=\sum_{i,j=1}^d\frac{\partial}{\partial x_j}\left(q_{ij}(x,t)\frac{\partial}{\partial x_i}\right)+\sum_{j=1}^dq_j(x,t)\frac{\partial}{\partial x_j}+\left(\sum_{j=1}^d\frac{\partial q_j}{\partial x_j}(x,t)\right)\mathbf{I},
\end{aligned}$$ ]{} for any $t\in[0,T]$. It follows therefore from [[(\[adjoint1\])]{}]{} that $\mathcal{D}(-A^*(t))=\mathcal{D}(-A(t))$ for all $t\in[0,T]$. It also follows from [[(\[adjoint1\])]{}]{} that the coefficients of $A^*(t)$ satisfy the same assumptions as that of $A(t)$. Therefore from [@Praha Example 6.1] or [@Amann; @Seely] it holds that $A^*(t)$ satisfies [[Assumption \[assumption2\]]{}]{} (iii). More precisely, for all $\alpha\in[0,1]$ and $t\in[0,1]$, $\mathcal{D}((-A^*(t))^{\alpha})=\mathcal{D}((-A^*(0))^{\alpha})$ and for all $v\in\mathcal{D}((-A^*(0))^{\alpha})$ it holds that $$\begin{aligned}
\label{adjoint2}
C^{-1}\Vert (-A^*(0))^{\alpha}v\Vert\leq \Vert (-A^*(t))^{\alpha}v\Vert\leq C\Vert (-A^*(0))^{\alpha}v\Vert,\quad t\in[0,T].
\end{aligned}$$ Note that for all $t\in[0,T]$, $(A^*(t))_h=A_h^*(t)$, where $(A^*(t))_h$ stands for the discrete operator associated to $A^*(t)$ and $A_h^*(t)$ is the adjoint of $A_h(t)$. Indeed using [[(\[discrete2\])]{}]{}, it holds that $$\begin{aligned}
\label{adjoint2b}
\langle (A^*(t))_hv,\chi\rangle_H&=&\langle A^*(t)v,\chi\rangle_H=\langle v, A(t)\chi\rangle_H=\langle A(t)\chi, v\rangle_H\nonumber\\
&=&\langle A_h(t)\chi,v\rangle_H=\langle \chi, A_h^*(t)v\rangle_H\nonumber\\
&=&\langle A_h^*(t)v,\chi\rangle_H,\quad \chi, v\in V_h, \quad t\in[0,T],
\end{aligned}$$ and therefore $(A^*(t))_h=A_h^*(t)$ for all $t\in[0,T]$. Let us recall the following equivalence of norms [@Stig2 (2.12)], where we replace $A$ by $A^*(t)$ $$\begin{aligned}
\label{adjoint2ad}
\Vert (-A_h^*(t))^{1/2}v\Vert\approx \Vert(- A^*(t))^{1/2}v\Vert,\quad v\in V_h, \quad t\in[0,T].\end{aligned}$$ Using [[(\[adjoint2\])]{}]{} and [[(\[adjoint2ad\])]{}]{} it holds that there exists a positive constant $K$ such that $$\begin{aligned}
\label{equisiam1}
K^{-1}\Vert (-(A_h^*(0))^{1/2}v\Vert\leq \Vert ((-A_h^*(t))^{1/2}v\Vert\leq K\Vert ((-A_h^*(0))^{1/2}v\Vert,\end{aligned}$$ for any $t\in[0,T]$ and $v\in V_h$. Following closely [@Stig2] or [@Stig3 (3.7)], it holds that $$\begin{aligned}
\label{adjoint3}
\Vert(- A_h(t))^{-1/2}v\Vert&=&\sup_{v_h\in V_h}\frac{\vert\langle (-A_h(t))^{-1/2}v,v_h\rangle_H\vert}{\Vert v_h\Vert}\nonumber\\
&=&\sup_{v_h\in V_h}\frac{\vert \langle v, (-A_h^*(t))^{-1/2}v_h\rangle_H\vert}{\Vert v_h\Vert}\nonumber\\
&=&\sup_{w_h\in V_h}\frac{\vert \langle v, w_h\rangle_H\vert}{\Vert (-A_h^*(t))^{1/2}w_h\Vert},\quad v\in V_h.\end{aligned}$$ Using [[(\[equisiam1\])]{}]{} yields $$\begin{aligned}
\label{eclair1}
\sup_{w_h\in V_h}\frac{\vert \langle v, w_h\rangle_H\vert}{K\Vert (-A_h^*(0))^{1/2}w_h\Vert}&\leq& \sup_{w_h\in V_h}\frac{\vert \langle v, w_h\rangle_H\vert}{\Vert (-A_h^*(t))^{1/2}w_h\Vert}\nonumber\\
&\leq& K\sup_{w_h\in V_h}\frac{\vert \langle v, w_h\rangle_H\vert}{\Vert (-A_h^*(0))^{1/2}w_h\Vert}\end{aligned}$$ Combining [[(\[adjoint3\])]{}]{} with [[(\[eclair1\])]{}]{} yields [$$\begin{aligned}
\label{equidiscrete3}
K^{-1}\Vert (-A_h(0))^{-1/2}v\Vert\leq \Vert (-A_h(t))^{-1/2}v\Vert\leq\Vert (-A_h(0))^{-1/2}v\Vert,\quad v\in V_h\end{aligned}$$ ]{} for all $t\in[0,T]$. Note that [[(\[equidiscrete3\])]{}]{} obviously holds if we replace $1/2$ by $0$ and by $1$. The proof of the lemma is therefore completed by interpolation theory.
For $t\in[0,T]$, we introduce the Ritz projection $R_h(t) :V\longrightarrow V_h$ defined by $$\begin{aligned}
\label{ritz1}
\langle -A(t)R_h(t)v,\chi\rangle_H=\langle -A(t)v,\chi\rangle_H=a(t)(v,\chi),\quad v\in V,\quad \chi\in V_h.\end{aligned}$$ Under the regularity assumptions on the triangulation [[(\[ritz0\])]{}]{} and in view of the V-ellipticity condition [[(\[ellip\])]{}]{}, it is well known (see e.g. [@Luskin (3.2)] or [@Ciarlet; @Suzuki]) that the following error estimate holds $$\begin{aligned}
\label{ritz2}
\Vert R_h(t)v-v\Vert+h\Vert R_h(t)v-v\Vert_{H^1(\Lambda)}\leq Ch^{r}\Vert v\Vert_{H^{r}(\Lambda)},\quad v\in V\cap H^{r}(\Lambda),\end{aligned}$$ for any $r\in[1,2]$. Moreover, using [[(\[ritz0a\])]{}]{} it holds that $$\begin{aligned}
\label{ritz0b}
\Vert R_h(t)v-v\Vert_{H^2(\Lambda)}\leq C\Vert v\Vert_2,\quad v\in V\cap H^2(\Lambda),\quad t\in[0,T].\end{aligned}$$ The following error estimate also holds (see e.g. [@Luskin (3.3)] or [@Ciarlet; @Suzuki]) [$$\begin{aligned}
\label{ritz3}
\Vert D_t\left(R_h(t)v-v\right)\Vert+h\Vert D_t\left(R_h(t)v-v\right)\Vert_{H^1(\Lambda)}\leq Ch^{r}\left(\Vert v\Vert_{H^{r}(\Lambda)}+\Vert D_tv\Vert_{H^{r}(\Lambda)}\right),\end{aligned}$$ ]{} for any $r\in[1,2]$ and $v\in V\cap H^r(\Lambda)$, where $D_t:=\frac{\partial }{\partial t}$. The following lemma will be useful in our convergence proof.
\[lemma0a\] Under [[Assumption \[assumption2\]]{}]{}, the following estimates hold [$$\begin{aligned}
\label{ref3}
\Vert (A_h(t)-A_h(s))(-A_h(r))^{-1}u^h\Vert&\leq& C\vert t-s\vert\Vert u^h\Vert,\quad r,s,t\in[0,T],\quad u^h\in V_h,\\
\label{ref2}
\Vert (-A_h(r))^{-1}\left(A_h(s)-A_h(t)\right)u^h\Vert&\leq& C\vert s-t\vert\Vert u^h\Vert,\quad r,s,t\in[0,T],\quad u^h\in V_h\cap D.\end{aligned}$$ ]{} Moreover for any $u^h\in V_h\cap \mathcal{D}\left(\left(-A(0)\right)^{1-\alpha_2}\right)$ the following estimate holds [$$\begin{aligned}
\label{ref4}
\Vert (-A_h(0))^{-\alpha_1}(A_h(t)-A_h(s))(-A_h(0))^{-\alpha_2}u^h\Vert&\leq& C\vert t-s\vert\Vert u^h\Vert,\quad s,t\in[0,T].\end{aligned}$$ ]{}
Using the definition of $A_h(t)$ and $A_h(s)$ yields $$\begin{aligned}
\label{ref5}
&&\Vert (A_h(t)-A_h(s))(-A_h(r))^{-1}u^h\Vert^2\nonumber\\
&=&\left\langle((A_h(t)-A_h(s))(-A_h(r))^{-1}u^h,((A_h(t)-A_h(s))(-A_h(r))^{-1}u^h\right\rangle_H \nonumber\\
&=&\left\langle((A(t)-A(s))(-A_h(r))^{-1}u^h,((A_h(t)-A_h(s))(-A_h(r))^{-1}u^h\right\rangle_H.\end{aligned}$$ Using Cauchy’s Schwartz inequality, the relation $A_h(r)R_h(r)=P_hA(r)$ (see e.g. [@Antonio1; @Stig2]), [[Assumption \[assumption2\]]{}]{} (ii) and the boundness of $R_h(r)$ yields $$\begin{aligned}
\label{ref6a}
&&\Vert (A_h(t)-A_h(s))(-A_h(r))^{-1}u^h\Vert\nonumber\\
&\leq& C\Vert ((A(t)-A(s))(-A_h(r))^{-1}u^h\Vert\nonumber\\
&=&C\Vert ((A(t)-A(s))(-A_h(r))^{-1}P_hu^h\Vert\nonumber\\
&=&C\Vert ((A(t)-A(s))R_h(r)(-A(r))^{-1}u^h\Vert\nonumber\\
&=&C\Vert ((A(t)-A(s))(-A(r))^{-1}(-A(r))R_h(r)(-A(r))^{-1}u^h\Vert\nonumber\\
&\leq&C\vert t-s\vert\Vert (-A(r))R_h(r)(-A(r))^{-1}u^h\Vert.\end{aligned}$$ Using triangle inequality and [[(\[ritz0b\])]{}]{} yields $$\begin{aligned}
\label{ref6b}
&&\Vert (-A(r))R_h(r)(-A(r))^{-1}u^h\Vert\nonumber\\
&\leq& \Vert (-A(r))R_h(r)(-A(r))^{-1}u^h-A(r)(-A(r))^{-1}u^h\Vert+\Vert A(r)(-A(r))^{-1}u^h\Vert\nonumber\\
&=&\left\Vert A(r)\left(R_h(r)(-A(r))^{-1}u^h-(-A(r))^{-1}u^h\right)\right\Vert+\Vert u^h\Vert\nonumber\\
&=&\Vert R_h(r)(-A(r))^{-1}u^h-(-A(r))^{-1}u^h\Vert_{H^2(\Lambda)}+\Vert u^h\Vert
\nonumber\\
&\leq &C\Vert (-A(r))^{-1}u^h\Vert_{H^2(\Lambda)}+\Vert u^h\Vert\nonumber\\
&\leq& C\Vert u^h\Vert.\end{aligned}$$ Substituting [[(\[ref6b\])]{}]{} in [[(\[ref6a\])]{}]{} yields $$\begin{aligned}
\label{ref6}
\Vert (A_h(t)-A_h(s))(-A_h(r))^{-1}u^h\Vert\leq C\vert t-s\vert \Vert u^h \Vert.\end{aligned}$$ This completes the proof of [[(\[ref3\])]{}]{}.
To prove [[(\[ref2\])]{}]{}, as in [@Antonio2] or [@Stig2] we set $V_{r}=\mathcal{D}(-A(r))$, $V_{r}^h=\mathcal{D}(-A_h(r))$, so $V'_{r}=\mathcal{D}\left((-A(r))^{-1}\right)$. Following [@Antonio2 (67)] or [@Stig2], we have [$$\begin{aligned}
\left\Vert (-A_h(r))^{-1}\left(A_h(s)-A_h(t)\right)u^h\right\Vert=\sup_{v_h\in V_r^h}\frac{\left\langle \left(A_h(s)-A_h(t)\right)u^h,(-A^*_h(r))^{-1}v_h\right\rangle_H}{\Vert v_h\Vert}\end{aligned}$$ ]{} Using the definition of $A_h(s)$ and $A_h(t)$, it holds that [$$\begin{aligned}
\left\Vert (-A_h(r))^{-1}\left(A_h(s)-A_h(t)\right)u^h\right\Vert&=&\sup_{v_h\in V_{r}^h}\frac{\left\langle \left(A(s)-A(t)\right)u^h,(-A^*_h(r))^{-1}v_h\right\rangle_H}{\Vert v_h\Vert}\nonumber\\
&=&\sup_{w_h\in V_{r}^h}\frac{\left\langle \left(A(s)-A(t)\right)u^h, w_h\right\rangle_H}{\Vert (-A^*_h(r))w_h\Vert}\nonumber\\
&\leq &C\sup_{w_h\in V_{r}^h}\frac{\left\langle \left(A(s)-A(t)\right)u^h, w_h\right\rangle_H}{\Vert w_h\Vert_{V_{r}}}\nonumber\\
&=&C\left\Vert \left(A(s)-A(t)\right)u^h\right\Vert_{-1}\nonumber\\
&=&C\left\Vert (-A(r))^{-1})\left(A(s)-A(t)\right)u^h\right\Vert\nonumber\\
&\leq& C\vert s-t\vert\, \Vert u^{h} \Vert ,\end{aligned}$$ ]{} where [[Assumption \[assumption2\]]{}]{} (ii) is used at the last step. This completes the proof of [[(\[ref2\])]{}]{}. The proof of [[(\[ref4\])]{}]{} follows from [[(\[ref2\])]{}]{} and [[(\[ref3\])]{}]{} by interpolation theory.
\[lemderiv\] Let [[Assumption \[assumption2\]]{}]{} be fulfilled. Then for any $u^h\in V_h\cap \mathcal{D}\left(\left(-A(0)\right)^{1-\alpha_2}\right)$ the following estimates hold $$\begin{aligned}
\label{ba1}
\Vert (-A_h(0))^{-\alpha_1}A_h'(t)(-A_h(0))^{-\alpha_2}u^h\Vert &\leq& C\Vert u^h\Vert,\quad t\in[0,T],\\
\label{ba2}
\Vert (-A_h(0))^{-\alpha_1}A_h''(t)(-A_h(0))^{-\alpha_2}u^h\Vert &\leq& C\Vert u^h\Vert,\quad t\in[0,T],\end{aligned}$$ where $\alpha_1$ and $\alpha_2$ are defined in [[Assumption \[assumption2\]]{}]{}.
Recall that $$\begin{aligned}
\label{ba3}
A_h'(t)=\lim_{\delta\longrightarrow 0}\frac{A_h(t+\delta)-A_h(t)}{\delta}.\end{aligned}$$ The proof of [[(\[ba1\])]{}]{} is completed by combining [[(\[ref4\])]{}]{} and [[(\[ba3\])]{}]{}. The proof of [[(\[ba2\])]{}]{} follows the same lines as that of [[Lemma \[lemma0a\]]{}]{}.
\[evolutionremark\] From [[Lemma \[lemma0a\]]{}]{}, it follows [@Pazy Theorem 6.1, Chapter 5] that there exists a unique evolution system $U_h :\Delta(T)\longrightarrow L(H)$, satisfying [@Pazy (6.3), Page 149] $$\begin{aligned}
\label{ref6}
U_h(t,s)=S^h_s(t-s)+\int_s^tS^h_{\tau}(t-\tau)R^h(\tau,s)d\tau,\end{aligned}$$ where $S^h_s(t):=e^{A_h(s)t}$, $R^h(t,s):=\sum\limits_{m=1}^{\infty}R^h_m(t,s)$, with $R^h_m(t,s)$ satisfying the following recurrence relation [@Pazy (6.22), Page 153] $$\begin{aligned}
R^h_{m+1}&=&\int_s^tR^h_1(t,s)R^h_m(\tau,s)d\tau,\\
R^h_1(t,s)&:=&(A_h(s)-A_h(t))S^h_s(t-s), \quad m\geq 1\end{aligned}$$ Note also that from [@Pazy (6.6), Chpater 5, Page 150], the following identity holds $$\begin{aligned}
\label{ref7}
R^h(t,s)=R_1^h(t,s)+\int_s^tR_1^h(t,\tau)R^h(\tau,s)d\tau.\end{aligned}$$ The mild solution of [[(\[semi1\])]{}]{} is therefore given by $$\begin{aligned}
\label{mild4}
u^h(t)=U_h(t,0)P_hu_0+\int_0^tU_h(t,s)P_hF(s,u^h(s))ds.\end{aligned}$$
\[pazylemma\] Under [[Assumption \[assumption2\]]{}]{}, the evolution system $U_h :\Delta(T)\longrightarrow H$ satisfies the following properties
- $U_h(.,s)\in C^1(]s,T]; L(H))$, $0\leq s\leq T$ and $$\begin{aligned}
\frac{\partial U_h}{\partial t}(t,s)=-A_h(t)U_h(t,s), \quad 0\leq s\leq t\leq T,\\
\Vert A_h(t)U_h(t,s)\Vert_{L(H)}\leq \frac{C}{t-s},\quad 0\leq s<t\leq T.\end{aligned}$$
- $U_h(t,.)u\in C^1([0,t[; H)$, $0<t\leq T$, $u\in\mathcal{D}(A_h(0))$ and $$\begin{aligned}
\frac{\partial U_h}{\partial s}(t,s)u=-U_h(t,s)A_h(s)u,\quad 0\leq s\leq t\leq T\\
\Vert A_h(t)U_h(t,s)A_h(s)^{-1}\Vert_{L(H)}\leq C, \quad 0\leq s\leq t\leq T.\end{aligned}$$
The proof is similar to that of [@Pazy Theorem 6.1, Chapter 5] by using [[(\[smooth1\])]{}]{}, [[(\[smooth2\])]{}]{}, [[Lemma \[lemma0a\]]{}]{} and [[Lemma \[lemma0\]]{}]{}.
\[evolutionlemma\] Let [[Assumption \[assumption2\]]{}]{} be fulfilled.
- The following estimates hold $$\begin{aligned}
\label{reste1}
\Vert R^h_1(t,s)\Vert_{L(H)}\leq C,&&\quad
\Vert R^h_m(t,s)\Vert_{L(H)}\leq \frac{C}{m!}(t-s)^{m-1},\quad m\geq 1,\\
\label{reste2}
\Vert R^h(t,s)\Vert_{L(H)}\leq C,&&\quad \Vert U_h(t,s)\Vert_{L(H)}\leq C,\quad 0\leq s\leq t\leq T.\end{aligned}$$
- For any $0\leq\gamma\leq\alpha\leq 1$ and $0\leq s\leq t\leq T$, the following estimates hold $$\begin{aligned}
\label{ae1}
\Vert (-A_h(r))^{\alpha}U_h(t,s)\Vert_{L(H)}&\leq& C(t-s)^{-\alpha},\quad r\in[0,T],\\
\label{ae3}
\Vert U_h(t,s)(-A_h(r))^{\alpha}\Vert_{L(H)}&\leq& C(t-s)^{-\alpha},\quad r\in[0,T],\\
\label{ae2}
\Vert (-A_h(r))^{\alpha}U_h(t,s)(-A_h(s))^{-\gamma}\Vert_{L(H)}&\leq& C(t-s)^{\gamma-\alpha}, \quad r\in[0,T].\end{aligned}$$
- For any $0\leq s\leq t\leq T$ the following useful estimates hold $$\begin{aligned}
\label{hen1}
\Vert \left(U_h(t,s)-\mathbf{I}\right)(-A_h(s))^{-\gamma}\Vert_{L (H)}&\leq& C(t-s)^{\gamma}, \quad 0\leq \gamma\leq 1,\\
\label{hen2}
\Vert \left (-A_h(r))^{-\gamma}(U_h(t,s)-\mathbf{I}\right)\Vert_{L (H)}&\leq& C(t-s)^{\gamma}, \quad 0\leq \gamma\leq 1.\end{aligned}$$
<!-- -->
- The proof of the first estimate of [[(\[reste1\])]{}]{} follows the same lines as [@Pazy Corollary 6.3, Page 153] by using [[(\[smooth2\])]{}]{}, Lemmas \[lemma0\] and \[lemma0a\]. The proof of the second estimate of [[(\[reste1\])]{}]{} follows the same lines as [@Pazy (6.23), Page 153]. The proof of the first estimate of [[(\[reste2\])]{}]{} is similar to [@Pazy (6.26), Page 153] and the proof of the second estimate of [[(\[reste2\])]{}]{} is similar to [@Pazy (6.27), Page 153].
- The estimate of [[(\[ae1\])]{}]{} for $\alpha=1$ is given in [[Lemma \[pazylemma\]]{}]{}. The proof of [[(\[ae1\])]{}]{} for the case $0\leq\alpha<1$ follows from the integral equation [[(\[ref6\])]{}]{}. In fact pre-multiplying both sides of [[(\[ref6\])]{}]{} by $(-A_h(s))^{\alpha}$, taking the norm in both sides, using [[Lemma \[lemma0\]]{}]{} and [[(\[smooth2\])]{}]{} yields $$\begin{aligned}
\label{paz1}
\Vert (-A_h(r))^{\alpha}U_h(t,s)\Vert_{L(H)}&\leq& \Vert (-A_h(r))^{\alpha}S_s^h(t-s)\Vert_{L(H)}\nonumber\\
&+&\int_s^t\Vert (-A_h(r))^{\alpha}S_{\tau}^h(t-\tau)\Vert_{L(H)}\Vert R^h(\tau,s)\Vert_{L(H)}d\tau\nonumber\\
&\leq& C(t-s)^{-\alpha}+C\int_s^t(t-\tau)^{-\alpha}d\tau\nonumber\\
&\leq& C(t-s)^{-\alpha}.
\end{aligned}$$ This proves [[(\[ae1\])]{}]{}. The proof of [[(\[ae2\])]{}]{} and [[(\[ae3\])]{}]{} are similar to that of [[(\[ae1\])]{}]{}.
- From [[(\[ref6\])]{}]{}, it holds that $$\begin{aligned}
\label{paz2}
(U_h(t,s)-\mathbf{I})(-A_h(r))^{-\gamma}&=&(-A_h(s))^{-\gamma}\left(e^{A(s)(t-s)}-\mathbf{I}\right)\nonumber\\
&+&\int_s^tS^h_{\tau}(t-\tau)R^h(\tau,s)(-A_h(s))^{-\gamma}d\tau.
\end{aligned}$$ Taking the norm in both sides of [[(\[paz2\])]{}]{}, using [[(\[smooth1\])]{}]{}, the boundness of $(-A_h(r))^{-\gamma}$ and [[Lemma \[evolutionlemma\]]{}]{} (i) yields $$\begin{aligned}
\Vert (U_h(t,s)-\mathbf{I})(-A_h(s))^{-\gamma}\Vert_{L(H)}&=&C(t-s)^{\gamma}+C\int_s^td\tau\leq C(t-s)^{\gamma}.\nonumber
\end{aligned}$$ This completes the proof of [[(\[hen1\])]{}]{}. The proof of [[(\[hen2\])]{}]{} is similar to that of [[(\[hen1\])]{}]{}.
The following space regularity of the semi-discrete problem [[(\[semi1\])]{}]{} will be useful in our convergence analysis.
\[regularitylemma\] Let [[Assumption \[assumption2\]]{}]{} (i)-(ii), [[Assumption \[assumption1\]]{}]{} and [[Assumption \[assumption3\]]{}]{} be fulfilled with the corresponding $0\leq \beta<2$. Then for all $\gamma\in[0,\beta]$ and $\alpha\in[0,2)$ the following estimates hold $$\begin{aligned}
\label{regular1}
\Vert (-A_h(r))^{\gamma/2}u^h(t)\Vert&\leq& C,\quad\hspace{2cm} 0\leq r,t\leq T,\\
\label{regular1a}
\Vert (-A_h(0))^{\alpha/2}u^h(t)\Vert&\leq& Ct^{\beta/2-\alpha/2},\quad t\in[0, T], \quad \beta\in[0,2],
$$
We first show that $$\begin{aligned}
\label{regular8a}
\Vert u^h(t)\Vert\leq C,\quad t\in[0,T].
\end{aligned}$$ Taking the norm in both side of [[(\[mild4\])]{}]{} and using the triangle inequality yields [$$\begin{aligned}
\label{regular3}
\Vert u^h(t)\Vert\leq \Vert U_h(t,0)P_hu_0\Vert+\left\Vert\int_0^tU_h(t,s)P_hF(s,u^h(s))ds\right\Vert ds
:=I_0+I_1.\end{aligned}$$ ]{} Using [[Lemma \[evolutionlemma\]]{}]{} (i) and the uniformly boundedness of $P_h$, it holds that $$\begin{aligned}
\label{regular4}
I_0\leq \Vert u_0\Vert\leq C.\end{aligned}$$ Using [[Lemma \[evolutionlemma\]]{}]{} (i), [[Assumption \[assumption3\]]{}]{} and the uniformly boundedness of $P_h$, it holds that $$\begin{aligned}
\label{regular5}
I_1&\leq& \int_0^t\Vert U_h(t,s)P_hF(s,u^h(s))\Vert\leq C\int_0^t\left(C+\Vert u^h(s)\Vert\right)ds\nonumber\\
&\leq& C+C\int_0^t\Vert u^h(s)\Vert ds.\end{aligned}$$ Substituting [[(\[regular5\])]{}]{} and [[(\[regular4\])]{}]{} in [[(\[regular3\])]{}]{} yields $$\begin{aligned}
\label{regular8}
\Vert u^h(t)\Vert\leq C+C\int_0^t\Vert u^h(s)\Vert ds.\end{aligned}$$ Applying the continuous Gronwall’s lemma to [[(\[regular8\])]{}]{} completes the proof of [[(\[regular8a\])]{}]{}. Let us now prove [[(\[regular1\])]{}]{}. Pre-multiplying [[(\[mild4\])]{}]{} by $(-A_h(r))^{\gamma/2}$, taking the norm in both sides and using triangle inequality yields $$\begin{aligned}
\label{regular9}
\left\Vert (-A_h(r))^{\gamma/2}u^h(t)\right\Vert&\leq& \left\Vert (-A_h(r))^{\gamma/2}U_h(t,0)P_hu_0\right\Vert_{L(H)}\nonumber\\
&+&\int_0^t\left\Vert (-A_h(r))^{\gamma/2}U_h(t,s)P_hF(s,u^h(s))\right\Vert ds
\nonumber\\
&=:&II_0+II_1.\end{aligned}$$ Inserting $(-A_h(0))^{-\gamma/2}(-A_h(0))^{\gamma/2}$, using [[Lemma \[evolutionlemma\]]{}]{} (ii) and [[Lemma \[lemma0\]]{}]{}, it holds that [$$\begin{aligned}
\label{premier}
II_0\leq \Vert (-A_h(r))^{\gamma/2}U_h(t,0)(-A_h(0))^{-\gamma/2}\Vert_{L(H)}\Vert (-A_h(0))^{\gamma/2}u_0\Vert\leq C.\end{aligned}$$ ]{} Using [[Lemma \[lemma0\]]{}]{}, [[Lemma \[evolutionlemma\]]{}]{} (ii), [[Assumption \[assumption3\]]{}]{} and [[(\[regular8a\])]{}]{} yields $$\begin{aligned}
\label{deuxieme}
II_1&\leq& C\left(\int_0^t\left\Vert(-A_h(s))^{\gamma/2}U_h(t,s)\right\Vert_{L(H)}ds\right)\sup_{r\in[0,T]}\left\Vert F\left(r,u^h(r)\right)\right\Vert \nonumber\\
&\leq& C\sup_{s\in[0,T]}\left(1+\Vert u^h(s)\Vert\right)\int_0^t(t-s)^{-\gamma/2}ds\leq C.\end{aligned}$$ Substituting [[(\[deuxieme\])]{}]{} and [[(\[premier\])]{}]{} in [[(\[regular9\])]{}]{} completes the proof of [[(\[regular1\])]{}]{}. The proof of [[(\[regular1a\])]{}]{} is similar to that of [[(\[regular1\])]{}]{}. This completes the proof of [[Lemma \[regularitylemma\]]{}]{}.
Let us consider the following deterministic problem: find $w\in V$ such that $$\begin{aligned}
\label{determ1}
w'=A(t)w,\quad w(\tau)=v,\quad t\in(\tau,T].\end{aligned}$$ The corresponding semi-discrete problem in space is: find $w_h\in V_h$ such that $$\begin{aligned}
\label{determ2}
w_h'(t)=A_h(t)w_h,\quad w_h(\tau)=P_hv,\quad t\in(\tau,T],\quad \tau\geq 0.\end{aligned}$$ Let us define the operator $$\begin{aligned}
T_h(t,\tau):=U(t,\tau)-U_h(t,\tau)P_h,\end{aligned}$$ so that $w(t)-w_h(t)=T_h(t,\tau)v$. The following lemma will be useful in our convergence analysis.
\[spaceerrorlemma\] Let $r\in[0,2]$ and $\gamma\leq r$. Let [[Assumption \[assumption2\]]{}]{} be fulfilled. Then the following error estimate holds for the semi-discrete approximation [[(\[determ2\])]{}]{} $$\begin{aligned}
\label{er0}
\Vert w(t)-w_h(t)\Vert=\Vert T_h(t,\tau)v\Vert\leq Ch^r(t-\tau)^{-(r-\gamma)/2}\Vert v\Vert_{\gamma},\end{aligned}$$ for any $v\in \mathcal{D}\left(\left(-A(0)\right)^{\gamma/2}\right)$.
As in [@Antonio1 (3.5)] or [@Stig2], we set $$\begin{aligned}
\label{espa0}
w_h(t)-w(t)&=&\left(w_h(t)-R_h(t)w(t)\right)+\left(R_h(t)w(t)-w(t)\right)\nonumber\\
&\equiv& \theta(t)+\rho(t).\end{aligned}$$ Using the definition of $R_h(t)$ and $P_h$, we can prove exactly as in [@Stig2; @Antonio1] that $$\begin{aligned}
\label{espacetamb}
A_h(t)R_h(t)=P_hA(t),\quad t\in[0,T].\end{aligned}$$ One can easily compute the following derivatives $$\begin{aligned}
\label{espace1a}
\theta_t&=&A_h(t)w_h(t)-R_h'(t)w(t)-R_h(t)A(t)w(t),\\
\label{espace1b}
D_t\rho&=&R_h'(t)w(t)+R_h(t)A(t)w(t)-A(t)w(t).\end{aligned}$$ Endowing $V$ and the linear subspace $V_h$ with the $\Vert .\Vert_{H^1(\Lambda)}$ norm, it follows from [[(\[ritz2\])]{}]{} that $R_h(t)\in L(V, V_h)$ for all $t\in [0, T]$. By the definition of the differential operator, it follows that $R_h'(t)\in L(V, V_h)$ for all $t\in[0, T]$. Hence $P_hR_h'(t)=R_h'(t)$ for all $t\in[0,T]$ and it follows from [[(\[espace1b\])]{}]{} that $$\begin{aligned}
\label{espace1c}
P_hD_t\rho=R_h'(t)w(t)+R_h(t)A(t)w(t)-P_hA(t)w(t).\end{aligned}$$ Adding and subtracting $P_hA(t)w(t)$ in [[(\[espace1a\])]{}]{} and using [[(\[espacetamb\])]{}]{}, it follows that $\theta$ satisfies the following equation $$\begin{aligned}
\label{espa1}
\theta_t=A_h(t)\theta-P_hD_t\rho,\quad t\in(\tau,T],\end{aligned}$$ Since $\{A_h(t)\}_{t\in[0,T]}$ generates an evolution system $\{U_h(t,s)\}_{0\leq s\leq t\leq T}$, it holds that $$\begin{aligned}
\label{espa2}
\theta(t)=U_h(t,\tau)\theta(\tau)-\int_{\tau}^tU_h(t,s)P_hD_s\rho(s)ds.\end{aligned}$$ Splitting the integral part of [[(\[espa2\])]{}]{} into two intervals and integrating by parts over the first interval yields $$\begin{aligned}
\label{espa3}
\theta(t)&=& U_h(t,\tau)\theta(\tau)+U_h(t,\tau)P_h\rho(\tau)-U_h\left(t,(t+\tau)/2\right)P_h\rho\left((t+\tau)/2\right)\nonumber\\
&+&\int_{\tau}^{(t+\tau)/2}\frac{\partial}{\partial s}\left(U_h(t,s)\right)P_h\rho(s)ds-\int_{(t+\tau)/2}^tU_h(t,s)P_hD_s\rho(s)ds.\end{aligned}$$ Using the expression of $\theta(\tau)$, $\rho(\tau)$ and the fact that $u_h(\tau)=P_hv$, it holds that $$\begin{aligned}
\label{espa4}
\theta(\tau)+P_h\rho(\tau)=0.\end{aligned}$$ Using [[(\[espa4\])]{}]{} reduces [[(\[espa3\])]{}]{} to $$\begin{aligned}
\label{espa5}
\theta(t)&=& -U_h(t,s)P_h\rho((t+\tau)/2)+\int_{\tau}^{(t+\tau)/2}\frac{\partial}{\partial s}\left(U_h(t,s)
\right)P_h\rho(s)ds\nonumber\\
&-&\int_{(t+\tau)/2}^tU_h(t,s)P_hD_s\rho(s)ds.\end{aligned}$$ Taking the norm in both sides of [[(\[espa5\])]{}]{}, using the uniformly boundedness of $P_h$, [[(\[smooth2\])]{}]{}, Lemma \[lemma0a\] and Lemma \[evolutionlemma\] (i) yields [$$\begin{aligned}
\label{espa6}
\Vert\theta(t)\Vert&\leq& C\Vert \rho((t+\tau)/2)\Vert+\int_{\tau}^{(t+\tau)/2}\left\Vert U_h(t,s)A_h(s)\right\Vert_{L(H)}\Vert \rho(s)\Vert ds+\int_{(t+\tau)/2}^t\Vert D_s\rho(s)\Vert ds\nonumber\\
&\leq& C\Vert \rho((t+\tau)/2)\Vert+\int_{\tau}^{(t+\tau)/2}(t-s)^{-1}\Vert \rho(s)\Vert ds+\int_{(t+\tau)/2}^t\Vert D_s\rho(s)\Vert ds.\end{aligned}$$ ]{} Using [[(\[ritz2\])]{}]{}, it holds that $$\begin{aligned}
\label{espa7}
\Vert \rho(s)\Vert\leq Ch^r\Vert w(s)\Vert_r.\end{aligned}$$ Note that the solution of [[(\[determ1\])]{}]{} is represented as follows. $$\begin{aligned}
\label{encore1}
w(s)=U(s,\tau)v,\quad s\geq \tau.\end{aligned}$$ Pre-multiplying both sides of [[(\[encore1\])]{}]{} by $(-A(s))^{r/2}$, inserting an appropriate power of $-A(\tau)$, using [[Lemma \[evolutionlemma\]]{}]{} (ii) and [@Antjd1 Lemma 1] yields $$\begin{aligned}
\label{encore2}
\Vert (-A(t))^{r/2}w(s)\Vert&\leq& \Vert (-A(s))^{r/2}U(s,\tau)(-A(\tau))^{-\gamma/2}\Vert_{L(H)}\Vert (-A(\tau))^{\gamma/2}v\Vert\nonumber\\
&\leq& C(s-\tau)^{-(r-\gamma)/2}\Vert (-A(\tau))^{\gamma/2}v\Vert\nonumber\\
&\leq& C(s-\tau)^{-(r-\gamma)/2}\Vert v\Vert_{\gamma}.\end{aligned}$$ Therefore it holds that $$\begin{aligned}
\label{espa8}
\Vert w(s)\Vert_r\leq C(s-\tau)^{-(r-\gamma)/2}\Vert v\Vert_{\gamma}, \quad 0\leq \gamma\leq r\leq 2,\quad \tau<s.\end{aligned}$$ Substituting [[(\[espa8\])]{}]{} in [[(\[espa7\])]{}]{} yields $$\begin{aligned}
\label{espa8a}
\Vert \rho(s)\Vert_r\leq Ch^r(s-\tau)^{-(r-\gamma)/2}\Vert v\Vert_{\gamma}.\end{aligned}$$ Using [[(\[ritz3\])]{}]{}, it holds that $$\begin{aligned}
\label{espa13}
\Vert D_s\rho(s)\Vert \leq Ch^r(\Vert w(s)\Vert_r+\Vert D_sw(s)\Vert_r).\end{aligned}$$ Taking the derivative with respect to $s$ in both sides of [[(\[encore1\])]{}]{} yields $$\begin{aligned}
\label{encore3}
D_sw(s)=A(s)U(s,\tau)v.\end{aligned}$$ As for [[(\[encore2\])]{}]{}, pre-multiplying both sides of [[(\[encore3\])]{}]{} by $(-A(s))^{r/2}$, inserting $(-A(\tau))^{-\gamma/2}(-A(\tau))^{\gamma/2}$ and using [[Lemma \[evolutionlemma\]]{}]{} (ii) yields $$\begin{aligned}
\label{encore4a}
\Vert D_sw(s)\Vert_r\leq C(s-\tau)^{-1-(r-\gamma)/2}\Vert v\Vert_{\gamma}.\end{aligned}$$ Substituting [[(\[espa8\])]{}]{} and [[(\[encore4a\])]{}]{} in [[(\[espa13\])]{}]{} yields $$\begin{aligned}
\label{espa15}
\Vert D_s\rho(s)\Vert&\leq& Ch^r\left((s-\tau)^{-(r-\gamma)/2}\Vert v\Vert_{\gamma}+(s-\tau)^{-1-(r-\gamma)/2}\Vert v\Vert_{\gamma}\right)\nonumber\\
&\leq& Ch^r(s-\tau)^{-1-(r-\gamma)/2}\Vert v\Vert_{\gamma}.\end{aligned}$$ Substituting [[(\[espa8a\])]{}]{} and [[(\[espa15\])]{}]{} in [[(\[espa6\])]{}]{} yields $$\begin{aligned}
\label{espa17}
\Vert\theta(t)\Vert&\leq& Ch^r(t-\tau)^{-(r-\gamma)/2}\Vert v\Vert_{\gamma}\nonumber\\
&+&Ch^r\int_{\tau}^{(t+\tau)/2}(t-s)^{-1}(s-\tau)^{-(r-\gamma)/2}\Vert v\Vert_{\gamma}ds\nonumber\\
&+&Ch^r\int_{(t+\tau)/2}^t(s-\tau)^{-1-(r-\gamma)/2}\Vert v\Vert_{\gamma}ds.\end{aligned}$$ Using the estimate [$$\begin{aligned}
\label{espa18}
\int_{\tau}^{(t+\tau)/2}(t-s)^{-1}(s-\tau)^{-(r-\gamma)/2}ds+\int_{(t+\tau)/2}^t(s-\tau)^{-1-(r-\gamma)/2}ds\leq C(t-\tau)^{-(r-\gamma)/2},\nonumber\end{aligned}$$ ]{} it follows that $$\begin{aligned}
\label{espa20}
\Vert \theta(t)\Vert\leq Ch^r(t-\tau)^{-(r-\gamma)/2}\Vert v\Vert_{\gamma}.\end{aligned}$$ Substituting [[(\[espa20\])]{}]{} and [[(\[espa13\])]{}]{} in [[(\[espa0\])]{}]{} yields $$\begin{aligned}
\Vert w(t)-w_h(t)\Vert\leq \Vert\theta(t)\Vert+\Vert \rho(t)\Vert\leq Ch^r(t-\tau)^{-(r-\gamma)/2}\Vert v\Vert_{\gamma}.\end{aligned}$$ This completes the proof of [[Lemma \[spaceerrorlemma\]]{}]{}.
\[remaksemi\] [[Lemma \[spaceerrorlemma\]]{}]{} generalizes [@Antonio1 Lemma 3.1] to time dependent problems. It also generalises [@Luskin Lemmas 3.2 and 3.3] and [@Thomee1 Theorems 3 and 4] to more general boundary conditions than only Dirichlet boundary conditions. Note that the fact that the solution vanishes at the boundary is fundamental in the proof of [@Luskin Lemmas 3.2 and 3.3] and [@Thomee1 Theorems 3 and 4], where authors used energy estimates arguments.
The following theorem gives the space convergence error of the semi-discrete solution in space toward the exact solution. It is fundamental in the proof of the convergence of the fully discrete scheme.
\[proposition2\] Let [[Assumption \[assumption2\]]{}]{}, [[Assumption \[assumption1\]]{}]{} and [[Assumption \[assumption3\]]{}]{} be fulfilled. Let $u(t)$ and $u^h(t)$ be the mild solution of [[(\[model\])]{}]{} and [[(\[semi1\])]{}]{} respectively.
- If $0<\beta<2$, then the following error estimate holds $$\begin{aligned}
\label{time1}
\Vert u(t)-u^h(t)\Vert\leq Ch^{\beta},\quad 0\leq t\leq T.\end{aligned}$$
- If $\beta=2$, then the following error estimate holds $$\begin{aligned}
\label{time2}
\Vert u(t)-u^h(t)\Vert\leq Ch^{2}\left(1+\max\left(0,\ln(t/h^2)\right)\right),\quad 0< t\leq T.\end{aligned}$$
- If $\beta=2$ and if further [[Assumption \[assumption4\]]{}]{} is fulfilled, then the following error estimate holds $$\begin{aligned}
\Vert u(t)-u^h(t)\Vert\leq Ch^2,\quad 0\leq t\leq T.\end{aligned}$$
Subtracting [[(\[mild4\])]{}]{} form [[(\[mild0\])]{}]{}, taking the norm and using triangle inequality yields $$\begin{aligned}
\label{estiI}
\Vert u(t)-u^h(t)\Vert&\leq& \left\Vert U(t,0)u_0-U_h(t,0)P_hu_0\right\Vert\nonumber\\
&+&\left\Vert\int_0^{t}\left[U(t,s)F\left(s,u(s)\right)-U_h(t,s)P_hF\left(s,u^h(s)\right)\right]
ds\right\Vert\nonumber\\
& =:&III_0+III_1.\end{aligned}$$ Using [[Lemma \[spaceerrorlemma\]]{}]{} with $r=\gamma=\beta$ yields $$\begin{aligned}
\label{jour1}
III_0\leq Ch^{\beta}\Vert u_0\Vert\leq Ch^{\beta}.\end{aligned}$$ Using [[Lemma \[spaceerrorlemma\]]{}]{} with $r=\beta$ (with $\beta<2$), $\gamma=0$, [[Assumption \[assumption3\]]{}]{}, [[Lemma \[regularitylemma\]]{}]{} and [[Lemma \[evolutionlemma\]]{}]{} yields $$\begin{aligned}
\label{estiI2}
III_1&\leq& \int_0^t\left\Vert U(t,s)F\left(s,u(s)\right)-U(t,s)F\left(s,u^h(s)\right)\right\Vert ds\nonumber\\
&+&\int_0^t\left\Vert U(t,s)F\left(s,u^h(s)\right)-U_h(t,s)P_hF\left(s,u^h(s)\right)\right\Vert ds\nonumber\\
&\leq& C\int_0^t\left\Vert u(s)-u^h(s)\right\Vert_{L^2(\Omega,H)}ds+Ch^{\beta}\int_0^t(t-s)^{-\beta/2}ds\nonumber\\
&\leq& Ch^{\beta}+C\int_0^{t}\left\Vert u(s)-u^h(s)\right\Vert ds.\end{aligned}$$ Substituting [[(\[estiI2\])]{}]{} and [[(\[jour1\])]{}]{} in [[(\[estiI\])]{}]{} yields $$\begin{aligned}
\label{lat1}
\left\Vert u(t)-u^h(t)\right\Vert\leq Ch^{\beta}+C\int_0^{t}\left\Vert u(s)-u^h(s)\right\Vert ds.\end{aligned}$$ Applying the continuous Gronwall’s lemma to [[(\[lat1\])]{}]{} prove [[(\[time1\])]{}]{}. The proof of [[(\[time2\])]{}]{} is straightforward. This completes the proof of [[Proposition \[proposition2\]]{}]{}.
The following lemma extends some results in [@Luskin] (see e.g. [@Luskin Lemma 2.4, (2.8)] and [@Luskin Lemma 2.6]) to the case of fully semilinear problem. It also extends [@Antjd2 Lemma 3.7] to the case of non-autonomous problems.
\[ancien\] Let [[Assumption \[assumption1\]]{}]{} (with $0<\beta<2$), [[Assumption \[assumption2\]]{}]{}, [[Assumption \[assumption3\]]{}]{} and [[Assumption \[assumption4\]]{}]{} be fulfilled.
- The following estimate holds $$\begin{aligned}
\Vert D_tu^h(t)\Vert\leq Ct^{-1+\beta/2},\quad t\in[0,T].\end{aligned}$$
- For any $\alpha\in(0, \beta)$, the following estimate holds $$\begin{aligned}
\left\Vert (-A_h(0))^{\alpha/2}D_tu^h(t)\right\Vert\leq Ct^{-1-\alpha/2+\beta/2},\quad t\in(0,T].\end{aligned}$$
- The following holds $$\begin{aligned}
\Vert D^2_tu^h(t)\Vert\leq Ct^{-2+\beta/2},\quad t\in(0,T].\end{aligned}$$
As in the proof of [@Stig2 Theorem 5.2] or [@Antjd2 Lemma 3.7], we set $v^h(t)=tD_tu^h(t)$, it follows that $D_tv^h(t)$ satisfies the following equation $$\begin{aligned}
D_tv^h(t)&=&A_h(t)v^h(t)+D_tu^h(t)+tA_h'(t)u^h(t)+tP_h\frac{\partial F}{\partial t}\left(t, u^h(t)\right)\nonumber\\
&+&tP_h\frac{\partial F}{\partial u}\left(t,u^h(t)\right)v^h(t).\end{aligned}$$ Therefore by the Duhammel’s principle, it holds that $$\begin{aligned}
\label{deri1}
v^h(t)&=&\int_0^tU_h(t,s)\left[D_tu^h(t)+sA_h'(s)u^h(s)+sP_h\frac{\partial F}{\partial s}\left(s, u^h(s)\right)\right]ds\nonumber\\
&+&\int_0^tsU_h(t,s)P_h\frac{\partial F}{\partial u}\left(s,u^h(s)\right)v^h(s)ds.\end{aligned}$$ Taking the norm in both sides of [[(\[deri1\])]{}]{}, using [[Assumption \[assumption3\]]{}]{} and [[Lemma \[pazylemma\]]{}]{} yields [$$\begin{aligned}
\label{inter1}
\Vert v^h(t)\Vert&\leq& \int_0^t\left\Vert U_h(t,s)D_su^h(s)\right\Vert ds+\int_0^ts\left\Vert U_h(t,s)A_h'(s)u^h(s)\right\Vert ds\nonumber\\
&+&\int_0^ts\left\Vert U_h(t,s)P_h\frac{\partial F}{\partial s}\left(s, u^h(s)\right)\right\Vert ds+\int_0^ts\left\Vert U_h(t,s) P_h\frac{\partial F}{\partial u}\left(s,u^h(s)\right)\right\Vert_{L(H)}\Vert v^h(s)\Vert ds\nonumber\\
&\leq& \int_0^t\left\Vert U_h(t,s)D_su^h(s)\right\Vert ds+\int_0^ts\left\Vert U_h(t,s)A_h'(s)u^h(s)\right\Vert ds+Ct^2\nonumber\\
&+&C\int_0^t\Vert v^h(s)\Vert ds.\end{aligned}$$ ]{} Using [[Lemma \[lemderiv\]]{}]{} and [[Lemma \[regularitylemma\]]{}]{} yields $$\begin{aligned}
\label{inter2}
&&\int_0^ts\left\Vert U_h(t,s)A_h'(s)u^h(s)\right\Vert ds\nonumber\\
&\leq&\int_0^ts\left\Vert U_h(t,s)\left(-A_h(0)\right)^{1-\beta/2}\right\Vert_{L(H)}\left\Vert\left(-A_h(0)\right)^{-1+\beta/2} A_h'(s)(-A_h(0))^{-\beta/2}\right\Vert_{L(H)}\nonumber\\
&&\left\Vert(-A_h(0))^{\beta/2}u^h(s)\right\Vert ds\nonumber\\
&\leq& Ct\int_0^t(t-s)^{-1+\beta/2}\left\Vert (-A_h(0))^{\beta/2}u^h(s)\right\Vert ds\nonumber\\
&\leq& Ct\int_0^t(t-s)^{-1+\beta/2}ds\leq Ct^{1+\beta/2}.\end{aligned}$$ Using [[Lemma \[pazylemma\]]{}]{} and [[Lemma \[regularitylemma\]]{}]{}, we obtain [$$\begin{aligned}
\label{inter3}
&&\left\Vert U_h(t,s)D_su^h(s)\right\Vert\nonumber\\
&\leq& \left\Vert U_h(t,s)A_h(s)u^h(s)\right\Vert+\Vert U_h(t,s)P_hF(u^h(s))\Vert\nonumber\\
&\leq& \left\Vert U_h(t,s)(-A_h(0))^{1-\beta/2}\right\Vert_{L(H)}\left\Vert (-A_h(0))^{\beta/2}u^h(s)\right\Vert+\Vert U_h(t,s)\Vert_{L(H)}\Vert P_hF(u^h(s))\Vert\nonumber\\
&\leq&C(t-s)^{-1+\beta/2}\Vert u_0\Vert_{\beta}+C\Vert u_0\Vert\nonumber\\
&\leq& C(t-s)^{-1+\beta/2}.\end{aligned}$$ ]{} Substituting [[(\[inter3\])]{}]{} and [[(\[inter2\])]{}]{} in [[(\[inter1\])]{}]{} yields $$\begin{aligned}
\label{inter4}
\Vert v^h(t)\Vert&\leq& C\int_0^t(t-s)^{-1+\beta/2}ds+Ct^2+C\int_0^t\Vert v^h(s)\Vert ds\nonumber\\
&\leq& Ct^{\beta/2}+C\int_0^t\Vert v^h(s)\Vert ds.\end{aligned}$$ Applying the continuous Gronwall’s lemma to [[(\[inter4\])]{}]{} yields $$\begin{aligned}
\Vert v^h(t)\Vert \leq t^{\beta/2}.\end{aligned}$$ Therefore we have $$\begin{aligned}
\Vert D_tu^h(t)\Vert\leq Ct^{-1+\beta/2}.\end{aligned}$$ Let us now prove (ii). It follows from [[(\[deri1\])]{}]{} that $$\begin{aligned}
\label{deri2}
D_tu^h(t)&=&t^{-1}\int_0^tU_h(t,s)\left[D_su^h(s)+sA_h'(s)u^h(s)+sP_h\frac{\partial F}{\partial s}\left(s, u^h(s)\right)\right]ds\nonumber\\
&+&t^{-1}\int_0^tU_h(t,s)sP_h\frac{\partial F}{\partial u}\left(s,u^h(s)\right)D_su^h(s)ds.\end{aligned}$$ Pre-multiplying both sides of [[(\[deri2\])]{}]{} by $(-A_h(0))^{\alpha/2}$ yields $$\begin{aligned}
\label{deri3}
&&(-A_h(0))^{\alpha/2}D_tu^h(t)\nonumber\\
&=&t^{-1}\int_0^t(-A_h(0))^{\alpha/2}U_h(t,s)\left[D_su^h(s)+sA_h'(s)u^h(s)P_h\frac{\partial F}{\partial s}\left(s,u^h(s)\right)\right]ds\nonumber\\
&+&t^{-1}\int_0^ts\left(-A_h(0)\right)^{\alpha/2}U_h(t,s)P_h\frac{\partial F}{\partial u}\left(s,u^h(s)\right)D_su^h(s)ds.\end{aligned}$$ Taking the norm in both sides of [[(\[deri3\])]{}]{} yields [$$\begin{aligned}
\label{inter5}
&&\left\Vert (-A_h(0))^{\alpha/2}D_tu^h(t)\right\Vert\nonumber\\
&\leq&t^{-1}\int_0^t(t-s)^{-\alpha/2}\left[\Vert D_su^h(s)\Vert+s\left\Vert \frac{\partial F}{\partial s}\left(s,u^h(s)\right)\right\Vert\right]ds\nonumber\\
&+&t^{-1}\int_0^t(t-s)^{-\alpha/2}s\left\Vert P_h\frac{\partial F}{\partial u}\left(s,u^h(s)\right)\right\Vert_{L(H)}\Vert D_su^h(s)\Vert ds\nonumber\\
&+&t^{-1}\int_0^ts\left\Vert\left(-A_h(0)\right)^{\alpha/2}U_h(t,s)A_h'(s)u^h(s)\right\Vert ds\nonumber\\
&\leq& Ct^{-1}\int_0^t(t-s)^{-\alpha/2}\left[s^{-1+\beta/2}+s\right]ds+t^{-1}\int_0^ts\left\Vert(-A_h(0)^{\alpha/2}U_h(t,s)A_h'(s)u^h(s)\right\Vert ds\nonumber\\
&\leq &Ct^{-1}\int_0^t(t-s)^{-\alpha/2}s^{-1+\beta/2}ds+\int_0^t\Vert(-A_h(0)^{\alpha/2}U_h(t,s)A_h'(s)u^h(s)\Vert ds.\end{aligned}$$ ]{} Using [[Lemma \[lemderiv\]]{}]{} and [[Lemma \[regularitylemma\]]{}]{}, it holds that $$\begin{aligned}
\label{inter6}
&&\int_0^t\left\Vert (-A_h(0))^{\alpha/2}U_h(t,s)A_h'(s)u^h(s)\right\Vert ds\nonumber\\
&\leq& \int_0^t\left\Vert (-A_h(0))^{\alpha/2}U_h(t,s)(-A_h(0))^{\epsilon}\right\Vert_{L\left(\mathcal{D}\left(-A(0)\right)^{\epsilon}, H\right)}\nonumber\\
&&\times\left\Vert (-A_h(0))^{-\epsilon}A_h'(s)(-A_h(0))^{-1+\epsilon}(-A_h(0))^{1-\epsilon}u^h(s)\right\Vert ds\nonumber\\
&\leq& C\int_0^t(t-s)^{-\alpha/2}\left\Vert (-A_h(0))^{1-\epsilon}u^h(s)\right\Vert ds\nonumber\\
&\leq& C\int_0^t(t-s)^{-\alpha/2-\epsilon}s^{\beta/2-1+\epsilon}ds\nonumber\\
&\leq & Ct^{\beta/2-\alpha/2}.\end{aligned}$$ Substituting [[(\[inter6\])]{}]{} dans [[(\[inter5\])]{}]{} yields $$\begin{aligned}
\label{inter7}
\left\Vert (-A_h(0))^{\alpha/2}D_tu^h(t)\right\Vert&\leq& Ct^{-1}\int_0^t(t-s)^{-\alpha/2}s^{-1+\beta/2}ds+Ct^{-\alpha/2-\epsilon}\nonumber\\
&\leq& Ct^{-1-\alpha/2+\beta/2}.\end{aligned}$$ This completes the proof of (ii). To prove (iii), as in [@Antjd2 Lemma 3.7] we set $w^h(t)=tD^2_tu^h(t)$. Taking the derivative with respect to $t$ in both sides of [[(\[semi1\])]{}]{} yields $$\begin{aligned}
\label{deri4}
D^2_tu^h(t)&=&A_h'(t)u^h(t)+A_h(t)D_tu^h(t)+P_h\frac{\partial F}{\partial t}\left(t,u^h(t)\right)\nonumber\\
&+&P_h\frac{\partial F}{\partial u}\left(t,u^h(t)\right)D_tu^h(t).\end{aligned}$$ Taking the derivative with respect to $t$ in both side of [[(\[deri4\])]{}]{} yields [$$\begin{aligned}
\label{deri5}
D^3_tu^h(t)&=&A_h''(t)u^h(t)+2A_h'(t)D_tu^h(t)+A_h(t)D^2_tu^h(t)\nonumber\\
&+&P_h\frac{\partial^2F}{\partial t^2}\left(t,u^h(t)\right)D_tu^h(t)+2P_h\frac{\partial^2F}{\partial t\partial u}\left(t,u^h(t)\right)D_tu^h(t)\nonumber\\
&+&P_h\frac{\partial^2F}{\partial t\partial u}\left(t,u^h(t)\right)D_tu^h(t)+P_h\frac{\partial^2F}{\partial u^2}\left(t,u^h(t)\right)\left(D_tu^h(t), D_tu^h(t)\right).\end{aligned}$$ ]{} Using [[(\[deri5\])]{}]{} and [[(\[deri4\])]{}]{} and rearranging yields [$$\begin{aligned}
\label{deri6}
D_tw^h(t)&=&D_t^2u^h(t)+tD_t^3u^h(t)\nonumber\\
&=&A_h(t)w^h(t)+A_h'(t)u^h(t)+A_h(t)D_tu^h(t)+P_h\frac{\partial F}{\partial t}\left(t,u^h(t)\right)\nonumber\\
&+&P_h\frac{\partial F}{\partial u}\left(t,u^h(t)\right)D_tu^h(t)+tA_h''(t)u^h(t)+2tA_h'(t)D_tu^h(t)\nonumber\\
&+&tP_h\frac{\partial^2F}{\partial t^2}\left(t,u^h(t)\right)D_tu^h(t)+2tP_h\frac{\partial^2F}{\partial t\partial u}\left(t,u^h(t)\right)D_tu^h(t)\nonumber\\
&+&tP_h\frac{\partial^2F}{\partial t\partial u}\left(t,u^h(t)\right)D_tu^h(t)+tP_h\frac{\partial^2F}{\partial u^2}\left(t,u^h(t)\right)\left(D_tu^h(t),D_tu^h(t)\right).\end{aligned}$$ ]{} By the Duhammel’s principle, it follows from [[(\[deri6\])]{}]{} that [$$\begin{aligned}
\label{deri7}
w^h(t)&=&\int_0^tU_h(t,s)\left[A_h'(s)u^h(s)+A_h(s)D_su^h(s)+sA_h''(s)u^h(s)
+2sA_h'(s)D_su^h(s)\right]ds\nonumber\\
&+&\int_0^tU_h(t,s)\left[P_h\frac{\partial F}{\partial s}\left(s,u^h(s)\right)+P_h\frac{\partial F}{\partial u}\left(s,u^h(s)\right)D_su^h(s)\right]ds\nonumber\\
&+&\int_0^tU_h(t,s)sP_h\left[sP_h\frac{\partial^2F}{\partial s^2}\left(s,u^h(s)\right)D_su^h(s)+3\frac{\partial^2F}{\partial s\partial u}\left(s,u^h(s)\right)D_su^h(s)\right]ds\nonumber\\
&+&\int_0^tsU_h(t,s)\frac{\partial^2F}{\partial u^2}\left(s,u^h(s)\right)\left(D_su^h(s), D_su^h(s)\right)ds.\end{aligned}$$ ]{} Taking the norm in both sides of [[(\[deri7\])]{}]{} yields [$$\begin{aligned}
\label{deri8}
\Vert w^h(t)\Vert&\leq& \int_0^t\left\Vert U_h(t,s)A_h'(s)u^h(s)\right\Vert ds+\int_0^t\left\Vert U_h(t,s)A_h(s)D_su^h(s)\right\Vert ds\nonumber\\
&+&t\int_0^t\left\Vert U_h(t,s)A_h''(s)u^h(s)\right\Vert ds+2t\int_0^t\left\Vert U_h(t,s)D_su^h(s)\right\Vert ds\nonumber\\
&+&C\int_0^t\Vert D_su^h(s)\Vert ds+C\int_0^ts\Vert D_su^h(s)\Vert ds+C\int_0^ts\Vert D_su^h(s)\Vert^2ds.\end{aligned}$$ ]{} Using [[Lemma \[lemderiv\]]{}]{}, [[Lemma \[pazylemma\]]{}]{} and [[Lemma \[regularitylemma\]]{}]{} yields $$\begin{aligned}
\label{inter8}
\int_0^t\left\Vert U_h(t,s)A_h'(s)u^h(s)\right\Vert ds\leq Ct^{\beta/2}.\end{aligned}$$ Using (ii) and [[Lemma \[regularitylemma\]]{}]{} yields $$\begin{aligned}
\label{inter9}
&&\int_0^t\left\Vert U_h(t,s)A_h(s)D_su^h(s)\right\Vert ds\nonumber\\
&\leq& \int_0^t\left\Vert U_h(t,s)(-A_h(s))^{1-\beta/2-\epsilon}\right\Vert_{L(H)}\left\Vert (-A_h(s))^{\beta/2-\epsilon}D_su^h(s)\right\Vert ds\nonumber\\
&\leq& C\int_0^t(t-s)^{-1+\beta/2-\epsilon}s^{\beta/2-\epsilon}ds\leq Ct^{-1+\beta/2-\epsilon}.\end{aligned}$$ Using [[Lemma \[lemderiv\]]{}]{} and [[Lemma \[regularitylemma\]]{}]{} yields $$\begin{aligned}
\label{inter10}
&&\int_0^t\left\Vert U_h(t,s)A_h''(s)u^h(s)\right\Vert ds\nonumber\\
&\leq&\int_0^t\left\Vert U_h(t,s)(-A_h(0))^{1-\beta/2+\epsilon}\right\Vert_{L(H)}\nonumber\\
&&\times\left\Vert (-A_h(0))^{-1+\beta-\epsilon}A_h''(s)(-A_h(0))^{-\beta/2+\epsilon}(-A_h(0))^{\beta/2-\epsilon}u^h(s)\right\Vert ds\nonumber\\
&\leq&C\int_0^t(t-s)^{-1+\beta/2-\epsilon}\left\Vert (-A_h(0))^{\beta/2-\epsilon}u^h(s)\right\Vert ds\nonumber\\
&\leq& C\int_0^t(t-s)^{-1+\beta/2-\epsilon}ds\nonumber\\
&\leq& Ct^{\beta/2-\epsilon}.\end{aligned}$$ Using (i) yields $$\begin{aligned}
\label{inter11}
\int_0^ts\Vert D_su^h(s)\Vert^2ds\leq C\int_0^ts^{-1+\beta}ds\leq Ct^{\beta}.\end{aligned}$$ Substituting [[(\[inter11\])]{}]{}, [[(\[inter10\])]{}]{}, [[(\[inter9\])]{}]{} and [[(\[inter8\])]{}]{} in [[(\[deri8\])]{}]{} yields $$\begin{aligned}
\Vert w^h(t)\Vert\leq Ct^{-1+\beta/2}.\end{aligned}$$ This completes the proof of the lemma.
For non commutative operator $H_j$ on Banach space, we define the following product $$\begin{aligned}
\prod_{j=k}^mH_j=\left\{\begin{array}{ll}
H_mH_{m-1}\cdots H_k\quad\quad \text{if}\quad m\geq k,\\
\mathbf{I}\quad \hspace{3cm}\text{if}\quad m<k.
\end{array}
\right.\end{aligned}$$ The following stability result is fundamental in our convergence analysis.
\[stabilite\] Let [[Assumption \[assumption1\]]{}]{}, [[Assumption \[assumption2\]]{}]{}, [[Assumption \[assumption3\]]{}]{} and [[Assumption \[assumption4\]]{}]{} be fulfilled. Then the following stability estimate holds [$$\begin{aligned}
\left\Vert \left(\prod_{j=k}^m e^{\left(A_{h,j}+J^h_j\right)\Delta t}\right)(-A_{h,k})^{\gamma}\right\Vert_{L(H)}&\leq& Ct_{m-k+1}^{-\gamma},\quad 0\leq k\leq m\leq M,\end{aligned}$$ ]{} for any $\gamma\in[0,1)$.
As in [@Ostermann1 Theorem 1], the main idea is to compare the composition of the perturbed operator with the frozen operator $$\begin{aligned}
\label{frozen1}
\prod_{j=k}^me^{\left(A_{h,k}+J^h_k\right)\Delta t}=e^{(t_{m+1}-t_k)\left(A_{h,k}+J^h_k\right)}.\end{aligned}$$ Using [@Antjd1 Lemma 9] yields the following estimate $$\begin{aligned}
\label{frozen2}
\left\Vert \prod_{j=k}^me^{\left(A_{h,k}+J^h_k\right)\Delta t}(-A_{h,k})^{\gamma}\right\Vert_{L(H)}&=&\left\Vert e^{\left(A_{h,k}+J^h_k\right)t_{m-k+1}}(-A_{h,k})^{\gamma}\right\Vert_{L(H)}\nonumber\\
&\leq& Ct_{m-k+1}^{-\gamma}.\end{aligned}$$ It remains to bound $\Delta^m_k(-A_{h,k})^{\gamma}$, where $\Delta^m_k$ is defined as follows $$\begin{aligned}
\label{frozen3}
\Delta^m_k:=\prod_{j=k}^me^{\left(A_{h,j}+J^h_j\right)\Delta t}-\prod_{j=k}^me^{\left(A_{h,k}+J^h_k\right)\Delta t}.\end{aligned}$$ Using the telescopic identity we obtain [$$\begin{aligned}
\label{frozen4}
&&\Delta^m_k\nonumber\\
&=&\sum_{j=k+1}^{m-1}\Delta ^m_{j+1}\left(e^{\left(A_{h,j}+J^h_j\right)\Delta t}-e^{\left(A_{h,k}+J^h_k\right)\Delta t}\right)e^{(t_j-t_k)\left(A_{h,k}+J^h_k\right)}\nonumber\\
&+&\sum_{j=k+1}^me^{(t_{m+1}-t_{j+1})\left(A_{h,k}+J^h_k\right)}\left(e^{\left(A_{h,j}+J^h_j\right)\Delta t}-e^{\left(A_{h,k}+J^h_k\right)\Delta t}\right)e^{(t_j-t_k)\left(A_{h,k}+J^h_k\right)}.\nonumber\\\end{aligned}$$ ]{} Using the variation of parameter formula [@Engel Chapter III, Corollary 1.7] yields $$\begin{aligned}
\label{frozen5}
e^{\left(A_{h,l}+J^h_l\right)\Delta t}=e^{A_{h,l}\Delta t}+\int_0^{\Delta t}e^{A_{h,l}(\Delta t-s)}J^h_le^{\left(A_{h,l}+J^h_l\right)s}ds.\end{aligned}$$ It follows therefore from [[(\[frozen5\])]{}]{} that $$\begin{aligned}
\label{frozen6}
\left(e^{\left(A_{h,j}+J^h_j\right)\Delta t}-e^{\left(A_{h,k}+J^h_k\right)\Delta t}\right)
&=&\left(e^{A_{h,j}\Delta t}-e^{A_{h,k}\Delta t}\right)\nonumber\\
&+&\int_0^{\Delta t}e^{A_{h,j}(\Delta t-s)}J^h_je^{\left(A_{h,j}+J^h_j\right)s}ds\nonumber\\
&-&\int_0^{\Delta t}e^{A_{h,k}(\Delta t-s)}J^h_ke^{\left(A_{h,k}+J^h_k\right)s}ds\nonumber\\
&=:&IV_1+IV_2+IV_3.\end{aligned}$$ Using the integral formula of Cauchy exactly as in [@Ostermann1 Lemma 1] yields $$\begin{aligned}
\label{frozen7}
\Vert IV_1\Vert_{L(H)}=\left\Vert \left(e^{A_{h,j}\Delta t}-e^{A_{h,k}\Delta t}\right)\right\Vert_{L(H)}\leq C\Delta t.\end{aligned}$$ Using [@Antjd1 Lemma 9], [[Assumption \[assumption2\]]{}]{} and [[Assumption \[assumption3\]]{}]{} yields [$$\begin{aligned}
\label{frozen8}
\Vert IV_2\Vert_{L(H)}+\Vert IV_3\Vert_{L(H)} &\leq&2\int_0^{\Delta t}\left\Vert e^{A_{h,k}(\Delta t-s)}\right\Vert_{L(H)}\Vert J^h_k\Vert_{L(H)}\left\Vert e^{\left(A_{h,k}+J^h_k\right)s}\right\Vert_{L(H)}ds\nonumber\\
&\leq& C\int_0^{\Delta t}ds\leq C\Delta t.\end{aligned}$$ ]{} Substituting [[(\[frozen8\])]{}]{} and [[(\[frozen7\])]{}]{} in [[(\[frozen6\])]{}]{} yields $$\begin{aligned}
\label{frozen9}
\left\Vert \left(e^{\left(A_{h,j}+J^h_j\right)\Delta t}-e^{\left(A_{h,k}+J^h_k\right)\Delta t}\right)\right\Vert_{L(H)}\leq C\Delta t.\end{aligned}$$ Inserting an appropriate power of $(-A_{h,k})^{\gamma}$ in [[(\[frozen4\])]{}]{}, using triangle inequality and [[(\[frozen9\])]{}]{} yields $$\begin{aligned}
\label{frozen9}
&&\left\Vert \Delta ^m_k(-A_{h,k})^{\gamma}\right\Vert_{L(H)}\nonumber\\
&\leq&\sum_{j=k+1}^{m-1}\Vert \Delta ^m_{j+1}(-A_{h,k})^{\gamma}\Vert_{L(H)}\Vert (-A_{h,k})^{-\gamma}\Vert_{L(H)}\nonumber\\
&&\times\left\Vert \left(e^{\left(A_{h,j}+J^h_j\right)\Delta t}-e^{\left(A_{h,k}+J^h_k\right)\Delta t}\right)\right\Vert_{L(H)}\left\Vert e^{(t_j-t_k)\left(A_{h,k}+J^h_k\right)}(-A_{h,k})^{\gamma}\right\Vert_{L(H)}\nonumber\\
&+&\sum_{j=k+1}^m\left\Vert e^{(t_{m+1}-t_{j+1})\left(A_{h,k}+J^h_k\right)}\right\Vert_{L(H)}\left\Vert \left(e^{\left(A_{h,j}+J^h_j\right)\Delta t}-e^{\left(A_{h,k}+J^h_k\right)\Delta t}\right)\right\Vert_{L(H)}\nonumber\\
&&\times \left\Vert e^{(t_j-t_k)\left(A_{h,k}+J^h_k\right)}(-A_{h,k})^{\gamma}\right\Vert_{L(H)}\nonumber\\
&\leq& C\Delta t\sum_{j=k+1}^{m-1}\Vert \Delta^m_{j+1}(-A_{h,k})^{\gamma}\Vert_{L(H)} t_{j-k}^{-\gamma}+C\Delta t\sum_{j=k+1}^mt_{j-k}^{-\gamma}\nonumber\\
&\leq& C+C\Delta t\sum_{j=k+1}^{m-1}t_{j-k}^{-\gamma}\Vert \Delta^m_{j+1}(-A_{h,k})^{\gamma}\Vert_{L(H)}.\end{aligned}$$ Applying the discrete Gronwall’s lemma to [[(\[frozen9\])]{}]{} yields $$\begin{aligned}
\label{frozen10}
\left\Vert \Delta ^m_k(-A_{h,k})^{\gamma}\right\Vert_{L(H)}\leq C.\end{aligned}$$ Using [[(\[frozen10\])]{}]{} and [[(\[frozen2\])]{}]{} completes the proof of [[Lemma \[stabilite\]]{}]{}.
\[stabilitysolution\] Let Assumptions \[assumption2\], \[assumption1\] and \[assumption3\] be fulfilled. Then the numerical scheme [[(\[semi2\])]{}]{} satisfies the following estimate $$\begin{aligned}
\Vert u^h_m\Vert\leq R,\quad m\in\{0, 1, \cdots, M\},
\end{aligned}$$ where $R>0$ is independent of $h$, $m$, $M$ and $\Delta t$.
Iterating the numerical solution [[(\[semi2\])]{}]{} by substituting $u^h_j$, $j=m-1, \cdots, 1$ only in the first term of [[(\[semi2\])]{}]{} by their expressions yields [$$\begin{aligned}
\label{dim1}
&&u^h_m=\left(\prod_{j=0}^{m-1}e^{\Delta t(A_{h,j}+J^h_j)}\right)u^h_0\\
&+&\sum_{k=0}^{m-1}\int_0^{\Delta t}\left(\prod_{j=m-k}^{m-1}e^{\Delta t(A_{h,j}+J^h_j)}\right)e^{(A_{h, m-k-1}+J^h_{m-k-1})(\Delta t-s)}a^h_{m-k-1}\left(t_{m-k-1}+\frac{\Delta t}{2}\right)ds\nonumber\\
&+&\sum_{k=0}^{m-1}\int_0^{\Delta t}\left(\prod_{j=m-k}^{m-1}e^{\Delta t(A_{h,j}+J^h_j)}\right)e^{(A_{h,m-k-1}+J^h_{m-k-1})(\Delta t-s)}\nonumber\\
&&G^h_{m-k-1}\left(t_{m-k-1}+\frac{\Delta t}{2}, u^h_{m-k-1}\right)ds.\nonumber\end{aligned}$$ ]{} Taking the norm in both sides of [[(\[dim1\])]{}]{}, using triangle inequality, [[Lemma \[stabilite\]]{}]{} and [[Assumption \[assumption3\]]{}]{} yields [$$\begin{aligned}
\label{dim2}
\Vert u^h_m\Vert&\leq&\left\Vert\left(\prod_{j=0}^{m-1}e^{\Delta t(A_{h,j}+J^h_j)}\right)\right\Vert_{L(H)}\Vert u^h_0\Vert\\
&+&\sum_{k=0}^{m-1}\int_0^{\Delta t}\left\Vert\left(\prod_{j=m-k}^{m-1}e^{\Delta t(A_{h,j}+J^h_j)}\right)\right\Vert_{L(H)}\left\Vert e^{(A_{h, m-k-1}+J^h_{m-k-1})(\Delta t-s)}\right\Vert_{L(H)}\nonumber\\
&\times&\left\Vert a^h_{m-k-1}\right\Vert\left(t_{m-k-1}+\frac{\Delta t}{2}\right)ds\nonumber\\
&+&\sum_{k=0}^{m-1}\int_0^{\Delta t}\left\Vert\left(\prod_{j=m-k}^{m-1}e^{\Delta t(A_{h,j}+J^h_j)}\right)\right\Vert_{L(H)}\left\Vert e^{\left(A_{h,m-k-1}+J^h_{m-k-1}\right)(\Delta t-s)}\right\Vert_{L(H)}\nonumber\\
&\times&\left\Vert G^h_{m-k-1}\right\Vert_{L(H)}\left\Vert\left(t_{m-k-1}+\frac{\Delta t}{2}, u^h_{m-k-1}\right)\right\Vert ds\nonumber\\
&\leq& C\Vert u^h_0\Vert+C\sum_{k=0}^{m-1}\int_0^{\Delta t}\left(t_{m-k-1}+\frac{\Delta t}{2}\right)ds\nonumber\\
&+&C\sum_{k=0}^{m-1}\int_0^{\Delta t}\left[\left(t_{m-k-1}+\frac{\Delta t}{2}\right)+u^h_{m-k-1}\right]ds.\end{aligned}$$ ]{} Using the fact that $t_{m-k-1}+\frac{\Delta t}{2}\leq T$ and $\Vert u^h_0\Vert\leq \Vert u_0\Vert$, it holds from [[(\[dim2\])]{}]{} that $$\begin{aligned}
\label{dim3}
\Vert u^h_m\Vert\leq C\Vert u_0\Vert +C+C\Delta t\sum_{k=0}^{m-1}\Vert u^h_k\Vert.\end{aligned}$$ Applying the discrete Gronwall"s lemma to [[(\[dim3\])]{}]{} yields $$\begin{aligned}
\label{dim4}
\Vert u^h_m\Vert\leq C(1+\Vert u_0\Vert)\leq R,\quad m\in\{0, \cdots, M\}.\end{aligned}$$ This completes the proof of [[Lemma \[stabilitysolution\]]{}]{}.
\[lemmestrategie1\] Let Assumptions \[assumption2\] and \[assumption3\] be fulfilled. Then the fractional powers of $-(A_{h,k}+J^h_k)$ exist and the following estimate holds $$\begin{aligned}
\Vert \left(-(A_{h, k}+J^h_k)\right)^{-\alpha}\Vert_{L(H)}\leq C,\quad \alpha>0,\end{aligned}$$ with $C$ independent of $h$ and $k$.
First of all we claim that $e^{(A_{h,k}+J^h_k)t}$ is uniformly exponentially stable. In fact, from the variation of parameters formula [@Engel Chapter 3, Corollary 1.7] or [@Pazy Page 77, Section 3.1] it holds that $$\begin{aligned}
\label{strategie2}
e^{(A_{h,k}+J^h_k)t}=e^{A_{h,k}t}+\int_0^te^{A_{h,k}(t-s)}J^h_ke^{(A_{h,k}+J^h_k)s}ds,\quad t\geq 0.
\end{aligned}$$ Taking the norm in both sides of [[(\[strategie2\])]{}]{}, inserting appropriately power of $(-A_{h,k})^{-\gamma}(-A_{h,k})^{\gamma}$ (with $\gamma\in(0, 1)$), using the uniformly boundedness of $(-A_{h,k})^{-\gamma}$, [[Assumption \[assumption3\]]{}]{} and [[(\[smooth2\])]{}]{} yields $$\begin{aligned}
\label{strategie3}
&&\Vert e^{(A_{h,k}+J^h_k)t}\Vert_{L(H)}\nonumber\\
&\leq& \Vert (-A_{h,k})^{-\gamma}\Vert_{L(H)}\Vert (-A_{h,k})^{\gamma}e^{A_{h,k}t}\Vert_{L(H)}\nonumber\\
&+&\int_0^t\Vert (-A_{h,k})^{-\gamma}\Vert_{L(H)}\Vert(-A_{h,k})^{\gamma} e^{A_{h,k}(t-s)}\Vert_{L(H)}\Vert J^h_k\Vert_{L(H)}\Vert e^{(A_{h,k}+J^h_k)s}\Vert_{L(H)}ds\nonumber\\
&\leq& Ct^{-\gamma}+C\int_0^t(t-s)^{-\gamma}\Vert e^{(A_{h,k}+J^h_k)s}\Vert_{L(H)}ds.
\end{aligned}$$ Applying the generalized Gronwall’s lemma [@Henry Lemma 3.5.2] to [[(\[strategie3\])]{}]{} yields $$\begin{aligned}
\label{strategie4}
\Vert e^{(A_{h,k}+J^h_k)t}\Vert_{L(H)}\leq Ct^{-\gamma}=\frac{C}{t^{\gamma}},\quad \gamma\in(0,1),\quad t\geq 0.\end{aligned}$$ Taking the limit as $t$ goes to $\infty$ in [[(\[strategie4\])]{}]{} yields $$\begin{aligned}
\lim_{t\longrightarrow\infty} \Vert e^{(A_{h,k}+J^h_k)t}\Vert_{L(H)}=0.
\end{aligned}$$ Employing [@Engel Proposition 1.7, Chapter V, Page 299], it follows that $e^{(A_{h, k}+J^h_k)t}$ is exponentially stable, i.e. there exists two positive constants $L_k$ and $\omega_k$ such that $$\begin{aligned}
\label{strategie5}
\Vert e^{(A_{h,k}+J^h_k)t}\Vert_{L(H)}\leq L_ke^{-\omega_kt},\quad t\geq 0.
\end{aligned}$$ Let $B[0, R]:=\{v\in H: \Vert v\Vert\leq R\}$, where $R$ is defined in [[Lemma \[stabilitysolution\]]{}]{}. More generally, for every $\tau\in[0, T]$ and $v\in B[0, R]$ there two positive constants $L_{\tau, v}$ and $\omega_{\tau,v}$ such that $$\begin{aligned}
\label{strategie7a}
\Vert e^{(A_{h}(\tau)+J^h_{\tau,v})t}\Vert_{L(H)}\leq L_{\tau,v}e^{-\omega_{\tau,v}t},\quad t\geq 0,
\end{aligned}$$ where $J^h_{\tau, v}:=P_h\frac{\partial F}{\partial v}(\tau, v)$. Note that the function $(\tau, v)\longmapsto \omega_{\tau, v}$ is continuous. This follows from the definition of the growth bound $\omega_{\tau, v}$ $$\begin{aligned}
\omega_{\tau, v}:=\inf_{t>0}\frac{1}{t}\log\left\Vert e^{(A_h(\tau)+J^h_{\tau, v})t}\right\Vert_{L(H)},\quad \tau\in[0, T],\quad v\in B[0, R].\end{aligned}$$ Due to [[(\[strategie7a\])]{}]{}, the following constant is well defined $$\begin{aligned}
\label{strategie7b}
L'_{\tau, v}:=\sup_{t\geq 0}\left\Vert e^{\left(A_h(\tau)+J^h_{\tau, v}\right)t}\right\Vert_{L(H)}e^{\omega_{\tau, v}t},\quad \tau\in[0, T],\quad v\in B[0, R].\end{aligned}$$ It follows from the above definition [[(\[strategie7b\])]{}]{} that the function $(\tau, v)\longmapsto L'_{\tau, v}$ is continuous. Therefore by Weierstrass’s theorem there exist two positive constants $L'$ and $\omega$ such that $$\begin{aligned}
L'=\sup_{\tau\in[0, T], v\in B(0, R)}L'_{\tau, v},\quad \omega=\inf_{\tau\in[0, T], v\in B(0, R)}\omega_{\tau, v}.
\end{aligned}$$ Consequently, we have $$\begin{aligned}
\label{strategie6}
\Vert e^{(A_{h,k}+J^h_k)t}\Vert_{L(H)}\leq L'e^{-\omega t},\quad t\geq 0,\quad k\in\{0,1,\cdots,M\}.
\end{aligned}$$ This proves the claim. Let us now finish the proof of [[Lemma \[lemmestrategie1\]]{}]{}. Assumptions \[assumption2\] and \[assumption3\] imply that $-(A_{h, k}+J^h_k)$ is a positive operator. Therefore its fractional powers are well defined and are given by $$\begin{aligned}
\label{strategie1}
\left(-(A_{h, k}+J^h_k)\right)^{-\alpha}=\frac{1}{\Gamma(\alpha)}\int_0^{\infty}t^{\alpha-1}e^{(A_{h,k}+J^h_k)t}dt,
\end{aligned}$$ where $\Gamma(\alpha)$ is a gamma function, see e.g. [@Henry; @Pazy; @Engel]. Taking the norm in both sides of [[(\[strategie1\])]{}]{} and using [[(\[strategie6\])]{}]{} yields $$\begin{aligned}
\left\Vert\left(-(A_{h, k}+J^h_k)\right)^{-\alpha}\right\Vert_{L(H)}&\leq&\frac{L'}{\Gamma(\alpha)}\int_0^{\infty}t^{\alpha-1}e^{-\omega t}dt=\frac{L'\omega^{2-\alpha}}{\Gamma(\alpha)}\int_0^{\infty}s^{\alpha-1}e^{-s}ds\nonumber\\
&=&L'\omega^{2-\alpha}<\infty.
\end{aligned}$$ This completes the proof of the lemma.
\[lemmestrategie2\] Let Assumptions \[assumption2\] and \[assumption3\] be fulfilled. Then the following estimate holds $$\begin{aligned}
\label{strategie8a}
\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-\alpha}(-A_{h,k})^{\alpha}\right\Vert_{L(H)}\leq C,\quad \alpha\in[0, 1]\\
\label{strategie8b}
\left\Vert (-A_{h,k})^{\alpha}\left(-(A_{h,k}+J^h_k)\right)^{-\alpha}\right\Vert_{L(H)}\leq C,\quad \alpha\in[0, 1].\end{aligned}$$
We only prove [[(\[strategie8a\])]{}]{} since the proof of [[(\[strategie8b\])]{}]{} is similar. For $\alpha=1$, using triangle inequality, [[Assumption \[assumption3\]]{}]{} and [[Lemma \[lemmestrategie1\]]{}]{} it holds that [$$\begin{aligned}
\label{strategie8}
&&\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1}(-A_{h,k})\right\Vert_{L(H)}\nonumber\\
&\leq& \left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1}\left(-(A_{h,k}+J^h_k)\right)\right\Vert_{L(H)}+\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1}\right\Vert_{L(H)}\Vert J^h_k\Vert_{L(H)}\nonumber\\
&\leq& C.\end{aligned}$$ ]{} Note that [[(\[strategie8a\])]{}]{} obviously holds for $\alpha=0$. As in [@Antjd1; @Antjd2; @Thomee2] the intermediates cases follow by interpolation technique.
\[utilemma\] For $k=0,\cdots, M-1$ and $t_k\leq t\leq t_{k+1}$, let us set $$\begin{aligned}
\label{reste1}
L^h_k(t)&:=&\left(A_h(t)-A_{h,k}\right)u^h(t)-a^h_k(t_k+t)\nonumber\\
&+&G^h_k\left(t,u^h(t)\right)-G^h_k\left(t_k+\frac{\Delta t}{2},u^h(t_k)\right).\end{aligned}$$ Under [[Assumption \[assumption1\]]{}]{}, [[Assumption \[assumption2\]]{}]{}, [[Assumption \[assumption3\]]{}]{} and [[Assumption \[assumption4\]]{}]{}, provided that $L^h_k$ is twice differentiable on $(t_k, t_{k+1})$, the following estimates hold $$\begin{aligned}
\label{interessant1}
\left\Vert (-A_h(0))^{-\epsilon}\left(L^h_k\right)'\left(t_k+\frac{\Delta t}{2}\right)\right\Vert&\leq& Ct_k^{-1+\beta/2},\quad k\geq 1,\\
\label{interessant2}
\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-\epsilon}\left(L^h_k\right)'\left(t_k+\frac{\Delta t}{2}\right)\right\Vert&\leq& Ct_k^{-1+\beta/2},\quad k\geq 1,\\
\label{interessant3}
\left\Vert \left(-A_h(0)\right)^{-1}\left(L^h_k\right)''\left(t\right)\right\Vert&\leq& Ct^{-2+\beta/2},\quad t>0,\\
\label{interessant4}
\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-\epsilon}\left(L^h_k\right)''\left(t\right)\right\Vert&\leq& Ct^{-2+\beta/2},\quad t>0,\end{aligned}$$ where $\epsilon>0$ is a positive number, small enough.
Let us start with the estimate of [[(\[interessant1\])]{}]{}. Taking the derivative in both sides of [[(\[reste1\])]{}]{}, using [[(\[remainder2\])]{}]{} and [[(\[remainder1\])]{}]{} yields $$\begin{aligned}
\label{vend1}
(L^h_k)'(t)&=& A_h'(t)u^h(t)+\left(A_h(t)-A_{h,k}\right)D_tu^h(t)+P_h\frac{\partial F}{\partial t}\left(t,u^h(t)\right)\nonumber\\
&+&P_h\frac{\partial F}{\partial u}\left(t,u^h(t)\right)D_tu^h(t)-P_h\frac{\partial F}{\partial u}\left(t_k+\frac{\Delta t}{2},u^h_k\right)D_tu^h(t)\nonumber\\
&-&P_h\frac{\partial F}{\partial t}\left(t_k+\frac{\Delta t}{2}, u^h_k\right)-a^h_k.\end{aligned}$$ Taking the norm in both sides of [[(\[vend1\])]{}]{}, using [[Lemma \[utilemma\]]{}]{}, [[Assumption \[assumption3\]]{}]{}, [[Lemma \[regularitylemma\]]{}]{}, [[Lemma \[lemderiv\]]{}]{}, [[Lemma \[ancien\]]{}]{} and the fact that $(-A_h(0))^{-\epsilon}$ is bounded yields [$$\begin{aligned}
\label{vend2}
&&\left\Vert (-A_h(0))^{-\epsilon}\left(L^h_k\right)'\left(t_k+\frac{\Delta t}{2}\right)\right\Vert\nonumber\\
&\leq& \left\Vert (-A_h(0))^{-\epsilon} A_h'\left(t_k+\frac{\Delta t}{2}\right)u^h\left(t_k+\frac{\Delta t}{2}\right)\right\Vert+C\left\Vert P_h\frac{\partial F}{\partial t}\left(t_k+\frac{\Delta t}{2}, u^h\left(t_k+\frac{\Delta t}{2}\right)\right)\right\Vert\nonumber\\
&+&C\left\Vert P_h\frac{\partial F}{\partial u}\left(t_k+\frac{\Delta t}{2}, u^h\left(t_k+\frac{\Delta t}{2}\right)\right)\right\Vert_{L(H)}\left\Vert D_tu^h\left(t_k+\frac{\Delta t}{2}\right)\right\Vert+C\left\Vert P_h\frac{\partial F}{\partial t}\left(t_k+\frac{\Delta t}{2}, u^h_k\right)\right\Vert\nonumber\\
&+&C\left\Vert P_h\frac{\partial F}{\partial u}\left(t_k, u^h_k\right)\right\Vert_{L(H)}\left\Vert D_tu^h\left(t_k+\frac{\Delta t}{2}\right)\right\Vert+C\left\Vert \frac{\partial F}{\partial t}\left(t_k+\frac{\Delta t}{2}, u^h_k\right)\right\Vert\nonumber\\
&\leq& \left\Vert (-A_h(0))^{-\epsilon}A_h'\left(t_k+\frac{\Delta t}{2}\right)(-A_h(0))^{-1+\epsilon}\right\Vert_{L(H)}\left\Vert (-A_h(0))^{1-\epsilon}u^h\left(t_k+\frac{\Delta t}{2}\right)\right\Vert\nonumber\\
&+&C+C\left\Vert u^h\left(t_k+\frac{\Delta t}{2}\right)\right\Vert+C\left\Vert D_tu^h\left(t_k+\frac{\Delta t}{2}\right)\right\Vert+C\nonumber\\
&\leq&C\left(t_k+\frac{\Delta t}{2}\right)^{-1+\epsilon+\beta/2}+C\left(t_k+\frac{\Delta t}{2}\right)^{-1+\beta/2}\leq Ct_k^{-1+\beta/2}.\end{aligned}$$ ]{} This completes the proof of [[(\[interessant1\])]{}]{}.
Let us now prove [[(\[interessant2\])]{}]{}. Inserting an appropriate power of $-A_{h, k}$, using [[(\[interessant1\])]{}]{}, Lemmas \[lemma0\] and \[lemmestrategie2\] yields $$\begin{aligned}
&&\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-\epsilon}\left(L^h_k\right)'\left(t_k+\frac{\Delta t}{2}\right)\right\Vert\nonumber\\
&\leq& \left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-\epsilon}(-A_{h,k})^{\epsilon}\right\Vert_{L(H)}\left\Vert\left(-A_{h,k}\right)^{-\epsilon}\left(L^h_k\right)'\left(t_k+\frac{\Delta t}{2}\right)\right\Vert\nonumber\\
&\leq& Ct_k^{-1+\beta/2}.\end{aligned}$$ This completes the proof of [[(\[interessant2\])]{}]{}. Let us complete the proof of the lemma with [[(\[interessant3\])]{}]{}. Taking the derivative in both sides of [[(\[vend1\])]{}]{} yields [$$\begin{aligned}
\label{vend3}
\left(L^h_k\right)''(t)&=&A_h''(t)u^h(t)+2A_h'(t)D_tu^h(t)+A_h(t)D_t^2u^h(t)
+P_h\frac{\partial^2F}{\partial t^2}\left(t,u^h(t)\right)\nonumber\\
&+&2P_h\frac{\partial^2F}{\partial t\partial u}\left(t, u^h(t)\right)D_tu^h(t)+P_h\frac{\partial^2F}{\partial u^2}\left(t,u^h(t)\right)\left(D_tu^h(t), D_tu^h(t)\right).\end{aligned}$$ ]{} Inserting $(-A_h(0))^{-1}$ in [[(\[vend3\])]{}]{}, taking the norm in both sides, using [[Lemma \[lemderiv\]]{}]{}, [[Lemma \[regularitylemma\]]{}]{}, [[Lemma \[ancien\]]{}]{} and the fact that $(-A_h(0))^{-1}$ is bounded yields $$\begin{aligned}
\label{vend4}
&&\Vert \left(-A_h(0))^{-1}(L^h_k\right)''(t)\Vert\nonumber\\
&\leq&\Vert (-A_h(0))^{-1}A_h''(t)\Vert_{L(H)}\Vert u^h(t)\Vert+2\Vert (-A_h(0))^{-1}A_h'(t)\Vert_{L(H)}\Vert D_tu^h(t)\Vert\nonumber\\
&+&\Vert (-A_h(0))^{-1}A_h(t)\Vert_{L(H)}\Vert D^2_tu^h(t)\Vert+C\left\Vert\frac{\partial^2F}{\partial t^2}\left(t,u^h(t)\right)\right\Vert\nonumber\\
&+&C\left\Vert\frac{\partial^2F}{\partial t\partial u}\left(t,u^h(t)\right)\right\Vert_{L(H)}\Vert D_tu^h(t)\Vert+C\left\Vert\frac{\partial^2F}{\partial t\partial u}\left(t,u^h(t)\right)\right\Vert_{L(H)}\left\Vert D_tu^h(t)\right\Vert\nonumber\\
&+&C\left\Vert\frac{\partial^2F}{\partial u^2}\left(t,u^h(t)\right)\right\Vert_{L(H\times H, H)}\left\Vert D_t^2u^h(t)\right\Vert\nonumber\\
&\leq& C+Ct^{-1+\beta/2}+Ct^{-2+\beta/2}\leq Ct^{-2+\beta/2}.\end{aligned}$$ The proof of [[(\[interessant4\])]{}]{} is similar to that of [[(\[interessant2\])]{}]{}. This completes the proof of [[Lemma \[utilemma\]]{}]{}.
\[bonlemma\] Let [[Assumption \[assumption2\]]{}]{} be fulfilled, let $m\in\{0, 1, \cdots, M\}$ and $0<t\leq T$. Then the following estimate holds $$\begin{aligned}
\label{utile1}
\left\Vert \left(-(A_{h,m}+J^h_m)\right)^{\alpha}e^{\left(A_{h,m}+J^h_m\right)t}\right\Vert_{L(H)}&=&
\left\Vert e^{\left(A_{h,m}+J^h_m\right)t}\left(-(A_{h,m}+J^h_m)\right)^{\alpha}\right\Vert_{L(H)}\nonumber\\
&\leq& Ct^{-\alpha},\quad \alpha\in[0, 1].\end{aligned}$$ Moreover, for $0\leq \alpha_1\leq\alpha_2\leq 1$ and any $0\leq t\leq T$, the following estimate holds [$$\begin{aligned}
\label{utile2}
\left\Vert (-(A_{h,m}+J^h_m))^{-\alpha_1}\varphi_1(\Delta t(A_{h,m}+J^h_m))(-(A_{h,m}+J^h_m))^{\alpha_2}
\right\Vert_{L(H)}\leq C\Delta t^{\alpha_1-\alpha_2}.\end{aligned}$$ ]{}
Let us start with the proof of [[(\[utile1\])]{}]{}. Note that for $\alpha=1$, using Assumption \[assumption2\] and \[assumption3\], we have $$\begin{aligned}
\label{utile1a}
\Vert e^{A_{h,m}t}(-(A_{h,m}+J^h_m))\Vert_{L(H)}&\leq& \Vert e^{A_{h,m}t}A_{h,m}\Vert_{L(H)}+\Vert e^{A_{h,m}t}J^h_m\Vert_{L(H)}\nonumber\\
&\leq& Ct^{-1}+C\leq Ct^{-1}.\end{aligned}$$ From [[(\[strategie2\])]{}]{}, it holds that $$\begin{aligned}
\label{utile1b}
e^{(A_{h,m}+J^h_m)t}(-(A_{h,m}+J^h_m))&=&e^{A_{h,m}t}(-(A_{h,m}+J^h_m))\\
&+&\int_0^te^{A_{h,m}(t-s)}J^h_me^{(A_{h,m}+J^h_m)s}(-(A_{h,m}+J^h_m))ds.\nonumber\end{aligned}$$ Taking the norm in both sides of [[(\[utile1b\])]{}]{} and using [[(\[utile1a\])]{}]{} yields $$\begin{aligned}
\label{utile1c}\left\Vert e^{(A_{h,m}+J^h_m)t}(-(A_{h,m}+J^h_m))\right\Vert_{L(H)}&\leq& Ct^{-1}\\
&+&C\int_0^t\left\Vert e^{(A_{h,m}+J^h_m)s}(-(A_{h,m}+J^h_m))\right\Vert_{L(H)}ds.\nonumber\end{aligned}$$ Applying the Gronwall’s lemma to [[(\[utile1c\])]{}]{} yields $$\begin{aligned}
\label{utile1d}
\left\Vert e^{(A_{h,m}+J^h_m)t}(-(A_{h,m}+J^h_m))\right\Vert_{L(H)}&\leq& Ct^{-1}.\end{aligned}$$ Note that [[(\[utile1\])]{}]{} obviously holds for $\alpha=0$. The intermediate cases therefore follow by interpolation technique and the proof of [[(\[utile1\])]{}]{} is completes. Let us now prove [[(\[utile2\])]{}]{}. From [[(\[phi1\])]{}]{}, it holds that $$\begin{aligned}
\label{utile4}
&&(-(A_{h,m}+J^h_m))^{-\alpha_1}\varphi_1\left(\Delta t\left(A_{h,m}+J^h_m\right)\right)(-(A_{h,m}+J^h_m))^{\alpha_2}\nonumber\\
&=&\frac{1}{\Delta t}\int_0^{\Delta t}e^{\left(A_{h,m}+J^h_m\right)(\Delta t-s)}(-(A_{h,m}+J^h_m))^{\alpha_2-\alpha_1}ds.\end{aligned}$$ Taking the norm in both sides of [[(\[utile4\])]{}]{} and using [[(\[utile1\])]{}]{} yields [$$\begin{aligned}
\Vert (-A_h(0))^{-\alpha_1}\varphi_1\left(\Delta t\left(A_{h,m}+J^h_m\right)\right)(-A_h(0))^{\alpha_2}\Vert_{L(H)}&\leq& C\Delta t^{-1}\int_0^{\Delta t}(\Delta t-s)^{\alpha_1-\alpha_2}ds\nonumber\\
&\leq& C\Delta t^{\alpha_1-\alpha_2}.\end{aligned}$$ ]{} This proves [[(\[utile2\])]{}]{}, and the proof of [[Lemma \[bonlemma\]]{}]{} is completed.
The following lemma can be found in [@Stig2].
\[lemmastig\] For all $\alpha_1, \alpha_2>0$ and $\alpha\in[0,1)$, there exist two positive constants $C_{\alpha_1,\alpha_2}$ and $C_{\alpha,\alpha_2}$ such that $$\begin{aligned}
\label{stig}
\Delta t\sum_{j=1}^mt_{m-j}^{-1+\alpha_1}t_j^{-1+\alpha_2}&\leq& C_{\alpha_1,\alpha_2}t_m^{-1+\alpha_1+\alpha_2},\\
\Delta t\sum_{j=1}^mt_{m-j}^{-\alpha}t_j^{-1+\alpha_2}&\leq& C_{\alpha,\alpha_2}t_m^{-\alpha+\alpha_2}.\end{aligned}$$
The proof of the first estimate of [[(\[stig\])]{}]{} follows from the comparison with the following integral $$\begin{aligned}
\int_0^t(t-s)^{-1+\alpha_1}s^{-1+\alpha_2}ds.\end{aligned}$$ The proof of the second estimate of [[(\[stig\])]{}]{} is a consequence of the first estimate.
Proof of [[Theorem \[mainresult1\]]{}]{} {#proof_main}
----------------------------------------
We split the error term in two parts via triangle inequality as follows $$\begin{aligned}
\Vert u(t_m)-u^h_m\Vert\leq \Vert u(t_m)-u^h(t_m)\Vert+\Vert u^h(t_m)-u^h_m\Vert=:V_1+V_2.\end{aligned}$$ The space error $V_1$ is estimated in [[Proposition \[proposition2\]]{}]{}. It remains to estimate the time error $V_2$. The initial value problem [[(\[semi\])]{}]{} in the subinterval $[t_m, t_{m+1}]$ can be written in the following form [$$\begin{aligned}
\frac{du^h}{dt}&=&\left[A_{h,m}+J^h_m\right]u^h(t)+a^h_mt+G^h_m\left(t_m+\frac{\Delta t}{2},u^h(t_m)\right)\nonumber\\
&+&\left(A_h(t)-A_{h,m}\right)u^h(t)+G^h_m(t,u^h(t))-G^h_m\left(t_m+\frac{\Delta t}{2},u^h(t_m)\right).\end{aligned}$$ ]{} Consequently, by the variation of constant formula, we have the following representation of the exact solution [$$\begin{aligned}
\label{semi3}
&&u^h(t_{m+1})\nonumber\\
&=&e^{\left(A_{h,m}+J^h_m\right)\Delta t}u^h(t_m)+\int_0^{\Delta t}e^{\left(A_{h,m}+J^h_m\right)(\Delta t-s)}L^h_m(s+t_m)ds\nonumber\\
&+&\int_0^{\Delta t}e^{\left(A_{h,m}+J^h_m\right)(\Delta t-s)}\left[G^h_m\left(t_m+\frac{\Delta t}{2},u^h(t_m)\right)+a^h_m(t_m+s)\right]ds\end{aligned}$$ ]{} where $L^h_k(t)$ is defined in [[Lemma \[utilemma\]]{}]{}. Let $e^h_{m+1}:=u^h_{m+1}-u^h(t_{m+1})$ be the time error at $t_{m+1}$ and $\delta^h_{m+1}$ be the defect defined by $$\begin{aligned}
\label{defect}
\delta^h_{m+1}:=\int_0^{\Delta t}e^{\left(A_{h,m}+J^h_m\right)(\Delta t-s)}L^h_m(s+t_m)ds.\end{aligned}$$ Taking the difference between [[(\[semi2\])]{}]{} and [[(\[semi3\])]{}]{} yields [$$\begin{aligned}
\label{recursion}
e^h_{m+1}&=&e^{\left(A_{h,m}+J^h_m\right)\Delta t}e^h_m-\delta^h_{m+1}\nonumber\\
&+&\Delta t\varphi
_1\left(\Delta t(A_{h,m}+J^h_m)\right)\left[G^h_m\left(t_m+\frac{\Delta t}{2},u^h_m\right)-G^h_m\left(t_m+\frac{\Delta t}{2},u^h(t_m)\right)\right].\end{aligned}$$ ]{} Iterating the error recursion [[(\[recursion\])]{}]{} and using the fact that $e^h_0=0$ yields [$$\begin{aligned}
e^h_{m}&=&\sum_{k=0}^{m-1}S^h_{m-1,k+1}\left[\Delta t\varphi_1\left(\Delta t(A_{h,k}+J^h_k)\right)\left(G^h_k\left(t_k+\frac{\Delta t}{2},u^h_k\right)-G^h_k\left(t_k+\frac{\Delta t}{2},u^h(t_k)\right)\right)-\delta^h_{k+1}\right]\nonumber\\
&=&\Delta t\sum_{k=0}^{m-1}S^h_{m-1,k+1}\varphi_1\left(\Delta t(A_{h,k}+J^h_k)\right)\left(G^h_k\left(t_k+\frac{\Delta t}{2},u^h_k\right)-G^h_k\left(t_k+\frac{\Delta t}{2},u^h(t_k)\right)\right)\nonumber\\
&-&\sum_{k=0}^{m-1}S_{m-1,k+1}^h\varphi_1\left(\Delta t\left(A_{h,k}+J^h_k\right)\right)\delta^h_{k+1}\nonumber\\
&=:&J_1+J_2,\end{aligned}$$ ]{} where $$\begin{aligned}
S^h_{m, k}:=\left(\prod_{j=k}^me^{\Delta t(A_{h, j}+J^h_j)}\right),\quad m, k\in\mathbb{N}.\end{aligned}$$ Using triangle inequality, [[Lemma \[stabilite\]]{}]{} and [[(\[remainder4\])]{}]{} yields $$\begin{aligned}
\Vert J_1\Vert\leq C\Delta t\sum_{k=0}^{m-1}\Vert S^h_{m-1,k+1}\Vert_{L(H)}\Vert e^h_k\Vert\leq C\Delta t\sum_{k=0}^{m-1}\Vert e^h_k\Vert.\end{aligned}$$ We therefore obtain the following estimate $$\begin{aligned}
\label{cast0}
\Vert e^h_{m}\Vert\leq C\Delta t\sum_{k=0}^{m-1}\Vert e^h_k\Vert+\Vert J_2\Vert.\end{aligned}$$ Assuming that the map $L^h_k$ is twice differentiable on $(t_k, t_{k+1})$, we obtain the following Taylor expansion [$$\begin{aligned}
\label{Taylor}
L^h_k(s+t_k)&=&\left(s-\frac{\Delta t}{2}\right)\left(L^h_k\right)'\left(t_k+\frac{\Delta t}{2}\right)\nonumber\\
&+&\left(s-\frac{\Delta t}{2}\right)^2\int_0^1(1-\sigma)\left(L^h_k\right)''\left(t_k+\frac{\Delta t}{2}+\sigma\left(s-\frac{\Delta t}{2}\right)\right)d\sigma,\end{aligned}$$ ]{} where $0<s<\Delta t$. Let the linear operator $\varphi_2$ be defined as follows $$\begin{aligned}
\label{phi2}
\quad\varphi_2\left(\Delta t\left(A_{h,m}+J^h_m\right)\right)&:=&\frac{1}{\Delta t^2}\int_0^{\Delta t}e^{\left(A_{h,m}+J^h_m\right)(\Delta t-s)}sds. \end{aligned}$$ The functions $\varphi_1$ and $\varphi_2$ satisfy the following relation $$\begin{aligned}
\label{relation}
\varphi_2(z)=\frac{\varphi_1(z)-1}{z}.\end{aligned}$$ Note that the operators $\varphi_1$ and $\varphi_2$ defined respectively in [[(\[phi1\])]{}]{} and [[(\[phi2\])]{}]{} also satisfy the following relation $$\begin{aligned}
\label{Oster}
&&\varphi_2\left(\Delta t\left(A_{h,m}+J^h_m\right)\right)-\frac{1}{2}\varphi_1\left(\Delta t\left(A_{h,m}+J^h_m\right)\right)\nonumber\\
&=&\Delta t\left(A_{h,m}+J^h_m\right)\chi\left(\Delta t\left(A_{h,m}+J^h_m\right)\right),\end{aligned}$$ where $\chi\left(\Delta t\left(A_{h,m}+J^h_m\right)\right)$ is a bounded linear operator. In particular, as in [@Ostermann1 (20)] or [@Gonza1 (2.8b)], one can easily check by using [@Antjd1 Lemma 9] that the following estimates hold for any $\gamma\geq 0$ [$$\begin{aligned}
\label{ganza1}
\Vert \varphi_1\left(\Delta t\left(A_{h,m}+J^h_m\right)\right)\Vert_{L(H)}+\Vert \varphi_2\left(\Delta t\left(A_{h,m}+J^h_m\right)\right)\Vert_{L(H)}&\leq& C,\\
\label{gonza2}
\Vert \left(-(A_{h,k}+J^h_k)\right)^{-\gamma}\chi\left(\Delta t\left(A_{h,m}+J^h_m\right)\right)\left(-(A_{h,k}+J^h_k)\right)^{\gamma}\Vert_{L(H)}&\leq& C.\end{aligned}$$ ]{} Taking in account [[(\[Taylor\])]{}]{} and [[(\[Oster\])]{}]{}, the defect [[(\[defect\])]{}]{} can be written as follows [$$\begin{aligned}
\label{repre1}
\delta^h_k&=&\Delta t^2\left(\varphi_2\left(\Delta t\left(A_{h,k}+J^h_k\right)\right)-\frac{1}{2}\varphi_1\left(\Delta t\left(A_{h,k}+J^h_k\right)\right)\right)\left(L^h_k\right)'\left(t_k+\frac{\Delta t}{2}\right)\\
&+&\int_0^{\Delta t}e^{(\Delta t-s)\left(A_{h,k}+J^h_k\right)}\left(s-\frac{\Delta t}{2}\right)^2\int_0^1(1-\sigma)\left(L^h_k\right)''\left(t_k+\frac{\Delta t}{2}+\sigma\left(s-\frac{\Delta t}{2}\right)\right)d\sigma ds.\nonumber\end{aligned}$$ ]{} Substituting [[(\[Oster\])]{}]{} in [[(\[repre1\])]{}]{} yields [$$\begin{aligned}
\label{repre2}
\delta^h_k&=&\Delta t^3\left(A_{h,k}+J^h_k\right)\chi\left(\Delta t\left(A_{h,k}+J^h_k\right)\right)\left(L^h_k\right)'\left(t_k+\frac{\Delta t}{2}\right)\nonumber\\
&+&\int_0^{\Delta t}e^{(\Delta t-s)\left(A_{h,k}+J^h_k\right)}\left(s-\frac{\Delta t}{2}\right)^2\int_0^1(1-\sigma)\left(L^h_k\right)''\left(t_k+\frac{\Delta t}{2}+\sigma\left(s-\frac{\Delta t}{2}\right)\right)d\sigma ds.\nonumber\\
&=:&\delta^{(1)h}_k+\delta^{(2)h}_k.\end{aligned}$$ ]{} Before proceeding further, we claim that $$\begin{aligned}
\label{fact2}
\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1}\delta^{(2)h}_k\right\Vert \leq C\Delta t^3t_k^{-2+\beta/2}.\end{aligned}$$ In fact, using [[Lemma \[utilemma\]]{}]{}, [[Lemma \[lemma0\]]{}]{} and [@Antjd1 Lemma 9] it holds that [$$\begin{aligned}
\label{mardi1}
&&\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1}\delta^{(2)h}_k\right\Vert\nonumber\\
&\leq& C\int_0^{\Delta t}\left\Vert e^{(\Delta t-s)\left(A_{h,k}+J^h_k\right)}\right\Vert_{L(H)}\left(s-\frac{\Delta t}{2}\right)^2\nonumber\\
&&\int_0^1(1-\sigma)\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1} \left(L^h_k\right)''\left(t_k+\frac{\Delta t}{2}+\sigma\left(s-\frac{\Delta t}{2}\right)\right)\right\Vert d\sigma ds\\
&\leq& C\int_0^{\Delta t}\left(s-\frac{\Delta t}{2}\right)^2\int_0^1(1-\sigma)\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1} \left(L^h_k\right)''\left(t_k+\frac{\Delta t}{2}+\sigma\left(s-\frac{\Delta t}{2}\right)\right)\right\Vert d\sigma ds.\nonumber\end{aligned}$$ ]{} Since $$\begin{aligned}
\frac{\Delta t}{2}+\sigma\left(s-\frac{\Delta t}{2}\right)\geq 0,\quad s\in[0,\Delta t],\quad \sigma\in[0,1],\end{aligned}$$ it follows from [[Lemma \[utilemma\]]{}]{} that $$\begin{aligned}
\label{mardi2}
\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1} \left(L^h_k\right)''\left(t_k+\frac{\Delta t}{2}+\sigma\left(s-\frac{\Delta t}{2}\right)\right)\right\Vert \leq Ct_k^{-2+\beta/2}, \end{aligned}$$ for $s\in[0,\Delta t]$ and $\quad \sigma\in[0,1]$. Substituting [[(\[mardi2\])]{}]{} in [[(\[mardi1\])]{}]{} yields $$\begin{aligned}
\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1} \delta^{(2)h}_k\right\Vert&\leq& C\int_0^{\Delta t}\int_0^1(1-\sigma)\left(s-\frac{\Delta t}{2}\right)^2t_k^{-2+\beta/2}d\sigma ds\nonumber\\
&\leq& C\Delta t^3t_k^{-2+\beta/2}.\end{aligned}$$ We can also easily check that $$\begin{aligned}
\label{very1}
\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1-\epsilon}\delta_k^{(1)h}\right\Vert\leq C\Delta t^3t_k^{-1+\beta/2}.\end{aligned}$$ In fact, employing [[Lemma \[utilemma\]]{}]{} and [[(\[gonza2\])]{}]{}, it holds that $$\begin{aligned}
\label{fact1}
&&\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1-\epsilon}\delta_k^{(1)h}\right\Vert\nonumber\\
&\leq& C\Delta t^3\left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-\epsilon}\chi\left(\Delta t\left(A_{h,k}+J^h_k\right)\right)\left(-(A_{h,k}+J^h_k)\right)^{\epsilon}\right\Vert_{L(H)}\nonumber\\
&&\times\left\Vert\left(-(A_{h,k}+J^h_k)\right)^{-\epsilon}\left(L^h_k\right)'\left(t_k+\frac{\Delta t}{2}\right)\right\Vert\nonumber\\
&\leq& C\Delta t^3t_k^{-1+\beta/2}.\end{aligned}$$ Note that $J_2$ can be recast in two terms as follows $$\begin{aligned}
\label{cast1}
J_2&=&-\sum_{k=0}^{m-1}S^h_{m-1,k+1}\varphi_1\left(\Delta t\left(A_{h,k}+J^h_k\right)\right)\delta^{(1)h}_{k+1}\nonumber\\
&-&\sum_{k=0}^{m-1}S^h_{m-1,k+1}\varphi_1\left(\Delta t\left(A_{h,k}+J^h_k\right)\right)\delta^{(2)h}_{k+1}\nonumber\\
&=:& J_{21}+J_{22}.\end{aligned}$$ Using [[Lemma \[bonlemma\]]{}]{}, [[(\[very1\])]{}]{}, [[Lemma \[stabilite\]]{}]{}, [[Lemma \[lemmastig\]]{}]{} and [[Lemma \[lemmestrategie2\]]{}]{} it holds that $$\begin{aligned}
\label{cast2}
&&\Vert J_{21}\Vert\nonumber\\
&\leq& \sum_{k=0}^{m-1}\left\Vert S^h_{m-1,k+1}(-A_h(0))^{1-\epsilon/2}\right\Vert_{L(H)}\nonumber\\
&&\times\left\Vert(-A_h(0))^{-1+\epsilon/2}\varphi_1\left(\Delta t\left(A_{h,k}+J^h_k\right)\right)\left(-(A_{h,k}+J^h_k)\right)^{1+\epsilon/2}\right\Vert_{L(H)}\nonumber\\
&&\times \left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1-\epsilon/2}\delta^{(1)h}_{k+1}\right\Vert\nonumber\\
&\leq& C\Delta t^{3}\sum_{k=0}^{m-1}t_{m-k-1}^{-1+\epsilon}t_{k+1}^{-1+\beta/2}\left\Vert(-A_h(0))^{-1+\epsilon/2}\left(-(A_{h,k}+J^h_k)\right)^{1-\epsilon/2}\right\Vert_{L(H)} \nonumber\\
&&\times\left\Vert\left(-(A_{h,k}+J^h_k)\right)^{-1+\epsilon/2}\varphi_1\left(\Delta t\left(A_{h,k}+J^h_k\right)\right)\left(-(A_{h,k}+J^h_k)\right)^{1+\epsilon/2}\right\Vert_{L(H)}\nonumber\\
&\leq& C\Delta t^{3-\epsilon}\sum_{k=0}^{m-1}t_{m-k-1}^{-1+\epsilon}t_{k+1}^{-1+\beta/2}\nonumber\\
&\leq& C\Delta t^{2-\epsilon}.\Delta t\sum_{k=0}^{m-1}t_{m-1-k}^{-1+\epsilon}t_{k+1}^{-1+\beta/2}\nonumber\\
&\leq& C\Delta t^{2-\epsilon}t_{m-1}^{-1+\beta+\epsilon}\leq C\Delta t^{2-\epsilon} \Delta t^{-1+\beta/2}\leq C\Delta t^{1+\beta/2-\epsilon}.\end{aligned}$$ Using [[Lemma \[bonlemma\]]{}]{}, [[(\[fact2\])]{}]{} and [[Lemma \[stabilite\]]{}]{}, it holds that $$\begin{aligned}
\label{cast3a}
\Vert J_{22}\Vert&\leq& \sum_{k=0}^{m-1}\left\Vert S^h_{m-1,k+1}(-A_h(0))^{1-\epsilon}\right\Vert_{L(H)}\nonumber\\
&&\times\left\Vert(-A_h(0))^{-1+\epsilon}\varphi_1\left(\Delta t\left(A_{h,k}+J^h_k\right)\right)\left(-(A_{h,k}+J^h_k)\right)\right\Vert_{L(H)}\nonumber\\
&&\times \left\Vert \left(-(A_{h,k}+J^h_k)\right)^{-1}\delta^{(2)h}_{k+1}\right\Vert\nonumber\\
&\leq& C\Delta t^3\sum_{k=0}^{m-1} t_{k+1}^{-2+\beta/2}\left\Vert(-A_h(0))^{-1+\epsilon}\left(-(A_{h,k}+J^h_k)\right)^{1-\epsilon}\right\Vert_{L(H)}\nonumber\\
&&\times\left\Vert\left(-(A_{h,k}+J^h_k)\right)^{-1+\epsilon}\varphi_1\left(\Delta t\left(A_{h,k}+J^h_k\right)\right)\left(-(A_{h,k}+J^h_k)\right)\right\Vert_{L(H)}\nonumber\\
&\leq& C\Delta t^3\sum_{k=0}^{m-1}\Delta t^{-\epsilon} t_{k+1}^{-2+\beta/2}\nonumber\\
&\leq&C\Delta t^{2-\epsilon}\,\Delta t\sum_{k=0}^{m-1}t_{k+1}^{-2+\beta/2}.\end{aligned}$$ Note that [$$\begin{aligned}
\label{cast3b}
\Delta t\sum_{k=0}^{m-1}t_{k+1}^{-2+\beta/2}=\Delta t^{-1+\beta/2}\sum_{k=0}^{m-1}(k+1)^{-2+\beta/2}=\Delta t^{-1+\beta/2}\sum_{k=1}^mk^{-2+\beta/2}.\end{aligned}$$ ]{} The sequence $v_k=k^{-2+\beta/2}$ is decreasing. Therefore, by comparison with the integral we have $$\begin{aligned}
\label{cast3c}
\sum_{k=1}^mv_k=\sum_{k=1}^mk^{-2+\beta/2}\leq 1+\int_1^mt^{-2+\beta/2}dt\leq 1+Cm^{-1+\beta/2}.\end{aligned}$$ Substituting [[(\[cast3c\])]{}]{} in [[(\[cast3b\])]{}]{} yields $$\begin{aligned}
\label{cast3d}
\Delta t\sum_{k=0}^{m-1}t_{k+1}^{-2+\beta/2}\leq C\Delta t^{-1+\beta/2}+Ct_m^{-1+\beta/2}.\end{aligned}$$ Substituting [[(\[cast3d\])]{}]{} in [[(\[cast3a\])]{}]{} yields $$\begin{aligned}
\label{cast3}
\Vert J_{22}\Vert\leq C\Delta t^{1+\beta/2-\epsilon}+C\Delta t^{2-\epsilon}t_m^{-1+\beta/2}\leq C\Delta t^{1+\beta/2-\epsilon}.\end{aligned}$$ Substituting [[(\[cast3\])]{}]{} and [[(\[cast2\])]{}]{} in [[(\[cast1\])]{}]{} yields $$\begin{aligned}
\label{cast4}
\Vert J_2\Vert\leq \Vert J_{21}\Vert+\Vert J_{22}\Vert\leq C\Delta t^{1+\beta/2-\epsilon}.\end{aligned}$$ Substituting [[(\[cast4\])]{}]{} in [[(\[cast0\])]{}]{} yields $$\begin{aligned}
\label{cast5}
\Vert e^h_m\Vert\leq C\Delta t^{1+\beta/2-\epsilon}+C\Delta t\sum_{k=0}^{m-1}\Vert e^h_k\Vert.\end{aligned}$$ Applying the discrete Gronwall’s inequality to [[(\[cast5\])]{}]{} yields $$\begin{aligned}
\Vert e^h_m\Vert\leq C\Delta t^{1+\beta/2-\epsilon}.\end{aligned}$$ This completes the proof of [[Theorem \[mainresult1\]]{}]{}.
Numerical simulations {#numericalexperiment}
=====================
We consider the following reactive advection diffusion reaction with diagonal difussion tensor $$\begin{aligned}
\label{reactiondif1}
\dfrac{\partial u}{\partial t}=\left[D(t) \left(\varDelta u-\nabla \cdot(\mathbf{v}u)\right)+\dfrac{e^{-t} u}{\vert u\vert +1}\right],\end{aligned}$$ with mixed Neumann-Dirichlet boundary conditions on $\Lambda=[0,L_1]\times[0,L_2]$. The Dirichlet boundary condition is $u=1$ at $\Gamma=\{ (x,y) :\; x =0\}$ and we use the homogeneous Neumann boundary conditions elsewhere. The initial solution is $u(0)=0$. To check our theoritical result in [[Theorem \[mainresult1\]]{}]{}, we use $D(t)=1+e^{-t}$. For comparaison with current exponential Rosenbrock method [@Antjd2] for constant operator $A$, we have taken $D(t)=1$. In [[Figure \[FIGII\]]{}]{}, we will use the following notations
- ’Magnus-Rosenbrock’ is used for the errors graph of the Magnus Rosenbrock scheme for the nonautonomous equation [[(\[reactiondif1\])]{}]{} corresponding to the coefficient $D(t)=1+e^{-t}$.
- ’C-Magnus-Rosenbrock’ is used for the errors graph of the novel Magnus Rosenbrock scheme for fixed coefficient $D(t)=1$ in [[(\[reactiondif1\])]{}]{} (constant operator linear operator).
- ’Exponential-Rosenbrock’ is used for the errors graph for the second order exponential Euler Rosenbrock scheme [@Antjd2] for fixed coefficient $D(t)=1$ in [[(\[reactiondif1\])]{}]{}(constant operator linear operator).
In all graphs, the reference solution or ’exact solution’ is numerical solution with the smaller time step $\Delta t= 1/4096$. The linear operator $A(t)$ is given by $$\begin{aligned}
A(t)=(1+e^{-t})\left(\varDelta(.)-\nabla. \mathbf{v}(.)\right),\quad t\in[0, T],\end{aligned}$$ where $\mathbf{v}$ is the Darcy velocity obtained as in [@Antonionews Fig 6]. Clearly $\mathcal{D}(A(t))=\mathcal{D}(A(0))$, $t\in[0, T]$ and $\mathcal{D}((-A(t))^{\alpha})=\mathcal{D}((-A(0))^{\alpha})$, $t\in[0, T]$, $0\leq \alpha\leq 1$. The function $q_{ij}(x,t)$ defined in [[(\[family\])]{}]{} is given by $q_{ii}(x,t)=1+e^{-t}$, and $q_{ij}(x,t)=0,\, i\neq j$ . Since $q_{ii}(x,t)$ is bounded below by $1+e^{-T}$, it follows that the ellipticity condition [[(\[ellip\])]{}]{} holds and therefore as a consequence of [[Section \[numscheme\]]{}]{}, it follows that $A(t)$ is sectorial. Obviously [[Assumption \[assumption2\]]{}]{} is fulfills. The nonlinear function $F$ is given by $F(t,v)= \dfrac{e^{-t} v}{1+\vert v\vert}$, $t\in[0, T]$, $v\in H$ and obviously satisfies [[Assumption \[assumption3\]]{}]{}. Let $f:[0, T]\times \Lambda\times\mathbb{R}\longrightarrow \mathbb{R}$ be defined by $f(t, x, z)=\frac{e^{-t}z}{1+\vert z\vert}$. We take $F :[0, T]\times H\longrightarrow H$ to be the Nemytskii operator defined as follows $$\begin{aligned}
(F(t, v))(x)=f(t, x, v(x)),\quad t\in[0, T],\quad x\in \Lambda,\quad v\in H.
\end{aligned}$$ One can easily check that $$\begin{aligned}
\frac{\partial f}{\partial z}(t, x, z)=-\frac{e^{-t}\vert z\vert}{(1+\vert z\vert)^2},\quad (t, x, z)\in[0, T]\times\Lambda\times\mathbb{R}.
\end{aligned}$$ Therefore $$\begin{aligned}
(F'(t, v))(u(x))=\frac{\partial f}{\partial z}(t, x, v(x)).u(x)=-\frac{e^{-t}\vert v(x)\vert}{(1+\vert v(x)\vert)^2}.u(x).
\end{aligned}$$ One can easily check that $$\begin{aligned}
\label{deriveebornee}
\frac{e^{-t}\vert v(x)\vert}{(1+\vert v(x)\vert)^2}\leq \frac{e^{-t}\vert v(x)\vert}{1+\vert v(x)\vert}\leq e^{-t}\leq C,\quad t\in[0, T],\quad x\in \Lambda,\quad v\in H.\end{aligned}$$ Therefore, it holds that $$\begin{aligned}
\left\Vert \frac{\partial F}{\partial u}(t,u)\right\Vert_{L(H)}\leq C,\quad -\left\langle F'(t, u)v, v\right\rangle_H\geq 0,\quad t\in[0, T],\quad u,v \in H.\end{aligned}$$ One can also obviously prove that $$\begin{aligned}
\left\Vert\frac{\partial^kF }{\partial t\partial u}(t,u)\right\Vert_{L(H)}\leq C,\quad\left\Vert\frac{\partial^2F}{\partial u^2}(t,u)\right\Vert_{L(H\times H; H)}\leq C,\nonumber\end{aligned}$$ for all $t\in[0, T]$ and $u\in H$. Hence [[Assumption \[assumption3\]]{}]{} is fulfilled.
![Convergence of the Magnus Rosenbrock scheme at final time $T=1$. For constant coefficient $D(t)=1$, we have compared the Magnus Rosenbrock scheme with the second order exponential Euler Rosenbrock scheme [@Antjd2]. The order of convergence in time is $1. 92$ Magnus Rosenbrock scheme (with $D(t)=1+e^{-t}$), $1.95$ for the Magnus Rosenbrock scheme (with $D(t)=1$)and $2.08$ for the second order exponential Euler Rosenbrock scheme. []{data-label="FIGII"}](Dmagnus.eps){width="60.00000%"}
In [[Figure \[FIGII\]]{}]{}, we can observe the convergence of the Magnus Rosenbrock scheme ($D(t)=1+e^{-t}$ and $D(t)=1$), and the second order exponential Euler Rosenbrock scheme ($D(t)=1$). The order of convergence in time is $1. 92$ for Magnus Rosenbrock scheme ($D(t)=1+e^{-t}$), $1.95$ for the Magnus Rosenbrock scheme ($D(t)=1$) and $2.08$ for the second order exponential Euler Rosenbrock scheme ($D(t)=1$). As we can also observe, the convergence orders in time of the Magnus Rosenbrock scheme are well in agreement with our theoretical result in [[Theorem \[mainresult1\]]{}]{} as the theoretical order is $2$ with order reduction $\epsilon$, which is very small here.
[00]{}
Amann, H.:
Blanes, S., Casa, F., Oteo, J. A., Ros, J.:
Blanes, S., Casas, F., Oteo, J. A., Ros, J.:
Blanes, S., Moan, P. C.:
Ciarlet, P. G.:
Elliot, C., Larsson, S.,
Engel, K. J., Nagel, R.:
Evans, L. C.:
Fujita, H., Suzuki, T.: Evolutions problems (part1). in: P. G. Ciarlet and J. L. Lions(eds.), Handb. Numer. Anal., vol. II, North-Holland, 789-928 (1991)
Gondal, M. A.:
González, C., Ostermann, A.:
González, C., Ostermann, A., . Thalhmmer, M.:
González, C., Ostermann, A., Palencia, C., Thalhammer, M.:
González, C., Thalhmmer, M.:
Hairer, E., Wanner, G.: ,
Henry, D.:
Hipp, D., Hochbruck, M., Ostermann, A.:
Hochbruck, M., Ostermann, A., Schweitzer, J.:
Hochbruck, M. Lubich, C.:
Hochbruck, M., Ostermann, A.:
Iserles, A., Munthe-Kaas, H. Z., Nørsett S. P., Zanna, A.:
Jentzen, A., Kloeden, P. E.:
Jentzen, A., Kloeden, P. E., Winkel, G.:
Kloeden, P. E., Platen, E.:
Larsson, S.:
Larsson, S.:
Leykekhman, D., Vexler, B.:
Lord G. J., Tambue, A.:
Lu, Y. Y.:
Lunardi, A.:
Luskin, M., Rannacher, R.:
Magnus, M.:
Mingyou. H., Thomee, V.:
Mukam, J. D., Tambue, A.:
Mukam, J. D., Tambue, A.:
Nambu, T.:
Ostermann, A., Thalhammer, M.:
Ostermann, A., Thalhammer, M.:
Pazy, A.: Seely, R.:
Seidler, J.:
Tambue, A., Lord G. J, Geiger, S.: Tambue, A., Ngnotchouye, J. M. T.:
Tanabe, H.:
Caliari,M., Vianello M., Bergamaschi L. Thomée, V.:
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Nonreflexive quantum mechanics is a formulation of quantum theory based on a non-classical logic termed [nonreflexive logic]{} (a.k.a. ‘non-reflexive’). In these logics, the standard notion of identity, as encapsulated in classical logic and set theories, does not hold in full. The basic aim of this kind of approach to quantum mechanics is to take seriously the claim made by some authors according to whom quantum particles are [non-individuals]{} in some sense, and also to take into account the fact that they may be absolutely indistinguishable (or indiscernible). The nonreflexive formulation of quantum theory assumes these features of the objects already at the level of the underlying logic, so that no use is required of symmetrization postulates or other mathematical devices that serve to pretend that the objects are indiscernible (when they are not: all objects that obey classical logic are [individuals]{} in a sense). Here, we present the ideas of the development of nonreflexive quantum mechanics and discuss some philosophical (mainly metaphysical) motivations and consequences of it.'
author:
- '[Décio Krause]{}[^1]'
- '[Jonas R. B. Arenhart]{}'
date: |
Research Group in Logic and Foundations of Science\
Department of Philosophy\
Federal University of Santa Catarina\
(Forthcoming in [Cadernos de História e Filosofia da Ciência]{}, Campinas - SP, Brazil)\
October 2015
title: 'Presenting Nonreflexive Quantum Mechanics: Formalism and Metaphysics'
---
Introduction
============
Indiscernibility of physical systems (or physical objects) appear already in classical physics. For instance, in 1808 John Dalton, the founder of modern atomism, claimed that there would be no differences among chemical elements of a same kind: “we may conclude that the ultimate particles of all homogeneous bodies are perfectly alike in weight, figure, etc.” [@dal08 p.142-3]. In 1900, Max Planck, still reasoning within the scope of classical physics, made the fundamental move to quantum mechanics by assuming that in distributing an integral number $N$ of energy elements over $C$ states, the indiscernibility of the elements should be taken into account. Planck’s formula $$Z = \frac{(N+C-1)!}{N! (C-1)!}$$ shows that the division by $N!$ makes the trick. Some years before, the so-called Gibbs paradox appeared in classical statistical mechanics showing that indiscernible particles are not only part of the physics, but also of the philosophical problem concerning the proper understanding of how an appropriate metaphysics for physical theories would look like. Indeed, should we take the natural approach suggested by the Sackur-Tetrode formula for the entropy of an ideal gas and accept that nature made particles indistinguishable and that there is no more talking about the topic? Or should we adopt a minimalist approach and concede that indistinguishability may help us in drawing the line between classical and quantum particles (see French and Krause [@frekra06 chap.2])? The first route would blur the difference between classical and quantum particles, both being absolutely indistinguishable in some cases. In the second case, we may argue that classical particles are different from their quantum counterparts for they may be considered to be individuals and distinguishable in some sense, while the matter is not so easily settled for quantum particles, which, due to their indiscernibility, may be taken as some kind of [non-individuals]{}.
The word ‘non-individuals’ has an historical use. Several forerunners of quantum theory have referred to quantum objects this way, mainly in the sense that quantum particles “are not individuals”, as suggested by Schrödinger [@sch57 p.206]. But the term ‘non-individual’ may confuse the reader in suggesting that these entities would be something different from that which we can refer to or speak about. In fact, we don’t know in precise terms what they are. Each theory gives its own characteristics to these entities (B. Falkenburg traces the ‘metamorphoses’ of the term ‘particle’ from classical physics to the most advanced physical theory of today [@fal07 Chap.6]). Being non-individuals does not prevent quantum particles from being able to be isolated by some time, but that they cannot be regarded as individuals [in the standard sense of the word]{}, as something that have an identity and can be identified in other times even if mixed with others of similar species. As it is well known, this does not happen with quantum particles. It is in this sense that we say that all protons are absolutely indiscernible, so are all electrons, all neutrons, and do on. Hence, all atoms of the same chemical isotope, made from protons, neutrons and electrons, are also indiscernible. Without this hypothesis, chemistry does not work, and quantum physics does not work either.
We shall take the second approach mentioned earlier and concede that there is a significant difference between quantum and classical particles. In classical mechanics we can at least in principle follow well-defined trajectories for each particle and by adding a reasonable assumption concerning their impenetrability we may grant that some form of spatio-temporal principle of individuality holds. When we turn to quantum mechanics, on the other hand, things get much more complicated; in fact, for quantum particles there is no trajectory in the classical sense, unless a hidden-variables approach is followed, which has its own shortcomings; the impenetrability assumption is also unavailable, for it may happen that particles are in entangled states. More than that, quantum mechanics encapsulates indistinguishability in the statistical treatment of aggregates of particles through the famous Bose-Einstein and Fermi-Dirac statistics. These are seen as grounding the fact that in some cases nothing whatsoever may distinguish quantum particles, for as is famously remarked, permutations of particles do not count in this statistics as giving rise to different states.
One of the difficulties with this fact — quantum indistinguishability as obtained through the statistics of the theory — concerns precisely the relationship between the experimental fact that is being expressed and its mathematical formulation. Experimentally, it is found that there are situations in which one cannot distinguish between two particles of the same kind, as the theory predicts; mathematically, though, since the formalism of the theory is constructed within classical set theory using a part of functional analysis, the indistinguishability must be brought into the theory through the assumption of some symmetry conditions that grant us the desired effect. In a standard set theory, let us emphasize, whenever we have [two]{} objects, they are necessarily discernible: they are [distinct]{}. Now, from a philosophical point of view the assumption of these symmetry conditions seems to be at odds with the idea of genuinely indistinguishable entities, for it seems to be a mechanism to mask the fact that the entities dealt with are not really indistinguishable, since they belong to classical logic and set theory (see the discussions in Krause and Arenhart [@kraare15]).
The difficulty involving standard mathematics and logic and their relation to ‘genuine’ indistinguishability can be expressed in a nutshell as follows: on the one hand, orthodox quantum mechanics says that there are situations according to which we cannot discern the quanta; on the other hand, within classical mathematics, we represent this fact using symmetry postulates, whose aim is to mask the fact that in standard logic and mathematics one can always know that wherever there are two items, they are discernible by purely logical means inside set theory. We may even opt for a realistic interpretation according to which, even conceding indistinguishability, the particles [remain there]{} with their hidden properties, and even in this case there is no way to tell them apart. So, concerning this situation, three options seem to be available: (i) either the problem concerns the relationship between the entities dealt with by QM and standard mathematics, so that those entities may not sit comfortably together with assumptions made in classical mathematics, or (ii) there is some kind of hidden variable which may confer individuality to quantum entities, but, being hidden, is not taken into account by the standard formalism, or (iii) the quanta must be treated mathematically as certain entities which are discernible but their discernibility cannot be expressed in qualitative terms by the vocabulary of the theory. Perhaps the last two may be seen as just one option, but nothing in principle prevents us from distinguishing them. In fact, (ii) allows that in principle we could extend the formalism to accommodate the hidden variables accounting for the individuality, something (iii) in this reading does not allow for; in this last case the individuality is there but the theory does not allow us to express how it is constituted.
All these options present their own challenges; after briefly discussing items (ii) and (iii) in sections \[ind\] and \[sta\], we turn to the first one. This is perhaps the place to say that we are not claiming that QM violates standard mathematics and logic; our point is that it seems that this theory can be formulated in different and non equivalent ways, most of them within these ‘classical’ frameworks. In saying that, we are not assuming that there is just one QM, we take it as a fact that there are different formulations of different theories which are empirically equivalent. What we intent to do is to explore an alternative approach and look for the (at least conceptual) gains we may have. As we shall see in section \[qsp\], there is an alternative formulation/construction of quantum mechanics that incorporates non-individuality right from the start and makes use of this fact in order to develop the formalism of the theory. It is our goal here to render this formulation more easily accessible and clear. Once that is done, the philosophical view that quantum particles are not individuals (along with some simple logical theory of quasi-set) is seen as giving rise to fruitful developments of the theory itself, as an instance of a kind of reflexive equilibrium procedure: the standard mathematical formulation and experiments suggest that quantum particles are not individuals; however, that mathematical apparatus does not reflect that non-individuality, so, we use that fact in order to develop a mathematical formalism that takes non-individuality into account; this formalism, now, gives rise to another version of quantum mechanics. We discuss this view in section \[dis\].
Indistinguishability as a quantum phenomenon {#ind}
============================================
To better understand the reasons why quantum particles are considered to be absolutely indistinguishable and how that absolute individuality is supposed to lead to their [non-individuality]{}, we may start by comparing quantum and classical statistics. This is the standard [trope]{} on this issue, and we believe it is a good strategy to be followed, because since we have no access to the particles themselves, it seems that their collective behavior may give us some clues as to their identity and (in)distinguishability (see French and Krause [@frekra06 chap.4]).
The first point to be noticed is that there is some sense in which both classical and quantum particles are taken to be indistinguishable: when they share all their intrinsic (or state independent) properties. Sometimes it is common to call particles sharing all their intrinsic properties ‘identical particles’. However, this is just an abuse of language to say that the particles are of the same kind. The similarities among quantum and classical stop there, however. The main feature behind the claim that classical particles are individuals after all and that they are not really indistinguishable in the same sense that quantum particles are said to be concerns their collective behavior; the way that classical statistical mechanics describes aggregates of classical particles differs significantly from the way quantum mechanics describes aggregates of quantum particles.
Classical particles behave collectively according to Maxwell-Boltzmann statistics (MB). Let us illustrate it by considering the particular case in which there are two particles, labeled $1$ and $2$, to be distributed in two different states, $A$ and $B$. Then, according to MB statistics, we can have the following cases:
1. A(1) and A(2);
2. B(1) and B(2);
3. A(1) and B(2);
4. A(2) and B(1).
Assuming the equiprobability hypothesis — the claim that none of these cases has more chances to occur than the others (a claim not without its difficulties, see French and Krause [@frekra06 chap. 2]) — implies that all these possibilities are assigned the same weight, that is, $\frac{1}{4}$, and despite being considered as indistinguishable in the sense discussed above, permutations of particles, as it happens in the cases 3 and 4, are counted as giving rise to distinct possibilities. Here lies the key where the alleged difference between classical and quantum particles concerning identity and individuality matters.
The fact that permutations should bee seen as giving rise to different situations is generally taken to witness for their individuality. In fact, even though the particles share all their intrinsic properties, this simple fact seems to speak in favor of some individuality being manifested by the particles as they are permuted. Their individuality accounts for the difference in cases 3 and 4. Another way to put it is in terms involving state descriptions as describing a possible world: a permutation of particles in case 3 leads to a distinct possible world, one in which it is particle 2 that is in A and 1 in B. What accounts for the differences in the two worlds? Not the properties of the particles, assuming them indiscernible. So, it must be the case that they have an individuality intrinsic to them that is responsible for this possibility. That individuality can be understood in terms of some underlying substratum, a [haecceity]{}, or even more simply in terms of the trajectories (spatio-temporal location) with the impenetrability assumption of which we have already spoken about earlier (see, for instance, Ladyman [@lad07] for a discussion in those terms).
In quantum mechanics, on the other hand, nothing like this happens. To keep with the case of two particles labeled once again $1$ and $2$ distributed in states $A$ and $B$, we have two cases, one for bosons and other for fermions. For the bosons, the following possibilities obtain:
1. A(1) and A(2);
2. B(1) and B(2);
3. A(\*) and B(\*).
Again assuming equiprobability for each of these cases, each of them will have the same probability of $\frac{1}{3}$ to occur. Notice that in case 3 we have employed $*$ instead of the particles labels; that accounts for the distinctive feature of quantum statistics: the fact that permutations do not give rise to a distinct situation. The $*$ then means [one]{} particle in $A$ and [one]{} in $B$. The same idea can be used for Fermi-Dirac statistics, the one employed for fermions:
1. A(\*) and B(\*).
For fermions, as for bosons, permutations change nothing, but differently from bosons, the fermions obey the famous Pauli Exclusion Principle, which roughly speaking grants that no two fermions can occupy the same state; thus there is only one option in this particular case. The great novelty of quantum statistics comes from the fact that permutations do not give rise to distinct states. In terms of the description involving possible worlds above, a permutation does not give rise to a distinct world. Contrarily to the classical case, there is nothing in the particles to ground the fact that the situation before the permutation is distinct from the situation after the permutation. Thus, they are the same situation. As a consequence, there is no intrinsic individuality principle doing that job.
Obviously, if that kind of reasoning is going to ground the claim that quantum particles have no individuality and are really indistinguishable, as we wish it to do, then it must be recognized that our language is inadequate to express that fact, for we began by labeling the particles, which amounts to individuate and identify them (something we cannot do for absolutely indistinguishable items), and then we were forced to ‘forget’ the labels and place an $*$ in their place to neutralize the fact that if a label was allowed in the counting we would have intuitively the same classical statistics given by Maxwell-Boltzmann. Now, this is precisely the conundrum between the “nature” of quantum particles and the language we use to talk about them. To deal more properly with this topic, let us see how the formalism of quantum mechanics deals with the statistics. This will help us to appreciate the relevance of the alternative formulation of quantum mechanics we introduce later.
The standard way to deal with indistinguishable objects in QM {#sta}
=============================================================
We would like to emphasize that in standard approaches to the formalism of quantum mechanics, given inside some classical set theory, quanta are individuals (as entities represented either by sets or by ur-elements) whose individuality is veiled by symmetry principles that are added by hand. In fact, the standard formalism of quantum mechanics masks the individuality of the particles through the use of a mathematical trick: one begins with labelled individuals (particles 1 and 2, say) and by convenient manipulation, a clever postulate mimics the behavior of indiscernible non-individuals, say by means of symmetric and anti-symmetric vectors/functions. The price to be paid for the use of such a [man[œ]{}uvre]{} is mainly philosophical, since physics works for all practical purposes: one wants to be committed with an ontology of non-individuals, but in the end, all we are left with are the same old individuals we started with, clothed as non-individuals by the symmetry postulates asserting the invariance of their permutability. Let us see precisely how it is done, for when we know what we wish to avoid, it is then easier to fully appreciate what is gained by the use of the [$\mathfrak{Q}$]{}-spaces introduced below. Also, this way of approaching quantum mechanics, through the use of labels (we don’t know how to proceed without them within such standard mathematical frameworks) is one of the signs of individuality.
Consider the quantum statistics mentioned above. To each particle $1$ and $2$ we associate a Hilbert space ${\mathcal}{H}_{1}$ and ${\mathcal}{H}_{2}$ respectively. If the particles are indiscernible, these spaces are in fact the same for both. For the system composed by both particles we associate the tensor product ${\mathcal}{H}_{1} \otimes {\mathcal}{H}_{2}$. As is usual, we write a typical vector of this space by $|\psi_1\rangle \otimes |\psi_2\rangle$, or simply $|\psi_1\rangle |\psi_2\rangle$ for short. Then, the possibilities available for the particles $1$ and $2$ in states $A$ and $B$ are stated as the following:
1. $|\psi_{1}^A\rangle |\psi_{2}^A\rangle$
2. $|\psi_{1}^B\rangle |\psi_{2}^B\rangle$
3. $\frac{1}{\sqrt{2}}(|\psi_{1}^A\rangle |\psi_{2}^B\rangle \pm |\psi_{2}^A\rangle |\psi_{1}^B\rangle)$.
We are here simplifying the notation: the third case splits in two, according to whether we deal with bosons or fermions; for bosons, we have cases 1-3 with a plus sign in the last one (symmetric vector), for fermions we only have the third possibility, now with the minus sign (anti-symmetric vector). What really matters is that to obtain the correct statistics one must employ symmetric and anti-symmetric vectors. This is the first part of the trick: label the particles and write the appropriate states, that is, symmetrical for bosons and anti-symmetrical for fermions. This means in particular that the asymmetrical states $|\psi_{1}^A\rangle |\psi_{2}^B\rangle$ and $|\psi_{2}^A\rangle |\psi_{1}^B\rangle$, like the ones employed in classical statistics, are not taken in consideration.[^2] Notice that the use of these states allows us to distinguish between the particles: we are able to attribute specific states to each one of them, although indiscernible (this is one way to substantiate the claim that quantum particles are indiscernible individuals; see French and Krause [@frekra06 chap.4] for the details).
But how can we make sure that only the states with the appropriate symmetry types are available? That is the second part of the trick, and it consists in postulating that only the appropriately symmetrized states obtain. The famous [Indistinguishability Postulate]{} (IP) does just this job:
$$\label{IP}
\langle \psi_{12} | \hat{O} | \psi_{12} \rangle = \langle P\psi_{21} | \hat{O} | P\psi_{21} \rangle.$$
What this principle says, roughly speaking, is that the result of a measurement before and after a permutation of the labels of particles results in the same quantity, so, permuting the labels amounts to no physical difference at all. In fact, a more technical reading of IP says that only observables compatible with the permutation operators $P$ (the ones whose purpose is to permute the labels given to the particles) are allowed. This amounts to the following relation:
$$[\hat{O}P,P\hat{O}] = \hat{O}P - P\hat{O} = 0$$
As one can check, only symmetrical and anti-symmetrical states satisfy these conditions. There is a whole discussion on what are the grounds to postulate IP and to accept it, and the different readings of this principle that may be compatible with it; in one of them, for example, which makes it weaker, the asymmetric states are not banned, but remain simply inaccessible to the particles (as we mentioned previously). In this last case, the particles may be regarded as individuals of some kind, but we shall not enter into these discussions here (but see French and Krause [@frekra06 chap.4] for further discussions). We just remark that this possibility shows that quantum particles [can]{} be considered as individuals too! (But, as we are trying to show, this is not the better way to see them).[^3]
So, this is in short how the trick is done in the standard formalism of quantum mechanics: to grant that permutations of particles do not give rise to different states, that is, to grant that the particles are absolutely indiscernible, one begins by labeling them in the first place, and then, by using only certain symmetrized vectors and by imposing a condition that only these vectors represent physical states, we achieve the desired result.
Each of these steps has its own difficulties for the philosopher interested in a commitment with ‘real’ indistinguishable entities, in particular for a defense of a metaphysics of nonindividuals. Really, to begin with, by assuming that there is some sense in labeling particles, we may consider them to have well defined identity conditions, for one of them was labeled $1$ and *the other* was labeled $2$. This idea runs counter to the whole enterprise being pursued, that is, that the particles must be [really]{} indiscernible. As they are generally understood, labels seem to demand a semantic theory committed with individuals and identity.
Here it would be interesting to comment that even the Fock space formalism (which will be deal with below) is not completely free from initially attached labels, as Redhead and Teller have claimed (see [@tel95]). In fact, in their usual construction, they begin with labeled Hilbert spaces (spaces ${\mathcal}{H}_{1}$ and ${\mathcal}{H}_{2}$, as we have been discussing); so, even if in the final presentation of a Fock space the labels are not mentioned, we need labels in Hilbert spaces for the usual constructions of the Fock spaces. The trick is not, after all, dispensed with! This criticism justifies our approach in the next section to construct Fock spaces inside a nonreflexive system of logic in which none of the mentioned tricks is required. We begin with indiscernible objects and build from there.
Note: — in fact, for $n$ indistinguishable particles, one takes the same Hilbert space $n$ times. But this does not eliminate the trick, for we still work with vectors of the form $| n_1, n_2, \ldots, n_k\rangle$, representing a state with $n_i$ quantum particles with a certain eigenvalue in a certain state (which would attest only to the quantity of them but not their individuality). In pursuing the details, however, we need to go back to sets and individuals. There is no scape within standard mathematics!
The [$\mathfrak{Q}$]{}-spaces {#qsp}
=============================
In order to avoid the (even hidden) assumption that we need to start with individuals, we introduce the notion of [$\mathfrak{Q}$]{}-spaces. We begin with an outline of quasi-set theory, the framework within which we develop our nonreflexive quantum mechanics.
Basic quasi-set theory [$\mathfrak{Q}$]{}
-----------------------------------------
Quasi-set theory is a ZFU-like first-order set theory comprising two kinds of atoms:[^4] the M-atoms, which behave like ZFU atoms and are denoted in the language with the help of a unary predicate $M$, and the m-atoms, which satisfy the unary predicate $m$. For the last kind of atoms identity is not defined, for according to our intuitive interpretation of the theory, we attribute to them the role of quantum particles: they must be not only indistinguishable, but more than that, identity statements must not make sense for them. To grant that both goals are achieved, that is, that there is one kind of objects that can be related by an indistinguishability relation and be such that identity and difference statements do not make sense for them, we must in [$\mathfrak{Q}$]{} depart from some of the original ZFU features. Let us check briefly how it can be done (for more details, see [@frekra06 chap.7], [@frekra10]).
Note: — before we begin with the theory, let us make another general remark. Why to advance the thesis that the notion of identity is senseless for the objects denoted by the m-atoms? The problem is that, if we assume that they obey the traditional theory of identity of classical logic (and set theory), once we have two of them, they are necessarily [different]{}. Well, in this case, they must have some difference, given by a property (since we are, as usual, avoiding to suppose notions of substratum)! Well, either the physical theory presents this property or not. In the first case, the entities could not be indiscernible; in the second case, the difference would be hidden in the underlying logic, but even so [it needs to be regarded as existing]{}. Then we have again two cases: either logic (or set theory) shows us the property or it does not. In both cases the entities turn to be individuals — we remark that even in standard mathematics there are entities (real numbers, say) that are different but whose difference cannot be described by the theory (an example would be the two least elements of two disjoint sets of real numbers) according to a given well-order, which can be assumed to exist in the presence of the axiom of choice, but which cannot be described by a formula of the theory of sets).
Let us begin with the underlying logic of [$\mathfrak{Q}$]{}: it is first-order logic *without* identity. From a syntactical point of view, the postulates of this logic are the same as those of classical logic, but semantically they are understood differently. To understand this restriction one needs only to remember that classical semantics is developed inside a classical set theory like ZFC. If this kind of semantics were employed to the underlying logic of [$\mathfrak{Q}$]{}, then we would be committing ourselves with precisely those features of classical logic we want to avoid, namely, that identity makes sense for every object and indistinguishability relations collapse in identity. Quantifier must be understood with some reservation: $\forall x \varphi(x)$ is to be not read as ‘for each object of the domain’, which would entail their identification, but simply as ‘for all objects of the domain’, meaning precisely this: for all objects of the domain. Similarly, $\exists x \varphi(x)$ means that [some]{} object of the domain satisfies the formula $\varphi(x)$, and we need not to identify it, say by a name; for details, see [@are14],[@arkra09].
Besides atoms, the theory is designed to grant that we can build collections. By definition, a quasi-set (q-set for short) is something that is not an atom, and we distinguish them for they obey a primitive unary predicate $Q$; that is, $Q(x)$ indicates that $x$ is a q-set. Such collections are assumed to be formed in stages, as it happens in the classical cumulative hierarchy; for some of these collections, in some step m-atoms may be involved as member, but for others it may happen that no m-atom appear in any step of the formation of the q-set. To designate this last kind of q-set we introduce the primitive predicate $Z$, denoting what we call a [classical q-set]{}, or simply a [set]{}, and they turn to be copies of the sets of ZFU. Technically speaking, sets are collections with no m-atoms in their transitive closure. We call the M-atoms and sets [classical things]{} of the theory. When restricted to the classical things, [$\mathfrak{Q}$]{} allows us to develop all the standard mathematics that can be developed inside ZFU.
Identity must be introduced by definition. This fact allows us to manufacture an identity relation restricted to q-sets and M-atoms only, but not holding among m-atoms (to fit the intuitive interpretation we have attributed them). The definition is the following:
$x = y {:=}(Q(x) \wedge Q(y) \wedge \forall z (z \in x \leftrightarrow z \in y)) \vee (M(x) \wedge M(y) \wedge \forall z (x \in z \leftrightarrow y \in z))$.
According to the definition, equal q-sets are those that have the same members, and equal M-atoms are those that belong to the same q-sets. If there are m-atoms involved, since the notion of identity does not make sense to them, it may be the case that we are unable to effectively know whether two q-sets are identical, for we have no means to identify their elements. But things here work in the conditional: [if]{} two q-sets have the same elements, [then]{} they are identical and reciprocally.[^5]
One can easily prove from this definition that identity is reflexive, and the substitution law is postulated as an axiom; then, identity (where it holds) has both properties that characterize it in first-order systems. What about indistinguishability? We postulate that it has the properties of an equivalence relation ([i.e.]{} it is reflexive, symmetric and transitive), but it is not compatible with the membership relation, that is, from $x\equiv y$ and $x \in z$ we cannot always derive that $y \in z$; this is done in order to grant that identity and indistinguishability do not collapse in the same concept. For the other properties $P$ of the language, though, we have that if $x \equiv y$ and $P(x)$, then $P(y)$. Also, for the classical things, identity and indistinguishability are equivalent.
The construction of q-sets proceeds in much the same way as in ZFU: by the iterated application of well-known set theoretical operations as power set, union, cartesian product, among others. One difference that is worth mentioning concerns the notion of unordered pair; in [$\mathfrak{Q}$]{} we cannot proceed as in the usual set theories, postulating that for every $x$ and $y$ there is a collection of objects equal to either $x$ or $y$, and the reason is simple: for m-atoms identity is not defined, so that to restrict ourselves to the classical definition would amount to the impossibility of pairs of m-atoms, something undesirable in a theory of collections. To mend the situation, the theory encompasses an unordered pair axiom stating that for every two items $x$ and $y$ there is a q-set $z$ containing them as elements (as in the classical case, there may be other elements in $z$ as well). We then apply the separation axiom (which has its quasi-set theoretical version similar to the classical formulation) to a q-set $z$ and separate a ‘pair’, in fact, which is the q-set of the elements of $z$ containing only those elements that are indistinguishable from either $x$ or $y$. This collection is denoted $[x, y]_{z}$; if $x \equiv y$ holds, then we write simply $[x]_{z}$ and call it the weak singleton of $x$. Notice that, intuitively speaking, the quantity of elements in weak pairs and weak singletons may be higher than 2 and 1 respectively, for $z$ may contain many elements indistinguishable from $x$ and/or $y$.
From unordered pairs we can then define a weak version of ordered pairs as follows, which stands for the q-set of the items indistinguishable from $x$ and from $y$ that belong to $z$:
\[wp\] $\langle x, y \rangle_z {:=}[[x]_z, [x,y]_z]_z$
Of course, any binary relation (or q-relation as we shall see) involving only m-atoms is symmetric for $\langle x, y \rangle_{z} = \langle y, x \rangle_{z}$. Then, the definition of cartesian product follows as usual:
$z \times w {:=}[\langle x, y \rangle_{z \cup w} : x \in z \wedge y \in w]$
If there are only m-atoms involved, then $z \times w = w \times z$ for all $z$ and $w$, as is easy to see. The theory has also a union axiom which works as in standard set theories; so, we shall use the symbols $\cup$ and $\cap$ taken then from granted from now on.
A q-set $R$ is a binary quasi-relation between q-sets $z$ and $w$ if its elements are weak ordered pairs of the form $\langle x , y \rangle_{z \cup w}$, with $x \in z$ and $y \in w$.
Now, for quasi-functions we cannot proceed as usual, for the classical concept of function uses the notion of identity. To generalize the usual concept, then, and allow that it applies also for q-sets of m-atoms, we require that a mapping does not attribute different [kinds]{} of values to indistinguishable arguments. Thus, we have:
\[qf\] $f$ is a quasi-function among q-sets $A$ and $B$ if and only if $f$ is quasi-relation between $A$ and $B$ such that for every $u \in A$ there is a $v \in B$ such that if $\langle u, v\rangle \in f$ and $\langle w,z\rangle \in f$ and $u \equiv w$ then $v \equiv z$.
What this definition grants us is that a quasi-function maps indistinguishable elements to indistinguishable elements. For classical objects, both quasi-relations and quasi-functions coincide with the classical definitions. With some restrictions, one can also define the concepts of injection, surjection and bijection, but we shall not present those definitions here (see [@frekra06 chap. 7]).
Another important concept in [$\mathfrak{Q}$]{} concerns the attribution of cardinality to q-sets. Since collections having m-atoms as elements cannot be well-ordered (for that concept also presupposes identity), one cannot proceed in the usual ways to grant that every q-set will have a cardinal number associated to it, for example, through the attribution of ordinals following the well-known von Neumann definition. The usual strategy has been to adopt a primitive concept of quasi-cardinal which generalizes the classical notion of cardinal: in the classical part of quasi-set theory one builds cardinals, as usual, and then, with the primitive quasi-function $qc$ one attributes cardinals to every q-set, making sure that classical sets will have as their quasi-cardinal (the cardinal attributed by $qc$) precisely the same cardinal attributed to them in the classical part of the theory.
The notion of quasi-cardinal also allows us to obtain in [$\mathfrak{Q}$]{} another interesting concept: the [strong singleton]{}. As we have seen, given any object $x$ whatever we may obtain the weak singleton of $x$, a q-set which in the case of $x$ being an m-atom may have quasi-cardinal greater than 1. The strong singleton of $x$ relative to a q-set $w$, denoted by $\llbracket x \rrbracket_w$ is simply a q-set containing an element that is indistinguishable from $x$ which belong to $w$ and whose quasi-cardinal is precisely 1. With the resources of [$\mathfrak{Q}$]{} however, if $x$ is an m-atom one cannot deduce that $x$ is precisely the only element of $\llbracket x \rrbracket_w$, for identity must be used for that. So, we can reason within [$\mathfrak{Q}$]{} that we may have in $\llbracket x \rrbracket_w$ [an m-atom]{} having determinate properties, without being able to specify which particular one it is.
The following theorem expresses the invariance by permutations in [$\mathfrak{Q}$]{}, a result we shall employ below[^6]:
Let $x$ be a finite q-set such that $\neg(x = [z]_t)$ for some $t$ and let $z$ be an m-atom such that $z \in x$. If $w \in t$, $w \equiv z$ and $w \notin x$, then there exists $\llbracket w \rrbracket _t$ such that $$(x - \llbracket z \rrbracket_t) \cup \llbracket w \rrbracket_t \equiv x$$
\[unobser\]
In words: two indiscernible elements $z$ and $w$, with $z \in x$ and $w \notin x$, expressed by their strong-singletons $\llbracket t \rrbracket_t$ and $\llbracket w \rrbracket_t$, are ‘permuted’ and the resulting q-set remains indiscernible from the original one. The hypothesis that $\neg(x = [z]_t)$ grants that there are in $t$ indiscernible elements from $z$ which do not belong to $x$ (for details and the proof of the Unobservability of Permutations theorem, see [@frekra06 pp.295,296]).
These are the central notions that will be important in our development of [$\mathfrak{Q}$]{}-spaces, to which we now turn below.
Quasi-functions
---------------
From now on, we shall follow [@domholkra08], where further details can be seen. We begin with a q-set of real numbers $\epsilon = \{\epsilon_{i}\}_{i \in I }$, where $I$ is an arbitrary collection of indexes, denumerable or not. Since it is a collection of real numbers, which may be constructed in the classical part of [$\mathfrak{Q}$]{} we have that $Z(\epsilon)$. Intuitively, the elements $\epsilon_{i}$ represent the eigenvalues of a physical observable $\hat{O}$, that is, they are the values such that $\hat{O}|\varphi_{i}\rangle=\epsilon_{i}|\varphi_{i}\rangle$, with $|\varphi_{i}\rangle$ the corresponding eigenstates. The fact that the observables are Hermitian operators grants us that the eigenvalues are real numbers, thus we are justified in assuming $\epsilon$ to be a set of real numbers. Consider then the quasi-functions $f:\epsilon \longrightarrow \mathcal{F}_{p}$, where $\mathcal{F}_{p}$ is the quasi-set formed of all finite and pure quasi-sets (that is, finite quasi-sets whose only elements are indistinguishable m-atoms). Each of these $f$ is a q-set of ordered pairs $\langle \epsilon_{i}, x\rangle$ with $\epsilon_{i}\in\epsilon$ and $x\in\mathcal{F}_{p}$. From $\mathcal{F}_{p}$ we select those quasi-functions $f$ which attribute a non-empty q-set only to a finite number of elements of $\epsilon$, the image of $f$ being $\emptyset$ for the other cases. We call $\mathcal{F}$ the quasi-set containing only these quasi-functions. Then, the quasi-cardinal of most of the q-sets attributed to elements of $\epsilon$ according to these quasi-functions is $0$. Now, elements of ${\mathcal}{F}$ are quasi-functions which we read as attributing to each $\epsilon_{i}$ a q-set whose quasi-cardinal we take to be the occupation number of this eigenvalue. We write these quasi-functions as $f_{\epsilon_{i_{1}}\epsilon_{i_{2}}\ldots\epsilon_{i_{m}}}$. According to the given intuitive interpretation, the levels $\epsilon_{i_{1}}\epsilon_{i_{2}}\ldots\epsilon_{i_{m}}$ are occupied. We say that if the symbol $\epsilon_{i_{k}}$ appears $j$-times, then the level $\epsilon_{i_{k}}$ has occupation number $j$. For example, the notation $f_{\epsilon_{1}\epsilon_{1}\epsilon_{1}\epsilon_{2}\epsilon_{3}}$ means that the level $\epsilon_{1}$ has occupation number $3$ while the levels $\epsilon_{2}$ and $\epsilon_{3}$ have occupation numbers $1$. The levels that do not appear have occupation number zero. Another point to be remarked is that since the elements of $\epsilon$ are real numbers, we can take the standard ordering relation over the reals and order the indexes according to this ordering in the representation $f_{\epsilon_{i_{1}}\epsilon_{i_{2}}\ldots\epsilon_{i_{m}}}$. \[or\] This will be important when we consider the cases for bosons and fermions.
The quasi-functions of ${\mathcal}{F}$ provide the key to the solution to the problem of labeling states. In fact, since we use pure quasi-sets as the images of the quasi-functions, there is simply no question of indexes for particles, for all that matters are the quasi-cardinals representing the occupation numbers. To make it clear that permutations change nothing, one needs only to notice that a quasi-function is a q-set of weak ordered pairs (see definitions \[wp\] and \[qf\]). Taking two of the pairs belonging to some quasi-function, let us say $\langle\epsilon_{i}, x \rangle$, $\langle \epsilon_{j}, y\rangle$, with both $x$ and $y$ non-empty, a permutation of particles would consist in changing elements from $x$ with elements from $y$. But, by the unobservability of permutations theorem \[unobser\], what we obtain after the permutation is a q-set indistinguishable from the one we began with. Remember also that a quasi-function attributes indistinguishable images to indistinguishable items; thus, the indistinguishable q-set resulting from the permutations will also be in the image of the same eigenvalue. To show this point precisely, we recall that by definition $\langle\epsilon_{i}, x\rangle$ abbreviates $[[\epsilon_{i}],[\epsilon_{i},x]]$,[^7] and an analogous expression holds for $\langle\epsilon_{j}, y\rangle$. Also by definition, $[\epsilon_{i},x]$ is the collection of all the items indistinguishable from $\epsilon_{i}$ or from $x$ (taken from a previously given q-set). For this reason, if we permute $x$ with $x'$, with $x\equiv x'$ we change nothing for $[\epsilon_{i},x] \equiv [\epsilon_{i},x']$. Thus, we obtain $\langle\epsilon_{i},x\rangle \equiv \langle\epsilon_{i},x'\rangle$ and the ordered pairs of the ‘permuted’ quasi-function will be indiscernible (the same if there are no m-atoms involved). Thus, the permutation of indistinguishable elements does not produce changes in the quasi-functions.
A Vector Space Structure
------------------------
Now, we wish to have a vector space structure to represent quantum states. To do that, we need to define addition and multiplication by scalars. Before we go on, we must notice that we cannot define these operations directly on the q-set ${\mathcal}{F}$, for there is no simple way to endow it with the required structure; our strategy here is to define $\star$ (multiplication by scalars) and $+$ (addition of vectors) in a q-set whose vectors will be quasi-functions from ${\mathcal}{F}$ to the set of complex numbers $\mathbb{C}$. Let us call $C$ the collection of quasi-functions which assign to every $f\in \mathcal{F}$ a complex number. Once again, we select from $C$ the sub-collection $C_{F}$ of quasi-functions $c$ such that every $c\in C_{F}$ attributes complex numbers $\lambda \neq 0$ for only a finite number of $f\in \mathcal{F}$. Over $C_{F}$ we can define a sum and a product by scalars in the same way as it is usually done with functions as follows.
Let $\gamma$ $\in {\mathcal}{C}$, and $c$, $c_{1}$ and $c_{2}$ be quasi-functions of $C_{F}$, then $$(\gamma\star c)(f) {:=}\gamma(c(f))$$ $$(c_{1}+ c_{2})(f) {:=}c_{1}(f) + c_{2}(f)$$
The quasi-function $c_{0}\in C_{F}$ such that $c_{0}(f)=
0$ for every $f\in {\mathcal}{F}$ acts as the null element for the sum operation. This can be shown as follows:
$$(c_{0}+c)(f)= c_{0}(f)+c(f)= 0+c(f)= c(f), \forall f.$$
With both the operations of sum and multiplication by scalars defined as above we have that $\langle C_{F},\mathbb{C}, +,\star \rangle$ has the structure of a complex vector space, as one can easily check. Some of the elements of $C_F$ have a special status though; if $c_{j}\in C_{F}$ are the quasi-functions such that $c_{j}(f_{i})= \delta_{ij}$ (where $\delta_{ij}$ is the Kronecker symbol), then the vectors $c_{j}$ are called the basis vectors, while the others are linear combinations of them. For notational convenience, we can introduce a new notation for the q-functions in $C_{F}$; suppose $c$ attributes a $\lambda \not= 0$ to some $f$, and $0$ to every other quasi-function in ${\mathcal}{F}$. Then, we propose to denote $c$ by $\lambda f$. The basis quasi-functions will be denoted simply $f_{i}$, as one can check. Now, multiplication by scalar $\alpha$ of one of these quasi-functions, say $\lambda f_{i}$ can be read simply as $(\alpha \cdot \lambda) f_{i}$, and sum of quasi-functions $\lambda f_{i}$ and $\alpha f_{i}$ can be read as $(\alpha + \lambda) f_{i}$. What about the other quasi-functions in $C_{F}$? We can extend this idea to them too, but with some care: if, for example $c_{0}$ is a quasi-function such that $c_{0}(f_{i})= \alpha$ and $c_{0}(f_{j})= \lambda$, attributing $0$ to every other quasi-function in ${\mathcal}{F}$, then $c_{0}$ can be seen as a linear combination of quasi-functions of a basis; in fact, consider the basis quasi-functions $f_{i}$ and $f_{j}$, (this is an abuse of notation, for they are representing quasi-functions in $C_F$ that attribute 1 to each of these quasi-functions). The first step consists in multiplying them by $\alpha$ and $\lambda$, respectively, obtaining $\alpha f_{i}$ and $\lambda f_{j}$ (once again, this is an abuse, for these are quasi-functions in $C_F$ that attribute the mentioned complex numbers to $f_{i}$ and to $f_{j}$). Now, $c_{0}$ is in fact the sum of these quasi-functions, that is, $c_{0} = \alpha f_{i} + \lambda f_{j}$, for this is the function which does exactly what $c_{0}$ does. One can then extend this to all the other quasi-functions in $C_F$ as well.
Inner Products
--------------
The next step in our construction is to endow our vector space with an inner product. This is a necessary step for we wish to calculate probabilities and mean values. Following the idea proposed in [@domholkra08], we introduce two kinds of inner products, which lead us to two Hilbert spaces, one for bosons and another for fermions. We begin with the case for bosons:
Let $\delta_{ij}$ be the Kronecker symbol and $f_{\epsilon_{i_{1}}\epsilon_{i_{2}}\ldots\epsilon_{i_{n}}}$ and $f_{\epsilon_{i'_{1}}\epsilon_{i'_{2}}\ldots\epsilon_{i'_{m}}}$ two basis vectors (as discussed above), then $$f_{\epsilon_{i_{1}}\epsilon_{i_{2}}\ldots\epsilon_{i_{n}}}\circ
f_{\epsilon_{i'_{1}}\epsilon_{i'_{2}}\ldots\epsilon_{i'_{m}}} {:=}\delta_{nm}\sum_{p}\delta_{i_{1}pi'_{1}}\delta_{i_{2}pi'_{2}}\ldots\delta_{i_{n}pi'_{n}}.$$
Notice that this sum is extended over all the permutations of the index set $i'=(i'_{1},i'_{2},\ldots,i'_{n})$; for each permutation $p$, $pi'=(pi'_{1},pi'_{2},\ldots,pi'_{n})$.
For the other vectors, the ones that can be seen as linear combinations in the sense discussed above, we have: $$(\sum_{k}\alpha_{k}f_{k})\circ(\sum_{k}\alpha'_{k}f'_{k}) {:=}\sum_{kj}\alpha_{k}^{\ast}\alpha'_{j}(f_{k}\circ f'_{j}),$$
where $\alpha^{\ast}$ is the complex conjugate of $\alpha$. Now, let us consider fermions. As remarked above in page , the order of the indexes in each $f_{\epsilon_{i_{1}}\epsilon_{i_{2}}\ldots\epsilon_{i_{n}}}$ is determined by the canonical ordering in the real numbers. Thus, we define another “$\bullet$” inner product as follows, which will do the job for fermions:
Let $\delta_{ij}$ be the Kronecker symbol and $f_{\epsilon_{i_{1}}\epsilon_{i_{2}}\ldots\epsilon_{i_{n}}}$ and $f_{\epsilon_{i'_{1}}\epsilon_{i'_{2}}\ldots\epsilon_{i'_{m}}}$ two basis vectors, then $$f_{\epsilon_{i_{1}}\epsilon_{i_{2}}\ldots\epsilon_{i_{n}}}\bullet
f_{\epsilon_{i'_{1}}\epsilon_{i'_{2}}\ldots\epsilon_{i'_{m}}} {:=}\delta_{nm}\sum_{p}\sigma_{p}\delta_{i_{1}pi'_{1}}\delta_{i_{2}pi'_{2}}\ldots\delta_{i_{n}pi'_{n}}$$ where: $\sigma_{p} = 1$ if p is even and $\sigma_{p} = -1$ if p is odd.
This definition can be extended to linear combinations as in the previous case.
Fock spaces using [$\mathfrak{Q}$]{}-spaces
-------------------------------------------
We begin with a definition to simplify the notation. For every function $f_{\epsilon_{i_{1}}\epsilon_{i_{2}}\ldots\epsilon_{i_{n}}}$ in ${\mathcal}{F}$, we put:
$$\alpha |\epsilon_{i_{1}}\epsilon_{i_{2}}\ldots\epsilon_{i_{n}}) {:=}\alpha f_{\epsilon_{i_{1}}\epsilon_{i_{2}}\ldots\epsilon_{i_{n}}}$$
Note that this is a slight modified version of the standard notation. We begin with the case of bosons.
Suppose a normalized vector $|\alpha \beta \gamma \ldots )$, where the norm is taken from the corresponding inner product. Let $\zeta$ stand for an arbitrary collection of indexes. We define $a_{\alpha}^{\dagger}|\zeta ) \propto |\alpha \zeta )$ in such a way that the proportionality constant satisfies $a_{\alpha}^{\dagger} a_{\alpha} |\zeta ) = n_{\alpha} |\zeta )$. From this it will follow, as usual, that: $$((\zeta|a_{\alpha}^{\dagger})(a_{\alpha}|\zeta)) = n_{\alpha}.$$
$a_{\alpha} | \ldots n_{\alpha} \ldots ) := \sqrt{n_{\alpha}} | \ldots n_{\alpha} - 1 \ldots )$
On the other hand $$a_{\alpha}a_{\alpha}^{\dagger} | \ldots n_{\alpha} \ldots ) = K \sqrt{n_{\alpha} + 1}|\ldots n_{\alpha} \ldots ),$$ where $K$ is a proportionality constant. Applying $a_{\alpha}^{\dagger}$ again, we have $$a_{\alpha}^{\dagger}a_{\alpha}a_{\alpha}^{\dagger} | \ldots n_{\alpha} \ldots ) = K^2 \sqrt{n_{\alpha} + 1}|\ldots n_{\alpha} + 1 \ldots ).$$
Using the fact that $a_{\alpha}^{\dagger} a_{\alpha} |\zeta ) = n_{\alpha} |\zeta )$, we have that: $$(a_{\alpha}^{\dagger}a_{\alpha})a_{\alpha}^{\dagger} | \ldots n_{\alpha} \ldots ) = \sqrt{n_{\alpha} + 1}K|\ldots n_{\alpha} + 1 \ldots ).$$ So, $K = \sqrt{n_{\alpha} + 1}$.
Then, we have:
$a_{\alpha}^{\dagger} | \ldots n_{\alpha} \ldots ) := \sqrt{n_{\alpha} + 1}| \ldots n_{\alpha} + 1 \ldots ).$
From this definition, with additional computations, we obtain: $(a_{\alpha}a_{\beta}^{\dagger} - a_{\beta}^{\dagger}a_{\alpha})|\psi) = \delta_{\alpha \beta}|\psi)$. In our language, this means the same as $$[a_{\alpha} ; a_{\beta}^{\dagger}] = \delta_{\alpha \beta} I.$$
In an analogous way, it can be shown that:$$[a_{\alpha}; a_{\beta}] = [a_{\alpha}^{\dagger};a_{\beta}^{\dagger}] = 0.$$
So, the bosonic commutation relation are the same as in standard Fock space formalism.
For fermionic states we use the antisymmetric product $\bullet$. We begin by defining the creation operator $C_{\alpha}^{\dagger}$:
If $\zeta$ is a collection of indexes of non-null occupation numbers, then $C_{\alpha}^{\dagger} := \alpha |\zeta)$
If $\alpha$ is in $\zeta$, then $|\alpha \zeta)$ is a vector of null norm. This implies that $(\psi|\alpha \zeta) = 0$, for every $\psi$. It follows that systems in states of null norm have no probability of being observed. Furthermore, their addition to another vector does not contribute with any observable difference. To take the situation into account, we define:
Two vectors $|\phi)$ and $|\psi)$ are similar if the difference between them is a linear combination of null norm vectors. We denote similarity of $|\phi)$ and $|\psi)$ by $|\phi) \cong |\psi)$.
Using the definition of $C_{\alpha}^{\dagger}$ we can describe what is the effect of $C_{\alpha}$ over vectors: $$(\zeta| C_{\alpha} := (\alpha \zeta|.$$
Then, for any vector $|\psi)$, $$(\zeta| C_{\alpha}|\psi) = (\alpha \zeta|\psi) = 0$$ for $\alpha \in \zeta$ or $(\psi|\alpha \zeta) = 0$. Then, if $|\psi) = |0)$, then $(\zeta| C_{\alpha}|0) = (\alpha \zeta|0) = 0$. So, $C_{\alpha}|0)$ is orthogonal to any vector that contais $\alpha$, and also to any vector that does not contain $\alpha$, so that it is a linear combination of null norm vectors. So, we can put by definition that $\vec{0} := C_{\alpha}|0)$. In an analogous way, if $ \sim \alpha$ denotes that $\alpha$ has occupation number zero, then we can also write $C_{\alpha}|(\sim \alpha) \ldots) = \vec{0}$, where the dots mean that other levels have arbitrary occupation numbers.
Now, using our notion of similar vectors, we can write $C_{\alpha}|0) \cong \vec{0}$ and $C_{\alpha}|(\sim \alpha) \ldots) \cong \vec{0}$. The same results are obtained when we use $\cong$ and the sign of identity. By making $|\psi) = |\alpha)$, we have $(\zeta| C_{\alpha}|\alpha) = (\alpha \zeta|\alpha) = 0$ in every case, except when $|\zeta) = |0)$. In that case, $(0| C_{\alpha}|\alpha) = 1$. Then, it follows that $C_{\alpha}|\alpha) \cong 0$. In an analogous way, we obtain $C_{\alpha}|\alpha \zeta) = \cong |(\sim \alpha) \zeta)$ when $\alpha \notin \zeta$. In the case $\alpha \in \zeta$, $|\alpha \zeta)$ has null norm, and so, for every $|\psi)$: $$(\alpha \zeta| C_{\alpha}^{\dagger}|\psi) = (\alpha \zeta|\alpha \psi) = 0.$$
It then follows that $$(\psi|C_{\alpha}|\alpha \zeta) = 0,$$ so that $C_{\alpha}|\alpha \zeta)$ has null norm too.
Now we calculate the anti-commutation relation obeyed by the fermionic creation and annihilation operators. We begin calculating the commutation relation between $C_{\alpha}$ and $C_{\beta}^{\dagger}$. We do that by studying the relationship between $|\alpha \beta)$ and $|\beta \alpha)$. Let us consider the sum $|\alpha \beta) + |\beta \alpha)$. The product of this sum with any vector distinct from $|\alpha \beta)$ is null. For the product with $|\alpha \beta)$ we obtain $(\alpha \beta|[|\alpha \beta) + |\beta \alpha)] = (\alpha \beta || \alpha \beta) + (\alpha \beta || \beta \alpha)$. By definition, this is equal to $\delta_{\alpha \alpha} \delta_{\beta \beta} - \delta_{\alpha \beta}\delta_{\beta\alpha} + \delta_{\alpha \beta}\delta_{\alpha\alpha} - \delta_{\alpha \alpha}\delta_{\beta\beta}$. This is equal to $1 - 0 + 0 - 1 = 0$.
The same conclusion holds if we multiply the sum $|\alpha \beta) + |\beta \alpha)$ by $(\beta \alpha|$. It then follows that $|\alpha \beta) + |\beta \alpha)$ is a linear combination of null norm vectors, which we denote by $|nn)$, so that $$|\alpha \beta) = - |\beta \alpha) + |nn).$$
Given that, we can calculate $$C_{\alpha}^{\dagger}C_{\beta}^{\dagger}|\psi) = |\alpha \beta\psi) = -|\beta \alpha |\psi) + |nn) = - C_{\beta}^{\dagger}C_{\alpha}^{\dagger}|\psi) + |nn).$$
From this it follows that $\{C_{\alpha}^{\dagger};C_{\beta}^{\dagger}\}|\psi) = |nn)$. We do not lose generality by setting $\{C_{\alpha}^{\dagger};C_{\beta}^{\dagger}\}|\psi) = 0 $. In an analogous way we conclude: $$\{C_{\alpha};C_{\beta}\}|\psi) = 0.$$
Now we calculate the commutation relation between $C_{\alpha}$ and $C_{\beta}^{\dagger}$. There are some cases to be considered. We first assume that $\alpha \neq \beta$. If $\alpha \notin \psi$ or $\beta \in \psi$ then $$\{C_{\alpha};C_{\beta}^{\dagger}\}|\psi) \approx \vec{0}.$$
If $\alpha \in \psi$ and $\beta \notin \psi$, assuming that $\alpha$ is the first symbol in the list of $\psi$, then $\{C_{\alpha};C_{\beta}^{\dagger}\}|\psi) = C_{\alpha}|\beta \psi) + C_{\beta}^{\dagger}|\psi(\sim \alpha))\cong -|\beta \psi (\sim \alpha)) + |\beta \psi (\sim \alpha)) = \vec{0}$.
Now, if $\alpha = \beta$ and $\alpha \in \psi$, then $\{C_{\alpha};C_{\alpha}^{\dagger}\}|\psi) = C_{\alpha}|\alpha \psi) + C_{\alpha}^{\dagger}|\psi(\sim \alpha))\cong \vec{0} + |\psi) = |\psi)$.
If $\alpha = \beta$ and $\alpha \notin \psi$, then $\{C_{\alpha};C_{\alpha}^{\dagger}\}|\psi) = C_{\alpha}|\alpha \psi) + C_{\alpha}^{\dagger}|\psi(\sim \alpha))\cong |\psi) + \vec{0} = |\psi)$.
In any case, we recover $\{C_{\alpha};C_{\alpha}^{\dagger}\}|\psi) \cong \delta_{\alpha \beta} |\psi)$. So, we can put $$\{C_{\alpha};C_{\alpha}^{\dagger}\} = \delta_{\alpha \beta}.$$
It then follows that the commutation properties in [$\mathfrak{Q}$]{}-spaces are the same as in traditional Fock spaces.
Using this formalism, we can adapt all the developments done in [@mat67 Chap.7] and [@mer70 Chap.20] for the number occupation formalism. But here, contrary to what happens in these books, no previous (even unconscious) assumptions about the individuality of quantum objects is taken into account.
Foundational and metaphysical remarks {#dis}
=====================================
Now that a non-standard version of quantum mechanics was sketched inside quasi-set theory, what are the main philosophical lessons that one can draw from this whole enterprise? Are we supposed to substitute traditional quantum mechanics by its nonreflexive version? The answer to the last question is an obvious NO, or not necessarily. What is illustrated by this construction is only that one can take a version of non-individuality seriously and, building from there, to develop the formalism of quantum mechanics anew, without labeling and individuality. We remark once again that in using standard mathematics (standard set theory), we need to assume that the represented objects are individuals at the start, and need to make some mathematical tricks in order to keep them with something that mimic non-individuality. Once indiscernible non-individuals are essential to quantum mechanics, our move illustrates that one can assume non-individuality and indistinguishability as central in a formalism and, along with some set theoretical tools compatible with such non-individuality ([i.e.]{} quasi-set theory), derive a version of the formalism of the theory.
This kind of move illustrates how one may profitably approach the cooperative work between the study of foundations of science and metaphysics. Metaphysics may be seen as contributing positively to the foundations of a scientific theory like quantum mechanics, by suggesting that due to the main features of the theory, we are dealing with a new category of entities, the non-individuals. Metaphysicians are entitled to describe the main metaphysical features that entities qualified as non-individuals could have. By searching for the logic that adequately describes those entities we are led to the development of [$\mathfrak{Q}$]{}-spaces. Quantum mechanics, on its own side, contributes to the development of metaphysics. By being our most successful theory, and by putting so much weight on indiscernible entities, it forces metaphysicians to develop an adequate notion of non-individuals and indiscernible entities compatible with the description provided by quantum mechanics.
What we aim now is to present how we understand this kind of collaboration between science and metaphysics in a more general picture. The following may be seen as a rudimentary approach to how a productive relation between science and metaphysics may be developed which does not focus only on science (such as Ladyman and Ross do in [@lr07]) and also does not rely so heavily on what metaphysicians have already achieved (as it seems to be the case of French’s approach in [@fre14 chap.3]). We intend to present how a really collaborative effort to understand quantum metaphysics, for instance, may proceed. This is not achieved by adopting neither a ‘physics first’ approach, and nor by providing a ‘metaphysics first’ approach. Both physics and metaphysics must be cooperatively involved in the enterprise. The ‘dialectics’ of such a cooperation may be roughly described as a case of a process aiming at a state of Reflective Equilibrium. Let us see briefly how.
We begin, as usual, with an actually available formulation of our scientific theory; in this case, quantum mechanics. In general there is an informal interpretation of the theory. Anyway, the mathematics of the theory may give us some clues as to how the entities being described behave. In general, we simply talk about the entities described by the theory by transferring our everyday categories of stable objects, properties, relations and things in general to the domain being described by the scientific theory. What our most successful theories in the beginning of the twentieth century have shown us is that this transference is not always justified. General relativity, for instance, demanded for a revision in our concepts of space and time, while quantum mechanics demanded a revision of many other concepts as well, such as stable objects possessing properties and having well defined identity conditions (it is enough to recall some no-go theorems like Kochen-Specker’s and Bell’s). That such is the case is, as is well-known, the main contention of Ladyman and Ross in [@lr07 Chap.1]: metaphysicians have not yet taken seriously the lessons from modern physics.
However, after perceiving that a scientific theory demands a revision in some fundamental concepts, what are metaphysicians supposed to do? In the case of quantum mechanics, in the specific case of non-individuals, how can we rigorously describe the situation? Well, we must first recognize that the mathematical apparatus employed to describe the theory encapsulates some of the items requiring revision, like labeling and identity. So, how should we proceed? Just as suggested by French [@fre14] with his [viking approach to metaphysics]{}, we may check what the metaphysicians can say about the new kind of entities. Have they already articulated a theory about the behavior of those entities? If they have, do those entities, thus categorized, mesh well with what the scientific theory says about the world? In the case of quantum entities, it seems, as we have argued in sections \[sta\] and \[ind\], the metaphysical ideas of non-individuals, of indiscernible classes of entities, do not fit very well with some of the assumptions that are encoded in the standard formalism of the theory. So, in this case, the search for an alternative logical foundations is a first required step.
Our suggestion is that we could then use the information brought by the scientific theory to develop a new logic, a logic that already encodes the basic information about the entities dealt with. Nonreflexive logics do just that, as we have briefly sketched in the beginning of section \[qsp\]. So, we use the actual formulation of a theory in order to discover what are the main features of the entities being dealt with. In case those features demand a revision of some of our most general categories, then we have at least two options: either (i) we reformulate the theory in order to keep the traditional categories as much as possible (as it happens for instance, in Bohmian mechanics) while still retaining empirical adequacy, or else (ii) we try to take those novelties seriously, at face value as it were, and look for the best place to implement them. Given that the idea of non-individuals, as we have been dealing with them, is a general category, related with identity, it seemed that the underlying logic was the most appropriate place to introduce that category. Also, that change of view proved to be very fruitful: we could not only account for the non-individuality of quantum particles, but also show how that non-individuality is able to generate a formalism of quantum mechanics anew! Another way to account for the novelties, still concerning this second case, is not to change the formalism or the logic, but to keep the mathematics of the theory intact and provide for an interpretation of the formalism that somehow accounts for the novelties of the theory. One of such options is now known as [wave function realism]{}. The main idea behind this interpretation seems to be that the wave function of the mathematical formalism is the reality of which the theory is about. In this case, the very concept of ordinary space is to be completely redefined, along with the concept of three dimensional objects. Even though this is an interesting approach, we shall not discuss it in what follows, keeping our focus on the notion of non-individuality (for more on wave function realism, see the discussion in the collection by Alyssa Ney and David Albert [@aly13], [@kraare14]).
So, by reformulating the formalism of Fock spaces in terms of [$\mathfrak{Q}$]{}-spaces, we are doing much more than merely paying lip service the a revisionary form of metaphysics: we show, somehow, that the concept of a non-individual is fundamental in at least one understanding of quantum mechanics. This is different from, for instance, the claim that some kind of paraconsistent logic may be useful for our understanding of quantum superpositions (as advanced in [@cos13]; see also the discussion in [@are15]). In this case, one interprets a physical fact in logico-metaphysical terms. Our claim is that the nonreflexive foundations of quantum mechanics developed here go one step further: we do not merely take the claims of quantum theory and re-interpret them in a non-classical logic system, but we begin with the non-classical system and inside it we develop the scientific theory itself. So, by reading some of the main features of the entities dealt with by the theory, we are able to use now those very features as the cornerstones for a new formulation of the theory, one that takes those features as central. This move closes a kind of circle: from a standard formulation of the theory we obtain some metaphysical feature of the entities. By laboring over those features, we develop a metaphysical approach to them that is rigorous enough to be encapsulated by a system of logic. Finally, by employing that system of logic we are able to develop a formalism that is able to express the theory.
Perhaps we could describe the kind of procedure developed here with the help of an analogy with the situation in foundations of mathematics, as presented by Kunen [@kun09 p.191]. In discussing the foundations of mathematics, Kunen claims that logic must be developed [twice]{}: we begin with our intuitions and intuitive rules of inference and, based solely on them, develop an informal set theory. Obviously, some logic is required in order to develop a set theory, and in the case of standard set theories this logic is informal classical logic. Now, inside the set theory we have just developed, we can begin anew and develop, in a completely rigorous way, *using set theoretical tools*, our very system of logic. Now, inside such a set theory can we make sure that a reasonable model theory, for instance, is available. So, logic must be built twice: first, as an informal system, capable of grounding the development of a set theory and such that it does not presuppose set theoretical concepts in its formulation. Then, inside the just developed set theory, we can develop, as a set theoretical entity, logic itself. In the second case, the tools of set theory are clearly available, and rigorous research is possible (for a discussion of that move concerning nonreflexive logic, see [@arkra09] and [@are14]).
A similar situation happens in the metaphysical foundations of quantum mechanics we have advanced. We begin with our intuitive conceptual apparatus, good enough for our everyday working of prediction and development of the theory. Perhaps that apparatus comprises a minimal constructive logic (as suggested by da Costa [@cos08], such a logic would have the minimum for our actual development of science). With that apparatus we can develop logic and the mathematics required for quantum mechanics. Now, as we have discussed, quantum mechanics seems to demand a revision of the apparatus we began with.[^8] Logic must be built again. Now, however, differently from Kunen’s description of the situation in mathematics, a new system of logic must be developed, not inside a classical set theory, but totally new, from scratch, taking into account the lessons from science. This is what we have attempted to provide here in the case of non-individuals. Obviously we do not require that every field of knowledge take nonreflexive logic into account; no, our everyday activities are rooted on our everyday experiences, for which identity does not seem to be eliminable, at least by this time. That can be interpreted in two ways: either nonreflexive logics takes the place of classical logic everywhere, with the proviso that it is the ‘classical’ part of [$\mathfrak{Q}$]{} that is being used, or else one accepts a pluralism in logic, claiming that distinct fields of knowledge demand distinct systems of logic. That is a interesting debate, but we shall not face it now.
Suffice to say, for now, that the construction of [$\mathfrak{Q}$]{}-spaces is only a beginning to the study of the non-reflexive foundations of quantum mechanics. Nonreflexive logics still present philosophical challenges and advance some interesting problems in the foundations of quantum mechanics. It is up to us to develop those systems and discover whether the system can face the challenges.
[99]{}
Arenhart, J. R. B., Semantic analysis of non-reflexive logics. [Logic Journal of the IGPL]{}, [22]{}(4), pp.565-584, 2014.
Arenhart, J. R. B., Krause, D., Quantifiers and the foundations of quasi-set theory, *Principia*, **13**(3), pp. 251-268, 2009.
Arenhart, J. R. B. and Krause, D. Contradictions, quantum mechanics, and the square of oppositions. Forthcoming in [Logique et Analyse]{}, 2015.
da Costa, N. C. A., [Ensaio sobre os fundamentos da lógica]{}. $3^{rd}$ ed. São Paulo: Hucitec, 2008.
da Costa, N. C. A. and de Ronde, C. 2013. The Paraconsistent Logic of Superpositions. [Foundations of Physics]{} **43**: 854-858.
Dalton, J. 1808. [A new system of chemical philosophy]{}. London: S. Russell.
Domenech, G., Holik, F. and Krause, D., Q-spaces and the foundations of quantum mechanics, *Foundations of Physics* 38 (11), pp.969-994, 2008.
Falkenburg, B. [Particle Metaphysics: A Critical Account of Subatomic Reality]{}. Berlin-Heidelberg: Springer-Verlag.
French, S., [The Structure of the World. Metaphysics and Representation]{}. Oxford: Oxford Un. Press, 2014.
French, S. and Krause, D., *Identity in Physics: A Historical, Philosophical, and Formal Analysis*, Oxford:Oxford University Press 2006.
French, S., Krause, D., Remarks on quasi-set theory, *Studia Logica*, **95**(1-2) pp. 101-124, 2010.
Krause, D., On a calculus of non-individuals: ideas for a quantum mereology. In Dutra, L. H. de A.; Luz, A. M. (orgs.), [Linguagem, Ontologia e Ação]{}. Florianópolis: NEL/UFSC, v. 10, Coleção Rumos da Epistemologia. pp. 92-106, 2011.
Krause, D. and Arenhart, J. R. B., Entangling and disentangling realism and wave function. Essay Review of [The Wave Function. Essays on the metaphysics of quantum mechanics]{}, edited by Alyssa Ney and David Z. Albert (Oxford: Oxford Un. Press, 2013). [Metascience]{} 23 (3): 467-73, 2014.
Krause, D., Arenhart, J. R. B., Individuality, Quantum Physics and a Metaphysics of Nonindividuals: the role of the formal. To appear in Alexandre Guay and Thomas Pradeau (eds.) [Individuals Across the Sciences]{}, Oxford Un. Press, 2015, pp. 61-80
Kunen, K. [The Foundations of Mathematics]{}. London: College Publications, 2009.
Ladyman, J., On the identity and diversity of objects in a structure. Proceedings of the Aristotelian Society Supplementary Volume LXXXI, pp.23-43, 2007.
Ladyman, J., Ross, D., [Everything Must Go: Metaphysics Naturalized]{}. Oxford: Oxford Un. Press, 2007.
Mattuck, R. D., [A Guide to Feynman Diagrams in the Many-Body Problem]{}. 2nd. ed. New York: McGraw-Hill, 1967. (Reprinted by Dover Pu., 1992).
Merzbacher, E., [Quantum Mechanics]{}. 2nd.ed., New York: John Wiley.
Ney, A., Albert, D., (eds.) [The Wave Function: Essays on the Metaphysics of Quantum Mechanics]{}. Oxford: Oxford Un. Press, 2013.
Schrödinger, E. What is an elementary particle? In Schrödinger, E. [Science, Theory and Man]{}. London: George Allen and Unwin Ltd., pp. 193-223.
Simons, P. [Parts: a Study in Ontology]{}. Oxford: Clarendon Press, 1987.
Teller, P. [An Interpretive Introduction to Quantum Field Theory]{}. Princeton: Princeton Un. Press, 1995.
[^1]: Partially supported by CNPq.
[^2]: They constitute [surplus formal structures]{}, according to a terminology introduced by M. Redhead, that is, elements of the formalism corresponding to nothing in the real world (see [@tel95 p.25]).
[^3]: In fact, quantum physics also [can]{} (and [is]{}) be developed within a standard set theory which comprises individuals; as we have already remarked, physics works fine, but the philosophical problems turn to light.
[^4]: We work with atoms because it seems to be more intuitive to regard physical objects not as sets. As is well known, atoms are entities that can belong to sets, but which have no elements (in the sense of the membership relation), yet they could be [composed]{} of other entities in a mereological sense. The problem is that a [quantum mereology]{} was not developed yet – regarding mereology, see [@sim87]; concerning some problems related to a quantum mereology, see [@kra11].
[^5]: We recall Bertrand Russell’s definition of pure mathematics (and [$\mathfrak{Q}$]{} is the pure counterpart of our quantum mechanics) as the class of propositions of the form ‘$p$ implies $q$’, given in his [The Principles of Mathematics]{}.
[^6]: Recall the motivations for our development of quantum mechanics inside [$\mathfrak{Q}$]{}: we want to use the features of non-individuals that are encapsulated in quasi-set theory to somehow obtain quantum theory
[^7]: We are leaving aside the subindices in this notation.
[^8]: Again, see the discussion in [@kraare15] for further information on those issues.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'For a three dimensional magnetohydrodynamic (MHD) plasma the dynamo action with ABC flow as initial condition has been studied. The study delineates crucial parameter that gives a transition from coherent nonlinear oscillation to dynamo. Further, for both kinematic and dynamic models at magnetic Prandtl number equal to unity the dynamo action is studied for driven ABC flows. The magnetic resistivity has been chosen at a value where the fast dynamo occurs and the growth rate shows no further variation with the change of magnetic Reynold’s number. The exponent of growth of magnetic energy increases, indicating a faster dynamo, if a higher wave number is excited compared to the one with a lower wave number. The result has been found to hold good for both kinematic and externally forced dynamic dynamos where the backreaction of magnetic field on the velocity field is no more negligible. In case of an externally forced dynamic dynamo, the super Alfvenic flows have been found to excite strong dynamos giving rise to the growth of magnetic energy of seven orders of magnitude. The back-reaction of magnetic field on the velocity field through Lorentz force term has been found to affect the dynamics of the velocity field and in turn the dynamics of magnetic field, leading to a saturation, when the dynamo action is very prominent.'
author:
- Rupak Mukherjee
- Rajaraman Ganesh
bibliography:
- 'biblio.bib'
title: 'Study of Dynamo Action in Three Dimensional Magnetohydrodynamic Plasma with Arnold-Beltrami-Childress Flow'
---
Introduction
============
One of the most interesting open questions of astrophysics is the birth of magnetic field in the cosmos. There are several theoretical models [@zeldovich:1983; @moffatt:1978] and some of them are tested in laboratory [@gailitis:2001; @stieglitz:2001; @monchaux:2007; @ravelet:2008] also, mimicing some aspects of the astrophysical plasma. Amongst the zoo of theoretical models, E N Parker’s [@parker:1955] theory of [*dynamo action*]{} is one widely celebrated model. The large-scale magnetic field generation in the ‘Sun’ or in galaxies are mostly attributed to mean-field-dynamo. On the other hand there are astrophysical evidences of small scale magnetic field generation through turbulent fluctuation dynamo [@tzeferacos:2018; @moll:2011; @lovelace:1994; @latif:2013; @seta:2015; @kumar:2014; @st:2018]. The origin of dynamo in three dimensional plasma is still poorly understood and still a matter of debate [@brandenburg:2012]. In general, the most of the theoretical models employed in this study use the basic equations of MagnetoHydroDynamics (MHD). The model governs the dynamics of each ‘[*fluid element*]{}’ - a collisional enough fundamental block of the medium. However, MHD equations describing the plasma in the continuum limit offer fundamental challenges to the analytical solution of the basic equations [@morrison:1998]. Hence it is interesting to ask whether there is any finite dimensional description of the subject exists?\
The authors have shown that in two spatial dimensions for incompressible flows a finite dimensional approach exists and the analytical results were found to fit well with the numerical results obtained earlier [@rupak:2018b]. However, the authors also delineate the regimes where the analytical description does not hold good [@rupak:2018b]. In three spatial dimensions, the problem becomes more critical to analyse analytically. The phase space of the system being infinite dimensional, in three dimensions, long time prediction of the chaotic trajectories are extremely challenging. But, it was previously shown by the authors that for some typical chaotic flows in three spatial dimensions, the flow and the magnetic field variables are found to reconstruct back to their initial condition - thereby getting trapped in the phase space of the system [@rupak:2018c]. The cause of such [*recurrence*]{} is believed to be the low dimensional behaviour of the single fluid plasma medium for some typical parameters. Most of the short scales were not excited in the system and thus the continuum was acting like a low degrees of freedom medium.\
Therefore in is natural to ask, is there any way to excite the short scales in a regulated manner as we continuously move in the parameter scale? Finally the question becomes, what happens when all the scales are excited?\
In the first part of this paper, we address the above questions. We propose a model distinctly showing a continuous transition to self-consistent dynamo from a non-linear coherent oscillation [@rupak:2018b]. Though, an analytical description identifying the exact process is still under development, the direct numerical simulation studies of three dimensional chaotic flows support the conjecture. It is known that, in case of a short-scale dynamo, the magnetic field lines frozen to the plasma flows get first stretched along the chaotic velocity flows - then gets twisted and folded back [@ott:1998]. Such processes introduce generation of short scales into the system giving birth of dynamos classified as STF dynamo [@vainshtein:1996; @vainshtein:1997]. Though it is well known that a small but non-zero resistivity affects the plasma relaxation because of reconnection process [@woltjer:1958; @taylor:1974; @parker:1979; @taylor:1986; @qin:2012; @rupak:2018d], we choose flows with finite viscosity and resistivity showing the robustness of our results. We see a continuous growth of magnetic energy at the cost of kinetic energy and thereafter decrease of the magnetic energy through reconnection process and converting back to kinetic energy. We move in parameters and find the reconnection to occur with less probability and growth of magnetic energy. Thus we move in parameter space and in one limit observe coherent nonlinear oscillation of kinetic and magnetic energy within the premise of single fluid magnetohydrodynamics and in the other limit observe dynamo action to occur.\
From our previous study [@rupak:2018c], we choose flows that do not recur and thus even though there is an energy exchange between the kinetic and the magnetic modes, the trapping in phase space does not occur even in the opposite limit of the dynamo.\
As a test case we choose Arnold-Beltrami-Childress (ABC) flow since it is already known to produce fastest dynamos in the kinematic regimes [@alexakis:2011] and is non-recurrent [@rupak:2018c]. We reproduce the works of Galloway and Frisch [@galloway:1986] in the kinematic regime and further check our results in the self-consistent regime with the work of Sadek [*et al*]{} [@alexakis:2016]. We use the well-benchmarked three dimensional MHD code G-MHD3D [@rupak:2018; @rupak:2018a] capable of direct numerical simulation of weakly compressible single fluid MHD equations.\
In the second part of the paper, the growth of magnetic energy (dynamo action) under the action of externally driven ABC flow has been studied. The analysis can be divided into two parts, [*i)*]{} the linear kinematic regime where the plasma flow stretches the magnetic field lines giving rise to an exponential growth of magnetic field and [*ii)*]{} the nonlinear regime where the magnetic field generates Lorentz force strong enough to modify the topology of the background plasma flow, resulting in a saturation of the growth rate of the magnetic field [@childress:2008; @brandenburg:2005].\
In the linear kinematic regime, the growth of magnetic field is found to exponentially rise without bound after crossing a critical threshold Reynold’s number ($Rm_c$). When the backreaction is not negligible, (we dub this case as [*dynamic dynamo*]{}) the dynamo saturates when it enters the nonlinear regime. At very low magnetic Prandtl number ($P_m \ll 1$) the dynamic dynamo has been studied in detail [@brandenburg:2001; @steenbeck:1966; @krause:2016; @mininni:2007; @mininni:2005; @childress:2008; @galanti:1992; @ponty:1995; @archontis:2003; @frick:2006]. A detailed review of the ABC flow leading to dynamo action is due to Galloway [@galloway:2012]. However, in the solar convection zone, unprecedentedly high resolution simultion with Prandtl number unity ($P_m = 1$) has been shown to produce global-scale magnetic field even in the regime of large Reynolds numbers [@hotta:2016]. For $P_m = 1$ and Reynold’s number $Rm$ $\textgreater$ $Rm_c$, we extend our search for faster dynamos (now dynamic one) when the initial velocity and forcing scales contain higher wave-number. We also address the cases where the Alfven speed and sonic speed differ significantly. We see that the backreaction of the magnetic field alters the ABC flow profile and thereby the growth of magnetic energy itself gets affected. We notice three distinct growth rates in the magnetic energy solely because of the inclusion of the backreaction of the magnetic field on the velocity field and vice-versa.\
In particular, the following aspects of dynamo action under ABC flow have been studied in detail:
- We identify crucial parameter that controls the generation of short scales and thereby lead to dynamo action.
- We find that, above the critical magnetic Raynold’s number ($Rm_c$), the growth rate of kinematic dynamo process increases with velocity scale containing higher wave numbers.
- We also find that, above $Rm_c$, the growth rate of self-consistent or dynamic dynamo also increases, with velocity and forcing scales containing higher wave numbers.
- For super Alfvenic flows, when strong dynamo action occur, the effect of interplay of energy between magnetic and kinetic modes leads to saturation of the growth of magnetic energy at late times.
- For ABC flow to start with, both for kinematic as well as dynamic dynamo action, the magnetic energy is primarily contained in the intermediate scales.
Governing Equations
===================
The single fluid magnetohydrodynamic (MHD) description of a plasma is quite incomplete but has been found to serve aptly to explain many phenomena observed in laboratory and astrophysical systems. Thus under certain criteria, the plasma dynamics is believed to be well modelled through single fluid MHD equations. The two different charge species (electrons and ions) are assumed to form a single fluid because of the negligible mass of the electrons. A fluid element is assumed to be much larger than the length scale of separation between the two different charge species. Also the timescale at which the phenomena are observed are quite longer than the gyrofrequency of each of the charge species. Thus no large scale electric field is produced or sustained in the timescale of interest.\
The basic equations governing the dynamics of the magnetohydrodynamic fluid are as follows: $$\begin{aligned}
&& \label{density} \frac{\partial \rho}{\partial t} + \vec{\nabla} \cdot \left(\rho \vec{u}\right) = 0\\
&& \frac{\partial (\rho \vec{u})}{\partial t} + \vec{\nabla} \cdot \left[ \rho \vec{u} \otimes \vec{u} + \left(P + \frac{B^2}{2}\right){\bf{I}} - \vec{B}\otimes\vec{B} \right]\nonumber \\
&& \label{velocity} ~~~~~~~~~ = \mu \nabla^2 \vec{u} + \rho\vec{f}\\
&& \label{Bfield} \frac{\partial \vec{B}}{\partial t} + \vec{\nabla} \cdot \left( \vec{u} \otimes \vec{B} - \vec{B} \otimes \vec{u}\right) = \eta \nabla^2 \vec{B}\\
&& \text{where} ~~ P = C_s^2 \rho ~ \text{and} ~ \vec{f} = \begin{pmatrix}
A \sin(k_f z) + C \cos (k_f y) \\
B \sin(k_f x) + A \cos (k_f z) \\
C \sin(k_f y) + B \cos (k_f x)
\end{pmatrix}. \nonumber \end{aligned}$$ In the above system of equations, $\rho$, $\vec{u}$, $P$ and $\vec{B}$ are the density, velocity, kinetic pressure and the magnetic field of a fluid element respectively. $\mu$ and $\eta$ denote the coefficients of kinematic viscosity and magnetic resistivity. We assume $\mu$ and $\eta$ are constants throughout space and time. The symbol “$\otimes$” represents the dyadic between the two vector quantities.\
The kinetic Reynold’s number ($Re$) and magnetic Reynold’s number ($Rm$) are defined by $Re = \frac{U_0 L}{\mu}$ and $Rm = \frac{U_0 L}{\eta}$ where $U_0$ is the maximum velocity of the fluid medium to start with and $L$ is the system length.\
We also define the sound speed of the fluid medium as $C_s = \frac{U_0}{M_s}$, where, $M_s$ is the sonic Mach number of the fluid. We assume it to be uniform throughout the space and time. The Alfven speed is determined from the relation $V_A = \frac{U_0}{M_A}$ where $M_A$ is the Alfven Mach number of the plasma medium. The initial magnetic field present in the plasma is determined from relation $B_0 = V_A \sqrt{\rho_0}$, where, $\rho_0$ is the initial density profile of the fluid.
Parameter Details {#sec:parameter}
=================
The first results of Galloway and Frisch [@galloway:1984] showed critical dependency on the magnitude of magnetic resistivity ($Rm$). Later this result was further tested and reproduced with much greater accuracy and resolution by in several other independent studies [@bouya:2013; @bouya:2015]. The observation that, within a $2 \pi$ periodic box, for algorithms depending on spectral solvers, the smallest features in magnetic field are on scales of order $Rm^{-1/2}$ indicates that the grid size required to resolve a given $Rm$ scales like $Rm^{1/2}$ [@galloway:1986]. We choose, $N = 64$ which resolves $Rm$ upto $4096$ and keep our parameters fixed at $Rm = 450$ (where the growth rate of dynamo was found to get saturated [@galloway:1986]) well within the resolution threshold. Also the result of Sadek [*et al*]{} [@alexakis:2016] confirms that most of the kinetic and magnetic energy content remains within the large scales, even when the driving wave-number is kept at intermediate scales (at least upto $k_f = 16$). This sets limit to our choice of maximum driving wave number ($k_f = 16$) at the grid resolution $N = 64$.
Throughout our simulation, we set $N = 64$, $L = 2 \pi$, $\delta t = 10^{-4}$, $\rho_0 = 1$. For some test runs the grid resolution is increased to $N = 128$ for both kinematic and dynamic cases but we found no significant variation of the physics results. The initial magnitude of density ($\rho_0$) is known to affect the dynamics and growth rate of an instability in a compressible neutral fluid [@bayly:1992; @terakado:2014]. However, in present case we keep the initial density fixed ($\rho_0 = 1$) for all the runs. We check our code with smaller time stepping ($\delta t$) keeping the grid resolution $N = 64$. No deviation from the results were observed with such test runs.
The kinematic viscosity is controlled through the parameters $Re$ and to guarantee similar decay of kinetic and magnetic energy, we set $Re = Rm$ everywhere. Next we vary the Alfven speed through $M_A$ and observe the effect of these parameters on the dynamo action. We also change the magnitude of forcing by controlling the values of $A, B, C$ and the length-scale of forcing through $k_f$.
The OpenMP parallel MHD3D code is run on 20 cores for 9600 CPU hours for a single run with parameters mentioned above and got the following results.
Simulation Results {#sec:simulation_results}
==================
Initial Profile of Density, Velocity and Magnetic Field
-------------------------------------------------------
We start with the initial condition $\rho = \rho_0$ as a uniform density fluid, the initial velocity profile as $u_x = U_0 [A \sin(k_f z) + C \cos (k_f y)]$, $u_y = U_0 [B \sin(k_f x) + A \cos (k_f z)]$, $u_z = U_0 [C \sin(k_f y) + B \cos (k_f x)]$ and the initial magnetic field as $B_x = B_y = B_z = B_0$. We keep the initial profiles of all the fields identical throughout our paper unless otherwise stated.
Transition to Dynamo
--------------------
For $M_A \sim 1$, a coherent nonlinear oscillation is reproduced as reported earlier [@rupak:2018b]. As $M_A$ is moved from unity, the oscillation persists alongwith the generation of other modes into the system. Thus as can be found from Fig. \[dynamo\_MA\], the magnetic energy does not come back to its initial value after one period of oscillation. Upon further increment of Alfven Mach number, the linear dependency of the frequency of oscillation breaks down and persistent magnetic field starts to generate. Finally, the growth of magnetic energy reaches a maximum. From Fig. \[dynamo\_MA\], it can be seen that, the normalised magnetic energy at $M_A = 10^2$ & $10^3$ does not differ significantly, indicating a saturation of the growth. However, such saturation does not occur in the driven cases where, the plasma is driven continuously using an external drive which pumps in kinetic energy to the system. Such phenomena is further explored in the next section of the paper.
![(Color online) Transition to dynamo with the increase of $M_A$ from coherent nonlinear oscillation.[]{data-label="dynamo_MA"}](dynamo_MA.pdf)
Kinematic Dynamo {#subsec:kinematic_dynamo}
----------------
The phenomenon of magnetic energy growing exponentially with time for a statistically steady flow, where the velocity field is held fixed in time, is called, kinematic dynamo action. Arnold-Beltrami-Childress (ABC) flow being a steady solution of Euler equation, sets the premise to study the kinematic dynamo problem. For ABC flow kinematic dynamo was first obtained by Arnold [*et al*]{} [@arnold:1983] at magnetic Reynold’s number ($Rm$) between $9$ and $17.5$. Galloway [*et al*]{} [@galloway:1984] found a more efficient dynamo effect with much higher growth rate after $Rm = 27$ breaking certain symmetries of the flow. Later on, the study had been extended for the parameters where $A$, $B$, $C$ are not equal [@galloway:1986]. The threshold $Rm$ for a kinematic dynamo has been well explored [@ponty:2005; @mininni:2007; @schekochihin:2005; @iskakov:2007]. The real part of the growth rate of the magnetic energy for increasing $Rm$ is found to increase while the imaginary part decreases continuously [@galloway:1986]. ABC flows with differnet forcing scales ($k_f \neq 1$) providing kinematic dynamo has been explored by Galanti [*et al*]{} [@galanti:1992] and more recently by Archontis [*et al*]{} [@archontis:2003].
First we reproduce the previous results and then choose an optimal set of working parameters. The motivation behind choosing the parameters are explained in the previous section. We give the following runs (Table \[parameter\_run\_kinematic\]) to explore the parameter regime of a kinematic dynamo problem.
Name $k_f$ $M_A$ $Rm$
------ ------- -------- -------
KF1 $1$ $1000$ $450$
KF2 $2$ $1000$ $450$
KF3 $4$ $1000$ $450$
KF4 $8$ $1000$ $450$
KF5 $16$ $1000$ $450$
KM1 $1$ $100$ $450$
KR1 $1$ $10$ $450$
KR2 $1$ $10$ $200$
KR3 $1$ $10$ $120$
KM2 $1$ $1$ $450$
KM3 $1$ $0.1$ $450$
: Parameter details with which the simulation has been run for kinematic dynamo problem.[]{data-label="parameter_run_kinematic"}
### Effect of Magnetic Resistivity
Effect of magnetic resistivity ($\eta$) through the magnetic Reynold’s number ($Rm$) has been widely studied in past [@galloway:1986; @galloway:1986; @galanti:1992] and in recent years [@bouya:2013]. First we reproduce the previous results by Galloway [*et al*]{} [@galloway:1986] using our code \[Runs: KR1, KR2, KR3\]. Similar to the previous study [@galloway:1986] we choose $U_0 = 1$, $A = B = C = 1$, $k_f = 1$. We time evolve only Eq. \[Bfield\] for the initial time data mentioned above for magnetic Reynold’s number $Rm = 120, 200, 450$ and obtain the identical growth of magnetic field as Galloway [*et al*]{} [@galloway:1986]. This result is shown in Fig \[Frisch\]. We also reproduce the real and imaginary part of the eigenvalue obtained previously [@galloway:1986]. The critical value of onset of kinetic dynamo action is found to be $Rm = 27$.\
![Kinematic Dynamo effect reproduced using the identical parameter regime ($A = B = C = 1$, $k_f = 1$) by Galloway [*et al*]{} [@galloway:1986]. The grid resolution is $64^3$ which is close to the value $60^3$ that was taken by Galloway [*et al*]{} [@galloway:1986]. The growth rates of magnetic energy $\left(\sum\limits_{V} \frac{B^2(x,y,z)}{2}\right)$ of kinematic dynamo are found to increase as $Rm$ is increased. The oscillation frequency of the magnetic energy is also found to be similar as Galloway [*et al*]{} [@galloway:1986][]{data-label="Frisch"}](dynamo.pdf)
We derive the energy spectra of the kinematic dynamo from ABC flow \[Fig.(\[Spectra\_Frisch\])\] and observe energy is not only contained in large scales rather, the energy contained in the intermediate scles are quite large. We observe a $k^{0.7}$ scaling of magnetic energy.
![Magnetic energy spectra at different time for ABC flow with the identical parameter described in Fig. (\[Frisch\]) with $Rm = 450$. The energy contained in the large scales at late times shows that the short scales are equally important for a kinematic dynamo obtained from ABC flow.[]{data-label="Spectra_Frisch"}](B_Spectra_Kinematic.pdf)
### Effect of Forcing Scale {#subsubsec:forcing_scale}
The effect of forcing scale on the growth rate of magnetic energy has been earlier studied by Galanti [*et al*]{} [@galanti:1992] for $k_f = 1$ to $10$ for $Rm$ values upto $Rm = 45$. For the kinematic dynamo case, Galloway and Frisch ([@galloway:1986]) have shown that, even though the critical value of $Rm$ for kinematic dynamo action for ABC flow is $Rm_c = 27$, the growth rate monotonically increases until $Rm = 350$. In the past work of Galanti [*et al*]{} [@galanti:1992] it was found that $k_f = 2$ has higher growth rate than $k_f =1$ for $Rm = 45$. Also changing $Rm$ from $12$ to $20$ did not affect the growth rate for different $k_f$ much. In our case, we keep the forcing length scale at $k_f = 1, 2, 4, 8, 16$ holding the $Rm = 450$ much above the critical value ($Rm_{crit} = 27$) of onset of kinetic dynamo for $k_f = 1$ (where imaginary part of the eigenvalue ($2 \gamma$) is undetectably small) and in the regime where the growth rate does not vary much with the further increment of $Rm$. \[Runs: KF1, KF2, KF3, KF4, KF5\] The growth rate of normalised magnetic energy, $\frac{B^2}{B^2_0}$, is found to increase as $k_f$ is increased \[Fig.\[kin\_kf\_U0\_1\], \[kin\_kf\_initial\]\]. However, the growth rate ($2 \gamma$) saturates as $k_f$ is increased for $U_0 = 0.1$\[\[kin\_kf\_initial\]. A similar saturation was also observed earlier though at $Rm = 12$ and $20$ [@galanti:1992].
The late time dynamics is found to be widely different for different driving frequencies ($k_f$) \[Fig.\[kin\_kf\_U0\_1\],\[kin\_kf\_initial\]\]. For a kinetic dynamo problem similar transient behaviour ($k_f = 16$ and $8$ in Fig. \[kin\_kf\_initial\]) starting from a typical initial condition has been addresses previously in detail [@bouya:2013]. It was found that when the fastest growing eigenmode is not excited, it takes some time for the fastest eigenmode to overcome the initially excited mode and hence the crossover happens at a later time. Even for a dynamic dynamo under external forcing, similar result was earlier obtained by Galanti [*et al*]{} [@galanti:1992] for $A = B = C = 1$, $k_f = 1$ and $Re = Rm = 12$.
![Kinematic Dynamo effect for different driving frequency ($k_f$). The magnetic energy is normalised with the initial magnitude at time $t = 0$. The parameters chosen are $A = B = C = 1$ and $Re = Rm = 450$ with $U_0 = 1$. The initial growth rates are found to grow as $k_f$ is increased.[]{data-label="kin_kf_U0_1"}](kin_kf_U0_1.pdf)
![Kinematic Dynamo effect for different driving frequency ($k_f$). The normalised magnetic energy is defined as $\left(\sum\limits_{V} \frac{B^2(x,y,z,t)}{2} - \sum\limits_{V} \frac{B^2(x,y,z,0)}{2}\right)$. The parameters chosen are $A = B = C = 1$ and $Re = Rm = 450$ with $U_0 = 0.1$. The initial growth rates are found to grow and saturate as the $k_f$ is increased.[]{data-label="kin_kf_initial"}](kin_kf.pdf)
### Effect of Alfven Speed
The Alfven speed is defined as $V_A = \frac{U_0}{M_A}$. If $M_A \textless 1$; $V_A \textless U_0$ and the plasma is called Sub-Alfvenic. Similarly if $M_A \textgreater 1$, the plasma is Super-Alfvenic. For kinetic dynamo problem, the growth rate of magnetic energy is found to be independent of the magnitude of $M_A$. \[Runs: KF1, KM1, KR1, KM2, KM3\] We check the growth rate for $M_A = 0.1, 1, 10, 100, 1000$ and for every case the growth rate of dynamo is found to be identical as shown in Fig. \[MA\_kin\_U0\_1\] and \[MA\_kin\] unlike the dynamic case discussed in the next subsection.
![Kinematic Dynamo effect for different driving frequency ($M_A$). The parameters chosen are $A = B = C = 1$ and $Re = Rm = 450$ with $U_0 = 1$. The growth rates are found to be identical for different $M_A$ values.[]{data-label="MA_kin_U0_1"}](MA_Kin_U0_1.pdf)
![Kinematic Dynamo effect for different driving frequency ($M_A$). The parameters chosen are $A = B = C = 1$ and $Re = Rm = 450$ with $U_0 = 0.1$. The growth rates are found to be identical for different $M_A$ values.[]{data-label="MA_kin"}](MA_Kin.pdf)
Dynamo with Back-reaction
-------------------------
A dynamic dynamo represents a situation where the magnetic energy grows exponentially for a plasma where the plasma itself evolves in time. Hence the velocity field is not externally imposed like a kinematic dynamo, rather it has a dynamical nature. The time evolution of the velocity field is generally governed by the Navier-Stokes equation including the magnetic feedback on the velocity field. In order to simulate such a scenario, we time evolve all the three equations, viz. Eq. \[density\], \[velocity\], \[Bfield\]. A result for parameters $M_A = 1000$ and $k_f = 1$ for initial flow profile ABC is given in Fig. \[Long\].\
![Dynamic dynamo growth of kinetic $\left(\sum\limits_{V} \frac{U^2(x,y,z)}{2}\right)$ and magnetic $\left(\sum\limits_{V} \frac{B^2(x,y,z)}{2}\right)$ energy for ABC flow with $M_A = 1000$ and $k_f = 1$ for a long time with initial flow profile as ABC flow.[]{data-label="Long"}](Long.pdf)
We change the forcing scale, Alfven velocity and compressibility to observe the effect on the dynamics of the fields.\
We turn on external forcing to the velocity field. We keep the nature of forcing as $f_x = A \sin(k_f z) + C \cos (k_f y)$, $f_y = B \sin(k_f x) + A \cos (k_f z)$, $f_z = C \sin(k_f y) + B \cos (k_f x)$. We keep $A = B = C = 0.1$ and $U_0 = 1$ throughout all the calculations and fix $Re = Rm = 450$. In case of an external forcing the initial memory is lost and hence the sensitivity to the initial condition is expected to be lost. We redo our numerical calculations for an initial random velocity field profile and find that the basic nature of dynamo effect does not get affected as shown in Fig. \[Rand\]. The saturation regime for both the kinetic (sum over all velocity modes) and magnetic (sum over all magnetic modes) energies remain the same though the two systems are evolved from different initial conditions. We perform the following runs (Table. \[parameter\_run\_for\_dyn\]) using our code to understand the externally forced ABC flow dynamo process.\
![Dynamic dynamo growth of kinetic and magnetic energy for two different initial conditions with $M_A = 1000$ and $k_f = 1$. The initial growth rate of magnetic energy for ABC initial flow profile is found to be identical with that of random field profile.[]{data-label="Rand"}](Rand.pdf)
Name $k_f$ $M_A$ $M_s$
------- ------- -------- -------
FDF1 $1$ $1000$ $0.1$
FDF2 $2$ $1000$ $0.1$
FDF3 $4$ $1000$ $0.1$
FDF4 $8$ $1000$ $0.1$
FDF5 $16$ $1000$ $0.1$
FDMA1 $1$ $100$ $0.1$
FDMA2 $1$ $10$ $0.1$
FDMA3 $1$ $1$ $0.1$
FDMA4 $1$ $0.1$ $0.1$
FDMS1 $1$ $100$ $0.2$
FDMS2 $1$ $100$ $0.3$
FDMS3 $1$ $100$ $0.4$
FDMS4 $1$ $100$ $0.5$
: Parameter details with which the simulation has been run for the externally forced dynamic dynamo problem.[]{data-label="parameter_run_for_dyn"}
Now we vary $U_0$ and the magnitude of $A, B, C$ keeping $A = B = C$ for all the cases. We run our simulation for $U_0 = 0.1, 0.2, 0.3, 0.4, 0.5$ keeping $A = B = C = 0.1$ and see the trend of dynamo action is identical for all values of $U_0$ \[Fig. \[U\_Comp\_Forced\]\]. Next we vary the values of $A = 0.1, 0.1, 0.3$ keeping $A = B = C$ and $U_0 = 0.1$. We see faster growth of dynamo with higher values of forcing through the magnitudes of $A$, $B$ and $C$ \[Fig.\[ABC\_Comp\_Forced\]\].
![Dynamic dynamo growth of kinetic and magnetic energy for five different initial velocities viz. $U_0 = 0.1, 0.2, 0.3, 0.4, 0.5$, with $M_A = 1000$ and $k_f = 1$. The initial growth rate of magnetic energy for ABC initial flow profile is found to be identical for all initial magnitude of velocities.[]{data-label="U_Comp_Forced"}](U_Comp_Forced.pdf)
![Dynamic dynamo growth of kinetic and magnetic energy for three different forcing magnitudes $A = 0.1, 0.2, 0.3$ having $A = B = C$, with $M_A = 1000$ and $k_f = 1$. The initial growth rate of magnetic energy for ABC initial flow profile is found to be increase with the increament of magnitude of forcing.[]{data-label="ABC_Comp_Forced"}](ABC_Comp_Forced.pdf)
### Effect of Forcing scale {#effect-of-forcing-scale}
A dynamic dynamo with external forcing has been studied earlier by Galanti [*et al*]{} [@galanti:1992] for incompressible plasma with $U_0 = 1$ $A = B = C = 0.1$, $Re$ & $Rm$ upto $20$ (below $Rm_c = 27$), and $k_f = 1, 2, 4$. We change the length scale of forcing ($k_f$) on the velocity field keeping $U_0 = 0.1$, $A = B = C = 0.1$, $Re = Rm = 450$, $M_s = 0.1$ and $M_A = 1000$ as fixed parameters. \[Runs: FDF1, FDF2, FDF3, FDF4, FDF5\] From Fig. \[kf\_F\_MA\_1000\] we find, the growth rate of magnetic energy increases while that of kinetic energy decreases as $k_f$ is increased. The case $k_f = 16$ in Fig. \[kf\_F\_MA\_1000\] shows a delayed dynamo action. A possible explanation of this late time dynamo action is the excitation of a slow eigenmode to start with, which gets overpowered by the fastest eigenmode excited later. The identical phenomena we have seen in the kinematic dynamo section \[Fig. \[kin\_kf\_initial\]\]. From Fig. \[kf\_F\_MA\_1000\] we also note that though externally forced, the saturation regime of both kinetic and magnetic eneries goes downwards as $k_f$ is increased. This is so, because the forcing scale also has a wave number term within it which helps to drain out energy through viscous dissipation, if a higher wavenumber ($k_f$) is excited.\
![Dynamic dynamo growth of kinetic and magnetic energy for $k_f = 1, 2, 4, 8, 16$ with $M_A = 1000$. The growth rate of magnetic energy is found to be steeper with the increase of $k_f$. The growth rate of kinetic energy due to external forcing decreases as $k_f$ is increased.[]{data-label="kf_F_MA_1000"}](kf_F_MA_1000.pdf)
### Effect of Alfven Speed
We change the Alfven Mach number ($M_A$) of the plasma, keeping $U_0 = 0.1$ $A = B = C = 0.1$, $Re = Rm = 450$, $M_s = 0.1$ and $k_f = 1$.
We analyse the runs: FDF1, FDMA1, FDMA2, FDMA3. By choosing $M_A = 1, 10, 100$ and $1000$ we set the Alfven Velocity $V_A = \frac{U_0}{M_A} = 10^{-1}, 10^{-2}, 10^{-3}$ and $10^{-4}$ respectively. For $\rho_0 = 1$, $V_A = B_0$, the initial magnitude of the seed magnetic field profile. As we start from a lower value of $B_0$, the growth rate of the magnetic energy increases rapidly. This is quite similar to the kinematic dynamo action with a distinct difference. In kinematic dynamo there was no saturation of magnetic energy. On the other hand, in forced dynamic dynamo, there is a saturation value of the magnetic field. This saturation is believed to be due to the backreaction of the magnetic field on the velocity field through the Lorentz force term. The strong magnetic field generated through the dynamo process, starts affecting the topology of the velocity field in turn affecting its dynamics. Thus the modified velocity field no longer remains a ABC flow and finally the dynamo saturates. The effect of such magnetic feedback on the velocity field is shown in Fig \[feedback\] for $M_A = 1,100, 1000$.\
We do the following observations from Fig \[feedback\].
$\bullet$ We notice that, for both the case $M_A = 100$ and $1000$ there exists three distinct slopes. At the beginning, the magnetic energy starts exponentialy increasing with time. Once it gets amplified by around four orders of magnitude, the exponent of increament suddenly falls down for both the cases $M_A = 100$ and $1000$. After that, the magnetic energy again starts increasing with higher exponent.
$\bullet$ It is also note-worthy that, the initial growth rate of the magnetic energy for $M_A = 100$ and $1000$ are identical though they differ later on. Thus we understand it as similar to Kinematic dynamo (\[MA\_kin\]) where the backreaction is negligible. However, at later time because of the difference in the strength of the backreaction, the slopes of increament of the magnetic energy in logarithmic scale differs.
$\bullet$ We also find that when the growth of the dynamo is several orders of magnitude (for higher values of $M_A$) the kinetic energy also grows faster though ultimately both kinetic and magnetic energies saturate at the same value.
$\bullet$ We see that, independent of the strength of the seed magnetic field, the saturation regime of the kinetic and magnetic energies are the same.
Thus we conclude from the above observations that, if the velocity field is ABC forced, whatever be the seed magnetic field, the dynamo effect becomes possible in super Alfvenic systems and the dynamo action is quite strong leading the final magnetic energy comparable to the kinetic energy.\
![Dynamic dynamo growth of kinetic and magnetic energy for $M_A = 1, 100, 1000$. The back-reaction of magnetic field on velocity field is found to affect the growth rate and dynamics of velocity field. This effect is captured in the time evolution of kinetic energy for different $M_A$.[]{data-label="feedback"}](Energy.pdf)
### Energy Spectra
Now we analyse the kinetic and magnetic energy spectra of the dynamic dynamo action at different times for $U_0 = 0.1$, $A = B = C = 0.1$, $k_f = 1$, $M_A = 1000$, $Re = Rm = 450$. Initially the energy content was limited to the fundamental mode only. But in course of time the kinetic energy shows a $k^{-5/3}$ spectra while the magnetic energy shows a $k^{0.7}$ spectra identical to the kinematic dynamo phenomena. However it is worth notable that the growth of magnetic energy in intermediate scales is much slower than the kinematic dynamo as can be found in Fig.\[Spectra\_Frisch\]\
![Kinetic energy spectra for dynamic dynamo with $U_0 = 0.1$, $A = B = C = 0.1$, $k_f = 1$, $M_A = 1000$, $Re = Rm = 450$.[]{data-label="E_Spectra"}](E_Spectra.pdf)
![Kinetic energy spectra for dynamic dynamo with $U_0 = 0.1$, $A = B = C = 0.1$, $k_f = 1$, $M_A = 1000$, $Re = Rm = 450$.[]{data-label="B_Spectra"}](B_Spectra.pdf)
Summary and Future Works
========================
In this work we have analysed several phenomena of a magnetohydrodynamic plasma under ABC flow.\
$\bullet$ First we study the kinematic dynamo effects where the velocity field is the ABC flow - a known solution of Euler equation. At different wave-numbers of the flow, we see that the growth rate of the magnetic energy in the kinematic dynamo case increases as $k_f$ is increased.\
$\bullet$ In case of an ABC forced velocity field for the dynamic dynamo problem seems to show similar variation with $k_f$, though now the dynamo action becomes very prominant. The magnetic energy grows upto the order of kinetic energy when we remain in the super Alfvenic regime.\
$\bullet$ The magnetic energy is found to be contained primarily in the intermediate scales in wave-number.\
The compressibility however has not been found to affect the results for the weakly compressible cases. The effect of variation of initial density ($\rho_0$) on the dynamo effect can be an interesting piece of study and will be explored elsewhere.\
Acknowledgement
===============
R.M. acknowledges several insightful discussions with Akanksha Gupta, Vikrant Saxena and Abhijit Sen at Institute for Plasma Research, India. The development as well as benchmarking of MHD3D has been done at Udbhav and Uday clusters at IPR.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In order to resolve and characterize anisotropy in turbulent plasma flows a proper estimation of the background magnetic field is crucially important. Various approaches to calculate the background magnetic fields, ranging from local fields to globally averaged fields, are commonly used in the analysis of turbulent data. Here we investigate how the uncertainty in the orientation of a scale dependent background magnetic field influences the ability to resolve anisotropy. Therefore we introduce a quantitative measure, the *angle uncertainty*, which characterizes the uncertainty of the orientation of the background magnetic field which turbulent structures are exposed to. The angle uncertainty can be used as a condition to estimate the ability to resolve anisotropy with certain accuracy. We apply our description to resolve spectral anisotropy in fast solar wind data. We show that if the angle uncertainty grows too large, the power of the turbulent fluctuations is attributed to false local magnetic field angles, which may lead to an incorrect estimation of spectral indices. In our results an apparent robustness of the spectral anisotropy to false local magnetic field angles is observed, which can be explained by a stronger increase of power for lower frequencies when the scale of the local magnetic field is increased. The frequency dependent angle uncertainty is a measure which can be applied to any turbulent system.'
author:
- 'F. Gerick'
- 'J. Saur'
- 'M. von Papen'
title: The uncertainty of local background magnetic field orientation in anisotropic plasma turbulence
---
Introduction {#sec:intro}
============
Turbulent flows in magnetized plasmas are anisotropic due to the presence of a magnetic field (see, e.g. reviews by @Horbury2012 and @Oughton2015). In contrast to the velocity field, for the magnetic field no Galileo transformation exists such that for a certain eddy (or turbulent structure), the magnetic field associated with larger eddies vanishes. Therefore the magnetic field of all larger scales directly influences the smaller scales of turbulence.
Deciphering the anisotropic structure of plasma turbulence is a major challenge and several models are debated in the literature (e.g., @Matthaeus1990 [@GS95; @Bieber1996; @Saur1999; @Galtier2005; @Boldyrev2006; @Galtier2006; @Beresnyak2008; @Howes2008; @Howes2011; @Boldyrev2012; @Narita2015]). For understanding the anisotropy of turbulence in magnetized plasmas, the spatial and temporal extents of the magnetic field controlling the orientation and the decay of the turbulent eddies of specific scales remains unclear, but is of crucial importance.
Two approaches are commonly used to characterize the controlling scale of the magnetic field, referred to as global and local frame [@Maron2001; @Horbury2008; @Beresnyak2009a; @Cho2009; @Tessein2009; @Chen2011; @Matthaeus2012]. In the global frame the magnetic field ${\mathbf}{B}(t)$ is averaged over scales much larger than the correlation length of the turbulent fluctuations to obtain the global mean field ${\mathbf}{B}_0$. In the local frame, on the other hand, it is assumed that a magnetic field at scales on the same order as those given by the individual turbulent structure or eddy control the anisotropy of the turbulence [@Cho2000; @Maron2001; @Cho2002; @Cho2004].
These considerations on the controlling scales are relevant for magnetized plasmas whether observed in space or generated in numerical simulations. They impose important questions if the turbulent flow contains fluctuations ${\mathbf}{\delta b}{=}{\mathbf}{B}-{\mathbf}{B}_0$ with a root mean square (RMS) similar or larger compared to the mean magnetic field obtained by averaging over global scales. Only in the case of ${\mathbf}{B}_0{\gg} {\mathbf}{\delta b}$, the problem simplifies because the local background field is approximately equal to the global mean field.
The solar wind is a medium where the large scale background magnetic field ${\mathbf}{B}_0$, averaged over hours, days or years is often on the same order as the RMS of the magnetic field fluctuations ${\mathbf}{\delta b}$. Solar wind studies using a global magnetic field frame only detected anisotropy in the power of the fluctuations, but no anisotropy in the spectral index [@Tessein2009]. On the other hand several studies using a local and scale dependent magnetic field for the analysis have revealed anisotropy in both power and spectral index $\kappa$ in the inertial range spectrum of solar wind data [@Horbury2008; @Alexandrova2008; @Podesta2009; @Chen2010; @Luo2010; @Wicks2010; @Wicks2011; @Podesta2013]. @Horbury2008, for the first time, analyzed the spectral index $\kappa$ with respect to a background magnetic field using such a local frame. The observed spectrum showed a spectral index of -2 parallel compared to -5/3 perpendicular to the local magnetic field (see Figure 2, lower panel, in @Horbury2008), which is in agreement with the predicted scalings of the critical-balance theory [@GS95]. Besides the spectral index, other anisotropic properties have been successfully analyzed using a local and scale dependent magnetic field (e.g., @Salem2012 [@He2013; @Bruno2015]).
Here we introduce a necessary geometrical condition for the scales of the magnetic field to observationally resolve anisotropy within measured or simulated data. This condition is given by the average uncertainty in the orientation of the local background magnetic field at a certain scale. This uncertainty in the orientation is measured by the angle of the local background magnetic field with respect to the orientation of the observed fluctuations. This is referred to as *angle uncertainty* in the remainder of this work. To quantify the angle uncertainty in spacecraft measurements, we use the *field to flow angle* $\theta$, which is defined as the angle between the magnetic field ${\mathbf}{B}$ and the unperturbed flow direction of the solar wind ${\mathbf}{v}_\mathrm{SW}$. In other systems the orientation of the field may be conveniently defined in a different way.
![\[fig:intro\]Schematic representation of the orientation of an eddy with respect to the orientation of two different background fields averaged at the scale of the eddy $s_e$ and averaged at some larger scale $s_b$.](fig1.pdf){width="\linewidth"}
The orientation of an elongated eddy within a magnetic vector field ${\mathbf}{B}$ is shown schematically in Figure \[fig:intro\]. Measuring along the dashed line the magnetic field averaged over scale $s_e$, which characterizes the size of an eddy (detailed definition in the following section), is associated with the angle $\theta_e$. If the associated background magnetic field is defined over a larger scale $s_b$ the field to flow angle is $\theta_b$. The fluctuations observed at the eddy scale are in this case associated with a different field to flow angle. We hypothesize that if the angle discrepancy between the scale at which the magnetic field is averaged and the eddy scale grows beyond a certain threshold, the angle of the local magnetic field is no longer well estimated. Therefore, the anisotropic properties of turbulent eddies might not be resolved under the assumption that the orientation of the eddies adjust locally to the magnetic field.
In the following we define the necessary scales, give a mathematical definition of the background magnetic field for different levels of localizations, and formally introduce the angle uncertainty as a measure for the orientation of an eddy within such averaged magnetic fields. Subsequently we apply it to 91 days of magnetic field measurements within the fast solar wind [@Wicks2010] and explore its suitability as a necessary condition to resolve observed or expected solar wind spectral anisotropy.
Analysis of spectral anisotropy {#sec:methods}
===============================
Relevant Scales & Wavelet Method {#ch:wavelets}
--------------------------------
To analyze spectral properties of turbulent fluctuations we use a method based on the wavelet transformation. We denote $B_i(t)$ with $i{=}R,T,N$ the magnetic field components measured as a function of time $t$ in the RTN-coordinate system[^1].
The wavelet transformation of the components $B_i(t)$ is calculated as $$W_i(t,\sigma)=\frac{1}{\sqrt{\sigma}}\int\limits B_i(t')\psi\left(\frac{t'-t}{\sigma}\right)dt',
\label{eq:wavelet}$$ where $\psi(\eta)$ is the mother wavelet and $\sigma$ the wavelet scale. The absolute squared values of the complex wavelet coefficients, $|W_i(t,\sigma)|^2$, give energy density at time $t$ and wavelet scale $\sigma$. In case of the Morlet wavelet $$\psi(\eta)=\pi^{-1/4}e^{-i\omega_0 \eta}e^{-\eta^2/2}
\label{eq:morlet}$$ the wavelet scale is the standard deviation of the Gaussian amplitude envelope of the wavelet displayed as dashed line in Figure \[fig:morlet\] [@TC98]. The width of the wavelet can thus be defined by the full width at half maximum of the Gaussian window $2\sqrt{2\ln(2)}\sigma$ (shown as light grey area in Figure \[fig:morlet\]). The wavelet packet given by Equation can be associated with two scales. One scale $s_e$ is associated with the period (frequency) of the fluctuations of the turbulent eddy that is to be analyzed. The other scale $s_b$ is associated with the full width at half maximum of the wavelet, which constrains the temporal resolution and which is used in the following section for definition of the local background magnetic field.
![Real part of the normalized Morlet wavelet in time domain with the amplitude envelope (dashed), the Gaussian with standard deviation $\sigma_e$ and the two characteristic time scales $s_e$ (dark grey area) and $s_b(\sigma)$ (light grey area).[]{data-label="fig:morlet"}](fig2.pdf){width="\linewidth"}
The translation from wavelet scale to frequency $$f_e=(\omega_0{+}\sqrt{2{+}\omega_0^2})/(4\pi\sigma)$$ depends on the number of oscillations $\omega_0$ within the wavelet [@Meyers1993]. Here we use $\omega_0{=}6$ so that $f_e{=}(1.033\sigma)^{-1}$. We define the scale $s_e$ of the eddy under consideration as one period of the frequency $1/f_e$ (shown in Figure \[fig:morlet\], dark grey area) independent of the choice of $\omega_0$. The energy density of the wavelet coefficients can be associated with the eddy frequency $f_e$, which is consistent with the classical Fourier analysis commonly used in turbulence analysis.
To associate a local background magnetic field to each wavelet coefficient one can use a Gaussian with standard deviation $\sigma_b{=}\sigma$. This is a reasonable choice as it describes the scale over which the energy density in the wavelet is calculated [@Horbury2008]. In the remainder of this work we also investigate background magnetic fields averaged over larger scales using Gaussian windows with standard deviation $\sigma_b {>} \sigma$. We therefore use a local, scale dependent background magnetic field $b_i$ for each component $i$ given by
$$b_i(t,s_b) = \int B_i(t') \exp\left(\frac{-(t'-t)^2}{2\sigma_b^2}\right) dt',\label{eq:magsmooth}$$
the convolution of the magnetic field with a Gaussian [@Horbury2008; @Podesta2009]. We introduce a dimensionless factor $\alpha$, so that $\sigma_b{=}\alpha \sigma$, to quantify the increase of the averaging width. In case of $\alpha{=}1$ the averaging width corresponds to the envelope of the wavelet. The total averaging scale is $s_b(\sigma_b){=}2\sqrt{2\ln(2)}\sigma_b$. Standard deviations smaller than $\sigma_b{=}\sigma$ should not be used to average the magnetic field since the energy density of the associated wavelet coefficients would correspond to wavelets larger than the averaged magnetic field.
The ratio of the smallest possible averaging scale and the eddy scale $$\frac{s_b(\sigma)}{s_e}=\frac{2\sqrt{2\ln{2}}(\omega_0+\sqrt{2+\omega_0^2})}{4\pi},\label{eq:se-sb-ratio}$$ depends only on the choice of $\omega_0$. This ratio is always larger than one and increases with larger $\omega_0$. The most local choice would be $\omega_0{=}6$ as $\omega_0{<}6$ fails the admissibility condition of wavelets [@Farge1992]. That is why there is a minimum difference between $s_e$ and $s_b$ for wavelet based analysis. For $\omega_0{=}6$ the minimum averaging scale $s_b(\sigma)$ is $2.28$ times larger than $s_e$.
Field to Flow Angles & Uncertainty
----------------------------------
To compute a field to flow angle $$\theta(t,s_b)=\cos^{-1}\left(\frac{{\mathbf}{b}(t,s_b)\cdot {\mathbf}{v}_\mathrm{sw}}{|{\mathbf}{b}(t,s_b)||{\mathbf}{v}_\mathrm{sw}|}\right),
\label{eq:theta}$$ one can use the local background magnetic field vector ${\mathbf}{b}(t,s_b)$ obtained from equation . This gives the angle between the (local) background magnetic field to the average solar wind velocity ${\mathbf}{v}_\mathrm{sw}$. The second angle which describes the orientation of the local background magnetic field is the azimuth angle, but studies have shown that spectral anisotropy is approximately azimuthally symmetric around the local background magnetic field [@Horbury2008; @Podesta2009]. We therefore only consider $\theta$ to characterize the variability of the orientation of the magnetic field.
The global power spectral density (PSD) at a distinct field to flow angle and with a temporal resolution $\Delta t$ can be obtained from the wavelet coefficients by $$P(f_e; \theta)=\sum_{i=R,T,N} P_i(f_e; \theta),$$ where $$P_i(f_e; \theta)=\frac{2\Delta t}{N}\sum_{j=1}^N\left|W_i(t_j,\sigma;\theta)\right|^2
\label{eq:psd}$$ is computed from $N$ wavelet coefficients $W_i(t_j,\sigma;\theta)$ associated with the angle $\theta(t_j,s_b)$. In our analysis we calculate average $P(f_e; \theta)$ within bins of $\theta = 0{-}10^\circ, 10{-}20^\circ, \dots ,80 {-} 90^\circ$.
We averaged the magnetic field time series according to Equation and calculated scale dependent angles that characterize the orientation of an associated local background magnetic field according to Equation . We now define the *angle uncertainty* $$\delta\theta(t,s_b,s_e)=\theta(t,s_e)-\theta(t,s_b). \label{eq:deltatheta01}$$ This angle quantifies the difference in magnetic field orientation between the eddy of scale $s_e$ and the (larger) scale $s_b$ over which the local magnetic field is defined.
We hypothesize that this newly introduced quantity is an indication for how local an averaged magnetic field is. The root mean square of the angle uncertainty $\delta\theta$ can be used to describe the average uncertainty of the orientation of eddies of size $s_e$ when their orientation is measured with respect to a larger local background magnetic field. A minimum uncertainty arises from the difference between the background magnetic field averaged at scale $s_b$ and the magnetic field averaged with a Gaussian of standard deviation $\sigma_e{=}(2\sqrt{2\ln(2)})^{-1}s_e$, which corresponds to the scale of the eddy fluctuations $s_e$ (shown in Figure \[fig:morlet\] as dotted line). The minimum uncertainty is an inevitable consequence of wavelet analysis as energy density at eddy scale $s_e$ is averaged over the width of the wavelet ${s_b(\sigma)}$. The frequency uncertainty of the wavelet transform as an additional factor in the angle uncertainty is neglected as it is found to be insignificant compared to the difference between $s_e$ and ${s_b(\sigma)}$.
In the following we use the root mean square ${\mathrm{RMS}(\delta\theta)}$ to analyze the influence of the angle uncertainty on plasma turbulence properties, i.e. spectral anisotropy at magnetohydrodynamic (MHD) scales discussed here.
Solar Wind Observations {#ch:results}
=======================
Angle Uncertainty & Spectral Anisotropy
---------------------------------------
![$\theta(t,s_b,s_e)$ (top) and $\delta\theta(t,s_b,s_e)$ (bottom) at frequency $f_e{=}0.1$ Hz for different averaging scales $s_b$ of the local background magnetic field in one hour of Ulysses solar wind data (1995, DOY 100, 01:00:55 to 02:00:55).[]{data-label="fig:timeseries"}](fig3.pdf){width="\linewidth"}
We now investigate the scale dependence of the RMS of the angle uncertainty $\delta\theta$ and how it is related to the ability to resolve spectral anisotropy. For that matter we use 91 days of fast solar wind data with a resolution of 1 s from the Ulysses spacecraft from 1995, days 100-190, during a polar orbit at around 1.4-1.9 AU [@Balogh1992; @McComas2000; @Wicks2010]. Similar and for comparison with @Horbury2008, the mean flow velocity of the solar wind is assumed to be in the radial direction. For the the time intervals used in this study, the deviation between the radial direction and the measured solar wind flow is on average 2 degree and can be neglected. We calculate the angle resolved PSD according to Equation and the angle uncertainty according to Equation for several different scales on which the average magnetic field is calculated ($\alpha{=}1{-}300$).
Figure \[fig:timeseries\] shows an example of $\theta(t,s_b)$ (top) and $\delta\theta(t,s_b,s_e)$ (bottom) at eddy frequency $f_e{=}0.1$ Hz in one hour of the solar wind data. Three different averaging scales $s_b(\alpha\sigma)$ are displayed. For $\alpha{=}1$, $\theta(t,s_b)$ is almost indistinguishable from $\theta(t,s_e)$ and $\delta\theta$ is small (RMS = $3^\circ$). The larger the averaging scale $s_b$ represented by the factor $\alpha$, the larger the values of $\delta\theta$. At a factor $\alpha{=}50$, the angle $\theta$ as a function of time becomes fairly smooth compared to the highly fluctuating $\theta(t,s_e)$. The resultant $\delta\theta$ for $\alpha{=}50$ vary strongly and shows values up to $90^\circ$.
The RMS of the angle uncertainty $\delta\theta$ for the complete data set is shown in Figure \[fig:dtheta\_factor\] as a function of eddy frequency $f_e$. We see that the RMS of the angle uncertainty increases with the width of background magnetic field expressed through the factor alpha. This increase is expected from the sample values shown in Figure \[fig:timeseries\]. ${\mathrm{RMS}(\delta\theta)}$ also increases as the frequency $f_e$ decreases, which may be explained by the power law increase of power towards lower frequencies in the turbulent cascade. The dotted lines mark the frequency range $15$ mHz ${<}f_e{<}$ 100 mHz considered to be the inertial range and used later to estimate the spectral index [@Horbury2008]. In this range ${\mathrm{RMS}(\delta\theta)}$ reaches values between 7–12$^\circ$ for $\alpha{=}5$. For small factors $\alpha$ around 2-3 the RMS of the angle uncertainty is below 10$^\circ$. At very large factors $\alpha{\geq}50$ the RMS is over 25$^\circ$ and averaging scales reach the length of the outer scale (indicated by dashed lines in Figure \[fig:dtheta\_factor\]) estimated to be $L \sim 6\cdot 10^6$ km by @Wicks2010. As the power depending on the angle is sorted into $10^\circ$ bins one might expect that the spectral anisotropy vanishes at factors $\alpha{\geq} 5$, because the RMS of the angle uncertainty grows larger than the angle bin. We will analyze this aspect in detail later in this paper.
![RMS of $\delta\theta$ as a function of eddy frequency $f_e$ at different factors $\alpha{=}1, ..., 300$ (indicated as numbers of lines) of the minimum averaging width for characterizing a local background magnetic field. The areas averaged at scales larger than the outer scales $L{\sim}1.5\cdot 10^6$ km [@Wicks2010] and $v_\mathrm{sw}{\approx}760$ km/s are indicated by dashed lines. The dotted vertical lines represent the boundary frequencies used for spectral index fitting and the colored dots aid comparison with Figures \[fig:psumhisto\] and \[fig:psds\].[]{data-label="fig:dtheta_factor"}](fig4.pdf){width="\linewidth"}
For the spectral index analysis we compute the power spectra $P(f_e;\theta)$ and determine the spectral indices in the range of $15 \mathrm{~mHz~}{<}f{<} 100 \mathrm{~mHz}$ for each $\theta(t,s_b(\alpha\sigma))$ bin. The spectral indices at low (0–10$^\circ$) and high (60–70$^\circ$) angles for each factor $\alpha$ are shown in Figure \[fig:correlation\_spec\_factor\]. Error bars denote the 95% confidence interval of the least squares fit in log-space. We show the spectral index at 60–70$^\circ$, because there are not enough coefficients with angles 80–90$^\circ$ to compute a meaningful average. However, the spectral index is approximately constant for angles $\theta{>}50^\circ$ [@Horbury2008; @Papen2015] and, therefore, the angle bin 60–70$^\circ$ represents the perpendicular case. For $\alpha{=}1$ the analysis is similar to those of @Horbury2008 and @Podesta2009 and the spectral indices are in agreement with the anisotropic scaling predicted by the critical-balance theory [@GS95]. It shows a spectrum $f^{-2}$ parallel and $f^{-5/3}$ perpendicular to the local background magnetic field (see Figure \[fig:correlation\_spec\_factor\], $\alpha{=}1$). We note that the spectral anisotropy is not maximal for $\alpha{=}1$ but for $\alpha{=}10$. However, the difference between the parallel spectral indices corresponding to these factors is small and within error bars. For factors $\alpha{>}10$ the anisotropy slowly decreases. This is also observed in the shorter 31 day data set used by @Horbury2008, which span DOY 100-130 of year 1995 and is shown in light red and grey lines in Figure \[fig:correlation\_spec\_factor\]. The spectral index of the perpendicular cascade is not affected by the scale of the background field. The reason is that for $\alpha{=}1$ the spectral index as a function of the field to flow angle $\theta$ near $0^\circ$ changes very rapidly with growing $\theta$, while it is almost constant at -5/3 for field to flow angles in the range of 30$^\circ$ to 90$^\circ$ (see Figure 2 in @Horbury2008). Additionally, the power at smaller angles is sufficiently smaller compared to the power at larger angles, which thus dominate the spectral contributions (again Figure 2 in @Horbury2008). The factors $\alpha{>}100$ correspond to a local field so large, namely averaged over $s_b{=}4.5$ h at 15 mHz, that it can be regarded as a global background field. Accordingly, the spectral anisotropy is not resolved for a global field. This is to our knowledge the first time that a gradual change of the spectral anisotropy from local to global field has been shown.
![Spectral index $\kappa$ at $\theta{=}0{-}10^\circ$ and $\theta{=}60{-}70^\circ$ as a function of increased averaging width by the factor $\alpha$ . Error bars show the 95% confidence interval of the least squares fit to $P(f_e;\theta)$. Light-colored lines indicate spectral indices from 31 day data (DOY 100-130) of Ulysses data [@Horbury2008]. We show angles between $\theta{=}60{-}70^\circ$ as a meaningful average for $\theta{=}80{-}90^\circ$ was not available for large averaging widths.[]{data-label="fig:correlation_spec_factor"}](fig5.pdf){width="\linewidth"}
The robustness of the spectral index for factors $5{<}\alpha{<}20$ is unexpected as the previously introduced RMS of the angle uncertainty suggests that anisotropy might not be resolved for factors $\alpha{\geq} 5$, since a strong variability of the spectral index for $\theta{<}30{-}40^\circ$ is observed and a resolution of $10^\circ$ is necessary (see Figure 2, lower panel, in @Horbury2008). To explain this discrepancy between the observed spectral anisotropy and the observed RMS of the angle uncertainty we now study how accurately the wavelet coefficients are associated with the angle bins under consideration.
Origin of Power in the Parallel Angle Bin
-----------------------------------------
In Figure \[fig:dtheta\_factor\] we have shown that large scales of the local background field lead to large angle uncertainties. This can be interpreted in the sense that a large scale local background magnetic field, i.e. a background magnetic field with factors $\alpha\geq5$, is not an adequate representation for the orientation of the turbulent fluctuations. Mathematically, this means that wavelet coefficients $W(t,f_e,\theta)$ are not assigned to the correct angle.
![Origin $\theta$ of power for wavelet coefficients originally within the $0{-}10^\circ$ bin for $\alpha{=}1$ at 15 mHz (top) and 100 mHz (bottom). Percentages of the power are with respect to the total power in the $0{-}10^\circ$ bin at $\alpha{=}1$. For larger factors $\alpha{>}1$ the power additionaly comes from larger field to flow angles. It can be seen that *false coefficients* first contribute to low frequencies. For $\alpha{=}10$, e.g., the maximum power comes from $20^\circ$ at 15 mHz but only from $10^\circ$ at 100 mHz.[]{data-label="fig:psumhisto"}](fig6.pdf){width="\linewidth"}
We now investigate if and how many wavelet coefficients associated with larger angles $\theta(\alpha{=}1){>}10^\circ$ at the most local scale are falsely assigned to the angle bin $0{-}10^\circ$ when the local background field is large ($\alpha{=}2{-}300$). In the following we refer to wavelet coefficients originating from higher angles and being assigned to the $0{-}10^\circ$ bin as *false* coefficients.
We compute the angular origin $\theta(\alpha{=}1)$ of larger scale ($\alpha{>}1$) coefficients within the $0{-}10^\circ$ bin for the upper (100 mHz) and lower (15 mHz) frequency boundary of the fit range. The result tells us how many *false* coefficients contribute to the power $\sum |W \left( \theta(\alpha) \right) |^2$ within the $0^\circ{\leq}\theta(\alpha){<}10^\circ$ bin and is shown in Figure \[fig:psumhisto\]. To aid visualization of the redistribution of the angle bins for growing $\alpha$ we choose a bin resolution of $2^\circ$. The histograms are normalized to the total power in the 0–10$^\circ$ bin at $\alpha{=}1$. It can be seen that if we use larger and larger averaging widths, the power spectra include increasingly more false coefficients with angles originally outside the $0{-}10^\circ$ bin. For $\alpha{=}10$ we observe that for 15 mHz the maximum of coefficients actually stems from angles around $20^\circ$ and thus the power has large contributions from *false coefficients*. The corresponding RMS of the angle uncertainties associated with this averaging width $\alpha{=}10$ are RMS$(\delta\theta){>}15^\circ$ (shown in Figure \[fig:dtheta\_factor\]).
To understand the influence of the origin of the power presented in Figure \[fig:psumhisto\] on the slope of the power spectra, we compute the PSD for several averaging widths in Figure \[fig:psds\]. For $\alpha{=}2$ the contribution of power from larger angles is low and, therefore, the spectral energy distribution $P(f;\theta{=}0{-}10^\circ)$ at $\alpha{=}2$ is almost identical to $P(f;\theta{=}0{-}10^\circ)$ at $\alpha{=}1$ (see Fig. \[fig:psds\], green and blue). For $\alpha{=}50$ most of the coefficients are falsely associated with contributions from angles of $20^\circ$ and $30^\circ$ for 15 mHz and 100 mHz, respectively. This shows how large averaging widths smooth out local small scale variations and may thus lead to false angle association. However, as power from larger angles at increasing factors does not contribute equally to the $0{-}10^\circ$ bin for 15 mHz and 100 mHz, the slope appears to be similar. The amount of power of larger angles associated with the $\theta{=}0{-}10^\circ$ bin at $\alpha{=}10,50$ (Fig. \[fig:psumhisto\], red and cyan) for 15 mHz is much larger than for 100 mHz. From this follows that the spectral index can still be as steep as -2 and even be steeper than at $\alpha{=}1$, but the total power clearly increased. Due to this unequal contribution of power of larger angles at different frequencies, the spectral index can remain as steep as -2 to factors of $\alpha{\approx}20{-}50$ even though power is added to the parallel spectrum. For an increased averaging width by a factor of $\alpha{\geq}50$ the spectral index still shows anisotropy, but is more shallow than -2. For $\alpha{\geq}200$ the parallel spectral index is -5/3 and no spectral anisotropy can be resolved any more (see Figure \[fig:correlation\_spec\_factor\]). Despite the fact that $\kappa$ stays around -2 for $\alpha$ up to 50, the magnetic field averaged at factors $\alpha{>}5$ should not be considered an appropriate local background magnetic field, as false coefficients contribute to the power.
![PSD and spectral index $\kappa$ at $\theta{=}0{-}10^\circ$ at the factors 1, 2, 10, 50 and 300 in blue, green, red, cyan and magenta, respectively. Only 15 mHz and 100 mHz frequencies are shown to aid comparison to Figure \[fig:psumhisto\] and visualization of the slopes.[]{data-label="fig:psds"}](fig7.pdf){width="\linewidth"}
Discussion & Conclusions
========================
We have introduced the angle uncertainty between the orientation of the averaged magnetic field and the orientation of the eddy fluctuations as a measure to describe the uncertainty of the orientation of a (local) background magnetic field and thus as a measure to resolve anisotropy of turbulent properties. We studied the scale dependent angle uncertainty and how it is related to the resolution of spectral anisotropy in fast solar wind data.
The RMS of the angle uncertainty ${\mathrm{RMS}(\delta\theta)}$ depends on the frequency/eddy size of the fluctuation. A finite RMS implies the existence of a basic, frequency dependent, uncertainty to resolve anisotropy. We investigated previously observed anisotropy with a resolution of 10$^\circ$ [@Horbury2008]. Only if the RMS of the angle uncertainty is lower than 10$^\circ$ the correct association of magnetic field orientation to the wavelet coefficients can be assured. Based on the results presented in Figure \[fig:dtheta\_factor\], such a correct association is obtained for averaging widths $\alpha\leq4$, which corresponds to $s_b{\leq}10s_e$. It is apparent that the definition of a local background magnetic field depends on the frequency range or eddy size under consideration and on the anisotropy to be resolved. Anisotropy sensitive to changes below 10$^\circ$ would require smaller averaging scales of the magnetic field to resolve such an anisotropy in case the distribution of scale dependent energy is similar to case studies here.
The solar wind observations presented in this work show that observed spectral anisotropy is not adequately resolved any more for $\alpha{\geq}50$ and vanishes for $\alpha{\geq}200$, i.e. an averaging width of more than 200 times larger than the eddy scale. Although the RMS of the angle uncertainty at factors $\alpha{\geq}5$ is larger than the width of the angle bin the spectral index remains anisotropic. This unexpected apparent robustness of the spectral anisotropy with respect to increased averaging width can be explained by a frequency dependent gain of power from wavelet coefficients of higher angles. Higher frequencies gain less power from wavelet coefficients associated with higher angles than lower frequencies (see Figure \[fig:psumhisto\]). The origin of the power, meaning the angles associated with the power when averaging with $\alpha{=}1$, at different frequencies is in agreement with the frequency dependent RMS of the angle uncertainty. Even though the total power at small angles clearly increases for increasing factors $\alpha$ (see Figure \[fig:psds\]), the slope of the PSD remains steep even for very large averaging widths. The apparent robustness to the increased averaging width should therefore not lead to an incorrect conclusion on the size of a local background magnetic field. The RMS of the angle uncertainty predicts the error in the association of wavelet coefficients and for $\alpha{\geq}5$ the total power in the parallel spectrum clearly increases. Only the anisotropy in the spectral index appears to be intact due to the effect of frequency dependent power gain.
Within the studied data set, a magnetic field averaged at $\alpha{\approx}5{-}50$, may be regarded as an intermediate background magnetic field, not local nor global. For an intermediate background magnetic field, although scale dependently averaged, the RMS of the angle uncertainty ${\mathrm{RMS}(\delta\theta)}$ within the frequency range under consideration is larger than the accuracy needed (10$^\circ$) and wavelet coefficients may not be linked to the correct angle.
At even larger averaging scales corresponding to $\alpha{\geq}50$, the background magnetic field approaches the global mean magnetic field. This is often referred to as a magnetic field averaged scale independently [@Oughton2015]. Even if averaged scale dependently, a magnetic field on the order of the outer scale $L{\sim}1.5\cdot 10^6$ km may be regarded as the global mean magnetic field [@Wicks2010]. The RMS of the angle uncertainty ${\mathrm{RMS}(\delta\theta)}$ within the analyzed frequency range, exceeds 25$^\circ$ for averages over scales larger than the outer scale (see Figure \[fig:dtheta\_factor\], dashed lines). In this case a significant amount of power cannot be linked correctly to a field to flow angle bin.
For comparison between local and global background magnetic field we also analyzed the data using a scale-independent background field (not shown). Here, the power of eddies at different scales is associated with the same background magnetic field. We were unable to observe spectra which scale with $f^{-2}$ parallel to the background magnetic field using such a global frame. The RMS of the angle uncertainty of such a scale-independent magnetic field also depends on the frequency under consideration. When averaging with a window width of 5 times the largest period ($5\cdot67$ s) of the frequencies within which spectral indices are calculated, the RMS of the angle uncertainty ranges from 7$^\circ$ at low frequencies to 10$^\circ$ at high frequencies (see Figure \[fig:dtheta\_factor\] at $\alpha{\approx}2$ for 15 mHz and $\alpha{\approx} 13$ for 100 mHz). Eddies of lower frequency with periods closer to the averaging scale might be represented well enough by such a background magnetic field. However, as higher frequencies are analyzed with the same background magnetic field, the power of these eddies is associated with an angle resulting from larger scale fluctuations. Consequently more power from *false* coefficients contribute to higher frequency fluctuations whereas very few (or none) contribute to the lower frequency fluctuations. Following this, the effect that the spectral index can remain steep, observed for scale dependent magnetic fields, does not hold for the global approach. Note, in the limit of a very strong background magnetic field, where $|{\mathbf}{B}_0|\gg |{\mathbf}{\delta b}|$, the values of $\delta\theta$ decrease and may have a negligible frequency dependency and the global approach will be applicable.
The angle uncertainty presented here provides an uncertainty measure of the orientation of turbulent structures/eddies as function of scale of the averaged magnetic field and of the associated frequency/period of the eddies under investigation. This method, however, does not constrain the scale of the wave numbers of the eddies in direction parallel and perpendicular to the associated background magnetic field. The reason is that under the assumption of Taylor’s hypothesis [@Taylor1938] and spatial and temporal stationarity of the magnetic field, wave vectors of different magnitude and orientation contribute to the spectral energy density at one frequency $f_e$ (e.g., @Fredricks1976 [@Papen2015]).
The angle uncertainty is physically controlled by two effects: a) the frequency/scale dependent amplitudes of the turbulent fluctuations and b) any non-turbulent contributions such as the magnetic field convected out from the solar corona. These contributions generate the total field, which controls the orientation of the turbulent eddies. Thus the contribution of the solar background field [@Parker1958] with respect to the amplitude of the fluctuations plays an important role for the angle uncertainty. For example, for magnetic field fluctuations ${\mathbf}{\delta b}$ much smaller than the amplitude of the global mean magnetic field ${\mathbf}{B}_0$, the angle uncertainty would tend to small values and anisotropy should be well resolved. The ability to resolve anisotropy as a function of scale is not universally equal, but depends on the turbulent system, for example on the values of the spectral slopes of the energy distribution. The angle uncertainty is a helpful measure that can be applied to various systems to evaluate the ability to resolve anisotropy with a certain degree.
natexlab\#1[\#1]{}\[1\][[\#1](#1)]{}
Alexandrova, O., Lacombe, C., & Mangeney, A. 2008, AnG, 26, 3585
Balogh, A., Beek, T., Forsyth, R., [et al.]{} 1992, A&AS, 92, 221
Beresnyak, A., & Lazarian, A. 2008, ApJ, 682, 1070
Beresnyak, A., & Lazarian, A. 2009, ApJ, 702, 460
Bieber, J. W., Wanner, W., & Matthaeus, W. H. 1996, JGRA, 101, 2511
Boldyrev, S. 2006, PhRvL, 96, 115002
Boldyrev, S., & Perez, J. C. 2012, ApJL, 758, L44
Bruno, R., & Telloni, D. 2015, ApJL, 811, L17
Chen, C. H. K., Mallet, A., Yousef, T. A., Schekochihin, A. A., & Horbury, T. S. 2011, MNRAS, 415, 3219
Chen, C. H. K., Wicks, R. T., Horbury, T. S., & Schekochihin, A. A. 2010, ApJL, 711, L79
Cho, J., & Lazarian, A. 2004, ApJ, 615, L41
Cho, J., & Lazarian, A. 2009, ApJ, 701, 236
Cho, J., Lazarian, A., & Vishniac, E. T. 2002, ApJ, 564, 291
Cho, J., & Vishniac, E. T. 2000, ApJ, 539, 273
Farge, M. 1992, AnRFM, 24, 395
Fredricks, R. W., & Coroniti, F. V. 1976, JGR, 81, 5591
Galtier, S. 2006, JPlPh, 72, 721
Galtier, S., Pouquet, A., & Mangeney, A. 2005, PhPl, 12, 092310
, P., & [Sridhar]{}, S. 1995, ApJ, 438, 763
He, J., Tu, C., Marsch, E., Bourouaine, S., & Pei, Z. 2013, ApJ, 773, 72
Horbury, T. S., Forman, M., & Oughton, S. 2008, PhRvL, 101, 175005
Horbury, T. S., Wicks, R. T., & Chen, C. H. K. 2012, SSRv, 172, 325
Howes, G. G., Cowley, S. C., Dorland, W., [et al.]{} 2008, JGRA, 113, 1
Howes, G. G., TenBarge, J. M., & Dorland, W. 2011, Physics of Plasmas, 18, 102305
Luo, Q. Y., & Wu, D. J. 2010, ApJ, 714, L138
Maron, J., & Goldreich, P. 2001, ApJ, 554, 1175
Matthaeus, W. H., Goldstein, M. L., & Roberts, D. A. 1990, JGRA, 95, 20673
Matthaeus, W. H., Servidio, S., Dmitruk, P., [et al.]{} 2012, ApJ, 750, 103
McComas, D. J., Barraclough, B. L., Funsten, H. O., [et al.]{} 2000, JGRA, 105, 10419
Meyers, S. D., Kelly, B. G., & O’Brien, J. J. 1993, MWRv, 121, 2858
Narita, Y. 2015, AnG, 33, 1413
Oughton, S., Matthaeus, W. H., Wan, M., & Osman, K. T. 2015, RSPTA, 373, 20140152
Parker, E. N. 1958, ApJ, 128, 664
Podesta, J. J. 2009, ApJ, 698, 986
Podesta, J. J. 2013, SoPh, 286, 529
Salem, C. S., Howes, G. G., Sundkvist, D., [et al.]{} 2012, ApJL, 745, L9
Saur, J., & Bieber, J. W. 1999, JGRA, 104, 9975
Taylor, G. 1938, RSPSA, 164, 476
Tessein, J. a., Smith, C. W., MacBride, B. T., [et al.]{} 2009, ApJ, 692, 684
Torrence, C., & Compo, G. P. 1998, BAMS, 79, 61
von Papen, M., & Saur, J. 2015, ApJ, 806, 116
Wicks, R. T., Horbury, T. S., Chen, C. H. K., & Schekochihin, A. A. 2010, MNRAS Letters, 407, L31
Wicks, R. T., Horbury, T. S., Chen, C. H. K., & Schekochihin, A. A. 2011, PhRvL, 106, 045001
[^1]: The unit vector ${\mathbf}{e}_R$ points radially away from the sun, ${\mathbf}{e}_T{=}{\mathbf}{e}_\Omega\times{\mathbf}{e}_R$ is perpendicular to ${\mathbf}{e}_R$ and the sun’s rotational axis ${\mathbf}{e}_\Omega$ and ${\mathbf}{e}_N{=}{\mathbf}{e}_R\times{\mathbf}{e}_T$ completes the right handed system.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'There are at least three different notions of degrees of freedom (DF) that are important in comparison of quantum and classical dynamical systems. One is related to the type of dynamical equations and inequivalent initial conditions, the other to the structure of the system and the third to the properties of dynamical orbits. In this paper, definitions and comparison in classical and quantum systems of the tree types of DF are formulated and discussed. In particular, we concentrate on comparison of the number of the so called dynamical DF in a quantum system and its classical model. The comparison involves analyzes of relations between integrability of the classical model, dynamical symmetry and separability of the quantum and the corresponding classical systems and dynamical generation of appropriately defined quantumness. The analyzes is conducted using illustrative typical systems. A conjecture summarizing the observed relation between generation of quantumness by the quantum dynamics and dynamical properties of the classical model is formulated.'
author:
- |
N. Buri' c[^1]\
Institute of Physics, University of Belgrade,\
PO BOX 68, 11000 Belgrade, Serbia.
title: Relations between different notions of degrees of freedom of a quantum system and its classical model
---
PACS: 03.65. Yz, 05.45.Mt
Introduction
============
Relation between quantum and the corresponding classical systems is indeed a complex one [@Landsman]. The differences between the two theories have induced major changes in our understanding of Reality [@dEsp] and have been used in major technological developments [@tech]. Perhaps equally interesting are various similarities or analogies between quantum and classical theories [@corr]. Such notions, with the apparently similar meaning and role in quantum and in classical mechanics, are notions of number and type of degrees of freedom that are relevant for the description of a system’s structure. However, there are at least three different concepts that can be justifiably called degrees of freedom. One of these is rather formal and mathematical since it is determined by the mathematical nature of the relevant evolution equation. Since the type of evolution equations of quantum mechanical systems, i.e. partial differential equation, is radically different from that of the corresponding classical mechanical systems, i.e. ordinary differential equations, the number of the degrees of freedom in this sense is also different. The second definition concentrates on an abstract definition of the system’s structure as formalized by the Lie algebra of distinguished dynamical variables. The structure of a quantum system and the corresponding classical system are characterized by the same Lie algebra, and the two system have the same number and type of the correspondingly defined degrees of freedom. The third notion of the number of degrees of freedom that we shall introduced and call the dynamical degrees of freedom, is fundamentally related to the dynamics displayed by a particular system. We shall see that the relation between the number of degrees of freedom in this third sense for a quantum and the corresponding classical system is rather nontrivial. In order to study this relation we shall need to use an appropriate notion of quantumness of a quantum state, and to compare dynamical changes of the quantumness with the properties of dynamics of the corresponding classical model. To this end, we shall numerically analyze quantum and classical dynamics for the examples of two typical systems. This analyzes will lead us to formulate our major conjecture concerning the relation between quantumness generation, dynamical properties and the numbers of dynamical degrees of freedom of a quantum system and its classical model.
In the next section we shall discuss the three notions of degrees of freedom. Algebraic definition of the structural degrees of freedom was introduced for quantum systems in references [@Cheng1; @Cheng2; @Cheng3]. We have slightly generalized the presentation so that it automatically applies on classical Hamiltonian systems as well as on the quantum systems. In this section we introduce and stress the importance of the third type, so called the dynamical degrees of freedom. The quantumness of states with respect to the selected set of the basic variables is introduced in section 3, and is similar in spirit with the notion of generalized entanglement presented in [@Viola1; @Viola2; @Viola3], and similarly in [@Klyasko] and [@Zanardi]. In section 4. we present the notion of a classical model as a constrained quantum system such that the quantumness is preserved minimal during the evolution. This way of defining the classical models was introduced in [@usPRA1] for a system of oscillators and in [@usPRA2] for spin systems. Section 5 treats examples in order to infer possible relations between the numbers of dynamical degrees of freedom for a quantum system and its classical model. The results suggest our major conjecture formulated in section 5.2. Here presented view of the general relation between the number of dynamical DF for quantum and the corresponding classical system and the dynamics of classical models of quantum systems and that of the generalized quantumness has not been discussed before. Summary is given in section 6.
Three types of degrees of freedom
=================================
We shall argue that it makes sense to introduce at least three different types of degrees of freedom corresponding to different intuitions about what a notion of degrees of freedom should mean. The three types will be called: a) formal DF (FDF); b) structural or algebraic DF (SDF) and c) dynamical DF (DDF). The first corresponds to the dimensionality of the manifold of inequivalent initial conditions and is thus related to the type of dynamical equations. The second is related to the structural properties of the system and is formulated using the [Lie]{} algebraic relations between the basic physical variables of the system. In fact, SDF become specified as a crucial part of the definition of the system under consideration. The Hamiltonian is also involved in this considerations, but without any reference to the system’s dynamics. In fact it is only the kinematical part of the Hamiltonian that is relevant for SDF. The third notion, which we call dynamical DF (DDF), is related to the dimension of typical invariant manifold of the dynamics. The most commonly used notion of degrees of freedom is perhaps closest to the notion of SDF. Together with FDF, SDF and DDF we shall use the notation NFDF, NSDF and NDDF to denote the numbers of the corresponding DF.
Formal degrees of freedom
--------------------------
Let us first define the notion of FDF. The [*type*]{} of a dynamical system under consideration is defined by the set of evolution or dynamical equations of first order in the evolution parameter. The Cauchy problem for different types of dynamical equations requires different phase spaces or manifolds of initial conditions. Hamiltonian mechanical system is formulated on a 2N-dimensional real manifold, evolution equation of a field $\psi(x), x\in {\mathbf R}$, like for example the Schr" odinger equation, is formulated on an infinite space of functions, and the evolution of an $N$ level quantum systems is formulated on an $N$-dimensional complex space.
[*Definition of FDF: The number of FDF of a dynamical system is defined as half the number of the real dimensions of the manifold of physically inequivalent initial conditions for the relevant evolution equations.*]{}
For a Hamiltonian system with a $2N$-dimensional real phase space ($N$ might be $\infty$ as for fields), and this also covers the quantum unitary evolution (please see the [**Remark 1**]{}) the number of FDF is $N$.
The specification of “physically inequivalent initial conditions” refers to the general theory and depends on the type of dynamical equations or in other words on the constraints imposed by the type of dynamics and does not refer to a particular system. For example, the unitary quantum evolution on a Hilbert space ${\cal H}$ preserves the norm of vectors $|\psi\rangle\in {\cal H}$ and is invariant under global phase change. These facts could, but need not, be explicitly incorporated into the appropriate dynamical equations. In any case, the manifold of inequivalent initial conditions is actually the projective space $P{\cal H}$.
This type of constraints should be clearly distinguished from those that are imposed as specific to a particular system. The latter are in fact a part of the specification of the system under consideration. Such constraints need not, but usually are, incorporated into an appropriate reformulation of the dynamical equations. An example of such constraints is the constraint of the constant length of a pendulum in a constant gravitational field, which, if taken explicitly into the account, reduces the dimensionality of the manifold of initial conditions from 6 to 2. However, as pointed out, these types of constraints are in fact a part of the definition of what is the system under consideration and are treated in the next section.
[**Remark 1**]{} Evolution of a quantum system on an $N$-dimensional Hilbert space ${\cal H}^N$ is equivalent to a linear symplectic flow on a symplectic manifold $P{\cal H}^{N-1}$ [@Ashtekar; @Brody; @Marmo]. Any such linear system as $N-1$ integrals of motion and is completely integrable. This can also be seen as a consequence of the fact that the group of automorphisms of such quantum system is $U(N)$, so that any pair of pure states can be connected by some unitary transformation. Any Hamiltonian, i.e. a Hermitian operator can be diagonalized using some $U(N)$ transformation.
Obviously, the notion of FDF is rather formal, and does not correspond to the one commonly used in Physics. The later is captured by consideration of the system’s structure and is precisely expressed using the relevant dynamical algebra.
Structural degrees of freedom
-----------------------------
The notion of structural degrees of freedom is inspired by intuitively justified counting of independent elementary motions in compound classical Hamiltonian systems of mechanical type. A 1D particle is naturally associated one degree of freedom. If there are $N$ 1D particles with the energy $H_i(q_i,p_i)=p_i^2/2m_i$ each, then a collection of such systems has the energy $H=\sum_i^N p_i^2/2m_i+V_{int}(q_1,q_2\dots)$, where the interaction is described by a function depending on some of the coordinates. Independently of the actual dynamics, the number of degrees of freedom of such a system is taken to be $N$. A quantum system obtained by quantization of such classical system is assumed to have the same number of degrees of freedom as the classical one. Let us stress that this notion of the number of degrees of freedom is not related to the inequivalent initial conditions for the dynamical equations, nor is related to the actual motion of the system. Rather, it is related to independent possible motions the system. A straightforward definition of the number of the structural degrees of freedom for the systems with energy of the above form might be given as the number of momenta that appear quadratically in the energy. However, many realistic models of quantum as well as classical systems have the energy expressed in terms of natural dynamical variables which is not of the above form. Examples of quantum systems are: the systems of spins, identical bosons or fermions etc..., and examples of classical systems are different ridged bodies and degenerate completely integrable systems written in the action-angle variables. Obviously, a more general definition of structural degrees of freedom is needed in order to apply the elementary intuition onto more complex systems.
[*Dynamical algebra of relevant dynamical variables*]{}
A particular physical system is specified, and thus distinguished from an abstract general framework (such as “Hamiltonian dynamical systems”, or “Unitary evolution”), by describing what can be measured on it, i.e. by specifying the set of basic dynamical variables, and by expressing the interactions within the system in terms of these dynamical variables. In other words, the class of physically relevant dynamical variables should be described and the evolution equations (the Hamiltonian) should be expressed in terms of these dynamical variables.
Structure of the physical system is mathematically described by relations between the basic dynamical variables which form a Lie algebra $g$. Each of the basic dynamical variables is associated with an elementary change of the system’s state, and successive independent changes are related by the Lie algebra structure of the set of basic dynamical variables. In classical mechanics the Lie algebra is realized by functions on a symplectic manifold and the Poisson bracket, and in quantum mechanics by operators on a Hilbert space and the commutator. In either case, by definition of what the dynamical algebra is, the manifold of the system’s states is such that the action of the corresponding Lie group $G$ generated by the Lie algebra $g$, is transitive. The Lie algebra defined in this way is traditionally called the systems dynamical algebra. In our context, more appropriate name would be the structural algebra, but we shall use the traditional terminology.
Description of a dynamical system amounts first to the specification of its dynamical algebra g and its manifold of states (providing an irreducible realization of $G$ as a group of transformations). Once this is done, the generator of the evolution, i.e. the Hamiltonian, is specified as an expression (possibly nonlinear) in terms of the basic dynamical variables belonging to g. Thus, in general the Hamiltonian is not necessarily an element of $q$.
In what follows the dynamical algebra $g$ will always be a Lie algebra, with rank $l$ and dimension $n$. The Lie group generated by the dynamical algebra will be denoted by $G$.
[*Definition of the structural degrees of freedom* ]{}
The following considerations restricted onto the quantum systems with unitary irreducible representation of the dynamical algebra have been provided in [@Cheng1; @Cheng2; @Cheng3]. We present the definition of SDF first on the purely algebraic level, with no specification of the system quantum or classical nature. We than indicate how the classical and quantum systems with a given dynamical algebra are constructed.
Consider a system (classical or quantum) with a given dynamical algebra $g$ of distinguished variables. The notion of SDF is defined in the same way for classical and for quantum systems, and is based only on the properties of the dynamical algebra. The dynamical algebra $g$ has $\gamma$ different chains of subalgebras: $g\supset g^l_{s^l}\supset g^{l}_{s^l-1}\dots \supset
g^{l}_{1},\>l=1,2,\dots \gamma$. Casimir elements of $g$ and all the algebras in (any of) the subalgebra chain form the relevant complete set of dynamical variables (CSDV) $Q_j, j=1,2\dots d$. There is $d=l+(n-l)/2$ of these, independently of the subalgebra chain. Some of these Casimir elements are fully degenerate.
[*Definition of SDF*]{} The number of non-fully degenerate elements in CSDV is $m\leq
(n-l)/2$ and is by definition the number of SDF. The non-fully degenerate Casimir elements in a particular chain are the dynamical variables that define $m$ SDF. The number of SDF is chain independent, although the elements that define SDF are chain dependent.
The previous definition of SDF is given in algebraic terms with no reference to the classical or quantum nature of the system. It remains to construct classical and quantum systems corresponding to the given dynamical algebra g, which would thus have the corresponding SDF. This is done using the KKS theory of orbits of the co-adjoined action of g [@KKS]. Given a dynamical Lie algebra g (from a certain class) one can show that orbits $O_g$ of the co-adjoined action of g are symplectic manifolds, which can be used as phase space ${\cal M}\equiv O_g$ of Hamiltonian dynamical systems with g as the dynamical algebra, and the corresponding SDF. To construct the quantum system with g as the dynamical algebra one can directly use the theory of unitary irreducible representations of g, or quantize the above symplectic manifolds $O_g$. Once the system phase space or the system Hilbert space are constructed, the system is finally specified by giving a Hamiltonian as an expression, possibly nonlinear, in terms of representatives of the elements of g. Observe that SDF are specified with no reference to a particular Hamiltonian, so that SDF are not related to the properties of the paricular system. Furthermore, ordering ambiguity, which occurs in the quantization of Hamiltonians, effects in no way the number and the type of SDF.
[**Remark**]{} In some particular realizations of an algebra the non-fully degenerate Casimir elements might not be independent. The physical system corresponding to this realization has the number of SDF smaller than it is for the abstract algebra (Please see section 5.1.4).
[**Remark: Composite quantum systems**]{} If the dynamical algebra $g$ of a quantum system $C$ can be represented as a direct sum of dynamical algebras of two systems $A$ and $B$, that is $g^C= g^A\oplus g^B$, then the tensor product of irreps of $G^A$ and $G^B$ is an irrep space of $G^C$, that is ${\cal H}^C={\cal H}^A\otimes {\cal H}^B$. If $l_{A,B}$ and $n_{A,B}$ are the ranks and dimensions of $g^A$ and $g^B$, then in general the number of SDF of $C$ is $m_C=m_A+m_B$. Thus, in the case $g^C= g^A\oplus g^B$ the system $C$ can be represented as a union of two systems and the number of SDF is additive. Of course, there are systems whose dynamical algebra is not decomposable in the above form, but which, nevertheless, have more than one SDF (Please see section 5.1.4). Quite in general, if the dynamical algebra $g$ of the system is semi-simple then it can be uniquely expressed as a direct sum of mutually commutative and orthogonal simple algebras: $g=\oplus_k g_k$ and the Hilbert space which is an irrep space of $g$ factors as $H=\otimes_k H_k$. Thus, in the case of semi-simple dynamical algebra the number of SDF is additive, but the number of SDF in all the factor systems with $g_k$ dynamical algebras need not be unity for each $g_k$. Analogous statements apply to classical systems with semi-simple Lie dynamical algebra.
[**Examples**]{} Following examples will also be used later. The first example is relevant for classical as well as quantum systems. The second example is commonly associated with genuinely quantum systems, but nevertheless has perfectly well defined realization as a (classical ) Hamiltonian dynamical system.
2\) A system with Heisenberg-Weyl $h_4$ dynamical algebra: The algebra is commonly given by the relation between its four basis elements $\{a^{\dag},a,n,I\}$ $$\begin{aligned}
&&[a,a^{\dag}] = I,\quad [n,a^{\dag}]=a^{\dag},\quad [n,a]=-a,\nonumber\\
&&[a^{\dag},I] = 0,\quad [a,I]=0,\quad [n,I]=0.\end{aligned}$$
$h_4$ has $rank=2$, and is not semisimple algebra. The algebra chain $h_4\supset u(1)\otimes u(1)$ determines the SDF. The Casimir elements of $h_4$ and $u(1)$ are proportional to unity, so there is only one SDF. In the classical case the phase space is $R^2$ with symplectic coordinates $(q,p)$. The Casimir element corresponding to one SDF is in fact the action variable of the harmonic oscillator $J=p^2+q^2$. In the quantum case the algebra $h_4$ is uniquely represented by multiplication and differentiation operators on $L_2({\bf R})$, which are related to the basis elements $a,a^{\dag}$ by $ a=(x+i\partial/\partial x)/\sqrt {2\hbar}, \>a^{\dag}=(x-i\partial/\partial x)\sqrt {2\hbar}$. The Casimir element corresponding to the one SDF is the number operator $\hat n$.
A classical mechanical systems, composed of two $h_4$ subsystems, has the dynamical algebra $h^1_{4}\otimes h^2_{4}$ and the canonical variables $p_{x_1},p_{x_2},x_1,x_2$ satisfying $$\{p_{x_i},x_j\}=\delta_{ij},\> \{x_i,x_j\}=0,\> \{p_{x_i},p_{x_j}\}=0,\quad i,j=1,2,$$ where $\{,\}$ denotes the Poisson bracket, and the indexes $i,j=1,2$ denote the first and the second system. The basic canonical variables are related to the algebra basis elements by $a_j=(x_j+ip_{x_j})/\sqrt {2}, \> a_j^{\dag}=(x_j-ip_{x_j})\sqrt {2},\>j=1,2,\>i=\sqrt {-1}$. The system has two SDF corresponding to the subalgebra $h^1_4\otimes h^2_4=u^1(1)\otimes u^1(1)\otimes u^2(1)\otimes u^2(1)$. The same applies to the corresponding quantum version.
Hamiltonian of two possibly interacting oscillators is expressed in terms of $p_{x_i},x_i,\>i=1,2$ as $H=p_{x_1}^2/2m_{1}+p_{x_2}^2/2m_2+V(x_1,x_2)$. In the case that there is no interaction $V=0$ the Hamiltonian is, in the classical as well as the quantum case, a linear expression of the algebra generators $H\sim\sum_{i=1,2}\omega_i n_i$. The variables corresponding to the chosen SDF, i.e. the algebra generators $n_i,\>=1,2$ are constants of motion for such Hamiltonian dynamics. In other words, the subalgebra that is used to define the SDF is in fact a dynamical symmetry of the system. This dynamical symmetry is usually broken if there is interaction between the variables corresponding to different SDF. However, if the interaction between the harmonic oscillators is of the form $V(x_1,x_2,p_{x_1},p_{x_2})=(x^2+p_{x_1}^2)(x_2^2+p_{x_2}^2)=n_1n_2$ then the subgroup generated by the subalgebra used to define the SDF is a dynamical symmetry.
2\) A system with $su(2)$ dynamical algebra: The algebra is generated by the Pauli spin matrices $\{I,\sigma_1,\sigma_2,\sigma_3\}$ that satisfy the relations $$[\sigma_k,\sigma_l]=2i\sum_m\epsilon_{k,l,m}\sigma_m,\quad [\sigma_{l},I]=0\quad k,l,m=1,2,3.$$ The dimension and the rank of $su(2)$ are $d=3,l=1$, so that the number of SDF of a system with this dynamical algebra is one. The subalgebra chain $su(2)\supset u(1)$ determines the SDF, given by the non-degenerate Casimir $\sigma_3$.
A pair of spins has the dynamical algebra $su^1(2)\oplus su^2(2)$. The number od SDF is two. The SDF could correspond to the subalgebra chain $su^1(2)\oplus su^2(2)\supset u^1(1)\oplus u^2(1)$ with non-degenerate Casimirs $\sigma^1_3$ and $\sigma^2_3$. Wether the subalgebra corresponding to the two SDF generates a dynamical symmetry depends on the Hamiltonian and the particular form of the interaction between the two spins. This is the case if the interaction involves only the subalgebra generators $\sigma^1_3$ and $\sigma^2_3$ of the form $\sigma^1_3\otimes \sigma^2_3$ (Please see section 5.1.2).
Dynamical degrees of freedom
-----------------------------
The notion of dynamical degrees of freedom is central in our work. Although the formal definition is quite simple, actual determination of the number of DDF, without previous solution of the dynamical problem, seams to be impossible. However, it is worth giving a precise definition and attempt to use it in characterizing the relation between the properties of quantum and classical dynamics. We have seen that the Hamiltonian need not be an element of the dynamical algebra i.e. might be given by a nonlinear expression in terms of the basic dynamical variables. In this case, the coadjoined orbits $O_g$ of the dynamical algebra $g$, discussed after the definition of SDF, need not be invariant under under the Hamiltonian dynamics. Manifolds which are dynamically invariant and irreducible might be of dimension larger than the number $N$ of SDF. For example, if the Hamilton’s function is the only constant of motion then a generic irreducible invariant manifold is $2N-1$ dimensional manifold of constant energy. Similarly, an irreducible invariant manifold of a composite quantum system might be of larger dimension than the total number of additive SDF. This happens for example if the interaction generates entanglement between the components. It makes sense to analyze, as an important dynamical property of the systems, the dimension of its generic invariant manifold.
[*Definition of DDF*]{} The number of DDF of a quantum or classical system with the SDF defined as above is by definition the dimension of its generic dynamically invariant and irreducible manifold.
In classical mechanics the generic dynamically invariant and irreducible manifold is understood as the manifold in which a generic orbit is embedded, so that the number of SDF is in fact the embedding dimension of a generic orbit. In quantum mechanics this is the manifold (subspace) on which the generic orbit is ergodic. The difference occurs because of the nonlinear character of the classical dynamical equations and the consequent fact that the generic orbit of the classical system might be a (fat) fractal [@Farmer; @usIAN]. This is impossible in quantum mechanics where all orbits of the state vector are either periodic, with different dimensionality, or quasi-periodic, i.e. ergodic on a torus of dimension equal to the half of FDF. Of course, from a practical point of view a periodic orbit with very long period is for some purposes like an ergodic one. However, for our purposes the ergodicity of quantum orbits is a qualitative property meant to characterize the orbits with generic behavior.
Consider an integrable Hamiltonian system with $N$ SDF. In the generic case the number of DDF is also $N$. However, the system might be degenerate in such a way that all frequencies of motion on all invariant dim-$N$ tori are commensurate so that all orbits (except few isolated ones) are periodic and therefore the system has only one DDF. Analogously, consider a quantum system with an $N$-dimensional Hilbert space (as a special case of an integrable system). If the energy spectrum of the system is known then the number of DDF can be discussed without specifying the dynamical algebra and the number of SDF. If $N-1$ ratios $e_i/e_1$ of the energy eigenvalues $e_i, i=1,2\dots N$ are irrationally related then typical orbits (except the stationary ones) are quasi-periodic, and the motion of the state vector is ergodic on an $N-1$ dimensional torus in ${\bf R}^{2N}$. The number of DDF is $N-1$, the same as the number of (quantum) FDF. If there are some rational relations the orbit in the union of the corresponding eigenspaces is periodic, and the number of DDF is smaller than $N-1$. In the special case of $N=\infty$ and when all eigenvalues are multiples of a single value, the generic orbit is a circle of dimension $1$. This corresponds to the harmonic oscillator with the dynamical algebra $h_4$ and the number of SDF and DDF both equal to unity. The number of FDF is of course infinite. It appears, from number-theoretic reasons and linearity, that a generic quantum system with $N$ dimensional Hilbert space should have $N-1$ DDF.
Obviously, the number of SDF is smaller or equal than the number of FDF, but the number of DDF can be smaller, equal or greater than the number of SDF. We have seen that the NSDF of a classical system and its quantization are equal. However, determination of the relation between NDDF of related quantum and classical systems is a challenging problem. The problem involves analyzes of the relation between classical integrability, separability in the fixed set of SDF, dynamical symmetry and generation of quantum entanglement between the fixed set of SDF. In what follows we shall contribute very little to the general solution of this quite difficult problem. We shall only use a couple of examples to indicate some important facts which certainly contribute to the impression that there should be a general relation between NDDF of related quantum and classical systems, but also show that the relation is not a simple one. In order to proceed with the analyzes we shall need to define a) a measure of quantumness of a given quantum state and b) a classical system which is naturally associated with an arbitrary quantum system with defined dynamical algebra and the Hamiltonian operator. The well known notion of generalized coherent states is needed in both of these steps, so we shall provide a brief recapitulation.
g-coherent states and the measure of quantumness
=================================================
The number of SDF is in the quantum case directly related to the dimensionality of the manifold of the coherent states corresponding to the system’s dynamical algebra. The coherent states corresponding to a dynamical algebra $g$ will be denoted as $g$-coherent states. Let us briefly recapitulate the definition of the generalized g-coherent states. This is also relevant for the definition of the classical model of a quantum system, to be introduced in the next section.
Consider first quantum systems with $g=h_4\oplus h_4\dots\oplus h_4$ dynamical algebra and the Hilbert space ${\cal H}=L_2(x_1)\otimes L_2(x_2)\dots L_2(x_N)$. It makes sense to define a level of quantumness of a state $|\psi\rangle\in {\cal H}^N$ by the expression $$\Delta_{h_4}(\psi)=\sum_i^N \Delta_{\psi} \hat x_i\Delta_{\psi}\hat p_{i},$$ where $\Delta_{\psi} \hat A$ denotes dispersion of $\hat A$ in $|\psi\rangle$. The states that have minimal quantumness $\Delta_{h_4}$ are $N$-products of Glauberg $h_4$-coherent states. This is one of the important properties of the $h_4$ coherent states $|\alpha_1\rangle$ carried over to the products of $h_4$-coherent states $|\alpha_1\rangle\alpha_2\rangle\dots|\alpha_N\rangle$. In the previous expressions $\alpha_i\in {\mathbf C}$ are complex parameters that parameterize the manifold of the coherent states for the $i$-th system. It is well known that all such states can be obtained by the action of a displacement operators $D(\alpha_1,\dots \alpha_N)=\exp \sum_i^N(\alpha_i\hat a_i^{\dag}-h.c.)$ onto the vacuum state. It is also well known that this property is used to define and construct generalized $g$-coherent states for systems with compact dynamical groups with finite ${\cal H}^N$. However, in this case the level of quantumness $\Delta_{g}(\psi)$ is defined in a different way.
Let us first, very briefly, recapitulate the construction of generalized $g$-coherent states. There are several generalizations of Glauber, i.e. $h_4$ coherent states. Perelomov [@Perelomov] and Gillmore [@Gilmor] independently introduced two different generalizations based on the group-theoretical structure of the $H_4$ coherent states. The essential ideas of both approaches are the same, the differences being in the class of Lie groups, and the corresponding available tools, and in the choice a reference state. In both approaches, the set of $g$-coherent states depends, besides the algebra $g$, also on the particular Hilbert space $H^{\Lambda}$ caring the irrep $\Lambda$ of $g$ and on the choice of an, in principle(Perelomov), arbitrary referencee state, denoted $|\psi_0>$. Here $\Lambda$ is a multi-index indexing irreps of $g$.
The subgroup $S_{\psi_0}$ of $G$ which leaves the ray corresponding to the state $|\psi_0>$ invariant is called the stability subgroup of $|\psi_0>$: $h|\psi_0>=|\psi_0>\exp i\chi(h), h\in S_{\psi_0}.$ Then, for every $g\in G$ there is a unique decomposition into the product of two elements, one from $S_{\psi_0}$ and one from the coset $G/S_{\psi_0}$ so that $g|\psi_0>= \Omega|\psi_0> \exp i\chi(h)$. The states of the form $|\Lambda, \Omega>=\Omega|\psi_0>$ for all $g\in G$ are the $g$ coherent states. The notation indicates the fixed arbitrary irrep by $\Lambda$ and the particular element from the coset $G/S_{\psi_0}$ by $\Omega$. The latter parameterizes the set of $g$-coherent states, obtained using the $\Lambda$ irrep of $g$. Geometrically the set of $g$-coherent state form a manifold with well defined Riemanien and symplectic structure.
Consider a quantum system with the dim-n dynamical algebra represented on a finite-dimensional Hilbert space ${\cal H}^N$. Denote by $L_i,\> i=1,2,\dots n$ the algebra generators in an arbitrary basis of the algebra. Level of quantumness of a pure state $|\psi>\in {\cal H}^N$ is defined as $$\Delta_g(\psi)=\sum_i^n<\psi|L_i^2|\psi>-<\psi|L_i|\psi>^2,$$ where the sum is taken over an orthonormal bases of the dynamical algebra $g$. It make sense to consider the quantity $\Delta_g(\psi)$ as a measure of quantumness of the state $\psi$ for the system with the dynamical algebra $g$. The general definition of $g$-coherent states is such that $\Delta_g(\psi)$ is minimized precisely by such coherent states.
In the case of systems with a semi-simple dynamical algebra $g$, studied by Gillmore, the irrep space is characterized by the unique highest weight state $|\Lambda,\Lambda>$ (or the lowest weight state $|\Lambda,-\Lambda>$). This vector is annihilated all $E_{\alpha},\> ((E_{-\alpha})$ where $E_{\alpha}$ ( $E_{-\alpha}$) belong to the standard Cartan basis of $g$: $\{H_i,E_{\alpha},E_{-\alpha}\}$.. The state $|\Lambda,\Lambda>$ is left invariant by operators in the Cartan subalgebra $H_i$. The set of $g$ coherent states can be represented in the form of an action of the displacement operator on the reference state $|\Lambda,\Lambda>$. $$|\alpha>=D(\alpha)|\Lambda,\Lambda>=\exp[\sum
\alpha_iE_i-h.c.]|\Lambda,\Lambda>,$$ where$\alpha$ is a multi-parameter standing for the set of complex parameters $\alpha_i$ and the sum extends over all $E_{-\alpha}$ that do not annihilate $|\Lambda,\Lambda>$. The stabilizer $S_{\psi_0}$ of the reference state $|0>$ is the subgroup generated by the Cartan subalgebra of $g$ The complex parameters $\alpha_i,i=1,2\dots M$ parameterize $2M$ dimensional manifold $G/S_{\psi_0}$ of $g$-coherent states.
It is obvious that the number of SDF of a quantum system is equal to half the number of dimensions of the manifold of g-coherent states. However, if the Hamiltonian is a nonlinear expression in terms of the dynamical algebra generators then the manifold of g-coherent states is not dynamically invariant. Therefore the system that has started from the manifold of $g$-coherent state, i.e. the states with minimal quantumness will necessarily leave this manifold and the quantumness will increase. Of course, later evolution, i.e. from non-coherent states, might lead to local (in time) increase or decrease of quantumness. In any case, $g$-quantumness is not preserved along typical orbits. Quantum systems that generate g-quantumness, i.e. such that $g$-quantumness is not a constant of motion, satisfy the inequality NDDF$>$NSDF. We see again that the number of DDF is often larger than the number of SDF.
Classical model of a quantum system
====================================
Consider a quantum system given by a dynamical algebra $g$, a Hilbert space ${\cal H}$ and a Hamiltonian $\hat H$. Classical Hamiltonian dynamical system on the symplectic manifold of g-coherent states $G/S_{|\psi_0>}$ and given by the Hamiltonian function $H(\alpha)=<\alpha|\hat H|\alpha>$ is called the classical model of the quantum dynamical system $({\cal H},g,{\hat H})$. Physically, the classical model corresponds to a certain type of coarse-grained description of the quantum system. The coarse-graining is modeled by imposing specific constraints on the Hamiltonian formulation of quantum dynamics [@usPRA1; @usPRA2].
In the Hamiltonian formulation of a quantum system $({\cal H}^N, g,\hat H)$ [@Ashtekar; @Brody; @Marmo], the Schr" odinger equation on ${\cal H}^N$ (the dimension of (${\cal N}^N$, can be $N=\infty$) is considered as a Hamiltonian dynamical system on the manifold ${\cal M}={\mathbf R}^{2N}$, with the simplectic and Riemannian structures given by the imaginary and the real part of the Hermitien scalar product. The Hamiltonian of the system is given as $H(X)=\langle \psi_X|\hat H|\psi_X\rangle$. The set of real and imaginary components $\Re c_i,\> \Im c_i$ of a state $|\psi\rangle$ in any basis gives a set of canonical coordinates $Q_i=\Re c_i/\sqrt 2$ and $P_i= \Im c_i/\sqrt 2$ of the point $X_{\psi}$. The dynamical algebra of the quantum system is $g$ with the generators $L_i$ and the corresponding $g$-coherent states. Consider a constrained quantum Hamiltonian system with the constraint that the quantumness $\Delta_g(\psi)$ is preserved minimal during the evolution, and call such system $\Delta_g$ constrained. The manifold of constraints $\Delta_g(\psi)=min$ is denoted $\Gamma_g$.
[*Definition of the classical model*]{} The classical model of the quantum system $({\cal H}^N, g,\hat H)$ is by definition the reduction of the $\Delta_g$ constrained Hamiltonian system $({\cal M}, \langle\hat H\rangle)$ onto the constrained manifold $\Gamma_g$.
The constraint $\Delta_g$ depends on the algebra of basic dynamical variables and is related to an equivalence relation between the quantum pure states. For a given algebra the space of quantum pure states can be partitioned into equivalence classes such that each equivalence class contains one and only one state from the constrained manifold, i.e. a state with minimal g-quantumness. The $\Delta_g$ constrained Hamiltonian system preserves the equivalence of states during the evolution.
One can use the theory of constrained Hamiltonian systems developed by Dirac, to study the Hamiltonian system with $\Gamma_g$ constraints. However, if $\Gamma_g$ is symplectic, the general procedure can be bypassed, with the known result that the reduced constrained system is in fact also Hamiltonian on $\Gamma_g$ with the Hamilton’s function given by $\langle \hat H\rangle|_{\Gamma_g}$. In our case, when $\Gamma_g$ is determined by $\Delta_g$, it is known that $\Gamma_g$ coincides with the manifold of $g$-coherent states and the latter is a symplectic manifold. Thus, a definition of the classical model equivalent to the previous one can be given as
[*Definition of the classical model*]{} The classical model of the quantum system $({\cal H}^N, g,\hat H)$ is the Hamiltonian dynamical system on the symplectic manifold parameterizing the $g$ coherent states $\alpha\in \Gamma_g$ with the Hamilton’s function $H(\alpha)=\langle\alpha|\hat H|\alpha\rangle$.
The classical model has the same dynamical algebra and the same number of SDF as the corresponding quantum system.
Special examples of the construction of classical models for a system of coupled unharmonic oscillators and spin-$j$ systems with arbitrary $j$ that is the systems with dynamical algebras $h_4\oplus h_4\oplus\dots h_4$ and $su(2)$ on ${\cal H}^{2j+1}$, have been introduced in [@usPRA1] and [@usPRA2] respectively. In the following we shall treat in detail the example of two coupled oscillators, and consider two more examples given by $su(2)\oplus su(2)$ and $su(3)$ dynamical algebras.
Classical limit of a quantum system, if i exists, is obtained from the corresponding classical model in some macro-limit. Given a quantum system, its classical model and its classical limit are not the same Hamiltonian dynamical systems. For the oscillators systems the macro-limit corresponds to $\hbar\rightarrow 0$ and for the spin system to $j\rightarrow \infty$. The macro-limit is well controlled because the classical model satisfies the constraint of constant and minimal dispersions. Dynamics of the classical model and the classical limit might be qualitatively different.
[**Remark**]{} Structural features of systems of identical bosons or fermions are also described by appropriate dynamical algebras. Consequently, the definition of SDF, construction of the corresponding g-coherent states and finally the construction of the corresponding classical model are all performed as in the general case. Many different dynamical algebras have been constructed using the canonical commutation relations (CCR) for a system of bosons or fermions $$[a_i,a_j^{\dag}]_{\pm}=\delta_{ij},\>[a_i^{\dag},a_j^{\dag}]_{\pm}=0,\>[a_i,a_j]_{\pm}=0,$$ and have been found useful in the treatment of dynamical problems. In (7) $a_i^{\dag}$ is the single particle creation operator of the $i$-th energy level, and + (-) indicates the anti-commutator (commutator). Generators of the relevant dynamical algebras are expressed in terms of single particle operators and their bilinear combinations. Most commonly used dynamical algebras, describing the structure of systems of $r$-level fermions or bosons, are $u(r)$ or $so(2r)$ and $(sp(2r)$ modeling physically different situations. For example $u(r)$ generators (not all independent) are given by the set $\{a_i^{\dag}a_j,\>i,j=1,2\dots r\}$, with the corresponding commutation relations. We shall provide few more details only in the $u(r)$ case.
Subalgebra structure, which determines the number of SDF, of the representations of these groups can be very reach, and strongly depends on representation. An example is given by the case when the elementary excitation operators for a system of $n$ $r$-level fermions are of the form $$\{a_i^{\dag}a_k,\> 1\leq k\leq n, n+1\leq i\leq r\}.$$ and these represent $n(r-n)$ SDF of such fermion system with $u(r)$ dynamical algebra.
Typical Hamiltonian of a many-body system of identical particles is expressed in terms of the algebra generators, by possibly nonlinear expression $$H=\sum^r_{i}\omega_i a_i^{\dag} a_i+\sum^r V_{ijkl}a_i^{\dag} a_j^{\dag} a_k a_l.$$ Ground state of such a Hamiltonian is denoted by $\psi_0$ and of course shares the symmetry of the system. This fact introduces crucial difference is the manifolds of coherent states for bosons or fermions with the same $u(r)$ dynamical algebra.
Generalized coherent states are given by the the action of the exponential mapping from the set of elementary excitations on the ground state $$|\alpha\rangle=D(\alpha)|\psi_0\rangle=\exp\sum_{ik} (\alpha_{ik}a_i^{\dag}a_k+\bar \alpha_{ik}a_k^{\dag}a_i)\>|\psi_0\rangle,$$ and, in the fermionic case, form the symplectic manifold $U(r)/U(n)\otimes U(r-n)$, which is the phase space of the classical model of the fermionic system. On the other hand the manifold of the coherent states, i.e. the phase space of the classical model in the bosonic case is $U(r)/U(1)\otimes U(r-1)$. Classical model share the symmetry of the ground state, that is the symmetry of the fermionic or bosonic system. In any case the coherent states are considered as the most classical states of the considered quantum system. These states minimize the quantumnes $\Delta_{u(r)}(\psi)$ defined as in the general case (5). Well defined classical model for fermions or bosons does not imply existence of an appropriate classical limit. The later is achieved as the number of particles $n\rightarrow \infty$ only for bosons. Of course the set of coherent states is not invariant on the quantum evolution generated by the Hamiltonian with general $V_{ijkl}\neq 0$, and the quantumness is not preserved by such evolution. Dynamics of the classical model is given by the general prescription. The Hamilton function on the phase space is the expectation $\langle \alpha|\hat H|\alpha\rangle$. Dynamics of examples of such classical Hamiltonian systems corresponding to bosons or fermions have been treated in [@Cheng3]. Relation between the evolution of quantumness and the qualitative properties of the classical model has not been studied.
Relation between DDF of a quantum system and its classical model
================================================================
Dynamics of classical models of quantum systems have been studied for various examples in [@Cheng1; @Cheng2; @Cheng3], but without understanding them as constrained Hamiltonian systems. Relation between dynamics of entanglement and the dynamics of classical models for a pair of qubits was studied in [@JaPhysRev06; @JaAnnPhys]. Here we want to argue in favor of existence of a general relation between the generation of g-quantumness and qualitative properties of the dynamics of the classical model as defined in the previous section, and to use this to infer relations between the corresponding NDDF.
Examples
--------
Our first example is rather trivial in that it illustrates the case when the quantum system and its classical model are the same Hamiltonian dynamical system. The second example considers a pair of $1/2$-spins as a genuinely quantum system, i.e. a quantum system which can not be obtained by quantization of a classical mechanical system. Nevertheless the classical model is well defined and the question of relation between NDDF for the two systems makes sense. Furthermore, in this case the question has a definite and complete answer, suggesting a general relation between the integrability of the classical model, the dynamical symmetry and the dynamical generation of quantumness. In this case, the NDDF of the spins system and the classical model are equal for any of the different dynamical regimes. However, this simple relation between integrability of the classical model and the lack of quantumness generation is not valid in general. The relation is more complicated as is illustrated by our third example of coupled nonlinear oscillators.
### von-Neumann case: $u(N)$ dynamical algebra
This is a rather trivial example because the quantum system and its classical model are identical. The classical model is integrable, the quantum Hamiltonian is an element of $u(N)$ dynamical algebra and all states of the quantum system have equal $u(N)$-quantumness.
The quantum system is described by $N$ dimensional Hilbert space ${\cal H}^N$ and the dynamical algebra $u(N)$, which means that every hermitian operator on $H^N$ has physical interpretation as a measurable quantity. In particular, any Hamiltonian is an element of the dynamical algebra, i.e. a Hermitian operator. Due to the normalization and global phase invariance the state space of the system is $CP^{N-1}$ which is topologically like $S^{2N-1}/S^1$, and represents a $2(N-1)$ manifold with Riemanien and symplectic structure. Geometrically, it should be natural to associate $N-1$ FDF with this system. The same number of SDF follows from $u(N)$ dynamical algebra. The Hilbert space is the fully symmetric irrep space of $u(N)$ with the highest weight: $\Lambda=(1,0,\dots 0)$. The basis can be labeled by the following chain of subalgebras: $u(N)\supset u(N-1)\dots\supset u(1)$ with the corresponding Casimir elements $C_i^{u(k)},\> i=1,2\dots k, \>k=1,2\dots N$. The $N-1$ non-fully degenerate operators are $C_i^{u(k)},\> i=1,2\dots
k, \>k=1,2\dots N-1$ and label the basis $|i\rangle=|0,0,\dots i,\dots
0\rangle,i=0,1,2,\dots N-1$. Explicitly: $C_k^{u(k)}|i>=\Theta(k-(N-i))|i\rangle$, and $\Theta(i)$ is the Heaviside function on $i=1,2\dots N-1$. Thus there is $N-1$ SDF, the same as the number of FDF.
Elementary excitation operators are given by: $E_{i0}|\psi_0\rangle=|i\rangle,i=1,2,\dots N-1$ where $|\psi_0\rangle$ is the lowest weight vector of the $\Lambda=(1,0,\dots 0)$ representation, and $u(N)$ coherent states are obtained as $|\alpha\rangle=exp(\sum \alpha_i E_{i1}-h.c)|0\rangle$. Coherent states are parameterized by the coset space $U(N)/U(N-1)\otimes U(1)$ which is isomorphic to $CP^{N-1}$. We see that all states are $u(N)$ coherent states and therefore all states have the same quantumness. Any Hamiltonian can be diagonalized by an $U(N)$ transformation and thus expressed as a combination of the Casimir elements.
It should be noticed that since any state is $u(N)$ coherent state the dynamics of the quantum system on $CP^{N-1}$ with the Hmiltonian $\hat H$, and its classical model with the Hamiltonian function $\langle\psi|\hat H|\psi\rangle$ on the phase space $U(N)/U(N-1)\otimes U(1)\sim CP^{N-1}$ are identical (and integrable) for any Hamiltonian. A quantum system with generic $u(N)$ Hamiltonian and the corresponding (isomorphic) classical model have equal number of DDF.
### Coupled spins: $g=su^1(2)\oplus su^2(2)$ semi-simple dynamical algebra
Consider a pair of spins with the Hilbert space $H=C^{2}\otimes
C^{2}$ and the Hamiltonian $$H=\omega(J_z^1+J_z^2)+\mu_x J_x^1J_x^2+\mu_z J_z^1J_z^2.$$ where $J^{i}_{x,y,z},\>i=1,2$ are operators satisfying the commutation relations of $su^1(2)\oplus su^2(2)$.
The dynamical group of the system is $SU^1(2)\otimes SU^2(2)$. A possible subgroup chain is $SU^1(2)\otimes SU^2(2)\supset SO(2)\otimes SO(2)$. The number and the nature of SDF are as in the $su^1(2)\oplus su^2(2)$ example of section 2.2. The set of g-coherent states is given by products of su(2)-coherent states for each of the spins. Topologically, they represents a Cartesian product of two 2D-spheres $S^1\times S^2$ where the superscripts correspond to the spins 1 and 2. The set of coherent states is parameterized by local symplectic coordinates on $S^1\times S^2$ as $|q,p\rangle\equiv |q_1,p_1,q_2,p_2\rangle$. They minimize quantumness (5). Non-coherent states are entangled with respect to SDF given by $J_z^1$ and $J_z^2$, with the quantumness larger than minimal. The radius of each of the spheres of the coherent states is $2\sqrt{J\hbar}$ where $2J+1$ is the dimension of the relevant $su(2)$ representation.
When $\mu_x=0,\mu_z\neq 0$ the Hamiltonian commutes with $J_z^1$ and $J_z^2$, and the system has the dynamical symmetry of $SO^1(2)\otimes SO^2(2)$ corresponding to the selected SDF. This is manifested in the fact that the system does not generate quantumness with respect to the considered SDF despite the interaction $\mu_z J_z^1J_z^2$ between the two spins.
If $\mu_x\neq 0$ the $SO^1(2)\otimes SO^2(2)$ dynamical symmetry is broken. The set of coherent states is not dynamically invariant. The system generates quantumness in the form of entanglement between the $SO^1(2)\otimes SO^2(2)$ dynamical degrees of freedom (please see fig. 1).
{width="100.00000%"}
[*Dynamics of the classical model*]{}
The phase space of the classical model is the manifold of $su^1(2)\oplus su^2(2)$ coherent states, i.e. the Cartesian product of two spheres. Local symplectic coordinates on $S^1\times S^2$ are denoted by $q_1,p_1,q_2,p_2$. The Hamilton’s function of the classical model is the coherent state expectation $H(q,p)=\langle p,q|\hat H| p,q\rangle$ of (11), and is expressed in terms of the expectation values of the generators $\langle p,q |J_{x,y,z}^{1,2}|p,q\rangle$ by the following formulas: $$\begin{aligned}
\langle\hat J^i_x\rangle(p,q)&=&{q_i\over 2}(4J-q_i^2-p_i^2)^{1/2},\nonumber\\
\langle\hat J^i_y\rangle(p,q)&=&-{p_i\over 2}(4J-q_i^2-p_i^2)^{1/2},\nonumber\\
\langle\hat J^i_z\rangle(p,q)&=&{1\over 2}(q_i^2+p_i^2-2J), \quad i=1,2\end{aligned}$$ where $J=1/2$ in our case of two $1/2$-spins.
The Hamiltonian of the classical model is given by $$\begin{aligned}
H&=&(p_1^2+p_2^2+q_1^2+q_2^2)+\mu_z (p_1^2+q_1^2-2J)(p_2^2+q_2^2-2J)\nonumber\\
&+&\mu_x q_1q_2[(4J-p_1^2-q_1^2)(4J-p_2^2-q_2^2)]^{1/2},
\end{aligned}$$ with $J=1/2$. Observe that the classical model involves only the coherent states expectations of $J^i_{x,y,z}$ and products of such expectations for different spins. No expectations of operators which are nonlinear expressions in terms of $J^i_{x,y,z}$ occur. It is often stated that the classical limit of the spin system is obtained by taking $J\rightarrow \infty$. Strictly speaking, this corresponds to the limit of classical models of a sequence of large quantum spins. The systems of 1/2-spins do not have the classical limit but do have classical model, i.e. (13) with $J=1/2$.
When $\mu_x=0,\mu_z\neq 0$, despite the interaction between the SDF, the Poisson bracket $\{H,\langle J_z^{1,2}\rangle\}$ is zero, being proportional to $\mu_x$. Like in the quantum case the system has dynamical $SO(2)\times SO(2)$ symmetry. The classical model is completely integrable with the obvious set of independent constants of motion.
When $\mu_x\neq 0$ the $SO(2)\otimes SO(2)$ dynamical symmetry is broken, and the coherent state expectations $\langle J_z^{1,2}\rangle$ are not constants of motion any more. The classical model might not be completely integrable. Chaotic orbits of the classical model with $J=1/2$ are easily found in numerical computations, and one such orbit is illustrated in fig.2.
{width="100.00000%"}
Thus, analyzes of the quantum system (11) and its classical model (13) suggest that the quantum system (11) generates quantumness if and only if the classical model (13) is not completely integrable. However, the following example will show that such relation is not generally true.
Nevertheless, one property of the quantum system and its classical model should be observed and stressed. The two interacting $1/2$-spins and their classical model simultaneously have the dynamical symmetry generated by the subalgebra used to define the SDF, and if the interaction is such then the quantum system does not generate quantumness. Otherwise if the dynamical symmetry is broken the quantum system generates quantumness. A particular feature of this example is that if the dynamical symmetry that corresponding to the SDF is broken then the classical model is not integrable. Of course, there are classical Hamiltonian systems with potential interaction that breaks the symmetry used to define the SDF, but that are nevertheless completely integrable. An example is provided next.
### Coupled oscillators: $h_4\oplus h_4$ dynamical algebra
The dynamical algebra and the subalgebra chain used to specify the SDF are the same as in the example 1 of section 2.2. Therefore the system has two SDF corresponding to the algebra elements given by the Casimir elements $n_1$ and $n_2$ of the subalgebra. Coherent states $|p,q\rangle$ are again separable and given by the product of coherent states $|p,q\rangle\equiv |p_1,q_1\rangle|p_2,q_2\rangle$ for each of the SDF.
Consider the quantum Hamiltonian $$\hat H=\sum_{i=1,2} \frac{1}{2}[\hat p_i^2+\hat q_i^2] +\mu_1 \hat q_1^2\hat q_2-\frac{\mu_2}{3} \hat q_2^3$$
Observe that $\hat q_i,\hat p_i$ denote the coordinate and the momentum operators, while $(q_i,p_i)$ are not their eigenvalues (which shall not occur) but are the parameters of the $h_4\oplus h_4$ coherent states.
Dynamics of the quantumness $\Delta_{h_4\oplus h_4}(\psi)=\sum_{i=1,2}\Delta_{\psi}\hat q_i\Delta_{\psi}\hat p_i$ is computed staring from a coherent state $|p,q\rangle$. Figure 3 demonstrate that when $(\mu_1=1,\mu_2=1)$ then $\Delta_{h_4\oplus h_4}(\psi(t))\neq const$ and when $(\mu_1= 0,\mu_2=0)$ or $(\mu_1= 0,\mu_2=1)$ then $\Delta_{h_4\oplus h_4}(\psi(t))= const=0.5$ is minimal all the time. Thus, the example suggests that as long as there is no interaction between the two SDF there isc no generation of quantumness, even if the dynamics of the separated SDF is governed by a non-quadratic Hamiltonian. On the other hand, the quantumness is generated if there is specific interaction between the SDF. Observe that the interaction in (14) is such that the system with interaction is not symmetric under the subgroup generated by the subalgebra used to define SDF. In other words, the Casimir operators $\hat n_1$ and $\hat n_2$, that are used to identify the two SDF do not commute with the interaction.
{width="100.00000%"}
[*Dynamics of the classical model*]{}
The Classical model is given by $$H_{p,q}=\sum_{i=1,2}\frac{1}{2}[ p_i^2+ q_i^2]+\langle p,q| \hat \mu_1 q_1^2\hat q_2+\frac{\mu_2}{3}\hat q_2^3|p,q\rangle$$
Using the general formula for a Hamiltonian of the form $\hat H=\sum_i\hat p_i^2/2m_i+\hat V(q_1,q_2)$ $$\label{e:H_sum}
H_{p,q}={\sum_i p_i^2/ 2m_i}+V(x)+\sum_{k=1}^{\infty} {1\over 2^k k!}{\hbar^k V^{(2k)}(x)\over (2m\omega)^k},$$ where $V^{2k}(x)$ denotes the sum of derivatives of order $2k$ of the potential $V(x)\equiv V(q_1,q_2)$, we obtain the explicit form of the Hamilton’s function for the classical model $$\begin{aligned}
H(q_1,q_2,p_1,p_2)&=&\frac{1}{2}[p_1^2+p_2^2+q_1^2+q_2^2]+\mu_1 q_1^2 q_2-\frac{\mu_2}{3} q_2^3\nonumber\\
&+&\frac{\hbar}{2}[1+\mu_1(q_1+q_2)-\mu_2 q_2].\end{aligned}$$
Many typical orbits are computed and it is demonstrated that when $(\mu_1=0,\mu_2=0)$ or $(\mu_2=1,\mu_1=0)$ then all orbits are periodic or quasi-periodic and when $(\mu_2=1,\mu_1=1)$ then there are irregular orbits. The results are illustrated in figure 4. by plotting $q_1(t)$ component of a periodic (fig.4a) and chaotic (fig. 4b) orbits. Other interesting cases of the parameter $(\mu_1,\mu_2)$ values will be discussed shortly.
{width="100.00000%"}
[*Dynamics of the classical model in the macro-limit*]{}
The Hamilton’s function of a classical system is obtained from the classical model in the macro-limit $\hbar\to 0$ $$H_{cl}=\frac{1}{2}[p_1^2+p_2^2+q_1^2+q_2^2]+ \mu_1 q_1^2 q_2-\frac{\mu_2}{3} q_2^3.$$ Orbits are computed and it is demonstrated that for $(\mu_1=0)$ (and arbitrary $\mu_2$) all orbits are periodic or quasi-periodic and when $(\mu_1=\mu_2 =1)$ then some orbits are irregular. In fact, it is known that the classical system, known as the Henon-Hiles model [@HH], is completely integrable for the following cases of the parameter values: a) $(\mu_1/\mu_2=0)$; b) $\mu_1/\mu_2=-1$ and c) $\mu_1/\mu_2=-1/6$. Observe that the classical model (17) for the parameter values a) is also integrable (separable) with some bounded orbits, and for these parameter values the quantum system does not generate quatumness. The last two integrable cases are of special interest for the comparison with the quantum model. The classical model with $\mu_1/\mu_2=-1$ is integrable, but the SDF are not separated and $H_4\otimes H_4$ is not a symmetry of the classical model. For these parameter values the quantum model generates quantumness. So, it is not just the complete integrability of the classical system which is enough to imply the lack of quantumness generation by the quantized system.
We have also tested a system of two harmonic oscillators with the simplest interaction $\hat V=\hat q_1 \hat q_2$, in which case the classical system and the classical model are the same. The classical system is integrable but the corresponding quantum system generates quantumness (please see fig. 5).
{width="100.00000%"}
The conclusion of all computations with the example (14) and its classical model (17) is as follows. If the classical model is not integrable the quantum system generates quantumness. However, the quantum system might generate quantumness even if the corresponding classical model is completely integrable. In the considered examples, this happens if the interaction between SDF is such that the subgroup generated by the subalgebra chain used to define the SDF is not a dynamical symmetry. The classical model is completely integrable, so there exist the corresponding action-angle variables. However, the action variables and the variables corresponding to the SDF are related by a nonlinear transformation. Therefore, the commutation relations between SDF variables and the quantum Hamiltonian do not reproduce the corresponding Poisson brackets and the quantum SDF variables do not generate a dynamical symmetry.
The last observation lead us to considered a system of coupled harmonic oscillators such that the subgroup generated by the subalgebra chain used to define the SDF is a dynamical symmetry. The Hamiltonian of the quantum system is $$\hat H=\sum_{i=1,2} \frac{1}{2}[\hat p_i^2+\hat q_i^2] +\frac{1}{4}[\hat p_1^2+\hat q_1^2][[\hat p_2^2+\hat q_2^2],$$
In this case the quantum system does not generate quantumness, despite the interaction between SDF. The classical model and the classical system differ by a constant only, and are completely integrable. More importantly they are in a subclass of integrable systems such that the Hamiltonian is expressed in terms of only the Casimir elements used to define SDF (the action variables of the harmonic oscillators). The reason for the lack of quantumness generation is by now quite clear: the subgroup generated by the subalgebra chain used to define the SDF is a dynamical symmetry.
Other examples that we have treated include systems with dynamical algebras $g=su(2)\oplus h_4$ and $g=su(3)$. These also support the general conjecture formulated in the next subsection. We shall briefly comment the case of systems providing realizations of $su(3)$ dynamical algebra, because they provide the opportunity to illustrate few interesting features.
### A simple system with more than one SDF: $su(3)$ dynamical algebra
The example of $su(3)$ dynamical algebra is used to illustrate the systems with more than one SDF which nevertheless can not be considered as composed of component systems with fewer number of SDF because the irrep space of states does not have the corresponding tensor product structure. Thus, the system should not be considered as simple in the sense that it is not composed of simpler systems. The example will also illustrate another important fact, namely the fact that the number of SDF might depend on the particular irrep that is carried by the system’s Hilbert space. However, the general conjecture about NDDF, concerning the relation between the dynamical symmetry, generation of quantumness and dynamics of the classical models seems to be confirmed also by the $su(3)$ examples.
The $su(3)$ Lie algebra has rank $2$ and dimension $8$. The basic commutation relations between the generators $E_{i,j}, \>i,j=1,2,3$ which are not independent are: $[E_{ij},E_{kl}]=\delta_{jk}E_{il}-\delta_{il}E_{kj}$ and can be realized in terms of bosonic creation and annihilation operators of three modes as follows: $E_{i,j}=a_i^{\dag}a_j,\>i,j=1,2,3$. The eight independent hermitian generators are given by: $X_1=(a_1^{\dag}a_1 -a_2^{\dag}a_2);\>X_2=(a_1^{\dag}a_1
-a_2^{\dag}a_2-2a_3^{\dag}a_3);\>Y_k=i(a_k^{\dag}a_j-a_j^{\dag}a_k);\>Z_k=(a_k^{\dag}a_j-a_j^{\dag}a_k),\>
k=1,2,3,\> j=k+1 \>({\rm mod} 3)$. These will be used in the formula (5) for the level of $su(3)$-quantumnes in a particular system with the corresponding SDF.
In order to determine the number of SDF we need to find the number of nonfully degenerate operators in any particular chain of subalgebras. We shall use the subalgebra chain: $su(3)\supset
su(2)\oplus u(1)\supset u(1)$ with five Casimir operators usually denoted by $C_2,C_3,Y,T^2,T_z$. $C_2$ and $C_3$ are the Casimir operators of the su(3) itself, $T^2$ and $T_z$ are the Casimir operators of $su(2)$ and u(1) and $Y$ corresponds to $u(1)$. Thus, in general there are three nonfully degenerate operator and consequently a system with $su(3)$ algebra has three SDF. However, the system is also characterized by its Hilbert space i.e. by a particular irrep and for some irrep all three DF might not be independent.
All irreps of the $su(3)$ algebra can be labeled by their highest weight: $\Lambda=\lambda_1f_1+\lambda_2 f_2$ where $f_1$ and $f_2$ are the highest weights of the two fundamental representations: $(1,0)$ and $(0,1)$. The fully symmetric representations correspond to $\lambda_1=0$ or $\lambda_2=0$. In the fully symmetric representation the operators $T^2$ and $Y$ are not independent and thus in this case the number of SDF is just 2. A system with such SDF has the $su(3)$ dynamical symmetry if its Hamiltonian is expressed in terms of $T^2$ and $T_z$ or $Y,T^2$ and $T_z$ in the two or three degrees of freedom cases.
The coherent states of the $SU(3)$ dynamical group are obtained as in the general case, using the highest weight vector as the reference state $|\psi_0\rangle$. In the general case the coherent states are parameterized by the six dimensional manifold: $SU(3)/U(1)\otimes U(1)$ and in the case of the fully symmetric irrep with two SDF by the four dimensional $SU(3)/U(2)$. As usual the coherent states are of the form $|\Lambda,\alpha>= D(\alpha)|\psi_0\rangle$, where $\Lambda$ is fixed by the irrep and $\alpha$ indexes different coherent states. According to the adopted definition the coherent states have minimal quantumness. The minimal quantumness is preserved by systems with Hamiltonians linear in terms of the algebra generators used to define the corresponding SDF. Dynamics od $su(3)$ quantumness with Hamiltonians nonlinear in the generators corresponding to SDF is illustrated in the following example.
Consider the system of $N$ particles with three possible $N_d$-degenerate energy levels. The following Hamiltonian for such a system is known as the Lipkin model: $$H=\sum_{i=1}^3\omega_i E_{ii}-\mu\sum_{i\neq j}^3E_{i,j}^2$$ where $E_{ij}$ satisfy $su(3)$ commutation relations. Dynamical symmetries of systems with such Hamiltonians and dynamics of the corresponding classical models were studied in [@Cheng2]. When $N\leq N_d$ the Hilbert space of the system is the carrier space of the fully-symmetric irrep and the system has two SDF. If $\mu=0$ there is the dynamical symmetry corresponding to $su(3)\supset su(2)\oplus u(1)\supset u(1)\times u(1)$, and the system does not generate quantumness . For $\mu\neq 0\neq \omega_i$ the dynamical symmetry corresponding to the SDF is broken and the system generates quantumness with respect to the relevant SDF. This is illustrated in figure 6.
{width="100.00000%"}
[**Remark**]{} Properties of quantum systems that indicate if the system is obtained by quantization of an integrable or non-integrable classical system have been studied intensively in the past ( Please see for example [@q_chaos]). For example, such properties are distributions of spectral levels and dynamics of quantum phase-space distributions including their zeros. However, contrary to the studies of classical dynamics where qualitative properties of orbits are central, in these studies of quantum systems properties of state-vector orbits are usually not considered. The reason for this is that orbits of any quantum system are either periodic or quasi-periodic, and no direct qualitative comparison with typically chaotic classical orbits is possible. On the other hand, the notion of DDF is directly related to the dynamics of generic orbits and is relevant for classical as well as quantum systems. Also, generation of quantumness is a property of quantum dynamics that directly reflects relevant properties of the corresponding classical model. Furthermore, classical models exist for a much larger class of quantum systems than those obtained by quantization of classical systems.
Formulation of the main conjecture
----------------------------------
Based on the above examples we formulate the following conjecture for quantum systems with a dynamical Lie algebra and the corresponding classical models
[*Conjecture*]{}
If the dynamics preserves dynamical variables that correspond to the non-degenerate Casimir elements of the subalgebra chain used to define the SDF than NSDF=NDDF for the quantum as well as the classical systems. In this case the quantum system does not generate quantumness and the quantum system and the classical model are given in terms of the Casimir elements corresponding to the SDF.
We have seen examples of classical models which are integrable but not in the form of the previous conjecture. In this case, the corresponding quantum system generates quantumness with respect to the considered SDF, and the NDDF $>$ NSDF for the quantum system. On the other hand, in general for the classical integrable systems NDDF= NSDF.
In the case of generation of quantumness and non-integrability of the classical model we know that NDDF$>$NSDF for the quantum and the classical case, but we can not make any prediction in general concerning the relation between NDDF of a quantum system and the corresponding classical model.
Summary
=======
We have discussed three different types of degrees of freedom (DF): formal, structural and dynamical, that are meaningful and useful in descriptions of quantum and classical systems. The formal (FDF) and the dynamical (DDF) are related to the dynamics; the formal to the type of evolution equation of the class of systems and the dynamical to the relevant properties of a particular system dynamics. On the other hand, the structural DF (SDF) represent what is commonly understood by DF, and describe structural properties of a system, and not its dynamics. SDF have been defined quite generally for systems with basic variables forming a realization of a Lie algebra. An appropriate Lie-algebra uniquely determines the number of SDF and a particular chain of subalgebras determines which are the SDF. Considerations of properties of typical orbits of generic Hamiltonian systems suggest that a notion of dynamical DF (DDF), different in number and type from SDF, is useful and important. Similarly, considerations of entanglement dynamics of typical quantum systems also suggest an analogous notion of DDF. We defined the notion of DDF, for classical as well as quantum systems, as the dimension of dynamically irreducible and invariant manifold generic for a given system. The number of DDF is generically larger than the number of SDF.
We then defined the notion of quantumness of a state with respect to some SDF. It turns out that the Lie algebraic generalized coherent states are the states with minimal quantumness. The generalized coherent states are also used to define an appropriate classical model of a quantum system, with the same number and type of SDF and coarse-grained dynamics.
Our next task was to examine the relation between the numbers of DDF and SDF for a quantum system and its classical model. To this end we have studied relations between generation of quantumness by a quantum system and the dynamics of its classical model. Analyzes of relevant examples suggested the general conjecture that: A quantum system with SDF determined by an algebra and its particular chain of subalgebras does not generate quantumness with respect to the SDF if and only if the subalgebras used to define the SDF generate dynamical symmetries. If the quantumness can not be dynamically generated than the Hamiltonians of the quantum system and its classical model is necessarily expressed solely in terms of the Casimir elements used to define the SDF. The classical systems with this property are certainly completely integrable. Of course, there are classical completely integrable systems with Hamiltonian depending on variables other than the Casimir elements related to the SDF. We have demonstrated examples of quantum systems with such integrable classical models which do generate quantumness with respect to the SDF. This shows that the above conjecture cannot be extended to include all completely integrable classical models.
0.5cm
[**Acknowledgements**]{} This work was supported in part by the Ministry of Education and Science of the Republic of Serbia, under project No. $171017$
[99]{}
[^1]: e-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study a mixture of two superfluids with density-density and current-current (Andreev-Bashkin) interspecies interactions. The Andreev-Bashkin coupling gives rise to a dissipationless drag (or entrainment) between the two superfluids. Within the quantum hydrodynamics approximation, we study the relations between speeds of sound, susceptibilities and static structure factors, in a generic model in which the density and spin dynamics decouple. Due to translational invariance, the density channel does not feel the drag. The spin channel, instead, does not satisfy the usual Bijl-Feynman relation, since the f-sum rule is not exhausted by the spin phonons. The very same effect on one dimensional Bose mixtures and their Luttinger liquid description is analysed within perturbation theory. Using diffusion quantum Monte Carlo simulations of a system of dipolar gases in a double layer configuration, we confirm the general results. Given the recent advances in measuring the counterflow instability, we also study the effect of the entrainment on the dynamical stability of a superfluid mixture with non-zero relative velocity.'
author:
- Jacopo Nespolo
- 'Grigori E. Astrakharchik'
- Alessio Recati
bibliography:
- 'biblio.bib'
title: 'Andreev-Bashkin effect in superfluid cold gases mixtures'
---
Introduction
============
Mixtures of different kinds of miscible superfluids arise in various areas of physics, starting from the first experiments on ${}^3$He-${}^4$He mixtures [@PhysRevLett.15.773], through possible applications to astrophysical objects [@AstrophysJ.282.533; @AstrophysJ.836.203], all the way to the more recent developments in the fields of superconductivity [@PhysRevB.89.104508], cold atoms [@PhysRevLett.89.190404; @PhysRevA.77.011603; @PhysRevLett.118.055301] and exciton-polariton condensates [@Science.326.974].
The statistics of each component of the mixture can be arbitrary, and Bose-Bose, Bose-Fermi, and Fermi-Fermi mixtures were all successfully realised experimentally in cold gases. In these experiments, also the chemical nature of the components can vary: the use of two different elements, of different isotopes of the same element, and of different internal states of a common isotope were demonstrated. The ability to reach simultaneous quantum degeneracy in such a wide variety of atomic species in cold gases experiments allows for the realisation of very diverse interactions between the two component superfluids.
One of the most elusive effect of coupled superfluids is the existence of a non-zero entrainment between them. The presence of mutual transport has been pointed out for the first time in the 1970s by Andreev and Bashkin [@JETP.42.164], correcting some previous work on three-fluid hydrodynamics [@JETP.5.542; @JETP.40.338]. The most prominent feature of such an effect, nowadays known as Andreev-Bashkin effect (AB), is that the superfluid current $\mathbf{j}_i$ of one component will in general depend also on the superfluid velocity ${\ensuremath{{\bf v}}}_j$ of the other component, or, in other words, that the superfluid density is a non-diagonal matrix, $\rho_{ij}$, namely, $$\label{eq:current}
{\ensuremath{{\bf j}}}_i = \rho_{ij} {\ensuremath{{\bf v}}}_j,$$ with the indices $i,j = \{1, 2\}$ labelling the species and implicit summation on repeated indices. We shall refer to the off-diagonal element ${\ensuremath{\rho_{12}}}$ of the superfluid density matrix as the *superfluid drag*. Phenomenologically, a nonzero ${\ensuremath{\rho_{12}}}$ carries important implication in the dynamics of vortices of a superfluid mixture. Notably, it is predicted that the circulation leading to the stable vortex configurations change abruptly as ${\ensuremath{\rho_{12}}}$ is varied, giving rise to stable multiply circulating vortex configurations [@JETP.60.741]
Despite its introduction was inspired by the problem of $^3$He and $^4$He superfluid mixtures, the low miscibility of these two fluids makes this system hardly achievable in experiments. The AB mechanism has instead recently found applications in the domains of astrophysics and of cold atom systems. In the astrophysical literature, it has been hypothesised that the AB effect could be the source of several peculiar behaviours in neutron stars cores [@AstrophysJ.282.533], which modern models predict to be composed of a mixture of neutrons and protons, both in a superfluid phase (see [@Lattimer536; @PhysRevD.70.043001] and reference therein). Cold atom experiments, on the other hand, thanks to their flexibility and tunability, could open the way to a direct measurement of superfluid drag, albeit this will still require some careful analysis and mitigation of common drawbacks. For instance, calculations within the Bogoliubov theory for Bose-Einstein condensates with typical repulsive interaction—where quantum fluctuations are depressed—predict the AB to be very small [@LowTempPhys.30.770; @PhysRevA.72.013616]. Quantum fluctuations can be enhanced by increasing the interactions, but this would also intensify three-body losses, which could be in turn suppressed by confining the system in low dimensional geometries or introducing an optical lattice, as studied, e.g., in [@PhysRevA.79.063610; @PhysRevA.86.033627]. However, optical lattices break translational invariance, thus strongly reducing the superfluid density even a $T=0$. We recall that, in continuous space (and with time-reversal symmetry), the superfluid density approaches the total density as $T$ is lowered to zero (see, e.g., [@Leggett1998]).
Aside from making the AB mechanism efficient, a very important question is how to measure experimentally its strength. The dynamical protocols typically proposed require the ability to initialise a superfluid current in one component and then observe the onset of dissipationless transport in the other one, initially at rest. For best results, these kind of measurements would likely require a ring geometry and can be of difficult interpretation, since a number of decay processes are present [@ZoranCurrent; @AlessioCurrent].
In the present work, we address some of the above mentioned issues. In particular, we derive some relations between the superfluid drag and other measurable quantities, such as the susceptibilities of the system and the speeds of sound. For systems with $\mathbb{Z}_2$ symmetry between the two species, in which spin and density channels decouple, the density channel follows the usual relations, whereas we show how the AB breaks the usual Bijl-Feynman relation for the spin channel. Our findings open the way to measuring the superfluid drag experimentally using standard static and dynamic observables.
To provide support to our theoretical predictions, we study quantitatively a specific model which can show large entrainment, i.e., a dipolar Bose gas trapped in a bilayer configuration. Using the diffusion quantum Monte Carlo method, we extract the dispersion relations, the susceptibilities, the structure factors and the superfluid densities. We show that they satisfy, in a proper regime, the expression derived in the general theory. In particular, it is shown that the standard expression relating the square of the spin speed of sound to the inverse of the susceptibility is inapplicable and it should be corrected by a factor proportional to the superfluid drag. Another possible quantity which could reveal the presence of a superfluid drag is the shift in the position of the dynamical instability. We report a general stability analysis of the mixture, and derive a simple analytical expression for the onset of the dynamical instability to linear order in the drag.
As mentioned above, the AB mechanism could play a more prominent role in low dimensionality. We discuss the modifications to Luttinger liquid theory necessary to describe coupled one dimensional superfluids. We find that, in analogy with the general description, the spin Luttinger parameters, as derived by means of perturbative or ab-initio calculations, receive a correction from the superfluid drag, which could become particularly relevant in the strongly interacting (Tonks-Girardeau) regime.
The paper is organised as follows. In Sec. \[sec:ab\_intro\] we recall the main aspects of the AB effect, which are then analysed within a minimal quantum hydrodynamic toy model in Sec. \[sec:quantum\_hydro\]. The relations among experimentally relevant observables are derived. The Luttinger liquid theory for one dimensional coupled superfluid is corrected for the presence of AB in the same section. Numerical evidence in support of our theoretical findings are reported in Sec. \[sec:numerics\], where we analyse the presence and magnitude of the superfluid drag in a bilayer system of dipolar bosons. In Sec. \[sec:dynamic\_stability\] we study the dynamical instability of the mixture with respect to the relative velocity between the two fluids. Conclusions and future perspectives are drawn in Sec. \[sec:conclusions\]. For the sake of completeness, the derivations of some relations used in the main text are postponed to the appendices without affecting the comprehension of the main results.
Andreev-Bashkin effect {#sec:ab_intro}
======================
Microscopically, the current drag originates from the interactions between two superfluids, leading to the formation of quasi-particles with nonzero content of either of the two species. It is then easy to understand that the transport properties of the two components are not independent: the flow of one component must be accompanied by mass transport of the other component [@JETP.42.164].
Some important relations concerning the superfluid densities in Eq. can be easily obtained by considering the kinetic energy contribution in the expansion of the ground state energy in terms of the superfluid velocities [@PhysRevLett.95.090403]. Due to Galilean invariance, if $\phi_i$ is the phase of the superfluid order parameter for component $i$, of mass $m_i$, its velocity is given by ${\ensuremath{{\bf v}}}_i= (\hbar/m_i)
\nabla \phi_i$ and the energy due to the superfluid velocities can be written as $$\label{eq:energy_Y}
\delta E = \int d^D x\, \sum_{ij}
\frac{\hbar^2}{2m_i m_j}\rho_{ij} \nabla \phi_i \cdot
\nabla \phi_j.$$ By performing a Galilean boost with velocity ${\ensuremath{{\bf V}}}$, the phases are shifted to $\phi'_i = \phi_i - (m_i / \hbar){\ensuremath{{\bf V}}} \cdot {\ensuremath{{\bf r}}}$, and the energy change, to first order in ${\ensuremath{{\bf V}}}$, is $\delta E' = \delta E - \int
{\ensuremath{{\bf P}}}\cdot{\ensuremath{{\bf V}}}\, d^Dx$, with ${\ensuremath{{\bf P}}}/\hbar = \sum_i
(\rho_i + {\ensuremath{\rho_{12}}}) / m_i \nabla \phi_i $ the momentum density. Since on the other hand one must have ${\ensuremath{{\bf P}}}=n_1 \nabla \phi_1 + n_2 \nabla
\phi_2$, with $n_{1,2}$ the number densities, the superfluid densities must satisfy $$\label{eq:density_condition}
m_i n_i = \rho_i + {\ensuremath{\rho_{12}}}.$$ Introducing the effective masses $m_{1,2}^*$ through $\rho_{ii}
\equiv n^{}_i m_i^2/ m_i^*$, we obtain $$\label{eq:drag_mstar}
{\ensuremath{\rho_{12}}}= n_i m_i \left( 1 - \frac{m_i }{ m^*_i} \right),$$ which provides the relation between the superfluid drag and the effective masses and a constraint for the effective mass ratio.
Quantum hydrodynamic model {#sec:quantum_hydro}
==========================
In general, the presence of effective masses changes the relation between static and dynamical properties of the system, and the possibility for some excitation modes to exhaust the sum rules. Let us consider a superfluid mixture with an energy density $e(n_1, n_2)$. Expanding the energy around its ground state value to second order in the density fluctuations $\Pi_i(x)$ and adding it to Eq. we obtain a hydrodynamic Hamiltonian for two miscible superfluids, $$\label{eq:Hhydro}
H=\frac{1}{2}\sum_{ij}\int\left(
\rho_{ij} \frac{\hbar \nabla \phi_i}{m_i} \cdot
\frac{\hbar \nabla \phi_j}{m_j}
+\alpha_{ij}\Pi_i\Pi_j\right)\, d^D x,$$ where the matrix $\alpha_{ij}=\partial^2e/\partial n_i\partial n_j$ contains the information on inter- and intra-species interactions of the two fluids. Hamiltonian by requiring that the fields $\phi_i$ and $\Pi_j$ satisfy canonical commutation relations for bosons, i.e., $[\phi_i(x), \Pi_j(y)] = i\hbar \delta_{ij}\delta(x-y)$.
For the sake of clarity, we take the two superfluids to be equal: $\rho_{ii}=\rho$, $\alpha_{ii}=\alpha$, $m_i = m$ and $n_i = n/2$, with $n
\equiv N/V$ the total number density of the system (see Appendix \[app:nonZ2\] for the non-symmetric case). Due to the assumed $\mathbb{Z}_2$ symmetry, the dynamics of this model decouples if we rewrite it in terms of the new fields $$\phi_{d(s)} = (\phi_1 \pm \phi_2)/\sqrt{2}, \quad
\Pi_{d(s)} = (\Pi_1 \pm \Pi_2)/\sqrt{2}.$$ The fields $\Pi_d$ and $\phi_d$ represent the fluctuations in total density and global phase, respectively. In a similar fashion, $\Pi_s$ and $\phi_s$ encode the fluctuations of the difference in density of the two species (magnetisation) and their relative phase (spin wave), respectively. We use the labels $d(s)$ to indicate the density (spin) channel of the system’s excitations. The new fields inherit the canonical commutation relations and act as two independent hydrodynamic modes, obeying the Hamiltonian $$\label{eq:Hhydro_normal_modes}
H = \frac{1}{2} \sum_{i=d,s} \int
\left[\rho_{i} \left(\frac{\hbar\nabla\phi_i}{m}\right)^2 +
\alpha_{i} \Pi_i^2 \right] \, d^Dx,$$ where $\rho_{d(s)}=\rho \pm {\ensuremath{\rho_{12}}}$ and $\alpha_{d(s)}=\alpha\pm
\alpha_{12}$. The Hamiltonian is now diagonal in the two channels, and the dispersion relations for the two modes are linear in the momentum $k$, of the form $$\begin{aligned}
\label{eq:hydro_dispersion}
(\hbar \omega)^2 = \frac{\alpha_i \rho_i}{m^2} (\hbar k)^2, \qquad (i=d,s).\end{aligned}$$ The quantity $\alpha_i \rho_i / m^2$ can be identified with the speed of sound of each mode. For the density mode, we have $$\label{eq:cd}
c_d^2 = \frac{(\rho + {\ensuremath{\rho_{12}}})}{m^2} (\alpha + \alpha_{12})
= \frac{n}{2m}(\alpha + \alpha_{12}),$$ where in the last equality we used that, at $T=0$, the total superfluid density is equal to the total mass density of the system. We note in particular that $c_d$ is independent of the superfluid drag. On the other hand, the spin speed of sound is $$\label{eq:cs}
c_s^2 = \frac{(\rho - {\ensuremath{\rho_{12}}})}{m^2} (\alpha - \alpha_{12})
= \frac{n}{2m}\left(\frac{2m}{m^*} - 1 \right)(\alpha - \alpha_{12}),$$ which explicitly depends on ${\ensuremath{\rho_{12}}}$.
From the Hamiltonian , the static response to density and to spin probes are simply given by $\alpha_{d(s)}$, which can be identified with the inverse compressibility $\kappa_d^{-1}$ and inverse magnetic susceptibility $\chi_s^{-1}$, respectively. We thus obtain the relations $$\begin{aligned}
c_d^2 &= \frac{n}{2m\kappa_d}, \label{eq:cd_kappa} \\
c_s^2 &= \frac{\rho-{\ensuremath{\rho_{12}}}}{m^2\chi_s}
= \frac{n}{2m\chi_s} \left(\frac{2m}{m^*} - 1 \right).
\label{eq:cs_chi}\end{aligned}$$ These relations suggest that, by independently measuring $c_s$ and $\chi_s$, it is possible to obtain the strength of the mass renormalisation, i.e., the magnitude of the superfluid drag. Note that $c_s$ is expected to vanish for $m^* = 2m$, which imposes a bound $m^* \leq 2m$. From Eq. , this bound translates into ${\ensuremath{\rho_{12}}}\leq mn / 4$, thus anticipating result , of which we will provide an additional derivation below.
The previous analysis has important consequences with respect to Bijl-Feynman relations (f-sum rule) linking the dispersion relations to the static structure factors (see, e.g., [@GiulianiVignale]). From the above discussion, it turns out that the f-sum rule for the density channel is exhausted by the phonon mode, while for the spin mode this is not the case, leading to the effective mass correction in the determination of the dispersion relation. In particular, the zero temperature spin structure factor at low momenta reads $$\label{eq:Ssk_vs_chi}
S_s(k) \stackrel{k\to0}{=} \frac{k}{2 m c_s} \frac{(\rho - {\ensuremath{\rho_{12}}})}{m}
=\frac{k}{2m} \sqrt{(\rho - {\ensuremath{\rho_{12}}}) \chi_s},$$ which does not satisfy the Bijl-Feynman relation. Notice that the linear term in $k$ of the $S_s(k)$ can vanish, either because of a vanishing susceptibility or because of a saturated drag, i.e., ${\ensuremath{\rho_{12}}}= \rho$. The former (latter) case corresponds to a vanishing (diverging) spin speed of sound. The fact that the drag and the interspecies interaction act independently on the spin speed of sound \[cf. Eq. \] is general, and applies beyond the $\mathbb{Z}_2$ symmetry we assumed in this section. In particular, the standard condition for the onset of phase separation (i.e., $\chi_s\rightarrow \infty$ for $\alpha=\alpha_{12}$) still holds (cf. Appendix \[app:nonZ2\]).
On the other hand, due to translational invariance, the density structure factor satisfy the Bijl-Feynman relation and it reads $$\label{eq:Sdk_vs_kappa}
S_d(k) \stackrel{k\to0}{=} \frac{n k}{4 m c_d}
=\frac{k}{2m} \sqrt{2m n \kappa_d}.$$
Let us conclude this section by briefly mentioning the effect of the AB physics on the specific heat of the mixture. At low but finite temperature, we may expect that thermal fluctuations do not change the low energy spectrum significantly. Then the low temperature dispersion relations are still linear, of the form, $\epsilon_i(k) = c_i k$, $(i=d,s)$, and we assume, within the hydrodynamic picture, that the highest momentum that can be thermally excited is $k_{T,i} = k_BT/c_i$. In the low temperature limit and $D$ spatial dimensions, these assumptions lead to the specific heat $$C_v \propto T^{D}\left(\frac{1}{c_d^D}+\frac{1}{c_s^D}\right),$$ which carries a dependence on ${\ensuremath{\rho_{12}}}$ through the sound velocity in the spin channel. A large superfluid drag will therefore lead to a strong increase of the specific heat.
One dimensional systems and Luttinger Liquid
--------------------------------------------
Since the superfluid drag is due to quantum fluctuations, one can think about increasing them by increasing the interactions, i.e., quantum depletion and mutual dressing. This can be easily seen in the weakly interacting regime, where analytical expressions for the superfluid drag have been nicely obtained within a Bogoliubov approach by Fil and Shevchenko [@PhysRevA.72.013616; @LowTempPhys.30.770]. In a three-dimensional system, three-body losses strongly limit the possible increase of the interaction strengths. On the other hand, in one dimension, it is possible to reach strong quantum regimes, including the so-called Tonks-Girardeau regime. The low-energy excitations of one-dimensional gases are described in terms of Luttinger liquids [@Giamarchi]. For the sake of simplicity, we consider two equal Luttinger liquids coupled together, with speed of sound $c_0$ and Luttinger parameter $K_0\ge 1$. By introducing both the density-density and the current-current couplings as a perturbation, we can write $$\begin{aligned}
H_{LL}&=&\sum_{i=1,2} \frac{c_0}{2}
\int [K_0 (\partial_x \phi_i)^2 + \frac{1}{K_0} \Pi_i^2] \nonumber \\
&+& \int [{\ensuremath{\rho_{12}}}\frac{\partial_x \phi_1 \partial_x \phi_2}{m^2}
+ g_{12} \Pi_1 \Pi_2].
\label{eq:H_luttinger}\end{aligned}$$ As before, we can easily diagonalise the Hamiltonian by introducing the fields for the in-phase and out-of-phase fluctuations. We obtain a standard expression for coupled Luttinger liquids $$H_{LL}=\sum_{i=d,s} \frac{c_i}{2}
\int [K_i (\partial_x \phi_i)^2 + \frac{1}{K_i} \Pi_i^2],$$ where the density parameters read $$\begin{aligned}
2mc_{d}^2 &= n (c_0/K_0+g_{12}), \\
2mK_{d}^2 &= n/(c_0/K_0+g_{12}),\end{aligned}$$ and for the spin sector we get $$\begin{aligned}
\label{spinLL}
2mc_{s}^2 &= (n-4{\ensuremath{\rho_{12}}}/m)(c_0/K_0-g_{12}), \\
2mK_{s}^2 &= (n-4{\ensuremath{\rho_{12}}}/m)/(c_0/K_0-g_{12}).\end{aligned}$$ In the above expressions, $n$ is the total density of the system, and we have used the fact that, for a translationally invariant system, $c_d K_d = n / 2m$ (see also [@condmat.9807366]), which implies that $c_0 K_0 + {\ensuremath{\rho_{12}}}/ m^2 = n / m$. Therefore, $c_d$ and $K_d$ do not depend on the off-diagonal superfluid density and for the compressibility we have $\kappa = K_d / c_d = n / (2m c_d^2)$. On the other hand, the spin channel parameters acquire a dependence on ${\ensuremath{\rho_{12}}}$, as seen before in the general case. In fact, the susceptibility reads $\chi_s = K_s / c_s = (n - 4 {\ensuremath{\rho_{12}}}) / (2 m c_s^2)$, to be compared with Eq. . The correction due to AB in Eq. , in the strongly interacting limit can therefore deeply modify the standard perturbative analysis [@NewJPhys.10.045025; @PhysRevA.77.013607] and the RG flow for coupled Bose Luttinger liquids. Note, once again, that the previous equations imply a bound on the value of the superfluid drag, ${\ensuremath{\rho_{12}}}\leq nm / 4$, which coincides with the one coming from Eq. in the previous section. Recent Monte-Carlo simulations on one dimensional Bose gases confirm our results [@parisi_giorgini_private].
Magnitude of the drag and numerical evidence {#sec:numerics}
============================================
The information on the superfluid drag can be extracted from quantum Monte-Carlo (QMC) simulations based on the path integral formalism. In this formalism, in fact, the superfluid density can be related to the statistics of winding numbers of particles’ paths around the simulation domain [@PhysRevB.36.8343]. By extending this result to two species in the same simulation box (see Appendix \[sec:app\_path\_integral\] for details), we obtain the relation $$\begin{aligned}
\label{eq:winding_drag}
\rho_T & = \frac{L^{2-D}}{\beta D}
\left[m_1^2 {\ensuremath{\langle W_1^2 \rangle}} + m_2^2 {\ensuremath{\langle W_2^2 \rangle}} + 2 m_1 m_2 {\ensuremath{\langle W_1 W_2 \rangle}}
\right] \nonumber \\
& = \rho_{1} + \rho_{2} + 2 {\ensuremath{\rho_{12}}},\end{aligned}$$ linking the total superfluid density $\rho_T$ to the winding numbers $W_{1,2}$ of the two species. Here we are considering a simulation volume $L^D$ at inverse temperature $\beta = 1/T$. For zero temperature results, $T$ is taken smaller than all the other energy scales of the system and the results are checked *a posteriori* for convergence.
From Eq. , the superfluid drag can be interpreted as the covariance between the superfluid densities of the two components. Then, thanks to Cauchy-Schwarz inequality, ${\ensuremath{\rho_{12}}}^2 \leq \rho_1 \rho_2$, and assuming the symmetric case, in which $\rho_1 = \rho_2$, we obtain an upper bound on the magnitude of the drag, $$\label{eq:drag_bound}
{\ensuremath{\rho_{12}}}\leq \frac{\rho_T}{4}, \qquad (\rho_1 = \rho_2)$$ which also bounds the effective mass to $m^* \leq
2m$. As already noted above, the condition of saturation of this bound corresponds to a vanishing speed of sound in the spin channel \[cf. Eqs.-\]. We shall also see below that the saturation of this bound is a limiting case in the dynamic stability of the mixture \[see Eq. \].
Quantum Monte Carlo results for bilayer dipolar Gases
-----------------------------------------------------
Lattice simulations already showed evidence of superfluid drag effects [@PhysRevA.86.033627; @PhysRevLett.95.090403]. It was shown that the drag depends on the lattice geometry, increases with the increase of interspecies interactions and attains its maximum for non-equal masses of the two particle species. The presence of the lattice explicitly breaks the translational invariance, thus deeply modifying the mechanism leading to a dissipationless drag. In particular, for incommensurate fillings, the drag between the two fluids is essentially mediated by the presence of vacancies [@PhysRevLett.95.090403].
The magnitude of the superfluid drag, normalised by the total superfluid density, spans the whole range allowed by bound . However, one must point out that the presence of the lattice causes the depletion of the total superfluid density; in particular, on a lattice, it is no longer true that the total superfluid density coincides with the total particle density at zero temperature [@Leggett1998]. In this context, it is noteworthy to mention the analytical results of Ref. [@PhysRevA.79.063610], which compute the superfluid drag starting from the physical parameters of the lattice in a weak coupling approximation. The authors report a superfluid drag ${\ensuremath{\rho_{12}}}/ nm$, normalised to the total mass density, of the order of $10^{-5}$–$10^{-4}$ for weak to moderate intercomponent scattering amplitude. Quantitatively similar QMC results are reported in [@PhysRevA.86.033627]. It is important to keep in mind that these low values are primarily due to the small total superfluid density on the lattice.
![Schematic representation of the bilayer dipolar bosonic model. Particles obeying Bose-Einstein statistics are confined in two parallel layers. The dipoles are pinned perpendicularly to the layers’ planes and parallel to each other, yielding entirely repulsive on-plane interactions and partially attractive out-of-plane interactions.[]{data-label="fig:BDB"}](bilayer_diagram)
In the following, we focus on a system of dipolar Bose gases confined in a bilayer geometry in continuous space, with the dipole orientation pinned perpendicular to the planes, as sketched in Fig. \[fig:BDB\]. This system is similar to the one studied in Ref. [@LowTempPhys.25.493], which pointed out the presence of entrainment between the superfluid currents of two charged superfluids in a bilayer configuration. The relative strength of interspecies interactions as compared to intraspecies ones can be tuned by changing the distance between the two layers. As it will be shown later in Fig. \[fig:dqmc\], for an extended range of this control parameter, the superfluid drag can reach very large values. Dipolar particles in a bilayer configuration are particularly advantageous under a variety of aspects. Confining the molecules in a two-dimensional geometry and imposing a repulsive dipolar interaction strongly reduces the detrimental two-body chemical reactions [@Ni2010]. At the same time it allows to exploit the anisotropy of the dipolar interaction, which is partially attractive between particles on different layers. Introducing the distance $h$ between the two layers, the interaction between two particle of mass $m$ and dipole moment $d$ on different layers can be written as $$\label{eq:dipolar_potential}
V(r,h)=d^2\frac{r^2-h^2}{(r^2+h^2)^{5/2}},$$ where $r$ is the relative distance in the plane of motion. For dipolar gases, it is very useful to introduce the characteristic length $r_0=md^2/\hbar^2$. The various regimes of the system are characterised by the interlayer parameter $h/r_0$ and the in-layer parameter $n_i r_0^2$, with $n_i$ the single layer density. Static and dynamic properties of this system were recently investigated in [@PhysRevA.90.043623; @PhysRevA.94.063630]. In particular, it has been found that a transition from two coupled superfluid (atomic phase) to a pair superfluid (molecular phase) takes place when the attractive interaction is strong enough. We will show that, by approaching the transition point while remaining in the atomic phase, the drag superfluidity becomes prominent.
To recover the description of Eq. , we point out that miscibility is here to be intended with respect to the position of the particles projected in the direction orthogonal to the layers’ planes. Dipolar Bose gases in a double layer configuration do not show any phase separation [@PhysRevA.90.043623]. This is intuitive, since any potential with $V(q)|_{q=0}\leq 0$ (in momentum space) admits a bound state in $2D$. (The interlayer potential \[eq:dipolar\_potential\] has the peculiarity to have $V(q)|_{q=0}=0$, which makes the bound and scattering states of the system at weak coupling rather peculiar [@PhysRevA.82.044701]). The dipolar bilayer Bose gas is moreover the first example of a two component Bose gas which can form pairs without collapsing (i.e. forming clusters) [@PhysRevA.90.043623] as it occurs, e.g., in mixtures with contact interaction only.
Besides serving as a testbed for the numerical study of the superfluid drag in a homogeneous geometry and being a new system showing the AB physics, the dipolar bilayer configuration can represent one of the best-case scenarios for the experimental observation of the presence of superfluid drag. Recent experiments using dipolar molecules consisting of two atoms of Erbium-168 demonstrated the availability of condensates with large magnetic moments, up to $r_0 \approx 1600\, {\rm a}_0$, with ${\rm a}_0$ the Bohr radius [@PhysRevLett.115.203201]. This value is still almost one order of magnitude smaller with respect to the typical wavelength of the lasers used to confine the Er$_2$ molecules in arrays of 2D layers. Experiments on stable polar Na-K molecules [@MartinNaK], which sport much larger $r_0$, may help in overcoming this problem. The recent proposal of sub-wavelength confinement [@PhysRevLett.115.140401] may further stretch the experimentally accessible range of values of $h/r_0$, albeit it is not of easy implementation for dipolar molecules. Therefore, experimental realisation of a bilayer system of strongly interacting dipolar superfluids is reasonably within reach of current or near-future technology.
We study the system by means of diffusion QMC, which allows us to extract both the thermodynamics and the low energy spectrum of the system. Diffusion QMC is based on solving the Schrödinger equation in imaginary time, thus projecting out the ground state of the system (for a general introduction on the method see, e.g., [@PhysRevB.49.8920]). The contributions of the excited states are exponentially suppressed and the ground-state energy is recovered in the limit of long propagation time. The simulations are performed for $60$ particles with the same parameters as in Refs. [@PhysRevA.90.043623; @PhysRevA.94.063630]. For some quantities, this number of particles is sufficiently large to be close to the thermodynamic limit; for some others, residual finite-size corrections must be taken into account, as it will be explained in more details later.
In this framework, a number of observables of interest can be obtained in a straightforward way. The value of the gap $\Delta$ and of the spin susceptibility $\chi_s$ are obtained from the dependence of the ground-state energy on the polarisation $P = (N_1-N_2) / N$. The latter is tuned by moving particles from layer $1$ to layer $2$ while keeping the total number of particles $N = N_1 +N_2$ constant. In the limit of small polarisation $P$, the energy can be expanded as $$\label{eq:susceptibility}
E(P) = E(0) + N \Delta \cdot P + N \frac{n}{2 \chi_s} \cdot P^2.$$ In the gapless phase ($\Delta=0$) the dependence on the polarisation is quadratic, while in the gapped phase it is linear. Similarly, the compressibility $\kappa_d$ at $T=0$ can be obtained from the volume dependence of the energy for an unpolarised gas, $$\label{eq:compressibility}
\kappa_d^{-1} = -\mathcal{V} \left(
\frac{\partial^2 E}{\partial \mathcal{V}^2}
\right),$$ where $\mathcal{V}$ is the $D$ dimensional volume of the system ($\mathcal{V}
= L^2$ in the $2D$ geometry at hand).
The study of structure factors provides a way of accessing the dynamic properties of the system. We use the technique of pure estimators [@PhysRevA.10.303; @PhysRevB.52.3654] to compute the intermediate scattering function $$S_{\alpha\beta}({\bf k},\tau) = \frac{1}{N}
\left\langle
\rho_\alpha ({\ensuremath{{\bf k}}}, \tau) \rho_\beta(-{\ensuremath{{\bf k}}}, 0)
\right\rangle,$$ with $\rho_\alpha({\ensuremath{{\bf k}}}, \tau) = \sum_j^{N_\alpha} \exp\lbrace -i{\ensuremath{{\bf k}}}
\cdot {\ensuremath{{\bf r}}}_{j\alpha}(\tau) \rbrace$ and ${\ensuremath{{\bf r}}}_{j \alpha}$ the position of particle $j$ in layer $\alpha$. The intermediate scattering function provides information on the correlations in imaginary time $\tau$ and is the main ingredient to compute the static structure factor $S_{\alpha \beta}({\ensuremath{{\bf k}}}) \equiv S_{\alpha
\beta}({\ensuremath{{\bf k}}}, 0)$. We consider a balanced system with $N_A = N_B$ and study the symmetric and antisymmetric structure factors, $$S_{d(s)}(k) = S_{11}(k) \pm S_{12}(k),$$ corresponding to the density and spin channels of the discussion above, respectively. The compressibility and the spin susceptibility can be compared to the respective static structure factors in the low momentum limit, in order to verify the sum rules. The structure factors further provide information on the excitation spectra: their long imaginary time asymptotic behaviour can be fitted to an exponential decay of the form $$\label{eq:omega_decay}
S_{d(s)}({\ensuremath{{\bf k}}}, \tau) \sim Z e^{ -\omega_{d(s)}(k) \tau}.\qquad
(\tau\to\infty)$$ When phononic excitations are present, $\omega_{d(s)}(k)$ is linear for small momenta, with the slope directly related to the speeds of sound of the density and spin channels, respectively, through $\hbar \omega_{d(s)}(k) \simeq
c_{d(s)} k$.
![Speeds of sound as a function of interlayer spacing $h$ for $nr_0^2 =
1$, as extracted from different observables. (a) Speed of sound of the density mode, $c_d$. (b) Speed of sound of the spin mode, $c_s$. The Feynman method makes use of the static structure factor $S(k) = \hbar k /
(2mc)$, computed at the smallest $k$ compatible with periodic boundary conditions. The speeds $c_{\alpha}[S_{\alpha}(k)]$, with $\alpha = \lbrace d,s \rbrace$ the channel index, are computed with this method. The data show that the f-sum rule is exhausted by the phonon mode in the density channel, whereas this does not hold in the spin channel, the arrow indicating the divergence of $c_s[S_s(k)]$. The speeds $c_{\alpha}(\omega_\alpha)$, computed from the excitation spectrum, assume a linear phononic dispersion relation $\omega(k) = ck$ with $\omega_\alpha(k)$ obtained from Eq. . The speeds $c_d(\kappa_d)$ and $c_s(\chi_s)$ are computed from $mc_d^2 = n
\kappa_d^{-1}$ and $mc_s^2 = n \chi_s^{-1}$, with $\kappa_d$ and $\chi_s$ obtained from Eqs. and , respectively. The speed of sound in the atomic limit coincides in the two channels. It is obtained from standard thermodynamic relations using the equation of state $E(nr_0^2)$ (taken from Ref. [@PhysRevA.75.063630]) of a single layer system with half the density of the bilayer system.[]{data-label="fig:dqmc:c"}](dqmc_c){width="8.5cm"}
It is instructive to show that, in a gapless system without the drag, exactly the same information on the speeds of sound can be recovered from the static structure factors $S_{d,s}(k)$, the low-momentum excitation spectra $\omega_{d,s}(k)$, the compressibility $\kappa_d$ and the susceptibility $\chi_s$. The speeds of sound obtained from the different methods are shown in Fig. \[fig:dqmc:c\] for the density (a) and spin (b) modes. The density mode is gapless for any value of the interlayer separation $h$. The speed of sound of this channel, as obtained from structural, energetic and thermodynamic quantities, always yields compatible values throughout the explored range of $h$. Finite-size effects reduce the speed of sound, which for large $h$ (decoupled layers) appears to lie below its asymptotic value. The latter is obtained from the equation of state of a model with a single species at half the density (so called “atomic” limit). In the computations, the dipolar interaction potential was truncated at a distance equal to half the size of the simulation box. By adding the missing “tail” correction to the compressibility, it is possible to recover correct atomic limit asymptotics, as shown in the figure. The situation is quite different for the spin channel, where the gap opens for $h/r_0 \lesssim 0.35$ and different methods cannot be consistent in that parameter range. For large values of the interlayer separation, $h/r_0 \gtrsim
0.6$, we recover once again the atomic limit and different quantities are consistent with one another. Manifestly, it is not the case for parameter range $0.35 \lesssim h / r_0 \lesssim 0.6$, which still corresponds to a gapless phase but is in the vicinity of the transition point. In this region there is no consistency between the speeds of sound obtained with different methods and, importantly, the f-sum rule is not satisfied. The reason for this is appearance of the superfluid drag which we analyse in more details below. It is interesting to note that the finite-size effects are more pronounced in the density mode compared to the spin mode. One way to understand this is that the spin mode probes the response to the polarisation, which does not change the system volume, while the density (compression) mode is the response to a change in volume. The tail correction, being sensitive to the change of the volume, is able to account for the finite-size discrepancy.
Finally, in order to directly probe the superfluidity properties of the system, we introduce the winding number related to species $\alpha$, $${\ensuremath{{\bf W}}}_{\alpha}(\tau) = \sum_{i(\alpha) = 1}^{N/2} \int_0^\tau d\tau'
\frac{d{\ensuremath{{\bf r}}}_{i(\alpha)}(\tau')}{d\tau'},$$ where $i(\alpha)$ indexes particles belonging to species $\alpha$ only. Taking the limit of long propagation time, the statistics of the winding numbers are related to the superfluid densities. In particular, we evaluate the symmetric and antisymmetric combinations $$\label{eq:qmc_winding}
\rho_1 + \rho_2 \pm 2{\ensuremath{\rho_{12}}}= \lim_{\tau \to \infty}
\frac{\langle [{\ensuremath{{\bf W}}}_1(\tau) \pm {\ensuremath{{\bf W}}}_2(\tau)]^2 \rangle}{2 N \tau}.$$ According to Eq. , when the plus sign is considered, the quantity above is an estimator of the total superfluid density of the system, $\rho_T$. In our zero temperature simulation, this quantity is always compatible with the total mass density, as it should be in continuous space, and in contrast with the low total superfluidity observed in lattice simulations. When the minus sign is considered, on the other hand, we directly probe the magnitude of the superfluid drag. This observable asymptotically attains the value $\rho_T$ in the non interacting ($h \to \infty$) limit. In the symmetric mixture case, borrowing the same notations of Sec. \[sec:quantum\_hydro\], this quantity reduces to $(\rho - {\ensuremath{\rho_{12}}}) / nm$, and is reported in Fig. \[fig:dqmc\]. For low interactions (large $h$), it is compatible with unity and drops to zero as interactions are ramped up. This corresponds to ${\ensuremath{\rho_{12}}}= \rho_T / 4$, i.e., with bound . It must be noted that, for $h < h_c
\approx 0.35 r_0$[@PhysRevA.90.043623], the system enters the molecular phase, so that not all the drop in $\rho - {\ensuremath{\rho_{12}}}$ can be ascribed to an increase of the drag, but one must also keep into account the emergence of the molecular condensate. Indeed, the description we put forward holds only as long as the system is still in the atomic phase. Interestingly, the saturation of the bound (or, alternatively $m^*\rightarrow 2m$) appears to coincide with the transition to the molecular phase in which bound state physics start dominating.
![ The quantity $(\rho-{\ensuremath{\rho_{12}}})$, extracted from diffusion QMC data using different estimators. The direct winding number estimator (diamonds and dashed line to guide the eye) is compared with indirect estimators. The latter make use of the relations derived in the quantum hydrodynamic model of Sec. \[sec:quantum\_hydro\] (cf. also Eq. ). The data sets are in satisfactory agreement with one another, with the exception of the estimator $c_s^2 [S_s(k)] \chi_s$, which tends to diverge as the molecular phase is approached (shaded region). The origin of the errorbars and of remaining finite-size effects are discussed in the main text.[]{data-label="fig:dqmc"}](dqmc)
By independently measuring $S_s(k)$, $\chi_s$ and $\omega_s$ (hence $c_s$), we can show that the usual Bijl-Feynman approximation is not applicable to systems in the presence of the superfluid drag. To this end, we use Eqs.- to express $(\rho-{\ensuremath{\rho_{12}}})$ in terms of the observables listed above, and then compare it with the direct winding number measurement. Figure \[fig:dqmc\] reports the data corresponding to the three independent expressions $$\label{eq:many_rho12}
\frac{\rho-{\ensuremath{\rho_{12}}}}{nm}
= \frac{4mS^2_s(k)}{\chi_s k^2}
= 2 \frac{c_s(\omega_s) S_s(k)}{k}
= c_s^2(\omega_s) \chi_s,$$ where we are indicating by $c_s(\omega_s)$ the speed of sound in the spin channel as extracted from the fit to Eq. . The data show fair agreement among the above expressions and between them and the direct measurement of $\rho - {\ensuremath{\rho_{12}}}$. A notable exception are the data for $c_s^2[S_s(k)]/\chi_s$, where $c_s[S_s(k)]$ is the speed of sound in the spin channel, computed *as if* Bijl-Feynman relation held. We observe that this expression leads to the wrong behaviour in the region where $\rho-{\ensuremath{\rho_{12}}}$ differs from one, i.e., where the drag effect is more prominent.
In both Figs \[fig:dqmc:c\] and \[fig:dqmc\], the errobars in the quantities extracted from the static structure factor and the winding number come from statistical averaging. Spectral frequency, compressibility and susceptibility have additional contributions to the error due to the use of fitting procedures, Eqs. and . The errors are effectively increased in some of the results, due to cancellation of opposite trends as a function of $h$. The finite-size effects are important for a quantitative agreement, as can be seen from Fig. \[fig:dqmc:c\]a. It is not obvious that different quantities have similar finite-size correction, which might eventually be responsible for some remaining differences between various estimations in Fig. \[fig:dqmc\].
Dynamic stability {#sec:dynamic_stability}
=================
Given the recent advances in measuring the spin superfluidity and its critical dynamics, we devote the final section to explore the consequences of the presence of a superfluid drag term on such phenomena. Both long living spin oscillations [@PhysRevA.94.063652] and critical spin superflow [@PhysRevLett.119.185302] for Bose-Bose mixtures as well as for Fermi-Bose superfluid mixtures [@PhysRevLett.115.265303] have been measured in the weakly interacting regime. The agreement with the available estimates [@ComptesRendusPhysique.16.241; @EurPhysJD.69.126] is reasonable but rather far from being quantitative.
In the following, we determine the critical relative velocity required to trigger the dynamical instability of a binary mixture of superfluids at zero temperature. To this end, we generalise Eqs. and and the results of Refs. [@JLowTempPhys.150.612; @JLowTempPhys.155.219; @EurPhysJD.69.126] by considering the energy functional $$\begin{aligned}
E[&n_1, n_2, \phi_1, \phi_2] = \nonumber \\
&= \int \left\lbrace \frac{\hbar^2}{2}
\sum_\alpha \left[ \frac{m_\alpha n_\alpha - {\ensuremath{\rho_{12}}}}{m_\alpha^2}
(\nabla \phi_\alpha)^2 \right]
\right.
\nonumber \\
&\left. \quad
+\frac{{\ensuremath{\rho_{12}}}}{m_1 m_2} \nabla \phi_1 \cdot \nabla \phi_2
+ e(n_1, n_2)\, \right\rbrace d^Dx\end{aligned}$$ with $e(n_1, n_2)$ the internal energy density. We subtracted ${\ensuremath{\rho_{12}}}$ from the diagonal kinetic terms, so that the condition on the total density is automatically satisfied. Considering $n_\alpha$ and $\phi_\alpha$ as conjugate variables, the Hamilton equations $m_\alpha \partial_t n_\alpha = \delta E / \delta \phi_\alpha$ and $m_\alpha \partial_t \phi_\alpha = - \delta E / \delta n_\alpha$, yield (implying $\alpha \neq \beta$) $$\begin{aligned}
& \partial_t n_\alpha =
-\nabla(n_\alpha {\ensuremath{{\bf v}}}_\alpha)
+ m_\alpha^{-1}\nabla \left[{\ensuremath{\rho_{12}}}({\ensuremath{{\bf v}}}_\alpha - {\ensuremath{{\bf v}}}_\beta)
\right], \label{eq:ndot} \\
& \partial_t {\ensuremath{{\bf v}}}_\alpha =
- \nabla \frac{v_\alpha^2}{2}
+ \nabla \left[
\frac{\eta_\alpha}{m_\alpha}
\left( \frac{v_\alpha^2}{2} - {\ensuremath{{\bf v}}}_1 \cdot {\ensuremath{{\bf v}}}_2 \right)
\right]
- \frac{\nabla \mu_\alpha}{m_\alpha},
\label{eq:vdot}\end{aligned}$$ where ${\ensuremath{{\bf v}}}_\alpha = (\hbar / m) \nabla \phi_\alpha$ are the superfluid velocities, $\mu_\alpha = \partial \epsilon / \partial n_\alpha$ the chemical potentials and we defined $\eta_\alpha \equiv \partial {\ensuremath{\rho_{12}}}/ \partial n_\alpha$, implicitly assuming that ${\ensuremath{\rho_{12}}}$ is a well-behaved function of the densities. The previous system of hydrodynamic equations is satisfied by a steady state solution of uniform velocity, matter and drag fields, such that all gradient terms vanish. With a slight change of notation, we perturb about this solution by expanding the density and velocity fields as $$\begin{aligned}
& n_\alpha \mapsto \bar{n}_\alpha + n_\alpha e^{i({\ensuremath{{\bf qr}}} - \omega t)}, \\
& {\ensuremath{{\bf v}}}_\alpha \mapsto \bar{{\ensuremath{{\bf v}}}}_\alpha + {\ensuremath{{\bf v}}}_\alpha
e^{i({\ensuremath{{\bf qr}}} - \omega t)},\end{aligned}$$ and only keep the first order in the fluctuations, whereas, since it is already small with respect to $n_\alpha m_\alpha$, we only keep the zero-th order in ${\ensuremath{\rho_{12}}}$. Substituting these expansions into Eqs. -, we obtain $$\begin{aligned}
\label{eq:n_expanded}
& n_\alpha [\omega - \bar{{\ensuremath{{\bf v}}}}_\alpha \cdot {\ensuremath{{\bf q}}}] =
\bar{n}_\alpha {\ensuremath{{\bf v}}}_\alpha \cdot {\ensuremath{{\bf q}}}
- \frac{{\ensuremath{\rho_{12}}}}{m_\alpha}
({\ensuremath{{\bf v}}}_\alpha - {\ensuremath{{\bf v}}}_\beta) \cdot {\ensuremath{{\bf q}}},\\
& {\ensuremath{{\bf v}}}_\alpha [\omega - \bar{{\ensuremath{{\bf v}}}}_\alpha \cdot {\ensuremath{{\bf q}}}] =
\frac{{\ensuremath{{\bf q}}}}{m_\alpha} \sum_{\gamma=1,2}
\mu_{\alpha\gamma} n_\gamma,
\label{eq:v_expanded}\end{aligned}$$ where we abbreviated $\mu_{\alpha \beta} \equiv \partial \mu_\alpha / \partial
\bar{n}_\beta$ (hence $\mu_{12} = \mu_{21}$, as expected by the symmetry of interactions). In Eq. , we neglected the term proportional to $\eta_\alpha$, as the leading order is $O({\ensuremath{\rho_{12}}}n_\alpha, {\ensuremath{\rho_{12}}}v_\alpha)$, and the zero-th order, homogeneous by hypothesis, vanishes under spatial differentiation.
Eliminating $n_\alpha$ and $v_\alpha$ in the system of equations - leads to
$$\begin{aligned}
& \left\lbrace \Omega_1^2 \Omega_2 -
\left[ c_1^2 \Omega_2 -
{\ensuremath{\rho_{12}}}\left( \frac{\mu_{11}\Omega_2}{m_1^2} -
\frac{\mu_{21}\Omega_1}{m_1 m_2} \right) \right]
q^2 \right\rbrace
\left\lbrace \Omega_2^2 \Omega_1 -
\left[ c_2^2 \Omega_1 -
{\ensuremath{\rho_{12}}}\left( \frac{\mu_{22}\Omega_1}{m_2^2} -
\frac{\mu_{12}\Omega_2}{m_1 m_2} \right) \right]
q^2 \right\rbrace \nonumber
= \\
& \qquad \qquad \qquad \qquad \qquad \qquad = q^4
\left[ \bar{n}_1 \frac{\mu_{12} \Omega_2}{m_1}
- {\ensuremath{\rho_{12}}}\left( \frac{\mu_{12} \Omega_2}{m_1^2} -
\frac{\mu_{22}\Omega_1}{m_1 m_2}
\right) \right]
\left[ \bar{n}_2 \frac{\mu_{21} \Omega_1}{m_2}
- {\ensuremath{\rho_{12}}}\left( \frac{\mu_{21} \Omega_1}{m_2^2} -
\frac{\mu_{11}\Omega_2}{m_1 m_2}
\right) \right]
, \label{eq:stability}\end{aligned}$$
where we set $\Omega_\alpha \equiv \omega - \bar{{\ensuremath{{\bf v}}}}_\alpha \cdot {\ensuremath{{\bf q}}}$ and $c_\alpha^2 \equiv \bar{n}_\alpha \mu_{\alpha \alpha} / m_\alpha$, the latter being the speed of sound of a single superfluid. We also define $$c_{12}^4 \equiv \frac{\bar{n}_1 \bar{n}_2}{m_1 m_2} \mu_{12}^2,$$ which has got the dimension of a speed and measures the strength of inter-species contact interactions. By assuming a linear dispersion $\omega =
Cq$, the one above is a sixth order equation for the speed of sound $C$ of the mixture, which determines its stability. In particular, the presence of complex roots flags a dynamical instability. By setting ${\ensuremath{\rho_{12}}}= 0$, Eq. reduces to the problem of the stability of a mixture interacting only via contact interactions, which was studied in depth in [@EurPhysJD.69.126].
This equation simplifies considerably when rewritten in a symmetric frame of reference (SFR) in which $v/2 = \bar{{\ensuremath{{\bf v}}}}_1 \cdot {\ensuremath{{\bf q}}} = -
\bar{{\ensuremath{{\bf v}}}}_2 \cdot {\ensuremath{{\bf q}}}$, so that the projection of the relative velocity of the two fluids along ${\ensuremath{{\bf q}}}$ is simply $v$. When $n_1 m_1 = n_2 m_2$, this choice corresponds to the frame of reference of the centre of mass of the system. A brief analysis of the case $v=0$ is reported in the Appendix \[app:nonZ2\].
Symmetric mixture
-----------------
Some insight into the dynamic stability of the mixture can be gained by considering a $\mathbb{Z}_2$ symmetric mixture, in which $\bar{n}_\alpha =
\bar{n}$, $m_\alpha = m$ and $\mu_{\alpha\alpha}$ are the same for both species (hence also the speeds of sound coincide, $c_\alpha = c$). The stability equation then simplifies to a biquadratic equation whose roots can be readily calculated. Restricting to positive relative velocities, the condition that the two-fluid speed of sound be real then yields the critical relative velocities for the stability of the mixture: $$\label{eq:vcrit_z2}
\frac{v_{{\rm c}1}}{c} = 2 \sqrt{ 1 - \frac{c_{12}^2}{c^2}}, \quad
\frac{v_{{\rm c}2}}{c} = 2 \sqrt{\left(1 +
\frac{c_{12}^2}{c^2}\right)\left(1-4\frac{{\ensuremath{\rho_{12}}}}{\rho_T}\right)}.$$ For $c_{12} = {\ensuremath{\rho_{12}}}= 0$, the system is unstable for $v > 2 c$, confirming the results of Refs. [@JLowTempPhys.150.612; @JLowTempPhys.155.219; @ComptesRendusPhysique.16.241; @EurPhysJD.69.126]. When the relative velocity $v$ lies within $v_{{\rm c}1}$ and $v_{{\rm c}2}$, the mixture becomes unstable, as schematically shown in Fig.\[fig:stability\].
It is worth reminding that the stability analysis presented here is valid within hydrodynamics, i.e., assuming a linear dispersion relation. The inclusion of non-linear terms can give rise to finite momentum instabilities above the upper critical velocity, as it was pointed out in the case of a Bose-Bose mixture without superfluid drag [@PhysRevA.83.063602]. Whether such finite momentum instability occurs in the same way also in presence of AB corrections is beyond the scope of the present paper and it will be discuss elsewhere. Nevertheless for large enough drag, $$\frac{{\ensuremath{\rho_{12}}}}{\rho_T} >
\frac{1}{4} \left[ 1 - \frac{c^2 - c_{12}^2}{c^2 + c_{12}^2}\right].$$ the lower critical velocity is $v_{{\rm c}2}$, which depends on the entrainment: in this regime, it should be possible observe a shift in the onset of the dynamical instability. In particular for ${\ensuremath{\rho_{12}}}=
\rho_T/4$, i.e., when the condition is saturated and the spin speed of sound vanishes, the mixture is unstable already at vanishing relative velocities.
For completeness, we compare the critical velocities for the dynamic stability with the speeds of sound of the density and spin modes, Eqs. -, respectively, which provide the critical velocity for the energetic (Landau) instability. With the notation of this section, they read $$\label{eq:energetic_instability_thresholds}
c_d^2 = c^2 + c_{12}^2, \qquad
c_s^2 = \left(1 - 4 \frac{{\ensuremath{\rho_{12}}}}{\rho_T} \right) (c^2 - c_{12}^2),$$ and they are also reported in Fig. \[fig:stability\]. Although it occurs at a lower critical velocity, the energetic instability would not obfuscate the dynamical instability. A counterflow experiment would only slightly trigger the Landau instability, which in any case would develop much more slowly than the dynamical one. Experimental methods based on the generation of soliton trains [@PhysRevA.84.041605; @PhysRevLett.106.065302; @PhysRevLett.119.185302] were shown to be able to cleanly identify dynamical instabilities of the kind discussed in this section.
![Stability diagram for a superfluid mixture with contact interactions such that $c_{12} = 0.6$ and all speeds expressed in units of $c$. The spin speed of sound $c_s$ is always below the density speed of sound $c_d$ (cf.Eq. ). The former delimits the energetically unstable (EU) region (solid filling). The dynamically unstable (DU) region (grid filling) lies between the critical velocities $v_{{\rm c}1}$ and $v_{{\rm c}2}$, cf. Eq. . The threshold relative velocities for both EU and DU vanish for ${\ensuremath{\rho_{12}}}/ \rho_T = 1/4$.[]{data-label="fig:stability"}](vcrit_c12_06)
Conclusions {#sec:conclusions}
===========
In the present work, we studied the physics of a superfluid mixture in the presence of current-current interactions, which lead to the so-called Andreev-Bashkin effect (AB). Our minimal quantum hydrodynamic theory highlights the consequences of the presence of the superfluid drag ${\ensuremath{\rho_{12}}}$ as an off-diagonal coefficient of the superfluid density matrix. In particular, for $\mathbb{Z}_2$ symmetric mixtures, in which the density and spin modes decouple, we predict the density channel to remain unaffected; the spin channel, on the other hand, exhibits corrections of linear order in ${\ensuremath{\rho_{12}}}$ in both static quantities, such as the spin susceptibility $\chi_s$, and in dynamic quantities, such as the speed of sound of the spin mode, $c_s$. A prominent consequence of these corrections is that Bijl-Feynman theory, relating $c_s$ to $\chi_s$, is no longer satisfied in the presence of AB corrections, as shown by Eq. . In light of these results, we identify the speeds of sound and the susceptibilities as being the most promising observables in order to experimentally study the AB effect. The quantum hydrodynamic theory further provides an upper bound to the value of ${\ensuremath{\rho_{12}}}$, which must remain less than a fourth of the total superfluid density $\rho_T = \rho_1 + \rho_2 + 2{\ensuremath{\rho_{12}}}$ (alternatively, in the language of the effective mass, $m^* \leq 2m$). The same bound appears also in the path integral formulation of superfluid densities, extended to a mixture of two superfluids, as well as a limiting case of the critical velocity for the dynamic stability of the mixture.
The minimal toy model employed in Sec. \[sec:quantum\_hydro\] also gives insight on low but finite temperature properties such as the specific heat, which is expected to depend on the drag through a factor $c_s^{-D}$ in $D$ dimensions. In general, finite temperature has a detrimental effect on the possibility of observing AB-related phenomena [@LowTempPhys.25.493; @LowTempPhys.30.770], due to the reduction of the superfluid density and to the emergence of dissipative drag between the two components. In this respect, the quantities we suggest to focus on to find experimental evidence of the presence of entrainment are much more suitable than trying to directly detect the current induced by one component on the other. It is worth mentioning that a recent experiment [@1708.03923] was able to record the dynamics of both superfluid and normal fractions of a weakly interacting Bose-Bose mixture, as well as the finite temperature polarisabilities. We believe that experiments such as this one pave the way towards the detection of subtle superfluid effects such as the AB drag.
As a check on the predictions of the quantum hydrodynamic model, we numerically investigated a system of $\mathbb{Z}_2$ symmetric bilayer dipolar bosons, with repulsive interactions within each layer and partially attractive interactions between the two layers. The diffusion quantum Monte Carlo method allows the simultaneous measurement of several observables, separately on the density and spin channels. By inverting the theoretical relations between the static and dynamic observables, and the superfluid drag, we can compare indirect estimators of ${\ensuremath{\rho_{12}}}$ with the direct estimator based on the winding numbers. Figure \[fig:dqmc\] shows a good agreement between direct and indirect measurements of $(\rho - {\ensuremath{\rho_{12}}})
/ nm$. A notable exception is when one extracts the speed of sound of the spin channel using the Bijl-Feynman relation, which, as aforementioned, breaks down if ${\ensuremath{\rho_{12}}}\neq 0$.
We finally analysed the stability of the superfluid mixture. The condition for the onset of the (static) phase separation instability only involves the density channel, and hence is not modified by ${\ensuremath{\rho_{12}}}$. On the other hand, when the two fluids are put in relative motion, for some values of the relative velocity the mixture becomes dynamically unstable. The critical velocities, reported in Eqs. , carry a dependence on the superfluid drag. This result extends the previous ones [@JLowTempPhys.150.612; @JLowTempPhys.155.219; @EurPhysJD.69.126; @ComptesRendusPhysique.16.241] which analysed a system with contact interactions only.
In the light of recent experimental works [@PhysRevLett.115.203201; @PhysRevLett.115.265303; @PhysRevLett.118.055301], it is reasonable to claim that, albeit the AB is expected to be a subdominant effect, it may still be within reach of current experimental technology. It further opens up the game to interesting new phenomenological effects: for instance, if one is able to phase imprint a vortex on one species of the mixture, one may expect at least part of the vorticity to be transferred in a dissipationless fashion to the other species. This *sympathetic stirring* of a superfluid mixture could become a useful tool in the study of vortex dynamics in superfluid mixtures. The stability conditions we put forward in Sec. \[sec:dynamic\_stability\] and Appendix \[app:nonZ2\] may also have consequences in astrophysics, and in particular on the rotation profiles of neutron star cores [@AstrophysJ.836.203].
We acknowledge useful discussions with Giovanni Barontini, Stefano Giorgini, Francesco Minardi, Lode Pollet and Sandro Stringari. G. E. A. acknowledges partial financial support from the MICINN (Spain) Grant No. FIS2014-56257-C2-1-P. The Barcelona Supercomputing Center (The Spanish National Supercomputing Center – Centro Nacional de Supercomputación) is acknowledged for the computational facilities provided. J. N. acknowledges support from the European Research Council through FP7/ERC Starting Grant No. 306897. The authors gratefully acknowledge the Gauss Centre for Supercomputing e.V. ([gauss-centre.eu](gauss-centre.eu)) for funding this project by providing computing time on the GCS Supercomputer SuperMUC at Leibniz Supercomputing Centre (LRZ, [lrz.de](lrz.de)).
Quantum hydrodynamics for a non-symmetric mixture {#app:nonZ2}
=================================================
We consider Eq. in the more general case in which the two species are not symmetric ($\rho_{11} \neq \rho_{22}$, $\alpha_{11} \neq
\alpha_{22}$ and $m_1 \neq m_2$). Hamilton equations lead to the equation of motion for phase fluctuations, $$\ddot{\phi_i}(x) = - \left( \frac{\hbar^2}{m_i m_l} \alpha_{il} \rho_{lj}
\right) k^2 \phi_j(x),$$ and the quantity in parentheses can be identified with the matrix $C_{ij}$ of the squared speeds of sound.
Written with respect to an eigenbasis of $C_{ij}$, the system exhibits two non-interacting modes which obey a linear dispersion relation $\omega_\pm =
c^2_\pm k^2$. The speeds of sound $c_\pm^2$, eigenvalues of $C_{ij}$, generalise Eqs. -. Obviously, contrary to Sec. \[sec:quantum\_hydro\], the two independent modes are not “pure" density and spin channels, due to the lack of $\mathbb{Z}_2$ symmetry.
In terms of the coefficients of matrix $C_{ij}$, the eigenspeeds of sound read $$\begin{aligned}
c^2_\pm = & \frac{C_{11} + C_{22}}{2} \nonumber \\
& \pm \frac{1}{2}
\sqrt{(C_{11} + C_{22})^2 - 4 (C_{11} C_{22} - C_{12} C_{21})},\end{aligned}$$ The system is stable as long as $c_\pm^2$ are positive, which leads to the inequality $C_{11}C_{22} - C_{12}C_{21} >0$, or, in terms of the model’s coefficients, $$(\alpha_{11} \alpha_{22} - \alpha_{12}^2)
(\rho_{11} \rho_{22} - {\ensuremath{\rho_{12}}}^2) > 0$$ The second factor in this equation is always positive, thanks to the bound on the magnitude of the superfluid drag (cf. discussion leading to Eq. ). We thus obtain a condition on the strength of the density-density interactions, namely $(\alpha_{11} \alpha_{22} - \alpha_{12}^2)>0$, thus recovering the well known criterion for the onset of the phase separation instability [@book_pitaevskii.stringari.2016].
Relation between winding numbers and superfluid densities {#sec:app_path_integral}
=========================================================
Following [@PhysRevB.36.8343], we imagine a slab of fluid sandwiched between two parallel boundaries, moving with respect to the fluid with velocity ${\ensuremath{{\bf v}}}$. We can write the density matrix of a two-species system in a frame comoving with the boundaries as $$\begin{aligned}
&\rho_v(R, R', \beta) = {\ensuremath{\langle R' |}} e^{-\beta H_v} {\ensuremath{| R \rangle}}, \\
&H_v = \sum_{\alpha=\{1,2\}}\sum_{j=1}^{N_\alpha} \frac{({\ensuremath{{\bf p}}}_j -
m_\alpha {\ensuremath{{\bf v}}})^2}{2m_\alpha} + V,\end{aligned}$$ where $V$ comprises both inter- and intra-species interactions, both assumed to depend on positions only. (Greek indices identify the species, while Latin indices count particles). The change in total momentum in the frame of reference of the boundaries is due to the response of the normal component $\rho_N$ of the fluid, hence, calling $\rho_0 = \rho_{v=0}$, the density matrix in the frame of rest of the fluid, $$\frac{\rho_N}{\rho_0} (N_1 m_1 + N_2 m_2){\ensuremath{{\bf v}}} =
\frac{\rho_N}{\rho_0} M {\ensuremath{{\bf v}}} = {\ensuremath{\langle {\ensuremath{{\bf P}}}_v \rangle}} =
\frac{\operatorname{tr}[{\ensuremath{{\bf P}}}_v \rho_v]}{\operatorname{tr}[\rho_v]},$$ and is equal to one minus the total superfluid fraction, which can then be written in terms of the variation of the free energy as $$\label{eq:rhos_free_energy_shift}
\frac{\rho_T}{\rho_0} = \frac{\partial
\mathcal{F}_v}{\partial(\frac{1}{2} M v^2)}.$$
We can write the density matrix in Fourier space and relate the density matrix $\rho_0$ in the frame of rest to $\rho_v$, $$\rho_v = \exp \left\lbrace i {\ensuremath{{\bf v}}} \cdot \sum_{\alpha, j}
(r_{j\alpha} - r'_{j\alpha}) m_\alpha\right\rbrace \tilde \rho_0.$$ Under periodic permutations of the particles, $$\rho_v = e^{i {\ensuremath{{\bf v}}} \cdot m_1 {\ensuremath{{\bf W_1}}} L}
e^{i {\ensuremath{{\bf v}}} \cdot m_2 {\ensuremath{{\bf W_2}}}L} \tilde \rho_0,$$ whence $$\begin{aligned}
e^{-\beta \Delta \mathcal{F}_v} &= \frac{1}{\mathcal{Z}}
\int \tilde \rho_0\, e^{i (m_1 {\ensuremath{{\bf W_1}}} + m_2 {\ensuremath{{\bf W_2}}})\cdot {\ensuremath{{\bf v}}}L}
\nonumber \\
& = {\ensuremath{\langle e^{i (m_1 {\ensuremath{{\bf W_1}}} + m_2 {\ensuremath{{\bf W_2}}})\cdot {\ensuremath{{\bf v}}}L} \rangle}}.\end{aligned}$$ Expanding to leading order the previous expression, one finally gets $$\begin{aligned}
\Delta \mathcal{F}_v = \frac{1}{2} \frac{v^2 L^2}{d\beta}
{\ensuremath{\langle (m_1 W_1 + m_2 W_2)^2 \rangle}}\end{aligned}$$ and, using Eq. and $\rho_0 = M L^{-d}$, $$\begin{aligned}
\rho_s^{tot} & = \frac{L^{2-d}}{d\beta} {\ensuremath{\langle (m_1 W_1 + m_2 W_2)^2 \rangle}},\end{aligned}$$ yielding Eq. .
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We show that the measure on markings of ${\mathbf{Z}}_n^d$, $d \geq 3$, with elements of $\{0,1\}$ given by iid fair coin flips on the range ${\mathcal {R}}$ of a random walk $X$ run until time $T$ and $0$ otherwise becomes indistinguishable from the uniform measure on such markings at the threshold $T = \tfrac{1}{2}T_{{\rm cov}}({\mathbf{Z}}_n^d)$. As a consequence of our methods, we show that the total variation mixing time of the random walk on the lamplighter graph ${\mathbf{Z}}_2 \wr {\mathbf{Z}}_n^d$, $d \geq 3$, has a cutoff with threshold $\tfrac{1}{2} T_{{\rm cov}}({\mathbf{Z}}_n^d)$. We give a general criterion under which both of these results hold; other examples for which this applies include bounded degree expander families, the intersection of an infinite supercritical percolation cluster with an increasing family of balls, the hypercube, and the Caley graph of the symmetric group generated by transpositions. The proof also yields precise asymptotics for the decay of correlation in the uncovered set.'
address:
- |
Stanford University\
Department of Mathematics\
Stanford, CA 94305\
\
- |
Microsoft Research\
Theory Group\
Redmond, WA 98052\
\
author:
-
-
bibliography:
- 'lamplighter.bib'
title: Uniformity of the Uncovered set of Random Walk and Cutoff for Lamplighter Chains
---
and
Introduction
============
Suppose $G = (V,E)$ is a finite, connected graph and $X$ is a lazy random walk on $G$. This means that $X$ is the Markov chain with state space $V$ and transition kernel $$p(x,y;G) = {\mathbf{P}}_x[X(1) = y] = \begin{cases}
\frac{1}{2} \text{ if } x = y,\\
\frac{1}{2\deg(x)} \text{ if } \{x,y\} \in E.
\end{cases}$$ Let $$\tau_{{\rm cov}}(G) = \min\{t \geq 1 : V \text{ is contained in the range of } X\big|[0,t]\}$$ be the *cover time* and let $T_{{\rm cov}}(G) = {\mathbf{E}}_\pi \tau_{{\rm cov}}(G)$ be the *expected cover time*. Here and hereafter, a subscript of $\pi$ indicates that $X$ is started from stationarity. Let $\tau(y) = \min\{t \geq 1 : X(t) = y\}$ be the first time $X$ hits $y$ and $$T_{{\rm hit}}(G) = \max_{x,y \in V} {\mathbf{E}}_x \tau(y)$$ be the *maximal hitting time*. If $(G_n)$ is a sequence of graphs with $T_{{\rm hit}}(G_n) = o(T_{{\rm cov}}(G_n))$ then a result of Aldous (Theorem 2, [@A91]) implies that $T_{{\rm cov}}(G_n)$ has a threshold around its mean: $T_{{\rm cov}}(G_n) / {\mathbf{E}}\tau_{{\rm cov}}(G_n) = 1 + o(1)$. Many sequences of graphs satisfy this condition, for example ${\mathbf{Z}}_n^d$ for $d \geq 2$, ${\mathbf{Z}}_2^n$, and the complete graph $K_n$. When Aldous’ condition holds, the set $${\mathcal {L}}(\alpha;G_n) = \{x \in V_n : \tau(x) \geq \alpha T_{{\rm cov}}(G_n)\}$$ of *$\alpha$-late points*, i.e. points hit after time $\alpha T_{{\rm cov}}(G_n)$, $\alpha \in (0,1)$, often has an interesting structure. The case $G_n = {\mathbf{Z}}_n^2$ was first studied by Brummelhuis and Hilhorst in [@BH91] where it is shown that ${\mathbf{E}}|{\mathcal {L}}(\alpha;{\mathbf{Z}}_n^2)|$ has growth exponent $2(1-\alpha)$ and that points in ${\mathcal {L}}_n(\alpha;{\mathbf{Z}}_n^2)$ are positively correlated. This suggests that ${\mathcal {L}}(\alpha;G_n)$ has a fractal structure and exhibits clustering. These statements were made precise by Dembo, Peres, Rosen, and Zeitouni in [@DPRZ_LATE06] where they show that the growth exponent of $|{\mathcal {L}}(\alpha;{\mathbf{Z}}_n^2)|$ is $2(1-\alpha)$ *with high probability* in addition to making a rigorous quantification of the clustering phenomenon.
If $G_n$ is either $K_n$ or ${\mathbf{Z}}_n^d$ for $d \geq 3$ then it is also true that $\log|{\mathcal {L}}(\alpha;G_n)| \sim (1-\alpha) \log |V_n|$ with high probability. In contrast to the case of ${\mathbf{Z}}_n^2$, ${\mathcal {L}}(\alpha;K_n)$ does not exhibit clustering and is “uniformly random” in the sense that conditional on $s_0 = |{\mathcal {L}}(\alpha;K_n)|$, all subsets of $K_n$ of size $s_0$ are equally likely. The rapid decay of correlation in ${\mathcal {L}}(\alpha;{\mathbf{Z}}_n^d)$ for $d \geq 3$ determined by Brummelhuis and Hilhorst [@BH91] indicates that the clustering phenomenon is also not present in this case and leads one to speculate that ${\mathcal {L}}(\alpha;{\mathbf{Z}}_n^d)$ is likewise in some sense “uniformly random.”
The purpose of this article is to quantify the degree to which this holds. We use as our measure of uniformity the following statistical test. Let ${\mathcal {R}}(\alpha;G)$ be the (random) subset of $V$ covered by $X$ at time $\alpha T_{{\rm cov}}(G)$ and let $\mu(\cdot;\alpha,G)$ be the probability measure on ${\mathcal {X}}(G) = \{f \colon V \to \{0,1\}\}$ given by first sampling ${\mathcal {R}}(\alpha;G)$ then setting $$f(x) = \begin{cases}
\xi(x) \text{ if } x \in {\mathcal {R}}(\alpha;G),\\
0 \text{ otherwise,}
\end{cases}$$ where $(\xi(x) : x \in V)$ is a collection of iid variables such that ${\mathbf{P}}[\xi(x) = 0] = {\mathbf{P}}[\xi(x) = 1] = \tfrac{1}{2}$. The question we are interested in is:
*How large does $\alpha \in (0,1)$ need to be so that $\mu(\cdot;\alpha,G)$ is indistinguishable from the uniform measure $\nu(\cdot;G)$ on ${\mathcal {X}}(G)$?*
It must be that $\alpha \geq 1/2$ in the case of ${\mathbf{Z}}_n^d$ for $d \geq 2$ since if $\alpha < 1/2$ then $$\frac{|{\mathcal {L}}(\alpha;{\mathbf{Z}}_n^d)| - \tfrac{1}{2}n^d} {n^{d/2}} \to \infty \text{ as } n \to \infty.$$ In particular, the deviations of the number of zeros from $n^d/2$ for such $\alpha$ far exceeds that in the uniform case. By Theorem 2 of [@A91], it is also true that $\alpha \leq 1$ since if $\alpha > 1$ then with high probability $|{\mathcal {L}}(\alpha;{\mathbf{Z}}_n^d)| = 0$. The main result of this article is that the threshold for indistinguishability for any sequence of graphs $(G_n)$ with uniformly bounded maximal degree and $\lim_{n \to \infty} |V_n| = \infty$ is $\alpha = \tfrac{1}{2}$ provided random walk on $(G_n)$ is *uniformly locally transient* and either satisfies a Harnack inequality or whose Green’s function exhibits sufficiently fast decay.
We need the following definitions in order to give a precise statement of our results. The *total variation mixing time* of $G$ is $$T_{{\rm mix}}(\epsilon;G) = \min\{t \geq 0 : \max_{x \in V} \| p^t(x,\cdot;G) - \pi \|_{TV} \leq \epsilon\}$$ where $p^t(x,y;G) = {\mathbf{P}}_x[ X(t) = y]$ is the $t$-step transition kernel of $X$ started at $x$, $$\| \mu - \nu\|_{TV} = \max_{A \subseteq V} | \mu(A) - \nu(A)| = \frac{1}{2} \sum_{x \in V} |\mu(x) - \nu(x)|$$ is the *total variation distance* between the measures $\mu,\nu$ on $V$, and $\pi$ is the stationary distribution of $X$. The *uniform mixing time* of $G$ is $$T_{{\rm mix}}^U(\epsilon;G) = \min\left\{ t \geq 0 : \max_{x,y \in V} \left| \frac{p^t(x,y;G)}{\pi(y)} - 1 \right| \leq \epsilon \right\}.$$ It is a basic fact ([@AF_MC], [@LPW08], see also Proposition \[prop::mixing\_decay\]) that $T_{{\rm mix}}^U(\epsilon;G)$ is within a factor of $\log|V|$ of $T_{{\rm mix}}(\epsilon;G)$, however, for many graphs this factor is constant. Whenever we omit $\epsilon$ and write $T_{{\rm mix}}(G), T_{{\rm mix}}^U(G)$ it is understood that $\epsilon = \tfrac{1}{4}$. The *Green’s function* of $G$ is $$g(x,y;G) = \sum_{t=1}^{T_{{\rm mix}}^U(G)} p^t(x,y;G),$$ i.e. the expected amount of time that $X$ spends at $y$ until time $T_{{\rm mix}}^U(G)$ when started at $x$. For $A \subseteq V$, we set $$g(x,A;G) = \sum_{y \in A} g(x,y;G).$$ We say that $(G_n)$ is *uniformly locally transient* with *transience function* $\rho \colon [0,\infty) \times [0,\infty) \to [0,\infty)$ if $$g(x,A;G_n) \leq \rho(d(x,A), {{\rm diam}}(A)) \text{ for all } n, x \in V, \text{ and } A \subseteq V.$$ Here, $d(\cdot,\cdot)$ is the graph distance, $d(x,A) = \min_{y \in A} d(x,y)$, and $\rho(\cdot,s)$ is assumed to be non-increasing with $\lim_{r \to \infty} \rho(r,s) = 0$ when $s$ is fixed. Let $\rho(r) = \rho(r,1)$, $${\overline}{\Delta}(G) = \max_{x \in V} \deg(x),\ {\underline}{\Delta}(G) = \min_{x \in V} \deg(x),\ \Delta(G) = \frac{{\overline}{\Delta}(G)}{{\underline}{\Delta}(G)}.$$
\[assump::graphs\] $(G_n)$ is a sequence of uniformly locally transient graphs with $|V_n| \to \infty$ and such that there exists $\Delta_0 > 0$ so that $\Delta(G_n) \leq \Delta_0$ for all $n$ and, for each $r > 0$,
1. $\log |B(x,r)| = o(\log |V_n|)$ as $n \to \infty$, and
2. $T_{{\rm mix}}^U(G_n) {\overline}{\Delta}^r(G_n) = o(|V_n|)$ as $n \to \infty$.
\[assump::far\_away\_harnack\] $(G_n)$ is a sequence of graphs satisfying either
1. \[assump::far\_away\] for every $\gamma > 0$ there exists $R_n^\gamma \to \infty$ as $n \to \infty$ satisfying $R_n^\gamma \leq \tfrac{1}{2}\max\{ R > 0 : \max_{x \in V_n} |B(x,R)| \leq |V_n|^{\gamma}\}$ such that for every $r > 0$, $$\frac{T_{{\rm mix}}^U(G_n)}{R_n^\gamma} \max_{d(x,A) \geq R_n^\gamma} g(x,A) = o(1) \text{ as } n \to \infty$$ uniformly $A \subseteq V_n$ with ${{\rm diam}}(A) \leq r$, or
2. \[assump::harnack\] a uniform Harnack inequality, i.e. for each $\alpha > 1$ there exists $C = C(\alpha) > 0$ such that for every $x,r,R > 0$ with $R/r \geq \alpha$ and positive harmonic function $h$ on $B(x,R)$ we have that $$\max_{y \in B(x,r)} h(y) \leq C \min_{y \in B(x,r)} h(y).$$
\[thm::threshold\] If $(G_n)$ satisfies Assumptions \[assump::graphs\] and \[assump::far\_away\_harnack\] then for every $\epsilon > 0$, $$\begin{aligned}
&\lim_{n \to \infty} \| \mu(\cdot;\tfrac{1}{2} + \epsilon,G_n) - \nu(\cdot;G_n) \|_{TV} = 0 \text{ and }\\
&\lim_{n \to \infty} \| \mu(\cdot;\tfrac{1}{2} - \epsilon,G_n) - \nu(\cdot;G_n) \|_{TV} = 1.\end{aligned}$$
\[rem::bounded\_degree\] If $(G_n)$ is a sequence with $|V_n| \to \infty$ and ${\overline}{\Delta}(G_n)$ is bounded in $n$, then Assumption \[assump::graphs\] is equivalent to the decay of $g(x,y;G_n)$ in $d(x,y)$ uniformly in $n$.
Many families of graphs satisfy Assumptions \[assump::graphs\] and \[assump::far\_away\_harnack\], for example ${\mathbf{Z}}_n^d$ for $d \geq 3$, random $d$-regular graphs whp, also for $d \geq 3$, and the hypercube ${\mathbf{Z}}_2^n$. We will discuss these and other examples in the next section.
The problem that we consider is closely related to determining the mixing time of the *lamplighter walk*, which we now introduce. If $G = (V,E)$ is a finite graph, the wreath product $G^\diamond = {\mathbf{Z}}_2 \wr G$ is the graph $(V^\diamond, E^\diamond)$ whose vertices are pairs $(f,x)$ where $f \in {\mathcal {X}}(G)$ and $x \in V^\diamond$. There is an edge between $(f,x)$ and $(g,y)$ if and only if $\{x,y\} \in E$ and $f(z) = g(z)$ for $z \notin \{x,y\}$. $G^\diamond$ is also referred to as the *lamplighter graph* over $G$ since it can be constructed by placing “lamps” at the vertices of $G$; the first coordinate $f$ of a configuration $(f,x)$ indicates the state of the lamps and the second gives the location of the lamplighter.
{width="50mm"}
The *lamplighter walk* $X^\diamond$ on $G$ is the random walk on $G^\diamond$. Its transition kernel $p(\cdot,\cdot;G^\diamond)$ can be constructed from $p(\cdot,\cdot;G)$ using the following procedure: given $(f,x) \in V^\diamond$,
1. sample $y \in V$ adjacent to $x$ using $p(x,\cdot;G)$,
2. randomize the values of $f(x)$, $f(y)$ using independent fair coin flips,
3. move the lamplighter from $x$ to $y$.
That both $f(x)$ and $f(y)$ are randomized rather than just $f(y)$ is necessary for reversibility.
Random walk on a sequence of graphs $(G_n)$ is said to have a (total variation) cutoff with threshold $(a_n)$ if $$\lim_{n \to \infty} \frac{T_{{\rm mix}}(\epsilon;G_n)}{a_n} = 1 \text{ for all } \epsilon > 0.$$ It is believed that many graphs have a cutoff, but establishing this is often quite difficult since it requires a delicate analysis of the behavior of the underlying random walk. The term was first coined by Aldous and Diaconis in [@AD86] where they prove cutoff for the top-in-at-random shuffling process. Other early examples include random transpositions on the symmetric group [@DS81], the riffle shuffle, and random walk on the hypercube [@A83]. By making a small modification to the proof of Theorems \[thm::threshold\], we are able to establish cutoff for the lamplighter walk on base graphs satisfying Assumptions \[assump::graphs\] and \[assump::far\_away\_harnack\].
Before we state these results, we will first summarize previous work related to this problem. The mixing time of $G^\diamond$ was first studied by Häggström and Jonasson in [@HJ97] in the case $G_n = K_n$ and $G_n = {\mathbf{Z}}_n$. Their work implies a cutoff with threshold $\tfrac{1}{2} T_{{\rm cov}}(K_n)$ in the former case and that there is no cutoff in the latter. The connection between $T_{{\rm mix}}(G^\diamond)$ and $T_{{\rm cov}}(G)$ is explored further in [@PR04], in addition to developing the relationship between the relaxation time of $G^\diamond$ and $T_{{\rm hit}}(G)$, and ${\mathbf{E}}2^{|{\mathcal {L}}(\alpha;G)|}$ and $T_{{\rm mix}}^U(G^\diamond)$. The results of [@PR04] include a proof of cutoff when $G_n = {\mathbf{Z}}_n^2$ with threshold $T_{{\rm cov}}({\mathbf{Z}}_n^2)$ and a general bound that $$\label{eqn::pr_bound}
\left[ \frac{1}{2} + o(1) \right] T_{{\rm cov}}(G_n) \leq T_{{\rm mix}}(G_n) \leq [1+o(1)] T_{{\rm cov}}(G_n)$$ whenever $(G_n)$ is a sequence of vertex transitive graphs with $T_{{\rm hit}}(G_n) = o(T_{{\rm cov}}(G_n))$. It is not possible to improve upon without further hypotheses since the lower and upper bounds are acheived by $K_n$ and ${\mathbf{Z}}_n^2$, respectively.
The bound applies to ${\mathbf{Z}}_n^d$ when $d \geq 3$ since $T_{{\rm hit}}({\mathbf{Z}}_n^d) \sim c_d n^d$ and $T_{{\rm cov}}({\mathbf{Z}}_n^d) = c_d' n^d (\log n)$ (see Proposition 10.13, Exercise 11.4 of [@LPW08]). This leads [@PR04] to the question of whether there is a threshold for $T_{{\rm mix}}({\mathbf{Z}}_n^d)$ and, if so, if it is at $\tfrac{1}{2} T_{{\rm cov}}({\mathbf{Z}}_n^d)$, $T_{{\rm cov}}({\mathbf{Z}}_n^d)$, or somewhere in between. By a slight extension of our methods, we are able to show that the threshold is at $\tfrac{1}{2} T_{{\rm cov}}({\mathbf{Z}}_n^d)$ when $d \geq 3$, and that the same holds whenever $(G_n)$ satisfies Assumptions \[assump::graphs\] and \[assump::far\_away\_harnack\].
\[thm::lamplighter\] If $(G_n)$ satisfies Assumptions \[assump::graphs\] and \[assump::far\_away\_harnack\] then $T_{{\rm mix}}(\epsilon;G_n^\diamond)$ has a cutoff with threshold $\tfrac{1}{2} T_{{\rm cov}}(G_n)$.
In order to prove Theorems \[thm::threshold\] and \[thm::lamplighter\] we need to develop a delicate understanding of both the process of coverage and the correlation structure of ${\mathcal {L}}(\alpha;G_n)$. The proof yields the following theorem, which gives a precise estimate of correlation decay in ${\mathcal {L}}(\alpha;G_n)$ under the additional hypothesis of vertex transitivity.
\[thm::correlation\_decay\] Suppose $(G_n)$ is a sequence of vertex transitive graphs satisfying Assumption \[assump::graphs\]. If $(x_n^i)$ for $1 \leq i \leq \ell$ is a family of sequences with $x_n^i \in V_n$ and $|x_n^i - x_n^j| \geq r$ for every $n$ and $i \neq j$, $$\begin{aligned}
\label{eqn::correlation_decay}
(1-\delta_{r,\ell}) |V_n|^{-\ell \alpha - \delta_{r,\ell}} \leq& {\mathbf{P}}[ x_n^i \in {\mathcal {L}}(\alpha; G_n) \text{ for all } i]\\
\leq& (1+\delta_{r,\ell}) |V_n|^{-\ell \alpha+\delta_{r,\ell}} \notag
\end{aligned}$$ where $\delta_{r,\ell} \to 0$ as $r \to \infty$ while $\ell$ is fixed. If ${\overline}{\Delta}(G_n) \to \infty$ we take $r = 1$ and $\delta_{1,\ell} = o(1)$ as $n \to \infty$.
Outline {#outline .unnumbered}
-------
The remainder of the article is structured as follows. We show in Section \[sec::examples\] that the hypotheses of Theorems \[thm::threshold\] and \[thm::lamplighter\] hold for a number of natural examples. In Section \[sec::prelim\], we collect several general estimates that will be used throughout the rest of the article; Proposition \[prop::tv\_bound\] is in particular of critical importance. Next, in Section \[sec::cov\_hit\] we will develop precise asymptotic estimates for the cover and hitting times of graphs $(G_n)$ satisfying Assumption \[assump::graphs\]. The key idea is that the process of hitting a point can be understood by looking at the number $N(x,t)$ of excursions of $X$ from $\partial B(x,r)$ to $\partial B(x,R)$ for $r < R$, then allowing the walk run for $\beta T_{{\rm mix}}^U(G)$ some $\beta > 0$ in order to remix. Uniform local transience implies that at the time $x$ is hit, $N(x,t)$ is typically quite large and concentrated around its mean. This condition also gives that $\tfrac{1}{k}\sum_{j=1}^k p_j(x)$ is well concentrated around its mean, where $p_j(x)$ is the probability that the $j$th excursion hits $x$ in time $\alpha T_{{\rm mix}}^U(G)$ after exiting $B(x,R)$, $\alpha \leq \beta$, conditional on its point of entry and the point of entry of the $(j+1)$st excursion. Decomposing the process of hitting into excursions between concentric spheres is not new, and is used to great effect in [@DPR_MAN03], [@DPRZ_THICK01], [@DPRZ_COV04], [@DPRZ_LATE06]. Our implementation of this idea is new since explicit representations of hitting probabilities and Green’s functions in addition to the approximate rotational invariance in ${\mathbf{Z}}_n^d$ are simply not available in the generality we consider. We prove Theorem \[eqn::correlation\_decay\] in Section \[sec::corr\_decay\]. This result, which may be of independent interest, is important in Section \[sec::total\_variation\] since it allows us to deduce that points in ${\mathcal {L}}(\tfrac{1}{2}; G_n)$ are typically “spread apart.” The article ends with a proof of Theorems \[thm::threshold\] and \[thm::lamplighter\] as well as a list of related open questions.
Examples {#sec::examples}
========
${\mathbf{Z}}_n^d$, $d \geq 3$ {#mathbfz_nd-d-geq-3 .unnumbered}
------------------------------
Although the simplest, this is the motivating example for this work. It is well-known (see Section 1.5 of [@LAW91]) that there exists a constant $c_d > 0$ so that $g(x,y;{\mathbf{Z}}_n^d) \leq c_d |x-y|^{2-d}$, which implies uniform local transience. Assumption \[assump::far\_away\_harnack\] part is also satisfied since it is also a basic result that random walk on ${\mathbf{Z}}_n^d$ satisfies a Harnack inequality (see [@LAW91], Section 1.4).
Super-critical percolation cluster {#super-critical-percolation-cluster .unnumbered}
----------------------------------
Suppose that $\eta_e$ is a collection of iid random variables indexed by the edges $e=(x,y)$ of ${\mathbf{Z}}^d$, $d \geq 3$, taking values in $\{0,1\}$ such that ${\mathbf{P}}[\eta_e = 1] = p \in [0,1]$. An edge $e$ is called open if $\eta_e = 1$. Let ${\mathcal {C}}(x)$ denote the subset of ${\mathbf{Z}}^d$ consisting of those elements $y$ that can be connected to $x$ by a path consisting only of open edges. Let $C_\infty$ denote the event that there exists an infinite open cluster and let $p_c = \inf\{ p > 0 : {\mathbf{P}}[ C_\infty] >0\}$. Suppose $p > p _c$. Then it is known that there exists a unique infinite open cluster ${\mathcal {C}}_\infty$ almost surely. Fix $x \in {\mathcal {C}}_\infty$ and consider the graph $G_n = B(x,n) \cap {\mathcal {C}}_\infty$. It follows from the works of Delmotte [@D99], Deuschel and Pisztora [@DP96], Pisztora [@P96], and Benjamini and Mossel [@BM03] that the heat kernel for continuous time random walk (CTRW) on $G_n$ has Gaussian tails whp when $n$ is large enough; see the discussion after the statement of Theorem A of [@B04]. Consequently, the Green’s function of the CTRW on $(G_n)$ has the same quantitative behavior as for $({\mathbf{Z}}_n^d)$, which easily implies the same is true for the lazy random walk, which in turn yields uniform local transience for $(G_n)$ whp when $n$ is sufficiently large. Therefore there exists $n_0 = n_0(\omega)$ such that $(G_n : n \geq n_0(\omega))$ almost surely satisfies Assumption \[assump::graphs\]. Furthermore, it is a result of Barlow [@B04] that there exists $n_1 = n_1(\omega)$ such that random walk on $(G_n : n \geq n_1(\omega))$ almost surely satisfies a Harnack inequality and hence Assumption \[assump::far\_away\_harnack\].
Bounded Degree Expanders {#bounded-degree-expanders .unnumbered}
------------------------
Suppose that $(G_n)$ is an expander family with uniformly bounded maximal degree such that $|V_n| \to \infty$. Then there exists $T_0 < \infty$ such that $T_{\rm rel}(G_n) \leq T_0$ for every $n$ where $T_{\rm rel}(G_n)$ is the relaxation time of lazy random walk on $G_n$. Equation (12.11) of [@LPW08] implies that $$p^t(x,y;G_n) \leq C \left( \frac{1}{|V_n|} + e^{-t/T_0}\right)$$ and Theorem 12.3 of [@LPW08] gives $T_{{\rm mix}}^U(G_n) = O(\log |V_n|)$. By Remark \[rem::bounded\_degree\], to check Assumption \[assump::graphs\] we need only show the uniform decay $g(x,y;G_n)$ in $d(x,y)$. If $t < d(x,y)$, then it is obviously true that $p^t(x,y;G_n) = 0$. Hence $$\label{eqn::expander_green_decay} g(x,y;G_n) \leq C\left(\frac{ O(\log |V_n|)}{|V_n|} + \sum_{t=d(x,y)}^{T_{{\rm mix}}^U(G_n)} e^{- t/ T_0}\right)
\leq C_1 e^{- d(x,y) / T_0} + o(1)$$ as $n \to \infty$. We will now argue that $(G_n)$ satisfies part of Assumption \[assump::far\_away\_harnack\]. Suppose that ${\overline}{\Delta} \geq \max_{x \in V_n} \deg(x)$ for every $n$. We can obviously take $R_n^\gamma = \gamma \log |V_n| / (2 \log {\overline}{\Delta})$, hence we have $T_{{\rm mix}}^U(G_n) / R_n^\gamma = O(1)$ as $n \to \infty$. Combining this with implies that $(G_n)$ satisfies Assumption \[assump::far\_away\_harnack\].
Random Regular Graphs {#random-regular-graphs .unnumbered}
---------------------
Suppose that $d \geq 3$ and let ${\mathcal {G}}_{n,d}$ denote the set of $d$-regular graphs on $n$ vertices. It is by now well-known [@BS89] that, whp as $n \to \infty$, an element chosen uniformly from ${\mathcal {G}}_{n,d}$ is an expander. Consequently, whp, a sequence $(G_n)$ where each $G_n$ is chosen independently and uniformly from ${\mathcal {G}}_{n,d}$, $d \geq 3$, almost surely satisfies the hypotheses of our theorems.
Hypercube {#hypercube .unnumbered}
---------
As in the case of super-critical percolation, for ${\mathbf{Z}}_2^n$ it is easiest to prove bounds for the CTRW which, as we remarked before, easily translate over to the corresponding lazy walk. The transition kernel of the CTRW is $$p^t(x,y;{\mathbf{Z}}_2^n) = \frac{1}{2^n} (1+e^{-2t/n})^{n-|x-y|} (1-e^{-2t/n})^{|x-y|},$$ where $|x-y|$ is the number of coordinates in which $x$ and $y$ differ. The spectral gap is $1/n$ (see Example 12.15 of [@LPW08]) which implies $\Omega(n) = T_{{\rm mix}}^U({\mathbf{Z}}_2^n) = O(n^2)$ (see Theorem 12.3 of [@LPW08]). Suppose that $A \subseteq {\mathbf{Z}}_2^n$ has diameter $s$ and $d(x,A) = r$. If $y \in A$, we have $$p^t(x,y;{\mathbf{Z}}_2^n) \leq \frac{1}{2^n} (1+e^{-2t/n})^{n-r} (1-e^{-2t/n})^r.$$ It is easy to see that $$p^t(x,y;{\mathbf{Z}}_2^n) \leq \begin{cases}
(C_\epsilon t/n)^r \exp[- (t / C_\epsilon n)(n-r)] \text{ if } t \leq \epsilon n,\\
e^{-\rho_\epsilon n} \text{ if } t > \epsilon n,
\end{cases}$$ provided $\epsilon > 0$ is sufficiently small. Consequently, $$g(x,A;{\mathbf{Z}}_2^n) \leq C n^{s-r}$$ and therefore ${\mathbf{Z}}_2^n$ is uniformly locally transient. The other hypotheses of Assumption \[assump::graphs\] are obviously satisfied. As for Assumption \[assump::far\_away\_harnack\], we note that in this case, we can take $R_n^\gamma = \gamma n/ (2\log_2 n)$. Thus if $r > 0$ it is easy to see that if ${{\rm diam}}(A) \leq s$ and $d(x,A) \geq R_n^\gamma$ we have that $$\sum_{y \in A} p^t(x,y;{\mathbf{Z}}_2^n) \leq n^s e^{-\rho_\epsilon n}.$$ if $t > \epsilon n$. On the other hand, if $t \leq \epsilon n$, then we have $$\sum_{y \in A} p^t(x,y;{\mathbf{Z}}_2^n) \leq n^{s} \left( \frac{ C_\epsilon t}{n} \right)^{\gamma n/ (2\log_2 n)} e^{- t/2 C_\epsilon}.$$ Hence it is not hard to see that ${\mathbf{Z}}_2^n$ satisfies Assumption \[assump::far\_away\_harnack\].
Caley Graph of $S_n$ Generated by Transpositions {#caley-graph-of-s_n-generated-by-transpositions .unnumbered}
------------------------------------------------
Let $G_n$ be the Caley graph of $S_n$ generated by transpositions. By work of Diaconis and Shahshahani [@DS81], $T_{{\rm mix}}(G_n) = \Theta(n (\log n))$, which by Theorem 12.3 of [@LPW08] implies $T_{{\rm mix}}^U(G_n) = O( n^2 (\log n)^2)$. We are now going to give a crude estimate of $p^t(\sigma,\tau;S_n)$. By applying an automorphism, we may assume without loss of generality that $\sigma = {\rm id}$. Suppose that $d({\rm id},\tau) = r$ and that $\tau_1,\ldots,\tau_r$ are transpositions such that $\tau_r \cdots \tau_1 = \tau$. Then $\tau_1,\ldots,\tau_r$ move at most $2r$ of the $n$ elements of $\{1,\ldots,n\}$, say, $k_1,\ldots,k_{2r}$. Suppose $k_1',\ldots,k_{2r}'$ are distinct from $k_1,\ldots,k_{2r}$ and $\alpha \in S_n$ is such that $\alpha(k_i) = k_i'$ for $1 \leq i \leq r$. Then the automorphism of $G_n$ induced by conjugation by $\alpha$ satisfies $\alpha \tau \alpha^{-1} \neq \tau$. Therefore the size of the set of elements $\tau'$ in $S_n$ such that there exists a graph automorphism $\varphi$ of $G_n$ satisfying $\varphi(\tau) = \tau'$ and $\varphi({\rm id}) = {\rm id}$ is at least $\begin{pmatrix} n\\ 2r \end{pmatrix} \geq 2^{-2r} n^{2r} ((2r)!)^{-1}$ assuming $n \geq 4r$. Therefore $$p^t(e,\tau;G_n) \leq \frac{2^{2r} (2r)!}{n^{2r}} \text{ and } g(e,\tau;G_n) \leq C(2^{2r} (2r)!) (\log n)^2 n^{2-2r}.$$ If ${{\rm diam}}(A) = s$ then trivially $|A| \leq n^{2s}$ from which it is clear that $(G_n)$ is uniformly locally transient. The other parts of Assumption \[assump::graphs\] are obviously satisfied by $G_n$. As for Assumption \[assump::far\_away\_harnack\], a simple calculation shows that we can take $R_n^\gamma \leq \gamma n/4 + O(1)$. Hence setting $R_n^\gamma = \sqrt{n}$, a calculation analogous to the one above gives that Assumption \[assump::far\_away\_harnack\] is satisfied.
Preliminary Estimates {#sec::prelim}
=====================
The purpose of this section is to collect several general estimates that will be useful for us throughout the rest of the article.
\[lem::tv\_l1\_bound\] If $\mu,\nu$ are measures with $\nu$ absolutely continuous with respect to $\mu$ and $$\int \frac{d\nu}{d\mu} d\nu = 1 + \epsilon$$ then $$\|\nu - \mu \|_{TV} \leq \frac{\sqrt{\epsilon}}{2}.$$
This is a consequence of the Cauchy-Schwarz inequality: $$\begin{aligned}
& \|\mu - \nu\|_{TV}^2
= \left( \frac{1}{2}\int \left| \frac{d\nu}{d\mu} - 1 \right| d\mu\right)^2
\leq \frac{1}{4}\int \left| \frac{d\nu}{d\mu} - 1 \right|^2 d\mu
= \frac{1}{4}\left(\int \frac{d\nu}{d\mu} d\nu - 1\right)\end{aligned}$$
Let $\nu$ denote the uniform measure on ${\mathcal {X}}(G) = \{f \colon V \to \{0,1\}\}$
\[prop::tv\_bound\] Suppose that $\mu$ is a measure on ${\mathcal {X}}(G)$ given by first sampling ${\mathcal {R}}\subseteq V$ according to a probability $\mu_0$ on $2^{|V|}$, then, conditional on ${\mathcal {R}}$ sampling $f \in {\mathcal {X}}(G)$ by setting $$f(x) = \begin{cases}
\xi(x) \text{ if } x \in {\mathcal {R}},\\
0 \text{ otherwise,}
\end{cases}$$ where $(\xi(x) : x \in V)$ is a collection of iid random variables with ${\mathbf{P}}[ \xi(x) = 0] = {\mathbf{P}}[\xi(x) = 1] = \tfrac{1}{2}$. Then $$\int \frac{d\mu}{d\nu} d\mu = \int \int 2^{|{\mathcal {R}}^c \cap {\mathcal {S}}^c|} d\mu_0({\mathcal {R}}) d\mu_0({\mathcal {S}}).$$
Suppose $f \in {\mathcal {X}}(G)$ is such that $f|{\mathcal {R}}^c \equiv 0$ for some ${\mathcal {R}}\subseteq V$. Letting $\mu(\cdot|{\mathcal {S}})$ be the conditional law of $\mu$ given ${\mathcal {S}}$, we have $$\begin{aligned}
\int \frac{d\mu}{d\nu} d\mu
&= 2^N \int \mu(\{ f\}) d\mu(f)\\
&= 2^N \int \int \left( \int \mu(\{f\} | {\mathcal {S}}) d\mu_0({\mathcal {S}}) \right) d\mu(f|{\mathcal {R}}) d\mu_0({\mathcal {R}}).\end{aligned}$$ Note that $$\mu(\{f\} | {\mathcal {S}}) = 2^{-|{\mathcal {R}}\cap {\mathcal {S}}| - |{\mathcal {S}}\setminus {\mathcal {R}}|}| {\mathbf{1}}_{\{f|{\mathcal {R}}\setminus {\mathcal {S}}\equiv 0\}}.$$ Hence, the above is equal to $$\begin{aligned}
&2^N \int \int \left( \int 2^{-|{\mathcal {R}}\cap {\mathcal {S}}|-|{\mathcal {S}}\setminus {\mathcal {R}}|} {\mathbf{1}}_{\{f|{\mathcal {R}}\setminus {\mathcal {S}}\equiv 0\}} d\mu_0({\mathcal {S}}) \right) d\mu(f|{\mathcal {R}}) d\mu_0({\mathcal {R}})\\
=& 2^N \int \int 2^{-|{\mathcal {R}}\cap {\mathcal {S}}| - |{\mathcal {S}}\setminus {\mathcal {R}}|} \left( \int {\mathbf{1}}_{\{f|{\mathcal {R}}\setminus {\mathcal {S}}\equiv 0\}} d\mu(f|{\mathcal {R}}) \right) d\mu_0({\mathcal {R}}) d\mu_0({\mathcal {S}})\\
=& 2^N \int \int 2^{-|{\mathcal {R}}\cap {\mathcal {S}}| - |{\mathcal {S}}\setminus {\mathcal {R}}|} 2^{-|{\mathcal {R}}\setminus {\mathcal {S}}|} d\mu_0({\mathcal {S}}) d\mu_0({\mathcal {R}}),\end{aligned}$$ where $N = |V|$. Simplifying the expression in the exponent gives the result.
Roughly speaking, the general strategy of our proof will be to show that if ${\mathcal {R}}, {\mathcal {R}}'$ denote independent copies of the range of random walk on $G_n$ run up to time $(\tfrac{1}{2} + \epsilon) T_{{\rm cov}}(G_n)$ and ${\mathcal {L}}= V \setminus {\mathcal {R}}$, ${\mathcal {L}}' = V \setminus {\mathcal {R}}'$ then $$\label{eqn::exp_moment}
{\mathbf{E}}\exp(\zeta |{\mathcal {L}}\cap {\mathcal {L}}'|) = 1 + o(1) \text{ as } n \to \infty$$ for $\zeta > 0$. This method cannot be applied directly, however, since this exponential moment blows up even in the case of ${\mathbf{Z}}_n^3$. To see this, suppose that $X,X'$ are independent random walks on ${\mathbf{Z}}_n^3$ initialized at stationarity. We divide the cover time $c_3 n^3 (\log n)$ into rounds of length $n^2$. In the first round, with probability $1/4$ we know that $X$ starts in ${\mathbf{L}}_1 = {\mathbf{Z}}_n^2 \times \{n/8,\ldots,3n/8\}$. In each successive round, $X$ has probability $\rho_0 > 0$ strictly bounded from zero in $n$ of not leaving ${\mathbf{L}}_2 = {\mathbf{Z}}_n^2 \times \{1,\ldots,n/2\}$ and ending the round in ${\mathbf{L}}_1$. Since there are $c_3 n (\log n)$ rounds, this means that $X$ does not leave ${\mathbf{L}}_1$ with probability at least $$\tfrac{1}{4} \rho_0^{c_3 n \log n} \geq c \exp(-\rho_1 n \log n).$$ Since $X'$ satisfies the same estimate, we therefore have $${\mathbf{E}}\exp(\zeta |{\mathcal {L}}\cap {\mathcal {L}}'|) \geq c \exp(\tfrac{\zeta}{2} n^3- 2\rho_1 n\log n) \to \infty \text{ as } n \to \infty.$$ The idea of the proof is to truncate the exponential moment in by replacing $\mu_0$ with ${\widetilde}{\mu}_0$, the law of ${\mathcal {R}}(\tfrac{1}{2} + \delta; G_n)$ *conditional on typical behavior* so that $$\| {\widetilde}{\mu}_0 - \mu_0 \|_{TV} = o(1) \text{ as } n \to \infty.$$ We do this in such a way that the uncovered set exhibits a great deal of spatial independence in order to make the exponential moment easy to estimate. To this end, we will condition on two different events. The first is that points in ${\mathcal {L}}(\tfrac{1}{2} + \delta;G_n)$ are well-separated: for any $x \in V_n$ we have $|{\mathcal {L}}(\tfrac{1}{2} + \delta;G_n) \cap B(x,R_n^\gamma)| \leq M = M_{\gamma,\delta}$. Given this event, we can partition ${\mathcal {L}}(\tfrac{1}{2} + \delta;G_n)$ into disjoint subsets $E_1,\ldots, E_M$ such that $x,y \in E_\ell$ distinct implies $d(x,y) \geq R_n^\gamma$. Observe: $${\mathbf{E}}\exp(\zeta |{\mathcal {L}}\cap {\mathcal {L}}' \cap E_\ell |) \leq {\mathbf{E}}\prod_{x \in E_\ell} \left(1 + e^\zeta \prod_{j=1}^{N'(x,T)} (1-q_j'(x)) \right).$$ where $N'(x,T)$ is the number of excursions of $X'$ from $\partial B(x,r)$ to $\partial B(x,R)$ by time $T$ and $q_j'(x)$ is the probability the $j$th such excursion hits $x$ conditional on its entrance and exit points. When $T$ is large, uniform local transience implies that $N'(x,T)$ and $\prod_{j=1}^k q_j'(x)$ can be estimated by their mean and, roughly speaking, this is the second event on which we will condition. Finally, we get control of the entire exponential moment by an application of Hölder’s inequality.
We finish the section by recording a standard lemma that bounds the rate of decay of the total variation and uniform distances to stationarity:
\[prop::mixing\_decay\] For every $s,t \in N$, $$\begin{aligned}
&\max_x \| p^{t+s}(x,\cdot) - \pi \|_{TV} \leq 4 \max_{x,y} \|p^t(x,\cdot) - \pi \|_{TV} \| p^s(y,\cdot) - \pi \|_{TV} \label{eqn::tv_decay}\\
&\max_{x,y} \left| \frac{p^{t+s}(x,y)}{\pi(y)} - 1\right| \leq \max_{x,y} \frac{p^{s}(x,y)}{\pi(y)} \max_x \| p^{t}(x,\cdot) - \pi \|_{TV} \label{eqn::uniform_decay}.\
\end{aligned}$$
The first part is a standard result; see, for example, Lemmas 4.11 and 4.12 of [@LPW08]. The second part is a consequence of the semigroup property: $$\begin{aligned}
& \frac{1}{\pi(z)} p^{t+s}(x,z)
= \frac{1}{\pi(z)} \sum_y p^t(x,y) p^s(y,z)\\
=& \frac{1}{\pi(z)} \sum_y [p^t(x,y) - \pi(y) + \pi(y)]p^s(y,z)\\
\leq& \left(\max_{y,z} \frac{p^s(y,z)}{\pi(z)}\right) \| p^t(x,\cdot) - \pi \|_{TV} + 1 \end{aligned}$$
Note that and give $$\begin{aligned}
&\max_{x} \| p^t(x,\cdot) - \pi \|_{TV} \leq c e^{-c\alpha} \text{ for } t \geq \alpha T_{{\rm mix}}(G)\\
&\max_{x,y} \left| \frac{p^{t+s}(x,y)}{\pi(y)} - 1 \right| \leq c e^{-c\alpha} \text{ for } t \geq T_{{\rm mix}}^U(G) + \alpha T_{{\rm mix}}(G) \label{eqn::uniform_exponential_decay}\end{aligned}$$ where $c > 0$ is a universal constant. We will often use without reference, and, for simplicity use that the same inequality holds when $T_{{\rm mix}}^U(G) + \alpha T_{{\rm mix}}(G)$ is replaced by $\alpha T_{{\rm mix}}^U(G)$, perhaps adjusting $c > 0$.
Hitting and Cover Times {#sec::cov_hit}
=======================
In this section we will develop a number of estimates useful for understanding the process of coverage via excursions of random walk between concentric spheres. Throughout, we assume that we have a sequence $(G_n)$ satisfying Assumption \[assump::graphs\] with transience function $\rho$. We will often suppress the index $n$ and refer to an element of $(G_n)$ as $G$.
Probability of Success
----------------------
Fix $R > r$ and let $X$ be a lazy random walk on $G$. Suppose $E = \{x_1,\ldots,x_\ell\} \subseteq V$ where $d(x_i,x_j) \geq 2R$ for $i \neq j$. Let $E(s) = \{ x \in V : d(x,E) \leq s\}$ where $d(x,E) = \min_{y \in E} d(x,y)$.
![The solid and dashed circles represent the boundaries of $E(r)$ and $E(R)$, respectively, and the small points are the elements of $E$. Note that $X$ may re-enter $E(r)$ during the interval $[\sigma_k^\beta(E), \tau_{k+1}^\beta(E)]$](excursions.pdf){width="80mm"}
.
Fix $\beta \geq 0$ and define stopping times $$\begin{aligned}
\tau_0(E) &= \min\{t \geq 0 : X(t) \in \partial E(r)\},\\
\sigma_0(E) &= \min\{ t \geq \tau_0(E) : X(t) \notin E(R)\}\end{aligned}$$ and inductively define $$\begin{aligned}
\tau_k^\beta(E) &= \min\{t \geq \sigma_{k-1}^{\beta}(E) + T_\beta^U : X(t) \in \partial E(r)\},\\
\sigma_k^\beta(E) &= \min\{ t \geq \tau_k^\beta(E) : X(t) \notin E(R)\},\end{aligned}$$ where $\partial E(r) = \{z : d(z,E) = r\}$.
Fix $\alpha \in [0,\beta]$. Let $S_j^{\alpha,\beta}(x;E)$ be the event that $X(t)$ hits $x$ in $[\tau_{j}^\beta(E),\sigma_{j}^\beta(E) + T_\alpha^U]$, $$p_j^{\alpha,\beta}(x;E) = {\mathbf{P}}[S_j^{\alpha,\beta}(x;E)|X(\tau_{j}^\beta(E)),X(\tau_{j+1}^\beta(E))],$$ and $$a_j^{\alpha,\beta}(x;E) = {\mathbf{E}}\left[ \sum_{t = \tau_{j}^\beta(E)}^{\sigma_{j}^\beta(E) + T_\alpha^U} {\mathbf{1}}_{\{X(t) = x\}} \bigg| X(\tau_{j}^\beta(E)) , X(\tau_{j+1}^\beta(E)) \right].$$ Finally, let ${\overline}{p}_{r,R}^{\alpha,\beta}(x;E) = {\mathbf{E}}_\pi p_0^{\alpha,\beta}(x;E)$ and ${\overline}{a}_{r,R}^{\alpha,\beta}(x;E) = {\mathbf{E}}_\pi a_0^{\alpha,\beta}(x;E)$. For $\beta \geq \alpha \geq 1$ note that $$\begin{aligned}
{\overline}{p}_{r,R}^{\alpha,\beta}(x;E) = {\overline}{p}_{r,R}^{1,1}(x;E) + O\left( \frac{T_\beta^U}{|V|}\right).\end{aligned}$$ Since ${\overline}{p}_{r,R}^{1,1}(x;E) \geq c {\overline}{\Delta}^{-r}(G)$ and $T_\beta^U {\overline}{\Delta}^r(G) / |V| = o(1)$ as $n \to \infty$, we therefore have $${\overline}{p}_{r,R}^{\alpha,\beta}(x;E) = (1+o(1)) {\overline}{p}_{r,R}^{1,1}(x;E).$$ From now on, we will write ${\overline}{p}_{r,R}(x;E)$ for ${\overline}{p}_{r,R}^{1,1}(x;E)$. It is also true that ${\overline}{a}_{r,R}^{\alpha,\beta}(x;E) = (1+o(1)){\overline}{a}_{r,R}^{1,1}(x;E)$ and we will also write ${\overline}{a}_{r,R}(x;E)$ for ${\overline}{a}_{r,R}^{1,1}(x;E)$.
\[lem::prob\_success\_bound\] For each $\delta > 0$ there exists $\gamma_0 > 0$ such that for $\beta-\alpha \geq \gamma_0$ and all $n$ large enough we have $$\begin{aligned}
\label{lem::pj_bound}
1-\delta \leq &\frac{p_j^{\alpha,\beta}(x;E)}{{\mathbf{P}}[ S_j^{\alpha,\beta}(x) | X(\tau_j^\beta(E))]} \leq 1+\delta,\\
1-\delta \leq &\frac{a_j^{\alpha,\beta}(x;E)}{ {\mathbf{E}}\left[ \sum_{t = \tau_{j}^\beta(E)}^{\sigma_{j}^\beta(E) + T_\alpha^U} {\mathbf{1}}_{\{X(t) = x\}} \bigg| X(\tau_{j}^\beta(E)) \right]} \leq 1+\delta.
\end{aligned}$$ In particular, $p_j^{\alpha,\beta}(x;E) \leq (1+\delta) \rho(r)$ and $a_j^\beta(x;E) \leq (1+\delta) \rho(0) \rho(r)$.
Note that $$\begin{aligned}
& {\mathbf{P}}[ X(\sigma_j^\beta(E) + T_\alpha^U) = z | X(\tau_j^\beta(E)) = z_j, X(\tau_{j+1}^\beta(E)) = z_{j+1}]\\
=& \frac{{\mathbf{P}}[ X(\sigma_j^\beta(E) + T_\alpha^U) = z, X(\tau_j^\beta(E)) = z_j, X(\tau_{j+1}^\beta(E)) = z_{j+1}]}{{\mathbf{P}}[ X(\tau_j^\beta(E)) = z_j, X(\tau_{j+1}^\beta(x;E)) = z_{j+1}]}\\
=& \left(\frac{{\mathbf{P}}[ X(\tau_{j+1}^\beta(E)) = z_{j+1} | X(\sigma_j^\beta(E) + T_\alpha^U) = z]}{{\mathbf{P}}[ X(\tau_{j+1}^\beta(E)) = z_{j+1}| X(\tau_j^\beta(E)) = z_j]}\right) \\
&\ \ \ \ \ \ \ \ {\mathbf{P}}[X(\sigma_j^\beta(E) + T_\alpha^U) = z| X(\tau_j^\beta(E)) = z_j].\end{aligned}$$ Mixing considerations imply $$\begin{aligned}
{\mathbf{P}}[ X(\tau_{j+1}^\beta(E)) = z_{j+1}| X(\tau_j^\beta(E)) = z_j] &= \big[ 1+ O(e^{-c\beta}) \big] {\mathbf{P}}_\pi[ X(\tau_0(E)) = z_{j+1}],
\end{aligned}$$ and $$\begin{aligned}
&{\mathbf{P}}[ X(\tau_{j+1}^\beta(E)) = z_{j+1}| X(\sigma_j^\beta(E) + T_\alpha^U) = z]\\
=& \big[ 1+ O(e^{-c(\beta-\alpha)}) \big] {\mathbf{P}}_\pi[ X(\tau_0(E)) = z_{j+1}].\end{aligned}$$ Consequently, if $\mu_j$ denotes the law of $X(\sigma_j^\beta(E) + T_\alpha^U)$ conditional on $X(\tau_j^\beta(E))$ and $X(\tau_{j+1}^\beta(E))$ and $\mu$ is the law of $X(\sigma_j^\beta(E) + T_\alpha^U)$ but conditional only on $X(\tau_j^\beta(E))$, we have $1-\delta \leq d\mu_j / d\mu \leq 1+\delta$ when $\beta - \alpha$ is large enough. Thus, $$\begin{aligned}
p_j^{\alpha,\beta}(x;E)
=& \int {\mathbf{P}}[ S_j^{\alpha,\beta}(x)| X(\tau_j^\beta(E)), X(\sigma_{j}^\beta(E) + T_\alpha^U) = z, X(\tau_{j+1}^\beta(E))] d\mu_j(z)\\
\leq& (1+\delta)\int {\mathbf{P}}[ S_j^{\alpha,\beta}(x)| X(\tau_j^\beta(E)), X(\sigma_j^\beta(E) + T_\alpha^U) = z] d\mu(z)\\
=& (1+\delta){\mathbf{P}}[S_j^{\alpha,\beta}(x) | X(\tau_{j}^\beta(E))]\end{aligned}$$ The lower bound for $p_j^{\alpha,\beta}(x;E)$ and the bounds for $a_j(x;E)$ are proved similarly.
\[lem::prob\_success\_concen\] Fix $r > 0$ and $\delta \in (0,1)$. There exists $\gamma_0 > 0$ depending only on $r,\delta$ such that for all $R \geq r$, $\beta - \alpha \geq \gamma_0$, and $n$ large enough we have $$\begin{aligned}
& {\mathbf{P}}\left[ \sum_{j=1}^k p_j^{\alpha,\beta}(x;E) \notin [1-\delta,1+\delta] {\overline}{p}_{r,R}(x;E) k \right]\\
&\ \ \ \ \ \ \ \ \ \ \ \leq 4\exp\left(-\frac{C \delta^2 {\overline}{p}_{r,R}(x;E)}{\rho(r)} k \right) \notag\end{aligned}$$ and $$\begin{aligned}
& {\mathbf{P}}\left[ \sum_{j=1}^k a_j^{\alpha,\beta}(x;E) \notin [1-\delta,1+\delta] {\overline}{a}_{r,R}(x;E) k \right]\\
&\ \ \ \ \ \ \ \ \ \ \ \leq 4\exp\left(-\frac{C \delta^2 {\overline}{a}_{r,R}(x;E)}{\rho(r)} k \right) \notag\end{aligned}$$ where $C > 0$ is independent of $r,R,\delta$.
Let $\mu$ be the measure on $\partial E(r)$ induced by the law of $X(\tau_0(E))$ given that $X$ has a stationary initial distribution. For each $\delta > 0$, let ${\mathcal {M}}(\delta)$ be the set of measures $\nu$ on $\partial E(r)$ satisfying $$\label{eqn::measure_bound}
\max_{z \in \partial E(r)} \left|\frac{\nu(z)}{\mu(z)} - 1 \right| + \max_{z \in \partial E(r)} \left|\frac{\mu(z)}{\nu(z)} - 1 \right| \leq \delta.$$ Let $\mu_y(z) = {\mathbf{P}}_y[ X(\tau^\gamma(E)) = z]$ where $\tau^\gamma(E) = \min\{ t \geq T_\gamma^U : X(t) \in \partial E(r)\}$. Mixing considerations imply that $\mu_y \in {\mathcal {M}}( C e^{-C \gamma})$ for some $C > 0$. Fix $\delta > 0$, $\delta' < \delta/2$, and take $\beta - \alpha = \gamma$ so large that $C e^{-C \gamma} \leq \delta'/2$. Let ${\overline}{\mu}$, ${\underline}{\mu}$ be elements of ${\mathcal {M}}(\delta'/2)$ such that ${\mathbf{P}}[S_0^{\alpha,\beta}(x) | X(\tau_0(E)) = Z]$ where $Z \sim {\overline}{\mu}, {\underline}{\mu}$ stochastically dominates from above and below, respectively, all other choices in ${\mathcal {M}}(\delta'/2)$. Assume that $\gamma_0$ is chosen sufficiently large so that the previous lemma applies for $\delta'/2$ when $n$ is sufficiently large.
Let $(U_j), (L_j)$ be iid sequences with laws ${\mathbf{P}}[S_0^{\alpha,\beta}(x) | X(\tau_0(E)) = Z]$, $Z \sim {\overline}{\mu}, {\underline}{\mu}$, respectively. With ${\overline}{U} = {\mathbf{E}}U_1$ and ${\overline}{L} = {\mathbf{E}}U$, obviously $$(1-\delta') {\overline}{p}_{r,R}(x;E) \leq {\overline}{L} \leq {\overline}{U} \leq (1+\delta') {\overline}{p}_{r,R}(x;E).$$ By construction, we can find a coupling of $U_j, L_j, p_j^{\alpha,\beta}(x;E)$ so that $$L_j \leq p_j^{\alpha,\beta}(x;E) \leq U_j \text{ for all } j.$$
Corollary 2.4.5 of [@DZ98] implies $${\mathbf{E}}e^{\lambda U_1} \leq \frac{1}{2\rho(r)} \big( {\overline}{U} e^{2 \lambda \rho(r)} + 2\rho(r)-{\overline}{U} \big)$$ hence Exercise 2.2.26 of [@DZ98] gives that the Fenchel-Legendre transform $\Lambda^*$ of the law of $U_1$ satisfies $$\Lambda^*(u) \geq {\widetilde}{\Lambda}^*(u)\equiv \frac{u}{2\rho(r)} \log\left(\frac{u}{{\overline}{U}}\right) + \left(1-\frac{u}{2\rho(r)}\right) \log \left(\frac{1-u/(2\rho(r))}{1-{\overline}{U}/(2\rho(r))}\right).$$ As $${\widetilde}{\Lambda}^*({\overline}{U}) = ({\widetilde}{\Lambda}^*)'({\overline}{U}) = 0 \text{ and } ({\widetilde}{\Lambda}^*)''(u) \geq \frac{1}{2\rho(r) u}$$ we have $$\inf_{u \geq (1+\delta') {\overline}{U}} \Lambda^*(u) \geq \frac{1}{4\rho(r) {\overline}{U}} (\delta')^2 {\overline}{U}^2 = \frac{(\delta')^2 {\overline}{U}}{4\rho(r)},$$ assuming $\delta' < 1$. Consequently, Cramer’s theorem (Theorem 2.2.3, part (c), of [@DZ98]) implies that $$\begin{aligned}
\label{eqn::p_concen}
{\mathbf{P}}\left[ \sum_{i=1}^k U_i \leq (1+\delta') {\overline}{U} k \right] \geq 1-2\exp\left(- \frac{(\delta')^2 {\overline}{U} k}{4\rho(r)} \right).\end{aligned}$$ An analogous estimate also holds for $(L_i)$ with ${\overline}{U}$ replaced by ${\overline}{L}$. The proof of concentration for the $a_j^{\alpha,\beta}(x;E)$ is the same.
Excursion Lengths
-----------------
In this subsection we will estimate the mean and prove concentration for the lengths $\tau_{k+1}^\beta(E) - \tau_{k}^\beta(E)$ of successive excursions. Before we do this, it will be helpful for us to estimate the Radon-Nikodym derivative of the law of random walk conditioned not to hit $E(r)$ with respect to the stationary measure $\pi$.
\[lem::radon\_control\] For $\alpha, s \geq 0$ we have $${\mathbf{P}}_y[ X(T_\alpha^U) = z | {\mathcal {A}}] = \big[ 1+ O( e^{-c \alpha} + |E| \rho(s,r)) + o(1) \big] \pi(z) \text{ as } n \to \infty$$ where ${\mathcal {A}}= \{ \tau(E) \geq T_\alpha^U, d(X(T_\alpha^U),E) \geq s\}$.
For $z \in V$ with $d(z,E) \geq s$, observe $$\begin{aligned}
& {\mathbf{P}}_y[X(T_\alpha^U) = z \big| {\mathcal {A}}]
= \frac{{\mathbf{P}}_y[X(T_\alpha^U) = z, \tau(E) \geq T_\alpha^U]}{{\mathbf{P}}_y[{\mathcal {A}}]}\\
=& \frac{{\mathbf{P}}_y[\tau(E) \geq T_\alpha^U \big| X(T_\alpha^U) = z] {\mathbf{P}}_y[X(T_\alpha^U) = z]}{{\mathbf{P}}_y[{\mathcal {A}}]}\\
=& (1+O(e^{-c\alpha}))\frac{{\mathbf{P}}_y[\tau(E) \geq T_\alpha^U \big| X(T_\alpha^U) = z] \pi(z)}{{\mathbf{P}}_y[{\mathcal {A}}]}.\end{aligned}$$ For $\alpha' < \alpha$, $$\begin{aligned}
& {\mathbf{P}}_y[\tau(E) \geq T_\alpha^U \big| X(T_\alpha^U) = z]\\
=& {\mathbf{P}}_y[\tau(E) \geq T_\alpha^U - T_{\alpha'}^U | X(T_\alpha^U) = z] -
{\mathbf{P}}_y[T_\alpha^U > \tau(E) \geq T_\alpha^U - T_{\alpha'}^U | X(T_\alpha^U) = z]\end{aligned}$$ We have $$\begin{aligned}
& {\mathbf{P}}_y[\tau(E) \geq T_\alpha^U - T_{\alpha'}^U | X(T_\alpha^U) = z]
= 1-\frac{{\mathbf{P}}_y[ \tau(E) < T_\alpha^U - T_{\alpha'}^U, X(T_\alpha^U) = z]}{{\mathbf{P}}_y[X(T_\alpha^U) = z]} \\
=& 1-\frac{1+O(e^{-c\alpha})}{\pi(z)} {\mathbf{P}}_y[X(T_\alpha^U) = z | \tau(E) < T_\alpha^U - T_{\alpha'}^U] {\mathbf{P}}_y[\tau(E) < T_\alpha^U - T_{\alpha'}^U] \\
=& {\mathbf{P}}_y[\tau(E) \geq T_\alpha^U - T_{\alpha'}^U] + O(e^{-c(\alpha-\alpha')}).\end{aligned}$$
Note that $$\begin{aligned}
& {\mathbf{P}}_y[T_\alpha^U > \tau(E) \geq T_\alpha^U - T_{\alpha'}^U | X(T_\alpha^U) = z]\\
=& \frac{1+ O(e^{-c\alpha})}{\pi(y)\pi(z)} {\mathbf{P}}_y[T_\alpha^U > \tau(E) \geq T_\alpha^U - T_{\alpha'}^U , X(T_\alpha^U) = z] \pi(y).
\end{aligned}$$ By reversibility, this is equal to $$\begin{aligned}
&\frac{1+ O(e^{-c\alpha})}{\pi(y)} {\mathbf{P}}_z[ \tau(E) \leq T_{\alpha'}^U, d(X(t),E) > r \text{ for all } T_{\alpha'}^U < t \leq T_\alpha^U, X(T_\alpha^U) = y]\\
\leq& \frac{1+O(e^{-c\alpha})}{\pi(y)}{\mathbf{P}}_z[X(T_\alpha^U) = y|\tau(E) \leq T_{\alpha'}^U]{\mathbf{P}}_z[\tau(E) \leq T_{\alpha'}^U]\\
=& (1+O(e^{-c(\alpha-\alpha')})){\mathbf{P}}_z[\tau(E) \leq T_{\alpha'}^U]\end{aligned}$$ A union bound implies this is of order $O\left(|E|\rho(s,r) + o(1)\right)$. With ${\mathcal {A}}_1 = \{d(X(T_\alpha^U),E) \geq s\}$, $$\begin{aligned}
& {\mathbf{P}}_y[{\mathcal {A}}] = {\mathbf{P}}_y[\tau(E) \geq T_\alpha^U, {\mathcal {A}}_1]\\
=& \big({\mathbf{P}}_y[\tau(E) \geq T_\alpha^U - T_{\alpha'}^U \big| {\mathcal {A}}_1] - {\mathbf{P}}_y[T_\alpha^U > \tau(E) \geq T_\alpha^U - T_{\alpha'}^U \big| {\mathcal {A}}_1] \big) {\mathbf{P}}_y[ {\mathcal {A}}_1]\\
=& {\mathbf{P}}_y[\tau(E) \geq T_\alpha^U - T_{\alpha'}^U] + O\left( e^{-c(\alpha-\alpha')}+ |E|\rho(s,r) +o(1) \right),\end{aligned}$$ the last line coming from a similar analysis as before. Consequently, $$\begin{aligned}
\frac{{\mathbf{P}}_y[\tau(E) \geq T_\alpha^U \big| X(T_\alpha^U) = z]}{{\mathbf{P}}_y[{\mathcal {A}}]}
&= 1 + O\left( e^{-c(\alpha-\alpha')} + |E|\rho(s,r) +o(1)\right).\end{aligned}$$ Taking $\alpha' = \alpha/2$ gives the lemma.
Let $\tau_k(E) = \tau_k^0(E)$, $\sigma_k(E) = \sigma_k^0(E)$, and $T_{r,R}(E) = {\mathbf{E}}_\pi[\tau_1(E) - \tau_0(E)]$.
\[lem::mean\_excursion\_length\] For every $r,\delta > 0$ there exists $R_0 > r$ such that $R \geq R_0$ implies $$(1-\delta) T_{r,R}(E) \leq \min_{y \notin E(R)} {\mathbf{E}}_y \tau_0(E) \leq \max_{y \notin E(R)} {\mathbf{E}}_y \tau_0(E) \leq (1+\delta) T_{r,R}(E)$$ for all $n$ large enough.
We have that $$\begin{aligned}
{\mathbf{E}}_\pi[ \tau_1(E) - \tau_0(E)]
&= {\mathbf{E}}_\pi[ \sigma_0(E) - \tau_0(E)] + {\mathbf{E}}_\pi[ \tau_1(E) - \sigma_0(E)].\end{aligned}$$ Obviously, $$\begin{aligned}
{\mathbf{E}}_\pi[\sigma_0(E) - \tau_0(E)]
&\leq \max_{y \in E(r)} {\mathbf{E}}_y \sigma_0(E)
\leq c T_{{\rm mix}}^U(G)\end{aligned}$$ for some $c > 0$ since in each interval of length $T_{{\rm mix}}^U(G)$, random walk started in $E(r)$ has probability uniformly bounded from below of leaving $E(R)$ provided $n$ is large enough. It is also obvious that $$\min_{y \notin E(R)} {\mathbf{E}}_y \tau_0(E) \leq {\mathbf{E}}_\pi[ \tau_1(E) - \sigma_0(E)] \leq \max_{y \notin E(R)} {\mathbf{E}}_y \tau_0(E).$$ The previous lemma implies $$(1- \delta) {\mathbf{E}}_\pi[\tau_0(E)] \leq {\mathbf{E}}_y[\tau_0(E) | {\mathcal {A}}] \leq T_\alpha^U + (1+\delta) {\mathbf{E}}_\pi[ \tau_0(E)]$$ for all $y \notin E(R)$ provided we choose $R,\alpha,s,n$ large enough to accommodate our choice of $\delta$. Hence, $$(1-\delta){\mathbf{E}}_\pi[\tau_0(E) ] \leq {\mathbf{E}}_y[\tau_0(E)] \leq (1+\delta){\mathbf{E}}_\pi[\tau_0(E) ],$$ as it is not difficult to see that $T_{{\rm mix}}^U(G) = o(T_{r,R}(E))$ as $n \to \infty$. Therefore $$\max_{y_1,y_2 \notin E(R)} \frac{{\mathbf{E}}_{y_1} \tau_0(E)}{{\mathbf{E}}_{y_2} \tau_0(E)} \leq 1+\delta,$$ which proves the lemma.
\[lem::excursion\_concen\] For each $\beta \geq 0$ and $r, \delta > 0$ there exists $R_0 > r$ such that $$\begin{aligned}
&{\mathbf{P}}_{y}\left[ \tau_k^\beta(E) \leq (1-\delta) T_{r,R}(E) k \right] \leq e^{-C \delta^2 k} \label{eqn::excursion_concen_lower}\\
&{\mathbf{P}}_{y}\left[ \tau_k^\beta(E) \geq (1+\delta) T_{r,R}(E) k \right] \leq e^{-C \delta^2 k} \label{eqn::excursion_concen_upper}\end{aligned}$$ for all $R \geq R_0$, $y \in V$, and $n$ large enough.
We adapt the proof of Lemma 2.4 of [@DPR_MAN03] to our setting. First of all, it follows from Lemma \[lem::mean\_excursion\_length\] that $$\begin{aligned}
\max_{y} {\mathbf{E}}_y[\tau_0(E)] \leq CT_{r,R}(E)\end{aligned}$$ for some $C > 0$ provided $R,n$ are sufficiently large. Consequently, Kac’s moment formula (see [@FP99], equation 6) for the first hitting time of a strong Markov process implies for any $j \in {\mathbf{N}}$ we have that $$\label{eqn::kac_moment}
\max_y {\mathbf{E}}_y[ (\tau_0(E))^j ] \leq j! c^j T_{r,R}^j(E)$$ for some $c > 0$. This implies that there exists $\lambda_0 > 0$ so that $$\max_{y} {\mathbf{E}}_y \exp[ \lambda \tau_0(E)/T_{r,R}(E)] < \infty \text{ for all } \lambda \in (0,\lambda_0).$$ Using ${\mathbf{E}}[\sigma_0(E) - \tau_0(E)] = o(T_{r,R}(E))$, a similar argument implies that, by possibly decreasing $\lambda_0$, $$\max_{y} {\mathbf{E}}_y \exp[\lambda \sigma_0(E)/T_{r,R}(E)] < \infty \text{ for all } \lambda \in (0,\lambda_0).$$ Combining the strong Markov property with $T_\beta^U = o(T_{r,R}(E))$ yields $$\max_{y} {\mathbf{E}}_y \exp[\lambda \tau_k^\beta(E) / T_{r,R}(E)] < \infty \text{ for all } \lambda \in (0,\lambda_0).$$ Let $R_0$ be large enough so that the previous lemma implies $$(1-\delta/2) T_{r,R}(E) \leq \min_{y \notin E(R)} {\mathbf{E}}_y \tau_0(E) \leq \max_{y \notin E(R)} {\mathbf{E}}_y \tau_0(E) \leq (1+\delta/2) T_{r,R}(E)$$ for $R \geq R_0$ and $n$ large enough. We compute, $$\begin{aligned}
&\max_{y \notin E(R)} {\mathbf{E}}_y e^{-\theta \tau_0(E)}
\leq 1 - \theta \min_{y \notin E(R)} {\mathbf{E}}_y \tau_0(E) + \theta^2 \max_{y \notin E(R)} {\mathbf{E}}_y \tau_0^2(E)\\
&\leq 1 - \theta(1-\delta/2) T_{r,R}(E) + \rho \theta^2 \leq \exp(\rho \theta^2 - \theta(1-\delta/2) T_{r,R}(E)).\end{aligned}$$ where $\rho = c T_{r,R}^2(E)$ for some $c > 0$. Since $\tau_0(E) \geq 0$, Chebychev’s inequality leads to as $$\begin{aligned}
& {\mathbf{P}}_{y}\left[ \tau_k^\beta(E) \leq (1-\delta) T_{r,R}(E) k \right]
\leq \exp(\theta(1-\delta) T_{r,R}(E) k) {\mathbf{E}}_{y} e^{- \theta \tau_k^\beta(E)}\\
\leq& \exp(\theta (1-\delta)T_{r,R}(E) k)\left[ \max_{y \notin E(R)} {\mathbf{E}}_y e^{-\theta \tau_0(E)} \right]^k\\
\leq& \exp(\theta (1-\delta)T_{r,R}(E) k) \exp(\rho \theta^2 k - \theta(1-\delta/2) T_{r,R}(E) k)\end{aligned}$$ Taking $$\theta = \frac{\delta T_{r,R}(E)}{ c_1\rho}$$ we get that $$\begin{aligned}
& {\mathbf{P}}_{y}\left[ \tau_k^\beta(E) \leq (1-\delta) T_{r,R}(E) k \right]
\leq \exp(\rho \theta^2 k -\theta T_{r,R}(E) k \delta/2)\\
\leq& \exp(\rho \delta^2 T_{r,R}^2(E) k / (c_1^2\rho^2)-\delta^2 T_{r,R}^2(E) k / (2 c_1\rho))
\leq \exp(-c \delta^2 k),\end{aligned}$$ provided we take $c_1$ sufficiently large.
To prove , we need to bound $$\begin{aligned}
&{\mathbf{P}}_{y}\left[ \tau_k^\beta(E) \geq (1+\delta) T_{r,R}(E) k \right]\\
\leq& \exp(-\theta (1+\delta) T_{r,R}(E) k ) \left( e^{ \theta T_\beta^U} \max_{y} {\mathbf{E}}_y e^{\theta \tau_0(E)} \max_{y \in E(r)} {\mathbf{E}}_y e^{\theta[\sigma_0(E) - \tau_0(E)]} \right)^k\end{aligned}$$ We note that $$\begin{aligned}
&\max_{y } {\mathbf{E}}_y e^{\theta \tau_0(E)}
\leq (1+o(1)) \max_{y \notin E(R)} {\mathbf{E}}_y e^{\theta \tau_0(E)}\\
\leq& \exp(\theta(1+\delta/2) T_{r,R}(E) + \rho \theta^2 + o(1)).\end{aligned}$$ Take $$\theta = \frac{\delta T_{r,R}(E)}{c_1 \rho},$$ with $c_1$ to be fixed shortly. Since $\max_{y \in E(r)} {\mathbf{E}}_y [\sigma_0(E) - \tau_0(E)] = o(T_{r,R}(E))$ as $n \to \infty$, Kac’s formula yields $$\max_{y \in E(r)} {\mathbf{E}}_y e^{\theta [\sigma_0(E) - \tau_0(E)]} = 1+o(1) \text{ as } n \to \infty.$$ Since $T_\beta^U = o(T_{r,R}(E))$ as $n \to \infty$ as well, we have $$\begin{aligned}
& {\mathbf{P}}_{y}[ \tau_k^\beta(E) \geq (1+\delta) T_{r,R}(E) k]\\
\leq& \exp(-\theta(1+\delta) T_{r,R}(E) k + \theta(1+\delta/2) T_{r,R}(E) k + \rho \theta^2 k + o(1) k)\\
\leq& \exp(-\theta \delta T_{r,R}(E) k/2 + \rho \theta^2 k + o(1) k).\end{aligned}$$ Taking $c_1 > 0$ large enough gives the result.
Hitting and Covering
--------------------
\[lem::hitting\_estimate\] For every $\delta > 0$ there exists $r_0$ such that for each $r \geq r_0$ there is an $R_0 > r$ so that if $R \geq R_0$ the following holds. If $E_n = \{x_{n1},\ldots,x_{n\ell}\} \subseteq V_n$ with $d(x_{ni}, x_{nj}) \geq 2R$ for $i \neq j$ and $y_n \in V_n$ is such that $d(x_{ni}, y_n) \geq 2R$ for all $n$, then $$\begin{aligned}
1- \delta &\leq \liminf_{n \to \infty} \frac{{\mathbf{E}}_{y_n} \tau(x_{ni})}{ T_{r,R}(E_n)/{\overline}{p}_{r,R}(x_{ni};E)}\\
&\leq \limsup_{n \to \infty} \frac{{\mathbf{E}}_{y_n} \tau(x_{ni})}{ T_{r,R}(E_n)/{\overline}{p}_{r,R}(x_{ni};E)} \leq 1+\delta.
\end{aligned}$$
We will omit the indices $n$ and $i$ and just write $x$ for $x_{ni}$, $y$ for $y_n$, and $E$ for $E_n$. Fix $r$ sufficiently large so that $\rho(r) < \delta^2/100$. Let $N(x;E) = \min\{ k \geq 1 : {\mathbf{1}}_{S_k^{\alpha,\beta}(x)} = 1\}$ and let $${\widetilde}{\tau}(x) = \min\{ t \geq 0 : X(t) = x \text{ and } t \in I\}$$ where $$I_k = [\tau_k^\beta(x), \sigma_k^\beta(x) + T_\alpha^U] \text{ and } I = \cup_k I_k.$$ Then $$\begin{aligned}
\tau_{N(x;E)}^\beta(E) \leq {\widetilde}{\tau}(x)
\leq \tau_{N(x;E)+1}^\beta(E).\end{aligned}$$ Let $$A(M;\delta) = \bigcap_{j \geq M} B(j;\delta) \equiv \bigcap_{j \geq M} \left\{ (1-\delta) T_{r,R}(E) j \leq \tau_j^\beta(E) \leq (1+\delta)T_{r,R}(E) j \right\}.$$ With $\| {\widetilde}{\tau}(x) \| = \max_z {\mathbf{E}}_z {\widetilde}{\tau}(x)$, note that $$\begin{aligned}
& {\mathbf{E}}_y {\widetilde}{\tau}(x) {\mathbf{1}}_{A^c(M;\delta)}
\leq \sum_{j \geq M} {\mathbf{E}}_y {\widetilde}{\tau}(x) {\mathbf{1}}_{B^c(j;\delta)} \notag\\
\leq& \sum_{j \geq M} \bigg[ {\mathbf{E}}_y \tau_j^\beta(E) {\mathbf{1}}_{B^c(j;\delta)} + \| {\widetilde}{\tau}(x) \| {\mathbf{P}}[ B^c(j;\delta)] \bigg] \notag\\
\leq& 2C \sum_{j \geq M} \bigg[ j T_{r,R}(E) + \| {\widetilde}{\tau}(x)\| \bigg] e^{-C \delta^2 j} \notag\\
\leq& C_1 \| {\widetilde}{\tau}(x)\| \sum_{j \geq M} (1+j) e^{-C \delta^2 j}
\leq C_2 \| {\widetilde}{\tau}(x)\| \frac{e^{-C \delta^2 M}}{\delta^4} \label{eqn::tilde_tau_trunc}.\end{aligned}$$ In the third step we used that $$\begin{aligned}
&{\mathbf{E}}_y \tau_j^\beta(E) {\mathbf{1}}_{B^c(j;\delta)}
\leq ({\mathbf{E}}_y [\tau_j^\beta(E)]^2)^{1/2} ({\mathbf{P}}[B^c(j;\delta)])^{1/2}\\
\leq&
\frac{2T_{r,R}(E)}{\lambda} j \left( {\mathbf{E}}_y \exp(\lambda \tau_j^\beta(E) / (j T_{r,R}(E))) \right)^{1/2} C e^{-C \delta^2 j},\end{aligned}$$ where $\lambda \in (0,\lambda_0)$, $\lambda_0$ as in the proof of Lemma \[lem::excursion\_concen\]. Uniform local transience implies $$\big| {\mathbf{E}}_y {\widetilde}{\tau}(x) - \| {\widetilde}{\tau}(x) \| \big| \leq \delta {\mathbf{E}}_y {\widetilde}{\tau}(x)$$ when $R$ is large enough. Consequently, there exists $M > 0$ large enough depending only on $\delta$ so that $${\mathbf{E}}_y {\widetilde}{\tau}(x) {\mathbf{1}}_{A(M;\delta)} \leq {\mathbf{E}}_y {\widetilde}{\tau}(x) \leq (1+\delta) {\mathbf{E}}_y {\widetilde}{\tau}(x) {\mathbf{1}}_{A(M;\delta)}.$$ Now, $$\begin{aligned}
& {\mathbf{E}}_y \tau_{N(x;E)+1}^\beta(E) {\mathbf{1}}_{A(M;\delta)}
= {\mathbf{E}}_y \left[ N(x;E) \left( \frac{\tau_{N(x;E)+1}^\beta(E)}{N(x;E)}\right) {\mathbf{1}}_{A(M;\delta)} \right]\\
\leq& (1+\delta) T_{r,R}(x) {\mathbf{E}}_y N(x;E) +
{\mathbf{E}}_y \tau_M^\beta(E)
\leq (1+\delta) T_{r,R}(E) {\mathbf{E}}_y N(x) + C M T_{r,R}(E).\end{aligned}$$ and, similarly, $${\mathbf{E}}_y \tau_{N(x;E)}(E) {\mathbf{1}}_{A(M;\delta)} \geq (1-\delta) T_{r,R}(E) {\mathbf{E}}_y N(x;E).$$ Therefore $${\mathbf{E}}_y {\widetilde}{\tau}(x) \leq (1+2\delta) T_{r,R}(E) {\mathbf{E}}_y N(x;E) + C M T_{r,R}(E).$$ By Lemma \[lem::prob\_success\_concen\], $$\begin{aligned}
&{\overline}{p}_{r,R}(x;E) \left[\exp(-(1+\delta) {\overline}{p}_{r,R}(x;E) j)- C\exp\left(-\frac{C \delta^2 {\overline}{p}_{r,R}(x;E)}{\rho(r)} j\right) \right]\\
\leq& {\mathbf{E}}_y p_{j+1}^{\alpha,\beta}(x;E) \exp\left(-[1+O(\rho(r))]\sum_{i=1}^j p_i^{\alpha,\beta}(x;E) \right)\\
\leq& {\overline}{p}_{r,R}(x;E) \left[\exp(-(1-\delta) {\overline}{p}_{r,R}(x;E) j) +C \exp\left(-\frac{C \delta^2 {\overline}{p}_{r,R}(x;E)}{\rho(r)} j\right)\right].\end{aligned}$$ Taking $r$ sufficiently large gives $$\begin{aligned}
& {\mathbf{E}}_y N(x;E)
= \sum_{j=1}^\infty j {\mathbf{P}}[ N(x;E) = j]\\
\leq& CM^2 \rho(r) + \sum_{j=M+1}^\infty j (1+o(1))\left({\overline}{p}_{r,R}(x;E) \exp(-(1-\delta) {\overline}{p}_{r,R}(x;E) j) \right)\\
\leq& 2C M^2 \rho(r) + \frac{1+\delta}{{\overline}{p}_{r,R}(x;E)}.\end{aligned}$$ Similarly, $${\mathbf{E}}_y N(x;E) \geq \frac{1-\delta}{{\overline}{p}_{r,R}(x;E)}.$$ Increasing $r$ if necessary so that $M^2 \rho(r) \leq \delta$ yields $$\frac{1-2\delta}{{\overline}{p}_{r,R}(x;E)} \leq {\mathbf{E}}_y N(x;E) \leq \frac{1+2\delta}{{\overline}{p}_{r,R}(x;E)}.$$ This proves that ${\mathbf{E}}_y {\widetilde}{\tau}(x) = (1+o(1)) \frac{T_{r,R}(E)}{{\overline}{p}_{r,R}(x;E)}$ as $n \to \infty$. By mixing considerations, the probability of the event $F_k$ that $X$ hits $E(r)$ in $J_k = [\sigma_k^\beta(E) + T_\alpha^U, \sigma_k^\beta(E) + T_\beta^U]$ is $O(T_\beta^U |E| {\overline}{\Delta}^r(G)/ |V|)$. With $F = \cup_{k=1}^{N(x;E)+1} F_k$, we have $${\mathbf{E}}_y {\widetilde}{\tau}(x) {\mathbf{1}}_{F^c} \leq {\mathbf{E}}_y \tau(x) \leq {\mathbf{E}}_y {\widetilde}{\tau}(x)$$ and, analogous to the proof of , $${\mathbf{E}}_y {\widetilde}{\tau}(x) {\mathbf{1}}_{F^c} = \bigg[ 1 + O\left( \frac{ T_\beta^U |E| {\overline}{\Delta}^r(G) }{|V| {\overline}{p}_{r,R}(x;E)} \right)^{1/2} \bigg]{\mathbf{E}}_y {\widetilde}{\tau}(x).$$ The lemma now follows since ${\overline}{p}_{r,R}(x;E) \geq C{\underline}{\Delta}^{-r}(G) $.
We will now specialize to the case $E = \{x\}$; for simplicity of notation we will omit $E$. Let $$O_{r,R}(x) = \frac{{\overline}{a}_{r,R}(x)}{T_{r,R}(x)} .$$
For every $\delta > 0$ there exists $r_0$ so if $r \geq r_0$ there is $R_0 > r$ such that $R \geq R_0$ implies $$(1-\delta)\pi(x) \leq O_{r,R}(x) \leq (1+\delta)\pi(x)$$ for all $n$ large enough.
Let $N(x,T) = \min\{ k : \tau_k^\beta(x) \geq T\}$, $J_k$ as in the previous lemma, $J = \cup_k J_k$, and ${\mathcal {G}}(x) = \sigma( X(\tau_j^\beta(x)) : j \geq 1)$. Then $$\sum_{j=1}^{N(x,T)} a_j^{\alpha,\beta}(x) \leq {\mathbf{E}}\left[ \sum_{t=1}^T {\mathbf{1}}_{\{X(t) = x\}} {\mathbf{1}}_{t \notin J} \bigg| {\mathcal {G}}(x) \right] \leq \sum_{j=1}^{N(x,T)+1} a_j^{\alpha,\beta}(x).$$ Lemmas \[lem::prob\_success\_concen\] and \[lem::excursion\_concen\] give that $$\begin{aligned}
(1-\delta) T_{r,R}(x) &\leq \frac{N(x,T)}{T} \leq (1+\delta) T_{r,R}(x) \text{ and }\\
(1-\delta) {\overline}{a}_{r,R}(x) &\leq \frac{\sum_{j=1}^k a_j^{\alpha,\beta}(x)}{k} (1+\delta) {\overline}{a}_{r,R}(x),\end{aligned}$$ with high probability as $T \to \infty$, for all $r,R,k,n,\beta-\alpha$ large enough. Consequently, using that $(a_j^{\alpha,\beta}(x) : j \geq 1)$ is uniformly bounded, it is not hard to see that $$(1-\delta)\frac{{\overline}{a}_{r,R}(x)}{T_{r,R}(x)} \leq \frac{1}{T} \sum_{j=1}^{N(x,T)} a_j^{\alpha,\beta}(x) \leq (1+\delta)\frac{{\overline}{a}_{r,R}(x)}{T_{r,R}(x)},$$ with high probability as $T \to \infty$, for all $r,R,n,\beta-\alpha$ large enough. The middle term converges to $\pi(x)$ as $T \to \infty$ since $$\lim_{T \to \infty} \frac{1}{T} {\mathbf{E}}\sum_{t=1}^T {\mathbf{1}}_{\{ X(t) \in E(r) \}} {\mathbf{1}}_{t \in J} = 0.$$
Uniform local transience implies that there exists constants $c,C > 0$ so that $c {\overline}{a}_{r,R}(x) \leq {\overline}{p}_{r,R}(x) \leq C {\overline}{a}_{r,R}(x)$; combining this with the previous lemma yields $$\frac{c \deg(x)}{|V|} \leq \frac{{\overline}{p}_{r,R}(x)}{T_{r,R}(x)} \leq \frac{C \deg(x)}{|V|}.$$ Let $\epsilon > 0$ and let $$H_{n,k}^\epsilon = \left\{ x \in V_n : \frac{{\underline}{\Delta}(G_n) k \epsilon}{|V_n|} < \frac{{\overline}{p}_{r,R}(x)}{T_{r,R}(x)} \leq \frac{{\underline}{\Delta}(G_n)(k+1)\epsilon}{|V_n|} \right\}$$ be a partition of $V_n$ into at most $\Delta_0 \epsilon^{-1} $ subsets, where $\Delta_0$ is the constant from Assumption \[assump::graphs\]. By passing to a subsequence, we may assume without loss of generality that $$d_k^\epsilon = \lim_{n \to \infty} d_{n,k}^\epsilon \equiv \lim_{n \to \infty} \frac{\log|H_{n,k}^\epsilon|}{\log |V_n|}$$ exists for every $k$. Note that $d_k^\epsilon \in [0,1]$ for those $k$ so that $|H_{n,k}^\epsilon| \neq 0$ for all $n$ large enough and, since the partition is finite, necessarily there exists $k$ so that $d_k^\epsilon = 1$. In particular, there exists $k$ so that $d_k^\epsilon \neq 0$. Let $$C_{n,k}^\epsilon = \frac{|V_n|}{ {\underline}{\Delta}(G_n) k \epsilon} d_k^\epsilon \log |V_n| \text{ and }
C_n^\epsilon = \max_k C_{n,k}^\epsilon.$$
\[lem::cover\_time\] For each $\delta > 0$ there exists $r_0,\epsilon_0$ so if $r \geq r_0$ there is $R_0 > r$ such that $R \geq R_0$ and $\epsilon \in (0,\epsilon_0)$ implies $$\label{eqn::subset_cover_bound}
1-\delta \leq \liminf_{n \to \infty} \frac{T_{{\rm cov}}(H_{n,k}^\epsilon)}{C_{n,k}^\epsilon} \leq \limsup_{n \to \infty} \frac{T_{{\rm cov}}(H_{n,k}^\epsilon)}{C_{n,k}^\epsilon} \leq 1+\delta.$$ for all $k$ with $d_k^{\epsilon} > 0$. Furthermore, $$\label{eqn::cover_time_bound}
1-\delta \leq \liminf_{n \to \infty} \frac{T_{{\rm cov}}(G_n)}{C_n^\epsilon} \leq \limsup_{n \to \infty} \frac{T_{{\rm cov}}(G_n)}{C_n^\epsilon} \leq 1+\delta$$
Suppose $k$ is such that $d_k^\epsilon > 0$. Then $|H_{n,k}^\epsilon| \to \infty$ as $n \to \infty$. Let $r,R,n > 0$ be sufficiently large so that Lem ma \[lem::hitting\_estimate\] applies with our choice of $\delta$. Since $\log |B(x,r)| = o(\log |V_n|)$, it follows that for all $n$ large enough there exists an $R$-net $E_{n,k}^\epsilon$ of $H_{n,k}^\epsilon$ such that $$\log |E_{n,k}^\epsilon| = \log |H_{n,k}^\epsilon| + o(1) \text{ as } n \to \infty.$$ The upper and lower bounds from the Matthews method ([@M88]; see also Theorem 11.2 and Proposition 11.4 of [@LPW08]) combined with the definition of $C_{n,k}^\epsilon$ imply . Theorem 2 of [@A91] implies that $$\lim_{n \to \infty} \frac{\tau_{{{\rm cov}}}(H_{n,k}^\epsilon)}{{\mathbf{E}}\tau_{{{\rm cov}}}(H_{n,k}^\epsilon)} = 1.$$ As $\tau_{{\rm cov}}(G_n) = \max_k \tau_{{\rm cov}}(H_{n,k}^\epsilon)$ and the maximum is over a finite set, it follows that $\tau_{{\rm cov}}(G_n) = (1+o(1))\max_k T_{{\rm cov}}(H_{n,k}^\epsilon)$. Taking expectations of both sides gives .
Correlation Decay {#sec::corr_decay}
=================
Note that vertex transitivity implies ${\overline}{p}_{r,R}(\cdot)$ and $T_{r,R}(\cdot)$ do not depend on their arguments.
First, assume that we are in the case of bounded maximal degree. Let $E$ be as in the previous section and let $\delta > 0$ be arbitrary. Fix $r$ so that $\rho(r) \leq \delta^3/100 C \ell$ and ${\overline}{p}_{r,R} \leq \delta^3$ where $\ell = |E|$. Let $R_0 > r$ and $\beta-\alpha$ be sufficiently large so that Lemmas \[lem::prob\_success\_concen\] and \[lem::excursion\_concen\] apply with our choice of $\delta,r$. Finally, let $N(x_i) = \min\{ k : S_k^{\alpha,\beta}(x_i) \text{ occurs}\}$ and ${\mathcal {G}}(E) = \sigma(p_j^{\alpha,\beta}(x;E) : x \in E, j \geq 1)$. Since $d(x_i,x_j) \geq 2R$, the probability that $X$ neither hits $x$ nor $x'$ in the interval $[\tau_j^\beta(x;E), \sigma_j^\beta(x;E) + T_\alpha^U]$ is $$\label{eqn::p_j_E_union_bound}
1 - [1+ O(\rho(R))][ p_j^{\alpha,\beta}(x;E) + p_j^{\alpha,\beta}(x';E)].$$ This holds more generally for any subset of $E$, hence $$\begin{aligned}
&{\mathbf{E}}\big[ {\mathbf{P}}[ N(x_1) > k_1,\ldots, N(x_\ell) > k_\ell | {\mathcal {G}}(E)] \big]\\
=& {\mathbf{E}}\prod_{i=1}^\ell \exp\left( -[1 + O(\rho(r))] \sum_{j=1}^{k_i} p_j^{\alpha,\beta}(x_i;E) \right)\\
=& \exp\left( -[1+O(\delta)] \sum_{i=1}^\ell {\overline}{p}_{r,R}(x_i;E) k_i \right) + \sum_{i=1}^\ell O(\exp(- {\overline}{p}_{r,R}(x_i;E) \ell k_i / \delta)),\end{aligned}$$ where the last equality followed from our choice of $r$ and Lemma \[lem::prob\_success\_concen\]. Combining this with Lemma \[lem::excursion\_concen\] and that the probability $X$ hits $E(r)$ in $J_k$ is at most $O(T_\beta^U|E| {\overline}{\Delta}^r(G) / |V|) = o({\overline}{p}_{r,R}(x;E))$ for any $x \in E$, we have $$\begin{aligned}
& {\mathbf{P}}[ \tau(x_1) \geq k T_{r,R}(E) / {\overline}{p}_{r,R}(x_1;E) ,\ldots, \tau(x_\ell) \geq k T_{r,R}(E) / {\overline}{p}_{r,R}(x_n;E)]\\
=& (1+o(1)) \exp\left( -[1+O(\delta)] \ell k \right) + O(\exp( - C\delta^2 k /\rho(r)))\\
=& (1+o(1)) \exp\left( -[1+O(\delta)] \ell k \right)\end{aligned}$$ By vertex transitivity, $$T_{{\rm hit}}(G) = (1+o(1)) \frac{T_{r,R}(x_i)}{{\overline}{p}_{r,R}(x_i)} = (1+o(1)) \frac{T_{r,R}(E)}{{\overline}{p}_{r,R}(x_i;E)}.$$ Using that the cover time is asymptotically $T_{{\rm hit}}(G) \log|V|$ gives the result.
This proof works also for unbounded degree, but is not quite sufficient for the statement of our theorem since we would like to allow for points in $E$ to be adjacent. There are two parts that break down. First, in Section \[sec::cov\_hit\] we proved the concentration of $p_j^{\alpha,\beta}(x;E)$ when $x \in E$ and we also assumed that $x,y \in E$ implies $d(x,y) \geq 2R$. To allow for $x,y$ adjacent, we define $$p_j^{\alpha,\beta}(y;E) = {\mathbf{P}}[ S_j^{\alpha,\beta}(y;E) | X(\tau_j^\beta(E)), X(\tau_{j+1}^\beta(E))]$$ for $y \in E(r/2)$. It is not difficult to see that for such $y$, $p_j^{\alpha,\beta}(y;E)$ exhibits nearly the same concentration behavior as for $y \in E$. Second, the estimate is no longer good enough since $\rho(1)$ does not decay in $n$. However, it is not difficult to see that the same probability satisfies the estimate $$1 - [1+ O({\overline}{\Delta}^{-1}(G)) ][ p_j^{\alpha,\beta}(x;E) + p_j^{\alpha,\beta}(x';E)],$$ which suffices since ${\overline}{\Delta}^{-1}(G_n) \to 0$ as $n \to \infty$. The rest of the proof is the same.
Vertex transitivity was used only to get that $T_{r,R}(x;E) / {\overline}{p}_{r,R}(x;E) = (1+o(1)) T_{{\rm hit}}(G)$. The same proof works more generally, but leads to more complicated formulae. However, it is not difficult to see that the upper bound takes a very similar form:
\[lem::corr\_decay\_upper\_bound\] If $(x_n^i)$ for $1 \leq i \leq \ell$ is a family of sequences with $x_n^i \in H_{n,k(i)}^\epsilon$ and $|x_n^i - x_n^j| \geq r$ for every $n$ and $i \neq j$, $$\begin{aligned}
\label{eqn::correlation_decay_upper_bound}
{\mathbf{P}}[ x_n^i \in {\mathcal {L}}(\alpha; G_n) \text{ for all } i]
\leq (1+\delta_{r,\ell}) |V_n|^{-\ell d_k^\epsilon \alpha+\delta_{r,\ell}}
\end{aligned}$$ where $\delta_{r,\ell} \to 0$ as $r \to \infty$ while $\ell$ is fixed. If ${\overline}{\Delta}(G_n) \to \infty$ then we take $r = 1$ and $\delta_{1,\ell} = o(1)$ as $n \to \infty$.
Total Variation Bounds {#sec::total_variation}
======================
Lower Bound
-----------
We will now prove the lower bound for Theorems \[thm::threshold\] and \[thm::lamplighter\]. This is actually just a slight extension of Theorem 4.1 of [@PR04], but we include it for the reader’s convenience. Recall from the introduction that $\mu(\cdot;\alpha,G)$ is the probability measure on ${\mathcal {X}}(G) = \{f \colon V \to \{0,1\}\}$ given by first sampling ${\mathcal {R}}(\alpha;G)$ then setting $$f(x) = \begin{cases}
\xi(x) \text{ if } x \in {\mathcal {R}}(\alpha;G),\\
0 \text{ otherwise,}
\end{cases}$$ where $(\xi(x) : x \in V)$ is a collection of iid variables such that ${\mathbf{P}}[\xi(x) = 0] = {\mathbf{P}}[\xi(x) = 1] = \tfrac{1}{2}$ and $\nu(\cdot;G)$ is the uniform measure on ${\mathcal {X}}(G)$.
\[lem::lower\_bound\] For every $\delta > 0$, $$\lim_{n \to \infty} \| \mu(\cdot;\tfrac{1}{2}-\delta, G_n) - \nu(\cdot;G_n)\|_{TV} = 1.$$
For $A \subseteq V$ and $m > 0$, let $\tau_{{{\rm cov}}}(A;m)$ be the first time all but $m$ of the vertices of $A$ have been visited by $X$. For each $k$ such that $d_k^\epsilon > 0$ we will show that $$\label{eqn::cover_bound} \lim_{n \to \infty} {\mathbf{P}}[ \tau_{{\rm cov}}( H_{n,k}^\epsilon, |H_{n,k}^\epsilon|^{\alpha}) < (1-\alpha-\delta) C_{n,k}^\epsilon] = 0$$ for each $\delta > 0$ and $\epsilon \in (0,\epsilon_0(\delta))$. If not, then for some such $k,\delta,\alpha$ we have $$\limsup_{n \to \infty} {\mathbf{P}}[ A_{n,k}(\alpha,\delta)] > 0$$ where $$A_{n,k}(\alpha,\delta) = \{\tau_{{\rm cov}}( H_{n,k}^\epsilon, |H_{n,k}^\epsilon|^{\alpha}) < (1-\alpha-\delta) C_{n,k}^\epsilon\}.$$ It follows from the Matthews method upper bound ([@M88]; see also Theorem 11.2 of [@LPW08]) that $$\begin{aligned}
&{\mathbf{E}}[ \tau_{{\rm cov}}(H_{n,k}^\epsilon) - \tau_{{{\rm cov}}}(H_{n,k}^\epsilon, |H_{n,k}^\epsilon|^\alpha) | A_{n,k}(\alpha,\delta)]\\
\leq& \alpha(1+O(\epsilon)) C_{n,k}^\epsilon \leq \alpha(1+\delta/4) C_{n,k}^\epsilon,\end{aligned}$$ where we take $\epsilon$ so small that the $O(\epsilon)$ term is at most $\delta/4$. Markov’s inequality now implies $${\mathbf{P}}[ \tau_{{{\rm cov}}}(H_{n,k}^\epsilon) < (1 - \delta/2) C_{n,k}^\epsilon | A_{n,k}(\alpha,\delta)] > 0.$$ This is a contradiction as Theorem 2 of [@A91] implies $\tau_{{{\rm cov}}}(H_{n,k}^\epsilon)/ C_{n,k}^\epsilon \to 1$ in probability.
For each $n$ let $k_0(n)$ be an index that achieves the maximum in $\max_k C_{n,k}^\epsilon$. Now, implies that whp at time $\tfrac{1}{2}(1-3\delta) T_{{\rm cov}}(G_n) = \tfrac{1}{2}(1-3\delta + O(\epsilon)) C_{n,k_0(n)}^\epsilon$ the size of the subset of $H_{n,k_0(n)}^\epsilon$ not visited by $X$ is at least $|H_{n,k_0(n)}^\epsilon|^{(1+2\delta + O(\epsilon))/2}$ but less than $|H_{n,k_0(n)}^\epsilon|^{(1+4\delta + O(\epsilon))/2}$. Thus the number of zeros in a marking of $H_{n,k_0(n)}^\epsilon$ sampled from $\mu(\cdot;\tfrac{1}{2}(1-3\delta),G_n)$ is whp at least $$\tfrac{1}{2}| H_{n,k_0(n)}^\epsilon| + (1+o(1))|H_{n,k_0(n)}^\epsilon|^{(1+2\delta + O(\epsilon))/2} \text{ as } n \to \infty.$$ This proves the lemma since the probability of having deviations of this magnitude from the mean tends to zero in a uniform marking.
Concentration of $q_j$
----------------------
Let $q_j(x) = {\mathbf{P}}[S_j(x) | X(\tau_j(x)), X(\sigma_j(x))]$. Note that $q_j(x)$ is the not the same as $p_j^{\alpha,\beta}(x)$ from Section \[sec::cov\_hit\]. Indeed, the excursions on which we condition are different since we do not allow the random walk to run for a multiple for $T_{{\rm mix}}^U(G)$ after exiting $\partial B(x,R)$ and we condition on the entrance and exit points of the current excursion rather than the entrance points of the current and successive excursion. While both of these changes may seem cosmetic, they affect the concentration behavior, since while $p_j^{\alpha,\beta}(x)$ satisfies , in locally tree-like graphs it can be that $q_j(x) = 1$ with positive probability.
Suppose that $(G_n)$ satisfies Assumption \[assump::far\_away\_harnack\] part . Let $\epsilon > 0$ be arbitrary, $R_n^\gamma$ be as in Assumption \[assump::far\_away\_harnack\], $\gamma > 0$ to be determined later, and let $E$ be a set of points in $V_n$ such that if $x,y$ are distinct in $E$ then $d(x,y) \geq 4 R_n^\gamma$. Fix $R > r > 0$ and let $\tau_{k+1}(E) = \min\{ t \geq \sigma_k(x) : X(t) \in \partial E(r)\}.$ Fix $\beta > 0$ and define indices $i(j,x)$ inductively as follows. Set $$i(1,x) = \min\{ k \geq 1 : \tau_{k+1}(E) - \sigma_k(x) \geq T_\beta^U \}$$ and, for each $j \geq 1$, let $$i(j+1,x) = \min\{ k \geq i(j,x) + 1 :
\tau_{k+1}(E) - \sigma_k(x) \geq T_\beta^U \}.$$ When $x$ is clear from the context we will write $i(j)$ for $i(j;x)$.
\[lem::iid\_replacement\] For each $\delta > 0$ and $r > 0$ there exists $R_0 > r$ such that for $R > R_0$ fixed there exists iid random variables $( I(j,x) : x \in E, j \geq 1)$ which stochastically dominate from above $(i(j,x) : x \in E, j \geq 1)$ and satisfy $${\mathbf{P}}[ I((1-\delta)j,x) \geq j] \leq C\exp(-C \delta^2 j)$$ for all $n$ large enough. Let ${\mathcal {G}}(j,x) = \sigma(\{q_{i(k)}(x) : k \neq j\} \cup \{ q_{i(k)}(y) : y \in E \setminus \{ x\}\})$. There exists iid random variables $(Q_j(x) : j \geq 1)$ taking values in $[0,2\rho(r)]$ such that $$1-O(e^{-c \beta}) \leq \frac{{\mathbf{E}}[ q_{i(j)} (x) | {\mathcal {G}}(j,x)]}{Q_j(x)} \leq 1+O(e^{-c\beta})$$ and $$1-O(e^{-c \beta}) \leq \frac{{\overline}{p}_{r,R}(x)}{{\mathbf{E}}Q_j(x)} \leq 1+ O(e^{-c \beta})$$ for all $n$ large enough. Furthermore, the families $\{ (Q_j(x) : j \geq 1) : x \in E\}$ are independent.
Define stopping times $$\begin{aligned}
\sigma_{k0}(E) &= \min\{ t \geq \sigma_k(x) : d(X(t),E) \geq 2R_n^\gamma\},\\
\tau_{k1}(E) &= \min\{ t \geq \sigma_{k0}(x) : d(X(t),E) \leq R_n^\gamma\}\end{aligned}$$ For $j \geq 1$, inductively set $$\begin{aligned}
\sigma_{k j }(E) &= \min\{ t \geq \tau_{kj}(E) : d(X(t),E) \geq 2R_n^\gamma \},\\
\tau_{k(j+1)}(E) &= \min\{ t \geq \sigma_{kj}(E) : d(X(t),E) \leq R_n^\gamma\}.\end{aligned}$$ Note that $\sigma_{kj}(E) - \tau_{kj}(E) \geq R_n^\gamma$. Thus, for $j_\beta = T_\beta^U / R_n^\gamma$ we have that $\tau_{kj_\beta}(E) \geq \sigma_k(x) + T_\beta^U$. Let $F_k(x) = \{ X(t) \in E(r) \text{ for } t \in [\sigma_k(x), \sigma_k(x) + T_\beta^U]\}$. Let $x_{kj}$ be the element in $E$ such that $d(X(\tau_{kj}(E)), x_{kj}) \leq R_n^\gamma$. Observe $$\begin{aligned}
&{\mathbf{P}}_{X(\tau_{kj}(E))}[ X(t) \in E(r) \text{ for } t \in [ \tau_{kj}(E), \sigma_{kj}(E)] \big| x_{kj}] \\
\leq& C\max_{d(y,x_{kj}) = R_n^\gamma} g(y, B(x_{kj}, r);G_n).\end{aligned}$$ Uniform local transience also yields $${\mathbf{P}}_{X(\sigma_k(x))}[ X(t) \in E(r) \text{ for } t \in [\sigma_k(x), \tau_{k0}(E)]] \leq C \rho(R,r) \leq \delta/2,$$ provided $R > r$ is large enough. A union bound thus gives $$\begin{aligned}
{\mathbf{P}}_{X(\tau_k(x))}[ F_k(x) ]& \leq \max_{z} \max_{d(y,z) = R_n^\gamma}g(y, B(z, r);G_n)\frac{T_\beta^U}{R_n^\gamma} + \delta/2 \notag\\
&\leq \delta/2 + o(1)\leq \delta \label{eqn::hit_net_bound},
\end{aligned}$$ as $n \to \infty$ by part of Assumption \[assump::far\_away\_harnack\]. Note that if $x_1,\ldots,x_\ell \in E$ and $j(1),\ldots,j(k)$ are such that $\tau_{j(k)}(x_k) \leq \tau_{j(k+1)}(x_{k+1})$ then we have $$\begin{aligned}
&{\mathbf{P}}[ F_{j(1)}(x_1),\ldots, F_{j(\ell)}(x_\ell)]
= {\mathbf{E}}[ {\mathbf{P}}_{X(\tau_{j(\ell))}(x_\ell)}[ F_{j(\ell)}(x_\ell)] {\mathbf{1}}_{F_{j(1)}(x_1)} \cdots {\mathbf{1}}_{F_{j(\ell-1)}(x_{\ell-1})}]\\
\leq& \delta {\mathbf{P}}[ F_{j(1)}(x_1),\ldots, F_{j(\ell-1)}(x_{\ell-1})]
\leq \dots \leq \delta^{\ell}.\end{aligned}$$ This can of course be repeated with any subset of the above events which implies the stochastic domination claim. It easily now follows from Cramer’s theorem that $${\mathbf{P}}\left[ I((1-\delta)k,x) \geq k \right] \leq 2 \exp(-C \delta^2 k).$$
For the second part of the lemma, we just need to get a bound on $\mu_x(z) / \pi(z)$ where $\mu_x$ is the law of random walk started at $x$ conditioned not to get within distance $r$ of $E$ by, say, time $T_{\beta/2}^U$. This can be done in exactly the same way as in the proof of Lemma \[lem::radon\_control\]. Indeed, the term $|E| \rho(s,r)$ in the statement of that lemma comes from a bound on the probability that random walk at distance $s$ from $E$ hits $E$ in time $T_\alpha^U$. In the situation of this lemma, the role of $s$ is replaced by $R_n^\gamma$ and we can use the scheme developed above to estimate the error contributed by this term by $O(\delta)$ provided $n$ is sufficiently large.
\[lem::prob\_success\_bound\] If $(G_n)$ satisfies part of Assumption \[assump::far\_away\_harnack\] then for each $r,\delta > 0$ there exists $R_0 > r$ such that $R \geq R_0$ implies $$\begin{aligned}
\label{eqn::q_concentration_harnack}
&{\mathbf{P}}\left[ \prod_{j=1}^{k} (1-q_j(x)) \geq (1-(1+\delta) {\overline}{p}_{r,R}(x))^{k(1+\delta)}\right]\\
& \ \ \ \ \ \ \ \ \ \leq C\big[ \exp\left( - C \delta^2 {\overline}{p}_{r,R}(x) k / \rho(r) \right) + \exp(- C \delta^2 k) \big]. \notag\end{aligned}$$ for all $n$ large enough.
The uniform Harnack inequality implies that $q_j(x) \leq 2C \rho(r)$ where $C = C(R/r)$ is the constant from the statement of part of Assumption \[assump::far\_away\_harnack\]. Let $F_j = \{\tau_j(x) - \sigma_{j-1}(x) \leq T_\beta^U\}$. Arguing as in the previous lemma and invoking uniform local transience, there exists iid random variables ${\widetilde}{F}_j(x)$ with ${\mathbf{P}}[{\widetilde}{F}_j(x) = 1] = \delta = 1-{\mathbf{P}}[{\widetilde}{F}_j(x) = 0]$ that stochastically dominate $({\mathbf{1}}_{F_j(x)} : j)$ provided $R$ is sufficiently large. We let $\iota(j)$ be the $j$th smallest index $i$ such that $F_i(x)$ occurs. The lemma now follows from an argument similar to that of Lemma \[lem::prob\_success\_concen\]. Indeed, we can stochastically dominate $q_{\iota(j)}(x)$ from below by iid random variables $L_j$ with ${\mathbf{E}}L_j \geq (1-\delta) {\overline}{p}_{r,R}(x)$ and $L_j \leq 10C \rho(r)$. By Cramer’s theorem, $$\begin{aligned}
{\mathbf{P}}\left[ \prod_{j=1}^{k} (1-L_j) \geq (1-(1+\delta) {\overline}{p}_{r,R}(x))^{k}\right]
\leq C\exp\left( - C \delta^2 {\overline}{p}_{r,R}(x) k / \rho(r) \right).\end{aligned}$$ The lemma now follows since, again by Cramer’s theorem, $${\mathbf{P}}[ \iota((1-\delta)k) \geq k] \leq C \exp(-C \delta^2 k).$$
Proof of Theorem \[thm::threshold\]
-----------------------------------
Let $\delta > 0$ be arbitrary and assume that $R > r,n_0,\epsilon$ have been chosen so that for all $n \geq n_0$ we have $$1-\delta \leq \frac{T_{{\rm cov}}(G_n)}{C_n^\epsilon} \leq 1+\delta.$$ We may assume without loss of generality that $d_k^\epsilon > 0$ for all relevant $k$ and, in particular, that $|H_{n,k}^\epsilon|^{-\delta} \to 0$ for every $k$. Indeed, Lemma \[lem::hitting\_estimate\] implies that $T_{{\rm hit}}(G_n) = \Theta(|V_n|)$, consequently if $\log| H_{n,k}^\epsilon| \to 0$ as $n \to \infty$ then $T_{{\rm cov}}(H_{n,k}^\epsilon)$ is negligible in comparison to $T_{{\rm cov}}(G_n)$. If $(G_n)$ satisfies Assumption \[assump::far\_away\_harnack\] Part we take $R_n^\gamma$ as given there. Otherwise, we take $R_n^\gamma = \max\{ R > 0 : \max_{x \in V_n} |B(x,R)| \leq |V_n|^{\gamma}\}$.
Let ${\mathcal {R}}(t)$ denote the range of random walk at time $t$ and ${\mathcal {L}}(t) = V \setminus {\mathcal {R}}(t)$. Letting $$M = \begin{cases}
20 \Delta_0 \sup_{n} {\overline}{\Delta}^R(G_n)/ (\delta \epsilon d^\epsilon) \text{ if } \sup_{n} {\overline}{\Delta}(G_n) < \infty,\\
20 \Delta_0 / (\delta \epsilon d^\epsilon) \text{ otherwise}
\end{cases}$$ and $$\begin{aligned}
{\mathcal {T}}_0 &= \min\left\{T \geq 0 : \max_{x} |{\mathcal {L}}(t) \cap B(x,R_n^\gamma)| \leq M \right\},\end{aligned}$$ we have that ${\mathbf{P}}[ {\mathcal {T}}_0 > \tfrac{1+5\delta}{2} T_{{\rm cov}}(G_n)] = o(1)$ provided $\gamma$ is sufficiently small, $R$ is so large that $\delta_{R,m} \leq 1$, $d^\epsilon = \min\{ d_k^\epsilon : d_k^\epsilon > 0\}$, and $m =20/d^\epsilon$. Furthermore, letting $${\mathcal {T}}_1 = \min\left\{T \geq 0 : |{\mathcal {L}}(t) \cap H_{n,k}^\epsilon| \leq |H_{n,k}^\epsilon|^{1/2-\delta} \text{ for all } k \right\}$$ we have that ${\mathbf{P}}[ {\mathcal {T}}_1 > \tfrac{1+5\delta}{2} T_{{\rm cov}}(G_n)] = o(1)$.
First suppose that $(G_n)$ has uniformly bounded maximal degree. Fix $R > r$ and let $E$ be an $R$-net of $H_{n,k}^\epsilon$. Fix $x \in H_{n,k}^\epsilon$ and suppose that $x_1,\ldots,x_\ell \in B(x,R_n^\gamma) \cap H_{n,k}^\epsilon \cap E$ are distinct. Lemma \[lem::corr\_decay\_upper\_bound\] gives us $${\mathbf{P}}[ x_1,\ldots, x_\ell \in {\mathcal {L}}( (1+\delta)/2;G_n)] \leq (1+\delta_{R,\ell}) |V_n|^{-(1+\delta)\ell d_k^\epsilon/2 + \delta_{R,\ell}}.$$ Consequently, a union bound yields $$\begin{aligned}
&{\mathbf{P}}[ |{\mathcal {L}}( (1+\delta)/2;G_n) \cap B(x,R_n^\gamma) \cap E| \geq \ell]\\
\leq& (1+\delta_{R,\ell}) |B(x,R_n^\gamma)|^\ell |V_n|^{-(1+\delta)\ell d_k^\epsilon/2 + \delta_{R,\ell}}\\
\leq& (1+\delta_{R,\ell}) |V_n|^{(\gamma - (1+\delta)d_k^\epsilon/2)\ell + \delta_{R,\ell}}.\end{aligned}$$ Hence choosing $\gamma \leq d^\epsilon/4$ the above is $O(|V_n|^{-3})$. Since the number of disjoint $R$-nets necessary to cover $H_{n,k}^\epsilon$ is at most ${\overline}{\Delta}^R(G_n)$, the result now follows from a union bound. In the case of unbounded maximal degree we can skip the step of subdividing the $H_{n,k}^\epsilon$ into $R$-nets since in this case $\delta_{1,m} \to 0$, otherwise the proof is the same. The second claim is immediate from Markov’s inequality and Lemma \[lem::corr\_decay\_upper\_bound\].
Let $N(x,T)$ be the number of excursions from $\partial B(x,r)$ to $\partial B(x,R)$ that have occurred by time $T$.
Let $$\begin{aligned}
{\mathcal {T}}_2 &= \min\left\{ T \geq 0: \max_{x \in H_{n,k}^\epsilon} \prod_{k=1}^{N(x,T)} (1-q_j(x)) \leq |H_{n,k}^\epsilon|^{-1/2-\delta} \text{ for all } k \right\}.\end{aligned}$$ and set $$\label{eqn::stopping_time}
{\mathcal {T}}= {\mathcal {T}}_0 \vee {\mathcal {T}}_1 \vee {\mathcal {T}}_2 \vee \left( \frac{1+5\delta}{2}\right) T_{{\rm cov}}(G).$$ Let $k_0(n)$ be a sequence so that $\liminf_{n \to \infty} d_{k(n)}^\epsilon \geq \delta_0 > 0$. For $x \in H_{n,k_0(n)}^\epsilon$ we have $$\left( \frac{1+3\delta}{2} \right) C_{n,k_0(n)}^\epsilon \geq \left(\frac{1+3\delta+O(\epsilon)}{2}\right) \frac{\delta_0 T_{r,R}(x) \log|V_n|}{4\rho(r)}$$ for all $n$ large enough. Thus letting $M_{n,k_0(n)}^\epsilon(x) =(1+3\delta)/2 \cdot C_n^\epsilon(x) / T_{r,R}(x)$, we have $$M_{n,k_0(n)}^\epsilon(x) \geq \left( \frac{1+3\delta + O(\epsilon)}{2} \right) \frac{\delta_0 \log |V_n|}{4 \rho(r)}.$$ Now, $$\begin{aligned}
&{\mathbf{P}}[ (1-\delta) T_{r,R}(x) M_{n,k_0(n)}^\epsilon(x) \leq \tau_{M_{n,k_0(n)}^\epsilon}(x) \leq (1+\delta) T_{r,R}(x) M_{n,k_0(n)}^\epsilon(x)]\\
\geq& 1- C\exp\left(- \frac{C \delta_0 \delta^2}{\rho(r)}\log|V_n|\right)
\geq 1 - O(|V|^{-100}),\end{aligned}$$ provided we choose $r$ large enough. Choosing $R > r$ sufficiently large, Lemma \[lem::prob\_success\_bound\] gives us $$\begin{aligned}
{\mathbf{P}}\left[ \prod_{j=1}^{M_{n,k_0(n)}^\epsilon(x)}(1-q_j(x)) \geq |H_{n,k_0(n)}^\epsilon|^{-1/2-\delta}\right]
\leq O(|V|^{-100}).\end{aligned}$$ Combining everything, $$\label{eqn::stopping_time_bound}
{\mathbf{P}}\left[ {\mathcal {T}}\neq \left(\frac{1+5\delta}{2}\right) T_{{\rm cov}}(G)\right] = o(1) \text{ as } n \to \infty.$$ Let $\mu$ be the probability on ${\mathcal {X}}(G)$ given by first sampling ${\mathcal {R}}\subseteq G$ according to $\mu_0$, the measure on subsets of $V$ given by running $X$ to time $(1+5\delta)/2 \cdot T_{{\rm cov}}(G)$, then sampling $f|{\mathcal {R}}$ by marking with iid fair coins and $f|V \setminus {\mathcal {R}}\equiv 0$. Define ${\widetilde}{\mu}$ similarly except by sampling ${\mathcal {R}}\subseteq G$ according to ${\widetilde}{\mu}_0$, the measure given by running $X$ up to time ${\mathcal {T}}$ rather than $(1+5\delta)/2 \cdot T_{{\rm cov}}(G)$. As a consequence of , $$\|\mu - {\widetilde}{\mu}\|_{TV} \leq {\mathbf{P}}\left[ {\mathcal {T}}\neq \left(\frac{1+5\delta}{2}\right) T_{{\rm cov}}(G)\right] = o(1) \text{ as } n \to \infty.$$
Suppose we have two independent random walks $X,X'$ on $G$, each with stationary initial distribution, and let ${\mathcal {T}},{\mathcal {T}}'$ be stopping times for each as in . Let ${\mathcal {R}},{\mathcal {R}}'$ be their ranges at time ${\mathcal {T}},{\mathcal {T}}'$, respectively, and ${\mathcal {L}}= V \setminus {\mathcal {R}}$, ${\mathcal {L}}' = V \setminus {\mathcal {R}}'$. Let $q_j'(x)$ be the quantity analogous to $q_j(x)$ for $X'$ and ${\mathcal {G}}= \sigma(q_j'(x) : j \geq 1)$. The previous lemma implies that we can divide ${\mathcal {L}}$ into $M$ disjoint sets $E_1,\ldots,E_M$ such that if $x,y \in E_\ell$ with $x \neq y$ then $d(x,y) \geq R_n^\gamma > R$. Consequently, letting ${\mathcal {G}}(E_\ell) = \otimes_{x \in E_\ell} {\mathcal {G}}(x)$ we have $$\begin{aligned}
{\mathbf{E}}[\exp(\zeta |{\mathcal {L}}\cap {\mathcal {L}}' \cap E_\ell|) | {\mathcal {G}}(E_\ell)]
\leq& \prod_{x \in E_\ell} \left(1 + e^{\zeta}\left(\prod_{j=1}^{N(x,{\mathcal {T}}')} (1-q_j'(x))\right)\right)\\
\leq& \exp\left(e^{\zeta} \sum_k |H_{n,k}^\epsilon|^{-\delta}\right).\end{aligned}$$ Since $E_1,\ldots,E_M$ cover ${\mathcal {L}}$, it follows from Hölder’s inequality that $$\begin{aligned}
{\mathbf{E}}\exp(\zeta |{\mathcal {L}}\cap {\mathcal {L}}'|)
&\leq \bigg[ \exp(e^{\zeta M} \sum_k |H_{n,k}^\epsilon|^{-\delta}) \bigg]^{1/M} \label{eqn::holder_bound}\\
&\leq 1 + 2\frac{\exp(\zeta M)}{M} \sum_k |H_{n,k}^{\epsilon}|^{-\delta} \notag.\end{aligned}$$
Let $${\mathcal {T}}_2 = \min\left\{T \geq 0 : \max_k \max_{x \in H_{n,k}^\epsilon} \frac{(1+2\delta)\log |H_{n,k}^\epsilon|}{2 N(x,T) {\overline}{p}_{r,R}(x)} \leq 1 \right\}$$ and $$\label{eqn::stopping_time_sep}
{\mathcal {T}}= {\mathcal {T}}_0 \vee {\mathcal {T}}_1 \vee {\mathcal {T}}_2 \vee \left( \frac{1+5\delta}{2}\right) T_{{\rm cov}}(G).$$ It follows from Lemmas \[lem::excursion\_concen\] and \[lem::cover\_time\] and the definition of $H_{n,k}^\epsilon$ that $$\label{eqn::stopping_time_bound_sep}
{\mathbf{P}}\left[ {\mathcal {T}}\neq \left(\frac{1+5\delta}{2}\right) T_{{\rm cov}}(G)\right] = o(1) \text{ as } n \to \infty.$$ Let $\mu$ be the probability on ${\mathcal {X}}(G)$ given by first sampling ${\mathcal {R}}\subseteq G$ according to $\mu_0$, the measure on subsets of $V$ given by running $X$ to time $(1+5\delta)/2 \cdot T_{{\rm cov}}(G)$, then sampling $f|{\mathcal {R}}$ by marking with iid fair coins and $f|V \setminus {\mathcal {R}}\equiv 0$. Define ${\widetilde}{\mu}$ similarly except by sampling ${\mathcal {R}}\subseteq G$ according to ${\widetilde}{\mu}_0$, the measure given by running $X$ up to time ${\mathcal {T}}$ rather than $(1+5\delta)/2 \cdot T_{{\rm cov}}(G)$. As a consequence of , $$\|\mu - {\widetilde}{\mu}\|_{TV} \leq {\mathbf{P}}\left[ {\mathcal {T}}\neq \left(\frac{1+5\delta}{2}\right) T_{{\rm cov}}(G)\right] = o(1) \text{ as } n \to \infty.$$
Suppose we have two independent random walks $X,X'$ on $G$, each with stationary initial distribution, and let ${\mathcal {T}},{\mathcal {T}}'$ be stopping times for each as in . Using the same notation as the previous proof, by the definition of ${\mathcal {T}}_2'$ we have $$\begin{aligned}
{\mathbf{E}}\big[ {\mathbf{E}}[\exp(\zeta |{\mathcal {L}}\cap {\mathcal {L}}' \cap E_\ell|) | {\mathcal {G}}(E_\ell)] \big]
\leq& {\mathbf{E}}\prod_{x \in E_\ell} \left(1 + e^{\zeta}\left(\prod_{j=1}^{N(x,{\mathcal {T}}')} (1-q_j'(x))\right)\right) \notag\\
\leq& {\mathbf{E}}\prod_{x \in E_\ell} \left(1 + e^{\zeta}\left(\prod_{j=1}^{N(x)} (1-q_j'(x))\right)\right) \label{eqn::exp_bound_sep}\end{aligned}$$ where $N(x) = (1+2\delta) \log |H_{n,k}^\epsilon| / 2 {\overline}{p}_{r,R}(x)$ and $k$ is such that $x \in H_{n,k}^\epsilon$. Let $${\widetilde}{N}(x) = (1-\delta) N(x) \geq \frac{(1+\delta/2) \log |H_{n,k}^\epsilon|}{2 {\overline}{p}_{r,R}(x)}.$$ Observe that is bounded by $$\begin{aligned}
& {\mathbf{E}}\prod_{x \in E_\ell} \left(1 + e^{\zeta}\left(\prod_{j=1}^{{\widetilde}{N}(x)} (1-q_{i(j)}'(x)) + {\mathbf{1}}_{\{I({\widetilde}{N}(x)) > N(x)\}}\right)\right)\end{aligned}$$ As $E_\ell$ satisfies the hypotheses of Lemma \[lem::iid\_replacement\], this is in turn bounded by $$\begin{aligned}
& {\mathbf{E}}\prod_{x \in E_\ell} \left(1 + e^{\zeta}\left(\prod_{j=1}^{{\widetilde}{N}(x)} (1-(1-\delta/4)Q_j'(x))\right) + O(|V|^{-100})\right)\\
\leq& \exp\left(e^{\zeta} \sum_k |H_{n,k}^\epsilon|^{-\delta}\right).\end{aligned}$$ The theorem now follows from Hölder’s inequality, as in the previous proof.
The Lamplighter
---------------
This is proved by making several small modifications to the proof of Theorem \[thm::threshold\]. Namely, rather than considering the range of the random walk run up to time ${\mathcal {T}}$ as in either or , one considers the range ${\mathcal {R}}(x)$ run up to time ${\mathcal {T}}$, then conditioned on hitting a given point $x$. Exactly the same argument shows that the total variation distance of the law ${\widetilde}{\mu}_x$ on markings ${\mathcal {X}}(G)$ induced by iid coin flips on ${\widetilde}{{\mathcal {R}}}(x)$ from the uniform measure on ${\mathcal {X}}(G)$ is $o(1)$. This implies that the law $\mu_x$ on markings of ${\mathcal {X}}(G)$ given by iid coin flips on the range ${\mathcal {R}}(x)$ run up to time $\frac{1+\epsilon}{2} T_{{\rm cov}}(G_n)$, and conditioned to hit $x$, and the uniform measure is $o(1)$. At time $\frac{1+2\epsilon}{2} T_{{\rm cov}}(G_n)$, the random walk is well mixed conditional on its position at $\frac{1+\epsilon}{2} T_{{\rm cov}}(G_n)$, from which the result is clear.
Further Questions
=================
1. Theorem \[thm::threshold\] yields a wide class of examples where the threshold for indistinguishability is at $\tfrac{1}{2} T_{{\rm cov}}$, and ${\mathbf{Z}}_n^2$ is an example where the threshold is at $T_{{\rm cov}}$. Does there exists a sequence $(G_n)$ of vertex transitive graphs where the threshold is at $\alpha T_{{\rm cov}}(G_n)$ for $\alpha \in (1/2,1)$?
2. Our statistical test for uniformity is only valid for $\alpha > 1/2$. For $\alpha \leq 1/2$, the natural reference measure is iid markings conditioned on the number of zeros being on the order of $|V|^{1-\alpha}$. Can analogous results be proved in this setting?
3. Our definition of uniform local transience is given in terms of the Green’s function summed up to the *uniform mixing time*. Does it suffice to assume only the uniform decay of $$g(x,y;G) = \sum_{t=1}^T p^t(x,y;G)$$ where $T = T_{{\rm mix}}(G)$ or even $T = T_{{\rm rel}}(G)$?
4. The complete graph $K_n$ does not satisfy the hypotheses of Theorem \[thm::threshold\] yet the lamplighter walk on $K_n$ has a threshold at $\tfrac{1}{2} T_{{\rm cov}}(K_n)$. Is there a more general theorem allowing for a unified treatment of this case?
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The origin of cosmic magnetic fields is an unsolved problem and magnetogenesis could have occurred in the early Universe. We study the evolution of such primordial magnetic fields across the cosmological recombination epoch via 3D magnetohydrodynamic numerical simulations. We compute the effective or net heating rate of baryons due to decaying magnetic fields and its dependence on the magnetic field strength and spectral index. In the drag-dominated regime ($z \gtrsim 1500$), prior to recombination, we find no real heating is produced. Our simulations allow us to smoothly trace a new transition regime ($600 \lesssim z \lesssim 1500$), where magnetic energy decays, at first, into the kinetic energy of baryons. A turbulent velocity field is built up until it saturates, as the net heating rate rises from a low value at recombination to its peak towards the end of the transition regime. This is followed by a turbulent decay regime ($z \lesssim 600$) where magnetic energy dissipates via turbulent decay of both magnetic and velocity fields while net heating remains appreciable and declines slowly. Both the peak of the net heating rate and the onset of turbulent decay are delayed significantly beyond recombination, by up to 0.5 Myr (until $z\simeq 600-700$), for scale-invariant magnetic fields. We provide analytic approximations and present numerical results for a range of field strengths and spectral indices, illustrating the redshift-dependence of dissipation and net heating rates. These can be used to study cosmic microwave background constraints on primordial magnetic fields.'
author:
- |
Pranjal Trivedi$^{1,2}$[^1], Johannes Reppin$^{1}$[^2], Jens Chluba$^{3}$[^3] and Robi Banerjee$^{1}$\
$^{1}$Hamburger Sternwarte, University of Hamburg, Gojenbergsweg 112, 21029, Hamburg, Germany\
$^{2}$Department of Physics, Sri Venkateswara College, University of Delhi, 110021 New Delhi, India\
$^{3}$Jodrell Bank Centre for Astrophysics, Alan Turing Building, University of Manchester, Manchester M13 9PL, UK
bibliography:
- 'Lit.bib'
title: 'Magnetic heating across the cosmological recombination era: Results from 3D MHD simulations'
---
cosmology:theory, early Universe, magnetic fields, magnetohydrodynamics, turbulence, cosmic background radiation
Introduction {#sec:Intro}
============
Magnetic fields are observed ubiquitously and play an important role in the physics of galaxies out to large scales in the Universe [@Beck2015; @Han2017]. Microgauss strength magnetic fields have also been detected in high redshift galaxies [e.g. @Bernet2008; @Kronberg2008; @Mao2017] and across scales of several megaparsecs in clusters of galaxies [@Clarke2001; @Bonafede2010; @Feretti2012]. There is also indirect evidence from $\gamma$-ray observations suggesting a lower limit of $B \sim 10^{-16}$ G on intergalactic magnetic fields that can fill most of the cosmic volume [@Neronov2010; @Tavecchio2011; @Dolag2011; @Taylor2011; @Dermer2011; @Fermi2018].
These observations motivate a primordial origin, i.e., from the very early Universe, for the seed magnetic field responsible for these observed large-scale fields. Theoretical models have been proposed for the generation of such a primordial magnetic field (PMF) which generically involves breaking of conformal invariance of the electromagnetic action. Primordial magnetogenesis could have occurred at inflation [@Turner1988; @Ratra1992; @Gasperini1995; @Martin2008; @Subramanian2010; @Caprini2014; @Sharma2017] or at a cosmological phase transition [@Vachaspati1991; @Sigl1997; @Grasso2001; @Copi2008] However, a complete theory of magnetogenesis and the PMF’s subsequent evolution remains an important unresolved issue in cosmology [reviewed in @Kandus2011; @DurrerNeronov2013; @Subramanian2016].
The temperature and polarization anisotropies of the cosmic microwave background (CMB) are a sensitive probe of PMFs since magnetic fields induce metric perturbations as well as fluid perturbations via the Lorentz force [@Subramanian2006; @Finelli2008; @Paoletti2009; @Shaw2010PMF]. Upper limits have been placed on the comoving PMF at megaparsec scales of order or just below $B\simeq1$ nG derived from comparison to *Planck* observations, based on the CMB power spectrum [@Chluba2015PMF; @Planck2015PMF; @Zucca2017] and CMB non-Gaussianity [@Trivedi2012; @Trivedi2014; @Shiraishi2014]. Similar PMF upper limits exist from a range of other probes of magnetic fields including Faraday rotation of radio sources [@Pshirkov2016] and changes to the cosmic ionization history [@Sethi2005; @Kunze2014; @Kunze2015; @Chluba2015PMF; @Planck2015PMF].
However, a significant gap of several orders of magnitude between the upper limits and the inferred lower limits on large-scale primordial magnetic fields remains. Also, many of the currently applied approximations for magnetic heating and dissipation that are used to incorporate the effect on the cosmic ionization history, remain rather unsatisfactory [e.g., see discussions in @Chluba2015PMF; @Planck2015PMF]. Thus, an improved understanding of PMF signatures is needed to clarify their current amplitude, evolution and origin. Motivated by the outstanding issues, we investigate the amount of decay and dissipation of a PMF and the resultant heating produced across the cosmological recombination era [@Zeldovich68; @Peebles68; @Sunyaev2009], one of the clearly anticipated phases in the evolution of PMFs.
PMFs are expected to evolve in a qualitatively distinct fashion across different cosmological epochs. Damping of PMFs can occur via neutrino- or photon-sourced fluid viscosity, via the development of magnetohydrodynamic (MHD) turbulent decay or processes such as ambipolar diffusion [@Jedamzik1998; @Subramanian1998; @Banerjee2004].
Prior to cosmological recombination, there exists a photon drag-dominated viscous regime for the baryon-photon fluid. Close to the recombination epoch, once the radiation drag has started to drop abruptly, the PMF can source significant fluid motions . This induces turbulence in the baryon fluid which ultimately dissipates some fraction of the energy contained in the PMF via a turbulent cascade. The energy dissipated from the PMF can heat electrons and baryons to alter the thermal history [@Sethi2005; @Kunze2014; @Kunze2015; @Chluba2015PMF]. We also note that a phase of MHD turbulent decay of the PMF can also occur subsequent to earlier phase transitions (eg. electroweak [@Kahniashvili2010]).
In this work, we report MHD simulations that describe the transfer and subsequent dissipation of magnetic energy across the epoch of recombination. Previous work has shown that turbulent decay will be important for PMF evolution after recombination [@Sethi2005; @Chluba2015PMF]. Here, using full 3D simulations, we are able to quantify both the sourcing of the velocity field, the development of turbulence and the decay of the PMF. Our simulations allow us to smoothly follow the PMF and baryon velocity field across three distinct regimes: the photon drag-dominated regime, a new transition regime around recombination and the post-recombination regime dominated by turbulent decay. In particular, we can trace the relatively rapid evolution of baryon velocities, magnetic fields and the heating rate (of baryons and electrons) due to magnetic decay during the transition regime. This enables us to connect the evolution of magnetic fields in the pre- and post-recombination epochs. We can then quantify the PMF decay and heating rate of the baryon fluid as a function of PMF parameters: amplitude and spectral index. The timescale for turbulence to develop after recombination also depends on the PMF parameters and this affects the epoch of peak heating caused by the PMF. The simulations and fits presented in this work refine our understanding of the evolution of the PMF and offer a way to potentially place improved constraints on the nature of primordial magnetic fields.
Quantifying the heating effect of PMF dissipation via decaying turbulence can be related to two further effects in the CMB caused by fluid heating. Heated electrons result in delayed recombination which in turn modifies the CMB power spectra [@Sethi2005; @Chluba2015PMF; @Kunze2015]. Secondly, the additional energy in electrons up-scatters photons to higher energies, producing a distortion in the CMB spectrum [@Jedamzik2000; @Chluba2011therm]. These effects are important when placing improved constraints on the PMF and its origin. In addition, the effects of PMF dissipation and heating are also relevant for assessing the role of the PMF in structure formation [@Wasserman1978; @Kim1996; @Subramanian1998; @Sethi2005], 21-cm signals [@Sethi2009; @Schleicher2009; @Shiraishi2014_21cm], reionization signals [@Sethi2009; @Bowman2018] and primordial chemistry [@Schleicher2008; @Schleicher2008b], thus warranting careful treatments of the process. Our simulation outputs can also be used as initial conditions for the small-scale velocity field in structure formation simulations addressing the effects of PMFs.
The subsequent sections are organized in the following manner: In Section \[sec:av\_B\] we give a brief description of the magnetic field properties in the early Universe and present an overview of the their evolution. In Section \[sec:analytics\] we revisit the analytical treatment of the different evolutionary stages, specifically for the free-streaming damping epoch and the turbulent decay after recombination. The MHD simulations are described in Section \[sec:simulations\], where we explain the numerical setup used to integrate the governing MHD equations and the techniques used for modeling the epoch of recombination. In Section \[sec:Results\] we present the results of the numerical simulations for magnetic field and velocity amplitudes, magnetic and kinetic power spectra and dissipation and net heating rates. Results are first described for the fiducial case with an initially near scale-invariant magnetic power spectrum. Subsequently, we present and analyze the trends in these physical results when varying the PMF parameters of the field amplitude and spectral index. We conclude the results section by providing a semi-analytical representation of the dissipation rate sourced by the decay of PMF in Section \[sec:semi-analytics\]. A discussion of our results and caveats follows in Section \[sec:discussion\] and conclusions are presented in Section \[sec:conclusion\]
Primordial stochastic magnetic field {#sec:av_B}
====================================
We start by defining how the root mean square (rms) value of the primordial stochastic magnetic field (PMF) is related to the initial power spectrum at any moment. We assume that the initial power spectrum is given by $P_B(k)= A k^{n_B}$, $n_B>-3$. Then the r.m.s.[*comoving*]{} magnetic field strength over some Gaussian filtering scale $\lambda_{\rm f}$ is related to $$\begin{aligned}
\label{eq:def_B_av}
B^2_{\rm f}=a^4 \mathcal{B}^2_{\rm f}=\int \frac{\id ^3k}{(2\pi)^2} P_B(k) \,\expf{-\lambda^2_{\rm f} k^2}=\frac{A\,\Gamma\left(\frac{n_B+3}{2}\right)}{4\pi^2 \lambda_{\rm f}^{n_B+3}},\end{aligned}$$ where $\mathcal{B}^2_{\rm f}\propto (1+z)^4$ denotes the mean square field strength in proper coordinates. Equation is simply a definition without any physical input at this point. For example, $\lambda_{\rm f}=1\,\Mpc$ is frequently used to define $B_\lambda$ at large cosmological scales [e.g., @Planck2015PMF]. One simple way to capture the evolution and damping of the rms magnetic field as a function of different epochs (e.g., for free-streaming damping) can be found by setting the filtering scale to $\lambda_{\rm f}=\sqrt{2}/k_{\rm d}(z)$, where $k_{\rm d}$ is a specified damping scale that depends on time. This is equivalent to saying that the initial power spectrum amplitude in Fourier space is modulated by a factor $\simeq \expf{-2k^2/k_{\rm d}^2(z)}$. We thus have $$\begin{aligned}
\label{eq:barB}
\bar{B}^2=\frac{A\,\Gamma\left(\frac{n_B+3}{2}\right)}{4\pi^2}\frac{k_{\rm d}^{n_B+3} (z)}{2^{(n_B+3)/2}}.\end{aligned}$$ In this way, the rms magnetic field strength is determined by the evolution of $k_{\rm d}(z)$. For near scale-invariant fields (e.g., $n_B=-2.9$), most of the energy is stored at the largest scales and even a significant evolution of $k_{\rm d}$ does not affect $\bar{B}^2$ very much. In contrast, for very blue initial PMF power spectra, most of the energy density is present at small scales and rapid loss is expected even if $k_{\rm d}(z)$ only evolves slowly. This simplified picture also provides a reasonable description of the magnetic field evolution in the turbulent decay phase, even if no exponential cut-off is expected at small scales, $k>k_{\rm d}(z)$ [@Jedamzik1998; @Banerjee2004].
We will discuss below the evolution of the rms magnetic field strength across the recombination era in more detail. It is expected that in the phase well after recombination but before reionization ($10\lesssim z \lesssim 10^3$) the magnetic field energy density changes relatively slowly [@Banerjee2004]. In this regime, [*turbulent decay*]{} of magnetic fields leads to a logarithmic evolution of the magnetic field strength with redshift. In particular for near scale-invariant initial power spectra, this implies a close to constant comoving magnetic field energy density, $\rho_{B}= a^{-4} \bar{B}^2/8\pi$.
For convenience, in our analytic treatment we will normalize all magnetic field power spectra with respect to the redshift at which turbulent decay is expected to begin in the post-recombination era, $z_{\rm td}\simeq 10^3$. We will determine $z_{\rm td}$ more precisely below, but with this definition we approximately have $$\begin{aligned}
\label{eq:barB_norm}
\bar{B}^2\approx B_0^2
\left(\frac{k_{\rm d} (z)}{k_{\rm td}}\right)^{n_B+3},\end{aligned}$$ with $B_0\equiv\bar{B}(z_{\rm td})$ and $k_{\rm td}=k_{\rm d} (z_{\rm td})$. We now discuss the pre-recombination heating, which is dominated by free-streaming damping at $z\gtrsim z_{\rm td}$, and the post-recombination heating separately, improving upon the previous treatments in the literature.
Analytical considerations in the different regimes {#sec:analytics}
==================================================
The damping of magnetic fields and heating of the medium occurs in alternating phases of turbulent and viscous regimes [@Jedamzik1998; @Banerjee2004]. In this section, we briefly recap and attempt to improve previous analytic approximations in the different eras. We start with the phase just before recombination and then discuss the turbulent decay regime in the post-recombination epoch ($z\lesssim 10^3$). These approximations are then compared with the numerical results in Sect. \[sec:Results\].
Free streaming damping {#sec:free-damp}
----------------------
At sufficiently small scales (below the photon mean free path) and also at later redshifts after recombination, photons stream freely relative to the baryons. In this case, the baryon Alfvén speed, $V_{\rm bA}$ (in units of the light speed), in a magnetized medium is given by $$\begin{aligned}
\label{def:Alfven_speed_etc}
V_{\rm bA}&=\frac{a^{-2}\bar{B}}{\sqrt{4\pi \rho_{\rm b}}}\equiv \sqrt{\frac{3}{2 R}}\,\sqrt{\frac{\rho_B}{\rho_\gamma}}
\approx\frac{\pot{3.8}{-4}}{\sqrt{R}}\left[\frac{\bar{B}}{1\,{\rm nG}}\right]
\nonumber\\
&
\approx \pot{1.5}{-5} (1+z)^{1/2}\left[\frac{\Omega_{\rm b} h^2}{0.022}\right]^{-1/2}\left[\frac{T_0}{2.726\,\Kel}\right]^{2}
\left[\frac{\bar{B}}{1\,{\rm nG}}\right]
\nonumber\\
R&=\frac{3\rho_{\rm b}}{4\rho_\gamma}\approx \frac{666}{1+z}\,\left[\frac{\Omega_{\rm b} h^2}{0.022}\right]\left[\frac{T_0}{2.726\,\Kel}\right]^{-4}
\nonumber\\
\rho_B&=\frac{a^{-4} \bar{B}^2}{8\pi}\approx \pot{9.5}{-8} \left[\frac{\bar{B}}{1\,{\rm nG}}\right]^2\,\rho_\gamma
\nonumber\\
\rho_\gamma&\approx 0.26 \,\eV\,\cm^{-3} \left[\frac{T_0}{2.726\,\Kel}\right]^{4} (1+z)^4,\end{aligned}$$ where $\bar{B}$ is the rms magnetic field strength, $R$ the baryon-loading of the fluid, $\rho_B$ and $\rho_\gamma$ the rms magnetic and photon energy densities respectively. The free-streaming photons exert a small drag force, $\vek{F}_{\rm D}\approx -\frac{4}{3} \Ne \sigT \rho_\gamma \vek{\varv}$, on the baryons, mediated by the occasional scatterings of photons with the free electrons. This leads to damping of magnetic fields at very small scales[^4], $\lambda\lesssim \lambda_{\rm fs}$. These interactions source small-scale photon perturbations, which gradually damp due to free-streaming mixing [e.g., see @Minoda2017 for recent study of generation of small-scale fluctuations by PMFs in clusters].
Estimates for the free-streaming scale {#sec:free-streaming-scale}
--------------------------------------
The free-streaming scale is approximately determined by the condition [@Subramanian2016] $$\begin{aligned}
\label{def:k_td}
\frac{1}{k_{\rm fs}^2}\approx \int \frac{V_{\rm bA}^2 R \id t}{\Ne \sigT a^2}\approx \pot{1.4}{-7}
\left[\frac{B_0}{1\,{\rm nG}}\right]^2 \int \left(\frac{k_{\rm fs}}{k_{\rm td}}\right)^{n_B+3}\!\frac{\id t}{\Ne \sigT a^2},\end{aligned}$$ where the baryon Alfvén speed, $V_{\rm bA}$, baryon loading factor of the fluid, $R$, and magnetic energy density, $\rho_B$, defined in Eq. give $V_{\rm bA}^2 R = 3\rho_B/ 2\rho_\gamma$. For convenience we introduce the integral $$\begin{aligned}
\label{eq:k_lambda_def}
\frac{1}{k^2_\lambda}=\int_0^t \frac{\id t'}{\Ne \sigT a^2}=\int^\infty_{z} \frac{\id z}{\Ne \sigT a H(z)},\end{aligned}$$ which can be computed numerically and only depends on the standard background cosmology. It is approximately given by (see Fig. \[fig:k\_values\]), $k_\lambda\approx \sqrt{8/45}\, \kD$, where the standard photon diffusion damping scale, $\kD$, is defined by [@Weinberg1971; @Kaiser1983] $$\begin{aligned}
\label{eq:k_D_gamma}
\frac{1}{k_{\rm D}^2}\approx \frac{8}{45}\int^\infty_z \frac{\id z}{\Ne \sigT a H}\left[\frac{1}{1+R}+\frac{15}{16}\frac{R^2}{(1+R)^2}\right],\end{aligned}$$ for shear viscosity $\eta_\gamma=\frac{16}{45}\frac{\rho_\gamma}{ \Ne \sigT}$, which includes the effect of polarization[^5]. We have $k_{\rm D}\approx \pot{4.1}{-6}(1+z)^{3/2}\,\Mpc^{-1}$ in radiation-dominated era, such that $k_\lambda\approx \pot{1.8}{-6}(1+z)^{3/2}\,\Mpc^{-1}$. In the matter-dominated regime, one has $k_\lambda\propto (1+z)^{5/4}$; however, in our computation we use the full numerical result, which captures the rapid changes of $k_{\rm D}$ and $k_\lambda$ around $z\simeq 10^3$ (Fig. \[fig:k\_values\]).
By separating variables in Eq. and solving for $k_{\rm fs}$ we obtain $$\begin{aligned}
\label{def:kfs_rad_II}
k_{\rm fs}&\approx k_{\rm td}\,\left( \pot{1.4}{-7}\,\frac{(n_B+5)}{2}
\left[\frac{B_0}{1\,{\rm nG}}\,\frac{k_{\rm td}}{k_\lambda}\right]^2\right)^{-\frac{1}{n_B+5}}
=k_{\rm td}\,\left[\frac{k_\lambda}{k_\lambda(z_{\rm td})}\right]^{\frac{2}{n_B+5}}
\nonumber\\
k_{\rm td}&=\sqrt{\frac{2}{n_B+5}}\,\frac{k_\lambda(z_{\rm td})}{\pot{3.8}{-4}}
\left[\frac{B_0}{1\,{\rm nG}}\right]^{-1},\end{aligned}$$ where we used $k_{\rm fs}\approx k_{\rm td}$ at $z_{\rm td}$. Usually, the time-dependence of $k_{\rm fs}/k_{\rm td}$ is omitted, such that the factor $\left(k_{\rm fs}/k_{\rm td}\right)^{n_B+3}$ in Eq. can taken outside the integral. This yields a slightly different solution $$\begin{aligned}
\label{def:kfs_rad_II_old}
k^*_{\rm fs}&\approx k^*_{\rm td}\,\left( \pot{1.4}{-7}
\left[\frac{B_0}{1\,{\rm nG}}\,\frac{k^*_{\rm td}}{k_\lambda}\right]^2\right)^{-\frac{1}{n_B+5}}
=k^*_{\rm td}\,\left[\frac{k_\lambda}{k_\lambda(z_{\rm td})}\right]^{\frac{2}{n_B+5}}
\nonumber\\
k^*_{\rm td}&=\frac{k_\lambda(z_{\rm td})}{\pot{3.8}{-4}}
\left[\frac{B_0}{1\,{\rm nG}}\right]^{-1},\end{aligned}$$ which does not include any dependence of $k_{\rm fs}$ and $k_{\rm td}$ on the PMF spectral index. To fix $k_{\rm fs}$ we still need to estimate $z_{\rm td}$. However, it is already clear that $z_{\rm td}\simeq 10^3$, thus, $k_\lambda(z_{\rm td})\simeq 0.01-0.1\,\Mpc^{-1}$, which puts us into the regime $k_{\rm td}\simeq 26-260\,\Mpc^{-1}$ for $n_{B}\simeq -3$ and $B_0\simeq 1\,{\rm nG}$. Due to resolution issues and to avoid shocks, in our simulations we are only able to treat fields with $B_0\lesssim 10^{-2}\,{\rm nG}$. Assuming $B_0= 10^{-2}\,{\rm nG}$, one thus expects $k_{\rm td}\simeq 2600-26000\,\Mpc^{-1}$. We will see that in fact the lower end of this range seems to be reproduced by the simulations (cf. Fig. \[fig:mag\_spectra\_k09\]).
![Illustration for redshift dependence of some of the relevant scales. Here, $\kD$ and $k_\lambda$ are defined by Eq. and , respectively. We also introduced $k_{\rm t}= 2\pi \Ne \sigT a/\sqrt{R}$ for convenience.[]{data-label="fig:k_values"}](./eps/k_values_nB.pdf){width="0.92\columnwidth"}
### Estimate for $z_{\rm td}$ {#sec:drag_analytic}
To determine the redshift $z_{\rm td}$ at which the turbulent decay resumes after recombination, we use the condition that the kinetic Reynolds number should be $R_{\rm e}\simeq 1$ at $k_{\rm fs}\approx k_{\rm td}$ and $z_{\rm td}$ and then $R_{\rm e} > 1$ afterwards. In the free streaming regime, we have $R_{\rm e}\approx V_{\rm bA}/\alpha_\gamma \ell_{\rm fs}\approx V_{\rm bA} R\, \ell_{\rm mfp}/\ell_{\rm fs}$, where the drag coefficient $\alpha_\gamma\approx \Ne \sigT / R$ and $\ell_{\rm mfp}=1/\Ne \sigT$ for the mean free path were used [@Banerjee2004; @Subramanian2016]. With $\ell_{\rm fs}=\lambda_{\rm fs} a = 2\pi a/k_{\rm fs}$, this yields[^6] the condition $$\begin{aligned}
\label{eq:def_td}
\int^\infty_{z_{\rm td}} \frac{\id z}{\Ne \sigT a H(z)}\approx \frac{2}{(n_B+5)}\,\left.\frac{R}{(2\pi\,\Ne \sigT a)^2}\right|_{\rm td},\end{aligned}$$ for the redshift $z_{\rm td}$. We also define $k_{\rm t}= 2\pi \Ne \sigT a/\sqrt{R}$, such that Eq. can be cast into the form $k_{\rm \lambda}(z_{\rm td})\approx \sqrt{(n_B+5)/2}\,k_{\rm t}(z_{\rm td})$.
In Fig. \[fig:k\_values\], we show a comparison of the different wave numbers relevant to the discussion. The solution of Eq. in our treatment here is independent of the magnetic field strength. For $n_B=-3$, we find $z_{\rm td} \approx 820$ and more generally $$\begin{aligned}
\label{eq:z_td_approx}
z_{\rm td}\approx 820\left[1-\pot{4.66}{-2} \,(n_B+3)+\pot{1.39}{-3}(n_B+3)^2\right]\end{aligned}$$ for $n_B\in[-3,3]$ to percent precision. This shows that in our treatment $z_{\rm td}$ drops gradually with increasing spectral index, $n_B$. We furthermore obtain $$\begin{aligned}
\label{eq:k_td_lambda}
k_\lambda(z_{\rm td})&\approx \pot{7.50}{-3}\,\Mpc^{-1}\left[1-0.295 \,(n_B+3)
\right.
\nonumber\\&\quad+\left.\pot{4.66}{-2}(n_B+3)^2-\pot{2.97}{-3}(n_B+3)^2\right],\end{aligned}$$ which determines $k_{\rm td}$ in Eq. . This shows a moderate dependence of $k_{\rm td}$ on $n_B$, giving $k_{\rm td}\approx 20\,\Mpc^{-1}\left[B_0/1\,{\rm nG}\right]^{-1}$ for $n_B\simeq -3$ to $k_{\rm td}\approx 3.3\,\Mpc^{-1}\left[B_0/1\,{\rm nG}\right]^{-1}$ for $n_B\simeq 2$ in our treatment. We will see below that both Eq. and Eq. do not capture the dependence of the turbulent regime on $B_0$ and $n_{B}$, in particular when increasing $n_{B}$ (see Sect. \[sec:var\_B0\] and Sect. \[sec:var\_nB\]).
Comparing our results to previous works, a value $k_{\rm td}\approx 200-300\,\Mpc^{-1}\left[B_0/1\,{\rm nG}\right]^{-1}$ is usually quoted [e.g., @Sethi2005; @Kunze2015; @Subramanian2016]. Also, the redshift for turbulent decay usually is set to $z_{\rm td}\simeq 1100$ [e.g., @Sethi2005; @Kunze2015], independent of the PMF spectral index or strength. Here, we find lower values in both cases. The main reason for the different value at $n_{B}\simeq -3$ is the factor of $2\pi$ in the conversion from $\lambda_{\rm fs}$ to $k_{\rm fs}$. This yields the modified condition, $k_{\rm \lambda}(z_{\rm td})\approx k_{\rm t}(z_{\rm td})/2\pi$, for $z_{\rm td}$, which due to the exponential dependence of $k_{\rm \lambda}$ and $k_{\rm t}$ on redshift (see Fig. \[fig:k\_values\]) results in $z_{\rm td}\simeq 1050$ instead. Overall, our simulations seem to suggest a scaling $k_{\rm td}\simeq 20\,\Mpc^{-1}\left[B_0/1\,{\rm nG}\right]^{-1}$ and $z_{\rm td}\simeq 800$ in agreement with our simple analytic estimate for $n_{B}=-3$.
### Magnetic field dissipation in the pre-recombination era {#sec:pic_pre_rec}
The wavenumber $k_{\rm td}$ is only relevant if we wish to compare the initial power spectrum amplitude, $A$, for different values of $B_0$ and $n_B$. By construction, we already ensured that at $z=z_{\rm td}$ we have $\bar{B}^2 \approx B_0^2$ for all $n_B$. However, for different $n_B$ the effective damping scale at $z_{\rm td}$ differs, so that also the respective values for $A$ at the initial time vary. Consequently, we can also write $$\begin{aligned}
\label{eq:B2_fs}
\bar{B}^2\approx B_0^2 \left(\frac{k_{\rm fs}(z)}{k_{\rm td}}\right)^{n_B+3}
=B_0^2 \left(\frac{k_{\lambda}(z)}{k_{\lambda}(z_{\rm td})}\right)^{2(n_B+3)/(n_B+5)}\end{aligned}$$ at $z\gtrsim z_{\rm td}$, where we used Eq. to eliminate $k_{\rm fs}(z)/k_{\rm td}$. We then have the magnetic dissipation rate[^7] at $z\gtrsim z_{\rm td}$ $$\begin{aligned}
\label{eq:QbarB_II}
\frac{\id (Q_{B}/\rho_\gamma)}{\id z}\approx \frac{2(n_B+3)}{n_B+5}\frac{\rho_B}{\rho_\gamma} \frac{\partial_z k_{\lambda}(z)}{k_{\lambda}(z)}
=\frac{(n_B+3)}{n_B+5}\, \frac{ k_{\lambda}^2(z) \,r_B(z)}{\Ne \sigT a H},\end{aligned}$$ where we used $-(1/2) k^2_\lambda \partial_z k^{-2}_\lambda=k^{-1}_\lambda\partial_z k_\lambda$ and the magnetic energy density fraction $r_B(z)=\rho_B(z)/\rho_\gamma(z)$.
In the radiation-dominated era, $\rho_B(z)\propto \rho_\gamma(z) \,k_\lambda^{2(n_B+3)/(n_B+5)}$ and $k_\lambda\approx \pot{1.8}{-6}(1+z)^{3/2}\,\Mpc^{-1}$ such that we find the redshift scaling $\id Q_{B}/\id z\propto \rho_\gamma(z) \,(1+z)^{2(n_B+2)/(n_B+5)}$. This is in excellent agreement with previous studies [@Banerjee2004; @Wagstaff2015], but differs slightly from @Kunze2014, who obtained $\id Q_{B}/\id z\propto \rho_\gamma(z) \,(1+z)^{(3 n_B+7)/2}$. The difference is because $k_{\rm fs}\propto \kD \propto (1+z)^{3/2}$ was used in their expression, which neglects the scaling of $k_{\rm fs}$ because of changes in $\rho_B$. For blue magnetic field power spectra, this overestimates the free-streaming heating, as also pointed out by @Wagstaff2015.
To compare with our numerical simulations, it is furthermore useful to give the scaling in the matter-dominated regime. Then $k_\lambda \propto (1+z)^{5/4}$ and $H\propto (1+z)^{3/2}$, such that one has $\id Q_{B}/\id z\propto \rho_\gamma(z) \,(1+z)^{(3n_B+5)/2(n_B+5)}$. This shows a slightly slower increase with redshift than for the radiation-dominated case. Our simulations reveal a scaling trend for magnetic dissipation which is similar and this is discussed in Sect. \[sec:semi-analytics\] with reference to Fig. \[fig:heat\_rate\_comp\_index\].
### Creation of spectral distortions in the drag regime {#sec:pic_pre_rec_dist}
In the drag-dominated phase, the dissipated magnetic field energy flows directly to the CMB photons, as these provide the viscosity in the free-streaming era. This causes a small $y$-distortion [@Jedamzik2000; @Sethi2005; @Wagstaff2015], which physically is due to photon mixing [@Chluba2012], and also sources CMB temperature anisotropies at ultra-small scales ($k \gtrsim 20\,\Mpc^{-1}\,[B_0/1\,{\rm nG}]^{-1}$). Since we are in the free-streaming regime, the photon mixing process is [*incomplete*]{} so that only a (small) faction of the magnetic field energy really creates a distortion. This is in stark contrast to the damping of primordial temperature fluctuation that erases all small-scale fluctuation due to photon diffusion, which allows $\simeq 1/3$ of the acoustic wave energy to create an all-sky average distortion [@Chluba2012].
We also highlight that a small $y$-distortion has no significant effect on the baryon temperature and the recombination history as long as it remains below the [*COBE/FIRAS*]{} limit [@Chluba2008c]. Thus, this part of the PMF dissipation can only be constrained by measuring CMB spectral distortions and ultra small-scale CMB anisotropies, in spite of opposing recent claims [@Kunze2017]. Simple estimates, assuming that all the energy creates distortions, suggest a Compton $y$-parameter, $y\simeq 10^{-8}-10^{-7}$, for $B_0\simeq 1\,{\rm nG}$ [@Wagstaff2015]. Due to the finite resolution of the beam, which essentially leads to mixing of blackbodies of different temperature and hence a distortion [@Chluba2004], these estimates provide useful upper limits; however, since many other sources of $y$-distortions at similar or larger level exist [e.g., @Refregier2000; @Chluba2011therm; @Hill2015], it will be hard to use future distortion measurements to improve existing PMF limits in this way.
[0.49]{}  \[fig:slice\_bb\_z3000\]
[0.49]{}  \[fig:slice\_bb\_z400\]
Turbulent decay after recombination {#sec:turb_decay}
-----------------------------------
In this section, we briefly recap the analytic approximations commonly used to describe the heating in the turbulent decay regime. We then provide an improved derivation that also includes the evolution of the magnetic field strength on the eddy turnover time. These approximations will be used later to represent the results from our numerical simulations (e.g., Sect. \[sec:semi-analytics\]).
### Review of previous analytical results {#sec:turb_decay_old}
As argued by @Banerjee2004, in the post-recombination era turbulent decay dominates the damping of the magnetic field energy density with a evolution law $$\begin{aligned}
E_B(z)\approx E_{B, \rm td} \left[\frac{t}{t_{\rm eddy}}\right]^{-2(n_B+3)/(n_B+5)}\end{aligned}$$ for $t\gg t_{\rm eddy}$. Here, $t_{\rm eddy}$ is the eddy turnover time of the largest turbulent mode at the initial time, $t_{\rm td}=t(z_{\rm td})$, and $E_B=|B|^2/8\pi=a^4 \rho_{B}$ is the corresponding comoving magnetic energy density.
How do we generalize this to the expanding Universe? First of all, we want to obtain the limit $t\rightarrow t_{\rm td}$ for which $E_B(z)\rightarrow E_{B,\rm td}$. @Sethi2005 suggested the generalization as $$\begin{aligned}
\label{eq:BK_sol}
E_B(z)\approx E_{B, \rm td} \left[1+\frac{t-t_{\rm td}}{t_{\rm eddy}}\right]^{-m}\end{aligned}$$ with $m=2(n_B+3)/(n_B+5)$, which indeed has the right limit towards $t\rightarrow t_{\rm td}$. In the expanding Universe one also needs to replace $$\begin{aligned}
\label{eq:t_tedd}
\frac{t-t_{\rm td}}{t_{\rm eddy}}\rightarrow \int \!\frac{\id t}{t_{\rm eddy}}
=\frac{1}{[t_{\rm eddy}H]_{\rm td}}\!\int_z^{z_{\rm td}} \!\frac{\id z}{1+z}
=\frac{1}{\tau_{\rm td}}\ln\left(\frac{1+z_{\rm td}}{1+z}\right).\end{aligned}$$ Here, we have taken into account the explicit redshift dependence of the eddy turnover time, $t_{\rm eddy}(z)\simeq \ell_{\rm eddy}/\varv_{\rm bA}\approx t_{\rm eddy}(z_{\rm td}) \,(a/a_{\rm td})^{3/2}$, using $\varv_{\rm bA}\propto a^{-1/2}$ and $\ell_{\rm eddy} \propto a$, and assumed $H\propto a^{-3/2}$ as appropriate for the matter-dominated era. This neglects any evolution of $\varv_{\rm bA}$ related to $\rho_{B}$. We also defined $\tau_{\rm td}=t_{\rm eddy}(z_{\rm td}) H(z_{\rm td})$, which is $\tau_{\rm td}\approx [R H/(\Ne\sigT c)]_{\rm td}\approx 0.063$ by construction. We thus obtain $$\begin{aligned}
\label{eq:rho_turn_old}
E_B(z)\approx E_{B,\rm td}
\left[1+\frac{1}{\tau_{\rm td}}\ln\left(\frac{1+z_{\rm td}}{1+z}\right)\right]^{-m},\end{aligned}$$ which agrees well with previously used expressions [cf., @Sethi2005]. With this we find the dissipation rate at $z\lesssim z_{\rm td}$ $$\begin{aligned}
\label{eq:QbarB_III}
\frac{\id (Q_{B}/\rho_\gamma)}{\id z}\approx - \frac{1}{(1+z)} \frac{m}{\tau_{\rm td}} \,\frac{r_{B, \rm td} }{\left[1+\frac{1}{\tau_{\rm td}}\ln\left(\frac{1+z_{\rm td}}{1+z}\right)\right]^{m+1}},\end{aligned}$$ which is usually used to describe the heating due to turbulent decay. Notice, that in contrast to the heating in the free-streaming regime, this term directly affects the baryons, raising their temperature and thus leading to changes of the cosmological recombination history.
### Phenomenological derivation for the turbulent decay law
The picture for decay of magnetic fields in the turbulent phase is that the time-scale on which energy is drained is given by the eddy turnover time around the critical scale that defines the comoving energy density of the fields, $E_B=|B|^2/8\pi=a^4\rho_B$. From Eq. it follows that $\lambda_{\rm c}\propto E_B^{-1/(n_B+3)}$. We then have $$\begin{aligned}
\label{eq:eddy_rate}
\frac{1}{t_{\rm eddy}}\approx \frac{\varv_{\rm bA}}{\lambda_{\rm c} a}\approx\frac{1}{t_{\rm eddy, td}}\,
\left(\frac{E_B}{E_{B, \rm td}}\right)^{1/m}\,\left(\frac{a}{a_{\rm td}}\right)^{-3/2},\end{aligned}$$ where we used $\varv_{\rm bA}\propto \sqrt{E_B/R}$ and set $m=2(n_B+3)/(n_B+5)$ as before. We normalized everything with respect to the values at $z_{\rm td}$. Then the evolution equation for the comoving magnetic energy density is given by $$\begin{aligned}
\label{eq:dE_dt_eddy}
\frac{\id E_B}{\id t}&\approx -\frac{E_B}{t_{\rm eddy}}
=-\frac{E_B}{t_{\rm eddy, td}}
\left(\frac{E_B}{E_{B, \rm td}}\right)^{1/m}\left(\frac{a_{\rm td}}{a}\right)^{3/2}
=-\frac{E_{B}^{(m+1)/m}}{\tau_{\rm eddy, td} \,a^{3/2}}\end{aligned}$$ with $\tau_{\rm eddy, td}=t_{\rm eddy, td}E_{B, \rm td}^{1/m}\,a_{\rm td}^{-3/2}$. Neglecting the expansion of the Universe, we can set $a=a_{\rm td}$ and then integrate from $t_{\rm td}$ to find $$\begin{aligned}
\label{eq:Ev_sol}
E_B(t)\approx E_{B, \rm td} \left[1+\frac{t-t_{\rm td}}{m t_{\rm eddy, td}}\right]^{-m},\end{aligned}$$ which is only valid for $n_B>-3$. Comparing with Eq. , shows that solution is a little different, with an extra factor $1/m$ modulating the eddy turnover time, which we will discuss below.
Now including the effect of expansion during the matter-dominated era yields $$\begin{aligned}
\label{eq:Eb_analytic}
E_B(t)\approx E_{B, \rm td} \left[1+\frac{1}{m\tau_{\rm td}}\,\ln\left(\frac{1+z_{\rm td}}{1+z}\right)\right]^{-m}\end{aligned}$$ in a similar manner. Comparing with Eq. , we can see that again an extra factor of $1/m$ appears. Taking the redshift derivative yields $$\begin{aligned}
\label{eq:QbarB_IV}
\frac{\id (Q_{B}/\rho_\gamma)}{\id z}\approx - \frac{1}{(1+z)} \,\frac{1}{\tau_{\rm td}}\,\frac{r_{B, \rm td} }{\left[1+\frac{1}{m\tau_{\rm td}}\ln\left(\frac{1+z_{\rm td}}{1+z}\right)\right]^{m+1}}\end{aligned}$$ for the magnetic dissipation rate in the turbulent decay regime. As we can see, the heating amplitude differs by $1/m$ with respect to Eq. and scales slightly faster with redshift for near scale invariant initial spectra, since we have $m\rightarrow 0$ for $n_B\rightarrow -3$.
Does the extra factor of $1/m$ in the evolution law make sense? For blue initial power spectra, the energy density is fully dominated by the smallest scales. For $n_B\rightarrow \infty $, we find $m\rightarrow 2$, so that the evolution law, Eq. , becomes $$\begin{aligned}
E_B(t)\approx E_{B, \rm td} \left[1+\frac{t}{2 t_{\rm eddy, td}}\right]^{-m}.\end{aligned}$$ This indicates that the relevant time for energy dissipation is about twice the eddy turnaround time, a result that is also found in typical turbulent flow simulations. For lower values of $n_B$, the evolution affects the total energy density of the magnetic fields to a smaller extent and the eddy turnover time also scales less strongly with time. In the limit of quasi-constant strength of fields at all scales ($n_B\approx -3$), we also have $t_{\rm eddy}(E_B)\approx {\rm const}$, which means that the field strength decays exponentially at a rate $1/t_{\rm eddy}$. This limit is more problematic, since in this case the true small-scale dissipation time-scale, $t_{\rm turb}$, will determine the rate of change of the energy density and not the time-scale at which energy is injected into the turbulent cascade ($\simeq t_{\rm eddy}$). However, for $n_B>-3$ our approximate model should still provide a reasonable description of the problem in the different regimes, as long as $ t_{\rm eddy}\gg t_{\rm turb}$.
A last argument why the factor $1/m$ seems appropriate is that we know that at the initial redshift $z_{\rm td}$ we should find $$\begin{aligned}
\frac{\id (Q_{B}/\rho_\gamma)_{\rm td}}{\id z}
&\equiv \frac{1}{[H (1+z)]_{\rm td}}\frac{\id a^4 \rho_B(z_{\rm td})}{a^4\id t}
\approx -\frac{1}{[H (1+z)]_{\rm td}}\, \frac{r_{B, \rm td}}{t_{\rm eddy, td}}
\nonumber
\\
&=-\frac{r_{B, \rm td}}{(1+z_{\rm td})\tau_{\rm td}}, \end{aligned}$$ as is directly obtained from Eq. for $z=z_{\rm td}$ but is [*not*]{} reproduced by the law in Eq. . This shows that the magnetic dissipation rate at the transition from free-streaming damping to turbulent decay for fixed $\rho_B(z_{\rm td})$ is independent of $n_B$.
We will use Eq. as a starting point for analytic representations of our numerical results. Even if it does not capture the physics completely, it does represent the late time evolution very well (see Fig. \[fig:heat\_rate\_comp\_index\]). Together with Eq. for the pre-recombination evolution, we obtain a very useful semi-analytic description over a wide range of PMF parameters (also see Sect. \[sec:semi-analytics\]).
[.49]{} ![Slices representing the baryon velocity absolute value $|u|=\sqrt{u_x^2 + u_y^2 + u_z^2}$ through the $x-z$ plane at $y=0$ in the simulation box at two different redshifts, $z=3000$ (upper panel) and $z=400$ (lower panel). Both $|u|$ and $u_\mathrm{rms}$ are in code units (i.e. divided by sound speed). Note the 50 times smaller intensity scale for $|u|$ at $z=3000$.[]{data-label="fig:slices_u"}](eps/uu_slice_k09_z3000.pdf "fig:") \[fig:slice\_uu\_z3000\]
[.49]{} ![Slices representing the baryon velocity absolute value $|u|=\sqrt{u_x^2 + u_y^2 + u_z^2}$ through the $x-z$ plane at $y=0$ in the simulation box at two different redshifts, $z=3000$ (upper panel) and $z=400$ (lower panel). Both $|u|$ and $u_\mathrm{rms}$ are in code units (i.e. divided by sound speed). Note the 50 times smaller intensity scale for $|u|$ at $z=3000$.[]{data-label="fig:slices_u"}](eps/uu_slice_k09_z400.pdf "fig:") \[fig:slice\_uu\_z400\]
Simulations of Heating by primordial magnetic fields {#sec:simulations}
====================================================
In this section, we describe the setup of our numerical magnetohydrodynamic (MHD) simulations to follow the evolution of cosmological magnetic and velocity fields from the photon drag dominated regime, across the recombination era, into the turbulent decay dominated regime. Our goal is to compute and follow the evolution of the net heating produced by the dissipation of magnetic energy. We first detail the numerical methods, the code employed, the form of the MHD equations and the integration scheme used. The choice of cosmological coordinates and initial conditions is then described and the evolution of the photon drag is presented. Conversion to physical units is also discussed. We then briefly discuss how to use the results to extract different dissipation and heating rates.
Numerical Method and Setup
--------------------------
### The Pencil-Code
We perform full 3-D simulations with the `Pencil-Code` [@PENCIL10], at a resolution of $N=1536^3$, to numerically investigate the details of heating and sourcing of turbulence by magnetic fields in the context of varying photon viscosity across the epoch of recombination. Two-dimensional slices of the simulated 3D magnetic and baryon velocity fields, near the start and end epochs, are shown in Figs. \[fig:slices\_b\] $\&$ \[fig:slices\_u\] and discussed in section \[subsec:fiducial\_u\_b\].
The `Pencil-Code` is a finite-differences code with a sixth-order integration scheme. It solves the equations that govern the flow of the plasma and its interactions with the magnetic field for every grid point individually using the values of three neighbouring cells for the derivatives. This method means we are limited to subsonic velocities and sub-sonic Alfvén velocities, which limits us to a physical magnetic field strength $B_{\rm phys}\lesssim \text{few}\times10^{-2}\,{\rm nG}$ (Sect. \[sec:phy\_units\]). In general, the density fluctuations are of the order $\simeq 1\%$, corresponding to (nearly) incompressible flows.
The code integrates the MHD equations in the following form, \[eq:MHD\] $$\begin{aligned}
\frac{\mathrm{D}\ln\rho}{\mathrm{D} t} &= -\nabla \cdot \mathbf{u} \label{eq:MHD1}
\\
\frac{\mathrm{D}\mathbf{u}}{\mathrm{D}t} \,\,\,&= -\frac{\nabla p}{\rho} + \frac{\mathbf{j}\times \mathbf{B}}{\rho} + \mathbf{f}_\mathrm{visc} \label{eq:MHD2}
\\
\frac{\partial\mathbf{A}}{\partial t} \,\,\,&= \,\,\,\mathbf{u}\times \mathbf{B} - \eta\ \mathbf{j} \label{eq:MHD3},\end{aligned}$$ where $\mathrm{D}/\mathrm{D}t= \partial /\partial t + \mathbf{u}\cdot \nabla$ is the convective derivative, $\rho$ describes the (mass) density, $\mathbf{u}$ the velocity field of the plasma and $\mathbf{j}=\nabla\times \mathbf{B}$ is the MHD current in the simulation. Equation (\[eq:MHD1\]) is the continuity equation of the plasma, Eq. (\[eq:MHD2\]) is the Navier-Stokes equation, and Eq. (\[eq:MHD3\]) is the induction equation of the magnetic field.
There also is a magnetic resistivity $\eta$ in the code, meaning we are working with *non-ideal* MHD. It is set in such a way that the magnetic Prandtl number, the ratio of viscosity to resistivity, is given by $\mathrm{Pr}=\nu_3/\eta_3=1$. The code uses the vector potential $\mathbf{A}$, with $\mathbf{B}=\nabla\times\mathbf{A}$ as a primitive variable, as it naturally ensures the divergence free condition $\nabla\cdot \mathbf{B}$ for the magnetic field.
We do not explicitly solve the energy equation in our simulation, since we use a isothermal equation of state where the pressure is only dependent on the density in the simulation with $p=c_s^2\rho$. Assuming a mono-atomic gas of particles, the sound speed is then given [e.g. @Ma1995] by $c_s= \sqrt{4/3\ k_\mathrm{B}\ T_\mathrm{b}/m_p}\approx 5.7\,{\rm km\,s^{-1}}\,\left[(1+z)/1100\right]^{1/2}$. We also assume a post-recombination neutral hydrogen gas in thermal equilibrium with the CMB photon field, $T_{\rm b}\simeq T_\gamma\propto a^{-1}$, which is valid until $z\lesssim 150$. At scales below the photon diffusion scale (most relevant to our problem), the expression employed for the sound speed is valid even in the pre-recombination era, as photons and baryons no longer behave as a single tightly-coupled fluid ($c_s = c/\sqrt{3}$) in the free-streaming regime [see @Jedamzik1998; @Banerjee2004 for a more detailed discussion]. At larger scales, corrections could become relevant, but the effect of damping is deemed negligible at those scales. We also neglect the effect of recombination on the sound speed, which is related to the change in the number of free particles. For instance, for a fully ionized electron-proton plasma, one would have[^8] $c^{\rm ep}_s\approx\sqrt{2}c_s\simeq 1.4 \,c_s$.
The last term in Eq. (\[eq:MHD2\]) describes the viscosity of the baryon-photon fluid. It includes the effect of photon drag, which we model as a viscosity acting on the fluid, and additional numerical viscosity to avoid numerical instabilities, as well as to mimic sub-grid dissipation effects in the fully turbulent regime: $$\begin{aligned}
\mathbf{f}_\mathrm{visc} &= \textbf{f}_\mathrm{hyper} + \textbf{f}_\mathrm{drag} \label{eq:visc_tot}\\
\mathbf{f}_\mathrm{hyper} &= \nu_3\nabla^6 \mathbf{u} \label{eq:visc_num}\ \\
\mathbf{f}_\mathrm{drag} &= -\alpha(z) \cdot \mathbf{u} \label{eq:visc_drag}.\end{aligned}$$ The effect of numerical viscosity $\mathbf{f}_\mathrm{hyper}$ is controlled by the variable $\nu_3$, which we optimize to avoid large pile-up of power at small-scales in the turbulent phase (see Sect. \[sec:num\_vis\] and App. \[sec:appendix1\] for a detailed discussion of convergence w.r.t. numerical viscosity in our simulations). We discuss the implementation of the photon drag term $\alpha(z)$ in Sect. \[sec:photon\_drag\]. Our simulations neglect all gravitational interactions of the plasma as well as any contributions from photon perturbations.
### Transformation to Cosmological Coordinates
The [Pencil]{}-Code works in dimensionless code units, with time expressed in units of the Hubble time, assuming the matter-dominated regime, $t_{\rm exp}=H^{-1}= t_0\,a^{3/2}\approx \pot {8.2}{17} (h/0.7)^{-1}\,a^{3/2}\,{\rm s}$. Here, we do not explicitly treat the radiation-dominated regime, which only changes the redshift-dependence of the drag-dominated dissipation rate (see Sect. \[sec:pic\_pre\_rec\]). We furthermore use the [*super-comoving*]{} coordinates [@Banerjee2004]: $$\begin{aligned}
&\tilde{\rho}_{\rm b}= a^3 \rho_{\rm b}
&\tilde{p}= a^4 p &
&\tilde{\vek{B}}=a^2 \vek{B} &
&\tilde{\vek{u}}=a^{1/2} \vek{u} &
\\
&\id \tilde{t}= \id t/t_{\rm exp}
&\tilde{\eta}= a^{-1/2} \eta &
&\tilde{\vek{A}}=a \vek{A} &
&\tilde{\vek{j}}=a^{7/2} \vek{j} &.\end{aligned}$$ This coordinate transformation is a rescaling that uses the conformal property of the MHD equations allowing us to describe MHD in an expanding FRW Universe via the form of the MHD equations in a flat spacetime (see Eq. \[eq:MHD\]). Only one additional term is introduced, a kinematic Hubble drag described in the next sub-section. Within the simulation, the comoving sound speed $\tilde{c}_s=1$, so that computational velocities or magnetic field amplitudes can be converted to physical units using $\vek{\varv} = c_s \tilde{\vek{u}}$ and $\vek{B} = \sqrt{4\pi\,\rho_{\rm b}}\, c_s \,\tilde{\vek{u}}_{\rm A}$, where $\tilde{\vek{u}}_{\rm A}$ is the Alfvén velocity obtained directly from the code. To map from $\tilde{t}$ to redshift, we use $\id \tilde{t}=\id t/t_{\rm exp}\equiv -\id \ln (1+z)$, where $\id t = -t_{\rm exp} a \id z$. We obtain $\tilde{t}=\ln[(1+z_0)/(1+z)]$, where $z_0$ is the starting redshift. This implies $(1+z)=(1+z_0)\,\expf{-\tilde{t}}$, which is then used to compute the evolving photon drag coefficient $\alpha(z)$.
![The evolution of the photon drag coefficient $\tilde{\alpha}$, Eq. , in the simulation code units (where rates are in units of Hubble rate) as well as the constant kinematic drag coefficient $H/2$ (or 1/2 in code units), which is important only at late epochs.[]{data-label="fig:effective_rate_k09"}](./eps/alpha_drag_incl_H2\logstr.pdf)
### Photon Drag {#sec:photon_drag}
The occasional scattering of photons off electrons in the free-streaming regime damps magnetic field power at small scales [@Jedamzik1998]. In physical units, the drag coefficient is given by [e.g., @Banerjee2004] $$\begin{aligned}
\alpha &= \frac{c}{R \ell_{\rm mfp}}=\frac{\sigT\Ne c}{R}=\frac{\sigT N_{\rm H}c}{R}\,X_{\rm e},\end{aligned}$$ where $N_\mathrm{H}\approx \pot{1.9}{-7}\,(1+z)^3\,{\rm cm^{-3}}$ is the hydrogen nuclei number density[^9] and $X_{\rm e}$ the free electron fraction, which is provided by `CosmoRec` [@Chluba2010b]. The time transformation $\id t \rightarrow \id \tilde{t}$, implies that the rescaled drag coefficient is $\tilde{\alpha}=t_{\rm exp}\alpha$, relevant in the computation. Inserting numerical values then yields $$\begin{aligned}
\label{eq:alpha_drag}
\tilde{\alpha}
&\approx \pot{4.6}{-6}
\,\left[\frac{\Omega_{\rm m} h^2}{0.14}\right]^{-1/2}\left[\frac{T_0}{2.726\,\Kel}\right]^{4}\,X_{\rm e}\,\,(1+z)^{5/2},\end{aligned}$$ where $\Omega_{\rm m} h^2$ sets the total matter density. The evolution of $\tilde{\alpha}$ is illustrated in Fig. \[fig:effective\_rate\_k09\]. At $z\gg 1100$, the photon drag is dominant and prevents turbulence from developing. As recombination proceeds, the drag coefficient drops quickly and the flow becomes turbulent (also see discussion of the magnetic Reynolds number cf. Fig. \[fig:reynold\_amplaa\_comp\]).
In the final form of the velocity equation , an extra kinematic drag term $-\frac{1}{2} H \vek{u}$ appears due to our choice of scalings corresponding to an expanding Universe [cf. @Banerjee2004]. This term is usually neglected; however, it becomes significant after recombination (see Fig. \[fig:effective\_rate\_k09\]) and can be incorporated by replacing $\tilde{\alpha}\rightarrow \tilde{\alpha} + \frac{1}{2}$ inside the code. In the computation of the net heating rate, this term has to be taken into account as it reduces the net heating rate noticeably.
### Physical Units {#sec:phy_units}
To convert our code units into physical units, we use the physical sound speed $c_s \approx 5.7\,{\rm km\,s^{-1}}\,\left[(1+z)/{1100}\right])^{1/2}$ at recombination, evaluated at the current epoch $ c_s \approx 0.17\, {\rm km\,s^{-1}}\,\left({1+z}\right)^{1/2} $ as a comoving sound speed. We assume $\Te=\Tg\propto (1+z)$ at all times. This approximation is valid in the standard recombination scenario until $z\simeq 150$; however, significant heating by magnetic fields could change this, an effect that we do not treat in our simulations.
We can then determine the comoving physical box size using the distance an Alfvén wave travels within a Hubble time $t_0$. Taking the Alfvén speed equal to the plasma sound speed gives a box size $L_{\rm phys} \approx t_0 c_s L_\mathrm{code} \approx 28 \, (h/0.7)^{-1}\,{\rm kpc}$, where $L_\mathrm{code}=2\pi$ for our simulation box. The wavenumber in physical units then can be calculated from the box size, $L_{\rm phys} \approx 2\pi/k_{\rm phys}$ yielding $k_{\rm phys} \approx 220\, (h/0.7)$ Mpc$^{-1}$. For the smallest resolved physical scale, we therefore have $k_{\rm res}\simeq (N/2)\,k_{\rm phys}\simeq \pot{1.7}{5}\, (h/0.7)$ Mpc$^{-1}$ with a typical number of grid points $N = 1536$.
The magnetic field strength from an Alfvén velocity equivalent to the sound speed is given by $B_{\rm phys} = \sqrt{4 \pi \,\rho_{\rm b}} \,c_s \approx 3.9 \times 10^{-2}$ nG. To avoid the requirement of including significant density perturbations and shocks, we must limit Alfvén velocities to $\varv_{\rm bA} \lesssim c_s/2$. Thus, for the simulations considered here (and presented in the subsequent figures) the comoving physical magnetic field strengths lie in the range $B_{\rm 0, phys} \simeq (0.086-2.0) \,\times \,10^{-2}$ nG, corresponding to Alfvén velocities in the range $\simeq (0.022-0.51)\,c_s$.
### Initial Conditions
In our simulations, we initialize the magnetic field with a specific power spectrum $P_M \propto k^n$ with a spectral index $n$, recalling that $n=n_\mathrm{B}+2$. This is done by initializing a Gaussian random field in $k$-space which is then multiplied with the chosen power-law and transformed back into real space. By doing this we also have full control over the helicity of the magnetic field $\mathcal{H}=\mathbf{A}\cdot \mathbf{B}$, which we set to zero at the initial epoch, focusing on non-helical fields. Since magnetic helicity is a conserved quantity (although strictly, only in ideal MHD) this remains constant for the entire simulation, except for numerical fluctuations. By choosing the amplitude of the magnetic power spectrum we can fix the $B_\mathrm{rms}$ in our simulation. This is also discussed in Sect. \[sec:av\_B\] (where magnetic power spectrum was denoted as $P_B(k)$ and the spectral index $n_\mathrm{B}$ differs by two: $n =n_\mathrm{B}+2$).
The velocity field is initialized with $\mathbf{u}\mid_{t=0}=0$. After a few time steps velocity fluctuations of the order $u_\mathrm{max}\approx c_s$ develop, which then damp out. This *ring-in* phase is usually restricted to the high redshifts $z \sim 5000$, but for very steep spectra or low magnetic field strengths, it can last until $z\simeq 3000$. We show our results and figures for times $z \le 3000$, after this phase has ended and the simulation has gone into a near steady state. Unless stated otherwise, we keep track of the evolution of the magnetic and velocity field during the initial phase and rescale our curves to their values at $z=3000$, taking that epoch as the relevant initial condition for physical interpretation.
### Numerical viscosity and real heating {#sec:num_vis}
We use the hyperviscosity implementation of the `Pencil-Code` since it gives the lowest numerical dissipation and largest Reynolds numbers in the simulation. A small amount of numerical viscosity is required but we do not want it to be stronger than the physical effects we wish to observe and follow. For a given resolution one can use a specific minimum viscosity pre-factor $\nu_3$. To further reduce this pre-factor, one has to increase the resolution $N$. For a given viscosity, $\nu_3$, the simulation is independent of resolution as long as one is above the minimum value of the viscosity, which is a function of resolution. On the other hand there must exist enough dissipation when the photon drag becomes negligible. A detailed discussion of how the viscosity pre-factor was chosen as well as the convergence tests of resolution can be found in App. \[sec:appendix1\]. The adopted value of $\nu_3=\pot{2.5}{-16}$ is used in all the studies below where when no explicit viscosity parameter values are mentioned.
Theoretical considerations for magnetic heating rates
-----------------------------------------------------
In our simulations, the PMFs are the origin of energy that can subsequently be dissipated. Physically, the PMFs drive baryon velocities through the Lorentz force, sourcing a kinetic energy density, $\rho_{\rm kin}=\frac{1}{2}\,\rho_{\rm b} \left<\varv_{\rm b}^2\right>$. This process builds up a flow, which itself loses energy through i) interactions with CMB photons (via photon drag) and ii) dissipative processes at ultra-small scales (via Coulomb interactions and plasma effects). Photon drag leads to small-scale perturbations in the photon temperature, which through photon mixing processes can cause $y$-type distortions (see Sect. \[sec:pic\_pre\_rec\_dist\]). The dissipative processes at ultra-small scales lead to heating of the baryons, which becomes important in the turbulent phase of the evolution.
Neglecting dissipation processes, the total (comoving) energy density[^10], $E_{\rm tot}=E_B+E_{\rm kin}$, related to magnetic fields is conserved. Here, $E_B=a^4 \rho_B\equiv \frac{1}{2}\,a^4 \rho_{\rm b}\left<\varv^2_{\rm A,b}\right>$ and $E_{\rm kin}=a^4 \rho_{\rm kin}=\frac{1}{2}\,a^4 \rho_{\rm b}\left<\varv^2_{\rm b}\right>$. Real changes to the total energy are then given by $$\begin{aligned}
\frac{\id E_{\rm tot}}{\id t}
&=\frac{\id E_B}{\id t}+\frac{\id E_{\rm kin}}{\id t}=-2\alpha E_{\rm kin}- H E_{\rm kin} - \left.\frac{\id E_{\rm tot}}{\id t}\right|_{\rm heat}.\end{aligned}$$ The first term on the r.h.s. is due to photon drag, which can be obtained as $\dot{E}_{\rm kin}|_{\rm drag}\simeq \frac{1}{2}\,a^4 \rho_{\rm b}\left<2\,\varv_{\rm b}\,\dot{\varv}_{\rm b}\right>\equiv-2\alpha\,E_{\rm kin}$ using $\dot{\varv}_{\rm b}=-\alpha \varv_{\rm b}$. The second term is related to the extra cooling from Hubble expansion. This term causes extra losses from the velocity field and has to be eliminated when computing the net heating rates. The last term is due to dissipative processes in the baryon plasma at ultra-small scales, which leads to real heating.
Around the recombination era, three phases exist: at redshift $z\gg z_{\rm td}$, the dissipation is dominated by photon drag, so that $$\begin{aligned}
\frac{\id E_{\rm tot}}{\id t} &\approx -2\alpha E_{\rm kin}.\end{aligned}$$ As is clear from Fig. \[fig:effective\_rate\_k09\], the Hubble term can be neglected here. In this phase, [*no real heating occurs*]{}, but perturbations in the photon fluid are generated at small scales through the Doppler effect. In this regime the drag is so strong that only very small velocity fluctuations are generated by magnetic fields (via the Lorentz force). The drag is strong enough to prevent magnetic fields from accelerating baryons up to the Alfvén-velocity, such that turbulence is highly suppressed (see Sect. \[sec:Results\] and Fig. \[fig:tser\_drag\_k09\]).
Around $z\simeq z_{\rm td}\simeq 10^3$, hydrogen atoms form and photon drag drops rapidly. The velocity field slowly builds up until a turbulent flow is formed. However, in the transition phase, little energy is actually dissipated, and most of the energy lost by the PMFs is dumped to increase the flow’s kinetic energy (see Fig. \[fig:tser\_drag\_k09\]).
In the turbulent phase ($z\lesssim z_{\rm td}$), photon drag can be neglected and a significant fraction of the energy lost from the magnetic fields is converted into heat by the turbulent cascade: $$\begin{aligned}
\frac{\id E_{\rm tot}}{\id t} & \approx - H E_{\rm kin} - \left.\frac{\id E_{\rm tot}}{\id t}\right|_{\rm heat}.\end{aligned}$$ In the simulations, we can thus compute the real heating of the medium in the various phases as $$\begin{aligned}
\label{eq:net_heating_rate_def}
\left.\frac{\id E_{\rm tot}}{\id t}\right|_{\rm heat}\approx -\frac{\id E_B}{\id t}-\frac{\id E_{\rm kin}}{\id t}- H E_{\rm kin} -2\alpha E_{\rm kin}.\end{aligned}$$ We will see below that the above picture is reproduced by the simulations. In particular, the transition phase between the drag-dominated and the turbulent regime is a new physical regime that was previously missing or treated as unrealistically abrupt.
### Extraction of heating rates from simulation
Within our simulations, the scales responsible for real heating of the medium (due to Coulomb interactions) are not resolved but instead mimicked by numerical viscosity to reproduce a quasi-Kolmogorov turbulent cascade with magnetic field power spectrum similar to $E_k\propto k^{-5/3}$ (see Sect. \[sec:num\_vis\]). To compute the total energy loss rate of the PMFs, we can use the definition of the Alfvén speed, $V_{\rm bA}$, Eq. for the ratio of the average magnetic to photon energy density. This yields $$\begin{aligned}
\label{eq:r_B}
\left<r_B\right>=\frac{\left<\rho_B \right>}{\rho_\gamma}=\frac{2 R}{3} \, \left<V_{\rm bA}^2\right>
=\frac{2 R}{3} \, \frac{c^2_{\rm s}}{c^2}\,\left<\tilde{u}^2_A\right>,\end{aligned}$$ where we used $V_{\rm bA}^2=(c^2_{\rm s}/c^2) \, \tilde{u}^2_A$. We recall that $\tilde{u}_A$ is the Alfvén velocity obtained directly from the code. Noting that $R \propto a$ and $c^2_{\rm s} \propto a^{-1}$, we can then compute the dissipation and heating rates in physical units in terms of the rate of change of Alfvén velocity $\tilde{u}_A$ or magnetic field amplitude $B_0$, both squared, and both in the same code units, $$\begin{aligned}
\label{eq:Q_{B comp}}
\frac{\id (Q_B/\rho_\gamma)}{\id \ln z}
=-\frac{\id \left<r_B\right>}{\id \ln z}
&\approx - \pot{1.4}{-10}\frac{\id \quad \,\,}{\id \ln z} \left<\tilde{u}^2_A\right> \nonumber \\
&\approx - \pot{3.7}{-11}\frac{\id \quad \,\,}{\id \ln z} \left(\left<\frac{B_0}{0.51}\right)^2\right>,\end{aligned}$$ where $c^2_{\rm s}/c^2\simeq \pot{3.2}{-13}(1+z)$ and the value $B_0=0.51$ corresponds to our fiducial case. Note that the magnetic dissipation rate ${\rm d}Q_B/{\rm d}\ln z$ or ${\rm d}E_{\rm mag}/{\rm d}\ln z$ in Fig. \[fig:heat\_rate\_k09\] is the negative rate of change of magnetic energy density ${\rm d}\langle \rho_B\rangle /{\rm d}\ln z$. Similarly, we can define the changes to the kinetic energy density of the fluid and net heating rates through Eq. .
Simulation Results for different PMF parameters {#sec:Results}
===============================================
The subsequent sections describe our results for the evolution of the amplitudes and power spectra of both magnetic and velocity fields as well as for the derived dissipation and heating rates. Initially, we present simulation results for a single value of magnetic field strength, chosen with a near scale-invariant magnetic power spectrum (Sect. \[sec:scale-invariant\]). Subsequently, we discuss results for varying magnetic field amplitude as well as varying magnetic spectral index (Sect. \[sec:var\_B0\] and \[sec:var\_nB\]). We also compare our results with analytical expressions, highlighting expected trends and differences, providing analytic fits where relevant.
Scale-invariant magnetic spectrum {#sec:scale-invariant}
---------------------------------
We first present simulation results for a single fiducial case with magnetic field amplitude $B_0 = 0.51$ (in code units of $\varv_A/c_s$) at $z=3000$. This field amplitude is equivalent to $B_{0, \rm phys} \approx 2.0 \times 10^{-2}$ nG. This fiducial case is chosen for a nearly scale-invariant power spectrum $n_B = -2.9$ (or alternatively, $n = n_\mathrm{B} + 2 = -0.9$), which is of particular importance in the study of primordial magnetic fields generated by inflation.
### Evolution of magnetic and kinetic amplitude {#subsec:fiducial_u_b}
A visual representation of the changing magnetic field strength and baryon velocity over 2D slices from our 3D simulations is shown in Figs. \[fig:slices\_b\] $\&$ \[fig:slices\_u\]. A video of the temporal evolution of both the magnetic and kinetic field intensity slices can also be viewed at <https://www.hs.uni-hamburg.de/research/cosmo-mf>. Note the varying intensity scale between $z=3000$ $z=400$, particularly for the velocity in Fig. \[fig:slices\_u\]. We can clearly see how the magnetic field decays slowly before recombination, evolves quicker over the transition regime and then spatially mixes and decays in the turbulent regime. The velocity field is boosted from zero initial value to saturation toward the end of the transition regime and subsequently exhibits turbulent mixing and decay.
To analyze the evolution of the magnetic and kinetic energy density over the course of our simulation, we plot the root mean square (rms) velocity and magnetic field strength in Fig. \[fig:tser\_drag\_k09\]. At higher redshifts prior to decoupling, the fluid is dominated by photon drag. During this drag dominated regime, there is little evolution of the magnetic field – it decreases by less than $10\%$ from its early value of $B_0=0.51$ at $z=3000$. The kinetic field, initially starting from zero, is still highly suppressed, by more than two orders of magnitude compared to the sound speed $c_s$, at $z=3000$. It then rises slowly from $u_\mathrm{rms}\approx \pot{1.5}{-3}$ to $\pot{5}{-3}$ at $z\approx 1500$ as the drag coefficient decreases (see Fig. \[fig:effective\_rate\_k09\] where the drag is shown separately). We fit this slower rise of the velocity field with a power law $u \, \sim \, (1+z)^{-\beta}$, with $\beta = 1.32$.
![Evolution of the r.m.s. amplitudes (in code units, i.e. relative to the sound speed) of the magnetic field $B_\mathrm{rms}$ and baryon velocity field $u_\mathrm{rms}$ for the fiducial case of a near scale-invariant magnetic power spectrum with amplitude $B_0 = 0.51$ at $z=3000$ (corresponding to $B_{0, \rm phys} \approx 2.0 \times 10^{-2}$ nG).[]{data-label="fig:tser_drag_k09"}](./eps/timeseries_k09_nu_2_5e-16\logstr.pdf)
Below a redshift of $z \, \sim \, 1400$, the transition regime sets in due to a sharp decrease in the photon drag (Fig. \[fig:tser\_drag\_k09\]). For this period we fitted the steep increase in the velocity field amplitude with an exponential function $u\, \sim \, e^{-\alpha z}$ yielding a value of $\alpha = \pot{5.4}{-3}$. This transition regime lasts until the velocity to magnetic field ratio approaches a maximum value of $u_\mathrm{rms}/B_\mathrm{rms} \approx 1/2$. However, actual equipartition, which one could expect, is not reached and could be related to the fields’ intermittent nature over the total volume [@Subramanian1998turbulentdynamo; @Federrath2011]. This is also similar to the departure from full equipartition observed in earlier simulations of turbulence with the `Pencil-Code` [e.g., @Reppin17].
At $z\approx 800$ turbulent decay starts to set in, leading to the continued dissipation of the magnetic field energy as well as a turnover and then gradual decrease in the kinetic energy of the baryon velocity field. During the turbulent period, the r.m.s. magnetic field strength further decreases such that by $z=100$ it is approximately $60\%$ of its original value at $z=3000$. This can be seen more clearly in the upper panel of Fig. \[fig:ts\_comp\_k09\]. The transition phase captured in our simulations introduces a significant delay in the onset of turbulence.
### Magnetic and kinetic power spectra
![Evolution of magnetic (solid lines) and kinetic (dashed lines) power spectra versus redshift for a scale-invariant initial magnetic field. Spectra are shown for early times $z=3000$, during the transition regime at $z=1200\ \&\ 900$, and once the fluid has become fully turbulent at $z=400$. The magnetic field strength is $B_0 = 0.51\, \Leftrightarrow \, B_{0,\mathrm{phys}} = \pot{2.0}{-2}\mathrm{nG}$ at the reference redshift $z=3000$[]{data-label="fig:mag_spectra_k09"}](./eps/spectra_k09_nu_25e-16_combined.pdf)
{width="1.9\columnwidth"}
In Fig. \[fig:mag\_spectra\_k09\] (upper panel), we show the magnetic power spectra $P_M$ for different redshifts in our simulation for an initial near scale-invariant magnetic field ($n=-0.9$ or $n_{B}=-2.9$). The power spectrum $P_M$ is defined to be $\varv_A^2$ integrated over the angular variables in configuration space. The total energy of the magnetic field can be computed as the integral of the power spectrum in $k$-space, $E = \int P_M \dk = \int k P_M \dlnk$ and one can also define an integral scale $k_I$. For a scale-invariant spectrum we have $P_M\propto k^{-0.9}$ and the integral scale in our simulation is the box size $L$. For the more blue cases, $P_M \propto k^{n},\ n=0,1,2$ the integral scale $k_I$ is at larger $k$, a little closer to the Nyquist wavenumber in the simulation.
For the magnetic power spectra, shown in Fig. \[fig:mag\_spectra\_k09\], the first spectrum is plotted at $z=3000$, when the ring-in phase for velocities has already ended. The spectrum follows a power-law $P_M \propto k^{-0.9}$ up to the high-$k$ regime. The small-scale fluctuations have been damped out and the spectrum has a steep decline towards the Nyquist wavenumber $k_\mathrm{Ny}=N/2$, where $N$ is the resolution. Close to the Nyquist scale, the magnetic power spectrum has an exponential cutoff with the shape $P_M\propto \expf{-2k^2/k_{\rm d}^2}$, with the dissipation scale $k_{\rm d}$ [@Jedamzik1998; @Subramanian2016]
We observe that the power spectrum is slowly reshaped during the transition period, as visible for $z$=1200 and 900. This is because the ionization fraction decreases and the drag force on the baryons becomes negligible. The scale-invariant initial spectrum is transformed into a broken power-law spectrum with a knee below which it is steeper. In the turbulent phase ($z\lesssim800$), this knee, where the $k^{-0.9}$ spectrum turns into a turbulent one, moves to smaller $k$ over time, which means ever larger scales become turbulent in the simulation. Also, due to the negligible drag force at $z\lesssim 800$, the viscosity becomes dominated by the small hyperviscosity we added to the simulation setup. This has the effect of extending the spectrum more and more towards the Nyquist scale $k_\mathrm{Ny}$ and reducing the strong exponential cutoff which is apparent in the earliest spectrum shown at $z=3000$. At the latest redshift shown, $z=400$, the fluid has become fully turbulent at the scales present in our simulation with the power spectrum steepened to an almost Kolmogorov spectrum (with an energy distributed as $E\propto k^{-5/3}$), though in our simulations we find a power-law exponent of $\simeq -1.44$, slightly flatter than the canonical value.
The kinetic power spectrum evolves in a clearly different manner (see the lower panel of Fig. \[fig:mag\_spectra\_k09\]). We start our simulation with zero initial velocity $\mathbf{u}_0=0$ and the first power spectrum of the kinetic field at $z=3000$ is during an epoch, when the fluid is heavily dominated by photon drag. At this epoch only a small fraction of the total energy is in the velocity field, so the power spectrum is approximately two orders of magnitude below the magnetic power spectrum. The kinetic power spectrum is peaked at a very large wavenumber, $k_\mathrm{peak, phys}\approx \pot{4}{4} \ (h/0.7)\ \mathrm{Mpc}^{-1}$.
During the transition period a stronger kinetic power spectrum develops. It has a roughly $E_k\propto k^2$ shape until it reaches its peak, which shifts to lower $k$ as the kinetic power spectrum builds up. At wave numbers larger than the peak, there is a turbulent spectrum, with $E_k\propto k^{-5/3}$. By $z=900$, the slope of the turbulent part of the power spectrum is found to be $-1.71$. At the final redshift, $z=400$, the slope becomes slightly flatter than Kolmogorov due to a small bottleneck effect at the highest $k$-values (as tuning of hyperviscosity $\nu_3$ was done for the *magnetic* power spectrum). As the ionization fraction decreases and becomes negligible, the turbulent part of the spectrum covers a wider range of scales, starting from lower and lower $k$-values. Towards the end of the simulation, $z=400$, this scale is at $k\simeq 1000\ (h/0.7)\ \mathrm{Mpc}^{-1}$. A video of the temporal evolution of both the magnetic and kinetic spectra can be viewed at <https://www.hs.uni-hamburg.de/research/cosmo-mf>.
### Energy dissipation rates and net heating
In Fig. \[fig:heat\_rate\_k09\], we show the dissipation rate, i.e the negative rate of change, per logarithmic redshift, of the magnetic energy density ${\rm d}E_\mathrm{mag}/{\rm d}\ln z$ (blue curve) calculated using Eq. . Also shown are the dissipation rates of the kinetic energy density (purple), total ($E_\mathrm{mag}+E_\mathrm{kin}$) energy density (orange), the photon drag force (per unit area) $2\tilde{\alpha} E_{\rm kin}$ (green), and finally the difference between the total energy dissipation rate and the drag rate, which can be interpreted as the *net heating rate* being imparted into the fluid (red curve). Note that the kinematic drag term, related to the Hubble expansion, $-(1/2)H\mathbf{u}$, is subtracted out from the net heating and not shown as included in the drag curve (green). At high redshifts we also show the expected net heating rate (red dashed) if we had very low viscosity at $z>1500$, taken from extrapolation from the net heating behaviour for lower $\nu_3$ shown in Fig. \[fig:heat\_rate\_highz\].
In Fig. \[fig:heat\_rate\_k09\], one can clearly identify the three stages discussed above: at early times, the net heating is very small and energy losses from the magnetic fields are created by the effect of photon drag. Starting from $z\simeq 1500$, baryon velocities build up, absorbing losses from the magnetic fields, until a turbulent flow is formed. Once turbulence develops below $z\sim700$, the velocity field starts losing energy via a turbulent cascade. Net heating is observed to gradually build up once baryon velocities start rising. The rise becomes faster as the drag decreases. Then the net heating peaks as the transition era gives way to turbulent decay, and net heating decreases throughout the turbulent phase down to low redshifts.
For the chosen fiducial case initial conditions, the magnetic energy dissipation rate peaks around recombination at a redshift of about $z\approx 1000$, whereas the net heating rate peaks at $z\approx 600$. This significant delay ($\sim 0.5$ Myr) is caused by the time taken for the fluid to transition into the turbulent state. The net heating rate has a broad distribution and is not a sharply peaked feature as was expected from earlier analytical studies. This means that the onset of modifications to the cosmological ionization history are expected to be more gradual, which will avoid numerical issues pointed out in @Chluba2015PMF.
Note that the magnetic energy dissipation rate curve lies above the total energy dissipation rate up until a redshift of $z\approx 700$. This is because up until then the kinetic energy density increases, adding a negative contribution (dot-dashed purple curve) to the total energy dissipation rate. At late times, both magnetic and kinetic fields decay turbulently, contributing positively to the total energy dissipation rate ${\rm d}E_\mathrm{tot}/{\rm d}\ln z$. In this turbulent phase, the real losses from the system lead to heating of matter.
Variation of magnetic amplitude {#sec:var_B0}
-------------------------------
In this section we present simulation results for a varying magnetic field amplitude over the range $B_0 = (0.022-0.51)$ corresponding to $B_{\rm 0, phys} \simeq (0.086-2.0) \,\times \,10^{-2}$ nG, investigating the behaviour of magnetic and kinetic amplitudes as well as dissipation and net heating rates as a function of varying $B_0$.
### Evolution of magnetic and kinetic amplitude {#evolution-of-magnetic-and-kinetic-amplitude}
![ Evolution of magnetic (top) and kinetic (bottom) field r.m.s. amplitudes versus redshift for different initial magnetic field strengths $B_0$. The initial field strength was varied by a factor of $\simeq 20$ (values denoted in the legend) and the magnetic amplitudes are all normalized to unity at $z=3000$. Velocities are rescaled to the magnetic field strengths $B_\mathrm{rms}$.[]{data-label="fig:ts_comp_k09"}](./eps/timeseries_amplaa_comparison_scaled\logstr.pdf)
In Fig. \[fig:ts\_comp\_k09\] we show different simulation runs with varying initial ($z=3000$) magnetic field strength amplitude $B_0$. The upper panel shows the evolution of the magnetic field strength $B_\mathrm{rms}$ on a linear vertical scale. The lower panel shows the velocity field strength $u_\mathrm{rms}$, rescaled to the magnetic field strength $B_\mathrm{rms}$, on a logarithmic vertical scale. For the lowest initial field strength $B_0=0.022$ there is a $20\%$ decrease of the initial rms field value over the entire redshift range $3000> z>100$ that goes into increasing the velocity field. This diminution or decay of magnetic fields becomes more appreciable reaching $\simeq 45\%$ for the strongest magnetic field case $B_0=0.51$, highlighting that turbulent time-scales become shorter with increasing magnetic field strength.
As we first observed for the fiducial case, we again see three distinct phases from the simulation results for different initial amplitudes. The first one is the drag dominated regime, where $u_\mathrm{rms} \ll 1$ (i.e., $\varv \ll {\rm v_{bA}}$). From the upper panel of Fig. \[fig:ts\_comp\_k09\], where $B_\mathrm{rms}(z)$ is depicted, we see that for the lower $B_0$ runs, much less than a percent of magnetic field energy is transferred to $u_\mathrm{rms}$ during the drag dominated phase $3000 >z > 1500$, whereas it is $\sim 1\%$ for the highest $B_0$ case. The $u_\mathrm{rms}$ amplitude (plotted in units of initial $B_0$) is an increasing function of the initial amplitude $B_0$ in the drag regime. The stronger the initial magnetic field amplitude $B_0$, the more energy is transferred to the velocity field, albeit at a relatively small fraction. In the subsequent transition regime, $B_\mathrm{rms}$ values decay most rapidly while the velocities rise more sharply, both effects due to the decreasing drag. However, by the end of the transition phase all curves catch up to reach a similar ratio of $u_\mathrm{rms}/B_\mathrm{rms} \approx 0.3$, as seen in Fig. \[fig:ts\_comp\_k09\] around $z\approx 650$.
We fit to the obtained velocity data (i) a power law $u \simeq (1+z)^{-\beta}$ for the drag regime and (ii) an exponential $u\simeq \exp(-\alpha z)$ for the transition regime. This enables us to derive simple semi-analytical fit formulae plotted in Fig. \[fig:fit\_values\] for the parameters $\alpha$ and $\beta$ of the growing velocity field, as a function of $B_0$. $$\begin{aligned}
\alpha &= (4.49 + 0.57\ B_0^{-0.42})\times 10^{-3} \nonumber \\
\beta &= 4.97 - 4.0\ B_0^{0.12}
\label{eq:alpha_beta}\end{aligned}$$ The coefficient $\alpha $ in the transition exponential fit is steeper for lower initial field strengths $B_0$, visible in the upper (orange) curve in Fig. \[fig:fit\_values\], while the drag fit power-law exponent $\beta$ plotted as the lower (blue) curve, varies less.
The last phase is that of turbulent decay where both $u_\mathrm{rms}$ and $B_\mathrm{rms}$ decrease slowly at $z<600$. This occurs for all simulation runs of the amplitude variation in a similar fashion, but with a greater overall decay for larger initial amplitude $B_0$. In the next section we compare the magnetic energy dissipation rates to those expected from turbulence. Although the rate of decline of $B_\mathrm{rms}$ was faster in the transition era, the overall decline in the turbulent decay regime till its end (at $z =100$ in the simulation) is comparable to the decline in the transition regime. However, the decline in the velocity field rms amplitude over the turbulent regime, by at most a few tens of a percent up to a factor of two, is much slower than its rapid rise by more than an order of magnitude in the two earlier regimes.
![Parameters of the fitted curves describing the evolution of the velocity field $u_\mathrm{rms}$ in Fig. \[fig:ts\_comp\_k09\]. We fitted a curve for the drag dominated regime with $u\simeq (1+z)^{-\beta}$ (triangles), one for the transition regime with $u\simeq \exp{(-\alpha z)}$ (circles). The data points were then fitted with a power-law to give an analytical representation of $\alpha$ and $\beta$, given in Eq. []{data-label="fig:fit_values"}](./eps/fitting_values.pdf)
![Magnetic Reynolds numbers as a function of redshift for simulation runs where $B_0$ is varied. The critical magnetic Reynolds number $\mathrm{Rm} =1$ is indicated with a grey dashed line. For different initial amplitudes the $\mathrm{Rm} =1$ redshift varies from $z_\mathrm{td}\simeq 650$ (fiducial $B_0$ = 0.51) to $z_\mathrm{td}\simeq 300$.[]{data-label="fig:reynold_amplaa_comp"}](./eps/reynolds_amplaa_comp\logstr.pdf)
The redshift at which turbulence becomes important for the velocity field is illustrated in Fig. \[fig:reynold\_amplaa\_comp\], where we plot the evolution of the magnetic Reynolds number $\mathrm{Rm} = (\mathrm{v_A}/L)/\tilde{\alpha}$, as a function of the initial magnetic field amplitude $B_0$. For the fiducial case, $B_0=0.51$, the magnetic Reynolds number rises above unity only at redshift $z\lesssim 650$ which matches the epoch where turbulent decay sets in for the velocity field, see Fig. \[fig:ts\_comp\_k09\]. For weaker fields, the transition to the turbulent regime occurs even later.
![Comparison of the dissipation rate of the magnetic energy density with varying amplitude of the initial magnetic field $B_0$. All curves were normalized to their peak amplitude. The $z\gtrsim 1500$ drag-dominated range is affected by numerical convergence issues for the lower amplitudes. Black dotted curves are fits (Eqs. \[eq:fit\_func\_k09\_turb\] $\&$ \[eq:fit\_func\_k09\_trans\]) for turbulent and transition regimes.[]{data-label="fig:heat_rate_comp_k09"}](./eps/heat_rate_amplaa_comparison_k09_scaled_loglog_fits.pdf)
![Net heating rates for different initial magnetic field strengths $B_0$. For the high $B_0$ runs the peak of the net heating is well-pronounced. The low amplitude runs are dominated by hyperviscosity at high-$z$, so dashed lines were plotted to show the physical decline of net heating for earlier times.[]{data-label="fig:net_heat_rate_comp_k09"}](./eps/net_heat_rate_comparison_amplaa_scaled\logstr.pdf)
![Physical normalization of the magnetic dissipation rate ${\rm d}Q_B/{\rm d}\ln z$ (Fig. \[fig:heat\_rate\_comp\_k09\]) at its peak (blue line) and the net heating rate (Fig. \[fig:net\_heat\_rate\_comp\_k09\]) at its peak (red line), for different initial field strengths $B_0$.[]{data-label="fig:heat_rate_amplaa_k09"}](./eps/heat_rate_amplaa_dependence.pdf)
### Dissipation rates with varying magnetic amplitude
We now investigate the difference in the magnetic dissipation and baryon heating rates as a function of magnetic amplitude with the aim of possibly placing constraints on primordial magnetic fields.
The magnetic energy density dissipation rate (per logarithmic redshift interval), normalized to the energy density of radiation ${\rm d}(Q_B/\rho_\gamma)/{\rm d}\ln z $ is plotted for varying $B_0$ in Fig. \[fig:heat\_rate\_comp\_k09\]. Note that in this plot the curves have all been normalized to their peak value, $\left[{\rm d}(Q_B/\rho_\gamma)/{\rm d}\ln z\right]_\mathrm{peak}$. The normalization in physical units as a function of $B_0$ is given separately in Fig. \[fig:heat\_rate\_amplaa\_k09\]. We find that there is a dependence on $B_0$ of the redshift at which the magnetic energy density dissipation peaks, particularly for the weaker fields. The peak redshift varies from $z\simeq 1020-820$ over the range of magnetic field amplitudes $B_0=0.51-0.022$ while exhibiting a non-linear $z_\mathrm{peak} \leftrightarrow B_0$ relation.
Another trend that was observed by varying the field strength amplitude was that the height of the peak compared to the magnetic energy dissipation rate values at $z\approx 1500$ in Fig. \[fig:heat\_rate\_amplaa\_k09\] is a monotonic function of $B_0$. We find an increase of ${\rm d}Q_B/{\rm d}\ln z$ by a factor of $\simeq 3$ for $B_0=0.51$ (blue line) and an increase by a factor of $\simeq 8$ for the run with lowest amplitude with $B_0=\pot{2.2}{-2}$ (red line). The shape of $\mathrm{d}Q_B/\mathrm{d}\ln z$ between $z=3000$ and $z=1500$ is less robust and also more susceptible to numerical issues so we ignore those values, since we do not extract any physical properties from the simulation during this period. For the $B_0=0.51$ the high-$z$ behaviour can be described by a power-law $$\left.\frac{\mathrm{d}(Q_B/\rho_\gamma)}{\mathrm{d}\ln z}\right|_\mathrm{drag}= 0.23\times (1+z)^{0.054}.$$ For the lower amplitudes the numerical noise is too strong at those early redshift, so we do not infer anything from the high-$z$ behaviour. For the magnetic dissipation rate, the peak position does not coincide with Rm $\simeq 1$, but rather with the epoch at or just before when magnetic and kinetic energy densities reach closest to equipartition (see Fig. \[fig:ts\_comp\_k09\]). The real importance of turbulence can only be appreciated when considering the net heating rates (see Fig. \[fig:net\_heat\_rate\_comp\_k09\] and discussion below), which show that for a large redshift range the PMF evolution for small magnetic fields is only weakly turbulent. This leads to a net heating rate with a very broad maximum, in contrast to previous analyses.
![Fitting parameters for the semi-analytic formula used to describe the magnetic dissipation rates (Fig. \[fig:heat\_rate\_comp\_k09\]) in the turbulent decay regime for different $B_0$. The redshift parameter $z_0$ saturates towards $z \sim 950$ whereas the parameter $1/\tau_0$ increases for larger $B_0$.[]{data-label="fig:amplaa_fitting_paras"}](./eps/amplaa_fitting_parameters.pdf)
The net heating rate is computed by subtracting the dissipation due to drag, which can be calculated as $2(\tilde{\alpha}+1/2) E_{\rm kin}$, from the total energy dissipation rate ${\rm d}E_\mathrm{tot}/{\rm d}\ln z$. In Fig. \[fig:heat\_rate\_k09\] we had displayed the rates of $E_{\rm mag},\ E_K,\ E_\mathrm{tot},$ drag dissipation rate and the net heating rate for our fiducial case, $\nu_3=\pot{2.5}{-16},\ B_0=0.51$. We now show the net heating rates $\left[ \mathrm{d} E_{\rm tot}/\mathrm{dln} z - 2(\tilde{\alpha}+1/2) E_{\rm kin} \right]/\rho_\gamma$ for the simulation runs with varying magnetic amplitudes in Fig. \[fig:net\_heat\_rate\_comp\_k09\]. The net heating rate peaks at similar redshifts $z\sim 550-650$, with most of the small shift in epoch occurring between $B_0$=0.51 and 0.31. It rises to a peak over a roughly similar redshift range towards the end of the transition regime, once the drag dissipation has become negligibly small, and the peak epoch does not strongly dependent on varying $B_0$. The peak epoch of the net heating is significantly later ($\sim 0.5$ Myr) compared to the peak of magnetic dissipation, for all $B_0$ values, as already mentioned above, due to the time taken to develop sufficiently turbulent velocity fields. Another new result from these simulations is that net heating generally does not peak sharply but over a broad range of redshifts, increasingly so for weaker initial amplitudes $B_0$. The strongest amplitude case $B_0=0.51$ (blue) lowest curve in Fig. \[fig:net\_heat\_rate\_comp\_k09\]) has the most pronounced peak at $z\approx 650$ in the heating. For the lower $B_0$ runs the peak is very broad and almost constant over a considerable range of redshifts $z\sim 500-700$.
In Fig. \[fig:heat\_rate\_amplaa\_k09\] we show the physical normalizations of the amplitude of the peak of ${\rm d}Q_B/{\rm d}\ln z$ (blue line), as well as the maximum value of the net heating rate (red line) as a function of the initial magnetic field strength $B_0$. We find that the magnetic energy dissipation rate has the following dependence of the initial magnetic field in our simulations: $$\begin{aligned}
\left. \frac{\mathrm{d} (Q_B/\rho_\gamma)}{\mathrm{d} \ln z }\right|_\mathrm{peak} = 3.17\times 10^{-11}\times B_0^{2.26}\\
\left.\frac{\mathrm{d}Q_\mathrm{net \,\,heat}}{\mathrm{d}\ln z}\right|_\mathrm{peak} = 8.04\times 10^{-12}\times B_0^{2.55}.\end{aligned}$$ Since the magnetic energy dissipation rate is simply the logarithmic derivative of the magnetic energy density $\propto B^2$, we expect a scaling of ${\rm d}Q_B/{\rm d}\ln z\propto B_0^2$. The observed scaling of the amplitude in our runs is slightly steeper than quadratic.
### Semi-analytic representation of magnetic dissipation
We also show fits to the turbulent and transition regime portions of the magnetic dissipation rates as a function of $B_0$. For the low redshift turbulent decay dominated part, we fit an inverse logarithmic function of the form derived phenomenologically in Eq. , for the case of a near scale-invariant magnetic spectrum $n_B=-2.9 \Leftrightarrow m = 0.095$: $$\left. \frac{\mathrm{d} (Q_B/\rho_\gamma)}{\mathrm{d} \ln z }\right|_{\mathrm{turb}} = \left(1 + \frac{1}{(0.095)\tau_0} \ln\frac{1+z_0}{1 +z}\right)^{-1.095}.
\label{eq:fit_func_k09_turb}$$ This fitting process for parameters $\tau_0$ and $z_0$ was applied over a range of low redshifts $z \in [100, z_\mathrm{max}]$, where $z_\mathrm{max}$ ranges from $500$ to $900$, for initial amplitude $B_0 = 0.51-0.022$. This fitting range was chosen to give better agreement in the lower redshift part of the turbulent regime. The resulting parameters for the semi-analytic magnetic energy dissipation rate are shown in Fig. \[fig:amplaa\_fitting\_paras\]. By applying a power-law fit to the obtained values we can give a description of the parameters $z_0$ and $\tau_0^{-1}$ as a function of the initial magnetic field strength $B_0$: $$\begin{aligned}
z_0 &= 959 -2.37\, B_0^{-1.08} \nonumber \\
\tau_0^{-1} &= 0.75 -0.30\, B_0^{-0.17}.\end{aligned}$$ We should note that the fitted coefficients and exponents given above are somewhat sensitive to the chosen redshift range in which the fit is performed.
In the transition regime, on the high redshift side of the maximum rate of change of the magnetic energy density, we fitted a power-law of the form: $$\left. \frac{\mathrm{d}(Q_B/\rho_\gamma)}{\mathrm{d}\ln z}\right|_\mathrm{trans} = \,\,A\times (1+z)^{-\beta},
\label{eq:fit_func_k09_trans}$$ where $\beta \in [3.6, 5.0]$. The two separate fits over the turbulent and transition regimes were co-added with a cubic exponent and over-plotted as black dotted lines in Fig. \[fig:heat\_rate\_comp\_k09\], $$\left. \frac{\mathrm{d}(Q_B/\rho_\gamma)}{\mathrm{d}\ln z}\right|_\mathrm{drag} = \,\,A\times (1+z)^{\gamma}.
\label{eq:fit_func_k09_drag}$$ With $A=0.23$ and $\gamma = 0.054$. For $B_0=0.51$. The simulation of the lower amplitudes $B_0$ have not fully equilibrated at high-$z$ and the behaviour of $\mathrm{d}Q_B/\mathrm{d}\ln z$ is dominated by numerical effects, resulting in slopes $\gamma$ that show non-monotonic behaviour as a function of $B_0$, so we only fit for the run with highest $B_0$.
Dependence on spectral index {#sec:var_nB}
----------------------------
Since we set up the magnetic field power spectrum as an initial condition, we have the freedom to vary its spectral shape. A scale-invariant magnetic power spectrum $P_M \propto k^{-0.9}$ is expected from models of inflationary magnetogenesis [@Subramanian2016]. We also know that $P_M \propto k^4$ is an initial power spectrum that represents a magnetic field generated in a causal process, like a first-order phase transition [@Durrer03]. Probing PMF evolution for different magnetic spectra thus has the potential to constrain and discern among different models of magnetogenesis. In the following subsections we describe the trend of physical results from simulation runs with varying spectral indices.
### Evolution of magnetic and kinetic field amplitudes
![Evolution of the rms amplitudes with different initial spectra. The solid lines give the magnetic field strength, the dashed lines are the *rms* velocity in the simulation. For the steep spectra the same initial integral scale was used.[]{data-label="fig:ts_comp_index"}](./eps/timeseries_index_comparison_scaled\logstr.pdf)
In Fig. \[fig:ts\_comp\_index\] we show the redshift evolution of the magnetic field $B_\mathrm{rms}$ in the upper panel and the velocity field $u_\mathrm{rms}$ in the lower panel. The different curves correspond to (from top to bottom in the upper panel) a near scale-invariant spectrum $P_M\propto k^{-0.9}$ (blue), two flatter spectra $P_M\propto k^{-0.5}$ (orange), $P_M\propto k^0$ (green) and two steep spectra with $P_M\propto k^1,\ P_M\propto k^2 $ which are plotted as red and purple lines, respectively. The curves for $B_\mathrm{rms}$ and $u_\mathrm{rms}$ were normalized at $z=3000$, when the ring-in phase of the velocity field died down. It should be kept in mind that the relative evolution between $z=3000$ and the starting redshift of the simulation depends on the spectral index.
For the magnetic field strength $B_\mathrm{rms}$ there is a clearer difference in its evolution as a function of its spectrum. The scale-invariant case shows the least decay: its $B_\mathrm{rms}$ value at $z=100$ is $60\%$ of the initial value at $z=3000$, with a greater fraction of the decay happening during the later phases $z<1000$. On the other hand, the steepest spectrum $P_M\propto k^4$ decays by almost an order of magnitude during the simulation, with more than half the decline occurring prior to $z = 1000$. In general, there seems to be a clear distinction between the $B_{\rm rms}$ evolutionary shape for the scale-invariant case and the steeply evolving most blue spectra. Steepening the spectral index among the blue spectra ($k^1, k^2$) only slightly changes their magnetic amplitude evolution and they exhibit a very similar decay shape.
For the velocity field there is no strong dependence on the spectral index of the initial magnetic field at high-$z$ in the drag and transition regimes. All curves show a similar slope and amplitude except the scale-invariant fiducial case where velocities grow slightly faster. Only at lower redshifts, below $z<800$ and in the turbulent regime do the differences become apparent. The larger the initial spectral index $n$ is, the closer one gets to equipartition, defined in our units as equality of $u_\mathrm{rms}$ and $B_\mathrm{rms}$. Also, the runs with steeper initial spectra show more pronounced turbulent decay of velocities.
### Magnetic power spectra
![Evolution of the magnetic power spectra with different spectral index $n$, at four redshifts $z = 3000, 1200, 900\ \&\ 400$, decreasing from top to bottom in each panel. The upper panel is for $B_{z=3000} = 0.51$ while the lower shows the curves for $B_{z=3000} = 0.31$. For $B_{z=3000}=0.31$, a causal initial spectrum, $P_M\propto k^4$, is also simulated and shown.[]{data-label="fig:spectra_comp_index"}](./eps/spectra_index_comparison.pdf "fig:")\
![Evolution of the magnetic power spectra with different spectral index $n$, at four redshifts $z = 3000, 1200, 900\ \&\ 400$, decreasing from top to bottom in each panel. The upper panel is for $B_{z=3000} = 0.51$ while the lower shows the curves for $B_{z=3000} = 0.31$. For $B_{z=3000}=0.31$, a causal initial spectrum, $P_M\propto k^4$, is also simulated and shown.[]{data-label="fig:spectra_comp_index"}](./eps/spectra_index_comparison_b03.pdf "fig:")
![Relative normalization of the magnetic power spectra (shown in Fig. \[fig:spectra\_comp\_index\]) at $k_{1,_\mathrm{phys}} = 280\,(h/0.7)$ Mpc$^{-1}$ as a function of spectral index $n_B = n-2$](./eps/spectrum_normalization_index.pdf)
. \[fig:spectra\_index\_norm\]
We show the magnetic power spectra $P_M(k)/P_M(0)$ in Fig. \[fig:spectra\_comp\_index\] evolving through the same redshifts $z=3000,1200,900\ \&\ 400$ as in Fig. \[fig:mag\_spectra\_k09\] but now for a range of initial spectral indices. These spectra can be contrasted with the evolving spectra already presented for our fiducial case $B_0=0.51$, $n=-0.9$ in Fig \[fig:mag\_spectra\_k09\]. We observe that for $B_0=0.51$ (upper panel of Fig. \[fig:spectra\_comp\_index\]), the spectrum with $P_M\propto k^{-0.5}$ shows an evolution of the spectrum into a turbulent power-law over virtually all its $k$-modes. However, this is not the case for steeper spectral indices. For the $P_M\propto k^{0}$ case, the lowest wavenumber at which the initial spectrum has evolved into a turbulent Kolmogorov-like spectrum $P_M\simeq k^{-5/3}$ is $k\approx \pot{3.0}{3} \, (h/0.7)\ \mathrm{Mpc}^{-1}$. The power spectrum case of $P_M\propto k^1$, the reddest spectrum for which an integral scale $k_I$ can be defined, displays $k_I\simeq \pot{3.8}{4} \, (h/0.7)\ \mathrm{Mpc}^{-1}$ at $z=400$. For the $P_M\propto k^2$ the integral scale at $z=400$ shifts to wavenumber $k_I\approx \pot{4.8}{4} \,(h/0.7)\ \mathrm{Mpc}^{-1}$. All the integral and peak scales for bluer spectra shown in Fig. \[fig:spectra\_comp\_index\] are presented in Table \[tab:spectra\_peaks\_integralscales\].
In Fig. \[fig:spectra\_comp\_index\] (lower panel) we show the magnetic power spectra for the runs with a lower initial magnetic field strength of $B_0=0.31$, where the steepest case of an initial causally-generated power spectrum, $P_M\propto k^4$, is also presented. We observe that this case too develops a turbulent high-$k$ tail for $k>k_I$ with $k_I\approx \pot{6.4}{4} \, (h/0.7)\ \mathrm{Mpc}^{-1}$. The evolution of power spectra presented in Fig. \[fig:spectra\_comp\_index\] clearly demonstrate the establishment of a turbulent power-law over time, whose $k$-range is largest for scale-invariant spectra $k^{-0.9}$ and progressively restricted only to higher $k$-scales as the spectra become more blue to $k^2$ and $k^4$. The blue spectrum simulation runs also exhibit the start of an inverse transfer effect for decaying non-helical turbulence. [@Reppin17] have shown that this can occur for steep spectra if Reynolds numbers are large enough and the integral scale $k_I$ is peaked at sufficiently large $k$, which is the case here [also cf. @Brandenburg2015].
The normalization of magnetic power spectra has been set by relating its integral to the chosen $B_0$. The relative normalization of the magnetic power spectra as a function of spectral index $n_B=n-2$ is shown in Fig. \[fig:spectra\_index\_norm\]. The normalization is computed at the first binned wave vector value of $k_1=1.29$ which corresponds to $k_{1,\mathrm{phys}}= 280\,(h/0.7)$ Mpc$^{-1}$. The dependence of this normalization on the spectral index $n_B$ follows an exponential form and Fig. \[fig:spectra\_index\_norm\] also gives a fit.
Magnetic spectrum integral scale $k_I$ peak scale
--------------------------------------------------------- ---------------------- ----------------
$\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!B_0 =0.51$
$n=1$ $\pot{3.8}{4}$ $\pot{9.0}{3}$
$n=2$ $\pot{4.8}{4}$ $\pot{2.4}{4}$
$\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!B_0 =0.31$
$n=1$ $\pot{4.0}{4}$ $\pot{1.2}{4}$
$n=2$ $\pot{5.1}{4}$ $\pot{2.8}{4}$
$n=4$ $\pot{6.4}{4}$ $\pot{4.6}{4}$
: The integral and peak scale in physical units of $(h/0.7)\ \mathrm{Mpc}^{-1}$, for the more blue magnetic spectra, measured at $z=400$ (spectra shown in Fig. \[fig:spectra\_comp\_index\])[]{data-label="tab:spectra_peaks_integralscales"}
![Comparison of dissipation rate of magnetic energy density when varying the spectral index $P_M k\propto k^n$, with $n=n_B+2$. The curves are normalized to the values at $z=3000$. The spectral index varies from $-0.9$ to $4$ from top curve to bottom. We show combined fits (dotted curves) to the magnetic energy dissipation rates for the drag regime at high-$z$ (Eq. \[eq:drag\_param\]), for the turbulent regime (Eq. \[eq:QbarB\_temp\_sols\]) and for the transition region (Eq. \[eq:QbarB\_transition\_temp\]).[]{data-label="fig:heat_rate_comp_index"}](./eps/heat_rate_index_comparison_loglog_fits.pdf)
![Net heating rate for different spectral indices $n$ and two values of $B_0$ =0.51 and 0.31. The physical normalizations for both cases are presented in Fig. \[fig:index\_scalings\]. The dashed lines show an extrapolation of the rates avoiding the numerical issues which dominate at large $z$ and cause an unphysical rise, especially of the bluer spectra.[]{data-label="fig:net_heat_index_comp"}](./eps/net_heat_rate_index_comparison\logstr.pdf "fig:")\
![Net heating rate for different spectral indices $n$ and two values of $B_0$ =0.51 and 0.31. The physical normalizations for both cases are presented in Fig. \[fig:index\_scalings\]. The dashed lines show an extrapolation of the rates avoiding the numerical issues which dominate at large $z$ and cause an unphysical rise, especially of the bluer spectra.[]{data-label="fig:net_heat_index_comp"}](./eps/net_heat_rate_index_comparison\logstr_b031.pdf "fig:")
![Physical normalization of the peak values (blue) and initial (z$=3000$) values (green) of the magnetic energy density dissipation rate from Fig. \[fig:heat\_rate\_comp\_index\] as well as the net heating rate peak values (red) from Fig. \[fig:net\_heat\_index\_comp\], all as a function of spectral index $n_B$ where $n_B=n-2$.[]{data-label="fig:index_scalings"}](./eps/heat_rate_normalization_index_comparison.pdf)
### Dissipation rates and net heating rate {#sec:total_PMF_loss_n}
We now examine the changes in the magnetic energy dissipation rate and the net heating rate for the simulation runs with varying spectral indices. Fig. \[fig:heat\_rate\_comp\_index\] depicts the rate of change of the magnetic energy density. Note that here we have normalized all curves to their initial value at $z=3000$. The peak redshift is virtually independent of spectral index $n$, tending slightly to earlier epochs for bluer spectra. In the drag dominated regime $z > 1500$, the slope of the energy dissipation rate becomes monotonically steeper with increasing $n$. This has the effect that the magnetic dissipation rate only rises above its initial value for spectra redder than $P_M \propto k^0$. In the turbulent regime, the dissipation rates all decline monotonically, by over a factor of 3 for the near scale-invariant spectrum and by around an order of magnitude for the bluest spectrum. We fit to each of these three regimes as described in the next subsection.
The net heating rate due to magnetic fields for different initial spectral indices $n$ = -0.9, -0.5, 0, 1 $\, \& \,$ 2 is displayed in Fig. \[fig:net\_heat\_index\_comp\]. Two different magnetic field amplitudes $B_0$=0.51 and 0.31 (including the additional spectral case of $n=4$) are shown in the upper and lower panels respectively. The epoch of peak net heating shows a striking shift to higher redshifts as the spectral index varies from $n$=-0.9 to 2 or 4. There is a clear distinction between the scale-invariant case $n=-0.9$, which peaks at a redshift $z\approx 650$, and the net heating for steeper initial spectra, which peak at earlier epochs up to $z \approx 925$. Thus magnetic spectral variation alone can delay the epoch of peak net heating by $\sim$ 0.4 Myr. This effect is even more pronounced as $B_0$ is reduced from 0.51 to 0.31, as expected from the amplitude variation of net heating described earlier in Fig. \[fig:net\_heat\_rate\_comp\_k09\].
Overall, we discern two clear trends for the shape and position of the net heating rate as a function of magnetic field parameters: from earlier, as the field strength grows stronger, the net heating maximum is more sharply peaked in time. The second trend we see here is that as the magnetic spectral index is lowered (to near scale-invariant), the net heating peak is delayed by a greater time interval of up to 0.5 Myr after the epoch of recombination. To compare, this time-lag is greater than the age of the Universe at recombination.
![Fitting parameters for the magnetic energy density dissipation rate with varying spectral index in Fig. \[fig:heat\_rate\_comp\_index\] in the turbulent decay regime. The rates were fitted using Eq. over the range $z\in[100,900]$. A power-law was then fitted to give a functional form of $z_0$ and $\tau_0^{-1}$ in Eq. .[]{data-label="fig:dQBdlnz_comp"}](eps/fitting_pars_dQBdlnz_index.pdf)
### Semi-analytic representation of the PMF losses {#sec:semi-analytics}
In this section, we construct a semi-analytic representation of the total magnetic energy density dissipation rates. We provide separate fits for the turbulent, transition and drag-dominated regimes which we then combine together to plot as the black dotted curves in Fig \[fig:heat\_rate\_comp\_index\].
In the turbulent regime, following the phenomenological derivation presented in Sect. , we use the functional form $$\begin{aligned}
\label{eq:QbarB_temp}
\left.\frac{\id (Q_{B}/\rho_\gamma)}{\id \ln z}\right|_{\rm turb}\approx \frac{r_{B, 0} }{\left[1+\frac{1}{m\tau_0}\ln\left(\frac{1+z_0}{1+z}\right)\right]^{m+1}},\end{aligned}$$ with free parameters $\tau_0$ and $z_0$ as a starting point. We recall that $m=2(n_B+3)/(n_B+5)=2(n+1)/(n+3)$. The late-time behavior of the curves shown in Fig. \[fig:heat\_rate\_comp\_index\] can then be represented by \[eq:QbarB\_temp\_sols\] $$\begin{aligned}
r_{B, 0}&\approx 1
\\
z_0&=532 + 549\,m^{0.56}
\\
\tau^{-1}_0&= 0.07 + 2.91\,m^{1.36}.\end{aligned}$$ Since the denominator in Eq. vanishes at $z_{\rm lim}\simeq z_0\,\expf{\,m\, \tau_0}$, Eq. is only valid at $z<z_{\rm lim}$, where $z_\mathrm{lim} \in [1190\ ; 1560]$, and the turbulent regime lies well within this range. The power-law fits for $z_0$ and $\tau_0^{-1}$ are plotted in Fig \[fig:dQBdlnz\_comp\] as a function of spectral index $n_B$
Next, the transition regime between drag-dominated ($z\gtrsim 1500$) and pure turbulent-decay ($z\lesssim 800$), can be represented using a simple power-law scaling: $$\begin{aligned}
\label{eq:QbarB_transition_temp}
\left.\frac{\id (Q_{B}/\rho_\gamma)}{\id \ln z}\right|_{\rm trans}\approx A\,\left[\frac{1300}{1+z}\right]^\beta.\end{aligned}$$ To represent the rising dissipation rates from the drag regime to their peak values in Fig. \[fig:heat\_rate\_comp\_index\], we find $$\begin{aligned}
\label{eq:QbarB_transition_temp_sols}
A&\approx 1.59 - 1.04\,m^{0.64}\\
\beta&=3.51 - 1.94 \, m\end{aligned}$$ to work very well. For the transition regime the power-law was fitted in a redshift range of $z \in [1500,1200]$.
To approximate the evolution at the high redshift drag-dominated regime, we fit a power-law of the following form $$\left.\frac{\mathrm{d}(Q_B/\rho_\gamma)}{\mathrm{d}\ln z}\right|_\mathrm{drag} = A \left(1+z\right)^\gamma$$ to the magnetic energy dissipation rate in Fig. \[fig:heat\_rate\_comp\_index\] for $z\in [1500,\, 2500]$. This gives the observed slopes shown in Table \[tab:drag\_slopes\]. If we parameterize the slope $\beta$ and the amplitude $A$ as a function of the initial spectral index $m$ we get: \[eq:drag\_param\] $$\begin{aligned}
A &= 1.38\, e^{-8.75\, m} \\
\gamma &= 0.015 + 1.01\, m^{1.23}.\end{aligned}$$ We normalize the curves at $z=3000$. Fig. \[fig:high\_z\_fits\] shows the parameterization of the fits to the high redshift range in Fig. \[fig:heat\_rate\_comp\_index\] .
In the pre-recombination drag dominated regime, the theoretical prediction derived in Sect. \[sec:drag\_analytic\] for the magnetic energy density dissipation rate $\id Q_{B}/\mathrm{dln}\, z \propto \rho_\gamma(z) \,(1+z)^{5(n_B+3)/2(n_B+5)}$ is qualitatively comparable to the trend of the fitted slopes $\gamma$ as a function of spectral index - see Table \[tab:drag\_slopes\] . We note that our numerical simulations do not explicitly match the predicted high-$z$ scaling (cf. Eq. \[eq:B2\_fs\]) because of numerical effects (discussed in Appendix \[sec:appendix1\]) and also due to our assumption of matter-domination (see discussion in Sect. \[sec:pic\_pre\_rec\]).
![Fitting parameters for the high-$z$ drag-dominated magnetic energy dissipation rate in Fig. \[fig:heat\_rate\_comp\_index\], as a function of spectral index. The fit equations for the power-law parameters $A$ and $\gamma$ are given in Eq. []{data-label="fig:high_z_fits"}](./eps/high_z_fitting_pars_index.pdf)
For a combined fit for the dissipation rates for each spectral index over all three regimes, we found that a cubic exponent gives a good approximation for combining the turbulent and transition fits at lower redshifts, $$\begin{aligned}
\label{eq:QbarB_tot}
\frac{\id (Q_{B}/\rho_\gamma)}{\id \ln z}\approx
\begin{cases}
\left[\left.\frac{\id (Q_{B}/\rho_\gamma)}{\id \ln z}\right|_{\rm turb}^{-3}+\left.\frac{\id (Q_{B}/\rho_\gamma)}{\id \ln z}\right|_{\rm trans}^{-3}\right]^{-1/3}& z\lesssim z_{\rm lim}
\\[2mm]
\left.\frac{\id (Q_{B}/\rho_\gamma)}{\id \ln z}\right|_{\rm trans}& z\lesssim 1300
\\[2mm]
{\rm max}\left(\left.\frac{\id (Q_{B}/\rho_\gamma)}{\id \ln z}\right|_{\rm trans},
\left.\frac{\id (Q_{B}/\rho_\gamma)}{\id \ln z}\right|_{\rm drag}\right)& z\gtrsim 1300.
\end{cases}\end{aligned}$$ With this we obtain the fit functions over plotted as the dotted black curves shown in Fig. \[fig:heat\_rate\_comp\_index\].
$n_B$ observed slope fitted $\gamma$ predicted slope
------- ---------------- ----------------- -----------------
-2.9 0.05 0.07 0.12
-2.5 0.39 0.34 0.50
-2.0 0.61 0.63 0.83
-1.0 0.99 1.03 1.25
0.0 1.31 1.28 1.50
: Comparison of the observed (Fig. \[fig:heat\_rate\_comp\_index\]) and fitted slope $\gamma$ (Eq. \[eq:drag\_param\]) versus the predicted slope $5(n_B+3)/2(n_B+5)$ for the dissipation rate of magnetic energy density in the high redshift drag-dominated regime.[]{data-label="tab:drag_slopes"}
Discussion {#sec:discussion}
==========
We discuss below some of the caveats of our computations. There are related to i) resolution and numerical viscosity, ii) baryon density perturbations and shocks for $B_0 > 10^{-2}\,{\rm nG}$, iii) self-gravity and photon perturbations, iv) changes induced to the drag coefficient.
Let us start with resolution issues. Our technique of solving the MHD equations on a grid means that we are limited to about three orders of magnitude in dynamic range in $k$-space with our resolution of $N=1536^3$. However, we can still adequately trace the evolution of the power spectra as well as the individual kinetic and magnetic energy densities. Going to higher resolution will greatly increase the computational costs but it will also increase the range in $k$-space that can be resolved. This would allow us to separate the drag-dominated damping scale more cleanly from the Nyquist scale, hence reducing numerical artifacts.
At the highest redshifts ($z>3000$) of our simulation, the evolution of the energy dissipation rates is still affected by numerical viscosity leading to a ring-in phase of the velocity field and a slight underestimation of the magnetic dissipation rate (since part of the energy branches into numerical heating). However, we performed several tests in this regime, strongly reducing the numerical viscosity. This showed that the PMF evolution is well understood and by the transition epoch, numerical viscosity no longer significantly affects our heating rates. Thus, we are able to reliably deduce the net heating rates that go into the plasma from the decay of the magnetic energy density at $z\simeq 3000$ to $z\simeq 100$.
In our simulation setup we are restricted to sub-sonic Alfvén velocities. This limits us to $B_\mathrm{0,phys}\lesssim \pot{4}{-2} \mathrm{nG}$ for physical values of the magnetic field in our runs. While still physically interesting and instructive, this level of PMFs is more than one order of magnitude below the level currently constrained using CMB data, $B_\mathrm{0,phys} \sim 1\,\mathrm{nG}$ [@Planck2015PMF]. In this regime of stronger magnetic fields, one has to consider significant perturbations in the plasma as well as shocks, which we have ignored in our case of low Mach numbers. We expect this to increase the heating efficiency of the PMFs, with kinetic energy being converted into heat more rapidly. We could expect that the transient regime becomes shorter and the net heating might also evolve accordingly. This will have further implications for structure formation, so a simulation that takes such effects into account will be needed for the future treatment of heating due to stronger magnetic fields.
Stronger magnetic fields could also source significant additional baryon density perturbations which can grow after recombination. While we have ignored these in the present simulation involving relatively weak magnetic fields, they could play a role with stronger magnetic fields: altering the Jeans scale for structure formation [@Subramanian1998; @Sethi2005] as well as potentially constraining the PMF via contribution to the small-scale matter power spectrum and further effects [@Shaw:2010ea; @Jedamzik2013; @Pandey2015; @Jedamzik2018].
Similarly, we neglected self-gravity and photon perturbations in our simulation. In the drag-dominated regime, small-scale photon perturbations will be created. Part of these perturbations could affect the subsequent evolution of the plasma in the turbulent phase. They will also become visible as ultra-small scale CMB anisotropies, well below the photon diffusion scale $k\gtrsim 0.2\,\Mpc^{-1}$. At late stages, we also expect the growth of dark matter halos to influence the dynamics of the fluid flow. This will be particularly important at $z\lesssim 100-200$, when ambipolar diffusion should also become relevant [@Jedamzik1998]. In refined treatments of the problem, these effects should be considered carefully.
In our simulations, we used the pre-computed standard recombination history obtained with [CosmoRec]{} [@Chluba2010b] to determine the drag coefficient. However, significant heating by PMFs can directly affect the recombination history [e.g., @Sethi2005; @Chluba2015PMF], which in turn would modify the drag terms. We expect these effects to only become noticeable for higher PMF strength ($B_0\simeq1 \,{\rm nG}$), when shocks also have to be considered. Therefore, the heating rates obtained here should be relatively robust with respect to this aspect.
Finally, in our simulations we only considered non-helical magnetic fields. However, PMFs may have been generated with substantial helicity [@Vachaspati2001] which would significantly alter their subsequent evolution [@Banerjee2004] as well as produce distinct cosmological signatures [@Pogosian2002; @Planck2015PMF; @Ballardini2015; @Long2015]. The influence of magnetic helicity conservation as well as inverse cascade on the evolution and decay of PMFs in this case would be interesting to explore and is left for future work.
Conclusions {#sec:conclusion}
===========
We performed 3D numerical simulations of PMFs along with baryon velocity fluctuations across the cosmological recombination era. Photon drag was included using the standard recombination history obtained with [CosmoRec]{} and the MHD equations were solved in an expanding medium. Our simulations allow us to trace the flow of energy from the magnetic field, through the baryon velocity field to heating via turbulent decay and dissipation. We are able to describe the net heating rates smoothly across the epoch of recombination. This has enabled a clean separation of real heating, which will lead to an increase of the matter temperature, from drag-dominated magnetic energy losses to the CMB photon field, which only lead to secondary CMB temperature perturbations and spectral distortions.
We supported our computations with analytic estimates and provide several useful expressions to represent our numerical results. We find that at redshifts below $z\lesssim 500$, in the regime of well-developed turbulent decay, our new analytic approximation eq. with an additional $1/m$ factor gives a good fit.
Three main evolutionary stages for magnetic heating are observed (Fig. \[fig:heat\_rate\_k09\]): i) an initial phase that is dominated by photon drag ($z\gtrsim 1500$), ii) an intermediate transition period around cosmological recombination, when the photon drag force drops rapidly, and iii) the final fully turbulent MHD phase ($z\lesssim 600$). Only in the later part of the transition and the turbulent phase do we find significant heating of the medium, a result that is important when deriving and interpreting constraints on PMFs from CMB measurements.
Our computations show that the growth of baryon velocity fluctuations is strongly suppressed at early times because of high photon drag, but builds up during the intermediate transition phase (Fig. \[fig:ts\_comp\_k09\]). After recombination, a velocity saturation state approaching equipartition between magnetic and kinetic energy is reached, after which the plasma dissipates energy in a turbulent cascade. In previous treatments, the build-up of the velocity field and its turbulent decay were assumed to proceed instantaneously or in a prescribed fashion at recombination ($z\simeq 1100$). Our simulations reveal a substantial delay for the onset of turbulence until redshift $z\simeq 600-900$ (this depends on the amplitude and shape of the initial magnetic power spectrum). In the transition phase, the magnetic field decay continuously sources the baryon velocity field until a well developed turbulent state is reached, a process that causes a delay of up to 0.5 Myr.
We also find that the shapes of the net heating rates obtained are broader than previously estimated semi-analytically. For weaker magnetic fields, their peak broadens to extend to significantly lower redshifts ($z\lesssim 400$). In these cases, the fluid flow reaches turbulence more gradually with the magnetic Reynolds number exceeding unity only in the very late stages (Fig. \[fig:reynold\_amplaa\_comp\]). In contrast, steeper spectra ($n \sim 1-4$) produce an earlier peak in the net heating rates at $z_\mathrm{peak}\simeq 900$. The shape, epoch and width of the peak of the net heating rate are seen to depend significantly on the initial amplitude of the magnetic fields and their spectral index (Fig. \[fig:net\_heat\_rate\_comp\_k09\] and Fig. \[fig:net\_heat\_index\_comp\]). Our simulations further reveal that the peak magnetic energy density dissipation rate occurs once the drag coefficient has already dropped appreciably and the fluid approaches its maximum kinetic energy density growth rate (Fig. \[fig:ts\_comp\_k09\]).
For the near scale-invariant case, almost the entire magnetic power spectrum (over two orders of magnitude) is reshaped to a turbulent power-law (Fig. \[fig:mag\_spectra\_k09\]) as a turbulent cascade removes and redistributes power. The turbulent slope obtained is $n \sim -1.4$, slightly shallower than Kolmogorov-type turbulence. On the other hand, we find that only the intermediate-scale wave modes $k>\pot{1.0}{4}\ \mathrm{Mpc}^{-1}$ are reprocessed for spectra with $P_M\propto k^1$ and only the smallest scale modes $k> \pot{5.0}{4}\ \mathrm{Mpc}^{-1}$ for the steepest spectrum $P_M\propto k^4$ (Fig. \[fig:spectra\_comp\_index\]). The direct observation of the spectral transformation from an initial power-law with a cut-off to a spectrum in its transitional phase and then to a processed spectrum with a turbulent slope is a novel feature of our computations and a benefit of the full 3D treatment of the MHD equations.
To be able to achieve a more complete cosmological treatment, more physics at late times would need to be included: below a redshift of $z\lesssim 200$ effects like ambipolar diffusion could dominate and can contribute substantially to the evolution of the magnetic heating history at later epochs. Stronger magnetic fields along with compressible fluid phenomena would extend the scope of these simulations. A simultaneous treatment of density perturbations induced by PMFs would complement the evolution of heating and its cosmological effects. The results presented above represent a first step in simulating magnetic heating in cosmology and it would be interesting to extend them to describe magnetic signatures and constraints arising from the CMB, structure formation, 21-cm signals as well as the epoch of reionization.
Acknowledgments {#acknowledgments .unnumbered}
===============
JC would like to cordially thank Ethan Vishniac and Geoff Vasil for useful discussions about magnetic turbulence. We would also like to thank Jacques Wagstaff for his contribution at an early stage of the project and Stefan Hackstein for comments. PT and JR are supported by the German Deutsche Forschungsgemeinschaft (DFG), Sonderforschungsbereich (SFB) 676 section C9. JC is supported by the Royal Society as a Royal Society University Research Fellow at the University of Manchester, UK.
Numerical viscosity tests {#sec:appendix1}
=========================
![Comparison of magnetic power spectra for varying hyperviscosity $\nu_3$. Using a value $\nu_3< 2\times 10^{-16}$ gives excess power at high $k$. This is called the bottleneck effect which we avoid by chosing $\nu_3=2.5\times 10^{-16}$[]{data-label="fig:heat_rate_visc_comp_inset"}](./eps/visc_comparison_inset.pdf)
![Dependence of the net heating rate for different $\nu_3$ runs. The heating rate goes toward zero for low viscosities. We expect this behavior since the only dissipation at high-$z$ should be from the drag. This curve is used to extrapolate the physical net heating rate for high-$z$ drag-dominated regime.[]{data-label="fig:heat_rate_highz"}](./eps/heat_rate_comparison_nu3_scaled\logstr.pdf)
To achieve a stable simulation over a wide range of redshifts, we tuned the hyperviscosity $\nu_3$ in such a way that the bottleneck effect does not show up at late times ($z<800$). Due to the drag being large at early times the simulation is stabilized and the magnetic field is frozen-in, whereas at late times hyperviscosity is needed in order to have a stable simulation run. In Fig. \[fig:heat\_rate\_visc\_comp\_inset\] we show the magnetic power spectra for $B_0=0.31$ at a redshift of $z=750$ for different hyperviscosity parameters $\nu_3$ ranging from $\nu_3= \pot{1}{-17}$ to $\nu_3=\pot{5}{-16}$. We chose a value of $\nu_3=\pot{2.5}{-16}$ to get a power-law over a large range of $k$ modes. Also, the highest value $\nu_3=\pot{5}{-16}$ already shows an exponential cutoff at $k\approx k_\mathrm{Ny}$, which causes the power spectrum not to be a smooth power-law. We wanted to produce as little dissipation as needed, so the highest value we tested in our runs turned out to be too large.
To ensure the results we obtain in our simulations do not depend on the viscosity parameter $\nu_3$ in our setup, we ran multiple runs with varying runtime parameters for $\nu_3$. In Fig. \[fig:heat\_rate\_highz\] we plot the net heating rate for different $\nu_3$ for the amplitude $B_0=0.31$. At high redshift $z>1000$ we see that there is some dependence of low values of the heating rate on the numerical viscosity. We expect the net heating rate to be very low at high $z$ since we are in the drag dominated regime and all dissipation should be due to the drag force.
Resolution Study
================
We demonstrate the convergence of our simulation with respect to different choices of numerical resolution. For this resolution convergence test the setup with $B_0=0.31$, $\nu_3=\pot{2.5}{-16}$ and a near scale-invariant initial magnetic power spectrum with four resolutions $N=512^3,\ N=768^3,\ N=1024^3,\ \&\ N=1536^3$ were run. In Fig. \[fig:ts\_resolution\] we show the redshift evolution of velocity & magnetic field strength $u_\mathrm{rms}\ \&\ B_\mathrm{rms}$ at late times, between $z=3000$ and $z=100$. We see that the runs converge nicely, with only the two lowest resolution runs with $N=512^3$ and $N=768^3$ showing a slight deviation from our *fiducial* run with $N=1536^3$.
In Fig. \[fig:spec\_resolution\] we show the power spectra of these different runs at early times $z=3000$ as well as late times $z=400$. The initial spectra differ slightly from each other since we start the simulation at $z>3000$ and have a *ring-in* phase that is not shown here. The two runs with highest resolution converge reasonably well. For the runs with lower resolution there is a pile-up of magnetic power at $k\approx k_\mathrm{Ny}$. This is the so-called bottleneck effect and shows that energy is not dissipated fast enough and the viscosity pre-factor $\nu_3$ should be higher, for that particular resolution, to increase dissipation. For our chosen highest resolution of $N=1536^3$, this bottleneck does not occur at a significant level and we obtain a power-law spectrum over all scales.
![Redshift evolution of the magnetic field strength $B_\mathrm{rms}$ and the velocity field $u_\mathrm{rms}$ as a function of the resolution. The high resolution runs converge very well and even for the low resolution there is agreement on the qualitative behaviour.[]{data-label="fig:ts_resolution"}](./eps/timeseries_resolution_comparison_scaled\logstr.pdf)
![Magnetic power spectra at early ($z=3000$) and late ($z=400$) times for simulation runs with different resolution.[]{data-label="fig:spec_resolution"}](./eps/spectra_resolution_comparison.pdf)
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
[^3]: E-mail: [email protected]
[^4]: We denote comoving scales with $\lambda$ and proper scales with $\ell=a \lambda$.
[^5]: Without this effect one has $\eta=\frac{4}{15}\frac{\rho_\gamma}{ \Ne \sigT}$ [@Weinberg1971].
[^6]: We use $R_{\rm e}^2=1$ to readily eliminate $B_0$ from the expression.
[^7]: For this we need to compute $-a^{-4}\id (a^4 \rho_B)/\id t=-\rho_\gamma(z)\id ( \rho_B/\rho_\gamma)/\id t$.
[^8]: The more general expression should read $c_s^2\approx\frac{4}{3}(1-Y_{\rm p})(1+f_{\rm He}+X_{\rm e})\frac{k\,T_{\rm b}}{m_p}$, where $f_{\rm He}\approx 0.079$ for $Y_{\rm p}=0.24$ and $T_{\rm b}\approx T_\gamma\propto a^{-1}$.
[^9]: Assuming helium fraction $Y_{\rm p}=0.24$.
[^10]: The photon energy density is not affected significantly and will be assumed to remain unaltered throughout the evolution.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Let $X=G/\Gamma$ be the quotient of a connected reductive algebraic ${{\mathbf{C}}}$-group $G$ by a finite subgroup $\Gamma$. We describe the topological fundamental group of the homogeneous space $X$, which is nonabelian when $\Gamma$ is nonabelian. Further, we construct an example of a homogeneous space $X$ and an automorphism $\sigma$ of ${{\mathbf{C}}}$ such that the topological fundamental groups of $X$ and of the conjugate variety $\sigma X$ are not isomorphic.
[Résumé.]{} [*Espaces homogènes complexes conjugués avec groupes fondamentaux non isomorphes.*]{} Soit $X=G/\Gamma$ le quotient d’un ${{\mathbf{C}}}$-groupe algébrique réductif connexe $G$ par un sous-groupe fini $\Gamma$. On décrit le groupe fondamental topologique de l’espace homogène $X$, qui est non abélien quand $\Gamma$ est non abélien. Puis on construit un exemple d’espace homogène $X$ et d’automorphisme $\sigma$ de ${{\mathbf{C}}}$ tels que les groupes fondamentaux topologiques de $X$ et de la variété conjuguée $\sigma X$ ne sont pas isomorphes.
address:
- 'Borovoi: Raymond and Beverly Sackler School of Mathematical Sciences, Tel Aviv University, 6997801 Tel Aviv, Israel'
- 'Cornulier: Département de Mathématiques, Bâtiment 425, Université Paris-Sud 11, 91405 Orsay, France'
author:
- Mikhail Borovoi
- Yves Cornulier
date:
- 'May 9, 2015'
-
title: |
[\
]{} Conjugate complex homogeneous spaces\
with non-isomorphic fundamental groups
---
Abridged French version {#abridged-french-version .unnumbered}
=======================
Soit $X$ une variété algébrique pointée définie sur le corps ${{\mathbf{C}}}$ des nombres complexes, supposée irréductible et quasi-projective. L’espace topologique pointé $X({{\mathbf{C}}})$ est alors connexe; on désigne par $\pi_1(X):=\pi_1^\top(X({{\mathbf{C}}}))$ son groupe fondamental, appelé groupe fondamental topologique de $X$. Soit $\sigma$ un automorphisme du corps ${{\mathbf{C}}}$ (pas forcément continu). En appliquant $\sigma$ aux coefficients des polynômes définissant $X$, on obtient une variété $\sigma X$ sur ${{\mathbf{C}}}$, dite variété conjuguée. Les complétés profinis des groupes $\pi_1(X)$ et $\pi_1(\sigma X)$ sont canoniquement isomorphes (comme groupes topologiques), car ils s’identifient naturellement au groupe fondamental étale de $X$. En revanche, les groupes $\pi_1(X)$ et $\pi_1(\sigma X)$ ne sont pas toujours isomorphes, par un résultat de Serre [@Serre]. Les exemples de Serre comprennent des surfaces projectives lisses. D’autres exemples ont été obtenus plus récemment: des variétés de Shimura dans [@MS; @R], et des surfaces projectives dans [@BCG; @GJ] pour des choix très généraux de l’automorphisme $\sigma$ (dans [@GJ] pour tout $\sigma$ dont la restriction à $\overline{{{\mathbf{Q}}}}$ diffère de l’identité et de la conjugaison complexe).
Dans cette note, nous donnons un exemple des [*espaces homogènes*]{} conjugués avec groupes fondamentaux topologiques non isomorphes. Le plan de la note est le suivant. Nous considérons, dans le §\[s:fundam\], les groupes fondamentaux de certains espaces homogènes topologiques de la forme $G/\Gamma$, où $G$ est un groupe de Lie réel connexe et $\Gamma\subset G$ est un sous-groupe discret. Nous en déduisons, dans le §\[s:algebraic\], une formule explicite pour décrire le groupe fondamental $\pi_1(G/\Gamma)$ dans le cas où $G$ est un groupe algébrique linéaire connexe défini sur ${{\mathbf{C}}}$, et $\Gamma$ est un sous-groupe fini de $G$. En utilisant cette formule, nous construisons dans le §\[s:example\] un exemple d’espace homogène affine $X=G/\Gamma$ défini sur ${{\mathbf{C}}}$ et un automorphisme $\sigma$ de ${{\mathbf{C}}}$ tels que les groupes fondamentaux topologiques $\pi_1(\sigma X)$ et $\pi_1(X)$ ne sont pas isomorphes. Précisément, on choisit $G=\SL(n,{{\mathbf{C}}})\times{{\mathbf{C}}}^*$ avec $n\ge 5$, et $\Gamma$ un sous-groupe non abélien fini d’ordre 55. L’inclusion de $\Gamma$ dans $G$ est donnée par un plongement arbitraire de $\Gamma$ dans $\SL(n,{{\mathbf{C}}})$ et par un homomorphisme non trivial de $\Gamma$ dans ${{\mathbf{C}}}^*$. Notre formule permet de vérifier que $\pi_1(X)$ et isomorphe à $({{\mathbf{Z}}}/11{{\mathbf{Z}}})\rtimes_4{{\mathbf{Z}}}$, où la notation signifie que le générateur $1$ de ${{\mathbf{Z}}}$ agit sur ${{\mathbf{Z}}}/11{{\mathbf{Z}}}$ par multiplication par 4, tandis que pour $\sigma$ envoyant $\zeta=\exp 2\pi i/5$ sur $\zeta^2$, le groupe fondamental $\pi_1(\sigma X)$ de la variété conjuguée est isomorphe à $({{\mathbf{Z}}}/11{{\mathbf{Z}}})\rtimes_9{{\mathbf{Z}}}$. Un argument simple permet de vérifier que ces deux groupes ne sont pas isomorphes.
Introduction {#s:Intro}
============
Let $X$ be a pointed algebraic variety defined over ${{\mathbf{C}}}$. We assume that $X$ is irreducible and quasi-projective. The pointed topological space $X({{\mathbf{C}}})$ is then connected, and we denote by $\pi_1(X)$ the topological fundamental group of $X({{\mathbf{C}}})$, i.e., $\pi_1(X):=\pi_1^\top(X({{\mathbf{C}}}))$. Let $\sigma$ be a field automorphism of ${{\mathbf{C}}}$, not necessarily continuous. On applying $\sigma$ to the coefficients of the polynomials defining $X$, we obtain a conjugate algebraic variety $\sigma X$ over ${{\mathbf{C}}}$. Though the profinite completions of $\pi_1(X)$ and $\pi_1(\sigma X)$ are isomorphic, the groups $\pi_1(X)$ and $\pi_1(\sigma X)$ themselves are not necessarily isomorphic. Serre [@Serre] obtained the first examples of conjugate varieties $X$ and $\sigma X$ with $\pi_1(\sigma X)\not\simeq\pi_1(X)$. Serre’s examples include smooth projective surfaces. More examples were obtained recently: Shimura varieties in [@MS] and [@R], and smooth projective surfaces in [@BCG] and [@GJ] for a very general choice of $\sigma$ (in [@GJ] for any $\sigma$ whose restriction to $\overline{{{\mathbf{Q}}}}$ differs from the identity and the complex conjugation).
In this note we give an example of conjugate [*homogeneous spaces*]{} with non-isomorphic topological fundamental groups. The outline of the note is as follows. In Section \[s:fundam\] we consider topological homogenous spaces of the form $G/\Gamma$, where $G$ is a connected real Lie group and $\Gamma\subset G$ is a discrete subgroup. In Section \[s:algebraic\] we write an explicit formula for $\pi_1(G/\Gamma)$ when $G$ is a complex linear algebraic group and $\Gamma\subset G$ is a finite subgroup. In Section \[s:example\] using this formula we construct an example of an affine homogeneous space $X=G/\Gamma$ over ${{\mathbf{C}}}$ and an automorphism $\sigma$ of ${{\mathbf{C}}}$ such that $\pi_1(\sigma X)$ is not isomorphic to $\pi_1(X)$. In our example $G=\SL(n,{{\mathbf{C}}})\times {{\mathbf{C}}}^*$ with $n\ge 5$, and $\Gamma$ is a nonabelian finite subgroup of order 55.
The quotient of a Lie group by a discrete subgroup {#s:fundam}
==================================================
Let $$1\to S{\xrightarrow{\makebox[1.5em]{\scriptsize ${i}$}}} G{\xrightarrow{\makebox[1.5em]{\scriptsize ${\tau}$}}} T\to 1$$ be a short exact sequence of connected real Lie groups. Let $\Gamma\subset G$ be a discrete subgroup such that the projection $\Lambda=\tau(\Gamma)\subset T$ is discrete. Our goal is to describe $\pi_1(G/\Gamma)$, where $G/\Gamma$ is viewed as a pointed manifold with base point the image of 1.
Set $\Gamma_S=\Gamma\cap S$. The homomorphism $\tau\colon G\to T$ induces a fibration $ G/\Gamma\to T/\Lambda$ with fiber $S/\Gamma_S$, which gives rise to an exact sequence in homotopy groups $$\pi_1(S/\Gamma_S){\xrightarrow{\makebox[1.5em]{\scriptsize ${i_{*}}$}}}\pi_1(G/\Gamma){\xrightarrow{\makebox[1.5em]{\scriptsize ${\tau_{*}}$}}} \pi_1(T/\Lambda)\to 1.$$ The fibration $G\to G/\Gamma$ with fiber $\Gamma$ gives rise to an exact sequence in homotopy groups $$1\to\pi_1(G)\to\pi_1(G/\Gamma){\xrightarrow{\makebox[1.5em]{\scriptsize ${f}$}}} \Gamma\to 1,$$ where $f$ is a homomorphism by Lemma \[l:connecting\] below. Considering the above fibrations and also the fibrations $S\to S/\Gamma_S$, $T\to T/\Lambda$ and $G\to T$, we obtain the following commutative diagram of groups and homomorphisms with exact rows and columns: $$\xymatrix@R=18pt{
1\ar[r] &\pi_1(S)\ar[d]\ar[r] &\pi_1(S/\Gamma_S)\ar[d]^(0.44){i_{*}}\ar[r] &\Gamma_S\ar[d]^(0.44){i} \ar[r] &1 \\
1\ar[r] &\pi_1(G)\ar[d]\ar[r] &\pi_1(G/\Gamma)\ar[d]^(0.44){\tau_{*}}\ar[r]^-f &\Gamma\ar[d]^(0.44){\tau}\ar[r] &1 \\
1\ar[r] &\pi_1(T)\ar[d]\ar[r] &\pi_1(T/\Lambda)\ar[d]\ar[r]^-{f_T} &\Lambda\ar[d]\ar[r] &1\\
&1 &1 &1
}$$ From this diagram we obtain homomorphisms $$\chi\colon \pi_1(S)\to\pi_1(S/\Gamma_S){\xrightarrow{\makebox[1.5em]{\scriptsize ${i_*}$}}} \pi_1(G/\Gamma)\quad\text{and}\quad
\phi\colon\pi_1(G/\Gamma)\to \pi_1(T/\Lambda)\underset{\Lambda}{\times} \Gamma,$$ where the fiber product $\pi_1(T/\Lambda)\times_\Lambda\hs \Gamma$ is the group of pairs $(x,\gamma)\in \pi_1(T/\Lambda)\times \Gamma$ such that $f_T(x)=\tau(\gamma)$. The homomorphism $\phi$ takes $y\in\pi_1(G/\Gamma)$ to the pair $(\tau_*(y),f(y))\in \pi_1(T/\Lambda)\times_\Lambda\hs \Gamma$.
\[t:pi-1\] With the above notation, the sequence $$\pi_1(S){\xrightarrow{\makebox[1.5em]{\scriptsize ${\chi}$}}}\pi_1(G/\Gamma){\xrightarrow{\makebox[1.5em]{\scriptsize ${\phi}$}}} \pi_1(T/\Lambda)\underset{\Lambda}\times \Gamma\to 1$$ is exact. In particular, if $S$ is simply connected, then $\phi$ is an isomorphism.
[**Proof.**]{} We prove the theorem by diagram chasing. Clearly $\phi\circ\chi=1$. We show that ${{\rm ker}}\,\phi\subset{{\rm im}}\,\chi$. Let $y\in{{\rm ker}}\,\phi\subset \pi_1(G/\Gamma)$, then $f(y)=1$ and $\tau_*(y)=1$. Then $y$ comes from some element $z\in\pi_1(G)$, whose image in $\pi_1(T)$ is 1. Hence $z$ comes from some element $u\in\pi_1(S)$. We see that $y=\chi(u)$, as required.
We show that $\phi$ is surjective. Let $(x,\gamma)\in \pi_1(T/\Lambda)\times_\Lambda \Gamma$, i.e., $x\in\pi_1(T/\Lambda)$, $\gamma\in\Gamma$, and $f_T(x)=\tau(\gamma)$. We can lift $x$ to some element $y\in\pi_1(G/\Gamma)$, then $\tau(f(y))=\tau(\gamma)$. Set $z=f(y)\gamma^{-1}$, then $\tau(z)=1$, hence $z$ comes from some element of $\Gamma_S$ and from some element $u$ of $\pi_1(S/\Gamma_S)$. Set $y'=i_*(u)^{-1}y\in\pi_1(G/\Gamma)$, then $f(y')=\gamma$ and $\tau_*(y')=\tau_*(y)=x$. We see that $(x,\gamma)=\phi(y')$, as required.
The following lemma, which we used above, is well-known.
\[l:connecting\] Let $G$ be a connected Lie group, $\Gamma\subset G$ be a (closed) Lie subgroup, not necessarily connected. Then the connecting map $f\colon \pi_1(G/\Gamma)\to \pi_0(\Gamma)$ in the exact sequence $$\pi_1(\Gamma)\to\pi_1(G)\to\pi_1(G/\Gamma){\xrightarrow{\makebox[1.5em]{\scriptsize ${f}$}}}\pi_0(\Gamma)\to 1$$ is a homomorphism.
[**Proof.**]{} Denote by $\lambda\colon \Gamma\to\pi_0(\Gamma)$ the canonical epimorphism. Consider two based loops $\theta_i\colon[0,1]\to G/\Gamma$ in $G/\Gamma$ ($i=1,2$). Let $\tiltheta_i\colon [0,1]\to G$ be a path lifting the loop $\theta_i$ to $G$ with $\tiltheta_i(0)=1$, and set $\gamma_i=\tiltheta_i(1)\in \Gamma$. By definition $f([\theta_i])=\lambda(\gamma_i)\in\pi_0(\Gamma)$, where $[\theta_i]$ denotes the class of the based loop $\theta_i$ in $\pi_1(G/\Gamma)$. Then $\gamma_1\tiltheta_2$ is a path in $G$ from $\gamma_1$ to $\gamma_1\gamma_2$ mapping in $G/\Gamma$ to the loop $\theta_2$, hence the concatenation of $\tiltheta_1$ and $\gamma_1\tiltheta_2$ is a path in $G$ from $1$ to $\gamma_1\gamma_2$ mapping in $G/\Gamma$ to the loop obtained by concatenation of $\theta_1$ and $\theta_2$. Thus $f([\theta_1]\cdot[\theta_2])=\lambda(\gamma_1 \gamma_2)=\lambda(\gamma_1)\,\lambda(\gamma_2)=f([\theta_1])\, f([\theta_2])$, as required.
The quotient of a complex algebraic group by a finite subgroup {#s:algebraic}
==============================================================
Let $G$ be a connected linear algebraic group over ${{\mathbf{C}}}$. Let $\Gamma\subset G$ be a finite subgroup. Set $X=G/\Gamma$. We wish to compute the topological fundamental group $\pi_1(X)$.
Let $U$ denote the unipotent radical of $G$, then $G':=G/U$ is reductive. The canonical epimorphism $\rho\colon G\to G'$ induces a fibration $G/\Gamma\to G'/\Gamma'$ with fiber $U$, where $\Gamma'=\rho(\Gamma)$, and hence, the induced homomorphism $\rho_*\colon\pi_1(G/\Gamma)\to\pi_1(G'/\Gamma')$ is an isomorphism. Therefore, we may and shall assume that $G$ is reductive. Replacing the reductive group $G$ by a finite cover and $\Gamma$ by its inverse image, we may and shall assume that the semisimple group $S:=[G,G]$ is simply connected. Let $\Lambda$ denote the image of $\Gamma$ in the algebraic torus $T:=G/S$, then $T/\Lambda$ is also an algebraic torus, hence $\pi_1(T/\Lambda)$ is a free abelian group isomorphic to ${{\mathbf{Z}}}^{{\rm dim}\, T}$. The next corollary, which follows immediately from Theorem \[t:pi-1\], describes $\pi_1(G/\Gamma)$ in terms of $\Gamma$ and the free abelian group $\pi_1(T/\Lambda)$.
\[c:pi-1-alg\] Let $G$ be a connected reductive algebraic group over ${{\mathbf{C}}}$ such that the commutator subgroup $S$ of $G$ is simply connected. Set $T=G/S$. Let $\Gamma\subset G$ be a finite subgroup, and let $\Lambda$ denote the image of $\Gamma$ in $T$. Then there is a canonical isomorphism $$\pi_1(G/\Gamma){\overset{\sim}{\longrightarrow}}\pi_1(T/\Lambda)\underset{\Lambda}\times \Gamma,$$ where $\pi_1(T/\Lambda)\times_{\Lambda}\hs \Gamma$ is the fiber product with respect to the epimorphism $\pi_1(T/\Lambda)\to\Lambda$ of Lemma \[l:connecting\] and the canonical epimorphism $\Gamma\to\Lambda$.
Example {#s:example}
=======
Let $A={{\mathbf{Z}}}/m{{\mathbf{Z}}}$, the additive group of residues modulo $m$. Let $B\subset ({{\mathbf{Z}}}/m{{\mathbf{Z}}})^*$ be a [*cyclic*]{} subgroup of some order $r$ in the multiplicative group of invertible residues modulo $m$. The group $B$ acts naturally on $A$ by multiplication: an element $b\in B\subset({{\mathbf{Z}}}/m{{\mathbf{Z}}})^*$ acts by $a\mapsto ba$. Set $$H=A\rtimes B$$ (the semidirect product). We regard $B$ as a subgroup of $H$. Consider an embedding $\varphi\colon B{\hookrightarrow}{{\mathbf{C}}}^*$, then $\varphi(B)=\mu_r\subset {{\mathbf{C}}}^*$, the group of $r$-th roots of unity.
Choose an embedding $\alpha\colon H{\hookrightarrow}\SL(n,{{\mathbf{C}}})$ for some natural number $n$. Set $$G= \SL(n,{{\mathbf{C}}})\times {{\mathbf{C}}}^*.$$ For $(a,b)\in A\rtimes B=H$ set $$\psi(a,b)=(\alpha(a,b),\varphi(b))\in \SL(n,{{\mathbf{C}}})\times{{\mathbf{C}}}^*.$$ We obtain an embedding $\psi=\psi_{\alpha,\varphi}\colon H{\hookrightarrow}G$. Set $\Gamma=\psi(H),$ $X=X_{\alpha,\varphi}=G/\Gamma$. Then $X$ is an affine algebraic variety over ${{\mathbf{C}}}$.
Let $b\in B$. Write $A\rtimes_b {{\mathbf{Z}}}$ for the semidirect product of $A$ and ${{\mathbf{Z}}}$, where the generator $1$ of ${{\mathbf{Z}}}$ acts on $A$ by multiplication by $b$. Set $\zeta=\exp 2\pi i/r\in \mu_r$.
\[p:pi-1-X\] $\pi_1(X_{\alpha,\varphi})\simeq ({{\mathbf{Z}}}/m{{\mathbf{Z}}})\rtimes_{\varphi^{-1}(\zeta)}{{\mathbf{Z}}}$.
[**Proof.**]{} Set $S=\SL(m,{{\mathbf{C}}})$, $T={{\mathbf{C}}}^*$. Let $\tau\colon G=S\times T\to T$ denote the projection, then $\tau(\psi(a,b))=\varphi(b)$ for $(a,b)\in A\rtimes B =H$. Set $\Lambda=\tau(\Gamma)=\tau(\psi(H))\subset T$, then $\Lambda=\varphi(B)=\mu_r\subset{{\mathbf{C}}}^*=T$.
Consider the following universal covering of $T={{\mathbf{C}}}^*$: $$\ve\colon {{\mathbf{C}}}\to{{\mathbf{C}}}^*=T,\quad z\mapsto \exp 2\pi iz \ \text{ for }\ z\in{{\mathbf{C}}},$$ it induces a universal covering of $T/\Lambda$: $${{\mathbf{C}}}{\xrightarrow{\makebox[1.5em]{\scriptsize ${\ve}$}}}{{\mathbf{C}}}^*\to{{\mathbf{C}}}^*/\mu_r=T/\Lambda\simeq{{\mathbf{C}}}^*.$$ We identify $\pi_1(T/\Lambda)$ with $\ve^{-1}(\mu_r)=\rth{{\mathbf{Z}}}\subset{{\mathbf{C}}}$, then the homomorphism $\pi_1(T/\Lambda)\to\Lambda=\mu_r$ of Lemma \[l:connecting\] is the restriction of $\ve$ to $\rth{{\mathbf{Z}}}$, hence it takes the generator $\rth\in\rth{{\mathbf{Z}}}=\pi_1(T/\Lambda)$ to the element $\ve(\rth)=\zeta\in\mu_r$.
Since $S=\SL(n,{{\mathbf{C}}})$ is simply connected, by Corollary \[c:pi-1-alg\] we have $$\pi_1(X_{\alpha,\varphi})=\pi_1(G/\Gamma)=\pi_1(T/\Lambda)\underset{\Lambda}{\times}\Gamma\simeq\rth{{\mathbf{Z}}}\underset{\mu_r}{\times}H,$$ where the homomorphism $\rth{{\mathbf{Z}}}\to\mu_r$ takes $\rth$ to $\zeta$ and the homomorphism $H\to\mu_r$ takes $(a,b)\in H$ to $\tau(\psi(a,b))=\varphi(b)$. Since $\rth{{\mathbf{Z}}}$ is a free abelian group, the group extension $$1\to \{0\}\times A\to \rth{{\mathbf{Z}}}\underset{\mu_r}{\times}H\to \rth{{\mathbf{Z}}}\to 1$$ splits, hence $\pi_1(X_{\alpha,\varphi})\simeq A\rtimes\rth{{\mathbf{Z}}}$. The action of $\rth{{\mathbf{Z}}}$ on $A$ in this semidirect product decomposition is the canonical action of the quotient group $\rth{{\mathbf{Z}}}$ of $\rth{{\mathbf{Z}}}{\times}_{\mu_r}\hs H$ on the normal abelian subgroup $A$. Since the element $\rth\in\rth{{\mathbf{Z}}}$ has image $\zeta$ in $\mu_r$, which lifts to $\varphi^{-1}(\zeta)\in B\subset H$, we see that $\rth\in\rth{{\mathbf{Z}}}$ lifts to $(\rth,\varphi^{-1}(\zeta))\in \rth{{\mathbf{Z}}}{\times}_{\mu_r}\hs B\subset \rth{{\mathbf{Z}}}{\times}_{\mu_r}\hs H$, hence $\rth$ acts as $\varphi^{-1}(\zeta)$ on $A$. Identifying $\rth{{\mathbf{Z}}}$ with ${{\mathbf{Z}}}$ via $x\mapsto rx$ for $x\in\rth{{\mathbf{Z}}}$, we obtain the assertion of the proposition.
Now let us take $m=11$, then $A={{\mathbf{Z}}}/11{{\mathbf{Z}}}$. We take $B=({{\mathbf{Z}}}/11{{\mathbf{Z}}})^{*2}$, the group of nonzero quadratic residues modulo 11. The group $B$ is a cyclic group of order 5, namely, $B=\{\bar1,\bar4,\bar9,\bar5,\bar3\}$. Then $H=A\rtimes B$ is a finite nonabelian group of order 55. Let $n\ge 5$, then there exists an embedding $\alpha\colon H{\hookrightarrow}\SL(n,{{\mathbf{C}}})$ . For $b\in B$, $b\neq \bar1$, let $\varphi_b$ denote the embedding $B{\hookrightarrow}{{\mathbf{C}}}^*$ taking the generator $b$ of $B$ to $\zeta$, then $\varphi_b^{-1}(\zeta)=b$. We write $X_{\alpha,b}$ for $X_{\alpha,\varphi_b}$. Let $\sigma$ be any field automorphism of ${{\mathbf{C}}}$ taking $\zeta$ to $\zeta^2$. Consider the conjugate variety $\sigma X_{\alpha,b}$.
\[t:neq\] For $A={{\mathbf{Z}}}/11{{\mathbf{Z}}}$, $B=({{\mathbf{Z}}}/11{{\mathbf{Z}}})^{*2}$, $\sigma\in{{\rm Aut}}({{\mathbf{C}}})$ taking $\zeta$ to $\zeta^2$, the groups $\pi_1(X_{\alpha,4})$ and $\pi_1(\sigma X_{\alpha,4})$ are not isomorphic.
[**Proof.**]{} We have $\sigma(\zeta)=\zeta^2$. The homomorphism $\sigma\circ\varphi\colon B\to {{\mathbf{C}}}^*$ takes $\bar4$ to $\sigma(\zeta)=\zeta^2$, hence it takes $\bar4^3=\bar9$ to $(\zeta^2)^3=\zeta$. Thus $\sigma\circ\varphi_{4}=\varphi_{9}$.
For our group $G$ defined over ${{\mathbf{Q}}}$ and for $X=G/\Gamma$, we have $\sigma X= G/\sigma(\Gamma)$, where $\sigma$ acts on $\SL(n,{{\mathbf{C}}})$ and on ${{\mathbf{C}}}^*$ via the action on ${{\mathbf{C}}}$. For an embedding $\varphi\colon B{\hookrightarrow}{{\mathbf{C}}}^*$ we have $$\sigma X_{\alpha,\varphi}=G/(\sigma\circ\psi_{\alpha,\varphi})(H)=
G/\psi_{\sigma\circ\alpha,\sigma\circ\varphi}(H)=X_{\sigma\circ\alpha,\sigma\circ\varphi}\,.$$ We obtain that $$\sigma X_{\alpha,4}=\sigma X_{\alpha,\varphi_{4}}=X_{\sigma\circ\alpha,\sigma\circ\varphi_{4}}=X_{\sigma\circ\alpha,\varphi_{9}}=X_{\sigma\circ\alpha,9}.$$ By Proposition \[p:pi-1-X\] we have $$\pi_1(X_{\alpha,b})\simeq ({{\mathbf{Z}}}/11{{\mathbf{Z}}})\rtimes_b {{\mathbf{Z}}},$$ hence $$\pi_1(X_{\alpha,4})\simeq ({{\mathbf{Z}}}/11{{\mathbf{Z}}})\rtimes_{4}{{\mathbf{Z}}}\quad\text{and}\quad
\pi_1(\sigma X_{\alpha,4})=\pi_1(X_{\sigma\circ\alpha,9})\simeq ({{\mathbf{Z}}}/11{{\mathbf{Z}}})\rtimes_{9}{{\mathbf{Z}}}.$$ Now the theorem follows from the next Lemma \[l:neq\].
\[l:neq\] $({{\mathbf{Z}}}/11{{\mathbf{Z}}})\rtimes_{4}{{\mathbf{Z}}}\not\simeq ({{\mathbf{Z}}}/11{{\mathbf{Z}}})\rtimes_{9}{{\mathbf{Z}}}$.
We first need the following group-theoretic fact.
\[p:Gamma-4-9\] Let $A$ be any group without nonzero homomorphisms into ${{\mathbf{Z}}}$. When $\vk$ is an automorphism of $A$, we write $A\rtimes_\vk{{\mathbf{Z}}}$ for the semidirect product of $A$ and ${{\mathbf{Z}}}$, where the generator $t=1$ of ${{\mathbf{Z}}}$ acts on $A$ by $\vk$. Fix two automorphisms $\vk_1,\vk_2\in{{\rm Aut}}(A)$, and denote by $\vkbar_i$ the image of $\vk_i$ in the group of outer automorphisms $\Out(A)$. If $\vkbar_1$ is conjugate to neither $\vkbar_2$ nor $\vkbar_2^{-1}$ in $\Out(A)$, then the semidirect products $G_1=A\rtimes_{\vk_1} {{\mathbf{Z}}}$ and $G_2=A\rtimes_{\vk_2} {{\mathbf{Z}}}$ are not isomorphic.
[**Proof.**]{} By contraposition, let $\lambda\colon G_1{\overset{\sim}{\longrightarrow}}G_2$ be an isomorphism. Since for each of $i=1,2$, the subgroup $A$ is equal to the kernel of some/any nonzero homomorphism $G_i\to{{\mathbf{Z}}}$, we have $\lambda(A)=A$. Let $\vk\in{{\rm Aut}}(A)$ denote the restriction of $\lambda$ to $A$. For the generator $t\in {{\mathbf{Z}}}\subset G_1$, write $\lambda(t)$ as $a\hs t^e\in G_2$ with $a\in A$ and $e\in{{\mathbf{Z}}}$. Since $t$ generates $G_1/A$, we see that $\lambda(t)$ generates $G_2/A$ and hence $e=\pm 1$. Then for all $a'\in A$, writing $\gamma_a(a')=aa'a^{-1}$ we have $$\vk(\vk_1(a'))=\lambda(ta't^{-1})=a\hs t^{e} \vk(a')\hs t^{-{e}}a^{-1}=\gamma_a(\vk_2^{e}(\vk(a'))),$$ whence $\vk_1=\vk^{-1}\gamma_a\vk_2^{\hs{e}}\hs\vk$. Hence $\vkbar_1=\vkbar^{\,-1}\ \vkbar_2^{\,{e}}\ \vkbar$. Thus $\vkbar_1$ and $\vkbar_2^{\,{e}}$ are conjugate in $\Out(A)$.
[**Proof of Lemma \[l:neq\].**]{} We use Lemma \[p:Gamma-4-9\] when $A={{\mathbf{Z}}}/11{{\mathbf{Z}}}$, in which case ${{\rm Aut}}(A)=\Out(A)$, is abelian and can be identified with $({{\mathbf{Z}}}/11{{\mathbf{Z}}})^*$. Hence the assumption of Lemma \[p:Gamma-4-9\] in this case is just that $\vk_1$ and $\vk_2^{\pm 1}$ are distinct as elements of $({{\mathbf{Z}}}/11{{\mathbf{Z}}})^*$. Here $\vk_1=\bar4$ and $\vk_2=\bar9$. Since modulo $11$ we have $\bar9\neq \bar4$ and $\bar9^{-1}=\bar5\neq \bar4$, Lemma \[p:Gamma-4-9\] applies and we see that $({{\mathbf{Z}}}/11{{\mathbf{Z}}})\rtimes_{4}{{\mathbf{Z}}}\not\simeq ({{\mathbf{Z}}}/11{{\mathbf{Z}}})\rtimes_{9}{{\mathbf{Z}}}$. This completes the proofs of Lemma \[l:neq\] and Theorem \[t:neq\].
[Acknowledgements.]{} We are grateful to Boris Kunyavskiĭ for helpful comments. This note was completed during a stay of the first-named author at the Max-Planck-Institut für Mathematik, Bonn, and he is grateful to this institute for hospitality, support and excellent working conditions.
[BCG]{}
I. Bauer, F. Catanese, and F. Grunewald, [*Faithful actions of the absolute Galois group on connected components of moduli spaces,*]{} Invent. Math. [**199**]{} (2015), no. 3, 859–888.
G. González-Diez and A. Jaikin-Zapirain, [*The absolute Galois group acts faithfully on regular dessins and on Beauville surfaces,*]{} Preprint, <http://www.uam.es/personal_pdi/ciencias/ajaikin/preprints/galoisactionshort.pdf>.
J. S. Milne and J. Suh, [*Nonhomeomorphic conjugates of connected Shimura varieties,*]{} Amer. J. Math. [**132**]{}(3) (2010), 731–750.
C. S. Rajan, [*An example of non-homeomorphic conjugate varieties,*]{} Math. Res. Lett. [**18**]{} (2011), 937–943.
J.-P. Serre, [*Exemples de variétés projectives conjuguées non homéomorphes,*]{} C. R. Acad. Sci. Paris [**258**]{} (1964), 4194–4196.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Doubly charged scalars are common figures in several beyond the Standard Model studies, especially those related to neutrino masses. In this work, we estimate the High-Luminosity (HL-LHC) and High-Energy LHC (HE-LHC) sensitivity to doubly charged scalars assuming they decay promptly and exclusively into charged leptons. Our study focuses on the fit to the same-sign dilepton mass spectra and it is based on proton-proton collisions at $13$ TeV, $14$ TeV and $27$ TeV with integrated luminosity of $\mathcal{L}=139 fb^{-1}, 3 $ab$^{-1}$ and $15$ab$^{-1}$. We find that HL-LHC may probe doubly charged scalars masses up to $2.3$ TeV, whereas HE-LHC can impressively probe masses up to $3$ TeV, conclusively constituting a complementary and important probe to signs of doubly charged scalars in lepton flavor violation decays and lepton-lepton colliders.'
author:
- 'Tessio B. de Melo$^{1,2}$'
- 'Farinaldo S. Queiroz$^1$'
- Yoxara Villamizar$^1$
bibliography:
- 'ref.bib'
title: 'Doubly Charged Scalar at the High-Luminosity and High-Energy LHC'
---
Introduction
============
Doubly charged scalars are quite popular because they are present in many beyond the Standard Model studies such as the type II seesaw mechanism, which adds a $SU(2)$ scalar triplet to the SM spectrum and represents one of the most common ways to explain neutrino masses. Such scalar triplet features a doubly charged scalar[@Muhlleitner:2003me; @Akeroyd:2005gt; @Hektor:2007uu; @Perez:2008zc; @Chaudhuri:2013xoa; @Lindner:2016bgg; @Primulando:2019evb] which is entitled to interesting phenomenological signatures in the context of lepton flavor violation. Moreover, many models in the literature based on the Left-Right [@Pati:1974yy; @Mohapatra:1974hk; @Senjanovic:1975rk; @Dutta:2014dba; @Dev:2016vle; @Borah:2016hqn; @Dev:2018kpa] and 3-3-1 symmetries [@CiezaMontalvo:2006zt; @Alves:2011kc; @Alves:2012yp; @Machado:2018sfh; @CarcamoHernandez:2019vih; @Ferreira:2019qpf] also have this doubly charged scalar in their spectra. Other extended sectors such as Higgs triplets [@Gunion:1989ci; @Akeroyd:2012ms; @deMedeirosVarzielas:2017glw; @Ghosh:2018jpa; @Chala:2018opy; @Chabab:2018ert; @Chakraborty:2019uxk], little higgs [@ArkaniHamed:2002qx; @Hektor:2007uu], Georgi-Machacek model [@Georgi:1985nv; @Sun:2017mue; @Chiang:2018cgb], and other alternative attempts to explain neutrino masses commonly advocate for the existence of doubly charged scalars [@Zee:1985id; @Nebot:2007bc; @Camargo:2018uzw]. In summary, doubly charged scalars are indeed present in a wealth of beyond the Standard Model studies.
For this reason, several collider searches have been conducted for such particles [@Mitra:2016wpr; @Du:2018eaw], specially at the LHC [@Huitu:1996su; @Han:2007bk; @Chao:2008mq; @Akeroyd:2010ip; @Chiang:2012dk; @delAguila:2013mia; @Kanemura:2013vxa; @King:2014uha; @kang:2014jia; @Kanemura:2014ipa; @Bambhaniya:2015wna]. The latter relies on the search for two prompt, isolated, and highly energetic same sign leptons. Such signal is rare in the Standard Model but it may happen with a higher rate in the aforementioned models due to the presence of the doubly charged scalar particle. The invariant mass of the same sign leptons is a good handle to discriminate potential SM background events from those stemming from the doubly charged scalar prompt decay.
No positive signal has been observed, which led to the derivation of stringent limits with $95\%$ confidence level on the mass of the doubly charged scalar. These bounds depend on the nature of the doubly charged interactions and decays. These scalars can couple to left-handed and right-handed fermions, as occurs in left-right models. The production cross-section when couplings to left-handed leptons are involved is roughly two times larger because of the interactions involving the Z boson. We highlight that even in such models the doubly charged scalar decays mostly into charged leptons as long as the vacuum expectation value of the scalar triplet is sufficiently small. A similar conclusion applies to the canonical type II seesaw model [@Perez:2008ha]. In 3-3-1 models, for instance, such doubly charged scalars decay nearly exclusively into charged leptons as well. Therefore, taking the branching ratio into charged lepton to be nearly equal to unity seems a good assumption. That said, both CMS and ATLAS collaborations have performed their search for doubly charged scalars based on the pair production of the doubly charged scalar via the resonant production (see Fig.\[diagram1\]) [@CMS:2012kua; @ATLAS:2012hi; @Chatrchyan:2012ya; @ATLAS:2014kca; @ATLAS:2016pbt; @CMS:2016cpz; @CMS:2017pet; @Aaboud:2017qph] assuming the scalar decays exclusively into charged leptons. Different setups where the doubly charged scalar decays into additional particles can be easily accounted for by rescaling the bounds. An additional assumption made is that the decay width is small compared to the detector resolution, in other words, the narrow width approximation is applied. It is important to note that this is not always valid because, in models where the doubly charged scalar appears as a singlet field under $SU(2)_L$, the width can be large. Moreover, in photon initiated processes which lead to the pair production of doubly charged scalars mediated by a doubly charged scalar in the t-channel can be also important especially in the large width case [@Babu:2016rcr; @Crivellin:2018ahj]. The inclusion of such effect might improve the LHC sensitivity to doubly charged scalars, but on the other side includes model dependent parameters. In this work, we will adopt the approach done by ATLAS and CMS collaboration which will render our study conservative.
Motivated by the LHC upgrade to High-Luminosity and High-Energy modes we obtain the sensitivity to doubly charged scalars over a wide range of decay modes. Assuming the branching ratio into charged leptons to be 100%, $BR(H^{\pm\pm}\rightarrow l^{\pm}l^{\pm})=100\%$, we vary the branching ratio into individual charged lepton flavors from 0% to 100% to assess the impact on the lower mass bounds. In agreement with [@Li:2018jns] we notice that as we lower the branching ratio into $ee$ pairs the lower mass bound improves because of the misidentification effect which is quite relevant for electrons with high transverse momentum [@Mondal:2016czu; @Queiroz:2016qmc]. In particular, we consider the scenarios with $13$ TeV, $14$ TeV and $27$ TeV center of mass energy and several integrated luminosity setups as follows:
- $E_{cm} =13$ TeV, $\mathcal{L}=139$ fb$^{-1}$
- $E_{cm} =14$ TeV, $\mathcal{L}=3$ ab$^{-1}$
- $E_{cm} =14$ TeV,$\mathcal{L}=15$ ab$^{-1}$
- $E_{cm} =27$ TeV, $\mathcal{L}=3$ ab$^{-1}$
- $E_{cm} =27$ TeV, $\mathcal{L}=15$ ab$^{-1}$
We obtain lower mass bounds for fifty different decay configurations for each collider setup outlined above. In summary, we aim with this work to provide HL-LHC and HE-LHC reach to the doubly charged scalars covering several different decays modes. We start in section II discussing the data set and our model assumptions. In section III we present our bounds and later draw our conclusions.
![Feynman diagram of the resonant pair production process of the doubly charged scalar decaying into charged lepton pairs. There are other processes that contribute to the doubly charged scalar production not shown in the figure.[]{data-label="diagram1"}](sara.pdf)
Data set
========
Before forecasting the HL-LHC and HE-LHC sensitivity we attempted to reproduce the CMS and ATLAS results for a center of energy of $13$ TeV with an integrated luminosity of $\mathcal{L}=36.1$ fb$^{-1}$ [@Aaboud:2017qph]. We followed the recipe described in [@Aaboud:2017qph] and required each electron in the dilepton events to have at least $30$ GeV of transfer momentum ($p_T$). We applied a similar cut for dimuon events. The same logic applies to events having both flavors. Concerning the angular distribution, only electron candidates with a pseudo-rapidity $|\eta| < 2.47$ are selected because outside of this region the calorimeter features poor reconstruction. Muon candidates identified in the muon spectrometer are required to have $|\eta| < 2.5$ [@Aad:2016jkr].
![Production cross section as a function of the doubly charged scalar mass assuming $BR (H^{--}\rightarrow ee)=30\%$,$BR (H^{--}\rightarrow \mu\mu)=40\%$ and $BR (H^{--}\rightarrow \mu e)=30\%$. We overlaid the current and projected ATLAS limits with 13TeV center of mass energy with $36.1fb^{-1}$ and $139fb^{-1}$ of integrated luminosity.[]{data-label="fig:my_label1"}](newbranching.pdf)
The usual isolation criteria for dielectron and dimuon events take place [@Patra:2015bga; @Lindner:2016lpp] to suppress the background coming from misidentified electrons as jets, which is prominent in the high $p_T$ regime. Moreover, jets within a cone of size $\sqrt{(\Delta \eta)^2 + (\Delta\phi)^2}=0.2$ are rejected, where $\phi$ is the azimuthal angle around the beam line. It is known that the use of the inner tracking detector and muon spectrometer help to reduce the muon charge misidentification to a negligible level, thus the muon misidentification is not an issue in this type of study. The electron charge misidentification rate is more challenging though.
Electron charge misidentification mostly results from bremsstrahlung. The emitted photon might convert into an electron-positron pair toughening the charge reconstruction of the primary electron. Moreover, the emitted photon may, in some cases, travel through the inner detector without leaving a track. In the latter case, the electron charge, which is determined from its curvature track, is harder to determine especially for high-energy electrons which approximately travel straight, i.e making the measurement of the curvature more difficult. In the end, the charge misidentification rate induces roughly a 10% systematic error in the analysis for electrons with $p_T > 100$ GeV [@ATLAS:2014kca]. The main backgrounds in this search are classified into three classes: prompt, misidentified leptons and non-prompt. The prompt background arises from SM processes producing two same-sign leptons from $ZZ$, $WZ$, and $W^{\pm} W^{\pm}$ production, for instance. Opposite-sign lepton pairs produced from $W^{\pm} W^{\mp}$ also produce important background when the charge of one of the leptons is misidentified as aforementioned. The non-prompt background refers to jets misidentified as electrons and semileptonic decays of hadrons. In summary, one can clearly notice that the background sources do depend on the final state of interest.
This is important to have in mind because the doubly charged Higgs decay width is found to be, $$\Gamma (H^{\pm \pm}\rightarrow l^{\pm}l^{'\pm})= \frac{k\, h_{l l'}}{16 \pi} m_{H^{\pm \pm}},$$where $k=1$ for $l\neq l'$ and $k=2$ for $l= l'$.
Therefore, the couplings $h_{l l'}$ dictate the branching ratio into charged leptons and consequently the relevant final states. $h_{l l'}$ is assumed to be large enough (greater than $10^{-6}$) to ensure prompt decays. It is known that if this coupling is very small, displaced vertex searches become very important [@Han:2007bk; @Dev:2018kpa; @Antusch:2018svb].
We emphasize that the decay into gauge bosons is assumed to be absent. In a more complete approach, this can be easily justified by taking the vacuum expectation value of the scalar triplet that gives rise to the doubly charged scalar to be sufficiently small. Moreover, in our analysis we will consider only decays into electron and muon final states motivated by the ATLAS and CMS collaborations approach. This assumption is very simplistic and deviations from it might induce a sizeable difference in the overall lower mass bounds. As highlighted earlier, the focus of this work is to project CMS and ATLAS sensitivity to such doubly charged scalars keeping the same assumption adopted by the collaborations.
That said, the relevant channel in our reasoning is the resonant pair production of doubly charged scalars as displayed in Fig.\[diagram1\]. As we are carrying out our study in a model independent way, we will assume similarly to ATLAS and CMS collaboration that the decay width of the doubly charged scalar is small compared to the detector resolution. In other words, we will be working in the narrow width approximation. Besides the channel shown in Fig.\[diagram1\] one could also have pair production of doubly charged scalars with photon-photon scattering where a doubly charged scalar would appear in the t-channel. However, in this case, the width of the doubly charged scalar becomes important and model dependent. We refer the reader to [@Babu:2016rcr; @Crivellin:2018ahj] where the assessment of this channel was carefully investigated. In the end, the inclusion of this production channel improves the overall sensitivity to doubly charged scalars. In our work, we are mainly focused on the High-Luminosity and High-Energy LHC sensitivity to doubly charged scalars without considering a particular model and therefore we will not include these effects here.
Concerning the pair production of doubly charged scalars decaying into $WW$ pairs which is present in the higgs triplet mode, this channel is suppressed by $v_{H}/M_W^2$, where $v_H$ is the vacuum expectation value of the scalar triplet. For sufficiently small values of the $v_H$ this channels is not relevant. There is another source of doubly charged scalar production which happens via its associated production with $H^-$, $q\bar{q'} \rightarrow H^{++}H^{-}$. This channels it is not particularly clean since some decays go into neutrinos, missing energy. In summary, for these reasons we will focus on the pair production of $H^{++}H^{--}$ since it gives rise to a clean same sign dilepton signal.
Having discussed the signal and the simplifying assumptions adopted in our study we present the bounds in the next section.
![95% C.L. lower mass bounds on the doubly charged scalar assuming BR $(H^{\pm\pm}\rightarrow l^{\pm}l^{\pm})=100\%$) for center of mass energy of $13$ TeV and $139\,fb^{-1}$ of integrated luminosity.[]{data-label="fig:my_label3"}](13TeV139fbv1.pdf)
![95% C.L. lower mass bounds on the doubly charged scalar assuming BR $(H^{\pm\pm}\rightarrow l^{\pm}l^{\pm})=100\%$) for center of mass energy of $14$ TeV and $3\,ab^{-1}$ of integrated luminosity.[]{data-label="fig:my_label4"}](14TeV3abv1.pdf)
![95% C.L. lower mass bounds on the doubly charged scalar assuming BR $(H^{\pm\pm}\rightarrow l^{\pm}l^{\pm})=100\%$) for center of mass energy of $14$ TeV and $15\,ab^{-1}$ of integrated luminosity.[]{data-label="fig:my_label5"}](14TeV15abv1.pdf)
![95% C.L. lower mass bounds on the doubly charged scalar assuming BR $(H^{\pm\pm}\rightarrow l^{\pm}l^{\pm})=100\%$) for center of mass energy of $27$ TeV and $3\,ab^{-1}$ of integrated luminosity.[]{data-label="fig:my_label6"}](27TeV3abv1.pdf)
![95% C.L. lower mass bounds on the doubly charged scalar assuming BR $(H^{\pm\pm}\rightarrow l^{\pm}l^{\pm})=100\%$) for center of mass energy of $27$ TeV and $15\, ab^{-1}$ of integrated luminosity.[]{data-label="fig:my_label7"}](27TeV15abv1.pdf)
Bounds
======
ATLAS performed a resonant production search for doubly charged scalars, based on the diagram shown in Fig.\[diagram1\] assuming $BR (H^{\pm \pm} \rightarrow e^{\pm}e^{\pm})+BR (H^{\pm \pm} \rightarrow e^{\pm}\mu^{\pm})+BR (H^{\pm \pm} \rightarrow \mu^{\pm}\mu^{\pm})=100\%$ taking into account all the relevant background discussed earlier and reported null results in agreement with CMS collaboration [@CMS:2017pet]. As a result, an upper limit on the production cross section was derived for different decay modes. In Fig.\[fig:my\_label1\] we exhibit the 95% C.L. limit on the production cross section derived by ATLAS using $36.1 fb^{-1}$ of data based on $13$ TeV proton-proton collisions. We computed the production cross section of the doubly charged scalar using MadGraph5 [@Alwall:2011uj] and included hadronization and detector effects via Pythia [@Sjostrand:2007gs] and Delphes [@deFavereau:2013fsa] interfaces considering all the cuts described previously to impose $m_{H^{\pm\pm}} > 840$ GeV, as one can see in Fig.\[fig:my\_label1\]. Our result is in excellent agreement with ATLAS collaboration.
In order to forecast lower mass bounds for new collider configurations, one would need to scale the backgrounds and then find a new point that yields the same number of background events, assuming the same detector efficiency and acceptance. As we are dealing with dilepton events, the number of signal and background events scale equally with energy and luminosity allowing us to forecast the future collider sensitivities. In other words, the upper limit obtained on the number of signal events at each mass point is derived as a function of the number of background events. This is valid if the signal acceptance and efficiency are nearly independent of the resonance mass and center of mass energy. We emphasize that our assumptions are only reasonable for resonance searches, which do not need to rely on shape analysis. For this reason, we can use the code described in [@Thamm:2015zwa] to obtain future proton-proton collider sensitivities. In simple terms, we need to solve an equation of the type $N(m_{H^{\pm\pm}}^2,E_{new},\mathcal{L}_{new})/N({\rm bound},13TeV,36.1 fb^{1})=1$, where $N$ is the number of signal events for a given parton distribution function which we assumed to be the NNPDF23-NNLO [@Ball:2014uwa].
Firstly, we carried out this procedure to obtain the LHC sensitivity with $139 fb^{-1}$ of integrated luminosity, which should be out in the next public LHC data release. We assumed that the doubly charged scalar decays $30\%$ into $ee$, $40\%$ into $\mu\mu$ and $30\%$ into $e\mu$ and derived the lower mass bound as exhibited in Fig.\[fig:my\_label1\]. We repeated this exercise for a multitude of possibilities varying the branching into $ee$, $\mu\mu$ and $e\mu$ as shown in Fig.\[fig:my\_label3\].
Looking at Fig.\[fig:my\_label3\] one can conclude that the larger the branching ratio into $ee$ the weaker the bound. This has to do partially with the effects we described earlier such as the large electron misidentification rate which has shown to be important in other collider sensitivity studies [@Lindner:2016lxq; @Queiroz:2016qmc]. In particular, in case of null results one might impose $m_{H^{\pm\pm}}>1.2$ TeV assuming the doubly charged scalar decaying entirely into $e\mu$ pairs, i.e. with $BR(H^{--}\rightarrow ee)=0$ and $BR(H^{--}\rightarrow \mu\mu)=0$. Although, notice that if we keep $BR(H^{--}\rightarrow ee)=0$ and increase the branching ratio into muons the limits weakens. Therefore, our simple interpretation of the observed limits is not entirely related to the electron misidentification rate. If one cares to compare the expected limits and the observed limits obtained by ATLAS collaboration [@Aaboud:2017qph] one will clearly conclude that electron misidentification is the driving force behind the weakening of the bounds as branching ratio into $ee$ increases. If we keep $BR(H^{--}\rightarrow ee)=0\%$ and start from $BR(H^{--}\rightarrow \mu\mu)=100\%$ we can notice that as we decrease the branching ratio into $\mu\mu$ the bound strenthens and this is because we are inscreasing the branching ratio into the lepton flavor violation channel which suffers from a relatively small Standard Model background. Therefore, the combinantion of electron misidentification rate with the decay into lepton flavor violating decays helps us understand the lower mass bounds in the figures.
Now considering the High-Luminosity LHC (HL-LHC) setup, which is based on $14$ TeV center of mass energy and $\mathcal{L}=3 ab^{-1}$ we derived Fig.\[fig:my\_label4\]. The HL-LHC can probe masses up to $2.3$ TeV going down to $2.1$ TeV when the branching ratio into $ee$ is around $100\%$. We also derived the HL-LHC sensitivity when the luminosity is ramped up to $\mathcal{L}=15 ab^{-1}$ exhibited in Fig.\[fig:my\_label5\]. The High-Energy LHC (HE-LHC) reach was chosen to be the $27$ TeV center of mass energy proton-proton collision with $\mathcal{L}=15 ab^{-1}$, according to the report [@CidVidal:2018eel], but for completeness we also found the HE-LHC reach for $\mathcal{L}=3 ab^{-1}$.
Looking at Fig.\[fig:my\_label6\]-\[fig:my\_label7\] we see that HE-LHC will be able to explore doubly charged scalars with masses of $3.1$ TeV with $\mathcal{L}=3 ab^{-1}$, all the way up to $4.6$ TeV with $\mathcal{L}=3 ab^{-1}$. These projected bounds certainly constitute the strongest ones on the doubly charged scalar mass surpassing those coming from lepton flavor violation searches [@Lindner:2016bgg; @Dev:2018sel; @Ferreira:2019qpf], particularly in the case where the coupling constant involving $e\mu$ is negligible.
We highlight that if doubly charged scalar has sizeable couplings to $e\mu$, searches for lepton flavor violating muon decays lead to stronger bounds on the doubly charged scalar mass [@Lindner:2016bgg]. Collider bounds are more relevant otherwise.
Having in mind that other future collider projections aim at probing doubly charged scalars up to $2$ TeV [@Dev:2019hev] we conclude that in the absence of lepton flavor violation decays HE-LHC will give rise to the most stringent lower mass bounds in case of null results.
Photon-Photon Fusion
====================
In this section we will present our forecast for the HL-LHC and HE-LHC for the case where photon-photon fusion processes are included in the analysis. In this case there are many diagrams that contribute. Adding them we inscrease the production cross section up to roughly 60% [@Babu:2016rcr]. We will keep assuming that only ($e^{\pm} e^{\pm},\mu^{\pm} \mu^{\pm}$, and $e^{\pm}\mu^{\pm}$) might appear in the final states, though. This time we will focus on a few benchmark cases. Our lower mass bounds when the Drell-Yann and photon initiated processes are summarized in the table I.
-------------------------------------- ---------- ----------
Scenario HL-LHC HE-LHC
BR$(H^{--}\rightarrow ee)=100\%$ 3.13 TeV 4.84 TeV
BR$(H^{--}\rightarrow e\mu)=100\%$ 3 TeV 4.72 TeV
BR$(H^{--}\rightarrow \mu\mu)=100\%$ 3.1 TeV 4.79 TeV
-------------------------------------- ---------- ----------
: The HL-LHC and HE-LHC sensitivity to doubly charged scalar when the Drell-Yann and photon-photon fusion processes are included. HL-LHC and HE-LHC refer to the LHC running at $14$ TeV and $27$ TeV with $\mathcal{L}=15 ab^{-1}$, respectively.
One can easily conclude tha the inclusion of photon-photon mediated processes significantly improve the overall lower mass bounds. For instance, in the HE-LHC setup for $BR(H^{--}\rightarrow ee)=100\%$ the limits went from $4.265$ TeV to $4.84$ TeV. One should keep in mind that at those energies the proton PDF brings a significant uncertainty to the photon-photon mediated fusion processes, thus this bounds should be taken with care.
Comparison with future Colliders
================================
In this section we compare our projected lower mass limits with the existing ones stemming from future colliders [@Dev:2018upe]. At the International Linear Collider, there are two processes that lead to stringent collider bounds on the doubly charged scalar. The on-shell pair production of doubly charged scalars will take place for $m_{H^{\pm\pm}} <500$ GeV, so one can automatically conclude that HL-LHC and HE-LHC constitues a better probe, leaving to ILC the role for more precise measurements [@Nomura:2017abh]. The off-shell production occurs via u-channel exchange of the doubly charged scalar [@Swartz:1989qz]. In this case the constraints on the doubly charge scalar mass depend very much in the yukawa couplings. Anyway, the overall ILC sensitivity is not comparable to the HE-LHC. In other words, HL-LHC and HE-LHC still represent more promissing probes for such scalars at the TeV scale [@Nomura:2017abh]. Regarding the 100TeV collider, the situation changes for obvious reasons. The great improvement in the center-of-mass energy allows this collider to probe doubly charged scalars up to $10$ TeV, and even claim discovery for masses below $4.5$ TeV [@Du:2018eaw], something not feasible within the HE-LHC.
Conclusions
===========
Doubly charged scalars are present in several new physics models such as Higgs triplets, type II seesaw mechanism, left-right and 3-3-1 models. They play an important role in neutrino masses and lepton flavor violation observables. Motivated by their overall importance we obtained collider bounds under reasonable decay assumptions, similarly adopted by ATLAS collaboration. We started forecasting the LHC reach to doubly charged scalars for an integrated luminosity of $\mathcal{L}=139fb^{-1}$ varying the branching ratio into charged leptons. Moreover, we obtained the High-Luminosity and High-Energy LHC sensitivity to doubly charged scalars running with $3 ab^{-1}$ and $15ab^{-1}$ of integrated luminosity. As we varied the branching ratio into charged leptons for each collider configuration studied, in the end, we ended up covering fifty possible decay channels and consequently deriving fifty lower mass bounds in each collider setup considered. In particular, we showed that the HE-LHC will be able to probe doubly charged scalar masses up to $\sim 4.33$ TeV when it decays exclusively into ee and $\mu\mu$ final states, and masses of $4.66$ TeV when it decays entirely into $e\mu$. The latter case of flavor changing decays is complementary to the one stemming from lepton flavor violation. Anyway, it is clear that the HE-LHC will constitute a powerful probe for doubly charged scalars.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors thank Diego Restrepo and Clarissa Siqueira for discussions. FSQ acknowledges support from CNPq grants 303817/2018-6 and 421952/2018-0, UFRN, MEC and ICTP-SAIFR FAPESP grant 2016/01343-7. TM and YV acknowledge CAPES fellowship.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Contour dynamics is a classical subject both in physics and in complex analysis. We show that the dynamics provided by the Löwner-Kufarev ODE and PDE possesses a rigid algebraic structure given by the Virasoro algebra. Namely, the ‘positive’ Virasoro generators span the holomorphic part of the complexified vector bundle over the space of univalent functions, smooth on the boundary. In the covariant formulation they are conserved by the Löwner-Kufarev evolution. The ‘negative’ Virasoro generators span the antiholomorphic part. They contain a conserved term and we give an iterative method to obtain them based on the Poisson structure of the Löwner-Kufarev evolution. The Löwner-Kufarev PDE provides a distribution of the tangent bundle of non-normalized univalent functions, which forms the tangent bundle of normalized ones. It also gives an explicit correspondence between the latter bundle and the holomorphic eigen space of the complexified Lie algebra of vector fields on the unit circle. Finally, we give Hamiltonian and Lagrangian formulations of the motion within the coefficient body in the field of an elliptic operator constructed by means of Virasoro generators. We also discuss relations between CFT and SLE.'
address: 'Department of Mathematics, University of Bergen, Johannes Brunsgate 12, Bergen 5008, Norway'
author:
- 'Irina Markina and Alexander Vasil’ev'
title: 'Virasoro Algebra in Löwner-Kufarev contour dynamics'
---
[^1]
Introduction
============
The challenge of structural understanding of non-equilibrium interface dynamics has become increasingly important in mathematics and physics. Dynamical interfacial properties, such as fluctuations, nucleation and aggregation, mass and charge transport, are often very complex. There exists no single theory or model that can predict all such properties. Many physical processes, as well as complex dynamical systems, iterations and construction of Lie semigroups with respect to the composition operation, lead to the study of growing systems of plane domains. Recently, it has become clear that one-parameter expanding evolution families of simply connected domains in the complex plane in some special models has been governed by infinite systems of evolution parameters, conservation laws. This phenomenon reveals a bridge between a non-linear evolution of complex shapes emerged in physical problems, dissipative in most of the cases, and exactly solvable models. A sample problem is the Laplacian growth, in which the harmonic (Richardson’s) moments are conserved under the evolution, see e.g., [@Mineev; @Vas]. The infinite number of evolution parameters reflects the infinite number of degrees of freedom of the system, and clearly suggests to apply field theory methods as a natural tool of study. The Virasoro algebra provides a structural background in most of field theories, and it is not surprising that it appears in soliton-like problems, e.g., KdV or Toda hierarchies, see [@Faddeev; @Gervais].
Another group of models, in which the evolution is governed by an infinite number of parameters, can be observed in controllable dynamical systems, where the infinite number of degrees of freedom follows from the infinite number of driving terms. Surprisingly, the same structural background appears again for this group. We develop this viewpoint in the present paper.
One of the general approaches to the growing contour evolution was provided by Löwner and Kufarev [@Loewner; @Pommerenke2]. The contour evolution is described by a time-dependent conformal parametric map from a canonical domain, the unit disk in most of the cases, onto the domain bounded by the contour for each fixed instant. In fact, these one-parameter conformal maps satisfy the Löwner-Kufarev partial differential equation. A characteristic equation to this PDE represents an infinite dimensional controllable system for which the infinite number of conservation laws is given by the Virasoro generators in their covariant form.
Recently, Friedrich and Werner [@FriedrichWerner], and independently Bauer and Bernard [@BB], found relations between SLE (stochastic or Schramm-Löwner evolution) and the highest weight representation of the Virasoro algebra.
All above results encouraged us to conclude that the Virasoro algebra is a common structural basis for these and possibly other types of contour dynamics and we present the development in this direction here. For the first time, a construction, which appeared in the field theory plays the algebraic structural background for the contour evolution in classical complex analysis.
The structure of the paper is as follows. Sections 2 and 3 contain the necessary background on the Virasoro algebra and the Löwner-Kufarev equations. The main results are contained in Sections 4 and 5. In Section 4 we construct the Poisson structure on the cotangent bundle of the space of univalent functions smooth on the boundary and the Hamiltonian system generated by the Löwner-Kufarev equation in ordinary derivatives. We establish that the holomorphic Virasoro generators in the covariant formulation are conserved under the Löwner-Kufarev evolution (Theorem 2). The antiholomorphic generators are proved to contain a conserved term and we give an iterative method to obtain them based on the Poisson structure of the Löwner-Kufarev evolution. The Löwner-Kufarev PDE is shown to provide a distribution of the tangent bundle of non-normalized univalent functions, which forms the tangent bundle of normalized ones. It also gives an explicit correspondence between the latter bundle and the holomorphic eigen space of the complexified Lie algebra of vector fields on the unit circle. In Section 5, we give Hamiltonian and Lagrangian formulations of the motion within the coefficient body in the field of an elliptic operator constructed by means of Virasoro generators. The solutions with constant velocity coordinates are found. We prove that the norm of the driving function in the Löwner-Kufarev theory gives the minimal energy of the motion. The short Section 6 we add for completeness. We briefly review the connections between conformal field theory and the Schramm-Löwner evolution following [@BB; @FriedrichWerner].
[**Acknowledgements.**]{} We are thankful to Hélène Airault, Ludwig Faddeev, Paul Malliavin, and Yurii Neretin for many helpful discussions concerning the Virasoro algebra and its representations.
Virasoro Algebra
================
The Virasoro algebra $Vir$ plays a prominent role in modern mathematical physics, both in field theories and solvable models. It appears in physics literature as an algebra obeyed by the stress-energy tensor and associated with the conformal group, the Virasoro-Bott group, of the worldsheet in two dimensions, see e.g., [@Polchinski]. It is a unique central extension of the Lie algebra for the Lie-Fréchet group ${\text{\rm Diff }}S^1$ of sense-preserving diffeomorphisms of the unit circle $S^1$, and it is an infinite-dimensional real vector space. The extension is characterized by a real parameter $c$, so the Virasoro algebra refers to a class of isomorphic Lie algebras corresponding to different values of $c$. At the same time the Virasoro algebra is intrinsically related to the KdV canonical structure where the Virasoro brackets become the Magri brackets for the Miura transformations of elements of the phase space of the KdV hierarchy (see, e.g., [@Faddeev; @Gervais]).
The complex hull $\mathbb CVir$ of the Virasoro algebra can be realized as a central extension by $\mathbb C$ of the Witt algebra, a complex Lie algebra of derivations (or Leibnitz rule) of the algebra $\mathbb C[z,z^{-1}]$ of complex Laurent polynomials. The Witt algebra is spanned by the generators $L_n=z^{n+1}\frac{\partial}{\partial z}$ on $\mathbb C\setminus \{0\}$. The operators $L_n$ plus a central element $c$ are called the Virasoro generators. Under any irreducible representation of $\mathbb CVir$, the quantity $c$ is realized as a complex scalar and is called the central charge. The generators satisfy the commutation relations given by $$\{L_m,L_n\}_{Vir}=
(n-m)L_{m+n}+\frac{c}{12}n(n^2-1)\delta_{n,-m},\quad \{L_n,c\}_{Vir}=0,\quad n,m\in \mathbb Z,$$ where $c\in \mathbb C$ is the central charge. Considering the Virasoro algebra as an operator algebra, the generators $L_n$ become the coefficients in a formal Laurent series for the analytic component of the stress-energy tensor in 2-D field theory. The attribution ‘Virasoro algebra’ is due to a Virasoro’s seminal paper [@Virasoro].
Mathematically, the Virasoro algebra appeared for the first time as a central extension by the [*Gelfand-Fuchs cocycle*]{} [@GelfandFuchs] of the Lie algebra ${\text{\rm Vect }}S^1$ of smooth vector fields $\phi\frac{d}{d\theta}$ on the unit circle $S^1$ (see [@GelfandFuchs]), where the Lie bracket is defined to be the commutator of vector fields $$\label{Lie1}
[\phi_1,\phi_2]={\phi}_1{\phi}'_2-{\phi}_2{\phi}'_1.$$ Each element of the Lie-Fréchet group ${\text{\rm Diff }}S^1$ is represented as $z=e^{i\alpha(\theta)}$ with a monotone increasing $C^{\infty}$ real-valued function $\alpha(\theta)$, such that $\alpha(\theta+2\pi)=\alpha(\theta)+2\pi$. The Lie algebra for this group is identified with ${\text{\rm Vect }}S^1$. The relation of this Lie algebra to ${\text{\rm Diff }}S^1$ is subtile because the exponential map is not even locally a homeomorphism.
Canonical identification
------------------------
The entire necessary background of unitary representations of ${\text{\rm Diff }}S^1$ is found in the study of Kirillov’s homogeneous Kählerian manifold ${\text{\rm Diff }}S^1/S^1$. We deal with the analytic representation of ${\text{\rm Diff }}S^1/S^1$. Let ${\text{\bf S}}$ stand for the whole class of univalent functions $f$ in the unit disk $U$ normalized by $f(z)=z(1+\sum_{n=1}^{\infty}c_nz^n)$ about the origin and $C^{\infty}$-smooth on the boundary $S^1$ of $U$. Given a map $f\in {\text{\bf S}}$ we construct the adjoint univalent meromorphic map $$g(z)=d_1z+d_0+\frac{d_{-1}}{z}+\dots,$$ defined in the exterior $U^*=\{z:\,|z|>1\}$ of $U$, and such that $\hat{\mathbb C}\setminus\overline{f(U)}=g(U^*)$. Both functions are extendable onto $S^1$. This conformal welding gives the identification of the homogeneous manifold ${\text{\rm Diff }}S^1/S^1$ with the space ${\text{\bf S}}$: ${\text{\bf S}}\ni f\leftrightarrow f^{-1}\circ g|_{S^1}\in{\text{\rm Diff }}S^1/S^1$, or with the smooth contours $\Gamma=f(S^1)$ that enclose univalent domains $\Omega$ of conformal radius 1 with respect to the origin and such that $\infty\not\in \Omega$, $0\in\Omega$, see [@Airault], [@KY1]. So one can construct complexification of ${\text{\rm Vect }}S^1$ and further projection of the holomorphic part to the set $\mathcal M\subset \mathbb C^{\mathbb N}$, which is the projective limit of the coefficient bodies $\mathcal M=\lim_{n\leftarrow \infty}\mathcal M_n$, where $$\mathcal M_n=\{(c_1,\dots,c_n):\,\,f\in {\text{\bf S}}\}.\label{Mn}$$ The holomorphic Virasoro generators can then be realized by the first order differential operators $$L_j=\partial_j+\sum\limits_{k=1}^{\infty}(k+1)c_{k}\partial_{j+k},\quad j\in \mathbb N,$$ in terms of the affine coordinates of $\mathcal M$, acting over the set of holomorphic functions, where $\partial_{k}=\partial/\partial{c_k}$. We explain the details in the next subsection.
Complexification
----------------
Let us introduce local coordinates on the manifold $\mathcal M={\text{\rm Diff }}S^1/S^1$ in the concordance with the local coordinates on the space ${\text{\bf S}}$ of univalent functions smooth on the boundary. Observe that $\mathcal M$ is a real infinite-dimensional manifold, whereas ${\text{\bf S}}$ is a complex manifold. We are aimed at a complexification of $T\mathcal M$ which admits a holomorphic projection to $T{\text{\bf S}}$, where ${\text{\rm Vect }}_0 S^1={\text{\rm Vect }}\, S^1/{\text{\rm const}}$ is a module over the ring of smooth functions, which is associated with the tangent bundle $T\mathcal M$.
Given a real vector space $V$ the complexification $V_{\mathbb C}$ is defined as the tensor product with the complex numbers $V\otimes_{\mathbb R}\mathbb C$. Elements of $V_{\mathbb C}$ are of the form $v\otimes z$. In addition, the vector space $V_{\mathbb C}$ is a complex vector space that follows by defining multiplication by complex numbers, $\alpha(v\otimes z)=v\otimes \alpha z$ for complex $\alpha$ and $z$ and $v\in V$. The space $V$ is naturally embedded into $V\otimes \mathbb C$ by identifying $V$ with $V\otimes 1$. Conjugation is defined by introducing a canonical conjugation map on $V_{\mathbb C}$ as $\overline{v\otimes z}=v\otimes \bar z$.
An almost complex structure $J$ on $V$ can be extended by linearity to the complex structure $J$ on $V_{\mathbb C}$ by $J(v\otimes z)=J(v)\otimes z$. Observe that $$\overline{J(v\otimes z)}=J(\overline{v\otimes z}).$$
Eigenvalues of extended $J$ are $\pm i$, and there are two eigenspaces $V^{(1,0)}$ and $V^{(0,1)}$ corresponding to them given by projecting $\frac{1}{2}(1\mp iJ)v$. $V_{\mathbb C}$ is decomposed into the direct sum $V_{\mathbb C}=V^{(1,0)}\oplus V^{(0,1)}$, where $V^{(1,0)}=\{v\otimes 1- J(v)\otimes i\big| v\in V \}$ and $V^{(0,1)}=\{v\otimes 1+ J(v)\otimes i\big| v\in V \}$ are the eigen spaces corresponding to $\pm i$.
An almost complex structure on ${\text{\rm Vect }}_0 S^1$ may be defined as follows (see [@Airault]). We identify ${\text{\rm Vect }}_0 S^1$ with the functions with vanishing mean value over $S^1$. It gives $$\phi(\theta)=\sum\limits_{n=1}^{\infty}a_n\cos\,n\theta+b_n\sin\,n\theta.$$ Let us define an almost complex structure by the operator $$\label{CompStruct}
J(\phi)(\theta)=\sum\limits_{n=1}^{\infty}-a_n\sin\,n\theta+b_n\cos\,n\theta.$$ On ${\text{\rm Vect }}_0S^1\otimes \mathbb C$, the operator $J$ diagonalizes and we have the identification $${\text{\rm Vect }}_0S^1\ni \phi\leftrightarrow v:=\frac{1}{2}(\phi-iJ(\phi))=\sum\limits_{n=1}^{\infty}(a_n-ib_n)e^{in\theta}\in ({\text{\rm Vect }}_0S^1\otimes \mathbb C)^{(1,0)},$$ and the latter extends into the unit disk as a holomorphic function.
The Kirillov infinitesimal action [@Kir2] of ${\text{\rm Vect }}_0 S^1$ on ${\text{\bf S}}$ is given by a variational formula due to Schaeffer and Spencer [@Schaeffer page 32] which lifts the actions from the Lie algebra ${\text{\rm Vect }}_0
S^1$ onto ${\text{\bf S}}$. Let $f\in{\text{\bf S}}$ and let $\phi(e^{i\theta}):=\phi(\theta)\in {\text{\rm Vect }}_0 S^1$ be a $C^{\infty}$ real-valued function in $\theta\in(0,2\pi]$. The infinitesimal action $\theta \mapsto \theta+\varepsilon \phi(e^{i\theta})$ yields a variation of the univalent function $f^*(z)= f+\varepsilon\,\delta_{v}f(z)+o(\epsilon)$, where $$\delta_{v}f(z)=\frac{f^2(z)}{2\pi
}\int\limits_{S^1}\left(\frac{wf'(w)}{f(w)}\right)^2\frac{v(w)dw}{w(f(w)-f(z))} ,\label{var}$$ and $\phi\leftrightarrow v$ by the above identification. Kirillov and Yuriev [@KY1], [@KY2] (see also [@Airault]) established that the variations $\delta_{\phi}f(\zeta)$ are closed with respect to the commutator (\[Lie1\]), and the induced Lie algebra is the same as ${\text{\rm Vect }}_0
S^1$. The Schaeffer-Spencer operator is linear.
Treating $T\mathcal M$ as a real vector space, the operator $\delta_{\phi}$ transfers the complex structure $J$ from ${\text{\rm Vect }}_0 S^1$ to $T\mathcal M$ by $J(\delta_{\phi}):=\delta_{J(\phi)}$. By abuse of notation, we denote the new complex structure on $T\mathcal M$ by the same character $J$. Then it splits the complexification $T\mathcal M_{\mathbb C}$ into two eigenspaces $T\mathcal M_{\mathbb C}=T\mathcal M^{(1,0)}\oplus T\mathcal M^{(0,1)}$. Therefore, $\delta_{v}=\delta_{\phi-iJ(\phi)}:=\delta_{\phi}-iJ(\delta_{\phi})\in T\mathcal M^{(1,0)}$. Observe that $2z\partial_z=-i\partial_{\theta}$ on the unit circle $z=e^{i\theta}$, and $L_k=z^{k+1}d/d z=-\frac{1}{2}ie^{ik\theta}d/d \theta$ on $S^1$. Let us take the basis of ${\text{\rm Vect }}_0 S^1\otimes \mathbb C$ in the form $\nu_k=-ie^{ik\theta}$ in order to keep the index of vector fields the same as for $L_k$. Then, the commutator satisfies the Witt relation $\{\nu_m, \nu_n\}=(n-m)\nu_{n+m}$. Taking elements $\nu_k=-iw^k$, $|w|=1$ in the integrand of (\[var\]) we calculate the residue in (\[var\]) and obtain so called Kirillov operators $$L_j[f](z)=\delta_{\nu_j}f(z)=z^{j+1}f'(z), \quad j=1,2,\dots,$$ so that these $L_j$ are the holomorphic coordinates on $T\mathcal M^{(1,0)}$. In terms of the affine coordinates in $\mathcal M$ we get the Kirillov operators as $$L_j=\partial_j+\sum\limits_{k=1}^{\infty}(k+1)c_{k}\partial_{j+k},$$ where $\partial_k=\partial/\partial c_k$. They satisfy the Witt commutation relation $$\{L_m,L_n\}=(n-m)L_{n+m}.$$ For $k=0$ we obtain the operator $L_0$, which corresponds to the constant vectors from ${\text{\rm Vect }}\,S^1$, $L_{0}[f](z)=zf'(z)-f(z)$. The elements of the Fourier basis $-ie^{-i\theta k}$ with negative indices (corresponding to $T\mathcal M^{(0,1)}$) are extended into $U$ by $-iz^{-k}$. Substituting them in (\[var\]) we get very complex formulas for $L_{-k}$, which functionally depend on $L_k$ (see [@Airault], [@Kir2]), and which are dual to $L_k$ with respect to the action of $J$. The first two operators are calculated as $$\begin{aligned}
L_{-1}[f](z)&=&f'(z)-2c_1 f(z)-1,\\
L_{-2}[f](z)&=&\frac{f'(z)}{z}-\frac{1}{f(z)}-3c_1+(c_1^2-4c_2)f(z),\end{aligned}$$ see [@KY2].
This procedure gives a nice links between representations of the Virasoro algebra and the theory of univalent functions. The Löwner-Kufarev equations proved to be a powerful tool to work with univalent functions (the famous Bieberbach conjecture was proved [@Branges] using Löwner method). In the following section we show how Löwner-Kufarev equations can be used in a representation of the Virasoro algebra. In particular, we identify $T\mathcal M^{(1,0)}$ with $T\mathcal M$, equipped with its natural complex structure given by coefficients of univalent functions, by means the Löwner-Kufarev PDE.
Löwner-Kufarev Equations
========================
A time-parameter family $\Omega(t)$ of simply connected hyperbolic univalent domains forms a [*Löwner subordination chain*]{} in the complex plane $\mathbb
C$, for $0\leq t< \tau$ (where $\tau$ may be $\infty$), if $\Omega(t)\varsubsetneq \Omega(s)$, whenever $t<s$. We suppose that the origin is an interior point of the Carathéodory kernel of $\{\Omega(t)\}_{t=0}^{\tau}$.
A Löwner subordination chain $\Omega(t)$ is described by a time-dependent family of conformal maps $z=f(\zeta,t)$ from the unit disk $U=\{\zeta:\,|\zeta|<1\}$ onto $\Omega(t)$, normalized by $f(\zeta,t)=a_1(t)\zeta+a_2(t)\zeta^2+\dots$, $a_1(t)>0$, $\dot{a}_1(t)>0$. After Löwner’s 1923 seminal paper [@Loewner] a fundamental contribution to the theory of Löwner chains was made by Pommerenke [@Pommerenke1; @Pommerenke2] who described governing evolution equations in partial and ordinary derivatives, known now as the Löwner-Kufarev equations due to Kufarev’s work [@Kufarev].
One can normalize the growth of evolution of a subordination chain by the conformal radius of $\Omega(t)$ with respect to the origin by $a_1(t)=e^t$.
Löwner [@Loewner] studied a time-parameter semigroup of conformal one-slit maps of the unit disk $U$ arriving then at an evolution equation called after him. His main achievement was an infinitesimal description of the semi-flow of such maps by the Schwarz kernel that led him to the Löwner equation. This crucial result was then generalized in several ways (see [@Pommerenke2] and the references therein).
We say that the function $p$ is from the Carathéodory class if it is analytic in $U$, normalized as $p(\zeta)=1+p_1\zeta+p_2\zeta^2+\dots,\quad
\zeta\in U,$ and such that ${\text{\rm Re }}p(\zeta)>0$ in $U$. Pommerenke [@Pommerenke1; @Pommerenke2] proved that given a subordination chain of domains $\Omega(t)$ defined for $t\in [0,\tau)$, there exists a function $p(\zeta,t)$, measurable in $t\in [0,\tau)$ for any fixed $z\in U$, and from the Carathéodory class for almost all $t\in [0,\tau)$, such that the conformal mapping $f:U\to \Omega(t)$ solves the equation $$\frac{\partial f(\zeta,t)}{\partial t}=\zeta\frac{\partial
f(\zeta,t)}{\partial \zeta}p(\zeta,t),\label{LK}$$ for $\zeta\in U$ and for almost all $t\in [0,\tau)$. The equation (\[LK\]) is called the Löwner-Kufarev equation due to two seminal papers: by Löwner [@Loewner] who considered the case when $$p(\zeta,t)=\frac{e^{iu(t)}+\zeta}{e^{iu(t)}-\zeta},\label{yadro}$$ where $u(t)$ is a continuous function regarding to $t\in [0,\tau)$, and by Kufarev [@Kufarev] who proved differentiability of $f$ in $t$ for all $\zeta$ from the kernel of $\{\Omega(t)\}$ in the case of general $p$ in the Carathéodory class.
Let us consider a reverse process. We are given an initial domain $\Omega(0)\equiv \Omega_0$ (and therefore, the initial mapping $f(\zeta,0)\equiv f_0(\zeta)$), and a function $p(\zeta,t)$ of positive real part normalized by $p(\zeta,t)=1+p_1\zeta+\dots$. Let us solve the equation (\[LK\]) and ask ourselves, whether the solution $f(\zeta,t)$ defines a subordination chain of simply connected univalent domains $f(U,t)$. The initial condition $f(\zeta,0)=f_0(\zeta)$ is not given on the characteristics of the partial differential equation (\[LK\]), hence the solution exists and is unique but not necessarily univalent. Assuming $s$ as a parameter along the characteristics we have $$\frac{dt}{ds}=1,\quad \frac{d\zeta}{ds}=-\zeta
p(\zeta,t), \quad \frac{df}{ds}=0,$$ with the initial conditions $t(0)=0$, $\zeta(0)=z$, $f(\zeta,0)=f_0(\zeta)$, where $z$ is in $U$. Obviously, $t=s$. Observe that the domain of $\zeta$ is the entire unit disk. However, the solutions to the second equation of the characteristic system range within the unit disk but do not fill it. Therefore, introducing another letter $w$ (in order to distinguish the function $w(z,t)$ from the variable $\zeta$) we arrive at the Cauchy problem for the Löwner-Kufarev equation in ordinary derivatives $$\frac{dw}{dt}=-wp(w,t),\label{LKord}$$ for a function $\zeta=w(z,t)$ with the initial condition $w(z,0)=z$. The equation (\[LKord\]) is a non-trivial characteristic equation for (\[LK\]). Unfortunately, this approach requires the extension of $f_0(w^{-1}(\zeta,t))$ into the whole $U$ ($w^{-1}$ means the inverse function) because the solution to (\[LK\]) is the function $f(\zeta,t)$ given as $f_0(w^{-1}(\zeta,t))$, where $\zeta=w(z,s)$ is a solution of the initial value problem for the characteristic equation (\[LKord\]) that maps $U$ into $U$. Therefore, the solution of the initial value problem for the equation (\[LK\]) may be non-univalent.
Solutions to the equation (\[LKord\]) are holomorphic univalent functions $w(z,t)=e^{-t}z+a_2(t)z^2+\dots$ in the unit disk that map $U$ into itself. Every function $f$ from the class ${\text{\bf S}}$ can be represented by the limit $$f(z)=\lim\limits_{t\to\infty}e^t w(z,t),\label{limit}$$ where $w(z,t)$ is a solution to with some function $p(z,t)$ of positive real part for almost all $t\geq 0$ (see [@Pommerenke2 pages 159–163]). Each function $p(z,t)$ generates a unique function from the class ${\text{\bf S}}$. The reciprocal statement is not true. In general, a function $f\in {\text{\bf S}}$ can be obtained using different functions $p(\cdot,t)$.
Now we are ready to formulate the condition of univalence of the solution to the equation (\[LK\]), which can be obtained by combination of known results of [@Pommerenke2].
\[ThProkhVas\][[@Pommerenke2; @ProkhVas]]{} Given a function $p(\zeta,t)$ of positive real part normalized by $p(\zeta,t)=1+p_1\zeta+\dots$, the solution to the equation (\[LK\]) is unique, analytic and univalent with respect to $\zeta$ for almost all $t\geq 0$, if and only if, the initial condition $f_0(\zeta)$ is taken in the form , where the function $w(\zeta,t)$ is the solution to the equation with the same driving function $p$.
Recently, we started to look at Löwner-Kufarev equations from the point of view of motion in the space of univalent functions where Hamiltonian and Lagrangian formalisms play a central role (see, [@Vasiliev]). Some connections with the Virasoro algebra were also observed in [@MPV; @Vasiliev]. The present paper generalizes these attempts and gives their closed form. The main conclusion is that the Löwner-Kufarev equations are naturally linked to the holomorphic part of the Virasoro algebra. Taking holomorphic Virasoro generators $L_n$ as a basis of the tangent space to the coefficient body for univalent functions at a fixed point, we see that the driving function in the Löwner-Kufarev theory generates generalized moments for motions within the space of univalent functions. Its norm represents the energy of this motion. The holomorphic Virasoro generators in their co-tangent form will become conserved quantities of the Löwner-Kufarev ODE. The Löwner-Kufarev PDE becomes a transition formula from the affine basis to Kirillov’s basis of the holomorphic part of the complexified tangent space to $\mathcal M$ at any point. Finally, we propose to study an alternate Löwner-Kufarev evolution instead of subordination.
Witt algebra and the classical Löwner-Kufarev equations
=======================================================
In the following subsections we reveal the structural role of the Witt algebra as a background of the classical Löwner-Kufarev contour evolution. As we see further, the conformal anomaly and the Virasoro algebra appear as a quantum or stochastic effect in SLE.
Löwner-Kufarev ODE
------------------
Let us consider the functions $$w(z,t)=e^{-t}z\left(1+\sum\limits_{n=1}^{\infty}c_n(t)z^n\right),$$ satisfying the Löwner-Kufarev ODE $$\label{LKord1}
\frac{dw}{dt}=-wp(w,t),$$ with the initial condition $w(z,0)=z$, and with the function $p(z,t)=1+p_1(t)z+\dots$ which is holomorphic in $U$ and measurable with respect to $t\in [0,\infty)$, such that ${\text{\rm Re }}p>0$ in $U$. The function $w(z,t)$ is univalent and maps $U$ into $U$.
Let the function $w(z,t)$ be a solution to the Cauchy problem for the equation with the initial condition $w(z,0)=z$. If the driving function $p(\cdot,t)$, being from the Carathéodory class for almost all $t\geq 0$, is $C^{\infty}$ smooth in the closure $\hat{U}$ of the unit disk $U$ and summable with respect to $t$, then the boundaries of the domains $B(t)=w(U,t)\subset U$ are smooth for all $t$.
Observe that the continuous and differentiable dependence of the solution to a differential equation $\dot{x}=F(t,x)$ on the initial condition $x(0)=x_0$ is a classical problem. One can refer, e.g., to [@Volpato] in order to assure that summability of $F(\cdot, x)$ regarding to $t$ for each fixed $x$ and continuous differentiability ($C^1$ with respect to $x$ for almost all $t$) imply that the solution $x(t,x_0)$ exists, is unique, and is $C^1$ with respect to $x_0$. In our case, the solution to exists, is unique, analytic in $U$, and moreover, $C^1$ on its boundary $S^1$. Let us differentiate inside the unit disk $U$ with respect to $z$ and write $$\log w' =-\int\limits_{0}^{t}(p(w(z,\tau),\tau)+w(z,\tau)p'(w(z,\tau),\tau))d\tau,$$ choosing the branch of the logarithm such as $\log w'(0,t)=-t$. This equality is extendable onto $S^1$ because the right-hand side is, and therefore, $w'$ is $C^1$ and $w$ is $C^2$ on $S^1$. We continue analogously and write the formula $$w''=-w'\int\limits_{0}^{t}(2w'(z,\tau)p'(w(z,\tau),\tau)
+w(z,\tau)w'(z,\tau)p''(w(z,\tau),\tau))d\tau,$$ which guarantees that $w$ is $C^3$ on $S^1$. Finally, we come to the conclusion that $w$ is $C^\infty$ on $S^1$.
Let $f(z,t)$ denote $e^tw(z,t)$. The limit $\lim_{t\to\infty}f(z,t)$ is known [@Pommerenke2] to be a representation of all univalent functions.
Let the driving term $p(z,t)$ in the Löwner-Kufarev ODE be from the Carathéodory class for almost all $t\geq 0$, $C^{\infty}$ smooth in $\hat{U}$, and summable with respect to $t$. Then the domains $\Omega(t)=w(U,t)$ have smooth boundary $\partial
\Omega(t)$. So the Löwner equation can be extended onto the closed unit disk $\hat U=U\cup S^1$.
Consider the Hamiltonian function given by $$\label{Ham3}
H=\int\limits_{z\in S^1}f(z,t)(1-p(e^{-t}f(z,t),t))\bar \psi(z,t)\frac{dz}{iz},$$ on the unit circle $z\in S^1$, where $\psi(z,t)$ is a formal series $$\psi(z,t)=\sum_{n=-k}^{\infty}\psi_nz^{n},$$ defined about the unit circle $S^1$ for any $k\geq 0$. The Poisson structure on the symplectic space $(f, \bar\psi)$ is given by the canonical brackets $$\{P, Q\}=\frac{\delta P}{\delta f}\frac{\delta Q}{\delta \bar \psi}-\frac{\delta P}{\delta \bar \psi}\frac{\delta Q}{\delta f},$$ or in coordinate form (only $\psi_n$ for $n\ge 1$ are independent co-vectors corresponding to the tangent vectors $\partial_n$ with respect to the canonical Hermitean product for analytic functions) $$\{p, q\}=\sum_{n=1}^{\infty}\frac{\partial p}{\partial c_n}\frac{\partial q}{\partial \bar \psi_n}-\frac{\partial p}{\partial \bar \psi_n}\frac{\partial q}{\partial c_n}.$$ Here $$P(t)= \int\limits_{z\in S^1}p(z,t)\frac{dz}{iz},\quad Q(t)= \int\limits_{z\in S^1}q(z,t)\frac{dz}{iz}.$$ The Hamiltonian system becomes $$\label{sys1}
\frac{d f(z,t)}{dt}=f(1-p(e^{-t}f,t))=\frac{\delta H}{\delta \overline{\psi}}=\{f,H\},$$ for the position coordinates and $$\label{sys2}
\frac{d\bar \psi}{dt}=-(1-p(e^{-t}f,t)-e^{-t}fp'(e^{-t}f,t))\bar
\psi=\frac{-\delta H}{\delta f}=\{\overline{\psi},H\},$$ for the momenta, where $\frac{\delta}{\delta f}$ and $\frac{\delta}{\delta \overline{\psi}}$ are the variational derivatives. So the phase coordinates $(f,\bar{\psi})$ play the role of the canonical Hamiltonian pair.
The coefficients $c_n$ are the complex local coordinates on $\mathcal M$, so in these coordinates we have $$\begin{aligned}
\dot{c}_n & = &\frac{d c_n}{dt}=c_n-\frac{e^t}{2\pi i}\int\limits_{S^1}w(z,t)p(w(z,t),t)\frac{dz}{z^{n+2}}, \nonumber\\
&=&-\frac{1}{2\pi i}\int\limits_{S^1}\sum\limits_{k=1}^ne^{-kt}(e^tw)^{k+1}p_k\frac{dz}{z^{n+2}},\quad n\geq 1.\nonumber\end{aligned}$$ Let us fix some $n$ and project the infinite dimensional Hamiltonian system on an $n$-dimensional $\mathcal M_n$. The dynamical equations for momenta governed by the Hamiltonian function are $$\dot{\bar{\psi}}_j=-\bar{\psi}_j+\frac{1}{2\pi i}\sum\limits_{k=1}^n\bar{\psi}_k\int\limits_{S^1}(p+wp')\frac{dz}{z^{k-j+1}},\quad j= 1,\dots, n-1,\label{psi1}$$ and $$\dot{\bar{\psi}}_n=0.$$ In particular, $$\begin{aligned}
\dot{c}_1 & = & -e^{-t}p_1,\\
\dot{c}_2 & = & -2e^{-t}p_1c_1-e^{-2t}p_2,\\
\dot{c}_3 & = & -e^{-t}p_1(2c_2+c_1^2)-3e^{-2t}p_2c_1-e^{-3t}p_3,\\
\dots& & \dots\end{aligned}$$ for $n=3$ we have $$\begin{aligned}
\dot{\bar{\psi}}_1 & = & 2e^{-t}p_1\bar{\psi}_2+(2e^{-t}p_1c_1+3e^{-2t}p_2)\bar{\psi}_3,\\
\dot{\bar{\psi}}_2 & = & 2e^{-t}p_1\bar{\psi}_3,\\
\dot{\bar{\psi}}_3 & = & 0.\end{aligned}$$
Let us set the function $L(z):=f'(z,t)\bar\psi(z,t)$. Let $(L(z))_{<0}$ mean the part of the Laurent series for $L(z)$ with negative powers of $z$, $$(L(z))_{<0}=(\bar\psi_1+2c_1\bar\psi_2+3c_2\bar\psi_3+\dots)\frac{1}{z}+(\bar\psi_2+2c_1\bar\psi_3+\dots)\frac{1}{z^2}+\dots=\sum\limits_{k=1}^{\infty}\frac{L_k}{z^{k}}.$$ Then, the functions $L(z)$ and $(L(z))_{< 0}$ are time-independent for all $z\in S^1$.
It is easily seen that, passing from the cotangent vectors $\bar\psi_k$ to the tangent vectors $\partial_k$, the coefficients $L_k$ of $(L(z))_{<0}$ defined on the tangent bundle $T\mathcal M^{(1,0)}$ are exactly the Kirillov vector fields $L_k$. The corresponding fields $L_k$ in the covariant form are conserved by the Löwner-Kufarev ODE because $\dot L_k=\{L_k, H\}=0$. The above Poisson structure coincides with that given by the Witt brackets introduced for $L_k$ previously. For finite-dimensional grades this result was obtained in [@MPV].
Let us formulate the result as a theorem.
Let the driving term $p(z,t)$ in the Löwner-Kufarev ODE be from the Carathéodory class for almost all $t\geq 0$, $C^{\infty}$ smooth in $\hat{U}$, and summable with respect to $t$. Then the Kirillov fields in the covariant form are the conserved quantities for the Hamiltonian system (\[sys1\]–\[sys2\]) generated by the Löwner-Kufarev ODE.
Another way to construct a Hamiltonian system could be based on the symplectic structure given by the Kählerian form on ${\text{\rm Diff }}S^1/S^1$. However, there is no explicit expression for such form in terms of functions $f\in {\text{\bf S}}$. Moreover, there must be a Hamiltonian formulation in which the Löwner-Kufarev equation becomes an evolution equation. This remains an open problem.
At a first glance the situation with an ODE with a parameter is quite simple. Indeed, if we solve an equation of type $\dot{f}(t, e^{i\theta})=F(f(t, e^{i\theta}),t)$, then fixing $\theta$ we have an integral of motion $C=I(f(t,\cdot),t)={\text{\rm const}}$. Then, releasing $\theta$, we have $C(e^{i \theta})=I(f(t,e^{i\theta}),t)$. Expanding $C(e^{i\theta})$ into the Fourier series, we obtain an infinite number of conserved quantities, but they do not manifest an infinite number of degrees of freedom that govern the motion as in the field theory where the governing equations are PDE. In our case, we have not only one trajectory fixing the initial condition but a pensil of trajectories because our equation has an infinite number of control parameters, the Taylor coefficients of the function $p(z,t)$, which form a bounded non-linear set of admissible controls. Therefore, we operate with sections of the tangent and co-tangent bundles to the infinite dimensional manifold $\mathcal M$ instead of vector fields along one trajectory as in usual ODE.
No linear combinations $L^*_k$ of $L_1,\dots,L_n,\dots$ allows us to reduce the system of $\{L_k\}$ to a new system of involutory $\{L_k^*\}$ in order to claim the Liouville integrability of our system. Observe that the coefficients in these linear combinations must be constants to keep conservation laws.
Construction of $L_0$ and $L_{-n}$
----------------------------------
Consider again the generating function $L(z)=f'(z,t)\bar\psi(z,t)$ and the ‘non-negative’ part $(L(z))_{\ge 0}$ of the Laurent series for $L(z)$, $$(L(z))_{\ge 0}= (\bar\psi_0+2c_1\bar\psi_1+3c_2\bar\psi_2+\dots)+ (\bar\psi_{-1}+2c_1\bar\psi_0+3c_2\bar\psi_1+\dots)z+\dots$$ $$=\sum\limits_{k=0}^{\infty}\mathcal L_{-k}z^{k}.$$ All $\mathcal L_{-k}$ are conserved by the construction. Define $\bar\psi_0^*=-\sum_{n=1}^{\infty}c_k\bar\psi_k$, and $$L_0=\mathcal L_0-(\bar\psi_0-\bar\psi^*_0).$$ The operator $L_0$ acts on the class ${\text{\bf S}}$ by $L_0[f](z)=zf'(z)-f(z)$. Next define $L_{-1}=\mathcal L_{-1}-(\bar\psi_{-1}-\bar\psi_{-1}^*)-2c_1(\bar\psi_0-\bar\psi_0^*)$, where $\bar\psi_{-1}^*=0$. Then, $$L_{-1}[f](z)=f'(z)-2c_1 f(z)-1$$ Finally, $$L_{-2}=\mathcal L_{-2}-(\bar\psi_{-2}-\bar\psi_{-2}^*)-2c_1(\bar\psi_{-1}-\bar\psi_{-1}^*)-3{c_2}(\bar\psi_0-\bar\psi_0^*).$$ We choose $\bar\psi_{-2}^*=(c_3-3c_1c_2+c_1^3)\bar\psi_1+\dots$, so that $$\bar\psi_{-2}^*[f](z)=\frac{1}{z}-\frac{1}{f(z)}-c_1-(c_2-c_1^2)f(z),$$ and $$L_{-2}[f](z)=\frac{f'(z)}{z}-\frac{1}{f(z)}-3c_1+(c_1^2-4c_2)f(z).$$ An important fact is that $$L_{0}=c_1 \bar\psi_1+2c_2 \bar\psi_2+\dots,$$ $$L_{-1}=(3c_2-2c_1^2)\bar\psi_1+\dots,$$ $$L_{-2}=(5c_3-6c_1c2+2c_1^3) \bar\psi_1+\dots,$$ are linear with respect to $ \bar\psi_k$, $k\geq 1$, and therefore, are sections of $T^*\mathcal M$, which are dual to Kirillov’s vector fields. Equivalently, $$L_{0,-1,-2}[f](z)=\mbox{function}(c_1,c_2,\dots)z^2+\dots, \quad z^k=\frac{\partial f}{\partial c_{k-1}}.$$ All other co-vectors we construct by our Poisson brackets as $$L_{-n}=\frac{1}{n-2}\{L_{-n+1},L_{-1}\}=\frac{1}{n-4}\{L_{-n+2},L_{-2}\}.$$ The form of the Poisson brackets guarantees us that all $L_{-n}$ are linear with respect to $\bar\psi_1,\bar\psi_2,\dots$ and span the anti-holomorphic part of the co-tangent bundle ${T^{(0,1)}}^*\mathcal M$.
Let us summarize the above in the following conclusion. We considered a non-linear contour dynamics given by the Löwner-Kufarev equation. It turned out to be underlined by an algebraic structure, namely, by the Witt algebra spanned by the Virasoro generators $L_n$, $n\in\mathbb Z$.
- $L_n$, $n = 1,2,\dots$ are the holomorphic Virasoro generators. They span the holomorphic part of the complexiÞed tangent bundle over the space of univalent functions, smooth on the boundary. In the covariant formulation they are conserved by the Löwner-Kufarev evolution.
- $L_0$ is the central element.
- $L_{-n}$, $n = 1,2,\dots$ are the antiholomorphic Virasoro generators. They span the antiholomorphic part of the decomposition. They contain a conserved term and we give an iterative method to obtain them based on the Poisson structure of the Löwner-Kufarev evolution.
Löwner-Kufarev PDE
------------------
The Löwner equation in partial derivatives is $${ \dot{w}(\zeta,t)=\zeta w'(\zeta,t)p(\zeta,t)}, \quad {\text{\rm Re }}p(\zeta,t)>0, \quad |\zeta|<1.$$ with some initial condition $w(z,0)=f_0(z)$. Let us consider the one-parameter family of functions $f(z,t)=e^{-t}w(z,t)=z(1+\sum_{n=1}^{\infty}c_n(t)z^n)$, $f(z,0)=f_0(z)$ as a $C^1$ path in ${\text{\bf S}}$. At the initial point $f_0(z)$ we have that $T_{f_0}{\text{\bf S}}=T_{f_0}\mathcal M^{(1,0)}=T_{f_0}\mathcal M$. A path in the coefficient body $\mathcal M$ in the neighbourhood of $f_0$ is $(c_1(t),\dots, c_n(t),\dots)$ with the velocity vector $\dot c_1\partial_1+\dots+\dot c_n\partial_n+\dots \in T_{f_0}\mathcal M$.
Taking the Virasoro generators $\{L_k\}$, $k\geq 1$, as a basis in $T_{f_0}\mathcal M^{(1,0)}$ we wish the velocity vector written in this new basis to be $${ \dot c_1\partial_1+\dots+\dot c_n\partial_n+\dots=u_1L_1+\dots
u_n L_n+\dots}.\label{e1}$$ We compare (\[e1\]) with the Löwner-Kufarev equation $${ \dot f=\dot c_1\partial_1+\dots+\dot
c_n\partial_n+\dots=zf'p(z,t)- f=L_0+u_1L_1+\dots u_n L_n+\dots},\label{e2}$$ where $p(z,t)=1+u_1z+\dots+u_nz^n+\dots$, and $L_0f=zf'-f$. In view of similarity between these two expressions and , we notice that
- a new term $L_0$ appears in the Löwner-Kufarev equation;
- the function $p(z,t)$ with positive real part corresponds to subordination, whereas for generic trajectories it may have real part of arbitrary sign. We call this an [*alternate Löwner-Kufarev evolution*]{};
- the vector $L_0$ corresponds exactly to the rotation: $$e^{i\varepsilon}f(e^{-i\varepsilon}z)=f(z)-i\varepsilon(zf'(z)-f(z))+o(\varepsilon).$$
Let us consider the set ${\text{\bf S}}_0$ of non-normalized smooth univalent functions of the form $F(z,t)=a_0(t)z+a_1(t)z^2+\dots$, with a tangent vector $\dot a_0\partial_0+\dots+\dot a_n\partial_n+\dots$, where $\partial_k=\partial/\partial a_k$, $k=0,1,2, \dots$. Our aim is to define two different distributions for the tangent bundle $T{\text{\bf S}}_0$, that form a sub-bundle of co-dimension 1, which is the tangent bundle $T{\text{\bf S}}$. This will be realized by means of formulas and . Notice that $\partial_k F=z^{k+1}$. Setting $L_k(F):=z^{k+1}F'$ we get $$\dot F=\dot a_0\partial_0+\dots+\dot
a_n\partial_n+\dots=zf'p(z,t)=u_0L_0+u_1L_1+\dots u_n L_n+\dots,$$ where $p(z,t)=u_0+u_1z+\dots+u_nz^n+\dots$. This alternate Löwner-Kufarev equation represents recalculation of the tangent vector in the new basis $${ \dot a_0\partial_0+\dots+\dot a_n\partial_n+\dots=u_0L_0+\dots
u_n L_n+\dots},$$ where $L_k=a_0\partial_k+2a_1\partial_{k+1}+\dots$.
Let us present the distributions. We start with $F\in {\text{\bf S}}_0$, then we define $f\in {\text{\bf S}}$. The necessary distribution is the map $${\text{\bf S}}_0\ni F\to T_f{\text{\bf S}}\hookrightarrow T_F{\text{\bf S}}_0.$$
The analytic form of the first distribution is the following factorization $f_1(z,t)=\frac{1}{a_0}F(z,t)=z+\frac{a_1}{a_0}z^2+\dots$, so that $$\dot f_1=zf_1'p(z,t)-\frac{\dot a_0}{a_0}f_1,\label{F1}$$ where $u_0=\frac{\dot a_0}{a_0}$. Then we obtain $${ \dot c_1\partial_1+\dots+\dot c_n\partial_n+\dots=\hat L_0+u_1\hat L_1+ \dots +u_n \hat L_n+\dots}$$ where $\hat L_0f_1=u_0(zf_1'-f_1)$, $\hat L_kf_1=z^{k+1}f_1'$, $c_k=\frac{a_k}{a_0}$, $\partial_k=\frac{\partial}{\partial c_k}$. In particular, $a_0=e^t$ implies the Löwner-Kufarev equation for arbitrary sign of ${\text{\rm Re }}p$.
The analytic form of the second distribution becomes $f_2(z,t)=F(\frac{1}{a_0}z,t)=z+\frac{a_1}{a^2_0}z^2+\dots$, so that $$\dot f_2=zf_2'p(\frac{z}{a_0},t)- \frac{\dot a_0}{a_0} zf'_2,\label{F2}$$ where again $u_0=\frac{\dot a_0}{a_0}$. In the coefficient form we get $${ \dot c_1\partial_1+\dots+\dot c_n\partial_n+\dots=u_1\tilde L_1+ \dots +u_n \tilde L_n+\dots}$$ where $\tilde L_kf_2=z^{k+1}f_2'$, $c_k=\frac{a_k}{a^{k+1}_0}$, $\partial_k=\frac{\partial}{\partial c_k}$.
Observe that the equation gives an identification of $T\mathcal M^{(1,0)}$ with $T\mathcal M$.
Finally, let us make an explicit calculation of $\hat{L}_0$, which for $a_0=e^t$ we continue to denote by $L_0$. Using Kirilov’s basis $L_1,L_2,\dots$ as a linear combination we write $$L_0=\sum_{m=1}^{\infty}\Pi_mL_m.$$ The coefficients $\Pi_m$ are polynomials, which can be obtained using the following recurrent formulas $$K_1=0,\quad K_m=-\sum_{j=1}^{m-1}j(m-j+1)c_{m-j}c_j,\quad
\Pi_m=mc_m+\sum_{j=1}^{m}K_{m-j+1}P_{j-1},$$ where $P_k$ are polynomials $$\label{pol1}
P_0=1,\quad P_1=-2c_1,\quad P_2=4c^2_1-3c_2,\quad P_k=-\sum_{j=1}^{k}(j+1)c_jP_{k-j},$$
Let us summarize the above considerations in the following theorem.
The Löwner-Kufarev PDE gives the distribution for the tangent bundle $T{\text{\bf S}}_0$ of non-normalized smooth univalent functions ${\text{\bf S}}_0$, that forms a sub-bundle of co-dimension 1, which is the tangent bundle $T{\text{\bf S}}$.
The equation gives another distribution, and moreover, it makes the explicit correspondence between the natural complex structure of $T{\text{\bf S}}$, as ${\text{\bf S}}$ embedded into $\mathbb C^{\mathbb N}$, and the complex structure of $T\mathcal M^{(1,0)}$ at each point $f\in {\text{\bf S}}$ defined by .
One of the reason to consider the alternate Löwner-Kufarev PDE is the regularized canonical Brownian motion on smooth Jordan curves. For all Sobolev metrics $H^{\frac{3}{2}+\varepsilon}$, the classical theory of stochastic flows allows to construct Brownian motions on $C^1$ diffeomorphism group of $S^1$. The case 3/ 2 is critical. Malliavin [@Malliavin] constructed the canonical Brownian motion on the Lie algebra ${\text{\rm Vect }}S^1$ for the Sobolev norm $H^{3/2}$. Another construction was proposed in [@Fang]. Airault and Ren [@AiraultRen] proved that the infinitesimal version of the Brownian flow is Hölder continuous with any exponent $\beta<1$.
The regularized canonical Brownian motion on ${\text{\rm Diff }}S^1$ is a stochastic flow on $S^1$ associated to the It[ô]{} stochastic differential equation $$dg^r_{x,t}=d\zeta^r_{x,t}(g^r_{x,t}),$$ $$\zeta^r_{x,t}(\theta)=\sum_{n=1}^{\infty}\frac{r^{n}}{\sqrt{n^3-n}}(x_{2n}(t)\cos n\theta-x_{2n-1}(t)\sin n\theta),$$ where $\{x_k\}$ is a sequence of independent real-valued Brownian motions and $r\in (0,1)$ and the series for $\zeta^r_{x,t}(\theta)$ is a Gaussian trigonometric series. Kunita’s theory of stochastic flows asserts that the mapping $\theta\to g^r_{x,t}(\theta)$ is a $C^{\infty}$ diffeomorphism and the limit $\lim\limits_{r\to 1^{-}}g^r_{x,t}=g_{x,t}$ exists uniformly in $\theta$. The random homeomorphism $g_{x,t}$ is called [*canonical Brownian motion*]{} on ${\text{\rm Diff }}S^1$, see [@AiraultRen; @Fang; @Malliavin; @RenZhang]. It was shown in [@AiraultRen; @Fang], that this random homeomorphism is Hölder continuous.
The canonical Brownian motion can be defined not only on ${\text{\rm Diff }}S^1$, but also on the space of $C^{\infty}$-smooth Jordan curves by conformal welding. This leads to dynamics of random loops which are not subordinated.
Elliptic operators over the coefficient body
============================================
The Kirillov first order differential operators $L_k$ generate the elliptic operator $\sum |L_k|^2$. In this section we construct the geodesic equation and find geodesics with constant velocity coordinates in the field of this operator. In particular, we shall prove that the norm of the driving function in the Löwner-Kufarev theory gives the minimal energy of the motion in this field.
Dynamics within the coefficient body
------------------------------------
Let us recall the geometry of the coefficient body $\mathcal M_n$ for finite $n$. The affine coordinates are introduced by projecting $$\mathcal M\ni f=z\Big(1+\sum\limits_{k=1}^{\infty}c_kz^k\Big)\mapsto (c_1,\ldots, c_n)\in\mathcal M_n.$$ The manifold $\mathcal M_n$ was studied actively in the middle of the last century, see e.g., [@Babenko; @Schaeffer]. We compile some important properties of $\mathcal M_n$ below:
- $\mathcal M_n$ is homeomorphic to a $(2n-2)$-dimensional ball and its boundary $\partial \mathcal M_n$ is homeomorphic to a $(2n-3)$-dimensional sphere;
- every point $x\in \partial \mathcal M_n$ corresponds to exactly one function $f\in
{\text{\bf S}}$ which is called a [*boundary function*]{} for $\mathcal M_n$;
- boundary functions map the unit disk $U$ onto the complex plane $\mathbb C$ minus piecewise analytic Jordan arcs forming a tree with a root at infinity and having at most $n$ tips,
- with the exception for a set of smaller dimension, at every point $x\in \partial \mathcal M_n$ there exists a normal vector satisfying the Lipschitz condition;
- there exists a connected open set $X_1$ on $\partial \mathcal M_n$, such that the boundary $\partial \mathcal M_n$ is an analytic hypersurface at every point of $X_1$. The points of $\partial \mathcal M_n$ corresponding to the functions that give the extremum to a linear functional belong to the closure of $X_1$.
Properties (ii) and (iii) imply that the functions from ${\text{\bf S}}$ deliver interior points of $\mathcal M_n$. The Kirillov operators $L_j$ restricted onto $\mathcal M_n$ give truncated vector fields $$L_j=\partial_j+\sum\limits_{k=1}^{n-j}(k+1)c_k\partial_{j+k},$$ which we, if it causes no confusion, continue denoting by $L_j$ in this section. In [@MPV] based on the Löwner-Kufarev representation, we showed that these $L_j$ can be obtained from a partially integrable Hamiltonian system for the coefficients in which the first integrals coincide with $L_j$.
Let $c(t)=\big(c_1(t),\ldots,c_{n}(t)\big)$ be a smooth trajectory in $\mathcal M_n$; that is a $C^1$ map $c:[0,1]\to \mathcal M_n$. Then the velocity vector $\dot c(t)$ written in the affine basis as $\dot c(t)=\dot c_1(t)\operatorname{\partial}_1+\ldots+\dot c_{n}(t)\operatorname{\partial}_{n}$ can be also represented in the basis of vector fields $L_1,\ldots,L_{n}$ (compare with ) as $$\begin{aligned}
\label{eq4}
\dot c(t) & =\dot c_1(t)\operatorname{\partial}_1+\ldots+\dot c_{n}(t)\operatorname{\partial}_{n}\\ & = u_1L_1+u_2L_3+\ldots+u_{n}L_{n}, \nonumber \end{aligned}$$ where the coefficients $u_k$ can be written in the recurrent form as $$\label{eq3}
u_1=\dot c_1,\qquad u_k=\dot c_k-\sum_{j=1}^{k-1}(j+1)\dot c_ju_{k-j}.$$ Expressing $u_k$ in terms of $c_k$ and $\dot c_k$, we get $$\label{eq8}
u_k=\dot c_k+\sum_{j=1}^{k-1}P_{j}\dot c_{k-j}.$$
One may notice that these polynomials are the first coefficients of the holomorphic function $1/f'(z)$, where $f\in {\text{\bf S}}$. In the infinite dimensional case this follows from the Löwner-Kufarev equation with $a_0=e^t$. Kirillov’s fields $L_k$ act over these polynomials as $$L_kP_n=(n-2k-1)P_{n-k}\quad n\geq k\quad\text{and}\quad L_kP_n=0 \quad n< k.$$
We define $$\begin{aligned}
\label{eq2}
\omega_1 & = & dc_1,\nonumber \\
\omega_2 & = & dc_2-2c_1\omega_1,\nonumber\\
\ldots & \ldots & \ldots\ldots\ldots\ldots\ldots,\nonumber\\
\omega_n & = & dc_n-\sum_{j=1}^{n-1}(j+1)c_j\omega_{n-j}.\end{aligned}$$ Then, $\{\omega_1,\dots,\omega_n\}$ is a conjugate to $\{L_1,\ldots,L_n\}$ basis of one-forms. Namely, $$\omega_n(L_n)=1,\quad \omega_n(L_k)=0\ \ \text{if}\ \ k\neq n.$$
If $k>n$, then the vector fields $L_k$ do not contain $\operatorname{\partial}_n$. Since the form $\omega_n$ depends only on $dc_j$ with $j<n$, then $$\omega_n(L_k)=\operatorname{\partial}_n(L_k)-\sum_{j=1}^{n-1}(j+1)c_j\omega_{n-j}(L_k)=0\ \ \text{for}\ \ k>n>n-j.$$ If $n=k$, then $$\omega_n(L_n)=\operatorname{\partial}_n(L_n)-\sum_{j=1}^{n-1}(j+1)c_j\omega_{n-j}(L_n)=1+0\ \ \text{for}\ \ n>n-j.$$ To prove the case $k<n$ we apply the induction. Let us show for $L_1$. We have $$\omega_2(L_1)=dc_2(L_1)-2c_1(L_1)=2c_1-2c_1=0.$$ We suppose that $\omega_n(L_1)=0$. Then $$\omega_{n+1}(L_1)=dc_{n+1}(L_1)-\sum_{j=1}^{n}(j+1)c_j\omega_{n+1-j}(L_1)=(n+1)c_{n}-(n+1)c_n\omega_1(L_1)=0.$$ The same arguments work for $\omega_n(L_k)$ with $k<n$.
In the affine basis the forms $\omega_k$ can be written making use of the polynomials $P_n$. We observe that one-forms $\omega_k$ are defined in a similar way as the coordinates $u_k$ with respect to the Kirillov vector fields $L_k$. Thus, if we develop the recurrent relations and collect the terms with $dc_n$ we get $$\omega_k=dc_k+\sum\limits_{j=1}^{k-1}P_jdc_{k-j}.\quad
k=1,\ldots,n.$$
By the duality of tangent and co-tangent bundles the information about the motion is encoded by these one-forms.
Hamiltonian equations
---------------------
There exists an Hermitian form on $T\mathcal M_n$, such that the system $\{L_1,\ldots,L_n\}$ is orthonormal with respect to this form. The operator $L=\sum |L_k|^2$ is elliptic, and we write the Hamiltonian function $H(c,\bar c, \psi, \bar\psi)$ defined on the co-tangent bundle, corresponding to the operator $L$ as $H(c,\bar c,
\psi, \bar\psi)=\sum_{k=1}^{n}|l_k|^2$, where $$l_k=\bar \psi_k+\sum_{j=1}^{n-k}(j+1)c_j\bar \psi_{k+j}.$$ The corresponding Hamiltonian system admits the form $$\begin{aligned}
\dot c_1 & = & \frac{\operatorname{\partial}H}{\operatorname{\partial}\bar\psi_1}= \bar l_1 \\
\ldots & = & \ldots\ldots\ldots\ldots \\
\dot c_n & = & \frac{\operatorname{\partial}H}{\operatorname{\partial}\bar\psi_n}=\bar l_n+\sum_{j=1}^{n-1}(j+1) c_j\bar l_{n-j}\\
\dot{\bar\psi}_p & = & -\frac{\operatorname{\partial}H}{\operatorname{\partial}c_p}=-(p+1)\sum_{k=1}^{n-p} l_k\bar \psi_{k+p}\\
\ldots & = & \ldots\ldots\ldots\ldots \\
\dot{\bar\psi}_n & = & -\frac{\operatorname{\partial}H}{\operatorname{\partial}c_n}=0.\end{aligned}$$ Let us observe that $$\label{eq6}\dot l_k=\sum_{j=1}^{n-k}(j-k)\bar l_jl_{j+k}.$$ Expressing $\bar l_k$ from the first $n$ Hamiltonian equations we get $$\label{eq7}\bar l_k=\dot c_k+\sum_{j=1}^{k-1}P_{j}\dot c_{k-j},\quad k=1,\ldots,n.$$ We can decouple the Hamiltonian system making use of and which leads us to the following non-linear differential equations of the second order $$\ddot c_k=\dot{\bar l}_k+\sum_{j=1}^{k-1}(j+1)c_j\dot{\bar l}_{k-l}+\sum_{j=1}^{k-1}(j+1)
\dot c_j\bar l_{k-l},$$ where $\dot{ l}_k$ are expressed in terms of the product of $\bar l_jl_{j+k}$ by , and the last products depend on $P_j$, $\bar P_j$ and $\dot c$, $\dot{\bar c}_j$ for the corresponding indices $j$ by . For example, $$\ddot c_1 = \dot{\bar l}_1=\sum_{j=1}^{n-1}(j-1)\Big(\dot
c_j+\sum_{p=1}^{j-1}P_p \dot c_{j-p}\Big)\overline{\Big(\dot
c_{j+1}+\sum_{q=1}^{j}P_q\dot c_{j+1-q}\Big)}.$$
Comparing and , we conclude that $\bar
l_k=u_k$ and $u_k$ satisfy the differential equations $$\label{eq9}\dot u_k=\sum_{j=1}^{n-k}(j-k)\bar u_ju_{j+k},$$ on the solution of the Hamiltonian system. Observe that any solution of has a velocity vector of constant length. It is easy to see from the following system $$\begin{aligned}
\label{skewsym}
\bar u_1\dot u_1 & = & 0 \bar u_1\bar u_1u_2+\bar u_1\bar u_2u_3+2\bar u_1\bar u_3u_4+3\bar u_1\bar u_4u_5+4\bar u_1\bar u_5u_6+\ldots,\nonumber\\
\bar u_2\dot u_2 & = & -1 \bar u_1\bar u_2u_3+ 0 \bar u_2\bar u_2u_4+1\bar u_2\bar u_3u_5+2\bar u_2\bar u_4u_6+\ldots,\nonumber\\
\bar u_3\dot u_3 & = & -2 \bar u_1\bar u_3u_4-1 \bar u_2\bar u_3u_5+0\bar u_3\bar u_3u_6+\ldots,\\
\bar u_4\dot u_4 & = & -3 \bar u_1\bar u_4u_5-2 \bar u_2\bar u_4u_6+\ldots,\nonumber\\
\bar u_5\dot u_5 & = & -4 \bar u_1\bar u_5u_6+\ldots,\nonumber\\
\bar u_6\dot u_6 & = & \ldots\nonumber\end{aligned}$$ Then, $$\frac{d|u|^2}{dt}=2\sum_{k=1}^{n}(\bar u_k\dot u_k+u_k\dot{\bar u}_k)=0,$$ for any $n$, thanks to the cut form of our vector fields and the skew symmetry of . The simplest solution may be deduced for constant driving terms $u_k$, $k=1,\ldots,n$. The Hamiltonian system immediately gives the geodesic $$\begin{aligned}
c_1 & = & \bar u_1(0)s+c_1(0),\\
c_2 & = & \bar u_1^2(0)s^2+\bar u_2(0)s+c_2(0),\\
c_3 & = & 3\bar u_1(0)\big(\bar u_1^2(0)\frac{s^3}{3}+\bar u_2(0)\frac{s^2}{2}+c_2(0)\big)+2\bar u_2(0)
\big(\bar u_1(0)\frac{s^2}{2}+c_1(0)s\big)+\bar u_3(0)s+c_2(0),\\
\ldots & = & \ldots\ldots\ldots\ldots\ldots\ldots\end{aligned}$$ In general, $c_n$ becomes a polynomial of order $n$ with coefficients that depend on the initial data $c(0)$ and on the initial velocities $\bar u(0)$.
The Lagrangian $\mathcal L$ corresponding to the Hamiltonian function $H$ can be defined by the Legendre transform as $$\mathcal L=(\dot c,\bar
\psi)-H=\sum_{k=1}^{n}\Big(\bar l_k\bar \psi_k+\bar
\psi_k\sum_{j=1}^{k-1}(j+1)c_j\bar
l_{k-j}\Big)-\frac{1}{2}\sum_{k=1}^{n}|l_k|^2.$$ Taking into account that $$\bar \psi_k\dot c_k =\sum\limits_{j=1}^{k-1} (j+1)c_j\bar{\psi}_k\bar{l}_{k-j}+\bar{\psi}_k\bar{l}_k.$$ Summing up over $k$, we obtain $(\dot c,\bar
\psi)=\sum_{k=1}^{n}l_k\bar l_k=\sum_{k=1}^{n}\bar u_ku_k$, that gives us $$\mathcal L(c,\dot c)=\frac{1}{2}\sum_{k=1}^{n}|u_k|^2.$$ All these considerations can be generalized for $n\to\infty$. Thus, we conclude that the coefficients of the function $p(z,t)$ in the Löwner-Kufarev PDE play the role of generalized moments for the dynamics in $\mathcal M_n$ and $\mathcal M$ with respect to the Kirillov basis on the tangent bundle. Moreover, the $L^2$-norm of the function $p$ on the circle $S^1$ is the energy of such motion.
SLE and CFT
===========
In this section we briefly review for completeness the connections between conformal field theory (CFT) and Schramm-Löwner evolution (SLE) following, e.g., [@BB; @FriedrichWerner]). SLE (being, e.g., a continuous limit of CFT’s archetypical Ising model at its critical point) gives an approach to CFT which emphasizes CFT’s roots in statistical physics.
SLE$_{\varkappa}$ is a $\varkappa$-parameter family of covariant processes describing the evolution of random sets called the SLE$_{\varkappa}$ hulls. For different values of $\varkappa$ these sets can be either a simple fractal curve $\varkappa\in [0,4]$, or a self-touching curve $\varkappa\in (4,8)$, or a space filling Peano curve $\varkappa\geq 8$. At this step we deal with the chordal version of SLE. The complement to a SLE$_{\varkappa}$ hull in the upper half-plane $\mathbb H$ is a simply connected domain that is mapped conformally onto $\mathbb H$ by a holomorphic function $g(z,t)$ satisfying the equation $$\label{SLE}
\frac{dg}{dt}=\frac{2}{g(z,t)-\xi_t}, \quad g(z,0)=z,$$ where $\xi_t=\sqrt{\varkappa}B_t$, and $B_t$ is a normalized Brownian motion with the diffusion constant $\varkappa$. The function $g(z,t)$ is expanded as $\displaystyle g(z,t)=z+\frac{2t}{z}+\dots$. The equation is called the Schramm-Löwner equation and was studied first in [@LSW1]–[@LSW3], see also [@RohdeSchramm01] for basic properties of SLE. Special values of $\varkappa$ correspond to interesting special cases of SLE, for example $\varkappa=2$ corresponds to the loop-erasing random walk and the uniform spanning tree, $\varkappa=4$ corresponds to the harmonic explorer and the Gaussian free field. Observe, that the equation is not a stochastic differential equation (SDE). To rewrite it in a stochastic way (following [@BB], [@FriedrichWerner]) let us set a function $k_t(z)=g(z,t)-\xi_t$, where $k_t(z)$ satisfies already the SDE $$dk_t(z)=\frac{2}{k_t(z)}dt-d\xi_t.$$ For a function $F(z)$ defined in the upper half-plane one can derive the Itô differential $$\label{Ito}
dF(k_t)=-d\xi_t L_{-1}F(k_t)+dt (\frac{\varkappa}{2}L_{-1}^2-2L_{-2})F(k_t),$$ with the operators $L_{-1}=-\frac{d}{d z}$ and $L_{-2}=-\frac{1}{z}\frac{d}{d z}$. These operators are the first two Virasoro generators in the ‘negative’ part of the Witt algebra spanned by the operators $-z^{n+1}\frac{d}{d z}$ acting on the appropriate representation space. All other generators can be obtained by the commutation relation $$[L_m,L_n]=(n-m)L_{n+m}.$$For any state $|\psi\rangle$, the state $L_{-1}|\psi\rangle$ measures the diffusion of $|\psi\rangle$ under SLE, and $(\frac{\varkappa}{2}L_{-1}^2-2L_{-2})|\psi\rangle$ measures the drift. The states of interest are drift-less, i.e., the second term in vanishes. Such states are annihilated by $\frac{\varkappa}{2}L_{-1}^2-2L_{-2}$, which is true if we choose the state $|\psi\rangle$ as the highest weight vector in the highest weight representation of the Virasoro algebra with the central charge $c$ and the conformal weight $h$ given by $$c=\frac{(6-\varkappa)(3\kappa-8)}{2\varkappa},\quad h=\frac{6-\varkappa}{2\varkappa},$$ and the operators $L_{-1}$ and $L_{-2}$ are taken in the corresponding representation. It was obtained in [@BB] and [@FriedrichWerner], that $F(k_t)$ is a martingale if and only if $(\frac{\varkappa}{2}L_{-1}^2-2L_{-2})F(k_t)=0$. We define a CFT with a boundary in $\mathbb H$ such that the boundary condition is changed by a boundary operator. The random curve in $\mathbb H$ defined by SLE is growing so that it has states of one type to the left and of the other type to the right (the simplest way to view this is the lattice Ising model with the states defined as spin positions up or down). The mapping $g$ satisfying ‘unzips’ the boundary. The primary operator that induces the boundary change with the conformal weight $h$ is drift-less, and therefore, its expectation value does not change in time under the boundary unzipping. Hence all correlators computing with this operator remain invariant. Analogous considerations one may provide for the ‘radial’ version of SLE in the unit disk, slightly modifying the above statements.
[99]{}
H. Airault, P. Malliavin, [*Unitarizing probability measures for representations of Virasoro algebra*]{}, J. Math. Pures Appl. [**80**]{} (2001), no. 6, 627–667.
H. Airault, J. Ren, [*Modulus of continuity of the canonic Brownian motion “on the group of diffeomorphisms of the circle*]{}, J. Funct. Anal. [**196**]{} (2002), 395–426.
K. I. Babenko, [*The theory of extremal problems for univalent functions of class $S$*]{}, Proc. Steklov Inst. Math., No. 101 (1972). Transl. American Mathematical Society, Providence, R.I., 1975.
M. Bauer, D. Bernard, [*Conformal field theories of stochastic Loewner evolutions*]{}, Comm. Math. Phys. [**239**]{} (2003), no. 3, 493–521.
L. de Branges, [*A proof of the Bieberbach conjecture*]{}, Acta Math. [**154**]{} (1985), no. 1-2, 137–152.
L. D. Faddeev, [*Discretized Virasoro algebra*]{}, Contemp. Math., 391, Amer. Math. Soc., Providence, RI, 2005, 59–67.
S. Fang, [*Canonical Brownian motion on the diffeomorphism group of the circle*]{}, J. Funct. Anal. [**196**]{} (2002), 162–179.
R. Friedrich, W. Werner, [*Conformal restriction, highest-weight representations and SLE*]{}, Comm. Math. Phys. [**243**]{} (2003), no. 1, 105–122.
I. M. Gel’fand, D. B. Fuchs, [*Cohomology of the Lie algebra of vector fields on the circle*]{}, Functional Anal. Appl. [**2**]{} (1968), no.4, 342–343.
J.-L. Gervais, [*Infinite family of polynomial functions of the Virasoro generators with vanishing Poisson brackets*]{}, Phys. Lett. B [**160**]{} (1985), no. 4-5, 277–278.
J.-L. Gervais, A. Neveu, [*Dual string spectrum in Polyakov’s quantization. II. Mode separation*]{}, Nuclear Phys. B [**209**]{} (1982), no. 1, 125–145.
A. A. Kirillov, [*Geometric approach to discrete series of unirreps for Vir*]{}, J. Math. Pures Appl. [**77**]{} (1998), 735–746.
A. A. Kirillov, D. V. Yuriev, [*Kähler geometry of the infinite-dimensional homogeneous space $M={\rm Diff}\sb +(S\sp
1)/{\rm Rot}(S\sp 1)$*]{}, Functional Anal. Appl. [**21**]{} (1987), no. 4, 284–294.
A. A. Kirillov, D. V. Yuriev, [*Representations of the Virasoro algebra by the orbit method*]{}, J. Geom. Phys. [**5**]{} (1988), no. 3, 351–363.
D. J. Korteweg, G. de Vries, [*On the change of form of long waves advancing in a rectangular channel, and a new type of long stationary waves*]{},- Phil. Mag. [**38**]{} (1895), 422–443.
P. P. Kufarev, [*On one-parameter families of analytic functions*]{}, Rec. Math. \[Mat. Sbornik\] N.S. [**13(55)**]{} (1943), 87–118.
G. Lawler, O. Schramm, W. Werner, *Values of Brownian intersections exponents I : half-plane exponents*, Acta Mathematica **187** (2001) 237-273.
G. Lawler, O. Schramm, W. Werner, *Values of Brownian intersections exponents II : plane exponents*, Acta Mathematica **187** (2001) 275-308.
G. Lawler, O. Schramm, W. Werner, *Values of Brownian intersections exponents III : two-sided exponents*, Ann. Inst. Henri Poincaré **38** (2002) 109-123.
K. Löwner, [*Untersuchungen über schlichte konforme Abbildungen des Einheitskreises*]{}, Math. Ann. [**89**]{} (1923), 103–121.
P. Malliavin, [*The canonic diffusion above the diffeomorphism group of the circle*]{}, C. R. Acad. Sci. Paris, Sér. I Math. [**329**]{} (1999), 325–329.
I. Markina, D. Prokhorov, A. Vasil’ev, [*Sub-Riemannian geometry of the coefficients of univalent functions*]{}, J. Funct. Analysis [**245**]{} (2007), no. 2, 475–492.
M. Mineev-Weinstein, P. B. Wiegmann, A. Zabrodin, [*Integrable structure of interface dynamics*]{}, Phys. Rev. Letters [**84**]{} (2000), no. 22, 5106–5109.
J. Polchinski, [*String theory*]{}, Cambridge Univ. Press, 1998.
Ch. Pommerenke, [*Über die Subordination analytischer Funktionen*]{}, J. Reine Angew. Math. [**218**]{} (1965), 159–173.
Ch. Pommerenke, [*Univalent functions, with a chapter on quadratic differentials by G. Jensen*]{}, Vandenhoeck & Ruprecht, Göttingen, 1975.
D. Prokhorov, A. Vasil’ev, [*Univalent functions and integrable systems*]{}, Comm. Math. Phys. [**262**]{} (2006), no. 2, 393–410.
J. Ren, X. Zhang, [*Stochastic flow for SDEs with non-Lipschitz coefficients*]{}, Bull. Sci. Math. [**127**]{} (2003), 739–754.
S. Rohde and O. Schramm, *Basic properties of SLE*, Ann. Math. **161** (2005), 879–920.
A. C. Schaeffer, D. C. Spencer, [*Coefficient Regions for Schlicht Functions (With a Chapter on the Region of the Derivative of a Schlicht Function by Arthur Grad)*]{}, American Mathematical Society Colloquium Publications, Vol. 35. American Mathematical Society, New York, 1950.
L. A. Takhtajan, L.-P. Teo, [*Weil-Petersson metric on the universal Teichmüller space*]{}, Mem. Amer. Math. Soc. [**183**]{} (2006), no. 861, 119pp.
A. Vasil’ev, [*Energy characteristics of subordination chains*]{}, Arkiv Mat. [**45**]{} (2007), 141–156.
A. Vasil’ev, [*From the Hele-Shaw experiment to integrable systems: a historical overview*]{}, Complex Anal. Oper. Theory, [**3**]{} (2009), no. 2, DOI 10.1007/s11785-008-0104-8.
M. A. Virasoro, [*Subsidiary conditions and ghosts in dual-resonance models*]{}, Phys. Rev. , D1 (1970), 2933–2936.
M. Volpato, [*Sulla derivabilit[à]{}, rispetto a valori iniziali ed a parametri, delle soluzioni dei sistemi di equazioni differenziali ordinarie del primo ordine*]{}, Rend. Sem. Mat. Univ. Padova [**28**]{} (1958), 71–106.
V. Zakharov, L. Faddeev, [*The Korteweg-de Vries equation is a fully integrable Hamiltonian system*]{}, Funct. Anal. Appl. [**5**]{} (1971), 280–287.
[^1]: The authors were supported by the grant of the Norwegian Research Council \#177355/V30, by the NordForsk network ‘Analysis and Applications’ grant \#080151, and by the European Science Foundation Research Networking Programme HCAA
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper presents a control synthesis algorithm for dynamical systems to satisfy specifications given in a fragment of linear temporal logic. It is based on an abstraction-refinement scheme with nonuniform partitions of the state space. A novel encoding of the resulting transition system is proposed that uses binary decision diagrams for efficiency. We discuss several factors affecting scalability and present some benchmark results demonstrating the effectiveness of the new encodings. These ideas are also being implemented on a publicly available prototype tool, ARCS, that we briefly introduce in the paper.'
author:
- 'Oscar Lindvall Bulancea[^1]'
- 'Petter Nilsson[^2]'
- 'Necmiye Ozay[^3]'
title: 'Nonuniform abstractions, refinement and controller synthesis with novel BDD encodings[^4]'
---
=10000 = 10000
Introduction
============
Automatic synthesis of embedded control software that meets its specifications by construction provides a rigorous means for the design of cyber-physical control systems. Abstraction-based techniques, where one creates a finite transition system (FTS) corresponding to the continuous or hybrid system to be controlled, and solves a discrete control synthesis problem, has attracted considerable attention in the past decade [@belta2017formal; @tabuada2009verification].
Various software tools have been developed for correct-by-construction control synthesis. These tools differ by the class of systems (e.g., discrete-time vs. continuous-time; linear, piecewise affine or nonlinear) or specifications that they can handle (e.g., simple safety or reachability [@mazo2010pessoa; @rungger2013specification; @Rungger:2016:STS:2883817.2883834]; expressive fragments of linear temporal logic [@filippidis2016; @Wongpiromsarn11]), the abstraction techniques that are used (e.g., uniform grid-based [@mazo2010pessoa; @Rungger:2016:STS:2883817.2883834], multi-scale [@mouelhi2013cosyma], or partition-based [@filippidis2016]), the way they represent the FTSs internally (symbolic or explicit) and the synthesis techniques implemented.
The key challenge in control synthesis is scalability. Factors affecting scalability include the number of discrete states in the FTS, efficiency of computation of transitions between the discrete states based on continuous dynamics, the representation of the FTS, the complexity of the specification, and the complexity of the resulting controller. For instance, structural properties of dynamics such as linearity or monotonicity [@coogan2015efficient] are shown to make the computation of transitions easy. Sparsity of the dynamics has recently been exploited together with binary decision diagrams (BDDs), a compact (i.e., memory-efficient) representation of FTSs, to obtain abstractions efficiently [@gruber2017sparsity]. However, it is in general unclear how optimizing different factors individually would affect the efficiency of solving the end-to-end synthesis problem.
This paper builds on the abstraction-refinement based incremental synthesis approach by [@nilsson2017augmented] that handles a slightly more general class of specifications than most of the earlier tools listed above. In particular, the class includes safety, recurrence and persistence components, while allowing augmented finite transition systems as the discrete model, thereby handling fairness-like assumptions. To mitigate the state-explosion problem, a nonuniform partition of the continuous state space is used. The main contribution of the present paper is a novel BDD encoding of the states that takes into account the topology of the partition and that makes it convenient to add new states in the refinement process while preserving structure. The effectiveness of the new encoding is demonstrated with examples. A prototype tool, ARCS, that implements some of these ideas is also introduced.
Overview
========
In this section we formally state the control synthesis problem and give an overview of the solution methodology. The first ingredient of the synthesis problem is a dynamical system model $$\label{eq:sys}
x^+ = f(x, u, d),$$ where $x \in
\mathcal{X}$ is the state, $u\in \mathcal{U}$ is the control input, and $d \in \mathcal{D}$ is the disturbance. The notation $x^+$ either stands for the value of $x$ in the next time step in the discrete-time setting or the derivative in the continuous-time setting. The second ingredient is the specification, which in this paper is restricted to the following fragment of linear temporal logic: $$\label{eq:spec}
\varphi = \square A \land \lozenge \square B \land \left( \bigwedge_{i \in I} \square \lozenge G^i \right),$$ where $A$, $B$, and $G^i$’s are propositions that indicate the membership of the state in a certain subset of the state space $\mathcal{X}$. The specification $\varphi$ roughly mandates that the state trajectory of should never leave the states indicated by $A$ (*safety*); that it should eventually reach the states indicated by $B$ and remain there indefinitely (*persistence*); and that it should visit the states indicated by each $G^i$ infinitely often (*recurrence*). In comparison to other fragments of linear temporal logic, formula slightly generalizes the guarantee part of a Generalized Reactivity (1) specification [@bloem2012synthesis] and is similar to the guarantee part of a Generalized Rabin (1) formula [@ehlers2011generalized]. For a more complete treatment of semantics of LTL, we refer the reader to [@baier2008principles]. Given these ingredients, the control synthesis problem can be stated as follows.
Given a system of the form and a specification of the form , find a **control policy** $\mu$, possibly dependent on the state history, and a set of initial conditions $\mathcal{X}_0\subset\mathcal{X}$ called the **winning set**, such that all closed-loop trajectories starting in $\mathcal{X}_0$ satisfy $\varphi$.
Ideally, we want to solve for the largest winning set $\mathcal{X}_0$ but since this is in general hard, we opt for trying to incrementally expand it until it is large enough, or a certificate for its maximality is obtained.
Our solution methodology consists of the following steps:
1. Construction of an abstraction in the form of a (potentially augmented) FTS: This step requires partitioning the state space into cells based on propositions, finding transitions among cells, and, in case of an augmented FTS, finding transient cells for progress groups.[^5]
2. Representing the abstraction using BDDs: This step requires deciding on an encoding of the states and representing transitions and progress groups as a BDD as they are generated in step (1).
3. Solving the discrete synthesis problem on the abstraction via fixed-point algorithms that generate a winning set.
4. Refining the abstraction: If the winning set is not satisfactory, additional cells are added to the partition using the information from the fixed points in step (3). The BDD representation of the abstraction is updated accordingly with new states, transitions, and possibly progress groups.
5. Extracting a controller from the resulting fixed point.
For the first step, we follow the optimization-based procedures in [@nilsson2017augmented]. Our tool ARCS currently supports polynomial $f$ for the dynamics, a finite set $\mathcal{U}$ for the inputs, and rectangular sets for $\mathcal{X}$ and $\mathcal{D}$. It can however be extended to any setting where computing or over-approximating reachable sets (required for encoding transitions in the FTS) and certifying transience properties (required for progress groups in an augmented FTS) are possible. This paper is primarily concerned with steps (2) through (5) above; in the rest of the paper we introduce our novel ideas and demonstrate resulting computational gains.
Representing the Abstract System
================================
As mentioned in earlier sections the continuous system is abstracted to an FTS. Formally, an FTS is a tuple $\mathcal{T} = (\mathcal{Q}, \mathcal{U},\rightarrow_\mathcal{T}, \mathcal{L})$, where $\mathcal{Q}$ is a finite set of states, $\mathcal{U}$ is a finite set of inputs, $\rightarrow_\mathcal{T}\subset \mathcal{Q}\times \mathcal{U} \times \mathcal{Q}$ is a transition relation, and $\mathcal{L}$ is a labeling function mapping each state in $\mathcal{Q}$ to a subset of propositions appearing in the formula . In order to find a control policy $\mu_\mathcal{T}$ for the FTS $\mathcal{T}$, we need to represent it with a data structure suitable for both storage (memory efficiency) and processing (time efficiency). In what follows, we discuss different representations available in ARCS, their advantages and disadvantages, with a particular focus on a novel BDD encoding.
List Representation
-------------------
The perhaps most obvious way to represent an FTS is by encoding states and actions as integers, and transitions as an array $$\label{eq:listrep}
\mathcal L = \left[(q_0, u_0, q_0'), (q_1, u_1, q_1'), \ldots, (q_{|\mathcal T|}, u_{|\mathcal T|}, q'_{|\mathcal T|})\right].$$
Evidently, this choice requires $\mathcal O(|\mathcal T|)$ memory and standard array operations such as access, insertion, search, and deletion can be done in at most $\mathcal O(|\mathcal T|)$ time, where $|\mathcal{T}|$ is the number of transitions in the system.
Eq. can be viewed as a representation of a sparse matrix in coordinate (COO) format, that has non-zero entries $q_k'$ at positions $(q_k, u_k)$. There are several other methods for sparse matrix representations that have different benefits. For instance, the compressed sparse row (CSR) format allows for efficient matrix-vector product computation. $\blacksquare$
This representation is very simple and thus easy to implement, and scales linearly in both space and time. Although this might seem like an acceptable complexity, the size of a grid-based abstraction $\mathcal T$ scales at least exponentially with the dimension of the concrete system. To illustrate the potential for improvement consider an $n$-dimensional linear system $\dot x = Ax$: it requires $n^2$ numbers (the entries of $A$) to represent in its canonical ODE form, whereas the size of a finite abstraction based on a list scales exponentially with $n$. This results from the fact that the semantics of an ODE encodes more side information than the semantics of a transition system, thus allowing the former to be more succinct. The idea of this paper is to explore whether information such as geometrical relationships embedded in an ODE can be stored as part of the encoding of $\mathcal T$ by working with more sophisticated representations.
Binary Decision Diagrams
------------------------
[In this section, we present an alternative representation of transition systems based on Binary Decision Diagrams (BDDs). We briefly overview how certain operations on transition systems can be performed with BDDs. A key design choice for this type of representations is how to encode the state and action sets as binary variables. As the main contribution of the paper, we present a novel choice for the encodings that attempts to capture underlying geometrical relationships.]{}
A BDD is a data structure for representing *boolean functions* $$B : \{0,1\}^n \rightarrow \{0,1\},$$ taking binary variables, $\binvar_1$, …, $\binvar_n$, defined with an order of evaluation $\binvar_1 < \dots < \binvar_n$. To represent a finite set $C$ with a BDD, one needs an *encoding* $E: C \to \{0,1\}^n$, an injective map from elements of $C$ to truth assignments of the variables $\{\binvar_i\}$, following the order of evaluation. The boolean function $B_C$ is then said to represent the set $C$ if such an encoding $E$ is defined on all possible elements $c$ and $$\label{eq:BDD repr. def 2}
C = \{c: B_C(E(c)) = 1\},$$ i.e. the BDD forms the characteristic function of the set [@Bryant1992BDDs]. For a given encoding such a function can easily be obtained for singletons: if the encoding of $q_k$ is the binary array $E(q_k) = (b_{k,1}, b_{k,2}, \dots, b_{k,n})$, the boolean function for that element can be constructed as $$\label{eq:Element BDD}
B_{q_k}(\binvar_1,\binvar_2,\dots, \binvar_n) = \bigwedge\limits_{i=1}^n \left\{\begin{aligned}
\binvar_i& &b_{k, i} = 1\\
\overline{\binvar}_i& &b_{k,i} = 0
\end{aligned}\right\},$$ where $\overline{\binvar}_i$ denotes negation of the variable $\binvar_i$. Then a boolean function for the whole set $C$ can be formed as $B_C = \lor_{k} B_{q_k}$.
To construct BDDs for the elements of a transition mapping $\rightarrow_\mathcal{T}$, an encoding has to be chosen to represent elements $(q_k, u_k, q'_k) \in \mathcal{Q} \times \mathcal{U} \times \mathcal{Q}$. To construct singleton BDDs according to , one needs to separate the logical variables for the different parts of the elements, while also separating those used for the initial and final transition states $q_k$ and $q'_k$. Therefore $2n + m$ variables are defined: $\binvar_{q,1},\dots, \binvar_{q,n}$ to represent the set initial states $\mathcal{Q}$, $\binvar_{u,1},\dots, \binvar_{u,m}$ for the action set, and $\binvar_{q',1},\dots,\binvar_{q',n}$ for the final state set $\mathcal{Q}'$. Having defined encodings $E_\mathcal{Q}$ and $E_\mathcal{U}$ for the set of states $\mathcal{Q}$ and set of actions $\mathcal{U}$, an encoding for the transition can be chosen as $$E_{\mathcal{T}}(q_k, u_k, q'_k) = (E_\mathcal{Q}(q_k), E_\mathcal{U}(u_k), E_\mathcal{Q}(q'_k)).$$ With such an encoding, the BDD for one transition $t = (q_k, u_k, q'_k)$ can be constructed as in : $$B_t = B_{q_k} \land B_{u_k} \land B_{q'_k}.$$ The BDD representing the entire set $\rightarrow_{\mathcal T}$ is then constructed on disjunctive normal form from the singleton BDDs as $$B_\mathcal{T} = \bigvee\limits_{i=1}^{|\mathcal{T}|} B_{t_i} = \bigvee\limits_{i=1}^{|\mathcal{T}|} (B_{q_i} \land B_{u_i} \land B_{q'_i}).$$
Working with BDDs as representations of sets, one is limited to the logical operations of boolean functions, among which $\{\land, \lor, \neg, \exists, \forall\}$ are useful. However, using BDDs is analogous to working with the sets themselves, as the basic set operations {$\cup$, $\cap$, $\setminus$} have as counterparts logical operations on the corresponding BDDs. It follows from that for sets $C$ and $D$ represented by BDDs $B_C$ and $B_D$: $$\begin{aligned}
B_{C \cup D} &= B_C \lor B_D,\\
B_{C \cap D} &= B_C \land B_D,\\
B_{C \setminus D} &= B_C \land \neg B_D.\end{aligned}$$ In less trivial set definitions, the description might make use of existential and universal quantifiers, $\exists$ and $\forall$. As logical operators, these are defined as acting on a boolean function with one of its variables according to $$\begin{aligned}
\exists_{\binvar_1} B(\binvar_1, \binvar_2) &= B(0, \binvar_2) \lor B(1, \binvar_2) \label{eq:exists def},\\
\forall_{\binvar_1} B(\binvar_1, \binvar_2) &= B(0, \binvar_2) \land B(1, \binvar_2) \label{eq:forall def}.\end{aligned}$$ But as set notation, they are used to signify conditions which have to hold for at least one, or all, elements of a certain set. The BDD equivalent of quantification over elements of a set is to use logical quantification over all variables used in describing the corresponding set BDD. We therefore define the quantifiers $\exists_C$, $\forall_C$ as being $$\begin{aligned}
&\exists_C = \exists_{\binvar_1}\exists_{\binvar_2}\dots \exists_{\binvar_n} \label{eq:exists},\\
&\forall_C = \forall_{\binvar_1}\forall_{\binvar_2}\dots \forall_{\binvar_n} \label{eq:forall},\end{aligned}$$ where the variables $\vec{z} = (\binvar_1, \dots, \binvar_n)$ are used to describe the elements in a set $C$. By expansion of $\eqref{eq:exists}$ and $\eqref{eq:forall}$ using $\eqref{eq:exists def}$ and $\eqref{eq:forall def}$ it can be seen that these operators act on a BDD $B(\vec{z}, \vec{z}')$, where the variables $\vec{z}$ are quantified, by creating new BDDs $(\exists_C B) (\vec{z}')$ and $(\forall_C B) (\vec{z}')$. These new boolean functions have the following properties: ($\exists_C B)(\vec{z}') = 1$ if and only if there exists at least one assignment for $\vec{z}$ such that $B(\vec{z},\vec{z}') = 1$, and, respectively, $(\forall_C B) (\vec{z}') = 1$ if and only if for all assignments to $\vec{z}$, $B(\vec{z}, \vec{z}') = 1$. As assignments are merely the representations of set elements in our definitions, the operators - can be used to construct BDDs that represent sets defined through quantification. For the operations above, the complexity as functions of the size of BDDs involved is as follows:
- Conjunction/disjunction of two BDDs $B_1$ and $B_2$ requires $\mathcal{O}(|B_1||B_2|)$ time, producing a BDD with the same bound in size [@Bryant1986BDDs].
- Negation and assignment of a number of variables in a BDD $B$ requires $\mathcal{O}(|B|)$ time. Negation does not change the BDD size, but the size after variable assignment is bounded by the original BDD size $|B|$ [@Bryant1992BDDs].
- Following the previous remarks regarding and , quantification of a single variable on a BDD $B$, can be achieved in $\mathcal{O}(|B|^2)$ and results in a size bounded by $\mathcal{O}(|B|^2)$.
Each operation only takes time, and produces a BDD of size, that is polynomial in the input sizes, but successive applications of these operations are required in BDD manipulation. For instance, the quantifications and have a worst-case complexity that is exponential in the input size. However, these worst-case complexities are seldom encountered in practice.
The BDD data structure is based on a reduced binary tree whose size, i.e. number of nodes, varies not only with the number of elements it represents but also with the encodings and the evaluation order defined for the variables. Choices of evaluation order and encodings are therefore vital when using BDDs and deserves careful consideration, as time and memory used by the logical operations are dependent on the size of the BDD structures involved [@Bryant1986BDDs].
As for the choice of variable ordering, an optimal choice can result in a BDD of size linear in the number of binary variables, and a bad choice can give a size exponential in the number of variables [@Bryant1986BDDs]. Finding the optimal variable ordering is a computationally hard problem [@Bollig1996] and cannot be solved exactly for any large number of variables in reasonable time, although several heuristics exist (e.g. [@Bollig95simulatedannealing], [@SomenziHeuristics]).
The choice of element encoding $E$ involves two aspects: The number of variables used in the encoding, and how each element is mapped. We investigate two kinds of encodings for the states in the abstracted system. One memory-efficient encoding that minimizes the number of variables used, and one encoding that attempts to capture the structure of the partition after iterated refinement.
#### State Encodings:
The first type of encoding, which we refer to as the *log encoding*, assumes a numbering of the states from 1 to some number $|\mathcal{Q}|$ and uses these to define the mapping in the form $$\label{eq:log encoding map}
E_\text{log} : \left\{\begin{aligned}
&\mathcal{Q} \rightarrow \{0,1\}^n,\\
&q \mapsto \operatorname{Bin}(q-1),
\end{aligned}
\right.$$ where $\operatorname{Bin}(q)$ is the binary representation of the number $q$. When the state space $\mathcal{Q}$ is expanded at refinement, a state with number $k$ is split into two new states. One of these is numbered by $k$ and the other by $|\mathcal{Q}|+1$, after which they are encoded according to . The number of variables is also incremented if $|\mathcal{Q}| = 2^n$ before refinement, i.e. all encodings for $n$ variables are used by the present states. With this encoding, the absolute minimum of $\lceil \log_2(|\mathcal{Q}|)\rceil$ variables are used to encode the states. As it simply uses the least amount of variables, it is the encoding to prefer when nothing obvious can be stated about the structure of the problem. This is the standard encoding used in some tools [@filippidis2016; @gruber2017sparsity].
The novel encoding we propose in this paper—denoted the *split encoding*—is based on the splitting procedure during refinement. As the partition grows increasingly non-uniform with time, with a possibly small area becoming increasingly fine in contrast to others, we believe that an encoding that reflects this structure can lead to computational gains.
![Example of change in split encoding after refinement of initial partition using $k=2$ variables. Refinement of $q_4$ also reaches new largest refinement depth. Bits appended during expansion, but not later assigned, are shown in red.[]{data-label="fig:refinement_example"}](figures/Split_enc_illustration){width="\columnwidth"}
Starting with a coarse initial abstraction and an encoding $E$ for the states using $k$ variables, new states created from the splitting procedure have their encoding chosen based on that of its predecessor and refinement depth. The refinement depth is a measure of how many refinements have been performed on the domain the state contains. Every cell resulting from splitting a cell with depth $d$ has a depth of $d+1$, and cells in the initial partition are defined to have depth zero. When a state $q$ with depth $d$ is split into two others, $q'$, $q''$, the new states keep the encoding of their predecessor, with a modified bit at position $k+d+1$. This bit is set to 0 for $q'$ and to 1 for $q''$. In the event that the partition reaches a new largest splitting depth, a new variable first has to be created to describe the new states, effectively expanding all encodings by one bit. The default value of the appended bit is chosen as 0. An example of both cases is shown in Fig. \[fig:refinement\_example\]. If the states of the initial partition are given unique initial encodings before refinement, this procedure also results in a unique assignment of state encodings.
To summarize, having a partition with an encoding $E$ and largest refinement depth $D$; the state $q$ with a refinement depth $d$ and encoding $E(q) = (b_{q,1},\dots, b_{q,k},\dots,b_{q,k+D})$, is split into two states labeled $q'$ and $q''$. The encodings of the new states are chosen as $$\begin{aligned}
&E(q') = E(q)\big|_{b_{k+d+1} = 0},\label{eq:split enc 0}\\
&E(q'') = E(q)\big|_{b_{k+d+1} = 1}\label{eq:split enc 1}.\end{aligned}$$ But if $d = D$, then all encodings are first expanded by a 0 bit, i.e. $$E(q) = (E(q), \thinspace 0) \quad \forall q \in \mathcal{Q},$$ before applying and with the newly expanded encodings.[^6]
With this choice of encoding, neighboring cells have similar encodings and our hypothesis is that such similarity admits a compact BDD representation by capturing the geometry of the underlying vector fields. Smaller BDDs typically result in computational gains in the synthesis step, an effect we see in simulations. However, the number of binary variables is generally larger than with the log encoding so the theoretical worst-case complexity is higher.
Solving Synthesis Problems
==========================
The general way of solving a finite LTL synthesis problem involves translating the LTL specification to a Rabin automaton and computing fixed points on the product of the transition system and the Rabin automaton. Certain LTL fragments do however admit winning sets defined as fixed points on the transition system, which avoids the potentially expensive construction of the product system. This is the case for the GR(1) fragment, as well as the fragment considered in this paper. In the following, we present the fixed-point mappings associated with and how they can be evaluated symbolically when sets are represented as BDDs.
The fundamental component of the fixed-point mappings is the backwards controlled reachability operator $\operatorname{Pre}_{\sharp_1, \sharp_2}$ defined as follows: $$\label{eq:pre def}
\operatorname{Pre}_{\sharp_1, \sharp_2}(X) \hspace{-1mm} = \hspace{-1mm} \left\{ q \hspace{-0.5mm} : \hspace{-0.5mm} (\sharp_1 u \hspace{-0.5mm} \in \mathcal U), (\sharp_2 (q, u, q') \in \rightarrow_\mathcal{T}), q' \hspace{-0.5mm} \in \hspace{-0.5mm} X \right\}.$$ Here, $\sharp_1$ and $\sharp_2$ is either $\exists$ or $\forall$ and reflect the controllability assumptions: $\operatorname{Pre}_{\exists, \forall}$ corresponds to $u$ being controllable and nondeterminism uncontrollable, whereas $\operatorname{Pre}_{\forall, \exists}$ corresponds to uncontrollable $u$ but controllable nondeterminism.
#### Computing $\operatorname{Pre}$ with list representation:
For synthesis algorithms the fundamental operator is the $\operatorname{Pre}_{\exists, \forall}(X)$ operator, which can be computed as follows:
1. Find the set $C = \operatorname{Pre}_{\exists, \exists}(X)$ of all $q$ such that there exists $(q,u,q') \in \rightarrow_{\mathcal T}$ for $q' \in X$.
2. For each $q \in C$, for each action $u$, find the set $C_{q,u} = \{ q' : \exists (q, u, q') \in \rightarrow_{\mathcal T} \}$.
3. Now for all $q$, $q \in \operatorname{Pre}_{\exists, \forall}(X)$ if and only if $q \in C$ and for some $u$, $C_{q,u} \neq \emptyset$ and $C_{q,u} \subset X$.
The procedure can be slightly modified to account for other combinations of quantifiers (i.e. $\forall, \forall$). The first step can be done via one traversal of $\mathcal L$, and the second via $|C| |\mathcal U|$ traversals. Thus the complexity for computing $\operatorname{Pre}$ is upper bounded by $\mathcal O \left( |\operatorname{Pre}_{\exists, \exists}(X)| |\mathcal U| |\mathcal T| \right)$. However, since the same sets $C_{q,u}$ are typically computed many times when evaluating a fixed point, step 2 can be discounted across multiple $\operatorname{Pre}$ operations by storing the sets $C_{q,u}$, which improves the time complexity of a single $\operatorname{Pre}$ computation towards $\mathcal O(|\mathcal T|)$ at the expense of a larger memory footprint.
#### Computing $\operatorname{Pre}$ with BDD representation:
When the set of final states $X$ is represented as a BDD $B_X$ using an encoding $E$, the set $\operatorname{Pre}_{\sharp_1, \sharp_2}(X)$ can be represented as the binary mapping $$\label{eq:pre BDD}
B_{\operatorname{Pre}_{\sharp_1, \sharp_2}(X)} = \left\{\begin{aligned}
&{\sharp_1}_{\mathcal{U}}{\exists}_{\mathcal{Q}'}(B_{\mathcal T} \land B_X), &\sharp_2 = \exists,\\
&{\sharp_1}_{\mathcal{U}}{\forall}_{\mathcal{Q}'}(\neg B_{\mathcal T} \lor B_X), &\sharp_2 = \forall,
\end{aligned}\right.$$ which can be computed symbolically from $B_X$ via quantifier elimination.[^7] The runtime of this operation ultimately depends on the sizes of intermediate results, but considering the complexity and worst-case size result of each operation involved, an upper bound can be obtained as $\mathcal{O}((|B_X||B_\mathcal{T}|)^{2^{n_u + n_q}})$, when using $n_u$ action variables and $n_q$ end state variables. The complexity can also be expressed solely in the number of variables used, if one assumes the maximum size of each intermediate BDD. Then the theoretically largest possible runtime is that of the $2^{n_q-1}$ conjunctions when expanding ${\sharp_2}_{\mathcal{Q}'}$, acting on BDDs having assigned all but $n_q + n_u$ variables, thus having complexity $\mathcal{O}(2^{3n_q + 2n_u})$.
Equipped with the $\operatorname{Pre}$ operator, we give fixed-point characterizations for the winning set of . We borrow notation from $\mu$-calculus for succinct expression of fixed points. Let $\kappa : 2^{\mathcal{Q}} \rightarrow 2^{\mathcal{Q}}$ be a mapping that is monotone with respect to set inclusion, i.e., $V \subset W \implies \kappa(V) \subset \kappa(W)$. Then the *greatest fixed point of* $\kappa$, written $\nu V \; \kappa(V)$, is the value after convergence of the set sequence $$\label{eq:largefp}
V_0 = \mathcal{Q}, \quad V_{k+1} = \kappa(V_k).$$ Correspondingly, the *smallest fixed point of* $\kappa$, written $\mu V \; \kappa(V)$, is the value after convergence of $$\label{eq:smallfp}
V_0 = \emptyset, \quad V_{k+1} = \kappa(V_k).$$ Due to monotonicity and finiteness of $\mathcal{Q}$, both these sequences converge in a finite number of steps. With this notation, the winning set of is as follows: $$\label{eq:synth_fp}
\begin{aligned}
\operatorname{Win}_{\exists, \forall} & \left( \varphi \right) = \mu V_2 \; \nu V_1 \; \bigcap_{i \in I} \mu V_0 \; \operatorname{Pre}_{\exists, \forall}(V_2) \\ & \hspace{-3mm} \cup \left( B \cap G^i \cap \operatorname{Pre}_{\exists, \forall}(V_1) \right) \cup \left( B \cap \operatorname{Pre}_{\exists, \forall}(V_0) \right).
\end{aligned}$$
Specification-Guided Abstraction Refinement
-------------------------------------------
In the event that the winning set $\operatorname{Win}_{\exists, \forall}(\varphi)$ computed via is empty, or otherwise not satisfactory (e.g., it does not cover an expected initial condition), the abstraction can be refined in an attempt to extract more information about the underlying concrete system. Instead of doing this blindly, we select refinement regions guided by the internals of the fixed point computation . Loosely speaking, for a greatest fixed point we perform refinement in the set $\operatorname{Pre}_{\exists, \exists}(V_\infty) \setminus V_\infty$ just outside the fixed point $V_\infty$, with the hope that the greatest fixed point will be enlarged in the refined system. For a smallest fixed point, refinement is instead done in $V_1 \setminus V_\infty$, where $V_k$ is the $k$’th iteration of . These ideas are illustrated in Fig. \[fig:refinement\]. For multi-level fixed points such as we select the refinement regions as the union of the refinement regions corresponding to lower-level fixed points. For more details see [@nilsson2017augmented].
(3.5,0) circle (1.5); (3.5,0) circle (1.5); (3.5,0) circle (0.75);
(0,0) circle (1.5); (0,0) circle (1.2); (0,0) circle (1.2); (0,0) circle (0.5);
in [0,45,...,360]{} [ (:0.6) – ++ (:0.5); ($(3.5,0) + (\s:1.4)$) – ++ (:-0.6); ]{}
at (0,0) [$V_1$]{}; at (0.3,0.8) [$V_\infty$]{}; at (3.5,0) [$V_\infty$]{}; at (4,1.05) [$V_1$]{};
Controller Extraction
---------------------
In addition to computing the winning set, in practice also a controller that enforces the specification inside the winning set is required. Fundamentally, such a controller can be extracted by saving the set of $u$’s satisfying the quantification in low-level calls to $\operatorname{Pre}$ in , and storing these $u$’s in a memory hierarchy whose structure depends on the type of fixed point. For instance, invariance or reachability controllers can be memoryless (i.e. state feedback), but a recurrence controller must maintain an internal memory state that switches between different reachability objectives. [Thus certain specifications trivialize controller extraction and as a result this issue is not discussed in detail for tools that handle only invariance specifications. However, control extraction adds to the overall complexity when specifications of more general form such as are considered, or when AFTSs are used.]{}
Results and Comparisons
=======================
[In this section, we present results comparing different representations of transition systems. Our toolbox ARCS, available at <https://github.com/pettni/abstr-refinement>, implements the examples in this section. ARCS has a MATLAB front-end for handling continuous dynamics, computation of transitions, and list representations, and a C back-end for BDD operations using the CUDD library [@cudd].]{}
As benchmarks we consider hydronic radiant systems for buildings, in which chilled water is run through concrete slabs to regulate the temperature of the rooms to which they are connected. The systems are controlled by turning on/off flow to any one slab, thus changing the temperature of zone $i$ according to the heating dynamics $$c_i\dot{T}_i = \sum_{j\neq i}\frac{1}{R_{ij}}(T_i - T_j) + k_i,$$ where the sum is taken over all temperature zones in the system, including other rooms, slabs, supply water sources, and the outside (the last two are assumed to have constant temperature). Thermal capacitances $c_i$, thermal resistances $R_{ij}=R_{ji}$ and nominal heat gains $k_i$ are determined by sets of parameters that together define room types. In this article we use parameters for the outside and water temperature and two room types as defined in [@nilsson2017augmented], slightly extending the framework to take into account adjacency with multiple rooms of different types, and multiple slabs in different configurations. A setup with $n_r$ rooms and $n_s$ slabs results in a dynamical system with $n_r+n_s$ continuous states (room and slab temperatures), and $2^{n_s}$ discrete control inputs (water flow through each slab turned on or off). We construct a collection of benchmark systems and perform three types of run time tests:
1. Synthesis on systems of varying size and topology,
2. Synthesis on the same system but at different levels of refinement,
3. An end-to-end synthesis-refinement procedure on one system.
For these tests a persistence specification $\varphi = \lozenge \square B$ is considered, with $B$ being the proposition corresponding to all rooms and slabs having temperatures in $[22,25]^\circ C$ and $[21 , 27]^\circ C$ respectively, in a total domain of $([20, 28] ^\circ C)^{n_r + n_s}$.
In the first test, different configurations of adjacent rooms and connecting slabs, as shown in Fig. \[fig:configs\], are used. After refining the abstraction of each system 2000 times, resulting in roughly 2000 states in each abstraction, a final synthesis is performed and timed. The final synthesis is done using the BDD representation with the suggested log and split encoding, both as is and after reordering of the BDD variables using the simulated annealing algorithm implemented in CUDD. Fig. \[fig:figure1\] shows the resulting run times for each system and representation scenario. It is worth noting that for all systems except 3, which is special in the sense that one room lacks a controller, the order from slowest to fastest run times is consistently measured as log, reordered log, split and reordered split encodings.
![Test configurations based on different building layouts. Rooms of different types (a or b) having different heating dynamics, controlled by cooling slabs (red).[]{data-label="fig:configs"}](figures/test_illustration.tikz){width="\columnwidth"}
In the second test, System 2 in Fig. \[fig:configs\] is used in measuring the synthesis time at different levels of refinement. An abstraction is obtained after a number of synthesis-refinement iterations between 500 and 4000, and the run time of synthesis on that abstraction is measured. In Fig. \[fig:complexity\] the run time of synthesis is plotted against the number of transitions present in the abstraction. Just as in the first test, the split encoding is seen to have lower run time than the log encoding for both the ordered and unordered case. It is also interesting to note how for this test, the slope of the reordered split encoding graph in this loglog-plot is close to linear and less than that of the list encoding toward the larger number of transitions.
As a final test, the synthesis-refinement procedure is ran on System 2 for one hour using the list representation, log and split encoding, and cumulative time together with runtime of each synthesis-refinement iteration is measured. In Fig. \[fig:end\_to\_end\], the cumulative time is plotted against the total number of iterations achieved, and in Fig. \[fig:end\_to\_end\_histo\], the runtime of the iterations are shown as averages over intervals of 50 iterations. As seen in Fig. \[fig:end\_to\_end\], the split encoding manages a few hundred more iterations in total than the log encoding, especially during the time the abstraction is more refined and transition relations are more complex. And in Fig. \[fig:end\_to\_end\_histo\], the average iteration time for log encoding grows at a faster rate than for the split encoding, when the partition is increasingly refined. [As a side note, overall timing for each of the methods can be improved by warm-starting synthesis at each refinement step (see [@nilsson2017augmented]) but this is not currently implemented in ARCS, and will not affect the comparative results significantly. Finally, we also note that in terms of memory requirements, the number of nodes allocated by CUDD to represent the BDDs for split encoding is approximately twice as much as that for the log encoding, with $2\%$ to $10\%$ extra memory usage at iteration steps 2000 and 3000, respectively. This extra memory usage is not surprising given that the number of binary variables in split encoding is more than the minimal number achieved by the log encoding. On the other hand, the split encoding achieves a better compression in terms of memory used per number of binary variables as shown in Fig. \[fig:memory\]. Moreover, this redundancy in representation seems to improve computation times significantly.]{}
Finally, we point out how non-uniformly the partition of the state space evolves through the abstraction-refinement process. Figure \[fig:init\_partition\] shows an initial partition of the domain for System 1 and the partition after 200 refinement steps for the same system. It is hard to capture the structure in the refined partition when starting with a log encoding of the initial state and appending one state to the end of the list of states at every refinement step, where the list is eventually encoded via . On the other hand, the split encoding is designed to capture the topological relations in the partition naturally.
Conclusions
===========
In this paper, we presented an abstraction-refinement based controller synthesis framework and, specifically, discussed several ways of representing the transition systems resulting from abstractions. We proposed a novel BDD-based encoding, namely *split encoding*, of the states of the transition system that takes into account the geometry of the underlying continuous states and how they evolve with refinement. A comparative study of various representations shows the effectiveness of the new encoding. The presented ideas are implemented in a toolbox, ARCS, which is made publicly available.
[10]{}
C. Baier and J.-P. Katoen. . MIT press, 2008.
C. Belta, B. Yordanov, and E. A. Gol. . Springer, 2017.
R. Bloem, B. Jobstmann, N. Piterman, A. Pnueli, and Y. Saʼar. Synthesis of reactive (1) designs. , 78(3):911–938, 2012.
B. Bollig, M. Löbbing, and I. Wegener. Simulated annealing to improve variable orderings for obdds. In [*In Int’l Workshop on Logic Synth*]{}, pages 5–5, 1995.
B. Bollig and I. Wegener. Improving the variable ordering of obdds is np-complete. , 45(9):993–1002, 1996.
R. E. Bryant. Graph-based algorithms for boolean function manipulation. , (8):677–691, 1986.
R. E. Bryant. Symbolic boolean manipulation with ordered binary-decision diagrams. , 24(3):293–318, Sept. 1992.
S. Coogan and M. Arcak. Efficient finite abstraction of mixed monotone systems. In [*Proceedings of the 18th International Conference on Hybrid Systems: Computation and Control*]{}, pages 58–67. ACM, 2015.
R. Ehlers. Generalized rabin (1) synthesis with applications to robust system synthesis. In [*NASA Formal Methods Symposium*]{}, pages 101–115. Springer, 2011.
I. Filippidis, S. Dathathri, S. C. Livingston, N. Ozay, and R. M. Murray. Control design for hybrid systems with tulip: The temporal logic planning toolbox. In [*Proc. Multi-conference on Systems and Control (MSC)*]{}, pages 1030–1041, 2016.
F. Gruber, E. S. Kim, and M. Arcak. Sparsity-sensitive finite abstraction. , 2017.
M. Mazo Jr, A. Davitian, and P. Tabuada. Pessoa: A tool for embedded controller synthesis. In [*Computer Aided Verification*]{}, pages 566–569. Springer, 2010.
S. Mouelhi, A. Girard, and G. G[ö]{}ssler. Cosyma: a tool for controller synthesis using multi-scale abstractions. In [*Proc. of the 16th International Conference on Hybrid Systems: Computation and Control (HSCC)*]{}, pages 83–88, 2013.
P. Nilsson, N. Ozay, and J. Liu. Augmented finite transition systems as abstractions for control synthesis. , 27(2):301–340, 2017.
S. Panda, F. Somenzi, and B. F. Plessier. Symmetry detection and dynamic variable ordering of decision diagrams. In [*Proceedings of the 1994 IEEE/ACM International Conference on Computer-aided Design*]{}, ICCAD ’94, pages 628–631, Los Alamitos, CA, USA, 1994. IEEE Computer Society Press.
M. Rungger, M. Mazo Jr, and P. Tabuada. Specification-guided controller synthesis for linear systems and safe linear-time temporal logic. In [*Proceedings of the 16th international conference on Hybrid systems: computation and control*]{}, pages 333–342. ACM, 2013.
M. Rungger and M. Zamani. Scots: A tool for the synthesis of symbolic controllers. In [*Proceedings of the 19th International Conference on Hybrid Systems: Computation and Control*]{}, HSCC ’16, pages 99–104, New York, NY, USA, 2016. ACM.
F. Somenzi. Cudd: Cu decision diagram package release 3.0.0, 2015.
P. Tabuada. . Springer Science & Business Media, 2009.
T. Wongpiromsarn, U. Topcu, N. Ozay, H. Xu, and R. M. Murray. : a software toolbox for receding horizon temporal logic planning. In [*HSCC*]{}, pages 313–314, 2011.
[^1]: KTH Royal Institute of Technology, 100 44, Stockholm, Sweden (e-mail: [email protected])
[^2]: California Institute of Technology, Pasadena, CA 91125, USA (e-mail: [email protected])
[^3]: University of Michigan, Ann Arbor, MI 48109, USA (e-mail: [email protected])
[^4]: This work is supported in part by DARPA grant N66001-14-1-4045, and NSF grants CNS1446298 and ECCS-1553873.
[^5]: Augmented finite transition systems (AFTSs) are FTSs extended with progress groups that capture a type of fairness property. In particular, progress groups identify state-action sets such that the system is guaranteed to eventually leave the corresponding state set when the actions from the set are persistently selected. This is useful in abstractions to encode sets of transient states that do not contain any invariant sets. Our results are valid for AFTSs but for simplicity of exposition we focus on FTSs in the rest of the paper. For details of AFTSs, the reader is referred to [@nilsson2017augmented].
[^6]: Appending the encodings with $0$ to have all states represented with the same number of bits is just for coding convenience. One can alternatively change the co-domain of the function $E$ to $\cup_{i=k}^n \{0,1\}^i$.
[^7]: This is kept simple for the reader, however one has to be careful when performing negations, as the domain of possible assignments could be larger than the domain of elements. As such, one would need to modify , to $B_\mathcal{Q} \land \sharp_\mathcal{U}(\forall_{\mathcal{Q}'}(\neg B_\mathcal{T} \lor B_X))$ if $\sharp_2 = \forall$, further replace $\exists_\mathcal{U}(\dots)$ with $\exists_\mathcal{U}(B_\mathcal{U} \land \dots)$ if also $\sharp_1 = \exists$, and if $(\sharp_1, \sharp_2) = (\forall, \exists)$ replace $\forall_\mathcal{U}(\dots)$ with $\forall_\mathcal{U}(\neg B_\mathcal{U} \lor \dots)$, to not include assignments corresponding to non-existent states.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
-
title: |
**Very High Angular Resolution Science with the\
Square Kilometre Array**
---
Introduction
============
The future of radio astronomy at cm-wavelengths lies with the Square Kilometre Array (SKA) — a radio interferometer currently in the design stages, that will have a total collecting area in the order of one square kilometre [see e.g. @schilizzi07]. The current design stipulates that 25$\%$ of the total collecting area will reside in a number of remote array stations at distances of between 180 km to 3000 km from the centre of the array. It is envisaged that each of these remote stations will comprise several dishes with single pixel receivers operating in the approximate frequency range 0.5 - 10 GHz [e.g. @schilizzi07]. Such an instrument will provide very high sensitivity at angular resolutions ranging from several arcseconds to one milliarcsecond.
In the following sections, we highlight the fact that high angular resolution requiring baselines greater than 1000 km provides a rich science case with projects from many areas of astrophysics, including important contributions to key SKA science. This was presented in SKA Memo 135 [@godfrey11], and the following represents a subset of the science discussed in that work.
We note that the SKA remote stations will not include the sparse or dense aperture array technologies proposed for the SKA core, nor will they involve phased array feeds on the dishes. For the remote stations, only the standard technologies — dishes with single pixel feeds, will be involved. Much of the proposed high angular resolution science does not require access to very wide fields of view or very low frequencies ($< 500$ MHz). Therefore, dishes with single pixel receivers operating in the approximate frequency range 0.5 — 10 GHz are adequate for the vast majority of proposed high angular resolution science.
The science case will continue to develop as the SKA design proceeds over the coming years, as part of the Pre-construction phase Project Execution Plan or PEP [@schilizzi11].
In §\[sec:science\] we compile a list of science drivers for the high angular resolution component of the SKA. The science cases are extracted largely from the book “Science with the Square Kilometre Array" [@carilli04], and the Design Reference Mission for SKA-mid and SKA-lo [@drm], and we include more recent advances and additional science cases. In §\[sec:conclusions\] we present the conclusions and closing remarks resulting from this work.
High Angular Resolution\
Science Case {#sec:science}
========================
Strong Field Tests of Gravity {#sec:pulsar_BH_distances}
-----------------------------
One of the major science goals of the SKA is to test relativistic theories of gravity in the strong field regime via precision timing of pulsars. This will be achieved by (1) timing relativistic binary systems — e.g. pulsar-neutron star binaries and any pulsar-black hole binaries discovered in the future, including pulsars in orbit around the Galactic Centre; and (2) monitoring an array of millisecond pulsars (a pulsar timing array) to detect gravitational waves with nanoHertz frequencies [@kramer04; @cordes04]. This science goal is discussed in the Design Reference Mission [@drm] chapters 16 and 17. So far the discussions of the strong field gravity tests using pulsars have concentrated on the critical precision timing information, but the importance of high angular resolution has not been emphasised. Here, we highlight the importance of high angular resolution to achieve the aims of this important science goal.
Approximately 100 compact relativistic binaries are expected in the SKA Galactic pulsar census [@smits09], of which some fraction ($\gtrsim 5 - 25$) are expected to be in stellar-mass black hole binary systems [@lipunov05]. The likelihood of dynamic interactions in globular clusters means that the chances of finding exotic binaries such as millisecond pulsar-BH systems is enhanced in these environments [e.g. @sigurdsson03]. However, the most common BH-pulsar binary system are likely to be normal rather than recycled (millisecond) pulsars [see @pfahl05; @lipunov05].
In this section dealing with strong field tests of gravity, we focus on relativistic binaries in which the pulsar companion is a stellar-mass black hole, neutron star, or white dwarf. The discovery of a pulsar in a sufficiently compact orbit around the supermassive black hole at the Galactic centre (GC) would also enable tests of relativistic gravity that are complementary to those enabled by pulsars in compact orbits around stellar-mass black holes. The prospects of probing the space-time of the supermassive black hole at the Galactic centre via pulsar timing measurements are discussed in detail by @liu11. For GC pulsars, an orbital period $\lesssim$ 0.3 yrs would be required to ensure that perturbations caused by the mass distribution around Sgr A\* are negligible [@liu11]. Furthermore, frequencies $\gtrsim$ 15 GHz would be required to optimise the timing precision [@liu11] which is strongly affected by pulse-broadening caused by the extreme interstellar scattering at the GC.
### Accurate pulsar distances are\
essential
Precise measurements of the proper motion and distance to each of the relativistic binaries detected in the pulsar census are essential for these systems to be used as laboratories for testing theories of gravity. Accurate distance and proper motion measurements are required in order to correct for the acceleration terms that affect the spin and orbital period derivatives. The latter parameter is of particular relevance for testing alternative theories of gravity [@cordes04; @stairs10; @kramer10] and potentially detecting, or at least constraining, extra spatial dimensions [@simonetti11]. Let $P_b$ be the binary period, $\dot{P_b}$ the corresponding time derivative, c the speed of light, d the distance and $\mu$ the proper motion of the system. The so-called Shklovskii-effect [@shklovskii70] contributes to the observed period derivative an amount $$\frac{\dot{P}_b}{P_b} = \frac{\mu^2 d}{c}$$ This effect, if not precisely accounted for, limits the precision with which theories of gravity may be tested in relativistic binary pulsars. In some cases, the magnitude of the Shklovskii-effect can be comparable to, or greater than the intrinsic orbital period derivative due to gravitational radiation [see e.g. @bell96]. A similar effect arises due to the differential acceleration of the solar system and the pulsar in the gravitational potential of the Galaxy [@damour91]. The determination of this Galactic acceleration term requires precise knowledge of the pulsar’s spatial position, as well as the Galactocentric radius ($R_0$) and speed of the solar system ($v_0$). To underline the importance of precise distance measurements, it is worth noting that the tests of relativistic gravity in the Hulse-Taylor binary system B1913+16, which currently provides one of the most precise constraints of this kind, are limited by the uncertainty in the distance, which has been determined using the pulsar’s dispersion measure to a precision of $\sim 30\%$ [@weisberg08].
As noted above, the Galactic constants $R_0$ and $v_0$ are of fundamental importance in correcting for the acceleration terms that impact the observed binary period derivative. The high angular resolution component of the SKA could provide a measurement of $R_0$ with $1\%$ precision via parallax measurements of Sgr A\* [@fomalont04].
### Trigonometric parallax measurements are required to maximise the science return
Pulsar distances, in some cases, may be determined by timing measurements alone via the method of timing parallax. The orbital motion of the Earth causes a 6 monthly variation in the pulse arrival times due to the curvature of the wavefront, and consequent periodic change in the path length from the pulsar to Earth. The amplitude of this timing parallax signature is very small: $\Delta t_{\pi} \approx 1.2 \rm{\mu s} \times \cos \beta ~ d_{\rm kpc}^{-1}$, where $\beta$ is the ecliptic latitude, and $d_{\rm kpc}$ is the pulsar distance in kpc [@ryba91]. Therefore, accurate timing parallax measurements are limited to a subset of pulsars with very high timing precision; that is, millisecond pulsars with stable timing characteristics, and preferably low ecliptic latitude [@smits11]. In contrast, the ability to determine trigonometric (imaging) parallax (Figure \[fig:deller\_pulsar\_parallax\]) depends only on the flux density and distance of the source, and is therefore applicable to a much wider range of systems.
@smits11 simulated and compared the accuracy of trigonometric parallax measurements with various methods of timing parallax distance determination, and concluded that both timing parallax and trigonometric parallax capabilities will be required to enable precision tests of gravity in the strong field regime. The results of the simulations (Figure \[fig:smits11\_histograms\]) suggest that the SKA can potentially measure the trigonometric parallax distances for $\sim$9000 pulsars up to a distance of 13 kpc with an error of 20$\%$ or better, and timing parallax distances for only about 3600 millisecond pulsars out to 9 kpc, with an error of 20$\%$ or better.
It is highly likely that some of the most interesting relativistic binary systems will not provide sufficient timing precision to allow accurate timing parallax distance determination, but could still provide excellent tests for relativistic theories of gravity. This is possible because, despite the limited timing precision, accurate measurement of long term secular trends such as the orbital period derivative, $\dot{P}_b$, can still be achieved, given a long enough time. For example, the measured uncertainty in $\dot{P}_b$ decreases approximately as T$^{-2.5}$, where T is the total time span of data for the system [@damour92].
A good example of this is the pulsar-white dwarf relativistic binary system, J1141-6545. Owing to the asymmetry in self-gravitation between the pulsar and white dwarf companion, this system provides a unique laboratory for testing alternative theories of gravity [@bhat08]. However, the young pulsar in this system exhibits significant “timing noise" which limits the timing precision [@bailes05]. Despite the timing noise, J1141-6545 is likely to provide some of the most stringent tests of alternative theories of gravity: already four post-Keplerian parameters have been measured, and the orbital period derivative for this system is expected to be determined to better than $2\%$ by 2012, at which point uncertainty in the kinematic Doppler term, or Shklovskii-effect (the term involving the pulsar distance and proper motion) will dominate the errors [@bhat08]. With this example in mind, it should be noted that many of the pulsar-black hole binaries are likely to be normal pulsars (and probably young pulsars like J1141-6545, due to evolution of the systems), rather than recycled (millisecond) pulsars [see @pfahl05; @lipunov05]. This suggests that trigonometric (imaging) parallax measurements will be required to determine accurate distances for a large fraction of pulsar-black hole binaries.
### Why is the SKA required?
The high sensitivity of the long baseline SKA is required not only to detect weak and distant pulsars, but also to provide a high density of calibrator sources surrounding the pulsars that will enable multi-view in-beam calibration, and therefore high precision astrometry [@rioja09; @fomalont04]. Owing to its high sensitivity, the long baseline component of the SKA will be able to perform multi-view in-beam calibration using several compact, closely spaced calibrator sources, the closest of which will be in the order of several arcminutes from the target [see @godfrey11]. This technique will provide extremely accurate phase calibration at the position of the target, and provide astrometric precision of order 15$\mu$as at 1.4 GHz [@fomalont04]. Observations at frequencies below $\sim$ 5 GHz are affected by ionospheric refraction, but the ionospheric effects may be calibrated out using a wide bandwidth [@brisken00]. Only with the substantial improvement in sensitivity provided by the SKA will high precision astrometry on weak pulsars (and other weak sources) be possible.
### Benefits of high angular resolution to the pulsar timing array
High angular resolution could also be important in establishing the pulsar timing array (PTA) [@smits11]. Accurate astrometric information reduces the amount of observing time required to obtain a coherent timing solution by breaking the degeneracies between position uncertainty and pulsar spin-down [@smits11]. In the absence of accurate positional information, this can take 12 months or more. Therefore, the high angular resolution component of the SKA will assist the selection of stable millisecond pulsars to be included in the pulsar timing array.
Further, the PTA may detect the gravitational wave signal from individual nearby binary black holes. In that case, precise distances to the pulsars in the PTA are required to enable a precise measurement for the gravitational wave source location [@lee11].
Lastly, the high angular resolution component of the SKA will compile a significant sample of SMBH binaries (see §\[sec:SBBH\]). The identification of a large sample of SMBH binaries would enable statistical studies of the inspiral rates in various phases of the binary evolution. The inspiral rates, and the possible existence of a “stalling radius" are important factors in the interpretation of the gravitational wave background that will be investigated with the pulsar timing array [@jaffe03].
Modelling the Large-Scale Galactic Magnetic Field using Pulsars {#sec:magnetic_fields}
---------------------------------------------------------------
Wavelet tomography using a grid of thousands of pulsars with known rotation measures (RMs), dispersion measures (DMs) and distances will provide the best possible map of the Galactic magnetic field and electron density on large ($\gtrsim 100$ pc) scales [@stepanov02; @noutsos09; @gaensler04; @beck04; @gaensler06]. So far the discussions of mapping the Milky Way magnetic field using pulsars have concentrated on the ability to search for and identify many thousands of pulsars, but the importance of high angular resolution has not been emphasised. Here, we highlight the importance of high angular resolution to achieve the aims of this important science goal.
The DM and RM for a grid of thousands of pulsars will be obtained via the SKA Galactic pulsar census. The final ingredient to enable accurate tomographic models of the large scale Galactic magnetic field – accurate distance estimates to each of the pulsars – will require trigonometric parallax measurements to thousands of pulsars. Currently, distance estimates to pulsars are most commonly obtained via the pulsar’s dispersion measure combined with the galactic electron density model. Distance estimates using this method are typically uncertain by tens of percent, and can be in error relative to accurate parallax measurements by more than a factor of 2, due to the large uncertainty in the electron density model [@deller09b]. Precise pulsar distances will require either parallax distance measurements, or an improved electron density model, which itself will require parallax distance measurements to a large sample of pulsars [@cordes04]. Therefore, precision astrometry is a requirement for the SKA to enable the best possible model of the large scale Galactic magnetic field. Mapping the magnetic field of the Milky Way provides an excellent opportunity to address the issues surrounding the generation and preservation of galactic magnetic fields. The importance of understanding the large scale Galactic magnetic field configuration in the context of fundamental questions of astrophysics is discussed at length in e.g. @gaensler04 [@beck04 and refs. therein].
Imaging protoplanetary disks at centimetre wavelengths {#sec:ppd}
------------------------------------------------------
The scientific motivation for obtaining high angular resolution radio images of protoplanetary disks (Figure \[fig:ppd\_simulation\]) is three-fold. Firstly, it will enable imaging of various structures in the disk such as density waves and radial gaps formed by the interaction of the disk with a planetesimal [@wilner04]. Secondly, it will enable studies of the spatial dependence of spectral signatures relating to different grain properties in the disk [@greaves09]. Thirdly, imaging the line emission will probe the kinematics and effects of photoevaporation in the disk surface layers [@kamp07].
![Image of surface density structure in a protoplanetary disk from a smooth particle hydrodynamics simulation. This image shows the surface density structure of a 0.3 M$_\odot$ disk around a 0.5 M$_\odot$ star. A single dense clump has formed in the disk (upper right), at a radius of 75 AU and with a mass of $\sim$ 8 M$_{\rm Jupiter}$. Figure reproduced from Greaves et al., 2009, MNRAS: Letters, Vol. 391, pg. L76, with permission from John Wiley and Sons. \[fig:ppd\_simulation\]](f3.pdf)
Grains in protoplanetary disks grow from sub-micron sizes up to mm sizes by sticking together in low-velocity collisions. Larger grains tend to shatter in collisions rather than sticking together. How, and under what conditions, do the mm-sized grains overcome this barrier to become pebble sized grains? This question is the subject of ongoing debate, and is a question that may be addressed with the high angular resolution component of the SKA. Dust particles emit inefficiently at wavelengths larger than their size, and therefore emission at cm-wavelengths provides evidence for pebble sized grains, which in turn provides evidence for significant progress towards planet formation. The high angular resolution component of the SKA will address the following questions: Where does the growth of decimetre-sized grains occur within the disk? Are the grains clumping into protoplanets? In what environments do these large grains occur (stellar age, spectral type, etc.)? Such information will benefit our understanding of planet formation and improve models of protoplanetary disks [@wilner04; @wilner05; @greaves09; @natta07]. The reader is referred to @greaves09 for a more detailed discussion of the science case for imaging protoplanetary disks at cm-wavelengths.
Imaging protoplanetary disks with the SKA was initially proposed for frequencies in the range 20 – 35 GHz [@wilner04]. However, studies of protoplanetary disks can be carried out in the frequency range $\lesssim$ 10 GHz [@hoare09; @greaves10; @greaves09]. It is expected that the SKA will be able to image in detail the distribution of large dust particles in the disks around hundreds of nearby young stars at $\nu \lesssim 10$ GHz [@wilner05]. Initial estimates of the technical requirements indicate the need for very high sensitivity ($\sim$ 100 nJy/beam) on long ($\sim$ 1000 km) baselines [@greaves10]. This would enable $\sim$ 5 - 10 GHz observations of Earth analogues forming in southern star clusters at $\sim$ 20 - 60 pc (the $\beta$ Pic, TW Hya, AB Dor, Tuc/Hor groups). The e-MERLIN Legacy Project “PEBBLES" is aimed at studying the centimetre emission from pebble sized dust grains to show where and when planet-core growth is proceeding, and to identify accreting protoplanets. The initial results of the PEBBLES e-MERLIN survey will help to inform the scientific and technical requirements for this project with the SKA.
@kamp07 propose that mapping the line in nearby systems will also be an important tool for studying circumstellar disks with the high angular resolution component of the SKA. Neutral Hydrogen 21cm line emission traces a layer near the disk surface that is directly exposed to soft UV irradiation from the parent star, but shielded from the ionising UV and X-ray ($h \nu > 13.6$ eV) radiation by the outer layer of the disk. High angular resolution SKA observations of 21cm line emission will probe the kinematics of protoplanetary disks, as well as the effects of irradiation and photoevaporation at the surface layer.
In addition to these primary scientific motivations, high angular resolution could potentially be used to pin-point the location of any extra-terrestrial intelligence (ETI) signals detected from planets orbiting relatively nearby stars [@morganti06], by direct imaging and measuring the orbit of the planet.
Resolving AGN and Star\
Formation in Galaxies
-----------------------
At sub-mJy flux densities, the radio source counts at GHz frequencies are thought to be dominated by star-forming galaxies, as opposed to AGN which dominate source counts at higher flux densities [e.g. @seymour08]. Without morphological information or a measurement of brightness temperature, it is generally not possible to determine, for a given galaxy, whether the observed radio flux is dominated by emission from a compact, nuclear starburst or an active galactic nucleus [@norris90]. The brightness temperature of a radio source indicates which process, AGN or star formation, dominates the radio emission: starbursts are typically limited to brightness temperatures of $T_b \lesssim 10^5$ K, and this clearly distinguishes them from the compact cores of AGN, which exhibit brightness temperatures $T_b >> 10^5$ K [@norris90; @condon92]. Baselines longer than 3000 km are required to unambiguously distinguish AGN and star-formation in sources up to redshift $z = 7$ with flux densities down to at least 30 $\mu$Jy [@drm chapter 2]. Discriminating between AGN and starburst galaxies will be possible in most cases based on the morphological information provided by high angular resolution images [eg. @garrett99].
It is widely believed that AGN play an important role in the growth and evolution of galaxies. The interaction between the AGN and the surrounding medium may promote star formation at high redshift [e.g. @klamer04; @elbaz09] and/or suppress star formation at lower redshifts [e.g. @croton06]. A powerful approach to addressing questions on the relationship between AGN activity, black hole growth, and galaxy evolution, will be deep, high-resolution imaging with the SKA to detect and distinguish between the first starburst galaxies and the first AGN jets, and to determine the frequency of occurrence of low luminosity AGN in different galaxy types [@drm chapter 2]. This will enable a determination of the full range of SMBH masses and accretion rates and how these relate to galaxy histories.
This aspect of high angular resolution SKA science is discussed in detail in the SKA Design Reference Mission [@drm chapter 2]. The goal will be to conduct a high angular resolution SKA survey to obtain a statistically significant sample of galaxies through which to explore the contribution and role of AGNs versus star formation in galaxy evolution. The high angular resolution SKA survey will be coordinated with other multi-wavelength surveys to maximise the scientific return, and an additional benefit will be in studying the cosmic evolution of AGN activity, which will address important questions relating to radio AGN, such as the lifetimes, duty-cycles, fuelling and triggering mechanisms.
The first generation of AGN jets
--------------------------------
The discovery of powerful distant quasars at $z \gtrsim 6$ indicates that supermassive black holes $>10^9 M_\odot$ existed at that time. This suggests that the first supermassive black holes formed before, or during, the epoch of reionisation. Indeed, it has been suggested that AGN jets may have played a key role in the formation of some of the first stars and galaxies in the universe, through jet-induced star formation [@klamer04; @silk05; @elbaz09; @elbaz10].
@falcke04a suggest that the first generation of AGN jets produced by accreting supermassive black holes will be strongly confined by their dense environment and appear as distant Gigahertz Peak Spectrum (GPS)-like sources — that is, faint, compact sources with unusually low turn-over frequencies. The turn-over frequency, $\nu_{\rm peak}$ and linear size, $L$, of GPS and Compact Steep Spectrum (CSS) sources are found to follow an expression of the form $$\nu_{\rm peak} = 0.62 \left( \frac{L}{\rm kpc} \right)^{-0.65} \> \rm{GHz},$$ which results from the basic properties of synchrotron self-absorption [@falcke04a]. Since the source size and turn-over frequency of GPS sources are correlated but angular size and frequency scale differently with redshift, the first AGN jets should stand out from their low redshift counterparts in the parameter space defined by angular size, turn-over frequency, and flux density (see Figure \[fig:first\_SMBH\]).
![Plot of a combination of the turn-over frequency and angular size (size $\times \nu_{\rm peak}^{1.54}$) versus the peak flux density for a sample of GPS sources. Size, turn-over frequency, and flux density roughly form a fundamental plane for GPS radio galaxies. Standard GPS sources found at $z \sim 1$ occupy the upper right of the plot. High redshift “GPS-like" sources are expected to stand out from their low redshift counterparts, and occupy the lower left portion of the plot. See @falcke04a for details. Figure reproduced from New Astronomy Reviews, Vol. 48, Falcke et al., “Compact radio cores: from the first black holes to the last", pg. 1169, Copyright 2004, with permission from Elsevier. \[fig:first\_SMBH\]](f4.pdf)
@falcke04a suggest the following strategy for finding the first generation of AGN jets in the universe:
- a shallow all-sky multi-frequency survey in the range 100 – 600 MHz down to 0.1 mJy at arcsecond resolution;
- identification of compact, highly peaked spectrum sources in that frequency range;
- identification of empty fields in the optical;
- re-observation to exclude variable sources;
- observations with long baselines and resolutions of $\sim$10 mas to determine sizes and to pick out the ultra-compact low-frequency peaked (ULP) sources;
- spectroscopic confirmation of remaining candidates with observations or by other means.
The stated goal of 10 mas resolution, at a frequency of 1.4 GHz, would require baseline lengths up to $\sim$ 4000 km.\
### Radio/CO Studies of high redshift AGN Jets
@klamer04 reviewed molecular gas observations for a sample of $z>3$ galaxies, and found that the gas and dust are often aligned with the radio emission. Based on these results, they proposed a scenario in which CO is formed at the sites of star formation that are triggered by relativistic jets, as is seen in some nearby sources (eg. Cen A, 3C40). High sensitivity, high angular resolution imaging of high redshift radio galaxies will be required to complement high redshift CO imaging with ALMA, in order to study the relationship between radio jets and early star formation. Resolution of order tens of mas will likely be required at low frequency ($\lesssim 1.4$ GHz) to map the radio structures in detail.
Exploration of the Unknown
--------------------------
The Exploration of the Unknown has been identified as an important guiding principle for the design of the SKA [@carilli04; @wilkinson04]. This recognises the discovery potential provided by instruments that are capable of probing unexplored regions of parameter space. Whilst high angular resolutions are reached with existing radio telescopes, this domain has not been explored at the sensitivity of the SKA. The combination of high sensitivity and high angular resolution with the SKA will increase the observational phase space being searched, by opening up a large, unexplored region of the flux density–angular size plane. Observations at mas-scale resolution will, for the first time, be possible for thermal and non-thermal emission regions with brightness temperatures as low as hundreds of Kelvin. Current VLBI networks are, in general, limited to non-thermal sources with brightness temperatures $\gtrsim 10^6$ K. The combination of high sensitivity with a broad range of angular resolution up to mas-scales will provide greater discovery potential for the SKA. Furthermore, the ability to perform high angular resolution follow-up of transient radio sources will maximise the science return of transient searches, as discussed below.
### Transients
High angular resolution will play an important role in localising, identifying and understanding transient radio sources. Arcsecond resolution may be sufficient to identify the host galaxies of extragalactic fast transients, and follow-up spectroscopy of the host galaxies would provide the redshifts. However, mas-scale resolution could potentially localise transient sources on a much finer scale and help to determine their nature. High angular resolution follow-up observations of newly discovered classes of radio source would be of great benefit to understanding the source physics. Resolving the source morphology and its evolution could provide information on the energetics of the event and environment of the source. For the slower transient sources (with time-scales of weeks or longer), high angular resolution would enable measurement of the source proper motion which could discriminate between Galactic and extragalactic events. This would be particularly important if sources were found to be unresolved with no optical counterparts.
Long baselines are also an excellent discriminant between RFI and genuine astronomical events [@wayth11; @thompson11]. A triggered buffer [e.g. @macquart10b] would allow for off-line analysis of the transient sources, and would function as follows:
- Data from antennas on long baselines would be stored for a couple of minutes in a rolling buffer.
- A transient source detected within the long baseline field-of-view (effectively the 15m antenna primary beam) would trigger the download of this buffer for post-processing.
- The station beams could then be formed in the direction of the transient source whose location would be determined by the SKA core to within a few arcseconds. A rolling buffer is not required for the antennas of the SKA core, since these antennas will have access to the whole field of view, and the standard output would enable the transient source position to be determined to within a few arcseconds.
A pilot survey (V-FASTR) for VLBI detection of fast transients using a triggered buffer is currently being implemented on the VLBA [@wayth11]. The results of the V-FASTR survey will inform the technical requirements for this experiment with the SKA.
Binary Supermassive Black Holes {#sec:SBBH}
-------------------------------
Binary supermassive black holes play an important role in a number of areas of astrophysics, including the formation and evolution of galaxies, galactic dynamics, and gravitational wave science. Hierarchical structure formation models predict that a significant fraction of supermassive black holes reside in binary systems [@volonteri03], and these systems will have a strong impact on the central galactic environment [e.g. @merritt06]. Simulations of binary black hole evolution in a galactic environment suggest that the inspiral efficiency (that is, the rate of decay of the binary orbital radius) may decrease at an orbital radius of $0.001~ {\rm pc} \lesssim r \lesssim 10$ pc [@yu02], potentially leaving a fraction of SMBH binaries “stalled" for extended periods of time at these orbital radii.
The identification of a large sample of SMBH binaries would enable statistical studies of the inspiral rates in various phases of the binary evolution. This will be an important step in studies of galaxy merger rates and understanding the dynamical processes responsible for removing angular momentum from these systems, and delivering them to the gravitational wave dominated phase of evolution. The inspiral rates, and the possible existence of a “stalling radius" are important factors in the interpretation of the gravitational wave background that will be observed by the pulsar timing array [@jaffe03]. Statistical studies would also allow measurements of the influence of accretion versus mergers in SMBH growth, and lead to a more precise estimate of binary merger rates [@burke-spolaor11].
Nearby binary systems that are sufficiently massive may generate gravitational radiation strong enough to enable the object to be resolved above the stochastic background [e.g. @sesana09]. Whilst it may not carry a high probability [@burke-spolaor11b; @sesana09], the detection of both electromagnetic and gravitational wave emission from a nearby SMBH binary system would have a great scientific impact. Identification of the sky position and rough orbital solution for a nearby binary would not only raise the sensitivity of the pulsar timing array to the object manyfold [@jenet04], but allow a study of the impact of the binary system on the host galaxy dynamics.
High angular resolution imaging is an effective method of searching for SMBH binaries over a wide range of orbital radii, at both high and low redshift. Binary BH candidates may be identified by surveying a large number of radio-emitting AGN (which could be initially identified in existing, lower resolution surveys) to look for dual, compact, weakly polarised, flat-spectrum radio cores [@burke-spolaor11]. Particular classes of AGN thought to harbour binary black holes may be targeted [@tingay11]. Some SMBH binary merger models predict ejected AGN, which could be revealed by astrometric measurements of AGN showing an offset from the optical host’s kinetic centre.
Source statistics are rather uncertain, since the binary black hole inspiral time-scale, and the probability that both black holes in the binary system will be radio loud, are unknown factors. At present, only one paired supermassive black hole system at a separation much less than 1 kpc is known (and supported by multi-wavelength evidence). This system, 0402+379 [@rodriguez06], was first identified via VLBI imaging as a candidate binary supermassive black hole (see Figure \[fig:binary\_SMBH\]). In a search of archival VLBA data aimed at SMBH binary detection, this was the only binary detected from a sample of more than 3000 radio loud AGN [@burke-spolaor11]. The results indicate that the VLBA is limited by (u, v)-coverage, sensitivity and dynamic range, rendering a large statistical study unfeasible. However, these crucial capabilities are realised by the SKA. The great improvement in sensitivity and dynamic range will increase the detection efficiency by allowing weaker binary companions to be identified, and weaker AGN to be searched. Improved sample selection may also significantly improve the binary detection efficiency.
It is likely that tens of thousands of AGN must be surveyed in order to compile a significant sample of SMBH binaries, and the sensitivity of the SKA will be crucial in this regard, reducing the required integration time per source, and enabling a much larger sample of objects to be searched. Such a survey could be done in combination with a strong gravitational lens survey and absorption against AGN survey. Angular resolution of $\sim$ 1 mas could resolve projected separations of 8.5 pc at all redshifts, and sub-pc separations for the nearest galaxies. This survey could feasibly be carried out at frequencies $\nu \sim 5 - 10$ GHz.
X-ray binary systems and\
relativistic jets {#sec:X-ray_binaries}
-------------------------
### Jet Formation and Evolution
Understanding the connection between accretion and jet production has implications for the understanding of AGN and $\gamma$-ray bursts as well as X-ray binaries. X-ray binary systems (XRBs) provide a unique tool to study the coupling between jet production and accretion flow, due to the rapid evolution of the systems through a wide range of characteristic accretion states (on the time-scale of weeks to months), and the associated rapid changes in jet characteristics [@fender04].
There exist two different classes of X-ray binary jet that show dramatically different spectral and morphological characteristics. These are the compact, steady “hard-state" jets, and transient “flaring-state" jets [see e.g. @fender06; @fender10]. The two different jet classes are associated with different characteristic X-ray states, and a typical XRB will transition between the two jet classes on varying timescales, in unison with transitions between X-ray spectral states.
Transient jets in flaring XRBs are produced during outbursts in which bright, optically thin jet components are seen moving at relativistic speeds away from the core. This type of radio jet is associated with a transition from a hard power-law X-ray spectrum to a softer power-law X-ray spectrum, and later, a thermal disk blackbody dominated X-ray spectrum. Unlike the compact, steady “hard-state" jets, transient jets are typically resolved at VLBI scale resolution, and the optically thin, relativistic jet components are often observed to move away from the core on a timescale of hours. Relativistic ejections in X-ray binary systems can exhibit significant amplitude and structural changes over the course of a typical observation of several hours. Thus, high sensitivity, high angular resolution radio observations with good snapshot (u, v)-coverage are required to enable high time resolution “movies" of these relativistic outflows and avoid the problems that arise from rapid evolution of the jet morphology and brightness within a single observation [@tingay95; @mioduszewski01]. Such observations are crucial in order to tie jet ejection events to X-ray timing and spectral changes in the accretion flow. The high angular resolution component of the SKA will be of fundamental importance in this regard, particularly in the case of transient jets from neutron star XRBs, and even accreting white dwarf systems, about which very little is currently known, and which may be fainter than transient jets from black hole XRBs. It is certainly the case that the compact hard-state jets from neutron star XRBs are inherently fainter than black hole XRBs at the same X-ray luminosity [@migliari06]. By comparing the jets produced by accreting black hole, neutron star and white dwarf systems, the relationship between jet formation and system parameters (e.g. depth of the potential well, stellar surface, stellar magnetic field, black hole spin etc.) can be determined. Understanding the similarities and differences between disk-jet coupling in black hole, neutron star and white dwarf systems is a crucial step in understanding the jet production mechanisms and the role played by various physical parameters.
High angular resolution will also be required to resolve X-ray binary systems in nearby galaxies from the background emission [@fender04].
### Precise distances and luminosities
Determining the physical characteristics of an object depends critically on knowing its distance. At present, X-ray binary distances, and therefore luminosities, have significant fractional uncertainties. Distances are typically only known to within a factor of two [@jonker04]. Due to the limited sensitivity of existing VLBI arrays only three X-ray binary parallax distances have been measured to date [@bradshaw99; @miller-jones09b; @reid11]. The high sensitivity and astrometric precision of the SKA will enable precise parallax distances to be measured for a large number of X-ray binary systems, and thereby enable a number of fundamental questions to be addressed, for example, by what factor can Galactic X-ray binary systems exceed their Eddington luminosities? This issue is relevant to the interpretation of ultra-luminous X-ray sources (ULXs), from which the existence of intermediate-mass black holes has been inferred. Furthermore, it is claimed that a discrepancy between the quiescent luminosities of black hole and neutron star X-ray binaries provides evidence for the existence of event horizons in black holes [@garcia01]. Accurate luminosity measurements are required to test this claim. Accurate distances will also enable more precise estimates of the basic physical parameters such as component masses, orbital orientation, and black hole spins, as recently demonstrated by @reid11 and @gou11.
### The formation of stellar mass black holes
Compiling the full 3-dimensional space velocities for a large sample of X-ray binaries will provide constraints on theoretical models of stellar mass black hole formation. It is generally accepted that neutron stars receive a “kick" during their formation, due to intrinsic asymmetries in the supernova explosion or the recoil due to the associated mass ejection [see e.g. @nordhaus10 and refs therein]. Supernova kicks are invoked to explain the anomalously high space velocities that are common among pulsars. It is currently not known whether all stellar mass black holes receive kicks during formation. Theoretical models predict that the highest mass black holes are formed via direct collapse of the progenitor star, with little mass ejection. These systems are not expected to show anomalously high space velocities. The less massive systems are thought to form in two stages: initially a neutron star is created in a supernova explosion, followed by fallback of ejected material which pushes the compact object over the stable mass limit, resulting in the formation of a stellar mass black hole [@fryer99; @fryer01]. These lower mass black holes are expected to exhibit high velocities, similar to neutron stars.
Do the most massive black holes form via direct collapse, rather than a two stage process involving an initial supernova explosion? What is the mass threshold between these two black hole formation mechanisms? How does binarity and compact object mass affect the supernova explosion? Accurate distances and proper motions for a large sample of accreting black holes, when combined with data in other wavebands, can address these important questions [@miller-jones09a].
Very few X-ray binaries have accurate distance [@bradshaw99; @miller-jones09b; @reid11] or proper motion measurements [e.g. @mirabel01; @dhawan07]. Astrometry on X-ray binaries is only possible during the so-called “low-hard" or quiescent states. During these states, in which X-ray binaries spend the majority of their time, there exists a faint, steady, often unresolved radio jet, and there is no contribution from the bright mas-scale jet that exists during the flaring states. The closest black hole X-ray binary in quiescence, A0620-00, at a distance of 1.06 kpc [@cantrell10], has a flux density of 51 $\mu$Jy/beam at 8.5 GHz and lies on the fundamental plane of black hole activity [@gallo06] that defines the relationship between black hole mass, X-ray luminosity and radio luminosity [@merloni03]. Scaling from the flux density of A0620-00 using $S_{\nu} \propto 1 / d^2$ with the best current distance estimates [@jonker04] for the known population of X-ray binaries [@remillard06], we expect that the SKA will be capable of compiling accurate proper motions and parallaxes for several tens of black hole X-ray binaries. This would enable statistical studies of the velocities of stellar mass black holes. These conclusions are based on an assumed SKA image sensitivity of $\lesssim$ 100 nJy/beam. More robust estimates of the expected outcomes will be considered in an upcoming paper (Miller-Jones et al., in preparation).
![VLA 5 GHz image overlaid on an H$\alpha$ image of the nearest relatively face-on spiral, M33. Crosses mark the locations of young, dense regions. Due to limited sensitivity, the VLA only samples the top of the initial mass function — the SKA will distinguish UCregions and probe more completely the region population. Figure reproduced from New Astronomy Reviews, Vol. 48, Hoare, “Star formation at high angular resolution", pg. 1332, Copyright 2004, with permission from Elsevier. \[fig:M33\_UCHII\]](f6.pdf)
Mapping high mass star formation in nearby galaxies
---------------------------------------------------
The combined sensitivity and high angular resolution of the SKA will permit detailed studies of extragalactic regions for the first time. Global aspects of massive star formation, as traced by the region population, are best studied in nearby, face-on spirals such as M33 [@hoare04]. Figure \[fig:M33\_UCHII\] shows the location of some regions in M33 discovered with the VLA overlaid on an H$\alpha$ image of the spiral galaxy. Only the top end of the initial mass function is accessible with the EVLA due to the limited sensitivity. The increased sensitivity of the SKA will be of fundamental importance in this regard. The high angular resolution component of the SKA will be able to distinguish ultra-compact regions, which are young and therefore most relevant to identifying conditions at the star’s birth-place. Ultra Compact (UC) regions are typically deeply embedded in their parent molecular cloud and so cannot be studied at optical or near-IR wavelengths. The SKA will be capable of detecting individual UCregions out to a distance of nearly 50 Mpc, and will be able to resolve UCregions from their surrounding environment out to a distance of 1 Mpc [@johnson04]. Questions such as “what triggers high mass star formation?" are much easier to answer in nearby spirals than in the Milky Way because a wider range of conditions can be investigated and there are no line-of-sight issues with everything lying in the Galactic plane [@hoare04; @johnson04].
The SKA will determine the exact location of massive star formation relative to other protostars, density enhancements in the molecular gas, shock fronts and other features of the ISM, and will enable an investigation into the relationship between properties of star formation and environmental parameters such as metallicity, pressure, turbulence, stellar density, triggering scenarios, and how star formation differs in “burst" and quiescent modes [@johnson04; @hoare04].
Other Science
-------------
The preceding sections represent just a selection of the science case, and demonstrate that the high angular resolution component of the SKA provides a rich science case covering many areas of astrophysics, including important contributions to key SKA science. Due to length constraints, we do not consider every possible scientific application. Further science enabled by the high angular resolution component of the SKA includes [see @godfrey11]:
- Strong gravitational lensing;
- Small-scale structure and evolution in AGN Jets;
- absorption against AGN;
- Absolute Astrometry and Geodesy;
- Relative Astrometry: Parallax and Proper Motions;
- Stellar winds/outflows;
- Star-formation via studies of astrophysical masers;
- Stellar atmospheres:
- Imaging stellar atmospheres;
- Resolving stellar radio flares;
- Parallax and proper motions of radio detected normal stars.
- Spatial and temporal changes in the fundamental constants;
- Ultra High Energy Particle Astronomy at $\gtrsim$ 2 degree angular resolution via the Lunar Cherenkov technique;
- Scattering:
- Probing the Intergalactic Medium via Angular Broadening;
- Resolving AU-scale structure in the ISM via diffractive scintillation;
- Extreme scattering events.
- Spacecraft tracking.
This is by no means meant as an exhaustive list of science enabled by the high angular resolution component of the SKA, and the science case will continue to develop over the coming years, as part of the Pre-construction phase Project Execution Plan or PEP [@schilizzi11]. .
Conclusions {#sec:conclusions}
===========
High angular resolution requiring baselines greater than 1000 km provides a rich science case with projects from many areas of astrophysics, including important contributions to key SKA science. Much of the high angular resolution science can be achieved within the approximate frequency range 0.5 — 10 GHz, or can be recast for this frequency range, and the vast majority of high angular resolution science does not require access to wide fields of view.
Acknowledgments {#acknowledgments .unnumbered}
===============
We wish to thank the following people for helpful discussions: Joe Lazio, Clancy James, Jane Greaves, Richard Dodson, Maria Rioja and Tim Colegate.
Bailes, M. 2005, Binary Radio Pulsars, Proceedings of IAU Symposium 328, ASP Conference Series, 328, 33 Beck, R., & Gaensler, B. M. 2004, New Astronomy Reviews, 48, 1289 Bell, J. F., & Bailes, M. 1996, ApJL, 456, L33 Bhat, N. D. R., Bailes, M., & Verbiest, J. P. W. 2008, PhRvD, 77, 124017 Bradshaw, C. F., Fomalont, E. B., & Geldzahler, B. J. 1999, ApJL, 512, L121 Brisken, W. F., Benson, J. M., Beasley, A. J., Fomalont, E. B., Goss, W. M., & Thorsett, S. E. 2000, ApJ, 541, 959 Burke-Spolaor, S. 2011a, MNRAS, 410, 2113 Burke-Spolaor, S. 2011b, PhD Thesis, Supermassive black hole binaries and transient radio events: studies in pulsar astronomy, <http://hdl.handle.net/1959.3/196878> Cantrell, A.G., Bailyn, C.D., Orosz, J.A., McClintock, J.E., Remillard, R.A., Froning, C.S., Neilsen, J., Gelino, D.M. & Gou, L., 2010, ApJ, 710, 1127 Carilli, C. & Rawlings, S., 2004, “Science with the Square Kilometre Array" , New Astronomy Reviews, Vol. 48 Condon, J. J. 1992, ARA&A, 30, 575 Cordes, J. M., Kramer, M., Lazio, T. J. W., Stappers, B. W., Backer, D. C., & Johnston, S. 2004, New Astronomy Reviews, 48, 1413 Croton, D. J., et al. 2006, MNRAS, 365, 11 Damour, T., & Taylor, J. H. 1991, ApJ, 366, 501 Damour, T., & Taylor, J. H. 1992, Phys. Rev. D., 45, 1840 Deller, A. T., Bailes, M., & Tingay, S. J. 2009a, Science, 323, 1327 Deller, A. T., Tingay, S. J., Bailes, M., & Reynolds, J. E. 2009b, ApJ, 701, 1243 Dhawan, V., Mirabel, I. F., Rib[ó]{}, M., & Rodrigues, I. 2007, ApJ, 668, 430 Elbaz, D., Jahnke, K., Pantin, E., Le Borgne, D., & Letawe, G. 2009, A&A, 507, 1359 Elbaz, D. 2010, IAU Symposium, 267, 17 Falcke, H, Kording, E. & Nagar, N.M. 2004, New Astronomy Reviews, 48, 1157 Fender, R. 2004, New Astronomy Reviews, 48, 1399 Fender, R. 2006, Compact stellar X-ray sources, ed. W. Lewin & M. van der Klis (Cambridge Astrophysics Series, No. 39), 381 Fender, R. 2010, Lecture Notes in Physics, Berlin Springer Verlag, 794, 115 Fomalont, E & Reid, M. 2004, New Astronomy Reviews, 48, 1473 Fryer, C. L., & Kalogera, V. 2001, ApJ, 554, 548 Fryer, C. L. 1999, ApJ, 522, 413 Gaensler, B. M., Beck, R., & Feretti, L. 2004, New Astronomy Reviews, 48, 1003 Gaensler, B. M. 2006, Astronomische Nachrichten, 327, 387 Gallo, E., Fender, R. P., Miller-Jones, J. C. A., Merloni, A., Jonker, P. G., Heinz, S., Maccarone, T. J., & van der Klis, M. 2006, MNRAS, 370, 1351 Garcia, M. R., McClintock, J. E., Narayan, R., Callanan, P., Barret, D., & Murray, S. S. 2001, ApJL, 553, L47 Garrett, M. A., 2000, Perspectives on Radio Astronomy: Science with Large Arrays, ed. M.P. van Haarlem (ASTRON, The Netherlands), 139 (arXiv:astro-ph/9908350v1) Godfrey, L., Bignall, H. & Tingay, S., 2011, SKA Memo 135, <http://www.skatelescope.org/uploaded/16339_135_Memo_Godfrey.pdf> Gou, L., McClintock, J., Reid, M., et al. 2011, Bulletin of the American Astronomical Society, \#410.17 Greaves, J. S., Richards, A. M. S., Rice, W. K. M., & Muxlow, T. W. B. 2008, MNRAS, 391, L74 Greaves, J. et al., 2009, “PEBBLES" e-MERLIN Legacy Project Proposal, <http://www.e-merlin.ac.uk/legacy/proposals/e-MERLIN_Legacy_PEBBLES.pdf> Greaves, J., 2010, “PEBBLES: Protoplanetary disks with eMERLIN and SKA", presented at the SKA 2010 Science and Engineering Meeting, Manchester Hoare, M. G. 2004, New Astronomy Reviews, 48, 1327 Hoare, M. G. 2009, “Protoplanetary disks with SKA-mid", presented at the prepSKA work package 2 Annual Meeting, Manchester Jaffe, A. H., & Backer, D. C. 2003, ApJ, 583, 616 Jenet, F. A., Lommen, A., Larson, S. L., & Wen, L. 2004, ApJ, 606, 799 Johnson, K. E. 2004, New Astronomy Reviews, 48, 1337 Jonker, P. G., & Nelemans, G. 2004, MNRAS, 354, 355 Kamp, I., Freudling, W., & Chengalur, J. N. 2007, ApJ, 660, 469 Klamer, I. J., Ekers, R. D., Sadler, E. M., & Hunstead, R. W. 2004, ApJL, 612, L97 Kramer, M., Backer, D. C., Cordes, J. M., Lazio, T. J. W., Stappers, B. W., & Johnston, S. 2004, New Astronomy Reviews, 48, 993 Kramer, M. 2010, IAU Symposium, 261, 366 Lee, K. J., Wex, N., Kramer, M., Stappers, B. W., Bassa, C. G., Janssen, G. H., Karuppusamy, R., & Smits, R. 2011, MNRAS, 414, 3251 Lipunov, V. M., Bogomazov, A. I., & Abubekerov, M. K. 2005, MNRAS, 359, 1517 Liu, K., Wex, N., Kramer, M., Cordes, J.M., & Lazio, T.J.W., submitted Macquart, J.-P., et al. 2010, PASA, 27, 272 Merloni, A., Heinz, S., & di Matteo, T. 2003, MNRAS, 345, 1057 Merritt, D. 2006, ApJ, 648, 976 Migliari, S., & Fender, R. P. 2006, MNRAS, 366, 79 Miller-Jones, J. C. A., Jonker, P. G., Nelemans, G., Portegies Zwart, S., Dhawan, V., Brisken, W., Gallo, E., & Rupen, M. P. 2009a, MNRAS, 394, 1440 Miller-Jones, J. C. A., Jonker, P. G., Dhawan, V., Brisken, W., Rupen, M. P., Nelemans, G., & Gallo, E. 2009b, ApJL, 706, L230 Mioduszewski, A. J., Rupen, M. P., Hjellming, R. M., Pooley, G. G., & Waltman, E. B. 2001, ApJ, 553, 766 Mirabel, I. F., Dhawan, V., Mignani, R. P., Rodrigues, I., & Guglielmetti, F. 2001, Nature, 413, 139 Morganti, R., Hoare, M., Garrett, M. in “The magnificent memo series: Trade-offs between science and engineering for the square kilometre array", SKA Memo No. 82, August 2006 Natta, A., Testi, L., Calvet, N., Henning, T., Waters, R., & Wilner, D. 2007, Protostars and Planets V, 767 Nordhaus, J., Brandt, T. D., Burrows, A., Livne, E., & Ott, C. D. 2010, PhRvD, 82, 103016 Norris, R. P., Kesteven, M. J., Troup, E. R., Allen, D. A., & Sramek, R. A. 1990, ApJ, 359, 291 Noutsos, A. 2009, IAU Symposium, 259, 15 Pfahl, E., Podsiadlowski, P., & Rappaport, S. 2005, ApJ, 628, 343 Reid, M. J., McClintock, J. E., Narayan, R., et al. 2011, Bulletin of the American Astronomical Society, 43, \#223.01 Remillard, R. A., & McClintock, J. E. 2006, ARA&A, 44, 49 Rioja, M., Dodson, R., Porcas, R. W., Ferris, D., Reynolds, J., Sasao, T., & Schilizzi, R. 2009, Proceedings of the 8th International e-VLBI Workshop, 14, <http://pos.sissa.it/archive/conferences/082/014/EXPReS09_014.pdf> Rodriguez, C., Taylor, G. B., Zavala, R. T., Peck, A. B., Pollack, L. K., & Romani, R. W. 2006, ApJ, 646, 49 Ryba, M. F., & Taylor, J. H. 1991, ApJ, 371, 739 Schilizzi, R.T. et al., “Preliminary specifications for the SKA", SKA Memo No. 100, December 2007 Schilizzi, R.T. et al., “Project Execution Plan Pre-Construction Phase for the Square Kilometre Array", SKA Memo No. 130, January 2011 Sesana, A., Vecchio, A., & Volonteri, M. 2009, MNRAS, 394, 2255 Seymour, N., Dwelly, T., Moss, D., McHardy, I., Zoghbi, A., Rieke, G., Page, M., Hopkins, A., & Loaring, N. 2008, MNRAS, 386, 1695 Shklovskii, I. S. 1970, Soviet Astron.–AJ, 13, 562 Sigurdsson, S. 2003, Radio Pulsars, 302, 391 Silk, J. 2005, MNRAS, 364, 1337 Simonetti, J. H., Kavic, M., Minic, D., Surani, U., & Vejayan, V., ApJL, 737, L28. SKA Science Working Group, 2010, “The Square Kilometre Array Design Reference Mission: SKA-mid and SKA-lo v. 1.0" Smits, R., Kramer, M., Stappers, B., Lorimer, D. R., Cordes, J., & Faulkner, A. 2009, AAP, 493, 1161 Smits, R., Tingay, S. J., Wex, N., Kramer, M., & Stappers, B. 2011, A&A, 528, A108 Stairs, I. H. 2010, IAU Symposium, 261, 218 Stepanov, R., Frick, P., Shukurov, A., & Sokoloff, D. 2002, A&A, 391, 361 Thompson, D. R., Wagstaff, K. L., Brisken, W. F., Deller, A. T., Majid, W. A., Tingay, S. J., & Wayth, R. B. 2011, ApJ, 735, 98 Tingay, S. J., et al. 1995, Nature, 374, 141 Tingay, S. J., & Wayth, R. B. 2011, AJ, 141, 174 Volonteri, M., Haardt, F., & Madau, P. 2003, ApJ, 582, 559 Wayth, R. B., Brisken, W. F., Deller, A. T., Majid, W. A., Thompson, D. R., Tingay, S. J., & Wagstaff, K. L. 2011, ApJ, 735, 97 Weisberg, J. M., Stanimirovi[ć]{}, S., Xilouris, K., Hedden, A., de la Fuente, A., Anderson, S. B., & Jenet, F. A. 2008, ApJ, 674, 286 Wilkinson, P. N., Kellermann, K. I., Ekers, R. D., Cordes, J. M., & Lazio, T. J. W. 2004, New Astronomy Reviews, 48, 1551 Wilner, D. J., D’Alessio, P., Calvet, N., Claussen, M. J., & Hartmann, L. 2005, ApJL, 626, L109 Wilner, D. J. 2004, New Astronomy Reviews, 48, 1363 Yu, Q. 2002, MNRAS, 331, 935
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $X$ be the blow-up of a smooth projective 4-fold $Y$ along a smooth curve $C$ and let $E$ be the exceptional divisor. Assume that $X$ is a Fano manifold and has an elementary extremal contraction $\varphi:X\to Z$ of (3,1)-type (i.e. the exceptional locus of $\varphi$ is a divisor and its image is a curve) such that $E$ is $\varphi$-ample. We show that if the exceptional divisor of $\varphi$ is smooth, then $Y$ is isomorphic to ${\mathbb{P}}^{4}$ and $C$ is an elliptic curve of degree 4 in ${\mathbb{P}}^{4}$.'
author:
- Toru Tsukioka
title: 'A remark on Fano 4-folds having (3,1)-type extremal contractions'
---
Introduction
============
As an application of the extremal contraction theory, S. Mori and S. Mukai classified smooth Fano 3-folds with Picard number greater than or equal to 2 ([@MM]). We observe that many of examples in the Mori-Mukai’s list are obtained by blowing up other smooth projective 3-folds. In fact, 78 types among 88 types of smooth Fano 3-folds with $\rho\geq 2$ have $E_{1}$-type or $E_{2}$-type extremal contractions. In [@BCW] the authors classified smooth Fano varieties (defined over ${\mathbb{C}}$) obtained by blowing-up a smooth point, in any dimension. A next step is to consider the following problem:
[*Problem.*]{} Let $Y$ be a smooth projective variety. Let $\pi:X\to Y$ be the blow-up along a smooth curve $C$. Classify pairs $(Y,C)$ such that $X$ is Fano.
Remark that for the toric case, the classification is done in any dimension by [@S].
By the Cone and Contraction Theorem, we can take an extremal contraction $\varphi:X\to Z$ to normal projective variety such that the exceptional divisor $E$ of $\pi$ is $\varphi$-ample (see Lemma \[kisolem\] below). It is easy to show that any fiber of $\varphi$ is at most of dimension 2. The author studied the case where $\varphi$ is a del Pezzo surface fibration and gave a complete classification ([@T2]). In higher dimensions, it seems difficult to classify the case where $\varphi$ is birational. However, in dimension 4, there are several results on the birational extremal contractions, which may be applied to solve our problem.
In this paper, we investigate the case where $\varphi$ is of $(3,1)$-type contraction. Recall that in general, an extremal contraction $\varphi:X \to Z$ is said to be $(a,b)$-type, if $\dim($Exc$(\varphi))=a$ and $\dim(\varphi($Exc$(\varphi)))=b$. So, a $(3,1)$-type contraction for a 4-fold is a birational contraction which contracts a divisor $F$ to a curve $B$. The extremal contractions of $(3,1)$-type for smooth 4-folds are completely classified by [@Tk]. In particular, it is shown that the exceptional divisor $F$ is normal and $B$ is smooth. Moreover, $\varphi_{|F}:F\to B$ is either a ${\mathbb{P}}^{2}$-bundle or a $Q_{2}$-bundle (see [@Tk] Main Theorem). [^1]
In section \[example\], we first give an example. Let $C\subset {\mathbb{P}}^{4}$ be a smooth complete intersection of one hyperplane and two hyperquadrics. Then, we see that $X={\operatorname{Bl}}_{C}({\mathbb{P}}^{4})$ has a (3,1)-type extremal contraction to a complete intersection of two hyperquadrics (singular along a line) in ${\mathbb{P}}^{6}$. The section \[proof\] is devoted to show that this is the only example if we assume that Exc$(\varphi)$ is smooth. More precisely, we prove the following:
\[theorem\] Let $\pi:X\to Y$ be the blow-up of a smooth projective 4-fold $Y$ defined over ${\mathbb{C}}$, along a smooth curve $C$. Assume that $X$ is a Fano manifold and has an elementary extremal contraction $\varphi:X\to Z$ of (3,1)-type such that the exceptional divisor $E$ of $\pi$ is $\varphi$-ample. Let $F$ be the exceptional divisor of $\varphi$. If $F$ is smooth, then $Y$ is isomorphic to ${\mathbb{P}}^{4}$ and $C$ is a smooth complete intersection of a hyperplane and two hyperquadrics.
We will use the following lemma, which is essentially the same as in [@BCW](Lemme 2.1). For reader’s convinience, we include here the statement with its proof.
\[kisolem\] Let $X$ be a Fano manifold and let $E$ be a non-zero effective divisor on $X$. Then there exists an extremal ray ${\mathbb{R^{+}}}[f]\subset\overline{{\operatorname{NE}}}(X)$ such that $E\cdot f>0$.
[*Proof.*]{} Since $X$ is projective, we can take a curve $\Gamma$ on $X$ such that $E\cdot \Gamma>0$. By the Cone Theorem, there exist positive real numbers $a_{i}$, and extremal rational curves $f_{i}$ such that $\Gamma\equiv \sum a_{i}f_{i}$ (finite sum). Hence $$0<E\cdot \Gamma=\sum a_{i}(E\cdot f_{i}).$$ This implies that one of extremal rational curves satisfies $E\cdot f_{i}>0$.
Throughout this paper, we shall assume that the base field is the complex numbers. For a Cartier divisor $D$ and a 1-cycle $\alpha$ on a variety $X$, we denote the intersection number by $D\cdot \alpha$, but we also write $(D\cdot \alpha)_{X}$ when we need to clarify the variety in which the intersection number is taken.
An example {#example}
==========
We give an example of a smooth Fano 4-fold $X$ obtained by blowing up along a curve such that $X$ has another $(3,1)$-type extremal contraction.
[**Example**]{} Let $C\subset {\mathbb{P}}^{4}$ be a smooth complete intersection of a hyperplane and two hyperquadrics, $\pi:X\to {\mathbb{P}}^{4}$ the blow-up along $C$, and $E$ the exceptional divisor. Let $F$ be the strict transform of the hyperplane containing $C$. Remark that $F\simeq {\operatorname{Bl}}_{C}({\mathbb{P}}^{3})$ is a $Q_{2}$-bundle over ${\mathbb{P}}^{1}$. Let $e$ be a line in a fiber of the ${\mathbb{P}}^{2}$-bundle $\pi_{|E}:E\to C$, and let $f$ be the strict transform of a line in ${\mathbb{P}}^{4}$ intersecting $C$ at two points. Then we have $$\overline{{\operatorname{NE}}}(X)={\mathbb{R^{+}}}[e]+{\mathbb{R^{+}}}[f].$$
The extremal contraction associated to the ray ${\mathbb{R^{+}}}[e]$ is of course the blow-up $\pi:X\to {\mathbb{P}}^{4}$. Let $L:=\pi^{*}{{\cal O}}_{{\mathbb{P}}^{4}}(1)$. The linear system $|2L-E|$ is base-point-free and defines the extremal contraction $\varphi:X\to Z$ of the ray ${\mathbb{R^{+}}}[f]$. Indeed, we have $(2L-E)\cdot f=0$. Note that $B:=\varphi(F)$ is isomorphic to ${\mathbb{P}}^{1}$ and $\varphi_{|F}:F\to B$ is a $Q_{2}$-bundle. Thus, $\varphi$ is a (3,1)-type extremal contraction whose exceptional divisor is $F$. More precisely, the image $Z$ is a complete intersection of two hyperquadrics in ${\mathbb{P}}^{6}$, singular along $B\simeq{\mathbb{P}}^{1}$. To see this, we calculate $h^{0}(X,{{\cal O}}_{X}(2L-E))$ and $(2L-E)^{4}$.
Consider the exact sequence: $$0\to {{\cal O}}_{X}(2L-E)\to{{\cal O}}_{X}(2L)\to {{\cal O}}_{E}(2L)\to 0.$$ Remark that $A:=-K_{X}+(2L-E)=(5L-2E)+(2L-E)=7L-3E$ is ample by Kleiman’s criterion, because $A\cdot e=3>0$ and $A\cdot f=1>0$. Therefore, by the Kodaira vanishing, $H^{1}(X,{{\cal O}}_{X}(2L-E))=0$. On the other hand, we get $h^{0}(X,{{\cal O}}_{X}(2L))=h^{0}({\mathbb{P}}^{4},{{\cal O}}_{{\mathbb{P}}^{4}}(2))=15$. Since ${{\cal O}}_{E}(2L)\simeq(\pi_{|E})^{*}{{\cal O}}_{C}(2)$, we have $h^{0}(E,{{\cal O}}_{E}(2L))=h^{0}(C,{{\cal O}}_{C}(2))=\deg({{\cal O}}_{C}(2))=8$ (recall that $\pi_{|E}$ is a ${\mathbb{P}}^{2}$-bundle and $g(C)=1$). Hence, $$h^{0}(X,{{\cal O}}_{X}(2L-E))=h^{0}(X,{{\cal O}}_{X}(2L))-h^{0}(E,{{\cal O}}_{E}(2L))=7$$ and $|2L-E|$ defines a morphism $\varphi:X\to {\mathbb{P}}^{6}$. Now we determine the image of $X$. Note that we have $L^{2}\cdot E\equiv 0$, $L\cdot E^{3}=\deg C=4$, and $E^{4}=\deg N_{C/{\mathbb{P}}^{4}}=20$. Thus, $$(2L-E)^{4}=(2L)^{4}-8L\cdot E^{3}+E^{4}=4.$$ Consider the exact sequence $$0\to {{\cal O}}_{{\mathbb{P}}^{6}}(2)\otimes I_{Z}\to{{\cal O}}_{{\mathbb{P}}^{6}}(2)\to{{\cal O}}_{Z}(2)\to 0.$$ Since $h^{0}(Z,{{\cal O}}_{Z}(2))=h^{0}(X,{{\cal O}}_{X}(4L-2E))=26$, we obtain $$h^{0}({\mathbb{P}}^{6},{{\cal O}}_{{\mathbb{P}}^{6}}(2)\otimes I_{Z})
\geq h^{0}({\mathbb{P}}^{6},{{\cal O}}_{{\mathbb{P}}^{6}}(2))-h^{0}(Z,{{\cal O}}_{Z}(2))=28-26=2.$$ It follows that there exist two linearly independent hyperquadrics in ${\mathbb{P}}^{6}$ containing $Z$. Since $\deg Z=(2H-E)^{4}=4$, $Z$ is a complete intersection of two hyperquadrics.
Proof of Theorem \[theorem\] {#proof}
============================
Denote by $e$ a line in a fiber of the ${\mathbb{P}}^{2}$-bundle $\pi_{|E}:E\to C$. The key to prove Theorem \[theorem\] is the following:
\[lemma\] We have $F\cdot e=1$.
[*Proof.*]{} We denote by $(e)$ the corresponding point in Hilb$(X)$. Let $T$ be the reduced part of the irreducible component of Hilb$(X)$ containing $(e)$. Note that $T$ is a ${\mathbb{P}}^{2}$-bundle over $C$ whose fiber $T_{c}$ ($c\in C$) parametrizes lines in $E_{c}:=\pi^{-1}(c)\simeq {\mathbb{P}}^{2}$. In particular, $T$ is smooth and of dimension 3.
[*Step 1.*]{} For all $(e)\in T$ such that $e\not\subset F$, we have $\sharp(F\cap e)=1$. [^2] Assume the contrary, i.e. there exists $(e_{0})\in T$ such that $e_{0}\not\subset F$ and $\sharp(F\cap e_{0})\geq 2$. Remark that $\varphi(e_{0})\neq B$. Let $x_{i}$ $(i=1,2)$ be two distinct points in $F\cap e_{0}$ and let $b_{i}:=\varphi(x_{i})$. Consider the incidence graph: $$\begin{CD}
V @>p >> X \\
@VqVV \\
T
\end{CD}$$ We define $V_{i}:=p^{-1}(E\cap \varphi^{-1}(b_{i}))$ and $T_{i}:=q(V_{i})$ for $i=1,2$. Note that $\dim V_{i}=\dim (E\cap \varphi^{-1}(b_{i}))+1$ because $p$ is a ${\mathbb{P}}^{1}$-bundle. We observe that $q_{|V_{i}}$ is a finite map. Indeed, if not, there exists $t\in T_{i}$ such that $q^{-1}(t)\subset V_{i}$. Then $e_{t}:=p(q^{-1}(t))$ is contracted by $\varphi$. This contradicts to our assumption that $E$ is $\varphi$-ample. It follows that $\dim T_{i}=\dim V_{i}=2\ (i=1,2)$. Note also that $(e_{0})\in T_{1}\cap T_{2}$. Now, we have $$\dim(T_{1}\cap T_{2})\geq \dim T_{1}+\dim T_{2}-\dim T=2+2-3=1.$$ So, we can take an irreducible curve $A\subset T_{1}\cap T_{2}$ passing through $(e_{0})$. Then $q^{-1}(A)$ is a ruled surface having two exceptional curves $V_{i}\cap S\ (i=1,2)$, a contradiction.
[*Step 2.*]{} Consider $M:=(F\cap E)_{red}$. By Step 1, we see that for each $c\in C$, $e_{c}:=(F\cap E_{c})_{red}$ is a line in $E_{c}\simeq {\mathbb{P}}^{2}$. So, $\pi_{|M}:M\to C$ is a ${\mathbb{P}}^{1}$-bundle. In particular $M$ is irreducible. We can write $E_{|F}=mM$ with $m\in {\mathbb{Z}}^{+}$. We have $$(mM\cdot e_{c})_{F}=(E_{|F}\cdot e_{c})_{F}
=(E\cdot e_{c})_{X}=-1$$
By assumption, $F$ is smooth. So, $M\subset F$ is a Cartier divisor and $(M\cdot e_{c})_{F}$ is integer. It follows that $m=1$, i.e. the intersection $F\cap E$ is transversal. We conclude that $F\cdot e=\sharp(F\cap e)=1$.
By the proof of Lemma \[lemma\], $\pi_{|M}:M\to C$ is a ${\mathbb{P}}^{1}$-bundle and $(M\cdot e_{c})_{F}=-1$. So, $\pi_{|F}:F\to F':=\pi(F)$ is the blow-up along $C$, and $F'$ is smooth. On the other hand, by [@Tk], $\varphi_{|F}:F\to B$ is either a ${\mathbb{P}}^{2}$-bundle or a $Q_{2}$-bundle. Therefore $F$ is a Fano 3-fold with $\rho(F)=2$. By assumption, $F$ is smooth. So, by the Mori-Mukai’s list, the pair $(F',C)$ is one of the following:
- $F'\simeq {\mathbb{P}}^{3}$ and $C$ is a line;
- $F'$ is a hyperquadric $Q_{3}\subset{\mathbb{P}}^{4}$ and $C=H\cap H'$ with $H$, $H'\in |{{\cal O}}_{Q_{3}}(1)|$;
- $F'\simeq {\mathbb{P}}^{3}$ and $C=Q\cap Q'$ with $Q$, $Q'\in |{{\cal O}}_{{\mathbb{P}}^{3}}(2)|$.
In the case (3), $C$ is an elliptic curve. So, $Y$ is a Fano manifold by [@Wis] (Proposition 3.5). In the cases (1) and (2), we have $N_{C/F'}\simeq {{\cal O}}_{C}(1)^{\oplus 2}$. Since there exists an inclusion of normal bundles $N_{C/F'}\subset N_{C/Y}$, $N_{C/Y}$ cannot be isomorphic to ${{\cal O}}_{{\mathbb{P}}^{1}}(-1)^{\oplus 3}$. So, $Y$ is a Fano manifold again by [@Wis].
Now, by Lemma \[kisolem\], we can take an extremal ray ${\mathbb{R^{+}}}[m]$ such that $F'\cdot m>0$. Then, by Proposition \[proposition\] below, we have $\rho(Y)=1$. In paticular $F'$ is ample. Let $f$ be a minimal rational curve of the extremal contraction $\varphi$. We obtain the following table of intersection numbers (due to [@Tk] and [@MM]): $$
case $F\cdot f$ $E\cdot f$
------ -------------- ------------
(1) $-1$ or $-2$ $1$
(2) $-1$ $1$
(3) $-1$ $2$
$$
Let $f':=\pi_{*}f$. Note that $F'\cdot f'=(\pi^{*}F')\cdot f=(F+E)\cdot f$. In the cases (1) and (2), we have $F'\cdot f'\leq 0$, a contradiction because $F'$ is ample. So, only the case (3) (in which we have $F'\cdot f'=1$) is possible, and $(Y,F')\simeq ({\mathbb{P}}^{4},{{\cal O}}_{{\mathbb{P}}^{4}}(1))$. Consequently, $C$ is the complete intersection $F'\cap Q\cap Q'$ with $F'\in |{{\cal O}}_{{\mathbb{P}}^{4}}(1)|$ and $Q, Q'\in |{{\cal O}}_{{\mathbb{P}}^{4}}(2)|$.
It remains to prove the following:
\[proposition\] Let $Y$ be a smooth projective variety of dimension $n\geq 4$ and $D$ a prime divisor on $Y$ with $\rho(D)=1$. Assume that there exists an extremal contraction $\mu:Y\to V$ of ray ${\mathbb{R^{+}}}[m]$ with $D\cdot m>0$, $m$ being a minimal rational curve of the ray. If there exsits a smooth curve $C\subset D$ such that the blow-up $X:={\operatorname{Bl}}_{C}(Y)$ is a Fano manifold, then we have $\rho(Y)=1$. Moreover, if $D$ is isomorphic to ${\mathbb{P}}^{n-1}$, then we have $(Y,D)\simeq ({\mathbb{P}}^{n},{{\cal O}}_{{\mathbb{P}}^{n}}(1))$.
[*Proof.*]{} We shall consider two cases:
- there exists $v_{0}\in V$ such that $\dim (\mu^{-1}(v_{0})\cap D)\geq 1$;
- $\dim (\mu^{-1}(v)\cap D)=0$ for all $v\in V$.
In the case (1), there exsits a curve $B\subset \mu^{-1}(v_{0})\cap D$. So, we can write $B\equiv bm$ with $b\in {\mathbb{R^{+}}}$. Since $\rho(D)=1$, any curve in $D$ is numerically equivalent to a multiple of $m$. Hence, $\mu(D)$ is a point. We also have $D\cdot B>0$. Now, by Proposition 4 of [@T2], we conclude that $\rho(Y)=1$.
We show that the case (2) is impossible. In this case, any fiber of $\mu$ is at most of dimension 1. So, by [@Ando] (see also [@Wis] Theorem 1.2), $\mu$ is either, a ${\mathbb{P}}^{1}$-bundle, a conic bundle, or a blow-up along a smooth subvariety of codimension 2 in a smooth projective variety. If $\mu$ is a ${\mathbb{P}}^{1}$-bundle, take a fiber $m$ passing through a point on $C$. Let $\tilde{m}$ be the strict transform by the blow-up $\pi:X\to Y$. For the exceptional divisor $E$, we have $E\cdot \tilde m\geq 1$, so that $$K_{X}\cdot \tilde{m}=K_{Y}\cdot m+(n-2)E\cdot\tilde{m}\geq -2+(n-2)=n-4\geq 0,$$ which is absurd because $X$ is a Fano manifold.
If $\mu$ is a conic bundle, the extremal rational curve $m$ is a component of a singular fiber of $\mu$. Let $\Delta$ be the discriminant locus and let $\tilde{\Delta}:=\mu^{-1}(\Delta)$. The assumption $D\cdot m>0$ implies $\tilde{\Delta}\cap D \neq \varnothing$. Since $\rho(D)=1$, the non-zero effective Cartier divisor $\tilde{\Delta}_{|D}$ is ample. Therefore, $$(\tilde{\Delta}\cdot C)_{Y}=(\tilde{\Delta}_{|D}\cdot C)_{D}>0,$$ so that $\tilde{\Delta}\cap C\neq \varnothing$. Now, we can take a sigular fiber $\mu^{-1}(v_{0})$ ($v_{0}\in \Delta$) meeting $C$. Let $m_{0}\subset\mu^{-1}(v_{0})$ be a component such that $m_{0}\cap C\neq \varnothing$. Then, we have a contradiction as in the case of ${\mathbb{P}}^{1}$-bundle. The case of a blow-up along a centre of codimension 2, can be ruled out by using a same argument for the exceptional divisor of $\mu$ in place of $\tilde{\Delta}$.
Consequently, only the case (1) is possible, so that we have $\rho(Y)=1$. If $D\simeq{\mathbb{P}}^{n-1}$, by [@BCW](Lemme 4) we conclude that $(Y,D)\simeq ({\mathbb{P}}^{n},{{\cal O}}_{{\mathbb{P}}^{n}}(1))$
Our assumption that $F={\operatorname{Exc}}(\varphi)$ is smooth, is used in the proof of Lemma \[lemma\] (only for [*Step.2*]{}) and in the proof of Theorem \[theorem\] in oder to apply to $F$ the Mori-Mukai’s classification of smooth Fano 3-folds. So, it is natural to ask whether Theorem \[theorem\] remains true without the smoothness of $F$. Concerning to this question, it is worth seeing the following:
[**Example**]{} (A degenerate case of the example in Section \[example\]) We consider the union of two smooth conics $C=C_{1}\cup C_{2}\subset Y:={\mathbb{P}}^{4}$ obtained as complete intersection of a hyperplane and two hyperquadrics. We assume that $C_{1}$ and $C_{2}$ meet at two distinct points. Let $\pi:X\to {\mathbb{P}}^{4}$ be the blow-up along the ideal $I_{C_{1}\cup C_{2}}$ and $E$ the exceptional divisor. Let $F$ be the strict transform of the hyperplane containing $C=C_{1}\cup C_{2}$. Then $F$ is a $Q_{2}$-bundle over ${\mathbb{P}}^{1}$ having exactly two ordinary double points. Remark that $F$ is isomorphic to the blow-up of ${\mathbb{P}}^{3}$ along the ideal $I_{C_{1}\cup C_{2}}$. Moreover, $F$ can be realized as divisor in ${\mathbb{P}}^{1}\times {\mathbb{P}}^{3}$ by the equation $sX_{2}X_{3}+t(X_{0}^{2}+X_{1}^{2}+X_{2}^{2}+X_{3}^{2})=0$, where $(s:t)$ (resp. $(X_{0}:X_{1}:X_{2}:X_{3})$) is the homogeneous coodinates of ${\mathbb{P}}^{1}$ (resp. ${\mathbb{P}}^{3}$). The fiber over $(1:0)$ is two planes $P_{i}$ $(i=1,2)$ and the two ordinary double points lie on the line $P_{1}\cap P_{2}$.
As in Section \[example\], we see that the linear system $|\pi^{*}{{\cal O}}_{{\mathbb{P}}^{4}}(2)-E|$ defines a (3,1)-type contraction $\varphi:X \to Z$ to complete intersection of two hyperquadrics in ${\mathbb{P}}^{6}$, and its exceptional divisor is $F$. This gives an example of $(Y,C)$ such that $F={\operatorname{Exc}}(\varphi)$ is singular. However $X$ is also singular along two rational curves over the two intersection points of $C_{1}\cap C_{2}$.
Related results
===============
Let $X$ be a Fano manifold and let $\iota_{X}$ be its pseudo-index, i.e. the minimum of the anti-canonical degrees $(-K_{X}\cdot C)$ for rational curves $C$ on $X$. In [@BCDD], the authors discuss the inequality (“generalized Mukai conjecture”): $$\rho(X)(\iota_{X}-1)\leq \dim X$$ and prove it in dimension 4. The essential part is to show that if $\iota_{X}=2$, then $\rho(X)\leq 4$. Concerning to this, we have the following:
\[intersection\] Let $\pi:X\to Y$ be the blow-up of a smooth projective variety $Y$ of dimension $n\geq 4$ along a smooth curve $C$ and let $E$ be the exceptional divisor. Assume that $X$ is a Fano manifold and there is another blow-up $\varphi:X\to Z$ (different from $\pi$) along a smooth curve $B$. Let $F$ be the exceptional divisor of $\varphi$. Then, we have $E\cap F=\varnothing$.
[*Proof.*]{} Assume $E\cap F\neq \varnothing$. Take $a\in C$ and $b\in B$ such that $E_{a}\cap F_{b}\neq\varnothing$. Then we obtain $\dim (E_{a}\cap F_{b})\geq \dim E_{a}+\dim F_{b}-\dim X=n-4$. So, if $n\geq 5$, there is a curve contained in $E_{a}\cap F_{b}$ and then contracted by both $\pi$ and $\varphi$. This is absurd because we assume $\pi\neq \varphi$. Therefore, we have $n=4$. By (the proof of) Theorem \[theorem\], $\varphi_{|F}:F\to B$ cannot be a ${\mathbb{P}}^{2}$-bundle. So, the case $E\cap F\neq \varnothing$ is impossible.
We are now able to give a simple proof of a result in [@BCDD].
Let $X$ be a Fano manifold of dimension $\geq 4$ whose birational contractions are all blow-ups along smooth curves in smooth projective varieties. Assume that $X$ has at least one birational contraction. Then, we have $\rho(X)\leq 3$.
[*Proof.*]{} Let $E$ be an exceptional divisor on $X$. By Lemma \[kisolem\], we can take an extremal ray ${\mathbb{R^{+}}}[f]\subset \overline{{\operatorname{NE}}}(X)$ such that $E\cdot f>0$ . Then, by assumption and by Proposition \[intersection\] above, the associated contraction $\mu:={\operatorname{cont}}_{{\mathbb{R^{+}}}[f]}:X\to Z$ is of fiber type. So, there is a surjection $\mu_{|E}:E\to Z$. Hence, we have $\rho(Z)\leq\rho(E)=2$. Consequently, $\rho(X)=\rho(Z)+1\leq 3$.
[9]{}
T. Ando, On extremal rays of the higher-dimensional varieties. Invent. Math. **81,** (1985) 347–357.
L. Bonavero, C. Casagrande, O. Debarre and S. Druel, Sur une conjecture de Mukai. Comment. Math. Helv. **78,** (2003) 601–626.
L. Bonavero, F. Campana and J. Wisniewski, Variétés complexes dont l’éclatée en un point est de Fano. C. R. Math. Acad. Sci. Paris **334,** (2002) 463–468.
S. Mori and S. Mukai, Classification of Fano $3$-folds with $B\sb{2}\geq 2$. Manuscripta Math. **36,** (1981/82) 147–162.
Erratum: Manuscripta Math. **110,** (2003) 407.
H. Sato, Toric Fano varieties with divisorial contractions to curves. Math. Nachr. **261/262,** (2003) 163–170.
H. Takagi, Classification of extremal contractions from smooth fourfolds of $(3,1)$-type. Proc. Amer. Math. Soc. **127,** (1999) 315–321.
T. Tsukioka, Del Pezzo surface fibrations obtained by blow-up of a smooth curve in a projective manifold. C. R. Acad. Sci. Paris **340,** (2005) 581–586.
T. Tsukioka, Classification of Fano manifolds containing a negative divisor isomorphic to projective space. Geometriae Dedicata **123,** (2006) 179–186.
J. Wiśniewski, On contractions of extremal rays of Fano manifolds. J. Reine Angew. Math. **417,** (1991) 141–157.
[^1]: Remark that our $F$ and $B$ correspond to $E$ and $C$ in [@Tk].
[^2]: We mean by $\sharp(F\cap e)$ the number of points on $F\cap e$ without multiplicity.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper studies power allocation for distributed estimation of an unknown scalar random source in sensor networks with a multiple-antenna fusion center (FC), where wireless sensors are equipped with radio-frequency based energy harvesting technology. The sensors’ observation is locally processed by using an *uncoded amplify-and-forward* scheme. The processed signals are then sent to the FC, and are coherently combined at the FC, at which the best linear unbiased estimator (BLUE) is adopted for reliable estimation. We aim to solve the following two power allocation problems: 1) minimizing distortion under various power constraints; and 2) minimizing total transmit power under distortion constraints, where the distortion is measured in terms of mean-squared error of the BLUE. Two iterative algorithms are developed to solve the non-convex problems, which converge at least to a local optimum. In particular, the above algorithms are designed to jointly optimize the *amplification coefficients*, *energy beamforming*, and *receive filtering*. For each problem, a suboptimal design, a single-antenna FC scenario, and a *common harvester* deployment for colocated sensors, are also studied. Using the powerful *semidefinite relaxation* framework, our result is shown to be valid for any number of sensors, each with different noise power, and for an arbitrarily number of antennas at the FC.'
author:
- |
Vien V. Mai, Won-Yong Shin, *Senior Member*, *IEEE*,\
and Koji Ishibashi, *Member*, *IEEE*\
[^1] [^2] [^3] [^4]
title: Wireless Power Transfer for Distributed Estimation in Sensor Networks
---
Amplify-and-forwarding, best linear unbiased estimator (BLUE), distributed estimation, mean-squared error (MSE), wireless power transfer (WPT).
Introduction
============
Distributed inference in wireless sensor networks (WSNs) has been extensively studied for applications such as environmental monitoring, weather forecasts, health care, and home automation (see, e.g., [@YMG:08:CN; @VV:97:PROC; @XCL:08:SP; @KV:13:IT; @BEJ:14:SP; @JCS:14:SP; @MJS:16:COML] and references therein). Sensors in WSNs are powered typically by batteries, and hence the network lifetime is highly limited. In practice, periodically replacing or recharging batteries may be hard or even impossible (due to the fact that sensors are located inside toxic environments, building structures, or human bodies [@ZH:13:WCOM]). Therefore, although there have been many efforts in power management policies, the network lifetime remains a performance bottleneck and limits the wide-range deployment of WSNs.
Previous Work
-------------
The optimal power allocation strategies for distributed estimation in WSNs have received a great research interest both from analog and digital encoding perspectives [@CXG:07:SP; @ST:08:SP; @XCL:08:SP; @LA:09:SP; @FL:09:SP; @KV:13:IT; @WW:13:WCOM; @BEJ:14:SP]. Among encoding schemes, the *uncoded amplify-and-forward* scheme has been extensively studied due to its simplicity and information-theoretic-optimality properties under certain conditions [@GAS:08:IT]. In particular, the authors in [@CXG:07:SP] studied power allocation for orthogonal multiple access channels (MACs), when the best linear unbiased estimator (BLUE) is adopted. The same problem was considered in [@XCL:08:SP] for a coherent MAC. The effects of channel estimation error were reported in [@ST:08:SP] for orthogonal MACs adopting a linear minimum mean-squared error estimator, while in [@WW:13:WCOM], the sensing noise uncertainty was investigated by adopting the BLUE. Recently, the optimal transmit strategy for cooperative linear estimation was studied in [@KV:13:IT].
The tremendous performance gains achieved by multiple-antenna techniques highly motivate us to integrate this technology into future wireless systems including WSNs. The benefits of such technology in the context of WSNs have been recently studied for distributed inference [@BSTS:12:WCOM; @BEJ:14:SP; @NPC:14:SP; @JCSL:15:SP; @SDC:16:SP; @MJS:16:COML]. For a large-scale fusion center (FC) over a Rayleigh fading channel, it has been shown in [@JCSL:15:SP] that the detection/estimation performance remains asymptotically constant if the transmit power at each sensor decreases proportionally with increasing number of antennas at the FC. The benefits of the multiple-antenna FC in distributed detection were analyzed in terms of asymptotic error exponents in [@MJS:16:COML]. Power allocation strategies for distributed estimation were studied for the correlated source case [@SDC:16:SP] and for the correlated noise case [@BEJ:14:SP].
Although the network life span can be prolonged by applying the aforementioned strategies, one needs a disruptive design that fundamentally changes the limitation of a WSN. One of the promising solution is the so-called *energy harvesting* (EH), in which sensors scavenge energy from the ambient environment (e.g., solar, wind, and vibration) that can guarantee an infinite life span in theory [@NDA:15:JSAC]. However, due to the unpredictable nature of energy sources, EH is typically uncontrolled, and thus can be critical for some reliable-sensitive applications. In addition to commonly used energy sources such as solar and wind, ambient radio-frequency (RF) signals can be a viable new source for energy scavenging. Most of the researches on *wireless power transfer* (WPT) have been focused on *cellular* networks, where user terminals replenish energy from the received signals sent by the base station via the far-field RF-based WPT [@GS:10:ISIT; @ZH:13:WCOM; @ZZH:13:COM; @SLZ:14:WCOM; @XLZ:14:SP; @LZC:14:SP; @SLC:14:SP]. For example, the fundamental trade-off between the achievable rate and the transferred power was characterized in [@GS:10:ISIT]. Several practical receiver architectures for simultaneous information and power transfer were investigated in [@ZH:13:WCOM; @ZZH:13:COM]. Exploiting multiple antenna technologies in WPT has been widely studied: multiple-input-multiple-output broadcast channels [@ZH:13:WCOM], beamforming designs for multiuser multiple-input-single-output (MISO) [@XLZ:14:SP], physical-layer security problems for multiuser MISO [@LZC:14:SP], and multiple-antenna interference channels [@SLC:14:SP]. On the other hand, there are a relatively limited number of studies on WPT for *WSNs*; different WPT technologies for addressing energy/lifetime problems in WSNs were reviewed in [@XSH:13:WC; @KVB:14:PROC]; in [@YTG:15:GlobalSIP], the authors studied a distributed estimation system in which some of the multiple-antenna sensors, named super sensors, are capable of WPT to its neighboring sensors via beamforming; and in [@HH:15:GlobalSIP], several multiple-antenna RF-based chargers were used to replenish the wireless sensors and then to switch to the information transmission phase, where each sensor sent a quantized version of its measurement to the FC for estimation.
Main Contributions
------------------
For distributed estimation in WSNs, an important question is how to intelligently exploit multiple-antenna technologies and WPT to improve both the inference performance and network lifetime. In this paper, we devote to studying the optimality of WPT and the optimal allocation of harvested energy for distributed estimation of an unknown random source in WSNs with a multiple-antenna FC. Our main contributions are summarized as follows:
- When the BLUE is adopted at the FC for estimation, we jointly optimize the amplification coefficients, energy beamforming, and receiver filtering by adopting alternative minimization methods (see Algorithms \[alg:1\] and \[alg:2\]). To that end, we first solve the mean-squared error (MSE) minimization problem under the total power constraint at the FC as well as the causal power constraint at each sensor. Then, we solve a converse problem where the total transmit power at the FC is minimized subject to an MSE requirement.
- A key ingredient of our algorithms is the so-called *semidefinite relaxation*. We show that such a relaxation does not sacrifice the optimality of the relaxed problems. We derive the properties of the optimal solutions (see Theorems \[thrm:1\] and \[thrm:2\]).
- A special deployment of WPT in WSNs is also discussed, where a common energy harvester is used to collect energy from the FC. We show that the optimization problems are significantly simplified in this case. The optimal power–distortion trade-off is also characterized (see Theorem \[thrm:3\]).
Organization
------------
The rest of the paper is organized as follows. The system model and problem formulation are described in Section \[sec:2\].Section \[sec:3\] studies the problem of minimizing the MSE subject to power constraints. In Section \[sec:4\], the converse problem in Section \[sec:3\] is studied. The numerical results are shown in Section \[sec:5\]. Finally, we conclude the paper in Section \[sec:6\].
Notations
---------
The operators $\left(\cdot\right)^\top$, $\left(\cdot\right)^\ast$, $\left(\cdot\right)^\dag$ are the transpose, complex conjugate, and transpose conjugate, respectively. The notation ${{\mathbf{I}}_{n}}$ denotes the $n \times n$ identity matrix; $\operatorname{\mathrm{tr}}\left({\mathbf{A}}\right)$ denotes the trace of a matrix ${\mathbf{A}}$; ${\textsf{rank}\left({\mathbf{A}}\right)}$ denotes the rank of a matrix ${\mathbf{A}}$; ${\mathrm{diag}\left({{\mathbf{a}}}\right)}$ denotes a diagonal matrix with vector ${\mathbf{a}}$ being its diagonal, ${\mathbf{A}}\succeq{\mathbf{0}}$ denotes the positive semidefinite ${\mathbf{A}}$; ${\operatorname{\mathds{E}}\left\{{\cdot}\right\}}$ denotes the expectation operator; $\textsf{dim}\left({\mathbf{A}}\right)$ denotes a dimension of the subspace ${\mathbf{A}}$. We use the Bachmann–Landau notation: $ f\left(x\right)
=O\left(g\left(x\right)\right)$ if $
\lim_{x \to x_0}\frac{f\left(x\right)}{g\left(x\right)}=c < \infty$. Finally, we use the notation ${\left[n\right]}$ to denote the set of positive natural numbers up to $n$, i.e., ${\left[n\right]}=\{i:i=1,2,\ldots,n\}$.
System Model and Problem Formulation {#sec:2}
====================================
System Model
------------
As illustrated in Fig. \[fig:sm\], we consider a distributed estimation system where an ${n_{\mathrm{r}}}$-antenna FC collects data from ${n_{\mathrm{s}}}$ spatially distributed sensors. Let $\theta$ be an unknown scalar random parameter (source) with variance of ${\delta_{\theta}}^2$ to be estimated.[^5] Examples of such a parameter include physical phenomena such as pressure, temperature, sound intensity, radiation level, pollution concentration, seismic activity, etc.. We assume that all sensors do not have conventional energy supplies and hence need to harvest energy from the RF signal transferred by the FC for future use. We also assume that there is no cooperation among the sensors since they are spatially distributed. In this paper, we adopt a time-switching *harvest-then-forward* protocol [@LZC:14:COM] in which for each $\tau\, T$ amount of time, where $T$ is the length of one time slot, the FC transmits its energy signal to the sensors, and for the remaining $\left(1-\tau\right)T$ amount of time, the sensors observe and forward their observations to the FC for estimation while using the harvested energy from the RF signal. For analytical convenience, we set $\tau=1/2$ in the sequel unless otherwise specified.[^6]
In the first phase (i.e., the energy harvesting phase) of a time slot, the FC broadcasts its energy signal to the sensors through energy beamforming. More precisely, ${n_{\mathrm{b}}}\leq {n_{\mathrm{r}}}$ energy beams are assigned to ${n_{\mathrm{s}}}$ sensors without loss of generality. The energy signal received at the $k$th sensor is then given by $$\begin{aligned}
{\mathsf{r}}_k={{{\pmb{g}}}_{k}}^\dag {{\pmb{x}}}_{\mathrm{e}}+{\mathsf{m}}_k = {{{\pmb{g}}}_{k}}^\dag\sum_{i=1}^{{n_{\mathrm{b}}}}{{{\pmb{w}}}}_i {\mathsf{s}}_i +{\mathsf{m}}_k,\end{aligned}$$ where ${{\pmb{x}}}_{\mathrm{e}}=\sum_{i=1}^{{n_{\mathrm{b}}}}{{{\pmb{w}}}}_i {\mathsf{s}}_i$ is the energy signal transmitted from the FC; ${\mathsf{s}}_i$ is the energy-carrying signal for the $i$th energy beam fulfilling $\mathds{E}\{{\left|{{\mathsf{s}}_i}\right|}^2\}=1$ and ${\operatorname{\mathds{E}}\left\{{{\mathsf{s}}_i {\mathsf{s}}_j}\right\}}=0$ for $i \neq j$, which can be any arbitrary random signal provided that its power spectral density satisfies certain regulations on microwave radiation [@XLZ:14:SP]; ${{{\pmb{w}}}}_i \in {\mathbbmss{C}}^{{n_{\mathrm{r}}}\times
1}$ is the $i$th energy bemforming vector; ${{{\pmb{g}}}_{k}} \in {\mathbbmss{C}}^{{n_{\mathrm{r}}}\times 1}$ is the channel between the FC and $k$th sensor; and ${\mathsf{m}}_k$ is the additive noise at the $k$th sensor. By ignoring the background noise for the sake of simplicity, the harvested energy at the $k$th sensor in each slot is given by [@ZH:13:WCOM] $$\begin{aligned}
\label{eq:energy}
E_k = \frac{\zeta_k T}{2} \sum_{i=1}^{{n_{\mathrm{b}}}}{\left|{{{{\pmb{w}}}}_i^\dag{{{\pmb{g}}}_{k}}}\right|}^2,
\end{aligned}$$ where $0\leq \zeta_k \leq 1$ is the energy harvesting efficiency at the $k$th sensor. Then, the average power $P_k$ available for the information transmission phase at the $k$th sensor can be expressed as $$\begin{aligned}
\label{eq:transmit:power}
P_{k}=\frac{2\left(E_k-E^{\mathrm{cir}}_k\right)}{T}=\zeta_k \sum_{i=1}^{{n_{\mathrm{b}}}}{\left|{{{{\pmb{w}}}}_i^\dag{{{\pmb{g}}}_{k}}}\right|}^2-\frac{2E^{\mathrm{cir}}_k}{T},\end{aligned}$$ where $E_k^{\mathrm{cir}} \geq 0$ is the circuit energy consumption at the $k$th sensor, which is assumed to be constant over time slots. Similarly as in [@LZC:14:COM; @SLC:14:SP], we simply assume $\zeta_k=1$ and unit slot duration in the rest of this work (note that using an arbitrary $\zeta_k$ does not fundamentally change our power allocation problems). Similarly as in [@LZC:14:COM; @JZ:14:WCOM], for easy of presentation, we also assume that $\{E_k^{\mathrm{cir}}\}_{k=1}^{{n_{\mathrm{s}}}}=0$ to focus on the transmit power of the sensors.[^7] The FC has a total transmit power constraint $P$; we thus have $$\begin{aligned}
\label{eq:tot:const}
{\operatorname{\mathds{E}}\left\{{{{\pmb{x}}}_{\mathrm{e}}^\dag{{\pmb{x}}}_{\mathrm{e}}}\right\}}=\sum_{i=1}^{{n_{\mathrm{b}}}}{\left\|{{{{\pmb{w}}}}_i}\right\|}^2 \leq P.
\end{aligned}$$ Now, let us turn to describing the second phase (i.e., the information transmission phase) of a time slot. The observation at the $k$th sensor can be expressed as $$\begin{aligned}
x_k=\theta + u_k, \quad k=1,\ldots,{n_{\mathrm{s}}},\end{aligned}$$ where $u_k$ is the additive noise at the $k$th sensor with variance ${\sigma_{\mathrm{u},k}}^2$. The noise at each sensor is assumed to be independent of each other. In this paper, we adopt an analog uncoded amplify-and-forward scheme, i.e., the $k$th sensor just simply amplifies its observation by a factor $\alpha_k$. Therefore, by stacking the transmit signals from all sensors into a single vector ${{\pmb{t}}}$, it can be expressed as $$\begin{aligned}
\label{eq:transmit:signal}
{{\pmb{t}}}={{\mathbf{A}}}{\mathbf{1}}\theta + {{\mathbf{A}}}{{{\pmb{u}}}},\end{aligned}$$ where ${{\mathbf{A}}}= {\mathrm{diag}\left({{\alpha}_1},{\ldots},{{\alpha}_{{n_{\mathrm{s}}}}}\right)} \in
{\mathbbmss{C}}^{{n_{\mathrm{s}}}\times {n_{\mathrm{s}}}}$ is the amplification matrix; ${{{\pmb{u}}}}=\left[u_1
\, u_2 \, \cdots \, u_{{n_{\mathrm{s}}}} \right]^\top\in {\mathbbmss{C}}^{{n_{\mathrm{s}}}\times 1}$ is the noise vector at the sensors with zero mean and covariance matrix ${{{\pmb{R}}}_{\mathrm{s}}}={\mathrm{diag}\left({{\sigma_{\mathrm{u},1}}^2},{{\sigma_{\mathrm{u},2}}^2},{\ldots},{{\sigma_{\mathrm{u},{n_{\mathrm{s}}}}}^2}\right)}$; and ${\mathbf{1}}$ is the all one vector. Then, the received signal ${{{\pmb{z}}}}\in {\mathbbmss{C}}^{{n_{\mathrm{r}}}\times 1}$ at the FC can be written as $$\begin{aligned}
{{{\pmb{z}}}}={{{\pmb{H}}}}{{\mathbf{A}}}{\mathbf{1}}\theta + {{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{u}}}}+ {{{\pmb{n}}}},\end{aligned}$$ where ${{{\pmb{H}}}}\in {\mathbbmss{C}}^{{n_{\mathrm{r}}}\times {n_{\mathrm{s}}}}$ is the channel between the sensors and FC; and ${{{\pmb{n}}}}\in {\mathbbmss{C}}^{{n_{\mathrm{r}}}\times 1}$ is the noise vector at the FC with zero mean and covariance matrix ${{{\pmb{R}}}_{\mathrm{n}}}={\mathrm{diag}\left({{\sigma_{\mathrm{n},1}}^2},{{\sigma_{\mathrm{n},2}}^2},{\ldots},{{\sigma_{\mathrm{n},{n_{\mathrm{r}}}}}^2}\right)}$. Here, the random quantities $\theta$, ${{{\pmb{u}}}}$, and ${{{\pmb{n}}}}$ are statistically independent.
![The distributed estimation system with an ${n_{\mathrm{r}}}$-antenna FC and ${n_{\mathrm{s}}}$ spatially distributed sensors where ${{\pmb{G}}}=\left[{{{\pmb{g}}}_{1}}\,\,{{{\pmb{g}}}_{2}}\cdots\,\,{{{\pmb{g}}}_{{n_{\mathrm{s}}}}}\right]^\dag$. []{data-label="fig:sm"}](sm.eps){width="82.00000%"}
Since we consider a coherent MAC, we assume that there is perfect synchronization between the sensors and the FC. All wireless channels are assumed to be quasi-static flat fading, i.e., once each channel is realized, it remains fixed during each time slot and changes independently between slots. We further assume that full channel state information (CSI) is available at the FC. In practice, the sensors-to-FC channel ${{{\pmb{H}}}}$ can be estimated at the FC via periodic pilot transmissions from the sensors, while the FC-to-sensors channels $\{{{{\pmb{g}}}_{i}}\}_{i=1}^{{n_{\mathrm{s}}}}$ can be acquired owing to channel reciprocity between the sensors-to-FC and FC-to-sensors channels when the system operates in time-division-duplex mode.
Problem Formulation
-------------------
The received signal ${{{\pmb{z}}}}$ is constructively combined at the FC by a filtering vector ${{{\pmb{v}}}}\in {\mathbbmss{C}}^{{n_{\mathrm{r}}}\times 1}$. Then, by adopting the well-known BLUE [@Kay:93:Book Theorem 6.1], the FC estimates the parameter $\theta$ based on the minimal sufficient statistic ${\mathsf{y}}={{{\pmb{v}}}}^\dag {{{\pmb{z}}}}$ as follows:[^8]. $$\begin{aligned}
\label{eq:blue}
\hat{\theta}
=
\left[
{{{\pmb{a}}}}^\dag
{{{\pmb{H}}}}^\dag
{{{\pmb{v}}}}{\sigma_{\mathrm{tot}}}^{-2}
{{{\pmb{v}}}}^\dag
{{{\pmb{H}}}}{{{\pmb{a}}}}\right]^{-1}
{{{\pmb{a}}}}^\dag
{{{\pmb{H}}}}^\dag
{{{\pmb{v}}}}{\sigma_{\mathrm{tot}}}^{-2}
{\mathsf{y}},\end{aligned}$$ where ${\sigma_{\mathrm{tot}}}^{2}={{{\pmb{v}}}}^\dag\left[{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag{{{\pmb{H}}}}^\dag+{{{\pmb{R}}}_{\mathrm{n}}}\right]{{{\pmb{v}}}}$ is the total noise power after post-processing at the FC; and ${{{\pmb{a}}}}=\left[{\alpha}_1\,{\alpha}_2\,\ldots \,{\alpha}_{{n_{\mathrm{s}}}}\right]^\top$. The MSE of the BLUE can be written as $$\begin{aligned}
\label{eq:mse:def}
\textsf{mse}
&=
{\operatorname{\mathds{E}}\left\{{|\theta-\hat{\theta}|^2}\right\}}
=
\left[{{{\pmb{a}}}}^\dag
{{{\pmb{H}}}}^\dag
{{{\pmb{v}}}}{\sigma_{\mathrm{tot}}}^{-2}
{{{\pmb{v}}}}^\dag
{{{\pmb{H}}}}{{{\pmb{a}}}}\right]^{-1}
\nonumber\\
&=
\left[
\frac{
{\left|{{{{\pmb{v}}}}^\dag{{{\pmb{H}}}}{{{\pmb{a}}}}}\right|}^2
}{
{{{\pmb{v}}}}^\dag
\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)
{{{\pmb{v}}}}}
\right]^{-1}.\end{aligned}$$
Since the three quantities ${{{\pmb{a}}}}$, $\{{{{\pmb{w}}}}_k\}_{k=1}^{{n_{\mathrm{b}}}}$, and ${{{\pmb{v}}}}$ critically affect both the power requirement and estimation performance of the entire system, we jointly design the optimum sensor amplification coefficients ${{{\pmb{a}}}}$, receive filtering vector ${{{\pmb{v}}}}$, and energy beamforming $\{{{{\pmb{w}}}}_k\}_{k=1}^{{n_{\mathrm{b}}}}$ under practical constraints. To that end, we solve two types of minimization problems: 1) minimizing the MSE of the BLUE under causal individual power constraints at the sensors and a total power constraint at the FC; and 2) minimizing the total power consumed at the FC given a minimum requirement of the MSE. In particular, we aim to find the solution to the first problem, named $(\textsf{P1})$, by solving the following optimization problem. $$\begin{aligned}
\begin{aligned}
&{(\textsf{P1})}:\\
& \underset{{{{\pmb{v}}}},{{{\pmb{a}}}},\{{{{\pmb{w}}}}_i\}_{i=1}^{{n_{\mathrm{b}}}}}{\text{maximize}}
& & \frac{
{\left|{{{{\pmb{v}}}}^\dag{{{\pmb{H}}}}{{{\pmb{a}}}}}\right|}^2
}{
{{{\pmb{v}}}}^\dag
\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)
{{{\pmb{v}}}}}
\\
& \text{subject to}
&& {\left|{\alpha_k}\right|}^2\left({\delta_{\theta}}^2+{\sigma_{\mathrm{u},k}}^2\right) \leq \sum_{i=1}^{{n_{\mathrm{b}}}}{\left|{{{{\pmb{w}}}}_i^\dag{{{\pmb{g}}}_{k}}}\right|}^2, \quad \forall k \in {\left[{n_{\mathrm{s}}}\right]}
\\
&&&\sum_{i=1}^{{n_{\mathrm{b}}}}{\left\|{{{{\pmb{w}}}}_i}\right\|}^2 \leq P.
\end{aligned}\end{aligned}$$ As a counterpart of ${(\textsf{P1})}$, for a given MSE threshold $\textsf{mse}=1/\gamma$, the second optimization problem is stated as follows. $$\begin{aligned}
\begin{aligned}
&{(\textsf{P2})}:\\
& \underset{{{{\pmb{v}}}},{{{\pmb{a}}}},\{{{{\pmb{w}}}}_i\}_{i=1}^{{n_{\mathrm{b}}}}}{\text{minimize}}
& & \sum_{i=1}^{{n_{\mathrm{b}}}}{\left\|{{{{\pmb{w}}}}_i}\right\|}^2
\\
& \text{subject to}
&& \frac{
{\left|{{{{\pmb{v}}}}^\dag{{{\pmb{H}}}}{{{\pmb{a}}}}}\right|}^2
}{
{{{\pmb{v}}}}^\dag
\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)
{{{\pmb{v}}}}} \geq \gamma,
\\
&&& {\left|{\alpha_k}\right|}^2\left({\delta_{\theta}}^2+{\sigma_{\mathrm{u},k}}^2\right) \leq \sum_{i=1}^{{n_{\mathrm{b}}}}{\left|{{{{\pmb{w}}}}_i^\dag{{{\pmb{g}}}_{k}}}\right|}^2, \quad \forall k \in {\left[{n_{\mathrm{s}}}\right]}.
\end{aligned}\end{aligned}$$
Note that closed-form solutions to the global optimization of these two problems are generally unknown. Indeed, both problems are non-convex due to the coupled amplification vector ${{{\pmb{a}}}}$ and receive filtering ${{{\pmb{v}}}}$. Therefore, we turn to a simple approach—alternative minimization—which guarantees convergence, at least to a local optimum.
Minimizing MSE under Power Constraints {#sec:3}
======================================
In this section, we propose an alternative minimization algorithm to obtain the minimum solution to problem ${(\textsf{P1})}$. We also study the MSE performance for a large-scale antenna FC as well as a single-antenna FC.
Proposed Solution to Problem ${(\textsf{P1})}$
----------------------------------------------
Since problem ${(\textsf{P1})}$ is non-convex due to a non-concave objective function, we solve ${(\textsf{P1})}$ by using the alternative minimization method. Our goal is to progressively increase the objective function in ${(\textsf{P1})}$ by iteratively optimizing ${(\textsf{P1})}$ over ${{{\pmb{a}}}}$ and $\{{{{\pmb{w}}}}_i\}_{i=1}^{{n_{\mathrm{b}}}}$ for given ${{{\pmb{v}}}}$, and then over ${{{\pmb{v}}}}$ for given ${{{\pmb{a}}}}$. In order to find ${{{\pmb{v}}}}$, we first fix ${{{\pmb{a}}}}$ and solve the following unconstrained optimization problem: $$\begin{aligned}
\label{Pb:1:rbf}
\begin{aligned}
& \underset{{{{\pmb{v}}}}}{\text{maximize}}
& & \frac{
{\left|{{{{\pmb{v}}}}^\dag{{{\pmb{H}}}}{{{\pmb{a}}}}}\right|}^2
}{
{{{\pmb{v}}}}^\dag
\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)
{{{\pmb{v}}}}},
\end{aligned}\end{aligned}$$ which is a Rayleigh quotient and hence can be recasted as $$\begin{aligned}
\label{Pb:1:rbf}
\begin{aligned}
& \underset{{{{\pmb{v}}}}}{\text{minimize}}
& & {{{\pmb{v}}}}^\dag
\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)
{{{\pmb{v}}}}\\
& \text{subject to}
&& {{{\pmb{v}}}}^\dag{{{\pmb{H}}}}{{{\pmb{a}}}}=1.
\end{aligned}\end{aligned}$$ Solving the above problem, we obtain $$\begin{aligned}
\label{eq:rbf}
{{{\pmb{v}}}}^+=\kappa\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)^{-1}
{{{\pmb{H}}}}{{{\pmb{a}}}}.\end{aligned}$$ Note that the value of $\kappa$ is chosen to guarantee the equality constraint in . However, any selected value of $\kappa$ will not affect the objective function in ${(\textsf{P1})}$, and thus we simply choose $\kappa=1$ without loss of optimality. For a given ${{{\pmb{v}}}}$ in , we are now ready to find an update of ${{{\pmb{a}}}}$ and $\{{{{\pmb{w}}}}_i\}_{i=1}^{{n_{\mathrm{b}}}}$ in ${(\textsf{P1})}$. To facilitate the calculations, we define ${{{\pmb{f}}}}=\left[{{{\pmb{v}}}}^\dag{{{\pmb{h}}}_{1}} \, {{{\pmb{v}}}}^\dag{{{\pmb{h}}}_{2}} \, \ldots
\,{{{\pmb{v}}}}^\dag{{{\pmb{h}}}_{{n_{\mathrm{s}}}}}\right]^\top$ and ${{{\pmb{F}}}}={\mathrm{diag}\left({{{\pmb{f}}}}\right)}$, where ${{{\pmb{h}}}_{i}}$ is the $i$th column of the matrix ${{{\pmb{H}}}}$. Then, for a fixed receive filtering ${{{\pmb{v}}}}$, problem ${(\textsf{P1})}$ can be expressed as $$\begin{aligned}
\label{Pb:1:1}
\begin{aligned}
&\hspace{-0.2cm} \underset{{{{\pmb{a}}}},\{{{{\pmb{w}}}}_i\}_{i=1}^{{n_{\mathrm{b}}}}}{\text{maximize}} \hspace{-0.1cm}
& & \frac{
{\left|{{{{\pmb{a}}}}^\top{{{\pmb{f}}}}}\right|}^2
}{
{{{\pmb{a}}}}^\top
{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag
{{{\pmb{a}}}}^*
+
{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}}
\\
&\hspace{-0.2cm} \text{subject to} \hspace{-0.1cm}
&& {\left|{\alpha_k}\right|}^2\left({\delta_{\theta}}^2+{\sigma_{\mathrm{u},k}}^2\right) \leq \sum_{i=1}^{{n_{\mathrm{b}}}}{\left|{{{{\pmb{w}}}}_i^\dag{{{\pmb{g}}}_{k}}}\right|}^2, \, \forall k \in {\left[{n_{\mathrm{s}}}\right]}
\\
&&&\sum_{i=1}^{{n_{\mathrm{b}}}}{\left\|{{{{\pmb{w}}}}_i}\right\|}^2 \leq P.
\end{aligned}\end{aligned}$$ We remark that even with a fixed receive filtering ${{{\pmb{v}}}}$, problem ${(\textsf{P1})}$ is still non-convex, and thus needs to be transformed to a simple form. We further introduce ${{{\pmb{Q}}}}={{{\pmb{a}}}}^* {{{\pmb{a}}}}^\top$, ${{{\pmb{W}}}}=\sum_{i=1}^{{n_{\mathrm{b}}}}{{{\pmb{w}}}}_i{{{\pmb{w}}}}_i^\dag$, ${\mathbf{\Sigma}}={{{\pmb{f}}}}{{{\pmb{f}}}}^\dag$, ${\mathbf{\Psi}}={{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag$, ${{{\pmb{G}}}}_k={{{\pmb{g}}}_{k}}{{{\pmb{g}}}_{k}}^\dag$, and ${{\mathbf{D}}}_k=\mathrm{diag}({0},\ldots,{{\delta_{\theta}}^2+{\sigma_{\mathrm{u},k}}^2},\ldots,0)$. Then, we can rewrite the optimization problem as $$\begin{aligned}
\label{Pb:1:2}
\begin{aligned}
& \underset{{{{\pmb{Q}}}},{{{\pmb{W}}}}}{\text{maximize}}
& & \frac{
\operatorname{\mathrm{tr}}\left({{{\pmb{Q}}}}{\mathbf{\Sigma}}\right)
}{
\operatorname{\mathrm{tr}}\left(
{{{\pmb{Q}}}}{\mathbf{\Psi}}
\right)
+
{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}}
\\
& \text{subject to}
&& \operatorname{\mathrm{tr}}\left({{\mathbf{D}}}_k{{{\pmb{Q}}}}\right)-\operatorname{\mathrm{tr}}\left({{{\pmb{G}}}}_k{{{\pmb{W}}}}\right) \leq 0, \quad \forall k \in {\left[{n_{\mathrm{s}}}\right]}
\\
&&& \operatorname{\mathrm{tr}}\left({{{\pmb{W}}}}\right) \leq P
\\
&&& {{{\pmb{W}}}}\succeq {\mathbf{0}}, {{{\pmb{Q}}}}\succeq {\mathbf{0}}
\\
&&& {\textsf{rank}\left({{{\pmb{Q}}}}\right)}=1.
\end{aligned}\end{aligned}$$ In , if there exist a rank one solution of the optimal ${{{\pmb{Q}}}}={{{\pmb{Q}}}}^\star$ and a rank ${n_{\mathrm{b}}}$ solution of the optimal ${{{\pmb{W}}}}={{{\pmb{W}}}}^\star$, then one can recover the optimal ${{{\pmb{a}}}}^\star$ and $\{{{{\pmb{w}}}}_i^\star\}_{i=1}^{{n_{\mathrm{b}}}}$ by taking the eigenvalue decomposition of the matrices ${{{\pmb{Q}}}}^\star$ and ${{{\pmb{W}}}}^\star$, respectively. Note that problem is non-convex due to the linear fractional structure of its objective function. However, we can use the Charnes-Cooper transformation [@CC:62:NRLQ] to reformulate the quasi-convex objective function in into a simpler form as follows:[^9] $$\begin{aligned}
\label{Pb:1:3}
\begin{aligned}
& \underset{{\bar{{{\pmb{Q}}}}},{\bar{{{\pmb{W}}}}},\eta}{\text{maximize}}
&& \operatorname{\mathrm{tr}}\left({\bar{{{\pmb{Q}}}}}{\mathbf{\Sigma}}\right)
\\
& \text{subject to}
&&\operatorname{\mathrm{tr}}\left({\bar{{{\pmb{Q}}}}}{\mathbf{\Psi}}\right)+\eta{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}=1
\\
&&& \operatorname{\mathrm{tr}}\left({{\mathbf{D}}}_k{\bar{{{\pmb{Q}}}}}\right)-\operatorname{\mathrm{tr}}\left({{{\pmb{G}}}}_k{\bar{{{\pmb{W}}}}}\right) \leq 0, \quad \forall k \in {\left[{n_{\mathrm{s}}}\right]}
\\
&&& \operatorname{\mathrm{tr}}\left({\bar{{{\pmb{W}}}}}\right) \leq \eta P
\\
&&& {\bar{{{\pmb{W}}}}}\succeq {\mathbf{0}}, {\bar{{{\pmb{Q}}}}}\succeq {\mathbf{0}}, \eta > 0
\\
&&& {\textsf{rank}\left({\bar{{{\pmb{Q}}}}}\right)}=1.
\end{aligned}\end{aligned}$$ Note that $\eta=0$ is not feasible because from the third constraint, we must have ${\bar{{{\pmb{W}}}}}={{\pmb{0}}}$ if $\eta=0$. Thus, from the second constraint for any $k$, it follows that ${\bar{{{\pmb{Q}}}}}={{\pmb{0}}}$, which however violates the first constraint in .
If $({\bar{{{\pmb{Q}}}}}^\star,{\bar{{{\pmb{W}}}}}^\star,\eta^\star)$ is the optimal solution to problem , then $({\bar{{{\pmb{Q}}}}}^\star/\eta^\star,{\bar{{{\pmb{W}}}}}^\star/\eta^\star)$ is feasible to problem and achieves the same objective value as that of problem . On the other hand, let $t^\star=\operatorname{\mathrm{tr}}\left({{{\pmb{Q}}}}^\star{\mathbf{\Psi}}\right)+{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}$. Then, if $({{{\pmb{Q}}}}^\star,{{{\pmb{W}}}}^\star)$ is the optimal solution to problem , then $({{{\pmb{Q}}}}^\star/t^\star,{{{\pmb{W}}}}^\star/t^\star,1/t^\star)$ is feasible to problem and achieves the same objective value as that of problem . This implies that the Charnes-Cooper transform is a one-to-one mapping between the feasible sets of problems and . We can thus obtain the optimal solution to problem by solving problem , which has a simpler form in the sense that the non-convexity of the objective function in problem is eliminated.
Note that problem is still non-convex due to the rank constraint, which makes problem intractable in general. Hence, we will solve a relaxed version of by ignoring the rank constraint on ${{{\pmb{Q}}}}$, which leads to the semidefinite relaxation (SDR) of problem . $$\begin{aligned}
\begin{aligned}
&{(\textsf{SDR1})}:\\
& \underset{{\bar{{{\pmb{Q}}}}},{\bar{{{\pmb{W}}}}},\eta}{\text{maximize}}
&& \operatorname{\mathrm{tr}}\left({\bar{{{\pmb{Q}}}}}{\mathbf{\Sigma}}\right)
\\
& \text{subject to}
&&\operatorname{\mathrm{tr}}\left({\bar{{{\pmb{Q}}}}}{\mathbf{\Psi}}\right)+\eta{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}=1
\\
&&& \operatorname{\mathrm{tr}}\left({{\mathbf{D}}}_k{\bar{{{\pmb{Q}}}}}\right)-\operatorname{\mathrm{tr}}\left({{{\pmb{G}}}}_k{\bar{{{\pmb{W}}}}}\right) \leq 0, \,\,\, \forall k \in {\left[{n_{\mathrm{s}}}\right]}
\\
&&& \operatorname{\mathrm{tr}}\left({\bar{{{\pmb{W}}}}}\right) \leq \eta P
\\
&&& {\bar{{{\pmb{W}}}}}\succeq {\mathbf{0}}, {\bar{{{\pmb{Q}}}}}\succeq {\mathbf{0}}, \eta > 0.
\end{aligned}\end{aligned}$$ The relaxed problem ${(\textsf{SDR1})}$ is now convex—indeed semidefinite program (SDP)—whose optimal solution can be found, for example, by using the interior-point method (e.g., CVX [@cvx]). The following theorem characterizes the properties of the optimal solution to problem ${(\textsf{SDR1})}$.
\[thrm:1\] Let $\nu^\star$ and $\beta^\star$ be the optimal dual solutions associated with the first and third constraint in ${(\textsf{SDR1})}$, respectively. We also let ${\bar{{{\pmb{Q}}}}}^\star$ and ${\bar{{{\pmb{W}}}}}^\star$ be the optimal primal solutions to problem ${(\textsf{SDR1})}$. Then, the following three properties are fulfilled:
1. $\nu^\star >0$, $\beta^\star>0$;
2. ${\textsf{rank}\left({\bar{{{\pmb{W}}}}}^\star\right)}\leq\min\left({n_{\mathrm{s}}},{n_{\mathrm{r}}}\right)$;
3. ${\textsf{rank}\left({\bar{{{\pmb{Q}}}}}^\star\right)}=1$.
See Appendix \[appdx:1\].
The condition $\beta^\star>0$ implies that the total power constraint at the FC must be satisfied with equality, while property 2) implies that at most ${n_{\mathrm{b}}}=\min\left({n_{\mathrm{s}}},{n_{\mathrm{r}}}\right)$ energy beams are required for the optimal solution of problem ${(\textsf{SDR1})}$. It is worth noting that for fixed ${{{\pmb{v}}}}$, at the optimal solution $({\bar{{{\pmb{Q}}}}}^\star,{\bar{{{\pmb{W}}}}}^\star,\eta^\star)$, the individual power constraints in ${(\textsf{SDR1})}$ are not necessarily all tight, i.e., there may exist some $k$ such that $\operatorname{\mathrm{tr}}\left({{\mathbf{D}}}_k{\bar{{{\pmb{Q}}}}}\right)-\operatorname{\mathrm{tr}}\left({{{\pmb{G}}}}_k{\bar{{{\pmb{W}}}}}\right) < 0$. This fact reveals that the sensors do not always transmit all the power budget harvested from the energy harvesting phase, but power control is required to guarantee the MSE optimality. A similar observation was made in throughput optimization for multiple-antenna multiuser cellular systems in [@LZC:14:COM].
\[rmk:eqv:1:1:3:sdr\] We remark that since problem ${(\textsf{SDR1})}$ is a relaxed version of problem , in general, the solution to problem ${(\textsf{SDR1})}$ provides an upper bound on the optimal solution to problem , or equivalently, an upper bound on problem ${(\textsf{P1})}$ for a given ${{{\pmb{v}}}}$. Fortunately, we can show that the optimal solution to ${(\textsf{SDR1})}$ is also optimal to . To do that, let ${\Phi_{\eta}\left({\bar{{{\pmb{Q}}}}},{\bar{{{\pmb{W}}}}}\right)}$ be the objective function of problem or ${(\textsf{SDR1})}$ for a given feasible $\eta$, and $({\bar{{{\pmb{Q}}}}}^\star,{\bar{{{\pmb{W}}}}}^\star)$ and $({\bar{{{\pmb{Q}}}}}_\star,{\bar{{{\pmb{W}}}}}_\star)$ be the optimal solutions to problems ${(\textsf{SDR1})}$ and , respectively. Since the optimization problem ${(\textsf{SDR1})}$ is a relaxation of problem , we must have $$\begin{aligned}
\label{eq:rmk:1}
{\Phi_{\eta}\left({\bar{{{\pmb{Q}}}}}^\star,{\bar{{{\pmb{W}}}}}^\star\right)} \geq {\Phi_{\eta}\left({\bar{{{\pmb{Q}}}}}_\star,{\bar{{{\pmb{W}}}}}_\star\right)}.\end{aligned}$$ On the other hand, since ${\textsf{rank}\left({\bar{{{\pmb{Q}}}}}^\star\right)}=1$, the solution $({\bar{{{\pmb{Q}}}}}^\star,{\bar{{{\pmb{W}}}}}^\star,\eta)$ is also a feasible solution to problem . Therefore, we have $$\begin{aligned}
\label{eq:rmk:2}
{\Phi_{\eta}\left({\bar{{{\pmb{Q}}}}}^\star,{\bar{{{\pmb{W}}}}}^\star\right)} \leq {\Phi_{\eta}\left({\bar{{{\pmb{Q}}}}}_\star,{\bar{{{\pmb{W}}}}}_\star\right)}.\end{aligned}$$ From and , it follows that ${\Phi_{\eta}\left({\bar{{{\pmb{Q}}}}}^\star,{\bar{{{\pmb{W}}}}}^\star\right)} = {\Phi_{\eta}\left({\bar{{{\pmb{Q}}}}}_\star,{\bar{{{\pmb{W}}}}}_\star\right)}$. In other words, $({\bar{{{\pmb{Q}}}}}^\star,{\bar{{{\pmb{W}}}}}^\star)$ is also optimal solution to problem . Note that the above equivalence holds for any feasible $\eta$, and hence it holds for the optimal $\eta$.
Remark \[rmk:eqv:1:1:3:sdr\] suggests that we can solve the original problem ${(\textsf{P1})}$ for a given ${{{\pmb{v}}}}$ by equivalently solving the relaxed problem ${(\textsf{SDR1})}$ without loss of optimality. Finally, we summarize the overall procedure for solving problem ${(\textsf{P1})}$ in Algorithm \[alg:1\] below. In this algorithm, the FC iteratively updates ${{{\pmb{v}}}}$, ${{{\pmb{a}}}}$ and $\{{{{\pmb{w}}}}_i\}_{i=1}^{{n_{\mathrm{b}}}}$ in Step 3 and 4, respectively. The convergence and complexity of Algorithm \[alg:1\] are analyzed in the following remark.
\[rmk:convergence\] Note that the objective function in ${(\textsf{P1})}$ is increased in each step of Algorithm \[alg:1\]. Moreover, the objective function is upper-bounded by a certain value due to the finite total power at the FC, which implies that the algorithm must converge. However, the algorithm may converge to a local optimum due to the non-convex nature of the optimization problem. We now provide the complexity analysis of the proposed algorithm. Specifically, in each iteration of Algorithm \[alg:1\], the worst-case computational complexity for solving the generic convex problem in ${(\textsf{SDR1})}$, corresponding to Step 4 in Algorithm \[alg:1\], using the interior point method is given by $O\left(\left(2 {n_{\mathrm{s}}}+{n_{\mathrm{r}}}\right)^{1/2}\left({n_{\mathrm{s}}}^4 + {n_{\mathrm{r}}}^3{n_{\mathrm{s}}}+ {n_{\mathrm{r}}}^2{n_{\mathrm{s}}}^2\right)\log\left(\frac{1}{\xi}\right)\right)$ for an $\xi$-optimal solution [@AN:01:Book Chapter 6.6.3]. For the receive filtering update, based on the elementary vector matrix calculation [@Hun:07:tchrp], one can show that the computational complexity of Step 3 in Algorithm \[alg:1\] is $O\left({n_{\mathrm{r}}}^3+{n_{\mathrm{r}}}^2 {n_{\mathrm{s}}}\right)$. In Step 6, the amplification vector is constructed by using eigenvalue decomposition of rank one matrix ${\bar{{{\pmb{Q}}}}}$, and hence the complexity is $O\left({n_{\mathrm{s}}}^2\right)$. Thus, the overall complexity per iteration of Algorithm \[alg:1\] is at most $O\left(\left(2
{n_{\mathrm{s}}}+{n_{\mathrm{r}}}\right)^{1/2}\left({n_{\mathrm{s}}}^4 + {n_{\mathrm{r}}}^3{n_{\mathrm{s}}}+
{n_{\mathrm{r}}}^2{n_{\mathrm{s}}}^2\right)\log\left(\frac{1}{\xi}\right)+{n_{\mathrm{r}}}^3+{n_{\mathrm{r}}}^2
{n_{\mathrm{s}}}\right)$. We remark that although the complexity of the alternative minimization algorithms are typically unknown [@GCJ:11:IT; @SRL:11:SP], it is observed via simulations that they converge within 10 to 20 iterations in general.
Large-Scale Antenna FC
----------------------
The following proposition shows the property of the asymptotic MSE of the BLUE.
\[prop:asym\] Consider the distributed estimation system in Section \[sec:2\], where the channel matrix ${{{\pmb{H}}}}$ is a random matrix with independent and identical elements, each of which has zero mean and unit variance. As the number of antennas at the FC tends to infinity, the MSE defined in converges to that of centralized estimation systems.[^10] That is, as ${n_{\mathrm{r}}}\to \infty$, we have $$\begin{aligned}
\label{eq:mse:bmk}
\textsf{mse}
\overset{\textsf{a.s}}\to
\left[{\mathbf{1}}^\top {{{\pmb{R}}}_{\mathrm{s}}}^{-1} {\mathbf{1}}\right]^{-1},\end{aligned}$$ where $\overset{\textsf{a.s}} \to$ denotes the almost sure convergence.
Given the receive filtering in , the MSE can be written as $$\begin{aligned}
\label{eq:prop1:eq1}
\textsf{mse}
= \left[
{{{\pmb{a}}}}^\dag
{{{\pmb{H}}}}^\dag
\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)^{-1}
{{{\pmb{H}}}}{{{\pmb{a}}}}\right]^{-1}.\end{aligned}$$ Therefore, it suffices to prove that $$\begin{aligned}
{{{\pmb{a}}}}^\dag
{{{\pmb{H}}}}^\dag
\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)^{-1}
{{{\pmb{H}}}}{{{\pmb{a}}}}\overset{\textsf{a.s}} \to
{\mathbf{1}}^\top {{{\pmb{R}}}_{\mathrm{s}}}^{-1} {\mathbf{1}}\end{aligned}$$ as ${n_{\mathrm{r}}}$ tends to infinity. Using the matrix inversion lemma, we can show that $$\begin{aligned}
\label{eq:prop1:eq2}
&\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)^{-1}
\nonumber\\
&\hspace{0.25cm}
=
{{{\pmb{R}}}_{\mathrm{n}}}^{-1}
-
{{{\pmb{R}}}_{\mathrm{n}}}^{-1}
{{{\pmb{H}}}}\left(
{{\pmb{K}}}^{-1}
+
{{{\pmb{H}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}^{-1}
{{{\pmb{H}}}}\right)^{-1}
{{{\pmb{H}}}}^\dag
{{{\pmb{R}}}_{\mathrm{n}}}^{-1},\end{aligned}$$ where ${{\pmb{K}}}={{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag$. Substituting into , we obtain $$\begin{aligned}
\textsf{mse}^{-1}
&=
{{{\pmb{a}}}}^\dag
{{{\pmb{H}}}}^\dag
{{{\pmb{R}}}_{\mathrm{n}}}^{-1}
{{{\pmb{H}}}}{{{\pmb{a}}}}\nonumber\\
&\hspace{-0.5cm}
-
{{{\pmb{a}}}}^\dag
{{{\pmb{H}}}}^\dag
{{{\pmb{R}}}_{\mathrm{n}}}^{-1}
{{{\pmb{H}}}}\left(
{{\pmb{K}}}^{-1}
+
{{{\pmb{H}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}^{-1}
{{{\pmb{H}}}}\right)^{-1}
{{{\pmb{H}}}}^\dag
{{{\pmb{R}}}_{\mathrm{n}}}^{-1}
{{{\pmb{H}}}}{{{\pmb{a}}}}.\end{aligned}$$ Note that as ${n_{\mathrm{r}}}\to \infty$, we have [@Mar:10:WCOM] $$\begin{aligned}
\frac{1}{{n_{\mathrm{r}}}}{{{\pmb{H}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}^{-1}{{{\pmb{H}}}}\overset{\textsf{a.s}} \to {{{\pmb{R}}}_{\mathrm{n}}}^{-1}.\end{aligned}$$ Using this identity, we obtain $$\begin{aligned}
\label{eq:asym:mse}
\frac{\textsf{mse}^{-1}}{{n_{\mathrm{r}}}}
& \overset{\textsf{a.s}} \to
{{{\pmb{a}}}}^\dag
{{{\pmb{R}}}_{\mathrm{n}}}^{-1}
{{{\pmb{a}}}}-
{{{\pmb{a}}}}^\dag
{{{\pmb{R}}}_{\mathrm{n}}}^{-1}
\left(
\frac{{{\pmb{K}}}^{-1}}{{n_{\mathrm{r}}}}
+
{{{\pmb{R}}}_{\mathrm{n}}}^{-1}
\right)^{-1}
{{{\pmb{R}}}_{\mathrm{n}}}^{-1}
{{{\pmb{a}}}}\nonumber\\
&=
{{{\pmb{a}}}}^\dag
{{{\pmb{R}}}_{\mathrm{n}}}^{-1}
\left(
{{\mathbf{I}}_{}}
-
\left(
\frac{{{{\pmb{R}}}_{\mathrm{n}}}{{\pmb{K}}}^{-1}}{{n_{\mathrm{r}}}}
+
{{\mathbf{I}}_{}}
\right)^{-1}
\right)
{{{\pmb{a}}}}\nonumber\\
&=
\frac{1}{{n_{\mathrm{r}}}}
{{{\pmb{a}}}}^\dag
\left({{\pmb{K}}}+\frac{{{{\pmb{R}}}_{\mathrm{n}}}}{{n_{\mathrm{r}}}}\right)^{-1}
{{{\pmb{a}}}}.\end{aligned}$$ where the second equality follows from the matrix inversion lemma. From and the definitions of the matrices ${{\pmb{K}}}$ and ${{\mathbf{A}}}$, as ${n_{\mathrm{r}}}\to \infty$, we finally have $$\begin{aligned}
\textsf{mse}
\overset{\textsf{a.s}}\to
\left[{\mathbf{1}}^\top {{{\pmb{R}}}_{\mathrm{s}}}^{-1} {\mathbf{1}}\right]^{-1},\end{aligned}$$ which concludes the proof of the proposition.
\[alg:1\] **Initialization**: set $n :=0$, and generate ${{{\pmb{a}}}}^{(0)}$ and ${{\mathbf{A}}}^{(0)}$.\
**repeat**\
${{{\pmb{v}}}}^{(n)}=\left({{{\pmb{H}}}}{{\mathbf{A}}}^{(n)}{{{\pmb{R}}}_{\mathrm{s}}}{{{\mathbf{A}}}^{(n)}}^\dag{{{\pmb{H}}}}^\dag+{{{\pmb{R}}}_{\mathrm{n}}}\right)^{-1}{{{\pmb{H}}}}{{{\pmb{a}}}}^{(n)}$\
Solve problem ${(\textsf{SDR1})}$ with ${{{\pmb{v}}}}={{{\pmb{v}}}}^{(n)}$ to obtain the\
optimal solution $({\bar{{{\pmb{Q}}}}}^\star,{\bar{{{\pmb{W}}}}}^\star,\eta^\star)$.\
Set $({\bar{{{\pmb{Q}}}}}^{(n+1)},{\bar{{{\pmb{W}}}}}^{(n+1)},\eta^{(n+1)}):=({\bar{{{\pmb{Q}}}}}^\star,{\bar{{{\pmb{W}}}}}^\star,\eta^\star)$.\
Construct $\{{{{\pmb{a}}}}^{(n+1)},{{\mathbf{A}}}^{(n+1)}\}$ from ${\bar{{{\pmb{Q}}}}}^{(n+1)}/\eta^{(n+1)}$.\
Update $n := n+1$.\
**until convergence**\
**Output:** $({{{\pmb{Q}}}}={\bar{{{\pmb{Q}}}}}^{(n)}/\eta^{(n)},{{{\pmb{W}}}}={\bar{{{\pmb{W}}}}}^{(n)}/\eta^{(n)},{{{\pmb{v}}}}={{{\pmb{v}}}}^{(n)})$
Proposition \[prop:asym\] implies that as the number of antennas grows large, the effects of fading and noise at the FC disappear, and hence the performance benchmark is determined by the sensing quality. From , if the sensing noise at the sensors is equal to ${{{\pmb{R}}}_{\mathrm{s}}}=\sigma_{\mathrm{n}}^2{{\mathbf{I}}_{{n_{\mathrm{s}}}}}$, then it follows that $\left[{\mathbf{1}}^\top {{{\pmb{R}}}_{\mathrm{s}}}^{-1}
{\mathbf{1}}\right]^{-1}=\frac{\sigma_{\mathrm{n}}^2}{{n_{\mathrm{s}}}}$. This means that the MSE linearly decreases according to ${n_{\mathrm{s}}}$.
Single-Antenna FC
-----------------
It is of importance to study the single-antenna FC scenario separately not only for comparison but also because the problem is remarkably simplified. Specifically, for a single-antenna FC, the design of energy beamforming and receive filtering is neglected, and thus we aim to simply find the optimal amplification coefficients ${{{\pmb{a}}}}$ that minimize the MSE of the BLUE. During the energy transmission phase, we assume that the FC transmits an energy signal ${\mathsf{s}}$ such that $\mathds{E}\{{\left|{{\mathsf{s}}}\right|}^2\}=P$. In this case, the harvested energy at the $k$th sensor is given by $$\begin{aligned}
E_k=\frac{P{\left|{{\mathsf{g}}_k}\right|}^2}{2}.\end{aligned}$$ The MSE of the BLUE is boiled down to $$\begin{aligned}
\label{eq:mse:def:fc}
\textsf{mse}
=
\left[
\frac{
{{{\pmb{a}}}}^\top{{{\pmb{h}}}_{}}{{{\pmb{h}}}_{}}^\dag{{{\pmb{a}}}}^*
}{
{{{\pmb{a}}}}^\top
{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag
{{{\pmb{a}}}}^*
+
{\sigma_{\mathrm{n}}}^2
}
\right]^{-1},\end{aligned}$$ where ${{{\pmb{h}}}_{}}^\dag \in {\mathbbmss{C}}^{1 \times {n_{\mathrm{s}}}}$ is the channel between the sensors and the FC; ${{{\pmb{F}}}}={\mathrm{diag}\left({{{{\pmb{h}}}_{}}}\right)}$; and $n \in {\mathbbmss{C}}$ is the additive noise at the FC. Given the MSE in , we aim to solve the following problem $$\begin{aligned}
\label{Pb:mse:fc}
\begin{aligned}
& \underset{{{{\pmb{a}}}}}{\text{maximize}}
& & \frac{
{{{\pmb{a}}}}^\top{{{\pmb{h}}}_{}}{{{\pmb{h}}}_{}}^\dag{{{\pmb{a}}}}^*
}{
{{{\pmb{a}}}}^\top
{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag
{{{\pmb{a}}}}^*
+
{\sigma_{\mathrm{n}}}^2
}
\\
& \text{subject to}
&& {\left|{\alpha_k}\right|}^2\left({\delta_{\theta}}^2+{\sigma_{\mathrm{u},k}}^2\right) \leq P{\left|{{\mathsf{g}}_k}\right|}^2, \,\,\,\, \forall k \in {\left[{n_{\mathrm{s}}}\right]}.
\end{aligned}\end{aligned}$$ The above problem—quadratically constrained ratio of quadratic functions (QCRQ)—has been studied for parameter tracking using the Kalman filter at the FC [@JCS:14:SP], where the optimal solution is given by $$\begin{aligned}
\label{eq:av:fc}
{{{\pmb{a}}}}^\star=\frac{1}{\sqrt{{\left({{\pmb{P}}}^\star\right)_{{n_{\mathrm{s}}}+1,{n_{\mathrm{s}}}+1}}}}\bar{{{{\pmb{a}}}}}^*.\end{aligned}$$ Here, ${\left({{\pmb{P}}}^\star\right)_{i,j}}$ is the $\left(i,j\right)$th element of the matrix ${{\pmb{P}}}^\star$; $\bar{{{{\pmb{a}}}}}$ is the vector satisfying $\bar{{{{\pmb{a}}}}}^*\bar{{{{\pmb{a}}}}}^\top={{\pmb{P}}}_{{n_{\mathrm{s}}}}^\star$; ${{\pmb{P}}}_{{n_{\mathrm{s}}}}^\star$ is the ${n_{\mathrm{s}}}$th order leading principal submatrix of ${{\pmb{P}}}^\star$ obtained by excluding the $({n_{\mathrm{s}}}+1)$th row and column; and ${{\pmb{P}}}^\star \in
{\mathbbmss{C}}^{\left({n_{\mathrm{s}}}+1\right)\times \left({n_{\mathrm{s}}}+1\right)}$ is the optimal solution to the following problem $$\begin{aligned}
\label{Pb:mse:fc:sdp}
\begin{aligned}
& \underset{{{\pmb{P}}} \succeq 0}{\text{maximize}}
& & \operatorname{\mathrm{tr}}\left({{\pmb{P}}}\,{{\pmb{\Xi}}}\right)
\\
& \text{subject to}
&& \operatorname{\mathrm{tr}}\left({{\pmb{P}}}\,{{\pmb{C}}}\right)=1 \\
&&& \operatorname{\mathrm{tr}}\left({{\pmb{P}}}\,\bar{{{\mathbf{D}}}}_k\right) \leq P{\left|{{\mathsf{g}}_k}\right|}^2, \quad \forall k \in {\left[{n_{\mathrm{s}}}\right]},
\end{aligned}\end{aligned}$$ where ${{\pmb{P}}}=\left[t\,{{{\pmb{a}}}}^\top\,\,
t\right]^\dag\left[t\,{{{\pmb{a}}}}^\top\,\, t\right]$; $ {{\pmb{\Xi}}}=\left(
\begin{array}{cc}
{{{\pmb{h}}}_{}}{{{\pmb{h}}}_{}}^\dag & {\mathbf{0}} \\
{\mathbf{0}} & 0 \\
\end{array}
\right) $; ${{\pmb{C}}}=\left(
\begin{array}{cc}
{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag & {\mathbf{0}} \\
{\mathbf{0}} & t\,{\sigma_{\mathrm{n}}}^2 \\
\end{array}
\right) $; $ \bar{{{\mathbf{D}}}}_k=\left(
\begin{array}{cc}
{{\mathbf{D}}}_k & {\mathbf{0}} \\
{\mathbf{0}} & -P{\left|{{\mathsf{g}}_k}\right|}^2 \\
\end{array}
\right)$; and $t$ is an auxiliary variable.
Note that the solution in is indeed global optimum. This is different from the multiple-antenna FC case in which we may only achieve a local optimum.
A Common Energy Harvester
-------------------------
We now consider a special deployment case in WPT-enabled sensor networks, where a common energy harvester is used to collect energy from the FC.[^11] Assume that the common energy harvester is equipped with a single antenna, then the optimization problem can be stated as $$\begin{aligned}
\begin{aligned}
{(\textsf{P1--Sum})}: \quad & \underset{{{{\pmb{v}}}},{{{\pmb{a}}}},{{{\pmb{w}}}}}{\text{maximize}}
& & \frac{
{\left|{{{{\pmb{v}}}}^\dag{{{\pmb{H}}}}{{{\pmb{a}}}}}\right|}^2
}{
{{{\pmb{v}}}}^\dag
\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)
{{{\pmb{v}}}}}
\\
& \text{subject to}
&& {{{\pmb{a}}}}^\dag{{\mathbf{D}}}{{{\pmb{a}}}}\leq {\left|{{{{\pmb{w}}}}^\dag {{{\pmb{h}}}_{\mathrm{e}}}}\right|}^2
\\
&&&{\left\|{{{{\pmb{w}}}}}\right\|}^2 \leq P,
\end{aligned}\end{aligned}$$ where ${{{\pmb{h}}}_{\mathrm{e}}}$ is the channel between the FC and the common harvester and ${{\mathbf{D}}}={\mathrm{diag}\left({{\delta_{\theta}}^2+{\sigma_{\mathrm{u},1}}^2},{\cdots},{{\delta_{\theta}}^2+{\sigma_{\mathrm{u},{n_{\mathrm{s}}}}}^2}\right)}$. Since the objective function in problem ${(\textsf{P1--Sum})}$ is monotonically increasing with the norm of ${{{\pmb{a}}}}$, the sum power constraint (i.e., the first constraint) should be satisfied with equality and the right-hand side of this constraint should be as large as possible to be the optimal solution. This implies that the optimal energy beamforming is ${{{\pmb{w}}}}^\star=\sqrt{P}\frac{{{{\pmb{h}}}_{\mathrm{e}}}}{{\left\|{{{{\pmb{h}}}_{\mathrm{e}}}}\right\|}}$. Similarly as in problem ${(\textsf{P1})}$, we iteratively solve the above problem for ${{{\pmb{v}}}}$ and ${{{\pmb{a}}}}$, where ${{{\pmb{v}}}}$ is given in . Let $E=\frac{1}{2}|{{{\pmb{h}}}_{\mathrm{e}}}^\dag
{{{\pmb{w}}}}^\star|^2=\frac{1}{2}P{\left\|{{{{\pmb{h}}}_{\mathrm{e}}}}\right\|}^2$ be the harvested energy at the harvester. For a given ${{{\pmb{v}}}}$, the optimal ${{{\pmb{a}}}}$ is then the solution of the following problem: $$\begin{aligned}
\label{Pb:mse:sum:av}
\begin{aligned}
& \underset{{{{\pmb{a}}}}}{\text{maximize}}
& & \frac{
{\left|{{{{\pmb{a}}}}^\top{{{\pmb{f}}}}}\right|}^2
}{
{{{\pmb{a}}}}^\top
{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag
{{{\pmb{a}}}}^*
+
{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}}
\\
& \text{subject to}
&& {{{\pmb{a}}}}^\dag{{\mathbf{D}}}{{{\pmb{a}}}}= P{\left\|{{{{\pmb{h}}}_{\mathrm{e}}}}\right\|}^2,
\end{aligned}\end{aligned}$$ where ${{{\pmb{f}}}}=\left[{{{\pmb{v}}}}^\dag{{{\pmb{h}}}_{1}} \, {{{\pmb{v}}}}^\dag{{{\pmb{h}}}_{2}} \, \cdots
\,{{{\pmb{v}}}}^\dag{{{\pmb{h}}}_{{n_{\mathrm{s}}}}}\right]^\top$ and ${{{\pmb{F}}}}={\mathrm{diag}\left({{{\pmb{f}}}}\right)}$. The problem is equivalent to $$\begin{aligned}
\label{Pb:mse:sum:av:Rayleigh}
\begin{aligned}
& \underset{{{{\pmb{a}}}}}{\text{maximize}}
& & \frac{
{{{\pmb{a}}}}^\top {{\pmb{X}}} {{{\pmb{a}}}}^*
}{
{{{\pmb{a}}}}^\top
{{\pmb{Y}}}
{{{\pmb{a}}}}^*
},
\end{aligned}\end{aligned}$$ where ${{\pmb{X}}}={{{\pmb{f}}}}{{{\pmb{f}}}}^\dag$ and ${{\pmb{Y}}}={{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag+\frac{{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}}{P{\left\|{{{{\pmb{h}}}_{\mathrm{e}}}}\right\|}^2}{{\mathbf{D}}}$. Note that ${{\pmb{X}}} \succeq {\mathbf{0}}$ and ${{\pmb{Y}}} \succ {\mathbf{0}}$, and problem is indeed Rayleigh quotient, thus the optimal solution can be expressed as $$\begin{aligned}
{{{\pmb{a}}}}^\star
=
\sqrt{\frac{P{\left\|{{{{\pmb{h}}}_{\mathrm{e}}}}\right\|}^2}{{{{\pmb{f}}}}^\dag{{\pmb{Y}}}^{-1}{{\mathbf{D}}}{{\pmb{Y}}}^{-1}{{{\pmb{f}}}}}}
{{\pmb{Y}}}^{-1}
{{{\pmb{f}}}}^*.\end{aligned}$$ Then, the optimal value in is given by $$\begin{aligned}
\label{eq:mse:sum:obj}
\hspace{-0.3cm}
\max_{{{{\pmb{a}}}}} \frac{
{\left|{{{{\pmb{a}}}}^\top{{{\pmb{f}}}}}\right|}^2
}{
{{{\pmb{a}}}}^\top
{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag
{{{\pmb{a}}}}^*
+
{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}}
=
{\lambda_{\mathrm{max}}}\left({{\pmb{Y}}}^{-1}{{\pmb{X}}}\right)
= {{{\pmb{f}}}}^\dag {{\pmb{Y}}}^{-1} {{{\pmb{f}}}},\end{aligned}$$ where ${\lambda_{\mathrm{max}}}\left(\cdot\right)$ denotes the maximum eigenvalue of a matrix. It can be seen that the sum power constraint enables to significantly reduce the complexity of the optimization problem.
Minimizing Power under an MSE Constraint {#sec:4}
========================================
In this section, we study the power minimization for distributed estimation with an MSE constraint.
Proposed Solution to Problem ${(\textsf{P2})}$
----------------------------------------------
Similarly as in problem ${(\textsf{P1})}$, we adopt an alternative minimization method to iteratively solve problem ${(\textsf{P2})}$. Specifically, we first solve problem ${(\textsf{P2})}$ over ${{{\pmb{v}}}}$ for given ${{{\pmb{a}}}}$ by finding a solution to the following feasibility problem: $$\begin{aligned}
\label{Pb:2:rbf}
\begin{aligned}
& \underset{{{{\pmb{v}}}}}{\text{minimize}}
& & 0
\\
& \text{subject to}
&& \frac{
{\left|{{{{\pmb{v}}}}^\dag{{{\pmb{H}}}}{{{\pmb{a}}}}}\right|}^2
}{
{{{\pmb{v}}}}^\dag
\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)
{{{\pmb{v}}}}} \geq \gamma.
\end{aligned}\end{aligned}$$ Since the left-hand side (LHS) of the constraint in is increasing with the norm of ${{{\pmb{a}}}}$, one should choose ${{{\pmb{v}}}}$ such that the LHS term is as large as possible. Hence, problem can be rewritten as an unconstrained optimization problem as follows: $$\begin{aligned}
\label{Pb:2:rbf:non-cst}
\begin{aligned}
& \underset{{{{\pmb{v}}}}}{\text{maximize}}
& & \frac{
{\left|{{{{\pmb{v}}}}^\dag{{{\pmb{H}}}}{{{\pmb{a}}}}}\right|}^2
}{
{{{\pmb{v}}}}^\dag
\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)
{{{\pmb{v}}}}}.
\end{aligned}\end{aligned}$$ Solving the above problem, we obtain $$\begin{aligned}
\label{eq:rbf:P2}
{{{\pmb{v}}}}^+=\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)^{-1}
{{{\pmb{H}}}}{{{\pmb{a}}}}.\end{aligned}$$ For fixed ${{{\pmb{v}}}}$ given in , we now solve problem ${(\textsf{P2})}$ over ${{{\pmb{a}}}}$ and $\{{{{\pmb{w}}}}_i\}_{i=1}^{{n_{\mathrm{b}}}}$ as in the following: $$\begin{aligned}
\label{Pb:2:1}
\begin{aligned}
&\hspace{-0.2cm} \underset{{{{\pmb{a}}}},\{{{{\pmb{w}}}}_i\}_{i=1}^{{n_{\mathrm{b}}}}}{\text{minimize}} \hspace{-0.1cm}
& & \sum_{i=1}^{{n_{\mathrm{b}}}}{\left\|{{{{\pmb{w}}}}_i}\right\|}^2
\\
& \hspace{-0.2cm} \text{subject to} \hspace{-0.1cm}
&& \frac{
{\left|{{{{\pmb{a}}}}^\top{{{\pmb{f}}}}}\right|}^2
}{
{{{\pmb{a}}}}^\top
{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag
{{{\pmb{a}}}}^*
+
{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}} \geq \gamma,
\\
&&& \hspace{-0.12cm} {\left|{\alpha_k}\right|}^2\left({\delta_{\theta}}^2+{\sigma_{\mathrm{u},k}}^2\right) \leq \sum_{i=1}^{{n_{\mathrm{b}}}}{\left|{{{{\pmb{w}}}}_i^\dag{{{\pmb{g}}}_{k}}}\right|}^2, \, \forall k \in {\left[{n_{\mathrm{s}}}\right]},
\end{aligned}\end{aligned}$$ where ${{{\pmb{f}}}}=\left[{{{\pmb{v}}}}^\dag{{{\pmb{h}}}_{1}} \, {{{\pmb{v}}}}^\dag{{{\pmb{h}}}_{2}} \, \cdots
\,{{{\pmb{v}}}}^\dag{{{\pmb{h}}}_{{n_{\mathrm{s}}}}}\right]^\top$; ${{{\pmb{h}}}_{i}}$ is the $i$th column of the matrix ${{{\pmb{H}}}}$; and ${{{\pmb{F}}}}={\mathrm{diag}\left({{{\pmb{f}}}}\right)}$. We remark that for a fixed ${{{\pmb{v}}}}$, the MSE constraint (i.e., the first constraint) at the optimal solution to problem must be fulfilled with equality. We prove it by contradiction. Assume that the MSE constraint is satisfied with a strict inequality at the optimal solution $({{{\pmb{a}}}}^\star,\{{{{\pmb{w}}}}_i^\star\}_{i=1}^{{n_{\mathrm{b}}}})$. By letting $\bar{{{{\pmb{a}}}}}=t {{{\pmb{a}}}}^\star$ for $0<t< 1$, we can choose a sufficient large $t$ such that $$\begin{aligned}
\hspace{-0.3cm}
\frac{
{\left|{{{{{\pmb{a}}}}^\star}{{{\pmb{f}}}}}\right|}^2
}{
{{{{\pmb{a}}}}^\star}^\top
{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag
{{{{\pmb{a}}}}^\star}^*
+
{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}}
>
\frac{
{\left|{\bar{{{{\pmb{a}}}}}^\top{{{\pmb{f}}}}}\right|}^2
}{
\bar{{{{\pmb{a}}}}}^\top
{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag
\bar{{{{\pmb{a}}}}}^*
+
{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}}
\geq
\gamma.\end{aligned}$$
\[alg:2\] **Initialization**: set $n :=0$, and generate ${{{\pmb{a}}}}^{(0)}$ and ${{\mathbf{A}}}^{(0)}$.\
**repeat**\
${{{\pmb{v}}}}^{(n)}=\left({{{\pmb{H}}}}{{\mathbf{A}}}^{(n)}{{{\pmb{R}}}_{\mathrm{s}}}{{{\mathbf{A}}}^{(n)}}^\dag{{{\pmb{H}}}}^\dag+{{{\pmb{R}}}_{\mathrm{n}}}\right)^{-1}{{{\pmb{H}}}}{{{\pmb{a}}}}^{(n)}$\
Solve problem ${(\textsf{SDR2})}$ with ${{{\pmb{v}}}}={{{\pmb{v}}}}^{(n)}$ to obtain the\
optimal value $({{{\pmb{Q}}}}^\star,{{{\pmb{W}}}}^\star)$.\
Set $({{{\pmb{Q}}}}^{(n+1)},{{{\pmb{W}}}}^{(n+1)}):=({{{\pmb{Q}}}}^\star,{{{\pmb{W}}}}^\star)$.\
Construct $\{{{{\pmb{a}}}}^{(n+1)},{{\mathbf{A}}}^{(n+1)}\}$ from ${{{\pmb{Q}}}}^{(n+1)}$.\
Update $n := n+1$.\
**until convergence**\
**Output:** $({{{\pmb{Q}}}}={{{\pmb{Q}}}}^{(n)},{{{\pmb{W}}}}={{{\pmb{W}}}}^{(n)},{{{\pmb{v}}}}={{{\pmb{v}}}}^{(n)})$
When $\bar{{{{\pmb{w}}}}}_i=t{{{\pmb{w}}}}_i^\star$, $(\bar{{{{\pmb{a}}}}},\{\bar{{{{\pmb{w}}}}}_i\}_{i=1}^{{n_{\mathrm{s}}}})$ can also be a feasible solution to problem with the new objective value $t^2\sum_{i=1}^{{n_{\mathrm{b}}}}{\left\|{{{{\pmb{w}}}}_i^\star}\right\|}^2$, which is definitely smaller than the optimal value when the optimal solution is $({{{\pmb{a}}}}^\star,\{{{{\pmb{w}}}}_i^\star\}_{i=1}^{{n_{\mathrm{s}}}})$. This contradicts to the assumption that $({{{\pmb{a}}}}^\star,\{{{{\pmb{w}}}}_i^\star\}_{i=1}^{{n_{\mathrm{s}}}})$ is optimal. Therefore, the MSE constraint must hold with equality. Let ${{{\pmb{Q}}}}={{{\pmb{a}}}}^* {{{\pmb{a}}}}^\top$, ${{{\pmb{W}}}}=\sum_{i=1}^{{n_{\mathrm{b}}}}{{{\pmb{w}}}}_i{{{\pmb{w}}}}_i^\dag$, ${\mathbf{\Sigma}}={{{\pmb{f}}}}{{{\pmb{f}}}}^\top$, ${\mathbf{\Psi}}={{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag$, ${{{\pmb{G}}}}_k={{{\pmb{g}}}_{k}}{{{\pmb{g}}}_{k}}^\dag$, and ${{\mathbf{D}}}_k=\mathrm{diag}({0},\ldots,{{\delta_{\theta}}^2+{\sigma_{\mathrm{u},k}}^2},\ldots,0)$. Then, as in problem , we will omit the rank constraint on ${{{\pmb{Q}}}}$ and solve a relaxed version of , which leads to $$\begin{aligned}
\begin{aligned}
&{(\textsf{SDR2})}:\\
& \underset{{{{\pmb{Q}}}},{{{\pmb{W}}}}}{\text{minimize}}
&& \operatorname{\mathrm{tr}}\left({{{\pmb{W}}}}\right)
\\
& \text{subject to}
&&\operatorname{\mathrm{tr}}\left({{{\pmb{Q}}}}{\mathbf{\Sigma}}\right)=\gamma\operatorname{\mathrm{tr}}\left({{{\pmb{Q}}}}{\mathbf{\Psi}}\right)+\gamma {{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}\\
&&& \operatorname{\mathrm{tr}}\left({{\mathbf{D}}}_k{{{\pmb{Q}}}}\right)-\operatorname{\mathrm{tr}}\left({{{\pmb{G}}}}_k{{{\pmb{W}}}}\right) \leq 0, \quad \forall k \in {\left[{n_{\mathrm{s}}}\right]}
\\
&&& {{{\pmb{W}}}}\succeq {\mathbf{0}}, {{{\pmb{Q}}}}\succeq {\mathbf{0}}.
\end{aligned}\end{aligned}$$ The following theorem characterizes the properties of the optimal solution to problem ${(\textsf{SDR2})}$.
\[thrm:2\] Let $\beta^\star$ be the dual optimal solutions associated with the equality constraint in ${(\textsf{SDR2})}$. We also let ${{{\pmb{Q}}}}^\star$ and ${{{\pmb{W}}}}^\star$ be the primal optimal solutions to ${(\textsf{SDR2})}$. Then the following three properties are fulfilled:
1. $\beta^\star>0$;
2. ${\textsf{rank}\left({{{\pmb{W}}}}^\star\right)}\leq\min\left({n_{\mathrm{s}}},{n_{\mathrm{r}}}\right)$;
3. ${\textsf{rank}\left({{{\pmb{Q}}}}^\star\right)}=1$.
The proof can be found using the similar steps to the proof for Theorem \[thrm:1\].
Similarly as in Section \[sec:3\], we summarize the overall procedure for solving problem ${(\textsf{P2})}$ in Algorithm \[alg:2\]. In this algorithm, the objective value is monotonically reduced in each step, and for a given feasible threshold $\gamma$, it is lower-bounded by a certain value. As a result, the algorithm converges at least to a local optimum. Finally, it can be verified that the computational complexity of Algorithm 2 is same as that of Algorithm \[alg:1\].
Single-Antenna FC
-----------------
It would also be of interest to study problem ${(\textsf{P2})}$ for the single-antenna FC scenario. In this case, problem ${(\textsf{P2})}$ can be rewritten as $$\begin{aligned}
\label{Pb:pow:fc}
\begin{aligned}
& \underset{{{{\pmb{a}}}},P}{\text{minimize}}
& & P
\\
& \text{subject to}
&& \frac{
{{{\pmb{a}}}}^\top{{{\pmb{h}}}_{}}{{{\pmb{h}}}_{}}^\dag{{{\pmb{a}}}}^*
}{
{{{\pmb{a}}}}^\top
{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag
{{{\pmb{a}}}}^*
+
{\sigma_{\mathrm{n}}}^2
}
=
\gamma
\\
&&& {\left|{\alpha_k}\right|}^2\left({\delta_{\theta}}^2+{\sigma_{\mathrm{u},k}}^2\right) \leq P {\left|{{\mathsf{g}}_k}\right|}^2, \,\, \forall k \in {\left[{n_{\mathrm{s}}}\right]}.
\end{aligned}\end{aligned}$$ Define the matrices ${{\pmb{\Omega}}}=\left(
\begin{array}{cc}
{{{\pmb{a}}}}^*{{{\pmb{a}}}}^\top & {\mathbf{0}} \\
{\mathbf{0}} & P \\
\end{array}
\right) $; $ {{\pmb{P}}}=\left(
\begin{array}{cc}
{\mathbf{0}} & {\mathbf{0}} \\
{\mathbf{0}} & 1 \\
\end{array}
\right);
$ $ \bar{{{\mathbf{D}}}}_k=\left(
\begin{array}{cc}
\frac{1}{{\left|{{\mathsf{g}}_k}\right|}^2}{{\mathbf{D}}}_k & {\mathbf{0}} \\
{\mathbf{0}} & -1 \\
\end{array}
\right); \,\text{and}\,\, {{\pmb{E}}}={{{\pmb{h}}}_{}}{{{\pmb{h}}}_{}}^\dag-\gamma
{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag. $ Then, problem can be recast as $$\begin{aligned}
\label{Pb:pow:fc:1}
\begin{aligned}
& \underset{{{\pmb{\Omega}}}}{\text{minimize}}
& & \operatorname{\mathrm{tr}}\left({{\pmb{\Omega}}}{{\pmb{P}}}\right)
\\
& \text{subject to}
&& \operatorname{\mathrm{tr}}\left({{\pmb{\Omega}}}{{\pmb{E}}}\right)
=
\gamma{\sigma_{\mathrm{n}}}^2
\\
&&& \operatorname{\mathrm{tr}}\left({{\pmb{\Omega}}}\bar{{{\mathbf{D}}}}_k\right) \leq 0, \quad \forall k \in {\left[{n_{\mathrm{s}}}\right]}
\\
&&& {\textsf{rank}\left({{\pmb{\Omega}}}\right)}=1.
\end{aligned}\end{aligned}$$ By dropping the rank constraint on ${{\pmb{\Omega}}}$, problem is a SDP and thus can be solved efficiently. If we denote ${{\pmb{\Omega}}}^\star$ by the optimal solution to the relaxed problem of , then ${\textsf{rank}\left({{\pmb{\Omega}}}^\star\right)}=1$ and the optimal ${{{\pmb{a}}}}$ and $P$ can be found from ${{\pmb{\Omega}}}^\star$. Particularly, $$\begin{aligned}
P^\star&={\left({{\pmb{\Omega}}}^\star\right)_{{n_{\mathrm{s}}}+1,{n_{\mathrm{s}}}+1}}\\
{{{\pmb{a}}}}^\star &= \sqrt{ \operatorname{\mathrm{tr}}({{\pmb{\Omega}}}^\star_{{n_{\mathrm{s}}}})}\,{{{\pmb{u}}}}_1^*,\end{aligned}$$ where ${{\pmb{\Omega}}}_{{n_{\mathrm{s}}}}^\star$ is the ${n_{\mathrm{s}}}$th order leading principal submatrix of ${{\pmb{\Omega}}}^\star$ obtained by excluding the $({n_{\mathrm{s}}}+1)$th row and column and ${{{\pmb{u}}}}_1$ is the eigenvector associated with the largest eigenvalue of ${{\pmb{\Omega}}}^\star_{{n_{\mathrm{s}}}}$. Similarly as in problem ${(\textsf{P1})}$, in this case, the optimal solution $(P^\star, {{{\pmb{a}}}}^\star)$ is indeed a global optimum.
A Common Energy Harvester
-------------------------
Now, we consider the converse problem of ${(\textsf{P1--Sum})}$, in which we aim to minimize the transmit power at the FC subject to a minimum requirement of the MSE performance, $$\begin{aligned}
\begin{aligned}
{(\textsf{P2--Sum})}: \quad
& \underset{{{{\pmb{v}}}},{{{\pmb{a}}}},{{{\pmb{w}}}}}{\text{minimize}}
& & {\left\|{{{{\pmb{w}}}}}\right\|}^2
\\
& \text{subject to}
&& \frac{
{\left|{{{{\pmb{v}}}}^\dag{{{\pmb{H}}}}{{{\pmb{a}}}}}\right|}^2
}{
{{{\pmb{v}}}}^\dag
\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)
{{{\pmb{v}}}}}
\geq
\gamma
\\
&&& {{{\pmb{a}}}}^\dag{{\mathbf{D}}}{{{\pmb{a}}}}\leq {\left|{{{{\pmb{w}}}}^\dag {{{\pmb{h}}}_{\mathrm{e}}}}\right|}^2.
\end{aligned}\end{aligned}$$ If we multiply ${{{\pmb{w}}}}$ and ${{{\pmb{a}}}}$ by a scalar $\alpha>1$ and $\beta<1$, respectively, then the left-hand side of the MSE constraint (i.e., the first constraint) is strictly increased while the right-hand side of the sum power constraint (i.e., the second constraint) as well as the objective function are strictly decreased. Thus, the optimality for ${(\textsf{P2--Sum})}$ is achieved when all the above constraints are satisfied with equality. Problem ${(\textsf{P2--Sum})}$ can be formulated as a SDP, and hence solved efficiently by CVX. In the following, we establish a fundamental relationship between two problems ${(\textsf{P1--Sum})}$ and ${(\textsf{P2--Sum})}$.
\[thrm:3\] For a distributed estimation system using the BLUE with a common energy harvester, if we assume that the alternative algorithms solving ${(\textsf{P1--Sum})}$ and ${(\textsf{P2--Sum})}$ are initialized with ${{{\pmb{a}}}}^{(0)}$, then the optimal *power–distortion trade-off* is given by $$\begin{aligned}
\label{eq:PD}
\frac{1}{\textsf{mse}}= {{{\pmb{f}}}}^\dag \left({{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag+\frac{{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}}{P {\left\|{{{{\pmb{h}}}_{\mathrm{e}}}}\right\|}^2}{{\mathbf{D}}}\right)^{-1}{{{\pmb{f}}}}.\end{aligned}$$
See Appendix \[appdx:2\].
Theorem \[thrm:3\] is important since it enables to (numerically) find the *power–distortion trade-off* for distributed estimation in the cumulative power constraint case.
Numerical Results {#sec:5}
=================
In this section, we provide numerical examples by evaluating our proposed algorithms in Sections \[sec:3\] and \[sec:4\]. In the simulations, we consider the widely used 915 MHz frequency band in WSNs [@Gut:01:Standard] for both energy and information transmissions. For energy transmission, we consider the use of both the commercially available power transmitter (Powercast TX91501) with transmit power $P=1$W (30 dBm) and the RF power harvester (Powercast P2110). The detailed system parameters are summarized in Table \[table:parameter\]. To model a small-scale fading, we assume that the elements of the channel matrices are drawn independently from the Gaussian distribution with zero mean and unit variance. To further evaluate the effectiveness of the proposed algorithms, we also perform comparisons to low-complexity baseline schemes specified below.
Baseline Schemes
----------------
### Suboptimal Design for ${(\textsf{P1})}$
We divide the optimization procedure into two phases. In the first phase, the energy beamforming vectors $\{{{{\pmb{w}}}}_i\}_{i=1}^{{n_{\mathrm{b}}}}$ are designed such that the total harvested energy is maximized, which leads to the following maximization problem: $$\begin{aligned}
\label{Pb:mse:sub:phase1}
\begin{aligned}
& \underset{\{{{{\pmb{w}}}}_i\}_{i=1}^{{n_{\mathrm{b}}}}}{\text{maximize}}
& & \sum_{k=1}^{{n_{\mathrm{s}}}}\beta_k\left(\sum_{i=1}^{{n_{\mathrm{b}}}}{\left|{{{{\pmb{w}}}}_i^\dag{{{\pmb{g}}}_{k}}}\right|}^2\right)
\\
& \text{subject to}
&& \sum_{i=1}^{{n_{\mathrm{b}}}}{\left\|{{{{\pmb{w}}}}_i}\right\|}^2 \leq P.
\end{aligned}\end{aligned}$$ Here, $\{\beta_k\}_{k=1}^{{n_{\mathrm{s}}}}$ denote the energy weights indicating the priority (e.g., sensors with weaker channels can be assigned to a higher weight to guarantee fairness) of the corresponding sensors. It has been shown in [@ZH:13:WCOM] that the optimal strategy is to allocate all the power budget to the direction of ${{\pmb{\eta}}}$—the eigenvector associated with the largest eigenvalue of the matrix $\sum_{k=1}^{{n_{\mathrm{s}}}}\beta_k{{{\pmb{g}}}_{k}}{{{\pmb{g}}}_{k}}^\dag$. The optimal value in problem is achieved when ${{{\pmb{w}}}}_i^\star=\sqrt{p_i} {{\pmb{\eta}}}$ with $p_i\geq 0$ such that $\sum_{i=1}^{{n_{\mathrm{b}}}}p_i=P$.
In the second phase, we find the amplification vector ${{{\pmb{a}}}}$ and the receive filtering ${{{\pmb{v}}}}$ in terms of minimizing the MSE subject to the energy harvested in the first phase. In particular, we solve the following problem: $$\begin{aligned}
\label{Pb:mse:sub:phase2}
\begin{aligned}
& \underset{{{{\pmb{a}}}},{{{\pmb{v}}}}}{\text{maximize}}
& & \frac{
{\left|{{{{\pmb{a}}}}^\top{{{\pmb{f}}}}}\right|}^2
}{
{{{\pmb{a}}}}^\top
{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag
{{{\pmb{a}}}}^*
+
{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}}
\\
& \text{subject to}
&& {\left|{\alpha_k}\right|}^2\left({\delta_{\theta}}^2+{\sigma_{\mathrm{u},k}}^2\right) \leq P_k, \quad \forall k \in {\left[{n_{\mathrm{s}}}\right]},
\end{aligned}\end{aligned}$$ where ${{{\pmb{f}}}}=\left[{{{\pmb{v}}}}^\dag{{{\pmb{h}}}_{1}} \, {{{\pmb{v}}}}^\dag{{{\pmb{h}}}_{2}} \, \cdots
\,{{{\pmb{v}}}}^\dag{{{\pmb{h}}}_{{n_{\mathrm{s}}}}}\right]^\top$; ${{{\pmb{F}}}}={\mathrm{diag}\left({{{\pmb{f}}}}\right)}$; and $P_k={\left|{{{{\pmb{g}}}_{k}}^\dag{{{\pmb{w}}}}_i^\star}\right|}^2$. Problem corresponds to problem ${(\textsf{P1})}$ without the total power constraint and can be solved by iteratively updating ${{{\pmb{v}}}}$ and ${{{\pmb{a}}}}$.
### Suboptimal Design for ${(\textsf{P2})}$
To reduce the computational burden of the joint optimization for ${(\textsf{P2})}$, we propose a suboptimal design, in which the optimization procedure is divided into two phases. In the first phase, we aim to solve the following problem: $$\begin{aligned}
\label{Pb:pow:sub:phase1}
\begin{aligned}
& \underset{{{{\pmb{a}}}},{{{\pmb{v}}}}}{\text{minimize}}
& & {{{\pmb{a}}}}^\dag{\mathbf{D}}\,{{{\pmb{a}}}}\\
& \text{subject to}
&& \frac{
{\left|{{{{\pmb{v}}}}^\dag{{{\pmb{H}}}}{{{\pmb{a}}}}}\right|}^2
}{
{{{\pmb{v}}}}^\dag
\left(
{{{\pmb{H}}}}{{\mathbf{A}}}{{{\pmb{R}}}_{\mathrm{s}}}{{\mathbf{A}}}^\dag
{{{\pmb{H}}}}^\dag
+
{{{\pmb{R}}}_{\mathrm{n}}}\right)
{{{\pmb{v}}}}} \geq \gamma.
\end{aligned}\end{aligned}$$ We note that the objective function in problem is the total transmit power of the sensors. Since the receive filtering ${{{\pmb{v}}}}$ appears only in the constraint, we can iteratively solve problem for ${{{\pmb{a}}}}$ and ${{{\pmb{v}}}}$. Since the left-hand side of the constraint is nondecreasing with the norm of ${{{\pmb{a}}}}$, the constraint must be satisfied with equality. For a fixed ${{{\pmb{v}}}}$, the above problem can be expressed as follows: $$\begin{aligned}
\label{Pb:pow:sub:phase1:1}
\begin{aligned}
& \underset{{{{\pmb{a}}}},{{{\pmb{v}}}}}{\text{minimize}}
& & {{{\pmb{a}}}}^\top{\mathbf{D}}\,{{{\pmb{a}}}}^*
\\
& \text{subject to}
&& {{{\pmb{a}}}}^\top{{\pmb{E}}}\,{{{\pmb{a}}}}^*=\gamma\,{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}},
\end{aligned}\end{aligned}$$ where ${{\pmb{E}}}={{{\pmb{f}}}}{{{\pmb{f}}}}^\dag-\gamma {{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag$. To guarantee the feasibility of problem , the value of $\gamma$ must be chosen such that ${\left|{{{{\pmb{a}}}}^\dag{{{\pmb{f}}}}}\right|}^2 \geq
\gamma\, {{{\pmb{a}}}}^\top{{{\pmb{F}}}}{{{\pmb{R}}}_{\mathrm{s}}}{{{\pmb{F}}}}^\dag\,{{{\pmb{a}}}}^*$. Since the quantities ${{{\pmb{a}}}}^\top{\mathbf{D}}\,{{{\pmb{a}}}}^* \geq 0$ and ${{{\pmb{a}}}}^\top{{\pmb{E}}}\,{{{\pmb{a}}}}^* \geq 0$ are positive, problem can be rewritten as $$\begin{aligned}
\label{Pb:pow:sub:phase1:2}
\begin{aligned}
& \underset{{{{\pmb{a}}}},{{{\pmb{v}}}}}{\text{maximize}}
& & \frac{{{{\pmb{a}}}}^\top{{\pmb{E}}}\,{{{\pmb{a}}}}^*}{{{{\pmb{a}}}}^\top{\mathbf{D}}\,{{{\pmb{a}}}}^*}
\\
& \text{subject to}
&& {{{\pmb{a}}}}^\top{{\pmb{E}}}\,{{{\pmb{a}}}}^*=\gamma\,{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}\end{aligned},\end{aligned}$$ which is a Rayleigh quotient. Thus, the optimal solution to problem is given by $$\begin{aligned}
{{{\pmb{a}}}}^\star
=
\sqrt{\frac{\gamma\,{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}}{{{{\pmb{u}}}}_1^\dag {\mathbf{D}}^{-1/2} {{\pmb{E}}} {\mathbf{D}}^{-1/2}{{{\pmb{u}}}}_1}}
{\mathbf{D}}^{-1/2}
{{{\pmb{u}}}}_1^*,\end{aligned}$$ where ${{{\pmb{u}}}}_1$ denotes the unit-norm eigenvector associated with the largest eigenvalue of the matrix ${\mathbf{D}}^{-1/2} {{\pmb{E}}}
{\mathbf{D}}^{-1/2}$, ${\lambda_{\mathrm{max}}}\left({\mathbf{D}}^{-1/2} {{\pmb{E}}}
{\mathbf{D}}^{-1/2}\right)$. It follows that the minimum total transmit power of the sensors, $P_\textsf{s}^\star$, required to achieve the MSE of $1/\gamma$ is given by $$\begin{aligned}
P_\textsf{s}^\star=\frac{\gamma\,{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}}{{\lambda_{\mathrm{max}}}\left( {\mathbf{D}}^{-1/2} {{\pmb{E}}} {\mathbf{D}}^{-1/2}\right)}.\end{aligned}$$
In the second phase, we aim to minimize the total transmit power at the FC with the amplification coefficients $\{\alpha_k\}_{k=1}^{{n_{\mathrm{s}}}}$ that are the solutions to problem . In other words, we find the optimal solution to the following minimization problem: $$\begin{aligned}
\label{Pb:pow:sub:phase2}
\begin{aligned}
& \hspace{-0.3cm} \underset{\{{{{\pmb{w}}}}_i\}_{i=1}^{{n_{\mathrm{b}}}}}{\text{minimize}} \hspace{-0.2cm}
& & \sum_{i=1}^{{n_{\mathrm{b}}}}{\left\|{{{{\pmb{w}}}}_i}\right\|}^2
\\
&\hspace{-0.3cm} \text{subject to} \hspace{-0.2cm}
&& \sum_{i=1}^{{n_{\mathrm{b}}}}{\left|{{{{\pmb{w}}}}_i^\dag{{{\pmb{g}}}_{k}}}\right|}^2 \geq {\left|{\alpha_k}\right|}^2\left({\delta_{\theta}}^2+{\sigma_{\mathrm{u},k}}^2\right) , \, \forall k \in {\left[{n_{\mathrm{s}}}\right]}.
\end{aligned}\end{aligned}$$ Problem can be effectively solved by CVX. In the following subsections, we use these suboptimal designs as the baseline schemes to assess the effectiveness of our proposed algorithms.
Parameter Value
------------------------------------- ----------------------------------------------------------------------------------------------------------
Network Topology [$10 \textsf{m} \times 10 \textsf{m}$ square box]{}
Located at the origin $\left(0,0\right)$
Transmission power $30$ dBm
\[0.0cm\]\[0.0cm\][Fusion Center]{}
\[0.2cm\] (effective noise bandwidth $2$ MHz
\[-0.1cm\] and noise figure 7 dB)
Placed uniformly over $\{x,y| x,y \in \left[-10,10\right]\}$
Energy harvesting efficiency $51\%$
Path loss [@SR:92:AP] ${\textsf{PL}\left(d\right)}=31.7 +27.6 \log_{10}\left(d_{[\textsf{meters}]}\right) \quad [\textsf{dB}]$
: System Parameters []{data-label="table:parameter"}
MSE Minimization
----------------
Figure \[fig:1\] shows the average MSE for distributed estimation versus iteration index when ${n_{\mathrm{s}}}=5$, ${{{\pmb{R}}}_{\mathrm{s}}}=10^{-2}{{\mathbf{I}}_{{n_{\mathrm{s}}}}}$, $P=30$ dBm, ${\delta_{\theta}}=1$, and ${n_{\mathrm{r}}}=5,10,15,20$. One can see that the average MSE monotonically decreases while the algorithm converges within a few iterations. It can be obviously seen that the MSE performance is improved with the increasing number of antennas at the FC, ${n_{\mathrm{r}}}$. In this figure, we also plot a benchmark ideal case for distributed estimation, where all the observations at the sensors are assumed to be directly available at the FC, which will give a lower bound on the MSE performance. One can see that the average MSE evaluated via our simulation tends to approach the benchmark value, $\left[{\mathbf{1}}^\dag{{{\pmb{R}}}_{\mathrm{s}}}^{-1}{\mathbf{1}}\right]^{-1}$, as ${n_{\mathrm{r}}}$ increases. In Fig. \[fig:2\], the average MSE for distributed estimation is shown as a function of ${n_{\mathrm{s}}}$ for the optimal and suboptimal solutions when $P=30$ dBm, ${n_{\mathrm{r}}}=5$, ${{{\pmb{R}}}_{\mathrm{s}}}=0.1{{\mathbf{I}}_{{n_{\mathrm{s}}}}}$, ${{{\pmb{R}}}_{\mathrm{n}}}=0.5{{\mathbf{I}}_{{n_{\mathrm{r}}}}}$, and ${\delta_{\theta}}=1$. As expected, the MSE performance is improved as ${n_{\mathrm{s}}}$ increases. In this example, we can see that the suboptimal solution shows a reasonable performance compared to the optimal one.
![The average MSE for distributed estimation versus iteration index when ${n_{\mathrm{s}}}=5$, ${{{\pmb{R}}}_{\mathrm{s}}}=0.1{{\mathbf{I}}_{{n_{\mathrm{s}}}}}$, $P=30$ dBm, ${\delta_{\theta}}=1$, and ${n_{\mathrm{r}}}=5,10,15,20$. []{data-label="fig:1"}](fig1.eps){width="52.00000%"}
![The average MSE for distributed estimation as a function of ${n_{\mathrm{s}}}$ when ${n_{\mathrm{r}}}=5$, $P=30$ dBm, ${{{\pmb{R}}}_{\mathrm{s}}}=0.1{{\mathbf{I}}_{{n_{\mathrm{s}}}}}$, and ${\delta_{\theta}}=1$.[]{data-label="fig:2"}](fig2.eps){width="52.00000%"}
Sensor index Harvested power \[dBm\] Transmit power \[dBm\]
-------------- ------------------------- ------------------------
1 -31.449 -31.449
2 -27.687 -34.347
\[0.3ex\] 3 -30.737 -30.737
\[0.3ex\] 4 -32.865 -32.865
\[0.3ex\] 5 -13.847 -48.067
\[0.3ex\] 6 -29.886 -31.999
\[0.3ex\] 7 -28.307 -32.964
\[0.3ex\]
: Power Control []{data-label="table:1"}
![The average minimum transmit power at the FC for distributed estimation versus iteration index when ${n_{\mathrm{s}}}=10$, ${{{\pmb{R}}}_{\mathrm{s}}}=0.1{{\mathbf{I}}_{{n_{\mathrm{s}}}}}$, $\gamma^{-1}=0.015$, ${\delta_{\theta}}=1$, and ${n_{\mathrm{r}}}=5,10,15,20$. []{data-label="fig:4"}](fig3.eps){width="52.00000%"}
![The average minimum transmit power at the FC for distributed estimation as a function of the distortion target $\gamma^{-1}$ when ${n_{\mathrm{s}}}=10$, ${{{\pmb{R}}}_{\mathrm{s}}}=0.1{{\mathbf{I}}_{{n_{\mathrm{s}}}}}$, ${\delta_{\theta}}=1$, and ${n_{\mathrm{r}}}=5$. []{data-label="fig:5"}](fig5.eps){width="52.00000%"}
In Table \[table:1\], in order to elaborate on the attributes of the optimal solution to the MSE minimization problem, we present the values of the harvested and transmit power of each sensor at the optimal solution to problem ${(\textsf{P1})}$. In this example, we set $P=30$ dBm, ${{{\pmb{R}}}_{\mathrm{s}}}=0.1{{\mathbf{I}}_{{n_{\mathrm{s}}}}}$, ${n_{\mathrm{s}}}=7$, and ${n_{\mathrm{r}}}=2$. One can see that some of the sensors do not use their maximum power harvested from the FC, which implies that power control is needed to guarantee the optimal solution. In other words, some of individual power constraints (i.e., the first constraint) in problem ${(\textsf{P1})}$ may not be fully utilized, or equivalently, the corresponding dual variables may be zero. This is attributed from the fact that transmission with the full power may increase the interference level at the FC, which in turn reduces the estimation reliability. In this example, sensors 2, 5, 6, and 7 use only a fraction of their harvested power.
![The distortion–power trade-off for distributed estimation with a common energy harvester when ${{{\pmb{R}}}_{\mathrm{s}}}=10^{-2}{{\mathbf{I}}_{{n_{\mathrm{s}}}}}$, ${\delta_{\theta}}=1$, ${n_{\mathrm{s}}}={n_{\mathrm{r}}}=4$, and ${n_{\mathrm{s}}}={n_{\mathrm{r}}}=8$. []{data-label="fig:6"}](fig6.eps){width="52.00000%"}
Total Power Minimization
------------------------
Figure \[fig:4\] illustrates the average minimum transmit power at the FC for distributed estimation versus iteration index at the distortion target of $\gamma^{-1}=0.015$ when ${n_{\mathrm{s}}}=10$, ${{{\pmb{R}}}_{\mathrm{s}}}=0.1{{\mathbf{I}}_{{n_{\mathrm{s}}}}}$, ${\delta_{\theta}}=1$, and ${n_{\mathrm{r}}}=5,10,15,20$. As shown in this figure, the proposed algorithm converges quickly, and the transmit power is reduced as the number of antennas, ${n_{\mathrm{r}}}$, increases. In Fig. \[fig:5\], the average minimum transmit power at the FC for distributed estimation is shown as a function of the distortion target $\gamma^{-1}$ when ${n_{\mathrm{s}}}=10$, ${{{\pmb{R}}}_{\mathrm{s}}}=0.1{{\mathbf{I}}_{{n_{\mathrm{s}}}}}$, ${\delta_{\theta}}=1$, and ${n_{\mathrm{r}}}=5$. It is clear that the more strict distortion requirement is, the more power is needed. Note that the distortion target must be no smaller than the benchmark MSE value such that the optimization problem is feasible. One can also see that the optimal scheme should be used for power saving. In this example, we can save the amount of transmit power of 7.44, 9.46, and 9.05 dBm at $\gamma^{-1}=0.02, 0.03, 0.04$, respectively, compared to the suboptimal case.
A Common Energy Harvester
-------------------------
Finally, we validate the performance of the distributed estimation system with a common energy harvester. Specifically, the power–distortion trade-off is ascertained by referring to Fig. \[fig:6\], where the optimal MSE is depicted as a function of the minimum transmit power $P$ for distributed estimation when ${{{\pmb{R}}}_{\mathrm{s}}}=10^{-2}{{\mathbf{I}}_{{n_{\mathrm{s}}}}}$, ${\delta_{\theta}}=1$, ${n_{\mathrm{s}}}={n_{\mathrm{r}}}=4$, and ${n_{\mathrm{s}}}={n_{\mathrm{r}}}=8$. In this figure, the region above each trade-off curve is achievable. As $P$ tends to infinity, the MSE converges to that of centralized estimations, i.e., $\left[{\mathbf{1}}^\dag{{{\pmb{R}}}_{\mathrm{s}}}^{-1}{\mathbf{1}}\right]^{-1}$, plotted with the dotted curve. Moreover, as expected, the achievable region gets broader for a larger $({n_{\mathrm{s}}},{n_{\mathrm{r}}})$ pair.
Concluding Remarks {#sec:6}
==================
Using the SDR, we developed a new framework for solving the network lifetime problem of a WSN. To that end, we adopted the notion of RF-based WPT as well as the multiple-antenna technology so that both the life span and the estimation performance are substantially improved. In this paper, two optimization problems were formulated and iteratively solved by two proposed algorithms, which turned out to guarantee the convergence at least to a local optimum. We showed that power control is indeed required at the optimal solution. It was also shown that having multiple antennas at the FC provides a significant improvement in the estimation performance. Especially, it was shown that as the number of antennas grows large, the MSE of the distributed estimation with the BLUE approaches that of centralized estimations.
Proof of Theorem \[thrm:1\] {#appdx:1}
===========================
We start by proving the first property of Theorem \[thrm:1\]. We exploit the strong duality and then examine the Karush-Kuhn-Tucker condition of ${(\textsf{SDR1})}$. Let $\nu$, $\{\lambda_k\}_{k=1}^{{n_{\mathrm{s}}}}$, and $\beta$ be the dual variables of problem ${(\textsf{SDR1})}$. The Lagrangian of problem ${(\textsf{SDR1})}$ is defined as $$\begin{aligned}
&\mathcal{L}\left({\bar{{{\pmb{Q}}}}},{\bar{{{\pmb{W}}}}},\eta,\nu,{\lambda_k},\beta\right)
=
-\operatorname{\mathrm{tr}}\left({\bar{{{\pmb{Q}}}}}{\mathbf{\Sigma}}\right)
+
\beta\left(\operatorname{\mathrm{tr}}\left({\bar{{{\pmb{W}}}}}\right)-\eta P\right)
\nonumber\\
&\hspace{-0.2cm}
+
\sum_{k=1}^{{n_{\mathrm{s}}}}
\lambda_k
\operatorname{\mathrm{tr}}\left(
{{\mathbf{D}}}_k
{\bar{{{\pmb{Q}}}}}-
{{{\pmb{G}}}}_k
{\bar{{{\pmb{W}}}}}\right)
+
\nu\left(
\operatorname{\mathrm{tr}}\left({\bar{{{\pmb{Q}}}}}{\mathbf{\Psi}}\right)
+
\eta{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}-
1
\right).\end{aligned}$$ Then, the dual function of problem ${(\textsf{SDR1})}$ is given by $$\begin{aligned}
\min_{{\bar{{{\pmb{Q}}}}}\succeq {\mathbf{0}},{\bar{{{\pmb{W}}}}}\succeq {\mathbf{0}},\eta>0}
\mathcal{L}\left({\bar{{{\pmb{Q}}}}},{\bar{{{\pmb{W}}}}},\eta,\nu,{\lambda_k},\beta\right),\end{aligned}$$ which can be equivalently expressed as $$\begin{aligned}
\label{eq:lgr}
\min_{{\bar{{{\pmb{Q}}}}}\succeq {\mathbf{0}},{\bar{{{\pmb{W}}}}}\succeq {\mathbf{0}},\eta>0}
\operatorname{\mathrm{tr}}\left({\bar{{{\pmb{Q}}}}}{{{\pmb{Y}}}}\right)
+
\operatorname{\mathrm{tr}}\left({\bar{{{\pmb{W}}}}}{{{\pmb{Z}}}}\right)
+
\eta\xi
-
\nu,\end{aligned}$$ where $$\begin{aligned}
\xi&=\nu{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}-\beta P\\
{{{\pmb{Y}}}}&=-{\mathbf{\Sigma}}+\sum_{k=1}^{{n_{\mathrm{s}}}}\lambda_k{{\mathbf{D}}}_k+\nu{\mathbf{\Psi}}\\
{{{\pmb{Z}}}}&=-\sum_{k=1}^{{n_{\mathrm{s}}}}\lambda_k{{{\pmb{G}}}}_k+\beta{{\mathbf{I}}_{}}.\end{aligned}$$ When we let $\nu^\star$, $\{\lambda_k^\star\}_{k=1}^{{n_{\mathrm{s}}}}$, and $\beta^\star$ be the optimal dual solutions to problem ${(\textsf{SDR1})}$, we define $$\begin{aligned}
\label{eq:Y}
{{{\pmb{Y}}}}^\star&=-{\mathbf{\Sigma}}+\sum_{k=1}^{{n_{\mathrm{s}}}}\lambda_k^\star{{\mathbf{D}}}_k+\nu^\star{\mathbf{\Psi}}\\
\label{eq:Z}
{{{\pmb{Z}}}}^\star&=-\sum_{k=1}^{{n_{\mathrm{s}}}}\lambda_k^\star{{{\pmb{G}}}}_k+\beta^\star{{\mathbf{I}}_{}}.\end{aligned}$$ Then, the optimal ${\bar{{{\pmb{Q}}}}}^\star$ must be the solution to the following problem: $$\begin{aligned}
\label{Pb:nQ}
\begin{aligned}
& \underset{{\bar{{{\pmb{Q}}}}}\succeq {\mathbf{0}}}{\text{minimize}}
&& \operatorname{\mathrm{tr}}\left({\bar{{{\pmb{Q}}}}}{{{\pmb{Y}}}}^\star\right).
\end{aligned}\end{aligned}$$ To guarantee a bounded optimal value, we must have ${{{\pmb{Y}}}}^\star
\succeq {\mathbf{0}}$, and hence we obtain the optimal value $\operatorname{\mathrm{tr}}({\bar{{{\pmb{Q}}}}}^\star{{{\pmb{Y}}}}^\star)=0$, which implies that $$\begin{aligned}
\label{eq:Y:cond}
{\bar{{{\pmb{Q}}}}}^\star{{{\pmb{Y}}}}^\star={\mathbf{0}}.\end{aligned}$$
In the same manner, it follows that ${\bar{{{\pmb{W}}}}}^\star{{{\pmb{Z}}}}^\star={\mathbf{0}}$ and $\eta^\star \xi^\star=0$, or equivalently $\xi^\star=0$ since $\eta^\star >0$, where $\xi^\star=\nu^\star{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}-\beta^\star P$. From , the dual problem to problem ${(\textsf{SDR1})}$ can be rewritten as follows: $$\begin{aligned}
\label{Pb:dual}
\begin{aligned}
& \underset{\nu,\{\lambda_k\}_{k=1}^{{n_{\mathrm{s}}}},\beta}{\text{minimize}}
&& \nu
\\
& \text{subject to}
&& {{{\pmb{Y}}}}\succeq {\mathbf{0}}, {{{\pmb{Z}}}}\succeq {\mathbf{0}}, \xi \geq 0
\\
&&& \beta \geq 0, \lambda_k \geq 0, \quad \forall k \in {\left[{n_{\mathrm{s}}}\right]}.
\end{aligned}\end{aligned}$$ Since the duality gap between problem ${(\textsf{SDR1})}$ and is zero, $\nu^\star$ is equal to the optimal value of problem ${(\textsf{SDR1})}$, which is positive. Thus, we conclude that $\nu^\star>0$. Next, we will show that $\beta^\star >0$. First, if there exits a $\lambda_k>0$, then from the condition ${{{\pmb{Z}}}}\succeq
{\mathbf{0}}$, it follows that $\beta^\star >0$. From the condition $\xi^\star=0$ and the facts that ${{{\pmb{R}}}_{\mathrm{n}}}\succ {\mathbf{0}}$ and $\nu^\star
>0$, we also conclude that $\beta^\star >0$.
To prove the second property of Theorem \[thrm:1\], we use the fact that for any two matrices of the same size ${\mathbf{A}}$ and ${\mathbf{B}}$, ${\textsf{rank}\left({\mathbf{A}}-{\mathbf{B}}\right)}\geq
{\left|{{\textsf{rank}\left({\mathbf{A}}\right)}-{\textsf{rank}\left({\mathbf{B}}\right)}}\right|}$ [@HJ:85:Book]. Since $\beta^\star>0$ and ${\textsf{rank}\left(\sum_{k=1}^{{n_{\mathrm{s}}}}\lambda_k^\star{{{\pmb{G}}}}_k\right)}\leq {n_{\mathrm{s}}}$, it follows from that ${\textsf{rank}\left({{{\pmb{Z}}}}^\star\right)} \geq {\left|{{n_{\mathrm{r}}}-{n_{\mathrm{s}}}}\right|} \geq
{n_{\mathrm{r}}}-{n_{\mathrm{s}}}$. Let $\textsf{Null}\left({{{\pmb{Z}}}}^\star\right)$ be the null space of ${{{\pmb{Z}}}}^\star$. Then from the condition ${\bar{{{\pmb{W}}}}}^\star{{{\pmb{Z}}}}^\star={\mathbf{0}}$, we must have ${\bar{{{\pmb{W}}}}}^\star \in
\textsf{Null}\left({{{\pmb{Z}}}}^\star\right)$. Since ${\textsf{rank}\left({{{\pmb{Z}}}}^\star\right)} \geq
{n_{\mathrm{r}}}-{n_{\mathrm{s}}}$ and ${\textsf{rank}\left({\bar{{{\pmb{W}}}}}^\star\right)} \leq
\textsf{dim}\left(\textsf{Null}\left({{{\pmb{Z}}}}^\star\right)\right)$, it follows that ${\textsf{rank}\left({\bar{{{\pmb{W}}}}}^\star\right)}\leq {n_{\mathrm{s}}}$. Using the fact that ${\bar{{{\pmb{W}}}}}^\star$ is an ${n_{\mathrm{r}}}\times {n_{\mathrm{r}}}$ matrix, we conclude that ${\textsf{rank}\left({\bar{{{\pmb{W}}}}}^\star\right)}\leq {n_{\mathrm{r}}}$.
Finally, we prove the property of the optimal solution ${\bar{{{\pmb{Q}}}}}^\star$. Since ${\mathbf{\Psi}} \succ {\mathbf{0}}$, $\sum_{k=1}^{{n_{\mathrm{s}}}}\lambda_k^\star{{\mathbf{D}}}_k \succeq {\mathbf{0}}$, and $\nu^\star > 0$, we obtain $$\begin{aligned}
{\textsf{rank}\left(\nu^\star{\mathbf{\Psi}} + \sum_{k=1}^{{n_{\mathrm{s}}}}\lambda_k^\star{{\mathbf{D}}}_k\right)}={n_{\mathrm{s}}}.\end{aligned}$$ Hence, from the definition of ${{{\pmb{Y}}}}^\star$ in , it follows that $$\begin{aligned}
\label{eq:Y:rank}
{\textsf{rank}\left({{{\pmb{Y}}}}^\star\right)} \geq {n_{\mathrm{s}}}-{\textsf{rank}\left({\mathbf{\Sigma}}\right)}={n_{\mathrm{s}}}-1.\end{aligned}$$ From the condition , ${\bar{{{\pmb{Q}}}}}^\star$ must lie in the null space of ${{{\pmb{Y}}}}^\star$. Therefore, ${\textsf{rank}\left({\bar{{{\pmb{Q}}}}}^\star\right)} \leq
\textsf{dim}\left(\textsf{Null}\left({{{\pmb{Y}}}}^\star\right)\right)$, which is upper-bounded by one due to . Now, assume that ${\textsf{rank}\left({{{\pmb{Y}}}}^\star\right)}={n_{\mathrm{s}}}$. Then from , it follows that ${\bar{{{\pmb{Q}}}}}^\star={\mathbf{0}}$, which cannot be the optimal solution to problem ${(\textsf{SDR1})}$. In consequence, we must have ${\textsf{rank}\left({{{\pmb{Y}}}}^\star\right)}={n_{\mathrm{s}}}-1$ and thus ${\textsf{rank}\left({\bar{{{\pmb{Q}}}}}^\star\right)}=1$, which completes the proof of Theorem \[thrm:1\].
Proof of Theorem \[thrm:3\] {#appdx:2}
===========================
We derive the result in by showing that the optimal MSE in ${(\textsf{P1--Sum})}$ and the optimal power in ${(\textsf{P2--Sum})}$ are the inverse of each other. We start the proof by introducing the following lemma.
\[lem:PD\] For a given ${{{\pmb{v}}}}$ in , let $({{{\pmb{a}}}}_1^\star,{{{\pmb{w}}}}_1^\star)$ and $({{{\pmb{a}}}}_2^\star,{{{\pmb{w}}}}_2^\star)$ be the optimal solutions to problems ${(\textsf{P1--Sum})}$ and ${(\textsf{P2--Sum})}$, respectively. We also let ${f_{1}\left(\cdot,\cdot\right)}$ and ${f_{2}\left(\cdot,\cdot\right)}$ be the objective functions in ${(\textsf{P1--Sum})}$ and ${(\textsf{P2--Sum})}$, respectively. Then, they obey the property: if $\gamma={f_{1}\left({{{\pmb{a}}}}_1^\star,{{{\pmb{w}}}}_1^\star\right)}$, then $({{{\pmb{a}}}}_2^\star,{{{\pmb{w}}}}_2^\star):=({{{\pmb{a}}}}_1^\star,{{{\pmb{w}}}}_1^\star)$; and if $P={f_{2}\left({{{\pmb{a}}}}_2^\star,{{{\pmb{w}}}}_2^\star\right)} $, then $({{{\pmb{a}}}}_1^\star,{{{\pmb{w}}}}_1^\star):=({{{\pmb{a}}}}_2^\star,{{{\pmb{w}}}}_2^\star)$.
First, it is worth noting that all the inequality constraints in ${(\textsf{P1--Sum})}$ and ${(\textsf{P2--Sum})}$ are satisfied with equality at the optimal solutions. For a given ${{{\pmb{v}}}}$, we have ${\left\|{{{{\pmb{w}}}}_1^\star}\right\|}^2=P$. We will prove that $({{{\pmb{a}}}}_1^\star,{{{\pmb{w}}}}_1^\star)$ is also a solution to ${(\textsf{P2--Sum})}$, i.e., $({{{\pmb{a}}}}_2^\star,{{{\pmb{w}}}}_2^\star):=({{{\pmb{a}}}}_1^\star,{{{\pmb{w}}}}_1^\star)$. We prove it by contradiction. Assume that $({{{\pmb{a}}}}_1^\star,{{{\pmb{w}}}}_1^\star)$ is not a solution to ${(\textsf{P2--Sum})}$, that is, there exists a feasible solution $({{{\pmb{a}}}}_1',{{{\pmb{w}}}}_1')$ to ${(\textsf{P2--Sum})}$ such that ${\left\|{{{{\pmb{w}}}}_1'}\right\|} < {\left\|{{{{\pmb{w}}}}_1^\star}\right\|}$. In other words, we can find a constant $c>1$ such that $$\begin{aligned}
\label{eq:lem:1}
{\left\|{{{{\pmb{w}}}}_1'}\right\|} < c{\left\|{{{{\pmb{w}}}}_1'}\right\|} \leq {\left\|{{{{\pmb{w}}}}_1^\star}\right\|}.\end{aligned}$$ Since the objective function in ${(\textsf{P1--Sum})}$ is monotonically increasing with the norm of ${{{\pmb{a}}}}$, it follows that ${f_{1}\left(c\,
{{{\pmb{a}}}}_1',c \,{{{\pmb{w}}}}_1'\right)} > {f_{1}\left({{{\pmb{a}}}}_1',{{{\pmb{w}}}}_1'\right)}$. Since $({{{\pmb{a}}}}_1',{{{\pmb{w}}}}_1')$ is feasible to ${(\textsf{P1--Sum})}$, we also have that $
{f_{1}\left({{{\pmb{a}}}}_1',{{{\pmb{w}}}}_1'\right)}\geq {f_{1}\left({{{\pmb{a}}}}_1^\star,{{{\pmb{w}}}}_1^\star\right)}$. Thus, we obtain $$\begin{aligned}
{f_{1}\left(c\, {{{\pmb{a}}}}_1',c\, {{{\pmb{w}}}}_1'\right)}
>
{f_{1}\left({{{\pmb{a}}}}_1',{{{\pmb{w}}}}_1'\right)}
\geq
{f_{1}\left({{{\pmb{a}}}}_1^\star,{{{\pmb{w}}}}_1^\star\right)}.\end{aligned}$$ From and the second constraint in ${(\textsf{P1--Sum})}$, the solution $(c\, {{{\pmb{a}}}}_1',c\, {{{\pmb{w}}}}_1')$ is also feasible to ${(\textsf{P1--Sum})}$, and yields a higher objective value than the optimal $({{{\pmb{a}}}}_1^\star,{{{\pmb{w}}}}_1^\star)$ does. This contradicts to the assumption that $({{{\pmb{a}}}}_1^\star,{{{\pmb{w}}}}_1^\star)$ is optimal to ${(\textsf{P1--Sum})}$, and thus $({{{\pmb{a}}}}_1^\star,{{{\pmb{w}}}}_1^\star)$ must be a solution to ${(\textsf{P2})}$.
The proof for the second claim can be found using the similar steps to the proof for the first one, and hence is omitted here.
From , the optimal $\textsf{mse}$ satisfies . Since the optimal $P$ and $\textsf{mse}$ are the inverse of each other, it follows that optimal $P$ also satisfies . Due to the fact that the above property holds for any ${{{\pmb{v}}}}$, it holds for the optimal ${{{\pmb{v}}}}$ as well, which completes the proof of Theorem \[thrm:3\].
[10]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{}
J. Yick, B. Mukherjee, and D. Ghosal, “Wireless sensor network survey,” *Comput. Networks*, vol. 52, no. 12, pp. 2292–2330, Aug. 2008.
R. Viswanathan and P. K. Varshney, “Distributed detection with multiple sensors: [Part I]{}–fundamentals,” *Proc. [IEEE]{}*, vol. 85, no. 1, pp. 54–63, Jan. 1997.
J. J. Xiao, S. Cui, Z. Q. Luo, and A. J. Goldsmith, “Linear coherent decentralized estimation,” *[IEEE]{} Trans. Signal Process.*, vol. 56, no. 2, pp. 757–770, Feb. 2008.
S. Kar and P. K. Varshney, “Linear coherent estimation with spatial collaboration,” *[IEEE]{} Trans. Inf. Theory*, vol. 59, no. 6, pp. 3532–3553, Jun. 2013.
A. S. Behbahani, A. M. Eltawil, and H. Jafarkhani, “Decentralized estimation under correlated noise,” *[IEEE]{} Trans. Signal Process.*, vol. 62, no. 21, pp. 5603–5614, Nov. 2014.
F. Jiang, J. Chen, and A. L. Swindlehurst, “Optimal power allocation for parameter tracking in a distributed amplify-and-forward sensor network,” *[IEEE]{} Trans. Signal Process.*, vol. 62, no. 9, pp. 2200–2211, May 2014.
V. V. Mai, Y. Jeong, and H. Shin, “Error exponents for distributed detection,” *[IEEE]{} Commun. Lett.*, vol. 20, no. 1, pp. 121–124, Jan. 2016.
R. Zhang and C. K. Ho, “[MIMO]{} broadcasting for simultaneous wireless information and power transfer,” *[IEEE]{} Trans. Wireless Commun.*, vol. 12, no. 5, pp. 1989–2001, May 2013.
S. Cui, J. J. Xiao, A. J. Goldsmith, Z. Q. Luo, and H. V. Poor, “Estimation diversity and energy efficiency in distributed sensing,” *[IEEE]{} Trans. Signal Process.*, vol. 55, no. 9, pp. 4683–4695, Sep. 2007.
H. Şenol and Tepedelenlioǧlu, “Performance of distributed estimation over unknown parallel fading channels,” *[IEEE]{} Trans. Signal Process.*, vol. 56, no. 12, pp. 6057–6068, Dec. 2008.
J. Li and G. AlRegib, “Distributed estimation in energy-constrained wireless sensor networks,” *[IEEE]{} Trans. Signal Process.*, vol. 57, no. 10, pp. 3746–3758, Oct. 2009.
J. Fang and H. Li, “Power constrained distributed estimation with correlated sensor data,” *[IEEE]{} Trans. Signal Process.*, vol. 57, no. 8, pp. 3292–3297, Aug. 2009.
J. Y. Wu and T. Y. Wang, “Power allocation for robust distributed best-linear-unbiased estimation against sensing noise variance uncertainty,” *[IEEE]{} Trans. Wireless Commun.*, vol. 12, no. 6, pp. 2853–2869, Jun. 2013.
M. Gastpar, “Uncoded transmission is exactly optimal for a simple [Gaussian]{} sensor networks,” *[IEEE]{} Trans. Inf. Theory*, vol. 54, no. 11, pp. 5247–5251, Nov. 2008.
M. K. Banavar, A. D. Smith, C. Tepedelenlioǧlu, and A. Spanias, “On the effectiveness of multiple antennas in distributed detection over fading [MACs]{},” *[IEEE]{} Trans. Wireless Commun.*, vol. 11, no. 5, pp. 1744–1752, May 2012.
I. Nevat, G. W. Peters, and I. B. Collings, “Distributed detection in sensor networks over fading channels with multiple antennas at the fusion centre,” *[IEEE]{} Trans. Signal Process.*, vol. 62, no. 3, pp. 671–683, Feb. 2014.
F. Jiang, J. Chen, A. L. Swindlehurst, and J. A. López-Salcedo, “Massive [MIMO]{} for wireless sensing with a coherent multiple access channel,” *[IEEE]{} Trans. Signal Process.*, vol. 63, no. 12, pp. 3005–3017, Jun. 2015.
A. Shirazinia, S. Dey, D. Ciuonzo, and P. [Salvo Rossi]{}, “Massive [MIMO]{} for decentralized estimation of a correlated source,” *[IEEE]{} Trans. Signal Process.*, vol. 64, no. 10, pp. 2499–2512, May 2016.
M. Nourian, S. Dey, and A. Ahlén, “Distortion minimization in multi-sensor estimation with energy harvesting,” *[IEEE]{} J. Sel. Areas Commun.*, vol. 33, no. 3, pp. 524–539, Mar. 2015.
P. Grover and A. Sahai, “Shannon meets [Tesla: Wireless]{} information and power transfer,” in *Proc. ISIT’10*, Austin, TX, Jun 2010, pp. 2363–2367.
X. Zhou, R. Zhang, and C. K. Ho, “Wireless information and power transfer: [Architecture]{} design and rate-energy tradeoff,” *[IEEE]{} Trans. Commun.*, vol. 61, no. 11, pp. 4754–4767, Nov. 2013.
Q. Shi, L. Liu, W. Xu, and R. Zhang, “Joint transmit beamforming and receive power splitting for [MISO SWIPT]{} systems,” *[IEEE]{} Trans. Wireless Commun.*, vol. 13, no. 6, pp. 3269–3280, Jun. 2014.
J. Xu, L. Liu, and R. Zhang, “Multiuser [MISO]{} beamforming for simultaneous wireless information and power transfer,” *[IEEE]{} Trans. Signal Process.*, vol. 62, no. 18, pp. 4798–4810, Sep. 2014.
L. Liu, R. Zhang, and K. C. Chua, “Secrecy wireless information and power transfer with [MISO]{} beamforming,” *[IEEE]{} Trans. Signal Process.*, vol. 62, no. 7, pp. 1850–1863, Apr. 2014.
C. Shen, W. C. Li, and T. H. Chang, “Wireless information and energy transfer in multi-antenna interference channel,” *[IEEE]{} Trans. Signal Process.*, vol. 62, no. 23, pp. 6249–6264, Dec. 2014.
L. Xie, Y. Shi, Y. T. Hou, and A. Lou, “Wireless power transfer and applications to sensor networks,” *[IEEE]{} Wireless Commun. Mag.*, vol. 20, no. 4, pp. 140–145, Aug. 2013.
S. Kim, R. Vyas, J. Bito, K. Niotaki, A. Collado, A. Georgiadis, and M. M. Tentzeris, “Ambient [RF]{} energy-harvesting technologies for self-sustainable standalone wireless sensor platforms,” *[IEEE]{} Trans. Signal Process.*, vol. 102, no. 11, pp. 1649–1666, Nov. 2014.
G. Yang, W. P. Tay, and Y. L. Guan, “Optimal wireless power transfer and harvested power allocation for diffusion [LMS]{} in wireless sensor networks,” in *Proc. GlobalSIP’15*, Orlando, FL, Dec 2015, pp. 1–5.
T. C. Hsu and Y. W. P. Hong, “Wireless power transfer for distributed estimation in wireless passive sensor network,” in *Proc. GlobalSIP’15*, Orlando, FL, Dec 2015, pp. 48–52.
L. Liu, R. Zhang, and K. C. Chua, “Multi-antenna wireless powered communication with energy beamforming,” *[IEEE]{} Trans. Commun.*, vol. 62, no. 12, pp. 4349–4361, Dec. 2014.
T. M. Cover and J. Thomas, *Elements of Information Theory*, 2nd ed. 1em plus 0.5em minus 0.4emNew York, NY, USA: Wiley.
O. Orhan, D. G[ü]{}nd[ü]{}z, and E. Erkip, “Source-channel coding under energy, delay, and buffer constraints,” *[IEEE]{} Trans. Wireless Commun.*, vol. 14, no. 7, pp. 3836–3849, Jul. 2015.
M. Gastpar and M. Vetterli, *Source-channel communication in sensor networks*, ser. Lecture Notes in Computer Science.1em plus 0.5em minus 0.4emNew York, NY, USA: Springer, 2006, vol. 2634.
H. Ju and R. Zhang, “Throughput maximization in wireless powered communication networks,” *[IEEE]{} Trans. Signal Process.*, vol. 13, no. 1, pp. 418–428, Jan. 2014.
S. M. Kay, *Fundamentals of Statistical Signal Processing: Estimation Theory*.1em plus 0.5em minus 0.4emNJ: Prentice Hall, 1993.
A. Charnes and W. W. Cooper, “Programming with linear fractional functionals,” *Naval Res. Logist. Quarter.*, vol. 9, pp. 181–186, 1962.
M. Grant and S. Boyd, “[CVX]{}: Matlab software for disciplined convex programming, version 2.1,” <http://cvxr.com/cvx>, Mar. 2014.
B.-T. Aharon and A. Nemirovski, *Lectures on Modern Convex Optimization: [Analysis]{}, Algorithms, and Engineering Applications*, ser. MOS-SIAM Series on Optimization.1em plus 0.5em minus 0.4emPhiladelphia, PA: SIAM, 2001.
R. Hunger, “Floating point operations in matrix-vector calculus,” Technische Universitat Munchen, Munchen, Germany, Tech. Rep., 2007.
K. Gomadam, V. R. Cadambe, and S. A. Jafar, “A distributed numerical approach to interference alignment and applications to wireless interference networks,” *[IEEE]{} Trans. Inf. Theory*, vol. 57, no. 6, pp. 3309–3322, Jun. 2011.
Q. Shi, M. Razaviyayn, Z. Q. Luo, and C. He, “An iteratively weighted [MMSE]{} approach to distributed sum-utility maximization for a [MIMO]{} interfering broadcast channel,” *[IEEE]{} Trans. Signal Process.*, vol. 59, no. 9, pp. 4331–4340, Sep. 2011.
T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base station antennas,” *[IEEE]{} Trans. Wireless Commun.*, vol. 9, no. 11, pp. 3590–3600, Nov. 2010.
R. Horn and C. R. Johnson, *Matrix Analysis*.1em plus 0.5em minus 0.4emCambridge, U.K.: Cambridge Univ. Press, 1985.
S. Y. Seidel and T. S. Rappaport, “914 MHz path loss prediction models for indoor wireless communications in multifloored buildings,” *[IEEE]{} Trans. Antennas and Propagation.*, vol. 40, no. 2, pp. 207–217, Feb. 1992.
J. A. Gutierrez, M. Naeve, E. Callaway, M. Bourgeois, V. Mitter, and B. Heile, “IEEE 802.15.4: A Developing Standard for Low-Power, Low-Cost Wireless Personal Area Networks,” *[IEEE]{} Network*, vol. 15, no. 5, Sept. 2001, pp. 12–19.
[^1]: This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (2014R1A1A2054577) and by the Ministry of Science, ICT & Future Planning (MSIP) (2015R1A2A1A15054248).
[^2]: V. V. Mai was with Dankook University, Yongin 448-701, Republic of Korea. He is now with the Department of Automatic Control, KTH Royal Institute of Technology, SE-100 44 Stockholm, Sweden. (E-mail: [email protected]).
[^3]: W.-Y. Shin (corresponding author) is with the Department of Computer Science and Engineering, Dankook University, Yongin 448-701, Republic of Korea (E-mail: [email protected]).
[^4]: K. Ishibashi is with the Advanced Wireless & Communication Research Center (AWCC), The University of Electro-Communications, Tokyo 182-8585, Japan (E-mail: [email protected]).
[^5]: Our WPT framework is not designed only for a scalar source. Note that there is no restriction to apply it for the vector case even if finding a theoretical optimal solution with no approximation for estimating vector-valued sources remains an open problem.
[^6]: For a Gaussian sensor network using the source-channel encoding strategy[@GAS:08:IT], the rate-distortion theorems (e.g., [@CT:06:Book Theorem 10.3.3] for orthogonal MACs and [@GAS:08:IT Section IV] for a coherent MAC) enable us to characterize the effect of $\tau$ (via the rate expressions) on the distortion performance (see, e.g., [@OGE:14:WCOM] and references therein). In this work, we adopt an *analog uncoded amplify-and-forward* scheme without bandwidth expansion, in which the nature of information is in an analog form, but not in a bitwise form [@GAS:08:IT]. As a result, the power-distortion tradeoff (e.g., [@GV:03:Book Theorem 1]) is independent of $\tau$, and hence in our work, the value of $\tau$ is assumed to be a constant. In practice, the value $T_0=\left(1-\tau\right)T$ corresponds to the amount of time that each sensor needs for observing, amplifying, and forwarding its observation to the FC.
[^7]: Otherwise, we can rewrite our problem along with a power offset, i.e., $E_k^{\mathrm{cir}}>0$, as a problem without any power offset for a smaller $\zeta_k$.
[^8]: The *minimal sufficient statistic* is defined in the sense that we no longer need the individual sample since all the information has been captured by the sufficient statistic [@Kay:93:Book]
[^9]: Here, we use the transformations $\eta^{-1}=\operatorname{\mathrm{tr}}\left(\left({{{\pmb{Q}}}}{\mathbf{\Psi}}\right)+{{{\pmb{v}}}}^\dag{{{\pmb{R}}}_{\mathrm{n}}}{{{\pmb{v}}}}\right)$, ${\bar{{{\pmb{Q}}}}}=\eta{{{\pmb{Q}}}}$, and ${\bar{{{\pmb{W}}}}}=\eta{{{\pmb{W}}}}$.
[^10]: We use the term centralized estimation to refer to the case for which the sensors’ data are perfectly available at the FC, which serves as a performance benchmark.
[^11]: This differs from what we have considered so far, where each sensor has its own energy harvester. This approach reduces the hardware complexity of sensors. However, it is feasible only if the sensors are closely, or even colocated.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We analyze the existences, properties and stabilities of the self-localized solutions of the nonlinear quantum harmonic oscillator (NQHO) using spectral renormalization method (SRM). We show that self-localized single, dual and triple soliton solutions of the NQHO do exists, however only single and dual soliton solutions satisfy the necessary Vakhitov and Kolokolov slope condition, therefore triple soliton solution is found to be unstable, at least for the parameter ranges considered. Additionally, we investigate the stability characteristics of the single and dual soliton solutions using a split-step Fourier scheme. We show that single and dual soliton solutions are pulsating during time stepping. We discuss our findings and comment on our results.'
author:
- Cihan Bayindir
title: 'Self-Localized Solutions of the Nonlinear Quantum Harmonic Oscillator'
---
\[sec:level1\] Introduction
===========================
Quantum harmonic oscillator (QHO) is one of the fundamental models in quantum mechanics [@Schrodinger; @Griffiths; @Liboff; @Pauli; @Messiah]. It can be viewed as the quantum mechanical analog of the simple harmonic oscillator of the classical vibration theory. The vibrations of atomic particles and molecules under the effect of restoring spring like potential due to molecular bonding are modeled within its frame [@Schrodinger; @Griffiths; @Liboff; @Pauli; @Messiah]. QHO admits exact solutions in terms of Hermite polynomials and can be extended to N-dimensions to model multidimensional atomic and molecular vibrations [@Schrodinger; @Griffiths; @Liboff; @Pauli; @Messiah].
Nonlinear quantum mechanics studies, on the other hand, are becoming increasingly popular recently [@Nattermann; @Reinisch; @Wang_Nl_entang; @Bay_Zeno; @Bay_TWMS2017; @Chia_Vedral]. Majority of the studies on nonlinear quantum phenomena are modeled in the frame of dynamic equations, i.e. the nonlinear Schrödinger equation (NLSE). Compared to the linear Schrödinger equation, NLSE can adequately model the cubic nonlinearity effects on the wavefunction. Such nonlinearities give rise to many interesting quantum mechanical phenomena including but are not limited to solitons, rogue waves, nonlinear quantum entanglement and quantum chaos. Analogs of these phenomena may also appear in the macroscopic physical environment.
Various studies which investigate the effects of nonlinearity on quantum oscillations do also exist in the literature [@Kivshar; @Bay2019_qho_rogue; @Carinena2007; @Zheng; @Ranada; @SchulzeHalberg2012; @SchulzeHalberg2013]. As discussed in the relevant literature, nonlinearity can arise in different ways. One possible way that gives rise to nonlinearity is the nonlinear behavior of bonding spring like stiffness and its corresponding potential. Another source of nonlinearity, which is investigated in this paper, arises due to strong electric and magnetic fields, which eventually leads to the cubic nonlinear term in the NQHO models.
In this paper, we consider the NQHO model first proposed in [@Kivshar] and recently generalized by us [@Bay2019_qho_rogue]. This model equation can only be solved numerically, but for some limiting cases exact analytical solutions do exist [@Kivshar]. We study the self-localized solutions of this NQHO model using the spectral renormalization method (SRM). More specifically, we obtain the self-localized single, dual and triple soliton solutions of the NQHO using the SRM and discuss their properties. Additionally, we investigate the stability characteristics of those self-localized solutions using a split-step Fourier scheme which is used to perform the time stepping. We show that self-localized single and dual soliton solutions of the NQHO are stable and have pulsating behavior in time, at least for some of the parameter ranges considered in this paper. We also show that the self-localized triple soliton solution of the NQHO is not stable since it violates the necessary Vakhitov and Kolokolov slope condition for stability, at least for the parameter ranges considered.
\[sec:level2\]A Nonlinear Quantum Harmonic Oscillator Model
===========================================================
Linear quantum harmonic oscillator’s (LQHO) Hamiltonian (energy) can be given by [@Schrodinger; @Griffiths; @Liboff; @Pauli; @Messiah] $$\widehat{H}=\frac{\widehat{p}^2 }{2 m}+\frac{1}{2}k x^2=\frac{\widehat{p}^2 }{2 m}+\frac{1}{2}m \omega^2 x^2
\label{eq01}$$ In this formula $\widehat{H}$ denotes the Hamiltonian of the LQHO, $m$ denotes the particle mass and $k$ is the bonding stiffness of the atomic particle, which is analogous to spring constant in a classical mass-spring-dashpot system. In this formulation, the momentum operator can be given by $\widehat{p}=-i \hbar \partial / \partial x$ where $\hbar$ is the reduced Planck’s constant. Thus, the unsteady Schrödinger equation can be derived using the Hamiltonian formalism as $$i \hbar \frac{\partial \psi}{\partial t}+\frac{\hbar ^2 }{2 m}\frac{\partial^2 \psi}{\partial x^2}-\frac{1}{2}m \omega^2 x^2 \psi=0
\label{eq02}$$ where $i$ denotes the imaginary unity, $t$ is time variable, $x$ is the position and $\psi (x,t)$ is the wavefunction. This form of the LQHO is commonly studied in the literature [@Schrodinger; @Griffiths; @Liboff; @Pauli; @Messiah]. While majority of the studies on quantum harmonic oscillations utilizes linear models, few different forms of NQHO models are proposed in [@Kivshar; @Bay2019_qho_rogue; @Carinena2007; @Zheng; @Ranada; @SchulzeHalberg2012; @SchulzeHalberg2013]. Generally, two forms of nonlinearity are considered in these studies due to different nonlinear behaviors. One of them is the nonlinearly behaving molecular bond, which is commonly modeled using spring constant analogy, which is represented by different forms of the potential function than the commonly used trapping-well potential. However, in order to account for the effects of high-order electric and magnetic fields on wavefunction, a NQHO model is proposed in [@Kivshar]. The form of the NQHO proposed in [@Kivshar] can be given as $$i\psi_t + \frac{\partial^2 \psi}{\partial x^2}- x^2 \psi + \sigma \left|\psi \right|^2 \psi=0
\label{eq03}$$ In here $\sigma$ is a constant which controls the strength of the nonlinearity. This equation can be derived by applying the non-dimensional parameters given in [@Kivshar] to the LQHO and including the nonlinear term of the NLSE. Setting, $\sigma=0$, NQHO can be linearized and reduced to the LQHO, which admits solutions in the form of $$\psi(x,t)=U(x) e^{-i \mu t}
\label{eq04}$$ The analytical solution in this form can only be derived for the discrete spectrum of $\mu_n=1+2n$ where $n=0,1,2,...$ [@Kivshar]. For this discrete spectrum, the amplitude functions can be derived as $U_n=(2^n n! \sqrt{\pi} )^{-1/2} e^{-x^2/2}H_n(x)$ where $H_n(x)$ are the Hermite polynomials. These polynomials are described by $$H_n(x)=(-1)^n e^{x^2/2} \frac{d^n (e^{-x^2/2})}{dx^n}
\label{eq05}$$ giving $H_0=1$, $H_1=2x$, $H_2=4x^2-2$, $H_3=8x^3-12x$, ... etc. [@Kivshar; @Abramowitz; @Ryshik]. Eq.(\[eq03\]) requires numerical solution except for some limiting cases [@Kivshar], and a continuous frequency spectrum is considered in [@Kivshar] for its numerical solution.
In this paper, we study a slightly more general version of the NQHO given by Eq.(\[eq03\]) which was first proposed by Kivshar [@Kivshar] and later extended by us recently [@Bay2019_qho_rogue]. To model the effects of variable potential due to varying bonding (spring) stiffness, we use a potential well constant, $\alpha$. Thus, the form of the non-dimensional NQHO equation studied in this paper can be written as $$i\psi_t + \frac{\partial^2 \psi}{\partial x^2}-\alpha x^2 \psi + \sigma \left|\psi \right|^2 \psi=0
\label{eq06}$$ As before, the $t$ denotes the non-dimensional temporal parameter and $x$ denotes the non-dimensional spatial parameter. In the next sections of this paper, we study the existences and properties of the self-localized solutions of the NQHO given by Eq.(\[eq06\]) using the spectral renormalization method (SRM). Additionally, using a split-step Fourier scheme, we study the stability characteristics of such solutions of the NQHO.
Spectral Renormalization Method for Finding the Self-Localized Solutions of the Nonlinear Quantum Harmonic Oscillator
=====================================================================================================================
There are few different techniques used to find the self-localized solutions of nonlinear systems. Some of these techniques are the shooting, self-consistency and relaxation techniques [@Petviashvili; @Ablowitz; @Yang; @Fibich; @Bay_CSRM]. One of the most commonly used methods for this purpose is the Petviashvili’s method, in which the governing nonlinear equation is transformed into Fourier space similar to the other Fourier spectral methods using FFT routines [@Canuto; @Karjadi2010; @Karjadi2012; @trefethen; @BayPRE1; @BayPRE2; @BayTWMS2016; @Demiray2015; @BayPLA; @Baysci; @BayTWMS2015; @Bay_arxNoisyTun; @Bay_arxNoisyTunKEE; @Bay_arxEarlyDetectCS; @BayMS]. Then, a convergence factor is used in accordance with the degree of the nonlinear term [@Petviashvili; @Ablowitz]. Petviashvili was the person who proposed this approach and he also applied this method to the Kadomtsev-Petviashvili equation [@Petviashvili]. In order to treat various forms of the homogeneities different than the fixed ones, Petviashvili’s method is extended to the spectral renormalization method (SRM), which can be used to find the self-localized solutions of more general nonlinear equations [@Ablowitz; @Yang; @Fibich]. Later, another extension of the Petviashvili’s method is proposed by us [@Bay_CSRM], which is capable of finding the self-localized solutions in nonlinear waveguides under missing spectral information. This method is named as the compressive spectral renormalization method (CSRM) [@Bay_CSRM]. The SRM transforms the governing equation into wavenumber space using the Fourier transforms and then couples it to a nonlinear integral equation. The iterations are performed in the wavenumber space. Due to the coupling of these equations, the energy is conserved and the initial conditions converge to the self-localized solutions of the system studied [@Ablowitz]. Details and other possible uses of the SRM can be seen in [@Ablowitz].
In this section we apply the SRM to the NQHO model given by Eq.(\[eq06\]) in order to study the self-localized solutions of the NQHO. We start with rewriting the NQHO model given by Eq.(\[eq06\]) in the form of $$i\psi_t + \psi_{xx}-V(x)\psi +\sigma \left| \psi \right|^2 \psi=0
\label{eq07}$$ where $V(x)=\alpha x^2$ is the trapping-well potential function. Eq. (\[eq07\]) can be rewritten as $$i\psi_t + \psi_{xx} -V(x)\psi+ N(\left| \psi \right|^2) \psi =0
\label{eq08}$$ where $N(\left| \psi \right|^2)=\sigma \left| \psi \right|^2$ is the nonlinear term of the NQHO. Using the ansatz, $\psi(x,t)=\eta(x,\mu) \textnormal{exp}(i\mu t)$, Eq. (\[eq08\]) becomes $$-\mu \eta + \eta_{xx} -V(x)\eta+ N(\left| \eta \right|^2) \eta =0
\label{eq09}$$ where $\mu$ is the soliton eigenvalue. Iterations performed using the spectral representation of Eq. (\[eq09\]) may become singular [@Ablowitz]. In order to avoid the singularity of the scheme, a $p \eta$ term with $p>0$ can be added and substracted from Eq. (\[eq09\]) [@Ablowitz]. With this modification, the 1D Fourier transform of Eq. (\[eq09\]) can be found using the definition of the 1D Fourier transform as $$\widehat{\eta} (k)=F[\eta(x)] = \int_{-\infty}^{+\infty} \eta(x) \exp[i(kx)]dx
\label{eq10}$$ thus becomes $$\widehat{\eta} (k)=\frac{(p+| \mu|)\widehat{\eta}}{p+\left| k \right|^2} -\frac{F[V \eta]-F \left[ N(\left| \eta \right|^2) \eta \right]}{p+\left| k \right|^2}
\label{eq11}$$
The iteration formula given in Eq. (\[eq11\]) can be used to find the self-localized solutions of the NQHO model equation, however the iterations may diverge or it may tend to zero [@Ablowitz]. This problem can be solved by defining a new variable as $\eta(x)=\beta \xi(x)$, which has the Fourier transform given by $\widehat{\eta}(k)=\beta \widehat{\xi}(k)$. Using these substitutions, Eq. (\[eq11\]) can be rewritten as $$\begin{split}
\widehat{\xi}_{j+1} (k) &=\frac{(p+| \mu|)}{p+\left| k \right|^2}\widehat{\xi_j}-\frac{F[V \xi_j]}{p+\left| k \right|^2} \\
& +\frac{1}{p+\left| k \right|^2} F\left[\sigma { \left| \beta_j \right|^2 \left| \xi_j \right|^2} {\xi_j} \right] =R_{\beta_j}[\widehat{\xi}_j (k)]
\label{eq12}
\end{split}$$ which is the iteration equation of the SRM in wavenumber domain. The algebraic condition on the parameter $\beta$, which prevents the scheme from diverging or tending to zero, can be derived using the energy conservation principle. Multiplying Eq. (\[eq12\]) with the $\widehat{\xi}^*(k)$ term, where $^*$ shows the complex conjugation, and integrating the resulting equation to evaluate the total energy of the system, it is possible to derive the algebraic condition as $$\int_{-\infty}^{+\infty} \left|\widehat{\xi} (k)\right|^2 dk= \int_{-\infty}^{+\infty} \widehat{\xi}^* (k) R_{\beta}[\widehat{\xi} (k)]dk
\label{eq13}$$ This becomes the normalization constraint, which prevents the scheme from diverging or tending to zero. The method summarized above is the SRM [@Ablowitz] and applied to NQHO in this paper. Starting the simulations using a single or multi-Gaussians as initial conditions, Eq. (\[eq11\]) and Eq. (\[eq13\]) are applied iteratively to find the self-localized solutions of the NQHO. The iterations are continued until the parameter $\beta$ convergences, for which the cut-off criteria is given for different simulations in the next part of this paper.
\[sec:level3\]Results and Discussion
====================================
In this section we apply the SRM summarized above to find the self-localized solutions of the NQHO model equation. In our simulations throughout this paper we use $N=1024$ spectral components. Starting with a single humped Gaussian in the form of $\exp{(-(x-x_0)^2)}$, where $x_0$ is taken as $0$, SRM converges to single humped self-localized soliton solutions of the NQHO. We plot these self-localized solutions of the NQHO model equation in Fig. \[fig1\] for various values of, $\alpha$, the trapping-well potential coefficient. Other parameters are selected as $p=30, \sigma=1, \mu=10.8$ for this simulation. The convergence criteria for this simulation is taken as the normalized change of $\beta$ to be less than $10^{-15}$. The single humped Gaussian converges to self-localized solution of the NQHO rapidly. The exact form of the analytical solution is unknown, however the profile resembles solitary waves in shape, however they have an asymmetric structure. This result can also be verified with the findings presented in [@Kivshar]. As one can realize from the figure, the increase in trapping-well potential strength results in bigger waves. One possible explanation for this result is that, bigger $\alpha$ values represents the trapping of the solitons in a more confined well, thus such bigger solitons are formed.
![Self-localized single solitons for different trapping well potential strength, $\alpha$.[]{data-label="fig1"}](fig1.eps){width="3.4in"}
Next we turn our attention to investigate the effects of nonlinearity coefficient, $\sigma$, on the existence and properties of the self-localized solitons of the NQHO. With this motivation, we depict Fig. \[fig2\]. As before, the simulation parameters are selected as $p=30, \mu=10.8$ and $\alpha=1$. Various values of $\sigma$ are used as indicated in figure and again, the cut-off criteria for the SRM iterations are selected as normalized change of $\beta$ being less than $10^{-15}$. As Fig. \[fig2\] confirms, self-localized solutions of the NQHO exists for various values of $\sigma$ as well and they tend to be smaller as $\sigma$ grows larger.
![Self-localized single solitons for different nonlinearity strength, $\sigma$.[]{data-label="fig2"}](fig2.eps){width="3.4in"}
It is important to discuss the stability characteristics of the self-localized solutions of the NQHO model equation found using the SRM. For a soliton to be stable two conditions must be satisfied. The first conditions is known as the slope condition, $dP(\mu)/{d\mu}<0$, as first derived by Vakhitov and Kolokolov [@VakhitovStability; @SivanStability]. In here $P=\int \left| \psi \right|^2 dx$ denotes the soliton power. Therefore the soliton power as a function of soliton eigenvalue must be examined to analyze the stability characteristics of the self-localized solutions of the NQHO. With this aim, we plot the soliton power as a function of soliton eigenvalue in Fig. \[fig3\] for the single humped self-localized solution of the NQHO using parameters of $p=30, \alpha=1, \sigma=1$. We investigate the stability characteristics of such solitons within the soliton eigenvalue interval of $\mu=0-50$.
![Self-localized single soliton power as a function of soliton eigenvalue, $\mu$.[]{data-label="fig3"}](fig3.eps){width="3.4in"}
The figure clearly indicates that the Vakhitov and Kolokolov slope condition, $dP(\mu)/{d\mu}<0$, is satisfied piecewise. This results suggests that self-localized solitons of the NQHO may be stable for some ranges of the soliton eigenvalue, $\mu$, such as $\mu \approx [10.7-10.8]$ for the parameters considered. Being a necessary condition for the soliton stability, the Vakhitov and Kolokolov slope condition is not a sufficient one. The second condition for the soliton stability is the spectral condition. The spectral conditions states that the operator $L_+= -\Delta + V -1[N(\eta)-2\eta^2 N'(\eta)]-\mu$ for the NQHO problem that we analyze [@WeinsteinStability; @SivanStability], should have at most one eigenvalue which should be nonzero [@SivanStability]. In here, $\Delta$ shows the diffraction term, $V$ is the trapping-well potential and $N(\eta)$ is the nonlinear term given above. Spectral condition can be analyzed analytically or numerically. For various nonlinear models studied in the literature, a numerical approach is the more commonly used one.
With this aim, we investigate the temporal stability of self-localized solutions of the NQHO model equation given by Eq.(\[eq06\]) using a split-step Fourier method (SSFM). This SSFM is recently proposed by us [@Bay2019_qho_rogue] and it is validated using the analytical solutions of the some limiting cases of the NQHO model equation given by Eq.(\[eq06\]). It is also tested against a $4^{th}$ order Runge-Kutta integrator. The SSFM, splits the governing NQHO model equation into linear and nonlinear parts. As a possible splitting, the nonlinear part can be written as $$i\psi_t= -(-\alpha x^2 + \sigma \left| \psi \right|^2)\psi
\label{eq15}$$ which can be exactly solved. This integrations gives $$\tilde{\psi}(x,t_0+\Delta t)=e^{i(-\alpha x^2 + \sigma \left| \psi_0 \right|^2)\Delta t}\ \psi_0
\label{eq16}$$ In here, $\psi_0=\psi(x,t_0)$ denotes the initial condition. $\Delta t$ indicates the time step, which is selected as $\Delta t=5\times 10^{-5}$ throughout this study, which does not cause stability problems. The remaining linear part of the NQHO model equation is $$i\psi_t=-\psi_{xx}
\label{eq17}$$ One can compute the linear part of the NQHO equation in periodic domain using spectral techniques. Using the most commonly utilized Fourier spectral technique, the linear part can be evaluated as $$\psi(x,t_0+\Delta t)=F^{-1} \left[e^{-i k^2\Delta t}F[\tilde{\psi}(x,t_0+\Delta t) ] \right]
\label{eq18}$$ In here $k$ is the wavenumber parameter [@Bay_CSRM]. Inserting Eq.(\[eq16\]) into Eq.(\[eq18\]), the complete form of the SSFM iteration formula for the numerical solution of the NQHO model equation can be derived as $$%\begin{split}
\psi(x,t_0+\Delta t)= F^{-1} \left[e^{-ik^2\Delta t} F[ e^{i(-\alpha x^2+ \sigma \left| \psi_0 \right|^2 )\Delta t}\ \psi_0 ] \right]
%\end{split}
\label{eq19}$$ Starting from the initial conditions, we perform the integration of the NQHO model equation using this procedure. In order to study the temporal stabilities of the self-localized solutions of the NQHO model equations, the initial conditions are taken as the self-localized solutions found by the SRM, such as the ones depicted in Fig. \[fig1\] and Fig. \[fig2\]. Using the normalized self-localized soliton obtained by SRM for $p=30, \alpha=1, \sigma=1, \mu=10.8$ as the initial condition, we observe a pulsating behavior during the time stepping. This can also be understood by checking Fig. \[fig4\], where the soliton power as a function of time is depicted.
![Self-localized single soliton power as a function of time, $t$.[]{data-label="fig4"}](fig4.eps){width="3.4in"}
![Self-localized single soliton at two different times, $t=0$ and $t=500$.[]{data-label="fig5"}](fig5.eps){width="3.4in"}
In order to illustrate the pulsation properties of the self-localized solutions of the NQHO, we depict Fig. \[fig5\]. In this figure, the initial self-localized soliton profile and the same soliton after an integration time of $t=500$ is given. In our simulations we observe the pulsating recurrence type behavior between these two profiles, that is these forms are interchanging to each other gradually during time stepping in a recursive way.
![Self-localized dual solitons for different trapping well potential strength, $\alpha$.[]{data-label="fig6"}](fig6.eps){width="3.4in"}
Next, we turn our attention to dual humped self-localized solitons of the NQHO model equation. Using $N=1024$ spectral components as before, selecting the computation parameters are $p=150, \sigma=1, \mu=10.8$, and specifying the convergence as the normalized change of $\beta$ to be less than $10^{-15}$ as before, we depict the dual humped self-localizes solutions of the NQHO model equation in Fig. \[fig6\] for various values of $\alpha$. The initial condition for the SRM is selected as $\exp{(-(x-x_0)^2)}+\exp{(-(x-x_1)^2)}$ for which the locations of the Gaussians are selected as $-x_0=x_1=10$.
![Self-localized dual solitons for different nonlinearity strength, $\sigma$.[]{data-label="fig7"}](fig7.eps){width="3.4in"}
Setting $\alpha=1$ and keeping the other parameters as before, the dual humped self-localized solitons are obtained by the SRM for various values of the nonlinearity coefficient, $\sigma$, and they are depicted in Fig. \[fig7\]. As Fig. \[fig6\] and Fig. \[fig7\] confirms, the dual humped self-localized solutions of the NQHO model equation to also exist. As in the case of single humped solitons, the dual humped solitons are also asymmetric about the vertical axis, similar to the soliton profiles given in [@Kivshar].
![Self-localized dual soliton power as a function of soliton eigenvalue, $\mu$.[]{data-label="fig8"}](fig8.eps){width="3.4in"}
In order to check the stability characteristics of dual humped self-localized solitons, we depict soliton power as a function of soliton eigenvalue, $\mu$, in Fig. \[fig8\]. The parameters of computation are selected as $p=150, \alpha=1, \sigma=1$ and the soliton eigenvalue interval of $\mu=0-50$ is scanned. As indicated in Fig. \[fig8\], the Vakhitov and Kolokolov slope condition is only satisfied in a small interval of $\mu \approx 0.65-1.25$, thus stable solitons can be found in this range for the parameters considered.
![Self-localized dual soliton power as a function of time, $t$.[]{data-label="fig9"}](fig9.eps){width="3.4in"}
![Self-localized dual soliton at two different times, $t=0$ and $t=200$.[]{data-label="fig10"}](fig10.eps){width="3.4in"}
In order to check the temporal stability of the dual humped self-localized soliton we again use the SSFM summarized above. Starting the time stepping using the normalized dual humped self-localized soliton obtained for $p=150, \alpha=1,\sigma=1, \mu=1$, as the initial condition, the soliton power as a function of time is depicted in Fig. \[fig9\], and the dual humped soliton at two different times of $t=0$ and $t=200$ is depicted in Fig. \[fig10\]. Similar to the single humped self-localized solitons, the dual humped soliton also exhibits a pulsating behavior, the form is gradually and recursively interchanging from the form given at $t=0$ to the one given at $t=200$.
![Self-localized triple solitons for different trapping well potential strength, $\alpha$.[]{data-label="fig11"}](fig11.eps){width="3.4in"}
Lastly, we investigate the properties of triple humped self-localized solutions of the NQHO model equation obtained by the SRM. For the SRM to converge to such soliton solutions, parameters needed to be relaxed and are selected as $p=1000, \alpha=1, \sigma=1, \mu=1$. The convergence criteria for the SRM is also relaxed to be the normalized change of $\beta$ being less than $10^{-1}$. The initial condition used in SRM to obtain the triple humped solitons is selected as superimposed three Gaussians in the form of $\exp{(-(x-x_0)^2)}+\exp{(-(x-x_1)^2)}+\exp{(-(x-x_2)^2)}$ where $x_0=-10, x_1=0, x_2=10$. The triple humped self-localized solutions of the NQHO model equation are depicted in Fig. \[fig11\] and Fig. \[fig12\], for various values of $\alpha$ and $\sigma$, respectively.
![Self-localized triple solitons for different nonlinearity strength, $\sigma$.[]{data-label="fig12"}](fig12.eps){width="3.4in"}
![Self-localized dual soliton power as a function of soliton eigenvalue, $\mu$.[]{data-label="fig13"}](fig13.eps){width="3.4in"}
Contrary to results presented for single and dual humped self-localized solitons of the NQHO model equation earlier in this paper, the soliton power vs soliton eigenvalue graph is depicted in Fig. \[fig13\] indicates that Vakhitov-Kolokolov slope condition is not satisfied for the triple humped self-localized solutions of the NQHO model equation. We should note that scanned range of the soliton eigenvalue is $\mu=0-50$, however only some part of it is depicted for illustrative purposes. Since the graph increases monotonically, it is possible to conclude that triple humped self-localized solitons of the NQHO model equation are unstable, at least for the parameter ranges considered in this study.
Findings and the computational approach based on the SRM and the SSFM we have proposed in this paper for the investigations of the self-localized solitons of the NQHO model equation can be used to analyze nonlinear quantum harmonic oscillations. Additionally atomic vibration resonances, bonding and bond breaking strength of molecules under the effect of nonlinear electric and magnetic fields and trapping well potentials having variable strengths can also be studied within this frame. The procedure proposed in this paper and our findings can also find many possible applications in the macroscopic level, such as Bose-Einstein condensation.
\[sec:level1\]Conclusion and Future Work
========================================
In this paper we analyzed the existences and properties of the self-localized solitons of a nonlinear quantum harmonic oscillator model. In order to do so, we implemented a numerical approach based on the spectral renormalization scheme and showed that single, dual and triple self-localized soliton solutions of NQHO do exists. We compared our findings with the results existing in the literature. We also studied stability characteristics of such soliton solutions of the nonlinear quantum harmonic oscillator model using a split-step Fourier scheme. We showed that single and dual humped self-localized solutions of nonlinear quantum harmonic oscillator model equation are stable within some subintervals of the soliton eigenvalue, which may be considered as a piecewise continuous spectrum. Time stepping analysis showed that single and dual humped self-localized solitons of the nonlinear quantum harmonic oscillator model equation are pulsating in time. However, the triple humped self-localized soliton solution of the nonlinear quantum harmonic oscillator turned out to be unstable since it violates the necessary Vakhitov-Kolokolov slope condition. Our findings can be used to analyze nonlinear quantum harmonic oscillations under varying molecular bond stiffness and/or nonlinear field effects. Some other similar phenomena observed at the macroscopic level, such as in the Bose-Einstein condensation, can also be investigated using our results and the framework proposed in this paper.
[00]{}
E. Schrödinger, Annalen d. Physik (4), 79, 489 (1926).
D. J. Griffiths, *Introduction to Quantum Mechanics* (Prentice Hall, Harlow, 2004).
R. L. Liboff, *Introductory Quantum Mechanics* (Addison–Wesley, New York, 2002).
W. Pauli, *Wave Mechanics: Volume 5 of Pauli Lectures on Physics* (Dover, New York, 2000).
A. Messiah, *Quantum Mechanics*, (North-Holland, Amsterdam, 1967).
P. Nattermann, Sym. Nonl. Math. Phys., [**[2]{}**]{}, 270 (1997).
G. Reinisch, Phys. A: Stat. Mech. Appl., [**[206]{}**]{}, 229 (2001).
G. Wang, L. Huang, Y. C. Lai, and C. Grebogi, Phys. Rev. Lett., [**[112]{}**]{}, 110406 (2004).
C. Bayindir and F. Ozaydin. Opt. Commun., [**[413]{}**]{}, 141 (2018).
C. Bayindir. TWMS Journal of Applied and Engineering Mathematics, [**[7]{}**]{}, 236 (2017).
A. Chia, M. Hajdušek, R. Fazio, L. C. Kwek and V. Vedral, arXiv Preprint, arXiv:1711.07376, 2017.
Y. S. Kivshar, T. J. Alexander and S. K. Turitsyn, Phys. Lett. A, [**[278]{}**]{}, 225 (2001).
C. Bayindir, arXiv Preprint, arXiv:1902.08823, 2019.
J. F. Cariñena, M. F. Rañada and M. Santander, Ann. Phys., [**[322]{}**]{}, 434 (2007).
L. Zheng, T. Wang, X. Zhang and L. Ma, Appl. Math. Lett., [**[26]{}**]{}, 463 (2013).
M. F. Ranada, J. Math. Phys., [**[55]{}**]{}, 082108 (2014).
A. Schulze-Halberg and J. R. Morris, J. Phys. A, [**[45]{}**]{}, 305301 (2012).
A. Schulze-Halberg and J. R. Morris, J. Math. Phys., [**[54]{}**]{}, 112107 (2013).
M. Abramowitz and I. Stegun, *Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables*, (Dover Publications, New York, 1964).
I. M. Ryshik and I. S. Gradstein, *Tables of Integrals, Series, and Products*, (Academic Press, New York, 1965).
V. I. Petviashvili. Soviet Journal of Plasma Physics JETP, [**[2]{}**]{}, 257 (1976).
M. J. Ablowitz and Z. H. Musslimani. Optics Letters, [**[30]{}**]{}, 16, 2140 (2005).
T.I. Lakoba and J. Yang. Journal of Computational Physics, [**[226]{}**]{}, 1668 (2007).
G. Fibich. The Nonlinear Schrodinger Equation: Singular Solutions and Optical Collapse, Springer-Verlag (2015).
C. Bayindir. TWMS: Journal of Applied and Engineering Mathematics, [**[8-2]{}**]{}, 425 (2018).
C. Canuto. *Spectral Methods: Fundamentals in Single Domains* (Springer-Verlag, 2006).
E. A. Karjadi, M. Badiey and J. T. Kirby. The Journal of the Acoustical Society of America, [**[127]{}**]{}, 1787 (2010).
E. A. Karjadi, M. Badiey, J. T. Kirby and C. Bayindir. IEEE Journal of Oceanic Engineering, [**[37-1]{}**]{}, 112 (2012).
L. N. Trefethen. *Spectral Methods in [MATLAB]{}*, (SIAM, 2000).
C. Bayindir. Phys. Rev. E, [**[93]{}**]{}, 032201 (2016).
C. Bayindir. Phys. Rev. E, [**[93]{}**]{}, 062215 (2016).
C. Bayindir. TWMS: Journal of Applied and Engineering Mathematics, [**[6-1]{}**]{}, 135 (2016).
H. Demiray and C. Bayindir. Phys. of Plasm., [**[22]{}**]{}, 092105 (2015).
C. Bayindir. Physics Letters A, [**[380]{}**]{}, 156 (2016).
C. Bayindir. Sci. Rep., [**[6]{}**]{}, 22100 (2016).
C. Bayindir. TWMS Journal of Applied and Engineering Mathematics, [**[5-2]{}**]{}, 298 (2015).
C. Bayindir. TWMS Journal of Applied and Engineering Mathematics, [**[7-2]{}**]{}, 236 (2017).
C. Bayindir. 12th International Congress on Advances in Civil Engineering, Istanbul, Turkey (2016). (arXiv Preprint, arXiv:1602.05339).
C. Bayindir. TWMS: Journal of Applied and Engineering Mathematics, [**[9-1]{}**]{}, (2019). (to be published) (arXiv Preprint, arXiv:1602.00816 (2016)).
C. Bayindir. MS Thesis, University of Delaware, (2009).
N. G. Vakhitov and A. A. Kolokolov. Radiophys. Quantum Electron., [**[16]{}**]{}, 783 (1973).
Y. Sivan, G. Fibich, B. Ilan and M. I. Weinstein. Phys. Rev. E, [**[78]{}**]{}, 046602 (2008).
M. I. Weinstein. SIAM J. Math. Anal., [**[16-3]{}**]{}, 472 (1985).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper formulates optimal control problems for rigid bodies in a geometric manner and it presents computational procedures based on this geometric formulation for numerically solving these optimal control problems. The dynamics of each rigid body is viewed as evolving on a configuration manifold that is a Lie group. Discrete-time dynamics of each rigid body are developed that evolve on the configuration manifold according to a discrete version of Hamilton’s principle so that the computations preserve geometric features of the dynamics and guarantee evolution on the configuration manifold; these discrete-time dynamics are referred to as Lie group variational integrators. Rigid body optimal control problems are formulated as discrete-time optimization problems for discrete Lagrangian/Hamiltonian dynamics, to which standard numerical optimization algorithms can be applied. This general approach is illustrated by presenting results for several different optimal control problems for a single rigid body and for multiple interacting rigid bodies. The computational advantages of the approach, that arise from correctly modeling the geometry, are discussed.'
author:
- Taeyoung Lee
- Melvin Leok
- 'N. Harris McClamroch'
bibliography:
- 'cis.bib'
- 'tylee.bib'
title: Computational Geometric Optimal Control of Rigid Bodies
---
Introduction
============
This paper utilizes methods from geometric mechanics and optimal control to develop new computational procedures for geometric optimal control of rigid bodies. The emphasis is on formulating a discrete-time optimal control problem that inherits important conservation properties of rigid body dynamics; this is achieved by combining variational integrators [@MaWe2001] and Lie group methods [@IsMuNoZa2000] to evolve the mechanical configuration. This approach leads to Lie group variational integrators that define the discrete-time rigid body dynamics which the optimal control computations are based upon [@CMA07; @CMDA07].
Most of the prior work related to optimal control of a rigid body is based on local coordinates on ${\ensuremath{\mathrm{SO}(3)}}$ or quaternions [@Bil.JoGCaD1993; @Bye.JoGCaD1993; @Scr.JoGCaD1994; @Sey.JoGCaD1993]. Minimal representations of the attitude of a rigid body, such as Euler angles, exhibit coordinate singularities, and require manipulating complicated trigonometric expressions. Nonminimal representations such as quaternions have no coordinate singularities, but they also introduce certain complications. In particular, the group of unit quaternions $\mathsf{SU}(2)\simeq{\ensuremath{\mathsf{S}}}^3$ double covers ${\ensuremath{\mathrm{SO}(3)}}$, so there is an ambiguity in representing an attitude of a rigid body. Furthermore, the Hamiltonian structure of rigid body attitude dynamics is unnecessarily complicated when it is expressed in terms of quaternions [@LeRe2004].
By considering rigid body translation and rotation as evolution on a Lie group, optimal control problems defined on Lie groups were introduced by Roger Brockett [@BrockSIJC72; @BrockSIAM73] and by John Baillieul [@BailJOTA78]. They emphasized the use of Lie group structures to characterize controllability and existence of optimal controls; they also obtained analytical results for the solution of certain types of optimal control problems. An optimal control problem for a generalized rigid body on $\mathrm{SO}(n)$ was considered in [@Blo.SaCL1996], and a general theory of optimal control problems on a Lie group was developed in [@Jur.1998a; @Jur.1998; @Jur.BK1997] together with reachability and controllability conditions. Although these papers viewed rigid body translation and rotation as motion on a Lie group, their results are limited to optimal control problems that can be formulated solely in terms of kinematics. In particular, they do not include dynamics in their analysis, and assume that the controls enter directly at the level of the Lie algebra.
The approach of computational geometric optimal control is focused on developing numerical algorithms, for optimal control problems, that preserve the geometric properties of the dynamics and the optimal control problem [@Lee.2008]. The essential idea is to apply geometric optimal control theory to discrete-time mechanical systems obtained using geometric numerical integrators. A discrete-time version of the generalized rigid body equations and their formulation as an optimal control problem are presented in [@Blo.ICoDaC1998; @Blo.N2002], and discrete-time optimal control problems for the dynamics of a rigid body are considered in [@Blo.JoDaCS2007; @Lee.JoDaCS2007; @Lee.PotICoDaC2006]. A direct optimal control approach is applied to discrete-time mechanical systems in [@Jun.IC2005], and it is referred to as *Discrete Mechanics and Optimal Control*.
This paper presents the approach of computational geometric optimal control for the dynamics of rigid bodies on a Lie group. We take the same geometric perspective as in the work of Roger Brockett [@BrockSIJC72; @BrockSIAM73], viewing evolution on a Lie group as fundamental. However, the emphasis in the present paper is on geometric formulations of both the kinematics and dynamics in the optimal control formulation and the role of geometric methods in optimal control computations.
The development in the paper makes clear that there are important advantages in formulating the optimal control problem as a discrete-time optimal control problem using Lie group variational integrators and then applying standard computational methods to solve the resulting discrete-time optimization problem. This is in contrast with approaches that construct continuous-time necessary conditions and then make use of computational methods to solve these necessary conditions. The paper demonstrates that for for the optimal control of rigid bodies, the proposed approach exhibits important advantages.
The main contributions of this paper can be summarized as follows: (i) the analytical and computational results presented in this paper are coordinate free; they avoid the singularities, ambiguity, and complications associated with local coordinates, and they provide a global insight into rigid body dynamics, (ii) a geometric optimal control problem is formulated for nontrivial rigid body dynamics that evolve on a Lie group, and (iii) a computational geometric optimal control approach is developed based on a geometric numerical integrator.
[Section \[sec:MFOCRB\]]{} provides a summary of Lie group variational integrators for rigid bodies that evolve on a Lie group. The resulting discrete-time rigid body dynamics are used as a basis for formulating a discrete-time optimal control problem. In [Section \[sec:optsrbdy\]]{} and \[sec:optmrbdy\], four different examples of rigid body optimal control problems are studied in some detail. First, optimal orbit and attitude maneuvers for a rigid dumbbell spacecraft in orbit about a large central body are studied. Then, optimal attitude maneuvers for a 3D pendulum acting under uniform gravity are studied; the control input conserves the component of the vertical component of the angular momentum thereby requiring a careful computational treatment that avoids numerical ill-conditioning. The third example is a 3D pendulum attached to a cart that can move in a horizontal plane; optimal reconfiguration maneuvers are studied for this cart and pendulum system. The fourth example involves optimal attitude maneuvers of two rigid bodies connected by a universal joint; the control input conserves angular momentum and the resulting controlled system exhibits a symmetry that has to be taken into account in the numerical approach in order to avoid numerical ill-conditioning.
Mathematical formulation for optimal control of rigid bodies {#sec:MFOCRB}
============================================================
The dynamics of rigid bodies exhibit important geometric features. The configuration of a rigid body can be described by the position vector of its center of mass in the Euclidean space ${\ensuremath{\mathbb{R}}}^3$ and by the attitude of the rigid body represented by a rotation matrix in the special orthogonal group ${\ensuremath{\mathrm{SO}(3)}}=\{ R\in {\ensuremath{\mathbb{R}}}^{3\times 3}\,|\, R^TR=I,\det{R}=1\}$. Thus, the general motion of a rigid body is described by the special Euclidean group ${\ensuremath{\mathrm{SE}(3)}}={\ensuremath{\mathrm{SO}(3)}}\textcircled{s}{\ensuremath{\mathbb{R}}}^3$. The configuration manifold for the class of multiple rigid bodies can be represented as a product involving ${\ensuremath{\mathbb{R}}}^3,{\ensuremath{\mathrm{SO}(3)}}$, and ${\ensuremath{\mathrm{SE}(3)}}$. Therefore, the configuration manifold of rigid bodies is a Lie group. Furthermore, the dynamics of rigid bodies, viewed as Lagrangian or Hamiltonian systems, are characterized by symplectic, momentum and energy preserving properties. These geometric features determine the qualitative behavior of the rigid body dynamics.
In this paper, we study optimal control problems for rigid bodies while carefully considering the geometric features of the dynamics in both the analysis and numerical computations. In particular, discrete-time dynamics of rigid bodies are developed that evolve on the configuration manifold according to a discrete version of Hamilton’s principle. The resulting geometric numerical integrator, referred to as a Lie group variational integrator, preserves geometric features of the dynamics and guarantees evolution on the configuration manifold. Based on the discrete-time rigid bodies dynamics, a discrete-time optimal control problem for rigid bodies is formulated. Standard numerical optimization algorithms can then be applied to solve this discrete-time optimal control problem.
Thus, our approach to discrete-time optimal control is characterized by discretizing the continuous-time optimal control problem at the problem formulation stage using Lie group variational integrators. This is in contrast to traditional techniques wherein discretization only arises at the last stage when numerically solving the continuous-time optimality conditions. Since the geometric properties of the dynamics of rigid bodies are preserved by using a Lie group variational integrator, this optimal control approach yields geometrically-exact optimal control inputs and accurate trajectories that are efficiently computed [@Blo.ICoDaC1998; @Jun.IC2005; @Lee.JoDaCS2007; @Lee.PotICoDaC2006].
In this section, we first describe the fundamental procedure to develop a Lie group variational integrator and its computational properties. Then, a discrete-time optimal control problem is formulated using the Lie group variational integrator, and computational approaches are presented to solve it numerically.
Lie group variational integrator {#subsec:lgvi}
--------------------------------
Geometric numerical integrators are numerical integration algorithms that preserve features of the continuous-time dynamics such as invariants, symplecticity, and the configuration manifold [@HaLuWa2006]. The geometrically exact properties of the discrete-time flow generate improved qualitative behavior. In this paper, we view a Lie group variational integrator as an intrinsically discrete-time dynamical system.
Numerical integration methods that preserve the simplecticity of a Hamiltonian system have been studied extensively [@San.AN92; @LeRe2004]. One traditional approach is to carefully choose the coefficients of a Runge-Kutta method to satisfy a simplecticity criterion and order conditions in order to obtain a symplectic Runge-Kutta method. However, it can be difficult to construct such integrators, and it is not guaranteed that other invariants of the system, such as momentum maps, are preserved. Alternatively, variational integrators are constructed by discretizing Hamilton’s principle, rather than discretizing the continuous Euler-Lagrange equation [@MosVes.CMP91; @MaWe2001]. The resulting integrators have the desirable property that they are symplectic and momentum preserving, and they exhibit good energy behavior for exponentially long times. Lie group methods are numerical integrators that preserve the Lie group structure of the configuration manifold [@IsMuNoZa2000]. Recently, these two approaches have been unified to obtain Lie group variational integrators that preserve the geometric properties of the dynamics as well as the Lie group structure of the configuration manifold without the use of local charts, reprojections, or constraints [@MaPeSh1999; @Leo.Phd04; @CMA07; @CMDA07].
(22.3,14)(0,-14) (0.0,-2.0)[(5.3,2.0)\[c\]]{} (2.65,-2.0)[(0,-1)[1.0000]{}]{} (0.0,-5.0)[(5.3,2.0)\[c\]]{} (2.65,-5.0)[(0,-1)[1.0000]{}]{} (2.65,-5.5)[(1,0)[5.5]{}]{}(8.15,-5.5)[(0,-1)[3.5]{}]{}(0.0,-8.0)[(5.3,2.0)\[c\]]{} (2.65,-8.0)[(0,-1)[1.0000]{}]{} (0.0,-11.0)[(5.3,2.0)\[c\]]{} (2.65,-11.0)[(0,-1)[1.0000]{}]{} (0.0,-14.0)[(5.3,2.0)\[c\]]{}(5.5,-11.0)[(5.3,2.0)\[c\]]{} (8.15,-11.0)[(0,-1)[1.0000]{}]{} (5.5,-14.0)[(5.3,2.0)\[c\]]{} (11.5,-2.0)[(5.3,2.0)\[c\]]{} (14.15,-2.0)[(0,-1)[1.0000]{}]{} (11.5,-5.0)[(5.3,2.0)\[c\]]{} (14.15,-5.0)[(0,-1)[1.0000]{}]{} (14.15,-5.5)[(1,0)[5.5]{}]{}(19.65,-5.5)[(0,-1)[3.5]{}]{}(11.5,-8.0)[(5.3,2.0)\[c\]]{} (14.15,-8.0)[(0,-1)[1.0000]{}]{} (11.5,-11.0)[(5.3,2.0)\[c\]]{} (14.15,-11.0)[(0,-1)[1.0000]{}]{} (11.5,-14.0)[(5.3,2.0)\[c\]]{}(17,-11.0)[(5.3,2.0)\[c\]]{} (19.65,-11.0)[(0,-1)[1.0000]{}]{} (17,-14.0)[(5.3,2.0)\[c\]]{}
We now summarize the derivation of a Lie group variational integrator. In Lagrangian mechanics, the equations of motion are derived by finding the path that extremizes the action integral, which is the integral of the Lagrangian over time. The Legendre transformation provides an alternative description that leads to Hamilton’s equations. Discrete-time Lagrangian and Hamiltonian mechanics, referred to as variational integrators, have been developed by reformulating Lagrangian and Hamiltonian mechanics in a discrete-time setting [@MaWe2001].
Discrete-time mechanics has a parallel structure with the mechanics described in continuous-time, as summarized in [Figure \[fig:el\]]{}. The phase variables of the continuous-time Lagrangian are replaced by two copies of the discrete-time configuration variables and a discrete-time Lagrangian that approximates a segment of the action integral is chosen. An action sum is defined using the discrete-time Lagrangian such that it approximates the action integral. This is the only approximation made in the development of discrete-time mechanics. Discrete-time Euler-Lagrange equations are obtained by setting the variation of the action sum to zero. The discrete-time Legendre transformation yields the equivalent of Hamilton’s equations. Lie group variational integrators are developed to preserve the structure of the Lie group configurations as well as the geometric properties of the continuous-time dynamics. The basic idea for all Lie group methods is to express the update map for group elements in the configuration manifold in terms of the group operation, so that the group structure is preserved automatically without need of parameterizations, constraints, or reprojections.
More explicitly, consider a mechanical system whose configuration manifold is a Lie group $G$ and is described by a Lagrangian $L:{\ensuremath{\mathsf{T}}}G\rightarrow {\ensuremath{\mathbb{R}}}$. The discrete update for the configuration is chosen as $$\begin{aligned}
g_{k+1} = g_k f_k,\label{eqn:g1}\end{aligned}$$ where $g_{k},g_{k+1}\in G$ are configuration variables, and the subscript $k$ denotes the value of a variable at the time $t=kh$ for a fixed timestep $h\in{\ensuremath{\mathbb{R}}}$. The discrete-time update map is represented by a right group action of $f_k \in G$ on $g_k$. Since the group element is updated by a group action, the group structure is preserved.
The expression for the flow map in discrete-time is obtained from the discrete variational principle on a Lie group, as presented in [Figure \[fig:el\]]{}. A discrete Lagrangian $L_d:G\times G\rightarrow {\ensuremath{\mathbb{R}}}$ approximates the integral of the Lagrangian over a time step along the solution of the Euler-Lagrange equation $$\begin{aligned}
L_d(g_k,f_k) & \approx \int_{kh}^{(k+1)h} L(g(t),\dot g(t)) \,dt,\end{aligned}$$ where a curve $g(t):[kh,(k+1)h]\rightarrow G$ satisfies the Euler–Lagrange equation in the time interval $[k,(k+1)h]$ with boundary conditions $g(kh)=g_k$ and $g((k+1)h)=g_k f_k=g_{k+1}$. Analogous to the action integral, the action sum is defined as $$\begin{aligned}
\mathfrak{G}_d = \sum_{k=0}^{N-1} L_d(g_k,f_k).\end{aligned}$$ The discrete Lagrange-d’Alembert principle, which is a modification of Hamilton’s principle to include the effect of control inputs, states that the sum of the variation of the action sum and the virtual work done by the control inputs is zero. But, the infinitesimal variation of a Lie group element must be carefully expressed to respect the structure of the Lie group. For example, it can be expressed in terms of the exponential map $\exp:\mathfrak{g}\rightarrow G$ as $$\begin{aligned}
\delta g = \frac{d}{d\epsilon}\bigg|_{\epsilon=0} g \exp {\epsilon\eta} = g\eta,\label{eqn:delg}\end{aligned}$$ for a Lie algebra element $\eta\in\mathfrak{g}$. From the discrete Lagrange-d’Alembert principle, we obtain $$\begin{aligned}
\delta \mathfrak{G}_d +\sum_{k=0}^{N-1}{\ensuremath{\left[ u^+_k \cdot \eta_{k+1} + u^-_k \cdot \eta_k \right]}}=0\end{aligned}$$ for any $\delta g_k$, and for given discrete Lagrangian forces $u^+_{d_k},u^-_{d_k}\in \mathfrak{g}^*$. This yields the generalized discrete Euler–Poincaré equation $$\begin{gathered}
{\ensuremath{\mathsf{T}}}_e^* {\ensuremath{\mathsf{L}}}_{f_{k}} \cdot {\ensuremath{\mathbf{D}}}_2 L_d (g_{k},f_{k}) - \mathrm{Ad}^*_{f_{k}}\cdot({\ensuremath{\mathsf{T}}}_e^* {\ensuremath{\mathsf{L}}}_{f_{k+1}}\cdot {\ensuremath{\mathbf{D}}}_2 L_d (g_{k+1},f_{k+1}))\\
+ {\ensuremath{\mathsf{T}}}_e^* {\ensuremath{\mathsf{L}}}_{g_{k+1}}\cdot {\ensuremath{\mathbf{D}}}_1 L_d (g_{k+1},f_{k+1})+u^+_{d_{k-1}}+u^-_{d_k}=0.\label{eqn:DEP}\end{gathered}$$ Here ${\ensuremath{\mathsf{L}}}_f:G\rightarrow G$ denotes the left translation map given by ${\ensuremath{\mathsf{L}}}_f g = fg$ for $f,g\in G$, ${\ensuremath{\mathsf{T}}}_g {\ensuremath{\mathsf{L}}}_f:{\ensuremath{\mathsf{T}}}_g G\rightarrow {\ensuremath{\mathsf{T}}}_{fg} G$ is the tangent map for the left translation and $\mathrm{Ad}_g:\mathfrak{g}\rightarrow\mathfrak{g}$ is the adjoint map. A dual map is denoted by a superscript ${}^*$ (see [@Lee.2008] for detailed definitions and developments).
This approach has been applied to the rotation group $\mathrm{SO}(3)$ and to the special Euclidean group $\mathrm{SE}(3)$ for dynamics of rigid bodies in [@CMA07; @CCA05; @CMDA07] and the generalization to abstract Lie groups are summarized here, thereby generating a unified geometric integrator for the class of multiple generalized rigid bodies whose configuration manifold can be expressed as a Lie group, which includes products involving ${\ensuremath{\mathbb{R}}}^3$, ${\ensuremath{\mathrm{SO}(3)}}$, and ${\ensuremath{\mathrm{SE}(3)}}$ as special cases.
Discrete-time optimal control
-----------------------------
Optimal control problems involve finding a control input such that a certain optimality objective is achieved under prescribed constraints. Here, the control inputs are parameterized by their values at each discrete time step, and the discrete-time equations of motion, including the control inputs, are obtained from [(\[eqn:DEP\])]{}. Any standard numerical algorithm for constrained optimization can be applied to this discrete-time system.
An indirect approach to solving a discrete-time optimal control problem is based on solving discrete-time necessary conditions for optimality. The resulting two-point boundary value problem can be solved by using standard numerical root finding techniques; one such approach is the shooting method that iterates on initial values of the multipliers. Alternatively, a direct approach formulates the discrete-time optimal control problem as a nonlinear programming problem, which is solved using standard numerical optimization algorithms such as a sequential quadratic programming algorithm; one such approach is the DMOC (Discrete Mechanics and Optimal Control) approach [@Jun.IC2005].
Explicit time-discretization prior to numerical optimization has significant computational advantages. As discussed in the previous section, the discrete-time dynamics are faithful representations of the continuous-time dynamics, and consequently more accurate solutions to the optimal control problems are typically obtained. The external control inputs may break the Lagrangian and Hamiltonian system structure; for example, the total energy may not be conserved for a controlled mechanical system. But, the computational superiority of the discrete mechanics formulation still holds for controlled systems. In particular, it has been demonstrated in [@MaWe2001] that the discrete-time dynamics derived from the discrete Lagrange-d’Alembert principle accurately computes the energy dissipation rate of controlled systems. For example, this feature is extremely important in accurately computing optimal trajectories for spacecraft orbit and attitude maneuvers for which the control authority is low and the maneuver time is large.
The proposed discrete-time optimal control formulation provides a framework for accurate computations. In most indirect optimal control approaches, the optimal solutions are sensitive to small variations in the initial values of the multipliers. This may cause difficulties, such as numerical ill-conditioning, in solving the necessary conditions for optimality expressed as a two-point boundary value problem. Numerically computed sensitivity derivatives, using Lie group variational integrators, do not exhibit numerical dissipation, which typically arises in conventional numerical integration schemes. Thus, the proposed approach leads to numerical robustness and efficient numerical computations. This indirect computational approach exhibits the quadratic convergence rate that is typical of Newton methods when it is applied to an optimal attitude control problem [@ACC07.opt]; the error in satisfaction of the optimality condition converges to machine precision superlinearly. For the direct optimal control approach, the optimal control inputs can be parametrized using fewer degrees of freedom, thereby reducing the computational overhead.
Several optimal control problems involving rigid bodies have been previously studied by the authors. Minimum-fuel and time-optimal control of spacecraft large-angle attitude maneuvers are studied in [@ACC06; @HusMelSan.CDC06; @Lee.JoDaCS2007; @ACC08]. The optimal orbit transfer of a dumbbell spacecraft, wherein the rotational attitude dynamics are non-trivially coupled to the translational dynamics, is studied in [@Lee.PotICoDaC2006]. An underactuated optimal control problem for the attitude maneuver of a 3D pendulum is studied in [@ACC07.opt]. An optimal formation reconfiguration of multiple rigid body spacecraft is studied in [@CDC07.opt]. An optimal control problem for a dynamic system evolving on an abstract Lie group is developed in [@Lee.2008], thereby generating a unified approach for optimal control problems of multiple rigid bodies.
In this paper, we summarize results for two optimal control problems for a single rigid body in Section \[sec:optsrbdy\] and results for two optimal control problems for multiple rigid bodies in Section \[sec:optmrbdy\]. Each of these optimal control problems treats complex dynamics of a single or multiple rigid bodies, demonstrating the value of the proposed geometric optimal control approach.
Optimal control problems for a single rigid body {#sec:optsrbdy}
================================================
Optimal maneuver of a dumbbell spacecraft on ${\ensuremath{\mathrm{SE}(3)}}$ {#subsec:optse}
----------------------------------------------------------------------------
We develop an optimal 3D translational and rotational maneuver of a rigid dumbbell spacecraft in orbit about a large central body. The dumbbell spacecraft is composed of two spheres connected by a massless rod. An interesting feature of the dumbbell spacecraft is that there is coupling between its translational dynamics and its rotational dynamics due to the presence of both gravity forces and gravity moments that act on the dumbbell spacecraft.
The configuration manifold is the special Euclidean group ${\ensuremath{\mathrm{SE}(3)}}={\ensuremath{\mathrm{SO}(3)}}\,\textcircled{s}\,{\ensuremath{\mathbb{R}}}^3$. For $(R,x)\in{\ensuremath{\mathrm{SE}(3)}}$, the linear transformation from the body-fixed frame to the inertial frame is denoted by the rotation matrix $R\in{\ensuremath{\mathrm{SO}(3)}}$, and the position of the mass center in the inertial frame is denoted by a vector $x\in{\ensuremath{\mathbb{R}}}^3$. The vectors $\Omega,v\in{\ensuremath{\mathbb{R}}}^3$ are the angular velocity in the body-fixed frame, and the translational velocity in the inertial frame, respectively. Let $m\in{\ensuremath{\mathbb{R}}}$ and $J\in{\ensuremath{\mathbb{R}}}^{3\times 3}$ be the mass and the moment of inertia matrix of a rigid body. We assume that external control force $u^f\in{\ensuremath{\mathbb{R}}}^3$ and control moment $u^m\in{\ensuremath{\mathbb{R}}}^3$ act on the dumbbell spacecraft. Control inputs are parameterized by their values at each time step.
Define a $f_k=(F_k,Y_k)\in{\ensuremath{\mathrm{SE}(3)}}$ such that $g_{k+1}=(R_{k+1},x_{k+1})$ is equal to $g_kf_k$, i.e. $(R_{k+1},x_{k+1})=(R_k,x_k)\circ(F_k,Y_k)= (R_kF_k, x_k+R_kY_k)$. The rotation matrix $F_k$ represent the relative update of the attitude between integration steps. The gravitational potential is denoted by $U:{\ensuremath{\mathrm{SE}(3)}}\rightarrow{\ensuremath{\mathbb{R}}}$. We choose the following discrete Lagrangian $$\begin{aligned}
L_d(R_k,x_k,F_k,Y_k) = \frac{1}{2h}mY_k^TY_k+\frac{1}{h}{\mbox{tr}\ensuremath{\negthickspace{\ensuremath{\left[ (I-F_k)J_d \right]}}}}-hU(R_{k}F_k,x_{k}+R_kY_k),\end{aligned}$$ where $J_d\in{\ensuremath{\mathbb{R}}}^{3\times 3}$ is a non-standard moment of inertia matrix defined as $J_d=\frac{1}{2}\mathrm{tr}[J]I_{3\times 3}-J$. Substituting this discrete Lagrangian into [(\[eqn:DEP\])]{}, we obtain the following discrete equations of motion (see [@Lee.2008] for detailed development). $$\begin{gathered}
h \widehat{J\Omega_{k}}=F_{k}J_{d}-J_{d}F_{k}^T,\label{eqn:findF}\\
R_{k+1}=R_{k}F_{k},\label{eqn:Rkp}\\
x_{k+1} = x_{k} + hv_k,\label{eqn:xkp}\\
J\Omega_{{k+1}}=F_{k}^T J\Omega_{k}+h (M_{{k+1}}+u^m_{k+1})
\label{eqn:Omegakp},\\
m v_{k+1}=m v_k-h{\ensuremath{\frac{\partial U_{k+1}}{\partial x_{k+1}}}}+hu^f_{k+1},\label{eqn:vkp}\end{gathered}$$ where the hat map $\hat\cdot$ is an isomorphism from ${\ensuremath{\mathbb{R}}}^3$ to $3\times 3$ skew-symmetric matrices ${\ensuremath{\mathfrak{so}(3)}}$, defined such that $\hat x y =x\times y$ for any $x,y\in{\ensuremath{\mathbb{R}}}^3$. The moment $M\in{\ensuremath{\mathbb{R}}}^3$ due to the potential is given by, $$\begin{aligned}
\hat M & = {\ensuremath{\frac{\partial U}{\partial R}}}^TR-R^T{\ensuremath{\frac{\partial U}{\partial R}}},\end{aligned}$$ where the matrix ${\ensuremath{\frac{\partial U}{\partial R}}}\in{\ensuremath{\mathbb{R}}}^{3\times 3}$ is defined by $[{\ensuremath{\frac{\partial U}{\partial R}}}]_{ij}={\ensuremath{\frac{\partial U}{\partial [R]_{ij}}}}$ for $i,j\in{\ensuremath{\left\{ 1,2,3 \right\}}}$, and the $i,j$-th element of a matrix is denoted by $[\cdot]_{ij}$.
For a given $(R_k,x_k,\Omega_k,v_k)$, we solve the implicit equation [(\[eqn:findF\])]{} to find $F_k\in{\ensuremath{\mathrm{SO}(3)}}$. Then, the configuration at the next step $(R_{k+1},x_{k+1})$ is obtained from [(\[eqn:Rkp\])]{} and [(\[eqn:xkp\])]{}. Using the computed moment $M_{k+1}$ and force $-{\ensuremath{\frac{\partial U_{k+1}}{\partial x_{k+1}}}}$, velocities $\Omega_{k+1},v_{k+1}$ are obtained from [(\[eqn:Omegakp\])]{} and [(\[eqn:vkp\])]{}. This defines a discrete flow map, $(R_k,x_k,\Omega_k,v_k)\mapsto(R_{k+1},x_{k+1},\Omega_{k+1},v_{k+1})$, and this process can be repeated.
Since this Lie group variational integrator is obtained by discretizing Hamilton’s principle, it is symplectic and preserves the momentum map associated with the symmetry of the Lagrangian. In the absence of external forces and moments, the total energy oscillates around its initial value with small bounds on a comparatively short timescale, but there is no tendency for the mean of the oscillation in the total energy to drift (increase or decrease) from the initial value for exponentially long times.
The discrete flow map also preserves the group structure. By using the given computational approach, the matrix $F_k$, representing the change in relative attitude change over a time step, is guaranteed to be a rotation matrix. The rotation matrix $R_{k+1}$ is obtained by the group operation in [(\[eqn:Rkp\])]{}, so that it evolves on ${\ensuremath{\mathrm{SO}(3)}}$. Therefore, the orthogonal structure of the rotation matrices is preserved, and the attitude of each rigid body is determined accurately and globally.
This geometrically exact numerical integration method yields a highly efficient computational algorithm. The self-adjoint discrete Lagrangian used to derive this Lie group variational integrator guarantees that this integrator has second-order accuracy, while requiring only one function evaluation per integration step. Higher-order methods can be easily constructed using a composition method [@HaLuWa2006].
An implicit equation [(\[eqn:findF\])]{} must be solved at each time step to determine the attitude update. However the computational effort to solve each implicit equation is negligible; the relative attitude update is expressed at the Lie algebra level isomorphic to ${\ensuremath{\mathbb{R}}}^3$, and the corresponding Newton iteration converges to machine precision within two or three iterations. This method could be considered *almost explicit* when the computational cost is compared with explicit integrators with the same order of accuracy [@CMA07].
### Optimal control problem {#optimal-control-problem .unnumbered}
The objective is to transfer the spacecraft from a given initial condition $(R_0,x_0,\Omega_0,v_0)$ to a desired terminal condition $(R^f,x^f,\Omega^f,v^f)$ during a fixed maneuver time $Nh$, while minimizing the square of the $l_2$ norm of the control inputs. $$\begin{gathered}
\min_{u_{k+1}} {\ensuremath{\left\{ \mathcal{J}=\sum_{k=0}^{N-1}
\frac{h}{2}(u^f_{k+1})^TW_fu^f_{k+1}+
\frac{h}{2}(u^m_{k+1})^TW_mu^m_{k+1} \right\}}},\end{gathered}$$ where $W_f,W_m\in{\ensuremath{\mathbb{R}}}^{3\times 3}$ are symmetric positive-definite matrices.
### Necessary conditions for optimality {#necessary-conditions-for-optimality .unnumbered}
An indirect optimization method is used to determine the optimal solution, based on necessary conditions for optimality derived using variational arguments; the optimal control is characterized as a solution of a two-point boundary value problem. The augmented cost function to be minimized is $$\begin{aligned}
\mathcal{J}_a =
\sum_{k=0}^{N-1}&\frac{h}{2}(u^f_{k+1})^TW^fu^f_{k+1}+
\frac{h}{2}(u^m_{k+1})^TW^mu^m_{k+1}\nonumber\\
& +\lambda_k^{1,T}{\ensuremath{\left\{ -x_{k+1}+x_k+h v_k \right\}}} +\lambda_k^{2,T}{\ensuremath{\left\{ -m v_{k+1} + mv_k-h{\ensuremath{\frac{\partial U_{k+1}}{\partial x_{k+1}}}}+hu^f_{k+1} \right\}}}\nonumber\\
& +\lambda_k^{3,T}{\ensuremath{\left( \mathrm{logm}(F_k-R_{k}^TR_{k+1}) \right)}}^{\vee} +\lambda_k^{4,T}{\ensuremath{\left\{ -J\Omega_{k+1} + F_k^T J\Omega_k +
h{\ensuremath{\left( M_{k+1}+u_{k+1}^m \right)}} \right\}}},\end{aligned}$$ where $\lambda_k^1,\lambda_k^2,\lambda_k^3,\lambda_k^4\in {\ensuremath{\mathbb{R}}}^3$ are Lagrange multipliers. The matrix logarithm is denoted by $\mathrm{logm}:{\ensuremath{\mathrm{SO}(3)}}\rightarrow{\ensuremath{\mathfrak{so}(3)}}$ and the vee map $\vee:{\ensuremath{\mathfrak{so}(3)}}\rightarrow{\ensuremath{\mathbb{R}}}^3$ is the inverse of the hat map. The logarithmic form of [(\[eqn:Rkp\])]{} is used, and the constraint [(\[eqn:findF\])]{} is implicitly imposed using constrained variations. Using similar expressions for the variations given in [(\[eqn:delg\])]{}, the infinitesimal variation of the cost can be written as $$\begin{aligned}
\delta\mathcal{J}_a & = \sum_{k=1}^{N-1} h\delta
u_{k}^{f,T}{\ensuremath{\left\{ W_fu^f_{k}+\lambda_{k-1}^2 \right\}}}
+h\delta u_{k}^{m,T}{\ensuremath{\left\{ W_mu^m_{k}+\lambda_{k-1}^4 \right\}}}
+z_k^T{\ensuremath{\left\{ -\lambda_{k-1}+A_k^T \lambda_k \right\}}},\end{aligned}$$ where $\lambda_k=[\lambda_k^1;\lambda_k^2;\lambda_k^3;\lambda_k^4]\in{\ensuremath{\mathbb{R}}}^{12}$ is the vector of Lagrange multipliers, and $z_k\in{\ensuremath{\mathbb{R}}}^{12}$ represents the infinitesimal variation of $(R_k,x_k,\Omega_k,v_k)$, given by $z_k=[\mathrm{logm}(R_k^T \delta R_k)^\vee;\delta x_k,\delta\Omega_k,\delta v_k]$. The matrix $A_k\in{\ensuremath{\mathbb{R}}}^{12\times 12}$ is expressed in terms of $(R_k,x_k,\Omega_k,v_k)$ [@Lee.PotICoDaC2006]. Thus, necessary conditions for optimality are given by $$\begin{aligned}
u^f_{k+1} &= -W_{f}^{-1}\lambda_{k}^2,\label{eqn:ufkp}\\
u^m_{k+1} &= -W_{m}^{-1}\lambda_{k}^4,\label{eqn:umkp}\\
\lambda_{k} &= A_{k+1}^T \lambda_{k+1}\label{eqn:updatelam}\end{aligned}$$ together with the discrete equations of motion and the boundary conditions.
### Computational approach {#computational-approach .unnumbered}
Necessary conditions for optimality are expressed in terms of a two-point boundary problem. This problem is to find the optimal discrete flow, multipliers, and control inputs that simultaneously satisfies the equations of motion, optimality conditions, multiplier equations, and boundary conditions. We use a neighboring extremal method [@Bry.BK75], and choose a nominal solution satisfying all of the necessary conditions except the boundary conditions. The unspecified initial multiplier is updated by successive linearization so as to satisfy the specified terminal boundary conditions in the limit. This is also referred to as a shooting method. The main advantage of the neighboring extremal method is that the number of iteration variables is small.
The difficulty is that the extremal solutions are sensitive to small changes in the unspecified initial multiplier values. The nonlinearities also make it hard to construct an accurate estimate of sensitivity, thereby resulting in numerical ill-conditioning. Therefore, it is important to compute the sensitivities accurately in the neighboring extremal method. Here, the optimality conditions [(\[eqn:ufkp\])]{} and [(\[eqn:umkp\])]{} are substituted into the equations of motion and the multiplier equations, which are linearized to obtain $$\begin{aligned}
\begin{bmatrix} z_N \\ \delta \lambda_N \end{bmatrix}
=\begin{bmatrix} \Psi^{11} & \Psi^{12} \\ \Psi^{21} & \Psi^{22}\end{bmatrix}
\begin{bmatrix} z_0 \\ \delta \lambda_0 \end{bmatrix},\end{aligned}$$ where $\Psi^{ij}\in{\ensuremath{\mathbb{R}}}^{6\times 6}$ for $i,j\in\{1,2\}$ represents a computable linear operator. For the given two-point boundary value problem, $z_0=0$ since the initial condition is fixed. The terminal multipliers are free. Thus, we obtain $$\begin{aligned}
z_N = \Psi^{12} \delta\lambda_0.\end{aligned}$$ The linear operator $\Psi^{12}$ represents the sensitivity of the specified terminal boundary conditions with respect to the unspecified initial multiplier. Using this sensitivity, a guess of the unspecified initial multipliers is iterated to satisfy the specified terminal conditions in the limit. Any type of Newton iteration can be applied. We use a line search with backtracking algorithm, referred to as the Newton-Armijo iteration [@Kel.BK95].
### Numerical example {#numerical-example .unnumbered}
We study a maneuver of a rigid spacecraft under a central gravity field. We assume that the mass of the spacecraft is negligible compared to the mass of a central body, and we consider a fixed frame attached to the central body as an inertial frame. The resulting model is a Restricted Full Two Body Problem (RF2BP) [@Sch.CMDA02].
The spacecraft is modeled as a dumbbell, which consists of two equally massive spheres and a massless rod. The gravitational potential is given by $$\begin{aligned}
U(R,x)=-\frac{GMm}{2} \sum_{q=1}^2 \frac{1}{{\ensuremath{\left\| x+R\rho^q \right\|}}},\end{aligned}$$ where $G\in{\ensuremath{\mathbb{R}}}$ is the gravitational constant, $M,m\in{\ensuremath{\mathbb{R}}}$ are the mass of the central body, and the mass of the dumbbell, respectively. The vector $\rho^q\in{\ensuremath{\mathbb{R}}}^3$ is the position of the $q$th sphere from the mass center of the dumbbell expressed in the body fixed frame ($q\in {\ensuremath{\left\{ 1,2 \right\}}}$). The mass, length, and time dimensions are normalized by the mass of the dumbbell, the radius of a reference circular orbit, and its orbital period.
Initially, the spacecraft is on a circular orbit. The desired maneuver is to increase the orbital inclination by $60^\circ$. We explicitly consider the coupling effect between the orbital motion and the rotational attitude maneuver of the spacecraft. The maneuver time is chosen to be a quarter of the orbital period of the initial circular orbit. The boundary conditions are as follows, $$\begin{aligned}
{2}
x_0&=[1,0,0],&\quad x^f&=[-0.3536,0.3536,0.8660],\\
R_{0}&=\begin{bmatrix}
0& -1& 0\\
1& 0& 0\\
0& 0& 1\end{bmatrix},&
\quad R^f&=\begin{bmatrix}
-0.7071& 0.3535& 0.6123\\
-0.7071& -0.3535& -0.6123\\
0& -0.8660& 0.5\end{bmatrix},\\
\dot x_0&=[0,0.9835,0],&\quad \dot x^f&=[-0.6954,-0.6954,0],\\
\Omega_0&=[0,0,0.9835],&\quad\Omega^f&=[0,0,0.9835].\end{aligned}$$
[Figure \[fig:OptRB\]]{} illustrates the optimal spacecraft maneuver, convergence rate, and optimal control inputs. The optimal cost and the violation of the terminal boundary conditions are $13.03$, and $9.32\times 10^{-15}$ respectively. [Figure \[fig:OptRB\_err\]]{} shows the violation of the terminal boundary conditions versus the number of iterations on a semi-logarithmic scale. Red circles denote outer iterations of the Newton-Armijo iteration where the sensitivity derivatives are computed, and inner iterations correspond to backtracking in the line search routine. The initial guess of the unspecified initial multipliers is arbitrarily chosen. The error in satisfaction of the terminal boundary condition converges quickly to machine precision after the 20th iteration. These convergence results are consistent with the quadratic convergence rates expected of Newton methods with accurately computed gradients.
The shooting method may be prone to numerical ill-conditioning, as a small change in the initial multiplier can cause highly nonlinear behavior of the terminal conditions. However, as shown in [Figure \[fig:OptRB\_err\]]{}, the computational geometric optimal control approach exhibits excellent numerical convergence properties. This is because the proposed computational algorithms are geometrically exact and numerically accurate. There is no numerical dissipation introduced by the numerical algorithm, and therefore, the sensitivity derivatives are more accurately computed.
Optimal attitude reorientation of an underactuated 3D pendulum on ${\ensuremath{\mathrm{SO}(3)}}$ [@ACC07.opt] {#subsec:3dpend}
--------------------------------------------------------------------------------------------------------------
A 3D pendulum is a rigid body supported by a fixed frictionless pivot acting under the influence of a uniform gravitational field [@SheSanNal.CDC04]. The rigid body has three rotational degrees of freedom, and the configuration manifold is ${\ensuremath{\mathrm{SO}(3)}}$. The linear transformation from the body fixed frame and the inertial frame is denoted by $R\in{\ensuremath{\mathrm{SO}(3)}}$, and the angular velocity represented in the body fixed frame is denoted by $\Omega\in{\ensuremath{\mathbb{R}}}^3$. Let $e_3\in{\ensuremath{\mathbb{R}}}^3$ be the gravity direction in the inertial frame, and $J\in{\ensuremath{\mathbb{R}}}^{3\times 3}$ be the moment of inertia matrix of the rigid body with respect to the pivot point. The vector from the pivot point to the mass center, represented in the body fixed frame is given by $\rho\in{\ensuremath{\mathbb{R}}}^3$.
The Lagrangian of the 3D pendulum is invariant under a rotation about the gravity direction, and therefore the 3D pendulum has a $S^1$ symmetry action. Consequently, the angular momentum about the gravity direction, represented by $e_3^T RJ\Omega$, is preserved.
We study an optimal attitude control of the 3D pendulum with symmetry. An external control moment is chosen such that it does not have any component about the gravity direction. The structure of the control moment is chosen as $R^T e_3 \times u$ for a control parameter $u\in{\ensuremath{\mathbb{R}}}^3$. Thus, the angular momentum about the gravity direction is conserved along the controlled dynamics of the 3D pendulum. Such control inputs are physically realized by actuation mechanisms, such as point mass actuators, that change the center of mass of the 3D pendulum.
The discrete Lagrangian of the 3D pendulum is chosen to be $$\begin{aligned}
L_d(R_k,F_k) = \frac{1}{h} {\mbox{tr}\ensuremath{\negthickspace{\ensuremath{\left[ (I-F_k)J_d \right]}}}} + h mg e_3^T R\rho.\end{aligned}$$ The resulting Lie group variational integrator, including an external control input, is given by $$\begin{gathered}
h \widehat{J\Omega_{k}}=F_{k}J_{d}-J_{d}F_{k}^T,\label{eqn:findF2}\\
R_{k+1}=R_{k}F_{k},\label{eqn:Rkp2}\\
J\Omega_{{k+1}}=F_{k}^T J\Omega_{k}+hM_{{k+1}}+hR^T e_3 \times u_{k+1}.\label{eqn:Omegakp2}\end{gathered}$$
### Optimal control problem {#optimal-control-problem-1 .unnumbered}
The objective of the optimal control problem is to transfer the 3D pendulum from a given initial condition $(R_0,\Omega_0)$ to a desired terminal condition $(R^f,\Omega^f)$ during a fixed maneuver time $Nh$, while minimizing the square of the $l_2$ norm of the control inputs. $$\begin{gathered}
\min_{u_{k+1}} {\ensuremath{\left\{ \mathcal{J}=\sum_{k=0}^{N-1}
\frac{h}{2}(u_{k+1})^TWu_{k+1} \right\}}},\end{gathered}$$ where $W\in{\ensuremath{\mathbb{R}}}^{3\times 3}$ is a symmetric positive-definite matrix. In particular, we choose attitude maneuvers that can be described by rest-to-rest rotations about the unactuated gravity direction. The resulting optimal attitude maneuver exhibits the geometric phase effect [@MarRat.BK99], which in the zero group momentum case directly relates the group motion to the curvature enclosed by the trajectory in shape space.
### Necessary conditions for optimality {#necessary-conditions-for-optimality-1 .unnumbered}
We solve this optimal control problem by using an indirect optimization method, where necessary conditions for optimality are derived using variational arguments, and a solution of the corresponding two-point boundary value problem provides the optimal control. The augmented cost function to be minimized is $$\begin{aligned}
\mathcal{J}_a =
\sum_{k=0}^{N-1}& \frac{h}{2}u_{k+1}^TWu_{k+1} +\lambda_k^{1,T}{\ensuremath{\left( \mathrm{logm}(F_k-R_{k}^TR_{k+1}) \right)}}^{\vee}\nonumber \\ & +\lambda_k^{2,T}{\ensuremath{\left\{ -J\Omega_{k+1} + F_k^T J\Omega_k +
hM_{k+1}+hR_{k+1}^Te_3\times u_{k+1} \right\}}},\end{aligned}$$ where $\lambda_k^1,\lambda_k^2\in {\ensuremath{\mathbb{R}}}^3$ are Lagrange multipliers. The infinitesimal variation can be written as $$\begin{aligned}
\delta\mathcal{J}_a & = \sum_{k=1}^{N-1} h\delta u_{k}^{T}{\ensuremath{\left\{ Wu_{k}-R_k^Te_3\times \lambda_{k-1}^2 \right\}}}
+z_k^T{\ensuremath{\left\{ -\lambda_{k-1}+A_k^T \lambda_k \right\}}},\end{aligned}$$ where $\lambda_k=[\lambda_k^1;\lambda_k^2]\in{\ensuremath{\mathbb{R}}}^{6}$, and $z_k\in{\ensuremath{\mathbb{R}}}^{6}$ represents the infinitesimal variation of $(R_k,\Omega_k)$, given by $z_k=[\mathrm{logm}(R_k^T \delta R_k)^\vee;\delta\Omega_k]$. The matrix $A_k\in{\ensuremath{\mathbb{R}}}^{6\times 6}$ can be expressed in terms of $(R_k,\Omega_k),\lambda_k$. Thus, necessary conditions for optimality are given by $$\begin{aligned}
u_{k+1} &= W^{-1}(R_{k+1}^Te_3\times \lambda_{k}^2),\\
\lambda_{k} &= A_{k+1}^T \lambda_{k+1}\end{aligned}$$ together with the discrete equations of motion and the boundary conditions.
### Computational approach {#computational-approach-1 .unnumbered}
We apply the neighboring extremal method described in Section \[subsec:optse\]; the optimality condition is substituted into the equations of motion and the multiplier equation, and sensitivity derivatives of the optimal solution with respect to the initial multiplier are obtained, and the initial multiplier is iterated to satisfy the terminal boundary condition.
However, the underactuated control input, that respects the symmetry of the 3D pendulum, causes a fundamental singularity in the sensitivity derivatives, since the controlled system inherits the $S^1$ symmetry, and the cost functional is invariant under the lifted action of $S^1$. Consequently, the sensitivity derivatives vanish in the group direction. At each iteration, we need to compute inverse of a matrix of sensitivity derivatives to update the initial multiplier. However, the sensitivity matrix has a theoretical rank deficiency of one since the vertical component of the inertial angular momentum is conserved regardless of the initial multiplier variation. Therefore, this matrix inversion is numerically ill-conditioned.
We present a simple numerical scheme to avoid the numerical ill-conditioning caused by this symmetry. At each step, we decompose the matrix of sensitivity derivatives into a symmetric part and an anti-symmetric part. The symmetric part describes the sensitivity of the conserved angular momentum component due to the symmetry, and therefore it is zero and does not depend on the initial multiplier values. An update for the initial multipliers is determined using the matrix inverse of the anti-symmetric part; this matrix inverse is not ill-conditioned. This approach removes the singularity in the sensitivity derivatives completely, and the resulting optimal control problem is no longer ill-conditioned.
### Numerical example {#numerical-example-1 .unnumbered}
Properties of the 3D pendulum are chosen as, $$m=1\,\mathrm{kg},\quad J=\mathrm{diag}[0.13,0.28,0.17]\,\mathrm{kgm^2},\quad \text{and}\quad\rho=[0,0,0.3]\,\mathrm{m}.$$ The desired maneuver is a $180^\circ$ rotation about the vertical axis from a hanging equilibrium to another hanging equilibrium. The corresponding boundary conditions are given by $$\begin{aligned}
{2}
R_0&=I,&\quad R^f&=\mathrm{diag}[-1,-1,1],\\
\Omega_0&=[0,0,0],&\quad \Omega^f&=[0,0,0].\end{aligned}$$ The maneuver time is $1$ second, and the time step is $h=0.001$. Since the vertical component of the angular momentum is zero, the rotation is a consequence of the geometric phase effect [@MarRat.BK99]. This problem is challenging in the sense that the desired maneuvers are rotations about the gravity direction, but the control input does not directly generate any moment about the gravity direction.
[Figure \[fig:Opt3DP\]]{} illustrates the optimal pendulum maneuver, convergence rate, and optimal control inputs. The optimal cost and the violation of the terminal boundary conditions are $7.32$, and $4.80\times 10^{-15}$ respectively. As shown in [Figure \[fig:Opt3DP\_err\]]{}, the error in satisfaction of the terminal boundary condition converges to machine precision after the 50th iteration. The condition number of the decomposed sensitivity derivative varies from $10^0$ to $10^5$. If the sensitivity derivative is not decomposed, then the condition numbers are at the level of $10^{19}$, and the numerical iterations fail to converge. This numerical example demonstrates the excellent numerical convergence properties of the computational geometric optimal control approach that is achieved by incorporating a modification that eliminates the numerical ill-conditioning introduced by the symmetry.
Optimal control problems for multiple rigid bodies {#sec:optmrbdy}
==================================================
Optimal maneuver of a 3D pendulum on a 2D cart on ${\ensuremath{\mathrm{SO}(3)}}\times{\ensuremath{\mathbb{R}}}^2$
------------------------------------------------------------------------------------------------------------------
Consider a 3D pendulum whose pivot is attached to a cart that can translate on a horizontal plane. This is a generalization of the popular planar pendulum on a cart model (see, for example, [@Blo.PotICoDaC2005]), where the pendulum has three rotational degrees of freedom, and the cart moves on a two dimensional horizontal plane.
We define two frames; an inertial frame and a body fixed frame for the 3D pendulum whose origin is located at the moving pivot point. Define
---------------------------------------- ----------------------------------------------------------------------------------------------
$x\in{\ensuremath{\mathbb{R}}}$ Displacement of the cart along the $e_1$ direction in the reference frame
$y\in{\ensuremath{\mathbb{R}}}$ Displacement of the cart along the $e_2$ direction in the reference frame
$R\in{\ensuremath{\mathrm{SO}(3)}}$ Rotation matrix from the body fixed frame to the reference frame
$\Omega\in{\ensuremath{\mathbb{R}}}^3$ Angular velocity of the pendulum represented in the body fixed frame
$d\in{\ensuremath{\mathbb{R}}}^3$ Vector from the pivot to the mass center of the pendulum represented in the body fixed frame
$m\in{\ensuremath{\mathbb{R}}}$ Mass of the pendulum
$M\in{\ensuremath{\mathbb{R}}}$ Mass of the cart
---------------------------------------- ----------------------------------------------------------------------------------------------
The configuration manifold is ${\ensuremath{\mathrm{SO}(3)}}\times{\ensuremath{\mathbb{R}}}^2$. We assume that external control forces $u_x,u_y\in{\ensuremath{\mathbb{R}}}$ are applied to the cart.
The Lagrangian of the 3D pendulum on a cart is invariant under a rotation about the gravity direction. Therefore, it has a symmetry of ${\ensuremath{\mathsf{S}}}^1$ action, and the total angular momentum about the gravity direction is preserved. The external control forces acting on the cart break this symmetry, and the controlled system is not symmetric. In particular, the total angular momentum is not preserved in the controlled dynamics. Therefore, this optimal control problem should be distinguished from the optimal control of a 3D pendulum with symmetry, discussed in Section \[subsec:3dpend\], where a symmetry-preserving control input is chosen.
The discrete Lagrangian for the 3D pendulum on a 2D cart is $$\begin{aligned}
L_d &(R_k,x_k,y_k,R_{k+1},x_{k+1},y_{k+1})=\frac{1}{2h}(M+m) ((x_{k+1}-x_k)^2 + (y_{k+1}-y_k)^2)\nonumber\\ &+\frac{1}{h}{\mbox{tr}\ensuremath{\negthickspace{\ensuremath{\left[ (I-F_k)J_d \right]}}}} +\frac{m}{h}(x_{k+1}-x_k) e_1^T (R_{k+1}-R_k) d + \frac{m}{h}(y_{k+1}-y_k) e_2^T (R_{k+1}-R_k) d\nonumber\\ &+\frac{h}{2}mg e_3^T R_kd+\frac{h}{2}mg e_3^T R_{k+1}d.\label{eqn:Ld}\end{aligned}$$ From [(\[eqn:DEP\])]{}, the Lie group variational integrator for the 3D pendulum on a cart is given by the discrete-time equations [$$\begin{aligned}
p_{x_k} & = \frac{1}{h}(M+m) (x_{k+1}-x_k) +\frac{m}{h}e_1 (R_{k+1}-R_k) d,\label{eqn:pxk}\\
p_{y_k} & = \frac{1}{h}(M+m) (y_{k+1}-y_k) +\frac{m}{h}e_2 (R_{k+1}-R_k) d,\label{eqn:pyk}\\
\hat p_{\Omega_k} & = \frac{1}{h} (F_kJ_d -J_d F_k^T) + {\ensuremath{\left\{ \frac{m}{h}(x_{k+1}-x_k) \hat d R_k^T e_1+\frac{m}{h}(y_{k+1}-y_k) \hat d R_k^T e_2 - \frac{h}{2} m g \hat d R_k^T e_3 \right\}}}^\wedge,\label{eqn:pwk}\\
R_{k+1} & =R_kF_k,\label{eqn:Rkp3}\\
p_{x_{k+1}} & = p_{x_k}+hu_{x_{k+1}},\label{eqn:pxkp}\\
p_{y_{k+1}} & = p_{y_k}+hu_{y_{k+1}},\label{eqn:pykp}\\
\hat p_{\Omega_{k+1}} & = \frac{1}{h} (J_dF_k -F_k^T J_d ) + {\ensuremath{\left\{ \frac{m}{h}(x_{k+1}-x_k) \hat d R_{k+1}^T e_1 +\frac{m}{h}(y_{k+1}-y_k) \hat d R_{k+1}^T e_2 + \frac{h}{2} m g \hat d R_{k+1}^T e_3 \right\}}}^\wedge.\label{eqn:pwkp}\end{aligned}$$]{} The momenta variables $p_\Omega\in{\ensuremath{\mathbb{R}}}^3$, $p_x,p_y\in{\ensuremath{\mathbb{R}}}$ are given by $$\begin{aligned}
\begin{bmatrix}
p_\Omega \\ p_x \\ p_y
\end{bmatrix}
=
\begin{bmatrix}
J & m\hat d R^T e_1 & m\hat d R^T e_2\\
-me_1^T R \hat d & M+m & 0\\
-me_2^T R \hat d & 0 & M+m
\end{bmatrix}
\begin{bmatrix}
\Omega \\ \dot x \\ \dot y
\end{bmatrix}
.\label{eqn:p}\end{aligned}$$ The detailed derivation of this Lie group variational integrator is available in [@Lee.2008]. For given $(R_k,x_k,y_k,\Omega_k,\dot x_k,\dot y_k)$, we compute $(p_{\Omega_k},p_{x_k},p_{y_k})$ by [(\[eqn:p\])]{}. We use a fixed-point iteration to compute $R_{k+1}$. For an initial guess for $R_{k+1}$, the corresponding $x_{k+1},y_{k+1}$ are obtained by using [(\[eqn:pxk\])]{},[(\[eqn:pyk\])]{}. Then, we can find $F_k$ by solving [(\[eqn:pwk\])]{}. The updated value for $R_{k+1}$ is given by [(\[eqn:Rkp3\])]{}. This is repeated until $R_{k+1}$ converges. Then, $x_{k+1},y_{k+1}$ are obtained from [(\[eqn:pxk\])]{},[(\[eqn:pyk\])]{}, and $(p_{\Omega_{k+1}},p_{x_{k+1}},p_{y_{k+1}})$ are obtained by [(\[eqn:pxkp\])]{},[(\[eqn:pykp\])]{}, and [(\[eqn:pwkp\])]{}. The velocities $(\Omega_{k+1},\dot x_{k+1},\dot y_{k+1})$ are obtained from [(\[eqn:p\])]{}. This yields a flow map, $$(R_k,x_k,y_k,\Omega_k,\dot x_k,\dot y_k)\mapsto(R_{k+1},x_{k+1},y_{k+1},\Omega_{k+1},\dot x_{k+1},\dot y_{k+1}).$$
### Optimal control problem {#optimal-control-problem-2 .unnumbered}
The objective of the optimal control problem is to transfer the 3D pendulum on a cart from a given initial condition $(R_0,x_0,y_0,\Omega_0,\dot x_0,\dot y_0)$ to a desired terminal condition $(R^f,x^f,y^f,\Omega^f,\dot x^f,\dot y^f)$ during a fixed maneuver time $Nh$, while minimizing the square of the $l_2$ norm of the control inputs. $$\begin{gathered}
\min_{u_{k+1}} {\ensuremath{\left\{ \mathcal{J}=\sum_{k=0}^{N-1}
\frac{h}{2}u_{k+1}^T W u_{k+1} \right\}}},\label{eqn:J3dpendcart}\end{gathered}$$ where $u_{k}=[u_{x_k};u_{y_k}]\in{\ensuremath{\mathbb{R}}}^2$, and $W\in{\ensuremath{\mathbb{R}}}^{2\times 2}$ is a symmetric positive-definite matrix. The 3D pendulum on a cart is underactuated, since only the planar motion of the cart in its horizontal plane is actuated.
### Computational approach {#computational-approach-2 .unnumbered}
We apply a direct optimal control approach. The control inputs are parameterized by several points that are uniformly distributed over the maneuver time, and control inputs between these points are approximated using a cubic spline interpolation. For given control input parameters, the value of the cost is given by [(\[eqn:J3dpendcart\])]{}, and the terminal conditions are obtained by the discrete-time equations of motion given by [(\[eqn:pxk\])]{}-[(\[eqn:pwkp\])]{}. The control input parameters are optimized using constrained nonlinear parameter optimization to satisfy the terminal boundary conditions while minimizing the cost.
This approach is computationally efficient when compared to the usual collocation methods, where the continuous-time equations of motion are imposed as constraints at a set of collocation points. Using the proposed discrete-time optimal control approach, optimal control inputs can be obtained by using a large step size, thereby resulting in efficient total computations. Since the computed optimal trajectories do not have numerical dissipation caused by conventional numerical integration schemes, they are numerically more robust. Furthermore, the corresponding gradient information is accurately computed, which improves the convergence properties of the numerical optimization procedure.
### Numerical example {#numerical-example-2 .unnumbered}
Properties of the 3D pendulum and the cart are chosen as, $$M=m=1\,\mathrm{kg},\quad J=\mathrm{diag}[1.03,1.04,0.03]\,\mathrm{kgm^2},\quad\text{and}\quad d=[0,0,1]\,\mathrm{m}.$$ The desired maneuver is a rest-to-rest $180^\circ$ rotation of the pendulum about the vertical axis, while the cart returns to the initial location at the terminal time. The corresponding boundary conditions are given by $$\begin{gathered}
R_0=I,\quad \Omega_0=[0,0,0],\quad x_0=y_0=0,\quad \dot x_0=\dot y_0=0,\\
R^f=\mathrm{diag}[-1,-1,1],\quad \Omega^f=[0,0,0],\quad x^f=y^f=0,\quad \dot x^f=\dot y^f=0.\end{gathered}$$ The maneuver time is $2$ seconds, and the time step is $h=0.01$. Since only the planar motion of the cart is actuated, the rotation of the 3D pendulum is caused by the nonlinear coupling between the cart and the pendulum.
Each component of the control inputs is parameterized by 7 points. The resulting 14 control input parameters are optimized using sequential quadratic programming. [Figure \[fig:Opt3DPC\]]{} illustrates the optimal maneuver of the pendulum and the cart, angular velocity, and optimal control inputs. The blue circles denote the optimized control input parameters. The optimal cost and the violation of the terminal boundary conditions are $297.43$, and $1.83\times10^{-8}$, respectively. The optimal motion of the cart on the horizontal plane consists of a triangular-shaped loop, and the optimal maneuver of the 3D pendulum consists of large angle rotations. This also demonstrates the advantages of the computational geometric optimal control approach: it is difficult to study this kind of aggressive maneuvers of a multibody system using local coordinates, due to the coordinate singularities and the complexity of the equations in local coordinates. The presented computational geometric optimal control approach accurately characterizes the nonlinear coupling between the cart and the pendulum dynamics to obtain a nontrivial optimal maneuver of the 3D pendulum on a cart.
Optimal attitude reorientation of two connected rigid bodies on ${\ensuremath{\mathrm{SO}(3)}}\times{\ensuremath{\mathrm{SO}(3)}}$
----------------------------------------------------------------------------------------------------------------------------------
Consider two rigid bodies connected with a ball joint that has three rotational degrees of freedom. This represents a freely rotating system of coupled rigid bodies. The relative equilibria structure of this rigid body dynamics has been studied in [@Wan.1990]. We introduce three frames; an inertial frame and two body-fixed frames. Define
--------------------------------------- --------------------------------------------------------------------------------------------
$x\in{\ensuremath{\mathbb{R}}}^3$ Position of the ball joint in a reference frame
$R_i\in{\ensuremath{\mathrm{SO}(3)}}$ Rotation matrix from the $i$-th body-fixed frame to a reference frame
$d_i\in{\ensuremath{\mathbb{R}}}^3$ Vector from the joint to the mass center of the $i$-th body in the $i$-th body-fixed frame
$m_i\in{\ensuremath{\mathbb{R}}}$ Mass of the $i$-th body
--------------------------------------- --------------------------------------------------------------------------------------------
for $i\in\{1,2\}$. The configuration manifold is ${\ensuremath{\mathrm{SO}(3)}}\times{\ensuremath{\mathrm{SO}(3)}}\times{\ensuremath{\mathbb{R}}}^3$. In the absence of the potential field, the connected rigid body model has two symmetries; a symmetry of the translational action of ${\ensuremath{\mathbb{R}}}^3$, and a symmetry of the rotational action of ${\ensuremath{\mathrm{SO}(3)}}$.[^1] Due to these symmetries, the total linear momentum and the total angular momentum are preserved, and the configuration manifold can be reduced to a quotient space.
In this optimal control problem, we reduce the configuration manifold to ${\ensuremath{\mathrm{SO}(3)}}\times{\ensuremath{\mathrm{SO}(3)}}$ using the symmetry of the translational action of ${\ensuremath{\mathbb{R}}}^3$. The corresponding value of the total linear momentum is set to zero. The resulting connected rigid bodies model with a fixed mass center is closely related to the falling cat problem [@Eno.BK1993]. An appropriate cyclic change in the shape of the body yields a rotation in the orientation of the cat in accordance with the geometric phase effect [@Mon.1991]. In contrast to other models of the falling cat, which typically introduce two one-dimensional rotational joints, with a shape space given by $S^1\times S^1$, we consider instead a single ball joint with a shape space given by $SO(3)$.
Similar to the falling cat problem, we assume that an internal control moment $u\in{\ensuremath{\mathbb{R}}}^3$ is applied at the joint, so that it controls the relative attitude between two rigid bodies. More precisely, the control input $u$ represents the control moment applied to the first rigid body, represented in the reference frame. The equal and opposite control moment is applied to the second rigid body. Therefore the control moment changes the shape of the system. The total angular momentum is conserved for the controlled dynamics as the control input is an internal moment of the connected rigid bodies system. This optimal control problem is similar to the optimal control problem of the 3D pendulum discussed in Section \[subsec:3dpend\], as the control input respects the symmetry, and the corresponding momentum is preserved in the controlled dynamics.
The discrete Lagrangian for the two connected rigid bodies is $$\begin{aligned}
L_d (R_{1_k},F_{1_k}, R_{2_k},F_{2_k},x_k,x_{k+1}) =& \frac{m_1+m_2}{2h}(x_{k+1}-x_k)\cdot(x_{k+1}-x_k) + \frac{1}{h}{\mbox{tr}\ensuremath{\negthickspace{\ensuremath{\left[ (I_{3\times 3}-F_{1_k})J_{d_1} \right]}}}}\nonumber\\
&+\frac{1}{h}{\mbox{tr}\ensuremath{\negthickspace{\ensuremath{\left[ (I_{3\times 3}-F_{2_k})J_{d_2} \right]}}}}
+\frac{1}{h}{\mbox{tr}\ensuremath{\negthickspace{\ensuremath{\left[ m_1R_{1_k}(F_{1_k}-I_{3\times 3})d_1(x_{k+1}-x_k)^T \right]}}}}\nonumber\\
&+\frac{1}{h}{\mbox{tr}\ensuremath{\negthickspace{\ensuremath{\left[ m_2R_{2_k}(F_{2_k}-I_{3\times 3})d_2(x_{k+1}-x_k)^T \right]}}}}.\end{aligned}$$ From [(\[eqn:DEP\])]{}, we obtain the Lie group variational integrator, viewed as discrete-time equations of motion on ${\ensuremath{\mathrm{SO}(3)}}\times{\ensuremath{\mathrm{SO}(3)}}\times{\ensuremath{\mathbb{R}}}^3$. Since we are only interested in rotational maneuvers, we derive the following reduced equations of motion on ${\ensuremath{\mathrm{SO}(3)}}\times{\ensuremath{\mathrm{SO}(3)}}$ using the fact that the linear momentum is conserved. [$$\begin{aligned}
\hat p_{1_k} = & \frac{1}{h} {\ensuremath{\left\{ F_{1_k} (J_{d_1}-\alpha m_1 d_1 d_1^T) - (J_{d_1}-\alpha m_1 d_1 d_1^T ) F_{1_k}^T \right\}}}\nonumber\\
&-\beta\frac{m_1}{h}( R_{1_k}^T R_{2_k} F_{2_k} d_2 d_1^T - d_1 d_2^T F_{2_k}^T R_{2_k}^T R_{1_k})
+ \beta\frac{m_1}{h} (R_{1_k}^T R_{2_k} d_2 d_1^T -d_1 d_2^T R_{2_k}^T R_{1_k}) ,\label{eqn:p1kd}\\
\hat p_{2_k} = & \frac{1}{h}{\ensuremath{\left\{ F_{2_k} (J_{d_2} -\beta m_2 d_2d_2^T) -(J_{d_2}-\beta m_2 d_2d_2^T)F_{2_k}^T \right\}}}\nonumber\\
&-\alpha\frac{m_2}{h} ( R_{2_k}^T R_{1_k} F_{1_k} d_1 d_2^T - d_2 d_1^T F_{1_k}^T R_{1_k}^T R_{2_k}) +\alpha\frac{m_2}{h} (R_{2_k}^T R_{1_k} d_1 d_2^T -d_2 d_1^T R_{1_k}^T R_{2_k}) ,\label{eqn:p2kd}\\
R_{1_{k+1}}= &R_{1_k}F_{1_k},\label{eqn:R1kp}\\
R_{2_{k+1}}= &R_{2_k}F_{2_k},\label{eqn:R2kp}\\
p_{1_{k+1}} = & F_{1_k}^T (p_{1_k} -(B_{1_k}-B_{1_k}^T)^\vee)+hR_{1_{k+1}}^T u_{k+1},\label{eqn:p1kp}\\
p_{2_{k+1}} = & F_{2_k}^T (p_{2_k} -(B_{2_k}-B_{2_k}^T)^\vee)-hR_{2_{k+1}}^T u_{k+1},\label{eqn:p2kp}\end{aligned}$$]{} where $\alpha=\frac{m_1}{m_1+m_2}$, $\beta=\frac{m_2}{m_1+m_2}\in{\ensuremath{\mathbb{R}}}$, and the matrix $B_{i_k}\in{\ensuremath{\mathbb{R}}}^{3\times 3}$ for $i\in\{1,2\}$ is defined as $$\begin{aligned}
B_{i_k} = \frac{m_i}{h}(F_{i_k}-I)d_i{\ensuremath{\left\{ -\alpha R_{1_k} (F_{1_k}-I)d_1 -\beta R_{2_k} (F_{2_k}-I)d_2 \right\}}}^T R_{i_k}.\label{eqn:Bik}\end{aligned}$$ The momenta variables $p_{1},p_{2}\in{\ensuremath{\mathbb{R}}}^3$ are given by $$\begin{aligned}
\begin{bmatrix} p_{1} \\
p_{2} \end{bmatrix}=
\begin{bmatrix} J_1+\alpha m_1\hat d_1^T & \beta m_1 \hat d_1 R_{1}^T R_{2}\hat d_2\\
\alpha m_2\hat d_2 R_{2}^T R_{1} \hat d_1 & J_2 + \beta m_2 \hat d_2^2\end{bmatrix}
\begin{bmatrix} \Omega_{1} \\ \Omega_{2} \end{bmatrix}.\label{eqn:LTp}\end{aligned}$$
For given $(R_{1_k},R_{2_k},\Omega_{1_k},\Omega_{2_k})$, we find $p_{1_k},p_{2_k}$ by [(\[eqn:LTp\])]{}. We solve the implicit equations [(\[eqn:p1kd\])]{}, [(\[eqn:p2kd\])]{} to obtain $F_{1_k},F_{2_k}$. Then, $R_{1_{k+1}},R_{2_{k+1}}$ are obtained from [(\[eqn:R1kp\])]{},[(\[eqn:R2kp\])]{}, and $p_{1_{k+1}},p_{2_{k+1}}$ are obtained by [(\[eqn:p1kp\])]{},[(\[eqn:p2kp\])]{}. Finally, $\Omega_{1_{k+1}},\Omega_{2_{k+1}}$ are computed from [(\[eqn:LTp\])]{}. This yields a discrete flow map $(R_{1_k},R_{2_k},\Omega_{1_k},\Omega_{2_k})\mapsto(R_{1_{k+1}},R_{2_{k+1}},\Omega_{1_{k+1}},\Omega_{2_{k+1}})$.
### Optimal control problem {#optimal-control-problem-3 .unnumbered}
The objective of the optimal control problem is to transfer the connected rigid bodies from a given initial condition $(R_{1_0},R_{2_0},\Omega_{1_0},\Omega_{2_0})$ to a desired terminal condition $(R_1^f,R_2^f,\Omega_1^f,\Omega_2^f)$ during a fixed maneuver time $Nh$, while minimizing the square of the $l_2$ norm of the control inputs. $$\begin{gathered}
\min_{u_{k+1}} {\ensuremath{\left\{ \mathcal{J}=\sum_{k=0}^{N-1}
\frac{h}{2}u_{k+1}^T W u_{k+1} \right\}}},\label{eqn:J4}\end{gathered}$$ where $W\in{\ensuremath{\mathbb{R}}}^{3\times 3}$ is a symmetric positive-definite matrix. In particular, we choose an attitude maneuver that is described by a rest-to-rest rotation of the entire system while the relative attitude configuration at the terminal time is the same as at the initial time.
### Computational approach {#computational-approach-3 .unnumbered}
We apply a direct optimal control approach. For a given control input, the value of the cost is given by [(\[eqn:J4\])]{}, and the terminal conditions are obtained by the discrete-time equations of motion given by [(\[eqn:p1kd\])]{}-[(\[eqn:Bik\])]{}. We use constrained nonlinear parameter optimization to minimize the cost function subject to the terminal boundary condition obtained by the discrete-time equations of motion.
Since the total angular momentum is conserved regardless of the control input, the terminal constraints introduces a singularity due to the rotational symmetry. This ill-conditioning can be avoided by disregarding the terminal angular velocity constraint for the second body. For the given boundary conditions, the terminal angular velocity condition is automatically satisfied if the remaining terminal constraints are satisfied, due to the angular momentum conservation property. By formulating the optimization process this way, we eliminate the source of numerical ill-conditioning. This is similar to the modified computational approach discussed in Section \[subsec:3dpend\].
### Numerical example {#numerical-example-3 .unnumbered}
Properties of the rigid bodies are chosen as $$\begin{aligned}
m_1 = 1.5\mathrm{kg},\quad J_1=\begin{bmatrix}0.18 & 0.32 & 0.32\\0.32 & 1.88 & -0.06\\0.32 & -0.06& 1.86\end{bmatrix}\mathrm{kg\cdot m^2},\quad d_1=[-1.08,0.20,0.20]\mathrm{m},\\
m_2 = 1\mathrm{kg},\quad J_2=\begin{bmatrix}0.11 &-0.18 &-0.18\\-0.18 & 0.89 & -0.04\\-0.18 & -0.04 & 0.88\end{bmatrix}\mathrm{kg\cdot m^2},\quad d_2=[0.9,0.2,0.2]\mathrm{m}.\end{aligned}$$ The desired maneuver is a rest-to-rest $180^\circ$ rotation about the $x$ axis. $$\begin{gathered}
R_{1_0}=I,\quad \Omega_{1_0}=0,\quad R_{2_0}=I,\quad \Omega_{2_0}=0,\\
R_1^f=\mathrm{diag}[1,-1,-1] ,\quad \Omega_1^f=[0,0,0],\quad R_2^f=\mathrm{diag}[1,-1,-1],\quad \Omega_2^f=[0,0,0].\end{gathered}$$ The maneuver time is $4$ seconds, and the step size is $h=0.01$.
We parameterize each component of the control input at $7$ discrete points, and the control inputs are reconstructed by cubic spline interpolation. The resulting $21$ control input parameters are optimized by a sequential quadratic programming method to satisfy the terminal boundary conditions while minimizing the cost function.
[Figure \[fig:OptTRB\]]{} shows the optimal maneuver of the rigid bodies, angular velocity, and optimal control inputs. The blue circles denote the optimized control input parameters. The optimal cost and the violation of the terminal boundary conditions are $0.574$, and $2.48\times10^{-8}$, respectively.
The optimal maneuver consists of large angle rotations of the two rigid bodies. Throughout this complicated maneuver, the total angular momentum is zero, and the rotation about the $e_1$ axis depends on the geometric phase effect. This also demonstrates the advantages of the computational geometric optimal control approach. The Lie group variational integrator computes the weak geometric phase effect accurately, so that the iterations converge to a nontrivial optimal maneuver of the coupled rigid bodies.
Conclusions
===========
In this paper, a computational geometric approach for an optimal control problem of rigid body dynamics has been developed. The essential idea is formulating a discrete-time optimal control problem using a structure-preserving geometric numerical integrator, referred to as a Lie group variational integrator, and applying standard optimal control approaches, such as an indirect optimal control and a direct optimal control, to discrete-time equations of motion. This method is in contrast to the usual optimal control approach, where the discretization appears only in the last stage when numerically computing the optimal control inputs.
The computational geometric optimal control approach has substantial advantages in terms of preserving the geometric properties of optimality conditions. The discrete flow of Lie group variational integrators has desirable geometric properties, such as symplecticity and momentum preservation, and it is more reliable and robust over longer time periods. The computational geometric optimal control approach inherits the desirable properties of the Lie group variational integrator. In the necessary conditions for optimality, the multiplier equations are dual to the linearized equations of motion. Since the linearized flow of a Lagrangian/Hamiltonian system is also symplectic, the multiplier equations inherit certain geometric properties. The discrete-time necessary conditions preserve the geometric properties of the optimality conditions, as they are derived from a discrete-time analogue of Hamilton’s variational principle that yields a symplectic discrete-time flow.
The computational geometric optimal control approach allows us to find the optimal control input more efficiently. In the indirect optimal control, the shooting method may be prone to numerical ill-conditioning, as a small change in the initial multiplier can cause highly nonlinear behavior of the terminal condition. However, as shown in [Figure \[fig:OptRB\_err\]]{} and [Figure \[fig:Opt3DP\_err\]]{}, the computational geometric optimal control approach exhibits excellent numerical convergence properties. This is because the proposed computational algorithms are geometrically exact and numerically accurate. There is no numerical dissipation introduced by the numerical algorithm, and therefore, we are more accurately characterizing the sensitivities along the solution.
Another advantage of the computational geometric optimal control of rigid bodies is that the method is directly developed on a Lie group. There is no ambiguity or singularity in representing the configuration of rigid bodies globally, and the resulting equations of motion are more compact than those written in terms of local coordinates. As illustrated by [Figure \[fig:Opt3DPC\_3d\]]{} and [Figure \[fig:OptTRB\_3d\]]{}, the presented computational geometric optimal control approach utilizes the effects of the nonlinear coupling and the weak geometric phase of a multibody system to obtain nontrivial aggressive maneuvers of the rigid bodies. These results are independent of a specific choice of local coordinates, and they completely avoid any singularity, ambiguity, and complexity associated with local coordinates. Furthermore, the numerical results are group-equivariant, and are independent of the choice of inertial frame, which is in contrast to methods based on local coordinate representations. By formulating the problem in a global and intrinsic fashion, the algorithms presented are able to explore the space of control strategies which extend beyond a single coordinate chart, thereby providing a deeper insight into the global controlled dynamics of systems of rigid bodies.
[^1]: These can be considered as a single symmetry of the translational and rotational action of ${\ensuremath{\mathrm{SE}(3)}}$, but they are considered separately in this optimal control problem. By the general theory of reduction by stages [@CeMaRa2001], the two approaches are equivalent.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
This paper provides an in-depth study of the fundamental problems of finding small subgraphs in distributed dynamic networks.
While some problems are trivially easy to handle, such as detecting a triangle that emerges after an edge insertion, we show that, perhaps somewhat surprisingly, other problems exhibit a wide range of complexities in terms of the trade-offs between their round and bandwidth complexities.
In the case of triangles, which are only affected by the topology of the immediate neighborhood, some end results are:
- The bandwidth complexity of $1$-round dynamic triangle detection or listing is $\Theta(1)$.
- The bandwidth complexity of $1$-round dynamic triangle membership listing is $\Theta(1)$ for node/edge deletions, $\Theta(n^{1/2})$ for edge insertions, and $\Theta(n)$ for node insertions.
- The bandwidth complexity of $1$-round dynamic triangle membership detection is $\Theta(1)$ for node/edge deletions, $O(\log n)$ for edge insertions, and $\Theta(n)$ for node insertions.
Most of our upper and lower bounds are *tight*. Additionally, we provide almost always tight upper and lower bounds for larger cliques.
author:
- |
Matthias Bonne [^1]\
Department of Computer Science, Technion, Israel
- |
Keren Censor-Hillel [^2]\
Department of Computer Science, Technion, Israel
bibliography:
- 'related.bib'
title: Distributed Detection of Cliques in Dynamic Networks
---
Introduction
============
A fundamental problem in many computational settings is to find small subgraphs. In distributed networks it is particularly vital for various reasons, among which is the ability to perform some tasks much faster if, say, triangles do not occur in the underlying network graph (see, eg., [@Pettie2015; @Hirvonen2017]).
Finding cliques is a *local* task that trivially requires only a single communication round if the message size is unbounded. However, its complexity may dramatically increase when the bandwidth is restricted to the standard $O(\log n)$ bits, for an $n$-node network. For example, the complexity of detecting 4-cliques is at least $\Omega(n^{1/2})$ [@Czumaj2018]. For triangles, the complexity is yet a fascinating riddle, where only recently the first non-trivial complexities of $\tilde{O}(n^{2/3})$ and $\tilde{O}(n^{3/4})$ have been given for detection and listing, respectively [@Izumi2017], and the current state-of-the-art is the $\tilde{O}(n^{1/2})$-round algorithm of [@Chang2019]. For listing, there is an $\tilde{\Omega}(n^{1/3})$ lower bound [@Izumi2017; @Pandurangan0S18]. The only non-trivial lower bounds for detection say that a single round is insufficient, as given in [@Abboud2017] and extended for randomized algorithms in [@Fischer2018]. In [@Abboud2017] it was also shown that $1$-bit algorithms require $\Omega(\log^* n)$ rounds, improved in [@Fischer2018] to $\Omega(\log{n})$.
In this paper, we address the question of detecting small cliques in *dynamic* networks of limited bandwidth. We consider a model that captures real-world behavior in networks that undergo changes, such as nodes joining or leaving the network, or communication links between nodes that appear or disappear. Various problems have been studied in many variants of such a setting (see Section \[subsec:related\]).
The task of finding cliques is unique, in that it is trivial if the bandwidth is not restricted, and it can be easily guaranteed that at the end of each round all nodes have the correct output. This implies that one does not have to wait for *stabilization* and does not have to assume that the network is *quiet* for any positive number of rounds. If, however, the bandwidth is restricted, the solution may not be as simple, although some problems can still be solved even with very small bandwidth.
As a toy example, consider the case of triangle listing when a new edge is inserted to the graph. The endpoints of the inserted edge simply broadcast a single bit that indicates this change to all of their neighbors, and hence if a new triangle is created then its third endpoint detects this by receiving two such indications.
Nevertheless, we show that this simplified case is far from reflecting the general complexities of clique problems in such a dynamic setting. For example, the above algorithm does not solve the problem of *membership-listing* of triangles, in which *each* node should list all triangles that contain it. Indeed, we prove that this stronger variant *cannot* be solved with constant bandwidth, and, in fact, every solution must use at least $\Omega(\sqrt{n})$ bits.
Our contributions provide an in-depth study of various detection and listing problems of small cliques in this dynamic setting, as we formally define and summarize next.
Our contributions
-----------------
For a subgraph $H$ we categorize four types of tasks: Detecting an appearance of $H$ in the network, for which it is sufficient that a single node does so, listing all appearances of $H$, such that every appearance is listed at least by a single node, and their two *membership* variants, membership-detection and membership-listing, for which each node has to detect whether it is a member of a copy of $H$, or list all copies of $H$ to which it belongs, respectively.
The model is explicitly defined in Section \[section:prelim\]. In a nutshell, there can be one topology change in every round, followed by a standard communication round among neighboring nodes, of $B$ bits per message, where $B$ is the bandwidth. An algorithm takes $r$ rounds if the outputs of the nodes are consistent with a correct solution after $r$ communication rounds that follow the last change. In particular, a $1$-round algorithm is an algorithm in which the outputs are correct already at the end of the round in which the topology change occurred. Hence, $1$-round algorithms are very strong, in the sense that they work even in settings in which no round is free of changes. We note that in what follows, all of our algorithms are deterministic and all of our lower bounds hold also for randomized algorithms.
\
**Triangles:** Our upper and lower bounds for triangles ($H=K_3$) are displayed in Table \[tbl:summary\].
Node deletions Edge deletions Edge insertions Node insertions
---------------------- ---------------- ---------------- -------------------- -----------------
Detection/Listing $0$ $\Theta(1)$ $\Theta(1)$ $\Theta(1)$
Membership detection $0$ $\Theta(1)$ $O(\log n)$ $\Theta(n)$
Membership listing $0$ $\Theta(1)$ $\Theta(\sqrt{n})$ $\Theta(n)$
: The bandwidth complexities of $1$-round algorithms for dynamic triangle problems. []{data-label="tbl:summary"}
Most of the complexities in this table are shown to be tight, by designing algorithms and proving matching lower bounds. The one exception is for membership detection with edge insertions, for which we show an algorithm that uses $O(\log n)$ bits of bandwidth, but we do not know whether this is tight. However, we also show a $1$-round algorithm for this problem that works with a bandwidth of $O((\Delta\log{n})^{1/2})$, where $\Delta$ is the maximum degree in the graph, implying that if our logarithmic algorithm is optimal, then a proof for its optimality must exhibit a worst case in which the maximum degree is $\Omega(\log{n})$.
A single round is sufficient for solving all clique problems, given enough (linear) bandwidth. Nevertheless, for the sake of comparison, we show that with just one additional communication round, all of the bandwidth complexities in Table \[tbl:summary\] drop to $\Theta(1)$, apart from membership listing of triangles, whose bandwidth complexity for $r$-round algorithms is $\Theta(n/r)$.
\
**Larger Cliques:** We also study the bandwidth complexities for finding cliques on $s>3$ nodes. Here, too, a single round is sufficient for all problems, and the goal is to find the bandwidth complexity of each problem. Some of the algorithms and lower bounds that we show for triangles carry over to larger cliques, but others do not. Yet, with additional techniques, we prove that for cliques of constant size, almost all of the $1$-round bandwidth complexities are the same as their triangle counterparts.
\
**Combining types of changes:** In most cases, one can obtain an algorithm that handles several types of changes, by a simple combination of the corresponding algorithms. However, intriguingly, sometimes this is not the case. A prime example is when trying to combine edge insertions and node insertions for triangle detection: a node obtaining 1-bit indications of a change from two neighbors cannot tell whether this is due to an edge inserted between them, or due to an insertion of a node that is a neighbor of both. In some of these cases we provide techniques to overcome these difficulties, and use them to adjust our algorithms to cope with more than one type of change.
Challenges and techniques
-------------------------
\
**Algorithms:** The main challenge for designing algorithms is how to convey enough information about the topology changes that occur, despite non-trivial (in particular, sublinear) bandwidth. Consider, e.g., edge insertions. As described, while listing triangles is trivial with a single indication bit, this fails for membership detection or membership listing.
For membership detection we can still provide a very simple algorithm for which a logarithmic bandwidth suffices, by sending the identity of the new edge, and by *helping* neighbors in the triangle to know that they are such. For membership listing even this is insufficient. To overcome this challenge, we introduce a technique for sending and collecting *digests* of neighborhood information. When all digests from a given neighbor have been collected, one can determine the entire neighborhood of this neighbor. The caveat in using such an approach in a straightforward manner is that a node needs to list its triangles with a newly connected neighbor already at the same round in which they connected, and cannot wait to receive all of its neighbor’s digests. By our *specific choice* of a digests, we ensure that a newly-connected neighbor has enough information to give a correct output after a single communication round, despite receiving only a single digest from its latest neighbor.
\
**Lower bounds:** For the lower bounds, our goal is to argue that in order to guarantee that all nodes give a correct output, each node must receive enough information about the rest of the graph. To do this, we identify sequences of topology changes that exhibit a worst-case behavior, in the sense that a node cannot give a correct output if it receives too little information.
One approach for doing this is to look at a particular node $x$, and define as many sequences as possible such that the correct output of $x$ is different for each sequence. At the same time, we ensure that the number of messages that $x$ receives from the other nodes is as small as possible. These two requirements are conflicting — if $x$ can have many different outputs, it must have many different neighbors, which implies that it receives many messages with information. Still, we are able to find such a family of sequences for each problem, and we wrap-up our constructions by using counting arguments to prove the desired lower bounds. In some cases, e.g., membership-listing under edge insertions, even this is insufficient, and we construct the sequences such that one *critical* edge affects the output of $x$, but it is *added last*, so that it conveys as little information as possible.
The above can be seen as another step in the spirit of the *fooling views* framework, which was introduced in [@Abboud2017] for obtaining the first lower bounds for triangle detection under limited bandwidth. After being beautifully extended by [@Fischer2018], our paper essentially gives another indication of the power of the fooling views framework in proving lower bounds for bandwidth-restricted settings.
On the way to constructing our worst-case graph sequences, we prove combinatorial lemmas that show the existence of graphs with certain desired properties. To make our lower bounds apply also for randomized algorithms, we rely on Yao’s lemma and on additional machinery that we develop.
Additional related work {#subsec:related}
-----------------------
**Dynamic distributed algorithms:** Dynamic networks have been the subject of much research. A central paradigm is that of *self-stabilization* [@Dolev2000], in which the system undergoes various changes, and after some quiet time needs to converge to a stable state. The model addressed in this paper can be considered as such, but our focus is on algorithms that do not require any quiet rounds (though we also address the gain in bandwidth complexity if the system does allow them, for the sake of comparison). Yet, we assume a single topology change in each round. A single change in each round and enough time to recover is assumed in recent algorithms for maintaining a maximal independent set [@CensorHillel2016; @Assadi2019; @GuptaK18; @AssadiOSS18; @DuZ2018] and matchings [@Solomon16], and for analyzing amortized complexity [@ParterPS16]. Highly-dynamic networks, in which multiple topology changes can occur in each round are analyzed in [@Bamberger2018; @CensorHillelDKPS19], and it is an intriguing open question how efficient can subgraph detection be in such a setting. Various other dynamic settings in which the topology changes can be almost arbitrary have been proposed, but the questions that arise in such networks address the flow of information and agreement problems, as they are highly non-structured. A significant paper that relates to our work is [@KonigW13], which differs from our setting in the bandwidth assumption. Clique detection in the latter model is trivial, and indeed their paper addresses other problems.
**Distributed subgraph detection:** In the CONGEST model and in related models, where the nodes synchronously communicate with $O(\log{n})$ bits, much research is devoted to subgraph detection, e.g., [@Fraigniaud2017; @Gonen2017; @Korhonen2017; @Drucker2014; @Abboud2017; @Izumi2017; @Chang2019; @Pandurangan0S18].
A related problem is property-testing of subgraphs, where the nodes need to determine, for a given subgraph $H$, whether the graph is $H$-free or *far* from being $H$-free [@CensorHillel2019; @EvenFischer2017; @EvenLevi2017; @Fischer2017; @Fraigniaud2016; @Brakerski2011].
Preliminaries {#section:prelim}
-------------
In the dynamic setting, the network is a sequence of graphs. The initial graph, whose topology is known to the nodes, represents the state of the network at some starting point in time, and every other graph in the sequence is either identical to its preceding graph or obtained from it by a single topology change. The network is synchronized, and in each round, each node can send to each one of its neighbors a message of $B$ bits, where $B$ is the bandwidth of the network. Each node has a unique ID and knows the IDs of all its neighbors. We denote by $n$ the number of possible IDs. Hence, $n$ is an upper bound on the number of nodes that can participate in the network at all times.
Note that the nodes do not receive any indication when a topology change occurs. A node can deduce that a change has occurred by comparing the list of its neighbors in the current round and in the previous round. This implies that a node $u$ cannot distinguish between the insertion of an edge $uv$ and the insertion of a node $v$ which is connected to $u$, as the list of neighbors of $u$ is the same in both cases.
An algorithm can be designed to handle edge insertions or deletions or node insertions or deletions, or any combination of these. We say that an algorithm is an $r$-round algorithm if the outputs of the nodes after $r$ rounds of communication starting from the last topology change are correct. The (deterministic or randomized) $r$-round bandwidth complexity of a problem is the minimum bandwidth for which an $r$-round (deterministic or randomized) algorithm exists. We denote by [$\textsf{MemList}(H)$]{}, [$\textsf{MemDetect}(H)$]{}, [$\textsf{List}(H)$]{}, and [$\textsf{Detect}(H)$]{}, respectively, the problems of membership listing, membership detection, listing and detection of $H$. The following is a simple observation.
\[obs:probrel\]
Let $r$ be an integer and let $H$ be a graph. Denote by $B_{{\textsf{MemList}}}$, $B_{{\textsf{MemDetect}}}$, $B_{{\textsf{List}}}$ and $B_{{\textsf{Detect}}}$ the (deterministic or randomized) $r$-round bandwidth complexities of [$\textsf{MemList}(H)$]{}, [$\textsf{MemDetect}(H)$]{}, [$\textsf{List}(H)$]{}, and [$\textsf{Detect}(H)$]{}, respectively, under some type of topology changes. Then: $B_{{\textsf{Detect}}} \leq B_{{\textsf{List}}} \leq B_{{\textsf{MemList}}}$ and $B_{{\textsf{Detect}}} \leq B_{{\textsf{MemDetect}}} \leq B_{{\textsf{MemList}}}$.
Due to this observation, our exposition starts with the more challenging task of membership-listing, then addresses membership-detection, and concludes with listing and detection. Within each case, we first provide algorithms for each type of topology change, and then prove lower bounds.
Triangle problems
=================
Membership listing
------------------
Table \[tbl:k3\_mlist\] summarizes our results for membership listing of triangles.
Node deletions Edge deletions Edge insertions Node insertions
------------ ---------------- ---------------- -------------------- -----------------
$r = 1$ $0$ $\Theta(1)$ $\Theta(\sqrt{n})$ $\Theta(n)$
$r \geq 2$ $0$ $\Theta(1)$ $\Theta(1)$ $\Theta(n/r)$
: Bandwidth complexities of [$\textsf{MemList}(K_3)$]{}. []{data-label="tbl:k3_mlist"}
### Upper bounds {#subsubsec:tri-mlist-upper}
We start by showing that if only node deletions are allowed, no communication is required.
\[th:k3\_mlist\_ub\_mv\] The deterministic $1$-round bandwidth complexity of [$\textsf{MemList}(K_3)$]{} under node deletions is $0$.
The algorithm is very simple: the output of a node $u$ is the set of all triangles $uvw$ that exist in the initial graph, such that $v$ and $w$ are both neighbors of $u$ in the final graph.
Since only node deletions are allowed, every triangle $uvw$ that exists in the final graph must also exist in the initial graph, and all three nodes $u$, $v$ and $w$ must exist and be connected to each other. Conversely, if a triangle $uvw$ exists in the initial graph and $u$ is still connected to $v$ and $w$, then the triangle $uvw$ must exist in the final graph. Therefore, the output of $u$ is exactly the set of all triangles that contain $u$ in the final graph.
Next, we show how to handle edge deletions with $O(1)$ bits of bandwidth.
\[th:k3\_mlist\_ub\_me\] The deterministic $1$-round bandwidth complexity of [$\textsf{MemList}(K_3)$]{} under edge deletions is $O(1)$.
The idea of the algorithm is simple: every node $u$ needs to maintain the list of all triangles that contain it. In order to do this, $u$ needs to be able to determine, for every two of its neighbors $v$ and $w$, whether or not the edge $vw$ exists.
This can be achieved with only a single bit, as follows. Every node sends to all its neighbors an indication whether or not it lost an edge on the current round. We denote this bit by `DELETED`. Thus, for every triangle $uvw$, whenever one of the edges of the triangle is deleted (say, $vw$), the other node ($u$ in this case) receives $\texttt{DELETED} = 1$ from both $v$ and $w$. Since the only possible change in the graph is the deletion of an edge, $u$ can deduce that the edge $vw$ was deleted.
The above algorithm does not work as-is if both edge deletions and node deletions are allowed. For example, assume there exists a triangle $uvw$. When we remove the edge $vw$, both $v$ and $w$ send $\texttt{DELETED} = 1$ to $u$, thus $u$ knows that the edge $vw$ was deleted and the triangle $uvw$ no longer exists. However, $u$ receives the exact same input if some other node $x$, which is connected to both $v$ and $w$, but not to $u$, is deleted. Here, too, both $v$ and $w$ lose a neighbor, thus they both send $\texttt{DELETED} = 1$ to $u$, and $u$ deduces — now incorrectly — that the triangle $uvw$ no longer exists and removes it from its output.
This problem can be fixed, by observing that when the edge $vw$ is deleted, both $v$ and $w$ know whether or not $x$ is connected to $u$ (i.e., whether or not the triangle $vxu$, or $wxu$, exists). Therefore, we can have each node $v$ send $\texttt{DELETED} = 1$ to each neighbor $u$ only if the lost neighbor was connected to $u$; that is, $v$ sends $\texttt{DELETED} = 1$ to $u$ if it lost an edge to some node $x$ such that the triangle $vux$ existed on the previous round. This information allows $u$ to distinguish between the two cases described above and give the correct output. Hence the following corollary.
\[cor:k3\_mlist\_ub\_me\_mv\] The deterministic $1$-round bandwidth complexity of [$\textsf{MemList}(K_3)$]{} under node/edge deletions is $O(1)$.
Handling edge insertions and node insertions is much more complicated. We start by showing an algorithm that handles edge insertions.
\[th:k3\_mlist\_ub\_pe\] The deterministic $1$-round bandwidth complexity of [$\textsf{MemList}(K_3)$]{} under edge insertions is $O(\sqrt{n})$.
Let $N_u(r)$ be the set of neighbours of $u$ on round $r$. Note that $N_u(r)$ can be encoded as an $n$-bit string, which indicates, for every node $x$, whether or not $x$ is a neighbour of $u$ on round $r$.
The algorithm is as follows. When a new edge $uv$ is inserted on round $r$, both $u$ and $v$ send to all their neighbors the identity of their new neighbor, denoted by `NEWID`.
In addition, $u$ sends a bitmask of ${\left\lceil\sqrt{n}\right\rceil}$ bits to $v$, indicating, for every one of the previous ${\left\lceil\sqrt{n}\right\rceil}$ rounds, whether or not one of $u$’s neighbors has sent a `NEWID` to $u$. We denote this information by $\texttt{LAST}_u$. Node $v$ sends to $u$ the same information.
Finally, $u$ encodes $N_u(r)$ as an $n$-bit string, denoted $\texttt{ALL}_u(r)$, and starts sending it to $v$ in chunks of ${\left\lceil\sqrt{n}\right\rceil}$ bits per round. This process begins on round $r$ and ends on round $r + {\left\lfloor\sqrt{n}\right\rfloor}$, when the entire string has been sent. During these rounds, $u$ keeps sending `NEWID` to $v$, and to all its other neighbors, as described above. Node $v$ does the same. It should be noted that this continuous communication between $u$ and $v$ is not intended for allowing them to detect triangles that appear by the insertion of the edge $uv$, as these are detected immediately due to previous information. Rather, communicating `ALL` allows $u$ and $v$ to detect triangles that appear by other topology changes that may occur in subsequent rounds, as we show below.
Overall, `NEWID` requires $O(\log n)$ bits, while `LAST` and `ALL` require $O(\sqrt{n})$ bits. Thus, the required bandwidth is $O(\sqrt{n})$.
We show that at the end of each round, every node $u$ has enough information to determine, for every two of its neighbors $v$ and $w$, whether or not the edge $vw$ exists.
First, since only edge insertions are considered, if the edge $vw$ exists in the initial graph, it exists throughout. Also, if $vw$ is inserted when at least one of the edges $uv$ or $uw$ already exists, then $u$ receives this information through `NEWID`. The only other case is when $vw$ does not exist in the initial graph, and is inserted when $u$ is not yet connected to either $v$ or $w$. That is, the initial graph does not contain any of these three edges, and $vw$ is inserted before the other two. W.l.o.g. assume that $uv$ is inserted before $uw$. Now, let $t_v$ be the round in which $uv$ is inserted, and let $t_w$ be the round in which $uw$ is inserted.
If $t_w - t_v \leq {\left\lceil\sqrt{n}\right\rceil}$, then when $uv$ is inserted, $v$ sends a `NEWID` to $w$. Therefore, when $uw$ is inserted, $u$ can determine from $\texttt{LAST}_w$ that in round $t_v$ a neighbor of $w$ has sent a `NEWID` to $w$. Since the only edge inserted in that round is $uv$, $u$ determines that $v$ is a neighbor of $w$. If $t_w - t_v > {\left\lceil\sqrt{n}\right\rceil}$, then, in round $t_w$, $v$ has already sent the entire string $\texttt{ALL}_v(t_v)$ to $u$, which indicates that $w$ is a neighbor of $v$. Therefore, in all cases, $u$ determines whether or not the edge $vw$ exists, as claimed.
The algorithm given for Theorem \[th:k3\_mlist\_ub\_pe\] can be extended to handle edge deletions and node deletions as well, by simply having each node send to all its neighbors an indication whether or not it lost a neighbor on the current round, along with the ID of the lost neighbor (if any). This requires only $O(\log n)$ bits of bandwidth, and one can verify that this extended algorithm allows every node to give the correct output in all cases, as required. Therefore:
\[cor:k3\_mlist\_ub\_pe\_mvme\] The deterministic $1$-round bandwidth complexity of [$\textsf{MemList}(K_3)$]{} under node deletions and edge deletions/insertions is $O(\sqrt{n})$.
Also, if we are promised a quiet round, the problem can be solved with $O(1)$ bits of bandwidth, as the next theorem shows.
\[th:k3\_mlist\_ub\_r2\_pe\] The deterministic $2$-round bandwidth complexity of [$\textsf{MemList}(K_3)$]{} under edge insertions is $O(1)$.
On every round, for every two neighbors $u$ and $v$, $u$ sends two bits to $v$:
1. An indication whether or not $u$ has a new neighbor $w$ (i.e., the edge $uw$ has been inserted on the current round). We denote this information by `NEW`.
2. An indication whether or not any other neighbor of $u$ has sent $\texttt{NEW} = 1$ on the previous round. We denote this information by `LAST`.
Consider a triangle $uvw$, and assume that the edge $uv$ is the last one inserted. When $uv$ is inserted, both $u$ and $v$ send $\texttt{NEW} = 1$ to $w$. Since only one edge can be inserted on a single round, $w$ knows that the edge $uv$ has been inserted, and therefore the triangle $uvw$ exists.
On the next round, since $w$ received $\texttt{NEW} = 1$ from $v$, it sends $\texttt{LAST} = 1$ to $u$, indicating that some other neighbor of $w$ has sent $\texttt{NEW} = 1$ on the last round. $u$ knows that the edge $uv$ is the only edge that has been inserted on the last round, hence $v$ must be a neighbor of $w$. Therefore, $u$ can deduce that the triangle $uvw$ exists. Likewise, $w$ sends $\texttt{LAST} = 1$ to $v$, from which $v$ can deduce that the triangle $uvw$ exists.
It follows that all three nodes can determine that the triangle $uvw$ exists after two rounds of communication, as required.
The above algorithm can be combined with a variant of the algorithm of Corollary \[cor:k3\_mlist\_ub\_me\_mv\], in order to handle node deletions and edge deletions as well. In that algorithm, every node $v$ sends a bit to every neighbor $u$, indicating whether $v$ lost an edge on the current round. We denoted this information by `DELETED`. We can do the same thing here; however, we have the same problem that we encountered in Corollary \[cor:k3\_mlist\_ub\_me\_mv\] — when a node $u$ receives $\texttt{DELETED} = 1$ from two neighbors $v$ and $w$ on the same round, it cannot determine whether the edge between $v$ and $w$ was deleted, or some other node $x$, which was connected to both $v$ and $w$, was deleted. The triangle $uvw$ exists in the latter case, but not in the former case. Therefore, $u$ must be able to distinguish between these two cases.
In Corollary \[cor:k3\_mlist\_ub\_me\_mv\] we solved this problem by having every node $v$ send $\texttt{DELETED} = 1$ to a neighbor $u$ only if it lost a neighbor which is currently connected to $u$. This is possible, because the algorithm of Corollary \[cor:k3\_mlist\_ub\_me\_mv\] guarantees that on every round, $v$ knows the list of all common neighbors of $v$ and $u$.
In our case there is no such guarantee — $u$ will know whether or not its missing neighbor was connected to $v$ only in the next round. However, by using the extra quiet round, $u$ can easily distuinguish between the two problematic cases described above, by checking whether it received $\texttt{DELETED} = 1$ from some other node on the previous round.
Thus, we extend the algorithm of Theorem \[th:k3\_mlist\_ub\_r2\_pe\] as follows. Whenever a node $u$ loses an edge, it sends $\texttt{DELETED} = 1$ to all its neighbors. On the next round, $u$ sends to all its neighbors an indication whether or not it received $\texttt{DELETED} = 1$ from any other neighbor on the previous round. One can verify that this information allows every node to give the correct output after $2$ rounds of communication. Like the original algorithm, this uses $O(1)$ bits of bandwidth on every round, thus the total bandwidth is still $O(1)$ bits. This implies the following corollary, which is clearly optimal.
\[cor:k3\_mlist\_ub\_r2\_pe\_mvme\] The deterministic $2$-round bandwidth complexity of [$\textsf{MemList}(K_3)$]{} under node deletions and edge deletions/insertions is $O(1)$.
Node insertions are harder to handle than the other types of changes. If we are promised $r-1$ quiet rounds, a simple algorithm exists that uses only $O{\left(\frac{n}{r}\right)}$ bits of bandwidth. As we show in Section \[subsubsec:tri-mlist-lower\], this is tight.
\[th:k3\_mlist\_ub\_r\_any\] For every $r$, the deterministic $r$-round bandwidth complexity of [$\textsf{MemList}(K_3)$]{} under any type of change is $O(n/r)$.
The algorithm is as follows: on every round, every node $u$ prepares an $n$-bit message that specifies its current list of neighbors. Then, it breaks this message into $B$-bit blocks, where $B={\left\lceil\frac{n}{r}\right\rceil}$, and sends it to all its neighbors, one block on every round.
Additionally, on every round, $u$ sends to all its neighbors one bit indicating whether or not its list of neighbors has changed on the current round. Whenever the list of neighbors changes, $u$ builds its new list of neighbors, breaks it into blocks, and starts sending it again.
After at most $r-1$ quiet rounds, all nodes are guaranteed to finish sending their list to all their neighbors, thus every node can determine whether any pair of its neighbors are connected to each other or not, as required.
### Lower bounds {#subsubsec:tri-mlist-lower}
The 1-round bandwidth complexities for node and edge deletions are clearly tight. Next, we show that handling edge insertions in $1$ round requires at least $\Omega(\sqrt{n})$ bits of bandwidth. By Theorem \[th:k3\_mlist\_ub\_pe\], this bound is tight. We first prove the following lemma, which shows that a sufficiently dense bipartite graph includes a large enough complete bipartite subgraph. This has the same spirit as the results of Erdös [@Erdos1964], used in [@Abboud2017; @Fischer2018] for bounding the number of single-bit rounds for detecting triangles, but here the sides can be of different sizes.
\[lm:densebip\] For every $\varepsilon \in (0,1)$ there exist $\alpha,\beta,
\gamma \in (0,1)$, such that for every bipartite graph $G=(L \cup R,E)$ having at least $(1-\varepsilon) |L| |R|$ edges, there are subsets $A \subseteq L$ and $B \subseteq R$, whose sizes are $|A| \geq \alpha \cdot |L|$ and $|B| \geq \beta \cdot \gamma^{|L|} \cdot |R|$, such that $uv \in E$ for every $u \in A$ and $v \in B$ (i.e., the induced subgraph on $A \cup B$ is a complete bipartite graph).
Let $\alpha = \frac{1-\varepsilon}{6}$, and let $\mathcal{A}$ be the set of all subsets of $L$ whose size is exactly ${\left\lceil\alpha \cdot |L|\right\rceil}$. For every $A \in \mathcal{A}$, we denote by $N_A$ the set of all vertices in $R$ which are connected to every vertex in $A$. Consider the sum $S = \sum_{A \in \mathcal{A}} |N_A|$. Let $M =
\max \{ |N_A| : A \in \mathcal{A} \}$. Then: [$$\begin{aligned}
\label{eq:mlist_pe_sum_leq}
S
\leq
\sum_{A \in \mathcal{A}} M
=
\binom{|L|}{{\left\lceil\alpha \cdot |L|\right\rceil}} \cdot M\end{aligned}$$]{} On the other hand, the sum $S$ can be computed by counting, for every $v \in R$, the number of sets $A \in \mathcal{A}$ such that $v \in N_A$: [$$\begin{aligned}
S
&=
\sum_{A \in \mathcal{A}} |N_A|
=
\sum_{A \in \mathcal{A}} \sum_{v \in N_A} 1
=
\sum_{v \in R}
\sum_{\substack{A \in \mathcal{A} : \\ v \in N_A}} 1
=
\sum_{v \in R} | \{ A \in \mathcal{A} : v \in N_A \} |\end{aligned}$$]{}
For $p \in (0,1)$, let $k_p$ be the number of vertices in $R$ whose degree is at least $p \cdot |L|$. For every $v \in R$ whose degree is $d(v)$ we have $| \{ A \in \mathcal{A} : v \in N_A \} | =
\binom{d(v)}{{\left\lceil\alpha \cdot |L|\right\rceil}}$. Therefore we can bound the above sum $$S
=
\sum_{v \in R}
\binom
{d(v)}
{{\left\lceil\alpha \cdot |L|\right\rceil}}
\geq
\sum_{\substack{v \in R : \\ d(v) \geq p \cdot |L|}}
\binom
{d(v)}
{{\left\lceil\alpha \cdot |L|\right\rceil}}
\geq
\sum_{\substack{v \in R : \\ d(v) \geq p \cdot |L|}}
\binom
{{\left\lceilp \cdot |L|\right\rceil}}
{{\left\lceil\alpha \cdot |L|\right\rceil}}
=
k_p \cdot
\binom
{{\left\lceilp \cdot |L|\right\rceil}}
{{\left\lceil\alpha \cdot |L|\right\rceil}}
.$$ Combining this with gives $
\binom{|L|}{{\left\lceil\alpha \cdot |L|\right\rceil}} \cdot M
\geq
k_p \cdot
\binom
{{\left\lceilp \cdot |L|\right\rceil}}
{{\left\lceil\alpha \cdot |L|\right\rceil}},
$ which implies: [$$\begin{aligned}
\label{eq:mlist_pe_Mbound}
M
\geq
k_p
\cdot
\binom
{{\left\lceilp \cdot |L|\right\rceil}}
{{\left\lceil\alpha \cdot |L|\right\rceil}}
/
\binom
{|L|}
{{\left\lceil\alpha \cdot |L|\right\rceil}}\end{aligned}$$]{}
We can bound $k_p$ as follows: [$$\begin{aligned}
|E|
&=
\sum_{v \in R} d(v)
=
\sum_{\substack{v \in R : \\ d(v) \geq p \cdot |L|}}
d(v)
+
\sum_{\substack{v \in R : \\ d(v) < p \cdot |L|}}
d(v)
\leq
\sum_{\substack{v \in R : \\ d(v) \geq p \cdot |L|}}
|L|
+
\sum_{\substack{v \in R : \\ d(v) < p \cdot |L|}}
p \cdot |L|
\\ &=
k_p \cdot |L|
+
(|R| - k_p) \cdot p \cdot |L|
=
(1-p) \cdot k_p \cdot |L|
+
p \cdot |L| \cdot |R|\end{aligned}$$]{} On the other hand we have $|E| \geq (1-\varepsilon) \cdot |L| \cdot |R|$, and therefore $
(1-p) \cdot k_p \cdot |L| + p \cdot |L| \cdot |R|
\geq
(1-\varepsilon) \cdot |L| \cdot |R|
$, which implies $k_p \geq \frac{1-\varepsilon-p}{1-p} \cdot |R|$. Setting $p = \frac{1-\varepsilon}{2}$ gives $
k_p \geq \frac{1-\varepsilon}{1+\varepsilon} \cdot |R|
$, and substituting this into gives: [$$\begin{aligned}
M
\geq
\frac
{
1-\varepsilon
}
{
1+\varepsilon
}
\cdot
\frac
{
\binom
{{\left\lceilp \cdot |L|\right\rceil}}
{{\left\lceil\alpha \cdot |L|\right\rceil}}
}
{
\binom
{|L|}
{{\left\lceil\alpha \cdot |L|\right\rceil}}
}
\cdot
|R|\end{aligned}$$]{} Finally, we bound the binomial fraction on the right-hand size as follows. For simplicity, define $
a = |L|,
b = {\left\lceilp \cdot |L|\right\rceil},
c = {\left\lceil\alpha \cdot |L|\right\rceil}
$. Note that $a \geq b \geq c$. Now: [$$\begin{aligned}
\frac
{
\binom{b}{c}
}
{
\binom{a}{c}
}
&=
\frac
{
\prod_{i=1}^c (i+b-c)
}
{
\prod_{i=1}^c (i+a-c)
}
=
\prod_{i=1}^c \frac{i+b-c}{i+a-c}
=
\prod_{i=1}^c (1 - \frac{a-b}{i+a-c})
\\ &\geq
\prod_{i=1}^c (1 - \frac{a-b}{a-c})
=
\prod_{i=1}^c \frac{b-c}{a-c}
=
{\left(\frac{b-c}{a-c}\right)}^c\end{aligned}$$]{} Therefore: [$$\begin{aligned}
\label{eq:mlist_pe_binfrac}
\frac
{
\binom
{{\left\lceilp \cdot |L|\right\rceil}}
{{\left\lceil\alpha \cdot |L|\right\rceil}}
}
{
\binom
{|L|}
{{\left\lceil\alpha \cdot |L|\right\rceil}}
}
\geq
{\left(
\frac
{
{\left\lceilp \cdot |L|\right\rceil}
-
{\left\lceil\alpha \cdot |L|\right\rceil}
}
{
|L|
-
{\left\lceil\alpha \cdot |L|\right\rceil}
}
\right)}^{
{\left\lceil\alpha \cdot |L|\right\rceil}
}\end{aligned}$$]{}
In order to choose appropriate values for $\beta$ and $\gamma$, we now distinguish between two cases for the graph $G$. If $|L| \leq
\frac{1}{\alpha}$, then $\alpha |L| \leq 1$. Since there must be a vertex in $|L|$ having at least $(1-\varepsilon) |R|$ neighbors in $R$, the claim holds for every $\beta \leq 1$ and $\gamma \leq
1-\varepsilon$.
If $|L| > \frac{1}{\alpha}$, we can further develop the right hand side of as follows: [$$\begin{aligned}
{\left(
\frac
{
{\left\lceilp |L|\right\rceil}-{\left\lceil\alpha |L|\right\rceil}
}
{
|L|-{\left\lceil\alpha |L|\right\rceil}
}
\right)}^{
{\left\lceil\alpha |L|\right\rceil}
}
&\geq
{\left(
\frac
{
p |L|-(\alpha |L|+1)
}
{
|L|-\alpha |L|
}
\right)}^{
\alpha |L|+1
}
=
{\left(
\frac
{p-\alpha}
{1-\alpha}
-
\frac
{1}
{(1-\alpha) |L|}
\right)}^{
\alpha |L|+1
}
\\ &>
{\left(
\frac
{p-\alpha}
{1-\alpha}
-
\frac
{1}
{(1-\alpha) \frac{1}{\alpha}}
\right)}^{
\alpha |L|+1
}
=
{\left(
\frac
{p-2\alpha}
{1-\alpha}
\right)}^{
\alpha |L|+1
}
=
{\left(
\frac
{1-\varepsilon}
{5+\varepsilon}
\right)}^{
\alpha |L|+1
}\end{aligned}$$]{}
To sum it all up, we now have $
M
>
{\left(
\frac
{1-\varepsilon}
{5+\varepsilon}
\right)}^{
\alpha \cdot |L|+1
}
$. Recalling the definition of $M$, this inequality implies that there exists $A \subseteq L$ whose size is exactly ${\left\lceil\alpha \cdot |L|\right\rceil}$, such that $
|N_A|
>
{\left(
\frac
{1-\varepsilon}
{5+\varepsilon}
\right)}^{
\alpha \cdot |L|+1
}
$, i.e., there are more than $
{\left(
\frac
{1-\varepsilon}
{5+\varepsilon}
\right)}^{
\alpha \cdot |L|+1
}
$ vertices in $R$ which are connected to every vertex in $A$. Therefore, the claim holds for every $\beta$ and $\gamma$ such that: $
\beta
\leq
\frac
{1-\varepsilon}
{5+\varepsilon},
\gamma
\leq
{\left(
\frac
{1-\varepsilon}
{5+\varepsilon}
\right)}^{\alpha}
=
{\left(
\frac
{1-\varepsilon}
{5+\varepsilon}
\right)}^{
\frac
{1-\varepsilon}
{6}
}
$. Thus, given $\varepsilon \in (0,1)$, the following values of $\alpha,
\beta,\gamma$ satisfy the claim for every $G$: [$$\begin{aligned}
\alpha =
\frac
{1-\varepsilon}
{6},
\beta =
\frac
{1-\varepsilon}
{5+\varepsilon},
\gamma =
\min {\left\{
1-\varepsilon,
{\left(
\frac
{1-\varepsilon}
{5+\varepsilon}
\right)}^{
\frac
{1-\varepsilon}
{6}
}
\right\}}
.
\qedhere\end{aligned}$$]{}
\[th:k3\_mlist\_lb\_pe\] The randomized $1$-round bandwidth complexity of [$\textsf{MemList}(K_3)$]{} under edge insertions is $\Omega(\sqrt{n})$.
Let $A$ be a (randomized) $1$-round algorithm that solves [$\textsf{MemList}(K_3)$]{} under edge insertions using bandwidth $B$ with error probability $\varepsilon$. Let $t$ be a parameter to be defined later, and consider a tripartite graph with $n$ nodes as in Figure \[fig:tri\].
+=\[inner sep=0pt\]
(20,-30) ellipse (6 and 14); (40,-30) ellipse (3 and 8);
(20,-10) node [$W$]{}; (20,-20) node [$w_1$]{}; (20,-25) node [$\vdots$]{}; (20,-30) node [$\vdots$]{}; (20,-35) node [$\vdots$]{}; (20,-40) node [$w_{n-t-1}$]{};
(40,-10) node [$U$]{}; (40,-25) node [$u_1$]{}; (40,-30) node [$\vdots$]{}; (40,-35) node [$u_t$]{};
(60,-30) node [$v$]{};
(59,-30) – (42.3,-25); (59,-30) – (43,-30); (59,-30) – (42.3,-35); (37,-30) – (26,-30); (60,-31) .. controls (60,-54) and (0,-54) .. (14,-30);
Let $\mathcal{C}$ be the set of all possible bipartite graphs with vertex sets $W$ and $U$. Note that $|\mathcal{C}| = 2^{t(n-t-1)}$. For every $C \in \mathcal{C}$ and every $w \in W$, we define a sequence of changes $S_{C,w}$ as follows. We start with no edges. Then, we insert edges between $U$ and $W$ to get the bipartite graph $C$. During the next $t$ rounds, we connect $v$ to every $u \in U$, one by one. Finally, we insert the edge $vw$. In the end of $S_{C,w}$, after $1$ additional round of communication, node $v$ must output a list of all triangles containing $v$. By construction, this list is uniquely determined by the set of neighbors of $w$ in $U$. Thus, node $v$ needs to know the set of all neighbors of $w$ after $1$ additional round of communication.
We assume that the output of $v$ is correct with probability at least $1-\varepsilon$. Therefore, by Yao’s lemma, there exists a deterministic algorithm $A'$ that solves the same problem correctly for at least $1-\varepsilon$ of all inputs. We define a bipartite graph with node sets $\mathcal{C}$ and $W$, such that the edge $w-C$ exists if and only if the output of $v$ is correct for the sequence $S_{C,w}$.
We have $|W| = n-t-1$ and $|\mathcal{C}| = 2^{t(n-t-1)}$. Also, by assumption, this graph contains at least $(1-\varepsilon)$ of all possible edges. Therefore, by Lemma \[lm:densebip\], there exists $W_1
\subseteq W$ of size at least $\alpha \cdot (n-t-1)$ and $\mathcal{C}_1
\subseteq \mathcal{C}$ of size at least $\beta \cdot \gamma^{n-t-1}
\cdot 2^{t(n-t-1)}$, for some $\alpha,\beta,\gamma \in (0,1)$ that depend only on $\varepsilon$, such that the output of $v$ is correct for the sequence $S_{C,w}$ for every $C \in \mathcal{C}_1$ and every $w \in
W_1$.
Now, consider the input of node $v$ during any sequence $S_{C,w}$. During the first stage of building the bipartite graph $C$, $v$ is isolated and receives no input. Then, it receives a set of messages from the nodes in $U$, and finally one additional message from $w$. Note that the messages $v$ receives from the nodes in $U$ depend only on $C$, and not on $w$, since the nodes in $U$ cannot know the identity of $w$ until the final round of communication.
Now, on every round, every node in $U$ can send to $v$ any of $2^B$ possible messages. Altogether, during the entire sequence, the nodes of $U$ send to $v$ a set of $\frac{t(t+3)}{2}$ messages. Hence, the number of possible inputs from the nodes of $U$ to $v$ is $
2^{
B \cdot \frac{t(t+3)}{2}
}
$. Therefore, there exists $\mathcal{C}_2 \subseteq \mathcal{C}_1$ of size at least $
2^{
- B \cdot \frac{t(t+3)}{2}
}
\cdot
|\mathcal{C}_1|
$, such that $v$ receives the same input from all nodes in $U$ for every sequence $S_{C,w}$ for $C \in \mathcal{C}_2$ and every $w \in W_1$. Thus, we have: $$\begin{aligned}
\label{eq:mlist_ae_lb_1}
|\mathcal{C}_2|
&\geq
2^{
- B \frac{t(t+3)}{2}
}
|\mathcal{C}_1|
\geq
2^{
- B \frac{t(t+3)}{2}
}
\beta
\gamma^{n-t-1}
2^{t(n-t-1)}
=
\beta
\gamma^{n-t-1}
2^{
t(n-t-1) - B \frac{t(t+3)}{2}
}\end{aligned}$$
On the other hand, we can bound the size of $\mathcal{C}_2$ by considering the number of possible neighbors of $w$ in any $C \in
\mathcal{C}_2$, for every $w \in W$. Recall that during the sequence $S_{C,w}$, $v$ receives only one message from $w$. Also, for every $w
\in W_1$, $v$ must determine the set of all neighbors of $w$ in $U$. Since there are only $2^B$ possible inputs $v$ can receive from $w$, every $w \in W_1$ can have at most $2^B$ possible sets of neighbors in any $C \in \mathcal{C}_2$.
Every $w \in W \setminus W_1$ can have any subset of $U$ as its set of neighbors, hence it can have at most $2^t$ possible sets of neighbors. Now, since every $C \in \mathcal{C}_2$ is uniquely determined by set of neighbors of every $w \in W$, we have: $$\begin{aligned}
\label{eq:mlist_ae_lb_2}
|\mathcal{C}_2|
\leq
{\left(
\prod_{w \in W_1} 2^B
\right)}
{\left(
\prod_{w \in W \setminus W_1} 2^t
\right)}
=
2^{
B \cdot |W_1|
}
\cdot
2^{
t \cdot (|W| - |W_1|)
}
=
2^{
(B-t) \cdot |W_1|
+
t(n-t-1)
}\end{aligned}$$ Combining and gives: $
\beta
\cdot
\gamma^{n-t-1}
\cdot
2^{
t(n-t-1) - B \cdot \frac{t(t+3)}{2}
}
\leq
2^{
(B-t) \cdot |W_1| + t(n-t-1)
}
$, and setting $t = {\left\lceil\sqrt{n}\right\rceil}$ gives, with some algebraic manipulations, $B \geq \Omega(\sqrt{n})$.
Finally, for node insertions, we show that every $r$-round algorithm must use at least $\Omega(n/r)$ bits of bandwidth, which is tight by Theorem \[th:k3\_mlist\_ub\_r\_any\].
\[th:k3\_mlist\_lb\_r\_pv\] For every $r$, the randomized $r$-round bandwidth complexity of [$\textsf{MemList}(K_3)$]{} under node insertions is $\Omega{\left(\frac{n}{r}\right)}$.
Let $A$ be a (randomized) $r$-round algorithm that solves [$\textsf{MemList}(K_3)$]{} under node insertions using bandwidth $B$ with error probability $\varepsilon$. We show that $r \cdot B = \Omega(n)$.
Let $u$ be any node, and let $\mathcal{C}$ be the set of all possible graphs on the other $n-1$ nodes. For every $C \in \mathcal{C}$ we define the sequence $S_C$ as follows:
- Start with an empty graph (no nodes and no edges).
- During $n-1$ rounds, insert one node on each round and connect it to the nodes which have been already inserted, according to the edges in $C$. After $n-1$ rounds, the graph is identical to the graph $C$.
- On round $n$ insert $u$ and connect to all the other nodes.
After $r-1$ quiet rounds $u$ needs to output the list of all triangles that contain $u$. Since $u$ is connected to all the other nodes, this implies that $u$ needs to know the graph $C$. For every $C \in
\mathcal{C}$, the output of $u$ is guaranteed to be correct for the sequence $S_C$ with probability at least $(1-\varepsilon)$. Therefore, by Yao’s lemma, there exists a deterministic algorithm, $A'$, that guarantees that the output of $u$ is correct for at least $(1-\varepsilon)$ of all sequences. That is, there exists a subset $\mathcal{C}_1 \subseteq \mathcal{C}$, whose size is at least $(1-\varepsilon) \cdot |\mathcal{C}|$, such that $A'$ guarantees the correct output of $u$ for sequences $S_C$ for all $C \in \mathcal{C}_1$.
Now, the number of possible graphs on $n-1$ nodes is $2^{\binom{n-1}{2}}$, hence the size of $\mathcal{C}_1$ is at least $(1-\varepsilon) \cdot 2^{\binom{n-1}{2}}$. Since $A'$ is deterministic, and $u$ needs to distinguish correctly between all possible $C \in
\mathcal{C}_1$, this implies that the input $u$ receives from its neighbors must have at least $(1-\varepsilon) \cdot 2^{\binom{n-1}{2}}$ possible values. Every neighbor of $u$ can send any of $2^B$ messages on every round, thus the number of possible inputs $u$ can receive on a single round is $2^{B \cdot (n-1)}$. Therefore, during $r$ rounds of communication, the number of possible inputs to $u$ is $2^{r \cdot B \cdot (n-1)}$.
Combining the above we get that $
2^{
r \cdot B \cdot (n-1)
}
\geq
(1-\varepsilon) \cdot 2^{\binom{n-1}{2}}
$, which can be simplified to $
r \cdot B
\geq
\frac
{n}
{2}
+
\frac
{\log (1-\varepsilon)}
{n-1}
$, and it follows that $r \cdot B = \Omega(n)$.
Membership detection
--------------------
Table \[tbl:k3\_mdtct\] summarizes the results of this subsection.
Node deletions Edge deletions Edge insertions Node insertions
------------ ---------------- ---------------- ----------------- -----------------
$r = 1$ $0$ $\Theta(1)$ $O(\log n)$ $\Theta(n)$
$r \geq 2$ $0$ $\Theta(1)$ $\Theta(1)$ $\Theta(1)$
: Bandwidth complexities of [$\textsf{MemDetect}(K_3)$]{} []{data-label="tbl:k3_mdtct"}
### Upper bounds {#upper-bounds}
The upper bounds for node deletions and edge deletions follow from Observation \[obs:probrel\]. The following shows that edge insertions can be handled with $O(\log n)$ bits of bandwidth.
\[th:k3\_mdtct\_ub\_pe\] The deterministic $1$-round bandwidth complexity of [$\textsf{MemDetect}(K_3)$]{} under edge insertions is $O(\log n)$.
The algorithm works as follows. On every round, every node sends to all its neighbors an indication whether or not it got a new neighbor on the current round, along with the ID of the new neighbor (if any). We denote this information by `NEWID`. Note that with just this information, for every pair of adjacent edges $u-v-w$, at least one of $u$ and $w$ know that these two edges exist (specifically, the first one connected to $v$ knows that the other one is also connected to $v$).
Additionally, every node $u$ sends to every neighbor $v$ one bit, indicating whether $u$ knows that $v$ is part of some triangle. We denote this bit by `ACCEPT`.
Suppose the edge $uv$ is inserted and creates at least one triangle $uvw$. As explained above, in this case at least one of $u$ and $v$ knows that the triangle exists, therefore it sends $\texttt{ACCEPT} = 1$ to the two other nodes. Thus all three nodes know that they are part of a triangle. It follows that on each round all nodes can determine whether they are in a triangle.
We can extend the algorithm of Theorem \[th:k3\_mdtct\_ub\_pe\] to handle node/edge deletions as well. Now, in addition to `NEWID`, every node sends to all its neighbors an indication whether or not it has lost a neighbor on the current round, along with the ID of the lost neighbor (if any). Also, every node $u$ sends every neighbor $v$ one additional bit, indicating whether $u$ received $\texttt{NEWID} = v$ from any other node on the last round. We denote this bit by `LAST`.
Now, suppose that on round $t$ the edge $uv$ is inserted and creates a triangle $uvw$. As in Theorem \[th:k3\_mdtct\_ub\_pe\], after one round of communication all three nodes can determine that they are part of a triangle. Moreover, both $u$ and $v$ send the appropriate value of `NEWID` to all their neighbors. Then, on round $t+1$, every common neighbor of $u$ and $v$ sends to both of them $\texttt{LAST} = 1$. From this information, $u$ and $v$ can deduce the list of all of their common neighbors.
It follows that on every round, every node knows the exact set of triangles that contained it on the previous round. This allows every node to handle node/edge deletions appropriately, and to determine when all triangles that contained it are gone.
\[cor:k3\_mdtct\_ub\_pe\_mvme\] The deterministic $1$-round bandwidth complexity of [$\textsf{MemDetect}(K_3)$]{} under node deletions and edge deletions/insertions is $O(\log n)$.
If we only consider sequences of graphs with bounded degree $\Delta$, and only allow edge insertions, the complexity of Theorem \[th:k3\_mdtct\_ub\_pe\] can be improved to $O(\sqrt{\Delta \log n})$, using an algorithm similar to that of Theorem \[th:k3\_mlist\_ub\_pe\], but with $O(\sqrt{\Delta \log n})$ bits instead of $O(\sqrt{n})$, as follows. On every round, every node sends to all its neighbors an indication whether or not it has a new neighbor. We denote this information by `NEW`. Additionally, whenever a new edge $uv$ is inserted, the following happens:
- $u$ sends to $v$ an indication whether or not it knows about a triangle that contains $v$.
- For every round $i$ within the last ${\left\lceil\sqrt{\Delta \log n}\right\rceil}$ rounds, $u$ sends to $v$ an indication whether or not it has had a new neighbor on round $i$.
- For every round $i$ within the last ${\left\lceil\sqrt{\Delta \log n}\right\rceil}$ rounds, $u$ sends to $v$ an indication whether or not any of its neighbors has sent $\texttt{NEW} = 1$ on round $i$.
- $u$ computes the list of IDs of all its current neighbors, and starts sending it to $v$, ${\left\lceil\sqrt{\Delta \log n}\right\rceil}$ bits on every round. Since $u$ has at most $\Delta$ neighbors, after $O(\sqrt{\Delta \log n})$ rounds $v$ has the complete list. Note that by the time this process completes, the list may not be up-to-date.
All of this requires a $O(\sqrt{\Delta \log n})$ bandwidth, and it can be shown, similarly to Theorem \[th:k3\_mlist\_ub\_pe\], that this allows every node to give the correct output in all cases.
For a quiet round we have an upper bound of $O(1)$ bits from Observation \[obs:probrel\], Theorem \[th:k3\_mlist\_ub\_r2\_pe\], and Corollary \[cor:k3\_mlist\_ub\_r2\_pe\_mvme\].
Finally, for node insertions, we have a trivial upper bound of $O(n)$ bits. For a quiet round, the next theorem shows an upper bound of $O(1)$ bits.
\[th:k3\_mdtct\_ub\_r2\_pv\] The deterministic $2$-round bandwidth complexity of [$\textsf{MemDetect}(K_3)$]{} under node insertions is $O(1)$.
On every round, every node $u$ sends to every neighbor $v$ two bits:
- One bit, denoted by `NEW`, indicates whether or not $u$ has a new neighbor which has been inserted on the current round.
- One bit, denoted by `LAST`, indicates whether or not any neighbor of $u$ other than $v$ has sent $u$ $\texttt{NEW} = 1$ on the last round.
Consider a triangle $uvw$ and assume that $w$ is the last node inserted. When $w$ is inserted, $u$ and $v$ send $\texttt{NEW} = 1$ to each other. Since only one node can be inserted on a single round, both $u$ and $v$ know that $w$ is a common neighbor, hence the triangle $uvw$ exists.
On the next round, both $u$ and $v$ send $\texttt{LAST} = 1$ to $w$. This implies that some neighbor of $u$ (resp. $v$) has sent $\texttt{NEW} = 1$ on the last round, i.e., some neighbor of $u$ (resp. $v$) has had a new neighbor inserted on the last round. Since only one node can be inserted on a single round, $w$ can determine that the new neighbor must be $w$ itself, that is, some neighbor of $u$ (resp. $v$) is now also a neighbor of $w$. Hence $w$ can determine that it is part of at least one triangle, and output $1$ as required.
The above algorithm can be executed in parallel with the algorithm of Corollary \[cor:k3\_mlist\_ub\_r2\_pe\_mvme\] to handle all four types of changes. However, we now have a problem combining node insertions and edge insertions, just as we had a problem in Corollary \[cor:k3\_mlist\_ub\_r2\_pe\_mvme\] when we combined edge deletions and node deletions. The problem is that when a node $v$ has a new neighbor $w$, it does not know whether the edge $vw$ was inserted on the current round, or $w$ is a new node which was inserted on the current round and may be connected to other nodes as well. For example, assume $u$ has two neighbors, $v$ and $w$, and the edge $vw$ does not exist. Then, either the edge $vw$ is inserted, or a new node $x$ is inserted and connected to both $v$ and $w$. In both cases, $v$ and $w$ send $\texttt{NEW} = 1$ to $u$, and $u$ cannot distinguish between the former case (in which it should output 1), and the latter case (in which it should output 0).
To solve this problem, we can have every new node indicate to all its neighbors that it is new. Thus, whenever a node $u$ has a new neighbor $x$, after one round of communication $u$ can determine whether $x$ is a new node or not and send this information to all its neighbors. This allows all nodes to distinguish between the two problematic cases we described, and give the correct output in all cases.
\[cor:k3\_mdtct\_ub\_r2\_pv\_mvmepe\] The deterministic $2$-round bandwidth complexity of [$\textsf{MemDetect}(K_3)$]{} under node/edge deletions/insertions is $O(1)$.
### Lower bounds {#lower-bounds}
For node deletions and edge deletions, we have trivial constant lower bounds, which are tight, as shown above. For edge insertions, we have shown a general algorithm that uses $O(\log n)$ bits, and also an algorithm that uses $O(\sqrt{\Delta \log n})$ bits for graphs with bounded degree $\Delta$. The latter algorithm implies that showing a lower bound of $B$ bits for the bandwidth complexity of this problem would require looking at sequences of graphs with degree at least $\Omega(\frac{B^2}{\log n})$. In particular, in order to show that the algorithm of Theorem \[th:k3\_mdtct\_ub\_pe\] is optimal, one has to consider sequences of graphs with degree at least $\log n$. Other than that, it remains an open question whether or not the algorithm of Theorem \[th:k3\_mdtct\_ub\_pe\] can be improved.
What is the $1$-round bandwidth complexity of [$\textsf{MemDetect}(K_3)$]{} under edge insertions?
Finally, for node insertions, the following theorem shows a lower bound of $\Omega(n)$ bits.
\[th:k3\_mdtct\_lb\_pv\] The randomized $1$-round bandwidth complexity of [$\textsf{MemDetect}(K_3)$]{} under node insertions is $\Omega(n)$.
Let $A$ be a (randomized) $1$-round algorithm that solves [$\textsf{MemDetect}(K_3)$]{} under node insertions using bandwidth $B$ with error probability $\varepsilon$. Fix $x \in V$, let $U = V \setminus \{x\}$, and let $\mathcal{C}$ be the set of all possible graphs on the nodes of $U$. For every $C \in \mathcal{C}$ and every $u,v \in U$, define the sequence $S_{C,u,v}$ as follows:
- Start with an empty graph (no nodes and no edges).
- During $n-1$ rounds, insert one node of $U$ on each round and connect it to the nodes which have been already inserted, according to the edges in $C$. After $n-1$ rounds, the graph is identical to the graph $C$.
- On round $n$ insert $x$ and connect it to $u$ and $v$.
Then, $x$ should output 1 iff the edge $uv$ exists in $C$. By our assumption, for every $C$ and every $u,v \in U$, the output is correct with probability at least $1-\varepsilon$. Note that during the final round, $x$ receives only one message from each of $u$ and $v$. Also, during this final round, $u$ and $v$ do not know the identity of the other neighbor of $x$. Hence, the message received from $u$ depends only on the identity of $u$ and the graph $C$, and not on the identity of $v$, and the message received from $v$ depends only on the identity of $v$ and the graph $C$.
Now, consider the following experiment for a given graph $C \in
\mathcal{C}$. First, we run the sequence $S_{C,u,v}$ for some $u,v \in
U$, and stop just before the last round, in which $x$ is connected to $u$ and $v$. Then, every node $w \in U$ generates a message to be sent to $x$, as if it has been connected to $x$ on the final round. For every $w \in U$, let $m_w$ be the message generated by $w$ (note that $m_w$ is a random variable).
Note again, that the messages generated by the nodes of $U$ depend only on the graph $C$, and not on the other nodes that may have been connected to $x$ on the last round. Therefore, for every $u,v \in U$, given the messages generated by $u$ and $v$, $x$ should be able to determine, with probability at least $1-\varepsilon$, whether or not the edge $uv$ exists in $C$.
Next, for every pair of nodes $u,v \in U$, let $I_{uv}$ be the output of $x$ given the two messages $m_u$ and $m_v$. Note that for nodes $u,v,w$, the variables $I_{uv}$ and $I_{uw}$ are not necessarily independent. Now, let $C'$ be the graph on the nodes of $U$, in which the edge $uv$ exists if and only if $I_{uv}=1$. We consider $C'$ to be the result of the experiment.
Let $p_C$ denote the probability that at the end of the experiment we have $C'=C$. Since, for every $u,v \in U$, the value of $I_{uv}$ corresponds to the edge $uv$ in the graph $C$ with probability at least $1-\varepsilon$, we have $
p_C \geq (1-\varepsilon)^{\binom{n-1}{2}}
$. Summing the above for every $C \in \mathcal{C}$ we get: $$\begin{aligned}
\label{eq:mdtct_pv_geq}
\sum_C p_C
\geq
|\mathcal{C}| \cdot (1-\varepsilon)^{\binom{n-1}{2}}
=
(2-2\varepsilon)^{\binom{n-1}{2}}\end{aligned}$$
On the other hand, consider the set of messages generated by the nodes of $U$ at the end of the first stage of the experiment. For every possible set of messages $M$ generated by the nodes of $U$ during the first stage of the experiment, denote by $\phi_C(M)$ the probability for generating exactly the messages of $M$. Also, for every $C' \in
\mathcal{C}$, let $\Psi_M(C')$ denote the probability for the result of the experiment to be equal to $C'$, given the set of generated messages $M$. We have: [$$\begin{aligned}
\sum_C p_C
=
\sum_C \sum_M \phi_C(M) \cdot \Psi_M(C)
\leq
\sum_C \sum_M \Psi_M(C)
=
\sum_M \sum_C \Psi_M(C)
=
\sum_M 1
=
|\mathcal{M}|\end{aligned}$$]{}
where $\mathcal{M}$ is the set of all possible values of $M$. Since every message has exactly $2^B$ possible values, the number of possible sets of $(n-1)$ messages is $
|\mathcal{M}| = 2^{B(n-1)}
$, and hence $
\sum_C p_C \leq 2^{B(n-1)}
$. Combining this with gives $
2^{B(n-1)}
\geq
(2-2\varepsilon)^{\binom{n-1}{2}}
$. After some simplifications we get the desired bound: $$B
\geq
\log (2-2\varepsilon) \cdot \frac{n-2}{2}
=
\Omega(n)
\qedhere$$
Listing and detection
---------------------
Table \[tbl:k3\_listdtct\] summarizes the results of this subsection.
Node deletions Edge deletions Edge insertions Node insertions
---------------- ---------------- ----------------- -----------------
$0$ $\Theta(1)$ $\Theta(1)$ $\Theta(1)$
: Bandwidth complexities of [$\textsf{List}(K_3)$]{} and [$\textsf{Detect}(K_3)$]{} []{data-label="tbl:k3_listdtct"}
### Upper bounds {#upper-bounds-1}
The upper bounds for node deletions and edge deletions follow from Observation \[obs:probrel\] and the results of the last subsection.
For edge insertions, the algorithm of Theorem \[th:k3\_mlist\_ub\_r2\_pe\], which solves [$\textsf{MemList}(K_3)$]{} in $2$ rounds, guarantees that on every round, every triangle is known to at least one of its nodes. Hence, it can be used to solve [$\textsf{List}(K_3)$]{} in $1$ round.
\[cor:k3\_list\_ub\_pe\] The deterministic $1$-round bandwidth complexity of [$\textsf{List}(K_3)$]{} under edge insertions is $O(1)$.
The same is true for the algorithm of Corollary \[cor:k3\_mlist\_ub\_r2\_pe\_mvme\], where we also handle node/edge deletions. Therefore:
\[cor:k3\_list\_ub\_pe\_mvme\] The deterministic $1$-round bandwidth complexity of [$\textsf{List}(K_3)$]{} under node deletions and edge deletions/insertions is $O(1)$.
Considering node insertions, the problem can be solved with $O(1)$ bits of bandwidth.
\[th:k3\_list\_ub\_pv\] The deterministic $1$-round bandwidth complexity of [$\textsf{List}(K_3)$]{} under node insertions is $O(1)$.
The algorithm is as follows: on every round, every node sends to each one of its neighbors an indication whether or not it got a new neighbor on the current round. Thus, whenever a node $w$ is inserted and creates a triangle $uvw$, $u$ receives this indication from $v$ (and vice versa). Since $u$ also got connected to $w$ on the current round, and since only one node can be inserted on every round, $u$ can deduce that the triangle $uvw$ exists.
Note that $v$ also receives the same indication from $u$, and can deduce that the triangle $uvw$ exists. Thus, this algorithm guarantees that every triangle is known to at least two of its nodes.
In order to handle node deletions and edge deletions, we can combine the above algorithm with the technique used in Corollary \[cor:k3\_mlist\_ub\_r2\_pe\_mvme\] to distinguish between the deletion of an edge and the deletion of a node. This technique ensures that when either a noe or an edge is delete on round $t$, all incident nodes can determine, on round $t+1$, whether an edge or a node was deleted. This, combined with the guarantee that every triangle is known to at least two of its nodes, ensures that on every round at least one node can determine whether the triangle still exists, even if a node or an edge was deleted on the current round. Thus, listing all triangles is possible on every round, as required.
\[cor:k3\_list\_ub\_pv\_mvmepe\] The deterministic $1$-round bandwidth complexity of [$\textsf{List}(K_3)$]{} under node/edge deletions/insertions is $O(1)$.
By Observation \[obs:probrel\], all upper bounds shown here for [$\textsf{List}(K_3)$]{} also hold for [$\textsf{Detect}(K_3)$]{}.
### Lower bounds {#lower-bounds-1}
It is not clear whether it is possible to handle both node insertions and edge insertions with $O(1)$ bits of bandwidth. The algorithm of Corollary \[cor:k3\_mdtct\_ub\_pe\_mvme\] can be used to handle all four types of changes with $O(\log n)$ bits. Other than that, the exact bandwidth complexity of this combination is an open question.
What is the $1$-round bandwidth complexity of [$\textsf{List}(K_3)$]{} under node/edge insertions?
Larger cliques
==============
Throughout this section, let $s \geq 4$ be some constant, and denote by $K_s$ a clique on $s$ nodes.
Membership listing
------------------
Table \[tbl:ks\_mlist\] summarizes the results of this subsection.
Node deletions Edge deletions Edge insertions Node insertions
------------ ---------------- ---------------- -------------------- -----------------
$r = 1$ $0$ $\Theta(1)$ $\Theta(\sqrt{n})$ $\Theta(n)$
$r \geq 2$ $0$ $\Theta(1)$ $\Theta(1)$ $\Theta(n/r)$
: Bandwidth complexities of [$\textsf{MemList}(K_s)$]{} []{data-label="tbl:ks_mlist"}
### Upper bounds {#upper-bounds-2}
For every node $u$, the set of all cliques that contain $u$ is completely determined by the set of all triangles that contain $u$. Therefore, every algorithm that solves [$\textsf{MemList}(K_3)$]{} can be used to solve [$\textsf{MemList}(K_s)$]{} under the same set of changes. All upper bounds follow thus from the upper bounds for [$\textsf{MemList}(K_3)$]{} given in Section \[subsubsec:tri-mlist-upper\].
### Lower bounds {#lower-bounds-2}
For node deletions and edge deletions we have the trivial lower bounds of $0$ and $\Omega(1)$, respectively, both of which are tight.
The lower bound for edge insertions is shown next.
\[th:ks\_mlist\_lb\_pe\] The randomized $1$-round bandwidth complexity of [$\textsf{MemList}(K_s)$]{} under edge insertions is $\Omega(\sqrt{n})$.
The proof is almost identical to the proof of Theorem \[th:k3\_mlist\_lb\_pe\], with only one difference: instead of starting with an empty tripartite graph, we add a clique on $s-3$ nodes, which are connected to every node in the graph except $v$. The initial graph now looks like this:
+=\[inner sep=0pt\]
(20,-30) ellipse (6 and 14); (20,-10) node [$W$]{}; (20,-20) node [$w_1$]{}; (20,-25) node [$\vdots$]{}; (20,-30) node [$\vdots$]{}; (20,-35) node [$\vdots$]{}; (20,-40) node [$w_{n-t-s+2}$]{};
(40,-30) ellipse (3 and 8); (40,-10) node [$U$]{}; (40,-25) node [$u_1$]{}; (40,-30) node [$\vdots$]{}; (40,-35) node [$u_t$]{};
(60,-30) node [$v$]{};
(36,-47) circle (4); (36,-47) node [$K$]{};
(59,-30) – (42.3,-25); (59,-30) – (43,-30); (59,-30) – (42.3,-35); (37,-30) – (26,-30); (60,-31) .. controls (60,-64) and (0,-64) .. (14,-30);
(39.10,-44.45) – (59.28,-30.51); (36.92,-43.09) – (38.4,-36.8); (33.28,-44.11) – (25.44,-35.78);
with edges between every node in $K$ and every other node except $v$.
As in Theorem \[th:k3\_mlist\_lb\_pe\], for every bipartite graph $C$ between the nodes of $U$ and $W$ and every node $w \in W$, we define a sequence of changes $S_{C,w}$ as follows. We start with the initial graph defined above. Then, we insert edges between $U$ and $W$ to get the bipartite graph $C$. Then, during the next $t+s$ rounds, we connect $v$ to every node in $U \cup K$, one by one. Finally, we insert the edge $vw$. Note that now, since $v$ is connected to all nodes in $U$ and $K$, the set of instances of $K_s$ that contain $v$ is uniquely determined by the set of neighbors of $w$ in $U$. Therefore, after a single round of communication, $v$ should be able to determine the set of all neighbors of $w$ in $U$.
The proof proceeds as in Theorem \[th:k3\_mlist\_lb\_pe\]. The size of $W$ is now $n-t-s+2$, and the number of messages that $v$ receives from the nodes of $U$ and $K$ is now $\frac{(t+s)(t+s+3)}{2}$. We end up with the following inequality, where $|W_1| \geq \alpha |W|$ for some constants $\alpha,\beta,\gamma \in (0,1)$: [$$\begin{aligned}
\beta
\cdot
\gamma^{n-t-s+2}
\cdot
2^{
t(n-t-s+2)
-
B \cdot \frac{(t+s)(t+s+3)}{2}
}
&\leq
2^{(B-t) \cdot |W_1| + t(n-t-s+2)}\end{aligned}$$]{}
Setting $t = {\left\lceil\sqrt{n}\right\rceil}$ gives, as in Theorem \[th:k3\_mlist\_lb\_pe\], $B \geq \Omega(\sqrt{n})$.
For $r \geq 2$, since edge insertions require communication, we have a trivial lower bound of $\Omega(1)$ bits.
For node insertions, we again use a similar construction to the one we used for triangles to show a lower bound of $\Omega(n/r)$ bits.
\[th:ks\_mlist\_lb\_r\_pv\] For every $r$, the randomized $r$-round bandwidth complexity of [$\textsf{MemList}(K_s)$]{} under node insertions is $\Omega{\left(\frac{n}{r}\right)}$.
The proof is almost identical to Theorem \[th:k3\_mlist\_lb\_r\_pv\], with the only difference being that we limit the possible graphs that are constructed on the first $n-1$ rounds to a certain family of graphs.
Fix $x \in V$, and fix a set of $s-3$ other nodes $K \subseteq
V \setminus \{x\}$. Let $U = V \setminus (K \cup \{x\})$ be the set of all other nodes, and let $U = L \cup R$ be a partition of $U$ into two subsets of equal size (if $|U|$ is odd, let $|L| = |R| + 1$). Let $\mathcal{C}$ be the set of all possible bipartite graphs on the sets $L$ and $R$. For every $C \in \mathcal{C}$ we define a sequence of changes $S_C$ as follows:
- Start with an empty graph (no nodes and no edges).
- During $n-(s-2)$ rounds, insert one node of $U$ on each round and connect it to the nodes which have been already inserted, according to the edges in $C$. After $n-(s-2)$ rounds, the graph is identical to the graph $C$.
- During the next $s-3$ rounds, insert one node of $K$ on each round and connect it to all other nodes in the graph. After this stage, the graph consists of the graph $C$, a clique $K$, and all possible edges between the nodes of $C$ and the nodes of $K$.
- On round $n$ insert $x$ and connect it to all other nodes in the graph.
After one round of communication, $x$ should be able to determine the set of all isntances of $K_s$ that contain it. Note that, for $C,C' \in
\mathcal{C}$, if an edge $uv$ exists in $C$ but not in $C'$ then the set $\{x,u,v\} \cup K$ is a clique in the final graph of $S_C$, but not in the final graph of $S_{C'}$. Therefore, the correct output of $x$ is different for every two distinct $C,C' \in \mathcal{C}$.
The proof proceeds as in Theorem \[th:k3\_mlist\_lb\_r\_pv\]. The number of possible sequences is now $
2^{|L| \cdot |R|}
=
2^{
{\left\lfloor\frac{n-s+2}{2}\right\rfloor}
\cdot
{\left\lceil\frac{n-s+2}{2}\right\rceil}
}
$, and we end up with the following inequality: [$$\begin{aligned}
2^{r \cdot B \cdot (n-1)}
\geq
(1-\varepsilon)
\cdot
2^{
{\left\lfloor\frac{n-s+2}{2}\right\rfloor}
\cdot
{\left\lceil\frac{n-s+2}{2}\right\rceil}
}\end{aligned}$$]{}
After some simplifications we obtain $r \cdot B = \Omega(n)$, as claimed.
Membership detection
--------------------
Table \[tbl:ks\_mdtct\] summarizes the results of this subsection.
Node deletions Edge deletions Edge insertions Node insertions
------------ ---------------- ---------------- ----------------- -----------------
$r = 1$ $0$ $\Theta(1)$ $O(\sqrt{n})$ $\Theta(n)$
$r \geq 2$ $0$ $\Theta(1)$ $\Theta(1)$ $\Theta(1)$
: Bandwidth complexities of [$\textsf{MemDetect}(K_s)$]{} []{data-label="tbl:ks_mdtct"}
### Upper bounds {#upper-bounds-3}
All upper bounds follow from Observation \[obs:probrel\] and the results of the last subsection, except for the upper bound of $O(1)$ in $2$ rounds under node insertions, which can be solved by the same algorithm used in Theorem \[th:k3\_mdtct\_ub\_r2\_pv\] and its corollary for triangles.
### Lower bounds {#lower-bounds-3}
The lower bounds for node deletions and edge deletions are trivial.
As for edge insertions, it is not clear whether the upper bound of $O(\sqrt{n})$ is tight, and in general, what is the exact bandwidth complexity of the problem.
Note that the algorithm of Theorem \[th:k3\_mdtct\_ub\_pe\], which solves [$\textsf{MemDetect}(K_3)$]{} with $O(\log n)$ bits of bandwidth, does not work for larger cliques. To understand why, consider an almost-clique on $s$ nodes, with one missing edge $uv$. That is, inserting the edge $uv$ would create a clique on $s$ nodes.
For triangles, this is simply a node which is connected to both $u$ and $v$. In this case, we can guarantee that at least one of the nodes $u$ and $v$ knows that this construct exists. Thus, when the edge $uv$ is inserted, either $u$ or $v$ can indicate to the two other nodes that they are part of a triangle. For larger cliques, we have more than one common neighbor and more than two edges, and it is not clear how to provide the same guarantee. Therefore, the bandwidth complexity of this problem remains an open question.
What is the $1$-round bandwidth complexity of [$\textsf{MemDetect}(K_s)$]{} under edge insertions?
For node insertions, we show that at least $\Omega(n)$ bits of bandwidth are required.
\[th:ks\_mdtct\_lb\_pv\] The randomized $1$-round bandwidth complexity of [$\textsf{MemList}(K_s)$]{} under node insertions is $\Omega(n)$.
Our construction is a generalization of the construction used in Theorem \[th:k3\_mdtct\_lb\_pv\] for triangles.
Let $A$ be a (randomized) $1$-round algorithm that solves [$\textsf{MemDetect}(K_3)$]{} under node insertions using bandwidth $B$ with error probability $\varepsilon$. Fix $x \in V$, and fix a set of $s-3$ other nodes $K
\subseteq V \setminus \{x\}$. Let $U = V \setminus (K \cup \{x\})$, and let $U = L \cup R$ be a partition of $U$ into two subsets of equal size (if $|U|$ is odd, let $|L| = |R| + 1$). Let $\mathcal{C}$ be the set of all possible bipartite graphs on the setst $L$ and $R$. For every $C \in
\mathcal{C}$ and every $u \in L, v \in R$, define the sequence $S_{C,u,v}$ as follows:
- Start with an empty graph (no nodes and no edges).
- During $n-(s-2)$ rounds, insert one node of $U$ on each round and connect it to the nodes which have been already inserted, according to the edges in $C$. After $n-(s-2)$ rounds, the graph is identical to the graph $C$.
- During the next $s-3$ rounds, insert one node of $K$ on each round and connect it to all other nodes in the graph. After this stage, the graph consists of the graph $C$, a clique $K$, and all possible edges between the nodes of $C$ and the nodes of $K$.
- On round $n$ insert $x$ and connect it to $u$, $v$, and all nodes of $K$.
In the final graph, $x$ is connected to exactly $s-1$ nodes, all of which are connected to each other except maybe $u$ and $v$. Thus, $x$ is part of a clique of size $s$ if and only if the edge $uv$ exists. Therefore, after one round, $x$ should output 1 if the edge $uv$ exists in $C$, or 0 otherwise. By our assumption, for every $C$ and every $u,v
\in U$, the output is correct with probability at least $1-\varepsilon$.
During the final round, $x$ receives only one message from $u$, one message from $v$, and $s-3$ messages from the nodes of $K$. As in Theorem \[th:k3\_mdtct\_lb\_pv\], when $x$ receives these messages from $u$, $v$, and the nodes of $K$, these nodes do not know the identity of $u$ and $v$. Therefore, the message received from $u$ depends only on the identity of $u$ and the graph $C$, and not on the identity of $v$. Likewise, the message received from $v$ depends only on the identity of $v$ and the graph $C$, and the $s-3$ messages received from the nodes of $K$ depend only on the graph $C$.
Now, consider the following experiment for a given graph $C \in
\mathcal{C}$. First, we run the sequence $S_{C,u,v}$ for some $u,v \in
U$, and stop just before the last round, in which $x$ is connected to $u$ and $v$. Then, every node $w \in U \cup K$ generates a message to be sent to $x$, as if it has been connected to $x$ on the final round. For every $w \in U$, let $m_w$ be the message generated by $w$, and let $m_K$ be the set of messages generated by the nodes of $K$ (note that $m_w$ and $m_K$ are random variables).
Note again, that the messages generated by the nodes of $U$ depend only on the graph $C$, and not on the other nodes that may have been connected to $x$ on the last round. Therefore, for every $u \in L, v \in
R$, given the messages generated by $u$, $v$, and all nodes of $K$, $x$ should be able to determine, with probability at least $1-\varepsilon$, whether or not the edge $uv$ exists in $C$.
Next, for every pair of nodes $u \in L, v \in R$, let $I_{uv}$ be the output of $x$ given the messages $m_u$, $m_v$, and $m_K$. Now, let $C'$ be the graph on the nodes of $U$, in which the edge $uv$ exists if and only if $I_{uv}=1$. We consider $C'$ to be the result of the experiment.
We now compute the probability that the $x$ has the correct output for all pairs $u \in L, v \in R$. For $C \in \mathcal{C}$, let $p_C$ denote the probability that at the end of the experiment we have $C'=C$. For every $u \in L, v \in R$, the value of $I_{uv}$ corresponds to the edge $uv$ in the graph $C$ with probability at least $1-\varepsilon$. Thus we have: [$$\begin{aligned}
p_C
\geq
(1-\varepsilon)^{|L| \cdot |R|}\end{aligned}$$]{}
Summing the above for every $C \in \mathcal{C}$ we get: $$\begin{aligned}
\label{eq:ks_mdtct_pv_geq}
\sum_C p_C
\geq
|\mathcal{C}| \cdot (1-\varepsilon)^{|L| \cdot |R|}
=
(2-2\varepsilon)^{|L| \cdot |R|}\end{aligned}$$
On the other hand, consider the set of messages generated by the nodes of $U \cup K$ at the end of the first stage of the experiment. For every possible set of messages $M$ generated by the nodes of $U \cup K$ during the first stage of the experiment, denote by $\phi_C(M)$ the probability for generating exactly the messages of $M$. Also, for every $C' \in
\mathcal{C}$, let $\Psi_M(C')$ denote the probability for the result of the experiment to be equal to $C'$, given the set of generated messages $M$. We have: [$$\begin{aligned}
\sum_C p_C
&=
\sum_C \sum_M \phi_C(M) \cdot \Psi_M(C)
\leq
\sum_C \sum_M \Psi_M(C)
=
\sum_M \sum_C \Psi_M(C)
\\ &=
\sum_M 1
=
|\mathcal{M}|\end{aligned}$$]{}
where $\mathcal{M}$ is the set of all possible values of $M$. Since every message has exactly $2^B$ possible values, the number of possible sets of $(n-1)$ messages is: [$$\begin{aligned}
|\mathcal{M}| = 2^{B(n-1)}\end{aligned}$$]{}
Hence: [$$\begin{aligned}
\sum_C p_C \leq 2^{B(n-1)}\end{aligned}$$]{}
Combining this with gives: [$$\begin{aligned}
2^{B(n-1)} \geq (2-2\varepsilon)^{|L| \cdot |R|}\end{aligned}$$]{}
Both $L$ and $R$ contain $\Theta(n)$ nodes. Therefore, after some simplifications we get the desired bound: $$B
\geq
\log (2-2\varepsilon)
\cdot
\frac{|L| \cdot |R|}{n-1}
=
\Omega(n)
\qedhere$$
Listing and detection
---------------------
Table \[tbl:ks\_listdtct\] summarizes the results of this subsection.
Node deletions Edge deletions Edge insertions Node insertions
---------------- ---------------- ----------------- -----------------
$0$ $\Theta(1)$ $\Theta(1)$ $\Theta(1)$
: Bandwidth complexities of [$\textsf{List}(K_s)$]{} and [$\textsf{Detect}(K_s)$]{} []{data-label="tbl:ks_listdtct"}
### Upper bounds {#upper-bounds-4}
The upper bounds for node deletions and edge deletions follow from Observation \[obs:probrel\] and the results of the last subsection.
For edge insertions, the algorithm used in Corollary \[cor:k3\_list\_ub\_pe\] for triangle listing does not work for larger cliques. However, we can use the algorithm of Theorem \[th:k3\_mlist\_ub\_r2\_pe\], which solves membership-listing of triangles in $2$ rounds, to solve listing of larger cliques in $1$ round.
\[th:ks\_list\_ub\_pe\] The deterministic $1$-round bandwidth complexity of [$\textsf{MemDetect}(K_s)$]{} under edge insertions is $O(1)$.
As said, we use the exact same algorithm of Theorem \[th:k3\_mlist\_ub\_r2\_pe\]. This algorithm allows every node $u$ to determine, on round $t$, the list of all triangles that contained $u$ on round $t-1$. Hence, each node can determine the set of all cliques — of any size — that contained it on the previous round.
We show that every clique of size $s$ is known to at least one node. For this, we recall that by the algorithm of Theorem \[th:k3\_mlist\_ub\_r2\_pe\], on every round, every node sends to all its neighbors an indication whether or not it got a new neighbor on the current round. This information is denoted by `NEW`.
Now, assume that the insertion of an edge $uv$ on round $t$ creates a clique $K$ of size $s$, and let $x$ be a node in $K$ other than $u$ and $v$. We show that on round $t$, $x$ can determine that $K$ is a clique. On round $t-1$, all nodes of $K$ are connected by an edge except $u$ and $v$. In particular, on round $t-1$, $K \setminus \{u\}$ and $K \setminus
\{v\}$ are both cliques that contain $x$. Hence, as stated above, after round $t$, $x$ knows that these two cliques exist. Moreover, on round $t$, both $u$ and $v$ send $\texttt{NEW} = 1$ to $x$, from which $x$ can deduce that the edge $uv$ has just been inserted. Hence, $x$ can determine on round $t$ that the nodes of $K$ are all connected to each other, as required.
Combining this with node/edge deletions can be done similarly with the algorithm of Corollary \[cor:k3\_mlist\_ub\_r2\_pe\_mvme\].
\[cor:ks\_list\_ub\_pe\_mvme\] The deterministic $1$-round bandwidth complexity of [$\textsf{MemDetect}(K_s)$]{} under node deletions and edge deletions/insertions is $O(1)$.
For node insertions, we can use the same algorithm used for triangles in Theorem \[th:k3\_list\_ub\_pv\].
\[th:ks\_list\_ub\_pv\] The deterministic $1$-round bandwidth complexity of [$\textsf{List}(K_s)$]{} under node insertions is $O(1)$.
We use the same algorithm as in Theorem \[th:k3\_list\_ub\_pv\]. By this algorithm, on every round, each node sends to all its neighbors an indication whether or not it has a new neighbor on the current round. We show that this guarantees that every clique on $s$ nodes is known to at least one node.
Consider a clique $K$ with $s$ nodes, and assume that $u$ is the first node of $K$ that has been inserted (if some of the nodes of $K$ exist in the initial graph, let $u$ be any one of them). We show that $u$ can determine that $K$ is a clique $1$ round after the last node of $K$ is inserted.
By induction, we show that for every non-empty $L \subseteq K$, $u$ can determine that $L$ is a clique $1$ round after the last node of $L$ is inserted. The claim then follows for $L=K$.
The claim is trivial for $L=1$, since it contains only $u$. It is also trivial if $L$ is the set of nodes of $K$ that exist in the initial graph. Now, assume that a new node $v \in K$ is inserted, along with edges to all nodes of $L$. When this happens, $u$ is connected to all nodes of $L$, therefore it receives from each of them an indication that they are connected to the new node. Thus, after a single round of communication, $u$ can determine that the set $L \cup \{v\}$ is a clique, as claimed.
By Observation \[obs:probrel\], all upper bounds shown here for [$\textsf{List}(K_s)$]{} also hold for [$\textsf{Detect}(K_s)$]{}.
We note that all results of this subsection hold for any value of $s$, even $s=n-O(1)$.
### Lower bounds {#lower-bounds-4}
The lower bounds for each type of change are either $0$ or $\Omega(1)$, which is trivial.
The algorithm of Theorem \[th:ks\_list\_ub\_pv\] also works in combination with either node deletions or edge deletions, by simply having each node send an indication whether it lost a neighbor on the current round. However, unlike the case of triangles, this does not work if we allow both node deletions and edge deletions. This is because this algorithm guarantees that every clique is known to at least two of its nodes. For triangles, this suffices to guarantee that when either a node or an edge is deleted, at least one of these two nodes know whether the triangle still exists. For larger cliques, this is not the case — for example, consider a clique on $s \geq 4$ nodes, and let $u,v,x,y$ be four nodes in that clique. Assume that only $u$ and $v$ know that the clique exists. Now, either the edge between $x$ and $y$ is deleted, or some other node is deleted which is connected to both $x$ and $y$, but not to either $u$ or $v$. In both cases $x$ and $y$ can send an indication that they lost a neighbor, but until the next round, $u$ and $v$ cannot distinguish between the two cases, and therefore cannot determine whether the clique still exists.
This can be fixed by having every node send, together with an indication whether or not it has a new neighbor or a missing neighbor, the ID of the new or missing neighbor. However, this requires $O(\log n)$ bits of bandwidth, which is significantly more than the $O(1)$ bits required to handle each type of change alone. Thus, the exact bandwidth complexity of this combination remains an open question.
What is the $1$-round bandwidth complexity of [$\textsf{List}(K_s)$]{} under node insertions and node/edge deletions?
Also, as is the case for triangles, it is an open question whether it is possible to combine edge insertions and node insertions with $O(1)$ bits of bandwidth.
What is the $1$-round bandwidth complexity of [$\textsf{List}(K_s)$]{} under node/edge insertions?
#### Acknowledgements.
This project has received funding from the European Union’s Horizon 2020 Research And Innovation Program under grant agreement no. 755839.
[^1]: [email protected]
[^2]: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This is a review of recent developments in Monte Carlo methods in the field of ultra cold gases. For bosonic atoms in an optical lattice we discuss path integral Monte Carlo simulations with worm updates and show the excellent agreement with cold atom experiments. We also review recent progress in simulating bosonic systems with long-range interactions, disordered bosons, mixtures of bosons, and spinful bosonic systems. For repulsive fermionic systems determinantal methods at half filling are sign free, but in general no sign-free method exists. We review the developments in diagrammatic Monte Carlo for the Fermi polaron problem and the Hubbard model, and show the connection with dynamical mean-field theory. We end the review with diffusion Monte Carlo for the Stoner problem in cold gases.'
address: 'Department of Physics, Arnold Sommerfeld Center for Theoretical Physics and Center for NanoScience, University of Munich, Theresienstrasse 37, 80333 Munich, Germany'
author:
- Lode Pollet
title: 'Recent developments in Quantum Monte-Carlo simulations with applications for cold gases'
---
Introduction
============
Although the Schr[ö]{}dinger equation describes the evolution of any quantum system, solving the many-body problem remains a daunting task. The growth of the Hilbert space is exponential with the number of particles. Exact diagonalization hence requires an exponential amount of computational resources, since an exponentially growing set of numbers needs to be stored in order to describe the state vector.
The premise of Monte Carlo methods is that, under very mild mathematical requirements, physical properties can be computed stochastically converging as $\sim 1 / \sqrt{N}$, where $N$ is the number of independent samples, [*irrespective of the dimension of the (Hilbert) space*]{}. The task of a Monte Carlo developer is to find clever ways to generate independent samples for typically strongly interacting systems and arbitrary topologies. Quantum Monte Carlo simulations are classical Monte Carlo simulations of quantum systems that have been represented in a way amenable to Monte Carlo simulations. Such a quantum-to-classical mapping is necessary because the spectrum of a quantum system is a priori unknown. For example, using Feynman’s path integral description for thermodynamic systems, a quantum particle is represented by a polymer or worldline, describing the propagation in imaginary time. We will see in Sec. \[sec:PIMC\_lattice\] how this mapping of the $d$ dimensional quantum system to a classical system in $(d+1)$ dimensions is at the heart of [*path integral Monte Carlo*]{}, a method very successful for bosonic systems. In the best cases, an algorithm that scales linearly with the system size has been found; in other words, technical advances in computer hardware can directly be exploited to study larger systems and/or lower temperatures. We will see a couple of examples of this in Sec. \[sec:PIMC\_lattice\] and Sec. \[sec:PIMC\_continuous\]. However, the main drawback of stochastic methods is that they only work efficiently for a subclass of physical systems, namely those systems that are bosonic or can be mapped to bosonic systems. Generic fermionic systems cannot be sampled this way: the infamous sign problem occurs, in which a negative weight is associated with certain physical configurations. This does not prevent the application of the Monte Carlo method, but brings back the exponential scaling of the required resources with system size. That the sign problem is most likely unsolvable, has a deep physical meaning and follows from the proven NP-hardness of the problem [@Troyer2005]. In such cases one resorts to approximations, or tries to find expansions (based on an analytical understanding of the problem) that converge sufficiently fast such that the sign problem remains manageable. We will also see a couple of examples of this in Sec. \[sec:diagrams\] of this review. In the last section, we review the application of diffusion Monte Carlo to the Stoner problem in cold gases.
Stochastic methods are used in every branch of research. Already in condensed matter physics alone the available literature is too vast and diverse for a single review. In order to have a better defined framework for this text, we restrict ourselves to those Monte Carlo algorithms and applications that have been used in the context of cold gases over the past 10 years. We will omit variational Monte Carlo and be very concise on diffusion Monte Carlo and auxiliary field Monte Carlo. This does not mean that these methods are unimportant or unsuccessful (the latter was recently able to identify a spin liquid in the Hubbard model on the honeycomb lattice for example [@Meng2010]), but this decision is only determined by space constraints. Among the existing reviews on Monte Carlo methods we mention Ref. [@Bajdic2009; @Kolorenc2011] for variational and diffusion Monte Carlo methods for condensed matter physics combined with material descriptions, Ref. [@Georges1992; @Maier2005; @Kotliar2006; @Gull_review] for dynamical mean-field theory and impurity solvers, and Ref. [@Assaad2001] for determinantal methods based on auxiliary field decompositions using discrete time. This review assumes that the reader is familiar with the basics of classical Monte Carlo simulations, as can be found in many excellent textbooks such as for example Refs. [@LandauBinder2000; @Krauth2006; @JunLiu2001]. We are also focusing on three-dimensional (and occasionally two-dimensional) systems, omitting one-dimensional systems, for which the Density Matrix Renormalization Group (DMRG) is often the method of choice [@Schollwoeck2011]. Ref. [@Cazalilla2011] is a recent review covering the one-dimensional world of bosonic systems, including the relevant Monte Carlo methods.
The applications in this review are all situated in the field of ultra cold gases. These are dilute (a typical density is $\sim 10^{15}$ cm$^{-3}$), neutral atomic alkali-gases in a metastable state that has a lifetime of the order of a few seconds. Following the advances in laser trapping and cooling, bosonic atoms were cooled to quantum degeneracy in 1995. They were weakly interacting and well described by Bogoliubov’s theory of the weakly interacting Bose gas. There are two ways to make the system strongly interacting: one is to tune a magnetic field close to a Feshbach resonance where the scattering length diverges; the other is to load them into an optical lattice where the physics is dominated by tunneling as the atoms become more localized when ramping up the lattice laser [@Bloch2008]. Following the proposal of Jaksch [*et al.*]{} [@Jaksch1998], Greiner [*et al.*]{} [@Greiner2002] demonstrated experimentally the superfluid to Mott insulator transition for ultracold bosonic atoms subject to an optical lattice. This demonstrated that condensed matter models could be implemented in cold gas experiments. Compared to traditional condensed matter physics systems, ultracold gases are very clean, controllable, and have tunable system parameters [@Bloch2008].
The demonstration by Greiner [*et al.*]{} [@Greiner2002] has led to the paradigm of [*quantum simulation*]{}: an unsolvable model with competing interactions, believed to describe a real material and intractable by numerical means, can be implemented and realized in a cold gas experiment, and analyzed this way. By changing the interaction, dispersion, and density of the atoms the phase diagram of the model can be revealed. This allows to assess how well the model describes the material. An optical lattice system would thus be equivalent to an analog quantum computer, which is specifically tailored to one task (i.e., one particular model), but is more powerful than a classical computer in the sense that it can perform [*quantum*]{} operations (since typical scales in those systems are of the order of kHz, we do not claim that they are fast computers). Before the quantum optical lattice simulator can be trusted as a reliable device, it needs to be benchmarked and validated against known results. This is where exact numerical methods, and in practice these are quite often Monte Carlo simulations, come in: because of the cleanliness and control over optical lattice systems, their complete Hamiltonians can be simulated in cases where there is no sign problem, such as bosonic cold gases. Excellent agreement has been reached between experiments and simulations this way [@Trotzky2010], after taking a number of technical details into account. This has sparked renewed interest in developing numerically exact solutions to models in parameter regimes that were hitherto of lesser importance; for example, the three-dimensional Hubbard model for temperatures approaching the N[é]{}el temperature and interaction strengths up to $1.5$ times the bandwidth have now been fully and controllably mapped out [@Fuchs2011]. For those, cluster dynamical mean-field theory simulations (so-called dynamical cluster approximation (DCA) simulations to be precise) were performed for cluster sizes of size $\sim 100$, enabling a reliable extrapolation in cluster size while still having a manageable sign problem. Now that optical lattice emulators have been validated and are gradually moving on to parameter regimes no longer tractable by traditional numerical means, will they be used as stand-alone machines? In our view, the importance of numerical support for explanations of cold gas experiments will not diminish in the near future. Instead, we believe that computational physicists (for whom, ironically, [*quantum simulation*]{} is historically understood as simulating quantum systems on classical computers) will develop a variety of new algorithms and convergence schemes tailored towards the problem of interest. The reward will be a direct test against those experiments, and the development of new algorithms that may be applicable in other fields.
It is my hope that newcomers to the field will find this text useful as a starting point in understanding the different expansion schemes, their similarities and differences, and learn how such schemes can be turned into powerful stochastic algorithms. Perhaps this text can lower the threshold to enter into this field. Codes, full algorithmic descriptions or expressions for detailed balance cannot be found in this review, but references are given. I also hope that this overview is useful for the researchers in the field of cold gases without numerical background, so that a single text can be used as an orientation platform that accurately describes the state of the art. I likewise hope that experts will find this text useful for bringing together at first sight totally different methods, but which may ultimately lead to new ideas and crosstalk among aficionados of different subfields, and extend the double paradigm of quantum simulation to new models and parameter regimes.
This paper is organized as follows. We start in Sec. \[sec:PIMC\_lattice\] by reviewing path integral Monte Carlo methods for bosonic lattice systems. The key ideas behind the worm algorithm are explained. The physics of the Bose-Hubbard model is discussed in a nutshell, after which we compare the theoretically ideal case with the experimental realization of cold gases in a trap. We talk about time of flight interference images and single site in-situ resolution measurement tools. We proceed with disordered bosonic systems and the physics of the Bose glass phase. Polar molecules are next, which are systems with richer phase diagrams because of the long-range interaction between the molecules. Mixtures of bosonic atoms and bosonic atoms with an internal spin degree of freedom are also covered. In Sec. \[sec:PIMC\_continuous\] we review path integral Monte Carlo simulations for bosonic systems in continuous space, covering the worm algorithm and the following applications: the weakly interacting Bose gas, disordered Bose gases and supersolids for atomic/molecular systems with long-range interactions. In Sec. \[sec:diagrams\] we leave the bosonic world for fermionic systems which in general do not have a sign-positive representation. We review determinantal methods which are often used when an additional symmetry allows for a positive expansion. In this context we discuss the critical temperature for the resonant Fermi gas at unitarity. Next we introduce diagrammatic Monte Carlo. Historically, this was first applied to the Fr[ö]{}hlich Hamiltonian for electron-phonon interactions. This model is in fact sign positive, but the Fermi-polaron problem, which comes next, is not. Also, the state of the art of diagrammatic Monte Carlo for the Hubbard model is reviewed. In the final parts of the review we show how dynamical mean-field theory (DMFT) and its cluster extensions have been used to map out the thermodynamics of the 3D Hubbard model for temperatures approaching the N[é]{}el temperature, and how DMFT can be combined with diagrammatic Monte Carlo, paving the way for future developments. Before concluding, we mention diffusion Monte Carlo and how it was used in understanding the Stoner transition (or absence thereof) for atoms on the repulsive branch of the Feshbach resonance.
Path Integral Monte Carlo: lattice models {#sec:PIMC_lattice}
=========================================
Bose-Hubbard model
------------------
Consider the Bose-Hubbard model [@Fisher1989] describing scalar bosons on a lattice, $$H = H_1 + H_0 = -t \sum_{\langle i,j \rangle} b_i^{\dagger}b_j + \frac{U}{2} \sum_i n_i(n_i-1) - \mu \sum_i n_i.
\label{eq:BoseHubbard}$$ The operators $b_j$ satisfy the bosonic commutation relations, $[ b_i, b_j^{\dagger}] = \delta_{i,j}$ and zero for commutators with two creation or two annihilation operators. The operator $n_i = b_i^{\dagger}b_i$ counts the number of bosons on site $i$. The first term describes the hopping of bosons between neighbouring sites with tunneling amplitude $t$. The second term describes the on-site repulsion with strength $U$, while the last term is proportional to the chemical potential $\mu$. The latter two terms are diagonal in the Fock basis of occupation numbers, $\vert \{ n_i \} \rangle$, and are combined into $H_0$. The kinetic term, which is a one-body operator, is not diagonal in this basis but will lead to transitions between Fock states with matrix elements $\langle \ldots n_i-1, n_j+1, \ldots \vert -tb_j^{\dagger}b_i \vert \ldots n_i, n_j \ldots \rangle = -t \sqrt{n_i(n_j+1)}$. The lattice spacing is set to unity. Unless otherwise specified we have a cubic lattice in mind of linear size $L$ with periodic boundary conditions.
The Bose-Hubbard model has three phases, which can easily be identified in the limiting cases [@Fisher1989; @Sachdev99]: First, in the limit of high temperature, the system is a normal liquid. At zero temperature and $t=0$, the system is a Mott insulator with fixed, integer density, a gap, and zero compressibility. For finite hopping, stable Mott lobes around the $t=0$ insulators are found, which are surrounded by a gapless, compressible superfluid. The superfluid phase also exists at finite temperature. The Bose-Hubbard model is the simplest model that describes a conductor-insulator transition for bosons. It can also describe the physics of the weakly interacting Bose gas.
![Graphical representation of a typical configuration in the Green function sector. Imaginary time goes from left to right in the figure, there are five sites depicted. World lines are denoted by single lines (site is once occupied), double lines (site has occupancy two) or dashed lines (site is not occupied). Interactions (hopping of a particle) are denoted by vertical lines. The two circles mark a discontinuity in the world lines and correspond to the worm operators. One of them creates an extra particle, the other one annihilates it. Closed worldlines are formed when the worldlines annihilate each other. The figure is taken from Ref. [@Pollet2007]. []{data-label="fig:worldlines"}](LOWA_conf){width="0.8\columnwidth"}
![Graphical representation of how a worm pair is inserted or removed, signaling a transition between the partition function sector and the Green function sector. For inserting the worm, an arbitrary site and arbitrary imaginary time are chosen. The occupation between the worm ends can in general be either higher or lower than the occupation outside. One of the worm ends remains stationary (called the worm tail), while the other one is mobile (called the worm head). The figure is taken from Ref. [@Pollet2007]. []{data-label="fig:LOWA_insert"}](LOWA_insert){width="0.8\columnwidth"}
![Graphical representation of how a worm can insert/remove a kink ([*i.e.,*]{} a hopping of bosons), shown for a 1D lattice. When going from configuration (a) to configuration (b) or (c), we assume that the worm is moving to right (that is to greater imaginary times) and a kink is inserted to one of the neighbouring sites. Detailed balance also requires that we also stay sometimes in configuration (a). In algorithms with a fixed direction of propagation, the worm will reverse its propagation direction and start moving to lower imaginary times in such cases. For the reverse updates, when being in configuration (b) and moving to the left, the worm has the possibility of (a) removing the kink, (b) staying in the present configuration but changing its direction of propagation, (c) relinking the kink and changing the direction of propagation. All updates remain local : there are no changes to the configuration other than the ones shown over this small imaginary time interval and sites in the vicinity of the worm head. The figure is taken from Ref. [@Pollet2007]. []{data-label="fig:LOWA_kink"}](LOWA_kink){width="0.8\columnwidth"}
Continuous time expansion schemes and path integral Monte Carlo
---------------------------------------------------------------
The starting point is the following decomposition for the partition function, $$Z = {\rm Tr} e^{-\beta H} = {\rm Tr} {\mathcal{T} }e^{-\beta H_0} \exp \left[ - \int_0^{\beta} d\tau H_1(\tau) \right],$$ where $H_1(\tau) = e^{\tau H_0}H_1 e^{-\tau H_0}$. The trace is taken over all Fock basis states specified above (in which $H_0$ is diagonal). The exponential is expanded into a time-ordered product [@Mahan; @NegeleOrland; @FetterWalecka], $$Z = {\rm Tr} {\mathcal{T} }e^{-\beta H_0} \left[ 1 - \int_0^{\beta} d\tau H_1(\tau) + \int_0^{\beta} d\tau_1 \int_{\tau_1}^{\beta} d\tau_2 H_1(\tau_1) H_1(\tau_2) + \ldots \right].
\label{eq:BH_expansion}$$ This expansion can graphically be represented by worldlines (thanks to the $U(1)$ symmetry of the Bose-Hubbard model ), similar to as what is shown in Fig. \[fig:worldlines\] except for the two circles (which are explained later). The inverse temperature $\beta = 1/T$ is understood as an imaginary time, where the matrix elements of the operators $\exp [-\Delta \tau H_0]$ act as propagators between the different states at $\tau_1, \tau_2, \ldots$, with $0 < \tau_1 < \tau_2 < \ldots < \tau_j < \ldots < \tau_n < \beta$. The perturbations $H_1$ change the states at those times. A worldline then describes the trajectory of a particle propagating in imaginary time.
The expansion in path integral Monte Carlo (PIMC) is always understood over a finite volume and finite imaginary time. There is no singularity caused by a phase transition, which can only be studied in a finite size scaling analysis. The expansion is written in terms of an entire function so that there are no non-physical divergencies in the PIMC formulation. The expectation value of an observable $A$ is given by $\langle A \rangle = \frac{1}{Z} {\rm Tr} A e^{-\beta H}$. In PIMC, a statistical interpretation is given to Eq. \[eq:BH\_expansion\] by introducing weights $w$ through $Z = {\rm Tr}_{\vert \{n_i \} \rangle} w({\vert \{ n_i \} \rangle})$. We now have to statistically generate all possible configurations according to their weights by generating all possible expansion orders and matrix elements, assign an (unnormalized) weight to each one of them, evaluate the observable $A$ in every configuration, and sum all these contributions. For instance, one can perform local updates by inserting pairs of hopping elements, in which a particle hops from a site to its neighbour, and back at a later time.
At high temperatures or deep in the Mott insulating phase, the kinetic energy is small compared to $U$ and/or $T$, and few perturbation orders are needed in Eq. \[eq:BH\_expansion\]. Such is not the case when the contributions to the free energy coming from the hopping are large: There is no reason to expect that the local updates would be efficient. Even worse, they are not ergodic: The low energy states in a superfluid are given by states with a different winding number. These are paths in which a particle winds around the full length of at least one direction before closing on itself again. Configurations with different winding numbers are topologically distinct and cannot be transformed into each other by local updates alone. The winding number $W$ is directly related to the superfluid density through $\rho_s = \frac{\langle W^2 \rangle L^{2-d} } {d \beta}$ [@Pollock1987], with $d$ the dimension of the system. Worldlines with non-zero winding number are caused by bosonic permutations.
The worm algorithm
------------------
The worm algorithm has completely solved those ergodicity problems [@Prokofev1998_worm]. Instead of working with the partition function $Z$ alone one also works in the Green function sector $Z_G$, $$Z_G = {\rm Tr} \mathcal{T} \{ b_i(\tau_0) b_j^{\dagger}(\tau) e^{-\beta H} \}.$$ Graphically, the operators $b_i(\tau_0)$ $b_j^{\dagger}(\tau)$ are open ends delimiting a segment of a worldline (see Fig. \[fig:worldlines\]) and are called the worm head and the worm tail. The worms can be on any site and any time. A correct algorithm allows for the transition between the Green function sector and the partition function sector (where measurements of observables such as the energy and the superfluid stiffness are done), and allows to move the worms around in configuration space. This is shown in Fig. \[fig:LOWA\_insert\]. Worms also have the ability to insert and remove hopping elements (kinks), as shown in Fig. \[fig:LOWA\_kink\]. Since the worm operators correspond to open ends on a world line segment they have no problem in exploring configurations with different winding numbers. All updates are local in the Green function sector, which means that all acceptance factors can be made of order unity. In phases with off-diagonal long-range order (either true long-range order such as seen in a Bose-Einstein condensate or quasi long-range order with correlation functions decaying as a power-law) the worms will preferentially be far away from each other in configuration space, [*i.e.*]{}, we are efficient in describing the physics of those phases.
In the literature, different implementations of the above ideas can be found. The minimal requirements for the updates are however nothing more than the ones shown in the Figs. \[fig:LOWA\_insert\] and \[fig:LOWA\_kink\], combined with a move update which moves the position of the worm head forward or backward in imaginary time without changing the kinks. We will therefore only briefly mention the key ideas and refer to the relevant papers. Prokof’ev, Svistunov and Tupitsyn were the first to introduce the worm algorithm [@Prokofev1998_worm]. They formulated it for the Bose-Hubbard model in the grand-canonical ensemble in the path integral representation. Sandvik and Sylju[å]{}sen introduced worm operators for spin models in the stochastic series representation with directed loop updates [@Sandvik1999; @Syljuasen2003]. It is worth remarking that the loop algorithm was the first algorithm formulated in continuous imaginary time ([*i.e.,* ]{} free of Trotter discretization error) to overcome the critical slowing down near the second order phase transition [@Evertz1993; @Beard1996; @Evertz2003], even before the worm algorithm was invented. Pollet [*et al.*]{} tried to combine the two algorithms and formulated the worm algorithm in the path integral representation but with directed loops [@Pollet2007]. Rombouts [*et al.*]{} formulated a worm algorithm in the canonical ensemble by letting the worm head and the worm tail propagate simultaneously in imaginary time, thereby creating and annihilating a particle on different sites simultaneously [@Rombouts2006; @VanHoucke2006]. These ideas were picked up by Rousseau [*et al.*]{} in formulating algorithms with multiple worms (called stochastic Green functions) [@Rousseau2007; @Rousseau2008]. In [@Kawashima2004] a review of world-line Monte Carlo methods up to 2003 can be found with a discussion on the loop algorithm, directed loops algorithm and the worm algorithm. Classical worm algorithms have also been formulated [@Prokofev2001; @Prokofev2009; @Alet2003]. These algorithms can, in particular, be used to efficiently simulate quantum criticality in cases when the universality class can be mapped onto a higher dimensional classical model (cf. Sec. \[sec:WIBG\_Tc\]).
The physics of the 3D Bose-Hubbard model in a nutshell
------------------------------------------------------
![Finite temperature phase diagram of the Bose-Hubbard model at unit filling on a simple cubic lattice. $T/t = 5.591$ is the critical temperature of a non-interacting Bose gas with the tight binding dispersion. Away from this point, the transition from the normal liquid (NL) to the superfluid phase (SF) belongs to the universality class of the $d$-dimensional $XY$-model. Strictly speaking, the Mott insulator (MI) occurs only at zero temperature, but can loosely be defined to the right of the grey line at low enough temperature [@Gerbier2007]. At commensurate densities, the transition from the SF to the MI belongs to the universality class of the $(d+1)$ dimensional $XY$-model. Reprinted figure with permission from Ref. [@CapogrossoSansone07]. Copyright (2007) by the American Physical Society. []{data-label="fig:bose_hubbard"}](fig_bosehubbard){width="0.8\columnwidth"}
The phase diagram of the 3D Bose-Hubbard model, Eq. \[eq:BoseHubbard\], is shown in Fig. \[fig:bose\_hubbard\] at finite temperature and unit density. The transition from the normal liquid to the superfluid phase belongs to the 3D $XY$-model [@Fisher1989]. At zero temperature, the superfluid phase undergoes a phase transition to the Mott insulating phase. This transition belongs to the $(d+1)$ dimensional XY model (and is hence of mean-field nature with logarithmic corrections): Near the tip of the Mott lobe, the particle and hole excitations have the same effective mass and become relativistic [@Fisher1989]. Away from commensurability, the transition between the superfluid and the Mott insulator belongs to the Gaussian type universality class [@WIBG; @Fisher1989; @Sachdev99].
In the Mott insulator, the dispersion of the particle and hole excitations can be found by evaluating the Green function at zero momentum in the large imaginary time limit. Specifically, one finds, using the Lehmann representation, that $G({\mathbf p}, \tau) \to Z_{\pm} e^{{-\epsilon({\mathbf p})}_{\pm}\tau}, \tau \to \infty$, with $Z_{\pm}$ the quasiparticle weights and $\epsilon(\mathbf p)_{\pm}$ the energies of a single particle (+) and hole (-) excitation. This was used in Ref. [@CapogrossoSansone07] to indeed see the emerging relativistic effects near the tip of the Mott lobe, but also to accurately determine the phase boundary of the Mott lobe. The semi-analytic predictions can also be used to study inhomogeneous models in the local density approximation [@CapogrossoSansone07].
The transition between the normal liquid and the superfluid can be studied by a finite size scaling analysis for the superfluid density [@Fisher1989]. (note: Also the tip of the Mott lobe, where the critical exponents take mean-field values $\nu = 1/2$ and $z=1$, is best found using a finite size scaling analysis of the superfluid density. No logarithmic corrections could be discerned in Ref. [@CapogrossoSansone07].) Writing the distance to the critical point as $\theta = (T-T_c)/T_c$, the superfluid density $n_s$ obeys the scaling equation $$n_s(\theta, L) = \xi^{-1} f_s(\xi/L) = L^{-1} g_s(\theta L^{1/\nu}),
\label{eq:FSS_gap}$$ with $L$ the system size, $\xi$ the correlation length, and $f_s$ and $g_s$ universal scaling functions. The first equality follows from the theory of critical phenomena; in the second one we have used that the correlation length will be cut off by the system size close enough to the transition point and that $\xi \sim \theta^{-\nu}$, with $\nu = 0.6717(1)$ [@Campostrini2006] for the universality class of the 3D XY model. Hence, plotting $T/t$ as a function of $n_sL$ for different system sizes will display a crossing in a single point (to leading order), which is the critical temperature. Alternatively, data collapse can be obtained when plotting $n_sL$ as a function of $\theta L^{1/\nu}$.
Bosons in a 3D optical lattice
------------------------------
Although cold bosonic atoms subject to an optical lattice realize the Bose-Hubbard model, there are a few differences between experiments and the model discussed thus far. We list the most important ones:
- The confining potential to trap the neutral atoms is well described by a parabolic potential. For a potential $V(r)$ with spherical symmetry, the local chemical potential is changed to $\mu - V(r)$, with $\mu$ the global chemical potential. Its influence on the phase diagram has been analyzed in great detail [@Kashurnikov02; @Batrouni02; @Wessel04; @Gygi06; @Ma08; @Mahmud11]. Because of the inhomogeneous chemical potential, spatial coexistence of superfluids and Mott insulating domains may occur. A sharp phase transition for a homogenous system is replaced by a gradual crossover in a trapped system. For shallow potentials, the local density approximation (LDA) is often used, where every site is treated as an independent system with its own chemical potential. It is a very good approximation when the correlation length in the system is small, i.e., away from phase transitions [@Pollet10_crit] (see also Sec. \[sec:singlesite\]).
- The fact that atoms are neutral particles makes detection much more difficult than in typical condensed matter systems. The most frequent detection tool is time-of-flight interference images, where the confining potential is switched off and the atoms expand ballistically, after which a laser shines on them and makes a picture on a camera. In the long-time limit the measured atomic interference pattern reflects the momentum distribution of the initial system. However, for typical expansion times of $15-20$ms, there remains a Fresnel diffraction term which is not negligible [@Gerbier2008], while interactions during the expansion can be neglected provided the atoms are well localized in the lattice and the density is low. This Fresnel diffraction term acts like a knife cutting off long-range correlations; typically correlations beyond 5-6 lattice spacings average out due to the fast oscillating phases. In particular, the $k=0$ condensate peak in an interference pattern is seriously affected (and cannot be used to infer the condensate fraction at all). The Fresnel term can fortunately be taken into account in Monte Carlo simulations [@Gerbier2008]. For the weakly interacting Bose gas on the other hand [*interactions*]{} during the expansion cannot be neglected [@Dalfovo1999].
- A major difference with condensed matter physics is that trapped cold gases are well described by an isolated system (recall for example the absence of a phonon bath). Despite the very low absolute temperatures, the involved scales are [*not*]{} orders of magnitude smaller than the bosonic condensation temperature (or the Fermi energy for fermionic systems). As a consequence, entropy is conserved but temperature changes when changing the system parameters in an ideal experiment, such as adiabatically ramping up the optical lattice [@Schmidt2006; @Ho2007; @Pollet2008]. The entropy cannot directly be computed in quantum Monte Carlo simulations. Most often it is computed by integrating the specific heat over temperature, but also thermodynamic integration and quantum Wang-Landau sampling to obtain the full partition function have been tried [@Pollet2008]. For a uniform system the entropy is exponentially suppressed ($\sim e^{-U/T}$) in the Mott insulator, meaning that the Mott insulator cannot be reached under adiabatic changes [@Ho2007] and that the system would heat into a normal phase. The normal phase, just as the Mott insulator, has no interference peaks in time-of-flight images. Fortunately, for a trapped system the entropy of the Mott lobes is not important since the bosons in the edges of the cloud are normal. The edges can hence serve as an entropy reservoir. In Ref. [@Ho2007] the width of the superfluid rings between the Mott plateaus in the density wedding cake structure of the trapped Bose-Hubbard model was shown to be related to the total entropy in the system. It was argued that by ramping up the optical lattice, the confinement gets substantially tighter, the width of the superfluid regimes shrinks, and temperature may raise dramatically. However, quantum Monte Carlo simulations showed that this scenario was too pessimistic and that Mott-like features (which require a temperature lower than the crossover scale $T/U < 0.2$ [@Gerbier2007]) can easily be reached for realistic parameter regimes [@Pollet2008]. Because the entropy is carried by the normal bosons and the largest volume fraction is found in the edges, one sees that for sufficiently deep lattices temperature scales with the on-site repulsion strength, $T \sim U$ [@Pollet2008; @Gerbier2005; @Ho2007]. Note that temperature cannot be measured directly in the lattice system, although it can be accurately determined before the atoms are loaded in the lattice by time-of-flight images, provided the condensate fraction is not too high [@Dalfovo1999].
- Since the system size is small and the number of measurement techniques limited, it is also difficult to determine ’transitions’ accurately. It was argued in Refs. [@Diener2007; @Kato2008; @Zhou09a] that an accurate determination of the critical temperature for the normal to superfluid temperature is complicated because of the existence of strong, though non-divergent, peaks already present in the normal phase. Together with the absence of a diverging length scale in the trapped system at the critical point, the precise determination of the critical point would be impossible. These arguments did not take the Fresnel diffraction term into account (which would rather enforce the argument). However, as we will discuss in detail below, Trotzky [*et al.*]{} could accurately determine the full finite temperature phase diagram at unit filling in the trap center by monitoring the visibility, the condensate peak and its width [@Trotzky2010], showing that the arguments of Refs. [@Diener2007; @Kato2008; @Zhou09a] are presently not a limiting factor for a realistic system.
![ Comparison between the integrated column densities from time-of-flight images obtained experimentally (lower row) and by quantum Monte Carlo simulations (upper row) for fully realistic system parameters. Five different temperatures (these are input parameters in the Monte Carlo simulations) are shown, $T = 11.9$nK, $T=19.1$nK, $T=26.5$nK, $T=31.8$nK, and $T= 47.7$nK, from left to right. The lattice depth is 8 recoil energies ($U/t=8.11$) and there are about $N=2.8 \times 10^5$ particles in the system. The figure is courtesy of Stefan Trotzky and results from the work reported in Ref. [@Trotzky2010]. It was taken from the website of I. Bloch, <http://www.quantum-munich.de/media/finite-temperature-comparison-of-experiments-and-qmc-simulations/>. []{data-label="fig:comparison_bosehubbard"}](fig_comparison_qmc_expt){width="0.8\columnwidth"}
In the introduction we mentioned that cold atom experiments can be seen as quantum simulators or quantum emulators, but before they can be trusted as such quantum analog computers, they should be benchmarked against known results. Trotzky [*et al.*]{} chose the superfluid-to-normal liquid transition of the Bose-Hubbard model at unit density in the middle of the trap [@Trotzky2010] for this benchmarking. They mapped the transition out experimentally on the basis of time-of-flight images, and compared the results with quantum Monte Carlo simulations with no free fitting parameter. The excellent agreement is shown in Fig. \[fig:comparison\_bosehubbard\]. In this analysis, it was crucial to take the conservation of entropy and the finite duration of the time-of-flight images into account (the knotted peaks in Fig. \[fig:comparison\_bosehubbard\] are a consequence of this). However, additional non-entropic heating was observable (and surprisingly large at the lowest temperatures), due to spontaneous emission, classical noise and lattice imperfections. This has stimulated studies from the quantum optics community to better estimate and find solutions for this problematic behaviour [@Pichler2010; @Gerbier2010].
Single-site resolution and addressability {#sec:singlesite}
-----------------------------------------
One of the most amazing advances in the last couple of years is the single site resolution and addressability of atoms in optical lattices [@Gemelke09; @Bakr09; @Wuertz09; @Bakr10; @Sherson10; @Weitenberg11a; @Weitenberg11b; @Endres11; @Bakr11; @Hung10; @Hung11]. Individual atoms can now be observed and addressed, typically in a 2D setup. When the measurement in the quantum gas microscope [@Bakr09] is performed, the atoms are quickly brought into a very deep lattice suppressing tunneling. The atoms are then illuminated with an optical molasses that serves to localize the atoms while fluorescence photons are collected by the high resolution optics. In addition, light assisted collisions immediately eject atoms in pairs from the lattice site, leaving behind an atom only if the initial occupation is odd. Remaining atoms scatter several thousand photons during the exposure time and can be detected with high fidelity. The density profiles were also shown to be in good agreement with Monte Carlo simulations [@Bakr10; @Endres11], and hence provide further evidence that thermodynamics is a valid description for these systems. Also density-density correlation functions are now accessible experimentally, which for instance was demonstrated by measuring the parity operator for the 1D Mott insulator [@Endres11]. We will now discuss two new opportunities for studying many-body physics which were made possible by the experimental advances in single-site resolution methods, namely the possibility to obtain the equation of state, and the study of critical phenomena.
- Single-site addressability could provide a means for obtaining the equation of state. In the edges the gas is in a normal phase, where measuring the density profile and comparing to high temperature series expansions [@Zhou09a; @Ho09; @Ma10] gives access to all thermodynamic quantities. If this procedure cannot be followed, then temperature can still be determined by using the fluctuation-dissipation theorem [@Zhou09b; @Ma10] $$T \frac{\partial n( \bm{r})}{\partial \mu} = \int d \bm{r'} \langle n(\bm{r}) n(\bm{r'}) \rangle - \langle n(\bm{r}) \rangle \langle n(\bm{r'}) \rangle.$$ which is valid in the local density approximation (LDA). The integral only needs to be evaluated over the density-density correlation length [@Ma10]. We note that number fluctuations were previously suggested as an effective thermometer [@Capo07; @Gerbier06].
- It was also suggested that single-site detection tools can be useful for studying critical phenomena, which remained a controversial topic because of the parabolic confining potential [@Zhou10; @Fang11; @Hazzard11; @Guan10; @Yin11; @Batchelor10; @Zhang11]. In Ref. [@Pollet10_crit] the superfluid to normal transition was analyzed by finding regions where the LDA approximation breaks down. LDA implies quasi-homogeneity of the system when the change of thermodynamic properties of the system is negligible at the distance of the order of the correlation radius. LDA breaks hence down in the critical region. It was found numerically, and supported by analytical arguments, that LDA violations are tiny for the density profiles (because of the low value of the critical exponent $\alpha$) and by extension for the compressibility, which is the derivative of the density with respect to the chemical potential, in contrast to earlier findings [@Zhou09a; @Pollet10_comment; @Zhou10_reply]. On the positive side, it was demonstrated that analyzing the transition using time-of-flight images can be done without fitting the time of flight images. One has to construct the amplitude of the critical signal as a function of temperature, $P_c(T) = n(k=0) - n(k_{\rm max})$ with $n(k)$ the interference patterns at momenta $k$ and $k_{\rm max}$ the momentum at which the absolute value of the first derivative $dn/dk$ has a maximum. By then plotting $Q(T) = P_c(T) k_{\rm max}^s(T)$ with some exponent $s > 2- \eta$ (with $\eta$ the critical exponent for the correlation function at the the critical point which is almost zero for the 3D XY transition), the critical temperature can be read out accurately where $Q(T)$ reaches a sharp minimum. The scheme exploits the fact that the momentum distribution behaves as $n(k) \sim \xi^{2 - \eta}$ for $k \ll \xi^{-1}$ and as $n(k) \sim k^{\eta - 2}$ for $\xi^{-1} \ll k \ll a^{-1}$ with $a$ the lattice spacing. In Monte Carlo simulations, this scheme was shown to work even when taking the finite time-of-flight duration and other optical resolution limiting effects into account, but has not been tried experimentally yet.
- Going into more detail about the issue of scaling theory, Campostrini and Vicari introduced the scaling exponent for the trap [@Campostrini09; @Campostrini10] which shows how critical phenomena for finite systems with a power-lap trap with potential $V(r) = v^p r^p = (r/l)^p$ can be analyzed when making the trap softer. $v$ and $p$ are positive constants, $l=1/v$ is the trap size. A harmonic potential corresponds to $p=2$. The authors focused mostly on the Mott insulator to superfluid transition driven by the chemical potential with dynamical exponent $z=2$ in 1d [@Campostrini10; @Campostrini2010b; @Ceccarelli2012] and more recently also in 2d [@Ceccarelli2012b]. Using hard-core bosons, there are two Mott insulators, namely for $\mu < -2t$ with $\langle n \rangle = 0$ and for $\mu > 2t$ with $\langle n \rangle = 1$ (in 1d), and a superfluid in between. The trap breaks particle-hole symmetry so that these two transitions are not necessarily equal. Following scaling theory, the simplest trap-size scaling Ansatz one can make for the free energy at $T=0$ in $d$ dimensions is, $$F(\mu, T, l, x) \to l^{-\theta(z+d)} \mathcal{F}( \bar{\mu}l^{\theta/\nu}, T l^{z\theta}, x l^{-\theta}).$$ Here, $\bar{\mu} = \mu - \mu_c$ is the detuning, $\theta$ the trap scaling exponent, $x$ the distance to the trap center, and the function $\mathcal{F}$ a universal scaling function insensitive to the microscopic details. Finite size effects (when using a finite box of length $L$) can be taken into account by adding an additional argument $L l^{-\theta}$. A simple analysis leads to $\theta = p / (p+2)$ [@Campostrini09; @Pollet10_crit]. Similar expressions can be written down for the density and density-density correlation function, $$\begin{aligned}
\rho(x) = \langle n_x \rangle & \to & l^{-d\theta} \mathcal{D}(\bar{\mu} l^{\theta / \nu}, T l ^{z \theta}, x l^{-\theta}), \nonumber \\
G(x,y) = \langle n_x n_y \rangle - \langle n_x \rangle \langle n_y \rangle & \to & l^{-2 \theta d} \mathcal{G}( \bar{\mu} l^{\theta / \nu}, T l ^{z \theta}, x l^{-\theta}, y l ^{-\theta}),\end{aligned}$$ and for the gap $$\Delta = l^{-\theta z} \mathcal{K}(\bar{\mu} l^{\theta / \nu}).$$ The above set of equations are often used in Monte Carlo simulations to study crtical phenomena: data collapse on the universal scaling functions can be oberved provided the data are rescaled with the correct RG exponents . Examples of this can be found in Eq. \[eq:FSS\_gap\] as well as in Refs. [@Campostrini10; @Campostrini2010b; @Ceccarelli2012] for trap-size scaling.
The trap-size scaling theory for the low density limit $(\langle n \rangle = 0)$ in 1d for hard-core bosons (at $T=0$) can also be derived analytically from a mapping to quadratic spinless fermions [@Campostrini2010b]. Peculiar behaviour such as discontinuities in the scaled particle density can be related to the quantum nature of the transition. When the filling of the corresponding homogeneous system is nonzero, an infinite number of level crossings occur when increasing the trap size [@Campostrini2010b]. This leads to a modified trap-size scaling which is still controlled by the trap size exponent $\theta$ but shows modulations with a phase measuring the distance to the closest even level crossing. The asymptotic trap-size dependence of the superfluid phase, whose corresponding continuum theory is a conformal field theory with $z=1$, appears also to be modulated and is characterized by two length scales: one scaling as $\xi \sim l$ related to smooth modes, and one scaling as $\xi \sim l^{p/(p+1)}$ involving modes at the Fermi scale $k_F = \pi f$, with $f$ the filling of the homogeneous system [@Campostrini2010b]. At finite temperature, the periodic oscillations in the vicinity of the $ \langle n \rangle = 1$ transition found at $T=0$ vanished rapidly [@Ceccarelli2012]. Similar conclusions hold in 2d [@Ceccarelli2012b]; the type of analysis discussed in this paragraph is expected to be applicable to other models as well.
The authors of Refs. [@Campostrini2010b; @Ceccarelli2012] also investigated the validity of the local density approximation. They find that the LDA approximation of the particle density becomes exact in the large trap-size limit, and corrections are controlled by the trap exponent, but they confirm that corrections to LDA become much larger in the critical region.
- The phase transition from vacuum to superfluid in 2D was studied experimentally in Ref. [@Zhang11]. Previously, universal scaling behaviour was observed in interacting Bose gases in three [@Donner07] and two dimensions [@Hung11b], and in Rydberg gases [@Loew09]. This transition is one of a dilute (in the sense of [@Sachdev99], thus non-interacting in 2D and 3D and the Tonks-Girardeau limit in 1D) Bose gas, and is exactly at its upper critical dimension. One thus expects mean-field exponents $\nu = 1/2$ and $z=2$ with logarithmic corrections that could experimentally not be discerned however. The equation of state follows the scaling $\tilde{N} = F(\tilde{\mu})$ in which $F(x)$ is a universal function, and $\nu$ the inverse of the RG dimension of $\mu$. $$\begin{aligned}
\tilde{N} & = & \frac{N-N_r}{ (T/t)^{d/z + 1 - 1/(z\nu)} } \nonumber \\ %d=2
\tilde{\mu} & = & \frac{ (\mu - \mu_0)/t}{ (T/t)^{1/(z\nu)}} \end{aligned}$$ are the scaled occupation number and scaled chemical potential. $N_r$ is the regular, non-universal part of the occupation number, which is zero for the vacuum to superfluid transition. Monte Carlo simulations for the superfluid to Mott insulator were found in excellent agreement with the experimental data [@Fang11], which allows to determine over which temperature range the zero temperature critical point can be felt.
For completeness, we mention that the worm algorithm was also successfully applied to study the phase diagrams of the 2D Bose-Hubbard model [@CapogrossoSansone08]. The reader can verify that the first path-integral Monte Carlo simulations without worm-type updates had large error bars on the determination of the critical point [@Krauth1991b]. Even with current computers the algorithmic slowdown would prevent an accurate determination of properties when superfluid fluctuations are large (that is certainly for large system sizes) when not using the worm updates ). In 1D, the phase diagram was accurately determined long ago by a density matrix renormalization group study [@Kuehner98; @Kuehner00]. State diagrams were calculated in the presence of a confining potential [@Rigol09; @Hen10; @Mahmud11]. Excitation spectra across the superfluid to Mott insulator transition in 1D were studied by using maximum entropy [@Jarrell96], showing a gapped mode in the strongly interacting superfluid phase [@Pippan09].
Disordered systems and the Bose glass phase
-------------------------------------------
The interplay between disorder and interactions is a long-standing problem in condensed-matter physics. Bosonic systems are especially difficult to handle since the limit of vanishing interactions is pathological: in the ground state, all bosons occupy the same lowest-energy localized state. In the presence of disorder, the concept of lakes is crucial: these are regions where the chemical potential is nearly homogeneous and mimic a uniform system. Such lakes can be arbitrarily large (so the quantization energy can be very low) but they are exponentially rare. Nevertheless, they are crucial for such properties as the existence of a gap in the spectrum. In the presence of disorder, a new phase is possible at zero temperature [@Giamarchi1987; @Giamarchi1988]: the Bose glass phase (BG), which is gapless, compressible but insulating. It defies intuitive notions about conductivity based on Fermi liquid theory. Fisher [*et al.*]{}, building on the one-dimensional work by Giamarchi and Schulz [@Giamarchi1987; @Giamarchi1988], argued the existence of the Bose glass phase in any dimension [@Fisher1989].
In the presence of a lattice, a commensurate system may be driven to a Mott insulator (MI). For a long time, there was controversy whether a direct transition between a superfluid (SF) and a Mott insulator (MI) in the presence of disorder was possible [@Freericks1996; @Scalettar1991; @Krauth1991; @Zhang1992; @Singh1992; @Makivic1993; @Wallin1994; @Pazmandi1995; @Pai1996; @Svistunov1996; @Pazmandi1998; @Kisker1997; @Herbut1997; @Sen2001; @Prokofev2004; @Wu_Phillips2008; @Bissbort2009; @Weichman2008_1; @Weichman2008_2]. Fisher [*et al.*]{} argued that this was unlikely, though not fundamentally impossible [@Fisher1989]. Curiously, a large number of direct numerical simulations and some approximate approaches observed the unlikely scenario. However, Fisher [*et al*]{} also showed that if the bound $\Delta$ on the disorder strength is larger than the half-width of the energy gap $E_{g/2}$ in the ideal Mott insulator, then the system is inevitably compressible and the transition is to the Bose glass phase. More recently, the [*theorem of inclusions*]{} proved that the Bose glass phase always has to intervene between the superfluid and the Mott insulator for generic disorder distributions [@Pollet2009_disorder; @Gurarie2009]. The proof proceeds in two steps by first showing that a direct transition between a gapless and a gapful phase is not possible, and second that the compressibility is finite at the superfluid to Bose glass transition (see Refs [@Pollet2009_disorder; @Gurarie2009] ). A direct consequence is that the transition of the Bose glass to the Mott insulator is of the Griffiths type: the vanishing of the gap at the critical point is due to a zero concentration of rare regions where extreme fluctuations of disorder mimic a regular gapless system, which was already conjectured by Fisher [*et al.*]{} [@Fisher1989]. A Mott glass can occur for zero compressibility, [*e.g.*]{}, for disorder in the hopping of hard-core bosons at half filling, which has particle-hole symmetry [@Pollet2009_disorder; @Prokofev2004].
![ Phase diagram of the disordered 3D Bose-Hubbard model at unit filling, obtained by a finite-size analysis of winding numbers. In the absence of disorder, the system undergoes a quantum phase transition between SF and MI phases. The presence of disorder allows for a compressible, insulating BG phase, which always intervenes between the MI and SF phases because of the theorem of inclusions [@Pollet2009_disorder; @Gurarie2009]. The transition between MI and BG is of the Griffiths type, as an exception implied by the theorem of inclusions [@Gurarie2009]. At $U/t \to 0$, the SF-BG transition line has an infinite slope [@Falco2009]. Reprinted figure with permission from Ref. [@Gurarie2009]. Copyright (2009) by the American Physical Society. []{data-label="fig:phasediagram_disorder"}](fig_phasediagram_disorder){width="0.8\columnwidth"}
The phase diagram of the 3D Bose-Hubbard model with box disorder in the chemical potential and bound $\Delta$ is shown in Fig. \[fig:phasediagram\_disorder\] for unit density. The BG phase always intervenes between the SF and MI, while the transition from the MI to the BG phase is of the Griffiths type. The transition line is then determined by measuring the gaps in the disorder-free MI. At $U/t \to 0$, the SF-BG transition line has an infinite slope going like $\Delta \sim U^{1-d/4}$ [@Falco2009], which is a consequence of Lifshitz-tails. Note that for low values of $U$ the superfluid region is extremely stable against disorder, and the transition to the BG phase reaches a maximum for $\Delta / t \approx 350$, a scale which was explained by percolation and local energy scales in Ref. [@Gurarie2009]. The superfluid densities in the finger of the phase diagram were low, indicating low transition temperatures.
In two dimensions, the topology of the phase diagram is the same as in 3D. Here too, reentrant behaviour is seen and the superfluid phase can exist to very high disorder strengths [@Soyler2011]. However, close to the superfluid to Mott-insulator transition for weak disorder, transition points for the SF-BG transitions could within error bars not be resolved from the transition point for a clean system [@Soyler2011]. In one dimension, Svistunov showed long ago that the BG phase always intervenes between the MI and SF [@Svistunov1996]. The phase diagram was computed by Monte Carlo simulations [@Prokofev1998_comment] and by the density matrix renormalization group [@Rapsch1999], with qualitative agreement. Interestingly, controversy remains about the nature of the transition in one dimension. Giamarchi and Schulz performed a lowest order RG scaling and found that the transition between SF and BG occurs universally at $K = 3/2$ (instead of $K=2$ for the SF-MI transition in the clean system) [@Giamarchi1988]. In the strong disorder scenario put forward in Refs. [@Altman2004; @Altman2008; @Altman2010] the transition for $\Delta/U \gg 1$ is believed to be non-universal with power law distributions of the Luttinger parameter. No simulations exist to date which distinguish between the two scenarios.
Experimentally, the Bose glass phase has never been detected unambiguously. The experiments of Helium-4 adsorbed on disordered substrates were not conclusive in establishing the existence of a Bose glass phase; the findings were better explained by a model which has a constant density of states for low and for high energies, with a gap in between [@Crowell97]. In the cold atom experiments of Ref. [@White09; @Pasienski09] where optical speckles are used to generate the disorder (and in practice only a single disorder realization is used), no distinction between a Mott insulator and a Bose glass phase could be made; only insulating phases could be distinguished from superfluid phases on the basis of time-of-flight interference images and transport. They found insulating phases for disorder strengths several hundred times the tunneling amplitude, in agreement with the quantum Monte Carlo simulations. However, they did not find an insulator to superfluid transition (missing the ’finger’ in Fig. \[fig:phasediagram\_disorder\]). Although the disorder distributions are different in experiment and in simulations, the topology of the phase diagram should be the same. The discrepancy is attributed to the low transition temperature (or equivalently, the low superfluid density at zero temperature in the finger [@Gurarie2009]) while the temperature in experiment is estimated to be well above it.
Finally, we wish to mention that there have been a number of recent experiments on disordered quantum antiferromagnets [@Yamada09; @Hong09; @Yu11]. They are, however, not without criticism either [@Zheludev09; @Wulff11].
Polar molecules
---------------
The influence of long-range interactions is currently also attracting a lot of interest in cold atomic and molecular gases: the first signatures of long-range interactions have been observed for magnetic interactions in $^{52}$Cr [@Griesmaier05_polar; @Lahaye07_polar; @Koch08_polar] with a dipole moment $D = 6\mu_B$ and $^{164}$Dy [@Lu11_polar], which is the most magnetic atom with a dipole moment $d = 10\mu_{\rm B}$. Electric dipole and van der Waals interactions between Rydberg states give rise to intriguing collective phenomena [@Heidemann07_polar]. In addition, there are huge experimental efforts towards the realization of quantum degenerate polar molecules [@Sage05_polar; @Ni07_polar; @Deiglmayr08_polar; @Aikawa10_polar], where the permanent dipole moment of the molecules gives rise to a strong and highly tunable electric dipole-dipole interaction [@Pupillo08_polar; @Buechler07_polar; @Menotti08_polar; @Menotti08b_polar].
We will not review the physics of dense molecular samples here, for which excellent recent reviews are available [@Carr09_polar; @Menotti08_polar; @Menotti08b_polar; @Dulieu09_polar; @Friedrich09_polar], and in particular for lattice systems there is Ref. [@Trefzger11_polar]. Rather, we will report on quantum Monte Carlo studies that explore the possibilities offered by current and future developments in this field. A central role so far was played by the possibility to observe a supersolid phase in these systems [@Pollet10_polar; @Capogrosso-Sansone10_polar].
We consider bosonic polar molecules in a strong electric field along the z direction, which induces the dipole moment $d_z \le d$; here $d$ denotes the permanent dipole moment of the heteronuclear molecule. The dominant interaction between the polar molecules is then given by the dipole-dipole interaction $V(R) = \frac{d_z^2}{4 \pi \epsilon_0} \frac{R^2 - 3z^2}{R^5}$ where the strength of the dipole interaction can continuously be tuned by the strength of the electric field. In addition, the polar molecules are confined into the $xy$ plane by a strong transverse harmonic confinement as can easily be achieved by a strong one-dimensional standing laser along the $z$ direction. The combination of strong transverse trapping and dipole interactions creates a repulsive barrier, which prevents the collapse naturally present in bosonic dipolar gases. The effective long-range two-dimensional potential is then found by integrating over the $z$ direction, and reduces to $V^{2D}_{\rm eff} \sim D/r^3$. We refer to Refs. [@Pupillo08_polar; @Buechler07_polar] for a detailed discussion on how such a potential can be tailored.
![Ground state phase diagram of the 2D Bose-Hubbard model with dipolar interactions with $t$ the tunneling amplitude and $V$ the interaction strength. The phases are a superfluid “SF”, supersolid “SS”, and a commensurate solid at filling factor $n=1/3$. With the double line we indicate a transition region of the Spivak-Kivelson bubble type (emulsions) gradually going over to a region of incommensurate, floating solids with increasing interaction strength. For large interaction strength, and starting around half filling, the supersolid phase is suppressed by emerging solid ordering (stripes at half filling and incommensurate, floating solids at other fillings). Reprinted figure with permission from Ref. [@Pollet10_polar]. Copyright (2010) by the American Physical Society. []{data-label="fig:phasediagram_polar"}](fig_polarmolecule){width="0.8\columnwidth"}
The ground state phase diagram on the triangular lattice is shown in Fig. \[fig:phasediagram\_polar\]. For $V/t > 7.5$ there is an insulating commensurate solid at filling factor $n=1/3$. For densities below $n < 1/3$ a superfluid phase is reached, similar to what is found for the short-range model with nearest-neighbour repulsion only [@Boninsegni05_polar; @Wessel05_polar; @Heidarian05_polar; @Melko05_polar; @Melko06_polar] , but the transition here is different (it’s first order in the short-range model because of domain-wall formation) and of the bubble type introduced by Spivak and Kivelson [@Spivak04_polar]: over a finite but narrow range of chemical potentials, small crystallites form an emulsion of bubbles inside a liquid. Such emulsions are expected to be discernible only on astronomical scales; on finite lattices simulations reveal no difference from a first-order transition. For $V \ge 30$ (not shown) we find the first evidence for additional plateaus at various fillings below $n=1/3$, which are not present in the short-range model. With increasing system size the number of plateaus grows and they are separated by small superfluid regions. It is expected that an incommensurate, floating solid is formed in the thermodynamic limit for strong interactions by analogy to the analysis of Ref. [@Isakov07_polar]. Note that in the classical limit of zero hopping the long-range model exhibits a devil’s staircase (see Refs. [@Bak82_polar; @Burnell09_polar] for 1D) of various solid phases.
Above the commensurate solid at $n=1/3$ we find a continuous second-order phase transition to a supersolid phase belonging to the universality class of the 3D XY model, similar to what occurs in the short-range model. While near the tip the supersolid phase exists only over a narrow filling factor range, it quickly extends ($V/t=15$) all the way to half filling. For larger interactions ($V/t > 20$) and close to half filling, the structure factor and the superfluid density go down and supersolidity is lost for $V/t = 30$ at and near half filling. At finite temperature, the Kosterlitz-Thouless transition to the superfluid phase for weak interactions posed no problem, but the simulations became very difficult at larger interaction strengths. The optimal transition temperature $T_c/t = 0.53(8)$ for supersolidity was found for a density $n=0.4$ and an interaction strength $V/t = 12$, corresponding to a temperature $T=0.8(2)$nK for LiCs.
The model was also studied on the square lattice in Ref. [@Capogrosso-Sansone10_polar]. They also found a superfluid, a solid, a supersolid and various incommensurate solid phases, but the supersolid transition temperature is generally lower than on the triangular lattice. A supersolid phase was found for both higher and lower densities than half filling, unlike short-range models with hard-core bosons [@Boninsegni05_polar; @Wessel05_polar; @Heidarian05_polar; @Melko05_polar; @Melko06_polar]. For hard-core bosons on the square lattice with also next-nearest neighbour hopping [@Chen08_polar; @Dang08_polar], a supersolid was found for densities $n < 1/4$ with star diagonal ordering, and it was also found for densities $0.25 < n < 0.5$ between a star and stripe solid at half filling [@Dang08_polar]. For soft-core bosons on a square lattice a supersolid phase exists for $n> 0.5$ with nearest-neighbour interactions only.
Much more theoretical work has been done on polar molecules with mean-field and analytical methods. We list here a few examples. Polar molecules in their ground state have been suggested to describe quantum magnetism [@Gorshkov11]. The phase diagram of a Bose-Einstein condensate with dipolar interactions loaded into an optical lattice with a staggered flux was studied in Ref. [@Thieleman11_polar]. Apart from uniform superfluid, checkerboard supersolid, and striped supersolid phases, several supersolid phases with staggered vortices were identified, which can be seen as combinations of supersolid phases found in earlier work on dipolar BECs [@Capogrosso-Sansone10_polar; @Pollet10_polar] and a staggered-vortex phase found for bosons in optical lattices with staggered flux.
Finally, we note that fermionic dipolar systems have also received a lot of theoretical interest (in the continuum). They cannot be treated by path integral Monte Carlo methods, and a discussion is beyond the scope of this review. We list a few exotic proposals: detection of a $p_x + ip_y$ fermionic superfluid [@Cooper09_polar], a nematic non-Fermi liquid [@Quantanilla09_polar; @Fregoso09_polar; @Carr10_polar], and an unusual dimerized “pseudogap” state at intermediate temperature for a layered system of polarized fermionic molecules [@Potter10].
Bose-Bose and Bose-Fermi mixtures
---------------------------------
The Mott phase of a single species of atoms suppresses any low-energy transport. If the Mott phase consists of at least two species, then the net number-of-atoms transport is still suppressed. However, counterflow, in which the currents of the two species are equal in absolute values but in opposite directions, can survive and become superfluid under certain conditions [@Kuklov03_twocomponent]. In case of a 2-component Bose-Bose mixture, the super-counter-flow (SCF) corresponds to an easy-plane ferromagnet in the strong coupling limit (or an easy-plane anti-ferromagnet for Fermi-Fermi mixtures). Such a phase has a resulting $U(1)$ symmetry, compared to the $U(1) \times U(1)$ symmetry of the full Hamiltonian. The symmetry breaking and phase transitions at double commensurate filling for various inter and intra-species interactions have been discussed in Ref. [@Kuklov04_twocomponent] for the $J-$current model, which is the $(d+1)$ dimensional classical analog of the $d$ dimensional quantum system. The possible phases are (1) $SF_A$ + $SF_B$ (or 2SF) , which are 2 miscible superfluids with non-zero order parameters $\langle \psi_A \rangle \neq 0$ and $\langle \psi_B \rangle \neq 0$, (2) $SF_A$ + $MI_B$ with $\langle \psi_A \rangle \neq 0$ and $\langle \psi_B \rangle = 0$ and vice versa, (3) $MI_A$ + $MI_B$ with $\langle \psi_A \rangle = 0$ and $\langle \psi_B \rangle = 0$ , and (4) SCF with $\langle \psi_A \rangle = 0$ and $\langle \psi_B \rangle = 0$ but $\langle \psi_A \psi_B^{\dagger} \rangle \neq 0$. For attractive intra-species interactions, it is possible to form a pair-superfluid phase (PSF) with order parameter $\langle \psi_A \psi_B \rangle \neq 0$ for equal densities (directly translating in a SCF phase for repulsive interactions at commensurate fillings). It was shown that the universal properties of the $2SF-PSF$ and $2SF-SCF$ quantum phase transitions can be mapped onto each other and that their universality class is identical to the one of th$(d+1)$ dimensional normal-superfluid transition in a single-component liquid [@Kuklov04_twocomponent]. Off-diagonal long-range order for the PSF and SCF phases is seen in the two-body density matrix. It is hence crucial to allow for updates in the Monte Carlo sampling such that the paths can wind together around the system volume, which can be accomplished by having worm operators for each species simultaneously in configuration space hereby sampling the two-body Green function. The worm operators $Q_{2b}$ in the two-body channel take the form (the worm operators in the one-body channel for each component separately can be kept, but do not lead to an ergodic algorithm on their own) $$Q_{2b} = \sum_{i,j} \int_0^{\beta} d\tau \int_0^{\beta} d\tau' (a_i^{\dagger}(\tau) b_j(\tau') + a_i^{\dagger}(\tau) b_j^{\dagger}(\tau') + a_i(\tau)b_j^{\dagger}(\tau') + a_i(\tau) b_j(\tau')$$ Note that this will result in an algorithm that is a factor $\beta L^d$ slower than the original worm algorithm for the single species case. However, off-diagonal long-range order in the two-body channel matters most when the imaginary times of the worm operators are close to each other. Hence, when inserting the worm-pair of the second species, one should insert this pair closely to either the worm head or tail of the other species. One can further restrict the sampling of the two-body Green function (and reducing the algorithmic slowdown) by considering a reduced worm pair [@Ohgoe11_twocomponent] $Q = \sum_i \int_0^{\beta} d\tau ( a_i^{\dagger}(\tau) b_i(\tau) + a_i(\tau) b_i^{\dagger}(\tau))$ (for SCF and analogously for PSF) and adding an appropriate set of one-body worm operations of the same type ($a$ or $b$).
![Ground state phase diagram of the 2D two-component Bose-Hubbard model at double half-integer filling. The intra-species on-site interactions are infinitely strong (hard-core limit), while the inter-species interaction, $U$, is finite. The respective hopping amplitudes are $t_a$ and $t_b$, and $z = 4$ is the coordination number. The revealed phases are (i) a checkerboard solid in both components or z-N[é]{}el phase (2CB), (ii) a checkerboard solid in one component and superfluid in the other (CB+SF), (iii) a superfluid in both components (2SF), (iv) a super-counter-fluid or XY-ferromagnet (SCF). The observed transition lines are: 2CB-SCF (first-order), SCF-2SF (second-order), 2CB-2SF (first-order), 2CB- CB+SF (second-order), and CB+SF-2SF (first-order). Reprinted figure with permission from Ref. [@Soyler09_twocomponent]. Copyright (2009) by the American Physical Society. []{data-label="fig:twocomponent_phasediagram"}](Twocomponent_phasediagram){width="0.8\columnwidth"}
The ground state phase diagram of the quantum two-component (’a’ and ’b’) Bose-Hubbard model for hard-core bosons with nearest-neighbour hopping and on-site intra-species interaction in the Hamiltonian $$H = -t_a \sum_{\langle i,j \rangle} a_i^{\dagger} a_j - t_b \sum_{\langle i,j \rangle} b_i^{\dagger} b_j + U \sum_i n^a_i n^b_i$$ at double half filling has been analyzed in Refs. [@Soyler09_twocomponent; @Capogrosso-Sansone10_twocomponent] and is shown in Fig. \[fig:twocomponent\_phasediagram\]. It displays all the insulating and superfluid phases anticipated in the previous paragraph. The transitions at finite temperature have also been analyzed in 2D and 3D [@Capogrosso-Sansone10_twocomponent]. For the Ising case in 3D (2CB), $T_c$ was found to be $T_c / t_b = 0.175(15)$ for $U/t_b = 11, t_a / t_b = 0.1$, which is where $T_c$ is expected to be the largest. This corresponds to an entropy per particle of $S/N = 0.5(1)$. For the melting of the $xy-$ferromagnet (SCF) in 3D, $T_c/t_b = 0.208(7)$ for $U/t_b = 21, t_a / t_b = 1$ was reported, with a critical entropy per particle of $S/N = 0.35(4)$ [@Capogrosso-Sansone10_twocomponent]. In 2D, the critical entropy for the Ising transition is about a factor of 2 lower than in 3D, for the $xy-$ferromagnetic transition it is an order of magnitude lower because of the Kosterlitz-Thouless nature of the transition. The entropies in 3D are within reach of current experimental setups; however the time scales needed for establishing magnetic ordering are large compared to the lifetime and remain challenging to achieve [@Capogrosso-Sansone10_twocomponent]. The phase diagram of the two-component Bose-Hubbard model was also obtained in the dynamical mean-field theory (see Sec. \[sec:bdmft\]) approximation [@Hubener09_twocomponent]. The change of the shape of the Mott lobe upon increasing the concentration of the second Mott lobe was studied in Ref. [@Guglielmino10_twocomponent], and the realization of such phases in a system with a parabolic trap was discussed in Ref. [@Guglielmino11_twocomponent].
Mixtures including conversion terms have been studied on a 2D lattice in Ref. [@Forges11_twocomponent]. When interspecies interactions are smaller than the intraspecies ones, the system is unpolarized, whereas in the opposite case the system is unpolarized in even Mott insulator lobes and polarized in odd Mott lobes and also in the superfluid phase. In the latter case, the transition between the Mott insulator of total density $2$ and the superfluid can be of either second or first order depending on the relative values of the interactions, whereas the transitions are continuous in all other cases.
The one-dimensional phase diagram of a two-component bosonic system where additionally pairs of type A can locally be annihilated and simultaneously pairs of type B be created was studied in Ref. [@Forges10_twocomponent]. The phase diagram of the ferromagnetic case (positive pair hopping) was found in close agreement to the phase diagram found with spinor bosons (see Sec. \[sec:spinor\]). In this case, the population is always balanced. In the antiferromagnetic case however, the superfluid phase is always polarized. The second Mott lobe has population balance, but a transition inside the first Mott lobe from a balanced population for strong interactions toward a polarized Mott insulator for weak interactions was found.
One-dimensional Bose-Fermi mixtures (with spin polarized fermions) are, after a Jordan-Wigner transformation on the fermions, also amenable for path integral Monte Carlo studies. Most of the reported works were done in the canonical ensemble using the worm algorithm variants of Ref. [@Pollet06_twocomponent]. The possible phases translate directly to their higher dimensional analogs mentioned above [@Pollet06_twocomponent; @Pollet08_twocomponent; @Hebert08_twocomponent; @Hebert07_twocomponent; @Zujev08_twocomponent; @Varney08_twocomponent]. In particular at double half filling with equal hopping amplitudes, phases consisting of two miscible Luttinger liquids (corresponding to 2SF), a spin density wave (corresponding to SCF), and a ferromagnet (phase separation) were found [@Pollet06_twocomponent]. Here too, in case the boson-boson and the fermion-boson repulsion is very large, a mapping to a spin model can be done. The [*XXZ*]{} Hamiltonian $H^{\rm XXZ} = \sum_i J (\sigma^x_i \sigma^x_{i+1} + \sigma^y_i \sigma^y_{i+1}) + J_z \sigma_i^z + \sigma_{i+1}^z$ with $J = - \frac{t_B t_F}{U_BF}$ and $J_z = \frac{t_B^2 + t_F^2}{2U_{\rm BF}} - \frac{t_b^2}{U_{BB}}$ has a first order transition from the gapless SDW phase toward the ferromagnet for $J=J^z$ or $U_{\rm BF} / U_{\rm BB} = 2$ for $t_b = t_F = 1$. When the hopping amplitudes are unequal, a transition between the gapless SDW phase and an Ising Neel phase is also possible [@Pollet06_twocomponent].
Cold atom experiments with a $^{87}$Rb - $^{41}$K Bose-Bose mixture [@Catani08_twocomponent] or a $^{87}$Rb - $^{40}$K Bose-Fermi mixture [@Guenter06_twocomponent; @Ospelkaus06_twocomponent] focused on the loss of coherence of the $^{87}$Rb bosonic atoms when adding the second species, notwithstanding their different statistics and interspecies interactions (even the sign!). By using a Feshbach resonance the interaction strength between the $^{87}$Rb bosons and the $^{40}$K fermions could be tuned [@Best09_twocomponent], where a strong asymmetry between the repulsive and the attractive side was found. For a bosonic mixture of the $\vert F=1, m_F = -1 \rangle$ and $ \vert F=2, m_F = -2 \rangle$ hyperfine levels of $^{87}$Rb atoms, it was found that the presence of a second component can reduce the apparent superfluid coherence, most significantly when the second component either experiences a strongly localizing lattice potential or none at all [@Gadway10_twocomponent]. Mixtures with different bosonic and fermionic isotopes of Yb were also recently studied experimentally in an optical lattice [@Sugawa11_twocomponent]. It was found that an interspecies interaction between bosons and fermions leads to drastic changes, causing effects that include melting, generation of composite particles, an anti-correlated phase and complete phase separation. A number of theoretical explanations which include heating in order to conserve entropy [@Guenter06_twocomponent; @Pollet06_twocomponent; @Cramer08_twocomponent; @Cramer11_twocomponent] , the multiband model [@Tewari09_twocomponent], the self-trapping effect [@Luehmann08_twocomponent; @Best09_twocomponent], changes in the chemical potential due to the presence of external harmonic confinement [@Buonsante08_twocomponent], and polaronic effects [@Guglielmino10_twocomponent; @Gadway10_twocomponent] have been put forward to explain the experimental observations. In general, it is fair to state that for multi-component systems a similar quantitative understanding as for the single species Bose-Hubbard model is still lacking.
In a series of experiments at MIT [@Weld09_twocomponent; @Weld10_twocomponent; @Medley11_twocomponent] a magnetic field gradient was used to initially separate the two components spatially. The region of overlap forms a domain wall, and its width can be used as a thermometer. It can be shown that the average magnetization $ \langle s \rangle = {\rm tanh}( - \beta \Delta {\mathbf \mu B}(x)/2)$, with $\Delta {\mathbf {\mu}}$ the difference in magnetic moment between the two components, and ${\mathbf B}(x)$ the position-dependent magnetic field. The temperature is hence directly measurable, and temperatures of the order of a few tens of pK have been reported. The latter is obtained by spin gradient demagnetization: the magnetic field gradient is slowly reduced, so that the two components can mix over a wider region where more states become available hereby reducing the temperature [@Medley11_twocomponent].
A different route to quantum magnetism was followed in Ref. [@Simon11_twocomponent], following a proposal by S. Sachdev [*et al.*]{} [@Sachdev02_twocomponent] in the context of the first superfluid-to-Mott insulator transition experiments [@Greiner2002] with single species. If a tilt is applied to the lattice, then atoms prepared in a Mott insulator, can only tunnel if the linear tilt $E$ equals the potential energy. Hence, if $E = U$, an atom can tunnel to a neighbouring site, provided the atom there has not tunneled yet. This creates an effective spin-spin interaction within the resonant subspace. The quantum phase transition from a paramagnetic to an antiferromagnet was observed in Ref. [@Simon11_twocomponent] in 1D. The main advantage of this scheme is that the relevant time scale is set by the hopping $t$ and not by the superexchange scale $\sim t^2/U$ and that the initial entropies in the Mott insulator can be very low. More sophisticated lattice geometries will produce frustrated systems with novel quantum liquid and dimer covered ground states [@Pielawa11_twocomponent].
Spinor bosons {#sec:spinor}
-------------
![Phase diagram of the spin-1 Bose-Hubbard model in 1D for $U_2 / U_0 = -0.1$. Reprinted figure with permission from Ref. [@Batrouni09]. Copyright (2009) by the American Physical Society. []{data-label="fig:spinor_phasediagram"}](Spinor_phasediagram){width="0.8\columnwidth"}
Bosons with an internal spin degree of freedom on a lattice are described by the following spin Bose-Hubbard model, $$H = -t \sum_{\langle i,j \rangle, \sigma} b_{i, \sigma}^{\dagger}b_{j,\sigma} + \frac{U_0}{2} \sum_i n_i(n_i-1) + \frac{U_2}{2} \sum_i ( F_i^2 - 2 n_i),
\label{eq:BoseHubbard_spin}$$ in standard lattice notation. The spin operator $F_i = \sum_{\sigma}^{\dagger} F_{\sigma, \sigma'} a_{\sigma'}$ with $F_{\sigma, \sigma'} $ the standard spin-F matrices contain contact interactions as well as interconversion terms between the spins. This model has been studied for $F=1$ in 1D in the ground state by using QMC [@Apaja04; @Batrouni09] and DMRG [@Rizzi05; @Bergkvist06]. When $U_2 < 0$ ferromagnetism is favoured. The resulting phase diagram is shown in Fig. \[fig:spinor\_phasediagram\]. Compared to the phase diagram of scalar bosons, the main difference is the shrinking of the base of the Mott lobes on the $y$-axis. The transition between the ferromagnetic Mott lobes and the ferromagnetic superfluid is continuous. When $U_2 > 0$ ([*e.g.,*]{} $^{23}$Na), low total spin states are favored. In the absence of hopping, the base of the even Mott lobes grow at the expense of the odd ones, which disappear entirely for $U_0 = 2U_2$. The transition between even(odd) lobes is first(second) order. Mean-field theory [@Demler02; @Snoek04; @Imambekov03; @Imambekov04; @Krutitsky04; @Kimura05; @Pai08] predicts for $2dU_2/U_0 < 0.1$ in $d=2,3$ that, when $t/U_0^{c1} \sim \sqrt{U_2/4dU_0}$, the Mott lobes of even order are comprised of two phases: (a) a singlet phase for $t/U_0 \le t/U_0^{c1}$ and (b) a nematic phase for $t/U_0^{c1} \le t/U_0 \le t/U_0^{c}$ where $t/U_0^{c}$ is the tip of the Mott lobe. The nematic phase breaks spin rotational symmetry but preserves time reversal symmetry and has gapless spin-wave excitations. The spin-singlet phase does not break spin symmetry and has a gap to all excitations. Inside the Mott lobe, mean-field theory predicts that the nematic-to-singlet transition is first order. Mean-field theory also predicts for $2dU_2/U_0 > 0.1$ in $d=2,3$ that even Mott lobes are entirely in the singlet phase and odd Mott lobes entirely nematic. The superfluid polar phase has broken spin rotational symmetry without breaking time reversal symmetry.
The 1D QMC simulations of [@Batrouni09] however find only signs of a crossover between the nematic and singlet phase inside the even Mott lobes by investigating the behaviour of $\langle F^2 \rangle \to 0$. The first Mott lobe was entirely dimerized, as was also found in an earlier study [@Apaja04] and in mean-field theory [@Imambekov03].
In experiments, the dipolar ferromagnetic interactions of $^{87}$Rb have given rise to spin textures after a rapid quench across the ferromagnetic phase transition [@Sadler06; @Vengalattore08], when simultaneous magnetic ordering and atomic superfluidity was observed. So far, no experiments have been performed on spinor bosons in an optical lattice.
It would be very interesting to extend the present QMC studies to higher dimensions and check in particular if there is a true nematic-to-singlet transition inside the even Mott lobes for $U_2 > 0$. Also the nature of the transitions between the phases remain poorly understood. Algorithmic advances are needed to study bigger system sizes than the $\sim 30$ sites studied so far. It would also be interesting to compute with QMC the entropy at the phase transitions (so far only mean-field results exist [@Mahmud10]), both with and without a confining potential, to see under what conditions an experimental realization of $F=1$ and possibly higher spin $F$ systems is feasible.
Path Integral Monte Carlo : continuous space models {#sec:PIMC_continuous}
===================================================
After having described path integral Monte Carlo simulations on a lattice with continuous imaginary time, we move on to path integral Monte Carlo methods in continuous space, which can only be formulated with discrete imaginary time. The applications will be similar to the ones discussed for lattice systems: scalar bosons and the superfluid to normal transition, disordered scalar bosons and supersolids for bosons with long-range interactions.
Methods
-------
![Illustration of the “swap” or “reconnect” update in the worm algorithm for continuous-space models. The x-axis denotes space, the y-axis imaginary time. In (a) there are two full paths, and two segments which terminate on the worm head (I) and the worm tail (M). The worm head is at time slice $j$. A slice $j + \bar{m}$ is chosen. After the update (b), the worm head has jumped to a neighbouring path, and a new segment is generated from time slice $j$ to time slice $j + \bar{m}$, connecting the old position of the worm head with the previously existing path. The previously existing path between $j$ and $j + \bar{M}$ is erased. Reprinted figure with permission from Ref. [@Boninsegni05]. Copyright (2006) by the American Physical Society. []{data-label="fig:PIMC_swap"}](fig_pimc_swap){width="0.8\columnwidth"}
We consider a many-body system with Hamiltonian $$H = H_1 + H_2 = -\frac{\hbar^2}{2m} \sum_i \nabla^2_i + \sum_{i < j} V( \vert {\mathbf r}_i - \mathbf{r}_j \vert ),$$ where $V$ is a pairwise interaction depending only on the relative distance between the particles. Note that the algorithm can straightforwardly be formulated for more general interactions, but we restrict ourselves to the standard case of a central potential for simplicity. $H_1$ is the kinetic energy term; $H_2$ the potential energy term. The notation ${\mathbf R} \equiv ({\mathbf r}_1, \mathbf{r}_2, \ldots \mathbf{r}_N)$ will be used to denote the positions of all particles in the system.
The partition function is the trace over the density operator, $Z = {\rm Tr} e^{-\beta H}$. The position-space density matrix is $\rho( {\mathbf R}, {\mathbf R'}, \beta) = \langle {\mathbf R} \vert e^{-\beta H} \vert {\mathbf R'} \rangle$. Note that the product of two density matrices is again a density matrix since $e^{(-\beta_1 + \beta_2)H} =e^{-\beta_1H} e^{-\beta_2H}$. When repeating this $M$ times, and defining the time step $\tau = \beta / M$, one has a discrete path. In general, the kinetic energy operator and the potential energy operator do not commute, but for $M$ large enough, the primitive approximation $$e^{-\tau (H_1 + H_2)} \approx e^{-\tau H_1} e^{-\tau H_2}$$ can be used. It has an error of order $\tau^2$ which can be neglected. According to the Trotter formula, $e^{-\beta (H_1 +H_2)} = \lim_{M \to \infty} [ e^{-\tau H_1 }e^{- \tau H_2}]^M$, valid for self-adjoint operators $H_1$, $H_2$ and $H_1 + H_2$ with a spectrum bounded below, it becomes exact for $M \to \infty$ [@Trotter59]. In practice one uses a higher-order scheme such as the Chin formula [@Chin02], but we continue with the primitive approximation in order not to overload the notation.
For the evaluation of the position-space density matrix in the primitive approximation, $$\rho( {\mathbf R}_0, {\mathbf R}_2, \tau) \approx \int d {\mathbf R}_1 \langle {\mathbf R}_0 \vert e^{-\tau H_1} \vert {\mathbf R}_1 \rangle \langle {\mathbf R}_1 \vert e^{-\tau H_2} \vert {\mathbf R}_2 \rangle,$$ we need to know the kinetic and the potential energy density matrices. Since the potential energy is diagonal in position space, its matrix element is trivial, $\langle {\mathbf R}_1 \vert e^{-\tau H_2} \vert {\mathbf R}_2 \rangle = e^{-\tau H_2 ( {\mathbf R}_1) } \delta({\mathbf R}_1 - {\mathbf R}_2)$. The kinetic energy term is a one-body operator and can be diagonalized by a Fourier transform. Replacing the finite sum over the momenta of a finite box by a continuous integral (see also Ref. [@Ceperley95]), $$\langle {\mathbf R}_0 \vert e^{-\tau H_1} \vert {\mathbf R}_1 \rangle = (4 \pi \hbar^2 \tau / (2m) )^{-dN/2} e^{-\frac{2m({\mathbf R}_0 - {\mathbf R}_1)^2}{4 \hbar^2 \tau} }.
\label{eq:PIMC_gauss}$$ It is the appearance of $\tau$ in the denominator of the exponential that prevents a formulation of the algorithm without Trotter error. We now arrive at a discrete path integral expression for the density matrix in the primitive approximation, $$\begin{aligned}
\rho( {\mathbf R}_0, {\mathbf R}_M, \beta) & = & \int d {\mathbf R}_1 \ldots d {\mathbf R}_{M-1} (4 \pi \hbar^2 \tau / (2m))^{-dNM/2} \nonumber \\
{} & {} & \exp \left( -\sum_{m=1}^{M} \left[ \frac{ 2m({\mathbf R}_{m-1} - {\mathbf R}_m)^2 }{4 \hbar^2 \tau} + \tau V({\mathbf R}_m) \right] \right).\end{aligned}$$ Because of the indistinguishability of bosons, the density matrix is always understood as the sum over all permutations, $$\rho( {\mathbf R}_0, {\mathbf R}_M, \beta) = \frac{1}{N!} \rho( {\mathbf R}_0, {\mathcal P} {\mathbf R}_M, \beta)$$ Evaluation of physical observables requires the sampling over all possible paths: all possible positions of the particles and all possible exchanges [@Ceperley95]. Superfluid properties for large systems can only be simulated efficiently using the same ’worm’-idea we have seen in Sec. \[sec:PIMC\_lattice\]: one works with open segments of paths (or worldlines) instead of closed worldlines, a procedure which directly samples the Green function. Off-diagonal long-range order can then be sampled efficiently since the one-body density matrix (that is the equal-time Green function) has a large weight there. For the updates we refer the reader to the full discussion in Refs. [@Boninsegni05; @Boninsegni06]. The crucial update is the “swap” or “reconnect” update, in which the worm head jumps to a neighbouring worldline, hereby performing a bosonic exchange. This is illustrated in Fig. \[fig:PIMC\_swap\]. The update can be done with acceptance factors of order unity [@Boninsegni05; @Boninsegni06]. An additional advantage of the worm algorithm is that it works directly in the grand-canonical ensemble.
When the worm algorithm for continuous space was introduced in Refs. [@Boninsegni05; @Boninsegni06] it was illustrated for the U(1) transition between normal and superfluid $^4$He in 2D (which is a Kosterlitz-Thouless transition) and 3D (which is the so-called [*lambda*]{} transition) with a few thousand particles, which is roughly a factor 100 higher than what was previously achievable. The transition temperature $T_c$ agreed better than $0.5\%$ with experiment. The deviation is explained by the neglect of three-body (and higher) interactions in the Aziz potential [@Aziz79]. Access to larger particle numbers was crucial in the study of solid $^4$He and the existence of a possible supersolid phase. Path integral Monte Carlo simulations have shown that the ideal hcp solid is an insulator [@Boninsegni2006_helium_1; @Clark2006] and that vacancies phase separate [@Boninsegni2006_helium_2], but that defects such as grain boundaries [@Pollet2007_helium] and dislocations [@Boninsegni2007_helium] may be superfluid depending on the elastic properties [@Pollet2008_helium]. The controversy on the issue of supersolid defects in Helium still continues but this discussion is beyond the scope of this review. For a recent review on the properties of (super)solid Helium, see Ref. [@Prokofev2007], and for a trend article see [@Kuklov2011_helium].
Weakly Interacting Bose gas {#sec:WIBG_Tc}
---------------------------
![ Critical temperature of the 3D dilute bose gas as function of the gas parameter $na^3$. The symbols labeled by PRA04 correspond to the results of Ref. [@Nho04], the ones labeled by PRL97 correspond to Ref. [@Grueter97]. The dashed line (green) is the expansion (1) of Ref. [@Kashurnikov01] and the dotted line (black) is the expansion of Ref. [@Arnold01a] including logarithmic corrections. The solid line (red) is a guide to the eye. Reprinted figure with permission from Ref. [@Pilati08]. Copyright (2008) by the American Physical Society.[]{data-label="fig:PIMC_dilutegas"}](fig_bosegas_tc){width="0.6\columnwidth"}
For a system such as superfluid $^4$He or solid $^4$He, the pairwise potential is typically the Aziz-potential [@Aziz79], which looks very similar to a Lennard-Jones potential. For cold atoms however, the true interatomic potential is often poorly known. Because of the diluteness of the system and the low momenta and energies involved in typical collisions, only the $s-$wave scattering length matters, provided the effective range is small enough. Different model potentials can then be used such that the low-energy scattering properties are reproduced. Typical choices include a hard-sphere potential ($V^{\rm HS}(r) = \infty, r < a$ and zero otherwise), a soft-sphere potential ($V^{\rm SS}(r) = V_0, r < R_0$ and $V_0 > 0$, and the potential is zero otherwise), or a negative power potential ($V^{\rm NP}(r) = \alpha/r^p$ with $\alpha > 0$ and $p > 3$). In all cases the scattering length $a$ can be computed analytically; in particular for hard-sphere potentials the cutoff is the same as the scattering length while for soft-sphere potentials it is always larger [@LandauLifshitz].
Instead of the primitive approximation (or a higher order scheme) it is often advantageous to use the pair-product approximation, $$\rho( {\mathbf R}, {\mathbf R}', \tau) = \prod_{i=1}^{N} \rho_1({\mathbf r}_i, {\mathbf r}'_i, \tau) \prod_{i<j} \frac{\rho_{\rm rel}({\mathbf r}_{ij}, {\mathbf r}'_{ij}, \tau )} {\rho^0_{\rm rel} ({\mathbf r}_{ij}, {\mathbf r}'_{ij}, \tau ) },$$ where $\rho_1$ is the single-particle density-matrix (Eq. \[eq:PIMC\_gauss\]) and $\rho_{\rm rel}$ is the two-body density matrix of the interacting system which depends only on the relative coordinates ${\mathbf r}_{\rm ij} = {\mathbf r}_i - {\mathbf r}_j$. The latter is divided by the corresponding ideal-gas term, which is given by Eq. \[eq:PIMC\_gauss\] with the replacement $m \to m/2$. The computation of the two-body density matrix requires only solving a (radial) Schr[ö]{}dinger equation [@Pollock84_PPA; @Pollock87_PPA; @Ceperley95; @Pilati06], but in the special case of hard-sphere potential the analytic approximation introduced by Cao and Berne [@CaoBerne] is very accurate and therefore used in practice, $$\frac{\rho^{\rm CB}_{\rm rel}({\mathbf r}_{ij}, {\mathbf r}'_{ij}, \tau } {\rho^0_{\rm rel} ({\mathbf r}_{ij}, {\mathbf r}'_{ij}, \tau )} = 1 - \frac{a(r+r') - a^2}{rr'} e^{- [rr' + a^2 -a(r+r') ] (1 + \cos \theta) m / (2 \hbar^2 \tau)},$$ with $a$ the cut-off (equivalent to the scattering length) of the hard-sphere potential and $\theta$ the angle between ${\mathbf r}$ and ${\mathbf r}$’. For a hard-sphere and soft-sphere potential there is no choice but to use the pair-product approximation since the primitive approximation is not valid when the potential has discontinuities.
This method has been used to compute the equation of state of the Bose gas both in the normal and the superfluid phase [@Pilati06]. It was also used in Ref. [@WIBG] to compare the properties of the dilute gas in 1D, 2D and 3D against an improved Beliaev diagrammatic technique. The worm algorithm also allowed for a more precise determination of the critical temperature of the 3D dilute Bose gas [@Pilati08]. Without interactions, the universality class of Bose-Einstein condensation belongs to the Gaussian complex-field universality class, but with interactions this changes to an XY model. Thus, the critical temperature $T_c$ with interactions cannot perturbatively be obtained from $T_c^0 = (2\pi\hbar^2/mk_B)[n/\zeta(3/2)]^{2/3}$ with $\zeta(3/2) = 2.612)$, which is the transition temperature for the non-interacting model [@Baym99]. The deviation from $T_c^0$ is parametrized as $$T_c = T_c^0(1 + c(an^{1/3})).
\label{eq:bosegas_tc}$$ The linear change in the scattering length was predicted by Lee and Yang in 1958 [@LeeYang], but no information on the numerical coefficient $c$ was provided, not even its sign. Rigorous upper bounds on $T_c$ can be proved [@Seiringer08], but they are much weaker (going as the root of $na^3$). Ref. [@Seiringer08] also provides an overview of different predictions for $T_c$. The numerical coefficient $c$ was calculated in Refs. [@Kashurnikov01; @Arnold01b] by solving the effective 3D classical $\vert \psi \vert^4$ model using lattice Monte Carlo simulations, clearly establishing the linear behaviour in Eq. \[eq:bosegas\_tc\]. The reported (universal) value is $c=1.29(5)$. The same classical model was used in Ref. [@Arnold01a] to calculate higher order logarithmic corrections to Eq. \[eq:bosegas\_tc\]. In Refs. [@Grueter97; @Nho04] conventional PIMC was used, but simulations suffered from small particle numbers and an inefficient calculation of superfluid properties, and even large discrepancies at high densities between the two simulations were found. Conventional path integral Monte-Carlo was also applied to the trapped system with in two dimensions [@Krauth1996]. The worm algorithm simulations of Ref. [@Pilati08] addressed considerably larger particle numbers (up to $10^6$ for $na^3 = 5 \times 10^{-3}$) than Refs. [@Grueter97; @Nho04] and number of time slices. For low values of the gas parameter (in the universal regime), the results agree with the classical field calculations. For higher densities $T_c$ first increases, reaches a maximum, and then decreases below $T_c^0$. Simulations with a hard-sphere and a soft-sphere potential gave the same answer for $na^3 \le 10^{-4}$, which is higher than the estimate $na^3 \le 10^{-6}$ for the validity of Eq. \[eq:bosegas\_tc\]. This is illustrated in Fig. \[fig:PIMC\_dilutegas\]. Also the Kosterlitz-Thouless transition of the dilute Bose gas in 2D was analyzed for large system sizes in Ref. [@Pilati08], with good agreement with classical $\vert \psi \vert^4$ theories up to quite large densities. The weakly interacting 2D Bose gas in a harmonic trap was investigated in Ref. [@Holzmann2008; @Holzmann10] with conventional path-integral Monte Carlo simulations, addressing the Kosterlitz-Thouless transition and the fluctuation regime and universal properties, respectively.
Disordered systems
------------------
Disordered bosons in continuous space were studied in Refs. [@Pilati_disorder1; @Pilati_disorder2]. Without a lattice there are no commensurability effects and hence the interplay between a Bose glass phase and a Mott insulator is absent. In Refs. [@Pilati_disorder1; @Pilati_disorder2] the suppression of $T_c$ caused by the disorder was addressed. The disorder was modeled by an isotropic 3D speckle potential. Unlike the disorder-free system $T_c$ changed considerably between $na^3 = 10^{-4}$ and $na^3 = 10^{-6}$ (An older study found no substantial drop in $T_c$ [@Gordillo00]). Agreement with a perturbative approach for $\delta$-correlated disorder could not unambiguously be established since the precision in the weak disorder limit was not high enough, and where deviations from $T_c^0$ become appreciable, the perturbative approach is no longer valid [@Lopatin02]. In the regime of weak interactions and strong disorder, the superfluid transition turns out to be well characterized by the existence of a mobility edge, separating localized from extended states, which is largely independent of temperature and interaction strength. In the regime of strong disorder, strong interactions and low temperatures, a phase where the gas is both normal and highly degenerate (with an energy scaling as $\sim T^2$) was identified, which should be related to the Bose glass phase predicted at $T = 0$.
Supersolids and long-range interactions
---------------------------------------
![ Snapshots of a system of bosons interacting via the potential $V(r) = D/a^3, r \le a$ and $V(r) = D/r^3, r > a$. The dimensionless interparticle distance is $r_s = 1/\sqrt{nr_0^2} = 0.14$ with $n$ the density and $r_0$ the characteristic length given by $r_0 = mD/\hbar^2$. The cutoff of the potential is set by $a/r_0 = 0.3$. Snapshots for four temperatures are shown, (a) $T = 200 D/r_0^3$, (b) $T = 20 D/r_0^3$, (c) $T = 1.0 D/r_0^3$, and (d) $T = 0.1 D/r_0^3$. At high temperature a classical gas is observed. When lowering the temperature, droplets form that become phase coherent at the lowest temperature. Reprinted figure with permission from Ref. [@Cinti10_SS]. Copyright (2010) by the American Physical Society.[]{data-label="fig:droplet"}](fig_droplet){width="0.6\columnwidth"}
Since the experiments by Kim and Chan on solid $^4$He [@KimChan04a; @KimChan04b], interest in the supersolid phase has never faded. It is now generally accepted that the ground state of $^4$He is an insulating, commensurate solid. The question can hence be asked whether there exist supersolids for a single species in continuous, uniform space interacting via a pair potential. Going back to the work of Gross [@Gross57; @Gross58], it is expected on the basis of a Gross-Pitaevskii picture [@Josserand07_SS; @Henkel10_SS] that a model with interactions $V(r) = V_0, r \le a$, and zero otherwise, can be supersolid for certain densities (the interparticle distance has certainly to be smaller than the cutoff $a$ and of the order of $a/2$). This is indeed the case. Several potentials with a smoother behaviour were also tried [@Cinti10_SS] such as $V(r) = 1/ (a+r)^3$ and $V(r) = V_0, r \le a$ and $V(r) = V_0/ (r/a)^{\alpha}, r > a$ with $\alpha = 3$ or $6$ all lead to supersolid phases. The picture is one of droplets that become phase coherent as one lowers the temperature, as can be seen in Fig. \[fig:droplet\]. All of these potentials can be tailored using polar molecules or Rydberg atoms [@Henkel10_SS; @Pupillo10_SS].
A pure $\sim 1/r^3$ interaction in 2D does not have a supersolid phase [@Buechler07_polar; @Astrakharchik07_SS; @Mora07_SS]. At weak interactions, a normal fluid is found at high temperature which undergoes a Kosterlitz-Thouless transition to a superfluid [@Filinov10_SS]. At stronger interactions in the ground state there is a commensurate triangular solid [@Buechler07_polar] which is reached from the superfluid phase by a Spivak-Kivelson type transition [@Spivak04_polar], but which has never directly been seen in simulations because of the large system sizes that are required. Note that emulsions of solid immersed in the liquid phase [*can*]{} be interpreted as a supersolid phase [@Kuklov2011_helium].
Dipolar systems have also been studied inside a parabolic potential [@Lozovik04_SS; @Pupillo10_SS; @Jain11_SS]. It was also shown that a mixture of equal-mass dipolar isotopes, in such a configuration, de-mixes at finite temperature due to quantum statistical effects [@Jain11b_SS].
Diagrammatic methods {#sec:diagrams}
====================
A diagrammatic expansion for some relevant quantity $Q$ ([*e.g.*]{}, a Green function) is a series of integrals with an ever increasing number of integration variables, $$Q(y) = \sum_{m=0}^{\infty} \sum_{\xi_m} \int \mathcal{D} (\xi_m, y, x_1, \ldots, x_m) dx_1 \ldots dx_m.
\label{eq:diagram_expansion}$$ Here, $y$ is a set of parameters on which the quantity $Q$ can depend, $\xi_m$ are indices for different terms of the same expansion order $m$, and the $x$ are integration variables [@Prokofev98_froehlich; @VanHoucke08]. In diagrammatic Monte Carlo, these integrals are not performed explicitly (unlike analytical methods) but configurations are generated for specific values of $m, \xi_m$ and $x_1 \ldots x_m$. The summation over these variables is done by Monte Carlo sampling. Almost all methods considered in this paper can be written in the form of Eq. \[eq:diagram\_expansion\]; they only differ in the choice of the expansion parameter, the representation, and the presence or absence of a positive-definite expansion (Frobenius theorem).
Path integral methods are based on a strong coupling expansion, which is an expansion in the kinetic energy. For bosonic systems this led to a sign-free expansion, but for fermionic or frustrated magnetic systems such an expansion is usually not sign positive. The sign problem is usually so bad (because of the scaling with the system volume) that such algorithms make no sense. This chapter is dedicated to methods based on an expansion in the two-body potential energy term $H_1$ of the Hamiltonian, or in $U$ for short, which offers new opportunities: (i) it is a standard perturbative many-body expansion which can be combined with textbook many-body techniques, and also formulated directly for real time and frequencies [@Mahan; @NegeleOrland; @FetterWalecka], (ii) the one-body problem $H_0$ is quadratic, so we can use Wick’s theorem, (iii) the problem can be formulated for a finite lattice or immediately in the thermodynamic limit. On the negative side, (i) the series convergence is not guaranteed and this is problem and parameter dependent ([*e.g.*]{}, for bosons in the thermodynamic limit, a naive expansion in $U$ diverges because attractive bosons lead to a collapse), (ii) the sign problem, which is often needed for establishing series convergence, makes the evaluation of higher order diagrams problematic, (iii) technical aspects are more difficult than in the case of path integral Monte Carlo (e.g., storage of a four-point vertex function can be a real issue).
Let’s consider the statistical operator expressed in the real space - imaginary time representation, $$\exp(- \beta H) = \exp(-\beta H_0) {\mathcal T} \exp \left( -\int_0^{\beta} d\tau H_1(\tau) \right),
\label{eq:diagram_stat_op}$$ with $\beta$ the inverse temperature, the Heisenberg operator $H_1(\tau) = \exp(\tau H_0) H_1 \exp(- \tau H_0)$, and $\mathcal{T}$ the time ordering operator. Expanding Eq. \[eq:diagram\_stat\_op\] in powers of $H_1$, the partition function takes the form (for the Hubbard model) $$\begin{aligned}
Z & = & \sum_{n=0}^{\infty} (-U)^n \sum_{x_1 \ldots x_n} \int_{0<\tau_1 < \tau_2 < \ldots < \beta} \prod_{j=1}^{n} d\tau_j {\rm Tr} \left[ \ldots \right] \nonumber \\
\left[ \ldots \right] & = & \left[ \e^{-\beta H_0} c_{\uparrow}^{\dagger} (x_j \tau_j) c_{\uparrow} (x_j \tau_j) c_{\downarrow}^{\dagger} (x_j \tau_j) c_{\downarrow}(x_j \tau_j) \right].
\label{eq:diagram_series}\end{aligned}$$ This expansion generates the diagrams that consist of the four-point vertices $U$ connected by single-particle propagators $G_{\sigma}^{(0)} (x_i - x_j, \tau_i - \tau_j) = - {\rm Tr} [ \mathcal{T} e^{-\beta H_0} c^{\dagger}_{\sigma} (x_i \tau_i) c_{\sigma} (x_j \tau_j) ]$. The $p$-th order diagram is graphically given by a set of $(p!)^2$ possible connections of vertices by propagators shown in the top row of Fig. \[fig:determinant\]. Historically, this expansion was only used in connection with determinantal methods where the sign problem was either absent (due to an additional particle-hole symmetry) or manageable. This is discussed in Sec. \[sec:determinant\] and illustrated for the resonant Fermi gas in Sec. \[sec:resonant\_fermions\]. In Sec. \[sec:diagmc\] we will see how, in cases the sign problem is too bad, a sampling over all diagrams can be tried, with applications for the Fr[ö]{}hlich polaron, the Fermi polaron and the Hubbard model. Also dynamical mean-field theory (DMFT) can be understood in this way (Sec. \[sec:dmft\]) and combined with diagrammatic Monte Carlo algorithms.
Determinantal methods {#sec:determinant}
---------------------
![Diagrammatic series for the partition function. The upper line is the graphical representation of the series Eq. \[eq:diagram\_series\], lower line depicts Eq. \[eq:diagram\_determinant\]. The diagram signs are shown explicitly. The figure is taken from Ref. [@Burovski2006b]. []{data-label="fig:determinant"}](fig_determinant){width="0.8\columnwidth"}
Determinantal methods are used to alleviate the sign problem of Eq. \[eq:diagram\_series\]. In case of attractive fermions with equal spin population, or for repulsive fermions at half filling on a bipartite lattice (which has an additional particle-hole symmetry), the sign problem is absent. These methods are formulated on a finite lattice, so just like in path integral Monte Carlo there is no real symmetry breaking. Also in DMFT (see Sec. \[sec:dmft\]) the use of determinantal methods is widespread to solve the impurity problem. Let’s consider all diagrams of a given order $p$ with a fixed vertex configuration [@Burovski2006; @Burovski2006b], $$\mathcal{S}_p = \left\{ (x_j, \tau_j), j = 1, \ldots p \right\},$$ and sum over all $(p!)^2$ ways of connecting vertices with propagators. Then Eq. \[eq:diagram\_series\] takes the form $$Z = \sum_{p=0}^{\infty} (-U)^p \sum_{x_1 \ldots x_n} \int_{0<\tau_1 < \tau_2 < \ldots < \beta} \prod_{j=1}^{n} d\tau_j \det A^{\uparrow}(\mathcal{S}_p) \det A^{\downarrow} (\mathcal{S}_p),
\label{eq:diagram_determinant}$$ with $A^{\sigma}(\mathcal{S}_p)$ the $p \times p$ matrices whose elements are the single-particle propagators, $A_{ij}^{\sigma} (\mathcal{S}_p) = G_{\sigma}^{(0)}(x_i - x_j, \tau_i - \tau_j), i,j = 1, \ldots p $. For an equal number of spin-up and spin-down particles, $\det A^{\uparrow} \det A^{\downarrow} = \vert \det A \vert^2$ is positive. The Feynman diagrams are represented by a collection of vertices shown in the bottom row in Fig. \[fig:determinant\].
The simplest Monte Carlo scheme consists just of inserting and removing vertices and reevaluating the determinant (using so-called fast updates) [@Rubtsov03; @Rubtsov04; @Rubtsov05; @Burovski2004], but also worm-type updates have been devised for the dilute gas regime in order to simultaneously measure the correlation function [@Burovski2006b]. (note: Worm-type updates are here not related to winding numbers as in path integral Monte Carlo, but the term is used in the weaker sense of going to an extended configuration space where physical constraints are broken. By going back to the physical configuration space, all physical constraints are restored, and a configuration that is strongly decorrelated from the previous one is reached.) We refer to the cited papers for a detailed description of the algorithm.
Since the single-particle propagators depend on the Manhattan distance between lattice sites, a finite lattice is used in these methods. A continuous-space version can also be formulated [@Burovski2008]. Also note that such methods are useless for bosons because the evaluation of a permanent (instead of the determinant for fermions) is a $\#P$-problem (note that this is circumvented by [*sampling*]{} over all bosonic permutations in path integral Monte Carlo).
Application: The balanced spin$-1/2$ Fermi gas at unitarity {#sec:resonant_fermions}
-----------------------------------------------------------
![The critical temperature $T_c$ in units of the Fermi energy $\epsilon_F$ versus filling factor $\nu$. The continuum limit corresponds to $\nu \to 0$. The linear extrapolation (solid line) of the seven data points at lowest filling factors (filled circles) yield a critical temperature $T_c/ \epsilon_F = 0.173(6)$. The dashed line corresponds to a quadratic fit through all data points. Reprinted figure with permission from Ref. [@Goulko2010]. Copyright (2010) by the American Physical Society.[]{data-label="fig:detqmc_tc"}](fig_detqmc_tc){width="0.8\columnwidth"}
For Fermi gases in the vicinity of a Feshbach resonance, the scattering length diverges. Universal results are found in the zero-range resonant limit: when the effective range $r_0$ of the interaction goes to zero, $r_0 \to 0$, the $s$-wave scattering length remains finite, [*ie,*]{} $k_Fa$ remains fixed for $k_F r_0 \to 0$. The nature of the interaction potential is then irrelevant and the only remaining energy scale is the Fermi energy. The chemical potential is related to the Fermi energy by a universal number, $\mu = \xi \epsilon_F$, where $\xi$ is known as the Bertsch parameter. Up to this parameter, the equation of state is the same as that of an ideal gas. A value smaller than one indicates attractive interactions. The unitary regime was studied in great detail in BCS-BEC crossover studies (for a review, see Ref. [@Giorgini2008; @Varenna]). We are focusing on the determination of the critical temperature at unitarity and the equation of state.
Using determinantal Monte Carlo simulations, Goulko and Wingate found the critical temperature at $T_c / \epsilon_F = 0.173(6)$ [@Goulko2010] using a linear extrapolation shown in Fig. \[fig:detqmc\_tc\]. They put their final value at $T_c / \epsilon_F = 0.171(5)$, a value a bit higher than the previous result from [@Burovski2006; @Burovski2006b], where $T_c/ \epsilon_F = 0.152(7)$ was found with an identical method, illustrating the difficulties with the extrapolations. However, the continuous space version of the method [@Burovski2008] found the same result as in [@Burovski2006]. Bulgac [*et al.*]{} used an auxiliary field Monte Carlo approach (see Sec. \[sec:auxiliary\]) which extracted the critical temperature from the finite-size scaling of the condensate fraction, using the same procedure as in [@Burovski2006; @Burovski2006b], found an upper bound $T_c / \epsilon_F \le 0.15(1)$ [@Bulgac2008]. Previously, the same group [@Bulgac2006; @Bulgac2007] claimed $T_c / \epsilon_F = 0.23(2)$. There are also results obtained with the restricted path integral Monte Carlo method [@Akinenni2007], $T_c / \epsilon_F \approx 0.245$. Results for the imbalanced case (which has a sign problem) are also available [@Goulko2010]. We thus see that different Monte Carlo methods agree with each other, but others produced results for $T_c$ that vary over more than $70\%$, well outside error bars. This shows that interpretation of data, extrapolations, and providing correct error bars remain a hard task, for which the utmost care is needed.
Using diagrammatic Monte Carlo (see Sec. \[sec:diagmc\]), Van Houcke [*et al.*]{} [@VanHoucke2011] determined the equation of state of the balanced Fermi gas at unitarity for temperatures up to 5 times $T_c$. The results were in excellent agreement with the MIT experiments [@VanHoucke2011; @Ku2011] and in fair agreement with the ENS experiments [@Nascimbene2010]. With this method, the difficult double extrapolation in system size and density of the lattice determinant methods [@Burovski2006; @Goulko2010] can be avoided, but an extrapolation in expansion order is needed (and these simulations are not sign-free). The MIT experiments put $T_c/T_F = 0.167(13) $ [@Ku2011]. Discrepancies of a similar magnitude existed for the chemical potential at unitarity. The MIT experiments find $\xi = 0.376(5)$ [@Ku2011] (in disagreement with the ENS experiments which found $\xi = 0.415(10)$). The MIT value is consistent with the upper bound $\xi < 0.383$ [@Forbes2011] and is close to $\xi = 0.36(1)$ from a selfconsistent T-matrix calculation [@Haussmann2007] (cf. the good agreement between the selfconsistent T-matrix calculation and the exact result for the Fermi polaron problem at unitarity discussed in Sec. \[sec:fermi\_polaron\]). It lies below the earlier estimates $\xi = 0.44(2)$ [@Carlson2003] and $\xi = 0.42(1)$ [@Astrakharchik2004] found in fixed-node quantum Monte-Carlo simulations which provide upper bounds.
Sampling of all Feynman diagrams {#sec:diagmc}
--------------------------------
Prokof’ev and Svistunov have introduced, in a number of different contexts, a diagrammatic Monte Carlo scheme with a scalar representation in which all diagrams are sampled instead of a determinant evaluated. The idea is straightforward though audacious: an algorithm is devised such that all Feynman diagrams (involving all topologies, all expansion orders and all allowed momenta and frequencies) are sampled in a Monte Carlo scheme. An advantage of Feynman diagrams over strong coupling expansions is the absence of symmetry factors and lattice embedding coefficients. Other advantages of the method include the flexibility of the scheme (it is in principle applicable to any Hamiltonian), the connection with analytical tools, the possibility to directly formulate the method in the thermodynamic limit, and the possibility to work directly with real time or real frequencies, sidestepping the ill-conditioned analytical continuation problem inherent to methods formulated in imaginary time. The drawbacks are the lack of a convergence guarantee, the sign problem (meaning that only low expansion orders are accessible), and the fact that broken phases require a separate treatment. Prokof’ev and Svistunov argue that these drawbacks are however acceptable: The method does not try to alleviate the sign problem; in fact, the sign ’blessing’ is often crucial for the series convergence. When the system is deep inside a well defined phase a few diagram expansion orders suffice to accurately describe the physics, and it is expected that the sign problem is still tolerable for these orders. Note that the sign problem does not scale with the system volume for these methods.
An often employed strategy is to reduce the space of the diagrams. While the method of Sec. \[sec:determinant\] sums up all possible diagrams for the full Green function, including disconnected ones, the present method focuses typically on the selfenergy $\Sigma(k,\omega)$ in combination with the Dyson equation, $$G( k, \omega)^{-1} = G_0(k, \omega)^{-1} - \Sigma(k,\omega),$$ In particular, disconnected diagrams should not be generated for the selfenergy. Instead of considering an expansion with bare propagators $G_0$, one can also consider a skeleton expansion in which the bare propagators are replaced by fully dressed propagators $G$, which should then be determined selfconsistently [@Prokofev07_bold]. This further reduces the space of diagrams which should be sampled (but does not prevent the exponential growth of the number of diagrams with expansion order) : if a diagram is one-particle reducible, [*i.e.*]{}, it contains a subdiagram obtained by cutting two propagator lines and is disconnected otherwise from the rest of the diagram, then this diagram has to be discarded in a skeleton expansion. This scheme is known as bold diagrammatic Monte Carlo [@Prokofev07_bold], and requires a selfconsistency loop since the unknown propagator $G$ has to be determined selfconsistently. In practice, the method is started with bare propagators after which statistics for the selfenergy are collected. Then, a Dyson equation is performed (involving (fast) Fourier transforms if the coordinate and the (imaginary) time representation are used) and a new propagator is obtained. This procedure is iterated until selfconsistency is reached. For the final run with a fixed propagator $G$ the usual Markov chain and Monte Carlo convergence properties hold. For the selfconsistency loop however there is always the possibility that a metastable solution is found (e.g., in the vicinity of a first order transition). This is inherent to any selfconsistency problem and not a property of a diagrammatic Monte Carlo process.
As previously mentioned, there is no mathematical guarantee that the series expansion in Eq. \[eq:diagram\_expansion\] is convergent. The series may well be asymptotic or even divergent. Weak divergences can be overcome by applying resummation techniques. One constructs the partial sums up to the maximum expansion order $N_*$, $$\Sigma(N_*) = \sum_{n=1}^{N_*} D_N F_N^{(N_*)},$$ where the requirement on the factors $F_N^{(N_*)}$ is such that they approach unity for $N \ll N_*$ for large $N_*$ and suppress higher order contributions rendering the series $\sum_{N=1}^{\infty} D_N F_N^{(N_*)}$ convergent. The crossover region from where the function is close to unity to where it is approximately zero also has to increase with $N_*$. There are infinitely many way of satisfying these conditions and we list only a few examples: Cesáro-Riesz, Borel and Lindel[ö]{}f resummation. Depending on the nature of the divergence, not all may work, but final results have to be independent of the choice of $F$ provided $F$ is strong enough to compensate the divergence [@VanHoucke08; @Prokofev07_bold].
From the generality of the above discussion, the reader will appreciate the potential of these methods but also understand that they are still under development and that their true potential is only gradually being discovered. Ther are still many unexplored ideas which may turn this method into a versatile tool and which could lead to interesting research. We now give a few examples of where these ideas have already been implemented successfully. For historical reasons, we start with Fr[ö]{}hlich polarons, proceed with the Fermi-polaron problem, and finally show results for the Hubbard model.
### Fr[ö]{}hlich polarons
![A typical diagram contributing to the polaron Green function of the Fr[ö]{}hlich Hamiltonian. Dashed lines denote phonon propagators, full backbone lines bare electron propagators. The length of the diagram is the total imaginary time of the Green function. In this linear representation, the first and last electron Green function are necessarily uncovered by phonon propagators. Reprinted figure with permission from Ref. [@Prokofev98_froehlich]. Copyright (1998) by the American Physical Society.[]{data-label="fig:froehlich_diagram"}](fig_froehlich_diagram){width="0.8\columnwidth"}
The Fr[ö]{}hlich Hamiltion describes optical phonons coupled to electrons via $$H_{\rm e-ph} = \sum_{k,{\bm q}} V(\bm{q}) (b_{\bm q}^{\dagger} - b_{-{\bm q}}) a_{\bm {k-q}}^{\dagger} a_{\bm k},$$ with $a_{\bm k}$ and $b_{\bm q}$ the electron and phonon annihilation operators with momenta $\bm{k}$ and $\bm{q}$, $V(\bm{q}) = i(2 \sqrt{2} \alpha \pi)^{1/2} / q$ the coupling strength and $\alpha$ a dimensionless coupling constant. The phonon propagator is independent of momentum, $D(\bm{q}, \tau) = \exp( - \omega_p \tau)$, with $\omega_p$ the frequency of the optical phonon. Electron propagators are given by $G_0(\bm{p}, \tau) = e^{- (p^2 / (2m) - \mu) \tau}$, with $\mu$ the chemical potential (here a tuning parameter).
The expansion for the polaron Green function in terms of bare electron Green functions and phonon propagators turns out to be positive definite. A typical diagram is shown in Fig. \[fig:froehlich\_diagram\]. An ergodic set of updates consists of just inserting and removing phonon propagators at arbitrary points in imaginary time, but other updates such as shifting the position in imaginary time of a vertex or changing the topology of the diagram by reconnecting the phonon propagators with different vertices can improve the sampling efficiency [@Prokofev98_froehlich; @Mishchenko00]. As always, one can think of many good ways of performing the sampling.
On a present-day laptop one can sample expansion orders up to $\sim 100$ in a couple of minutes with an accuracy of $\sim 1$%. In Ref. [@Mishchenko00] a cyclical representation was used, which allows for improved estimators.
The polaron Green function is the central quantity in this problem. It follows from the Lehmann representation that, if $$G(\bm{k}, \tau \gg \omega_p^{-1}) \to Z_k \exp [ - (E(\bm{k}) - \mu) \tau ],
\label{eq:polaron_green}$$ then for $k=0$ the energy $E_0$ is the ground state energy of the system and the factor $Z$ shows the fraction of the bare-electron state in the true eigenstate of the polaron. It also follows that one must choose $\mu < E_0$, but the closer it is tuned to $E_0$ the more accurate the exponential decay of the Green function can be resolved.
Since the physics of electron-phonon interactions is beyond the scope of this review, we refer to Refs. [@Prokofev98_froehlich; @Mishchenko00] for a general physics discussion and quantitative results for the binding energy, effective mass, structure of the polaronic cloud, and the spectral analysis (after analytic continuation) for any coupling strength (both small and large polarons). For completeness we mention that polarons in the Su-Schrieffer-Heeger model have also been studied in detail [@Marchand10].
### The Fermi-polaron problem in three dimensions {#sec:fermi_polaron}
![Polaron (black circles) and molecule (red triangles) effective mass as function of $k_F a$. The vertical dotted line stands for $(k_F a)_c = 1.11$, where the polaron energy becomes lower than the molecular energy. The dashed line is the contribution from the first order diagram [@Combescot07]. Reprinted figure with permission from Ref. [@Prokofev08_fermipolaron]. Copyright (2008) by the American Physical Society.[]{data-label="fig:diagmc_fermipolaron"}](fig_diagmc_fermipolaron){width="0.8\columnwidth"}
With ’Fermi polaron’, a spin-down fermion resonantly interacting with a sea of non-interacting spin-up fermions is understood. It is an idealization of a Fermi mixture with strong imbalance. The Fermi polaron problem was crucial in understanding the difference between the Rice and MIT (and ENS) experiments for imbalanced Fermi gases. We refer to Ref. [@Chevy2010] for a recent review on this interpretation, as well as for descriptions of the equation of state of imbalanced Fermi gases both on the BCS and the BEC side. We will restrict ourselves to the physics of the Fermi polaron problem at zero temperature.
For the Fermi polaron problem near unitarity, the nature of the interaction potential is irrelevant. Diagramatically, this implies that the sum $\Gamma(\tau, p)$ of all ladder diagrams for the interaction potential has to be considered as a separate diagrammatic element, which also takes care of the UV divergences. A closed form for $\Gamma$ in universal form exists [@Varenna], $$\begin{aligned}
\Gamma^{-1}(\eta, {\bm p}) & = & \frac{m}{4\pi a} - \frac{m}{8 \pi} \sqrt{p^2 - 4 m \eta} - \bar{\Pi}(\eta, \bm{p}), \\
\bar{\Pi}(\eta, \bm{p}) & = & \int_{q \le k_F} \frac{d^3 \mathbf{q} }{ (2\pi)^3} \frac{1}{q^2/ (2m) + ({\mathbf{p}} -{ \mathbf{q}})^2/(2m) - \eta},\end{aligned}$$ with $\eta = \omega + \epsilon_F + \mu + i0^{+}$. In Refs. [@Prokofev07_fermipolaron; @Prokofev08_fermipolaron] the T-matrix was calculated by applying bold diagrammatic Monte Carlo instead of converting the above formulae to the imaginary time domain.
Deep in the BEC limit (repulsive side) the impurity atom will form a polaron quasi-particle dressed by the majority fermions. Deep in the BCS limit (attractive) side, the impurity atom will form a bound state (a bosonic molecule) with a majority atom. Hence, one expects a transition between the polaronic and molecular regimes in the unitary regime.
On the BCS side, Chevy wrote down a variational Ansatz for the polaron wavefunction in the subspace of single particle-hole excitations created by the minority atom [@Chevy2006], $$\vert \psi \rangle = \left( \phi_0 d_0^{\dagger} + \sum_{\bm {k,q}} \phi_{\bm {k,q}} d^{\dagger}_{\bm{k} - \bm{q}} u_{\bm k}^{\dagger} u_{\bm{q}} \right) \vert \rm{FS} \rangle,$$ where the $\phi$ are variational parameters, the sum over $k$ ($q$) is restricted to be above(below) the Fermi surface (FS), $d$ annihilates the spin down atom and $u$ a spin up atom. The quasiparticle dispersion can then be written for small momenta ${\bm k}$ as $$E_k^{\rm pol} = A E_F + \frac{k^2}{2m^*},
\label{eq:fermi_polaron_variational}$$ in analogy to the one for a free atom, but with renormalized (variational) parameters $A_{\rm var} \approx -0.6$ and $m^*_{\rm var} = 1.17m$ for $a \to \infty$ [@Chevy2006; @Combescot07]. If a quasiparticle is well defined, the polaron energy and dispersion can be found in diagrammatic Monte Carlo simulations based on Eq. \[eq:polaron\_green\], while the molecular energy follows from a similar equation for the 2-particle propagator. In Refs. [@Prokofev07_fermipolaron; @Prokofev08_fermipolaron] the polaronic and molecular channel were treated on equal footing. The diagrams consist of a backbone impurity propagator dressed by T-matrices. The T-matrices have to be connected by majority fermion propagators in all possible ways, excluding (sub)diagrams that are already part of the T-matrix. A list of updates and the corresponding equations for detailed balance are described in detail in Ref. [@Prokofev08_fermipolaron] and will not be repeated here. Expansion schemes based on bare and bold propagators were also discussed.
The transition between the polaron and molecule regime does not happen at unitarity, but at a slightly larger value $k_F a = 1.11(2)$ [@Prokofev07_fermipolaron; @Prokofev08_fermipolaron] shown in Fig. \[fig:diagmc\_fermipolaron\], where also the polaronic and molecular effective masses are shown. Diagrammatic Monte Carlo finds $A = -0.61(1)$ and $m^* = 1.20(1)m$ at unitarity. Variational fixed node Monte Carlo simulations find $A = -0.59(1)$ and $m^* = 1.09m$ [@Lobo06]. It is remarkable that the first order diagram on the polaronic side with selfconsistent polaron propagators [@Chevy2006; @Combescot07] (equivalent to the variational Ansatz) is surprisingly close to the full answer due to a remarkable cancellation of higher order diagrams. Note that the equation of state of a resonant Fermi gas on the BCS side is accurately described by a gas of polarons [@Chevy2010]. On the other side of the polaron-to-molecule transition, the composed boson is interacting with the Fermi sea with a surprisingly accurate mean-field energy $g_{\rm ad} n_{\uparrow}$, where the atom-dimer coupling $g_{\rm ad}$ is related to the atom-dimer scattering length, $a_{\rm ad} = 1.18a$ [@STM]. Eq. \[eq:fermi\_polaron\_variational\] was extended to the molecular sector, where the dimer is dressed by single particle-hole pairs of the majority Fermi sea [@Mora09; @Punk09; @Combescot09].
Experimentally, the parameter $A$ can be determined from rf-spectroscopy. This requires that the imbalance between majority and minority atoms is large enough so that no superfluid core is formed. The MIT experiments found values for $A$ in agreement with the Monte Carlo predictions [@Schirotzek09]. The transition between the molecular and the polaronic transitions could not be determined because this transition is preempted by phase separation between the ideal Fermi gas and the polarized molecular superfluid. The effective mass can be determined from collective mode excitations. This was done in Ref. [@Nascimbene09] where $m^* = 1.17(10)m$ was found, in close agreement with the Monte Carlo values.
### Hubbard model
![Momentum dependence of the selfenergy at the lowest fequency for the 2D Hubbard model with $t=1, U=4, \mu = 3.1,$ and $T=0.4$. Comparison is made with single-site DMFT and a cluster DMFT method, DCA. The mean-field contribution (the Hartree term $Un_{\sigma} = 2.3t$ ) was subtracted. Arrows indicate the position of the Fermi momentum $p_F$. The figure is taken from Ref. [@Kozik2010]. []{data-label="fig:diagmc_hubbard"}](fig_diagmc_hubbard){width="0.8\columnwidth"}
For the Hubbard model, only a bare expansion for the selfenergy was explored thus far [@Kozik2010]. This was done for $U/t \le 4$ in the Fermi liquid regime, for temperatures down to $T/t = 1/40$. The momentum-dependent selfenergy is shown for the lowest frequency in Fig. \[fig:diagmc\_hubbard\], where comparison is made with the dynamical cluster approximation (DCA) method, see Sec. \[sec:dmft\]. A momentum cluster of size 32 is needed to obtain good agreement. Diagrams up to order 8 were generated, which is a limit set by the sign problem. Going to larger values of $U$ remains topic of investigation and will require a bold diagrammatic Monte Carlo approach, the use of a T-matrix (cf. Sec. \[sec:fermi\_polaron\]), the combination with DMFT [@Pollet_Anderson], or a combination thereof.
Continuous-time Auxiliary-Field methods {#sec:auxiliary}
---------------------------------------
Auxiliary-field methods use a Hubbard-Stratonovich decomposition of the two-body propagator into a set of propagators of one-body potentials containing an auxiliary variable which should be integrated over. Discrete-time versions of this idea, such as the Blankenbecler-Sugar-Scalapino algorithm [@Blankenbecler81] and the Hirsch-Fye algorithm [@HirschFye86] algorithm have been developed long ago (cf. [@Assaad2001]). The extrapotlation in the discrete time step is then mandatory, but can often be done reliably. Another advantage is the linear scaling of the discrete methods with $\beta$ (compared to $\beta^3$ for the continuous models). A fair comparison would also need to include the increase of the autocorrelation times, but no such systematic studies exist. The first continuous-time method for fermionic lattice models was an auxiliary-field decomposition scheme presented by Rombouts [*et al.*]{} [@Rombouts99] and applied to nuclear matter and small Hubbard lattices. It was reformulated 10 years later by Gull [*et al.*]{} [@Gull08; @Gull_review], whose formulation we now follow.
A non-zero constant is added to the two-body part, $$H_U = U \sum_i n_{\uparrow} n_{\downarrow} - K / \beta.$$ By expanding the exponential in powers of $H_U$ and applying the auxiliary field decomposition [@Rombouts99] $$1 - \frac{\beta U}{K} \sum_i \left( n_{i\uparrow} n_{i\downarrow} - \frac{n_{i\uparrow} + n_{i\downarrow}}{2} \right) = \frac{1}{2V} \sum_{i, s_i} e^{\gamma s_i ( n_{i\uparrow} - n_{i \downarrow} )},$$ where the last term on the left hand side denotes the usual shift in chemical potential, and with $$\cosh(\gamma) = 1 + \frac{U \beta V}{2K},$$ one arrives at a partition function written solely in one-body propagators $$\begin{aligned}
Z & = & \sum_{n=0}^{\infty} \prod_{j=1}^{n} \int_{\tau_{j-1}}^{\beta} d\tau_j \sum_{s_j = \pm 1} (\frac{K}{2 \beta V} )^n Z_n ( \{ s_j, \tau_j , x_j \} ), \nonumber \\
Z_n( \{ s_j, \tau_j, x_j \} ) & = & {\rm Tr} \prod_{i=n}^{1} \exp( - \Delta_i H_0) \exp(s_i \gamma (n_{x_j, \uparrow} - n_{x_j, \downarrow} ) ),\end{aligned}$$ with $\Delta_i = \tau_{i+1} - \tau_i$ for $i<n$ and $\Delta_n = \tau_1 + \beta - \tau_n$. The time arguments are continuous variables and not regularly spaced on $[0 , \beta [$, unlike the discrete-time methods [@Blankenbecler81; @HirschFye86]. Generalizing the derivation provided in Ref. [@Georges1992] the weights are expressed as $$\begin{aligned}
\frac{Z_n ( \{ s_j, \tau_j , x_j \} ) }{Z_0} = \prod_{\sigma} \det N_{\sigma}^{-1} ( \{ s_j, \tau_j , x_j \} ) \nonumber \\
\det N_{\sigma}^{-1} ( \{ s_j, \tau_j , x_j \} ) \equiv e^{V_{\sigma}^{ \{ s_j \}} } - G_{0 \sigma} ^{ \{ \tau_j, x_j \} } ( e^{V_{\sigma}^{ \{ s_j \}} } - 1) \nonumber \\
e^{V_{\sigma}^{ \{ s_j \}} } \equiv {\rm diag} \left( \e^{\gamma (-1)^{\sigma} s_1} , \ldots , e^{\gamma (-1)^{\sigma} s-k} \right), \end{aligned}$$ with the notations $(-1)^{\uparrow} = 1, (-1)^{\downarrow} = -1,$ and equal time evaluations are taken as $\tau = 0^{+}$. Rombouts used a fixed length representation, which leads to additional combinatorial factors in the weights. The simplest possible set of updates that fulfill ergodicty are inserting and removing new auxiliary variables $\sigma$, which can be balanced against each other. If the new time and site are chosen arbitrarily, and if the auxiliary variable is also chosen uniformly among all variables for the reverse update, then the acceptance factor reads $$R = \frac{K}{n+1} \frac{\det N_{\uparrow}(y) \det N _{\downarrow}(y) }{\det N_{\uparrow}(x) \det N _{\downarrow}(x)},$$ where $x$ denotes the old configuration, $y$ the configuration after inserting the new auxiliary variable, and $n$ the expansion order in configuration $x$. Efficient numerical schemes for such updates are discussed elsewhere [@Rombouts99; @Gull08; @Gull_review].
An application for this method is given in Sec. \[sec:DCA\_Hubbard\] in combination with DMFT. We also draw attention to one other application, namely enhanced Pomeranchuk cooling schemes of a SU($2N$) ultra-cold fermionic in optical lattices at half filling [@Cai2012] because of the large number of hyperfine-spin components.
Dynamical Mean-Field Theory (DMFT) {#sec:dmft}
----------------------------------
Dynamical Mean-Field Theory (DMFT) provides an approximate solution to a many-body problem, unlike the other methods covered in this paper. There are many excellent introductory texts and reviews on DMFT available, which goes beyond the scope of this review [@Georges1992; @Maier2005; @Kotliar2006; @Gull_review]. DMFT has however a close connection to diagrammatic methods, which we want to make clear. Like in any mean-field theory, it considers a single site coupled selfconsistently to the rest of the lattice, but it additionally retains dynamical information (retardation effects). In the diagrammatic language, DMFT sums up all skeleton diagrams contributing to the selfenergy built with purely local propagators. Technically, this is done by mapping the many-body problem onto an impurity problem, for which efficient numerical procedures are known to solve it [@Gull_review]. There exist apart from approximate methods (such as the iterative perturbation theory (IPT) and the non-crossing approximation (NCA)), also controllable and/or exact methods such as exact diagonalization, the numerical renormalization group, and Monte Carlo solvers (such as the weak-coupling expansion method (see Sec. \[sec:determinant\]), the auxiliary field method (see Sec. \[sec:auxiliary\]), and a strong coupling expansion method [@Werner2006]). They are all discussed in detail in Ref. [@Gull_review]. Crucially, for a single site impurity problem, the sign problem is absent.
### Formalism
The basic set of equations, specified here for the Hubbard model, can be formulated as follows. Let’s introduce the full local Green function obtained by integrating $G({\bf k},i \omega)$ over the Brillouin zone, $$G_{\rm loc}(i \omega) = \int \frac{d {\bf k}}{ (2\pi)^d} G( {\bf k}, i \omega).
\label{eq:dmft_Gloc}$$ We also introduce the functional integral representation of the partition function, $$Z = \int \mathcal{D} \psi_{\sigma}^{\dagger} \mathcal{D}\psi_{\sigma} e^{-S^{\rm imp}},$$ where the impurity action in imaginary time representation reads, $$\begin{aligned}
S^{\rm imp} & = & \int_0^{\beta} d\tau \psi_{\sigma}^{\dagger} (\tau) (\partial_{\tau} - \mu ) \psi_{\sigma}(\tau) + U n_{\uparrow}(\tau) n_{\downarrow}(\tau) \nonumber \\
{} & {} & + \int_0^{\beta} d\tau \int_0^{\beta} d\tau' \psi_{\sigma}^{\dagger}(\tau) \Delta(\tau - \tau') \psi_{\sigma}(\tau'),
\label{eq:dmft_impurity_action}\end{aligned}$$ with $\Delta(\tau - \tau')$ an unknown ’hybridization’ function. DMFT consists of solving the impurity problem $\Sigma^{\rm imp}[G_{\rm loc} (i \omega)]$ as a functional of the full local Green function such that the selfconsistency equation $$G(\bm{k}, i\omega)^{-1} = G_0({\bm k}, i\omega)^{-1} - \Sigma^{\rm imp}(i\omega),
\label{eq:dmft_G}$$ is fulfilled. Here, $G_0(\bm{k}, i \omega) = (i \omega - \mu + \epsilon_k )^{-1}$ is the non-interacting Green function of the many-body problem, with $\epsilon_k$ the dispersion. There are many ways to solve a selfconsistency equation, but in practice an iteration scheme is used: From an initial guess for the impurity selfenergy, a first guess for the (full) Green function is obtained via Eq. \[eq:dmft\_G\] and the local Green function $G_{\rm loc}$ is computed with Eq. \[eq:dmft\_Gloc\]. The sum over the Brillouin zone in Eq. \[eq:dmft\_Gloc\] is usually replaced by a one-dimensional integral over the density of states. The hybridization function is determined from the Dyson equation for the impurity problem, $\Delta(i \omega)^{-1} = G_{\rm loc}^{-1}(iw) + \Sigma^{\rm imp}(i \omega)$. The impurity problem for a given hybridization function $\Delta$ must then be solved, and a new $\Sigma_{\rm imp}$ is obtained, after which the scheme is repeated until convergence is reached. In this iteration process, the hybridization function (which is just an auxiliary function irrelevant for the underlying Hubbard model) is also determined. DMFT is a widely used method and still being developed further. We mention two main directions: cluster dynamical mean field theory both for real-space clusters and momentum clusters (we refer to Ref. [@Maier2005] for an extensive review) and the dual fermion approach, in which the irreducible vertex can be treated by using a dual set of variables [@Rubtsov2008]. It has been used in electronic structure calculations in combination with density functional theory methods [@Kotliar2006]. Close connections between cellular ([*i.e.*]{}, a real-space cluster) DMFT and cluster perturbation theory have also been revealed in the framework of self-energy functional methods [@Potthoff2003a; @Potthoff2003b].
### Application: cluster DMFT for the 3D Hubbard model {#sec:DCA_Hubbard}
![Phase diagram of the 3D Hubbard model at half filling. Shown are the predictions form second order perturbation theory (SOPT), Weiss mean-field theory, the Heisenberg limit prediction, dynamical mean-field theory (DMFA), dynamical cluster approximation (DCA) [@Kent2005], and the determinantal lattice quantum Monte Carlo simulations by Staudt [*et al.*]{} [@Staudt2000]. Reprinted figure with permission from Ref. [@Kent2005]. Copyright (2005) by the American Physical Society.[]{data-label="fig:hubbard_halffilling"}](fig_hubbard_halffilling){width="0.8\columnwidth"}
![Entropy per volume as a function of filling for a 3D Hubbard model at $T = 0.35t \approx T_N$ for $U/t = 8$. The inset shows the entropy per particle. Shown are the single site DMFT results and the DCA results extrapolated in the cluster size. The error bars are dominated by the extrapolation. Reprinted figure with permission from Ref. [@Fuchs2011]. Copyright (2011) by the American Physical Society.[]{data-label="fig:dca_hubbard"}](fig_dca_hubbard){width="0.8\columnwidth"}
Since the first experimental signatures of the Mott insulator in the 3D Fermi Hubbard model have been observed [@Joerdens2008; @Schneider2008], the questions of when and how to reach the antiferromagnetic transition became important, as well as how to observe such a phase [@DeLeo2008; @McKay2011]. The phase diagram at half filling, shown in Fig. \[fig:hubbard\_halffilling\], was known from the sign-free auxiliary field Monte Carlo calculation of Ref. [@Staudt2000], which was confirmed by a cluster DMFT (DCA) calculation by Kent [*et al.*]{} [@Kent2005]. The latter showed that so-called periodic Betts clusters [@Betts1997; @Betts1999] of size 48 were as accurate as the simulations on real-space clusters of size up to $L=10^3$ [@Staudt2000], and are optimal for finite size scaling.
The lowest entropies per particle in present experiments are about $S/N \approx \log(2)$, but typically they are a bit higher, about $S/N \approx 1.2$ [@Joerdens2010]. At these temperatures, the local physics becomes a very good approximation: DMFT and high temperature series expansions [@Oitmaa] give essentially identical answers [@DeLeo2011; @Fuchs2011; @Scarola2009].
The entropy per spin at the N[é]{}el temperature in the Heisenberg model (that is, the infinite $U$ case at half filling in the Hubbard model) is $S/N = 0.35$ [@Werner2005; @Wessel2010]; about half of $\log(2)$. Can this value be substantially higher at lower values of $U$?
Fuchs [*et al.*]{} provided the full thermodynamics of the 3D Hubbard model for temperatures approaching the N[é]{}el temperature (without breaking the symmetry), for any filling, and interaction strengths up to $1.5$ times the bandwidth by using DCA with an auxiliary-field Monte Carlo impurity solver. From entropy curves such as the one shown in Fig. \[fig:dca\_hubbard\] they could construct in the local density approximation the total entropy of a trapped system. They found that the maximal critical entropy per particle $S/N = 0.65(6)$ is found for $U/t = 8$, about 1.5 times as high as without a parabolic trapping potential, $S/N = 0.41(3)$ for $U/t=8$. So, just as in the case of 3D bosons, the liquid in the edges acts as a big entropy reservoir, primarily because of larger volume fractions in the edges compared to the middle of the trap. The value of the critical entropy was only weakly dependent on the value of $U/t$. We note that such entropies are nowadays easily reached in bosonic lattice systems. Fuchs [*et al.*]{} also found that the double occupancy changes little with temperature, while the nearest-neighbour spin-spin correlation functions show a stronger signal around the N[é]{}el temperature. All these findings were confirmed in Ref. [@Paiva2011], who employed a determinantal lattice Monte Carlo algorithm, and extrapolated in lattice size and Trotter time discretization step. They also went to lower temperatures inside the broken phase. In Ref. [@Gorelik2010] a real-space extension of single site DMFT was used, focusing on the double occupancy, which for large values of $U/T$ shows a strong increase in the broken phase when temperature is lowered from $T_N$ to zero. In two dimensions, the temperature and entropy scales for observing antiferromagnetism have also been determined [@Paiva2010].
### Incorporating DMFT in diagrammatic Monte Carlo
![Scheme illustrating how DMFT can be used to construct locally fully bold propagators as initial propagators for a general diagrammatic Monte Carlo simulation. The Matsubara frequency index has been suppressed to ease the notation. From a certain knowledge of the full Green function $G({\bm{p}})$ the local Green function is constructed (left). The impurity problem is then solved (it is assumed that this can be done in an efficient way), providing the selfconsistent solution for the local selfenergy. On the right, diagrammatic Monte Carlo samples all non-local diagrams contributing to the selfenergy. The local and non-local contributions to the selfenergy are then merged. The Dyson equation then gives us more information on the full Green function. This scheme is repeated until convergence is reached. Reprinted figure with permission from Ref. [@Pollet_Anderson]. Copyright (2011) by the American Physical Society.[]{data-label="fig:anderson"}](fig_incorporation){width="0.8\columnwidth"}
By explaining DMFT in terms of Feynman diagrams, as done here and in Ref. [@Kotliar2006], it follows that it is possible to combine DMFT with bold diagrammatic Monte Carlo: Since DMFT sums up all local skeleton diagrams of the self energy, it can be used to construct initial ’bold’ propagators for diagrammatic Monte Carlo, as is illustrated in Fig. \[fig:anderson\]. This was demonstrated in Ref. [@Pollet_Anderson] for Anderson’s model of localization [@Anderson58]. It is a one-particle quantum effect where destructive interference between all possible paths of the particle scattering off impurities can localize the particle. Anderson localization depends strongly on the dimension. In one dimension the wave function is always localized. In three dimensions, there is a mobility edge separating extended states from localized ones. Because of the single-particle Hamiltonian, Anderson localization is a problem that can be diagonalized and does not have an exponentially growing Hilbert space. In practice, the lattice sizes that can be fully diagonalized are small, and Anderson localization remains a tricky problem since it has no small parameter.
In Ref. [@Pollet_Anderson] non-interacting fermions on a 3D lattice with on every site a chemical potential distributed according to a (quenched) delta-correlated Gaussian distribution (which facilitates the diagrammatic technique) was studied. Spatial coordinates on a lattice and real time at zero temperature were used. The topology of the diagrams for the Green function built from bare propagators is then the same as in Fig. \[fig:froehlich\_diagram\]. The DMFT part reduces in this case to an algebraic equation and provides a solution very close to the correct answer: The magnitude of the contributions to the selfenergy that are not built from purely local propagators, is very small. Up to 50 expansion orders could be sampled. For strong disorder, expansion times up to $3-4$ times the hopping could be reached; for weak disorder expansion times up to $10$ times the hopping could be reached. In order to find the mobility threshold, vertex corrections are needed. This has not been implemented yet for this model. Another obvious extension of this idea would be to combine DMFT with diagrammatic Monte Carlo for the Hubbard model in the parameter regime of large values of $U$, so that the local physics can be summed up from the outset in the DMFT loop (also within the DMFT framework $U$ can be made non-local in imaginary time). Whether this approach would allow to provide controlled error bars for large values of $U$ has not been tried yet.
The cold gas community has also shown interest in Anderson localization in order to demonstrate localization with matter waves. The first experiments were done around 2007 with bosons in one dimension without lattice at low enough density such that interactions are negligible [@Billy08]. The disorder is generated with optical speckles. The longitudinal trap is switched off, and the BEC starts expanding. Then the expansion rapidly stops, and the density in the wings is exponential, typical for Anderson localization [@Billy08; @SanchezPalencia07]. At the same time there were also experiments performed with one-dimensional quasi-periodic lattices (the Aubry-Andr[é]{} model), a system which features a crossover between extended and exponentially localized states [@Roati08]. Localization was demonstrated by investigating transport properties, and spatial and momentum distributions. Anderson localization has also been studied with bosons in 3D [@Jendrzejeweski11] and with fermionic, spin-polarized $^{40}$K atoms [@Kondov11]. In the latter experiment, the cloud has a mobile component that expands ballistically, but more rapidly than a thermal gas. It also has a localized component that becomes fixed after a rapid initial expansion. A mobility edge was defined as the energy below which particles are localized. Although it increases with the disorder strength, it does not follow the self consistent Born prediction or predictions from weak scattering theory. We note that many aspects of this experiment remain unexplained, but in the absence of a small parameter, any theoretical description is difficult. It remains to be seen if diagrammatic Monte Carlo can provide more insight into this problem.
### Comments on real-time dynamics
Non-equilibrium dynamics remains extremely difficult to describe accurately numerically. For one-dimensional systems, time-adaptive DMRG can be used [@White2004; @Daley2004], but this fails in higher dimensions. We refer to Ref. [@Eckstein2010] for a review of the non-equilibrium flow equation method and DMFT. It also contains a discussion of an interaction quench in the Hubbard model with DMFT using Monte Carlo methods. Instead of a rapid thermalization, an intermediate prethermalized state was found. Gull [*et al.*]{} have formulated a diagrammatic Monte Carlo method on the Keldysh contour for impurity models [@Gull2011]. They pre-summed the class of non-crossing diagrams, sampled corrections to it, and could describe long-time and steady-state properties over a wide range of interaction strengths.
### Bosonic DMFT {#sec:bdmft}
![Ground state phase diagram of the 3D Bose-Hubbard model. The ground state phase diagram in computed in the mean-field approximation (’Mean Field’) [@Fisher1989], exactly with worm-type Monte Carlo path integral simulations (’Monte Carlo’) [@CapogrossoSansone07], and using B-DMFT with the density of states of a cubic lattice (’B-DMFT’) and on a Bethe lattice with coordination number 6 (’Bethe $z=6$’). Reprinted figure with permission from Ref. [@Anders2010]. Copyright (2010) by the American Physical Society.[]{data-label="fig:bdmft"}](fig_bdmft){width="0.8\columnwidth"}
The bosonic dynamical mean-field theory (B-DMFT) can be developed along the same lines as in Sec. \[sec:dmft\]. The only point of attention is that bosons can condense, which requires a similar extension of DMFT as in Ref. [@Chitra2001] (so-called EDMFT). As in any diagrammatic method, studying the condensed phase $\langle b \rangle = \phi \neq 0$ requires that one allows to break the symmetry by writing a generalized action for the impurity problem. The action consists of all the local terms (chemical potential and potential energy terms), to which the mean-field contribution is added, $-\kappa \phi \int d \tau (b(\tau) + b^{\dagger}(\tau))$ for a real, homogeneous and time-independent condensate, where $\kappa = zt$ is the coordination number times the hopping amplitude. Recalling the theory of the weakly interacting Bose gas, the introduction of normal and anomalous propagators is needed in the presence of a condensate. Switching to the Nambu-Gor’kov notation, $\bm{b^{\dagger}} = ( b^{\dagger}, b)$ and $\bm{\Phi} = (\phi, \phi)$, the hybridization part in the action can be written as, $$S_{\rm hyb} = -\frac{1}{2} \int_0^{\beta} \int _0^{\beta} d\tau d\tau' (\bm{b^{\dagger}}(\tau) - \bm{\Phi}) \bm{\Delta}(\tau - \tau') (\bm{b}(\tau) - \bm{\Phi}),$$ where $\Delta_{11}(\tau) = \Delta_{22}(-\tau)$ and $\Delta_{12}(\tau) = \Delta_{21}(\tau)$ are real functions describing the normal and anomalous hybridization functions. The above terms linear in $\Phi$ can be combined with the mean-field decoupling changing $\kappa$ to $\kappa = zt - \Delta_{11}(i \omega = 0) - \Delta_{12}(i \omega_n = 0)$ and bringing the hybridization action into a form suitable for Monte Carlo simulations since it contains the full $\bm{b}$ (and not $\bm{b}$ - $\bm{\Phi}$). A Monte Carlo solver based on a strong coupling expansion was introduced in Ref. [@Anders2010] for the impurity problem, which allows to compute the condensate, normal and anomalous Green function. The B-DMFT equations are closed by setting $\phi = \langle b \rangle$ calculated on the impurity site, extracting the connected Green functions (the ones for the depleted bosons) and selfenergies, from which a new hybridization matrix can be computed by the same selfconsistency relations as for fermions (a $2 \times 2$ matrix needs to be inverted because of the Nambu-Gor’kov formalism). The value of $\kappa$ also needs to be updated in every iteration step. Because the expansion is done in the hybridization and condensate terms, a sign problem occurs in the condensed phase because of the opposite sign of anomalous and normal Green functions. Full details of the algorithm can be found in Ref. [@Anders2011].
The above formalism works for any dispersion and interaction, and is stable for any phase. The phase diagram of the 3D Bose-Hubbard model was calculated with a precision better than $2\%$, as is shown in Fig. \[fig:bdmft\]. It does not describe universal behaviour of the phase transition correctly (as expected), but local physics are extremely well captured. Other results include the finding of non-universal critical exponents, the successful study of the weakly interacting Bose gas, the failure of the Hugenholtz-Pines relation, and the derivation of the DMFT equations in three different ways, as well as the connection with cavity methods for Bethe structures (see Ref. [@Zamponi] and the appendix in Ref. [@Anders2011]). For completeness, we mention that other authors claimed the correct B-DMFT formalism earlier [@Byczuk2008; @Hubener09_twocomponent; @Hu2009], despite apparent differences in formalism with Refs. [@Anders2010; @Anders2011], and failing to publish results in the broken phase for arbitrary system parameters [@Byczuk2008]. The importance of the development of B-DMFT lies in possible extensions to real-time dynamics, Bose-Fermi mixtures, spinful bosonic systems with spin $F \ge 2$, and perhaps as a starting point for more general diagrammatic Monte Carlo simulations of bosonic systems. Just as in the fermionic case, there is a close connection with variational cluster approximations [@Knap2011; @Arrigoni2011], which have been extended to non-equilibrium dynamics formulated on the Keldysh contour already [@Knap2011b]. Two-component bosonic mixtures were also studied in the DMFT approximation (but with an exact diagonalization solver for the impurity problem) in Refs. [@Li2011a; @Li2011b] .
Diffusion Monte Carlo {#sec:diffusion}
=====================
In the last section of this review we switch to diffusion Monte Carlo. It is not a diagrammatic method, although it shares some similarities with path integral Monte Carlo: in diffusion Monte Carlo a number of walkers are propagated forward in imaginary time in order to project out the ground state. The method has been reviewed in detail in Refs. [@Bajdic2009; @Kolorenc2011] and in relation to cold atoms in Ref. [@Giorgini2008]. We will therefore be rather schematic for the method, and focus on a single application, namely the controversial issue of Stoner ferromagnetism with cold fermionic atoms on the upper branch of the Feshbach resonance. For the BCS-BEC crossover which was studied much more intensely with this method, we refer to Ref. [@Giorgini2008].
Methods
-------
The position of the $N$ walkers is given by $\bm{R} = (\bm{r}_1, \ldots, \bm{r}_N)$, defined at every time $\tau = j \delta \tau, j = 1,2, \ldots$ where $\delta \tau$ is the time step. The ground state satisfies $\psi_0({\bm{r}}) = \lim_{j \to \infty} \langle \delta(\bm{r}_j - \bm{r}) \rangle$. The method is used almost exclusively in combination with importance sampling; one defines $f(\bm{R}, \tau) = \psi_T(\bm{R}) \psi(\bm{R}, \tau)$ as the product of the wave function $\psi$ with a time-independent trial wave function $\psi_T({\bm{R}})$. The trial wave function encodes physical knowledge we have about the system before the start of the simulation. It is not unique, but the closer it is to the true (unknown) ground state the more the simulation is enhanced. It should not be orthogonal to the ground state. The Schr[ö]{}dinger equation in imaginary time for $f(\bm{R}, \tau)$ is $$\begin{aligned}
- \frac{ \partial f( {\bm{R}}, \tau)}{\partial \tau} & = & - \frac{\hbar^2}{2m} \big( \nabla_{\bm{R}}^2 f(\bm{R}, \tau) - \nabla_{\bm{R}} [ \bm{F}(\bm{R}) f(\bm{R}, \tau)] \nonumber \\
{} & {} & + (E_L - E_{\rm ref}) f(\bm{R}, \tau) \big),\end{aligned}$$ with $E_L(\bm{R}) = \psi_T(\bm{R})^{-1} H \psi_T({\bm{R}})$ the local energy, $E_{\rm ref}$ a reference energy introduced to stabilize the numerics, and $\bm{F}({\bm{R}}) = 2 \psi_T^{-1}({\bm{R}}) \nabla_{\bm{R}} \psi_T({\bm{R}})$ the quantum drift term. The energy can be calculated as $$E = \frac{\int d\bm{R} E_L(\bm{R}) f(\bm{R}, \tau \to \infty)}{\int d\bm{R} f( \bm{R}, \tau \to \infty)}.$$ In every step, walkers propagate according to the drift term and a random diffusive term with variance $\hbar \delta \tau / m$. The potential energy is then evaluated, which modifies the weight of the walker. In order not to spend computer time on configurations with walkers that have an exponentially small weight, a killing and rebirth step of walkers is built in the algorithm in such a way that the average number of walkers remains constant [@Bajdic2009; @Kolorenc2011]. (note: the bias coming from the size of the population has never been systematically addressed [@Cuervo2005]. There exist however diffusion Monte Carlo variants such as PIGS (path integral ground state) that do not suffer from a finite population bias, which are close in spirit to path integral Monte Carlo (at finite temperature) and which have superior convergence properties than diffusion Monte Carlo [@Sarsa2000; @Cuervo2005; @Rossi2009]. Such a basic algorithm (with an appropriately chosen positive trial wave function) suffices for a bosonic system, and may be considered an alternative to path integral Monte Carlo. In our opinion path integral Monte Carlo is preferable since the bias coming from the trial wave function is not easy to filter out in practice. Another ground state method is reptation Monte Carlo which has been formulated both on the continuum [@Baroni1999] and on the lattice [@Carleo2010]. Methods such as PIGS and reptation Monte Carlo have not found widespread use in the field of cold gases however. For fermions, the infamous sign problem occurs again [@Troyer2005]. In such cases, the nodal surface is built into the trial wave function such that $f(\bm{R}, \tau) > 0$. Walkers should then (ideally) not cross the nodal surface, where the drift term is infinite and pushes the walkers away. Diffusion Monte Carlo is in such cases a variational method: if the exact knowledge of the nodal surface were known, the exact ground state energy can be found, while for any approximation of the nodal surface the obtained energy will be higher than the true ground state energy [@Reynolds82]. There also exist methods with nodal relaxation. It is crucial to obtain a good trial wave function and a good nodal surface. The most general trial wavefunction used in the studies of fermionic cold gases has the form $\psi_T({\bm{R}}) = \psi_J(\bm{R}) \psi_{\rm BCS}(\bm{r})$, with the Jastrow from $\psi_J(\bm{R}) = \prod_{i,j} f_{\sigma, \sigma'}( \vert \bm{r}_{i \sigma} - \bm{r}_{j \sigma'} \vert)$ describing the short-range correlations between particles of different spins at different positions, and the BCS part taken as a Slater determinant of orbitals $\phi(\bm{r}) = \alpha \sum_{k_{\alpha} < k_{\rm max}} e^{i \bm{k}_{\alpha}\cdot \bm{r}} + \phi_s(\bm{r})$, where $\alpha$ is a variational parameter (see Ref. [@Giorgini2008; @Astrakharchik05] and references therein). This can describe Fermi liquid regimes $(\phi_s = 0)$ and s-wave paired phases $(\phi_s \neq 0)$. In practice, simulations are done for $N=14$ to $N=64$ particles and results are then extrapolated to the thermodynamic limit.
Application: the Stoner model with atoms on the repulsive branch
----------------------------------------------------------------
![Growth rate of the pairing and Stoner ferromagnetic instabilities after a quench as a function of the final interaction strength $1/k_F a$. Final interactions with negative (positive) values of $1/k_F a$ correspond to the BCS (BEC) side of the Feshbach resonance. The Stoner instability simultaneously occurs in multiple channels. The most unstable channel is indicated by the solid red line, the others by dashed red lines. The “RPA Stoner” instability corresponds to the RPA result [@Babadi09] with bare as opposed to Cooperon-mediated interactions [@Pekker2011]. Inset: Schematic diagram of the pair creation process showing the binding energy (spring) being absorbed by the Fermi sea (arrows). Reprinted figure with permission from Ref. [@Pekker2011]. Copyright (2011) by the American Physical Society.[]{data-label="fig:Stoner"}](fig_stoner){width="0.8\columnwidth"}
Itinerant ferromagnetism, known from the transition metals such as Co, Ni and Fe, is explained in textbooks in terms of the Stoner mean-field criterion [@Stoner]: when the density of states at the Fermi level times the coupling $U$ becomes unity then the RPA spin susceptibility has a pole signaling a transition to a ferromagnetic phase. However, the mean-field treatment breaks down for strong interactions and mean-field is thus applied beyond its range of validity. The Stoner picture is consequently not without criticism. Kanamori argued that screening should be taken into account, which may even prevent the transition [@Kanamori]. The Stoner criterion also predicts a transition in one dimension where it violates the Lieb-Mattis theorem [@LiebMattis]. Although ferromagnetism is known since ancient times the basic model remains partly unsolved. A quantum simulation of the Stoner model with cold gases would hence answer a fundamental question in condensed matter physics.
The equilibrium state of cold atoms swept across the Feshbach resonance to the repulsive side is however a gas of BEC molecules. The only possibility to observe the Stoner transition is to quench the atoms across the transition such that they remain on the upper branch of the Feshbach resonance. A ferromagnetic state is then created dynamically if the rate at which ferromagnetic correlations develop is sufficiently faster than the rate of molecule formation, [*i.e.,*]{} there might be a window where spin domains can be observed because of the two different time scales for the competing pairing and magnetic instabilities.
There have been a number of theoretical studies of the repulsive two-component Fermi gas [@Houbiers1997; @ZhangDasSarma2005; @Duine2005; @Conduit2009a; @Conduit2009b; @Conduit2009c] assuming the (meta)stability of the repulsive gas. Mean field predicts a second order transition for $k_F a = \pi/2$ for a homogeneous system, with $k_F = (2 \pi^2 (n_{\uparrow} + n_{\downarrow}))^{1/3}$. Studies including next order corrections predict a smaller value of the critical density ($k_F a = 1.054$) and a discontinuous jump in the magnetization.
The MIT experiments on $^6$Li obtained indirect evidence of a Stoner transition [@Jo2009]: a minimum in the kinetic energy, a maximum in the volume and a maximum of atom loss rate (around $k_F a \approx 2.2$, larger than the mean-field prediction). No spin domains, which would be direct evidence, were observed however.
Monte Carlo simulations can only be done in thermal equilibrium. In Ref. [@Pilati2010] several interaction types are used: hard spheres, soft spheres and attractive square well potentials. The absence of the molecular bound states for the latter are implemented by choosing the Jastrow correlation term to be the scattering solution of the square well potential corresponding to positive energy. They calculated the equation of state for the unpolarized gas and found resuls independent of the interaction potential only for $k_F a < 0.4$. When the energy exceeds the energy of a phase separated gas, the gas is unstable to ferromagnetism. A partially ferromagnetic phase was also found in case of spin imbalance. The quantitative determination of the phase diagram is strongly model dependent. In Ref. [@Chang2010] model-specific backflow corrections, known to be important from electron gas and $^3$He studies, were added to the plane wave orbitals used to construct the Slater determinants in the trial wave function. They substantially lowered the energy of a hard-sphere gas, but mattered less for the upper branch. We recall that for $^3$He sophisticated trial wave functions have been developed [@Holzmann06], otherwise wrong phases (overestimating polar fluids) and bad quantitative agreement with experiment is found. We also note that the variational and diffusion Monte Carlo simulations in 2D of Ref. [@Drummond2011] found no ferromagnetic fluid between the paramagnetic fluid and a crystalline structure for hard-sphere interactions. Adding a $r^{-3}$ interaction did not significantly alter the phase diagram.
The competing magnetic and pairing instabilities were studied in a linear stability analysis of collective excitations [@Pekker2011]. When the Cooperon is taken into account, it can be seen in Fig. \[fig:Stoner\] that the pairing instability is dominating over the ferromagnetic instability on both sides of the resonance. In such a study the minimum in kinetic energy is also found in the vicinity of the maximum of the pairing rate. It seriously questions the interpretation of the MIT experiments in terms of a Stoner transition. Other papers also question the validity of the MIT interpretation [@Zhai2009; @ZhangHo2011; @Barth2011], the latter claiming that ferromagnetism on the upper branch with zero range interactions violates the Tan relations in combination with a variational argument for a gas in equilibrium.
The MIT experiments were repeated and improved with a faster change of the scattering length and explicitly measuring spin fluctuations [@Sanner2011]. No domains were visible however, even domains as small as consisting of 5 spins were absent. The molecule formation occurs very fast on a scale of $10 /\epsilon_F$ which is accompanied by strong local heating. The new MIT experiments rule out the existence of a ferromagnetic metastable phase in agreement with [@Pekker2011; @ZhangHo2011]. Hence, a ferromagnetic phase will have to be to specially prepared or sought in other systems, for instance with a narrow resonance, or with different dispersions or for mass or spin imbalance [@Conduit2010a; @Conduit2010b; @Keyserlingk2011].
On the lattice, ferromagnetism is well established. Adding a single hole to the Hubbard model at half filling for $U = \infty$ on a bipartite lattice leads to Nagaoka-ferromagnetism [@Nagaoka]. The ferromagnetic phase extends to finite doping, but has a very low critical temperature ( of the order of a percent of the Fermi energy at most) and is very sensitive to the dispersion according to a DMFT study [@Park07]. Similar conclusions are found in diffusion Monte Carlo (see Ref. [@Carleo2011] for the latest study and references therein for older work).
As a final application of diffusion Monte Carlo we mention that is a tool of preferene to compute exchange functionals used in density functional theory. For cold gases, this was done and combined with the Kohn-Sham equations in Ref. [@Tama_DFT], where the issue of ferromagnetism in a weak optically lattice was also studied.
Conclusion {#sec:conclusion}
==========
We have given an overview of the interplay between quantum simulation in the traditional sense by performing simulations on classical computers of quantum problems, and quantum simulation in the atomic-physics-quantum-optics meaning where an experiment simulates a prototypical model of condensed matter physics which is intractable numerically. We have provided a roadmap showing how different expansions lie at the heart of different types of algorithms, and provided references to the literature for a detailed description of each algorithm. We have seen how large-scale path-integral Monte Carlo simulations of bosonic systems have culminated in excellent agreement between theory and experiment for up to a million particles at experimentally relevant temperatures. For systems with long-range interactions and particles with an internal spin degree of freedom, questions remain however. For fermionic systems, no method with a positive-definite expansion exists. One resorts then to approximations such as density mean-field theory (DMFT), or tries to sample all possible Feynman diagrams (diagrammatic Monte Carlo) and hopes for fast convergence, possibly after analytical manipulations such as series resummations. We have seen examples of both (large scale DCA simulations in order to provide benchmarking for the 3D Hubbard model and the Fermi-polaron problem, to name just a few), and we showed that DMFT methods form a subclass of diagrammatic Monte Carlo methods from the diagrammatic point of view. DMFT, thanks to its widespread use and technical advantages, can hence be used as a promising starting point for diagrammatic Monte Carlo, which seems an interesting avenue for future research. We looked at diffusion Monte Carlo in relation to the controversial issue of ferromagnetism for atoms on the upper branch of a Feshbach resonance.\
[*Acknowledgements*]{} I wish to thank all my teachers, coworkers and colleagues who selflessly explained me the physics of cold gases and Monte Carlo simulations over the past decade. In particular, I wish to mention Peter Anders, Waseem Bakr, George Batrouni, Gianni Blatter, Immanuel Bloch (and coworkers), Massimo Boninsegni, Hanspeter B[ü]{}chler, Evgeni Burovski, Barbara Capogrosso-Sansone, Philippe Corboz, Andrew Daley, Eugene Demler, Tilman Esslinger (and coworkers), Simon F[ö]{}lling, Sebastian Fuchs, Fabrice Gerbier, Thierry Giamarchi, Markus Greiner (and coworkers), Daniel Greif, Emanuel Gull, Fabian Hassler, Sebastian Huber, Gregor Jotzu, Wolfang Ketterle, Corinna Kollath, Evgeny Kozik, Stefan Kuhr, Anatoly Kuklov, Ping Nang Ma, David Pekker, Trey Porto, Sebastiano Pilati, Guido Pupillo, Nikolay Prokof’ev, Stefan Rombouts, Anders Sandvik, Vito Scarola, Radjeep Sensarma, Manfred Sigrist, Ulrich Schneider, Ulrich Schollw[ö]{}ck, Boris Svistunov, Leticia Tarruell, Stefan Trotzky, Matthias Troyer, Felix Werner, Philipp Werner, Wilhelm Zwerger, and Martin Zwierlein (and coworkers). This work was partly supported by a grant from the Army Research Office with funding from the DARPA OLE program.\
[99]{} Troyer M and Wiese U-J 2005 [*Phys. Rev. Lett.*]{} [**94**]{} 170201 Meng Z. Y., Lang T. C., Wessel S., Assaad F. F., Muramatsu A., Nature [**464**]{}, 847 (2010). Bajdic M and Mitas L 2009 [*Acta Physica Slovaca*]{} [**59**]{}, No.2, 81-168 Kolorenc J and Mitas L 2011 [*Rep. Prog. Phys.*]{} [**74**]{} 026502 Georges A, Kotliar G, Krauth W and Rozenberg M J 1996 [*Rev. Mod. Phys.*]{} [**68**]{} 13 Maier T, Jarrell M, Pruschke T and Hettler M H 2005 [*Rev. Mod. Phys.*]{} [**77**]{} 1027 Kotliar G, Savrasov S Y, Haule K, Oudovenko V S, Parcollet O, and Marianetti C A 2006 [*Rev. Mod. Phys.*]{} [**78**]{} 865 Gull E, Millis A J, Lichtenstein A I, Rubtsov A N, Troyer M and Werner P 2011 [*Rev. Mod. Phys.*]{} [**83** ]{} 349 Assaad F F 2001 [*Quantum Monte Carlo Methods on Lattices: The Determinantal method, Lecture notes of the Winter School on Quantum Simulations of Complex Many-Body Systems: From Theory to Algorithms.*]{} (Edited by : J. Grotendorst, D. Marx und A. Muramatsu. Publication Series of the John von Neumann Institute for Computing (NIC). NIC series Vol. 10. ISBN 3-00-009057-6 Pages 99-155) Landau, D P and Binder K 2000 [*A guide to Monte Carlo simulations in statistical physics*]{} (Cambridge University Press) Krauth W (2006) [*Statistical mechanics : algorithms and computations*]{} (Oxford University Press) Liu J S 2001 [*Monte Carlo Strategies in Scientific Computing*]{} (New York: Springer Verlag) Schollw[ö]{}ck U 2011 [*Annals of Physics*]{} [**326**]{} 96 Cazalilla M A, Citro R, Giamarchi T, Orignac E and Rigol M 2011 [*Rev. Mod. Phys.*]{} [**83**]{} 1405 Bloch I, Dalibard J and Zwerger W 2008 [*Rev. Mod. Phys.*]{} [**80**]{} 885 Jaksch D, Bruder C, Cirac J I, Gardin C W and Zoller P 1998 [*Phys. Rev. Lett*]{} [**81**]{}, 3108 Greiner M, Mandel O, Esslinger T, H[ä]{}nsch T W and Bloch I 2002 [*Nature*]{} [**415**]{}, 39 Trotzky S, Pollet L, Gerbier F, Schnorrberger U, BlochI , Prokof’ev N V , Svistunov B and Troyer M 2010 Nature Phys. [**6**]{}, 998 Fuchs S, Gull E, Pollet L, Burovski E, Kozik E, Pruschke T and Troyer M 2011 [*Phys. Rev. Lett.*]{} [**106**]{}, 030401
Fisher M P A, Weichman P B, Grinstein G, Fisher D S 1989 [*Phys. Rev. B*]{} [**40**]{} 546 Sachdev S 1999 [*Quantum Phase Transitions*]{} (New York: Cambridge University Press) Pollet L, Van Houcke K and Rombouts S 2007 [*J. Comp. Phys.*]{} [**225**]{} 2249 Mahan G D 2000 [*Many Particle Physics, 3rd Ed.*]{} (Springer) Negele J W and Orland H 1988 [*Quantum Many-Particle Systems*]{} (Perseus books) Fetter A L and Walecka J D 1971 [*Quantum theory of many-particle systems*]{} (San Francisco: McGraw-Hill) Pollock E L and Ceperley D M 1987 [*Phys. Rev. B*]{} [**36**]{} 8343-8352
Prokof’ev N V, Svistunov B V and Tupitsyn I S 1998 [*JETP*]{} [**87**]{}, 310 Sandvik A W 1999 [*Phys. Rev. B*]{} [**59**]{} 14157R Sylju[å]{}sen O F and Sandvik A W 2002 [*Phys. Rev. E*]{} [**66**]{} 046701 Evertz H G, Lana G and Marcu M 1993 [*Phys. Rev. Lett.*]{} [**70**]{} 875 Beard B B, and Wiese U-J 1996 [*Phys. Rev. Lett.*]{} [**77**]{} 5130 Evertz H G 2003 [*Adv.Phys.*]{} [**52**]{} 1 Rombouts S MA , Van Houcke K and Pollet L 2006 [*Phys. Rev. Lett.* ]{} [**96**]{} 180603 Van Houcke K, Rombouts S M A and Pollet L 2006 [*Phys. Rev. E* ]{} [**73**]{} 056703 Rousseau V G 2007 [*Phys. Rev. E*]{} [**77**]{} 056705 Rousseau V G 2008 [*Phys. Rev. E* ]{} [**78**]{} 056707 Kawashima N and Harada K 2004 [*Phys. Soc. Jpn.*]{} [**73**]{} 1379 Prokof’ev N and Svistunov B 2001 [*Phys. Rev. Lett.*]{} [**87**]{} 160601 Prokof’ev N V and Svistunov B V 2010 [*“Worm Algorithm for Problems of Quantum and Classical Statistics” in “Understanding Quantum Phase Transitions”*]{} (Boca Raton, ed by Lincoln D. Carr, Taylor & Francis) Alet F, and S[ø]{}rensen E S 2003 Rev. E 67 015701(R).
Gerbier F 2007 [*Phys. Rev. Lett.*]{} [**99**]{} 120405 Capogrosso-Sansone B, Prokof’ev N V and Svistunov B V 2007 [*Phys. Rev. B*]{} [**75**]{} 134302 Capogrosso-Sansone B, Giorgini S, Pilati S, Pollet L, Prokof’ev N, Svistunov B and Troyer M 2010 [*New J. Phys.*]{} [**12**]{} 043010 Campostrini M, Hasenbusch M, Pelissetto A and Vicari E 2006 [*Phys. Rev. B*]{} [**74**]{} 144506 Kashurnikov V A, Prokof’ev N V and Svistunov B V 2002 [*Phys. Rev. A*]{} [**66**]{}031601 Batrouni G G, Rousseau V, Scalettar R T, Rigol M, Muramatsu A, Denteneer P J H and Troyer M 2002 [*Phys. Rev. Lett.*]{} [**89**]{} 117203 Wessel S, Alet F, Troyer M, and Batrouni G G 2004 [*Phys. Rev. A*]{} [**70**]{} 053615 Gygi O, Katzgraber H G, Troyer M, Wessel S and Batrouni G G 2006 [*Phys. Rev. A*]{} [**73**]{} 063606 Ma P N, Yang K Y, Pollet L, Porto J V, Troyer M and Zhang F C 2008 [*Phys. Rev. A*]{} [**78**]{} 023605 Mahmud K W, Duchon E N, Kato Y, Kawashima N, Scalettar R T and Trivedi N 2011 [*Phys. Rev. B*]{} [**84**]{} 054302 Pollet L, Svistunov B V and Prokof’ev N V 2010 [*Phys. Rev. Lett.*]{} [**104**]{} 245705 Gerbier F, Trotzky S, F[ö]{}lling S, Schnorrberger U, Thompson J D, Widera A, Bloch I, Pollet L, Troyer M, Capogrosso-Sansone B, Prokof’ev N V and Svistunov B V 2008 [*Phys. Rev. Lett.*]{} [**101**]{} 155303 Dalfovo F, Giorgini S, Pitaevskii L P and Stringari S 1999 [*Rev. Mod. Phys.* ]{} [**71**]{} 46 Schmidt K P, Reischl Aand Uhrig G S 2006 [*Eur Phys J D*]{} [**38**]{} 343 Ho T-L and Zhou Q 2007 [*Phys. Rev. Lett.*]{} [**99**]{} 120404 Pollet L, Kollath C, Van Houcke K and Troyer M 2008 [*New J. Phys.*]{} [**10**]{}, 065001 Gerbier F, Widera A, F[ö]{}lling S, Mandel O, Gericke T, and Bloch I 2005 [*Phys. Rev. A*]{} [**72**]{} 053606 Diener R B, Zhou Q, Zhai H and Ho T L 2007 [*Phys. Rev. Lett.*]{} [**98**]{} 180404 Kato Y, Zhou Q, Kawashima N and Trivedi N 2008 [*Nature Phys.*]{} [**4**]{} 617 Zhou Q, Kato Y, Kawashima N and Trivedi N 2009 [*Phys. Rev. Lett.*]{} [**103**]{} 085701 Pichler H, Daley A J and Zoller P 2010 [*Phys. Rev. A*]{} [**82**]{} 063605 Gerbier F and Castin Y 2010 [*Phys. Rev. A*]{} [**82**]{} 013615
Gemelke N, Zhang X, Hung D L and Chin C 2009 [*Nature*]{} [**460**]{} 995 Bakr W, Gillen J, Peng A, F[ö]{}lling S and Greiner M 2009 [*Nature*]{} [**462**]{} 74 W[ü]{}rtz P, Langen T, Gericke T, Koglbauer A and Ott H 2009 [*Phys. Rev. Lett.*]{} [**103**]{} 080404 Bakr W S, Peng A, Tai M E, Ma R, Simon J, Gillen J I, F[ö]{}lling S, Pollet L and Greiner M 2010 [*Science*]{} [**329**]{} 547 Sherson J F, Weitenberg C, Endres M, Cheneau M, Bloch I and Kuhr S 2010 [*Nature*]{} [**467**]{} 68 Weitenberg C, Endres M, Sherson J F, Cheneau M, Schauss P, Fukuhara T, Bloch I and Kuhr S 2011 [*Nature*]{} [**471**]{} 319 Weitenberg C, Schauss P, Fukuhara T, Cheneau M, Endres M, Bloch I and Kuhr S 2011 [*Phys. Rev. Lett.*]{} [**106**]{} 215301 Endres M, Cheneau M, Fukuhara T, Weitenberg C, Schauss P, Gross C, Mazza M, Banuls M C, Pollet L, Bloch I and Kuhr S 2011 [*Science*]{} [**334**]{} 200 Bakr W S, Preiss P M, Tai M E, Ma R, Simon J and Greiner M 2011 [*Nature*]{} [**480**]{} 500 Hung C L, Zhang X, Gemelke N and Chin C 2010 [*Phys. Rev. Lett.*]{} [**104**]{} 160403 Hung C L, Zhang X, Ha L C, Tung S K, Gemelke N and Chin C 2011 [*New J. Phys.*]{} [**13**]{} 075019 Ho T L and Zhou Q 2009 [*Nat. Phys.*]{} [**6**]{} 131 Ma P N, Pollet L and Troyer M 2010 [*Phys. Rev. A*]{} [**82**]{} 033627 Zhou A and Ho T L 2011 [*Phys. Rev. Lett.*]{} [**106**]{} 225301 Capogrosso-Sansone B, Kozik E, Prokof’ev N V and Svistunov B V 2007 [*Phys. Rev. A*]{} [**75**]{} 013619 Gerbier F, F[ö]{}lling S, Widera A, Mandel O and Bloch I 2006 [*Phys. Rev. Lett.* ]{} [**96**]{} 090401 Zhou Q and Ho T L 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 245702 Fang S, Chung C M, Ma P N, Chen P and Wang D W 2011 [*Phys. Rev. A*]{} [**83**]{} 031605 Hazzard K R A and Mueller E J 2011 [*Phys. Rev. A*]{} [**84**]{} 013604 Guan X and Ho T L 2011 [*Phys. Rev. A*]{} [**84**]{} 023616 Yin X, Guan X W, Chen S and Batchelor M T 2011 [*Phys. Rev. A*]{} [**84**]{} 011602(R) Batchelor M T, Foerster A, Guan X and Kuhn C C N 2010 [*J. Stat. Mech.*]{} P12014. Zhang X, Hung C L, Tung S K, Gemelke N and Chin C 2011 [*New J. Phys.*]{} [**13**]{} 045011 Pollet L, Prokof’ev N and Svistunov B 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 199601 Zhou Q, Kato Y, Kawashima N and Trivedi N 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 199602 Campostrini M and Vicari E 2009 [*Phys. Rev. Lett.*]{} [**102**]{} 240601 Campostrini M and Vicari E 2010 [*Phys. Rev. A*]{} [**81**]{} 023606 Campostrini M and Vicari E 2010 [*Phys. Rev. A*]{} [**81**]{} 063614 Ceccarelli G, Torrero C and Vicari E 2012 [*Phys. Rev. A*]{} [**85**]{} 023616 Ceccarelli G, Torrero C 2012 [*Phys Rev A*]{} [**85**]{} 053637 Donner T,Ritter S, Bourdel T, [Ö]{}ttl A, K[ö]{}hl M and Esslinger T 2007 [*Science*]{} [**315**]{} 1556 Hung C L, Zhang X, Gemelke N and Chin C 2011 [*Nature*]{} [**470**]{} 236 L[ö]{}w R, Weimer H, Krohn U, Heidemann R, Bendkowsky V, Butscher B, B[ü]{}chler H P and Pfau T 2009 [*Phys. Rev. A*]{} [**80**]{}, 033422 Capogrosso-Sansone B, S[ö]{}yler G, Prokof’ev N and Svistunov B 2008 [*Phys. Rev. A*]{} [**77**]{} 015602 Krauth W and Trivedi N 1991 [*Europhys. Lett.*]{} [**14**]{} 627 K[ü]{}hner T D and Monien H 1998 [*Phys. Rev. B*]{} [**58**]{} R14741 K[ü]{}hner T D, White S R and Monien H 2000 [*Phys. Rev. B*]{} [**61**]{} 12474 Rigol M, Batrouni G G, Rousseau V G and Scalettar R T 2009 [*Phys. Rev. A*]{} [**79**]{} 053605 Hen I and Rigol M 2010 [*Phys. Rev. A*]{} [**82**]{} 043634 Jarrell M and Gubernatis J 1996 [*Phys. Rep.*]{} [**269**]{} 133 Pippan P, Evertz H G and Hohenadler M 2009 [*Phys. Rev. A*]{} [**80**]{} 033612
Giamarchi T and Schulz H J 1987 [*Europhys. Lett.*]{} [**3**]{} 1287 Giamarchi T and Schulz H J 1988[*Phys. Rev. B*]{} [**37**]{} 325
Freericks J K and Monien H 1996 [*Phys. Rev. B*]{} [**53**]{} 2691 Scalettar R T, Batrouni G G and Zimanyi G T, 1991 [*Phys. Rev. Lett.*]{} [**66**]{} 3144 Krauth W, Trivedi N and Ceperley D 1991 [*Phys. Rev. Lett.*]{} [**67**]{} 2307 Zhang L and Ma M, 1992 [*Phys. Rev. B*]{} [**45**]{} 4855 Singh K G and Rokhsar D S 1992 [*Phys. Rev. B*]{} [**46**]{} 3002 Makivic M, Trivedi N, and Ullah 1993 [*Phys. Rev. Lett.*]{} [**71**]{} 2307 Wallin M, S[ø]{}rensen E S, Girvin S M, and Young A P 1994 [*Phys. Rev. B*]{} [**49**]{} 12115 Pazmandi F, Zimanyi G, and Scalettar R 1995 [*Phys. Rev. Lett.*]{} [**75**]{} 1356 Pai R V, Pandit R, Krishnamurthy H R and Ramasesha S 1996 [*Phys. Rev. Lett.*]{} [**76**]{}, 2937 Svistunov B V 1996 [*Phys. Rev. B*]{} [**54**]{} 16131 Pazmandi F and Zimanyi G T 1998 [*Phys. Rev. B*]{} [**57**]{} 5044 Kisker J and Rieger H 1997 [*Phys. Rev. B*]{} [**55**]{} R11 981 Herbut I F 1997 [*Phys. Rev. Lett.*]{} [**79**]{} 3502 (1997) Sen P, Trivedi N and Ceperley D M 2001 [*Phys. Rev. Lett.*]{} [**86**]{}, 4092 Prokof’ev N and B. Svistunov B 2004 [*Phys. Rev. Lett.*]{} [**92**]{} 015703 Wu J and Phillips P 2008 [*Phys. Rev. B*]{} [**78**]{}, 014515 Bissbort U and Hofstetter W 2009 [*EPL*]{} [**86**]{} 50007 Weichman P B and Mukhopadhyay R 2008 [*Phys. Rev. B*]{} [**77**]{} 214516 Weichman P B 2008 [*Mod. Phys. Lett. B*]{} [**22**]{}, 2623
Pollet L, Prokof’ev N V, Svistunov B V and Troyer M 2009 [*Phys. Rev. Lett.*]{} [**103**]{} 140402 Gurarie V, Pollet L, Prokof’ev N V, Svistunov B V and Troyer M 2009 [*Phys. Rev. B*]{} [**80**]{} 214519 Falco G M, Nattermann T and Pokrovsky V L 2009 [*Phys. Rev. B*]{} [**80**]{} 104515 S[ö]{}yler S G, Kiselev M, Prokof’ev N V, and Svistunov B V 2011 [*Phys. Rev. Lett.*]{} [**107**]{} 185301 (2011) Prokofev N V and Svistunov B V 1998 [*Phys. Rev. Lett.*]{} [**80**]{} 4355 Rapsch S, Schollw[ö]{}ck U and Zwerger W 1999 [*Europhys. Lett.*]{} [**46**]{} 559 Altman E, Kafri Y, Polkovnikov A and Refael G 2004 [*Phys.Rev. Lett.*]{} [**93**]{} 150402 Altman E, Kafri Y, Polkovnikov A and Refael G 2008 [*Phys.Rev. Lett.*]{} [**100**]{} 170402 Altman E, Kafri Y, Polkovnikov A and Refael G 2010 [*Phys.Rev. B*]{} [**81**]{} 174528
Crowell P A, Van Keuls F W and Reppy J D 1997 [*Phys. Rev. B*]{} [**55**]{} 12620 White M, Pasienski M, McKay C, Zhou S Q, Ceperley D and DeMarco B 2009 [*Phys. Rev. Lett.*]{} [**102**]{} 055301 Pasienski M, McKay D, White M and DeMarco B 2009 [*Disordered insulatro in an optical lattice*]{} Preprint arXv:0908.1182 (unpublished) Yamada F, Tanaka H, Ono T and Nojiri T 2011 [*Phys. Rev. B*]{} [**83**]{} 020409 Hong T, Zheludev A, Manaka H and Regnault L P 2010 [*Phys. Rev. B*]{} [**81**]{} 060410 Yu R, Yin L, Sullivan N S, Xia J S, Huan C, Paduan-Filho A, Oliveira Jr N F, Haas S, Steppke A, Miclea C F, Weickert F, Movshovich R, Mun E D, Zapf V S and Roscilde T 2011 [*Bose glass and Mott glass of quasiparticles in a doped quantum magnet*]{} Preprint arXiv:1109.4403v2 (unpublished) Zheludev A and H[ü]{}vonen D 2011 [*Phys. Rev. B*]{} [**83**]{} 216401 Wulf E, M[ü]{}hlbauer S, Yankova T and Zheludev A 2011 [*Phys. Rev. B*]{} [**84**]{} 174414
Griesmaier A, Werner J, Hensler S, Stuhler J and Pfau T 2005 [*Phys. Rev. Lett.*]{} [**94**]{} 160401 Lahaye T, Koch T, B. Fr[ö]{}hlich B, Fattori M, Metz J, Griesmaier A, Giovanazzi S and Pfau T 2007 [*Nature*]{} [**448**]{} 672 Koch T, Lahaye T, Metz J, Fr[ö]{}hlich B, Griesmaier A and Pfau T 2008 [*Nat. Phys.*]{} [**4**]{} 218 Lu M, Burdick N Q, Youn W H, Lev B L 2011 [*Phys. Rev. Lett.*]{} [**107**]{} 190401 Heidemann R, Raitzsch U, Bendkowsky V, Butscher B, L[ö]{}w R, Santos L and Pfau T 2007 [*Phys. Rev. Lett.*]{}[**99**]{} 163601 Sage J M, Sainis S, Bergeman T and DeMille D 2005 [*Phys. Rev. Lett.*]{} [**94**]{} 203001 Ni K K, Ospelkaus S, deMiranda M H G, Peer A, Neyenhuis B, Zirbel J J, Kotochigova S, Julienne P S, Jin D S and Ye J 2008 [*Science*]{} [**322**]{} 231 Deiglmayr J, Grochola A, Repp M, M[ö]{}rtlbauer K, Gl[ü]{}ck C, Lange J, Dulieu O, Wester R and Weidem[ü]{}ller M 2008 [*Phys. Rev. Lett.*]{} [**101**]{} 133004 Aikawa K, Akamatsu D, Hayashi M, Oasa K, Kobayashi J, Naidon P, Kishimoto T, Ueda M and Inouye S 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 203001 Pupillo G, Micheli A, B[ü]{}chler H P and Zoller P 2008 [*“Condensed Matter Physics with Cold Polar Molecules” in “Cold molecules: Creation and applications”*]{} (Ed. Krems R V, Friedrich B and Swalley W C, Taylor & Francis) B[ü]{}chler H P, Demler E, Lukin M, Micheli M, Prokof’ev N, Pupillo G and Zoller P 2007 [*Phys. Rev. Lett.*]{} [**98**]{} 060404 Menotti C and Lewenstein M 2008 [*Ultra-Cold Dipolar Gases, Recent Progress in Many-Body Theories*]{} (Singapore, Ed. Boronat J, Astrakharchik G and Mazzanti F, Series on Advances in Quantum Many-Body Theory Vol. 11, World Scientific Publishing Co.). Menotti C, Lewenstein M, Lahaye T and Pfau T, 2008 [*Dipolar Interaction in Ultra-Cold Atomic Gases, Dynamics and Thermodynamics of Systems with Long-Range Interactions: Theory and Experiments*]{} (New York, Ed. Campa A, Giansanti A, Morigi G and Sylos Labini F, AIP Conf. Proc. No. 970) Carr L D, DeMille D, Krems R V and Ye J 2009 [*New J. Phys.*]{} [**11**]{} 055049 Dulieu O and Gabbanini C 2009 [*Rep. Prog. Phys.*]{} [**72**]{} 086401 Friedrich B and Doyle J 2009 [*Chem.Phys,Chem.*]{} [**10**]{} 604 Trefzger C [*et al*]{} 2011 [*J. Phys. B: At. Mol. Opt. Phys.*]{} [**44**]{} 193001 Pollet L, Picon J D, B[ü]{}chler H P and Troyer M 2010 [*Phys. Rev. Lett.*]{} [**104**]{} 125302 Capogrosso-Sansone B, Trefzger C, Lewenstein M, Zoller P and Pupillo G 2010 [*Phys. Rev. Lett.*]{} [**104**]{} 125301 Boninsegni M and Prokof’ev N V 2005 [*Phys. Rev. Lett.*]{} [**95**]{} 237204 Wessel S and Troyer M 2005 [*Phys. Rev. Lett.*]{} [**95**]{} 127205 Heidarian D and Damle K 2005 [*Phys. Rev. Lett.*]{} [**95**]{} 127206 Melko R G, Paramekanti A, Burkov A A, Vishwanath A, Sheng D N and Balents L 2005 [*Phys. Rev. Lett.*]{} [**95**]{} 127207 Melko R G, Del Maestro A and Burkov A A 2006 [*Phys. Rev. B*]{} [**74**]{} 214517 Spivak V and Kivelson S A 2005 [*Phys. Rev. B*]{} [**70**]{} 155114 Isakov S, Chien H C, Wu J J, Chen Y C, Chung C H, Sengupta K and Kim Y B 2009 [*Europhys. Lett.* ]{} [**87**]{} 36002 Bak P and Bruinsma R 1982 [*Phys. Rev. Lett.*]{} [**49**]{} 249 Burnell F J, Parish M M, Cooper N R and Sondhi S L 2009 [*Phys. Rev. B*]{} [**80**]{} 174519 Chen Y C, Melko R G, Wessel S and Kao Y J 2008 [*Phys. Rev. B*]{} [**77**]{} 014524 Dang L, Boninsegni M and Pollet L 2008 [*Phys. Rev. B*]{} [**78**]{} 132512 Gorshkov A V, Manmana S R, Chen G, Ye J, Demler E, Lukin M D and Rey A-M 2011 [*Phys. Rev. Lett*]{} [**107**]{} 115301 Thieleman O, Lazarides A and de Morais Smith C 2011 [*Phys. Rev. A*]{} [**83**]{} 013627 Cooper N R and Shlyapnikov G V 2009 [*Phys. Rev. Lett.*]{} [**103**]{} 155302 Quintanilla J, Carr S T and Betouras J J 2009 [*Phys. Rev. A*]{} [**79**]{} 031601(R) Fregoso B M, Sun K, Fradkin E and Lev B 2009 [*New J. Phys.*]{} [**11**]{} 103003 Carr S T, Quintanilla J and Betouras J J 2010 [*Phys. Rev. B*]{} [**82**]{} 045110 Potter A C, Berg E, Wang D-W, Halperin B I and Demler E 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 220406
Kuklov A B and Svistunov B V 2003 [*Phys. Rev. Lett.*]{} [**90**]{} 100401 Kuklov A, Prokof’ev N and Svistunov B 2004 [*Phys. Rev. Lett.*]{} [**92**]{} 030403 Kuklov A, Prokof’ev N and Svistunov B 2004 [*Phys. Rev. Lett.*]{} [**92**]{} 050402 Ohgoe T and Kawashima N 2011[*Phys. Rev. A*]{} [**83**]{} 023622 S. G. S[ö]{}yler, Capogrosso-Sansone B, Prokof’ev N V and Svistunov B V 2011 [*New J. Phys.*]{} [**11**]{} 073036 Capogrosso-Sansone B, S[ö]{}yler S G, Prokof’ev N V and Svistunov B V 2010 [*Phys. Rev. A*]{} [**81**]{} 053622 Hubener A, Snoek M and Hofstetter W 2009 [*Phys. Rev. B*]{} [**80**]{} 245109 Guglielmino M, Penna V and Capogrosso-Sansone B 2010 [*Phys. Rev. A* ]{} [**82**]{} 021601 Guglielmino M, Penna V and Capogrosso-Sansone B 2011 [*Phys. Rev. A* ]{} [**84**]{} 031603(R) de Forges de Parny L, H[é]{}bert F, Rousseau V G, Scalettar R T and Batrouni G G 2011 [*Phys. Rev. B*]{} [**84**]{} 064529 de Forges de Parny L, Traynard M, H[é]{}bert F, Rousseau V G, Scalettar R T and Batrouni G G 2010 [*Phys. Rev. A*]{} [**82**]{} 063602 Pollet L, Troyer M, Van Houcke K and Rombouts S M A 2006 [*Phys. Rev. Lett.*]{} [**96**]{} 190402 Pollet L, Kollath C, Schollw[ö]{}ck U and Troyer M 2008 [*Phys. Rev. A*]{} [**77**]{} 023608 H[é]{}bert F, Batrouni G G, Roy X and Rousseau V G 2008 [*Phys. Rev. B*]{} [**78**]{} 184505 H[’e]{}bert F, Haudin F, Pollet L and Batrouni G G 2007 [*Phys. Rev. A*]{} [**76**]{} 043619 Zujev A, Baldwin A, Scalettar R T, Rousseau V G, Denteneer P J H and Rigol M 2008 [*Phys. Rev. A*]{} [**78**]{} 033619 Varney C N, Rousseau V G and Scalettar R T 2008 [*Phys. Rev. A*]{} [**77**]{} 041608(R) Catani J, De Sarlo L, Barontini G, Minardi F and Inguscio M 2008 [*Phys. Rev. A*]{} [**77**]{} 011603(R) G[ü]{}nter K, St[ö]{}ferle T, Moritz H, K[ö]{}hl M and Esslinger T 2006 [*Phys. Rev. Lett.*]{} [**96**]{} 180402 Ospelkaus S, Ospelkaus C, Wille O, Succo M, Ernst P, Sengstock K and Bongs K 2006 [*Phys. Rev. Lett.*]{} [**96**]{} 180403 Best T, Will S, Schneider U, Hackerm[ü]{}ller L, van Oosten D, Bloch I and L[ü]{}hmann D S 2009 [*Phys. Rev. Lett.*]{} [**102**]{} 030408 Gadway B, Pertot D, Reimann R and Schneble D 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 045303 Sugawa S, Inaba K, Taie S, Yamazaki R, Yamashita M and Takahashi Y 2011 [*Nature Phys.*]{} [**7**]{} 642 Cramer M, Ospelkaus S, Ospelkaus C, Bongs K, Sengstock K and Eisert J 2008 [*Phys. Rev. Lett.*]{} [**100**]{} 140409 Cramer M 2011 [*Phys. Rev. Lett.*]{} [**106**]{} 215302 Lutchyn R M, Tewari S and Das Sarma S 2009 [*Phys. Rev. A*]{} [**79**]{} 011606(R) L[ü]{}hmann D S, Bongs K, Sengstock K and Pfannkuche D 2008 [*Phys. Rev. Lett.*]{} [**101**]{} 050402 Buonsante P, Giampaolo S M, Illuminati F, Penna V and Vezzani A 2008 [*Phys. Rev. Lett.*]{} [**100**]{} 240402 Weld D M, Medley P, Miyake H, Hucul D, Pritchard D E and Ketterle W 2009 [*Phys. Rev. Lett.*]{} [**103**]{} 245301 Weld D M, Miyake H, Medley P, Pritchard D E and Ketterle W 2010 [*Phys. Rev. A*]{} [**82**]{} 051603 Medley P, Weld D M, Miyake H, Pritchard E E and Ketterle W 2011 [*Phys. Rev. Lett.*]{}[**106**]{} 195301 Simon J, Bakr W S, Ma R, Tai M E, Preiss P M and Greiner M 2011 [*Nature*]{} [**472**]{} 307 Sachdev S, Sengupta K and Girvin S M 2002 [*Phys. Rev. B*]{} [**66**]{} 075128 Pielawa S, Kitagawa T, Berg E and Sachdev S 2011 [*Phys. Rev. B*]{} [**83**]{} 205135
Apaja V and Syljuåsen O F 2004 [*Phys. Rev. A*]{} [**74**]{} 035601 Batrouni G G, Rousseau V G and Scalettar R T 2009 [*Phys. Rev. Lett.*]{} [**102**]{} 140402 Rizzi M, Rossini D, De Chiara G, Montangero S and Fazio R 2005 [**95**]{} 240404 Bergkivist S, McCuclloch I P and Rosengren A 2006 [*Phys. Rev. A*]{} [**74**]{} 053419 Demler E and Zhou F 2002 [*Phys. Rev. Lett*]{} [**88**]{} 163001 Snoek M and Zhou F 2004 [*Phys. Rev. B*]{} [**69**]{} 094410 Imambekov A, Lukin M and Demler E 2003 [*Phys. Rev. A*]{} [**68**]{} 063602 Imambekov A, Lukin M and Demler E 2004 [*Phys. Rev. Lett.*]{} [**93**]{} 120405 Krutitsky K V and Graham R 2004 [*Phys. Rev. A*]{} [**70**]{} 063610 Kimura T, Tsuchiya S and Kurihara S 2005 [*Phys. Rev. Lett.*]{} [**94**]{} 110403 Pai R V, Sheshadri K and Pandit R 2008 [*Phys. Rev. B*]{} [**77**]{} 014503 Sadler L E, Higbie J M, Vengalattore M and Stamper-Kurn D M 2006 [*Nature*]{} [**443**]{} 312 Vengalattore M, Guzman J, Leslie S, Serwane F and Stamper-Kurn D M 2009 [*Phys. Rev. Lett.*]{} [**100**]{} 170403 Mahmud K W, Batrouni G G and Scalettar R T 2010 [*Phys. Rev. A*]{} [**81**]{} 033609
Trotter H F 1959 [*Proc. Am. Math. Soc.*]{} [**10**]{} 545 Chin S A and Chen C R 2002 [*J. Chem. Phys.*]{} [**117**]{} 1409 Ceperley D M 1995 [*Rev. Mod. Phys.*]{} [**67**]{} 1601 Boninsegni M, Prokof’ev N V and Svistunov B V 2006 [*Phys. Rev. Lett.*]{} [**96**]{} 070601 Boninsegni M, Prokof’ev N V and Svistunov B V 2006 [*Phys. Rev. E*]{} [**74**]{} 036701 Aziz R A, Nain V P S, Carley J S, Taylor W L and McConville G T 1979 [*J. Chem. Phys.*]{} [**70**]{} 4330 Boninsegni M, Prokof’ev N V and Svistunov B V 2006 [*Phys. Rev. Lett.*]{} [**96**]{} 105301 Clark B K and Ceperley D M 2006 [*Phys. Rev. Lett.*]{} [**96**]{} 105302 Boninsegni M, Kuklov A B, Pollet L, Prokof’ev N V, Svistunov B V and Troyer M 2006 [*Phys. Rev. Lett.*]{} [**97**]{} 080401 Pollet L, Boninsegni M, Kuklov A B, Prokof’ev N V, Svistunov B V and Troyer M 2007 [*Phys. Rev. Lett.*]{} [**98**]{} 15301 Boninsegni M, Kuklov A B, Pollet L, Prokof’ev N V, Svistunov B V and Troyer M 2007 [*Phys. Rev. Lett.*]{} [**99**]{} 035301 Pollet L, Boninsegni M, Kuklov A B, Prokof’ev N V, Svistunov B V and Troyer M 2008 [*Phys. Rev. Lett.*]{} [**101**]{} 097202 Prokof’ev N 2007 [*Adv. Phys.*]{} [**56**]{} Issue 2, 381 Kuklov A B, Prokof’ev N V and Svistunov B V 2011 [*Physics*]{} [**4**]{} 109
Nho K and Landau D P 2004 [*Phys. Rev. A*]{} [**70**]{} 053614 Gr[ü]{}ter P, Ceperley D and Lalo[ë]{} F 1997 [*Phys. Rev. Lett.*]{} [**79**]{} 3549 Kashurnikov V A, Prokof’ev N V and Svistunov B V 2001 [*Phys. Rev. Lett.*]{} [**87**]{} 120402 Arnold P, Moore G and Tom[á]{}$\check{\rm s}$ik B 2001 [*Phys. Rev. A*]{} [**65**]{} 013606 Pilati S, Giorgini S and Prokof’ev N V 2008 [*Phys. Rev. Lett.*]{} [**100**]{} 140405 Landau L D and Lifshitz E M 1977 [*Quantum Mechanics (Non-relativistic Theory)*]{} (Oxford, Pergamon Press), p. 550 Pollock E L and Ceperley D M 1984 [*Phys. Rev. B*]{} [**30**]{} 2555 Pollock E L and Ceperley D M 1987 [*Phys. Rev. B*]{} [**36**]{} 8343 Pilati S, Sakkos K, Boronat J, Casulleras J and Giorgini S 2006 [*Phys. Rev. A*]{} [**74**]{} 043621 Cao J and Berne B J 1992 [*J. Chem. Phys.*]{} [**97**]{} 2382 Baym G, Blaizot J P, Holzmann M, Lalo[ë]{} F and Vautherin D 1999 [*Phys. Rev. Lett.*]{} [**83**]{} 1703 Lee T D and Yang C N 1958 [*Phys. Rev.*]{} [**112**]{} 1419 Seiringer R and Ueltschi D 2008 [*Phys. Rev. B*]{} [**80**]{} 014502 Arnold P and Moore G 2001 [*Phys. Rev. Lett.*]{} [**87**]{} 120401 Krauth W 1996 [*Phys. Rev. Lett.*]{} [**77**]{} 3695 Holzmann M and Krauth W 2008 [*Phys. Rev. Lett.*]{} [**100**]{} 190402 Holzmann M, Chevallier M and Krauth W 2010 [*Phys. Rev. A*]{} [**81**]{}
Pilati S, Giorgini S and Prokof’ev N V 2009 [*Phys. Rev. Lett.*]{} [**102**]{} 150402 Pilati S, Giorgini S, Modugno M and Prokof’ev N 2010 [*New J. Phys.*]{} [**12**]{} 073003 Gordillo M C and Ceperley D M 2000 [*Phys. Rev. Lett.* ]{} [**85**]{} 4735 Lopatin A V and Vinokur V M 2002 [*Phys. Rev. Lett.*]{} [**88**]{} 235503 Kim E and Chan M H W 2004 [*Nature*]{} [**427**]{} 225 Kim E and Chan M H W 2004 [*Science*]{} [**305**]{} 1941 Gross E P 1957 [*Phys. Rev.*]{} [**106**]{} 161 Gross E P 1958 [*Ann. Phys.*]{} [**4**]{} 57 Josserand C, Pomeau Y and Rica S 2007 [*Phys. Rev. Lett.*]{} [**98**]{} 195301 Henkel N, Nath T and Pohl T 2010 [*Phys. Rev. Lett.*]{} [**104**]{} 195302 Cinti F, Jain P, Boninsegni M, Micheli A, Zoller P and Pupillo G 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 135301 Pupillo G, Micheli A, Boninsegni M, Lesanovsky I and Zoller P 2010 [*Phys. Rev. Lett.*]{} [**104**]{} 223002 Astrakharchik G E, Boronat J, Kurbakov I L and Lozovik Y E 2007 [*Phys. Rev. Lett.*]{} [**98**]{}, 060405 Mora C, Parcollet O and Waintal X 2007 [*Phys. Rev. B*]{} [**76**]{} 064511 Filinov A, Prokof’ev N V and Bonitz M 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 070401 Lozovik Y, Volkov S and Willander M 2004 [*JETP Lett.* ]{} [**79**]{} 473 Jain P, Cinti F and Boninsegni M 2011 [*Phys. Rev. B*]{} [**84**]{} 014534 Jain P and Boninsegni M 2011 [*Phys. Rev. A*]{} [**83**]{} 023602
Prokof’ev N V and Svistunov B V 1998 [*Phys. Rev. Lett.*]{} [**81**]{} 2514 Van Houcke K, Kozik E, Prokof’ev N V and Svistunov B V 2008 [*Computer Simulation Studies in Condensed Matter Physics XXI, Eds. Landau D P, Lewis S P and Schuttler H B*]{} (Heidelberg, Springer Verlag)
Burovski E, Prokof’ev N, Svistunov B and Troyer M 2006 [*Phys. Rev. Lett.*]{} [**96**]{}, 160402 Burovski E, Prokof’ev N, Svistunov B and Troyer M 2006 [*New. J. Phys.*]{} [**8**]{}, 153 Rubtsov A N 2003 [*Quantum monte carlo determinantal algorithm without hubbard-stratonovich transformation: a general consideration*]{} Preprint cond-mat/0302228 (unpublished) Rubtsov A N and Lichtenstein A I 2004 Pis’ma v [*JETP*]{} [**80**]{} 67 Rubtsov A N, Savkin V V and Lichtenstein A I 2005 [*Phys. Rev. B*]{} [**72**]{} 035122 Burovski E, Prokof’ev N and Svistunov B 2004 Phys. Rev. B [**70**]{} 193101(R). Burovski E, Kozik E, Prokof’ev N V, Svistunov B V and Troyer M 2008 [*Phys. Rev. Lett.*]{} [**101**]{} 090402 Giorgini S, Pitaevskii L P and Stringari S 2008 [*Rev. Mod. Phys.*]{} [**80**]{} 1215 Ketterle W and Zwierlein M W 2008 in [*in Ultracold Fermi Gases, Proceedings of the International School of Physics “Enrico Fermi”, Course CLXIV, Varenna, 20 - 30 June 2006, Ed. Inguscio M, Ketterle W and Salomon C*]{} (Amsterdam, IOS)
Goulko O and Wingate M 2010 [*Phys. Rev. A*]{} [**82**]{} 053621 Bulgac A, Drut J and Magierski P 2008 [*Phys Rev A*]{} [**78**]{} 23625 Bulgac A, Drut J and Magierski P 2006 [*Phys. Rev. Lett.*]{} [**96**]{} 90404 Bulgac A, Drut J and Magierski P2007 [*Phys. Rev. Lett.* ]{}[**99**]{} 120401 Akinenni V K, Ceperley D M and Trivedi N 2007 [*Phys. Rev. B* ]{}[**76**]{} 165116 Van Houcke K, Werner F, Kozik E, Prokof’ev N, Svistunov B, Ku M, Sommer A, Cheuk L W, Schirotzek A and Zwierlein M W 2012 [*Nat. Phys.*]{} [**8**]{} 366 Ku M J H, Sommer A T, Cheuk L W and Zwierlein M W 2012 [*Science*]{} [**335**]{} 563 Nascimb[è]{}ne S, Navon N, Jiang K J, Chevy F and Salomon C 2010 [*Nature* ]{}[**463**]{} 1057 Forbes M M, Gandolfi S and Gezerlis A 2011 [*Phys. Rev. Lett.*]{} [**106**]{} 235303 Haussmann R, Rantner W, Cerrito S and Zwerger W 2007 [*Phys. Rev. A* ]{} [**75**]{} 023610 Carlson J, Chang S Y, Pandharipande V R and Schmidt K E 2003 [*Phys. Rev. Lett.*]{} [**91**]{} 050401 Astrakharchik G E, Boronat J, Casulleras J and Giorgini S 2004 [*Phys. Rev. Lett.*]{} [**93**]{} 200404
Prokof’ev N.V. and Svistunov B. V. 2007 [*Phys. Rev. Lett.*]{} [**99**]{} 250201
Mishchenko A S, Prokof’ev N V, Sakamoto A and Svistunov B V 2000 [*Phys. Rev. B*]{} [**62**]{} 6317 Marchand D J J, De Filippis G, Cataudella V, Berciu M, Nagaosa N, Prokof’ev N V, Mishchenko A S and Stamp P C E 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 266605
Chevy F and Mora C 2010 [*Rep. Prog. Phys.*]{} [**73**]{} 112401 Prokof’ev N and Svistunov B 2008 [*Phys. Rev. B*]{} [**77**]{} 020408 Prokof’ev N V and Svistunov B V 2008 [*Phys. Rev. B*]{} [**77**]{} 125101 Chevy F 2006 [*Phys. Rev. A*]{} [**74**]{} 063628 Combescot R, Recati A, Lobo C and Chevy F 2007 [*Phys. Rev. Lett.*]{} [**98**]{} 180402 Lobo C, Recati A, Giorgini S and Stringari S 2006 [*Phys. Rev. Lett.*]{} [**97**]{} 20040 Skorniakov G V and Ter-Martirosian K A 1957 [*Sov. Phys. JETP*]{} 4 Mora C and Chevy F 2009 [*Phys. Rev. A*]{} [**80**]{} 033607 Punk M, Dumitrescu P T and Zwerger W 2009 [*Phys. Rev. A*]{} [**80**]{} 053605 Combescot R, Giraud S and Leyronas X [*Europhys. Lett.*]{} [**88**]{} 60007 Schirotzek A, Wu C-H, Sommer A and Zwierlein M W 2009 [*Phys. Rev. Lett.*]{} [**102**]{} 230402 Nascimb[è]{}ne S, Navon N, Jiang K, Tarruell L, Teichmann M, McKeever J, Chevy F and Salomon C 2009 [*Phys. Rev. Lett.*]{} [**103**]{} 170402
Kozik E, Van Houcke K, Gull E, Pollet L, Prokof’ev N, Svistunov B and Troyer M 2010 [*Europhys. Lett.*]{} [**90**]{} 10040 Pollet L, Prokof’ev N V, and Svistunov B V 2011 [*Phys. Rev. B*]{} [**83**]{}, 161103 Blankenbecler R, Scalapino D J and Sugar R L 1981 [*Phys. Rev. D*]{} [**24**]{} 2278 Hirsch J E and Fye R M 1986 [*Phys. Rev. Lett.*]{} [**56**]{} 2521 Rombouts S M A, Heyde K and Jachowicz N 1999 [*Phys. Rev. Lett.*]{} [**82**]{} 4155 Gull E, Werner P, Parcollet O and Troyer M 2008 [*Europhys. Lett.*]{} [**82**]{} 57003 Cai Z, Hung H, Wang L, Zheng D and Wu C 2012 [*arXiv:1202.6323*]{} (unpublished)
Werner P, Comanac A, De Medici L, Troyer M and Millis A J 2006 [*Phys. Rev. Lett.*]{} [**97**]{} 076405 Rubtsov A N, Katsnelson M I and Lichtenstein A I 2008 [*Phys. Rev. B*]{} [**77**]{} 033101 Potthoff M 2003 [*Eur. Phys. J. B*]{} [**32**]{} 429 Potthoff M, Aichhorn M and Dahnken C 2003 [*Phys. Rev. Lett.*]{} [**91**]{} 206402
Anderson P W 1958 [*Phys. Rev.*]{} [**109**]{} 1492 Billy J, Josse V, Zuo Z, Bernard A, Hambrecht B, Lugan P, Cl[é]{}ment D, Sanchez-Palencia L, Bouyer P and Aspect A 2008 [*Nature*]{} [**453**]{} 891 Sanchez-Palencia L, Cl[é]{}ment D, Lugan P, Bouyer P, Shlyapnikov G V and Aspect A 2007 [*Phys. Rev. Lett.*]{} [**98**]{} 210401 Roati G, D’Errico C, Fallani L, Fattori M, Fort C, Zaccanti M, Modugno G, Modugno M and Inguscio M 2008 [*Nature*]{} [**453**]{} 895 Jendrzejewski F, Bernard A, Mueller K, Cheinet P, Josse V, Piraud M, Pezz[é]{} L, Sanchez-Palencia L, Aspect A, Bouyer P 2012 [*Nat. Phys.*]{} [**8**]{} 398 Kondov S S, McGehee W R, Zirbel J J and DeMarco B 2011 [*Science*]{} [**334**]{} 66
J[ö]{}rdens R, Strohmaier N, G[ü]{}nter K, Moritz H and Esslinger T 2008 [*Nature*]{} [**455**]{} 204 Schneider U, Hackerm[ü]{}ller L, Will S, Best T, Bloch I, Costi T A, Helmes R W, Rasch D and Rosch A 2008 [*Science*]{} [**322**]{} 1520 De Leo L, Kollath C, Georges A, Ferrero M and Parcollet O 2008 [*Phys. Rev. Lett.*]{} [**101**]{} 210403 McKay D and DeMarco B 2011 [*Rep. Prog. Phys.*]{} 74 054401 Staudt R, Dzierzawa M and Muramatsu A 2000 [*Eur. Phys. J. B*]{} [**17**]{} 411 Kent P R C, Jarrell M, Maier T A and Pruschke T 2005 [*Phys. Rev. B*]{} [**72**]{}, 060411 Betts D D and Stewart G E 1997 [*Can. J. Phys.*]{} [**75**]{} 47 Betts D D, Lin H Q and Flynn J S 1999 [*Can. J. Phys.*]{} [**77**]{} 353 J[ö]{}rdens R, Tarruell L, Greif D, Uehlinger T, Strohmaier N, Moritz H, Esslinger T, De Leo L, Kollath C, Georges A, Scarola V, Pollet L, Burovski E, Kozik E and Troyer M 2010 [*Phys. Rev. Lett.*]{} [**104**]{} 180401 Oitmaa J, Hamer C and Zheng W 2006 [*Series Expansion Methods for Strongly Interacting Lattice Models*]{} (Cambridge, Cambridge University Press) De Leo L, Bernier J S, Kollath C, Georges A and Scarola V W 2011 [*Phys. Rev. A*]{} [**83**]{} 023606 Scarola V W, Pollet L, Oitmaa J and Troyer M 2009 [*Phys. Rev. Lett.*]{} [**102**]{} 135302 Werner F, Parcollet O, Georges A and Hassan S R 2005 [*Phys. Rev. Lett.*]{} [**95**]{} 056401 Wessel S 2010 [*Phys. Rev. B*]{} [**81**]{} 052405 Paiva T, Loh Y L, Randeria M, Scalettar R T and Trivedi N 2011 [*Phys. Rev. Lett.*]{} [**107**]{} 086401 Gorelik E V, Titvinidze I, Hofstetter W, Snoek M and Bl[ü]{}mer N 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 065301 Paiva T, Scalettar R, Randeria M and Trivedi N 2010 [*Phys. Rev. Lett.*]{} [**104**]{} 066406 White S R and Feiguin A E 2004 [*Phys. Rev. Lett.*]{} [**93**]{} 076401 Daley A, Kollath C, Schollw[ö]{}ck U and Vidal G 2004 [*J. Stat. Mech.: Theory Exp.* ]{} P04005 Eckstein M, Hackl A, Kehrein S, Kollar M, M[ö]{}ckel M, Werner P and Wolf F A 2010 [*Eur. Phys. J. Special Topics*]{} [**180**]{} 217 Gull E, Reichman D R and Millis A J 2011 [*Phys. Rev. B*]{} [**84**]{} 085134 Chitra R and Kotliar G 2001 [*Phys. Rev. B*]{} [**63**]{}, 115110 Anders P, Gull E, Pollet L, Troyer M and Werner P 2010 [*Phys. Rev. Lett.*]{} [**105**]{}, 096402 Anders P, Gull E, Pollet L, Troyer M and Werner P 2011 [*New J. Phys*]{} [**13**]{} 075013 Semerjian G, Tarzio M and Zamponi F 2009 [*Phys. Rev. B*]{} [**80**]{} 014524 Byczuk K and Vollhardt D 2008 [*Phys. Rev. B*]{} [**77**]{} 235106 Hu W J and Tong N H 2009 [*Phys. Rev. B*]{} [**80**]{} 245110 Knap M, Arrigoni E and von der Linden W 2011 [*Phys. Rev. B*]{} [**83**]{}, 134507 Arrigoni E, Knap M and von der Linden W 2011 [*Phys. Rev. B*]{} [**84**]{}, 01453 Knap M, von der Linden W and Arrigoni E 2011 [*Phys. Rev. B*]{} [**84**]{} 11514 Li Y Q, Bakhtiari M R, He L and Hofstetter W 2011 [*Phys. Rev. B*]{} [**84**]{} 144411 Li Y Q, Bakhtiari M R, He L and Hofstetter W 2012 [*Phys. Rev. A*]{} [**85**]{} 023624
Cuervo J E, Roy P N and Boninsegni M 2005 [*J. Chem. Phys.*]{} [**122**]{} 114504 Sarsa A, Schmidt K E and Magro W R 2000 [*J. Chem. Phys.*]{} [**113**]{} 1366 Rossi M, Nava M, Reatto L and Galli D E 2009 [*J. Chem. Phys*]{} [**131**]{} 154108 Baroni S and Moroni S 1999 [*Phys. Rev. Lett.*]{} [**82**]{} 4745 Carleo G, Becca F, Moroni S and Baroni S 2010 [*Phys. Rev. E*]{} [**82**]{} 046710 Reynolds P J, Ceperley D M, Alder B J and Lester Jr W A 1982 [*J. Chem. Phys.*]{} [**77**]{} 5593 Astrakharchik G E, Boronat J, Casulleras J and Giorgini S 2005 [*Phys. Rev. Lett.*]{} [**95**]{} 230405
Babadi M, Pekker D, Sensarma R, Georges A, and Demler E 2009 [*Non-equilibrium dynamics of interacting Fermi systems in quench experiments*]{} Preprint arXiv:0908.3483 (unpublished) Pekker D, Babadi M, Sensarma R, Zinner N, Pollet L, Zwierlein M W and Demler E 2011 [*Phys. Rev. Lett.*]{} [**106**]{} 50402
Stoner E C 1933 [*Phil. Mag.* ]{} [**15**]{} 1018 Kanamori J 1963 [*Prog. Theor. Phys*]{} [**30**]{} 275 Lieb E and Mattis D 1962 [*Physical Review*]{} [**125**]{} 164 Houbiers M, Ferwerda R, Stoof H T C, McAlexander W I, Sackett C A and Hulet R G 1997 [*Phys. Rev. A*]{} [**56**]{} 4864 Zhang Y and Das Sarma S 2005 [*Phys. Rev. B*]{} [**72**]{} 115317 Duine R A and MacDonald A H 2005 [*Phys. Rev. Lett.*]{} [**95**]{} 230403 Conduit G J and Simons B D 2009 [*Phys. Rev. A*]{} [**79**]{} 053606 Conduit G J, Green A G and Simons B D 2009 [*Phys. Rev. Lett.*]{} [**103**]{} 207201 Conduit G J and Simons B D 2009 [*Phys. Rev. Lett.* ]{} [**103**]{} 200403 Jo G B, Lee YR, Choi J H, Christensen CA, Kim T H, Thywissen J H, Pritchard D E and Ketterle W 2009 [*Science*]{} [**325**]{} 5947 Pilati S, Bertaina G, Giorgini S and Troyer M 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 30405 Chang S Y, Randeria M and Trivedi N 2011 [*Proc. Nat. Acad. Sci.*]{} [**108**]{} 51 Holzmann M, Bernu B and Ceperley D M 2006 [*Phys. Rev. B*]{} [**74**]{} 104510 Drummond N D, Cooper N R, Needs R J and Shlyapnikov G V 2011 [*Phys. Rev. B*]{} [**83**]{} 195429 Zhai H 2009 [*Phys. Rev. A*]{} [**80**]{} 051605 Zhang S and Ho T L 2011 [*New J. Phys.*]{} [**13**]{} 055003 Barth M and Zwerger W 2011 [*Ann. Phys.*]{} [**326**]{} 2544 Sanner C, Su E J, Huang W, Keshet A, Gillen J and Ketterle W 2011 [*Correlations and Pair Formation in a Repulsively Interacting Fermi Gas*]{} Preprint arXiv:1108.2017 (unpublished) Conduit G J and Altman E 2010 [*Phys. Rev. A*]{} [**82**]{} 043603 Conduit G J 2010 [*Phys. Rev. A*]{} [**82**]{} 043604 von Keyserlingk C W and Conduit G J 2011 [*Phys. Rev. A*]{} [**83**]{} 053625
Nagaoka Y 1966 [*Phys. Rev.*]{} [**147**]{} 392 Park H, Haule K, Marianetti C A and Kotliar G 2008 [*Phys. Rev. B*]{} [**77**]{} 035107 Carleo G, Moroni S, Becca F and Baroni S 2011 [*Phys. Rev. B*]{} [**83**]{} 060411(R) Ma P N, Pilati S Troyer M and Dai X 2012 [it Nature Phys.]{} doi:10.1038/nphys2348
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Using the phenol-terminated polycarbonate blend as an example, we demonstrate that the hydrodynamic boundary conditions for a flow of an adsorbing polymer melt are extremely sensitive to the structure of the epitaxial layer. Under shear, the adsorbed parts (chain ends) of the polymer melt move along the equipotential lines of the surface potential whereas the adsorbed additives serve as the surface defects. In response to the increase of the number of the adsorbed additives the surface layer becomes thinner and solidifies. This results in a gradual transition from the slip to the no-slip boundary condition for the melt flow, with a non-monotonic dependence of the slip length on the surface concentration of the adsorbed ends.'
author:
- Xin Zhou
- Denis Andrienko
- Luigi Delle Site
- Kurt Kremer
bibliography:
- 'man.bib'
title: 'Flow boundary conditions for chain-end adsorbing polymer blends'
---
Introduction {#sec:intro}
============
The equations of continuum fluid mechanics are incomplete without appropriate boundary conditions. In most situations it is required that both normal and tangential components of the relative fluid velocity vanish at the surface. [@batchelor2000] This, so called stick or no-slip boundary condition, has successfully accounted for most of the experimental facts. It is, however, empirical by nature: there are no theoretical arguments in favor of the no-slip; moreover, it has been known since Maxwell’s times [@Maxwell67] that even a simple kinetic theory of gasses predicts the non-zero value of the tangential velocity at the wall.
Providing some of the insight into the question, kinetic theory fails already for simple fluids adjacent to a rigid solid. In a more general context, multiple scattering from the individual wall molecules remains the major problem of most analytical theories. In this situation computer simulation techniques are able to advance our knowledge of the processes occurring at the surfaces.
Indeed, molecular dynamics (MD) simulations demonstrated that, in case of simple fluids, the flow boundary conditions are sensitive to the fluid epitaxial order [@Thompson90] as well as the wall structure. [@cottinbizonne2003] The situation, however, becomes much more complicated for polymers at the surfaces, because of the much richer molecular arrangement (e. g. formation of brushes, [@klein1996] various adsorbed and depleted layers [@aubouy1996]) as well as much stronger correlation between the atoms right at the wall with the rest of the surface layer. [@ThompsonGR92; @harmandaris2005; @priezjev2004]
Recent MD studies of the end-adsorbing polymer melts, performed with a novel quantum based multiscale approach for the surface/polymer interaction, showed that at least two mechanisms contribute to the hydrodynamic boundary conditions. [@zhou2005a] The attached parts of the chains scatter on the surface potential while moving along its equipotential lines. This induces the density and the chain conformation modulation in the adsorbed layer, and energy is lost from these modulations through the coupling to the thermostat, similar to the situation observed for a generic model of adsorbed surface layers. [@cieplak1994] On the other hand, single-end grafted chains of polymer brushes undergo a coil/stretch transition and disentangle from the melt at a given shear rate, [@brochard1992; @ajdari1994; @smith2005; @brochardwyart1994] favoring the slip boundary condition. Though investigated for the special case of end adsorbing polycarbonate on nickel, a similar scenario will occur for block copolymers with an adsorbing block.
Valid for [*monodispersed*]{} melts, this picture does not account for the usual melt polydispersity, which is an outcome of all synthetic polymerization reactions. Moreover, in many situations the self-blending of polymer melts is used. [@CheahC03] A small amount of a lower weight polymer improves the melt processability without significantly affecting its mechanical properties. It, however, alters the structure of the surface layer [@andrienko2005a] modifying the hydrodynamic boundary conditions for the melt flow. These changes shall be taken into account at the later stages of melt processing. [@NamhataGA99]
In this work we focus on the hydrodynamic boundary conditions for polymer [*blends*]{} adsorbed on a solid substrate. For this purpose we consider a particular system, that is bisphenol-A polycarbonate (BPA-PC) melt sheared over a (111) nickel surface. Our choice of polycarbonate as a test system is twofold. First, it is widely used in various applications [@derudder2000] and, therefore, has been intensively studied both experimentally [@Morbitzer88] and theoretically. [@TsaiLC98; @TschopKBBH98; @TschopKHBB98; @DelleSiteAAK02; @DelleSiteLK04] In particular, shear of monodispersed melts [@zhou2005a] and the structure of the static epitaxial layer for monodispersed melts [@AbramsK03; @AbramsDK03] and blends [@andrienko2005a] have already been considered. Second, although it is a specific system, it has a number of important generic features, which also apply to a realistic description of many polymer brushes composed of block copolymers.
Our prime goal is to relate the structure of the adsorbed layer, which changes in the presence of the low molecular weight additive, to the hydrodynamic boundary conditions for the melt flow, specified by the slip length and the surface friction coefficient.
Simulation details
==================
![ Chemical structures: (a) phenol-terminated bisphenol-[*A*]{} polycarbonate (in this study $n=1,5,20$); (b) diphenyl carbonate (DPC); (c) phenol. []{data-label="fig:systems"}](fig1.eps){width="6cm"}
We consider four types of polymer mixtures. The host polymer (major component) is the phenol terminated BPA-PC of $N_1 = 20$ repeat units. The second (minor) component is one of the following: BPA-PC of $N_2 =
5$ repeat units, one repeat unit, diphenyl carbonate (DPC), or phenol, all shown in Fig. \[fig:systems\]. For all mixtures, we use $n_1 = 400$ chains of the major component; the number of molecules of the minor component, $n_2$, is then adjusted to provide (approximately) $5\%$ of the total weight of the system. Exact numbers are given in Table \[tab:systems\].
$N_1,N_2$ $n_1,n_2$ $L_z, \sigma$
----------- ----------- --------------- -- --
20 400 80.91
20,5 400,76 85.15
20,1 400,239 84.97
20,DPC 400,521 84.69
20,Phenol 400,1186 83.79
: Studied systems. $\sigma = 4.41 \AA$. []{data-label="tab:systems"}
To simulate the melt, we use the previously developed coarse-graining model, in which each monomer is replaced by four beads that correspond to isopropylidene, carbonate, and the two linking phenylenes. Interaction potentials, bead sizes and coarse-graining procedure are described in Refs. [@AbramsK03; @AbramsDK03]. The bead-wall interaction potentials are obtained from [*ab initio*]{} density functional calculations. All [*internal*]{} beads experience strong increasing repulsion below $3.2\, \AA$, either due to the nature of the molecular interaction or due to the steric hindrance by the other beads. Only the chain ends absorb on the wall. The wall-end interaction potential is expanded in 2D reciprocal lattice space of (111) nickel surface and has the following form $$\begin{aligned}
U(x,y,z) = \sum_{i} U_{i}(z) \, f_{i}(x,y),
\label{eq:potential}\end{aligned}$$ where $i=0,1,2$ corresponds to the reciprocal vectors of different lengths, $f_{0} = 1$, $f_{1}=\cos({\bar x}-{\bar y})+\cos({\bar
x}+{\bar y})+\cos 2{\bar y}$, and $f_{2}=\cos({\bar x}-3{\bar
y})+\cos({\bar x} + 3{\bar y})+ \cos 2 {\bar x}$, where $({\bar
x},{\bar y}) = \frac{2 \pi}{a}\,(x,\frac{y}{{\sqrt 3}})$,
$$\begin{aligned}
\nonumber
U_{0} &=& \left\{
\begin{array}{lll}
\frac{5}{3} {\epsilon}_{r}
\left[ \frac{2}{5}
\left( \frac{z_0}{z} \right)^{10} -
\left( \frac{z_0}{z} \right)^4 + \frac{3}{5} \right] - \epsilon_0,
& z < z_{0} \\
\frac{\epsilon_0}{2} \left[
\cos \left(\pi \frac{z_c - z}{z_c-z_0} \right)-1 \right],
& z_0 \le z < z_c
%\\ 0, & z \ge z_c ,
\end{array}
\right.\end{aligned}$$
$$\nonumber
U_{1,2} = \left\{
\begin{array}{lll}
- \epsilon_{1,2}, & z < z_0 \\
\frac{\epsilon_{1,2} }{2}
\left[ \cos \left( \pi \frac{z_c - z}{z_c-z_0} \right)-1 \right],
& z_0 \le z < z_c
%\\ 0, & z \ge z_c.
\end{array}
\right.$$
The interaction potential obtained using [*ab initio*]{} calculations is well reproduced by the following set of parameters: $\epsilon_{r} =
1.5\,
\rm eV$, $\epsilon_{0}= 0.7\, \rm eV$, $\epsilon_{1} = -7/45\, \rm
eV$, and $\epsilon_{2} = -2/45\, \rm eV$. For details, see Refs. [@DelleSiteAAK02; @DelleSiteAA03; @zhou2005a].
The melts are confined to a slit pore of thickness $L_z$ with the walls perpendicular to the $z$ axis. Periodic boundary conditions are employed in $x$ and $y$ directions. The $x$ and $y$ box dimensions are set to $L_x = 22.23\, \sigma$ and $L_y = 21.72\,\sigma$ with $\sigma = 4.41\,\AA$, which corresponds to a (111) hexagonal lattice of nickel with 39 and 22 unit cells. The number density of beads is $n = 0.85$, which corresponds to $1.05\,
\rm g/cc$, the experimental density at the processing temperature, $570\, \rm K$. The units are chosen such that $k_{B} T =1$ with $T=570\,\rm K$ and $\sigma$ is unity. [@AbramsK03; @AbramsDK03]
Starting configurations are generated by randomly placing the chains in the simulation box. A short run is then used to remove the bead-bead overlaps. [@AbramsDK03] The production run is performed in the $NVT$ ensemble with Langevin thermostat with friction $0.5
\tau^{-1}$, where $\tau$ is the unit of time in the simulations. The thermostat is switched off in the shear direction. The velocity-Verlet algorithm with the timestep $\Delta t = 0.005 \tau$ is used to integrate the equations of motion. After equilibration for about $2
\times 10^5$ $\tau$, the shear is applied by moving the top and bottom walls in opposite directions at a constant velocity $v_w$, so that the shear rate is $\dot{\gamma} = 2v_{w}/L_z$. The wall velocity is the same as in our previous studies of [*monodispersed*]{} BPA-PC melts, [@zhou2005a] $v_{w} \tau/\sigma =
0.01$. The corresponding shear rate can be obtained from the time mapping $1 \tau \approx 25\, \rm ps$, see Ref. [@leon2005] for details. Using this mapping we obtain $\gamma \approx 10^7 s^{-1}$. Taking into account that the average chain length in a BPA-PC melt is $N \sim 70$ the corresponding chain reptation time $\tau_d \sim
N^{3.4}$ is almost by two orders of magnitude larger than that of the $N=20$ chains. Equivalently, the shear rate would be reduced to $10^5\, s^{-1}$, close to the value used for industrial processing of BPA-PC, $\sim 10^4\, s^{-1}$. Note that smaller shear rates either have no significant effect on the adsorbed layer or are difficult to analyze, due to significant error bars for the velocity profiles.
Results
=======
Blends with different additives
-------------------------------
We first have a look at the structure of the surface layer. Fig. \[fig:end\_density\] shows a typical chain end density profile for a monodispersed melt ($N=20$). The chain end density has a sharp peak next to the wall, which is due to the strong adsorption of the ends; the region with no ends follows; finally, the bulk concentration is reached at a distance of the order of the radius of gyration, $R_g
\approx 6.2 \sigma = 27\, \AA$. The inset illustrates how this profile changes in the presence of the additives: the shorter the chains of the [*minor*]{} component, the more of them adsorb on the wall. This of course results in a decrease of the number of the adsorbed chain ends of the [*major*]{} component, in agreement with our earlier studies [@andrienko2005a] except for the phenol additive, for which we previously observed a weak increase in the number of the adsorbed chain ends of the major component compared to the DPC case.
![ Typical chain end density profile ($N=20$). The inset depicts the change in the surface density (number of the adsorbed chain ends per unit area) for the minor and major components. The log scale is used to show the bulk density and the depleted region on the same plot. []{data-label="fig:end_density"}](fig2.eps){width="8cm"}
The representative snapshots of the systems with and without shear are shown in Fig. \[fig:snapshots\]. Snapshots (a) and (c) illustrate the thinning of the adsorbed layer due to the decrease in the number of the adsorbed chain ends of the major component. After shear is applied, the one-end attached chains disentangle from the melt, stretch and form a thin lubricating layer between the bulk and the chains adsorbed with two ends (see snapshots (b) and (d)). Corresponding changes can also be detected by observing the center of mass profiles. [@zhou2005a]
![ Snapshots of the mixtures: 20:5 without (a) and with (b) shear, 20:phenol without (c) and with (d) shear. Polymer chains are divided into three populations: chains which adsorb both ends (green), only one end (red), and no ends on the surface (blue). []{data-label="fig:snapshots"}](fig3.eps){width="8cm"}
Having resolved the structure of the adsorbed layer, we shall turn our attention to the velocity profiles, which are shown in Fig. \[fig:velocity\]. For all mixtures, the profiles have features similar to those of monodispersed systems. [@zhou2005a] Next to the wall the velocity is practically constant, since the adsorbed chains are dragged by the wall. Immediately after the plateau, the velocity profile becomes a linear function of $z$. The velocity of the beads at the wall, $v_s$, is smaller than the wall velocity, $v_w$, i. e. the chain ends slide over the wall, moving between the hollow and bridge sites of the potential (the difference in the adsorption energies of these two sites is rather small, of the order of $2\,kT$, see the contour plot in Fig. \[fig:surface\]). In case of the monodispersed system and $20:5$ mixture we have $v_s/v_w \sim 0.5$. This ratio increases for the $20:1$ mixture and reaches $1$ for the diphenyl carbonate and phenol additives.
![ Velocity profiles normalized to the wall velocity. Only a part of the profiles in the vicinity of the wall is shown. The inverse of the surface friction coefficient $\beta^{-1} \sim 1 - v_s / v_w$ (circles) and the shear stress ${f}_x/S$ (squares) are shown in the inset. To calculate the velocity profiles we average over all beads, independent of the nature of a particular bead. []{data-label="fig:velocity"}](fig4.eps){width="8cm"}
We can write the stationary shear stress $\tau_{xz}$ (force per unit area) as [@leger1999] $$\begin{aligned}
\tau_{xz} \equiv f_x / S = \beta (v_w - v_s),\end{aligned}$$ where $\beta$ is the friction coefficient between the adsorbed layer and the wall, $S$ is the surface area. Since the density of the minor component in the bulk is small ($5\%$ in weight and most of it is adsorbed on the walls) the bulk properties of the melt do not change significantly from mixture to mixture. Indeed, the shear stress $\tau_{xz}$ is practically independent of the molecular weight of the additive (see the inset to Fig. \[fig:velocity\]). Therefore, the change of $v_s$ is due to the different values of the friction coefficient $\beta$, which increases significantly for short additives.
This increase is, in fact, rather unexpected: the adsorbed additives reduce the number of the attractive sites available for the major component and, in principle, shall screen the effective interaction of the melt with the surface. Weaker interaction should result in a smaller, compared to the monodispersed melt, friction coefficient. This, however, does not happen in practice: the systems with the phenol or diphenyl carbonate additives, which basically cover about 80% of the wall, have the biggest friction coefficient.
![ Snapshots of the chain ends adsorbed on the walls. Blue beads: major component, yellow beads: minor component. The bead sizes correspond to their van der Waals radii. The triangular grid depicts the lattice of the (111) nickel surface. Contour plot illustrates the bead-wall potential used in simulation, together with the possible paths for the adsorbed beads. The contour levels are in electron volts, $1\, {\rm eV} \approx 20\, kT$. []{data-label="fig:surface"}](fig5.eps){width="8cm"}
To understand why this happens, let us have a look at the structure of the beads adsorbed on the surface. Fig. \[fig:surface\] shows the snapshots of the representative mixtures. The difference between the systems is now clear: the adsorbed layer of the 20:5 mixture has a structure similar to a two-dimensional (2D) gas; the packing of the 20:1 mixture is more dense, but still disordered, similar to a 2D liquid; for the 20:DPC as well as 20:Phenol (not shown) mixtures the concentration of the adsorbed chain ends is so high that they form a 2D crystalline layer. The hexagonal lattice of this layer has the lattice constant of $2a$, where $a$ is a lattice constant of the (111) nickel surface, which is of the order of the van de Waals diameter of the adsorbed beads, i. e. the close-packed surface layer is not frustrated energetically, or, in other words, the surface potential and the surface layer have commensurable lattices.
Different concentrations of the phenol additive
-----------------------------------------------
To quantify our results even further, we have also studied mixtures with different concentrations of the phenol additive. In these mixtures the surface densities of the adsorbed chain ends and phenol molecules are monotonic functions of the total number of phenol molecules in the system. Hence, they can be varied with a good precision, which helps to analyze the systems in a more systematic way.
The systems were prepared in a similar manner: the number of the phenol molecules was adjusted to provide different percentages of the total weight of the system, in the range from $1\%$ to $20\%$. Because of the finite size of the simulation box and phenol adsorption, the concentration of the phenol molecules in the bulk is always smaller than the one used for the system preparation. We therefore used this number (from $1\%$ to $20\%$) only to label a particular system. The important quantities are of course the surface density of the adsorbed chain ends and phenol molecules.
In addition to the velocity profiles we have calculated the two-dimensional radial distribution function of the adsorbed molecules $$g(r) =
\frac{2}{n^2} \frac{L_x L_y}{S(r)}
\left<
\sum_{i=0}^{n-1} \sum_{j = i+1}^{n} p_{ij}(r)
\right>$$ where $r_{ij} = \sqrt{(x_i-x_j)^2 + (y_i-y_j)^2}$, $S(r) =
\pi[(r+\Delta r)^2 - \pi r^2]$ is the bin area, $\Delta r =
r_{max}/N_{bins}$ is the bin width, $p_{ij}(r) = 1$ if $r<r_{ij}<r+\Delta r$ and zero otherwise. The sum is performed over the adsorbed beads (chain ends and phenol molecules) only, i. e. $z_{i,j}
- z_{wall} < \sigma$; $n$ is the number of the adsorbed beads; $<...>$ denotes the ensemble average.
![ Two-dimensional radial distribution function of the adsorbed chain ends. Inset shows the surface densities (number of adsorbed beads per unit area) of the adsorbed ends as a function of the phenol concentration. []{data-label="fig:rdf"}](fig6.eps){width="8cm"}
To further characterize the quality of the hexagonal packing we have also calculated the orientational order parameter [@strandberg92] $$S_6 = \left<
\frac{
\left|
\sum_{i=0}^{n-1} \sum_{j = i+1}^{n} q_{ij} \exp(i6\phi_{ij})
\right|
}{
\sum_{i=0}^{n} \sum_{j = i+1}^{n} q_{ij}
}
\right>$$ where $\phi_{ij}$ is the angle between the vector $r_{ij}$ and a given axis in the $xy$ plane, $q_{ij} = 1$ if $r_{ij}$ belongs to the first peak of the radial distribution function (in our case between $0.5
\sigma$ and $1.5 \sigma$), and zero otherwise. Note that $S_6 = 1$ in the case of a perfect hexagonal order whereas $S_6 = 0$ indicates the complete lack of such order.
The radial distribution functions are shown in Fig. \[fig:rdf\]. The increase of the total number of the adsorbed molecules results in a gradual solidification of the surface layer. The $4\%$ system ($\rho_s
= 0.775 \sigma^{-2}$) has a two-dimensional solid layer formed at the surface, as it can be seen from the peaks at $2l$ and $\sqrt{3}l$, where $l$ is the position of the first peak. Further increase of the amount of adsorbed molecules merely improves the hexagonal close-packing and, after the $5\%$ system ($\rho_s = 0.851
\sigma^{-2}$) the close packing of the surface is reached, i. e. the radial distribution function does not change anymore. The ratio between the number of the adsorbed chain ends and the adsorbed phenol molecules, however, still decreases, as it can be seen from the inset to Fig. \[fig:rdf\]. In case of the blends, the solidification of the surface layer occurs for the 20:DPC mixture.
The force on the wall and the velocity of the adsorbed layer are shown in Fig. \[fig:v\_surf\], together with the orientational order parameter $S_6$. The orientational order parameter $S_6$ confirms that the solidification of the surface layer occurs for the $4\%$ system, and the saturation for the $5\%$ system, in agreement with the behavior of the radial distribution functions.
From the dependence of the relative velocity of the adsorbed layer on the surface density of the adsorbed molecules one can see that once the epitaxial layer is in a solid state, it follows the wall, i. e. we have the stick boundary conditions for the surface layer. However, before the solidification, the velocity of the adsorbed layer is smaller than the wall velocity and is roughly proportional to the total concentration of the adsorbed chain ends.
Similar to the relative velocity, the friction force on the wall increases with the increase of the total amount of the adsorbed beads. However, it does not reach a constant value at the point of solidification of the surface layer, as the relative velocity does, but starts to decrease, with the decrease rate proportional to the ratio of the number of the adsorbed chain ends to the number of the adsorbed phenol molecules. This is due to the fact that the amount of the [*long*]{} chains adsorbed on the wall still decreases and the surface layer disentangles from the bulk of the melt.
Finally, we have also calculated the slip length, which is defined as the extrapolation length of the linear velocity profile in the bulk to zero and is often taken as a parameter in macroscopic descriptions of a fluid flow. The concentration dependence of the slip length is shown in the inset to Fig. \[fig:v\_surf\]. Two regimes can be clearly seen. First, at a low concentration of the adsorbed chain ends (e. g. in case of a pure melt, $N=20$) the melt slips over the wall. If we increase the surface concentration of the adsorbed beads, the slip length decreases and becomes zero at some concentration, i. e. we effectively have the no-slip boundary conditions for the melt flow. This point, however, does not coincide with the actual locking of the surface layer, which we observe only when it solidifies. The reason can be seen from the shape of the velocity profiles: let us denote the thickness of the adsorbed layer as $\delta$ and the velocity of this layer as $v_s$. Then the bulk velocity can be written as $v = v_s
(L_z-2z)/(L_z-2\delta)$. The slip length $b$, obtained from the condition $v(z = -L_z/2 - b) = v_w$, reads $$b = \left( v_w / v_s -1 \right) L_z/2 - \delta v_w / v_s.
\label{eq:slip}$$ It can be seen that two mechanisms contribute to the total slip length $b$. The first one is due to the apparent slip of the adsorbed layer over the surface. The second, negative, contribution is due to the finite thickness of the adsorbed polymer layer. As we add the additives, the adsorbed layer shrinks, giving rise to a higher slip; the relative velocity of the adsorbed layer, however, drops down much faster. Combined, these two mechanisms result in a zero slip for some intermediate concentration of the adsorbed chain ends and a negative slip for higher surface concentrations.
After the solidification, the adsorbed layer follows the wall, i. e. $v_w = v_s$. However, if we further increase the amount of the phenol additive, the ratio between the number of the adsorbed chain ends of the long chains and phenol molecules will decrease, as well as the thickness of the adsorbed layer. As a result, the (negative) slip length will start to increase again.
![ Normalized (to the wall velocity) velocity of the adsorbed layer $v_s / v_w$ (circles), orientational order parameter $S_6$ (squares), and the force per unit area on the wall (triangles). Inset: slip length, defined as a distance at which the melt velocity extrapolates to zero. []{data-label="fig:v_surf"}](fig7.eps){width="8cm"}
Discussion and conclusions
==========================
Let us first turn to a brief discussion on the possible mechanisms of friction in our system. Recalling that the adsorbed chain ends move along the equipotential lines of the surface potential, we conclude that the role of the adsorbed additives is twofold: apart from screening the interaction of the melt with the wall, they also serve as additional obstacles which block the possible paths for the chain ends of the major component. If we ignore weak entanglement of the chains in the adsorbed layer as well as the motion of the adsorbed chain ends of the minor component over the surface, the situation reduces to a two-dimensional site percolation problem. Each vertex is designated open or closed at random, with probability $p$ to be closed which is proportional to the concentration of the chain ends of the minor component. Percolation theory predicts that for the site percolation on a hexagonal lattice the critical point of the percolation probability is $p_c =
1/2$. [@sykes63] Assuming that the transition happens for the $20:1$ system (see the inset of Fig. \[fig:velocity\]) we obtain $p_c \approx 0.47$. Of course, due to the applied shear we have an oriented or directed percolation: our lattice sites shall be assigned particular orientations, along which the percolative paths are biased.
On the other hand, gradual solidification of the adsorbed layer results in the increase of the friction coefficient, due to the collective motion of the adsorbed molecules: once the surface layer is solidified, the adsorbed molecules cannot move in the direction perpendicular to the shear direction. However, hopping between the hole and bridge sites always involves a motion perpendicular to the shear direction, or, in other words, the hexagonal symmetry of the lattice forbids the motion of the beads along the shear.
Once the surface layer is in a solid state, the relative velocity of the adsorbed layer does not change anymore. However, the ratio between the adsorbed chain ends of long and short molecules still changes. This affects the thickness of the adsorbed layer as well as its entanglement with the rest of the melt. As a result, the shear force on the wall decreases with the increase of the concentration of the additive. Similar scenario has already been discussed in the framework of the mean-field theory. [@brochard1992; @brochardwyart1994; @ajdari1994; @smith2005]
To summarize, we studied the effective boundary conditions for a polymer blend adsorbed on a structured surface. The slip boundary condition observed for a monodispersed melt changes to the no-slip at some concentration of the additive. Further increase of the concentration of the additive at the surface leads to the solidification and locking of the motion of the adsorbed surface layer.
Finally, we would like to comment on the importance of the multiscale modeling methods we employ in our studies. As underlined in the introduction, we consider a particular system, which is of high relevance to different fields of modern technology. Our results, despite the particular system considered, are the direct expression of the interplay between specific (electronic based) molecule-surface interactions and global statistical and dynamical properties of the system, i. e. the interplay between the different scales is the crucial ingredient of the description. The electronic and molecular resolution, implicit into the parameterization of the model, allows for a level of analysis which is beyond any other existing coarse-grained models or mean-field approaches. In this sense, this study goes beyond the specificity of the system considered and calls for extensions to other systems, and experimental tests. The present approach is a route which allows to detect important, otherwise not accessible, properties and sets a new link between theory, experiments and technology.
This work was supported by the Alexander von Humboldt foundation, Germany (X.Z.) and by the BMBF, under Grant No. 03N6015, and the Bayer Corporation. The advise of Vagelis Harmandaris is acknowledged.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We theoretically investigate strong coupling between a single molecule and a single metallic nanoparticle. A theory suited for the quantum-mechanical description of surface plasmon polaritons (SPPs) is developed. The coupling between these SPPs and a single molecule, and the modified molecular dynamics in presence of the nanoparticle are described within a combined Drude and boundary-element-method approach. Our results show that strong coupling is possible for single molecules and metallic nanoparticles, and can be observed in fluorescence spectroscopy through the splitting of emission peaks.'
author:
- Andreas Trügler
- Ulrich Hohenester
date: 'February 12, 2008'
title: Strong coupling between a metallic nanoparticle and a single molecule
---
Introduction
============
Quantum optics has recently made its way to the field of plasmonics. [@chang:06; @akimov:07] This is due to the the rapid progress in nanofabrication and measurement techniques. Recent experiments have demonstrated the controlled coupling of single molecules with metallic nanoparticles (MNPs) [@anger:06; @kuehn:06; @gerhardt:07; @akimov:07] and metallic surfaces, [@labeau:07] of coupled nanoparticles, [@rechberger:03; @danckwerts:07] and of donor and acceptor molecules accross metal films. [@andrew:04] Possible applications of such hybrid molecule-MNP systems range from biosensing [@stuart:05; @yin:05] to active plasmonic devices. [@fedutik:07]
A key element of the quantum-optics toolbox is the strong coupling between a quantum emitter and a resonator, where excitation energy is coherently transferred between emitter and resonator. Strong coupling was first observed for single atoms in high-finesse optical resonators, [@turchette:95; @raimond:01] and more recently for various solid state systems, such as semiconductor quantum dots [@hennessy:07; @press:07] or superconductor circuits. [@wallraff:04] Although strong coupling between ensembles of molecules, e.g, $J$-aggregates of dyes, with plasmons has been reported, [@specialissue.organic] it is unclear whether the strong coupling regime can be reached for single molecules coupled to MNPs. The reason for this lies in the intricate interplay of the molecule-MNP coupling strength with the molecular relaxation dynamics, which becomes heavily altered in the vicinity of the nanoparticle.
It is the purpose of this paper to theoretically investigate the strong coupling regime between a single quantum emitter, such as a molecule or collodial quantum dot, and a single MNP. We start by developing a theory suited for the quantum-mechanical description of surface plasmon polaritons (SPPs), the coupling between these SPPs and single molecules, and the modified molecular dynamics in presence of the MNP. We employ a Drude framework for the description of the metal dynamics, and compute the quantized SPP modes within a boundary element method approach. [@hohenester.prb:05; @gerber.prb:07] Our results show that strong coupling is possible for molecules and MNPs and could be observed in fluorescence spectroscopy through the splitting of emission peaks.
We have organized this paper as follows. In section \[sec:theory\] we show how to compute surface plasmon modes within a boundary-element-method approach, and introduce a suitable quantization scheme for the surface plasmons. We also present details of the theoretical description of the coupled molecule-MNP system in presence of scatterings. Sec. \[sec:results\] presents results of our model calculations. We explore the strong-coupling regime for a single molecule coupled to a MNP, and identify the pertinent parameters for strong coupling. We also discuss limitations of our model. Finally, in Sec. \[sec:conclusions\] we summarize and draw some conclusions.
Theory {#sec:theory}
======
Plasmon quantization
--------------------
Although SPPs are generally considered as bosonic quasiparticles, most theoretical work does not explicitly rely on such description. In linear response one can employ the fluctuation-dissipation theorem to relate the dielectric response to the dyadic Green tensor of Maxwell’s theory, [@vogel:06; @hohenester.ieee:08] where all details of the metal dynamics are embodied in the dielectric function, which can be obtained from either experiment [@johnson:72] or first principles calculations. This approach is no longer applicable in nonlinear response. Also the neglect of plasmon relaxation at small timescales, as proposed in other work [@bergman:03], is not suited for the investigation of strong coupling, which critically depends on the relative importance of coupling and SPP dephasing. In this work we thus follow the seminal work of Ritchie, [@ritchie:57] where the electron dynamics in the metal is described within the hydrodynamic model. [@barton:79] For the transition metals Ag and Au, electrons with particle density $n_0$ are assumed to move freely in a medium with background dielectric constant $\epsilon_0$, which accounts for the screening of $d$-band electrons. [@ladstaedter.prb:04; @remark.drude]
![(Color online) (a) Spectrum of cigar-shaped Ag MNP. The height of the particle is 40 nm and the height-to-diameter ratio is $5:1$. The solid and dashed lines show the calculated spectra for the dielectric function of Ref. and the Drude form, [@remark.drude] respectively. The triangles at the bottom indicate the energies of the plasmon modes (d) and (e), and the arrow at $\sim 1.7$ eV indicates the energy of the molecular state $1$. (b) Non-radiative (solid line) and radiative (dashed-dotted line) decay rate for a molecule located at a certain distance from the MNP \[see panel (c)\] in units of the radiative free-space decay rate $\gamma_r^0$. The inset reports the level scheme used in our caluclations. (c) Discretized particle surface as used in our calculations. (d–f) Surface charge distribution of SPP eigenmodes of lowest energy. Only mode (d) has a nonzero dipole moment. ](figure1){width="0.8\columnwidth"}
![(Color online) Simulation for molecule which is initially in state $1$ \[inset of Fig. 1(b)\]. The solid and dashed lines show the population of the molecular state $1$ and the SPP dipole mode \[Fig. 1(d)\], respectively, for two different molecule–MNP distances. We use a molecule dipole moment of 10 atomic units, which corresponds to a free-space decay time $1/\gamma_r^0$ of approximately one nanosecond. In weak coupling (corresponding to the 8 nm distance) the non-radiative decay rate can be also estimated from time-dependent perturbation theory, $\gamma_{nr}\approx |g_\lambda|^2/\gamma_0$, where $g_\lambda$ is the coupling constant between the molecule and the resonant plasmon mode. The dotted line in the figure shows the resulting decay. In our simulations we include the 40 plasmon modes of lowest energy. ](figure2){width="0.8\columnwidth"}
The energy of a classical electron plasma is the sum of kinetic and electrostatic energy [@ritchie:57; @barton:79] $$\label{eq:energyplasma}
H=\mbox{$\frac 12$}\int\left\{n_0(\nabla\Psi)^2+\rho\Phi\right\}d^3r\,.$$ Here $\rho(\bm r)$ is the charge density displacement from equilibrium, $\Phi(\bm r)$ is the electrostatic potential induced by $\rho(\bm r)$, and $\Psi(\bm r)$ is the velocity potential, whose derivative gives the velocity density $\bm v=-\nabla\Psi$. [@barton:79] Throughout we use Gauss and atomic units ($e=m=\hbar=1$). For the SPPs of our present concern we consider surface charge distributions $\sigma$ which are nonzero only at the surface of the MNP. As detailed in Appendix \[app:supplementary\], the Hamilton function can be rewritten in a boundary element method (BEM) approach [@hohenester.prb:05; @garcia:02] as $$\begin{aligned}
\label{eq:energyspp}
H&=&\frac 1{2n_0}\Bigl\{\dot\sigma^T\left(2\pi+F\right)^{-1}G\,\dot\sigma
\nonumber\\&&\quad
+\omega_p^2\,\sigma^T
\left[2\pi(\epsilon_0+\epsilon_b)+(\epsilon_0-\epsilon_b)F\right]^{-1}G\,
\sigma\Bigr\}\quad\,.\end{aligned}$$ Here $\sigma$ is the vector of the surface charges within the discretized surface elements (see inset of Fig. 1), $G$ is the free Green function matrix which connects two surface elements, $F$ is the corresponding surface derivative, [@hohenester.prb:05; @garcia:02] $\omega_p=(4\pi n_0)^{\frac 12}$ is the plasma frequency, $\epsilon_0$ is the background dielectric constant of the metal [@remark.drude], and $\epsilon_b$ the dielectric constant of the embedding medium. We can now determine the eigenmodes of Eq. and quantize the plasma oscillations via a canonical transformation, following the standard procedure outlined in Refs. . Within such an approach we obtain the plasmon Hamiltonian $H_{\rm pl}=\sum_\lambda \omega_\lambda\,a_\lambda^\dagger a_\lambda^{\phantom\dagger}$ in second-quantized form, with $\omega_\lambda$ being the energy and $a_\lambda^\dagger$ the creation operator for the plasmon mode $\lambda$. The field operator for the SPPs is of the form $$\label{eq:quantspp}
\sigma(\bm r)=\sum_\lambda \left(\frac{2n_0}{\omega_\lambda\beta_\lambda}\right)^{\frac 12}
u_\lambda(\bm r)\left(a_\lambda^{\phantom\dagger}+a_\lambda^\dagger\right)\,,$$ where $u_\lambda(\bm r)$ is the plasmon eigenfunction and $\beta_\lambda$ the corresponding normalization constant.
{width="1.5\columnwidth"}
Molecule–MNP coupling
---------------------
With the SPP quantization we have now opened the quantum optics toolbox. This allows us to study strong coupling according to the standard prescription [@walls:95]. As for the description of the molecule we follow Refs. who considered a generic few-level system. This approach is also best suited for other quantum emitters, such as collodial quantum dots. The inset of Fig. 1(b) shows the level scheme used in our calculations. It consists of the molecule ground state $0$ and two excited states $1$ and $2$. We assume that an external pump laser brings the molecule into the excited state $2$, where it decays non-radiativly with a given rate $\gamma_m$ to the optically active state $1$. This indirect process allows us to separate the excitation dynamics, which is not modified in presence of the MNP, from the relaxation dynamics of state $1$, which becomes strongly modified if the molecule and SPP are in resonance. The coherent part of the molecule-MNP dynamics is described by the Hamiltonian $$\label{eq:ham}
H=H_{\rm mol}+H_{\rm pl}+H_{\rm mol-pl}+H_{\rm pump}\,.$$ Here $H_{\rm mol}$ and $H_{\rm pl}$ describe the molecular states and the SPP modes, respectively, $H_{\rm mol-pl}$ is the coupling between the molecular dipole and the surface charge , and $H_{\rm pump}$ is the interaction of the molecule with the pump laser. The last two terms in Eq. are described within the usual rotating-wave approximation. [@walls:95] In addition, we account for the incoherent part of the dynamics through a master equation of Lindblad form [@walls:95; @xu:04] $$\label{eq:master}
\dot\rho=-i[H,\rho]-\mbox{$\frac 12$}\sum_i\left(L_i^\dagger L_i^{\phantom\dagger}\rho+
L_i^\dagger L_i^{\phantom\dagger}\rho-2L_i^{\phantom\dagger}\rho L_i^\dagger\right)\,,$$ where $\rho$ is the density matrix of the coupled molecule-SPP system. The Lindblad operators $L_i$ describe the various scattering channels of molecular decay, plasmon decay through Landau damping, and radiative decay [@remark.drude].
Results {#sec:results}
=======
In our calculations we consider the cigar-shaped Ag MNP shown in Fig. 1. Other MNP shapes and metals will be discussed at the end. Figure 1(a) shows the spectra computed within our BEM approach [@gerber.prb:07] for the Ag dielectric function of Ref. (solid line) and the Drude framework [@remark.drude] (dashed line). Both spectra are in nice agreement, thus justifying the use of the Drude model. The energies of the SPP eigenmodes are indicated by triangles. Panel (b) reports the non-radiative and radiative decay rates of the molecule as a function of molecule-MNP distance, which we compute in the weak-coupling regime according to the prescription of Refs. . One observes that the rates dramatically increase when the molecule approaches the MNP. Here the decay process becomes strongly altered by the nanoparticle, which acts as a supplemental antenna and converts part of the molecule’s near field into radiation and Ohmic dissipation.
We next turn to the results of our master-equation approach. Figure 2 shows simulations based on the solution of Eq. where the molecule is initially brought into the excited state $1$. Let us first consider the larger molecule-MNP distance of 8 nm (upper two lines). Through the coupling $H_{\rm mol-pl}$, the lowest SPP mode becomes populated and subsequently decays through Landau damping and radiation. After a transient at early times, both molecule and SPP population decay mono-exponentially with the same decay constant. In this regime the molecule drives the strongly damped plasmon mode and hereby constantly transfers energy to the MNP. Things change considerably when the molecule is brought closer to the MNP. For $z=4$ nm (lower two lines) one observes a pronounced population beating between the molecule and the surface plasmon, superimposed on a sub-picosecond decay due to efficient plasmon damping. This beating behavior is a clear signature of strong coupling [@walls:95] which occurs in a regime where the molecule-MNP coupling is stronger than the plasmon damping.
Although strong coupling is most apparent in the time domain, spectroscopy appears to be a more suitable tool for its experimental observation. We next turn to the study of the setup shown in the inset of Fig. 1(b), where a weak pump laser brings the molecule to the excited state $2$. This process is followed by an internal decay into the optically active state $1$ and a final relaxation to the groundstate. Again, the last process is strongly modified in presence of the MNP. In our calculations we use the master equation to compute the steady state solution. Once a stationary condition is reached, we can compute the fluorescence spectra from the Wiener-Khinchin theorem by means of the quantum regression theorem. [@walls:95; @xu:04] Figure 3 shows results of our simulations for three different molecular dipole moments. For the smallest dipole moment of panel (a), one observes that the line broadens when the molecule is brought closer to the MNP. We verified that the line broadening is precisely given by the sum of radiative $\gamma_r$ and non-radiative $\gamma_{nr}$ decay rates shown in Fig. 1(b). For the larger dipole moments investigated in panels (b,c), we observe that at a distance of a few nanometers the line splits, thus indicating the onset of strong coupling. Here excitation energy is coherently transfered between the molecule and the SPP.
In a generic model, where a quantum emitter is coupled to a single cavity mode, the polariton eigenmodes $\Omega_\pm$ of the coupled system are of the form [@andreani:99] $$\label{eq:strongcoupling}
\Omega_\pm+=\omega_0-
\frac i4(\gamma_c+\gamma_m)\pm
\sqrt{g^2-\left(\frac{\gamma_c-\gamma_m}4\right)^2}\,.$$ Here, $\omega_0$ is the enery of the isolated molecule and cavity, which are assumed to be in resonance, $\gamma_m$ and $\gamma_c$ are the decay rates of the molecule and cavity, respectively, and $g$ is the coupling constant. Strong coupling occurs for $g>|\gamma_c-\gamma_m|/4$ and corresponds to the formation of a dressed state with finite lifetime. It is an intrinsic property of the coupling between the molecule and the cavity, and manifests itself as a doublet splitting of the emission lines. Quite generally, for the coupled molecule–MNP system Eq. is too simple, because the molecule couples not only to the MNP dipole mode but also to all other modes, and one must use a more refined description as we have done in this work. Nevertheless, Eq. allows us to estimate the pertinent parameters for strong coupling. From the plasmon decay rate $\gamma_0\sim 30$ fs for silver we can estimate a critical coupling strength of $g\approx\gamma_0/4\sim 5$ meV for the onset of strong coupling. Indeed, this value is in agreement with the results of Fig. 3 \[as can be inferred, e.g., from the line broadening in panel (b) at the distance of 3 nm where the emission line starts to split\]. For a coupling of the order of a few meV the approximation of a two-level system is justified for both molecules and quantum dots, although the true lineshape might be additionally influenced by internal degrees of freedom (e.g., vibrations) of the quantum emitter.
![(Color online) Line positions and oscillator strengths of polariton modes for cigar-shaped (two lines on left-hand side) and disk-shaped (two lines on right-hand side) nanoparticles and Ag and Au. In Au the lines split at smaller distances because of the stronger plasmon damping. [@remark.drude] We use $d=10$ atomic units.](figure4){width="\columnwidth"}
When the molecule is brought even closer to the MNP, the oscillator strength of the high-energy line vanishes and the low-energy line becomes strongly red-shifted. In this regime, where the energy renormalization is of the order of several tens of meV, the description of the quantum emitter in terms of a generic few-level system is expected to break down. The strong redshift of the emission peaks is due to the attractive interaction between the molecule and the MNP, which is strongly enhanced at small distances, and the different oscillator strengths are associated to the different dipole moments of the predominantly MNP- and molecule-like polariton modes at lower and higher energy, respectively. We also found that moderate detunings between the molecule and SPP energies do not drastically change the behavior shown in the figures. Figure 4 shows that similar behavior is found for other MNP shapes and materials. We have also performed calculations for spherical nanoparticles. Unfortunately, for both silver and gold the resulting surface plasmons have energies in a spectral region where $d$-band scatterings set in, [@ladstaedter.prb:04] and where the Drude description becomes questionable. Our results (not shown) indicate that for nanospheres strong coupling occurs at smaller distances than for the particle shapes shown in Fig. 4, which might be due to the larger number of plasmon modes to which the molecule can couple. [@hohenester.ieee:08]
Summary and Conclusions {#sec:conclusions}
=======================
In conclusion, we have studied strong coupling between a single molecule and a metallic nanoparticle within a fully quantum-mechanical approach. We have demonstrated that strong coupling is possible for realistic molecule and nanoparticle parameters, despite the strong plasmon damping, and should be observable in fluorescence spectroscopy through the splitting of emission peaks. Strong coupling is an important ingredient for future plasmonic-based quantum information schemes, and might play a significant role in biosensor applications.
We gratefully acknowledge most helpful discussions with Joachim Krenn and Alfred Leitner.
{#app:supplementary}
In this appenix we derive Eq. and show how to quantize the SPP modes. Our starting point is the energy of a classical plasma, Eq. . Let us first consider the first term on the right-hand side which describes the kinetic energy. From the relation $\bm v=-\nabla\Psi$ between the velocity field $\bm v$ and $\Psi$, we obtain for the continuity equation $$\label{eq:continuity}
\partial_t n=-n_0\nabla\bm v=n_0\nabla^2\Psi\,,$$ which gives us the relation between the density displacement $n$ and the velocity potential $\Psi$. In the following we consider surface charge distributions $\sigma$ which are nonzero only at the surface of the MNP. Integration of the continuity equation over a small cylinder $\Omega$ (height $h\to 0$ and base $\delta S$), which encloses a small surface element, then gives for the right-hand side of Eq. $$\label{eq:intcontinuity}
\int_\Omega n_0\nabla^2\Psi\,dV=\int_{\partial\Omega} n_0\,\hat{\bm n}\cdot\nabla\Psi\,dS
\cong n_0\,\frac{\partial\Psi}{\partial\hat n}\,\delta S\,.$$ Here, $\frac{\partial\Psi}{\partial\hat n}=\hat{\bm n}\cdot\nabla\Psi$ denotes the surface derivative of the velocity potential. Together with the left-hand side of Eq. we find the link between $\Psi$ and $\sigma$, $$\label{eq:sigmapsi}
n_0\,\frac{\partial\Psi}{\partial\hat n}=\dot\sigma\,.$$ Using Green’s first identity we can rewrite the term for the kinetic energy in Eq. as $$\label{eq:S5}
\int_\Omega(\nabla\Psi)^2\,dV=
\int_{\partial\Omega}\Psi\frac{\partial\Psi}{\partial\hat n}\,dS
-\int_\Omega \Psi\nabla^2\Psi\,dV\,.$$ As evident from the continuity equation , for a pure surface charge distribution $\nabla^2\Psi$ is zero inside the metallic nanoparticle, and the second term on the right-hand side of thus vanishes. We can now use the boundary-element method [@garcia:02; @hohenester.prb:05; @hohenester.ieee:08] to relate $\Psi$ to $\frac{\partial\Psi}{\partial\hat n}$. Our starting point is $$\label{eq:S6}
\Psi(\bm r)=\int_{\partial\Omega}\left(
G(\bm r,\bm s')\frac{\partial\Psi(\bm s')}{\partial\hat n}-
\frac{\partial G(\bm r,\bm s')}{\partial\hat n}\Psi(\bm s')\right)\,\frac{dS'}{4\pi}\,,$$ where $G(\bm r,\bm r')=1/|\bm r-\bm r'|$ is the free-space Green function. Performing the limit $\bm r\to\bm s$ in Eq. according to the prescription given in Refs. and using the same notation as in these references, we obtain $$\label{eq:S7}
2\pi\Psi=G\Psi'-F\Psi\,.$$ Here $\Psi'$ and $F$ are the surface derivatives of the velocity potential $\Psi$ and the Green function $G$, respectively, and $G$, $F$ and $\Psi$, $\Psi'$ are assumed to be convoluted in space.
At this point it is convenient to switch to the boundary elements of the discretized MNP surface (see also inset of Fig. 1): $\Psi$ and $\sigma$ are vectors of the length of the number of surface elements, and $G$ and $F$ are matrices connecting the different surface elements. We can thus solve Eq. through inversion $\Psi=(2\pi+F)^{-1}G\,\Psi'$. Together with the relation , the term for the kinetic energy can then be brought into the final form $$\label{eq:S8}
\mbox{$\frac 12$}n_0\int_{\partial\Omega}\Psi\frac{\partial\Psi}{\partial\hat n}\,dS
\longrightarrow \mbox{$\frac 12$}n_0\,\dot\sigma^T(2\pi+F)^{-1}G\,\dot\sigma\,.$$ Here $\sigma^T$ denotes the transposed surface charge vector.
For the potential energy of Eq. we follow the procedure given in Ref. . We start from a relation similar to Eq. but the velocity potential $\Psi$ replaced by the electrostatic potential $\Phi$. Taking the surface derivative inside and outside the MNP, we obtain $$\begin{aligned}
(2\pi+F)\Phi&=&\phantom{-}G\,\Phi_1'\\
(2\pi-F)\Phi&=& -G\,\Phi_2'\,.\end{aligned}$$ Here $\Phi_1'$ and $\Phi_2'$ denote the surface derivatives of the potential inside and outside the MNP. Multiplying the first equation with the dielectric constant $\epsilon_0$ of the metal and the second one with the dielectric constant $\epsilon_b$ of the embedding medium, gives after substraction $$\label{eq:S9}
\bigl(2\pi(\epsilon_0+\epsilon_b)+(\epsilon_0-\epsilon_b)F\bigr)\,\Phi=
G(\epsilon_0\Phi_1'-\epsilon_b\Phi_2')=4\pi\,G\sigma\,.$$ To arrive at the last term of we have used the boundary condition $\hat{\bm n}\cdot(\bm D_2-\bm D_1)=-\epsilon_b \Phi_2'+\epsilon_0\Phi_1'=4\pi\sigma$ of Maxwell’s equation. Putting all above results together, we finally get Eq. .
To obtain the eigenmodes we first rewrite Eq. in the short-hand notation $$\label{eq:S11}
H=\frac 1{2n_0}\left(\dot\sigma^T B\,\dot\sigma+\sigma^T A\,\sigma\right)\,,$$ where the explicit form of the matrices $A$ and $B$ can be inferred from Eq. . The matrices $A$ and $B$ are symmetric and thus can be diagonalized simultaneously. Let $\omega_\lambda^2$ and $u_\lambda$ denote the eigenvalues and eigenvectors of the generalized eigenvalue problem $$\label{eq:S12}
A\,u_\lambda=\omega_\lambda^2\,B\,u_\lambda\,.$$ The eigenvectors $u_\lambda$ can be chosen real, and are orthogonal in the sense $$\label{eq:S13}
u_\lambda^T\,B\,u_{\lambda'}=\beta_\lambda\,\delta_{\lambda\lambda'}\,.$$ We can thus expand the surface charge distribution in terms of these eigenfunctions viz. $$\label{eq:S14}
\sigma=\sum_\lambda \gamma_\lambda e^{i\omega_\lambda t}a_\lambda\,u_\lambda\,.$$ Here $\gamma_\lambda=\sqrt{{2n_0}/({\omega_\lambda\beta_\lambda})}$, and $a_\lambda$ is an expansion coefficient for the eigenmode $\lambda$. Inserting this expression into equation and performing the standard quantization procedure via a canonical transformation, [@ritchie:57; @barton:79; @arista:01] then brings us to the plasmon Hamiltonian in second-quantized form $$\label{eq:S15}
H_{\rm pl}=\sum_\lambda\omega_\lambda\,a_\lambda^\dagger a_\lambda^{\phantom\dagger}\,,$$ with $a_\lambda^\dagger$ being the creation operator for the plasmon mode $\lambda$.
[35]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , , , ****, ().
, , , , , , , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , , , , , ****, ().
, , , , , ****, ().
, , , , , ****, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, ****, ().
, , , , , ****, ().
, , , , , ****, ().
, , , ****, ().
, , , , , , , , , ****, ().
, , , , , , , , ****, ().
, , , , , , , , , ****, ().
, ****, ().
, , , , , , ****, ().
, ** (, ).
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , ****, ().
, ****, ().
, ** (, , ).
, , , ****, ().
, , , , , ****, ().
, , , , , , ****, ().
, , , ****, ().
, ****, ().
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We study a token-based central queue with multiple customer types. Customers of each type arrive according to a Poisson process and have an associated set of compatible tokens. Customers may only receive service when they have claimed a compatible token. If upon arrival, more than one compatible token is available, an *assignment rule* determines which token will be claimed. The *service rate* obtained by a customer is state-dependent, i.e., it depends on the set of claimed tokens and on the number of customers in the system. Our first main result shows that, provided the *assignment rule* and the *service rates* satisfy certain conditions, the steady-state distribution has a product form. We show that our model subsumes known families of models that have product-form steady-state distributions including the order-independent queue of [@Krzesinski11] and the model of [@Visschers12]. Our second main contribution involves the derivation of expressions for relevant performance measures such as the sojourn time and the number of customers present in the system. We apply our framework to relevant models, including an M/M/K queue with heterogeneous service rates, the MSCCC queue, multi-server models with redundancy and matching models. For some of these models, we present expressions for performance measures that have not been derived before.\
Keywords: product form, token-based, order-independent queue, redundancy system, matching model
author:
- 'U. Ayesta'
- 'T. Bodas'
- 'J.L. Dorsman[^1]'
- 'I.M. Verloop'
title: 'A token-based central queue with order-independent service rates'
---
[^1]: Corresponding author: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study holographic dual descriptions of very special conformal field theories with the T(2) symmetry. After constructing solutions in effective five dimensional Einstein gravity coupled with massive two-form fields, we uplift them to the ten dimensional type IIB supergravity, via a consistent truncation ansatz, to derive new analytical solutions in string theory. From the Kaluza-Klein ansatz in terms of the internal Sasaki-Einstein space, we obtain their field theory interpretation with concrete realizations in a large class of holographic $\mathcal{N}=1$ supersymmetric conformal field theories. Null compactification of these theories yields holographic dual descriptions of non-relativistic critical systems with translational invariance but without rotational invariance such as the ones induced from a constant electromagnetic field.'
---
Gravity Dual for Very Special Conformal Field Theories
in type IIB Supergravity
\
[Yu Nakayama]{}
Department of Physics, Rikkyo University, Toshima, Tokyo 171-8501, Japan
Introduction
============
In high energy physics, the Poincaré symmetry is often regarded as a fundamental assumption, but it is theoretically possible that it is merely an accidental low energy symmetry. One way to break the Lorentz symmetry while preserving the translational symmetry is to choose a particular direction in space-time. Certainly in our universe, there is a special “time direction" which emerges from the evolution of the universe, resulting in the breaking of the Lorentz symmetry down to a spatial rotation. Mathematically, however, it is more interesting to choose a particular null direction to break the Lorentz symmetry, leading to more non-trivial residual symmetries than the mere spatial rotation. This was suggested by Cohen and Glashow sometime ago [@Cohen:2006ky][@Cohen:2006ir] and dubbed “very special relativity".
To address the question of the emergent Lorentz symmetry, we recall that our modern understanding of effective field theories is based on the idea of renormalization group, and it is instructive to see how quantum field theories with the very special relativity behaves under the renormalization group flow. In order to study the renormalization group flow, it is crucial to understand the structure of renormalization group fixed points with the very special relativity. In [@Nakayama:2017eof], we have classified all the conformal extension of the very special relativity as possible candidates of the renormalization group fixed points that are consistent with the very special relativity. Such theories were named “very special conformal field theories".
A remarkable feature of the construction of very special conformal field theories is that if we start with the four-dimensional Poincaré invariant conformal field theories, the deformation that preserves the very special conformal symmetry has the Poincaré conformal dimension of five, which means (i) it is an irrelevant deformation from the point of view of the Poincaré conformal scaling and (ii) it may appear to be “non-renormalizable" from the naive power-counting. Therefore, while it is easy to write down the effective action that preserves the very special conformal symmetry at the tree level, the study of their quantum nature is difficult by design.
In this paper, we offer an alternative view on the strongly interacting very special conformal field theories possibly realized at the ultraviolet fixed point by using the holography. The holographic dual or gravity dual descriptions of strongly coupled conformal field theories are successful not only with Poincaré invariance but also without Poincaré invariance. In our previous paper [@Nakayama:2017eof], we have found that one of such examples, i.e. holographic dual description of the Schrödinger invariant conformal field theories [@Nishida:2007pj][@Son:2008ye][@Balasubramanian:2008dm][@Adams:2008wt][@Maldacena:2008wh][@Herzog:2008wg], has many common features with the holographic dual description of a particular kind of very special conformal field theories with the E(2) symmetry. See also earlier works [@Bergman:2000cw][@Alishahiha:2003ru] on the holographic descriptions of these theories with the same symmetry. The main focus of this paper is to study the least symmetric case of the T(2) invariant very special conformal field theories and their gravity duals which have not been addressed before. As we will see, null compactification of these theories yields holographic dual descriptions of non-relativistic critical systems with translational invariance but without rotational invariance and may have potential applications in condensed matter physics.
The organization of the paper is follows. In section 2, we review the basic facts about the very special conformal field theories. In section 3, we construct the holographic duals in the effective five dimensional Einstein gravity coupled with two-form matter fields. In section 4, we uplift the solutions to ten dimensional type IIB supergravity and obtain new analytical solutions. In section 5, we give the field theory interpretation of our holographic dual solutions. In section 6, we conclude with further discussions.
Very Special Conformal Field Theories
=====================================
Very special conformal symmetry is a subgroup of the Poincaré conformal symmetry of $SO(4,2)$ that preserves a particular null direction and include at least one special conformal transformation. For our purpose, let us choose this null direction to be $x^-$, where $x^{\pm} = \frac{1}{\sqrt{2}}(x^0\pm x^1)$ with the metric convention $ds^2 = -2dx^+dx^- + (dx^i)^2$. Here $i=2,3$.
Based on the classification without the conformal symmetry [@Cohen:2006ky][@Cohen:2006ir], in [@Nakayama:2017eof] we argued that there are four interesting classes of very special conformal symmetry. The maximal case is SIM(2) very special conformal algebra spanned by $\{P_+,P_-,P_i,J_{+i},J_{+-},J_{ij},\tilde{D},K_+\}$, whose commutation relation can be inferred from the original Poincaré conformal group with the conventional notation of $P_\mu$ being translation, $J_{\mu\nu}$ being Lorentz transformation, $K_\mu$ being special conformal transformation, and $D$ being dilatation (while $\tilde{D} = D + J_{+-})$. If we abandon $J_{ij}$, we obtain HOM(2) very special conformal algebra, and if we abandon $J_{+-}$, we obtain E(2) very special conformal algebra. Finally, if we abandon $J_{+-}$ and $J_{ij}$ we obtain T(2) very special conformal algebra, which will be the main focus of this paper. Note that the names only refer to the commutation relations of the Lorentz part of the algebra. For completeness the commutation relations of T(2) very special algebra are summarized in Table 1.
$P_+$ $P_-$ $P_i$ $J_{+i}$ $K_+$ $\tilde{D}$
------------- ------- ------------- ----------- ---------- -------------- -------------
$P_+$ $0$ $0$ $0$ $0$ $0$ $0$
$P_-$ $0$ $0$ $0$ $P_i$ $-\tilde{D}$ $2P_-$
$P_i$ $0$ $0$ $0$ $P_+$ $J_{+i}$ $P_i$
$J_{+i}$ $0$ $-P_i$ $-P_+$ $0$ $0$ $-J_{+i}$
$K_+$ $0$ $\tilde{D}$ $-J_{+i}$ $0$ $0$ $-2K_+$
$\tilde{D}$ $0$ $-2P_-$ $-P_i$ $J_{+i}$ $2K_+$ $0$
: The commutation relation of the T(2) very special conformal algebra.
\[tab1\]
Very special conformal field theories possess a local energy-momentum tensor that satisfies the following conditions $$\begin{aligned}
\partial_+T_{+}^{\ +} + \partial_-T_{+}^{\ -} + \partial_i T_+^{\ i} &= 0 \cr
\partial_+T_{-}^{\ -} + \partial_- T_{-}^{\ -} + \partial_i T_-^{\ i} &= 0 \cr
\partial_+{T_j}^{\ +} + \partial_-T_{j}^{\ -} + \partial_i T_{j}^{\ i} &= 0 \cr
T_{i}^{\ -} + T_{+}^{\ i} & = 0 \cr
2T_{-}^{\ -} + T_{i}^{\ i} &=0 \ .\end{aligned}$$ so that one can construct the local current associated with the dilatation $\tilde{D}$ and the special conformal transformation $K_+$.
Holographic realizations
========================
The main purpose of this paper is to construct holographic descriptions of the T(2) very special conformal field theories while the corresponding holographic models for E(2) very special conformal field theories were constructed in [@Nakayama:2017eof].[^1] In the field theory side, one way to construct very special conformal field theories with T(2) symmetry is to start with the Poincaré invariant conformal field theories and deform them by using the dimension five primary operator in the anti-symmetric tensor representation of the Poincaré group. $$\begin{aligned}
S = S_0 + \int d^4x O^{-x} ,\end{aligned}$$ where $S_0$ is the action with Poincaré conformal invariance, and $O^{-x} = -O^{x-}$ transforms as an anti-symmetry tensor and it has $\Delta = 5$ under the Poincaré dilatation $D$.
To mimic the construction in the gravity side, we begin with the five dimensional Einstein-Hilbert action with negative cosmological constant, which supports the AdS space, coupled with a massive two-form tensor field, which is dual to the dimension five anti-symmetric tensor operator. The explicit action is $$\begin{aligned}
S_B = \int d^5x \sqrt{g} \left(\frac{1}{2}R - \Lambda - \frac{1}{6} H_{MNL} H^{MNL} - \frac{m^2}{2} B_{MN} B^{MN} \right) \label{effective}\end{aligned}$$ where $B_{MN} = -B_{NM}$ and $H_{MNL} = \partial_M B_{NL} + \partial_N B_{LM} + \partial_L B_{MN}$. We will set $\Lambda = -6$ and $m^2 = 9$, where we recall that with this cosmological constant the mass and the Poincaré conformal dimension $\Delta$ of the two-form field are related as $m^2 = (\Delta-2)^2$.
We find the solution of the equations motion that have the symmetry corresponding to the $T(2)$ very special conformal symmetry. The metric is given by $$\begin{aligned}
ds^2 = -c \frac{dx^{-2}}{z^4} + \frac{dz^2-2dx^+dx^{-} + (dx^2)^2 + (dx^3)^2}{z^2} \ ,\end{aligned}$$ where it is invariant under the dilatation $\tilde{D}$ $$\begin{aligned}
x^- \to \lambda^2 x^- \ , \ \ x^i \to \lambda x^i \ , \ \ z \to \lambda z \ , \ \ x^+ \to x^+ \end{aligned}$$ as well as the very special conformal transformation $K_+$ $$\begin{aligned}
x^- \to \frac{x^-}{1+ax^{-}} \ , \ \ x^i \to \frac{x^i}{1+a x^-} \ , \cr
z \to \frac{z}{1+a x^{-}} \ , \ \ x^+ \to x^+ - \frac{a}{2}\frac{x_i^2 + z^2}{1+ax^{-}} \ . \end{aligned}$$ Note that this metric has the larger symmetry than the T(2) symmetry because the rotation in $ij$ directions is not broken: it is invariant under the E(2) symmetry. One can argue that whenever one wants to preserve the T(2) symmetry in the five dimensional metric, it must possess the enhanced E(2) symmetry.
In order to break the isometry further down to the T(2) very special conformal symmetry, we need a non-trivial matter configuration. In our case, the massive two-form tensor has the condensation $$\begin{aligned}
B = \frac{1}{2} B_{MN} dx^{M} dx^N = b \frac{dx^{-}dx^2}{z^3} \ ,\end{aligned}$$ which is invariant only under the isometry of the T(2) very special conformal symmetry: $J_{ij}$ rotation is obviously broken by choosing a particular $x^2$ direction in the two-form. We can readily compute the matter energy-momentum tensor and it has only component in $--$. Then the $--$ component of the Einstein equation demands $c=6b^2$. Note that $c$ can be changed by the rescaling of $dx^-$ and $dx^+$ unless it vanishes, so its magnitude is not physically significant. We also note that although $B$ breaks the $ij$ rotation, the energy-momentum tensor constructed out of $B$ does not break the $ij$ rotation so that the Einstein equation is consistent with the above metric ansatz.
In $d=5$ dimensions, we may use an alternative way to introduce a massive two-form field from the topological coupling. Let us consider a complex two-form field $q_2$ with the action $$\begin{aligned}
S_q = \int \left(\frac{1}{2}R *1 - \Lambda *1 -\frac{i}{3}\bar{q}_2 \wedge dq_2 + \frac{i}{3}q_2 \wedge d\bar{q}_2 -2q_2\wedge *\bar{q}_2 \right) \ , \end{aligned}$$ where $*$ is the five dimensional Hodge dual operator. The equation of motion of $q_2$ gives the self-dual equation $$\begin{aligned}
dq_2 = 3i * q_2 \end{aligned}$$ which describes a propagating (real) massive two-form with $m^2 = 9$. To see this, we may substitute $q_2 = k_2 + i l_2$ with real two-form fields $k_2$ and $l_2$. The equations of motion tell that $l_2 = \frac{1}{3} * dk_2$ and $k_2$ (or accordingly $l_2$ as well) satisfies the massive two-form equation with $m^2 = 9$. The Einstein equation is $$\begin{aligned}
R_{MN} = -4 g_{MN} +2(q_{M}^{L} \bar{q}_{N L} + q_{N}^{L} \bar{q}_{M L} - \frac{1}{12}g_{MN} q_{L K} \bar{q}^{LK}) \end{aligned}$$
We find the following solutions with the symmetry corresponding to the T(2) very special conformal symmetry: the metric is given by $$\begin{aligned}
ds^2 = -c \frac{dx^{-2}}{z^4} + \frac{dz^2-2dx^+dx^{-} + (dx^2)^2 + (dx^3)}{z^2}\end{aligned}$$ and the massive two-form has the condensation $$\begin{aligned}
q_2 = q (\frac{dx^{-}dx^2}{z^3} + i \frac{dx^{-}dx^3}{z^3})\ .\end{aligned}$$ The $--$ component of the Einstein equation demands $c = \frac{4}{3} q^2$.
Uplifting to type IIB supergravity
==================================
In the previous section, we constructed the gravity dual for the T(2) very special conformal field theories in the bottom up approach. Here we would like to uplift them to the ten dimensional type IIB supergravity, allowing us to interpret them from the string theory viewpoint as we will discuss in the next section.
Our starting point is the $AdS_5 \times SE_5$ flux compactification of type IIB supergravity $$\begin{aligned}
ds_{10}^2 &= ds^2(AdS_5) + ds^2(SE_5) \cr
F_{(5)} &= 4 \mathrm{vol}(SE_5) + 4\mathrm{vol}(AdS_5) \cr
H_{(3)}&= F_{(3)}=0 \ \end{aligned}$$ with a constant axion-dilaton field. Here the Sasaki-Einstein space $SE_5$ has a preferred $U(1)$ isometry called Reeb Killing vector and the metric has the local decomposition of $$\begin{aligned}
ds^2(SE_5) = ds^2(KE_4) + \eta^2 ,\end{aligned}$$ where $\eta = d\psi + \mathcal{A}$ is the (contact) one-form dual to the Reeb Killing vector. On the Sasaki-Einstein space, there is an $SU(2)$ structure $(\eta,J,\Omega)$ which satisfies $$\begin{aligned}
J\wedge \Omega &= 0 \cr
\Omega \wedge \bar{\Omega} &= 2 J \wedge J = 4 \mathrm{vol}(KE_4) \cr
*_4 J &= J \cr
*_4 \Omega &= \Omega \cr
d\eta &= 2J \cr
d\Omega &= 3i \eta \wedge \Omega \ .\end{aligned}$$ Here $*_4$ is the Hodge dual on the (locally) Kahler Einstein base $KE_4$. The Kahler form $J$ and $(2,0)$ form $\Omega$ are induced from those of the Kahler-Einstein base $KE_4$. The holographic interpretation of these gravity solutions have been studied in the literature and they are given by $\mathcal{N}=1$ superconformal field theories realized as large $N$ quiver gauge theories [@Klebanov:1998hh][@Benvenuti:2004dy].
With these ingredients, our Kaluza-Klein ansatz is to deform the five-form flux as $$\begin{aligned}
ds^2_{10} &= ds^2(E_5) + ds^2(KE_4) + \eta^2 \cr
F_{(5)} &= 4 \mathrm{vol}(SE_5) + 4\mathrm{vol}(E_5) + (* q_2 \wedge \Omega + q_2 \wedge \Omega \wedge \eta + c.c.) \cr
F_{(3)} &= 0 \cr
H_{(3)} &= 0 \cr
C_{(0)} &= 0 \cr
\Phi &= \phi \ , \label{KK}\end{aligned}$$ where $*$ is the Hodge dual with respect to the five dimensional metric $ds^2(E_5)$. For the complete consistent truncation, we have to keep the other scalar degrees of freedom such as the breathing mode, but for our particular solution, we are not going to excite any scalar degrees of freedom as we will argue so that this ansatz will become consistent.
As discussed in [@Liu:2010sa][@Gauntlett:2010vu], the Kaluza-Klein reduction gives us the five dimensional Einstein gravity coupled with massive (complex) two-form field $q_2$ that has the topological mass term with $m^2 = 9$, which we have already discussed in the previous section. The non-trivial claim is that the Kaluza-Klein reduction is a consistent truncation and the five dimensional solution $$\begin{aligned}
ds^2 &= -c \frac{dx^{-2}}{z^4} + \frac{dz^2-2dx^+dx^{-} + (dx^2)^2 + (dx^3)}{z^2} \cr
q_2 &= q (\frac{dx^{-}dx^2}{z^3} + i \frac{dx^{-}dx^3}{z^3})\ .\end{aligned}$$ substituted into solves the ten-dimensional type IIB supergravity equations of motions when $c = \frac{4}{3} q^2$. Most of the relevant computation has been done in the literature [@Liu:2010sa][@Gauntlett:2010vu], but since it is a relatively simple solution, one can directly check that the ten dimensional equations of motion hold.
Explicitly, the five form equations of motion gives $$\begin{aligned}
dq_2 = 3i * q_2 \end{aligned}$$ and it is readily solved by our ansatz since it is the same equation as the self-dual equation of the complex two-form field. The self-dual condition of the five form was already encoded in our Kaluza-Klein ansatz. Then the $MN$ (i.e. five dimensional) component of the ten dimensional Einstein equation gives $$\begin{aligned}
R_{MN} = -4 g_{MN} +2(q_{M}^{L} \bar{q}_{N L} + q_{N}^{L} \bar{q}_{M L} - \frac{1}{12}g_{MN} q_{L K} \bar{q}^{LK}) \label{einstein}\end{aligned}$$ and this is exactly the same equations that we have already solved in the effective five dimensional gravity. The crucial point here is that the other equations of motion are not affected because in those equations we have to contract $q_2$ to make a scalar, but since $q_2$ has only a null component, the scalar contraction always gives zero. Another point is that the Einstein equation holds not only in average but pointwise in the internal space due to the special geometric feature of our Kaluza-Klein ansatz.
Field theory interpretation
===========================
A comparison of the Kaluza-Klein ansatz with the linearized harmonic analysis on the flux compactification of the Sasaki-Einstein space with the AdS/CFT dictionary allows us to interpret the holographic solutions we constructed from the dual field theory perspective. As we have already mentioned two-form fields with the mass $m^2=9$ in the gravity side correspond to dimension five anti-symmetric tensor operators in the dual field theories. More explicitly, in the Sasaki-Einstein compactification, it is given by $$\begin{aligned}
O = \sum_I c_I \mathrm{Tr} F_I^{x-}(\lambda^I \lambda^I + cc) \ ,\end{aligned}$$ where $\lambda^I$ are gauginos and $I$ runs over the gauge groups of the quiver gauge theory. See e.g. [@Ceresole:1999zs] for the case of $SE_5 = T_{1,1}$ and [@Gunaydin:1984fk][@Kim:1985ez] for the case of $SE_5=S_5$. In the general flux compactification of the Sasaki-Einstein space, there is one particular choice of $c_I$ over different gauge groups which have a protected scaling dimension so that one can use it as our dimension five anti-symmetric tensor deformations. In the case of $\mathcal{N}=4$ super Yang-Mills theory, it is in the $10_c$ rep of SU(4) R-symmetry from the choice of the four different fermions, and all of them can be used.
The operator $O$ is charged under the R-symmetry as we can also see from the Kaluza-Klein reduction (since $\Omega$ appearing in the ansatz transforms under the Lie derivative with respect to the Reeb Killing vector), and it breaks the $R$-symmetry once added. Note also that the operator breaks (at least part of) the supersymmetry because the protected operator $ \mathrm{Tr} F_{\dot{\alpha}\dot{\beta}} \lambda \lambda$ and its conjugate $\mathrm{Tr}F_{{\alpha}{\beta}} \bar{\lambda} \bar{\lambda}$ preserves the different supersymmetry. Even this is the case, the deformation is an exactly marginal deformation with respect to the very special conformal symmetry as demonstrated in the holographic description since the deformation parameter $c$ is arbitrary.
Discussions
===========
In this paper, we constructed the gravity dual for T(2) invariant very special conformal field theories. In particular, we have presented new exact solutions in type IIB supergravity with the isometry corresponding to the T(2) very special conformal symmetry. Such top-down constructions give precise field theory interpretation of our dual descriptions in terms of explicit $\mathcal{N}=1$ supersymmetric conformal field theories.
Our field theory interpretation reveals that we can construct very special conformal field theories from the Poincaré conformal field theory by introducing dimension five operator such as the gauge field strength multiplied by gaugino bilinear. The similar operator can be found in the standard model of particle physics. For example, in QED one can readily construct the dimension five anti-symmetric tensor operator such as $F^{x-} \bar{\Psi}_e\Psi_e$, where $\Psi_e$ is the electron field, and in this way one may imagine the Lorentz violating interaction with the T(2) very special conformal symmetry may exist in the (ultraviolet completion of the) nature.
Beside the particle physics application, we would like to point out that the gravity dual description of T(2) symmetric very special conformal field theories have potential applications in condensed matter physics. Suppose we compactify the null direction $x^+$, and regard the Kaluza-Klein momentum as a particle number. Then we obtain non-relativistic conformal field theories with $z=2$ scaling. Here $x^-$ direction is identified with the non-relativistic time, and $J_{+i}$ is regarded as the Galilean boost symmetry. Furthermore $K_+$ is regarded as the non-relativistic special conformal transformation. Now the difference between our (compactified) T(2) invariant non-relativistic critical systems and the ones studied in the literature (e.g. the Schödinger conformal field theories) is that we do not have the spatial rotation as a symmetry. Therefore, our theory is suitable to describe the system without rotational symmetry (but with the translational symmetry). Such critical systems are ubiquitous: one may realize them by introducing background electric field or magnetic field along a particular spatial direction. It would be interesting to investigate their properties such as thermal properties by constructing corresponding blackhole solutions in our setup.
Acknowledgements {#acknowledgements .unnumbered}
================
This work is in part supported by JSPS KAKENHI Grant Number 17K14301. The preliminary results of this paper were presented at KIAS-YITP 2017 “Strings, Gravity and Cosmology”. The author would like to thank J. Gauntlett and N. Kim for fruitful comments and lively discussions there.
[99]{} A. G. Cohen and S. L. Glashow, Phys. Rev. Lett. [**97**]{}, 021601 (2006) doi:10.1103/PhysRevLett.97.021601 \[hep-ph/0601236\]. A. G. Cohen and S. L. Glashow, hep-ph/0605036.
Y. Nakayama, Phys. Rev. D [**97**]{}, no. 6, 065003 (2018) doi:10.1103/PhysRevD.97.065003 \[arXiv:1707.05423 \[hep-th\]\].
Y. Nishida and D. T. Son, Phys. Rev. D [**76**]{}, 086004 (2007) doi:10.1103/PhysRevD.76.086004 \[arXiv:0706.3746 \[hep-th\]\].
D. T. Son, Phys. Rev. D [**78**]{}, 046003 (2008) doi:10.1103/PhysRevD.78.046003 \[arXiv:0804.3972 \[hep-th\]\]. K. Balasubramanian and J. McGreevy, Phys. Rev. Lett. [**101**]{}, 061601 (2008) doi:10.1103/PhysRevLett.101.061601 \[arXiv:0804.4053 \[hep-th\]\].
A. Adams, K. Balasubramanian and J. McGreevy, JHEP [**0811**]{}, 059 (2008) doi:10.1088/1126-6708/2008/11/059 \[arXiv:0807.1111 \[hep-th\]\]. J. Maldacena, D. Martelli and Y. Tachikawa, JHEP [**0810**]{}, 072 (2008) doi:10.1088/1126-6708/2008/10/072 \[arXiv:0807.1100 \[hep-th\]\]. C. P. Herzog, M. Rangamani and S. F. Ross, JHEP [**0811**]{}, 080 (2008) doi:10.1088/1126-6708/2008/11/080 \[arXiv:0807.1099 \[hep-th\]\]. A. Bergman and O. J. Ganor, JHEP [**0010**]{}, 018 (2000) doi:10.1088/1126-6708/2000/10/018 \[hep-th/0008030\]. M. Alishahiha and O. J. Ganor, JHEP [**0303**]{}, 006 (2003) doi:10.1088/1126-6708/2003/03/006 \[hep-th/0301080\]. Y. Nakayama, arXiv:1802.06489 \[hep-th\]. J. T. Liu, P. Szepietowski and Z. Zhao, Phys. Rev. D [**81**]{}, 124028 (2010) doi:10.1103/PhysRevD.81.124028 \[arXiv:1003.5374 \[hep-th\]\].
J. P. Gauntlett and O. Varela, JHEP [**1006**]{}, 081 (2010) doi:10.1007/JHEP06(2010)081 \[arXiv:1003.5642 \[hep-th\]\]. I. R. Klebanov and E. Witten, Nucl. Phys. B [**536**]{}, 199 (1998) doi:10.1016/S0550-3213(98)00654-3 \[hep-th/9807080\]. S. Benvenuti, S. Franco, A. Hanany, D. Martelli and J. Sparks, JHEP [**0506**]{}, 064 (2005) doi:10.1088/1126-6708/2005/06/064 \[hep-th/0411264\]. A. Ceresole, G. Dall’Agata, R. D’Auria and S. Ferrara, Phys. Rev. D [**61**]{}, 066001 (2000) doi:10.1103/PhysRevD.61.066001 \[hep-th/9905226\].
M. Gunaydin and N. Marcus, Class. Quant. Grav. [**2**]{}, L11 (1985). doi:10.1088/0264-9381/2/2/001 H. J. Kim, L. J. Romans and P. van Nieuwenhuizen, Phys. Rev. D [**32**]{}, 389 (1985). doi:10.1103/PhysRevD.32.389
[^1]: The cases with SIM(2) and HOM(2) very special conformal field theories are much more challenging because we have to abandon either unitarity or locality. See some ideas to construct holographic duals in [@Nakayama:2018fib].
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Gen Tatara$^1$, Hidetoshi Fukuyama$^2$'
title: Phasons and excitations in skyrmion lattice
---
Introduction
============
A magnetic skyrmion is a magnetization structure in magnetic materials with spins at the core and perimeter pointing up and down, respectively. It has a topological charge, defined in two-dimensions as $$\begin{aligned}
n\equiv \frac{1}{4\pi M^3}\int d^2r \Mv\cdot(\nabla_x\Mv\times\nabla_y\Mv), \end{aligned}$$ of 1, where $\Mv$ is a vector representing the magnetization.
Clusters of such structures forming a lattice were known in thin film ferromagnets under magnetic field, although the structures were called magnetic bubbles at that time [@Malozemoff79]. The lattice of magnetic bubbles is stabilized because of competition between a uniaxial magnetic anisotropy energy and dipolar interaction energy. Skyrmion systems in non-centrosymmetric magnets have been intensively studied recently. Skyrmion lattices in helimagnets were experimentally discovered by neutron scattering measurements on bulk MnSi by Mühlbauer et al.[@Muhlbauer09]. Electron transport measurements were carried out by Neubauer et al. and topological Hall effect due to the spin Berry’s phase, which is proportional to the skyrmion number, was detected in the skyrmion phase of bulk MnSi [@Neubauer09]. Real-space observation of a skyrmion lattice was carried out by Yu et al. on a thin film of Fe$_{0.5}$Co$_{0.5}$Si using Lorentz transmission electron microscopy and the phase diagram was obtained by measuring the skyrmion density [@Yu10]. The sample thickness, 20 nm, is smaller than the helix period of the system, 90 nm, and thus the skyrmion structure observed is a two-dimensional one. It was noted there that the skyrmion lattice in this thin film appears over a wide region of the phase diagram, including very low temperature. Stability of the skyrmion phase in thin films turned out to be a common feature, as was reported in other systems like FeGe [@Huang12]. A skyrmion lattice in a thin film of a chiral magnetic insulator, Cu$_2$OSeO$_3$, was observed by Lorentz TEM measurement by Seki et al. [@Seki12]. The material Cu$_2$OSeO$_3$ is multiferroic, and thus the electric control of skyrmion structures is possible [@White12]. Recent developments are reviewed in Ref. [@Nagaosa13].
Jonietz et al. succeeded to induce the rotation of a skyrmion lattice in bulk MnSi by applying an electric current density of $10^6$ A/m$^2$ [@Jonietz10], and Schulz et al. induced the translational motion at the current density of the same order [@Schulz12]. The current density is about $10^{5}$ times smaller than the typical current density needed to drive magnetic domain walls, and such low current operation is a notable feature of topological magnetic structures [@TKS_PR08]. Schulz et al. detected the motion by measuring the emergent electric field generated by the moving topological object, and demonstrated that the skyrmion system is an intriguing playground for studying the emergent electromagnetism.
The excitation modes of skyrmion lattice were observed in Cu$_2$OSeO$_3$ by microwave absorption by Onose et al. [@Onose12]. Excitation modes were found near 1 GHz and 1.5 GHz when the AC magnetic field is applied within and perpendicular to the skyrmion plane, respectively, and they are assigned to be the rotational and breathing modes proposed in Ref. [@Mochizuki12], respectively.
A theoretical ground for stabilization of skyrmion lattices was given by Mühlbauer et al., based on a representation of a skyrmion lattice as a superposition of three helices [@Muhlbauer09]. They discussed that the quartic term in the Ginzburg-Landau free energy, $M^4$, gives rise to a term cubic in magnetization, $BM_zM^2$, when an external magnetic field, $B$, is applied ($z$ represents the direction of the field), and have shown that this “cubic” term is crucial for stabilization of the three helices state. The important role of the “cubic“ term to stabilize the states with three density waves was pointed out originally in the case of charge-density waves (CDW) by McMillan in 1975 for transition-metal dichalcogenides where transition metals are on planar hexagonal lattice and band structure with sixfold symmetry in the basal plane [@McMillan75]. The excitation of skyrmion lattice was studied including the “cubic” term by Petrova and Tchernyshyov, who found that the two vibration modes of the lattice are mixed because of the topological nature of the skyrmion lattice, resulting in a gapped mode and a gapless mode with a quadratic dispersion[@Petrova11]. Numerical studies of skyrmion lattice were carried out in Refs. [@Mochizuki12; @Ohe13]. The effect of a uniaxial magnetic anisotropy on stability of a skyrmion lattice (a magnetic vortex state) was studied in Refs. [@Bogdanov94; @Wilson14]. Current-induced dynamics of skyrmions was studied by numerical simulations [@IwasakiNC13; @Iwasaki13; @Sampaio13; @Mochizuki14]. The velocity of skyrmion lattice was deduced from a numerical simulation in Ref. [@IwasakiNC13], and it was found that the current density needed to drive a skyrmion lattice is lower than that of domain walls by a factor of about $10^{5}$. This feature is mainly attributed to the absence of intrinsic pinning effect for general two-dimensional topological structures [@TKS_PR08]. Skyrmions may have advantage even over vortices because isolated skyrmions are repelled by sample edges at low current density [@IwasakiNC13; @Sampaio13], while vortices are larger and softer structures which tend to annihilate easily at the edges [@NSTKTM08]. The effect of a thermal gradient was studied numerically in Ref. [@Mochizuki14]
Although skyrmion lattices are unique systems realized in magnets without inversion symmetry, the behaviors of excitations have similarity to those in Wigner crystals and CDW in two-dimensions under a magnetic field. In fact, the field-induced splitting of the transverse and longitudinal phonon modes having a linear dispersion in two-dimensional Wigner crystal was discussed in Refs. [@Chaplik72; @Fukuyama75]. Under a magnetic field, one of the modes was shown to acquire a mass proportional to the magnetic field as a result of the cyclotron motion, while the other modes remain massless. In the case of CDW, such dynamics is properly described by field variables (not particle coordinates) called phasons, which were first introduced by Lee, Rice and Anderson [@Lee74] for one-dimesional CDW resulting from Peierls transition. In fact the similar coupling of modes in two-dimesional CDW under magnetic field has been identified in terms of phasons [@Fukuyama78a]. Phasons are also useful to study pinning due to impurities and commensurability [@FukuyamaTakayama85]. Regarding the pinning by random impurities, the existence of two typical cases of weak and strong pinning has been clarified for one-dimensional CDW in terms of phason by Fukuyama and Lee [@Fukuyama78; @FukuyamaTakayama85]. It was shown that even in the weak pinning case, a collective pinning occurs resulting in the appearance of a characteristic length scale above which the structure is disordered due to the random potentials. The idea was extended to the case of three-dimensions by Lee and Rice [@Lee79].
Similarly, in order to describe the dynamics of magnetic structures, collective coordinates representation extracting the low energy excitations is highly useful. The dynamics of each localized spin is governed by a torque equation, and thus the dynamics of the structure may seem complicated if described in terms of the local torque acting on each spin. In contrast, the low energy behavior becomes clearer by introducing correct collective coordinates. In the case of a planar domain wall, Slonczewski pointed out that two coordinates, the wall position and the tilt angle of the wall plane, are good variable and derived the equations of motion (called the Slonczewski’s equation) in the presence of an external magnetic field [@Slonczewski72]. The critical feature of the domain wall is that the two coordinates are canonically conjugate to each other. For a translational motion of the wall, therefore, tilting of the wall plane is necessary. This feature was shown to result, in the case driven by an electric current, in an intrinsic pinning effect, which hinders the wall motion at low current even in the absence of extrinsic pinning potentials [@TK04]. It was discussed that a magnetic vortex in a film is described by the two coordinates representing the position, $X$ and $Y$, of its core in the film, and that the canonical relation between $X$ and $Y$ arises due to the topological number of the vortex [@Thiele73]. The vortex is therefore free from the intrinsic pinning effect [@TKS_PR08].
In this paper, we study the dynamics of the skyrmion lattice in a two-dimensional helimagnet based on a collective coordinates description. The skyrmion lattice is described by use of three helices [@Muhlbauer09], and three phase modes (phasons) representing translational modes and three modes describing the excitations out-of-the-helix planes are introduced following the study by Petrova and Tchernyshyov [@Petrova11]. The coordinates for the skyrmion lattice are field variables that depend on spatial position and the time, in contrast to the case of a domain wall and a vortex. The field representing the phase fluctuation is called the phason. We derive the effective Lagrangian for the collective coordinates in the slowly-varying case. We show that the system has two independent excitation modes corresponding to phasons along the two orthogonal directions. Their effective Lagrangian is shown to be equivalent to the one for a two-dimensional charged field with a mass and under a magnetic field as has been pointed out previously [@Chaplik72; @Fukuyama75]. The low energy excitations are a gapless mode with quadratic dispersion and a massive mode. The quadratic dispersion is different from the dispersion in the case of the Wigner crystal, and is due to the charge neutrality of phasons in the present case. The massive mode corresponds to the rotational mode of skyrmion cores. It turns out also that there is another mode, $\varphi_0$, which is non-dynamic in the present approximation but governs the topological nature of skyrmion lattice. We show that the magnetic field applied perpendicular to the skyrmion plane suppresses the fluctuation of the $\varphi_0$ mode and this effect stabilizes the skyrmion lattice. When the field is weak, $\varphi_0$ fluctuation leads to a screening of the skyrmion topological number, and this screening effect affects various physical quantities such as the excitation energy, topological Hall effect and the depinning threshold current in metals.
Model
=====
We consider a two-dimensional skyrmion lattice in the $xy$ plane based on the Ginzburg-Landau free energy for a coarse-grained magnetization in the continuum [@Muhlbauer09] $$\begin{aligned}
H &= {\int\frac{d^2r}{a^2}}\lt[\frac{J}{2}(\nabla\Mv)^2
+ D\Mv\cdot(\nabla\times\Mv) \rt] +H_{M} +H_{\rm pin}+H_{\rm ST}
\nnr
&={\int\frac{d^2r}{a^2}}\lt[\frac{J}{2}(\nabla\Mv)^2
- D\sum_{\mu=x,y}\sum_{\nu\lambda}\epsilon_{\mu\nu\lambda}M_\nu \nabla_\mu M_\lambda \rt] +H_{M} +H_{\rm pin}+H_{\rm ST}, \label{Hamiltonian}\end{aligned}$$ where $\Mv$ a dimensionless vector representing the magnetization direction ($M$ is the magnetization divided by $\frac{2\mub}{a^3}$ (for the g-factor of 2), where $a$ is the atomic spacing and $\mub$ is the Bohr magneton), $J$ is the exchange interaction energy, and $D$ is the strength of the Dzyaloshinskii-Moriya (DM) interaction. We consider the case where the DM interaction is isotropic in the $xy$ plane. $H_{M}$ is the free energy for the uniform magnetization including the effect of a magnetic field, $B$, applied in the $-z$ direction, $$\begin{aligned}
H_{M} &= {\int\frac{d^2r}{a^2}}\lt[\frac{a_M}{2}M^2+\frac{b_M}{4}M^4 + \frac{2\mub}{a^3}BM_z\rt], \end{aligned}$$ where $a_M$ and $b_M$ are parameters generally dependent on the temperature. The effect of uniaxial magnetic anisotropy energy along the $z$-direction discussed recently [@Wilson14] is incorporated in the present formalism by the uniform component of magnetization (${M_{\rm f}}$ below) including the anisotropy. A pinning effect due to random impurities is represented by $H_{\rm pin}$ and driving force due to electric current in metals is represented by the spin-transfer term, $H_{\rm ST}$. Our study without the spin-transfer term applies both to metals and insulators.
A calculation is carried out based on the Lagrangian formalism, where the Lagrangian is defined as $$\begin{aligned}
L &=L_{\rm B}-H,\end{aligned}$$ where $L_{\rm B}$ is the spin Berry’s phase term, which describes the dynamics of spin systems [@Auerbach94]. It is usually represented using polar coordinates for $\Mv$, $\theta, \phi$, as $$\begin{aligned}
L_{\rm B}&={\int\frac{d^2r}{a^2}}\hbar M\dot{\phi}(\cos\theta-1).\label{LBdef}\end{aligned}$$
The DM interaction breaks the inversion symmetry and favors helix magnetization structures. For instance, in the absence of external magnetic field, there is a solution of a single helix with wave vector $\kv$ with magnetization direction, $\nv$, rotating within the plane perpendicular to $\kv$. The wave vector may point in any direction in the $xy$-plane and its magnitude is determined by the exchange energy and DM interaction as $$\begin{aligned}
k &= \frac{D}{J}.\end{aligned}$$ The chirality of the helix is determined by the DM interaction to satisfy $(\hat{\kv}\cdot\nabla)\nv=\kv\times \nv$, where $\hat{\kv}\equiv \kv/k$, and thus a single helix solution is $$\begin{aligned}
\nv_{\rm 1h} &= \hat{\zv} \cos(\kv\cdot\rv) +(\hat{\kv}\times \hat{\zv}) \sin(\kv\cdot\rv).\end{aligned}$$ As was shown in Ref. [@Muhlbauer09], the skyrmion lattice state has a lower energy than a single helix state for a finite region in the plane of the temperature and magnetic field. We proceed focusing on the case skyrmion lattice is realized as a ground state. The skyrmion lattice is represented by a superposition of three helices whose wave vectors form an equilateral triangle in the presence of a uniform magnetization component, ${M_{\rm f}}$, induced by the applied magnetic field. The uniform component and the interaction arising from the quartic term in $H_{M}$ are essential for the stability of three helices state (See Ref. [@Muhlbauer09] and Appendix \[SECGL\]). The three wave vectors are chosen as (Fig. \[FIGkabc\]) $$\begin{aligned}
{\kv_{a}}&=k(1,0,0) \nnr
{\kv_{b}}&=k\lt(-\frac{1}{2},\frac{\sqrt{3}}{2},0\rt) \nnr
{\kv_{c}}&=k\lt(-\frac{1}{2},-\frac{\sqrt{3}}{2},0\rt) ,\end{aligned}$$ A skyrmion lattice configuration is represented by $$\begin{aligned}
\Mv &= {M_{\rm f}}\hat{\zv}+\sum_{i=a,b,c}\Mv_{i}, \label{skxconfiguration}\end{aligned}$$ where $\Mv_{i}$ represents the magnetization vector of three helices denoted by $i=a,b,c$. When fluctuations are neglected, we have $\Mv_{i} = {M_{\rm h}}\nv_i^{(0)}$, where ${M_{\rm h}}$ is the magnitude of the helices and $$\begin{aligned}
\nv_i^{(0)} &\equiv \hat{\zv} \cos(\kv_i\cdot\rv) +(\hat{\kv}_i\times \hat{\zv}) \sin(\kv_i\cdot\rv).\label{nvi0def}\end{aligned}$$ A structure of $\Mv$ is shown in Fig. \[FIGskx\] for ${M_{\rm f}}/{M_{\rm h}}=-0.8$. The spacing between the skyrmion cores is $a_{\rm s}\equiv\frac{4\pi}{\sqrt{3}k}=\frac{2}{\sqrt{3}}\lambda$, where $\lambda\equiv \frac{2\pi}{k}$ is the helix period. Since the topological charge of one skyrmion is $2\pi\frac{\hbar}{e}$, and each triangle of area $\frac{\sqrt{3}}{4}a_{\rm s}^2$ contains a half skyrmion, the average density of the effective magnetic field is $\frac{h}{e}\frac{\sqrt{3}}{2\lambda^2}$.
![ The wave vectors of the three helices, ${\kv_{a}}$, ${\kv_{b}}$ and ${\kv_{c}}$. \[FIGkabc\]](kabc.eps){width="0.45\hsize"}
![ Plot of skyrmion structure $\Mv$ constructed as a superposition of three helices and uniform magnetization ${M_{\rm f}}$ with ${M_{\rm f}}/{M_{\rm h}}=-0.8$. Directions of ${\kv_{a}}$, ${\kv_{b}}$ and ${\kv_{c}}$ are shown by black arrows. \[FIGskx\]](skyrmion_k.eps){width="0.4\hsize"}
To describe the excitations and dynamics of the skyrmion lattice, we introduce two collective coordinates for each helix. One is $\varphi_i(\rv,t)$ representing the phase of the helices and the other is $\beta_i(\rv,t)$ representing the fluctuation out-of helix plane, i.e., the fluctuation along $\kv_i$. The magnetization of the helices then read $$\begin{aligned}
\Mv_{i}
&= {M_{\rm h}}\lt( \beta_i \hat{\kv}_i +\sqrt{1-\beta_i^2} \nv_{i}\rt),
\label{Mvi}\end{aligned}$$ where $\nv_{i}$ are $$\begin{aligned}
\nv_{i} &= \hat{\zv} \cos(\kv_i\cdot\rv+\varphi_i) +(\hat{\kv}_i\times \hat{\zv}) \sin(\kv_i\cdot\rv+\varphi_i). \label{nvidef}\end{aligned}$$ The three variables $\varphi_i$’s represent the phasons of helical waves.
![ Excitation modes of a helix represented by $\varphi$ and $\beta$. Spins excited (i.e., having non-vanishing $\varphi$ or $\beta$) are shown in red. The mode $\varphi$ corresponds to a local modification of the helix pitch and the mode $\beta$ describes the tilt of the spin along the helix wave vector. \[FIGphibeta\]](helix_phi_l.eps "fig:"){width="0.45\hsize"} ![ Excitation modes of a helix represented by $\varphi$ and $\beta$. Spins excited (i.e., having non-vanishing $\varphi$ or $\beta$) are shown in red. The mode $\varphi$ corresponds to a local modification of the helix pitch and the mode $\beta$ describes the tilt of the spin along the helix wave vector. \[FIGphibeta\]](helix_beta_l.eps "fig:"){width="0.45\hsize"}
Effective Lagrangian for collective coordinates
===============================================
We use Eq. (\[nvidef\]) to derive the effective Lagrangian for the collective coordinates, $\varphi_i$’s and $\beta_i$’s. We expand with respect to $\beta_i$’s to the second order, and neglect spatial variation of ${M_{\rm h}}$. We assume that the fluctuations are slowly varying compared to the wave vectors of helices, and drop rapidly oscillating terms when we integrate over space. Namely, we approximate $$\begin{aligned}
\int d^2r \cos\Phi_i \cos\Phi_j &=
\frac{1}{2}\int d^2r\lt[\cos( (\kv_i+\kv_j)\cdot\rv+\varphi_i+\varphi_j) + \cos( (\kv_i-\kv_j)\cdot\rv+\varphi_i-\varphi_j) \rt]\nnr
& \simeq \frac{1}{2}\delta_{ij} \int d^2r \cos(\varphi_i-\varphi_j) \nnr
& = \int d^2r \sin\Phi_i \sin\Phi_j ,\end{aligned}$$ and $$\begin{aligned}
\int d^2r \cos\Phi_i \sin\Phi_j &=0 ,\label{slowvarying}\end{aligned}$$ where $\Phi_i\equiv \kv_i\cdot\rv+\varphi_i$. The average of the exchange interaction term then becomes $$\begin{aligned}
{\int\frac{d^2r}{a^2}}\frac{J}{2}(\nabla\Mv)^2
& \simeq {\int\frac{d^2r}{a^2}}\frac{J{M_{\rm h}}^2}{2}\lt[
\frac{1}{2}\sum_{i\neq j} \lt(\nabla\beta_i-\nabla\beta_j\rt)^2
+\sum_i(1-\beta_i^2)[(k_i)^2+(\nabla\varphi_i)^2+2\kv_i\cdot\nabla\varphi_i]\rt].\end{aligned}$$ DM term similarly is calculated as $$\begin{aligned}
{\int\frac{d^2r}{a^2}}D\Mv\cdot(\nabla\times\Mv)
& \simeq {\int\frac{d^2r}{a^2}}{M_{\rm h}}^2 \lt[
-3Dk+Dk\sum_i\beta_i^2-D\sum_i(\hat{\kv}_i\cdot\nabla)\varphi_i\rt].\end{aligned}$$ The sum of the exchange and DM interactions thus reads $$\begin{aligned}
H_J+H_{\rm DM}
& \simeq {\int\frac{d^2r}{a^2}}{M_{\rm h}}^2\lt[
\frac{J}{2}\sum_i(k^2\beta_i^2+(\nabla\varphi_i)^2)
+\frac{J}{4}\sum_{i\neq j} \lt(\nabla\beta_i-\nabla\beta_j\rt)^2
\rt].
\end{aligned}$$ We see that the excitation mode described by $\beta_i$ has a mass proportional to $Jk^2$.
The contributions from $H_{M}$ are calculated similarly. We neglect here the contributions including massive mode, $\beta_i$. As noted in Ref. [@Muhlbauer09], the quartic term is essential for describing a skyrmion lattice. From the expression $$\begin{aligned}
M^4 &= [({M_{\rm f}}\hat{\zv}+{\Mv_{\rm h}})^2]^2 \nnr
&={M_{\rm f}}^4+{M_{\rm h}}^4+2{M_{\rm f}}^2{M_{\rm h}}^2+4{M_{\rm f}}^2({\Mv_{\rm h}}\cdot\hat{\zv})^2+ 4{M_{\rm f}}^3({\Mv_{\rm h}}\cdot\hat{\zv})
+4{M_{\rm f}}({\Mv_{\rm h}}\cdot\hat{\zv}){M_{\rm h}}^2,\end{aligned}$$ where ${\Mv_{\rm h}}\equiv\sum_i \Mv_i$ is the magnetization of three helices, we see that the last term linear in ${M_{\rm f}}$ describes the interaction among three helices. The integral of the term is evaluated for slowly varying case as $$\begin{aligned}
\int d^2r ({\Mv_{\rm h}}\cdot\hat{\zv}){M_{\rm h}}^2 &= \frac{9}{4}{M_{\rm h}}^3 \int d^2r \cos(\varphi_a+\varphi_b+\varphi_c),
\label{cubicterm}\end{aligned}$$ where we noted $\sum_{i=a,b,c}\Phi_i=\sum_{i}\varphi_i$ as a result of $\sum_{i}\kv_i=0$ and $$\begin{aligned}
\int d^2r \cos\Phi_a \cos\Phi_b\cos\Phi_c &= \frac{1}{4}\int d^2r \cos(\varphi_a+\varphi_b+\varphi_c) \nnr
&= - \int d^2r \cos\Phi_a \sin\Phi_b\sin\Phi_c.\end{aligned}$$
Spin Berry’s phase term \[SECsbp\]
----------------------------------
For studying dynamics of collective coordinates, spin Berry’s phase term, Eq. (\[LBdef\]), needs to be treated carefully, since this expression generally includes a physically irrelevant contribution arising from the fact that the term is not well-defined at $\cos\theta=-1$. To discuss physical contribution, it is useful to rewrite the spin Berry’s phase term to avoid unphysical divergence at $\cos\theta=-1$. To do this, we consider a change of ${\Mv_{\rm h}}$ when collective coordinates $\varphi_i$ ($i=a,b,c$) are changed from $\varphi_i^0$ to $\varphi_i$. This change corresponds to a shift of the origin of the skyrmion lattice and does not have physical effect, since only the relative phase is meaningful if pinning is neglected. Without losing generality, we chose $\varphi_i^0=0$. The change of helix magnetization, $\delta\nv$, reads $$\begin{aligned}
\delta\nv=\sum_{i=a,b,c}\frac{\partial \nv_i}{\partial \varphi_i}\varphi_i
=\sum_{i=a,b,c}(\hat\kv_i\times\nv_i)\varphi_i. \label{deltanv}\end{aligned}$$ By the change of phase, the spin Berry’s phase term, $L_{\rm B}$, is modified by the amount [@Auerbach94] $$\begin{aligned}
\delta L_{\rm B} &= \hbar {M_{\rm h}}\int d^2 r \nv\cdot(\dot{\nv}\times\delta\nv).\end{aligned}$$ This form of spin Berry’s phase contains the correct physical dynamics of $\varphi_i$ and $\beta_i$. In terms of $\beta_i$ and $\varphi_i$, $\dot{\nv}$ reads $$\begin{aligned}
\dot{\nv} &= \sum_i\lt[ \dot{\varphi}_i(\hat\kv_i\times\nv_i)+\dot{\beta}_i \hat \kv_i \rt],\end{aligned}$$ and thus by use of Eq. (\[deltanv\]) $$\begin{aligned}
\delta L_{\rm B} &= \hbar {M_{\rm h}}{\int\frac{d^2r}{a^2}}\sum_{ijk}
\lt[
\dot{\varphi}_i \varphi_j [(\hat\kv_i\times\nv_i)\times (\hat\kv_j\times\nv_j)]
+ \dot{\beta}_i \varphi_j [\hat\kv_i \times (\hat\kv_j\times\nv_j)]
\rt]\cdot\nv_k.\end{aligned}$$ The coefficient connecting $\dot{\varphi}_i$ and $\varphi_j$, $\int d^2r[(\hat\kv_i\cdot\nabla)\nv_i\times (\hat\kv_j\nabla)\nv_j]\cdot\nv_k$, is the topological skyrmion number for a triangular lattice unit defined by $\kv_i$ and $\kv_j$ up to a constant. It is easy to see that $$\begin{aligned}
[(\hat\kv_a\times\nv_a)\times (\hat\kv_b\times\nv_b)] \cdot\nv_c
&=
\frac{3\sqrt{3}}{8} \cos (\Phi_a+\Phi_b+\Phi_c) \nnr
&
+\frac{\sqrt{3}}{8} \lt[ \cos (\Phi_b+\Phi_c-\Phi_a)+ \cos (\Phi_a+\Phi_c-\Phi_b)- \cos (\Phi_a+\Phi_b-\Phi_c) \rt]
\nonumber\\
[\hat\kv_i \times (\hat\kv_j\times\nv_j)
] \cdot\nv_j
&= -(\kv_i\cdot\kv_j)= -\frac{3}{2}\delta_{i,j}+\frac{1}{2},\end{aligned}$$ and we thus obtain, dropping oscillating contributions, $$\begin{aligned}
\int d^2r [(\hat\kv_i\times\nv_i)\times (\hat\kv_j\times\nv_j)] \cdot\nv_k
&=
\int d^2r \frac{3\sqrt{3}}{8} \epsilon_{ijk} \cos (\varphi_a+\varphi_b+\varphi_c) ,
\label{SBPvalue}\end{aligned}$$ where $i,j,k$ runs over $a,b,c$. The result of spin Berry’s phase term is thus $$\begin{aligned}
\delta L_{\rm B} &= \hbar g{M_{\rm h}}{\int\frac{d^2r}{a^2}}\lt[
(\dot{\varphi}_a \varphi_b- \varphi_a \dot \varphi_b)
+(\dot{\varphi}_b \varphi_c- \varphi_b \dot \varphi_c)
+(\dot{\varphi}_c \varphi_a- \varphi_c \dot \varphi_a)\rt]
\cos (\varphi_a+\varphi_b+\varphi_c)
\nnr
&+
\hbar {M_{\rm h}}{\int\frac{d^2r}{a^2}}\lt[
\dot{\beta}_a \lt(\varphi_a- \frac{1}{2}\lt(\varphi_b + \varphi_c\rt)\rt)
+
\dot{\beta}_b \lt(\varphi_b- \frac{1}{2}\lt(\varphi_c + \varphi_a\rt)\rt)
+
\dot{\beta}_c \lt(\varphi_c- \frac{1}{2}\lt(\varphi_a + \varphi_b\rt)\rt)
\rt],\end{aligned}$$ where $$\begin{aligned}
g \equiv \frac{3\sqrt{3}}{8},\end{aligned}$$ represents the topological number density for $\varphi_i$’s.
We note that if we estimate the spin Berry’s phase term based on the familiar expression (\[LBdef\]), we obtain $$\begin{aligned}
\int d^2 r \cos\theta \dot{\phi}
&= \int d^2 r \frac{M_z}{M_x^2+M_y^2}(\Mv\times \dot{\Mv})_z \nnr
&= - \sum_{i}\dot\varphi_i\lt(\frac{3}{2}\beta_i-\frac{1}{2}\sum_j\beta_j\rt).\end{aligned}$$ This result is does not have important terms connecting $\dot\varphi$ and $\varphi$, and is not correct. This results from a straightforward treatment of $ \frac{1}{M_x^2+M_y^2}$, which diverges at $M_x=M_y=0$. The same goes for the spin-transfer torque term.
Lagrangian
----------
The total Lagrangian for the collective coordinates (in the absence of current and pinning) thus reads $$\begin{aligned}
L &= {M_{\rm h}}^2 {\int\frac{d^2r}{a^2}}\lt[ \frac{\hbar g}{{M_{\rm h}}} \lt[\dot{\varphi}_a(\varphi_b-\varphi_c)+\dot{\varphi}_b(\varphi_c-\varphi_a)+\dot{\varphi}_c(\varphi_a-\varphi_b)\rt]\cos(\varphi_a+\varphi_b+\varphi_c) \rt. \nnr
&
-\frac{\hbar}{{M_{\rm h}}}\lt( \dot\varphi_a\lt(\beta_a-\frac{1}{2}(\beta_b+\beta_c) \rt) + \dot\varphi_b\lt(\beta_b-\frac{1}{2}(\beta_c+\beta_a) \rt)
+ \dot\varphi_c\lt(\beta_c-\frac{1}{2}(\beta_a+\beta_b) \rt) \rt)
\nnr
&
\lt.
-\frac{J}{2}\sum_i(k^2\beta_i^2+(\nabla\varphi_i)^2)
-\frac{J}{4}\sum_{i\neq j} \lt(\nabla\beta_i-\nabla\beta_j\rt)^2
+ h \cos(\varphi_a+\varphi_b+\varphi_c) \rt] , \label{Lfree}\end{aligned}$$ where $h\equiv -\frac{9}{4}b{M_{\rm f}}{M_{\rm h}}(>0)$. The above expression was derived assuming slowly-varyingness. As for $\varphi_i$’s, we have not carried out a perturbative expansion to keep the periodic nature of the variable. In fact, $\varphi_i$’s form zero modes (in the absence of pinning) and they are not necessarily small in amplitude.
The Lagrangian, (\[Lfree\]), is the one in the representation using canonical coordinates and momenta. The time-derivative term of the Lagrangian is a product of canonical momentum, $p$, and time-derivative of canonical coordinate, $\dot{q}$. We thus see that the canonical momentum of $\varphi_a$ is a composite field, $g(\varphi_b-\varphi_c)\cos(\varphi_a+\varphi_b+\varphi_c)-\beta_a+\frac{1}{2}(\beta_b+\beta_c)$. The skyrmion lattice dynamics appears thus to be complicated if represented in terms of $\varphi_i$ and $\beta_i$. This means that these variables themselves are not good variables to describe low energy dynamics. Our next task is to find good variables, which we shall carry out in the next section.
Before proceeding further, let us look into the Lagrangian more closely. When $h$ is large, $\varphi_i$’s are constrained to satisfy $\cos(\varphi_a+\varphi_b+\varphi_c)\simeq1$, and the canonical relations are determined by the first term of Eq. (\[Lfree\]). (The second term connecting $\dot{\varphi}_i$ to $\beta_i$ turns out to be irrelevant in the low energy behavior as we shall see later.) For instance, the canonical momentum of ${\varphi}_a$, a shift of spiral phase in the $x$ direction, is $(\varphi_b-\varphi_c)$, a shift of phase in the $y$ direction. The motion of skyrmion lattice in the $x$ direction and $y$ direction are thus coupled in the same manner as a motion of charged particles in the presence of a magnetic field. This is due to the fact that the topological number (spin Berry’s phase) of skyrmion acts as an effective magnetic field. This feature of topological magnetic structures was noted in the case of magnetic vortices by Thiele [@Thiele73]. In contrast, when $h$ is small, $\varphi_i$’s may fluctuate independently, resulting in $\cos(\varphi_a+\varphi_b+\varphi_c)\sim0$ and the disappearance of the kinetic term proportional to $g$. The canonical momentum of $\varphi_a$ then reduces to $\lt(\beta_a-\frac{1}{2}(\beta_b+\beta_c) \rt)$ and the dynamics reduces to that of a phonon with a linear dispersion as we will show below. We therefore see that the parameter $h$ is essential in determining the skyrmion dynamics.
Excitations
===========
The Lagrangian we obtained, (\[Lfree\]), seems complicated as it is. We will demonstrate in this section that it becomes much simpler if we choose good dynamic variables. Finding good variables can be realized by looking into the form of the kinetic term. In the present case, from the terms proportional to $g$, we expect that differences such as $\varphi_a-\varphi_b$ would be a good variable. After some algebra, it is easy to see that another good variable is $(\varphi_a-\varphi_c)+(\varphi_b-\varphi_c)$. Let us thus define $$\begin{aligned}
\varphi_+ &\equiv \frac{1}{2}(\varphi_a+\varphi_b)-\varphi_c \nnr
\varphi_- &\equiv \frac{1}{2}(\varphi_a-\varphi_b).\end{aligned}$$ Note that both variables are written in terms of phase differences. The Berry’s phase term proportional to $g$ then is simplified as $$\begin{aligned}
\dot{\varphi}_a(\varphi_b-\varphi_c)+ \dot{\varphi}_b(\varphi_c-\varphi_a) + \dot{\varphi}_c(\varphi_a-\varphi_b)
&= 2( \varphi_+\dot{\varphi}_- - \varphi_- \dot{\varphi}_+).\end{aligned}$$ Similarly, the other Berry’s phase contribution connecting $\dot{\varphi}_i$ and $\beta_i$ is simplified by introducing $$\begin{aligned}
\beta_+ &\equiv \frac{1}{2}(\beta_a+\beta_b)-\beta_c \nnr
\beta_- &\equiv \frac{1}{2}(\beta_a-\beta_b),\end{aligned}$$ as $$\begin{aligned}
\dot\varphi_a\lt(\beta_a-\frac{1}{2}(\beta_b+\beta_c)\rt) +
\dot\varphi_b\lt(\beta_b-\frac{1}{2}(\beta_c+\beta_a)\rt) +
\dot\varphi_c\lt(\beta_c-\frac{1}{2}(\beta_a+\beta_b)\rt)
&= \beta_+ \dot{\varphi}_+ + 3 \beta_- \dot{\varphi}_-.\end{aligned}$$ Here we notice that, although we have originally introduced six variables, $\varphi_i$’s and $\beta_i$’s, there are only four independent dynamic variables, $\varphi_\pm$ and $\beta_{\pm}$, since the kinetic terms are expressed using only these four variables. Since two of the four are the canonical momenta, we now see that there are only two excitation modes in the present effective Lagrangian. This fact is natural since the excitations corresponding to a sliding of helix $\varphi_a$ necessarily induces the motion of the other two helices, since the the direction of the sliding of the three helices are not orthogonal. In other words, there are only two independent spin fluctuation propagation directions in two dimensions.
We have succeeded in finding the correct dynamic variables in the Lagrangian. There are, however, two other variables in the Lagrangian. Although not dynamic, these variables may have essential effects on the dynamics of $\varphi_\pm$ and $\beta_\pm$. Introducing two variables as $$\begin{aligned}
\varphi_0 & \equiv \varphi_a + \varphi_b + \varphi_c \nnr
\beta_0 & \equiv \beta_a + \beta_b +\beta_c ,\end{aligned}$$ we see that the Lagrangian is simplified. For instance, $$\begin{aligned}
\sum_i \beta_i^2 &= \frac{2}{3}(\beta_+^2+3\beta_-^2)+\frac{1}{3}\beta_0^2 \nnr
\sum_{\langle i,j\rangle}(\nabla\beta_i-\nabla\beta_j)^2
&= 2[(\nabla\beta_+)^2+3(\nabla\beta_-)^2 ].\end{aligned}$$ The Lagrangian thus reads $$\begin{aligned}
L &= {M_{\rm h}}^2 {\int\frac{d^2r}{a^2}}\lt[\frac{\hbar}{{M_{\rm h}}}\lt[ 2 g\cos\varphi_0( \varphi_+\dot{\varphi}_- - \varphi_- \dot{\varphi}_+)
-( \beta_+ \dot{\varphi}_+ + 3 \beta_- \dot{\varphi}_- ) \rt] \rt.\nnr
& -\frac{J}{3} \lt( (\nabla\varphi_+)^2+3(\nabla\varphi_-)^2 +\frac{1}{2}(\nabla\varphi_0)^2 \rt) \nnr
& \lt.
- \frac{J}{3}k^2 \lt( \beta_+^2+3\beta_-^2+\frac{1}{2}\beta_0^2 \rt)
- \frac{J}{2} [(\nabla\beta_+)^2+3(\nabla\beta_-)^2]
+h\cos\varphi_0 \rt]. \label{Lphibeta}\end{aligned}$$ We see that the Lagrangian is now expressed in terms of two pairs of two canonical variables, $\varphi_\pm$ and $\beta_\pm$, and two modes $\varphi_0$ and $\beta_0$. The meaning of $\varphi_\pm$ are as follows. The mode $\varphi_-$ describes the phase fluctuation (phason) along $\kv_-\equiv{\kv_{a}}-{\kv_{b}}=\frac{\sqrt{3}}{2}\lt(\sqrt{3},-1,0\rt)$, while the propagation direction of $\varphi_+$ is along $\kv_+\equiv{\kv_{a}}+{\kv_{b}}-2{\kv_{c}}=\frac{3}{2}\lt(1,\sqrt{3},0\rt)$ (Fig. \[FIGkabc\]). The two modes thus correspond to skyrmion propagation in the two orthogonal directions. A different factor of three in the Lagrangian is due to the difference of the magnitude of the wavelengths, $k_+$ and $k_-$. The mode represented by $\beta_\pm$ describes the deformation of skyrmion lattice into the configuration more like a single helix in the directions $\kv_\pm$, respectively (Fig. \[FIGbeta\]). The mode $\varphi_0$ affects the phason dynamics via the kinetic term (the first term of Eq. (\[Lphibeta\])). It is not a dynamic variable in the standard sense, since the Lagrangian does not have terms containing $\dot{\varphi}_0$. (We may rewrite the kinetic term by use of partial integration with respect to time and obtain higher-order kinetic terms like $\dot{\varphi}_0\varphi_+\varphi_- \sin\varphi_0 $, but non-linear canonical momenta obtained from such kinetic terms are neglected in standard treatments.) The mode $\beta_0$ is decoupled from other modes and is irrelevant.
![ Skyrmion lattice structures with $\beta_-=0$ (left) and $\beta_-=-0.5$ (right). It is seen that $\beta$ excitation corresponds to a shift of the core, i.e., a rotational mode. \[FIGbeta\]](skyrmion_beta00z.eps "fig:"){width="0.45\hsize"} ![ Skyrmion lattice structures with $\beta_-=0$ (left) and $\beta_-=-0.5$ (right). It is seen that $\beta$ excitation corresponds to a shift of the core, i.e., a rotational mode. \[FIGbeta\]](skyrmion_beta05z.eps "fig:"){width="0.45\hsize"}
From the kinetic term of (\[Lphibeta\]), we see that phason $\varphi_+$ has two canonical momenta, $\varphi_-$ and $\beta_+$, and canonical momenta for $\varphi_-$ are $\varphi_+$ and $\beta_-$. The first canonical relation, between $\varphi_+$ and $\varphi_-$, is induced by a topological effect, and indicates that for the motion of the phason mode $\varphi_+$, excitation of $\varphi_-$ mode is necessary. This behavior has been known in topological magnetic structures such as magnetic vortices [@Thiele73; @SNTKO06]. In fact, the center of mass of a vortex in two dimensions, $X$ and $Y$ are canonically conjugate to each other and thus when a force along $x$-axis is applied, motion in the $y$-direction is induced. The system is thus similar to that of charged particle under a magnetic field. The effective magnetic field is generated by topological number of the magnetic structure. The other canonical momentum for $\varphi_\pm$ is $\beta_\pm$. This feature is similar to the case of magnetic domain wall. In fact, for a domain wall to slide, a tilting of the wall plane is necessary [@Chikazumi97; @TK04; @TKS_PR08]. This is because the translational motion of a wall is induced when a torque rotating the magnetization within the wall plane is applied, and this torque needs to be generated if the magnetization component to the wall plane is induced. It is a unique feature of two-dimensional skyrmion lattice, described by two phason fields, that it has both properties of domain walls and vortices. As we saw from the derivation, the two canonical relations arise from the same spin Berry’s phase term in the Lagrangian. Spin Berry’s phase term in the Lagrangian is equivalent to imposing the commutation relation of the spin operators, and both relations are therefore different manifestations of the spin commutation relation.
Dispersion for the small amplitude case
---------------------------------------
Dispersion of the excitation is now easy to calculate. Let us first consider the dispersion in the case of small amplitude of $\varphi_i$’s, expanding $\cos\varphi_0$ to the second order in $\varphi_0$. The time-integral of the Lagrangian, i.e., the action, in the Fourier representation is $$\begin{aligned}
\int dt L = \int \frac{d\omega}{2\pi} \sum_{\qv}
{\bm \varphi}_{-\qv,-\omega}^{\rm t} {\cal L_\varphi} {\bm \varphi}_{\qv,\omega},
\label{Lphibeta2}\end{aligned}$$ where the basis is chosen as (${\rm t}$ stands for transpose) $$\begin{aligned}
{\bm \varphi}_{\qv,\omega}\equiv ( \varphi_+(\qv,\omega),\varphi_-(\qv,\omega),\beta_+(\qv,\omega),\beta_-(\qv,\omega),\varphi_0(\qv,\omega),\beta_0(\qv,\omega))^{\rm t},\end{aligned}$$ and the $6\times6$ matrix is $$\begin{aligned}
{\cal L_\varphi} \equiv
{M_{\rm h}}^2 \lt(
\begin{array}{cccccc}
\frac{-J}{3}q^2 & - \frac{ 2ig }{{M_{\rm h}}}\hbar \omega & - \frac{i}{2{M_{\rm h}}}\hbar \omega & 0 & 0 & 0 \\
\frac{ 2ig}{{M_{\rm h}}} \hbar \omega & -J q^2 & 0 & -\frac{3i}{2{M_{\rm h}}}\hbar \omega & 0 & 0 \\
\frac{i}{2{M_{\rm h}}}\hbar \omega & 0 & \frac{-J}{3}\left(k^2+\frac{3}{2}q^2\right) & 0 & 0 & 0 \\
0 & \frac{3i}{2{M_{\rm h}}}\hbar \omega & 0 & -J \left(k^2+\frac{3}{2}q^2\right) & 0 & 0 \\
0 & 0 & 0 & 0 & \frac{-J}{6}q^2-\frac{h}{2} & 0 \\
0 & 0 & 0 & 0 & 0 & \frac{-J}{6}k^2 \\
\end{array}
\right). \label{Lphibeta2matrix}\end{aligned}$$ In the matrix representation, it is clear that $\beta_0$ is orthogonal to other modes and is irrelevant. The mode $\varphi_0$ is also decoupled in the small amplitude case, but it has an important role of renormalizing the topological number, $g$, and $h$ via the factor of $\cos\varphi_0$. We shall study the effect of the renormalization in Sect. \[SECrenorm\] and focus here on the small amplitude case. The determinant of the matrix is (dropping an irrelevant constant) $$\begin{aligned}
(Jq^2+3h) [\omega^2 - (\omega_q^+)^2 ][\omega^2 - (\omega_q^-)^2 ], \label{determinant_1}\end{aligned}$$ where $$\begin{aligned}
\hbar\omega_q^\pm & \equiv
\frac{4\sqrt{2}}{3\sqrt{3}}gJ{M_{\rm h}}k^2 {\mu_q}^{\frac{1}{2}}
\left({\mu_q}+\frac{3q^2}{8k^2g^2}\right)^{\frac{1}{2}}
\lt[ 1
\pm \frac{ {\mu_q}^{\frac{1}{2}} \lt( {\mu_q}+\frac{3q^2}{4k^2g^2} \rt)^{\frac{1}{2}}} { {\mu_q}+\frac{3q^2}{8k^2g^2} }
\rt]^{\frac{1}{2}} ,
\label{dispersion}\end{aligned}$$ where ${\mu_q}\equiv \left(1+\frac{3q^2}{2k^2}\right)$.
![ Left: Plot of dispersion of excitations, $\omega_q$ in unit of $J{M_{\rm h}}k^2/\hbar$ as function of $\frac{q}{k}$ for $g=1, 0.5,0$. Right: Plot of dispersion of excitations for small $g$ ($g=0.1,0$). The crossover to linear dispersion occurs at $\frac{q}{k}\sim g$. \[FIGdispersion\]](omegalargeg.eps "fig:"){width="0.45\hsize"} ![ Left: Plot of dispersion of excitations, $\omega_q$ in unit of $J{M_{\rm h}}k^2/\hbar$ as function of $\frac{q}{k}$ for $g=1, 0.5,0$. Right: Plot of dispersion of excitations for small $g$ ($g=0.1,0$). The crossover to linear dispersion occurs at $\frac{q}{k}\sim g$. \[FIGdispersion\]](omegasmallg.eps "fig:"){width="0.45\hsize"}
If $g=0$, Eq. (\[dispersion\]) leads to $$\begin{aligned}
\hbar \omega_q = \frac{2}{3}J{M_{\rm h}}kq\sqrt{1+\frac{3q^2}{2k^2}} \equiv \omega_q^{\rm L}.
\label{omegaL}\end{aligned}$$ The degenerate linear dispersion indicates the existence of two vibration modes of spins in two-dimensions similar to the phonon system. Once $g\neq0$, there is a gapless mode with dispersion $\omega\propto q^2$ and a massive mode as $q\ra0$ (Fig. \[FIGdispersion\]) as seen from Eq. (\[dispersion\]) expanded to the order of $q^2$, $$\begin{aligned}
\hbar\omega_q^\mp&= \lt\{
\begin{array}{c}
\frac{J{M_{\rm h}}}{2\sqrt{3}g}q^2 \\
\frac{8gJ{M_{\rm h}}k^2}{3\sqrt{3}} +\frac{J{M_{\rm h}}(1+8g^2)}{2\sqrt{3}g}q^2.
\end{array} \rt. \label{dispersion_q2}\end{aligned}$$ The singular behavior of the prefactors of $q^2$ terms at $g\ra0$ is a signature of crossover from a quadratic dispersion to a linear one. Effects of finite $g$ are characterized by the parameter $q/(kg)$. When $g$ is finite, $\varphi_+$ and $\varphi_-$ starts to form canonical conjugates due to the topological nature of skyrmions. In other words, the effective magnetic field arising from the spin Berry’s phase induces the Lorentz force on the $\varphi_\pm$ modes and mixes the two modes, resulting in the softening of one of the linear modes and in the formation of a massive mode with a gap proportional to $g$.
The mechanism of softening in the presence of magnetic field is the same as the one in the case of the Wigner crystal discussed in Refs. [@Chaplik72; @Fukuyama75; @Fukuyama78a], but the quadratic dispersion is distinct from the case of the Wigner crystal having the $q^{\frac{3}{2}}$ dispersion, because of the neutral charge of the phasons in the present case. Another unique feature of the skyrmion lattice is that the effect of $g$ is non-perturvative in the sense that the high energy behavior at $q$ larger than the crossover value ($\sim kg$) is also affected by $g$; the linear dispersions at $g\ll1$ and $q\gtrsim kg$, $$\begin{aligned}
\hbar\omega_q^\mp&=
\frac{2}{3}{J{M_{\rm h}}}kq\lt(1\pm \frac{2}{\sqrt{3}}\frac{kg}{q} +O\lt(\lt(\frac{kg}{q}\rt)\rt)^2\rt) ,\end{aligned}$$ do not merge to a single line for small but finite $g$ as seen in Fig. \[FIGdispersion\], in contrast to the case of the Wigner crystal.
Renormalization of the topological term \[SECrenorm\]
=====================================================
As seen in the Lagrangian, (\[Lphibeta\]), the $\varphi_0$ mode has an important effect on the dynamics of $\varphi_\pm$ by renormalizing the topological ($g$-)term. To take account of this renormalization effect, we treat $\varphi_0$ beyond the second order expansion scheme by use of self-consistent harmonic approximation [@Dashen74; @Nakano80]. This approximation takes account of the renormalization effect by use of the expectation value of $\varphi_0^2$, $\overline{\varphi_0^2}$ and take account of dynamic fluctuation as small variable. We thus rewrite $\cos\varphi_0$ as $$\begin{aligned}
\cos({\varphi_0}) &= 1-\frac{1}{2}\varphi_0^2 +\cdots \nnr
&= 1-\frac{1}{2} \overline{\varphi_0^2}
-\frac{1}{2}(\varphi_0^2 - \overline{\varphi_0^2} ) +\cdots \nnr
&\simeq f
\lt(1-\frac{1}{2}(\varphi_0^2 - \overline{\varphi_0^2})
\rt),\end{aligned}$$ where $$\begin{aligned}
f\equiv e^{-\frac{1}{2} \overline{\varphi_0^2} }.\end{aligned}$$ The topological term then becomes $2\hbar fg( \varphi_+\dot{\varphi}_- - \varphi_- \dot{\varphi}_+)$, and the Lagrangian for $\varphi_0$ is $$\begin{aligned}
L_{{\varphi_0}}
& \equiv {\int\frac{d^2r}{a^2}}\frac{1}{2}\lt[ -\frac{J}{3}(\nabla \varphi_0)^2
- hf \varphi_0^2
\rt]. \label{Lvarphi0}\end{aligned}$$ The average $\overline{\varphi_0^2}$ at the temperature $T$ is calculated as (see Appendix \[SECphiint\]) $$\begin{aligned}
\overline{{\varphi_0}^2}
&= 6\kb T \sumqv \frac{1}{{J}q^2+3hf},
\label{averagephi2}\end{aligned}$$ where $\kb$ is the Boltzmann constant. We carry out the integration over $\qv$ in two-dimensions, introducing a large wavelength cutoff of $k$, as $$\begin{aligned}
\sum_{\qv}\frac{1}{Jq^2+3hf }
&= \frac{1}{4\pi}\int_0^{k^2} \frac{dq^2}{{J}q^2+3hf} \nnr
&= \frac{1}{4\pi J}\ln\lt(1+\frac{Jk^2}{3hf}\rt).\end{aligned}$$ The self-consistency condition, Eq. (\[averagephi2\]), thus reads $$\begin{aligned}
\overline{{\varphi_0}^2}
&=\frac{3\kb T}{2\pi J}\ln\lt(1+\frac{Jk^2}{3h e^{-\frac{\overline{{\varphi_0}^2}}{2}}}\rt)
\label{sceq1}\end{aligned}$$ or $$\begin{aligned}
f
&=\lt(1+\frac{Jk^2}{3hf}\rt)^{ -\frac{3\kb T}{4\pi J} }.
\label{schaf}\end{aligned}$$ Near $T=0$, $$\begin{aligned}
f
&=1- \frac{3\kb T}{4\pi J}\ln\lt(1+\frac{Jk^2}{3h}\rt)+O(T^2).\end{aligned}$$ The solutions of $\overline{{\varphi_0}^2}$ and $f$ are plotted in Figs. \[FIGphi\_hT\] and \[FIGf\_hT\].
![ Left: Plot of $\overline{\varphi_0^2}$ as function of scaled magnetic field, $h$, for $T=0.1$, $T=0.2$ and $T=0.5$. Right: Plot of $\overline{\varphi_0^2}$ as function of the temperature, $T$, for $h=0.1$, $h=0.2$ and $h=0.5$. Temperature is in unit of $J/\kb$ and $h$ is in unit of $Jk^2$. \[FIGphi\_hT\]](phi2h.eps "fig:"){width="0.4\hsize"} ![ Left: Plot of $\overline{\varphi_0^2}$ as function of scaled magnetic field, $h$, for $T=0.1$, $T=0.2$ and $T=0.5$. Right: Plot of $\overline{\varphi_0^2}$ as function of the temperature, $T$, for $h=0.1$, $h=0.2$ and $h=0.5$. Temperature is in unit of $J/\kb$ and $h$ is in unit of $Jk^2$. \[FIGphi\_hT\]](phi2T.eps "fig:"){width="0.4\hsize"}
![ Left: Plot of $f=e^{-\frac{\overline{\varphi_0^2}}{2}}={\overline{g}}/g$ as function of $h$ for $T=0.1$, $T=0.2$ and $T=0.5$. Right: Plot of $f$ as function of $T$ for $h=0.1$, $h=0.2$ and $h=0.5$. \[FIGf\_hT\]](fh.eps "fig:"){width="0.4\hsize"} ![ Left: Plot of $f=e^{-\frac{\overline{\varphi_0^2}}{2}}={\overline{g}}/g$ as function of $h$ for $T=0.1$, $T=0.2$ and $T=0.5$. Right: Plot of $f$ as function of $T$ for $h=0.1$, $h=0.2$ and $h=0.5$. \[FIGf\_hT\]](fT.eps "fig:"){width="0.4\hsize"}
The Lagrangian after self-consistent harmonic approximation reads $$\begin{aligned}
L &= {M_{\rm h}}^2 {\int\frac{d^2r}{a^2}}\lt[\frac{\hbar}{{M_{\rm h}}}\lt[ 2 {\overline{g}}( \varphi_+\dot{\varphi}_- - \varphi_- \dot{\varphi}_+)
-( \beta_+ \dot{\varphi}_+ + 3 \beta_- \dot{\varphi}_- ) \rt] \rt.\nnr
& \lt.
-\frac{J}{3} \lt( (\nabla\varphi_+)^2+3(\nabla\varphi_-)^2 \rt)
- \frac{J}{3} \lt( k^2 ( \beta_+^2+3\beta_-^2) \rt)
-\frac{J}{2} [(\nabla\beta_+)^2+3(\nabla\beta_-)^2]
\rt], \label{Lphibetascha}\end{aligned}$$ where ${\overline{g}}\equiv fg$ is the effective topological amplitude and $f$ is a solution of Eq. (\[schaf\]). The excitation energy including the renormalization effect is obtained by replacing $g$ in Eqs. (\[dispersion\])(\[dispersion\_q2\]) by ${\overline{g}}$.
Cubic term and stability of skyrmion lattice
--------------------------------------------
In the present scheme based on a description using three helices, the “cubic” term (Eq. (\[cubicterm\]) and the term proportional to $h$ in Eq. (\[Lphibeta\])) is essential for stabilization of the skyrmion lattice. In fact, if $h=0$, the amplitude of $\varphi_0$ diverges in two-dimensions according to Eq. (\[averagephi2\]), and the topological nature of the skyrmion lattice (represented by the $g$-term) is smeared out. Any interaction Hamiltonian containing the magnetization to the second order does not give rise to such interaction connecting the variables for the three helices in the slowly-varying case. These results are consistent with the observation by Mühlbauer et al. [@Muhlbauer09]. On the other hand, skyrmion lattice is realized in numerical simulations using the Landau-Lifshitz-Gilbert equation [@Mochizuki12; @Ohe13], and it appears that skyrmion lattice arises without the “cubic” term. We speculate that the relaxation process used to obtain thermalized state numerically is essential for the appearance of the skyrmion phase. In fact, thermalization corresponds to the imaginary-time development, which is equivalent to taking account of the Ginzburg-Landau free energy, and thus the cubic term may be effectively included in the simulation.
Effective Lagrangian and equations of motion
============================================
As we have noted, $\beta_\pm$ in the Lagrangian (\[Lphibetascha\]) play roles of canonical momenta for $\varphi_\pm$. By integrating out the two $\beta_\pm$ modes, we can write the Lagrangian in terms of $\varphi_\pm$ only. Integrals are easily calculated resulting in a contribution to the Lagrangian (see Appendix \[SECbetaint\] for derivation) $$\begin{aligned}
\delta L_\varphi ={\int\frac{d^2r}{a^2}}\frac{m_\varphi}{2}(\dot{\varphi}_+^2 +3 \dot{\varphi}_-^2), \label{betaintL}\end{aligned}$$ where $$\begin{aligned}
m_\varphi\equiv \frac{3\hbar^2}{2Jk^2},\end{aligned}$$ is the mass for $\dot{\varphi}_+$. The Lagrangian written in terms of $\varphi_\pm$ is thus $$\begin{aligned}
L &= {\int\frac{d^2r}{a^2}}\lt[2{\hbar}{M_{\rm h}}\overline{g} ( \varphi_+\dot{\varphi}_- - \varphi_- \dot{\varphi}_+) +\frac{m_\varphi}{2}(\dot{\varphi}_+^2 +3 \dot{\varphi}_-^2)
-\frac{J{M_{\rm h}}^2}{3} \lt( (\nabla\varphi_+)^2+3(\nabla\varphi_-)^2 \rt)
\rt], \label{Lphi2}\end{aligned}$$ The Lagrangian (\[Lphi2\]) indicates that we have two modes with different mass and exchange constant and coupled by a kinetic term first order in time-derivative. The factor of three difference is due to the difference of the magnitudes of the propagation vectors, ${\kv_{a}}+{\kv_{b}}-2{\kv_{c}}=-3{\kv_{c}}$ and ${\kv_{a}}-{\kv_{b}}=\sqrt{3} \hat{\zv}\times{\kv_{c}}$. In the Fourier representation, Eq. (\[Lphi2\]) is $$\begin{aligned}
L=\sumom\sumqv (\varphi_+,\varphi_-)_{-\qv,-\omega}
\lt(\begin{array}{cc}
\frac{J{M_{\rm h}}^2 q^2}{3}-\frac{m_\varphi \omega ^2}{2} & -2 i\hbar {M_{\rm h}}\overline{g} \omega \\
2 i\hbar {M_{\rm h}}\overline{g} \omega & 3 \left(\frac{J{M_{\rm h}}^2 q^2}{3}-\frac{m_\varphi \omega
^2}{2}\right) \end{array}
\rt)
\lt( \begin{array}{c} \varphi_+ \\ \varphi_-\end{array} \rt)_{\qv,\omega}
\label{matrixphi2}\end{aligned}$$ The energy dispersion determined from the determinant is $$\begin{aligned}
\hbar \omega_q^\pm &= \frac{2\sqrt{2}{M_{\rm h}}\hbar^2 {\overline{g}}}{\sqrt{3} m_\varphi}
\lt[
1+\frac{m_\varphi Jq^2}{4(\hbar {\overline{g}})^2} \pm \sqrt{ 1+\frac{m_\varphi Jq^2}{2(\hbar {\overline{g}})^2} } \rt]^{\frac{1}{2}}. \label{dispersionphi}\end{aligned}$$ We thus obtain $$\begin{aligned}
\hbar \omega_q^\pm &= \lt\{
\begin{array}{c} \frac{J{M_{\rm h}}}{2\sqrt{3}{\overline{g}}}q^2 +O(q^4) \\
\frac{8{\overline{g}}J{M_{\rm h}}k^2}{3\sqrt{3}}+ \frac{J{M_{\rm h}}}{2\sqrt{3}{\overline{g}}}q^2 .
\end{array} \rt.\end{aligned}$$ We see that the prefactor of the $q^2$ term of the massive mode is different from the one obtained in Eq. (\[dispersion\_q2\]). The discrepancy is due to the fact that we have neglected the contribution containing $q^2\partial_t^2$ in deriving Eq. (\[betaintL\]). We shall show in Appendix \[SECbetaint\] that the correct dispersion is reproduced even after the integration if we include the $q^2\partial_t^2$ contribution in $\delta L_\varphi$. The effective Lagrangian of Eq. (\[Lphi2\]) is of the form commonly used to discuss the low energy excitations, but it does not always describe the correct dispersion to the order of $q^2$.
Let us derive the equations of motion for $\varphi_\pm$. In discussing the magnetization dynamics, dissipation of angular momentum, described by the Gilbert damping parameter ${\alpha_{\rm s}}$, is essential [@Chikazumi97]. The Gilbert damping cannot be expressed in terms of a Lagrangian, but is represented by dissipation function [@TKS_PR08], $$\begin{aligned}
W\equiv \frac{\hbar{\alpha_{\rm s}}}{2} {\int\frac{d^2r}{a^2}}\dot{\Mv}^2,\end{aligned}$$ and considering the equation of motion defined as $\displaystyle \frac{d}{dt}\frac{\delta L}{\delta \dot{q}_i}-\frac{\delta L}{\delta q_i}=-\frac{\delta W}{\delta \dot{q}_i}$, where $q_i$ represent generalized coordinates. Dissipation function for the skyrmion lattice reads $$\begin{aligned}
W = \hbar{\alpha_{\rm s}}{\int\frac{d^2r}{a^2}}\lt[ \frac{1}{2} [(\dot{\beta}_+)^2 + 3(\dot{\beta}_-)^2 ]
+ \frac{1}{3} [(\dot{\varphi}_+)^2 + 3(\dot{\varphi}_-)^2 ]
+ \frac{1}{6} (\dot{\varphi}_0)^2 \rt].
\label{dissipationfunc}\end{aligned}$$ The equation of motion including the Gilbert damping is $$\begin{aligned}
-2\hbar {M_{\rm h}}{\overline{g}}\dot{\varphi}_+ -3m_\varphi \ddot{\varphi}_- -2\hbar{\alpha_{\rm s}}\dot{\varphi}_-
+2J{M_{\rm h}}^2\nabla^2\varphi_- &=0 \nnr
2\hbar {M_{\rm h}}{\overline{g}}\dot{\varphi}_- -m_\varphi \ddot{\varphi}_+ -\frac{2}{3}\hbar{\alpha_{\rm s}}\dot{\varphi}_+
+\frac{2}{3}J{M_{\rm h}}^2\nabla^2\varphi_+ &=0 .
\label{eqofmo0}\end{aligned}$$
Microwave response
==================
Let us discuss the microwave response of skyrmion lattice based on our low energy Lagrangian. We consider two cases with the AC magnetic field, $B_{\rm ac}$, with the angular frequency of $\omega$ applied parallel and perpendicular to the $xy$ plane, respectively. The applied magnetic field, $\Bv_{\rm ac}$, couples to the magnetization as $$\begin{aligned}
H_{B} = -\frac{g\mub}{a^3}{\int\frac{d^2r}{a^2}}\Bv_{\rm ac}\cdot\Mv,\end{aligned}$$ where the magnetization is expressed in terms of collective coordinates in Eqs. (\[skxconfiguration\])(\[Mvi\]).
In-plane field ($B_{\rm ac}\perp \zv$)
--------------------------------------
An in-plane AC field, $B_{\rm ac}^{\parallel}$, couples to the excitation as $$\begin{aligned}
H_{B} = -\frac{g\mub}{2a^3}{\int\frac{d^2r}{a^2}}B_{\rm ac}^{\parallel}
\lt[ (\beta_++3\beta_-)\cos\phi_B+ \sqrt{3}(\beta_+-\beta_-)\sin\phi_B \rt],\end{aligned}$$ where $\phi_B$ is the angle representing the field direction in the $xy$-plane. Note that the in-plane field does not directly excite $\varphi_\pm$ modes because of rapidly oscillating components of magnetization.
A response to a microwave is described by the correlation function, given by the inverse of the Lagrangian (\[Lphibetascha\]). The diagonal component of the correlation function for $\beta_\pm$, $\chi_{\beta_\pm}$ reads (including an imaginary part arising from the spin damping by $\hbar\omega\rightarrow\hbar\omega+i\eta$) $$\begin{aligned}
\chi_{\beta_+} &=\frac{\frac{12}{81}
\lt[3J{M_{\rm h}}q^2[(\hbar\omega)^2-(\hbar\omega_q^{\rm L})^2]+16(\hbar\omega)^2 J{M_{\rm h}}k^2{\overline{g}}^2{\mu_q}\rt]}
{[(\hbar\omega+i\eta)^2-(\hbar\omega_q^{+})^2] [(\hbar\omega+i\eta)^2-(\hbar\omega_q^{-})^2] },\end{aligned}$$ and $ \chi_{\beta_-}=\frac{1}{3} \chi_{\beta_+}$, where $\omega_q^{\pm}$ are given by Eq. (\[dispersion\]) with $g$ replaced by ${\overline{g}}$. The modes $\varphi_\pm$ are canonical conjugate of $\beta_\pm$ and thus the correlation functions for $\varphi_\pm$ have the same poles. For $q=0$, the response to the in-plane AC field thus has a peak at $\omega=\omega_{q=0}^{+}=J{M_{\rm h}}k^2f={M_{\rm h}}f \frac{D^2}{J}$ as well as at $\omega=0$. For $J/a^2=1 $ meV and $D/a=0.18$ meV [@Iwasaki13] the resonant frequency is $\nu\equiv \frac{\omega}{2\pi}=1.2$ GHz if ${M_{\rm h}}=1$ and $f=1$. The peak shifts to lower frequency when renormalization becomes larger (smaller $f$) by reducing the perpendicular magnetic field. The excitation mode described by $\beta_\pm$ corresponds to a shift of the skyrmion core breaking the rotationally symmetric skyrmion structure, and it can be thus identified with the rotational mode. Our analysis indicates that there is only one excitation energy for the rotation. In contrast, in numerical simulation by Mochizuki [@Mochizuki12], clockwise and counterclockwise two rotational modes having different frequencies were found. The origin of discrepancy is not clear at present, but the boundary effect in the simulation may affect.
Experimentally, Onose et al. [@Onose12] observed an excitation of skyrmion lattice in Cu$_2$OSeO$_3$ by microwave absorption, and found an absorption peak around 1 GHz when the AC magnetic field is within the plane of the skyrmion lattice. It was also found that the peak shifts to higher frequency when the external field in the perpendicular direction increases. Those observed behaviors seem consistent with the excitation of the present collective coordinate scenario, although our analysis is in two-dimensions while the experiment is in three-dimensional systems.
$B_{\rm ac}\parallel \zv$
-------------------------
When AC field is applied along the $z$ axis, our Lagrangian, Eq. (\[Lphibeta2matrix\]), indicates that the mode excited are $\varphi_0$ and ${M_{\rm f}}$, the uniform component of the magnetization. As we have seen, the mode $\varphi_0$ has no dynamic term in the Lagrangian, and thus its excitation occurs only at $\omega=0$. In contrast, the fluctuation of ${M_{\rm f}}$ results in a peak at finite frequency. The mode ${M_{\rm f}}$ corresponds to an expansion and contraction of the core, and is identified with breathing mode pointed out by Mochizuki [@Mochizuki12]. The existence of a single excitation mode for $B_{\rm ac}\parallel \zv$ is consistent with the experimental result of Onose et al. [@Onose12].
To analyze the response of ${M_{\rm f}}$ is beyond the scope of the present paper, since we have treated ${M_{\rm f}}$ as given. It can be carried out by minimizing the mean-field energy (See Appendix \[SECGL\]) and by treating ${M_{\rm f}}$ as a dynamic variable.
Pinning model
=============
We consider a case of pinning due to random impurities which modifies the local anisotropy energy. The interaction we consider is $$\begin{aligned}
H_{{\rm pin(z)}} &= - {\int\frac{d^2r}{a^2}}v_{\rm p} \sum_{\Rv_i} \delta(\rv-\Rv_i) (M_z)^2 ,\label{Hpinz0}\end{aligned}$$ where $\Rv_i$ is the position of random impurities and $v_{\rm p}$ is the pinning strength.
We here try to derive the effective pinning potential for $\varphi_i$’s following Fukuyama and Lee [@Fukuyama78; @FukuyamaJPSJ78]. We consider the case of modification of local anisotropy energy and write the integration over $\rv$ as a integral over a finite domain with area $S_0$ and summation over the domains (labeled by $\Rv$) as $$\begin{aligned}
H_{{\rm pin(z)}} &\simeq - v_{\rm p} \sum_{\Rv} \sum_{\Rv_i}
\lt( \sum_{j=a,b,c}\cos^2(\kv_j\cdot\Rv_i+\varphi_j(\Rv_i)) \rt. \nnr
&\lt. + \sum_{j,k=a,b,c}2\cos(\kv_j\cdot\Rv_i+\varphi_j(\Rv_i))\cos(\kv_k\cdot\Rv_i+\varphi_k(\Rv_i)) \rt)\nnr
&= - \frac{v_{\rm p}}{4} \sum_{\Rv} \sum_{\Rv_i}
\lt( \sum_{j=a,b,c}e^{i(2\kv_j\cdot\Rv_i+2\varphi_j(\Rv_i))}
+ \sum_{j,k=a,b,c}2e^{i((\kv_j+\kv_k)\cdot\Rv_i+\varphi_j(\Rv_i)+\varphi_k(\Rv_i))} +{\rm c.c.} \rt).\end{aligned}$$ The area $S_0$ is defined from the length scale where variation of $\varphi_i$ is smaller than unity. Then the summation over impurity positions, $\Rv_i$, is carried out in each domain noting the fact that $\varphi_i$’s are treated as constant in the domain and that the summation over impurity positions results in a random walk as $$\begin{aligned}
\sum_{\Rv_i} e^{i(2\kv_j\cdot\Rv_i+2\varphi_j(\Rv_i))} & \simeq
\sqrt{N_{\rm i}} e^{i (2\varphi_j-\xi_j)} \nnr
\sum_{\Rv_i} e^{i((\kv_j+\kv_k)\cdot\Rv_i+\varphi_j(\Rv_i)+\varphi_k(\Rv_i))}
& \simeq
\sqrt{N_{\rm i}} e^{i ((\varphi_j+\varphi_k)-\xi_{jk})} ,\end{aligned}$$ where $N_{\rm i}$ is the number of impurities in a domain and $\xi_j$ and $\xi_{jk}$ represent a phase arising from random walk [@FukuyamaJPSJ78]. The random potential (\[Hpinz0\]) thus results in $$\begin{aligned}
H_{{\rm pin(z)}} &\simeq
- \frac{v_{\rm p}}{4} \frac{\sqrt{N_{\rm i}}}{S_0} \int{d^2\Rv} \sum_{j=a,b,c}
\lt[\cos(2\varphi_j-\xi_j)+\sum_{k=a,b,c} 2 \cos ((\varphi_j+\varphi_k)-\xi_{jk})\rt].\end{aligned}$$ Since we have six random phase factors, $\xi_j$ and $\xi_{jk}$, we cannot derive explicit expressions for $\varphi_i$’s which minimize the pinning energy. This fact is in contrast to the case of charge density waves with simply a single random phase, and we cannot therefore proceed further in the same manner as in Ref. [@Fukuyama78; @Lee79]. Nevertheless, we may assume that the effective pinning potential of harmonic shape is a good approximation if one regards $\varphi_\pm$ as the deviations of the phase variables from the local equilibrium configuration determined by the random pinning potential. We thus consider an effective pinning potential $$\begin{aligned}
H_{{\rm pin}} &= {\int\frac{d^2r}{a^2}}\lt(\frac{K_+}{2}\varphi_+ ^2 + \frac{K_-}{2}\varphi_- ^2 \rt),\label{Hpin0}\end{aligned}$$ where $K_\pm$ represent the strength of pinning of $\varphi_\pm$, respectively, which are treated as phenomenological parameters. We assume further that the periodicity with respect to $\varphi_\pm$ persists in the pinning potential, namely, we assume $$\begin{aligned}
H_{{\rm pin}} &= - {\int\frac{d^2r}{a^2}}\lt(K_+ \cos\varphi_+ + K_-\cos\varphi_-\rt).\label{Hpin}\end{aligned}$$ We expect from the symmetry argument that $K_+=\frac{1}{3}K_-\equiv K_{\rm p}$, and we consider this case below. Note that variables $\varphi_\pm$ here represent modifications of phase variables with respect to the equilibrium configuration determined taking account of random pinning potential. The Lagrangian for skyrmion lattice including the pinning effect is $$\begin{aligned}
L &= {\int\frac{d^2r}{a^2}}\lt[ -2\hbar {M_{\rm h}}{\overline{g}}(\dot{\varphi}_+ \varphi_- - \dot{\varphi}_- \varphi_+)
+ \frac{m_\varphi}{2}(\dot{\varphi}_+^2 + 3 \dot\varphi_-^2)
- \frac{J{M_{\rm h}}^2}{3}[(\nabla\varphi_+)^2+3(\nabla\varphi_-)^2] \rt. \nnr
& \lt.
+K_{\rm p}\lt( \cos\varphi_+ + 3\cos\varphi_-\rt)\rt]. \label{Lphipin}\end{aligned}$$ The excitation energy of skyrmion is modified by the pinning. Considering a small oscillation in Eq. (\[Lphipin\]), we obtain the dispersion relation $$\begin{aligned}
\omega_q^\pm =\frac{1}{\sqrt{3}m_\varphi}
\lt[ \sqrt{ (\hbar {M_{\rm h}}{\overline{g}})^2+4m_\varphi\lt(K_{\rm p}+\frac{2}{3}J{M_{\rm h}}^2 q^2\rt) } \pm (\hbar {M_{\rm h}}{\overline{g}}) \rt].\end{aligned}$$ Both of the two excitations thus become massive when pinning is present, $$\begin{aligned}
\omega_q^\pm = \frac{ \hbar {M_{\rm h}}{\overline{g}}}{\sqrt{3}m_\varphi}
\lt[ \sqrt{ 1+\frac{3m_\varphi K_{\rm p}}{ (\hbar {M_{\rm h}}{\overline{g}})^2 } }\pm 1 \rt]+O(q^2).\end{aligned}$$ The phenomenological parameter of the pinning strength, $K_{\rm p}$, is thus accessible experimentally by observing the excitation energy by microwave absorption.
Spin transfer effect in metals
===============================
So far we have considered a general case applicable to both metals and insulators. In this section, we consider the metallic case and discuss the dynamics induced by an applied electric current. In ferromagnetic metals, spin textures are usually slowly-varying compared to the conduction electron wavelength and the $sd$ exchange interaction coupling the conduction electron spin and localized moment is strong. In this case, the conduction electron spin going through the magnetization texture is rotated to align along the local magnetization direction. This spin rotation exerts a torque on the magnetization called the spin-transfer torque [@Berger86; @Slonczewski96], and this torque is the main driving force for the slowly-varying structures.
The spin-transfer effect is represented by the Hamiltonian $$\begin{aligned}
H_{\rm ST} &\equiv \hbar{M_{\rm h}}\tilde{P} {\int\frac{d^2r}{a^2}}(\cos\theta-1) (\jv\cdot\nabla){\phi} ,\end{aligned}$$ where $\tilde{P}\equiv\frac{Pa^2}{2e{M_{\rm h}}}$, $P$ is the spin polarization of the current, and $\jv$ represents the current density [@TKS_PR08]. In the same manner as the spin Berry’s phase term in Sec. \[SECsbp\], we obtain an equivalent form suitable for studying the collective dynamics as $$\begin{aligned}
\delta H_{\rm ST} &= \hbar{M_{\rm h}}\tilde{P} {\int\frac{d^2r}{a^2}}\nv\cdot((\jv\cdot\nabla){\nv}\times\delta\nv).\end{aligned}$$ By use of $$\begin{aligned}
\nabla_\mu {\nv} &= \sum_i\lt[ (\kv_i)_\mu (\hat\kv_i\times\nv_i)+\nabla_\mu {\beta}_i \hat \kv_i \rt],\end{aligned}$$ where $\mu$ denotes spatial direction, we obtain $$\begin{aligned}
\delta H_{\rm ST} &=
\hbar{M_{\rm h}}\tilde{P}g {\int\frac{d^2r}{a^2}}\lt[
((\jv\cdot\kv_a) \varphi_b- (\jv\cdot\kv_b)\varphi_a )
+( (\jv\cdot\kv_b)\varphi_c- (\jv\cdot\kv_c) \varphi_b )
+( (\jv\cdot\kv_c) \varphi_a- (\jv\cdot\kv_a) \varphi_c )\rt] \nnr
& \times
\cos (\varphi_a+\varphi_b+\varphi_c)
\nnr
&+
\tilde{P} {\int\frac{d^2r}{a^2}}\lt[
(\jv\cdot\nabla{\beta}_a) \lt(\varphi_a- \frac{1}{2}\lt(\varphi_b + \varphi_c\rt)\rt)
+
(\jv\cdot\nabla{\beta}_b) \lt(\varphi_b- \frac{1}{2}\lt(\varphi_c + \varphi_a\rt)\rt)
\rt. \nnr
& \lt. +
(\jv\cdot\nabla{\beta}_c) \lt(\varphi_c- \frac{1}{2}\lt(\varphi_a + \varphi_b\rt)\rt)
\rt].\end{aligned}$$ The dominant terms are those including only $\varphi_i$’s, since $\beta$-modes have mass gap. Using the self-consistent harmonic approximation and in terms of $\varphi_\pm$, they read $$\begin{aligned}
\delta H_{\rm ST} &= \hbar{M_{\rm h}}\tilde{P} {\int\frac{d^2r}{a^2}}2{\overline{g}}(v_-\varphi_+ - v_+\varphi_- ),\end{aligned}$$ where $$\begin{aligned}
v_- & \equiv \frac{\tilde{P} }{2}\jv\cdot({\kv_{a}}-{\kv_{b}})=\tilde{P} \frac{\sqrt{3}}{2}\jv\cdot(\hat{\zv}\times{\kv_{c}}) \nnr
v_+ & \equiv \frac{\tilde{P} }{2} \jv\cdot({\kv_{a}}+{\kv_{b}}-2{\kv_{c}})=-\tilde{P} \frac{3}{2} \jv\cdot{\kv_{c}},\end{aligned}$$ are driving speed of $\varphi_\pm$. We see that the spin-transfer effect is renormalized by the fluctuation of $\varphi_0$ by a factor of $f$.
The total Lagrangian for a skyrmion lattice in low energy region in metals including the pinning and the spin-transfer effect is $$\begin{aligned}
L &= {\int\frac{d^2r}{a^2}}\lt[ -2\hbar {M_{\rm h}}{\overline{g}}(\dot{\varphi}_+ \varphi_- - \dot{\varphi}_- \varphi_+)
+ \frac{m_\varphi}{2}(\dot{\varphi}_+^2 + 3 \dot\varphi_-^2)
- \frac{J{M_{\rm h}}^2}{3}[(\nabla\varphi_+)^2+3(\nabla\varphi_-)^2] \rt. \nnr
& \lt.
-2 \hbar{M_{\rm h}}{\overline{g}}(v_-\varphi_+ - v_+\varphi_- )
+K_{\rm p}\lt( \cos\varphi_+ + 3\cos\varphi_-\rt)\rt]. \label{Lphitotal}\end{aligned}$$ We see that the external current couples to the phase $\varphi_\pm$ itself. This might seem unphysical since the absolute value of phase has no physical meaning, but is correct, as is known in the case of domain walls [@TKS_PR08; @TK06]. We shall indeed show below that the absolute value of phase does not appear in the equation of motion (see Eq. (\[eqofmo\])). This Lagrangian is essentially an extension of the single-particle Lagrangian for a single magnetic vortex [@SNTKO06] into the case of lattice structure if the two-dimensional coordinates $(X(t),Y(t))$ representing the core of vortex are replaced by two field variables, $\varphi_\pm(\rv,t)$. In fact the single skyrmion dynamics is essentially the same as a single magnetic vortex represented by the Thiele equation. The dynamics of a magnetic domain wall is, in contrast, very different form skyrmions and vortices. In fact, a significant feature of skyrmion lattices is that they can be driven at much lower current density than magnetic domain walls [@Schulz12]. The Lagrangian for a domain wall is described by two coordinates, $X(t)$ representing the position, and $\phi(t)$ representing the angle of the wall plane and these variables play roles of $\varphi_\pm$ [@TK04]. The translational motion of the wall therefore requires the angle $\phi$ to grow, and the development of $\phi$ costs energy because of hard-axis magnetic anisotropy energy. This results in the intrinsic pinning effect which prevents the wall to move at low current. Low threshold current is thus realized either by introducing non-adiabatic torque [@Thiaville05; @Li04st] or by lowering the hard-axis anisotropy [@Koyama11]. In contrast, in the case of vortices and skyrmions, translation modes in the two spatial directions form canonical conjugates. The motion induced is therefore without energy cost if in the absence of pinning [@TKS_PR08]. The lower current density for vortices and skyrmions is thus understood qualitatively based on the Lagrangian.
It has been known that the current generally induces another torque on magnetization, called the non-adiabatic torque, which is perpendicular to the spin-transfer torque [@TKS_PR08]. This torque arises from spin relaxation and non-adiabatic scattering of conduction electrons and it cannot be expressed in terms of a Hamiltonian similarly to the case of the Gilbert damping. The torque is represented in the equation of motion for spin by replacing the Gilbert damping term, ${\alpha_{\rm s}}\dot{\nv}$ by $[{\alpha_{\rm s}}\partial_t-{\beta_{\rm s}}\tilde{P}(\jv\cdot\nabla)]\nv$, where ${\beta_{\rm s}}$ is a dimensionless parameter representing the non-adiabaticity and spin relaxation [@TK04; @Zhang04; @Thiaville05; @TKS_PR08]. In terms of phason, this corresponds to replacing ${\alpha_{\rm s}}\dot{\varphi}_\pm$ by ${\alpha_{\rm s}}\dot{\varphi}_\pm-{\beta_{\rm s}}v_\pm$, since a spatial derivative gives rise to a factor proportional to the phason wave vector.
Considering the case where the current is applied for $\varphi_+$, i.e., $v_-=0$, the equation of motion including the Gilbert damping and non-adiabatic torque is therefore $$\begin{aligned}
-2\hbar {M_{\rm h}}{\overline{g}}( \dot{\varphi}_+ - v_+)-3m_\varphi \ddot{\varphi}_- -2\hbar{\alpha_{\rm s}}\dot{\varphi}_-
+2J{M_{\rm h}}^2\nabla^2\varphi_- - 3K_{\rm p} \sin \varphi_- &=0 \nnr
2\hbar {M_{\rm h}}{\overline{g}}\dot{\varphi}_- -m_\varphi \ddot{\varphi}_+
-\frac{2\hbar}{3}\lt({\alpha_{\rm s}}\dot{\varphi}_+ -{\beta_{\rm s}}v_+\rt)
+\frac{2}{3}J{M_{\rm h}}^2\nabla^2\varphi_+ - K_{\rm p} \sin \varphi_+ &=0 .
\label{eqofmo}\end{aligned}$$ From the equation of motion, Eq. (\[eqofmo\]), the threshold current $\jc$ for the motion of the skyrmion lattice, determined by $2\hbar {M_{\rm h}}{\overline{g}}v_+ =3K_{\rm p}$, is $$\begin{aligned}
\jc= \frac{2eK_{\rm p}}{\hbar Pa^2 k fg},\label{jcresult}\end{aligned}$$ where $e$ is the elementary electric charge. The threshold current thus increases when the renormalization factor $f$ decreases.
After depinning, there is a solution of a steady flow if mass term is neglected. In the case of $\nabla\varphi_\pm=0$, the solution is $$\begin{aligned}
\dot\varphi_+&=v_+ \frac{1+\frac{{\alpha_{\rm s}}{\beta_{\rm s}}}{3({M_{\rm h}}{\overline{g}})^2}}{1+\frac{{\alpha_{\rm s}}^2}{3({M_{\rm h}}{\overline{g}})^2}}\nnr
\dot\varphi_-&=v_+ \frac{{\alpha_{\rm s}}-{\beta_{\rm s}}}{3{M_{\rm h}}{\overline{g}}}\frac{1}{1+\frac{{\alpha_{\rm s}}^2}{3({M_{\rm h}}{\overline{g}})^2}}.\end{aligned}$$ The longitudinal and transverse velocities of the lattice is defined as $v_\parallel= \dot\varphi_+/k$ and $v_\perp= \dot\varphi_-/k$, respectively. we see that the transverse velocity is proportional to $1/f$ as a result of renormalization. (The result in the absence of renormalization effect was obtained previously in Ref. [@IwasakiNC13].) Since both ${\alpha_{\rm s}}$ and ${\beta_{\rm s}}$ are usually small (typically of the order of $10^{-2}$), the longitudinal terminal velocity appears to be insensitive to $\beta$, as noted in Ref. [@IwasakiNC13] and in Ref. [@Kasai06] in the case of magnetic vortex .
To reproduce the observed three-dimensional threshold current density, defined as $\jc/a$, of $10^6$ A/m$^2$ [@Schulz12], Eq. (\[jcresult\]) indicates a very weak collective pinning energy of $K_{\rm p}\sim 6\times 10^{-11}$ eV if $a=5$Å, $P=1$, $J/a^2=1$ meV, $D/a=0.18$ meV [@IwasakiNC13]. It was shown in bulk MnSi that the threshold current density increases rapidly as a function of the temperature near the critical temperature [@Schulz12]. It was discussed there that the behavior is not explained solely by the descrease of ${M_{\rm h}}$, and that the enhancement of the pinning force due to softening of the skyrmion lattice needs to be taken into account. The renormalization effect we have found, indicating that the effective pinning force, proportional to $K_{\rm p}/f$, increases at higher temperatures due to the factor of $1/f$ may explain partially the observed softening behavior. The opposite temperature dependence was observed in a thin FeGe and the result was argued to be due to the weakening of the pinning by the thermal fluctuation [@Yu12].
Topological Hall effect
=======================
The topological charge of magnetic structures induces the Lorentz force on the conduction electrons and induces the Hall effect, called the topological Hall effect. The Hall force in a film is expressed generally as [@TKS_PR08] $$\begin{aligned}
F_{{\rm H},i} &= \frac{\pi\hbar}{e}Pd \sum_j j_j \Phi_{ij},\end{aligned}$$ where $i$ denotes a spatial direction, $-e(<0)$ is the electron charge, $P$ is the spin polarization of the current, $d$ is the thickness of the film and $$\begin{aligned}
\Phi_{ij}\equiv \frac{1}{4\pi M^3}\int d^2 r \Mv\cdot(\nabla_i \Mv\times\nabla_j\Mv),\end{aligned}$$ is the topological charge defined in the $ij$-plane. The Hall resistivity is $$\begin{aligned}
\rho_{xy} =-\frac { F_{{\rm H},y} } {ej_x} =-\frac{\pi\hbar}{e^2}Pd\Phi_{xy}.\end{aligned}$$ In the case of skyrmion lattice, described by Eq. (\[nvidef\]), the topological charge is calculated as (see Eq. (\[SBPvalue\])) $$\begin{aligned}
\Phi_{xy}
&=g\cos \varphi_0 ={\overline{g}},\end{aligned}$$ within the self-consistent harmonic approximation. The magnitude of the topological Hall effect is renormalized by a factor of $f(={\overline{g}}/g)$ in the skyrmion lattice phase.
Conclusion
==========
We have presented a theoretical description of a two-dimensional skyrmion lattice realized in helical magnets by use of phason fields. The ground state skyrmion lattice is described in terms of three helices following the previous approaches by Mühlbauer et al. and Petrova et al. [@Muhlbauer09; @Petrova11], and we have introduced collective coordinates consisting of fluctuations of helix phases (phasons) and perpendicular fluctuations. By deriving an effective Lagrangian describing slowly-varying phasons, we have confirmed previous observations that there are two excitation modes, one gapless mode having quadratic dispersion and a massive mode. The vector nature of spin does not lead to an essential difference from the scalar field like in charge-density waves (CDW) as for the excitation concerns except for the fact that spin phason is always coupled to an effective magnetic field of the spin Berry’s phase. We have found that there is another phase variable, $\varphi_0$, governing the stability and the topological nature of the skyrmion lattice. We demonstrated that the fluctuation of this mode results in a screening of the topological charge of skyrmion lattice, and that the screening effect would be observable in various measurements such as the microwave absorption, and current-induced dynamics and topological Hall effect in metals.
The authors thank X. Z. Yu, S. Seki, J. Kishine, C. Marrows, M. Mochizuki, W. Koshibae, A. Beekman, D. Takahashi and M. Ogata for valuable comments and discussions. H. F. thanks N. Nagaosa for useful discussions in early stage and J. Kishine for drawing attention to Ref. [@Petrova11]. This work was supported by a Grant-in-Aid for Scientific Research (C) (Grant No. 25400344) and (A) (Grant No. 24244053) from Japan Society for the Promotion of Science and UK-Japanese Collaboration on Current-Driven Domain Wall Dynamics from JST.
Mean-field energy of skyrmion lattice \[SECGL\]
===============================================
In this section we summarize the mean-field energy without applied current and pinning estimated for the skyrmion lattice structure, represented by Eqs. (\[skxconfiguration\])(\[nvi0def\]). The energy per unit site evaluated for the free energy (\[Hamiltonian\]) without both pinning and current-induced torque is $$\begin{aligned}
E_{\rm sk} &= \frac{a_M}{2}{M_{\rm f}}^2+\frac{b_M}{4}{M_{\rm f}}^4+\frac{2\mub}{a^3}B{M_{\rm f}}-\frac{\alpha_M}{2}{M_{\rm h}}^2+\frac{\beta_M}{4}{M_{\rm h}}^4
+\gamma_1 {M_{\rm f}}{M_{\rm h}}^3 +\gamma_2 {M_{\rm f}}^2{M_{\rm h}}^2, \label{GLenergy}\end{aligned}$$ where $\alpha_M=3\lt(\frac{D^2}{J}-a_M\rt)$, $\beta_M=\frac{51}{4}b_M$, $\gamma_1=\frac{9}{4}b_M$ and $\gamma_2=3b_M$. The term $\gamma_1$ is essential to stabilize the skyrmion structure.
For comparison, the energy for a single helix is given also by Eq. (\[GLenergy\]) but with different parameters; $\alpha_M=\lt(\frac{D^2}{J}-a_M\rt)$, $\beta_M=\frac{1}{4}b_M$, $\gamma_1=0$ and $\gamma_2=b_M$.
Integrating out degrees of freedom
==================================
Here we describe briefly the integrating-out variables by carrying out path-integral over variables. Usually this is carried out to focus on variables describing energy dynamics, and it is also useful to switch between $p,q$-representation to $q,\dot{q}$-representation. It should be noted that the calculation keeps all the quantum fluctuations included in the result if done without approximation. It is not thus equivalent to deleting variables in the equations of motion.
We first consider a case of a particle in one-dimension, whose Hamiltonian is $H=\frac{p^2}{2m}+V(q)$, where $p$ is the canonical momentum for $q$, $m$ is the mass of the particle and $V$ is a potential. The equations of motion (Hamilton equations) read $$\begin{aligned}
\dot{p} &= - \frac{\delta H}{\delta q} = - \frac{d V(q)}{d q} \nnr
\dot{q} &= \frac{\delta H}{\delta p} = \frac{p}{m} . \label{hamiltoneq}\end{aligned}$$ The equations are also obtained from the Lagrangian, defined as $L(q,p)=p \dot{q}-H$. In the path-integral formalism [@Feynman65], the dynamics of $q$ and $p$ is represented by a functional integral (denoted by ${\cal D}$) called the partition function, defined as $$\begin{aligned}
Z & \equiv \int {\cal D}q {\cal D}p e^{i\int dt (p \dot{q}-H)}.\end{aligned}$$ In the present example, the Lagrangian is quadratic with respect to $p$ and thus the integration over $p$ results in $$\begin{aligned}
Z & = \int {\cal D}q e^{i\int dt L(\dot{q},q))},\end{aligned}$$ where $L(\dot{q},q)\equiv \frac{m}{2}\dot{q}^2-V(q)$ and we have dropped an irrelevant constant. We thus obtained the Lagrangian represented by $\dot{q}$ and $q$. Since the integral was carried out exactly, the resulting Lagrangian is quantum mechanically equivalent to the original Hamiltonian. At the classical level, the equation of motion obtained from $ L(\dot{q},q))$, $ m\ddot{q} = - \frac{d V(q)}{d q}$, is equivalent to Eq. (\[hamiltoneq\]).
Integrating-out $\varphi_0$ \[SECphiint\]
-----------------------------------------
The partition function for $\varphi_0$ is written in the imaginary-time path integral formalism as [@Sakita85] $$\begin{aligned}
Z_{\varphi_0} &= \int{\cal D}\varphi_0 \! e^{\beta L_{\varphi_0}},\end{aligned}$$ where we noted that the Hamiltonian for $\varphi_0$ is $-L_{\varphi_0}$ and $\beta\equiv (\kb T)^{-1}$. The expectation value $\overline{{\varphi_0}^2}$, defined as $$\begin{aligned}
\overline{{\varphi_0}^2}\equiv \frac{1}{ Z_{\varphi_0}}
\int{\cal D}\varphi_0 \varphi_0^2(\rv) e^{-\beta L_{\varphi_0}},\end{aligned}$$ reads by using Fourier transform, $$\begin{aligned}
\overline{{\varphi_0}^2}
&=\frac{1}{ Z_{\varphi_0}}
\int{\cal D}\varphi_0 \sumqv |\varphi_0(\qv)|^2
e^{ -\beta\sum_{\qv'} \frac{1}{2}\lt[\frac{J}{3}(q')^2+hf\rt]|\varphi_0(\qv')|^2 } \nnr
&=
- \frac{\delta \ln Z_{\varphi_0}} {\delta \lt[ \frac{\beta}{2}\lt(\frac{J}{3}q^2+hf\rt)\rt]}.\end{aligned}$$ The partition function is a Gaussian integral, and thus $$\begin{aligned}
Z_{\varphi_0} &= \prod_{\qv} \pi \lt[\frac{\beta}{2}\lt(\frac{J}{3}q^2+hf\rt)\rt]^{-1},\end{aligned}$$ and thus we obtain $$\begin{aligned}
\overline{{\varphi_0}^2}
&= 2\kb T \sumqv \frac{1}{\frac{J}{3}q^2+hf}.\end{aligned}$$
Integrating-out $\beta_\pm$ \[SECbetaint\]
------------------------------------------
In the Lagrangian (\[Lphibetascha\]), the contributions including $\beta_\pm$ are $$\begin{aligned}
L_\beta & \equiv {M_{\rm h}}^2 {\int\frac{d^2r}{a^2}}\lt[-\frac{\hbar}{{M_{\rm h}}}( \beta_+ \dot{\varphi}_+ + 3 \beta_- \dot{\varphi}_- )
- \frac{J}{3} \lt( k^2 ( \beta_+^2+3\beta_-^2) \rt)
-\frac{J}{2} [(\nabla\beta_+)^2+3(\nabla\beta_-)^2 ] \rt] .\end{aligned}$$ The partition function in the real-time is defined as $$\begin{aligned}
Z_\beta\equiv \int{\cal D}\beta_+ {\cal D}\beta_- e^{\frac{i}{\hbar}\int dt L_\beta}.\end{aligned}$$ Using Fourier transformation, we carry out the integral as $$\begin{aligned}
Z_\beta &= \int{\cal D}\beta_+ {\cal D}\beta_-
{\rm exp}\lt[-\frac{i}{\hbar}\int dt \sumqv \lt[\frac{J{M_{\rm h}}^2 k^2}{3} \lt(1+\frac{3q^2}{2k^2}\rt)
\rt.\rt.\nnr
& \times
\lt(\lt|\beta_+(\qv)-\frac{3\hbar\dot{\varphi}_+(\qv)}{2J{M_{\rm h}}k^2\lt(1+\frac{3q^2}{2k^2}\rt)}\rt|^2 +3 \lt|\beta_-(\qv)-\frac{3\hbar\dot{\varphi}_-(\qv)}{2J{M_{\rm h}}k^2\lt(1+\frac{3q^2}{2k^2}\rt)}\rt|^2\rt)
\nnr
& \lt. \lt.
-\frac{3\hbar^2}{4Jk^2\lt(1+\frac{3q^2}{2k^2}\rt)}(\dot{\varphi}_+^2 +3 \dot{\varphi}_-^2)\rt] \rt]\nnr
&\equiv e^{\frac{i}{\hbar}\int dt \delta L_\varphi} , \label{betaintresult}\end{aligned}$$ where $\delta L_\varphi$ is the effective Lagrangian arising from the $\beta$-integral. The integral over $\beta_+$ is carried out as $$\begin{aligned}
\int{\cal D}\beta_+
{\rm exp}\lt[-\frac{i}{\hbar}\int dt \sumqv \frac{J{M_{\rm h}}^2 k^2}{3} \lt(1+\frac{3q^2}{2k^2}\rt)
\lt|\beta_+(\qv)-\frac{3\hbar\dot{\varphi}_+(\qv)}{2J{M_{\rm h}}k^2\lt(1+\frac{3q^2}{2k^2}\rt)}\rt|^2 \rt] \nnr
= \prod_{\qv,t} \lt(\frac{3\pi\hbar}{i J{M_{\rm h}}^2 (k^2+3q^2)}\rt) , \end{aligned}$$ which is a constant independent on dynamic variables. Neglecting irrelevant constants, the Lagrangian is $$\begin{aligned}
\delta L_\varphi
= \sumqv \frac{m_\varphi}{2}\frac{1}{1+\frac{3q^2}{2k^2}} (\dot{\varphi}_+(\qv)^2 +3 \dot{\varphi}_-(\qv)^2), \label{betaintL0}\end{aligned}$$ where $m_\varphi\equiv \frac{3\hbar^2}{2Jk^2}$ is the mass for $\dot{\varphi}_+$. In the real-space representation, the Lagrangian is non-local, but for discussing low energy dynamics of $\varphi_\pm$, it is enough to neglect the order of $q^2$, resulting in Eq. (\[betaintL\]).
To discuss the massive excitation mode, the order of $q^2$ in Eq. (\[betaintL0\]) needs to be kept. In this case, we have $$\begin{aligned}
\delta L_\varphi
= {\int\frac{d^2r}{a^2}}\lt[ \frac{m_\varphi}{2}
(\dot{\varphi}_+(\qv)^2 +3 \dot{\varphi}_-(\qv)^2)
- \frac{3m_\varphi}{4k^2} ((\nabla\dot{\varphi}_+)^2 +3 (\nabla\dot{\varphi}_-)^2) \rt]
+O(q^4\omega^2). \label{betaintLexp}\end{aligned}$$ Although the second term containing $\nabla^2\partial_t^2$ is higher-order contribution when considering standard excitations, it modifies the dispersion relation in the present case. In fact, the matrix of Eq. (\[dispersionphi\]) now reads $$\begin{aligned}
\lt(\begin{array}{cc}
\frac{J{M_{\rm h}}^2 q^2}{3}-\frac{m_\varphi \omega ^2}{2\mu_q} & -2 i\hbar {M_{\rm h}}{\overline{g}}\omega \\
2 i\hbar {M_{\rm h}}{\overline{g}}\omega & 3 \left(\frac{J{M_{\rm h}}^2 q^2}{3}-\frac{m_\varphi \omega
^2}{2\mu_q}\right) \end{array}
\rt),\end{aligned}$$ where $\mu_q\equiv 1+\frac{3q^2}{2k^2}$. The determinant of the matrix is (dropping constants) $$\begin{aligned}
\lt({\hbar\omega}\rt)^4
-\lt({\hbar\omega}\rt)^2 \frac{64}{27}(J{M_{\rm h}}k^2{\overline{g}}\mu_q)^2\lt(1+\frac{3q^2}{8k^2{\overline{g}}^2\mu_q}\rt)
+\frac{16}{81}(J^2{M_{\rm h}}^2 k^2q^2\mu_q)^2.\label{determinant2}\end{aligned}$$ (This result is identical to Eq. (\[determinant\_1\]) if $g$ is replaced by ${\overline{g}}=fg$.) The energy dispersion determined from Eq. (\[determinant2\]) is $$\begin{aligned}
\hbar \omega_q^\pm &= \frac{4\sqrt{2}{M_{\rm h}}Jk^2 {\overline{g}}\mu_q}{3\sqrt{3}}
\lt[
1+\frac{3q^2}{8(k{\overline{g}})^2\mu_q} \pm \sqrt{ 1+\frac{3q^2}{4(k{\overline{g}})^2\mu_q} } \rt]^{\frac{1}{2}}.\end{aligned}$$ We thus obtain the massive mode as $$\begin{aligned}
\hbar \omega_q^\pm &= \frac{8{M_{\rm h}}Jk^2 {\overline{g}}\mu_q}{3\sqrt{3}}
\lt[
1+\frac{3q^2}{8(k{\overline{g}})^2\mu_q} \rt]
\nnr
&=\frac{8 {M_{\rm h}}Jk^2 {\overline{g}}\mu_q}{3\sqrt{3}} +\frac{J{M_{\rm h}}(1+8{\overline{g}}^2) }{2\sqrt{3} {\overline{g}}} q^2 +O(q^4).\end{aligned}$$ We therefore reproduce the result of Eq. (\[dispersion\]) obtained before the integration over $\beta_\pm$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The results of the seventh edition of the BioASQ challenge are presented in this paper. The aim of the BioASQ challenge is the promotion of systems and methodologies through the organization of a challenge on the tasks of large-scale biomedical semantic indexing and question answering. In total, 30 teams with more than 100 systems participated in the challenge this year. As in previous years, the best systems were able to outperform the strong baselines. This suggests that state-of-the-art systems are continuously improving, pushing the frontier of research.\
**2019**. This manuscript version is made available under the CC-BY-NC-ND 4.0 license http://creativecommons.org/licenses/by-nc-nd/4.0/
author:
- 'Anastasios Nentidis$^{1,2}$'
- 'Konstantinos Bougiatiotis$^{1,3}$'
- Anastasia Krithara$^1$
- 'Georgios Paliouras$^{1,4}$'
bibliography:
- 'BioASQ7\_overview.bib'
title: Results of the seventh edition of the BioASQ Challenge
---
Introduction
============
The aim of this paper is twofold. First, we aim to give an overview of the data issued during the BioASQ challenge in 2019. In addition, we aim to present the systems that participated in the challenge and evaluate their performance. To achieve these goals, we begin by giving a brief overview of the tasks, which took place from February to May 2019, and the challenge’s data. Thereafter, we provide an overview of the systems that participated in the challenge. Detailed descriptions of some of the systems are given in the workshop proceedings. The evaluation of the systems, which was carried out using state-of-the-art measures or manual assessment, is the last focal point of this paper, with remarks regarding the results of each task. The conclusions sum up this year’s challenge.
Overview of the Tasks
=====================
The challenge comprised two tasks: (1) a large-scale biomedical semantic indexing task (Task 7a) and (2) a biomedical question answering task (Task 7b). In this section a brief description of the tasks is provided focusing on differences from previous years and updated statistics about the corresponding datasets. A complete overview of the tasks and the challenge is presented in [@Tsatsaronis2015].
Large-scale semantic indexing - 7a
----------------------------------
In Task 7a the goal is to classify documents from the PubMed digital library into concepts of the MeSH hierarchy. Here, new PubMed articles that are not yet annotated by MEDLINE indexers are collected and used as test sets for the evaluation of the participating systems. Similarly to task 5a and 6a, articles from all journals were included in the test data sets of task 7a. As soon as the annotations are available from the MEDLINE indexers, the performance of each system is calculated using standard flat information retrieval measures, as well as, hierarchical ones. As in previous years, an on-line and large-scale scenario was provided, dividing the task into three independent batches of 5 weekly test sets each. Participants had 21 hours to provide their answers for each test set. Table \[tab:7a\_data\] shows the number of articles in each test set of each batch of the challenge. 14,200,259 articles with 12.69 labels per article, on average, were provided as training data to the participants.
[M[0.12]{}M[0.18]{}M[0.21]{}M[0.21]{}]{} **Batch** & **Articles** & **Annotated Articles** & **Labels per Article**\
&7,358 &7,194 & 11.67\
&7,166 &7,021 & 12.95\
&11,019 &10,831 & 13.04\
&5,566 &5,482 & 12.32\
&6,729 &6,353 & 12.96\
**Total** & 37,838& 36,881& 12.31\
&6,380 &6,098&12.51\
&6,785 &6,621&12.75\
&6,207 &5,927&12.75\
&7,382 &7,079&13.00\
&7,240 &6,756&12.65\
**Total** &33,994 &32,481&12.27\
&6,266& 5,835& 12.58\
&11,455& 10,386& 12.86\
&4,750& 3,947& 12.67\
&7,338&5,021& 12.70\
&6,920&4,554& 12.63\
**Total** &36,729& 29,743& 12.14\
Biomedical semantic QA - 7b
---------------------------
The goal of Task 7b was to provide a large-scale question answering challenge where the systems had to cope with all stages of a question answering task for four types of biomedical questions: “yes/no”, “factoid”, “list” and “summary” questions [@balikas13]. As in previous years, the task comprised two phases: In phase A, BioASQ released 100 questions and participants were asked to respond with relevant elements from specific resources, including relevant MEDLINE articles, relevant snippets extracted from the articles, relevant concepts and relevant RDF triples. In phase B, the released questions were enhanced with relevant articles and snippets selected manually and the participants had to respond with *exact answers*, as well as with summaries in natural language (dubbed *ideal answers*). The task was split into five independent batches and the two phases for each batch were run with a time gap of 24 hours. In each phase, the participants received 100 questions and had 24 hours to submit their answers. Table \[tab:7b\_data\] presents the statistics of the training and test data provided to the participants. The evaluation included five test batches.
[M[0.18]{}M[0.1]{}M[0.23]{}M[0.23]{}]{} **Batch** & **Size** & **Documents** & **Snippets**\
Train& 2,747& 11.14& 13.91\
Test 1& 100& 3.07& 3.93\
Test 2& 100& 2.64& 3.22\
Test 3& 100& 3.08& 4.05\
Test 4& 100& 2.78& 3.71\
Test 5& 100& 2.39& 2.62\
**Total**& 3,247& 9.85& 12.31\
Overview of Participants
========================
Task 7a
-------
For this task, 12 teams participated and results from 30 different systems were submitted. In the following paragraphs we describe those systems for which a description was available, stressing their key characteristics. An overview of the systems and their approaches can be seen in Table \[tab:7a\_sys\].
[M[0.3]{}M[0.5]{}]{} **System** & **Approach**\
ceb & CNN, embeddings, ensembles\
DeepMesh & d2v, tf-idf, MESHlabeler, attention scheme, PLT\
Iria & bigrams, Luchene Index, k-NN, ensembles, UIMA ConceptMapper\
MeSHProbeNet-P & Bidirectional RNN (GRU), attention scheme, encoder-decoder architecture\
Semantic NoSQL KE & UIMA ConceptMapper, par2vec, DeepLearning4j[^1]\
The National Library of Medicine (NLM) team, in its “*ceb*” systems [@Rae2019], adopts an end-to-end deep learning architecture with Convolutional Neural Networks (CNN) [@liu2017deep] to improve the results of the Medical Text Indexer (MTI) [@morkBioasq2014]. In particular, they combine text embeddings with journal information. They also consider information about the years of publication and indexing, to capture concept drift and variations in the MeSH vocabulary respectively. They also experiment with an ensemble of independently trained DL models.
The Fudan University team builds upon their previous “*DeepMeSH*” systems, which are based on document to vector (*d2v*) and tf-idf feature embeddings [@peng2016], the MESHLabeler system [@liu2015] and learning to rank (LTR). This year, they incorporate AttentionXML [@You2018], a deep-learning-based extreme multi-label text classification model, in the “*DeepMeSH*”framework. In particular, AttentionXML combines a multi-label attention mechanism, to capture label-specific information, with a shallow and wide probabilistic label tree (PLT) [@Jain2016], for improved efficiency.
The “*Iria*” systems [@ribadascole] are based on the same techniques used by their systems for the previous version of the challenge which are summarized in Table \[tab:7a\_sys\] and described in the corresponding challenge overview [@nentidis2017results].
The “*MeSHProbeNet-P*” systems are upgraded versions of MeSHProbeNet [@Xun2019], which participated in BioASQ6 with the name “*xgx*”. Their approach is based on an end-to-end deep learning model with an encoder-decoder architecture. The encoder consists of a recurrent neural network with multiple attentive MeSH probes to extract different aspects of biomedical knowledge from each input article. In “*MeSHProbeNet-P*” the attentive MeSH probes are also personalized for each biomedical article, based on the domain of each article as expressed by the journal where it has been published.
Finally, the “*Semantic NoSQL KE*” system variants [@Bernd2019] were developed extending previous year’s “*SNOKE*” systems. The systems are based on the ZB MED Knowledge Environment [@ZBMed2017], utilizing the Snowball Stemmer [@snowball:2000] and the UIMA [@tanenblatt2010] ConceptMapper to find matches between MeSH terms and words in the title and abstract of each target document, adopting different matching strategies. Paragraph Vectors [@Le2014] trained on the BioASQ corpus are used to rank and filter all the MeSH headings suggested by the UIMA-based framework for each document.
Similarly to the previous year, two systems developed by NLM to assist the indexers in the annotation of MEDLINE articles, served as baselines for the semantic indexing task of the challenge. MTI [@morkBioasq2014] with some enchantments introduced in [@zavorin2016] and an extension of it, incorporating features of the winning system of the first BioASQ challenge [@tsoumakasBioasq].
Task 7b
-------
The question answering task was tackled by 73 different systems, developed by 18 teams. In the first phase, which concerns the retrieval of information required to answer a question, 6 teams with 23 systems participated. In the second phase, where teams are requested to submit exact and ideal answers, 13 teams with 52 different systems participated. An overview of the technologies employed by each team can be seen in Table \[tab:7b\_sys\].
[M[0.2]{}M[0.1]{}M[0.5]{}]{} **Systems** & **Phase**& **Approach**\
AUTH & A, B & MetaMap, BeCAS, Lucene Index, ElasticSearch, Wordnet, ELMo, SentiWordnet, w2vec, BiLSTM\
AUEB & A & BM25, w2vec, BERT, DL (BCNN, PACRR, PDRMM)\
MindLab & A & ElasticSearch, BM25, QuickUMLS, w2vec, WMD, DL (CNN)\
\_sys & A & Word and Sentence embeddings, Pseudo Relevance Feedback, BM25, LSI\
BJUTNLP & B & SQUAD, GloVe, BiLSTM, Pointer Network\
BIOASQ\_VK & B & ELMo, DMN attention mechanisms, NLTK-VADER\
DMIS & B & BioBERT, SQUAD, transfer learning\
google & B & BERT, CoQA, Natural Questions\
L2PS & B & SQUAD, Quasar-T, DRQA (RNN, LSTM), PSPR (LSTM), BioBERT\
LabZhu & B & PubTator, Stanford POS tool, SPARQL\
MQU & B & w2vec, tf-idf, DL (LSTM), Reinforcement Learning\
UNCC & B & BioBERT, SQUAD, Stanford POS tool, AllenNLP entailment\
unipi-quokka-QA & B & ELMo, ELMo-PubMed, BERT, BioBERT, SciSpacy\
The “*AUTH*” team participated in both phases of Task 7B, with focus on phase B. For the document retrieval task, they experimented with approaches based on the BioASQ search services and ElasticSearch, querying with the conjunction of words in each question for the top 10 documents. In Phase B, for factoid and list questions they used updated versions of their BioASQ6 system [@Dimitriadis2019], based on word embeddings, MetaMap [@AronsonL10], BeCAS [@Nunes2013] and WordNet. For yes/no questions they experiment with different deep learning methods, based on ELMo embeddings [@Peters2018], SentiWordnet [@Esuli06sentiwordnet:a] and similarity matrices to represent the question/answer pairs and use them as input for different BiLSTM architectures [@Dimitriadis2019bioasq].
The “*AUEB*” team participated in Phase A on document and snippet retrieval tasks yielding great results. They built upon their BioASQ6 document retrieval systems [@Brokos2018; @mcdonald2018], which they modify to yield a relevance score for each sentence and experiment with BERT and PACRR [@mcdonald2018] for this task. For snippet retrieval, they utilize a BCNN [@yin2016abcnn] model and a model based on POSIT-DRMM (PDRMM) [@mcdonald2018]. They also introduce JPDRMM, a novel deep learning approach for joint document and snippet ranking, based on PDRMM [@Pappas2019].
Another approach based on deep learning methodologies for Phase A, focusing again on document and snippet retrieval, was proposed by the “*MindLaB*” team from the National University of Colombia [@Vargas2019]. For the document retrieval they use the BM25 model [@robertson:1976] and ElasticSearch [@gormley2015elasticsearch] for efficiency, along with a Word Mover’s Distance [@kusner2015word] based re-ranking scheme. For snippet retrieval, as in the previous approach, they utilized a very large collection of PubMed articles to train a CNN with similarity matrices of question-answer pairs. More specifically, they employ the BioNLPLab[^2] w2vec embeddings that take into account the Part of Speech of each word. Also, they deploy the QuickUMLS [@soldaini2016quickumls] tool to create a cui2vec embedding for each snippet.
The “*\_sys*” systems also participated in Phase A of Task 7B. These systems filter the queries, using stop-word lists and regular expressions, and expand them using word embeddings and pseudo-relevance feedback. Relevant documents are retrieved, utilizing Query Likelihood with bigrams and BM25, and reranked, based on Latent Semantic Indexing (LSI) and document vectors. In particular, document vectors based on averaging sentence embeddings are adopted. Finally, different lists of documents are merged to form the final result, considering the position of the documents in each list.
In phase B, most systems focused on using embeddings and deep learning methodologies to tackle the tasks. For example the “*BJUTNLP*” system utilizes the SQUAD Dataset for pre-training. The system uses both GloVe embeddings [@pennington2014glove] (fine tuned during training) and character-level word embeddings (through a 1-dimensional CNN) as input to a BiLSTM model and for each question a Pointer Network [@see2017get] is finally responsible for pinpointing the exact start and end position of the answer in the relevant snippets.
The “*BIOASQ\_VK*” systems were based on BioBERT [@lee2019biobert], but with novel modifications to allow the model to cope with yes/no, factoid and list questions [@Kanjirangat2019]. They pre-trained the model on the SQUAD dataset (for factoid and list questions) and SQUAD2 (for yes/no questions) to leverage the small size of the BioASQ dataset and by exploiting different pre-/post-processing techniques they obtained great results on all subtasks.
The “*DMIS*” systems focused on the importance of the information (words, phrases and sentences) for a given question [@Yoon2019]. To this end, sentence level embeddings based on ELMo embeddings [@Peters2018] and attention mechanisms facilitated by Dynamic Memory Networks (DMN) [@kumar2016ask] are deployed. Moreover, sentiment analysis is performed on yes/no questions to guide the classification (positive corresponds to yes) using the NLTK-VADER [@hutto2014vader] tool.
The “*google*” systems [@Hosein2019], focus on factoid questions and are based on BERT based models [@devlin2018bert], specifically the one in [@alberti2019bert] trained on the Natural Questions [@kwiatkowski2019natural] dataset, while also utilizing the CoQA [@reddy2019coqa] and the BioASQ datasets. They experiment with different input to the models, including the abstracts of relevant articles, the provided gold snippets and predicted relevant snippets. In particular, they focus on error propagation in end-to-end information retrieval and question answering systems, reaching the interesting conclusion that the information retrieval part is a bottleneck for such end-to-end QA systems.
Interesting results come from the “*L2PS*” team where they quantify the importance of pre-training and fine-tuning models for question answering and view the task under different regimes, namely Reading Comprehension (RC) and Open QA [@Kamath2019]. For the RC regime they use DRQA’s document reader [@chen2017reading] while for the Open QA they utilize the PSPR model [@lin2018denoising]. They experiment with different datasets (SQUAD [@rajpurkar2016squad] for RC and Quasar-T [@dhingra2017quasar] for Open QA) for fine-tuning the models, as well as BioBert [@lee2019biobert] embeddings to gain insights on the effect of the context length in this task.
The “*LabZhu*” [@zhang2015] systems improved upon their systems from BioASQ6, with focus on exact answer generation. In particular, for factoid and list questions they developed two distinct approaches. One based on traditional information retrieval approaches, involving candidate answer generation and ranking, and one Knowledge-Graph based approach. In the latter approach, the answer type and the topic entity of the question are predicted and a SPARQL query is generated based on them and used to retrieve some results from the Knowledge Graph. Finally, the results of the two approaches are combined for the final answer of the question.
The Macquarie University (“*MQU*”) team focused on ideal answers and approached the task under a classification approach for snippet relevance [@Molla2019]. Extending their previous work [@Diego2017; @molla2018macquarie] the snippets are marked as summary relevant or not, utilizing w2vec embeddings and tf-idf vectors of the question-sentence pairs, showcasing that a classification scheme is more appropriate than a regression one. Also, based on their previous work [@molla_REINFORCE:2017], they conduct experiments using reinforcement learning towards the ROUGE score of the ideal answers and a correlation analysis between various ROUGE metrics and the BioASQ human evaluation scores, observing poor correlation of the ROUGE-Recall score with human evaluation.
The “*UNCC*” team focused on factoid, list and yes/no questions [@Telukuntla2019]. Their work is based on the BioBERT [@lee2019biobert] embeddings fine-tuned on previous years of BioASQ. They also utilize the SQUAD dataset for factoid answers and incorporated the Lexical Answer Type (LAT) [@ferrucci2010building] and POS-tags along with hand made rules to address specific errors of the system. Furthermore, they incorporated the entailment of the candidate sentences in yes/no questions using the AllenNLP library[@Gardner2017AllenNLP].
Finally, the “*unipi-quokka-QA*” system tackled all the different question types in phase B [@Resta2019]. Their work focused on experimenting with different Transformer models and embeddings, namely: ELMo, ELMo-Pumbed, BERT and BioBERT. They used different strategies depending on the question type, such as ensembles on yes/no questions, biomedical named entity extraction (using SciSpacy [@Neumann2019ScispaCyFA]) on list questions and different pre-/post-processing procedures.
In this challenge too, the open source OAQA system proposed by [@yang2016learning] served as baseline for phase B. The system which achieved among the highest performances in previous versions of the challenge remains a strong baseline for the exact answer generation task. The system is developed based on the UIMA framework. ClearNLP is employed for question and snippet parsing. MetaMap, TmTool [@Wei2016], C-Value and LingPipe [@baldwin2003lingpipe] are used for concept identification and UMLS Terminology Services (UTS) for concept retrieval. The final steps include identification of concept, document and snippet relevance, based on classifier components and scoring, ranking and reranking techniques.
Results
=======
Task 7a
-------
Each of the three batches of Task 7a were evaluated independently. The classification performance of the systems were measured using flat and hierarchical evaluation measures [@balikas13]. The micro F-measure (MiF) and the Lowest Common Ancestor F-measure (LCA-F) were used to choose the winners for each batch [@kosmopoulos2015evaluation].
According to [@Demsar06] the appropriate way to compare multiple classification systems over multiple datasets is based on their average rank across all the datasets. On each dataset the system with the best performance gets rank 1.0, the second best rank 2.0 and so on. In case two or more systems tie, they all receive the average rank. Table \[tab:7a\_res\] presents the average rank (according to MiF and LCA-F) of each system over all the test sets for the corresponding batches. Note, that the average ranks are calculated for the 4 best results of each system in the batch according to the rules of the challenge.
[M[0.4]{}M[0.07]{}M[0.09]{}M[0.07]{}M[0.09]{}M[0.07]{}M[0.09]{}]{} **System** & & &\
& MiF & LCA-F & MiF & LCA-F & MiF & LCA-F\
DeepMeSH5 & - &- & 1,00 & 1,00 & 1 &1\
DeepMeSH4 & - & - & 9,50& 9,50& 2,25& 1,75\
DeepMeSH3 & 8,25& 8,50& 3,50& 5,00& 2,5& 2,75\
DeepMeSH1 & 5,00& 6,25& 2,00& 2,63& 3,75& 4,13\
DeepMeSH2 & 7,25& 7,25& 3,50& 4,50& 4,75& 4,38\
MeSHProbeNet-P2 & 2,63& 2,63& 4,63& 5,88& 6,5& 8,25\
MeSHProbeNet-P1 & 3,25& 2,13& 6,38& 4,25& 6,88& 6,5\
MeSHProbeNet-P3 & 5,00& 4,63& 8,38& 7,25& 7,5& 7,38\
MeSHProbeNet-P & 2,38& 3,25& 7,00& 4,38& 8,13& 7,75\
MeSHProbeNet-P0 & 1,50& 1,25& 6,25& 5,63& 8,75& 7,88\
ceb 1 ensemble & - & - & - & - & 11& 11\
Default MTI & 9,75& 8,75& 12,00& 11,75& 12,25& 12,25\
ceb1 & 8,75& 9,25& 11,00& 11,25& 12,25& 13,5\
MTI First Line Index & 11,50& 11,25& 13,00& 12,50& 13,25& 12\
iria-mix & - & - & 14,00& 14,00& 14,5& 14,75\
Semantic NoSQL KE 2& - & - & - & - & 16& 16\
Semantic NoSQL KE 1 & - & - & - & - & 17& 17,75\
The results in Task 7a show that in all test batches and for both flat and hierarchical measures, some systems outperform the strong baselines. In particular, The “*MeSHProbeNet-P*” systems achieve the best performance in the first batch, outperformed by the “*DeepMeSH*” systems in the last two batches. More detailed results can be found in the online results page[^3]. Comparison of these results with corresponding system results from previous years reveals the improvement of both the baseline and the top performing systems through the years of the competition as shown in Figure [\[fig:01\]]{}.
{width="100.00000%"}
Task 7b
-------
[M[0.31]{}M[0.12]{}M[0.12]{}M[0.12]{}M[0.12]{}M[0.12]{}]{} **System** & **Mean Precision** & **Mean Recall** & **Mean F-measure** & **MAP** & **GMAP**\
aueb-nlp-2 & 0.2060& 0.4039& 0.2365& **0.2114**& 0.0075\
aueb-nlp-1 & 0.2124& 0.4083& 0.2440& 0.2086& 0.0065\
aueb-nlp-5 & **0.2157**& **0.4235**& **0.2467**& 0.1821& **0.0098**\
MindLab QA Reloaded& 0.1587 & 0.2760& 0.1723& 0.1527& 0.0013\
Deep ML methods for & 0.1331 & 0.2692& 0.1589& 0.1234& 0.0009\
MindLab Red Lions++ & 0.1371& 0.2538& 0.1535& 0.1187& 0.0014\
aueb-nlp-3 & 0.1488& 0.3427& 0.1779& 0.1149& 0.0053\
MindLab QA System ++& 0.1288& 0.2049& 0.1364& 0.1136& 0.0010\
aueb-nlp-4 & 0.1520 &0.3237& 0.1791& 0.1116& 0.0056\
MindLab QA System& 0.1297& 0.2536& 0.1478& 0.1094& 0.0016\
lh\_sys1& 0.0399 &0.0810& 0.0478& 0.0178& 0.0001\
lh\_sys3 & 0.0233 &0.0437& 0.0266& 0.0151& 0.0001\
lh\_sys5 & 0.0233 &0.0437& 0.0266& 0.0151& 0.0001\
lh\_sys4 & 0.0233 & 0.0437& 0.0266& 0.0148& 0.0001\
lh\_sys2 & 0.0182& 0.0281& 0.0193& 0.0051& 0.0001\
[M[0.27]{}M[0.13]{}M[0.13]{}M[0.13]{}M[0.13]{}M[0.12]{}]{} **System** & **Mean Precision** & **Mean Recall** & **Mean F-measure** & **MAP** & **GMAP**\
aueb-nlp-4& 0.1750& **0.6266** & 0.2471& **0.1199** & 0.0151\
aueb-nlp-2& 0.1740& 0.6139& 0.2449& 0.1121& 0.0156\
aueb-nlp-5 & **0.3599**& 0.6128& **0.4034** & 0.1102& **0.0164**\
aueb-nlp-1 & 0.1700& 0.5912& 0.2380& 0.1041& 0.0118\
auth-qa-1 & 0.2675& 0.3896& 0.2894 & 0.1033& 0.0018\
aueb-nlp-3 & 0.1600& 0.5806& 0.2266& 0.0986& 0.0104\
lh\_sys4& 0.1420 & 0.5490& 0.2081 & 0.0920& 0.0069\
Ir\_sys1 & 0.1410& 0.5365& 0.2059& 0.0907& 0.0059\
lh\_sys1& 0.1420 & 0.5449& 0.2076 & 0.0881& 0.0063\
MindLab QA Reloaded& 0.1330& 0.5288& 0.1950& 0.0863& 0.0062\
[M[0.24]{}M[0.0852]{}M[0.0852]{}M[0.0852]{}M[0.0852]{}M[0.0852]{}M[0.0852]{}M[0.0852]{}M[0.0852]{}]{}
**System** & & &\
& Acc. & F1 & Str. Acc. & Len. Acc. & MRR & Prec. & Rec. & F1\
BioBERT-DMIS-3 & **0.8286**& **0.8250** & **0.2857** & 0.4286 & 0.3452 & **0.5653**& 0.4131& **0.4619**\
BioBERT-DMIS & 0.8000& 0.7822& 0.2571 & 0.4571& 0.3224& 0.5236 & 0.3714 & 0.4202\
unipi-quokka-QA-5 & 0.8000 & 0.7939 & 0.0857 & 0.1714 & 0.1152 & 0.1713 & **0.5873** &0.2537\
BioBERT-DMIS-2 & 0.7429 &0.7200& 0.2571 &0.4571 & 0.3271 & 0.5486 & 0.3992 & 0.4468\
BioBERT-DMIS-4 & 0.7429 & 0.7351& 0.2286& 0.4571& 0.3238& 0.5069 & 0.3575 & 0.4051\
google-gold-input-ab & 0.7143 & 0.6941 & 0.2286 &0.2857 & 0.2571 & 0.1774 & 0.4175 & 0.2415\
unipi-quokka-QA-4 & 0.7143 & 0.6941& 0.0857 & 0.1714& 0.1152 & 0.1713 & **0.5873** & 0.2537\
unipi-quokka-QA-3 & 0.6857 & 0.6578 & 0.0857 & 0.1714 & 0.1152 & 0.1713 &**0.5873** & 0.2537\
google-gold-input & 0.6571 & 0.6023& **0.2857** & 0.3714 & 0.3167 & 0.2159& 0.4452& 0.2824\
DMIS & 0.6571 & 0.6023& **0.2857**& **0.5143**& **0.3638**& 0.5050 & 0.3714 & 0.4124\
BioASQ\_Baseline & 0.4857& 0.4643 & 0.0571 & 0.1429& 0.0867& 0.2127 & 0.3619 & 0.2573\
**Phase A**: For phase A and for each of the four types of annotations: documents, concepts, snippets and RDF triples, we rank the systems according to the Mean Average Precision (MAP) measure. The final ranking for each batch is calculated as the average of the individual rankings in the different categories. In Tables \[tab:5bA\_res\_sni\] and \[tab:5bA\_res\_doc\] some indicative results from batches 3 and 4 are presented. Full results are available in the online results page of Task 7b, phase A[^4]. These results are preliminary. The final results for Task 7b, phase A will be available after the manual assessment of the system responses.
**Phase B**: In phase B of Task 7b the systems were asked to produce exact and ideal answers. For ideal answers, the systems will eventually be ranked according to manual evaluation by the BioASQ experts [@balikas13]. Regarding exact answers[^5], the systems were ranked according to accuracy, F1 score on prediction of yes answer, F1 on prediction of no and macro-averaged F1 score for the yes/no questions, mean reciprocal rank (MRR) for the factoids and mean F-measure for the list questions. Table \[tab:5bB\_res\] shows the results for exact answers for the last batch of Task 7b. These results are preliminary. The full results of phase B of Task 7b are available online[^6]. The final results for Task 7b, phase B will be available after the manual assessment of the system responses.
The results presented in Figure [\[fig:02\]]{} show that this year the performance of systems in the yes/no questions, has clearly improved. In batch 5 for example, presented in Table \[tab:5bB\_res\], some systems outperformed the strong baseline based on previous versions of the OAQA system, with the top system achieving almost double the score of the baseline. Some improvement is also observed in the performance of the top systems for factoid and list questions in the preliminary results. However, there is even more room for improvement in these types of question as can be seen in Figure [\[fig:02\]]{}.
{width="100.00000%"}
Conclusions
===========
In this paper, an overview of the seventh BioASQ challenge is presented. The challenge consisted of two tasks: semantic indexing and question answering. Overall, as in previous years, the best systems were able to outperform the strong baselines provided by the organizers. This suggests that advances over the state of the art were achieved through the BioASQ challenge but also that the benchmark in itself is challenging. Moreover, the shift towards systems that incorporate ideas based on deep learning models observed in the previous year, is even more clear. Novel ideas have been tested and state-of-the-art deep learning methodologies have been adapted to biomedical question answering with great results. Specifically, the breakthroughs in different NLP tasks using clever techniques with the advent of new language-models, such as BERT and gpt-2, gave birth to new approaches that significantly boost the performance of the systems. In the future, we expect novel methodologies, such as the newly proposed XLNet [@DBLP:journals/corr/abs-1906-08237], to further cultivate research in the biomedical information systems field. Consequently, we believe that the challenge is successfully pushing the research frontier of this domain. In future editions of the challenge, we aim to provide even more benchmark data derived from a community-driven acquisition process.
Acknowledgments {#acknowledgments .unnumbered}
===============
Google was a proud sponsor of the BioASQ Challenge in 2018. The seventh edition of BioASQ is also sponsored by the Atypon Systems inc. BioASQ is grateful to NLM for providing baselines for task 7a and to the CMU team for providing the baselines for task 7b. Finally, we would also like to thank all teams for their participation.\
[^1]: https://deeplearning4j.org/ Accessed June 2019
[^2]: http://bio.nlplab.org Accessed June 2019
[^3]: <http://participants-area.bioasq.org/results/7a/>
[^4]: <http://participants-area.bioasq.org/results/7b/phaseA/>
[^5]: For summary questions, no exact answers are required
[^6]: <http://participants-area.bioasq.org/results/7b/phaseB/>
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'M. Ajello'
- 'L. Baldini'
- 'G. Barbiellini'
- 'D. Bastieri'
- 'K. Bechtol'
- 'R. Bellazzini'
- 'B. Berenji'
- 'E.D. Bloom'
- 'E. Bonamente'
- 'A.W. Borgland'
- 'J. Bregeon'
- 'M. Brigida'
- 'P. Bruel'
- 'R. Buehler'
- 'S. Buson'
- 'G.A. Caliandro'
- 'R.A. Cameron'
- 'P.A. Caraveo'
- 'J.M. Casandjian'
- 'C. Cecchi'
- 'E. Charles'
- 'A. Chekhtman'
- 'J. Chiang'
- 'S. Ciprini'
- 'R. Claus'
- 'J. Cohen-Tanugi'
- 'J. Conrad'
- 'S. Cutini'
- 'A. de Angelis'
- 'F. de Palma'
- 'C.D. Dermer'
- 'E. do Couto e Silva'
- 'P.S. Drell'
- 'A. Drlica-Wagner'
- 'T. Enoto'
- 'C. Favuzzi'
- 'S.J. Fegan'
- 'E.C. Ferrara'
- 'Y. Fukazawa'
- 'P. Fusco'
- 'F. Gargano'
- 'D. Gasparrini'
- 'S. Germani'
- 'N. Giglietto'
- 'F. Giordano'
- 'M. Giroletti'
- 'T. Glanzman'
- 'G. Godfrey'
- 'P. Graham'
- 'I.A. Grenier'
- 'S. Guiriec'
- 'M. Gustafsson'
- 'D. Hadasch'
- 'M. Hayashida'
- 'R.E. Hughes'
- 'A.S. Johnson'
- 'T. Kamae'
- 'H. Katagiri'
- 'J. Kataoka'
- 'J. Knödlseder'
- 'M. Kuss'
- 'J. Lande'
- 'L. Latronico'
- 'A.M. Lionetto'
- 'F. Longo'
- 'F. Loparco'
- 'M.N. Lovellette'
- 'P. Lubrano'
- 'M.N. Mazziotta'
- 'P.F. Michelson'
- 'W. Mitthumsiri'
- 'T. Mizuno'
- 'C. Monte'
- 'M.E. Monzani'
- 'A. Morselli'
- 'I.V. Moskalenko'
- 'S. Murgia'
- 'J.P. Norris'
- 'E. Nuss'
- 'T. Ohsugi'
- 'A. Okumura'
- 'E. Orlando'
- 'J.F. Ormes'
- 'M. Ozaki'
- 'D. Paneque'
- 'M. Pesce-Rollins'
- 'M. Pierbattista'
- 'F. Piron'
- 'G. Pivato'
- 'S. Rainò'
- 'M. Razzano'
- 'S. Ritz'
- 'M. Roth'
- 'P.M. Saz Parkinson'
- 'J.D. Scargle'
- 'T.L. Schalk'
- 'C. Sgrò'
- 'E.J. Siskind'
- 'G. Spandre'
- 'P. Spinelli'
- 'D.J. Suson'
- 'H. Tajima'
- 'H. Takahashi'
- 'T. Tanaka'
- 'J.G. Thayer'
- 'J.B. Thayer'
- 'L. Tibaldo'
- 'M. Tinivella'
- 'D.F. Torres'
- 'E. Troja'
- 'Y. Uchiyama'
- 'T.L. Usher'
- 'J. Vandenbroucke'
- 'V. Vasileiou'
- 'G. Vianello'
- 'V. Vitale'
- 'A.P. Waite'
- 'B.L. Winer'
- 'K.S. Wood'
- 'M. Wood'
- 'Z. Yang'
- 'S. Zimmer'
title: 'Limits on Large Extra Dimensions Based on Observations of Neutron Stars with the *Fermi*-LAT'
---
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The gap between cell and module efficiency is a major challenge for all photovoltaic (PV) technologies. For monolithic thin film PV modules, a significant fraction of this gap has been attributed to parasitic shunts, and other defects, distributed across the module. In this paper, we show that it is possible to contain or isolate these shunt defects, using the state of the art laser scribing processes, after the fabrication of the series connected module is finished. We discuss three possible alternatives, and quantify the performance gains for each technique. We demonstrate that using these techniques, it is possible to recover up to 50% of the power lost to parasitic shunts, which results in 1-2% (absolute) increase in module efficiencies for typical thin film PV technologies.'
author:
- 'Sourabh Dongaonkar, and Muhammad A. Alam, [^1][^2]'
title: 'In-Line Post-Process Scribing for Reducing Cell to Module Efficiency Gap in Monolithic Thin Film Photovoltaics'
---
=1
shunts, scribing, module efficiency, yield, thin film PV.
Introduction
============
gap between cell and module efficiency is a universal feature of all PV technologies [@Green2012a]. This so called “solar gap” constitutes about 3-5% absolute reduction in module efficiency, compared to the corresponding record cell efficiency, for all technologies [@Levitan2012]. Many thin film PV (TFPV) technologies have relatively low cell efficiencies; therefore, a significant drop in efficiency at the module level, is often a critical factor limiting their commercial success \cite{}. Moreover, due to the monolithically integrated module manufacturing processes used for TFPV technologies, the problem of cell to module efficiency gap requires a different analysis and design perspective compared to conventional crystalline silicon PV [@Hegedus2003].
Several factors are responsible for the drop in power conversion efficiency from the cell to the module level. A universal loss mechanism arises from the finite sheet resistance of the transparent conducting oxide (TCO) and/or metal layers in TFPV, and the resistances of the emitter and the metal grid lines in crystalline cells. In order to minimize these losses in crystalline PV, a variety of cell and metal grid designs have been proposed [@Kerschaver2006; @Flat1979]. In monolithic TFPV modules, the resistive losses are minimized by carefully choosing the number of series connected cells, and width of each cell, for each technology type [@Johansson2008].
Other factors arise from the challenges of depositing thin films over large area ($ \sim m^2 $) substrates in TFPV module manufacturing. These include film thickness variation over large area substrates [@Yang2009b], variation of contact sheet resistance [@Hudaya2012], composition variations in chalcogenide cells [@Powalla2003]. These manufacturing challenges not only reduce the module efficiency, but also lead to performance variation at the module level, causing reduced process yield. These may be considered ‘macroscopic’ variations, because they cause performance difference over length scales of several cm, across the module surface. The development of strategies to control these macroscopic variations is the focus of significant process development and research in the TFPV community.
Besides these macroscopic sources of performance variation, there are additional efficiency losses caused by parasitic shunts [@dongaonkar:124509] and weak diodes [@Koishiyev2009e]. These ‘microscopic’ defects are formed randomly across the solar cell surface, and lead to efficiency loss, when cells are connected in series to form modules [@dongaonkar:124509; @Koishiyev2009e]. Moreover, these shunts exhibit a log-normal distribution from cell to cell, and cause significant efficiency variation at the module level, which limits the process yield [@Dongaonkar2013b]. In crystalline solar cells, it is possible to avoid the random cell efficiency fluctuations caused parasitic shunts, by sorting individual cells into efficiency bins before assembling the modules [@Wolden2011]. The monolithic manufacturing process in TFPV, however, means that it is not possible to selectively remove the shunted low efficiency cells in the module. Moreover, the wide cell geometry in TFPV module also leads to significant non-local effects of the shunts on neighboring regions. Thus, a technique of isolating these shunts, after the fabrication of a monolithically integrated module is needed in order to close the cell to module efficiency gap, and to improve process yield.
In this paper, we first describe the 2D SPICE circuit simulation setup in section II, which is used to analyze module performance in presence of parasitic shunts. Then, in section III, we illustrate the effect of random shunts on module performance and variability. Next, in section IV, we discuss the scribing isolation technique for reducing the shunt losses at module level, and use Monte-Carlo simulations to assess their effectiveness. Finally, we conclude, by discussing the practical aspects of implementation of this method.
![(a) Schematic of a typical TFPV module showing module dimensions with $ N_{series} $ series connected cells (rectangles), subdivided into $ N_{parallel} $ sub-cells (squares). (b) These sub-cells are connected using contact sheet resistances to create a 2D mesh representation of the module. (c) Each sub-cell is represented by a physical equivalent circuit for TFPV cells, including the voltage dependent photocurrent, and non-Ohmic shunt current values.[]{data-label="fig1"}](fig1){width="3.49in"}
Simulation Setup
================
In order to evaluate the module performance in presence of parasitic shunts distributed across the surface, we use a 2D circuit representation of a TFPV module [@Dongaonkar2012] , see Fig. \[fig1\](b). This distributed circuit simulation approach is necessary because, the elongated rectangle shaped cells in monolithic TFPV modules leads to important (spatial) interplay of parasitic shunts [@Pieters2011].
Figure 1(a) shows the schematic of a typical TFPV module, with $ N_{series} $ series connected cells. In order to create the circuit representation, we sub-divide each rectangular cell into $ N_{parallel} $ sub-cells. These sub-cells are represented by a physics based equivalent circuit for TFPV cells (Fig. \[fig1\](c)), and are connected in series and parallel using the contact sheet resistance values of the TCO and metal layers ($ R_S $). This equivalent circuit accounts for voltage dependent collection ($ J_{photo} $) [@Sokel1982], generation dependent recombination ($ J_{rec,ph} $) [@Merten1998], and the non-ohmic shunt current ($ J_{SH} $) [@Dongaonkar2010] components. Each sub-cell, otherwise identical, is assigned different shunt current value, drawn from the log-normal distribution [@Dongaonkar2013b]. The details of the simulation setup and validation are described in detail in [@Dongaonkar2012; @Dongaonkar]. See section S1 in supplementary materials for details regarding equivalent circuit model, and model parameters.
In the following simulations, we use single junction amorphous silicon (a-Si:H) technology as an illustrative example, for ensuring quantitative comparison of the scribing techniques. We use the record cell performance for best sub-cells in the module [@Benagli2009], and the module structure is obtained from manufacturer data sheet [@Solar2012]. This simulation setup assumes that the macroscopic sources of variability have been minimized through good process control. Therefore, we can analyze the effect of shunt variability, in particular, using realistic shunt statistics parameters from the literature [@Dongaonkar2013b].
Shunt Statistics and Module Efficiency
======================================
In this section, we discuss the effect of parasitic shunts on the module efficiency. We first analyze the effect of shunted sub-cells on their neighboring regions, by analyzing the distribution of sub-cell voltages and power outputs, obtained from 2D circuit simulations. We then simulate multiple industrial a-Si:H modules ($ L_{module}=104cm $, $ W_{module}=120cm $, and $ N_{series}=104 $) [@Solar2012], in presence of the log-normal shunt distribution, in a Monte-Carlo simulation approach, to quantify the impact of shunt statistics on module efficiency and variability.
Non-local Shunt Effect: Illustrative Example
--------------------------------------------
In order to analyze the effect of shunts in interconnected modules, we simulate a $ 11\times11cm^2 $ sub-module with 11 series connected cells. We first simulate the sub-module with all identical sub-cells, with no parasitic shunts. The resulting IV and PV curves are shown in Fig. 2(b). Note that in this case, the power output of the sub-module is the algebraic sum of the 121 sub-cell power output (including the sheet resistance losses), so that $ P_{submod}^{ideal}=\sum_{121} P_{subcell}^{ideal} = 1.0761W $.
![(a) Simulation of an $ 11\times11cm^2 $ sub-module, with 11 series connected cells and 3 shunted sub-cells (colored squares), while all other sub-cells are assumed identical. (b) IV (solid) and PV (dashed) curves of the sub-module, without (blue), and with (green) 3 shunts. The symbols denote the MPP values in each case. (c) Distribution of sub-cell voltages $ V_{sub} $ (colorbar in $ V $) at the sub-module MPP (green symbols in (b)), showing that larger shunts (number 3 and 2), lower the operating voltage of their neighboring sub-cells in parallel (highlighted) to $ \sim0.3V $ and $ \sim0.7V $, respectively. (d) The corresponding power output $ P_{sub} $ (colorbar in $ mW $), shows that the sub-cell with largest shunt (number 3), actually consumes ($ P_{sub}\approx-5mW<0 $) the power output of its neighboring regions (circled).[]{data-label="fig2"}](fig2.png){width="3.49in"}
To demonstrate the implications of shunts on ideal module performance, we now assign 3 different $ I_{SH0} $ values to 3 specific sub-cells, as highlighted in the color plot in Fig. 2(a). The resulting IV and PV curves of the sub-module with these shunts are plotted in Fig. 2(b). As expected, the power output of the sub-module drops in presence of shunts. The magnitude of efficiency loss ($ \Delta\eta\approx-7\% $), however, cannot be explained by simple sum of individual sub-cell outputs. Specifically, we note that even if we assume that the 3 shunted sub-cells produce no power at all, the sub-module output is less than the sum of the power output of the 118 ‘good’ sub-cells; i.e., $ P_{module}^{shunted} (=1.0041W) < \sum_{118}P_{sub}^{ideal} (=1.0494W) $. This means that the shunted sub-cells are modifying the operation of other good sub-cells due to their interconnection, and we must analyze this interaction carefully in order to understand the impact of shunt defects at the module level.
In order to evaluate this effect of shunts on other sub-cells, we analyze the distribution of sub-cell operating points at the sub-module maximum power point (MPP), marked in Fig. 2(b). Fig. 2(c) shows the operating points of all sub-cells plotted as color plot (color bar is in volts). Note that depending on their magnitude the operating voltage $ V_{sub} $ of the shunted sub-cells, as well as their neighbors is lowered. This is means that sub-cells in parallel to a badly shunted sub-cell, operate away from their MPP voltage, reducing their power output. The number of neighboring sub-cells which suffer from lower operating voltage due to a shunted sub-cell depends on the magnitude of the shunt, as well as the net sheet resistance of the metal/TCO contact layers (compare the highlighted cells in Fig. 2(c)). Note that this effect of shunt defects lowering the operating voltage of the entire cell has been observed experimentally using electroluminescence imaging [@Helbig2010; @Runai2011].
Another effect of the interconnection of the shunted and non-shunted sub-cells is that the shunted sub-cell may dissipate power instead of producing it, as seen in the middle cell in Fig. 2(d). In this scenario, the power output of the sub-cells, in parallel to the shunted sub-cell, is dissipated in the defective region instead of flowing to the terminals. Thus, we have a disproportionately large impact on module efficiency due to the effect of shunted sub-cells on their neighbors. First, the sub-cells adjacent to a shunted sub-cell can get biased away from their MPP and produce less power. Moreover, even this reduced power output may be dissipated in the shunted sub-cell. This is because, a heavily shunted sub-cell can get biased so that $ I_{sub} \times V_{sub} < 0 $, and therefore sinks the output current of its neighbors. As a result, even a few shunts can result in a significant reduction in the sub-module output power.
![CDF plot showing the results of Monte-Carlo circuit simulation of module efficiency (blue), for the a-Si:H technology; in presence of log-normally shunt distribution, across sub-cells. This distribution is compared to the maximum possible module efficiency (black), without any shunts, and the maximum sub-cell efficiency. The efficiency loss due to sheet resistance (red), and parasitic shunt distribution are also highlighted.[]{data-label="fig3"}](fig3){width="3.4in"}
Module Efficiency Loss in Realistic Case
----------------------------------------
The illustrative example discussed in the previous section explains how relatively few shunts can cause significant reduction in efficiency at a module level. In order to quantify these effects in realistic situations, we need to include the full statistics of shunt formation in thin film solar cells, and simulate full size commercial TFPV modules. It has been shown that shunt current distribution in TFPV cells is described by a log-normal distribution [@Dongaonkar2013b]. Each sub-cell is in the module is assigned different shunt current value, from the log-normal distribution. The results of the Monte-Carlo simulations are shown in Fig. \[fig3\], comparing the cell efficiency (red line), with the maximum module efficiency without shunts (black line), and the distribution of module efficiencies in presence of shunt variability (blue curve). This shows that if the sheet resistance loss and losses due to parasitic distribution of shunts is accounted for, it can reproduce the typical gap between cell and module efficiency ( 3% absolute here). Note that shunt variability accounts for more than half of this gap ( 2% absolute), and the rest ( 1% absolute) is caused by sheet resistance loss. It is therefore apparent that reducing the effect of parasitic shunts at the module level can significantly improve the module efficiency.
Techniques for Shunt Loss Mitigation
====================================
Process Solutions for Shunt Removal
-----------------------------------
As parasitic shunt formation is a known issue in all PV technologies, there have been many attempts towards alleviating shunt losses at cell and module level. Some of these methods have focused on process solutions for improving deposition uniformity [@Yang2009b], better substrate cleaning [@Johnston2009], or inclusion of resistive interlayers [@Ottoson2006]. Few post-process methods for shunt removal have also been reported. These include, shunt busting in a-Si:H cells [@Nostrand1979], which uses application of high reverse bias on finished cells to burn-up or ‘bust’ the shunt paths. Another approach involves electrolytic treatment of finished solar cell under reverse bias [@Nath1988], or under illumination [@Roussillon2004], for passivating the shunt forming regions, or other non-uniformities on solar cell surface.
All these techniques focus on modifying the properties of shunt paths, and reducing the leakage through those regions. Moreover, these techniques are indiscriminate, as the whole module is subjected to a particular treatment for shunt passivation, which increases the chances of adversely affecting some other cell property. Finally, these techniques rely on specific, materials, or structures, used in each TFPV technology to achieve shunt passivation, thereby limiting their application more generally. Instead, here we describe a general post-process, scribing isolation technique for mitigating shunt effects, which differs from these earlier approaches in two specific aspects. First, it relies on not removing or passivating the shunts themselves, but on reducing the effect of shunted regions on their neighboring regions. Second, this approach is applicable equally to all TFPV technologies, which use scribing and monolithic integration for module fabrication.
![(a) Schematic showing the top view of series connected cells in a module with direction of current flow as shown. (b) Side view of the TFPV module along XX’, showing the series connection created using adjacent P1/P2/P3 laser scribes (arrows). (c) Cross section along YY’ showing two isolating scribes created by overlapping the P1 and P3 lasers, to cut-off the region between the scribe lines.[]{data-label="fig4"}](fig4){width="3.49in"}
Scribing Isolation Methods
--------------------------
From the simulations in section III.A, we had seen that the significant reduction in module efficiency due to shunting arises from the non-local effect of shunted sub-cells on their neighbors. Therefore, a significant portion of shunt induced losses can be avoided if this interaction between the neighboring sub-cells can be suppressed. Fortunately, the laser scribing method used for creating the series connections in TFPV modules (see cross section XX’ in Fig. \[fig4\](b)) can be used to create insulating (open circuit) scribes, which will isolate the shunted region (see cross section YY’ in Fig. \[fig4\](c)). Note that while the series connections (Fig. \[fig4\](b)) are created using successive P1-P2-P3 laser scribes adjacent to each other; the isolating scribes are created by superposing P1 and P3 lasers to disconnect all layers (Fig. \[fig4\](c)), see [@Haas] for details.
### Periodic Lengthwise Scribing
A straightforward method for shunt isolation using scribing involves creating full length isolating scribes in the direction of current flow in a TF module. These vertical scribes divide the rectangular cells in the module into columns of smaller width, each of which are connected to the common bus bar (see the schematic in Fig. \[fig5\](a)). This ensures that the effects of shunted sub-cells are restricted to fewer neighboring sub-cells (one for the case shown in Fig. 2(a)). This approach does not require identification of the shunt locations beforehand. The scribe lines, however, need to span the entire module length, which ensures that regardless of the shunt location, the number of neighboring sub-cells in parallel it can influence is limited. Note that this full length scribing isolation at the sub-module level has been demonstrated earlier using mechanical scribing for edge shunt isolation [@Pern2011]. In this work, we propose to extend this approach, and create periodic scribes for constraining shunt effect at all locations.
### Selective Scribing Isolation
It is apparent from the previous section that lengthwise scribing will require very long scribe lines for a typical sized TFPV module. It would therefore be beneficial to first identify the shunts, and scribe only around the shunted sub-cells. There are a variety of imaging methods, which can be used to identify the spatial location of shunts. These can either be contactless methods like PL [@Johnston2010a] or EBIC [@Yacobi1984] imaging, contact methods like EL [@Helbig2010] or DLIT [@Rosch2012] imaging, or a combination thereof. Most of these techniques have been studied extensively, and are being actively adapted for inline metrology purposes [@Hoppe2010]. Once the images have been used to identify shunt location, an image processing routine can automatically extract their coordinates [@Miranda2010], which can then be used to position the scribe heads in the right location.
Once the shunt locations are known, we can either choose to partially isolate, or fully isolate the shunted sub-cell. The partial isolation is created using 2 isolating scribe lines in parallel to the direction of current flow, either side of the shunted sub-cell (schematic in Fig. \[fig5\](b)). These lines ensure that the shunted sub-cell is disconnected from its neighboring sub-cells in parallel, while maintaining the current continuity through the series connected cells. Note that the sub-cells in series with the shunted ones, which are inside the isolating scribes, will now operate away from their MPP voltage, due to the condition of current continuity. This partial isolation, however, does not require very precise determination of shunt location, or positioning of the scribing heads, and will be easier to implement in practice.
![Sub-module schematics showing the three shunted sub-cells, compare the different scribing isolation schemes (isolating scribes shown with dashed lines). (a) Full length isolating scribes at periodic intervals, in the direction of current flow (arrows). (b) Partial isolation using parallel scribes, after shunt location detection, with different scribe lengths providing different degree of isolation. (c) Full isolation of shunted sub-cells using box scribes encircling the shunted region inside the sub-cell.[]{data-label="fig5"}](fig5){width="3.49in"}
The second option isolates the shunt fully by enclosing the shunt using isolating scribes (see schematic in Fig. \[fig5\](c)). This ensures full isolation, at the cost of greater area penalty, as all the area enclosed by the scribe lines is disconnected from the module and cannot produce output power. This requires a more precise identification of shunt locations, as well as more accurate positioning of scribe heads, in order to ensure that the shunt is in fact enclosed by the scribes. The shunts in this case, however, are fully isolated, and do not affect any sub-cells in series or parallel after scribing.
From the previous discussion, it appears that this scribing isolation approach simply exchanges shunt losses with dead area loss. In section III.A, however, we saw that the largest shunts have a disproportionate impact on module efficiency. And, due to the log-normal nature of shunt statistics, only a small fraction of sub-cells are heavily shunted [@Dongaonkar2013b]. Therefore, we only need to isolate a small fraction of the total shunts (largest 3%), in order to achieve significant performance recovery. This is the reason why the scribing isolation can recover significant shunt losses, at a relatively small increase in dead area loss due to few extra scribe lines.
Comparison of Scribing Isolation Schemes
========================================
![(a) Color plot of sub-module $ V_{sub} $ (color bar in $ V $), in presence of full length isolating vertical scribe lines (dotted), showing that the operating voltages of the neighboring sub-cells in parallel improve with scribing. The non-shunted sub-cells inside the scribes are slightly forward biased to ensure series connected conduction. (b) Distribution of $ P_{sub} $ values (color bar in $ mW $), showing that majority of sub-cells now operate at MPP, and the power output in shunted sub-cells is also improved.[]{data-label="fig6"}](fig6){width="3.3in"}
Sub-module Operation with Shunt Isolation
-----------------------------------------
In this section we revisit our illustrative example of an $ 11\times11cm^2 $ sub-module with 3 shunts. We will apply the various scribing schemes on this sub-module, and analyze the resulting voltage and power distributions carefully to gain insight into how these schemes improve module performance.
### Periodic Lengthwise Scribing
We first analyze the case of full length periodic scribe lines at the sub-module level, as shown in Fig. \[fig5\](a). The $ V_{sub} $ and $ P_{sub} $ distributions obtained from the simulations are shown in Fig. \[fig6\](a) and Fig. \[fig6\](b), respectively. Note that in this scenario, the effective range of all shunted sub-cells is restricted to only one neighboring sub-cell in parallel. Consequently, the operating voltages and power outputs of the ‘good’ sub-cells improves in general. Note, however, that due to series connection, the voltage of each column of sub-cells (between two isolating scribes) must be equal. Therefore, the operating points of the sub-cells in series to the shunted sub-cells now must move to slightly higher $ V_{sub} $ values (see columns containing shunted sub-cells marked (2) and (3)). In spite of this limitation, note that the blocking of parallel conduction paths, results in all sub-cells producing power (i.e. $ P_{sub}>0 $ for all sub-cells in Fig. \[fig5\](b)). Therefore, we can see that this approach will lead to improved sub-module output.
### Selective Scribing - Partial Isolation
If the shunt positions are known beforehand, we can use parallel isolating scribes on either side, of varying lengths ($ L_{scribe} $). This ensures that the non-shunted sub-cells are not unnecessarily scribed. Fig. \[fig7\](a) shows the color plot $ V_{sub} $ values from the sub-module simulation, in presence of the scribes shown in Fig. \[fig5\](b). Note that the operating points of sub-cells in parallel to shunted sub-cells improve close to MPP values. The $ V_{sub} $ values of the sub-cells in series to the shunted sub-cells however, now increase slightly to ensure total sub-module voltage being same. The sub-cell power outputs shown in Fig. \[fig7\](b) also show the reduced impact of shunted regions on neighboring sub-cells in parallel, at the cost of slight reduction in $ P_{sub} $ values for the sub-cells in series with the shunted sub-cells. Finally, note that the length of scribe lines used for isolation affects the degree of isolation, and accordingly, the efficiency improvement obtained, at the cost of higher dead area loss.
![(a) Color plot of sub-module $ V_{sub} $ (color bar in $ V $), in presence of isolating vertical scribe lines of different length (dotted), showing that the operating voltages of the neighboring sub-cells in parallel are back to MPP values. The non-shunted sub-cells inside the scribes are slightly forward biased to ensure series connected conduction. (b) Distribution of $ P_{sub} $ values (color bar in $ mW $), showing that majority of sub-cells now operate at MPP, and the power output in shunted sub-cells is also improved. Both these plots show that longer scribes allow better isolation, and better sub-cell characteristics.[]{data-label="fig7"}](fig7){width="3.3in"}
### Selective Scribing - Full Isolation
Finally, we analyze the case with full shunt isolation, achieved by enclosing them completely using isolating scribe (as shown in Fig. \[fig5\](c)). This approach leads to slightly higher dead area loss, since the region inside the scribe lines will not contribute to the output. The sub-module simulation results with full isolation are shown in Fig. \[fig8\]. Note that the isolation around the shunts reduces the voltage discrepancy across the different sub-cells considerably (Fig. \[fig8\](a)). The isolated region is shown by dark squares, as it is completely cut-off from the module output. Correspondingly, all the non-shunted sub-cells now operate near their MPP (Fig. \[fig7\](b)). Depending on the total area inside the scribed region ($ A_{scribe} $), some portion of the shunted sub-cells will still be generating power (see Fig. \[fig7\](b)), and the fraction of area lost to scribing per cell ($ A_{scribe}/A_{sub} $) will determine the eventual benefit of this scheme.
![(a) Color plot of sub-module $ V_{sub} $ (color bar in $ V $), with full isolation of the three shunts (black squares), showing that the operating voltages of almost all sub-cells are restored to MPP values. (b) Distribution of $ P_{sub} $ values (color bar in $ mW $), showing that all non-shunted sub-cells now operate at MPP, and the power output in shunted sub-cells is also improved considerably. These plots show that the area mismatch caused by full isolation has a limited impact on sub-module performance as a whole.[]{data-label="fig8"}](fig8){width="3.3in"}
Scribing Isolation for Module Efficiency Improvement
----------------------------------------------------
From the previous discussion, we can see that various scribing techniques improve the module performance, by isolating the shunts to varying degrees, but the trade-off in terms of dead area loss is also different. We now analyze these tradeoffs more quantitatively, for typical size commercial modules with parasitic shunt distribution. We consider the highest efficiency a-Si:H module ($ \eta=7.22\% $) from the CDF in Fig. \[fig3\], which has a certain distribution of shunts on the surface responsible for an efficiency loss of $ \sim2\% $ (absolute). We will apply the various scribing schemes to this particular module to evaluate the improvement in efficiency obtained. This will allow us to evaluate the effectiveness of these schemes, under realistic circumstances. In addition to the typical module and sub-cell parameters used in section III.B, we use a realistic value for isolating scribe lines to be $ 150\mu m $ [@Haas]. Note that the dead area loss in the scribed sub-cells is accounted for in the SPICE simulation by modifying the active area of each cell, based on the presence of isolating scribes. This also ensures that any effects due to mismatches in sub-cell IV are accounted for when evaluating the effect of scribing schemes.
![Plot comparing the efficiency of a typical 7.22% a-Si:H module with parasitic shunts (horizontal line), after the module is treated using different scribing isolation techniques. For the lengthwise isolating scribes (triangles), without detection, the efficiency improves with more vertical scribe lines ($ L_{scribe} $), but the efficiency gains remain modest due to large dead area penalty. The efficiency enhancements are significantly better if selective scribing is used to isolated more and more shunted sub-cells (increasing $ N_{iso}/N_{sub} $) for both partial (circles) and full (squares) isolation case. In case of partial isolation, the efficiency improves for larger $ L_{scribe} $, due to better isolation, but ultimately limited by dead area losses (squares). For full isolation, the dead area losses are determined by the ratio $ A_{scribe}/A_{sub} $, which limits the efficiency enhancements obtained due to reduced shunt effects.[]{data-label="fig9"}](fig9){width="3.49in"}
For the lengthwise scribing, it is apparent from Fig. \[fig6\] that an increase in the number of vertical scribes will improve the module efficiency by containing the effect of shunted sub-cells. This, however, requires very long scribe lines, so the dead area penalty also increases, and the incremental gains in module efficiency with increasing number of scribes ($ N_{scribe} $) is positive but modest. As shown by the triangles in in Fig. \[fig9\], the module efficiency improves to $ \sim7.6\% $ for 30 full length vertical scribes. This efficiency improvement, however, rises slowly with increasing $ N_{scribe} $ to larger values.
In case of targeted shunt isolation after determination of shunt locations, we need to choose the number of shunted sub-cells to isolate. This is also determined by the accuracy of the detection scheme. As expected as we isolate more and more shunted sub cells, out of the total $ N_{sub} $ (i.e. increasing $ N_{iso}/N_{sub} $ in Fig. \[fig9\]), the module efficiency improves significantly (see circles for partial isolation, and squares for full isolation in Fig. \[fig9\]). As expected, the improvements are limited by increased dead area loss with additional scribing, and the curves saturate for larger $ N_{iso}/N_{sub} $ values. Interestingly, note that using the prior detection of shunts, it is possible to get $ \eta_{module} \geq 7.6\% $ by partially or fully isolating just the largest 1% shunt defects. This is a consequence of the heavy tailed log-normal shunt distribution, which means that few largest shunts dominate the overall efficiency loss at the module level (see [@Dongaonkar2013b] for details). Therefore, we can obtain significant improvements by focusing on these most critical parasitic defects.
In case of partial isolation technique another optimization parameter is the length of isolating scribe $ L_{scribe} $, used for each shunted sub-cell. As expected from Fig. \[fig7\], longer scribe lines provide better isolation, and the efficiency improves for the same $ N_{iso}/N_{sub} $ value by increasing $ L_{scribe} $ from 1 to 9 cm in steps of 2 cm (see circles in Fig. \[fig9\]). Understandably, the dead area loss limits these gains, and therefore, the module efficiency for $ L_{scribe}=7cm $ vs. $ L_{scribe}=9cm $ are virtually identical.
In case of full isolation of shunts, the efficiency gains will be limited by the ratio between scribed sub-cell area, and total sub-cell area $ A_{scribe}/A_{sub} $. In a practical scenario, this ratio will depend on the accuracy of the shunt detection, and scribing head positioning. We, therefore, vary the dead area loss ($ A_{scribe}/A_{sub} $) from 20% to 80% (see squares in Fig. \[fig9\]), and show that regardless of the dead area loss associated with full scribing, the full shunt isolation provides highest module efficiency out of the 3 suggested techniques. This is because, by fully isolating the shunts this approach removes the effect of shunted sub-cells on all its neighbors (series or parallel). Furthermore, the random distribution of shunts on a module surface ensures that the mismatches introduced by scribing area loss are not significant.
Scribing Isolation and Module Yield
-----------------------------------
Based on the discussion so far, we observe that all three scribing techniques enable improvement in module efficiency. We now show that these results hold true in a statistical sense, and compare the change in module efficiency distribution for different scribing methods. In order to compare the techniques statistically, we repeat the Monte-Carlo simulation of module efficiency in section III.B, applying the different scribing techniques for each module. For the lengthwise periodic scribing, we choose full length scribes every 5 cm; and, for the selective scribing schemes we choose$ N_{iso}/N_{sub}=3\% $. For the partial isolation method, $ L_{scribe}=5cm $, and $ A_{scribe}/A_{sub}=0.4 $ for the full isolation case. These numbers are chosen because, from the analysis in the previous section, we note that the incremental gain in module efficiency with scribing more shunts or using longer scribe is minimal. Moreover, these values are reasonable from a practical standpoint of implementing this strategy in a production environment. In practice, these quantities need to be optimized, depending upon the technology of choice, the shunting and scribing parameters of the process, and the required efficiency improvement.
![CDF plots of module efficiency distribution for typical case (blue), lengthwise periodic scribing (magenta), partial shunt isolation after detection (green), and full shunt isolation after detection (red). The improvement in mean efficiency ($ \bar{\eta}_N $ ) ̅, and reduction in standard deviation $ s_N $ are apparent from the values shown in inset. The full isolation of shunts can yield efficiency values close to the maximum possible module efficiency (black).[]{data-label="fig10"}](fig10){width="3.49in"}
Fig. \[fig10\] compares the CDFs of module efficiency without any scribing vs. each scribing method. Note that as expected from previous sections, the mean module efficiency $ \bar{\eta} $ improves by $ \sim0.3\% $ (absolute) for lengthwise periodic scribing, $ \sim0.7\% $ (absolute) for partial scribing and $ \sim1.1\% $ (absolute) for full scribing isolation method. Interestingly, note that simply by isolating only 3% of the shunts, and incurring significant area penalty for each isolating box, we can bring the module efficiency very close to “ideal” module efficiency of 8.9%, which is obtained for the case with no shunt defects. Equally important, the scribing isolation of shunts also leads to reduced standard deviation of the module efficiency distribution, see Fig. \[fig9\]. This means that in addition to the overall module efficiency improvement, the scribing techniques also reduce the performance variability associated with random shunt formation. This in turn will improve the overall yield of the manufacturing process, which is an important factor in PV manufacturing cost.
Conclusion
==========
In this paper we demonstrate an in-line method for module efficiency improvement in monolithic TFPV modules, by isolation of parasitic shunt defects using laser scribing. The approach utilizes state of the art techniques of shunt identification, and laser scribing, to isolate a small fraction of the largest shunts in the module. We describe two techniques, which involve either partial isolation, or full isolation of the shunt defects. We used 2D, self-consistent, SPICE circuit simulations of the modules in presence of shunts, to compare these techniques quantitatively and illustrate the involved optimizations. We demonstrate using typical parameters that these scribing techniques can recover more than half of the module efficiency loss caused by parasitic shunts.
Finally, we would like to emphasize that these techniques are equally applicable to all monolithic TFPV technologies, and owing to the similarities in shunt statistics [@Dongaonkar2013b], are expected to yield similar performance gains at the module level. Indeed, while this analysis is presented for effect of shunt leakage, the isolation scheme will be equally beneficial for any other type of parasitic localized leakage/non-uniformity, e.g., weak diodes [@Koishiyev2009e], which can affect their neighboring regions adversely. The implementation of these techniques in realistic production line, however, will have to be analyzed on a case by case basis; including the degree of shunt or other leakage losses, and economics of implementing an extra process step. In this regard, the scribing techniques is advantageous, as they only require the state-of-the-art manufacturing tools for implementation, and only add one extra, post-process step to the flow.
We believe that the potential advantages of these schemes for module efficiency and process yield enhancement, coupled with relatively straightforward implementation, and applicability to different thin film technologies [@Dongaonkar2013a], will motivate the PV community to explore these approaches in production environments.
Acknowledgment {#acknowledgment .unnumbered}
==============
We would like to acknowledge the Network for Computational Nanotechnology for the computational resources used in this work. We would like to thank B. J. Stanbery for helpful discussions.
[10]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{}
M. A. Green, K. Emery, Y. Hishikawa, W. Warta, and E. D. Dunlop, “[Solar cell efficiency tables (version 40)]{},” *Progress in Photovoltaics: Research and Applications*, vol. 20, no. 5, pp. 606–614, Aug. 2012.
D. Levitan, “[The Solar Efficiency Gap - IEEE Spectrum]{},” 2012. \[Online\]. Available: <http://spectrum.ieee.org/green-tech/solar/the-solar-efficiency-gap>
“[Low cost high-efficiency amorphous silicon solar cells with improved light-soaking stability]{}.” \[Online\]. Available: <http://dx.doi.org/10.1016/j.solmat.2011.11.023>
S. S. Hegedus and A. Luque, *[Handbook of photovoltaic science and engineering]{}*.1em plus 0.5em minus 0.4emChichester \[u.a.\]: Wiley, 2003.
E. V. Kerschaver and G. Beaucarne, “[Back-contact solar cells: a review]{},” *Progress in Photovoltaics: Research and Applications*, vol. 14, no. 2, pp. 107–123, Mar. 2006.
A. Flat and A. Milnes, “[Optimization of multi-layer front-contact grid patterns for solar cells]{},” *Solar Energy*, vol. 23, no. 4, pp. 289–299, Jan. 1979.
J. Johansson, “[Modelling and Optimization of CIGS Solar Cell Modules]{},” Ph.D. dissertation, Lunds University, 2008.
F. Yang, L. Zhang, T. Tanaka, Y. Zheng, F. Schimitt, R. Wang, A. Tso, K. Reinke, N. Hammond, M. Pandit, L. Li, L. Tsuei, Z. Yuan, and B. Shieh, “[Uniform growth of a-Si / $\mu$c-Si Tandem junction solar cells over 5.7m2 substrates]{},” in *Photovoltaic Specialists Conference (PVSC), 2009 34th IEEE*, 2009, pp. 1541–1545.
C. Hudaya, J. H. Park, and J. K. Lee, “” **, vol. 7, no. 1, p. 17, Jan. 2012.
M. Powalla, D. Hariskos, E. Lotter, M. Oertel, J. Springer, D. Stellbogen, B. Dimmler, and R. Schäffler, “[Large-area CIGS modules: processes and properties]{},” *Thin Solid Films*, vol. 431-432, pp. 523–533, May 2003.
S. Dongaonkar, J. D. Servaites, G. M. Ford, S. Loser, J. Moore, R. M. Gelfand, H. Mohseni, H. W. Hillhouse, R. Agrawal, M. A. Ratner, T. J. Marks, M. S. Lundstrom, and M. A. Alam, “[Universality of non-Ohmic shunt leakage in thin-film solar cells]{},” *Journal of Applied Physics*, vol. 108, no. 12, p. 124509, 2010.
G. T. Koishiyev and J. R. Sites, “[Effect of weak diodes on the performance of CdTe thin-film modules]{},” in *2009 34th IEEE Photovoltaic Specialists Conference (PVSC)*.1em plus 0.5em minus 0.4emIEEE, Jun. 2009, pp. 001978–001981.
S. Dongaonkar, S. Loser, E. J. Sheets, K. Zaunbrecher, R. Agrawal, T. J. Marks, and M. A. Alam, “,” **, vol. 6, pp. 782–787, 2013.
C. A. Wolden, J. Kurtin, J. B. Baxter, I. Repins, S. E. Shaheen, J. T. Torvik, A. A. Rockett, V. M. Fthenakis, and E. S. Aydil, “[Photovoltaic manufacturing: Present status, future prospects, and research needs]{},” *Journal of Vacuum Science & Technology A: Vacuum, Surfaces, and Films*, vol. 29, p. 030801, 2011.
S. Dongaonkar and M. A. Alam, “[End to end modeling for variability and reliability analysis of thin film photovoltaics]{},” in *2012 IEEE International Reliability Physics Symposium (IRPS)*.1em plus 0.5em minus 0.4emIEEE, Apr. 2012, pp. 4A.4.1–4A.4.6.
B. E. Pieters, “[Spatial Modeling of Thin-Film Solar Modules Using the Network Simulation Method and SPICE]{},” *IEEE Journal of Photovoltaics*, vol. 1, no. 1, pp. 93–98, Jul. 2011.
R. Sokel and R. Hughes, “[Numerical analysis of transient photoconductivity in insulators]{},” *Journal of Applied Physics*, vol. 53, no. 11, p. 7414, 1982.
J. Merten, J. Asensi, C. Voz, A. Shah, R. Platz, and J. Andreu, “[Improved equivalent circuit and analytical model for amorphous silicon solar cells and modules]{},” *IEEE Transactions on Electron Devices*, vol. 45, no. 2, pp. 423–429, 1998.
S. Dongaonkar, K. Y, D. Wang, M. Frei, S. Mahapatra, and M. A. Alam, “[On the Nature of Shunt Leakage in Amorphous Silicon p-i-n Solar Cells]{},” *IEEE Electron Device Letters*, vol. 31, no. 11, pp. 1266–1268, Nov. 2010.
S. Dongaonkar, C. Deline, and M. A. Alam, “[Performance and Reliability Implications of Two Dimensional Shading in Monolithic Thin Film Photovoltaic Modules]{},” *IEEE Journal of Photovoltaics*, 2013.
S. Benagli and D. Borrello, “[High-efficiency amorphous silicon devices on LPCVD-ZnO TCO prepared in industrial KAI-M R&D reactor]{},” in *24th European Photovoltaic Solar Energy Conference and Exhibition*, no. September, 2009, pp. 21–25.
“[Moserbaer Solar, Power Series QS: Amorphous Silicon Thin Film Modules]{}.” \[Online\]. Available: [http://www.moserbaersolar.com/resources/thinfilms-english QS-latest.pdf](http://www.moserbaersolar.com/resources/thinfilms-english QS-latest.pdf)
A. Helbig, T. Kirchartz, R. Schaeffler, J. H. Werner, and U. Rau, “[Quantitative electroluminescence analysis of resistive losses in Cu(In,Ga)Se2 thin-film modules]{},” *Solar Energy Materials and Solar Cells*, vol. 94, no. 6, pp. 979–984, Jun. 2010.
F. R. Runai, F. Schwable, T. Walter, A. Fidler, S. Gorse, T. Hahn, and I. Kotschau, “[Imaging and performance of CIGS thin film modules]{},” in *Photovoltaic Specialists Conference (PVSC), 2011 37th IEEE*, 2011, pp. 3399–3403.
S. W. Johnston, N. J. Call, B. Phan, and R. K. Ahrenkiel, “[Applications of imaging techniques for solar cell characterization]{},” in *34th IEEE Photovoltaic Specialists Conference (PVSC)*.1em plus 0.5em minus 0.4emIEEE, Jun. 2009, pp. 000276–000281.
K. Ottoson, “[The role of i-ZnO for shunt prevention in Cu(In,Ga)Se2-based solar cells]{},” Ph.D. dissertation, Uppsala, 2006. \[Online\]. Available: [internal-pdf:/2006\_Karin\_Ottosson - Thesis.pdf](internal-pdf:/2006_Karin_Ottosson - Thesis.pdf)
G. Nostrand and J. Hanak, “[Method of removing the effects of electrical shorts and shunts created during the fabrication process of a solar cell]{},” Sep. 1979. \[Online\]. Available: [www.google.com/patents/US4166918](www.google.com/patents/US4166918)
P. Nath, K. Hoffman, C. Vogeli, and S. R. Ovshinsky, “[Conversion process for passivating current shunting paths in amorphous silicon alloy solar cells]{},” *Applied Physics Letters*, vol. 53, no. 11, p. 986, 1988.
Y. Roussillon, D. M. Giolando, D. Shvydka, A. D. Compaan, and V. G. Karpov, “[Blocking thin-film nonuniformities: Photovoltaic self-healing]{},” *Applied Physics Letters*, vol. 84, pp. 616–618 ST – Blocking thin–film nonuniformities:, 2004.
S. Haas, S. Krumscheid, A. Bauer, A. Lambertz, and U. Rau, “,” **.
F. J. Pern, F. Yan, L. Mansfield, S. Glynn, M. Rekow, and R. Murison, “[Performance characterization and remedy of experimental CuInGaSe<inf>2</inf> mini-modules]{},” in *Photovoltaic Specialists Conference (PVSC), 2011 37th IEEE*, 2011, pp. 2792–2797.
S. Johnston, I. Repins, N. Call, R. Sundaramoorthy, K. M. Jones, and B. To, “[Applications of imaging techniques to Si, Cu(In,Ga)Se<inf>2</inf>, and CdTe and correlation to solar cell parameters]{},” in *2010 35th IEEE Photovoltaic Specialists Conference*.1em plus 0.5em minus 0.4emIEEE, Jun. 2010, pp. 001727–001732.
B. Yacobi, T. McMahon, and A. Madan, “[Electron-beam-induced current microcharacterization of fabrication defects in hydrogenated amorphous silicon solar cells]{},” *Solar Cells*, vol. 12, no. 3, pp. 329–335, 1984.
R. Rösch, F. C. Krebs, D. M. Tanenbaum, and H. Hoppe, “[Quality control of roll-to-roll processed polymer solar modules by complementary imaging methods]{},” *Solar Energy Materials and Solar Cells*, vol. 97, pp. 176–180, Feb. 2012.
H. Hoppe, J. Bachmann, B. Muhsin, K.-H. Drüe, I. Riedel, G. Gobsch, C. Buerhop-Lutz, C. J. Brabec, and V. Dyakonov, “[Quality control of polymer solar modules by lock-in thermography]{},” *Journal of Applied Physics*, vol. 107, no. 1, p. 014505, 2010.
E. Miranda, E. O’Connor, and P. Hurley, “[Analysis of the breakdown spots spatial distribution in large area MOS structures]{},” in *2010 IEEE International Reliability Physics Symposium*.1em plus 0.5em minus 0.4emIEEE, 2010, pp. 775–777.
S. Dongaonkar and M. A. Alam, “[Thin Film Photovoltaic Panels and Repair Methods]{},’US Patent Application 13/761,914’ 2013.
[^1]: S. Dongaonkar and M. A. Alam are with the School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN, 47907 USA e-mail: [email protected], [email protected].
[^2]: Manuscript received April 19, 2005; revised December 27, 2012.
|
{
"pile_set_name": "ArXiv"
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.