Stationary properties of the Gauss-Galerkin QMoM truncation of MV-SDEs

A. Alecio · 2025

Read Stationary properties of the Gauss-Galerkin QMoM truncation of MV-SDEs by A. Alecio (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-00605.txt.

Introduction

Alexander Alecio studies steady states of a moment-closure scheme for random equations. The Gauss-Galerkin quadrature method replaces a probability law by a finite moment set. Whether closed equations keep the true steady states was unclear. Alecio settles it without leaning on convergence results.

For polynomial mean-field equations, the closure keeps as many steady states as the potential has peaks and valleys, not just its minima. The full bifurcation picture survives truncation. A scaling property lets stability changes be probed directly. The Dawson-Shiino model serves as the test case, matching earlier studies of its steady measures. Moment methods can thus be trusted with multistable systems.

Catalog record

CREATORA. Alecio
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-00605

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Stationary properties of the Gauss-Galerkin QMoM truncation of MV-SDEs A. Alecio Abstract . The Gauss Galerkin Method/Quadrature method of moments (GG-QMoM) closure scheme, introduced by Dawson [ 7 ] , closes a truncated set of moment equations of an SDE by a Galerkin approximation of its law in the space of probability measures. Here, results are presented on stationary solutions of the closed equations, irrespective of the number of moments retained (thus not dependent on the convergence theorem). These are applied to polynomial MV-SDEs, with explicit dependence on its moments in the drift, which can possess multiple stationary solutions. Particularly, we show as σ ↓ 0 \sigma\downarrow 0 there are as many stationary solutions as extrema of the potential (critically, not just the minima), preserving the bifurcation diagram of the full equations. Further, a scaling property of solutions is proven, allowing changes of stability to be directly probed. Finally, this is applied to the GG-QMoM truncation of Dawson-Shiino model, whose stationary measures and change in stability has been previously studied. 1991 Mathematics Subject Classification 60H10; 60G10; 60J60; 82C22; 35Q83; 35Q84 For the one-dimensional time-homogenous stochastic differential equation (1) dX_{t}=a(X_{t})dt+\sigma b(X_{t})dW_{t},\quad\sigma\in\mathbb{R},|b|>0 a typical problem would be to calculate its aggregate properties from this path-wise description. This work is concerned with a method, first introduced in [ 7 ] , to calculate the moments of the law of process ( 1 ). Suppose SDE ( 1 ) has polynomial coefficients, and possesses a unique, possibly weak, solution. The evolution of moments of its law can be calculated applying Itô’s lemma with x^{i} , and taking the expected value: \dot{m}_{i}=\mathbb{E}(dX^{i}_{t})=i\mathbb{E}[a(X_{t})X_{t}^{i-1}dt+\sigma^{2}\frac{(i-1)}{2}b(X_{t})X_{t}^{i-2}dt+\sigma b(X_{t})X_{t}^{i-1}dB_{t}] The expectation of the martingale (final) term is null and, as a,\,b are polynomials in X_{t} , we are left with a polynomial in moments of the law. Equivalently (2) \dot{m}_{i}=\int\mathsf{L}x^{n}\rho(t)\mathrm{d}x where \mathsf{L}=a\partial_{x}+\frac{\sigma^{2}}{2}b^{2}\partial_{xx} , the generator of process ( 1 ). In this way, SDE ( 1 ) is converted into a denumerable set of ODEs, the moment evolution equations (MEEs), (3) \dot{m}_{i}=f_{i}(m_{1},\dots,m_{i+j}),\quad j>0 which are row finite ( j<\infty is independent of i ), where f_{i} are all polynomial functions. In the ‘lower-diagonal’ case, (j=0) , an arbitrary moment can be calculated with knowledge the lower order moments only. For instance, the MEEs of the Ornstein-Uhlenbeck process dX_{t}=-\alpha X_{t}dt+\sigma dW_{t} are lower diagonal with explicit solution (4) \begin{split}&m_{n}(t)=e^{-\alpha nt}m_{n}(0)+\frac{\sigma^{2}}{2}n(n-1)\int_{0}^{t}e^{-\alpha n(t-2)}m_{n-2}(s)ds\\ \end{split} Critically, as the equation for m_{1} is closed, the moments can be calculated to an arbitrary order. In the general row-finite case, it is not typically possible to find a closed expression for a finite number of moments as we did in ( 4 ). Moment closure methods circumvent this by replacing m_{n+j},\,j>0 in the n^{th} and lower order MEEs with closure functions g_{n+j}(m_{1},\dots,m_{n}) . The resulting lower diagonal system of equations are (5) \dot{m}_{i}=f_{i}(m_{1},\dots,m_{i},g_{i+1}(m_{1},\dots,m_{i}),\dots,g_{i+j}(m_{1},\dots,m_{i})) the first n of which we call n - closed moment evolution equations ( n -MEEs), and its solution the n - closed moments . The choice of g_{n+j}(m_{0},\dots,m_{n}) significantly affects the behaviour of the n -closed moments. A number of truncation/closure schemes have been developed [ 4 ] , including Gauss-Galerkin QMoM [ 7 ] . The priority of GG-QMoM is to close the MEEs in a way that guarantees the Hankel matrix of n -closed moments is positive semidefinite (a necessary condition of the moments of any probability measure), following a fotiori as the n -closed moments correspond to an actual probability measure. This measure (the n -approximant) is an approximation of the law of SDE ( 1 ) by a superposition of n -point masses with varying intensities \beta_{i} and nodes, x_{i} . Initialised with the Gauss-Christoffel approximation of the initial measure, the moments are evolved according to the first 2n moment evolution equations ( 3 ). Higher order moments are entirely determined by the lower, closing the system. In effect, this is a Galerkin approximation to the weak form of the Fokker-Planck equation, where the approximant is in a subset of the space of probability measures P_{M}(\mathbb{R}) for which polynomial test functions consitute a seperating set 1 1 1 More precisely, Definition B.1 [ 5 ] . With knowledge of the first n moments, [ 7 , 12 ] give explicit forms for g_{n+1} for f(X_{t}) as the solution of h_{n+1}=0 , where f is a positive diffeomorphism, and h_{n} the n^{th} Hankel determinant of the law of f(X_{t}) . Practically, [ 6 ] provides an efficient algorithm for the resolution of the GG-QMoM n -MEEs. For this work, rather than directly resorting to the moment equations, equations are derived for \beta_{i},\,x_{i} using the transformation m_{i}=\sum_{j}\beta_{j}x_{j}^{i} . The resulting Jacobian is an invertible confluent Vandermonde matrix [ 10 ] , from which the Gauss-Galerkin equations can be derived. Not only does GG-QMoM systematically yield positive definite n- closed moments, but the associated approximant \tilde{\rho}_{n} converges to the law of the true solution as more moment evolution equations are retained. Theorem 0.1 (Convergence of GG-QMoM approximates: Campillo, Dawson [ 6 , 7 ] ) . Suppose there exists constants \{k_{n}:n\geq 1\} and \theta_{0}>0 such that (6) \displaystyle m_{n}(t)\leq k_{n}\quad\mathrm{for}\quad 0\leq t\leq T (7) \displaystyle\sum_{n=1}^{\infty}\theta^{n}\frac{k_{n}}{n!}<\infty\quad\mathrm{for\,all}\quad 0<\theta<\theta_{0} for the MEEs associated to process X . Further, for suitable initial probability measure suppose process X possesses a unique law. Then \tilde{\rho}_{n}(t)\Rightarrow\rho(t) as N\rightarrow\infty , where \rho_{t}=\mathrm{Law}(X_{t}) , in the sense of weak convergence of probability measures. Proof. This is Theorem 5.1 in [ 7 ] . See also [ 6 ] . Both rely on the traditional trilogy of demonstrating relative compactness of approximants, identifying the limit, and proving uniqueness. ∎ 0.1. MV-SDEs and Outline McKean-Vlasov SDEs [ 14 ] (MV-SDEs) are SDEs with explicit dependence on its law. MV-SDEs with explicit dependence on moments of the law and polynomial drift and diffusion such as (8) dX_{t}=\big(-\bar{V}^{{}^{\prime}}(X_{t})+\theta\mathbb{E}[P^{{}^{\prime}}(X_{t})]\big)dt+\sigma k(X_{t})dB_{t} will yield row-finite MEEs, thus moment truncation schemes are equally applicable. Many foundational works in the field such as [ 9 , 13 ] took this approach, and indeed has been reprised by later ones, [ 3 ] . MV-SDEs ( 8 ) are known to exhibit critical behaviour - multiple stationary measures, the number of which is a function of the parameters of ( 8 ). MV-SDE ( 8 ) with the simple bistable potential and quadratic interaction ( \bar{V}^{{}^{\prime}}=x^{3} , P^{{}^{\prime}}=x ), extensively studied in [ 8 , 15 ] , is used to test the efficacy of several truncation methods fixed number of moments ( N=4 ) in [ 18 ] . Broadly, there are two factors which favour truncation schemes to study critical behaviour of MV-SDEs ( 8 ). Stationary solutions with our assumptions, are unimodal which happen to be particularly compatible with simple truncation schemes. Further, they all preserve low order MEEs, the first of which was demonstrated in [ 2 ] to be equivalent to the self-consistency equation, an auxiliary implicit integral equation stationary measures must satisfy. In an unpublished work, the author similarly studied GG-QMoM n -MEEs of this model and the related model studied in [ 1 ] , finding it performed comparably or outperformed the schemes of [ 18 ] , even in difficult parameter regimes. Particularly, the pitchfork bifurcation diagram of attainable stationary measures was well preserved

Read the complete text