Multivariate approximations in Wasserstein distance by Stein's method and Bismut's formula (Q2312685)

Let \(X_t\) be the solution of the additive SDE \[dX_t=g(X_t)+\sqrt dB_t.\] Under the assumptions on the upper bound of the Hessian of \(g\) and the assumption that \(g\) has a contractive property, this Langevin SDE has a unique invariant probability measure \(\mu\). Take, e.g., \(g(x)=-c(1+|x|^2)^(p/2)x\). Let \(W\) be a random variable satisfying the following condition: there exists a copy \(W'\) of \(W\) such that \(E(W'-W|W)=\lambda (g(W)+R_1)\), \(E((W'-W)(W'-W)^T|W)=2\lambda (Id+R_2)\). The authors obtain an upper bound for the Wasserstein one distance between \(\mu\) and the law of \(W\) in terms of the above data, with which they also arrive at a sampling algorithm. The key intermediate results are gradient and higher derivative bounds on the solution of the Poisson equation/Stein's equation with right hand side a Lipschitz continuous function. A similar estimate is obtained in [\textit{J. Gorham} et al., Ann. Appl. Probab. 29, No. 5, 2884--2928 (2019; Zbl 1439.60073)] using a Bismut-Elworthy-Li formula [\textit{K. D. Elworthy} and \textit{X. M. Li}, J. Funct. Anal. 125, No. 1, 252--286 (1994; Zbl 0813.60049)] invoking higher regularity condition on \(g\). The authors apply instead the more versatile formulation: the integration by parts formula (the two formulas are now known to be equivalent for e.g. compact spaces, see [\textit{K. D. Elworthy} and \textit{X.-M. Li}, in: Itô's stochastic calculus and probability theory. Tribute dedicated to Kiyosi Itô on the occasion of his 80th birthday. Tokyo: Springer. 15--30 (1996; Zbl 0881.60052)]). They then use an explicit formula for the solution of the Poisson equation to obtain the necessary gradient and Hessian estimates. This non-normal approximation theorem can be considered as an extension of a result of \textit{S. Chatterjee} and \textit{Q.-M. Shao} [Ann. Appl. Probab. 21, No. 2, 464--483 (2011; Zbl 1216.60018)].