Quantitative approximations of evolving probability measures and sequential Markov chain Monte Carlo methods (Q1950384)

The main goal of the paper is to evaluate some non-asymptotic bounds for particle system approximations using an \(L^p\) approach in order to control the approximation error. The particle system dynamics is a combination of \(N\) independent Markov chain moves defined by the Markov processes \((X_t^N,P)\) with initial distribution \(\mu_0\). An approximation error -- of a function \(f:S\longrightarrow \mathbb{R}\) (\(S\) finite) with \(L^p\)-norm less than \(1\) -- is the difference \(<f,\nu_t^N>-<f,\mu_t>\), where \((\mu_t)_{t\geq 0}\) is a family of mutually absolutely continuous probability measures, and \(\nu_t^N\) is an empirical distribution of \((X_t^N,P)\). The main difference from other similar results in the domain is that this paper treats the case of \(N\) finite (non-asymptotic case) and time \(t\) continuous. The particle systems have the scope to discretize the evolution equations for probability measures (called Fokker-Planck equations). All original results obtained are summarized in Section 2 and establish non-asymptotic bounds for the variance of the estimator \(<f,\nu_t^N>\) in different situations (like global Poincaré and logarithmic Sobolev inequalities). The paper is technical, very hard to be followed by readers which do not know the domain. Two points can be emphasized here: 1. 50\% of the paper (3 sections and an appendix: 18 pages from 36) are dedicated to proofs. 2. Almost all results follow ideas and reasoning from \textit{P. Del Moral} and \textit{L. Miclo} [Séminaire de Probabilités XXXIV. Berlin: Springer. Lect. Notes Math. 1729, 1--145 (2000; Zbl 0963.60040)]. A similar approach of error bounds is developed for the discrete time by \textit{N. Schweizer} [Non-asymptotic error bounds for sequential MCMC methods. Ph.D. Thesis, Universität Bonn (2011)].