Concentration inequalities for mean field particle models (Q549866): Difference between revisions

A nonlinear equation \(\eta_{n+1}=\Phi_{n+1}(\eta_{n})\) is considered, where \((\eta_n)_{n \geq 0}\) is a sequence of probability measures on a sequence of measurable spaces equipped with some \(\sigma\)-fields. The mean field-type interacting particle system associated with the equation is a sequence of empirical measures \((\eta_n^N)_{n \geq 0}\) associated with \(N\) independent variables. These measures can be thought of as obtained by a sampling algorithm and, roughly speaking, it is expected that \(\eta_n^N \to \eta_n\) as \(N \to \infty\). The article is concerned with the fluctuations and the concentration properties of \((\eta_n^N)_{n \geq 0}\). The fluctuations are described by a functional central limit theorem which states the convergence in law for the random fields \(V_n^N:=N^{1/2}[\eta_n^N-\eta_n]\). The concentration properties quantify exponentially small probabilities of deviation events between \(\eta_n^N\) and \(\eta_n\). These concentration inequalities generalize the classical Hoeffding, Bernstein and Bennet inequalities for independent random sequences of interacting particles. Many illustrative examples are given. They include McKean-Vlasov type-diffusion models, McKean collision-type models of gases and a class of Feynman-Kac distribution flows arising in stochastic engineering sciences and in molecular chemistry.