Empirical approximation to invariant measures for McKean-Vlasov processes: mean-field interaction vs self-interaction (Q6103256)

It is well known that the McKean-Vlasov process is a stochastic process described by a stochastic differential equation (SDE), whose coefficients depend on the distribution of the solution. Therefore it is not a Markov process in the usual sense. The equation is often referred as the McKean-Vlasov SDE or the mean-field SDE, deduced from the ``propagation of chaos'' of particle systems with mean-field interaction. In this paper, the authors consider the following problem for McKean-Vlasov processes: whether and how the invariant measure of a McKean-Vlasov process can be approximated by the empirical measures of the process itself or some other processes. It is proved that, under a monotonicity condition, the invariant probability measure of a McKean-Vlasov process can be approximated by weighted empirical measures of some processes including itself. The main contribution of the paper is that the authors construct a theoretical foundation for designing online algorithms to evaluate invariant measures for a class of McKean-Vlasov processes. The empirical approximation overcomes two shortages of the classical propagation of chaos. First, an unbiased approximation as time goes to infinity is provided, and even one trajectory is enough for computing the approximating measures. Second, the algorithm allows to make full use of the history information. To some extent, the authors compute infinite sample paths by admitting the equivalence between the temporal and spatial average. In addition, the advantages of the authors' methods, comparing to the ``propagation of chaos'', are presented, as well as a comparison with previous results in this area and with similar algorithms.