Nonasymptotic analysis of adaptive and annealed Feynman-Kac particle models (Q502900): Difference between revisions

Feynman-Kac (= FK) particle models are Monte Carlo methods designed for sampling from a sequence of high dimensional probability distributions \(\{\eta_n\}\), defined by the FK recursive formula. The idea of the FK particle algorithm is the approximation of the measure \(\eta_n\) via simulation of an interacting particle system \((\xi_n)_n=(\xi_n^1,\ldots,\xi_n^N)_n\) of size \(N,\) so that \[ \eta_n^N=N^{-1}\sum_{1\leq i\leq N}\delta_{\xi_n^i}\to \eta_n,\quad N\to \infty. \] Starting from \(N\) independent samples \(\xi_0=(\xi_0^1,\ldots,\xi_0^N)\) from \(\eta_0\), every particle firstly evolves according to the acceptance-rejection scheme with recycle mechanism. At the second step the selected particles evolve according to given Markov transitions. The paper studies statistical properties of the error of approximation \(\eta_n\) by the empirical measure \(\eta_n^N\). One of the main results in this direction is derivation of the upper bound for the \(L_p\) mean value error \[ \mathbb{E}(|\eta_n^N(f)-\eta_N(f)|^p)^{1/p}, \] on the space of bounded functions \(f\). The bound is expressed via the Dobrushin ergodic coefficients of the aforementioned Markov transitions.