Approximation gaussienne d'algorithmes stochastiques à dynamique markovienne. (Gaussian approximation of stochastic algorithms) (Q1102035)

This paper addresses a problem involving a class of stochastic approximation algorithms. The author studies in fact the asymptotic behaviour of the ``normalized'' difference between a process built from the successive values produced by the algorithm and the solution of an ordinary differential equation ``naturally'' associated with it. Two theorems are given which, under appropriate but lengthy assumptions, insure that the normalized differences converge weakly towards Gaussian diffusions. The difference between the two theorems pertains to different behaviours assumed for the ``steps'' in the algorithms. The assumptions bear on the transition kernels of the Markov processes which ``drive'' the algorithms, rather than on moment and mixing properties of the random driving mechanism.