Non-asymptotic confidence bounds for stochastic approximation algorithms with constant step size (Q753365)

The multidimensional Robbins-Monro procedure with independent errors and constant step-size is considered. A comparison result for the stationary distribution of Markovian processes is derived which allows to bound the level of confidence regions. A stopping time which is based on the inner products of subsequent gradient estimates is introduced. It is shown that this stoppig time leads to confidence regions with a coverage probability which is uniform in the starting value.