Stability of Stochastic Approximation under Verifiable Conditions

In this paper we address the problem of the stability and convergence of the stochastic approximation procedure θn+1=θn+γn+1[h(θn)+ξn+1]. The stability of such sequences {θn} is known to heavily rely on the behaviour of the mean field h at the boundary of the parameter set and the magnitude of the stepsizes used. The conditions typically required to ensure convergence, and in particular the boundedness or stability of {θn}, are either too difficult to check in practice or not satisfied at all. The most popular technique to circumvent the stability problem consists of constraining {θn} to a compact subset K in the parameter space. This is obviously not a satis-factory solution as the choice of K is a delicate one. In the present contribution we first prove a "deterministic" stability result which relies on simple conditions on the seauences {ξn} and {γn}. We then propose and analyze an algorithm based on projections on adaptive truncation sets which ensures that the aforementioned conditions required for stability are satisfied. We focus in particular on the case where {ξn} is a so-called Markov state-dependent noise.