On the stability of the stochastic gradient Langevin algorithm with dependent data stream

Abstract

We prove, under mild conditions, that the stochastic gradient Langevin dynamics converges to a limiting law as time tends to infinity, even in the case where the driving data sequence is dependent.