Title:Functional Central Limit Theorem for Two Timescale Stochastic Approximation

Abstract:Two time scale stochastic approximation algorithms emulate singularly perturbed deterministic differential equations in a certain limiting sense, i.e., the interpolated iterates on each time scale approach certain differential equations in the large time limit when viewed on the `algorithmic time scale' defined by the corresponding step sizes viewed as time steps. Their fluctuations around these deterministic limits, after suitable scaling, can be shown to converge to a Gauss-Markov process in law for each time scale. This turns out to be a linear diffusion for the faster iterates and an ordinary differential equation for the slower iterates.