Random Function Iterations for Stochastic Fixed Point Problems

TL;DR

This paper investigates the convergence behavior of random function iterations in stochastic fixed point problems, establishing conditions for invariant measures and convergence types, with applications to distributed algorithms and computational methods.

Contribution

It introduces a general framework for analyzing convergence of random iterations to invariant measures, extending previous work to inconsistent cases and weaker assumptions.

Findings

Convergence to invariant measures under weaker conditions

Linear/geometric convergence guarantees

Applicability to distributed computation and algorithms

Abstract

We study the convergence of random function iterations for finding an invariant measure of the corresponding Markov operator. We call the problem of finding such an invariant measure the stochastic fixed point problem. This generalizes earlier work studying the stochastic feasibility problem}, namely, to find points that are, with probability 1, fixed points of the random functions [Hermer, Luke, Sturm, 2019]. When no such points exist, the stochastic feasibility problem is called inconsistent, but still under certain assumptions, the more general stochastic fixed point problem has a solution and the random function iterations converge to an invariant measure for the corresponding Markov operator. There are two major types of convergence: almost sure convergence of the iterates to a fixed point in the case of stochastic feasibility, and convergence in distribution more generally. We show how common structures in deterministic fixed point theory can be exploited to establish existence of invariant measures and convergence of the Markov chain. We show that weaker assumptions than are usually encountered in the analysis of Markov chains guarantee linear/geometric convergence. This framework specializes to many applications of current interest including, for instance, stochastic algorithms for large-scale distributed computation, and deterministic iterative procedures with computational error. The theory developed in this study provides a solid basis for describing the convergence of simple computational methods without the assumption of infinite precision arithmetic or vanishing computational errors.