On Stein's method and perturbations

TL;DR

This paper extends Stein's perturbation method to a broad range of distributions, enabling improved approximation assessments across various metrics like Kolmogorov, Wasserstein, and total variation.

Contribution

It generalizes the perturbation approach of Stein's method to include normal, Poisson, compound Poisson, binomial, and Poisson process distributions.

Findings

Method applies to diverse distribution perturbations.

Effective in multiple metrics including Wasserstein and total variation.

Examples demonstrate practical applicability.

Abstract

Stein's (1972) method is a very general tool for assessing the quality of approximation of the distribution of a random element by another, often simpler, distribution. In applications of Stein's method, one needs to establish a Stein identity for the approximating distribution, solve the Stein equation and estimate the behaviour of the solutions in terms of the metrics under study. For some Stein equations, solutions with good properties are known; for others, this is not the case. Barbour and Xia (1999) introduced a perturbation method for Poisson approximation, in which Stein identities for a large class of compound Poisson and translated Poisson distributions are viewed as perturbations of a Poisson distribution. In this paper, it is shown that the method can be extended to very general settings, including perturbations of normal, Poisson, compound Poisson, binomial and Poisson process approximations in terms of various metrics such as the Kolmogorov, Wasserstein and total variation metrics. Examples are provided to illustrate how the general perturbation method can be applied.