Stein's density method for multivariate continuous distributions (Q6165208): Difference between revisions

The density approach has been very successfully applied as a framework for Stein's method for probability approximation in the univariate setting. In the present paper the authors generalise this to the multivariate setting, deriving appropriate operators and equations that can be used in the approximation by a multivariate target distribution for which a density exists. Notions of a Stein kernel and Stein discrepancy which have proved very useful in applications of Stein's method are introduced in conjunction with the operators defined herein. The authors also introduce the notion of a \emph{weak} Stein equation to circumvent problems associated with existence of a solution to the usual Stein equation, and use this to establish bounds in Wasserstein distance under the assumption of the existence of a suitable Poincaré constant. The work is illustrated throughout by a number of examples. These include the comparison of normal distributions, normal versus Student's \(t\), the Azzalini--Dalla Valle skew-normal distribution versus the normal, the influence of the prior on the posterior distribution in a Bayesian setting, and comparison of copulas on the unit square.