------------------------------------------------
(14:00-15:00 UTC) Speaker: Yvik Swan (Université Libre de Bruxelles)
Title: Stein’s density approach
Abstract: Stein’s method is a collection of tools allowing to obtain explicit (non-asymptotic) bounds on discrepancies between probability distributions. The heart of the method is a so-called “Stein operator”, and it is the objective of this talk to present the basic theory behind these operators, and also cover some applications of the theory. After presenting the general theory (with a focus on Stein’s so-called density approach), we will give details on the 1-dimensional case, hereby connecting with classical topics such as Covariance and Poincaré inequalities.
The material covered in the talk is closely related to work with Marie Ernst, Christophe Ley, Guillaume Mijoule and Gesine Reinert.
(15:00-16:00 UTC) Speaker: Gesine Reinert (University of Oxford)
Title: Stein’s method for multivariate continuous distributions and some statistical applications
Abstract: Stein’s method can be viewed as employing an operator which action characterises a distribution uniquely, relating this operator to test functions through a Stein equation, and comparing distributions by taking expectations of this Stein equation. Using the density approach for Stein’s method, this talk will present a general framework for Stein’s method for multivariate continuous distributions.
For any given nontrivial distribution this approach leads to an infinite collection of Stein characterisations which can be used to assess distributional distances. We shall apply this framework to compare posterior distributions which arise from the same model but using different priors.
Using the notion of a weak Stein equation, bounds on distributional distances based on Lipschitz test functions are obtained if the distribution admits a Poincare’ constant. This result will be used to compare copulas.
Finally, using the notion of a Stein kernel, we shall extend Stein’s shrinkage results and Stein’s Unbiased Risk Estimate (SURE) to non-Gaussian distributions.
This talk is based on joint work with Max Fathi, Larry Goldstein, Adrien Saumard, and of course Yvik Swan.
------------------------------------------------