SEM217: Haosui (Kevin) Duanmu, UC Berkeley

Tuesday, March 30th @ 11:00-12:30 PM (ONLINE)

On extended admissible decision procedures and their nonstandard Bayes risk

Haosui (Kevin) Duanmu, UC Berkeley

ABSTRACT: Nonstandard analysis, a powerful machinery derived from mathematical logic, has had many applications in probability theory as well as stochastic processes. Nonstandard analysis allows construction of a single object—a hyperfinite probability space—which satisfies all the first order logical properties of a finite probability space, but which can be simultaneously viewed as a measure-theoretical probability space via the Loeb construction. As a consequence, the hyperfinite/measure duality has proven to be particularly in porting discrete results into their continuous settings.

The connection between frequentist and Bayesian optimality in statistical decision theory is a longstanding open problem. For statistical decision problems with a finite parameter space, it is well known that a decision procedure is extended admissible (frequentist optimal) if and only if it is Bayes. Such connection becomes fragile for decision problems with an infinite parameter space and one must relax the notion of Bayes optimality to regain such equivalence between extended admissibility and Bayes optimality. Various attempts have been made in the literature but they are subject to technical conditions which often rule our semi-parametric and nonparametric problems. By using nonstandard analysis, we develop a novel notion of nonstandard Bayes optimality (Bayes with infinitesimal excess risk). We show that, without any technical condition, a decision procedure is extended admissible if and only if it is nonstandard Bayes. We conclude by showing that several existing standard results in the literature can be derived from our main result.