Maximum likelihood estimators uniformly minimize distribution variance among distribution unbiased estimators in exponential families

TL;DR

This paper demonstrates that maximum likelihood estimators in exponential families uniquely minimize distribution variance among unbiased estimators when viewed as random distributions, extending classical concepts with a distributional perspective.

Contribution

It introduces a distributional framework for estimation, defining distribution expectation and variance, and proves the MLE's optimality and robustness as the minimum distribution variance unbiased estimator.

Findings

MLE is distribution unbiased in exponential families.

MLE uniquely minimizes distribution variance among unbiased estimators.

MLE remains robust for model misspecification via KL projection.

Abstract

We employ a parameter-free distribution estimation framework where estimators are random distributions and utilize the Kullback-Leibler (KL) divergence as a loss function. Wu and Vos [J. Statist. Plann. Inference 142 (2012) 1525-1536] show that when an estimator obtained from an i.i.d. sample is viewed as a random distribution, the KL risk of the estimator decomposes in a fashion parallel to the mean squared error decomposition when the estimator is a real-valued random variable. In this paper, we explore how conditional versions of distribution expectation ($E^{\dagger}$) can be defined so that a distribution version of the Rao-Blackwell theorem holds. We define distributional expectation and variance ($V^{\dagger}$) that also provide a decomposition of KL risk in exponential and mixture families. For exponential families, we show that the maximum likelihood estimator (viewed as a random distribution) is distribution unbiased and is the unique uniformly minimum distribution variance unbiased (UMV$^{\dagger}$U) estimator. Furthermore, we show that the MLE is robust against model specification in that if the true distribution does not belong to the exponential family, the MLE is UMV$^{\dagger}$U for the KL projection of the true distribution onto the exponential families provided these two distribution have the same expectation for the canonical statistic. To allow for estimators taking values outside of the exponential family, we include results for KL projection and define an extended projection to accommodate the non-existence of the MLE for families having discrete sample space. Illustrative examples are provided.