Title:W-Kernel and Its Principal Space for Frequentist Evaluation of Bayesian Estimators

Abstract:Evaluating the variability of posterior estimates is a key aspect of Bayesian model assessment. In this study, we focus on the posterior covariance matrix W, defined using the log-likelihood of each observation. Previous studies have examined the role of the principal space of W in Bayesian sensitivity analysis, notably MacEachern and Peruggia (2002) and Thomas et al. (2018). In this work, we show that the principal space of W is also relevant for frequentist evaluation, using the recently proposed Bayesian infinitesimal jackknife (IJ) approximation Giordano and Broderick (2023) as a key tool. We next consider the relationship between the matrix W and the Fisher kernel. We show that the Fisher kernel can be regarded as an approximation to W; the matrix W, in itself, can be interpreted as a reproducing kernel, which we refer to as the W-kernel. Based on this connection, we examine the dual relationship between the W-kernel formulation in the data space and the classical asymptotic formulation in the parameter space. These ideas suggest a form of Bayesian-frequentist duality that emerges through the dual structure of kernel PCA, where posterior and frequentist covariances serve as inner products in their respective spaces. As an application, we consider an approximate bootstrap of posterior means based on posterior samples generated by MCMC. We show that the projection onto the principal space of W facilitates frequentist evaluation, particularly of the higher-order term in this procedure. In one of the appendices, we introduce incomplete Cholesky decomposition as an efficient method for computing the principal space of W and discuss the related concept of representative subsets of the observations.