A Computable Measure of Suboptimality for Entropy-Regularised Variational Objectives
This research introduces Kernel Gradient Discrepancy (KGD), a new measure to quantify how well an approximate distribution matches an unknown target distribution in variational inference. KGD is crucial because it allows researchers to judge the quality of approximations without needing the true target density. It provides a mathematical tool to guide and optimize sampling algorithms for complex models.