Estimating Differential Entropy under Gaussian Convolutions

This paper studies the problem of estimating the differential entropy h(S+Z), where S and Z are independent d-dimensional random variables with Z∼𝒩(0,σ^2 I_d). The distribution of S is unknown, but n independently and identically distributed (i.i.d) samples from it are available. The question is whether having access to samples of S as opposed to samples of S+Z can improve estimation performance. We show that the answer is positive. More concretely, we first show that despite the regularizing effect of noise, the number of required samples still needs to scale exponentially in d. This result is proven via a random-coding argument that reduces the question to estimating the Shannon entropy on a 2^O(d)-sized alphabet. Next, for a fixed d and n large enough, it is shown that a simple plugin estimator, given by the differential entropy of the empirical distribution from S convolved with the Gaussian density, achieves the loss of O((log n)^d/4/√(n)). Note that the plugin estimator amounts here to the differential entropy of a d-dimensional Gaussian mixture, for which we propose an efficient Monte Carlo computation algorithm. At the same time, estimating h(S+Z) via popular differential entropy estimators (based on kernel density estimation (KDE) or k nearest neighbors (kNN) techniques) applied to samples from S+Z would only attain much slower rates of order O(n^-1/d), despite the smoothness of P_S+Z. As an application, which was in fact our original motivation for the problem, we estimate information flows in deep neural networks and discuss Tishby's Information Bottleneck and the compression conjecture, among others.