A Scalable Shannon Entropy Estimator

We revisit the well-studied problem of estimating the Shannon entropy of a probability distribution, now given access to a probability-revealing conditional sampling oracle. In this model, the oracle takes as input the representation of a set $S$ and returns a sample from the distribution obtained by conditioning on $S$, together with the probability of that sample in the distribution. Our work is motivated by applications of such algorithms in Quantitative Information Flow analysis (QIF) in programming-language-based security. Here, information-theoretic quantities capture the effort required on the part of an adversary to obtain access to confidential information. These applications demand accurate measurements when the entropy is small. Existing algorithms that do not use conditional samples require a number of queries that scale inversely with the entropy, which is unacceptable in this regime, and indeed, a lower bound by Batu et al.(STOC 2002) established that no algorithm using only sampling and evaluation oracles can obtain acceptable performance. On the other hand, prior work in the conditional sampling model by Chakraborty et al.(SICOMP 2016) only obtained a high-order polynomial query complexity, $\mathcal{O}(\frac{m^7}{ε^8}\log\frac{1}δ)$ queries, to obtain additive $ε$-approximations on a domain of size $\mathcal{O}(2^m)$. We obtain multiplicative $(1+ε)$-approximations using only $\mathcal{O}(\frac{m}{ε^2}\log\frac{1}δ)$ queries to the probability-revealing conditional sampling oracle. Indeed, moreover, we obtain small, explicit constants, and demonstrate that our algorithm obtains a substantial improvement in practice over the previous state-of-the-art methods used for entropy estimation in QIF.