Huber-robust likelihood ratio tests for composite nulls and alternatives

Abstract: We propose an e-value based framework for testing composite nulls against composite alternatives when an $\epsilon$ fraction of the data can be arbitrarily corrupted. Our tests are inherently sequential, being valid at arbitrary data-dependent stopping times, but they are new even for fixed sample sizes, giving type-I error control without any regularity conditions. We achieve this by modifying and extending a proposal by Huber (1965) in the point null versus point alternative case. Our test statistic is a nonnegative supermartingale under the null, even with a sequentially adaptive contamination model where the conditional distribution of each observation given the past data lies within an $\epsilon$ (total variation) ball of the null. The test is powerful within an $\epsilon$ ball of the alternative. As a consequence, one obtains anytime-valid p-values that enable continuous monitoring of the data, and adaptive stopping. We analyze the growth rate of our test supermartingale and demonstrate that as $\epsilon\to 0$, it approaches a certain Kullback-Leibler divergence between the null and alternative, which is the optimal non-robust growth rate. A key step is the derivation of a robust Reverse Information Projection (RIPr). Simulations validate the theory and demonstrate excellent practical performance.