We characterize the asymptotic performance of nonparametric one- and two-sample testing. The exponential decay rate or error exponent of the type-II error probability is used as the asymptotic performance metric, and an optimal test achieves the maximum rate subject to a constant level constraint on the type-I error probability. With Sanov's theorem, we derive a sufficient condition for one-sample tests to achieve the optimal error exponent in the universal setting, i.e., for any distribution defining the alternative hypothesis. We then show that two classes of Maximum Mean Discrepancy (MMD) based tests attain the optimal type-II error exponent on \mathbb Rd, while the quadratic-time Kernel Stein Discrepancy (KSD) based tests achieve this optimality with an asymptotic level constraint. For general two-sample testing, however, Sanov's theorem is insufficient to obtain a similar sufficient condition. We proceed to establish an extended version of Sanov's theorem and derive an exact error exponent for the quadratic-time MMD based two-sample tests. The obtained error exponent is further shown to be optimal among all two-sample tests satisfying a given level constraint. Our work hence provides an achievability result for optimal nonparametric one- and two-sample testing in the universal setting. Application to off-line change detection and related issues are also discussed.

We characterize the asymptotic performance of nonparametric goodness of fit testing. The exponential decay rate of the type-II error probability is used as the asymptotic performance metric, and a test is optimal if it achieves the maximum rate subject to a constant level constraint on the type-I error probability. We show that two classes of Maximum Mean Discrepancy (MMD) based tests attain this optimality on $\mathbb R^d$, while the quadratic-time Kernel Stein Discrepancy (KSD) based tests achieve the maximum exponential decay rate under a relaxed level constraint. Under the same performance metric, we proceed to show that the quadratic-time MMD based two-sample tests are also optimal for general two-sample problems, provided that kernels are bounded continuous and characteristic. Key to our approach are Sanov’s theorem from large deviation theory and the weak metrizable properties of the MMD and KSD.

Universal hypothesis testing (UHT) refers to the problem of deciding whether samples come from a nominal distribution or an unknown distribution that is different from the nominal distribution. Hoeffding's test, whose test statistic is equivalent to the empirical Kullback-Leibler divergence (KL divergence), is known to be asymptotically optimal for distributions defined on finite alphabets. With continuous observations, however, the discontinuity of the KL divergence in the distribution functions results in significant complications for UHT. This paper introduces a robust version of the classical KL divergence, defined as the KL divergence from a distribution to the Lévy ball of a known distribution. This robust KL divergence is shown to be continuous in the underlying distribution function with respect to the weak convergence. The continuity property enables the development of an asymptotically optimal test for the university hypothesis testing problem with continuous observations. The optimality is in the same sense as that of the Hoeffding's test and stronger than that of Zeitouni and Gutman. Perhaps more importantly, the developed test statistic can be computed through convex programs, making it much more meaningful in practice. Numerical experiments are also conducted to evaluate its performance as compared with some kernel based goodness of fit test that has been proposed recently.

This letter examines a variation of the canonical distributed detection system: a sensor may overhear other sensors' transmissions and thus may choose to refine its output in the hope of achieving a better detection performance. We show that while this is indeed possible for the fixed sample size test, asymptotically (in the number of samples) there is no performance gain, as measured by the Kullback-Leibler distance achievable at the fusion center, provided that the observations are conditionally independent. For conditionally dependent observations, however, we demonstrate that asymptotic detection performance may indeed be improved when overhearing is utilized.

This paper considers the computation of Wyner's common information between outputs of additive Gaussian channels with a common input. The work is motivated by recent generalization of Wyner's common information to continuous random variables and the associated lossy source coding interpretation, as well as its application to statistical inference. It is shown that with independent and identically distributed Gaussian noises, Wyner's common information between channel outputs is precisely the same as the mutual information between the source input and the channel outputs regardless of the source distribution. The result extends the previous result when the source distribution is Gaussian. Generalization to additive channels with correlated noises and its application to statistical estimation are also presented.