Abstract
We propose a novel nonparametric two-sample test based on the Maximum MeanDiscrepancy (MMD), which is constructed by aggregating tests with differentkernel bandwidths. This aggregation procedure, called MMDAgg, ensures that testpower is maximised over the collection of kernels used, without requiringheld-out data for kernel selection (which results in a loss of test power), orarbitrary kernel choices such as the median heuristic. We work in thenon-asymptotic framework, and prove that our aggregated test is minimaxadaptive over Sobolev balls. Our guarantees are not restricted to a specifickernel, but hold for any product of one-dimensional translation invariantcharacteristic kernels which are absolutely and square integrable. Moreover,our results apply for popular numerical procedures to determine the testthreshold, namely permutations and the wild bootstrap. Through numericalexperiments on both synthetic and real-world datasets, we demonstrate thatMMDAgg outperforms alternative state-of-the-art approaches to MMD kerneladaptation for two-sample testing.