On the Decreasing Power of Kernel and Distance Based Nonparametric Hypothesis Tests in High Dimensions

This paper is about two related methods for two sample testing and independence testing which have emerged over the last decade: Maximum Mean Discrepancy (MMD) for the former problem and Distance Correlation (dCor) for the latter. Both these methods have been suggested for high-dimensional problems, and sometimes claimed to be unaffected by increasing dimensionality of the samples. We will show theoretically and practically that the power of both methods (for different reasons) does actually decrease polynomially with dimension. We also analyze the median heuristic, which is a method for choosing tuning parameters of translation invariant kernels. We show that different bandwidth choices could result in the MMD decaying polynomially or even exponentially in dimension.