Detecting non-uniform patterns on high-dimensional hyperspheres

Abstract: We propose a new probabilistic characterization of the uniform distribution on hyperspheres in terms of its inner product, extending the ideas of \cite{cuesta2009projection,cuesta2007sharp} in a data-driven manner. Using this characterization, we define a new distance that quantifies the deviation of an arbitrary distribution from uniformity. As an application, we construct a novel nonparametric test for the uniformity testing problem: determining whether a set of (n) i.i.d. random points on the (p)-dimensional hypersphere is approximately uniformly distributed. The proposed test is based on a degenerate U-process and is universally consistent in fixed-dimensional settings. Furthermore, in high-dimensional settings, it stands apart from existing tests with its simple implementation and asymptotic theory, while also possessing a model-free consistency property. Specifically, it can detect any alternative outside a ball of radius (n^{-1/2}) with respect to the proposed distance.