WorksMathematics
Anytime-valid goodness-of-fit test by -spacings
Abstract
To perform an anytime-valid goodness-of-fit test, we propose spacing-based -processes that control the type-I error rate under arbitrary stopping rules. Based on the maximum spacing of order statistics, we construct spacing-based -processes for nonparametrically testing distribution. The resulting procedures are simple and computationally efficient, and are power-consistent. We call them -spacing tests. This approach is especially powerful in detecting localized discrepancy between the data-generating density and the null density. Under mild conditions, the logarithm of the proposed -process is asymptotically equivalent to that of the oracle-optimal -process. We further generalize the construction from maximal spacing to maximum spacing set in the form of the fraction of the largest th order spacings, as a measure of the lower-level set of density. It boosts power under more global discrepancy from the null. Furthermore, we propose a greedy predictable scheme, via argmax or softmax rules, to combine multiple -spacing testing strategies. We prove that it optimally tunes the -spacing test. The proposed approach is readily generalizable to multivariate testing via a coordinate-wise Rosenblatt-transform. Simulation studies demonstrate that the proposed methods perform favorably relative to the considered anytime-valid competitors as well as various types of multiple test-adjusted Kolmogorov–Smirnov tests.