Distribution-Free Tests of Independence with Applications to Testing More Structures

Fang Han, Han Liu

Published 2014-10-15Version 1

We consider the problem of testing mutual independence of all entries in a d-dimensional random vector X=(X_1,...,X_d)^T based on n independent observations. For this, we consider two families of distribution-free test statistics that converge weakly to an extreme value type I distribution. We further study the powers of the corresponding tests against certain alternatives. In particular, we show that the powers tend to one when the maximum magnitude of the pairwise Pearson's correlation coefficients is larger than C(log d/n)^{1/2} for some absolute constant C. This result is rate optimal. As important examples, we show that the tests based on Kendall's tau and Spearman's rho are rate optimal tests of independence. For further generalization, we consider accelerating the rate of convergence via approximating the exact distributions of the test statistics. We also study the tests of two more structural hypotheses: m-dependence and data homogeneity. For these, we propose two rank-based tests and show their optimality.