A nonparametric distribution-free test of independence among continuous random vectors based on \texorpdfstring{$L_1$}{}-norm

TL;DR

This paper introduces a new nonparametric, distribution-free independence test for multidimensional vectors using the $L_1$-distance, which is robust, asymptotically normal, and effective against local alternatives without regularity assumptions.

Contribution

The authors develop a novel $L_1$-based independence test that is distribution-free, asymptotically normal, and robust to the form of the underlying density, with demonstrated superior finite-sample performance.

Findings

Test statistic's distribution is unaffected by the density form.

The test has nontrivial local power against specific alternatives.

Simulation shows the method outperforms existing tests.

Abstract

We propose a novel statistical test to assess the mutual independence of multidimensional random vectors. Our approach is based on the $L_1$-distance between the joint density function and the product of the marginal densities associated with the presumed independent vectors. Under the null hypothesis, we employ Poissonization techniques to establish the asymptotic normal approximation of the corresponding test statistic, without imposing any regularity assumptions on the underlying Lebesgue density function, denoted as $f(\cdot)$. Remarkably, we observe that the limiting distribution of the $L_1$-based statistics remains unaffected by the specific form of $f(\cdot)$. This unexpected outcome contributes to the robustness and versatility of our method. Moreover, our tests exhibit nontrivial local power against a subset of local alternatives, which converge to the null hypothesis at a rate of {${\tiny n^{\tiny -1/2}h_n^{\tiny -{d/4}}}$}, $d\geq 2$, where $n$ represents the sample size and $h_n$ denotes the bandwidth. Finally, the theory is supported by a comprehensive simulation study to investigate the finite-sample performance of our proposed test. The results demonstrate that our testing procedure generally outperforms existing approaches across various examined scenarios.