Copula versions of distance multivariance and dHSIC via the distributional transform -- a general approach to construct invariant dependence measures

TL;DR

This paper introduces copula-based versions of distance multivariance and dHSIC that are invariant, assumption-free, and applicable to mixed distributions, with empirical estimators and tests demonstrating improved power.

Contribution

It develops copula-based dependence measures using the distributional transform, ensuring invariance and no distributional assumptions, along with consistent empirical estimators and tests.

Findings

Dependence measures are invariant to scalings and translations.

Empirical estimators inherit the original tests' limiting distributions.

Tests can outperform existing copula dependence measures.

Abstract

The multivariate Hilbert-Schmidt-Independence-Criterion (dHSIC) and distance multivariance allow to measure and test independence of an arbitrary number of random vectors with arbitrary dimensions. Here we define versions which only depend on an underlying copula. The approach is based on the distributional transform, yielding dependence measures which always feature a natural invariance with respect to scalings and translations. Moreover, it requires no distributional assumptions, i.e., the distributions can be of pure type or any mixture of discrete and continuous distributions and (in our setting) no existence of moments is required. Empirical estimators and tests, which are consistent against all alternatives, are provided based on a Monte Carlo distributional transform. In particular, it is shown that the new estimators inherit the exact limiting distributional properties of the original estimators. Examples illustrate that tests based on the new measures can be more powerful than tests based on other copula dependence measures.