Distance Covariance in Metric Spaces: Non-Parametric Independence Testing in Metric Spaces (Master's thesis)

TL;DR

This thesis develops a non-parametric method using distance covariance in metric spaces to test independence between random elements, providing estimators and asymptotic properties for practical statistical testing.

Contribution

It introduces a novel approach to independence testing in metric spaces using distance covariance, including estimators and bootstrap methods for hypothesis testing.

Findings

Distance covariance indicates independence in metric spaces of strong negative type.

Constructed estimators are asymptotically consistent for independence testing.

Bootstrap methods effectively determine rejection thresholds for tests.

Abstract

The aim of this thesis is to find a solution to the non-parametric independence problem in separable metric spaces. Suppose we are given finite collection of samples from an i.i.d. sequence of paired random elements, where each marginal has values in some separable metric space. The non-parametric independence problem raises the question on how one can use these samples to reasonably draw inference on whether the marginal random elements are independent or not. We will try to answer this question by utilizing the so-called distance covariance functional in metric spaces developed by Russell Lyons. We show that, if the marginal spaces are so-called metric spaces of strong negative type (e.g. seperable Hilbert spaces), then the distance covariance functional becomes a direct indicator of independence. That is, one can directly determine whether the marginals are independent or not based solely on the value of this functional. As the functional formally takes the simultaneous distribution as argument, its value is not known in the posed non-parametric independence problem. Hence, we construct estimators of the distance covariance functional, and show that they exhibit asymptotic properties which can be used to construct asymptotically consistent statistical tests of independence. Finally, as the rejection thresholds of these statistical tests are non-traceable we argue that they can be reasonably bootstrapped.