Regularized Estimation of Information via High Dimensional Canonical Correlation Analysis

TL;DR

This paper introduces a regularized method for estimating information between high-dimensional data sources using canonical correlation analysis, enhancing interpretability and scalability with theoretical links to spectral analysis.

Contribution

It proposes a novel regularized estimation approach for information measures via high-dimensional canonical correlation analysis, connecting spectral analysis and simplifying computations.

Findings

Effective estimation of information using regularized CCA.

Scalability improvements for large datasets.

Strong dualities with spectral analysis demonstrated.

Abstract

In recent years, there has been an upswing of interest in estimating information from data emerging in a lot of areas beyond communications. This paper aims at estimating the information between two random phenomena by using consolidated second-order statistics tools. The squared-loss mutual information is chosen for that purpose as a natural surrogate of Shannon mutual information. The rationale for doing so is developed for i.i.d. discrete sources -mapping data onto the simplex space-, and for analog sources -mapping data onto the characteristic space-, highlighting the links with other well-known related concepts in the literature based on local approximations of information-theoretic measures. The proposed approach gains in interpretability and scalability for its use on large datasets, providing physical interpretation to the free regularization parameters. Moreover, the structure of the proposed mapping allows resorting to Szeg\"o's theorem to reduce the complexity for high dimensional mappings, exhibiting strong dualities with spectral analysis. The performance of the proposed estimators is analyzed using Gaussian mixtures.