Towards a better understanding of the dual representation of phi divergences

TL;DR

This paper investigates dual representations of phi-divergences for estimation, revealing non-robustness of traditional estimators and proposing a new robust class with theoretical and simulation validation.

Contribution

It introduces a new class of robust estimators based on dual phi-divergence representation, improving robustness over traditional methods.

Findings

Minimum phi-divergence estimators are generally non-robust.

The new dual phi-divergence estimators exhibit improved robustness.

Simulation results favor the new estimators over existing methods.

Abstract

The aim of this paper is to study different estimation procedures based on $\varphi-$divergences. The dual representation of $\varphi-$divergences based on the Fenchel-Legendre duality is the main interest of this study. It provides a way to estimate $\varphi-$divergences by a simple plug-in of the empirical distribution without any smoothing technique. Resulting estimators are thoroughly studied theoretically and with simulations showing that the so called minimum $\varphi-$divergence estimator (MD$\varphi$DE) is generally non robust and behaves similarly to the maximum likelihood estimator. We give some arguments supporting the non robustness property, and give insights on how to modify the classical approach. An alternative class of $\varphi-$divergences robust estimators based on the dual representation is presented. We study consistency and robustness properties from an influence function point of view of the new estimators. In a second part, we invoke the Basu-Lindsay approach for approximating $\varphi-$divergences and provide a comparison between these approaches. The so called dual $\varphi-$divergence is also discussed and compared to our new estimator. A full simulation study of all these approaches is given in order to compare efficiency and robustness of all mentioned estimators against the so-called minimum density power divergence, showing encouraging results in favor of our new class of minimum dual $\varphi-$divergences.