The distribution of entropy estimators based on maximum mean log-likelihood

R. Moddemeijer

University of Groningen, Department of Computing Science,
P.O. Box 800, NL-9700 AV Groningen, The Netherlands,
phone: +31.50.363 3940 - fax: +31.50.363 38005 - e-mail: rudy@cs.rug.nl

Abstract

Entropy estimation is often based on the Maximum Likelihood (ML) method. When the probability density function sufficiently models the reality, the maximum average log-likelihood is a good (negative-)entropy estimator.

Previous work from the author suggests that, under certain conditions, the variance of such a statistic consists of a basic variance, plus a number of statistically independent contributions corresponding with the independently adjustable parameters of the pdf. Conform this assumption we derive and justify under certain conditions the distribution of a ML-based entropy estimator.

This knowledge about the distribution can be used to bridge the gap between process and application of optimal models, in particular for the selection of an optimal probability density function in the presence of superfluous parameters.


Full paper


Poster


Published

Twenty-first Symposium on Information Theory in the Benelux, May 25-26, 2000, Wassenaar (NL), pp 231-238, eds. J. Biemond; Werkgemeenschap Informatie- en Communicatietheorie, Enschede, The Netherlands, and IEEE Benelux Chapter on Information Theory, ISBN 90-71048 BibTeX


other publications