Generalized Negentropy for Statistical Dependence

Abstract

In this paper, we propose a generalized form of negentropy using the density power divergence, which can compromise between robustness and effectiveness. The conventional form of negentropy can be considered as its special case. From the fact that statistical dependence measures such as the mutual information can be defined in terms of joint and marginal negentropies, we define a new statistical dependence measure using the generalized negentropy, and analyze its behavior. Also, an algorithm for independent component analysis is derived from this measure. In the experiments, we evaluated the performance of the derived algorithm on a variety of source distributions and compared it with well-known algorithms. The results show that the proposed measure consistently outperforms others and we can compromise between robustness and effectiveness.