Optimal coding for discrete random vector

Abstract

Based on the notion of mutual information between the components of a discrete random vector, we construct, for data reduction reasons, an optimal quantization of the support of its probability measure. More precisely, we propose a simultaneous discretization of the whole set of the components of the discrete random vector which takes into account, as much as possible, the stochastic dependence between them. Computationals aspects and example are presented.