On the entropy minimization of a linear mixture of variables for source separation

  • Authors:
  • Frédéric Vrins;Michel Verleysen

  • Affiliations:
  • Université Catholique de Louvain - UCL, Electrical Engineering Department, Microelectronics Laboratory, Machine Learning Group, Louvain-la-Neuve, Belgium;Université Catholique de Louvain - UCL, Electrical Engineering Department, Microelectronics Laboratory, Machine Learning Group, Louvain-la-Neuve, Belgium and Belgian National Fund for the Sci ...

  • Venue:
  • Signal Processing - Special issue: Information theoretic signal processing
  • Year:
  • 2005

Quantified Score

Hi-index 0.00

Visualization

Abstract

The marginal entropy h(Z) of a weighted sum of two variables Z = αX + βY, expressed as a function of its weights, is a usual cost function for blind source separation (BSS), and more precisely for independent component analysis (ICA). Even if some theoretical investigations were done about the relevance from the BSS point of view of the global minimum of h(Z), very little is known about possible local spurious minima.In order to analyze the global shape of this entropy as a function of the weights, its analytical expression is derived in the ideal case of independent variables. Because of the ICA assumption that distributions are unknown, simulation results are used to show how and when local spurious minima may appear. Firstly, the entropy of a whitened mixture, as a function of the weights and under the constraint of independence between the source variables, is shown to have only relevant minima for ICA if at most one of the source distributions is multimodal. Secondly, it is shown that if independent multimodal sources are involved in the mixture, spurious local minima may appear. Arguments are given to explain the existence of spurious minima of h(Z) in the case of multimodal sources. The presented justification can also explain the location of these minima knowing the source distributions. Finally, it results from numerical examples that the maximum-entropy mixture is not necessarily reached for the 'most mixed' one (i.e. equal mixture weights), but depends of the entropy of the mixed variables.