Nonlinearity and Separation Capability: Further Justification for the ICA Algorithm with A Learned Mixture of Parametric Densities

摘要

We discuss the relation between nonlinearity and separation capability in the information-theoretic ICA scheme. We propose with justification that a 'loose matching' between the nonlinearity and source distribution is needed. These results give further support to the implementation technique by a learned mixture of parametric densities.