Local Convergence Analysis of FastICA

Abstract

The FastICA algorithm can be considered as a selfmap on a manifold. It turns out that FastICA is a scalar shifted version of an algorithm recently proposed. We put these algorithms into a dynamical system framework. The local convergence properties are investigated subject to an ideal ICA model. The analysis is very similar to the well-known case in numerical linear algebra when studying power iterations versus Rayleigh quotient iteration. Show more