The existence of spurious equilibrium in FastICA

Abstract

FastICA is a fast fixed-point algorithm proposed by Hyvärinen for Independent Component Analysis (ICA). For the outstanding performance such as fast convergence and robustness, it is now one of the most popular estimate methods. The existence of spurious equilibrium in FastICA is addressed in this paper, which comes from its applications in Blind Source Separation (BSS). Two issues are involved. The first is on the object function and the second is about the fixedpoint algorithm, an approximate Newton's method. Analysis shows the existence of spurious equilibrium, which is the singular point introduced during the algorithm's derivation. Experimental results show the estimates of spurious equilibria, and improvements are proposed by revising the convergence condition. © Springer-Verlag 2004.