INDEPENDENT COMPONENT ANALYSIS IN THE PRESENCE OF GAUSSIAN NOISE BASED ON ESTIMATING FUNCTIONS

Motoaki Kawanabe

The problem of estimating the statistical model of in­ dependent component analysis in the presence of Gaus­ sian noise is considered. Because of the additive noise, a combination of factor analysis and a noise­free ICA algorithm doesn't give a consistent estimator of the mixing matrix. In this paper, following the semipara­ metric statistical approach to the noise­free ICA model by Amari and Cardoso [1], we propose a method of estimating the mixing matrix consistently even if the additive noise exists. The proposed algorithm consists of two stages: First find the factor subspace by means of factor analysis, and then determine the directions of independent components based on an estimating func­ tion in this semiparametric model.