Blind separation from ε-contaminated mixtures

This paper deals with the problem of Blind Source Separation (BSS). BSS algorithms typically require that observed data are prewhitened. The data are here assumed to be contaminated by highly deviating samples. Hence, covariance matrix used for whitening and determining the number of signals is estimated unreliably. We propose a method where data are first whitened in a robust manner. Sources are then separated using an iterative least squares algorithm. The proposed method is compared to a method based on sample estimates and the influence of outliers is analysed.