On-line tracking for blind source separation using zero-point probability

We have developed a novel on-line method for separating instantaneous, linear mixtures of super-Gaussian sources. The method uses a simple, constantly updating estimate of the central part of the probability distributions of the candidate mixed signals which can then be used to update the unmixing coefficients. The method is simple to implement and its concentration on the central part of the probability distribution makes it insensitive to outliers in the data. This is in contrast both with methods involving explicit estimates of higher order statistics, which are very sensitive to outliers, and implicit methods that raise signals to a high order power as part of their estimation process. In this paper we outline the details of the "zero-point probability" as a contrast for source separation and compare its resistance to outliers with standard fourth-order contrasts.