Blind Separation of Noncircular Correlated Sources Using Gaussian Entropy Rate

We introduce a new blind source separation (BSS) algorithm for correlated noncircular sources that uses only second-order statistics and fully takes the correlation structure into account. We propose a parametric entropy rate estimator that uses a widely linear autoregressive (AR) model for the sources, and derive the BSS algorithm by minimizing the mutual information of separated time series. We compare the performance of the new algorithm with competing algorithms and demonstrate its superior separation performance as well as its effectiveness in separation of non-Gaussian sources when the identification conditions are met.