ALPHABET-BASED DEFLATION FOR BLIND SOURCE EXTRACTION IN UNDERDETERMINED MIXTURES

The deflation approach to blind source extraction estimates the source signals one by one. The contribution of the latest source estimate is computed via linear regression and subtracted from the observations before performing a new extraction. In the context of digital communications, novel alphabet-based contrast criteria can naturally be defined, leading to the recently proposed parallel deflation concept. We analyse the use of such criteria in the challenging scenario of underdetermined mixtures, where the sources outnumber the sensors. Due to the limitations of linear extraction, projection on the signal alphabet before the regression-subtraction stage is shown to be capital for a successful source estimation. It is also demonstrated that alphabet-based criteria outperform the constant modulus (CM) principle, even for CM-type sources. More interestingly, classical deflation can improve on parallel deflation, but requires a refinement to render its performance robust to the extraction ordering.