Solution of high-dimensional linear separation problems

Blind source separation (BSS) has been one of the emerging research topics within the signal processing community in recent years. Particularly, the maximum squared kurtosis has been found to be a suitable criterion for many technical applications of BSS. Conventionally an elementary Givens rotation estimator is applied to all source pairs in a Jacoby-like algorithm. However, those methods suffer from an escalation of computational expenses as soon as the number of sources becomes large. This paper introduces a novel eigenvector deflation method. It allows the separation of complex and high-dimensional mixtures without such performance penalty.