Phase Locked Matrix Factorization

We present a novel approach to separate linearly mixed dependent sources that are phase-locked. The separation is done through a minimization problem involving three variables (the mixing matrix, the source time-dependent amplitudes, and their relative phases). Results obtained in toy data sets show that this algorithm is very fast, that it estimates the mixing matrix with remarkable precision even with considerable amounts of noise, and that the sources are also correctly estimated. We interpret these results as a “proof-of-concept” that this approach is valid and discuss the necessary improvements to deal with more general situations.