Blind Source Separation by Nuclear Norm Minimization and Local Recoverability Analysis
Takashi Tanaka, Cédric Langbort, Lalit K. Mestha, Alvaro E. Gil · IEEE Signal Processing Letters · 2013
We propose a new blind source separation (BSS) algorithm that is effective when Hankel matrices constructed from individual source signals are near low-rank and satisfy a certain near-orthogonality condition. Source separation is achieved by finding a nonsingular reverse-mixing operation that minimizes nuclear norms of Hankel matrices constructed from estimated source signals. The new formulation results in a non-convex optimization problem involving a reverse-mixing matrix. Preliminary analysis of local recoverability of source signals as well as few numerical simulations are presented in this letter.