BLOCK-COORDINATE RELATIVE NEWTON METHOD FOR BLIND SOURCE SEPARATION
Alexander M. Bronstein, Michael M. Bronstein, Michael Zibulevsky · 2003
We generalize the relative Newton method recently proposed for quasi-maximum likelihood blind source separation. Special structure of the Hessian allows performing block-coordinate Newton descent, which significantly boosts the algorithm performance and reduces its computational complexity. Simulation results show that even for large problems, the number of iterations required by the block-coordinate method to converge remains approximately constant.