Blind source separation based on a novel relative gradient

In this paper, a novel relative gradient is proposed to solve the blind source separation problem. An iterative method is introduced to solve the nonlinear matrix equation which is derived from the relative gradient where no learning rate is needed. Kernel density estimation is utilized to estimate the density functions as well as their first and second derivatives, which makes the algorithm adaptive to the unobserved sources. Computer experiments confirm the efficiency of the proposed method.