Is gradient descent appropriate for entropy-based source separation?

Many blind source separation (BSS) methods are based on minimization of entropy-based criteria. Gradient-based methods have been observed to converge slowly when applied to such criteria. In this paper, we argue that the non-convex “bird's beak” shape of such criteria is the reason why gradient BSS methods perform poorly. To overcome these limitations, we propose a novel local parameter optimization method based upon curve fitting to a density-based entropy measure. Simulations show that the novel method can be used as an effective refinement procedure for an existing BSS method.