Iterative spectral independent component analysis

Linear independent component analysis (ICA) is a fundamental problem in signal processing. In this work we study the Spectral ICA approach introduced by Singer that is based on the Diffusion Framework. We analyze its asymptotic optimality condition, related to the discretization error of the Graph Laplacian with respect to the continuous backward Fokker–Planck operator. Thus, we derive an iterative Diffusion Framework-based spectral ICA formulation, that is rigorously shown to reduce the discretization error of the Graph Laplacian by iteratively estimating and canceling-out ICA components. The proposed scheme is shown to compare favourably with contemporary state-of-the-art linear ICA schemes, when applied to the demixing of signals and images.