Independent component analysis by wavelets

This paper introduces a new approach in solving the ICA problem using a method that fits in the contrast and minimize paradigm, mostly found in the ICA literature. In our case, the contrast is a L2 norm dependence measure, which constitutes an alternative to the usual criteria, based on mutual information. We propose a non parametric evaluation of the L2 contrast, using a wavelet projection estimator. The mean square error of the procedure is bounded under Besov assumptions. Finally, we provide a set of simulations to show how the method performs in practice.