I. Introduction
Linear blind source separation (BSS) consists of recovering a set of original source signals from instantaneous linear mixtures of these sources. A standard approach to the problem is to assume that the sources are all statistically independent of one another. Under this assumption, independent component analysis (ICA) [1] then tries to find a linear transform of the mixtures such that the recovered signals are as independent as possible. However, without any further constraint this approach cannot be applied to similar nonlinear BSS problems for the following reason: If and are two independent random variables, then and are also independent for any and [2] , p. 132. Therefore, there always exist an infinity of solutions if the space of nonlinear functions is not limited by some further means [3]. Work exploiting further constraints beyond independence in the nonlinear domain has involved extracting known nonlinearities depending upon unknown parameters [4], incorporating volume conservation transforms [5], extracting nonlinear functions whose inverse are constrained to be well approximated by an a priori neural network [6], [7], and using postnonlinear mixtures with componentwise nonlinearities [8].