PCA neural models and blind signal separation

Neural models for the blind separation of signals (BSS) from their linear mixtures are traditionally based on higher order moments. For example, models based on PCA extensions, such as nonlinear PCA, perform analysis of signals into independent components. However, second order models have been also used for BSS under the assumptions of non-correlation (as opposed to independence) and of the different spectral coloring of the sources. Yet these models were either based on second order optimization or on linear extensions of PCA. We show that standard PCA neural models can perform BSS through the equation (temporal filtering+PCA=BSS), which states that BSS is PCA preceded by temporal filtering. This result is both shown theoretically and demonstrated by simulation. Although almost any temporal filter will work, the question on the optimal filter is still open. Some discussion on this issue for filters of length 2 is given