On information maximization and blind signal deconvolution

We investigate two algorithms for blind signal deconvolution that have been proposed in the literature. We derive a clear interpretation of the information theoretic objective function in terms of signal processing and show that only one is appropriate to solve the deconvolution problem, while the other will only work if the unknown filter is constrained to be minimum phase. Moreover we argue that the blind deconvolution task is more sensitive to a mismatch of the density model than currently expected. While there exist theoretical arguments and practical evidence that blind signal separation requires only a rough approximation of the signal density this is not the case for blind signal deconvolution. We give a simple example that supports our argumentation and formulate a sufficiently adaptive density model to properly solve that problem