Finite sample effects of the fast ICA algorithm
Neurocomputing
| Hi-index | 754.84 |
It is pointed out that certain criteria proposed in the paper by Shalvi and Weinstein (see ibid., vol.36, no.2, p.312-21, 1990) have been proposed before in the context of real-valued signals. Furthermore, it is noted that the assertion made in the paper that all local maxima of the proposed criteria correspond to the desired equalizer solution, holds true, strictly speaking, only for infinite length equalizers. There exist counterexamples which show existence of false local maxima when finite-length equalizers are used. Finally, it is also noted that in the presence of additive Gaussian noise of unknown power spectral density at the input to the equalizer (equivalently, at the channel output), it is not clear if the methods proposed in the paper will yield the desired response up to a constant gain