A "unimodal" blind equalization criterion

By reexamining some initialization strategies for blind equalizers, we develop an approximation to a common contrast function for equalization which enjoys a "unimodal" property: the solution obtained from the modified criterion is unique to within standard scale factor ambiguity, and moreover yields a perfect equalizer whenever such is attainable. In the more realistic situation in which perfect equalization is unattainable, the modified criterion yields a good approximation which is quantified herein, and is therefore justified as an initialization strategy. An on-line interpretation leads to an adaptive Volterra filter followed by a tensor product approximation which furnishes the coefficients of a linear equalizer, and ties in with some earlier work on equalizer design restricted to finite data sets and Volterra kernel approximations.