On blind decision feedback equalization

Derived from the predictive structure of (non-blind) minimum mean square error decision feedback equalizer (MMSE-DFE), two new schemes, referred to as DFE(CMA, SF) and DFE(CMA, CMA), are investigated. The DFE(CMA, SF) design optimizes the forward filter using the constant modulus criterion applied to its output and obtains the feedback filter by the spectral factorization of the observation process. In contrast, the DFE(CMA, CMA) uses only the constant modulus criterion to design both the forward and feedback filters. At high SNR, both approaches offer asymptotic near optimal performance (in mean-square error). They are different in their implementation complexities and efficiencies. We also show that, at high SNR, the feedback filter in DFE(CMA, CMA) converges to the global minimum at a rate equal to that of a nonblind case. At low SNR, however, CMA updates of the feedback filter is affected by error propagation.