Error propagation is a serious concern when a decision feedback equalizer (DFE) is used in a communication system. This paper describes a method of mitigating the effects of error propagation by constraining the feedback tap coefficients. It is shown that the most natural method of constraining the feedback taps is to constrain the 1-norm of the tap vector. The paper also considers a constraint on the 2-norm of the feedback tap vector. The proposed method is demonstrated using the trellis coded 8-VSB system used
by the ATSC terrestrial broadcast standard for digital television. Results show that the constraints do reduce error propagation in the DFE, but the performance is considerably better when a zero delay trellis decoder is used to determine the decisions in the feedback filter.
In this paper, we show how the convergence time of equalizers for 8-VSB based on the conjugate gradient (CG) algorithm can be considerably improved through initialization based on a channel estimate. We derive real and complex minimum mean-square error (MMSE) equalizers and implement them adaptively using the conjugate gradient, recursive least squares (RLS), and least mean squares (LMS) algorithms. We show that both CG and RLS have similar convergence times --- both are much faster than LMS. Since the CG algorithm is easily initialized, we compare several methods of initialization to determine how each affects convergence and then apply the best methods to initialize equalizers using channel estimates. We find that initializing the correlation matrices and filling the feedback taps with training symbols greatly speeds convergence of the CG adaptive equalizer, potentially approaching the rate of convergence when running the algorithm on the matrix equations using the actual channel.