On the learning behavior of decision feedback equalizers

The performance analysis of decision feedback equalizers (DFE), either fractional or symbol-spaced, is commonly based on the Wiener solution. This solution can only be computed once characteristic values of the channel and noise variance are known. In the "real" world, the solution needs to be estimated. In low-bit rate systems, the complexity of algorithms is usually not an issue and least-squares solutions approximating the Wiener solution with high to accuracy are possible. For higher bit rate systems however, a gradient-type procedure like the LMS algorithm seems unavoidable. Given a training sequence of limited length, the learning behavior of LMS can considerably worsen the performance of the DFE. This paper shows some insight into the undesired effects.