Initialization Techniques for Improved Convergence of LMS DFEs in Strong Interference Environments

The least-mean square (LMS) decision-feedback equalizer (DFE) was previously shown [1], [2] to possess an extended convergence time in an interference limited environment. In [1] it was shown that the convergence time can be significantly reduced by using the received samples and the training data to initialize (data-aided initialization) the LMS weights with an estimate for the Wiener weights. In this paper, two data- aided initialization techniques for equalization in the presence of severe narrowband interference are discussed and compared. The estimate of the Wiener filter is obtained from data-based averages of the autocorrelation matrix and the cross-correlation vector. The first technique is the Multistage Wiener Filter (MSWF) first proposed by Goldstein, et al. [3]. This algorithm provides a reduced complexity approach by approximating the Wiener filter in a lower dimensional subspace. The second technique is a parametric approximation to the Direct Matrix Inversion (DMI) solution based on the Gohberg-Semencul formula [4], [5] to obtain the inverse in a computationally efficient fashion. Both techniques were compared in terms of complexity (i.e. the number of multiplications required) and BER performance as compared to the theoretical Wiener filter for the DFE. The MSWF requires fewer training symbols than the approximation to the DMI solution in order to approach the BER performance of the theoretical Wiener filter. The parametric approximation to the DMI solution is computationally efficient but exhibits instability due to assumptions made on the structure of the correlation matrix and when the minimum eigenvalue is close to zero at high SNR.