A new regularized QRD recursive least M-estimate algorithm: Performance analysis and applications

This paper proposes a new regularized QR decomposition based recursive least M-estimate (R-QRRLM) adaptive filtering algorithm and studies its mean and mean square convergence performance and application to acoustic echo cancellation (AEC). The proposed algorithm extends the conventional RLM algorithm by imposing a weighted L₂ regularization term on the coefficients to reduce the variance of the estimator. Moreover, a QRD-based algorithm is employed for efficient recursive implementation and improved numerical property. The mean convergence analysis shows that a bias solution to the classical Wiener solution will be introduced due to the regularization. The steady-state excess mean square error (EMSE) is derived and it suggests that the variance will decrease while the bias will increase with the regularization parameter. Therefore, regularization can help to trade bias for variance. In this study, the regularization parameter can be adaptively selected and the resultant variable regularization parameter QRRLM (VR-QRRLM) algorithm can obtain both high immunity to input variation and low steady-state EMSE values. The theoretical results are in good agreement with simulation results. Computer simulation results on AEC show that the R-QRRLM and VR-QRRLM algorithms considerably outperform the traditional RLS algorithm when the input signal level is low or during double talk.