A blind detection-guided approach for normalised multi-modulus crosstalk estimator in sparse multi-user DSL channels

In high-speed digital subscriber lines (DSL), far-end crosstalk is the main limiting factor on data rates. However, most of the crosstalk is due to the neighbouring twisted pairs in the binder. Therefore, the crosstalk channel matrix is sparse. Using Level 3 of Dynamic Spectrum Management, users are co-ordinated at the central office to cancel the crosstalk. Means for estimating the crosstalk canceller matrix are of critical importance for the cancellation to prove effective. Preferably, the estimation procedure should have low overhead both in computation and bandwidth. Normalised least mean squares (NLMS) based adaptive crosstalk cancellers (Gujrathi et al. (2009) [1]) have a low computational overhead but use a training sequence to ensure they converge adequately. However, using a training sequence consumes some amount of bandwidth which can be avoided if an unsupervised or blind algorithm like a normalised multi-modulus algorithm (NMMA) is used instead. A limitation of NMMA is that its convergence time is often longer than that of the NLMS algorithm. Furthermore, this is made worse as the number of canceller coefficients is made larger. In application to adaptive crosstalk cancellation within the multi-user DSL binder-channel, we argue that the convergence time can be significantly decreased by using an activity detector to exclude canceller coefficients below an appropriate minimum. In this paper, we present an activity detector design using a thresholding criterion based on the least squares technique, Akaike's information criterion (Homer et al. (1998) [2]) and Donoho's universal thresholding principle (Donoho (1995) [3]). This enables us to identify the significant crosstalkers within a DSL binder for each user. We further incorporate this strategy within the blind estimation NMMA and propose an enhanced crosstalk canceller. Our simulations indicate this multi-modulus detection-guided crosstalk canceller demonstrates improved convergence speed and has a steady state error close to that of the standard (non-detection-guided) canceller.