Pathwise interference cancellation for a DS-CDMA uplink

This thesis is concerned with the application of interference cancellation for a DS-CDMA uplink and particularly, pathwise processing. There are essentially two ways of handling the interference cancellation, either before or after the various signal components (multipath, multiple antennas) are recombined. These methods are known as precombining interference cancellation and postcombining interference cancellation, respectively. It will be shown that the received signal can be factored into two components,
one of them relying only on slow parameters whereas the other depends on fast parameters also. This observation has motivated pathwise interference cancellation, which only requires knowledge of the slow parameters. Besides the advantage of less stringent adaptation requirements, such a filter allows improved estimation of the fastly varying complex amplitude coefficients since the estimated path components contain the signal of interest with an increased SINR compared to the received signal. While the pathwise linear interference cancellation approaches produce good performance, their major drawback is a high implementational complexity. We therefore consider lower complexity alternatives while maintaining the pathwise approach. In particular, we investigate the application of polynomial expansion receivers in a pathwise context. The main complexity arising from an LMMSE approach is an inverse in a correlation matrix, a fact that is well known. The principle of polynomial expansion is to approximate this inverse by a low order, weighted polynomial in the correlation matrix to be inverted. We will show that the introduction of carefully chosen diagonal weighting matrices (instead of scalars, as previously proposed) can
substantially improve the performance over previously proposed methods under power imbalances between users or paths. However, it is difficult to obtain analytical performance expressions for polynomial expansion using standard techniques since the performance is always a function of the correlation properties of the set of spreading codes used. Hence, we resort to methods only recently introduced in the communications community which tackle this issue by letting the system dimensions grow to infinity. Such a large system analysis allows to obtain quantitative expressions for the asymptotic output SINR of linear receivers based on the properties of certain classes of large random matrices.