MULTISTAGE INTERFERENCE CANCELLATION TECHNIQUES IN CDMA

Abstract

In this thesis, we look at some multistage suboptimal multiuser detectors that are simpler to implement than the optimal detector. We first study the iterative implementation of linear detectors. Then we study detectors based on the relaxation of the optimum multiuser detection problem to tractable nonlinear continuous problems. It will be shown that these two types of detectors take the form of linear and non-linear interference cancellation structures respectively. It has previously been shown that well-known iterations for solving a set of linear equations correspond to linear interference cancellation structures for short-code CDMA. Here, we suggest applying a block-wise iteration which consists of an outer and an inner iteration. The outer iteration used is the Gauss-Seidel method while for the inner iteration, we study the Jacobi and the Gauss-Seidel iterations, their over-relaxation extensions as well as the conjugate gradient iteration. It is then possible to relate all the iterations to cancellation structures directly in the time domain. The performance of the detectors is studied via computer simulations where it is found that the block-wise conjugate gradient canceller can perform as well as a windowed conjugate gradient iteration and marginally better than the block-wise Gauss-Seidel approach. The second part of the thesis focuses on the relaxation of the binary constrained maximum-likelihood (ML) problem. We will consider the box-constrained and sphere-constrained ML problems. Iterative algorithms will be used to solve these problems. For the box-constrained ML problem, special cases of its algorithm correspond to known, non-linear successive and parallel interference cancellation structures, using a clipped soft decision function for making tentative decisions. It is also shown that the optimal solution of the sphere-constrained ML problem is a linear detector. The BER performance of these detectors is studied via computer simulations and the expected performance improvement over unconstrained ML is verified.