Blind and semi-blind maximum likelihood techniques for multiuser multichannel identification

We investigate blind and semi-blind maximum likelihood techniques for multiuser multichannel identification. Two blind Deterministic ML methods based on cyclic prediction filters are presented [1]. The Iterative Quadratic ML (IQML) algorithm is used in [1] to solve it: this strategy does not perform well at low SNR and gives biased estimates due to the presence of noise. We propose a modification of IQML that we call DIQML to "denoise" it and explore a second strategy called Pseudo-Quadratic ML (PQML). As proposed in [2], PQML works well only at very high SNR. The solution we present here makes it work well at rather low SNR conditions and outperform DIQML. Like DIQML, PQML is proved to be consistent, asymptotically insensitive to the initialisation and globally convergent. Furthermore, it has the same performance as DML. A semi-blind extension combining these algorithms with training sequence based approaches is also studied. Simulations will illustrate the performance of the different algorithms which are found to be close to the Cramer-Rao bounds.