Asymptotic performance of ML methods for semi-blind channel estimation

Two channel estimation methods are often opposed: training sequence methods which use the information coming from known symbols and blind methods which use the information coming from the received signal without integrating the possible knowledge of symbols. Semiblind methods combine both information and appear more powerful than both methods separately. Two maximum-likelihood approaches to semi-blind SIMO channel estimation are presented, one based on a deterministic model and another on a Gaussian model. Their asymptotic performance are studied and compared to the Cramer-Rao bounds. The superiority of semi-blind over blind and training sequence methods, and of the Gaussian approach is demonstrated.