Two methods for autoregressive estimationin noise

The maximum-likelihood (ML) and the expectation-maximization criteria have been previously used in the problem of autoregressive estimation in noise. This paper presents a thorough comparative study of these techniques. Despite these criteria lead in both cases to apparently similar algorithms, the methodological differences and connections between both approaches are explored. Their performance, speed of convergence, and robustness of the solution are assessed with the help of simulated experiments. Further research work at increasing robustness in the ML approach is finally proposed.