Accelerating the EM algorithm by smoothing: A special case Original Research Article

The standard em (estimate, maximize) algorithm exhibits very slow convergence. In the special test case where the underlying positive linear system has a unique solution, we describe two iterations, based on adaptively choosing the smoothing in the ems (estimate, maximize, smooth) algorithm, which accelerate the convergence of the em algorithm. The resulting algorithms, called adaptive ems (a-ems) algorithms, thus overcome the inaccuracy of ems while retaining its more rapid convergence and reduced overall computational cost.