Convergence tradeoffs for EM-type algorithms
Alfred O. Hero · 2003
The author analyzes the convergence properties of the EM algorithm for iteratively approximating the maximum likelihood estimate. A radius of convergence is specified and the asymptotic rate of convergence of the algorithm derived via the multivariate Taylor expansion with remainder. The radius and rate of convergence generally depend on the choice of complete data. The results can be used to evaluate different choices of complete data space in terms of algorithm performance.>