Title: On the Convergence of the EM Algorithm: From the Statistical Perspective

Abstract: The Expectation-Maximization (EM) algorithm is an iterative method that is often used for parameter estimation in incomplete data problems. Despite much theoretical endeavors devoted to understand the convergence behavior of the EM algorithm, some ubiquitous phenomena still remain unexplained. As observed in both numerical experiments and real applications, the convergence rate of the optimization error of an EM sequence is data-dependent: it is more spread-out when the sample size is smaller while more concentrated when the sample size is larger; and for a fixed sample size, one observes random fluctuations of the convergence rate of EM sequences constructed from different sets of i.i.d. realizations of the underlying distribution of the model. In this paper, by introducing an adaptive optimal empirical convergence rate $\overline{K}_{n}$, we develop a theoretical framework that quantitatively characterizes the intrinsic data-dependent nature of the convergence behaviors of empirical EM sequences in open balls of the true population parameter $\theta^{*}$. Our theory precisely explains the aforementioned randomness in convergence rate and directly affords a theoretical guarantee for the statistical consistency of the EM algorithm, closing the gap between practical observations and theoretical explanations. We apply this theory to the EM algorithm on three classical models: the Gaussian Mixture Model, the Mixture of Linear Regressions and Linear Regression with Missing Covariates and obtain model-specific results on the upper bound and the concentration property of the optimal empirical convergence rate $\overline{K}_{n}$.