1: \begin{abstract}

2: Expectation Maximization (EM) is among the most popular algorithms for

3: estimating parameters of statistical models.

4: However, EM, which is an iterative algorithm based on the maximum likelihood

5: principle, is generally only guaranteed to find stationary points of the

6: likelihood objective, and these points may be far from any maximizer.

7: This article addresses this disconnect between the statistical principles behind

8: EM and its algorithmic properties.

9: Specifically, it provides a global analysis of EM for specific models in which

10: the observations comprise an i.i.d.~sample from a mixture of two Gaussians.

11: This is achieved by (i) studying the sequence of parameters from idealized

12: execution of EM in the infinite sample limit, and fully characterizing the limit

13: points of the sequence in terms of the initial parameters; and then (ii) based

14: on this convergence analysis, establishing statistical consistency (or lack

15: thereof) for the actual sequence of parameters produced by EM.

16: \end{abstract}

17: