33
38

EM Converges for a Mixture of Many Linear Regressions

Abstract

We study the convergence of the Expectation-Maximization (EM) algorithm for mixtures of linear regressions with an arbitrary number kk of components. We show that as long as signal-to-noise ratio (SNR) is Ω~(k)\tilde{\Omega}(k), well-initialized EM converges to the true regression parameters. Previous results for k3k \geq 3 have only established local convergence for the noiseless setting, i.e., where SNR is infinitely large. Our results enlarge the scope to the environment with noises, and notably, we establish a statistical error rate that is independent of the norm (or pairwise distance) of the regression parameters. In particular, our results imply exact recovery as σ0\sigma \rightarrow 0, in contrast to most previous local convergence results for EM, where the statistical error scaled with the norm of parameters. Standard moment-method approaches may be applied to guarantee we are in the region where our local convergence guarantees apply.

View on arXiv
Comments on this paper