EM's Convergence in Gaussian Latent Tree Models
This provides theoretical support for the practical use of maximum likelihood methods in latent tree models, extending global convergence guarantees for EM, but it is incremental as it builds on existing work.
The paper tackles the optimization landscape and convergence of the Expectation-Maximization (EM) algorithm in Gaussian latent tree models, showing that the unique non-trivial stationary point of the population log-likelihood is its global maximum and that EM is guaranteed to converge to it in the single latent variable case.
We study the optimization landscape of the log-likelihood function and the convergence of the Expectation-Maximization (EM) algorithm in latent Gaussian tree models, i.e. tree-structured Gaussian graphical models whose leaf nodes are observable and non-leaf nodes are unobservable. We show that the unique non-trivial stationary point of the population log-likelihood is its global maximum, and establish that the expectation-maximization algorithm is guaranteed to converge to it in the single latent variable case. Our results for the landscape of the log-likelihood function in general latent tree models provide support for the extensive practical use of maximum likelihood based-methods in this setting. Our results for the EM algorithm extend an emerging line of work on obtaining global convergence guarantees for this celebrated algorithm. We show our results for the non-trivial stationary points of the log-likelihood by arguing that a certain system of polynomial equations obtained from the EM updates has a unique non-trivial solution. The global convergence of the EM algorithm follows by arguing that all trivial fixed points are higher-order saddle points.