Toward Global Convergence of Gradient EM for Over-Parameterized Gaussian Mixture Models

The sublinear convergence rate is due to the algorithmic nature of learning over-parameterized GMM with gradient EM.