Stochastic Composite Least-Squares Regression with Convergence Rate O(1/n)

We consider the minimization of a function defined on a Riemannian manifold M accessible only through unbiased estimates of its gradients. We develop a geometric framework to transform a sequence of slowly converging iterates generated from stochastic gradient descent (SGD) on M to an averaged iterate sequence with a robust and fast O(1/n) convergence rate. We then present an application of our framework to geodesically-strongly-convex (and possibly Euclidean nonconvex) problems. Finally, we demonstrate how these ideas apply to the case of streaming k-PCA, where we show how to accelerate the slow rate of the randomized power method (without requiring knowledge of the eigengap) into a robust algorithm achieving the optimal rate of convergence.