When Do Gradients Suffice? A Characterization of First-Order Convex OptimizationResearch Paper
First-order methods (gradient descent, Mirror Descent, their stochastic and accelerated variants) are the default tools of large-scale convex optimization, and their guarantees are usually stated in a Euclidean or geometry. A basic question underlies all of them: in which normed spaces can a convex Lipschitz function on the unit ball be minimized to…