21
112

Unified Optimal Analysis of the (Stochastic) Gradient Method

Abstract

In this note we give a simple proof for the convergence of stochastic gradient (SGD) methods on μ\mu-convex functions under a (milder than standard) LL-smoothness assumption. We show that for carefully chosen stepsizes SGD converges after TT iterations as O(LR2exp[μ4LT]+σ2μT)O\left( LR^2 \exp \bigl[-\frac{\mu}{4L}T\bigr] + \frac{\sigma^2}{\mu T} \right) where σ2\sigma^2 measures the variance in the stochastic noise. For deterministic gradient descent (GD) and SGD in the interpolation setting we have σ2=0\sigma^2 =0 and we recover the exponential convergence rate. The bound matches with the best known iteration complexity of GD and SGD, up to constants.

View on arXiv
Comments on this paper