Lower Bounds for Non-Convex Stochastic Optimization

TL;DR

This paper establishes fundamental lower bounds on the number of stochastic gradient queries needed to find approximate stationary points in non-convex optimization, confirming the optimality of existing algorithms.

Contribution

It provides tight lower bounds for stochastic first-order methods in non-convex optimization, demonstrating the optimality of stochastic gradient descent and variance reduction techniques.

Findings

Any algorithm requires at least  queries to find an -stationary point.

The lower bound of  is tight and matches the performance of stochastic gradient descent.

In a more restrictive model, at least  queries are needed, confirming the optimality of variance reduction methods.

Abstract

We lower bound the complexity of finding $\epsilon$-stationary points (with gradient norm at most $\epsilon$) using stochastic first-order methods. In a well-studied model where algorithms access smooth, potentially non-convex functions through queries to an unbiased stochastic gradient oracle with bounded variance, we prove that (in the worst case) any algorithm requires at least $\epsilon^{-4}$ queries to find an $\epsilon$ stationary point. The lower bound is tight, and establishes that stochastic gradient descent is minimax optimal in this model. In a more restrictive model where the noisy gradient estimates satisfy a mean-squared smoothness property, we prove a lower bound of $\epsilon^{-3}$ queries, establishing the optimality of recently proposed variance reduction techniques.