Improved Complexity for Smooth Nonconvex Optimization: A Two-Level Online Learning Approach with Quasi-Newton Methods

TL;DR

This paper introduces a novel two-level online learning approach with quasi-Newton methods that improves the complexity of finding stationary points in smooth nonconvex optimization, especially in low-dimensional settings.

Contribution

It proposes a new algorithm combining online learning and quasi-Newton methods, achieving better gradient complexity bounds than existing methods for nonconvex optimization.

Findings

Achieves gradient query complexity of O(d^{1/4} ε^{-13/8})

First guarantee showing quasi-Newton methods can outperform gradient descent in nonconvex optimization

Provides a novel online learning reformulation for nonconvex stationary point finding

Abstract

We study the problem of finding an $\epsilon$-first-order stationary point (FOSP) of a smooth function, given access only to gradient information. The best-known gradient query complexity for this task, assuming both the gradient and Hessian of the objective function are Lipschitz continuous, is ${O}(\epsilon^{-7/4})$. In this work, we propose a method with a gradient complexity of ${O}(d^{1/4}\epsilon^{-13/8})$, where $d$ is the problem dimension, leading to an improved complexity when $d = {O}(\epsilon^{-1/2})$. To achieve this result, we design an optimization algorithm that, underneath, involves solving two online learning problems. Specifically, we first reformulate the task of finding a stationary point for a nonconvex problem as minimizing the regret in an online convex optimization problem, where the loss is determined by the gradient of the objective function. Then, we introduce a novel optimistic quasi-Newton method to solve this online learning problem, with the Hessian approximation update itself framed as an online learning problem in the space of matrices. Beyond improving the complexity bound for achieving an $\epsilon$-FOSP using a gradient oracle, our result provides the first guarantee suggesting that quasi-Newton methods can potentially outperform gradient descent-type methods in nonconvex settings.