Cubic Regularization is the Key! The First Accelerated Quasi-Newton Method with a Global Convergence Rate of $O(k^{-2})$ for Convex Functions

TL;DR

This paper introduces a novel accelerated Quasi-Newton method with a global convergence rate of O(k^{-2}) for convex functions, bridging the gap between practical efficiency and theoretical guarantees.

Contribution

It proposes the first accelerated Quasi-Newton method with a proven O(k^{-2}) convergence rate for convex functions, using cubic regularization and inexact Hessian approximations.

Findings

Achieves a convergence rate of O(k^{-2}) for general convex functions.

Outperforms existing first and second-order methods in practical experiments.

Provides adaptive methods with real-time control of approximation errors.

Abstract

In this paper, we propose the first Quasi-Newton method with a global convergence rate of $O(k^{-1})$ for general convex functions. Quasi-Newton methods, such as BFGS, SR-1, are well-known for their impressive practical performance. However, they may be slower than gradient descent for general convex functions, with the best theoretical rate of $O(k^{-1/3})$. This gap between impressive practical performance and poor theoretical guarantees was an open question for a long period of time. In this paper, we make a significant step to close this gap. We improve upon the existing rate and propose the Cubic Regularized Quasi-Newton Method with a convergence rate of $O(k^{-1})$. The key to achieving this improvement is to use the Cubic Regularized Newton Method over the Damped Newton Method as an outer method, where the Quasi-Newton update is an inexact Hessian approximation. Using this approach, we propose the first Accelerated Quasi-Newton method with a global convergence rate of $O(k^{-2})$ for general convex functions. In special cases where we can improve the precision of the approximation, we achieve a global convergence rate of $O(k^{-3})$, which is faster than any first-order method. To make these methods practical, we introduce the Adaptive Inexact Cubic Regularized Newton Method and its accelerated version, which provide real-time control of the approximation error. We show that the proposed methods have impressive practical performance and outperform both first and second-order methods.