Non-asymptotic Global Convergence Analysis of BFGS with the Armijo-Wolfe Line Search
Qiujiang Jin,Ruichen Jiang,Aryan Mokhtari
2024-04-26
Abstract:In this paper, we establish the first explicit and non-asymptotic global convergence analysis of the BFGS method when deployed with an inexact line search scheme that satisfies the Armijo-Wolfe conditions. We show that BFGS achieves a global convergence rate of $(1-\frac{1}{\kappa})^k$ for $\mu$-strongly convex functions with $L$-Lipschitz gradients, where $\kappa=\frac{L}{\mu}$ denotes the condition number. Furthermore, if the objective function's Hessian is Lipschitz, BFGS with the Armijo-Wolfe line search achieves a linear convergence rate only determined by the line search parameters and independent of the condition number. These results hold for any initial point $x_0$ and any symmetric positive definite initial Hessian approximation matrix $B_0$, although the choice of $B_0$ affects the iteration count required to attain these rates. Specifically, we show that for $B_0 = LI$, the rate of $O((1-\frac{1}{\kappa})^k)$ appears from the first iteration, while for $B_0 = \mu I$, it takes $d\log \kappa$ iterations. Conversely, the condition number-independent linear convergence rate for $B_0 = LI$ occurs after $O\left(\kappa\left(d +\frac{M \sqrt{f(x_0)-f(x_*)}}{\mu^{3/2}}\right)\right)$ iterations, whereas for $B_0 = \mu I$, it holds after $O\left(\frac{M \sqrt{f(x_0)-f(x_*)}}{\mu^{3/2}}\left(d\log \kappa + \kappa\right)\right)$ iterations. Here, $d$ denotes the dimension of the problem, $M$ is the Lipschitz parameter of the Hessian, and $x_*$ denotes the optimal solution. We further leverage these global linear convergence results to characterize the overall iteration complexity of BFGS when deployed with the Armijo-Wolfe line search.
Optimization and Control