1: \begin{abstract}  

2: Quasi-Newton algorithms are among the most popular iterative methods for solving unconstrained minimization problems, largely due to their favorable superlinear convergence property.  

3: %

4: However, existing results for these algorithms are limited as  they provide either (i) a global convergence guarantee with an \textit{asymptotic} superlinear convergence rate, or (ii) a \textit{local} non-asymptotic superlinear rate for the case that the initial point and the initial Hessian approximation are chosen properly. Furthermore, these results are not composable, since when the iterates of the globally convergent methods reach the region of local superlinear convergence, it cannot be guaranteed that the Hessian approximation matrix will satisfy the required conditions for a non-asymptotic local superlienar convergence rate.  

5: %

6: %

7: %

8: %

9: %

10: In this paper, we close this gap and present  

11: the first \textit{globally} convergent quasi-Newton method with an \emph{explicit non-asymptotic} superlinear convergence rate. 

12: Unlike classical quasi-Newton methods, we 

13: build our algorithm upon the hybrid proximal extragradient method and propose a novel \emph{online learning} framework for updating the Hessian approximation matrices. 

14: Specifically, guided by the convergence analysis, we formulate the Hessian approximation update as an online convex optimization problem in the space of matrices, and relate the bounded regret of the online problem to the superlinear convergence of our method. 

15: \end{abstract}

16: