1: \begin{abstract}

2: This paper proposes a novel class of block quasi-Newton methods for convex optimization which we call symmetric rank-$k$ (SR-$k$) methods.

3: Each iteration of SR-$k$ incorporates the curvature information with~$k$ Hessian-vector products achieved from the greedy or random strategy.

4: We prove SR-$k$ methods have the local superlinear convergence rate of $\OM\big((1-k/d)^{t(t-1)/2}\big)$ for minimizing smooth and strongly convex function, where $d$ is the problem dimension and $t$ is the iteration counter.

5: This is the first explicit superlinear convergence rate for block quasi-Newton methods, 

6: and it successfully explains why block quasi-Newton methods converge faster than ordinary quasi-Newton methods in practice.

7: 

8:  \end{abstract}

9: