1: \begin{abstract}

2:   We prove that the finite-difference based derivative-free descent (FD-DFD) methods have a capability to find the global minima for a class of multiple minima problems. Our main result shows that, for a class of multiple minima objectives that is extended from strongly convex functions with Lipschitz-continuous gradients, the iterates of FD-DFD converge to the global minimizer $x_*$ with the linear convergence $\|x_{k+1}-x_*\|_2^2\leqslant\rho^k \|x_1-x_*\|_2^2$ for a fixed $0<\rho<1$ and any initial iteration $x_1\in\mathbb{R}^d$ when the parameters are properly selected. Since the per-iteration cost, i.e., the number of function evaluations, is fixed and almost independent of the dimension $d$, the FD-DFD algorithm has a complexity bound $\mathcal{O}(\log\frac{1}{\epsilon})$ for finding a point $x$ such that the optimality gap $\|x-x_*\|_2^2$ is less than $\epsilon>0$. Numerical experiments in various dimensions from $5$ to $500$ demonstrate the benefits of the FD-DFD method.

3: 

4: \keywords{multiple minima problem \and global minima \and nonconvex \and derivative-free \and convergence rate \and complexity}

5: % \PACS{PACS code1 \and PACS code2 \and more}

6: \subclass{65K05 \and 68Q25 \and 90C26 \and 90C56}

7: \end{abstract}