1: \begin{abstract}

2:   We consider the problem of finding critical points of functions that

3:   are non-convex and non-smooth. Studying a fairly broad class of such

4:   problems, we analyze the behavior of three gradient-based methods

5:   (gradient descent, proximal update, and Frank-Wolfe update). For

6:   each of these methods, we establish rates of convergence for general

7:   problems, and also prove faster rates for continuous sub-analytic

8:   functions. We also show that our algorithms can escape strict saddle

9:   points for a class of non-smooth functions, thereby generalizing

10:   known results for smooth functions. Our analysis leads to a

11:   simplification of the popular CCCP algorithm, used for optimizing

12:   functions that can be written as a difference of two convex

13:   functions.  Our simplified algorithm retains all the convergence

14:   properties of CCCP, along with a significantly lower cost per

15:   iteration. We illustrate our methods and theory via applications to

16:   the problems of best subset selection, robust estimation, mixture

17:   density estimation, and shape-from-shading reconstruction.

18: \end{abstract}