Abstract
We consider the problem of minimizing a difference-of-convex (DC) function,which can be written as the sum of a smooth convex function with Lipschitzgradient, a proper closed convex function and a continuous possibly nonsmoothconcave function. We refine the convergence analysis in [38] for the proximalDC algorithm with extrapolation (pDCA$_e$) and show that the whole sequencegenerated by the algorithm is convergent when the objective is level-bounded,{\em without} imposing differentiability assumptions in the concave part. Ouranalysis is based on a new potential function and we assume such a function isa Kurdyka-{\L}ojasiewicz (KL) function. We also establish a relationshipbetween our KL assumption and the one used in [38]. Finally, we demonstrate howthe pDCA$_e$ can be applied to a class of simultaneous sparse recovery andoutlier detection problems arising from robust compressed sensing in signalprocessing and least trimmed squares regression in statistics. Specifically, weshow that the objectives of these problems can be written as level-bounded DCfunctions whose concave parts are {\em typically nonsmooth}. Moreover, for alarge class of loss functions and regularizers, the KL exponent of thecorresponding potential function are shown to be 1/2, which implies that thepDCA$_e$ is locally linearly convergent when applied to these problems. Ournumerical experiments show that the pDCA$_e$ usually outperforms the proximalDC algorithm with nonmonotone linesearch [24, Appendix A] in both CPU time andsolution quality for this particular application.