1: \begin{abstract}

2: In this paper, we propose a successive convex approximation framework

3: for sparse optimization where the nonsmooth regularization function

4: in the objective function is nonconvex and it can be written as the

5: difference of two convex functions. The proposed framework is based

6: on a nontrivial combination of the majorization-minimization framework

7: and the successive convex approximation framework proposed in literature

8: for a convex regularization function. The proposed framework has several

9: attractive features, namely, i) flexibility, as different choices

10: of the approximate function lead to different type of algorithms;

11: ii) fast convergence, as the problem structure can be better exploited

12: by a proper choice of the approximate function and the stepsize is

13: calculated by the line search; iii) low complexity, as the approximate

14: function is convex and the line search scheme is carried out over

15: a differentiable function; iv) guaranteed convergence to a stationary

16: point. We demonstrate these features by two example applications in

17: subspace learning, namely, the network anomaly detection problem and

18: the sparse subspace clustering problem. Customizing the proposed framework

19: by adopting the best-response type approximation, we obtain soft-thresholding

20: with exact line search algorithms for which all elements of the unknown

21: parameter are updated in parallel according to closed-form expressions.

22: The attractive features of the proposed algorithms are illustrated

23: numerically.

24: \end{abstract}