Ding Tao PENG, Qi TANG, Xian ZHANG
We consider the constrained group sparse regression problem, where the loss function is convex. In order to obtain the exact continuous relaxation of the problem, we use the group Capped-L1 regularization to relax the group sparse regularization. First, we introduce three types of stationary points for the relaxed problem, that is, D(irectional)-stationary points, C(ritical)-stationary points, and L(ifted)-stationary points. We provide some equivalent descriptions for the three types of stationary points, based on which, we investigate the relationship among the three types of stationary points and obtain some properties of them. Furthermore, we analyze the necessary and sufficient optimality conditions for the original group sparse problem and the relaxed problem. At last, we investigate the relationship of the solutions between the original group sparse problem and the relaxed problem, not only from the point of view of global minimizers but also from the point of view of local minimizers. The results reveal the equivalence of the original problem and the relaxed problem.