1: \begin{abstract}

2:   This paper concerns with the group zero-norm regularized least squares estimator

3:   which, in terms of the variational characterization of the zero-norm, can be obtained

4:   from a mathematical program with equilibrium constraints (MPEC). By developing

5:   the global exact penalty for the MPEC, this estimator is shown to arise from

6:   an exact penalization problem that not only has a favorable bilinear structure

7:   but also implies a recipe to deliver equivalent DC estimators such as the SCAD

8:   and MCP estimators. We propose a multi-stage convex relaxation approach

9:   (GEP-MSCRA) for computing this estimator, and under a restricted strong convexity

10:   assumption on the design matrix, establish its theoretical guarantees

11:   which include the decreasing of the error bounds for the iterates to the true

12:   coefficient vector and the coincidence of the iterates after finite steps with

13:   the oracle estimator. Finally, we implement the GEP-MSCRA with the subproblems

14:   solved by a semismooth Newton augmented Lagrangian method (ALM) and compare

15:   its performance with that of SLEP and MALSAR, the solvers for the weighted

16:   $\ell_{2,1}$-norm regularized estimator, on synthetic group sparse regression problems

17:   and real multi-task learning problems. Numerical comparison indicates that

18:   the GEP-MSCRA has significant advantage in reducing error and achieving

19:   better sparsity than the SLEP and the MALSAR do.

20:  \end{abstract}