1: \begin{abstract}

2: In this work we propose and analyze a novel approach for group sparse recovery. It is based on regularized least squares

3: with an $\ell^0(\ell^2)$ penalty, which penalizes the number of nonzero groups. One distinct

4: feature of the approach is that it has the built-in decorrelation mechanism within each group,

5: and thus can handle challenging strong inner-group correlation. We provide a complete analysis of

6: the regularized model, e.g., existence of a global minimizer, invariance property, support

7: recovery, and properties of block coordinatewise minimizers.  Further, the regularized

8: problem admits an efficient primal dual active set algorithm

9: with a provable finite-step global convergence. At each iteration, it involves solving a least-squares

10: problem on the active set only, and exhibits a fast local convergence, which makes the method extremely efficient for recovering group sparse

11: signals. Extensive numerical experiments are presented to illustrate salient features of the

12: model and the efficiency and accuracy of the algorithm. A comparative study indicates

13: its competitiveness with existing approaches.

14: \end{abstract}