1: \begin{abstract}\hspace{0.4cm}

2:  The estimation problem in a high regression model with

3: structured sparsity is investigated. An algorithm using a two steps

4: block thresholding procedure called GR-LOL is provided. Convergence

5: rates are produced: they depend on simple coherence-type indices of

6: the Gram matrix -easily checkable on the data- as well as sparsity

7: assumptions of the model parameters measured by a combination of

8: $l_1$ within-blocks with $l_q,q<1$ between-blocks norms. The

9: simplicity of the coherence indicator suggests ways to optimize the

10: rates of convergence when the group structure is not naturally given

11: by the problem and is unknown. In such a case, an auto-driven

12: procedure is provided to determine the regressors groups (number and

13: contents). An intensive practical study compares our grouping

14: methods with the standard LOL algorithm. We prove that the grouping

15: rarely deteriorates the results but can improve them very

16: significantly. GR-LOL is also compared with group-Lasso procedures

17: and exhibits a very encouraging behavior. The results are quite

18: impressive, especially when GR-LOL algorithm is combined with a

19: grouping pre-processing.

20: \end{abstract}

21: