1: \begin{abstract}

2:   We study iterative regularization for linear models, when the bias is convex but not necessarily strongly convex.

3:   We characterize the stability properties of a primal-dual gradient based approach, analyzing its convergence in the presence of worst case deterministic noise.

4:   As a main example, we specialize and illustrate the results for the problem of robust sparse recovery.

5:   Key to our analysis is a combination of ideas from regularization theory and optimization in the presence of errors.

6:   Theoretical results are complemented by experiments showing that state-of-the-art performances can be achieved with considerable computational speed-ups.

7: \end{abstract}

8: