1: \begin{abstract}

2: By analyzing accelerated proximal gradient methods under a local quadratic growth condition, we show that restarting these algorithms at any

3: frequency gives a globally linearly convergent algorithm. This result was previously

4:  known only for long enough frequencies. 

5: 

6: Then, as the rate of convergence depends on the match between the frequency and the quadratic error bound, we design

7: a scheme to automatically adapt the frequency of restart from

8: the observed decrease of the norm of the gradient mapping.

9: Our algorithm has a better theoretical bound than previously proposed

10: methods for the adaptation to the quadratic error bound of the objective.

11: 

12: We illustrate the efficiency of the algorithm on a Lasso problem and on a regularized logistic regression problem.

13: \end{abstract}

14: