1: \begin{abstract}

2: We study $\ell^1$ regularized least squares optimization problem in a separable Hilbert space. We show that the iterative soft-thresholding algorithm (ISTA) converges linearly, 

3: without making any assumption on the linear operator into play or on the problem.

4: The result is obtained combining two key concepts:  the notion of {\it extended support}, a finite set containing the support, and the notion of  {\it conditioning over finite dimensional sets}.

5: We prove that ISTA identifies the solution extended support after a finite number of  iterations, and we derive linear convergence from the conditioning property, 

6: which is always satisfied for $\ell^1$ regularized least squares problems. 

7: Our analysis extends to the the entire class of thresholding gradient algorithms, for which we provide a conceptually new proof of strong convergence, as well as  convergence rates.

8: 

9: \bigskip

10: 

11: \noindent \textsc{Keywords.} Forward-Backward method, support identification,  conditioning, convergence rates.

12: 

13: \medskip

14: 

15: \noindent \textsc{MSC.}  49K40, 49M29, 65J10, 65J15, 65J20, 65J22, 65K15, 90C25, 90C46.

16: %

17: %\smallskip

18: %

19: %\noindent\hrulefill

20: %

21: %\noindent This material is supported by the Center for Brains, Minds and Machines, funded by NSF STC award CCF-1231216, and the Air Force project FA9550-17-1-0390. 

22: %G. Garrigos is supported by the European Research Council (ERC project NORIA), and part of his work was done while being a postdoc at the LCSL-IIT@MIT.

23: %L. Rosasco acknowledges the financial support of the Italian Ministry of Education, University and Research FIRB project RBFR12M3AC. 

24: %S. Villa is supported by the INDAM GNAMPA research project 2017 Algoritmi di ottimizzazione ed equazioni di evoluzione ereditarie.

25: %L. Rosasco and S. Villa acknowledge the financial support from the EU project 777826 - NoMADS.

26: %

27: %\noindent\hrulefill

28: \end{abstract}

29: