1: \begin{abstract} 

2: We consider solving the $\ell_1$-regularized least-squares ($\ell_1$-LS) 

3: problem in the context of sparse recovery, for applications such as 

4: compressed sensing.

5: The standard proximal gradient method, also known as iterative 

6: soft-thresholding when applied to this problem, has low computational cost 

7: per iteration but a rather slow convergence rate.

8: Nevertheless, when the solution is sparse, it often exhibits fast linear 

9: convergence in the final stage. 

10: We exploit the local linear convergence using a homotopy continuation strategy,

11: i.e., we solve the $\ell_1$-LS problem for a sequence of decreasing values 

12: of the regularization parameter, and use an approximate solution at the end 

13: of each stage to warm start the next stage.

14: Although similar strategies have been studied in the literature, 

15: there have been no theoretical analysis of their global iteration complexity.

16: This paper shows that under suitable assumptions for sparse recovery, 

17: the proposed homotopy strategy ensures that 

18: all iterates along the homotopy solution path are sparse.

19: Therefore the objective function is effectively strongly convex along the 

20: solution path, and geometric convergence at each stage can be established.

21: As a result, the overall iteration complexity of our method is 

22: $O(\log(1/\epsilon))$ for finding an $\epsilon$-optimal solution, which

23: can be interpreted as global geometric rate of convergence.

24: We also present empirical results to support our theoretical analysis.

25: \end{abstract}