1: \begin{abstract}

2: Sparse reconstruction approaches using the re-weighted $\ell_1$-penalty

3: have been shown, both empirically and theoretically, to provide a significant 

4: improvement in recovering sparse signals in comparison to the  

5: $\ell_1$-relaxation. However, numerical optimization of such penalties involves 

6: solving problems with $\ell_1$-norms in the objective many times. 

7: Using the direct link of reweighted $\ell_1$-penalties to the concave 

8: log-regularizer for sparsity, we derive a simple prox-like algorithm for 

9: the log-regularized formulation.

10: The proximal splitting step of the algorithm has a closed form solution, and we call

11: the algorithm {\em log-thresholding} in analogy to soft thresholding for the  

12: $\ell_1$-penalty. We establish convergence results, and demonstrate 

13: that log-thresholding provides more accurate sparse reconstructions compared to 

14: both soft and hard thresholding. Furthermore, the approach can be 

15: directly extended to optimization over matrices with penalty for rank (i.e. 

16: the nuclear norm penalty and its re-weigthed version), where we suggest a 

17: {\em singular-value log-thresholding} approach.

18: \end{abstract}

19: