1: \begin{abstract}

2: The convergence rate is analyzed for the SpaSRA algorithm

3: (Sparse Reconstruction by Separable Approximation)

4: for minimizing a sum $f (\m{x}) + \psi (\m{x})$ where $f$ is

5: smooth and $\psi$ is convex, but possibly nonsmooth.

6: It is shown that if $f$ is convex, then the error in the

7: objective function at iteration $k$, for $k$ sufficiently large,

8: is bounded by $a/(b+k)$ for suitable choices of $a$ and $b$.

9: Moreover, if the objective function is strongly convex, then the

10: convergence is $R$-linear.

11: An improved version of the algorithm based on a

12: cycle version of the BB iteration and an adaptive line search is given.

13: The performance of the algorithm is investigated using applications

14: in the areas of signal processing and image reconstruction.

15: \end{abstract}

16: