1: \begin{abstract}

2: In many signal processing applications, one aims to reconstruct a

3: signal that has a simple representation with respect to a certain

4: basis or frame. Fundamental elements of the basis known as ``atoms''

5: allow us to define ``atomic norms'' that can be used to construct

6: convex regularizers for the reconstruction problem.  Efficient

7: algorithms are available to solve the reconstruction problems in

8: certain special cases, but an approach that works well for general

9: atomic norms remains to be found. This paper describes an optimization

10: algorithm called \cogent, which produces solutions with succinct

11: atomic representations for reconstruction problems, generally

12: formulated with atomic-norm constraints. \cogent combines a greedy

13: selection scheme based on the conditional gradient approach with a

14: backward (or ``truncation'') step that exploits the quadratic nature

15: of the objective to reduce the basis size. We establish convergence

16: properties and validate the algorithm via extensive numerical

17: experiments on a suite of signal processing applications. Our

18: algorithm and analysis are also novel in that they allow for

19: \emph{inexact} forward steps. In practice, \cogent significantly

20: outperforms the basic conditional gradient method, and indeed many

21: methods that are tailored to specific applications, when the

22: truncation steps are defined appropriately.  We also introduce several

23: novel applications that are enabled by the atomic-norm framework,

24: including tensor completion, moment problems in signal processing, and

25: graph deconvolution.

26: \end{abstract}

27: