Abstract
A real m × n matrix A and a vector y ∈ ℝm determine the discrete l 1-regularization (DLR) problem
where | · |1 denotes the l 1-norm of a vector and ρ ≥ 0 is a nonnegative parameter. In this paper, we provide a detailed analysis of this problem which include a characterization of all solutions to (0.1), remarks about the geometry of the solution set and an effective iterative algorithm for numerical solution of (0.1). We are specially interested in the behavior of the solution of (0.1) as a function of ρ and in this regard, we prove in general the existence of critical values of ρ between which the l 1-norm of any solution remains constant. These general remarks are significantly refined when A is a strictly totally positive (STP) matrix. The importance of STP matrices is well-established [5, 14]. Under this setting, the relationship between the number of nonzero coordinates of a distinguished solution of the DLR problem is precisely explained as a function of the regularization parameter for a certain class of vectors in ℝm. Throughout our analysis of the DLR problem, we emphasize the importance of the dual maximum problem by demonstrating that any solution of it leads to a solution of the DLR problem, and vice versa.
Similar content being viewed by others
References
Abdelmalek, N.N.: Linear L 1 approximation for a discrete point set and L 1 solutions of overdetermined linear equations. J. Assoc. Comput. Mach. 18, 41–47 (1971)
Bertsekas, D.P.: Nonlinear Programming. Athena Scientific, Belmont, MA (2003)
Cheney, E.W.: Introduction to Approximation Theory. AMS Chelsea Publishing, Providence, RI (1998)
Donoho, D.L.: For most large underdetermined systems of linear equations the minimal l 1-norm solution is also the sparsest solution. Commun. Pure Appl. Math. 59, 797–829 (2006)
Fallat, S.M., Johnson, C.R.: Totally Nonnegative Matrices. Princeton University Press, Princeton, NJ (2011)
Florenzano, M., Le Van, C.: Finite Dimensional Convexity and Optimization. Springer, New York (2001)
Fu, H., Ng, M.K., Nikolova, M., Barlow, J.L.: Efficient minimization methods of mixed l2-l1 and l1-l1 norms for image restoration. SIAM J. Sci. Comput. 27, 1881–1902 (2006)
Gasca, M., Micchelli, C.A.: Total Positivity and its Applications. Kluwer, Norwell, MA (1996)
Karlin, S.: Total Positivity, vol. I. Stanford University Press, Stanford, CA (1968)
Kress, R.: Numerical Analysis. Springer, New York (1998)
Micchelli, C.A., Pinkus, A.: On n-widths in L ∞ . Trans. Am. Math. Soc. 234, 139–174 (1977)
Micchelli, C.A., Pinkus, A.: On n-widths in l ∞ II. Some related extremal problems. Constr. Func. Theory 77, 403–416 (1980)
Micchelli, C.A., Shen, L., Xu, Y., Zeng, X.: Proximity Algorithms for Image Models II: L1/ TV Denoising (preprint)
Pinkus, A.: Totally Positive Matrices. Cambridge University Press, Cambridge, UK (2010)
Pinkus, A.: Sparse representations and approximation theory. J. Approx. Theory 163, 388–412 (2011)
Pinkus, A.: On L 1-Approximation. Cambridge University Press, Cambridge, UK (1989)
Schmidt, D.: Almost Chebyshev properties for L 1-approximation of continuous functions. J. Approx. Theory 61, 335–350 (1990)
Yang, J., Zhang, Y.: Alternating direction algorithms for ℓ1-problems in compressive sensing. SIAM J. Sci. Comput. 33, 250–278 (2011)
Zhang, Y.: On theory of compressive sensing via l 1 minimization: simple derivations and extensions. Technical Report TR08-11. Dept. of Computational and Applied Mathematics, Rice Univ. (2008)
Zhao, T.: l 1-regularization, duality and strictly total positivity. PhD thesis, School of Mathematics and Computational Science, Sun Yat-sen Univ. (in preparation)
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Yuesheng Xu.
C. A. Micchelli was partially supported by the US Air Force Office of Scientific Research under grant FA9550-09-1-0511 and by the US National Science Foundation under grants DMS-0712827 and DMS-1115523.
Rights and permissions
About this article
Cite this article
Micchelli, C.A., Zhao, T. On discrete ℓ1-regularization. Adv Comput Math 38, 441–454 (2013). https://doi.org/10.1007/s10444-011-9247-7
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10444-011-9247-7
Keywords
- Sparsity
- l 1 approximation
- Convex optimization
- Vector alternation
- Vector equioscillation
- Compressed sensing