A non-convex piecewise quadratic approximation of ℓ0 regularization: theory and accelerated algorithm

Non-convex regularization has been recognized as an especially important approach in recent studies to promote sparsity. In this paper, we study the non-convex piecewise quadratic approximation (PQA) regularization for sparse solutions of the linear inverse problem. It is shown that exact recovery o...

Full description

Saved in:
Bibliographic Details
Published in:Journal of global optimization Vol. 86; no. 2; pp. 323 - 353
Main Authors: Li, Qian, Zhang, Wei, Bai, Yanqin, Wang, Guoqiang
Format: Journal Article
Language:English
Published: New York Springer US 01.06.2023
Springer Nature B.V
Subjects:
Algorithms
Approximation
Compressibility
Computer Science
Convergence
Inverse problems
Iterative methods
Mathematics
Mathematics and Statistics
Operations Research/Decision Theory
Optimization
Real Functions
Recovery
Regularization
Sparsity
Exact recovery
Sparsity optimization
65K05
Non-convex approximation
Accelerated gradient algorithm
90C06
Iterative thresholding algorithm
90C26
ISSN:0925-5001, 1573-2916
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Non-convex regularization has been recognized as an especially important approach in recent studies to promote sparsity. In this paper, we study the non-convex piecewise quadratic approximation (PQA) regularization for sparse solutions of the linear inverse problem. It is shown that exact recovery of sparse signals and stable recovery of compressible signals are possible through local optimum of this regularization. After developing a thresholding representation theory for PQA regularization, we propose an iterative PQA thresholding algorithm (PQA algorithm) to solve this problem. The PQA algorithm converges to a local minimizer of the regularization, with an eventually linear convergence rate. Furthermore, we adopt the idea of accelerated gradient method to design the accelerated iterative PQA thresholding algorithm (APQA algorithm), which is also linearly convergent, but with a faster convergence rate. Finally, we carry out a series of numerical experiments to assess the performance of both algorithms for PQA regularization. The results show that PQA regularization outperforms ℓ 1 and ℓ 1 / 2 regularizations in terms of accuracy and sparsity, while the APQA algorithm is demonstrated to be significantly better than the PQA algorithm.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0925-5001
1573-2916
DOI:10.1007/s10898-022-01257-6