Iterative methods for projection onto the lp quasi-norm ball

Qi An Nana Zhang Shan Jiang a Department of Computer Science,The Open University of China,Beijing,People's Republic of Chinab College of Economics and Management,Zhejiang University of Water Resources and Electric Power,Hangzhou,People's Republic of Chinac School of Management,Xiamen University,Xiamen,People's Republic of ChinaQi An is currently a Lecturer with the Department of Computer Science,The Open University of China. She received the Ph.D. degree from the Edward P. Fitts Department of Industrial and Systems Engineering,North Carolina State University,USA. Her research interests include nonconvex optimization methods and data mining.Nana Zhang is currently a Lecturer with College of Economics & Management,Zhejiang University of Water Resources and Electric Power. She received the Ph.D. degree with the School of Economics and Management in the Dalian University of Technology. Her research interests include machine learning and supply chain management.Shan Jiang is currently an Associate Professor with School of Management,Xiamen University. She received the Ph.D. degree with the Edward P. Fitts Department of Industrial and Systems Engineering,North Carolina State University,USA. Her research interests include machine learning and optimization.

DOI: https://doi.org/10.1080/10556788.2024.2424539

2024-12-11

Optimization Methods and Software

Abstract:In this paper, we focus on computing the projection onto the lp quasi-norm ball. Motivated by the existing IRBP algorithm, which tackles lp quasi-norm ball projection through projections onto a sequence of simple and tractable weighted l1 norm balls, we propose an innovative and effective iteratively reweighting approach for solving the projection problem. Specifically, our method centres on an ε -approximation that yields better approximation accuracy. We propose an innovative non-smooth, yet continuously differentiable approximation to the lp quasi-norm function. Leveraging the concavity of our approximate model, we developed a novel variant of the iterative reweighted l1 norm ball projection algorithm. Through rigorous analysis, we demonstrated the global convergence properties of our proposed numerical approach. Numerical studies demonstrates the superior computational efficiency of our proposed algorithm. Moreover, our solution method can be further leveraged within current iterative algorithms designed to solve lp quasi-norm ball constrained optimization problems, especially in scenarios where rapid convergence is of utmost importance.

operations research & management science,mathematics, applied,computer science, software engineering