EUVIMEDEUVIMED entdecken
Crossref

Linear Convergence of ISTA and FISTA

Bo-Wen Li, Bin Shi, Ya-Xiang Yuan

Journal of the Operations Research Society of China · 2026-06 · 14 (2) · 345-363

DOI: 10.1007/s40305-024-00561-0 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of the Operations Research Society of China, online erschienen am 01.10.2024. Verfasst von Bo-Wen Li und 2 weiteren Autor:innen. Die Arbeit wurde bisher 4-mal zitiert (Crossref) und nennt 23 Literaturangaben. Gefördert von: Key Project of Frontier Science Research of Chinese Academy of Sciences, National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

In this paper, we revisit the iterative shrinkage–thresholding algorithm (ISTA) for solving linear inverse problems with sparse representation, commonly found in signal and image processing. We conduct numerical experiments and observe that the convergence behavior of ISTA tends to be linear instead of logarithmic, resulting in a flat approximation on the logarithmic-scale ordinate. Through meticulous observations, we find that assuming the smooth part to be strongly convex, rather than just convex, is more reasonable for the least-square model, even when dealing with ill-conditioned image matrices. We also improve the pivotal inequality for composite optimization by considering the smooth part to be strongly convex instead of general convex, which was first introduced in Li et al. (Proximal Subgradient Norm Minimization of ISTA and FISTA, arXiv:2211.01610, 2022). By utilizing this pivotal inequality, we extend linear convergence to composite optimization, both in terms of the objective value and the square of the proximal subgradient norm. To support our findings, we use a simple ill-conditioned matrix that allows for easy computation of the singular values, instead of the original blur matrix. The new numerical experiment demonstrates that the fast iterative shrinkage–thresholding algorithm for strongly convex functions has a faster linear convergence rate compared to ISTA. Furthermore, based on the tighter pivotal inequality, we also generalize the faster linear convergence rate to composite optimization, considering both the objective value and the square of the proximal subgradient norm. We achieve this by leveraging a well-constructed Lyapunov function with slight modifications and the phase-space representation based on the high-resolution ordinary differential equation framework from the implicit-velocity scheme.

Förderung

  • Key Project of Frontier Science Research of Chinese Academy of Sciences
  • National Natural Science Foundation of China

Zitieren

APA

Li, B., Shi, B., & Yuan, Y. (2026). Linear Convergence of ISTA and FISTA. Journal of the Operations Research Society of China, 14(2), 345-363. https://doi.org/10.1007/s40305-024-00561-0

Vancouver

Li B, Shi B, Yuan Y. Linear Convergence of ISTA and FISTA. Journal of the Operations Research Society of China. 2026;14(2):345-363. doi:10.1007/s40305-024-00561-0

BibTeX
@article{Li2026,
  author = {Li, Bo-Wen and Shi, Bin and Yuan, Ya-Xiang},
  title = {Linear Convergence of ISTA and FISTA},
  journal = {Journal of the Operations Research Society of China},
  year = {2026},
  volume = {14},
  number = {2},
  pages = {345--363},
  doi = {10.1007/s40305-024-00561-0}
}
Kostenlos starten →