Linear Convergence of ISTA and FISTA
Bo-Wen Li,Bin Shi,Ya-Xiang Yuan
DOI: https://doi.org/10.1007/s40305-024-00561-0
2024-10-02
Journal of the Operations Research Society of China
Abstract: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.
operations research & management science