no. 17 · Wednesday 18 November 2026
An Inexact Proximal Gradient Method for Composite Multiobjective Optimization with Absolute and Relative Stopping Criteria
Abstract
Proximal gradient methods are widely used for composite optimization. In multiobjective optimization, however, the resulting minimax subproblem generally requires an additional inner method to solve, even when the proximal mapping of each nonsmooth component is available in closed form. In addition, most existing convergence results are obtained only under the assumption that the involved subproblems are solved exactly, which is unrealistic and limits the applicability of the proximal gradient methods. To make the proximal gradient methods practical, we propose an implementable inexact proximal gradient method based on proximal computation. Considering an approximate optimality condition and a hybrid stopping criterion including absolute and relative rules for the subproblem, we address the computational challenge by developing an inner inexact projected gradient method that can exploit the proximal mappings of the individual components. The $\mathcal{O}(1/k)$ convergence rate is also established for our method under certain conditions. Numerical experiments illustrate the computational benefits of solving the subproblems inexactly.
- fields
- Optimization and Control