The Inexact Cyclic Block Proximal Gradient Method
Date:
Locations:
- International Conference on Continuous Optimization (ICCOPT 2022), Bethlehem, PA
We extend cyclic block proximal gradient methods for composite optimization to permit approximate gradient and proximal-map evaluations. We characterize how the size and decay of these errors affect convergence, including conditions under which the method retains the rate of its exact counterpart. The analysis uses a relationship between approximate proximal points and approximate subgradients to treat different sources of computational error within a common framework. Numerical experiments illustrate the savings available from fixed approximation tolerances and tolerances that decrease as the algorithm progresses.