Retraction-Based Gradient Projection Algorithms on Manifolds
En palabras de los autores
We introduce a framework for retraction-based convex optimization on Riemannian manifolds, which includes a notion of retraction-specific convex sets and retraction-based gradient projection algorithms. The standard theory of gradient projection algorithms generalizes easily to this framework. Within this framework, we establish convergence results for retraction-based gradient projection algorithms with various stepsize rules. As an application, we use our framework to study the weighted low-rank approximation. We also provide numerical validation of our convergence results on the image completion task.
Resultado principalEl resumen no menciona limitaciones.
Apareció: lunes, 28 de septiembre. arXiv. Preprint, todavía sin revisión por pares.