Colloquiums and Conferences
Convergence Analysis of Transformed Gradient Projection Algorithms on Compact Matrix Manifolds
Time: 2024-05-23 16:30:00
Title: Convergence Analysis of Transformed Gradient Projection Algorithms on Compact Matrix Manifolds
Speaker: Assoc. Research Fellow Jianze Li (Invited by Prof. Yannan Chen)
Affiliation: Shenzhen Research Institute of Big Data
Time: May 22, 10:30–11:30
Venue: Lecture Hall, Room 111, West Building, School of Mathematical Sciences
Speaker Biography:
Jianze Li has worked at the Shenzhen Research Institute of Big Data since May 2019 as a research scientist and associate research fellow. Previously he received his Ph.D. from the Chern Institute of Mathematics at Nankai University in June 2013, was a lecturer in the School of Mathematics at Tianjin University from July 2013 to January 2018, conducted academic visits and postdoctoral research at GIPSA-Lab, Université Grenoble Alpes in France from September 2016 to August 2018, and was a visiting scholar at Ryerson University in Canada from September 2018 to April 2019. His main research areas are tensor optimization and manifold optimization. He has published 15 SCI papers in journals including SIMAX, SIOPT, MCOM, JMAA and LAA, and has completed one NSFC Young Scientists Fund project, one Guangdong Provincial General Programme project and one basic research project of the Shenzhen Research Institute of Big Data.
Abstract:
To solve optimization problems on compact matrix manifolds, we introduce a new algorithmic framework based on the projection from Euclidean space onto compact matrix manifolds, called the Transformed Gradient Projection (TGP) algorithm. Compared with existing algorithms, TGP features new search directions and a variety of stepsizes, including Armijo, nonmonotone Armijo and fixed stepsizes, to generate the next iterate. We focus in particular on the Stiefel and Grassmann manifolds and show that many existing algorithms can be viewed as special cases of this framework. We then establish the convergence of TGP, including weak convergence, convergence rate and global convergence. Finally, a series of numerical experiments show that TGP outperforms the classical gradient projection algorithm and retraction-based line-search algorithms in a range of settings.