江苏省应用数学(中国矿业大学)中心系列学术报告
报告题目:Fast Doubly Block Sparse Kaczmarz Methods with Momentum for Sparse Signal Recovery
报告人:张建华 教授 单位:东华理工大学理学院
报告时间:2026年11月25日(周三)晚上19:30-20:30
腾讯会议:966-738-723
https://meeting.tencent.com/dm/LNyO1ZajCVMr
数学学院
报告人及报告内容简介:
张建华,教授,理学博士;现任东华理工大学理学院副院长、江西省数学会常务理事。主要从事数值代数与科学计算、机器(深度)学习中的优化算法研究;主持或参与完成国家自然科学基金项目、江西省自然科学基金项目等14项;现主持国家自然科学基金和参与江西省自然科学基金重点项目各1项。在《Journal of Computational Physics》《Applied Numerical Mathematics 》《Journal of Computational and Applied Mathematics》和《Journal of Computational Mathematics》等国内外学术刊物上发表论文45篇,其中SCI收录论文26篇。
Abstract: In this talk, we propose a fast doubly block sparse Kaczmarz method with momentum (FDBSK-M) for solving sparse-regularized linear systems. At each iteration, the new method uses adaptive row-wise residual information, cyclic column-block updates and momentum-accelerated Bregman iteration to efficiently approximate sparse solutions. For consistent systems, the method converges linearly; in the presence of noise, the iteration error remains bounded by the noise level. Numerical experiments demonstrate that FDBSK-M achieves high recovery rates and superior convergence performance in sparse signal reconstruction, and offers better computational advantagesover classical randomized block sparse Kaczmarz by averaging (RSKA) method, the fast block sparse Kaczmarz (FBSK) method, and the max-distance sketched Bregman projection (MDSBP) method. The numerical results further indicate that combining momentum with cyclic column-block strategies significantly improves convergence speed and stability, providing an efficient and scalable algorithm for large-scale sparse recovery.
