摘要
针对含有较大奇异值的矩阵秩最小化问题,采用对数行列式函数代替核范数作为秩函数的非凸近似,应用增广拉格朗日交替方向法求解矩阵秩最小化问题。当罚参数β>1时,证明此算法产生的迭代序列收敛到原问题的稳定点。最后利用实际数据和随机数据,通过数值实验验证所提出的算法较现有的求解核范数矩阵秩最小化问题的算法更高效。
To solve the matrix rank minimization problem with large singular values, the log-determinant function was used as a rank approximation instead of the nuclear norm and an augmented Lagrangian alternating direction method was proposed. When penalty parameter β〉1 , the sequence of iterations generated by the proposed algorithm was proved to be convergent to a stationary point of the original problem. Finally, numerical experiments were conducted based on real data and random data. The results demonstrate that the proposed algorithm is more efficient than the existing nuclear norm method in solving the problem of matrix rank minimization.
出处
《山东科技大学学报(自然科学版)》
CAS
2016年第4期106-113,共8页
Journal of Shandong University of Science and Technology(Natural Science)
基金
国家自然科学基金项目(11241005)
关键词
对数行列式函数
核范数
增广拉格朗日交替方向法
低秩矩阵表示
log-determinant function
nuclear norm
augmented Lagrangian alternating direction method
low rank representation
作者简介
陈勇勇(1989-),男,山东泰安人,硕士研究生,研究方向为系统优化及应用、低秩矩阵恢复、分布式优化算法等.
王永丽(1977-)女,山东栖霞人,副教授,博士,研究方向为非线性优化理论与算法、分布式优化算法及数据挖掘等。本文通信作者.E-mail:wangyongli@sdkd.net.cn