摘要
§1.引言
设A∈RM×N,定义增广矩阵
(A~)=(O A AT O),(1)
其中上标T表示转置.不失一般性,假设M≥N,设σi,i=1,2,…,N是A的奇异值,ui和ui分别是对应的左右奇异向量,奇异值按从小到大或从大到小的顺序排列,则A的特征值恰好为±σi,i=1,2,…,N和M-N个零,±σi对应的特征向量分别为1/√2(uT i,vT i)T和1/√2(uT i,-vT i)T.
We study some properties of the Ritz pairs of an argumented matrix with respect to a kind of special subspaces. It is proved that the projected eigenproblem can be reduced to a half dimensional singular value problem. As an important application, this equivalence allows one to compute a partial SVD of a large scale matrix and refined shifts for use within an implicitly restarted refined bidiagonal-ization Lanczos algorithm (IRRBL), so that the computational cost and the storage requirement can be saved significantly.
出处
《数值计算与计算机应用》
CSCD
北大核心
2003年第4期257-261,共5页
Journal on Numerical Methods and Computer Applications
基金
国家重点基础研究专项基金(G19990328)
关键词
Ritz对
增广矩阵
标准正交基
特征值
奇异值
subspace, argumented matrix, Ritz pair, refined Ritz vector, eigenvalue, singular value, eigenvector, singular vector, singular triplet