针对传统QR算法在处理某些大矩阵的奇异值分解时可能不收敛的本质原因，提出采用双向收缩、多次分割的解决对策。研究了在一般矩阵数值计算文献中被忽视的、然而对奇异值分解精度有重要影响的细节如从左至右、从下至上的非零元素直线驱逐算法，提出了矩阵分割时子阵首、末行搜索算法，在这些基础上实现了完整的针对大型矩阵奇异值分解的多次分割、双向收缩QR算法。通过实例比较和分析了不分割与多次分割双向收缩QR算法的收敛速度的差异，证实了多次分割双向收缩QR算法具有迭代次数少、迭代过程无停滞、收敛迅速等优点，解决了传统QR算法处理某些大矩阵的SVD时可能不收敛的问题，对任何大矩阵都可实现快速SVD运算。

Aimed at the essential reason of the algorithm when it is used to process the singular non-convergence of the traditional QR （Quadrature Right-triangle） value decomposition （SVD） of some large-scale matrixes, a doubledirection shrink and multi-partition method is proposed. In this method, the line dislodgment algorithms of nonzero element from left to right and from down to up, which greatly influence the accuracy of SVD, are investigated, and a searching algorithm for the first and the last rows of the sub-matrix is put forward to realize the partition of the main matrix. Thus, a multi-partition and double-direction shrink QR algorithm for the SVD of large-scale matrix is implemented. An example is then presented to reveal the difference of convergence speed between the non-partition and the multi-partition QR algorithms. The results indicate that the proposed algorithm realizes a smooth iteration process with less iteration number and high convergence speed, overcomes the non-convergence of the traditional QR algorithm, and realizes the high-speed SVD computation of any large-scale matrix.