结构线性方程组的迭代方法与扰动分析

【作者】 向华；

【导师】 曹志浩；

【作者基本信息】 复旦大学 ， 计算数学， 2006， 博士

【摘要】 论文主要分为两部分，讨论结构化线性方程组的迭代方法和扰动分析。第一，二章是关于迭代方法的。第一章讨论预条件技术，针对对流扩散问题和Oseen问题离散后系数矩阵所具有的特殊结构，用近似Kronecker积构造预条件子。从而改善系数矩阵的谱性质，加速迭代方法的收敛。第二章讨论非精确的Krylov子空间方法。当外迭代用Krylov子空间方法，内迭代可以用松弛策略，非精确地求解。重点分析了非精确的BiCGStab方法，并提出了相应的松弛策略。讨论了Schur补方程，相关方程用非精确Krylov子空间方法求解时的收敛行为，还提出了与Monte Carlo方法结合的思想。第三至第五章是关于扰动分析的。第三章讨论鞍点问题的结构化向后误差和条件数，给出了鞍点问题结构化向后误差的一般表达式，并用结构化条件数分析了解的敏感性。第四章用矩阵导数作为工具推导Cauchy矩阵，Vandermonde矩阵等结构化矩阵的混合型和分量型条件数。在第五章我们考察了带Kronecker积的线性系统，得到了与经典结果类似的条件数，并讨论了其二层条件数。第六章给出了关于子空间距离和奇异值极大极小性质的一个注记。更多还原

【Abstract】 We discuss the iterative methods and perturbation analysis of structured linear systems in this thesis.Chapter 1 and Chapter 2 mainly concern about the iterative solution methods. In Chapter 1, we discuss the preconditioning technique. According to the special structure of the coefficient matrix arising from the SUPG discretization of convection-diffusion problem, or the MAC discretization of the Oseen problem, we use Kronecker product approximation to design the preconditioner. So we can change the spectral properties of the coefficient matrix, and improve the convergence. We focus on the inexact Krylov method in Chapter 2. When we use Krylov subspace methods as the outer iteration, we can apply relaxation strategy to inner iteration and use inexact matrix-vector product. We analyze the inexact BiCGStab and provide its corresponding relaxation strategy. We then apply the inexact Krylov subspace method to the Schur complement equation and the related equation. We also propose a new idea of combining relaxation strategy with Monte Carlo method.We turn to perturbation analysis from Chapter 3 to Chapter 5. In Chapter 3, we discuss the structured backward error and condition numbers of saddle point problem. The explicit general expression of structured backward error is obtained, and the structured condition number is applied to analyze the sensitivity of the solution. In Chapter 4, we use matrix derivative to deduce the mixed and componentwise condition numbers of structured matrix, such as Cauchy matrix, Vandermonde matrix, etc. In Chapter 5, we investigate the linear systems involving Kronecker product. We analyze its condition numbers and the level-2 condition numbers.In Chapter 6, we give a note on the minimax representation for the subspace distance and singular values.更多还原