第八章-代数特征值问题
数值分析(30)代数特征值问题(QR方法)
* * *
数值分析
数值分析
二、化一般矩阵为上Hessenberg阵
称形如 h11 h12 h1n 1 h1n h h h h 22 2n 1 2n 21 h32 h33 h3 n H hnn 1 hnn 的矩阵为上海森堡(H es s enberg) 阵。如果此对角 线元hii 1 ( i 2, 3, , n)全不为零 ,则称该矩阵为不可 约的上H es s enberg矩阵。 讨论用 Householder 变换将一般矩阵A相似变 换成H es s enberg阵
21, 0, 0
T
数值分析
数值分析
数值分析
数值分析
数值分析
数值分析
数值分析
数值分析
数值分析
数值分析
数值分析
数值分析
数值分析
数值分析
习题
数值分析
数值分析
第三节 求矩阵全部特征值的QR方法 一、求矩阵全部特征值的QR方法
60年代出现的QR算法是目前计算中小型矩阵的 全部特征值与特征向量的最有效方法。 理论依据:任一非奇异实矩阵都可分解成一个正交 矩阵Q和一个上三角矩阵R的乘积,而且当R的对角元 符号取定时,分解是唯一的。
数值分析
数值分析
uuT H2 I 2 T u u 2 2 1 0 4 2 0 1 2 4 0 1 0 0 1 0 H 2 0 0 2 2 2 0 0 2
2 2 4 2 2 2 2 4 2 0 0 2 2 2 2
数值分析
数值分析
第八章__特征值问题
例 2 矩阵
(0,1,0)T , 1、 11 、 A(ε) 的
数学系 李继根( jgli@ )
二、盖尔(Gerschgorin)定理
把矩阵
A 分裂成 A = diag(a11,L , ann ) + B ? D B
A 的扰动矩阵 A(ε) ? D εB ,显然 A(0) = D, A(1) = A 我们有理由猜测,如果 ε 足够小, A(ε) 的特征 L 、a) 值将位于 A(0) 的特征值(即元素 a11、 nn
由此,我们可以将 Gerschgorin定理看成定理 8 的“推论” 。
数学系 李继根( jgli@ )
事实上,设矩阵 A 的特征值 Gerschgorin区域 ,则有
ℓ 不属于 A 的
| ℓ ai i | ⊳ Ri , i ≪1, 2,⋯ , n 因此矩阵 ℓ I A 严格对角占优,根据定理 8 det(ℓ I A ) 0 这与 ℓ 是 A 的特征值相矛盾。
遗憾的是矩阵的特征向量一般不是矩阵元素的 连续函数,因此不一定是稳定的。
1 0 0 A(ε) = 0 1+ ε 1 ε 0 1+ ε 1+ ε、 1,特征向量为 的特征值为 1+ ε、 T (1,1/ ε , 1) - 。而 A(0) 的特征值为 和 特征向量为 (0,1,0)T 和 (1,0,0)T 。矩阵 特征向量在 ε = 0 处不连续。
1
数学系 李继根( jgli@ )
根据定理8,严格对角占优矩阵 征值,而
A 没有零特
| 0 ai i | ≪| ai i | ⊳ Ri , i ≪1, 2,⋯ , n
这说明矩阵
A 的特征值 ℓ 可能满足 | ℓ ai i | ℂ Ri , i ≪1, 2,⋯ , n
̃ ≪ A E , E ≪(⊙ ) 的解。这里 A i j
代数特征值问题(课堂PPT)
qTkAkqk
qT0A2k1q0 qT0A2kq0
a2j
j1
n
a2j
2k1 j
,
2k j
j1
n
a
2 2k
jj
( j
1)
k 1 j2 n
,
a2j 2jk
j 1
Note:
sikn dissp { tq k a }sn ,p { x 1 } a n O 1 2k
例1.
2 3 13
A(
代数特征值问题
向华
武汉大学数学与统计学院
1
搜索引擎
GTxx
xTe 1
G: Google Matrix,
“the world’s largest matrix computation”. 4,300,000,000
x: PageRank vector
“The $25,000,000,000 Eigenvector”
Akq0 a11kx1 a2k2x2 anknxn
a11k
x1
n j2
aj a1
j 1
k xj
a11k x1
思考:如果恰好在x1分量上a1=0?
9
迭代格式
q kAk 1 q ,k 1 ,2 ,
qk a11 kx1(k ) 可视为关于特征值 1 的近似特征向量
当|λ1 <1或|λ1|>1,产生下溢或上溢.作规格化:
)
5 9 4
11 7 14
10 6 15
8 112
(1)比较α=30和α=-30时的迭代次数,
注意两种情景下 |λ2 /λ1|的大小.
11
(2)取α=16,此时 1 3 ,x 4 1 ( 1 /2 ,1 /2 ,1 /2 ,1 /2 ) T
特征值与特征向量(高等代数课件)
0
x0n
即
x x
01 0n
是线性方程组 (0E A )X0的解,
又
0,
x01 0, ∴
x0n
(0EA )X0有非零解.
所以它的系数行列式 0EA0.
7.4 特征值与特征向量
以上分析说明:
若 0 是 的特征值,则 0EA0.
反之,若 0 P 满足 0EA0,
则齐次线性方程组 (0E A )X0有非零解.
7.4 特征值与特征向量
练习2:已知3阶方阵A的特征值为:1、-1、2,
则矩阵 BA 32A 2的特征值为: 1,3,0 ,
行列式 B = 0 .
7.4 特征值与特征向量
谢谢!
( B n 1 B n 2 A ) B n 1 A ②
比较①、②两式,得
7.4 特征值与特征向量
B0 E
B1 B0A a1E
B2 B1A a2E
③
B
n
1
Bn B n
1
2A a A an
n
E
1
E
以A n,A n1, ,A ,E依次右乘③的第一式、第二式、
…、第n式、第n+1式,得
由此知,特征向量不是被特征值所唯一确定的, 但是特征值却是被特征向量所唯一确定的,即
若 ( ) 且 ( ) ,则 .
7.4 特征值与特征向量
二、特征值与特征向量的求法
分析: 设 d im V n , 1 ,2 , ,n是V的一组基,
线性变换 在这组基下的矩阵为A.
设 0 是 的特征值,它的一个特征向量 在基
B B2
1A An
n
2
B0An 1 B0An B1An
An
Chapter 8特征值问题的计算方法
ɶ ɶ ɶ ɶ Qm+1Rm+1 = AQmRm
ɶ R = A2Q R = ⋯= AmQ R = Am+1 ɶ ɶ Qm+1 ɶm+1 m−1 m−1 1 1
即A
m
ɶ ɶ = QmRm
QR方法的收敛性 方法的收敛 收敛性
方法的收敛性质) 收敛性质 ( Th8.4.1 QR方法的收敛性质) 设 A 的特征值满足 λ1 > λ 2 > ⋯ > λ n > 0 ,且 Y 的 左特征向量 向量; 分解, 第i行是 A 对应于 λi 的左特征向量;若 Y 有 LU分解, 行是 则由QR迭代算法产生的矩阵 Am = [a 下的元素趋于0, 下的元素趋于 ,同时对角元素 a
∗ 其中A2 = α 2 ∗ ɶ A3
1 0 ɶ 令 H2 = ˆ 0 H 2
∗ 1 0 ∗ ∗ 1 0 ∗ ɶ ɶ H2 A2 H2 = = ˆ ɶ 0 H ˆ AH ˆ ˆ H2α2 H2 ɶ3 ˆ 2 0 H2 α2 A3 2
第八章 特征值问题的计算方法
/*Computational Method of Eigenvalue Problem*/ 本章主要介绍矩阵的特征值和特征向量的计算方法。 本章主要介绍矩阵的特征值和特征向量的计算方法。 特征值 的计算方法
§1 基本概念与性质
特征值和特征向量的基本概念与性质 特征值和特征向量的
1 0 令 H2 = ɶ 0 H 2
∗ ∗ ɶ ɶ H2 A2 H2 = ˆ ˆ ɶ ˆ H2α2 H2 A H2 3
det(λ I − A) = (λ − λ1 ) (λ − λ2 ) ...(λ − λ p ) 其中 n1 + n2 + ... + np = n; λi ≠ λ j (i ≠ j )
特征值与特征向量定义与计算
特征值与特征向量特征值与特征向量的概念及其计算定义1. 设A是数域P上的一个n阶矩阵,λ是一个未知量,称为A的特征多项式,记ƒ(λ)=| λE-A|,是一个P上的关于λ的n次多项式,E是单位矩阵。
ƒ(λ)=| λE-A|=λn+α1λn-1+…+αn= 0是一个n次代数方程,称为A 的特征方程。
特征方程ƒ(λ)=| λE-A|=0的根 (如:λ0) 称为A的特征根(或特征值)。
n次代数方程在复数域内有且仅有n 个根,而在实数域内不一定有根,因此特征根的多少和有无,不仅与A有关,与数域P也有关。
以A的特征值λ0代入 (λE-A)X=θ,得方程组 (λ0E-A)X=θ,是一个齐次方程组,称为A的关于λ0的特征方程组。
因为 |λ0E-A|=0,(λ0E-A)X=θ必存在非零解X(0),X(0) 称为A的属于λ0的特征向量。
所有λ0的特征向量全体构成了λ0的特征向量空间。
一.特征值与特征向量的求法对于矩阵A,由AX=λ0X,λ0EX=AX,得:[λ0E-A]X=θ即齐次线性方程组有非零解的充分必要条件是:即说明特征根是特征多项式 |λ0E-A| =0的根,由代数基本定理有n个复根λ1, λ2,…, λn,为A的n个特征根。
当特征根λi(I=1,2,…,n)求出后,(λi E-A)X=θ是齐次方程,λi 均会使 |λi E-A|=0,(λi E-A)X=θ必存在非零解,且有无穷个解向量,(λi E-A)X=θ的基础解系以及基础解系的线性组合都是A的特征向量。
例1. 求矩阵的特征值与特征向量。
解:由特征方程解得A有2重特征值λ1=λ2=-2,有单特征值λ3=4对于特征值λ1=λ2=-2,解方程组 (-2E-A)x=θ得同解方程组 x1-x2+x3=0解为x1=x2-x3 (x2,x3为自由未知量)分别令自由未知量得基础解系所以A的对应于特征值λ1=λ2=-2的全部特征向量为x=k1ξ1+k2ξ2 (k1,k2不全为零)可见,特征值λ=-2的特征向量空间是二维的。
The algebraic eigenvalue problem代数特征值问题
CLARENDON PRESS • OXFORD 1965
Contents
1. THEORETICAL BACKGROUND Page
Introduction Definitions Eigenvalues and eigenvectors of the transposed matrix Distinct eigenvalues Similarity transformations Multiple eigenvalues and canonical forms for general matrices Defective system of eigenvectors The Jordan (classical) canonical form The elementary divisors Companion matrix of the characteristic polynomial of A Non-derogatory matrices The Frobenius (rational) canonical form Relationship between the Jordan and Frobenius canonical forms Equivalence transformations Lambda matrices Elementary operations Smith's canonical form The highest common factor offc-rowedminors of a A-matrix Invariant factors of (A —XI) The triangular canonical form Hermitian and symmetric matrices Elementary properties of Hermitian matrices Complex symmetric matrices Reduction to triangular form by unitary transformations Quadratic forms Necessary and sufficient conditions for positive definiteness Differential equations with constant coefficients Solutions corresponding to non-linear elementary divisors Differential equations of higher order Second-order equations of special form Explicit solution of By = —Ay Equations of the form (AB— XI)x — 0 The minimum polynomial of a vector The minimum polynomial of a matrix Cayley-Hamilton theorem Relation between minimum polynomial and canonical forms Principal vectors Elementary similarity transformations Properties of elementary matrices Reduction to triangular canonical form by elementary similarity transformations Elementary unitary transformations Elementary unitary Hermitian matrices Reduction to triangular form by elementary unitary transformations Normal matrices Commuting matrices
8、矩阵特征值问题计算
对应的特征向量x1, x2 ,, xm线性无关.
定理7(对称矩阵的正交约化 ) 设A R nn为对称矩阵 , 则
(1) A的特征值均为实数; (2) A有n个线性无关的特征向量; (3) 存在正交矩阵P使得
1 2 , P 1 AP n 且i (i 1,2,, n)为A的特征值, 而P (u1,u2 , ,un )的列 向量u j为对应于 j的特征向量.
k
k
k A v0 max(vk ) max max(Ak 1v ) 0 k k 2 n 1 maxa1 x1 a2 x2 an xn 1 1 k 1 k 1 2 n maxa1 x1 a2 x2 an xn 1 1 1 (k )
k k 1
lim
vk
a1 x1.
即vk 是1的近似的特征向量. 而主特征值 (vk 1 ) j 1 n (vk 1 ) j 1 , 或1 . (v k ) j n j 1 (v k ) j
定理12 设A R nn有n个线性无关的特征向量, 其特征值
1 2 n ,
并设A的主特征值是实根,且满足
1 2 n ,
现在讨论求1及x1的基本方法.
(2.1)
v0 a1 x1 a2 x2 an xn , (设a1 0)
v1 Av0 a11 x1 a22 x2 ann xn ,
k k 2 n k vk Avk 1 1 a1 x1 a2 x2 an xn . 1 1 k 当k很大时,k 1 a1 x1, vk 1 1vk , Avk 1vk, v
Algebraic eigenvalue problems
6.0. Introduction 113 Chapter 6Algebraic eigenvalue problemsDas also war des Pudels Kern! G OETHE.6.0. IntroductionDetermination of eigenvalues and eigenvectors of matrices is one of the most important problems of numerical analysis. Theoretically, the problem has been reduced to finding the roots of an algebraic equation and to solving linear homogeneous systems of equations. In practical computation, as a rule, this method is unsuitable, and better methods must be applied.When there is a choice between different methods, the following questions should be answered:(a)Are both eigenvalues and eigenvectors asked for, or are eigenvalues alone sufficient?(b)Are only the absolutely largest eigenvalue(s) of interest?(c)Does the matrix have special properties (real symmetric, Hermitian, and so on)?If the eigenvectors are not needed less memory space is necessary, and further, if only the largest eigenvalue is wanted, a particularly simple technique can be used. Except for a few special cases a direct method for computation of the eigenvalues from the equation is never used. Further it turns out that practically all methods depend on transforming the initial matrix one way or other without affecting the eigenvalues. The table on p. 114 presents a survey of the most important methods giving initial matrix, type of transformation, and transformation matrix. As a rule, the transformation matrix is built up successively, but the resulting matrix need not have any simple properties, and if so, this is indicated by a horizontal line. It is obvious that such a compact table can give only a superficial picture; moreover, in some cases the computation is performed in two steps. Thus a complex matrix can be transformed to a normal matrix following Eberlein, while a normal matrix can be diagonalized following Goldstine-Horwitz. Incidentally, both these procedures can be performed simultaneously giving a unified method as a result. Further, in some cases we have recursive techniques which differ somewhat in principle from the other methods.It is not possible to give here a complete description of all these methods because of the great number of special cases which often give rise to difficulties. However, methods which are important in principle will be treated carefully114 Algebraic eigenvalue problems6.1. The power method 115 and in other cases at least the main features will be discussed. On the whole we can distinguish four principal groups with respect to the kind of transformation used initially:1.Diagonalization,2.Almost diagonalization (tridiagonalization),3.Triangularization,4.Almost triangularization (reduction to Hessenberg form).The determination of the eigenvectors is trivial in the first case and almost trivialin the third case. In the other two cases a recursive technique is easily established which will work without difficulties in nondegenerate cases. To a certain amount we shall discuss the determination of eigenvectors, for example, Wilkinson's technique which tries to avoid a dangerous error accumulation. Also Wielandt's method, aiming at an improved determination of approximate eigenvectors, will be treated.6.1. The power methodWe assume that the eigenvalues of are where Now we let operate repeatedly on a vector which we express as a linear combination of the eigenvectors(6.1.1) Then we haveand through iteration we obtain(6.1.2). For large values of the vectorwill converge toward that is, the eigenvector of The eigenvalue is obtained as(6.1.3) where the index signifies the component in the corresponding vector. The rate of convergence is determined by the quotient convergence is faster the116 Algebraic eigenvalue problemssmaller is. For numerical purposes the algorithm just described can be formulated in the following way. Given a vector we form two other vectors, and(6.1.4)The initial vector should be chosen in a convenient way, often one tries vector with all components equal to 1.E XAMPLEStarting fromwe find thatandAfter round-off, we getIf the matrix is Hermitian and all eigenvalues are different, the eigenvectors, as shown before, are orthogonal. Let be the vector obtained after iterations:We suppose that all are normalized:6.1. The power method 117 Then we haveandFurther,When increases, all tend to zero,and with, we get Rayleigh's quotient(6.1.5) ExampleWith andwe obtain for and 3,,and respectively, compared with the correct value The corresponding eigenvector isThe quotients of the individual vector components give much slower convergence; for example,The power method can easily be modified in such a way that certain other eigenvalues can also be computed. If, for example,has an eigenvalue then has an eigenvalue Using this principle, we can produce the two outermost eigenvalues. Further, we know that is an eigenvalue of and analogously that is an eigenvalue of If we know that an eigenvalue is close to we can concentrate on that, since becomes large as soon as is close toWe will now discuss how the absolutely next largest eigenvalue can be calculated if we know the largest eigenvalue and the corresponding eigenvector Let be the first row vector of and form(6.1.6)Here is supposed to be normalized in such a way that the first component is Hence the first row of is zero. Now let and be an eigenvalue and the corresponding eigenvector with the first component of equal to Then118 Algebraic eigenvalue problems we havesince and(note that the first component of as well as of is 1).Thus is an eigenvalue and is an eigenvector of Since has the first component equal to 0, the first column of is irrelevant, and in fact we need consider only the-matrix, which is obtained when the first row and first column of are removed. We determine an eigenvector of this matrix, and by adding a zero as first component, we get a vector Then we obtain from the relationMultiplying with we find and hence When and have been determined, the process, which is called deflation, can be repeated.E XAMPLEThe matrixhas an eigenvalue and the corresponding eigenvectoror normalized,Without difficulty we findNow we need consider onlyand we find the eigenvalues which are also eigenvalues of6.1. The power method 119 the original matrix The two-dimensional eigenvector belonging to isand henceSince we get andWith we findand Hence andand all eigenvalues and eigenvectors are known.If is Hermitian, we have when Now suppose thatand form(6.1.7) It is easily understood that the matrix has the same eigenvalues and eigenvectors as exceptwhich has been replaced by zero. In fact, we haveand and so on. Then we can again use the power method on the matrix120 Algebraic eigenvalue problems With the starting vectorwe find the following values for Rayleigh's quotient:and compared with the correct valueIf the numerically largest eigenvalue of a real matrix is complex,then must also be an eigenvalue. It is also clear that if is the eigenvector belonging to then is the eigenvector belonging toNow suppose that we use the power method with a real starting vectorThen we form with so large that the contributions from all the other eigenvectors can be neglected. Further, a certain component of is denoted by Thenwhere and the initial component of corresponding to is Hencewhere we have put Now we formHence(6.1.8) Then we easily findIn particular, if that is, if the numerically largest eigenvalues are ofthe form with real then we have the simpler formula(6.1.10)6.2. Jacobi's methodIn many applications we meet the problem of diagonalizing real, symmetric matrices. This problem is particularly important in quantum mechanics.In Chapter 3 we proved that for a real symmetric matrix all eigenvalues are real, and that there exists a real orthogonal matrix such that is diagonal. We shall now try to produce the desired orthogonal matrix as a— product of very special orthogonal matrices. Among the off-diagonal elements6.2. Jacobi's method 121 we choose the numerically largest element:The elementsand form a submatrix which can easily be transformed to diagonal form. We putand get(6.2.1) Now choose the angle such that that is, tan This equationgives 4 different values of and in order to get as small rotations as possible we claimPuttingandwe obtain:since the angle must belong to the first quadrant if tan and to the fourth quadrant if tan Hence we have for the anglewhere the value of the arctan-function is chosen between After a few simple calculations we get finally:(6.2.2)(Note that andWe perform a series of such two-dimensional rotations; the transformation matrices have the form given above in the elements and and are identical with the unit matrix elsewhere. Each time we choose such values and that We shall show that with the notation the matrix for increasing will approach a diagonal122 Algebraic eigenvalue problems matrix with the eigenvalues of along the main diagonal. Then it is obvious that we get the eigenvectors as the corresponding columns of since we have that is, Let be the column vector of and the diagonal element of Then we haveIf is denoted by we know from Gershgorin's theorem that for some value of and if the process has been brought sufficiently far, every circle defined in this way contains exactly one eigenvalue. Thus it is easy to see when sufficient accuracy has been attained and the procedure can be discontinued.The convergence of the method has been examined by von Neumann and Goldstine in the following way. We put and, as before,The orthogonal transformation affects only the row and column and the row and column. Taking only off-diagonal elements into account, we find for and relations of the formand hence Thus will be changed only through the cancellation of the elements and that is,Since was the absolutely largest of all off-diagonal elements, we haveandHence we get the final estimate,(6.2.3)After iterations,has decreased with at least the factor and for a sufficiently large we come arbitrarily close to the diagonal matrix containing the eigenvalues.In a slightly different modification, we go through the matrix row by row performing a rotation as soon as Here is a prescribed tolerance which, of course, has to be changed each time the whole matrix has been passed. This modification seems to be more powerful than the preceding one. The method was first suggested by Jacobi. It has proved very efficient for diagonalization of real symmetric matrices on automatic computers.6.2. Jacobi's method 123 ExampleChoosing we obtain, tan andAfter the first rotation, we haveHere we take and obtain tan andAfter the second rotation we haveand after 10 rotations we haveAfter rotations the diagonal elements are and while the remaining elements are equal to to decimals accuracy. The sum of the diagonal elements is and the product in good agreement with the exact characteristic equation:Generalization to Hermitian matrices, which are very important in modern physics, is quite natural. As has been proved before, to a given Hermitian matrix we can find a unitary matrix such that becomes a diagonal matrix. Apart from trivial factors, a two-dimensional unitary matrix has the formA two-dimensional Hermitian matrix124 Algebraic eigenvalue problems is transformed to diagonal form by wherePutting we separate the real and imaginary parts and then multiply the resulting equations, first by and then by and and finally add them together. Using well-known trigonometric formulas, we get(6.2.4) In principle we obtain from the first equation and then can be solved from the second. Rather arbitrarily we demand and hencewhereSince the remaining equation has the solutionwith and Now we want to choose according to in order to get as small a rotation as possible which impliesThe following explicit solution is now obtained (note that and cannot both be equal to because then would already be diagonal):(6.2.5) As usual the value of the arctan-function must be chosen between and6.3. Givens' method 125The element can now be writtenand consequently:(6.2.6) If we get and recover the result in Jacobi's method.This procedure can be used repeatedly on larger Hermitian matrices, where the unitary matrices differ from the unit matrix only in four places. In the places and we introduce the elements of our two-dimensional matrix. The product of the special matrices is a new unitary matrix approaching when is increased.Finally we mention that a normal matrix (defined through can always be diagonalized with a unitary matrix. The process can be performed following a technique suggested by Goldstine and Horwitz which is similar to the method just described for Hermitian matrices. The reduction of an arbitrary complex matrix to normal form can be accomplished through a method given by Patricia Eberlein In practice, both these processes are performed simultaneously.6.3. Givens' methodAgain we assume that the matrix is real and symmetric. In Givens' method we can distinguish among three different phases. The first phase is concerned with orthogonal transformations, giving as result a band matrix with unchanged characteristic equation. In the second phase a sequence of, functions is generated, and it is shown that it forms a Sturm sequence, the last member of which is the characteristic polynomial. With the aid of the sign changes in this sequence, we can directly state how many roots larger than the inserted value the characteristic equation has. By testing for a number of suitable values we can obtain all the roots. During the third phase, the eigenvectors are computed. The orthogonal transformations are performed in the following order. The elements and define a two-dimensional subspace, and we start by performing a rotation in this subspace. This rotation affects all elements in the second and third rows and in the second and third columns. However, the quantity defining the orthogonal matrixis now determined from the condition and not, as in Jacobi's method, by We have and The next rotation is performed in the (2, 4)-plane with the new126 Algebraic eigenvalue problemsdetermined from that is, tan that the element was changed during the preceding Now all elements in the second and fourth rows and in the second and fourth columns are changed, and it should be particularly observed that the element is not affected. In the same way, we make the elements equal to zero by rotations in the-planes.Now we pass to the elements and they are all set to zero by rotations in the planesDuring the first of these rotations, the elements in the third and fourth rows and in the third and fourth columns are changed, and we must examine what happens to the elements and which were made equal to zero earlier. We findFurther, we get and By now the procedure should be clear, and it is easily understood that we finally obtain a band matrix, that is, such a matrix that In this special case we have Now we put(6.3.1)has been obtained from by a series of orthogonal transformations,with In Chapter it was proved that and have the same eigenvalues and further that, if is an eigenvector of and an eigenvector of(both with the same eigenvalue), then we have Thus the problem has been reduced to the computation of eigenvalues and eigenvectors of the band matrixWe can suppose that all otherwise could be split into two determinants of lower order Now we form the following sequence of functions:(6.3.2)with and We find at once that which can be interpreted as the determinant of the-element in the matrix6.3. Givens' method 127 Analogously, we have which is the-minor ofBy induction, it is an easy matter to prove that is the characteristic polynomial.Next we shall examine the roots of the equation For we have the only root.For we observe that, Hence we have two real roots and with, for example,For we will use a method which can easily be generalized to an induction proof. Then we write and obtain from (6.3.2):Now it suffices to examine the sign of in a few suitable points:We see at once that the equation has three real roots and such thatIn general, if has the roots and the roots thenwhereBy successively putting and we find that has different signs in two arbitrary consecutive points. Hence has real roots, separated by the roots ofWe are now going to study the number of sign changes in the sequenceIt is evident that and Suppose that and are two such real numbers that in the closed interval Then obviously First we examine what happens if the equation has a root in the interval. From it follows for thatHence and have different signs, and clearly this is also true in an interval Suppose, for example, that then we may have the following combination of signs:Hence, the number of sign changes does not change when we pass through a root of When however, the situation is different.128 Algebraic eigenvalue problemsSuppose, for example, that& odd. Denoting the roots of by and the roots of by we haveThen we see that Now we let increase until it reaches the neighborhood of |where we find the following scheme:Hence Then we let increase again (now a sign change of may appear, but, as shown before, this does not affect until we reach the neighborhood of where we haveand hence Proceeding in the same way through all the rootswe infer that the number of sign changes decreases by one unit each time a root is passed. Hence we have proved that if is the number of eigenvalues of the matrix which are larger than then(6.3.3) The sequence is called a The described technique makes it possible to compute all eigenvalues in a given interval ("telescope method").For the third phase, computation of the eigenvectors, we shall follow J. H. Wilkinson in Let be an exact eigenvalue of Thus we search for a vector such that Since this is a homogeneous system in variables, and since we can obtain a nontrivial solution by choosing equations and determine the components of(apart from a constant factor); the remaining equation must then be automatically satisfied. In practical work it turns out, even for quite well-behaved matrices, that the result to a large extent depends on which equation was excluded from the Essentially, we can say that the serious errors which appear on an unsuitable choice of equation to be excluded depend on numerical compensations; thus round-off errors achieve a dominant influence.Let us assume that the equation is excluded, while the others are solved by elimination. The solution (supposed to be exact) satisfies the equations used for elimination but gives an error when inserted into the6.3. Givens' method 129Actually, we have solved the system(We had to use an approximation instead of the exact eigenvalue.) Since constant factors may be omitted, this system can be written in a simpler way:(6.3.5)where is a column vector with the component equal to and the others equal to If the eigenvectors of are this vector can be expressed as a linear combination, that is,(6.3.6) and from (6.3.5) we get(6.3.7) Now let and we obtain(6.3.8)Under the assumption that our solution approaches as(apart from trivial factors). However, it may well happen that is of the same order of magnitude as(that is, the vector is almost orthogonal to), and under such circumstances it is clear that the vector in (6.3.8) cannot be a good approximation of Wilkinson suggests that (6.3.5) be replaced by(6.3.9)where we have the vector at our disposal. This system is solved by Gaussian elimination, where it should be observed that the equations are permutated properly to make the pivot element as large as possible. The resulting system is written:(6.3.10)As a rule, most of the coefficients are zero. Since the have been obtained from the which we had at our disposal, we could as well choose the constants deliberately. It seems to be a reasonable choice to take all130 Algebraic eigenvalue problems equal to no eigenvector should then be disregarded. Thus we choose(6.3.11) The system is solved, as usual, by back-substitution, and last, the vector is normalized. Even on rather pathological matrices, good results have been obtained by Givens' method.6.4. Householder's methodThis method, also, has been designed for real, symmetric matrices. We shall essentially follow the presentation given by Wilkinson The first step consists of reducing the given matrix to a band matrix. This is done by orthogonal transformations representing reflections. The orthogonal matrices, will be denoted by with the general structure(6.4.1) Here is a column vector such that(6.4.2) It is evident that is symmetric. Further, we havethat is,is also orthogonal.The matrix acting as an operator can be given a simple geometric interpretation. Let t operate on a vector from the left:In Fig. 6.4 the line is perpendicular to the unit vector in a plane defined by and The distance from the endpoint of to is and the mapping means a reflection in a plane perpendicular toFigure 6.46.4. Householder's method 131 Those vectors which will be used are constructed with the first components zero, orWith this choice we form Further, by (6.4.2) we haveNow put and form successively(6.4.3)At the first transformation, we get zeros in the positionsand in the corresponding places in the first column. The final result will become a band matrix as in Givens' method. The matrix contains elements in the row, which must be reduced to zero by transformation with this gives equations for theelements and further we have the condition that the sum of the squares must beWe carry through one step in the computation in an example:The transformation must now produce zeros instead of and Obviously, the matrix has the following form:Since in the first row of only the first element is not zero, for example, the -element of can become zero only if the corresponding element is zero already in Puttingwe find that the first row of has the following elements:Now we claim that(6.4.4) Since we are performing an orthogonal transformation, the sum of the squares132 Algebraic eigenvalue problems of the elements in a row is invariant, and hencePutting we obtain(6.4.5) Multiplying (6.4.5) by and (6.4.4) by and we getThe sum of the first three terms is and further Hence(6.4.6) Inserting this into (6.4.5), we find that and from (6.4.4), andIn the general case, two square roots have to be evaluated, one for and one for Since we have in the denominator, we obtain the best accuracy if is large. This is accomplished by choosing a suitable sign for the square-root extraction for Thus the quantities ought to be defined as follows:(6.4.7) The sign for this square root is irrelevant and we choose plus. Hence we obtain for and(6.4.8) The end result is a band matrix whose eigenvalues and eigenvectors are computed exactly as in Givens' method. In order to get an eigenvector of an eigenvector of the band matrix has to be multiplied by the matrix this should be done by iteration:(6.4.9) 6.5. Lanczos' methodThe reduction of real symmetric matrices to tridiagonal form can be accomplished through methods devised by Givens and Householder. For arbitrary matrices a similar reduction can be performed by a technique suggested by Lanczos. In this method two systems of vectors are constructed, and which are biorthogonal; that is, for we have6.5. Lanczos' method 133 The initial vectors and can be chosen arbitrarily though in such a way that The new vectors are formed according to the rulesThe coefficients are determined from the biorthogonality condition, and for we form:If we getAnalogouslyLet us now consider the numerator in the expression for whenbecause of the biorthogonality. Hence we have for and similarly we also have under the same condition. In this way the following simpler formulas are obtained:If the vectors are considered as columns in a matrix and if further a tridiagonal matrix is formed from the coefficients and with one's in the remaining diagonal:then we can simply write and provided the vectors are linearly independent134 Algebraic eigenvalue problemsIf similar matrices are formed from the vectors and from the coefficientswe getCertain complications may arise, for example, that some or may become zero, but it can also happen that even if and The simplest way out is to choose other initial vectors even if it is sometimes possible to get around the difficulties by modifying the formulas themselves.Obviously, Lanczos' method can be used also with real symmetric or Hermi tian matrices. Then one chooses just one sequence of vectors which must form an orthogonal system. For closer details, particularly concerning the determination of the eigenvectors, Lanczos' paper should be consulted; a detailed discussion of the degenerate cases is given by Causey and GregoryHere we also mention still one method for tridiagonalization of arbitrary real matrices, first given by La Budde. Space limitations prevent us from a closer discussion, and instead we refer to the original paper6.6. Other methodsAmong other interesting methods we mention the method. Starting from a matrix we split it into two triangular matrices with and then we form Since the new matrix has the same eigenvalues as Then we treat in the same way as and so on, obtaining a sequence of matrices which in general converges toward an upper triangular matrix. If the eigenvalues are real, they will appear in the main diagonal. Even the case in which complex eigenvalues are present can be treated without serious complications. Closer details are given in where the method is described by its inventor, H. Rutishauser. Here we shall also examine the more general eigenvalue problem,where and are symmetric and, further,is positive definite. Then we can split according to where is a lower triangular matrix. Henceand where Sincethe problem has been reduced to the usual type treated before.6.7. Complex matricesFor computing eigenvalues and eigenvectors of arbitrary complex matrices (also, real nonsymmetric matrices fall naturally into this group), we shall first discuss a triangularization method suggested by Lotkin and Greenstadt6.7. Complex matrices 135The method depends on the lemma by Schur stating that for each square matrix there exists a unitary matrix such that where is a (lower or upper) triangular matrix (see Section 3.7). In practical computation one tries to find as a product of essentially two-dimensional unitary matrices, using a procedure similar to that described for Hermitian matrices in Section 6.2. It is possible to give examples for which the method does not converge (the sum of the squares of the absolute values of the subdiagonal elements is not monotonically decreasing, cf.but in practice convergence is obtained in many cases. We start by examining the two-dimensional case and put(6.7.1)From we get Further, we suppose that whereand obtain(6.7.2) Clearly we have Claiming we find withand(6.7.3) Here we conveniently choose the sign that makes as small as possible; with and we get Hence is obtained directly from the elements and Normally, we must take the square root of a complex number, and this can be done by the formulawhere When has been determined, we get and from(6.7.4) Now we pass to the main problem and assume that is an arbitrary complex matrix We choose that element below the main diagonal which is largest。
第八章 特征值问题
n
| x p | ¹ 0 ,因此
a p k xk
k 1, k p
k 1, k p
n
| a p k | | xk |
| xp |
从而
n
| a p k | | x p | Rp
| app | Rp
例 5
矩阵
骣 5 0.8 20 琪 A = 琪 10 1 4 琪 琪 琪 2 10i 1 桫
工程计算中,求解特征值问题 的特征对 ( , x ) 时,由于数据往往带有误差, 因此我们计算出的特征对 ( , x ) ,实际上是 扰动后的特征值问题
Ax x
xx A
的解。这里 A A E, E ( i j )
我们希望知道矩阵元素的变化对特征对的影响。 | 或j | ||的某个上界, i E || 由于我们一般只知道 因此有必要研究如何利用这样的上界,尽可能 x 准确地估计 与 、 与 x之间的差距,从 而可确定特征值问题的稳定性。 由于矩阵的特征多项式的系数是矩阵元素的连 续函数,而多项式的根都是其系数的连续函数, 因此矩阵的特征值作为特征多项式的零点都连 续地依赖于矩阵的元素。因此矩阵元素的连续 变化时,必有对应特征值的连续变化。
骣 5 0.4 20 琪 B = D- 1 AD = 琪 10 0.5 4 琪 琪 琪 4 10i 2 桫
三个行Gerschgorin圆分别收缩为:
G1¢( A) = { z ? C | z 20 | G2¢( A) = { z ? C | z 10 | G3¢( A) = { z ? C | z 10i |
i , j 1 i j
n
三、特征值的界
首先,根据矩阵 A 的Cartesian分解,有
高中数学知识点精讲精析 特征值与特征向量
2.5特征值与特征向量1.设矩阵A=a bc d⎛⎫⎪⎝⎭,如果存在实数λ及非零向量ξ,使得Aξλξ=,则称λ是矩阵A的一个特征值。
2.ξ是矩阵A的属于特征值的一个特征向量。
3.如果ξ是矩阵A的属于特征值λ的一个特征向量,则对任意的非零常数k,kξ也是矩阵A 的属于特征值λ的特征向量。
4.几何观点:特征向量的方向经过变换矩阵A的作用后,保持在同一直线上。
λ>0方向不变;λ<0方向相反;λ=0,特征向量就被变换成零向量。
属于矩阵的不同特征值的特征向量不共线。
5.数学上,线性变换的特征向量(本征向量)是一个非退化的向量,其方向在该变换下不变。
该向量在此变换下缩放的比例称为其特征值(本征值)。
一个变换通常可以由其特征值和特征向量完全描述。
特征空间是相同特征值的特征向量的集合。
6.定义空间上的变换—如平移(移动原点),旋转,反射,拉伸,压缩,或者这些变换的组合;以及其它变换—可以通过它们在向量上的作用来显示。
向量可以用从一点指向另一点的箭头来表示。
变换的特征向量是指在变换下不变或者简单地乘以一个缩放因子的非零向量。
特征向量的特征值是它所乘的那个缩放因子。
特征空间就是由所有有着相同特征值的特征向量组成的空间,还包括零向量,但要注意零向量本身不是特征向量。
变换的主特征向量是对应特征值最大的特征向量。
特征值的几何重次是相应特征空间的维数。
有限维向量空间上一个变换的谱是其所有特征值的集合。
例如,三维空间旋转的特征向量是沿着旋转轴的一个向量,相应的特征值是1,相应的特征空间包含所有和该轴平行的向量。
该特征空间是一个一维空间,因而特征值1的几何重次是1。
特征值1是旋转的谱当中唯一的实特征值。
特征值方程从数学上看,如果向量v与变换满足则称向量v是变换的一个特征向量,λ是相应的特征值。
其中是将变换作用于v 得到的向量。
这一等式被称作“特征值方程”。
假设是一个线性变换,那么v可以由其所在向量空间的一组基表示为:其中vi是向量在基向量上的投影(即坐标),这里假设向量空间为n 维。
特征值与多项式
设A为n阶方阵,若存在数λ和非零的 n维列向量x,使得 Ax=λx (1) 则称数λ为矩阵 A的特征值,称 x为矩阵A对应于特征值λ的
特征向量.
设x是对应于特征值λ的特征向量,由于 A(kx)=k(Ax)=k(λx)= λ(kx) k≠0 , 所以,kx也是A的对应于特征值λ的特征向量.这说明特征向量不 是被特征值唯一决定的.但是,特征值是被特征向量唯一决定的. 因此一个特征向量只属于一个特征值.
(2) 1 2
(注 : trA称为矩阵A的迹)
证 (1)由于1 , 2 , , n为A的特征值, 故 | I A | ( 1 )( 2 ) ( n ) = n (1 2 令 0, 得 | A | () 12
n
n ) n1
(1)n 12
n
n ,即 | A | 12
n
(2) 由于 a11 a12 a21 a22 | I A |
an1 an 2
a1n a2 n
Байду номын сангаас ann
的行列式的展开中, 注对角线的乘积 ( a11 )( a22 ) ( ann ) 是其中的一项,再由行列式的定义可知:展开式中的其余项至多 包含n-2个主对角线上的元素,因此|I-A|中含 n与 n 1的项只能 在主对角线元素乘积项中出现,故有
得基础解系
p2 (1, 4, 0)T ,p3 (1, 0, 4)T
所以k2 p2 k3 p3 (k2 , k3不全为零)是对应于
2 3 2的全部特征向量.
由例题知道,矩阵A的特征值之和1 2 3 3, 而矩 A的特征值之积12 3 4, 而矩阵A的行列式 | A | 4,当
11_特征值问题的数值方法
证明: 设Ax = λx, e ̸= 0, 记D = diag(a11 , · · · , ann ), 则有 (A − D)x = (λI − D)x. 若λ = aii , 定理显然成立. 现设λ ̸= aii , i = 1, · · · , n, 则 (λI − D)−1 (A − D)x = x 两边取范数有
记S (λ) = |y H x|. 上式说明λ的变化率反比于S (λ). 称(S (λ))−1 为关于特征 值λ的条件数. 如果A是实对称矩阵, 则可选P 为正交阵cond2 (P ) = 1,则
λ∈σ (A)
min |λ − µ| ≤ ∥E ∥2 .
且对A + E 的每个特征值µk , 总有A的特征值λi 使得|λi − µk | ≤ ∥E ∥2 , 此 处E 不一定对称. 如果E 对称,则有现面的定理 定 理5. 设A, E ∈ Rn×n 是对称阵, A, E 和A + E 的特征值分别按顺序排列成 λ1 ≥ λ2 ≥ · · · ≥ λn , ν1 ≥ ν2 ≥ · · · ≥ νn , µ1 ≥ µ2 ≥ · · · ≥ µn .
则有λi + νn ≤ µi ≤ λi + ν1 , i = 1, · · · , n.
证明: 设A对应于特征值λ1 , · · · , λn 的规范正交特征向量为{x1 , · · · , xn }, 记i为空间 W1 = span{x1 , x2 , · · · , xi }, 有前面的结论有 µi ≥ ((A + E )x, x) (Ax, x) (Ex, x) ≥ min + min x∈W1 ,x̸=0 x∈W1 ,x̸=0 (x, x) x∈W1 ,x̸=0 (x, x) (x, x) (Ex, x) = λi + min ≤ λi + νn . x∈W1 ,x̸=0 (x, x) min
数值分析考试大纲
数值分析》考试大纲一、考试标准(命题原则)1、考察学生对数值分析的基础知识(包括基本概念、基本内容、基本定理)的掌握程度以及运用已掌握的知识分析和解决问题的能力,衡量学生的数值分析及计算的能力。
2、题型比例客观题(判断题、填空题与选择题)约30--40%解答题(包括证明题)约60--70%3、难易适度,难中易比例:容易:40%,中等:50%,偏难10%。
4、考试知识点复盖率达80%以上。
二、考试时间:120分钟(2个小时)三、考试对象:数学与应用数学专业本科生四、考核知识点第一章引论(一)、知识点§1 数值分析的研究对象§2 数值计算的误差§3 病态问题、数值稳定性与避免误差危害§4 矩阵、向量和连续函数的范数(二)、基本要求1、了解向量和矩阵范数的定义和计算2、了解误差分析第二章插值法(一)、知识点§1 Lagrange插值§2 均差与Newton插值公式§3 插值余项的Peano估计§4 差分与等距节点插值公式§5 Hermite插值§6 分段低次插值§7 三次样条插值的计算方法§8 三次样条插值函数的性质与误差估计§9 B-样条函数§10 二元插值(二)、基本要求1、理解插值概念和插值问题的提法2、熟练掌握插值基函数、拉格朗日插值公式,会用余项定理估计误差3、掌握差商的概念及其性质,熟练掌握用差商表示的牛顿插值公式4、掌握埃米尔特插值、分段插值的定义和特点第三章函数逼近(一)、知识点§1 正交多项式§2 函数的最佳平方逼近§3 最小二乘法§4 周期函数的最佳平方逼近§5 快速Fourier变换§6 函数的最佳一致逼近§7 近似最佳一致逼近多项式§8 Chebyshev节约化(二)、基本要求1.了解正交多项式定义2.理解函数的最佳平方逼近3.掌握最小二乘法4.掌握周期函数的最佳平方逼近5.了解快速Fourier变换6.理解函数的最佳一致逼近7.了解近似最佳一致逼近多项式8.掌握Chebyshev节约化第四章数值积分和数值微分(一)、知识点§1 Newton-Cotes求积公式§2 复合求积公式§3 Peano的误差表示§4 Gauss求积公式§5 Romberg求积公式§6 奇异积分与振荡函数的积分§7 二维近似求积(二)、基本要求1、理解数值求积的基本思想,代数精度的概念2、熟练掌握梯形、辛普生等低价牛顿-柯特斯求积公式3、掌握复化求积公式:复化梯形求积公式、复化辛普生求积公式4、掌握龙贝格求积公式5、掌握高斯求积公式的定义和特点6、掌握几个数值微分公式第五章解线性代数方程组的直接方法(一)、知识点§1 Gauss消去法§2 主元素消去法§3 直接三角分解方法§4 矩阵的奇异值和条件数,直接方法的误差分析§5 解的迭代改进§6 稀疏矩阵技术介绍(二)、基本要求1、了解向量和矩阵范数的定义和计算2、掌握高斯消去法、按列选主元的高斯消去法、三角分解法3、了解求解特殊方程组的追赶法和Cholesky平方根法第六章解线性代数方程组的迭代方法(一)、知识点§1 迭代法的基本概念§2 Jacobi迭代法和Gauss-Seidel迭代法§3 超松弛(SOR)迭代法§4 共轭梯度法(二)、基本要求1、掌握Jacobi迭代法、Gauss-Seidel迭代法和SOR迭代法2、了解方程组右端项和系数矩阵的扰动对解的影响、方程组解法的误差分析第七章非线性方程和方程组的数值解法(一)、知识点§1 单个方程的迭代法§2 迭代加速收敛的方法§3 Newton迭代法§4 割线法与Muller方法§5 非线性方程组的不动点迭代法§6 非线性方程组的Newton法和拟Newton法(二)、基本要求1.掌握单个方程的迭代法2.了解迭代加速收敛的方法3.掌握Newton迭代法4.掌握割线法与Muller方法第八章代数特征值问题计算方法(一)、知识点§1 特征值问题的性质和估计§2 正交变换及矩阵分解§3 幂迭代法和逆幂迭代法§4 正交相似变换化矩阵为Hessenberg形式§5 QR方法§6 对称矩阵特征值问题的计算(二)、基本要求1.了解特征值问题的性质和估计2.理解正交变换及矩阵分解3.掌握幂迭代法和逆幂迭代法4.了解正交相似变换化矩阵为Hessenberg形式5.掌QR方法6.掌握对称矩阵特征值问题的计算第九章常微分方程初值问题的数值解法(一)、知识点§1 基本概念、Euler方法和有关的方法§2 Runge-Kutta方法§3 单步法的收敛性、相容性与绝对稳定性§4 线性多步法§5 线性差分方程§6 线性多步法的收敛性与稳定性§7 一阶方程组与刚性方程组(二)、基本要求1、了解一阶常微分方程初值问题数值解法的一些基本概念:步长、差分格式、单步法、多步法、显式法、隐式法、局部截断误差、整体截断误差、方法的阶数2、掌握欧拉法、改进欧拉法、梯形格式3、掌握龙格--库塔法的定义和特点4、了解亚当姆斯线性多步法5、了解差分法的收敛性和稳定性概念6、了解常微分方程边值问题五、考试要求书面答卷,闭卷考试,自带计算器。
线性代数 特征值问题 课件
故 1是矩阵A 1的特征值, 且x是A 1对应于 1 的特征向量.
三、特征值和特征向量的性质
n , 则有
1. 设 n 阶方阵 A aij 的特征值为1 , 2 ,,
(1) 1 2 n a11 a22 ann
即 i tr ( A)
8 6 2 (4 )( 2 ) 解特征方程 A I 0
即得A的特征值为 1 2, 2 4.
当 1 2时, 对应的特征向量应满足 3 2 1 x1 0 1 3 2 x 2 0 x1 x 2 0, x1 x 2 0.
得A的特征值为1 1, 2 3 2.
当1 1时, 解方程 A I x 0.由
1 1 1 1 0 1 A I 0 3 0 ~ 0 1 0 , 4 1 4 0 0 0 1 得基础解系 p1 0 , 1
式,当各i不相等时, 该行列式不等于 , 从而该矩阵 0 可逆.于是有 x1 p1 , x2 p2 ,, xm pm 0,0,,0,
即 x j p j 0 j 1,2,, m .但 p j 0,故 x j 0 j 1,2,, m .
所以向量组 p1 , p2 ,, pm 线性无关.
即
解得 x1 x 2 , 所以对应的特征向量可 p c 1, c 0. 取为 1
当 2 4时,由
1
3 4 1 x1 0 1 1 x1 0 ,即 , 1 3 4 x 2 0 1 1 x 2 0
第8章 特征值问题的计算方法
设 uk 和 k均收敛,由算法知 Auk1 k uk
lim
k
Auk 1
lim
k
k
lim
k
uk
Ax 1x k 1
uk 1
幂法可以计算矩阵的模最大 的特征值和对应的特征向量
例1:利用幂法求下列矩阵A 的模 2 1 0
最大的特征值及相应的特征向量. A 1 3 1
u0 1x1 2 x2 n xn;i C
n
n
Aku0
j Ak x j
j
k j
x
j
j
n 2
j
(
j 1
)k )x j )
Ak u0
1k
(1x1
n j2
j
(
j 1
)k
)
x
j
)
1x1(k
)
说明:当k充分大时, 1的一个近似特征向量为
uk
Ak u0
1k
特征向量可以相差一个倍数
因为向量 uk
Ak u0
1k
中含有未知量 1,实际不能计算
但我们关心的仅是 uk 的方向,故作如下处理:
令 uk
Ak u0
k
Ak u0
其中 k为 Ak u0 的模最大分量
1k (1x1
3 4.92
u3 y3 3 (0.3659 0.8537 1)T
Step4 y4 Au3 (1.5854 3.9268
4.8537)T
4 4.8537 u4 y4 4 (0.3266 0.8090 1)T
8.1 代数特征值问题
( 0, 0,1) , 10 .
T 3 (0)
解: Z y
y
( 0)
( 0, 0,1) ,
T
(1)
AZ
(1)
( 0)
( 0, 1, 2) ,
T
C 2,
y Z ( 0, 0.5,1)T , C y ( 2 ) AZ (1) ( 0.5, 2, 2.5)T ,
反幂法计算的主要步骤: 1.对A进行三角分解PA LU 2.求整数r,使得 yr( k ) max yi( k ) , C yr( k ) ,
1 i n (k ) y 计算Z ( k ) C 3.解方程组 LW ( k ) PZ ( k 1)
Uy ( k ) W
反幂法的一个应用
三、反幂法
反幂法是计算矩阵按模最小的特征值及特征向 量的方法,也是修正特征值、求相应特征向量的最 有效的方法。 设A为n n阶非奇异矩阵, , u为A的特征值与
相应的特征向量,即 Ax x x A x
1
A x
1
1
x
此式表明,A1的特征值是A的特征值的倒数,而 相应的特征向量不变。因此,若对矩阵A1用幂法, 即可计算出A1的按模最大的特征值,其倒数恰为A 的按模最小的特征值。 这就是反幂法的基本思想。
第8章 矩阵特征值计算
k 1 ,2 , 3 ,
第8章 矩阵特征值问题
工程实践中有多种振动问题,如桥梁 或建筑物 的振动,机械机件、飞机机翼的振动,及 一些稳定 性分析和相关分析可转 化为求矩阵特征值与特征向 量的问题。
矩阵A aij
nn
的特征值是A的特征多项式
f ( ) I A 的n个零点.
- 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
- 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
- 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。
λn k λ2 k (2) + ( ) α 2 x + L + ( ) α n x (n) ] λ1 λ1
n λi k λi k (i) lim( ) α i x = θ ⇒ lim ∑ ( ) α i x ( i ) = θ k→∞ λ k→∞ i = 2 λ1 1
故只要k充分大,就有 y
(k )
数值分析
数值分析
定理3 定理3
设 A∈Rn×n 为对称矩阵,则:
(1) A 的特征值均为实数; (2) A n 有 个线性无关的特征向量; (3)存在一个正交矩阵 P使
λ 1 P−1 AP =
λ2
, O λn
且 λi (i =1 L, n) A 为 特征值, 而 P = (u , u2 ,L, un ) 的列 , 1 向量 uj 为 A的对应于 λj 的特征向量.
p
P
p
E p,
(1.5)
其中 • 为矩阵 p 的范数, =1 2, ∞. p , p .这时 D − µI非奇异,设 x 证明 只要考虑 µ ∉σ ( A) 是 A+ E对应于 µ的特征向量,由( A+ E − µI )x = 0左乘 P−1 可得
数值分析
数值分析
(D − µI )(P−1x) = −(P−1EP)(P−1x), P−1x = −(D − µI )−1(P−1EP)(P−1x),
n
∑
(k )
由 y ( k ) = A y ( k −1) = A k y ( 0 ) = ⇒ y = λ [α 1 x
k 1 (1 )
i=1
α i x ( i )。
∑
n
i =1
A k (α i x ( i ) ) = ∑ α i λ ik x ( i )
i =1
n
设 α 1 ≠ 0, 由 λ1 > λ i ( i = 2, 3, L , n ) 得
A的3个圆盘为 1
E : 1
λ − 4 ≤1E2 : ,
λ≤
19 E , 3: 9
λ + 4 ≤1.8.
数值分析
数值分析
这样,3个圆盘都成为了孤立圆盘,每一个圆盘都包 含A 的一个特征值(为实特征值)且有估计
3 ≤ λ ≤ 5, 1
19 19 − ≤ λ2 ≤ , 9 9
−5.8 ≤ λ3 ≤ −2.2.
数值分析
第八章 矩阵特征值问题计算
工程实践中有多种振动问题,如桥梁 或建筑物 工程实践中有多种振动问题, 的振动,机械机件、飞机机翼的振动, 的振动,机械机件、飞机机翼的振动,及 一些稳定 性分析和相关分析可转 化为求矩阵特征值与特征向 量的问题。 量的问题。
矩阵A = aij
( )
n ×n
的特征值是A的特征多项式
数值分析
数值分析
例1
估计矩阵
4 A = 1 1 1 0 1 0 −1 −4
特征值的范围. 解
D: 1
的3个圆盘为 A
λ − 4 ≤1 D : , 2
λ ≤ 2,D : 3
λ + 4 ≤ 2,
由定理5,可知 A 的3个特征值位于3个圆盘的并集中, 由于 D 是孤立圆盘,所以 D内恰好包含 A 的一个特征值 1 1 ( λ 为实特征值),即 3 ≤ λ ≤ 5 1 1
(4) 若A 非奇异,λ ≠ 0,那么 是A -1的特征值
1
λ
数值分析
数值分析
定理2 定理2 (1)设 A∈Rn×n可对角化,即存在非奇异矩阵
P, 使
λ 1 P−1 AP =
λ2
O
λn
的充要条件是 A具有 n 个线性无关的特征向量. (2) 如果 A有 m (m ≤ n) 个 不同的特征值 λ , λ2 ,L, λm , 1 则对应的特征向量 x1, x2 ,L, xm 线性无关.
数值分析
数值分析
定理4 定理4
设 A∈Rn×n为对称矩阵(其特征值次序记为
λ ≥ λ2 ≥L≥ λn ),则 1
( Ax, x) 1. λn ≤ ≤λ 1 ( x, x)
( Ax,x) 2. λ =m n ax ; 1 x R ∈ (x, x) x≠0
(∀x ∈Rn , x ≠ 0);
( Ax, x) λn = m n in . x∈ R (x, x) x≠0
数值分析
数值分析
证明
只就(1)给出证明.
设 λ 为 A的特征值,即
Ax = λ x,
x = ( x1, x2 ,L, xn )T ≠ 0.
∞
记 xk = max xi = x
1≤i ≤n
的第 k个方程, ≠ 0, 考虑Ax = λx
即
∑a
j= 1
n
kj
xj = λ xk ,
或
(λ − akk )xk = ∑akj xj ,
λ1 ≈ yi( k +1) / yi( k ) , 相应的特征向量为y ( k+1)。
数值分析
数值分析
不 妨 令 x ( i ) = 1 ( i = 1, 2 , L , n )。 由 于 x ( i ) 可 看 做 基 , 故 必 存 在 不 全 是 零 的 α i ( i = 1, 2 , L , n ) , 使 得 y (0) =
数值分析
数值分析
定理5 定理5 (格什戈林圆盘定理) (1) 设 A = (aij )n×n, 则A 的每一个特征值必属于下述某个圆盘之中
n
λ − aii ≤ ri = ∑ aij
j =1 j ≠i
(i =1 2,L n). , ,
(1.4)
或者说, A 的特征值都在复平面上 n个圆盘的并集中. (2) 如果 A有 m 个圆盘组成一个连通的并集 S , 且 S与余下n − m 个圆盘是分离的,则 S内恰包含 A m 的 个 特征值. 特别地,如果 A 的一个圆盘 D 是与其他圆盘分离的 i (即孤立圆盘),则 D 中精确地包含 A的一个特征值. i
j ≠k n
于是
λ − akk xk ≤ ∑ akj x≤ xk j
j ≠k
n
∑a
j ≠k
n
kj
,
数值分析
数值分析
即
λ − akk ≤∑ akj = rk .
j ≠k
n
这说明, A 的每一个特征值必位于 A的一个圆盘中, 并且相应的特征值 λ一定位于第 k个圆盘中. 其中 k是对应特征向量 x绝对值最大的分量的下标.
yi ( k ) 是 y ( k )的第i 分量。
按 上 面 式 子 计 算 矩 阵 A按 模 最 大 的 特 征 值 与相应的特征向量的方法称为幂法。 法的收 幂
λ2 敛速度依赖于比值 ,比值越小,收敛越快。 λ1
数值分析
数值分析
两点说明: 1) 如 果 y ( 0 )的 选 取 恰 恰 使 得 α 1 = 0, 幂 法 计 算 仍 能 进行。因为计算过程中舍入误差的影响,迭代若干 次 后 , 必 然 会 产 生 一 个 向 量 y ( k ) , 它 在 x (1) 方 向 上 的 分 量不为零,这样,以后的计算就满足所设条件。 2) 因 y ( k ) = λ1k α 1 x (1) , 计 算 过 程 中 可 能 会 出 现 溢 出 ( λ1 > 1)或 成 为 0( λ1 < 1)的 情 形 。 解 决 方 法 : 每 次 迭 代所求的向量都要归范化。因此,幂法实际使用的 Z (0) = y(0) 计算公式是
数值分析
数值分析
第一节 特征值的估计和数值稳定性
定理1 定理1 则 (1) cλ为 cA的特征值( c为常数 c ≠ 0); (2)λ − µ 为 A− µI的特征值,即( A− µI ) = (λ − µ)x; (3) λ k为 Ak的特征值; 设 λ 为 A∈Rn×n 的特征值, Ax = λx x ≠ 0, ,
数值分析
数值分析
由定理6可知 P−1 P = cond(P) 是特征值扰动的放大系 数,但将 A 对角化的相似变换矩阵 P不是唯一的,所以取
cond( P) 的下确界
ν ( A) = inf{ cond(P) P−1AP = diag (λ1, λ2 ,L, λn )},
称为特征值问题的条件数.
(1.6)
只要 ν ( A) 不很大,矩阵微小扰动只带来特征值的微小 扰动.但是ν ( A)难以计算,有时只对一个 P,用 cond(P) 代替ν ( A).
数值分析
数值分析
特征值问题的条件数和解线性方程组的条件数是两个 不同的概念,对于一个矩阵 A ,两者可能一大一小. 关于计算矩阵 A 的特征值问题,当 n = 2, 3时,还可以 按行列式展开的方法求特征方程的根.但当 n较大时,如果 按展开行列式的方法,首先求出 p(λ)的系数,再求 p(λ) 的根,工作量就很大,用这种方法求特征值是不切实际的, 需要研究求 A 的特征值及特征向量的数值方法.
数值分析
数值分析
第二节 幂法和反幂法
一、幂法
求矩阵的按模最大的特征值(主特征值) 求矩阵的按模最大的特征值(主特征值)与相应 的特征向量。它是通过迭代产生向量序列, 的特征向量。它是通过迭代产生向量序列,由此计算特 征值和特征向量。 征值和特征向量。 设n × n阶实矩阵A的特征值λi ( i = 1, 2,L , n)
= λ [α 1 x
k 1
(1)
λi k + ∑ ( ) α i x ( i ) ] ≈ λ1k α 1 x (1) i = 2 λ1
n
数值分析
数值分析
因此,可把y ( k ) 看作特征值λ1的特征向量的近似。 由 y ( k +1) ≈ λ1k +1α1 x (1) , y ( k ) ≈ λ1kα1 x (1) ⇒ yi( k +1) λ1 ≈ ( k ) ( i = 1, 2,L n) yi