正交矩阵及实对称矩阵的对角化

矩阵可对角化

nn 阶方阵 A\bm{A} 相似于一个对角矩阵,即存在可逆矩阵 P\bm{P} 使

P1AP=Λ\bm{P}^{-1}\bm{A}\bm{P} = \bm{\Lambda}

则称 A\bm{A} 可对角化

nn 阶方阵 A\bm{A} 可对角化的充要条件是 A\bm{A}nn 个线性无关的特征向量。


证明:

    \impliedby

已知 A\bm{A}nn 个线性无关的特征向量 η1,η2,,ηn\bm{\eta}_1,\, \bm{\eta}_2,\, \cdots,\, \bm{\eta}_n

{Aη1=λ1η1Aη2=λ2η2Aηn=λnηn\left\lbrace\begin{aligned} \bm{A}\bm{\eta}_1 &= \lambda_1\bm{\eta}_1 \\ \bm{A}\bm{\eta}_2 &= \lambda_2\bm{\eta}_2 \\ \vdots \\ \bm{A}\bm{\eta}_n &= \lambda_n\bm{\eta}_n \end{aligned}\right.

P=[η1η2ηn]\bm{P} = \begin{bmatrix} \bm{\eta}_1 & \bm{\eta}_2 & \cdots & \bm{\eta}_n \end{bmatrix},则 P0\left\lvert P \right\rvert\ne 0,即 P\bm{P} 可逆。

AP=A[η1η2ηn]=[Aη1Aη2Aηn]=[λ1η1λ2η2λnηn]=[η1η2ηn][λ1λ2λn]=PΛ\begin{aligned} \bm{A} \bm{P} &= \bm{A} \begin{bmatrix} \bm{\eta}_1 & \bm{\eta}_2 & \cdots & \bm{\eta}_n \end{bmatrix} \\ &= \begin{bmatrix} \bm{A}\bm{\eta}_1 & \bm{A}\bm{\eta}_2 & \cdots & \bm{A}\bm{\eta}_n \end{bmatrix} \\ &= \begin{bmatrix} \lambda_1\bm{\eta}_1 & \lambda_2\bm{\eta}_2 & \cdots & \lambda_n\bm{\eta}_n \end{bmatrix} \\ &= \begin{bmatrix} \bm{\eta}_1 & \bm{\eta}_2 & \cdots & \bm{\eta}_n \end{bmatrix} \begin{bmatrix} \lambda_1 & & & \\ & \lambda_2 & & \\ & & \ddots & \\ & & & \lambda_n \end{bmatrix} \\ &= \bm{P} \bm{\Lambda} \end{aligned}

所以

P1AP=Λ\bm{P}^{-1}\bm{A}\bm{P} = \bm{\Lambda}

    \implies

Λ=[λ1λ2λn]\bm{\Lambda} = \begin{bmatrix} \lambda_1 & & & \\ & \lambda_2 & & \\ & & \ddots & \\ & & & \lambda_n \end{bmatrix}

P,P0,P1AP=Λ\exist_{\bm{P},\, |\bm{P}| \ne 0},\,\bm{P}^{-1}\bm{A}\bm{P} = \bm{\Lambda},则

PA=PΛ\bm{P} \bm{A} = \bm{P}\bm{\Lambda}

P=[η1η2ηn]\bm{P} = \begin{bmatrix} \bm{\eta}_1 & \bm{\eta}_2 & \cdots & \bm{\eta}_n \end{bmatrix},则

A[η1η2ηn]=[η1η2ηn][λ1λ2λn][Aη1Aη2Aηn]=[λ1η1λ2η2λnηn]\begin{aligned} A \begin{bmatrix} \bm{\eta}_1 & \bm{\eta}_2 & \cdots & \bm{\eta}_n \end{bmatrix} &= \begin{bmatrix} \bm{\eta}_1 & \bm{\eta}_2 & \cdots & \bm{\eta}_n \end{bmatrix} \\ &\kern{1em}\begin{bmatrix} \lambda_1 & & & \\ & \lambda_2 & & \\ & & \ddots & \\ & & & \lambda_n \end{bmatrix}\\ \begin{bmatrix} \bm{A}\bm{\eta}_1 & \bm{A}\bm{\eta}_2 & \cdots & \bm{A}\bm{\eta}_n \end{bmatrix} &= \begin{bmatrix} \lambda_1\bm{\eta}_1 & \lambda_2\bm{\eta}_2 & \cdots & \lambda_n\bm{\eta}_n \end{bmatrix} \\ \end{aligned}

所以

{Aη1=λ1η1Aη2=λ2η2Aηn=λnηn\left\lbrace\begin{aligned} \bm{A}\bm{\eta}_1 &= \lambda_1\bm{\eta}_1 \\ \bm{A}\bm{\eta}_2 &= \lambda_2\bm{\eta}_2 \\ \vdots \\ \bm{A}\bm{\eta}_n &= \lambda_n\bm{\eta}_n \end{aligned}\right.

P\bm{P} 可逆,所以 η1,η2,,ηn\bm{\eta}_1,\, \bm{\eta}_2,\, \cdots,\, \bm{\eta}_n 线性无关。


A\bm{A} 特征值没有重根(即恰有 nn 个不同的特征值),则 A\bm{A} 可对角化[1]


  1. 之前已证过不同特征值对应的特征向量线性无关。 ↩︎

幂零矩阵不可对角化。


证明:

O\bm{O} 显然成立。

假设 AO\bm{A} \ne \bm{O} 可对角化,则存在可逆矩阵 P\bm{P} 使

P1AP=[λ1λ2λn]\bm{P}^{-1}\bm{A}\bm{P} = \begin{bmatrix} \lambda_1 & & & \\ & \lambda_2 & & \\ & & \ddots & \\ & & & \lambda_n \end{bmatrix}

又因为

P1AkP=Λk\bm{P}^{-1}\bm{A}^k\bm{P} = \bm{\Lambda}^k

由于 A\bm{A} 幂零,故 Ak=O\bm{A}^k = \bm{O},即

P1AkP=O\bm{P}^{-1}\bm{A}^k\bm{P} = \bm{O}

所以

λ1k=λ2k==λnk=0\lambda_1^k = \lambda_2^k = \cdots = \lambda_n^k = 0

λ1=λ2==λn=0\lambda_1 = \lambda_2 = \cdots = \lambda_n = 0,那么

A=PΛP1=O\bm{A} = \bm{P}\bm{\Lambda}\bm{P}^{-1} = \bm{O}

AO\bm{A} \ne \bm{O} 矛盾,所以 A\bm{A} 不可对角化。


除此以外,若 J\bm{J} 为幂零矩阵,则 kE+Jk \bm{E} + \bm{J}(若尔当分解型)也不可对角化。

λ1,λ1,,λm\lambda_1,\, \lambda_1,\, \cdots,\, \lambda_mnn 阶方阵 A\bm{A} 所有相异的特征值,对于每个 i=1,,mi = 1,\, \cdots,\, m,记 ηi1,ηi2,,ηisi\bm{\eta}_{i1},\, \bm{\eta}_{i2},\, \cdots,\, \bm{\eta}_{is_i}A\bm{A} 关于 λi\lambda_i 的所有线性无关的特征向量,其中 sis_iλi\lambda_i 的代数重数,则向量组

η11,η12,,η1s1,η21,η22,,η2s2,,ηm1,ηm2,,ηmsm\begin{aligned} &\bm{\eta}_{11},\, \bm{\eta}_{12},\, \cdots,\, \bm{\eta}_{1s_1},\,\\ &\bm{\eta}_{21},\, \bm{\eta}_{22},\, \cdots,\, \bm{\eta}_{2s_2},\, \\ &\cdots,\, \\ &\bm{\eta}_{m1},\, \bm{\eta}_{m2},\, \cdots,\, \bm{\eta}_{ms_m} \end{aligned}

也线性无关。


证明:

k11,k12,,k1s1,k21,k22,,k2s2,,km1,km2,,kmsm\exist_{k_{11},\, k_{12},\, \cdots,\, k_{1s_1},\, k_{21},\, k_{22},\, \cdots,\, k_{2s_2},\, \cdots,\, k_{m1},\, k_{m2},\, \cdots,\, k_{ms_m}} 使

(k11η11+k12η12++k1s1η1s1)β1+(k21η21+k22η22++k2s2η2s2)β2++(km1ηm1+km2ηm2++kmsmηmsm)βm=θ\begin{aligned} &\overbrace{\left(k_{11}\bm{\eta}_{11} + k_{12}\bm{\eta}_{12} + \cdots + k_{1s_1}\bm{\eta}_{1s_1}\right)}^{\bm{\beta}_1} &+ \\ &\overbrace{\left(k_{21}\bm{\eta}_{21} + k_{22}\bm{\eta}_{22} + \cdots + k_{2s_2}\bm{\eta}_{2s_2}\right)}^{\bm{\beta}_2} &+ \\ &\cdots &+ \\ &\overbrace{\left(k_{m1}\bm{\eta}_{m1} + k_{m2}\bm{\eta}_{m2} + \cdots + k_{ms_m}\bm{\eta}_{ms_m}\right)}^{\bm{\beta}_m} &= \bm{\theta} \end{aligned}

β1+β2++βm=θ(1)\bm{\beta}_1 + \bm{\beta}_2 + \cdots + \bm{\beta}_m = \bm{\theta}\tag{1}

其中

{Aβ1=λ1β1Aβ2=λ2β2Aβm=λmβm(2)\left\lbrace\begin{aligned} \bm{A} \bm{\beta}_1 &= \lambda_1 \bm{\beta}_1 \\ \bm{A} \bm{\beta}_2 &= \lambda_2 \bm{\beta}_2 \\ \vdots \\ \bm{A} \bm{\beta}_m &= \lambda_m \bm{\beta}_m \tag{2} \end{aligned}\right.

不妨设 βj1,βj2,,βjt\bm{\beta}_{j_1},\, \bm{\beta}_{j_2},\, \cdots,\, \bm{\beta}_{j_t}β1,β2,,βm\bm{\beta}_1,\, \bm{\beta}_2,\, \cdots,\, \bm{\beta}_m 中不为 θ\bm{\theta} 的向量,其中 tmt \le m,则由 (1)(1)

βj1+βj2++βjt=θ(3)\bm{\beta}_{j_1} + \bm{\beta}_{j_2} + \cdots + \bm{\beta}_{j_t} = \bm{\theta}\tag{3}

(2)(2)βj1,βj2,,βjt\bm{\beta}_{j_1},\, \bm{\beta}_{j_2},\, \cdots,\, \bm{\beta}_{j_t}A\bm{A} 关于 λj1,λj2,,λjt\lambda_{j_1},\, \lambda_{j_2},\, \cdots,\, \lambda_{j_t} 的特征向量,且 λj1,λj2,,λjt\lambda_{j_1},\, \lambda_{j_2},\, \cdots,\, \lambda_{j_t} 相异,所以 βj1,βj2,,βjt\bm{\beta}_{j_1},\, \bm{\beta}_{j_2},\, \cdots,\, \bm{\beta}_{j_t} 线性无关,从而与 (3)(3) 矛盾(线性无关的向量组无法被线性表示为零向量)。

所以 β1,β2,,βm\bm{\beta}_1,\, \bm{\beta}_2,\, \cdots,\, \bm{\beta}_m 中不为 θ\bm{\theta} 的向量个数 t=0t = 0,即 β1,β2,,βm\bm{\beta}_1,\, \bm{\beta}_2,\, \cdots,\, \bm{\beta}_m 全为 θ\bm{\theta}

i,βi=θ\forall_{i},\,\bm{\beta}_{i} = \bm{\theta},且有 ηi1,ηi2,,ηisi\bm{\eta}_{i1},\, \bm{\eta}_{i2},\, \cdots,\, \bm{\eta}_{is_i} 线性无关,所以 ki1=ki2==kisi=0k_{i1} = k_{i2} = \cdots = k_{is_i} = 0,得证。

nn 阶方阵 A\bm{A} 特征多项式有因式分解

λEA=(λλ1)n1(λλ2)n2(λλm)nm\left\lvert \lambda \bm{E} - \bm{A} \right\rvert = (\lambda - \lambda_1)^{n_1}(\lambda - \lambda_2)^{n_2}\cdots(\lambda - \lambda_m)^{n_m}

其中 λ1,λ2,,λm\lambda_1,\, \lambda_2,\, \cdots,\, \lambda_mA\bm{A} 的所有相异的特征值,n1,n2,,nmn_1,\, n_2,\, \cdots,\, n_m 分别为 λ1,λ2,,λm\lambda_1,\, \lambda_2,\, \cdots,\, \lambda_m代数重数ni1n_i \ge 1n1+n2++nm=nn_1 + n_2 + \cdots + n_m = n

方程 (λiEA)x=θ(\lambda_i \bm{E} - \bm{A}) \bm{x} = \bm{\theta} 的基础解系的个数,即 nr(λiEA) n - \rank(\lambda_i \bm{E} - \bm{A}),称为 λi\lambda_i几何重数

几何重数 \le 代数重数。

λ0\lambda_0A\bm{A}kk 重特征值,则其几何重数 k\le k,即 λ0\lambda_0 的线性无关的特征向量个数不超过 kk 个。


证明:

(λ0EA)x=θ(\lambda_0 \bm{E} - \bm{A})\bm{x} = \bm{\theta} 的基础解系为 η1,η2,,ηsRn\bm{\eta}_1,\, \bm{\eta}_2,\, \cdots,\, \bm{\eta}_s \in \R^{n}

{Aη1=λ0η1Aη2=λ0η2Aηs=λ0ηs\left\lbrace\begin{aligned} \bm{A} \bm{\eta}_1 &= \lambda_0 \bm{\eta}_1 \\ \bm{A} \bm{\eta}_2 &= \lambda_0 \bm{\eta}_2 \\ \vdots \\ \bm{A} \bm{\eta}_s &= \lambda_0 \bm{\eta}_s \\ \end{aligned}\right.

η1,η2,,ηs\bm{\eta}_1,\, \bm{\eta}_2,\, \cdots,\, \bm{\eta}_s 扩充为 Rn\R^{n} 的一组线性无关的向量组 η1,η2,,ηs,ηs+1,,ηn\bm{\eta}_1,\, \bm{\eta}_2,\, \cdots,\, \bm{\eta}_s,\, \bm{\eta}_{s+1},\, \cdots,\, \bm{\eta}_{n}[1]

Aηs+1=c1,s+1η1++cs,s+1ηs+cs+1,s+1ηs+1++cn,s+1ηnAηn=c1,nη1++cs,nηs+cs+1,nηs+1++cn,nηn\begin{aligned} \bm{A} \bm{\eta}_{s+1} &= c_{1, s+1} \bm{\eta}_1 + \cdots + c_{s, s+1} \bm{\eta}_s + c_{s+1, s+1} \bm{\eta}_{s+1} + \cdots + c_{n, s+1} \bm{\eta}_{n} \\ \vdots\\ \bm{A} \bm{\eta}_{n} &= c_{1, n} \bm{\eta}_1 + \cdots + c_{s, n} \bm{\eta}_s + c_{s+1, n} \bm{\eta}_{s+1} + \cdots + c_{n, n} \bm{\eta}_{n} \end{aligned}

[Aη1Aηn]=[η1ηsηs+1ηn][λ0c1,s+1c1,nλ0cs,s+1cs,ncs+1,s+1cs+1,ncn,s+1cn,n]=[η1ηsηs+1ηn][λ0EsC1OC2]\begin{aligned} \begin{bmatrix} \bm{A} \bm{\eta}_1 & \cdots & \bm{A} \bm{\eta}_n \end{bmatrix} &= \begin{bmatrix} \bm{\eta}_1 & \cdots & \bm{\eta}_s & \bm{\eta}_{s+1} & \cdots & \bm{\eta}_{n} \end{bmatrix}\\ &\kern{1em}\begin{bmatrix} \begin{array}{ccc:ccc} \lambda_0 & & & c_{1, s+1} & \cdots & c_{1, n} \\ & \ddots & & \vdots & \ddots & \vdots \\ & & \lambda_0 & c_{s, s+1} & \cdots & c_{s, n} \\ \hdashline & & & c_{s+1, s+1} & \cdots & c_{s+1, n} \\ & & & \vdots & \ddots & \vdots \\ & & & c_{n, s+1} & \cdots & c_{n, n} \end{array} \end{bmatrix}\\ &= \begin{bmatrix} \bm{\eta}_1 & \cdots & \bm{\eta}_s & \bm{\eta}_{s+1} & \cdots & \bm{\eta}_{n} \end{bmatrix} \begin{bmatrix} \lambda_0 \bm{E}_s & \bm{C}_1 \\ \bm{O} & \bm{C}_2 \end{bmatrix} \end{aligned}

P=[η1ηsηs+1ηn]\bm{P} = \begin{bmatrix} \bm{\eta}_1 & \cdots & \bm{\eta}_s & \bm{\eta}_{s+1} & \cdots & \bm{\eta}_{n} \end{bmatrix},则

AP=P[λ0EsC1OC2]P1AP=[λ0EsC1OC2]=B\begin{aligned} \bm{A} \bm{P} &= \bm{P} \begin{bmatrix} \lambda_0 \bm{E}_s & \bm{C}_1 \\ \bm{O} & \bm{C}_2 \end{bmatrix}\\ \bm{P}^{-1} \bm{A} \bm{P} &= \begin{bmatrix} \lambda_0 \bm{E}_s & \bm{C}_1 \\ \bm{O} & \bm{C}_2 \end{bmatrix}\\ &= \bm{B} \end{aligned}

A\bm{A}B\bm{B} 相似。

λEB=(λλ0)EsC1OλEnsC2=(λλ0)EsλEnsC2=(λλ0)sλEnsC2=(λλ0)sfC2(λ)\begin{aligned} \left\lvert \lambda \bm{E} - \bm{B} \right\rvert &= \begin{vmatrix} (\lambda - \lambda_0)\bm{E}_s & -\bm{C}_1 \\ \bm{O} & \lambda \bm{E}_{n-s} - \bm{C}_2 \end{vmatrix}\\ &= \left\lvert (\lambda - \lambda_0) \bm{E}_s\right\rvert \left\lvert \lambda \bm{E}_{n-s} - \bm{C}_2 \right\rvert\\ &= (\lambda - \lambda_0)^s \left\lvert \lambda \bm{E}_{n-s} - \bm{C}_2 \right\rvert\\ &= (\lambda - \lambda_0)^s f_{\bm{C}_2}(\lambda) \end{aligned}

λ0\lambda_0 至少为 B\bm{B}ss 重特征值。

λEA=(λλ0)kh(λ)\left\lvert \lambda \bm{E} - \bm{A} \right\rvert = (\lambda - \lambda_0)^{k}h(\lambda)[2]λ0\lambda_0 不是 h(λ)=0h(\lambda)=0 的根,因为 λ0\lambda_0 只是 A\bm{A}kk 重特征值),且 λEA=λEB\left\lvert \lambda \bm{E} - \bm{A} \right\rvert = \left\lvert \lambda \bm{E} - \bm{B} \right\rvert,所以 sks \le k。否则若 s>ks > k,则 λ0\lambda_0 至少为 A\bm{A}ss 重特征值,与 λ0\lambda_0A\bm{A}kk 重特征值矛盾。


锚点有点问题,Admonition 里的锚点是单独放在 Admonition 里面的,然而锚点地址是一样的,导致这里的第一个锚点(以及后面的唯一锚点)都会跳转到上面的锚点。这个问题,以及 Admonition 内代码块渲染的问题都是因为 Admonition 里面的内容是单独渲染的,我也没暂时没有解决方案,只能说 Admonition 里的锚点尽量少用吧。


  1. 可扩充的证明:span{η1,η2,,ηs}={k1η1+k2η2++ksηski,kiR}\operatorname{span}\left\lbrace \bm{\eta}_1,\, \bm{\eta}_2,\, \cdots,\, \bm{\eta}_s \right\rbrace = \left\lbrace k_1\bm{\eta}_1 + k_2\bm{\eta}_2 + \cdots + k_s\bm{\eta}_s \mid \forall_{k_i},\,k_i \in \R \right\rbrace,则 ηs+1,ηs+1span{η1,η2,,ηs}\exist_{\bm{\eta}_{s+1}},\,\bm{\eta}_{s+1} \notin \operatorname{span}\left\lbrace \bm{\eta}_1,\, \bm{\eta}_2,\, \cdots,\, \bm{\eta}_s \right\rbrace,则 η1,η2,,ηs,ηs+1\bm{\eta}_1,\, \bm{\eta}_2,\, \cdots,\, \bm{\eta}_s,\, \bm{\eta}_{s+1} 线性无关,以此类推。 ↩︎

  2. h(λ)h(\lambda) 与前面的 fC2(λ)f_{\bm{C}_2}(\lambda) 都是关于 λ\lambda 的多项式。 ↩︎

大分块矩阵源码
1
2
3
4
5
6
7
8
9
10
11
\begin{bmatrix}
\begin{array}{cccc:ccc}
\lambda_0 & & & c_{1, s+1} & \cdots & c_{1, n} \\
& \ddots & & \vdots & \ddots & \vdots \\
& & \lambda_0 & c_{s, s+1} & \cdots & c_{s, n} \\
\hdashline
& & & c_{s+1, s+1} & \cdots & c_{s+1, n} \\
& & & \vdots & \ddots & \vdots \\
& & & c_{n, s+1} & \cdots & c_{n, n}
\end{array}
\end{bmatrix}

正交矩阵

α,β\bm{\alpha},\, \bm{\beta}nn 维向量,按列向量表示为

α=(a1a2an),β=(b1b2bn)\bm{\alpha} = \begin{pmatrix} a_1 \\ a_2 \\ \vdots \\ a_n \end{pmatrix},\quad \bm{\beta} = \begin{pmatrix} b_1 \\ b_2 \\ \vdots \\ b_n \end{pmatrix}

α,β\bm{\alpha},\, \bm{\beta}实向量,则称

i=1naibi\sum_{i=1}^n a_i b_i

α,β\bm{\alpha},\, \bm{\beta}内积,记为 (α,β)(\bm{\alpha}, \bm{\beta})

(α,β)=αβ(\bm{\alpha}, \bm{\beta}) = \bm{\alpha}^\intercal \bm{\beta}

α,β\bm{\alpha},\, \bm{\beta}复向量,则称

i=1naibˉi\sum_{i=1}^n a_i \bar{b}_i

α,β\bm{\alpha}, \bm{\beta}内积,记为 (α,β)(\bm{\alpha}, \bm{\beta})[1]

(α,β)=αβˉ(\bm{\alpha}, \bm{\beta}) = \bm{\alpha}^\intercal \bar{\bm{\beta}}


  1. bˉi\bar{b}_i 表示 bib_i共轭复数。使用共轭复数,可保证对复向量 α=(a1,a2,,an)\bm{\alpha} = \left( a_1, a_2, \cdots, a_n \right)^\intercal,有 (α,α)=i=1nai2(\bm{\alpha}, \bm{\alpha}) = \displaystyle \sum_{i=1}^{n}\left\lvert a_i \right\rvert^2↩︎

后面讨论的大都是实向量。而实向量内积的性质,都比较容易,不再赘述。

然而遗憾的是,一个比较基本的性质,即可交换性,对于复向量内积来说,却不再成立,也即

(α,β)(β,α)(\bm{\alpha}, \bm{\beta}) \neq (\bm{\beta}, \bm{\alpha})

这也难怪,看复向量内积的定义,看起来就不太对称的样子。而复向量内积那样定义,是为了保证模定义的有意义性。

并记

α=(α,α)\|\bm{\alpha}\| = \sqrt{(\bm{\alpha}, \bm{\alpha})}

α\bm{\alpha}

(α,β)=0(\bm{\alpha}, \bm{\beta}) = 0,则称 α,β\bm{\alpha},\, \bm{\beta} 正交垂直)。

α,β\bm{\alpha},\, \bm{\beta} 不为零向量,则称

θ=arccos((α,β)αβ),0θπ\theta = \arccos \left(\frac{(\bm{\alpha}, \bm{\beta})}{\|\bm{\alpha}\| \boldsymbol{\cdot} \|\bm{\beta}\|}\right),\quad 0 \le \theta \le \pi

α,β\bm{\alpha},\, \bm{\beta}夹角

若一个不含零向量的向量组 {αi}\left\lbrace \bm{\alpha}_i \right\rbrace 中的任意两个向量都正交,则称 {αi}\left\lbrace \bm{\alpha}_i \right\rbrace正交向量组

若一个不含零向量的向量组 {αi}\{\bm{\alpha}_i\} 中的任意两个向量都正交,且 {αi}\{\bm{\alpha}_i\} 中任意向量的模都为 11,则称 {αi}\{\bm{\alpha}_i\}标准正交向量组,简称法正交组,还可表示为

(αi,αj)={1,i=j0,ij(\bm{\alpha}_i, \bm{\alpha}_j) = \begin{cases} 1, & i = j \\ 0, & i \neq j \end{cases}

{αi}m\left\lbrace \bm{\alpha}_i \right\rbrace_m 为正交向量组,则 {αi}m\left\lbrace \bm{\alpha}_i \right\rbrace_m 线性无关。

k1α1+k2α2++kmαm=θ(1)k_1 \bm{\alpha}_1 + k_2 \bm{\alpha}_2 + \cdots + k_m \bm{\alpha}_m = \bm{\theta} \tag{1}

对式 (1)(1) 左乘 αi\bm{\alpha}_i^\intercal,得

j=1n(αi,kjαj)=0    ki(αi,αi)=0    ki=0\begin{aligned} \sum_{j=1}^{n}(\bm{\alpha}_i, k_j \bm{\alpha}_j) &= 0 & \implies \\ k_i (\bm{\alpha}_i, \bm{\alpha}_i) &= 0 & \implies \\ k_i &= 0 \end{aligned}

从而 i=1,2,,m,ki=0\forall_{i = 1, 2, \cdots, m},\, k_i = 0,即 {αi}m\left\lbrace \bm{\alpha}_i \right\rbrace_m 线性无关。

αRn\bm{\alpha} \in \R^n 与法正交组 {ei}n\left\lbrace \bm{e}_i \right\rbrace_n,有

α=i=1n(α,ei)ei\bm{\alpha} = \sum_{i=1}^{n}(\bm{\alpha}, \bm{e}_i) \bm{e}_i

施密特正交化

施密特正交化

{αi}n\left\lbrace \bm{\alpha}_i \right\rbrace_n 为线性无关的向量组,目标是构造出等价的正交向量组 {βi}n\left\lbrace \bm{\beta}_i \right\rbrace_n

不妨取 β1=α1θ\bm{\beta}_1 = \bm{\alpha}_1 \ne \bm{\theta}

α2\bm{\alpha}_2β1\bm{\beta}_1 的投影向量为 (α2,β1)β1β1β1\dfrac{(\bm{\alpha}_2, \bm{\beta}_1)}{\| \bm{\beta}_1 \|} \dfrac{\bm{\beta}_1}{\| \bm{\beta}_1 \|}(第一个分式是投影长度,第二个分式是投影方向的单位向量),于是取

β2=α2(α2,β1)β12β1\bm{\beta}_2 = \bm{\alpha}_2 - \frac{(\bm{\alpha}_2, \bm{\beta}_1)}{\| \bm{\beta}_1 \|^2} \bm{\beta}_1

β2\bm{\beta}_2β1\bm{\beta}_1 正交,同时显然 β2θ\bm{\beta}_2 \ne \bm{\theta},否则 α1,α2\bm{\alpha}_1,\, \bm{\alpha}_2 共线,与 {αi}n\left\lbrace \bm{\alpha}_i \right\rbrace_n 线性无关矛盾。

同理设 β3=α3c1β1c2β2\bm{\beta}_3 = \bm{\alpha}_3 - c_1 \bm{\beta}_1 - c_2 \bm{\beta}_2,其中 c1,c2c_1,\, c_2 为待定系数,使得 β3\bm{\beta}_3β1,β2\bm{\beta}_1,\, \bm{\beta}_2 都正交,即

(β3,β1)=0    (α3c1β1c2β2,β1)=0    (α3,β1)c1(β1,β1)c2(β2,β1)=0    (α3,β1)c1β12=0    c1=(α3,β1)β12\begin{aligned} (\bm{\beta}_3, \bm{\beta}_1) = 0 &\implies (\bm{\alpha}_3 - c_1 \bm{\beta}_1 - c_2 \bm{\beta}_2, \bm{\beta}_1) = 0 \\ &\implies (\bm{\alpha}_3, \bm{\beta}_1) - c_1 (\bm{\beta}_1, \bm{\beta}_1) - c_2 (\bm{\beta}_2, \bm{\beta}_1) = 0 \\ &\implies (\bm{\alpha}_3, \bm{\beta}_1) - c_1 \| \bm{\beta}_1 \|^2 = 0 \\ &\implies c_1 = \frac{(\bm{\alpha}_3, \bm{\beta}_1)}{\| \bm{\beta}_1 \|^2} \end{aligned}

同理可得 c2=(α3,β2)β22c_2 = \dfrac{(\bm{\alpha}_3, \bm{\beta}_2)}{\| \bm{\beta}_2 \|^2},于是取

β3=α3(α3,β1)β12β1(α3,β2)β22β2\bm{\beta}_3 = \bm{\alpha}_3 - \frac{(\bm{\alpha}_3, \bm{\beta}_1)}{\| \bm{\beta}_1 \|^2} \bm{\beta}_1 - \frac{(\bm{\alpha}_3, \bm{\beta}_2)}{\| \bm{\beta}_2 \|^2} \bm{\beta}_2

显然 β3\bm{\beta}_3β1,β2\bm{\beta}_1,\, \bm{\beta}_2 都正交,同时 β3θ\bm{\beta}_3 \ne \bm{\theta}

以此类推,设 βk=αkc1β1c2β2ck1βk1\bm{\beta}_k = \bm{\alpha}_k - c_1 \bm{\beta}_1 - c_2 \bm{\beta}_2 - \cdots - c_{k-1} \bm{\beta}_{k-1}β1,β2,,βk1\bm{\beta}_1,\, \bm{\beta}_2, \cdots, \bm{\beta}_{k-1} 都正交,且 βkθ\bm{\beta}_k \ne \bm{\theta},则

(βk,β1)=0    (αkc1β1c2β2ck1βk1,β1)=0    (αk,β1)c1(β1,β1)c2(β2,β1)ck1(βk1,β1)=0    (αk,β1)c1β12=0    c1=(αk,β1)β12\begin{aligned} (\bm{\beta}_k, \bm{\beta}_1) = 0 &\implies (\bm{\alpha}_k - c_1 \bm{\beta}_1 - c_2 \bm{\beta}_2 - \cdots - c_{k-1} \bm{\beta}_{k-1}, \bm{\beta}_1) = 0 \\ &\implies (\bm{\alpha}_k, \bm{\beta}_1) - c_1 (\bm{\beta}_1, \bm{\beta}_1) - c_2 (\bm{\beta}_2, \bm{\beta}_1) - \cdots - c_{k-1} (\bm{\beta}_{k-1}, \bm{\beta}_1) = 0 \\ &\implies (\bm{\alpha}_k, \bm{\beta}_1) - c_1 \| \bm{\beta}_1 \|^2 = 0 \\ &\implies c_1 = \frac{(\bm{\alpha}_k, \bm{\beta}_1)}{\| \bm{\beta}_1 \|^2} \end{aligned}

同理可得 c2=(αk,β2)β22,,ck1=(αk,βk1)βk12c_2 = \dfrac{(\bm{\alpha}_k, \bm{\beta}_2)}{\| \bm{\beta}_2 \|^2},\, \cdots,\, c_{k-1} = \dfrac{(\bm{\alpha}_k, \bm{\beta}_{k-1})}{\| \bm{\beta}_{k-1} \|^2},于是取

βk=αk(αk,β1)β12β1(αk,β2)β22β2(αk,βk1)βk12βk1=αki=1k1(αk,βi)βi2βi\begin{aligned} \bm{\beta}_k &= \bm{\alpha}_k - \frac{(\bm{\alpha}_k, \bm{\beta}_1)}{\| \bm{\beta}_1 \|^2} \bm{\beta}_1 - \frac{(\bm{\alpha}_k, \bm{\beta}_2)}{\| \bm{\beta}_2 \|^2} \bm{\beta}_2 - \cdots - \frac{(\bm{\alpha}_k, \bm{\beta}_{k-1})}{\| \bm{\beta}_{k-1} \|^2} \bm{\beta}_{k-1}\\ &= \boxed{\bm{\alpha}_{\textcolor{FF0099}{k}} - \sum_{\textcolor{82B1F5}{i}=1}^{\textcolor{FF0099}{k - 1}} \frac{(\bm{\alpha}_{\textcolor{FF0099}{k}}, \bm{\beta}_{\textcolor{82B1F5}{i}})}{\| \bm{\beta}_{\textcolor{82B1F5}{i}} \|^2} \bm{\beta}_{\textcolor{82B1F5}{i}}} \end{aligned}

碎碎念

最后的公式在 light 模式看不清,主要是因为我自己主要看 dark 模式,而且调一下 light 和 dark 都明显的颜色挺麻烦的,所以就算了。

还是找了一下,根据一个 StackExchange 的回答分别用了 #FF0099#0099FF#FF0099 挺完美的,效果很理想,#0099FF 不太行,还稍微牺牲了一下 dark 模式的对比度,只不过 Ta 给的颜色里也没找到更好的了。

我自己改了一下,把 #0099FF 改为了 #82B1F5,这样就不会牺牲 dark 模式的对比度了,而且 light 模式也不会太难看,唯一有个缺点就是不如 #FF0099 那么醒目,但是我觉得这个缺点可以接受。

若实方阵 A\bm{A} 满足 AA=E\bm{A}^\intercal \bm{A} = \bm{E},则称 A\bm{A}正交矩阵

对方阵 A\bm{A},以下表述等价:

  1. A\bm{A} 为正交矩阵
  2. AA=E\bm{A}^\intercal \bm{A} = \bm{E}
  3. A=A1\bm{A}^\intercal = \bm{A}^{-1}
  4. A\bm{A} 的列向量构成法正交组
  5. A\bm{A} 的行向量构成法正交组

证明:

前三者等价性显然。

(4) (5)     \implies (1):(这里演示 (4)     \implies (1),(5)     \implies (1) 同理)

对于法正交组 {αi}n\left\lbrace \bm{\alpha}_i \right\rbrace_n,设 A=[α1α2αn]\bm{A} = \begin{bmatrix} \bm{\alpha}_1 & \bm{\alpha}_2 & \cdots & \bm{\alpha}_n \end{bmatrix},则

A=[α1α2αn]\bm{A}^\intercal = \begin{bmatrix} \bm{\alpha}_1^\intercal \\ \bm{\alpha}_2^\intercal \\ \vdots \\ \bm{\alpha}_n^\intercal \end{bmatrix}

从而

AA=[α1α2αn][α1α2αn]=[α1α1α1α2α1αnα2α1α2α2α2αnαnα1αnα2αnαn]=[100010001]=E\begin{aligned} \bm{A}^\intercal \bm{A} &= \begin{bmatrix} \bm{\alpha}_1^\intercal \\ \bm{\alpha}_2^\intercal \\ \vdots \\ \bm{\alpha}_n^\intercal \end{bmatrix} \begin{bmatrix} \bm{\alpha}_1 & \bm{\alpha}_2 & \cdots & \bm{\alpha}_n \end{bmatrix} \\ &= \begin{bmatrix} \bm{\alpha}_1^\intercal \bm{\alpha}_1 & \bm{\alpha}_1^\intercal \bm{\alpha}_2 & \cdots & \bm{\alpha}_1^\intercal \bm{\alpha}_n \\ \bm{\alpha}_2^\intercal \bm{\alpha}_1 & \bm{\alpha}_2^\intercal \bm{\alpha}_2 & \cdots & \bm{\alpha}_2^\intercal \bm{\alpha}_n \\ \vdots & \vdots & \ddots & \vdots \\ \bm{\alpha}_n^\intercal \bm{\alpha}_1 & \bm{\alpha}_n^\intercal \bm{\alpha}_2 & \cdots & \bm{\alpha}_n^\intercal \bm{\alpha}_n \end{bmatrix} \\ &= \begin{bmatrix} 1 & 0 & \cdots & 0 \\ 0 & 1 & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \cdots & 1 \end{bmatrix} \\ &= \bm{E} \end{aligned}

从而 A\bm{A} 为正交矩阵。

(1)     \implies (4):(这里演示 (1)     \implies (4),(1)     \implies (5) 同理)

A\bm{A} 为正交矩阵,即 AA=E\bm{A}^\intercal \bm{A} = \bm{E},设 A=[α1α2αn]\bm{A} = \begin{bmatrix} \bm{\alpha}_1 & \bm{\alpha}_2 & \cdots & \bm{\alpha}_n \end{bmatrix},则同理

AA=E    [α1α1α1α2α1αnα2α1α2α2α2αnαnα1αnα2αnαn]=[100010001]    αiαj={1,i=j0,ij\begin{aligned} \bm{A}^\intercal \bm{A} &= \bm{E} &\implies \\ \begin{bmatrix} \bm{\alpha}_1^{\intercal} \bm{\alpha}_1 & \bm{\alpha}_1^{\intercal} \bm{\alpha}_2 & \cdots & \bm{\alpha}_1^{\intercal} \bm{\alpha}_n \\ \bm{\alpha}_2^{\intercal} \bm{\alpha}_1 & \bm{\alpha}_2^{\intercal} \bm{\alpha}_2 & \cdots & \bm{\alpha}_2^{\intercal} \bm{\alpha}_n \\ \vdots & \vdots & \ddots & \vdots \\ \bm{\alpha}_n^{\intercal} \bm{\alpha}_1 & \bm{\alpha}_n^{\intercal} \bm{\alpha}_2 & \cdots & \bm{\alpha}_n^{\intercal} \bm{\alpha}_n \end{bmatrix} &= \begin{bmatrix} 1 & 0 & \cdots & 0 \\ 0 & 1 & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \cdots & 1 \end{bmatrix} &\implies \\ \bm{\alpha}_i^\intercal \bm{\alpha}_j &= \begin{cases} 1, & i = j \\ 0, & i \neq j \end{cases} \end{aligned}

{αi}n\left\lbrace \bm{\alpha}_i \right\rbrace_n 为法正交组。

设二阶方阵

A=[a11a12a21a22]\bm{A} = \begin{bmatrix} a_{11} & a_{12} \\ a_{21} & a_{22} \end{bmatrix}

为正交矩阵,则由 A=A1\bm{A}^\intercal = \bm{A}^{-1} 可得

[a11a21a12a22]=1A[a22a12a21a11]\begin{bmatrix} a_{11} & a_{21} \\ a_{12} & a_{22} \end{bmatrix} = \dfrac{1}{\left\lvert \bm{A} \right\rvert} \begin{bmatrix} a_{22} & -a_{12} \\ -a_{21} & a_{11} \end{bmatrix}

Case 1: A=1\left\lvert \bm{A} \right\rvert = 1

从而

{a11=a22a21=a12\left\lbrace\begin{aligned} a_{11} &= a_{22} \\ a_{21} &= -a_{12} \end{aligned}\right.

A=a11a22a12a21=a112+a212=1\begin{aligned} \left\lvert \bm{A} \right\rvert &= a_{11} a_{22} - a_{12} a_{21} \\ &= a_{11}^2 + a_{21}^2 \\ &= 1 \end{aligned}

从而存在 θ[0,2π)\theta \in \left[ 0, 2\pi \right),使得

{a11=cosθa21=sinθ\left\lbrace\begin{aligned} a_{11} &= \cos \theta \\ a_{21} &= \sin \theta \end{aligned}\right.

A=[cosθsinθsinθcosθ]\bm{A} = \begin{bmatrix} \cos \theta & -\sin \theta \\ \sin \theta & \cos \theta \end{bmatrix}

A\bm{A} 代表的线性变换为绕原点逆时针旋转 θ\theta

Case 2: A=1\left\lvert \bm{A} \right\rvert = -1

从而

{a11=a22a21=a12\left\lbrace\begin{aligned} a_{11} &= -a_{22} \\ a_{21} &= a_{12} \end{aligned}\right.

A=a11a22a12a21=a112a212=1\begin{aligned} \left\lvert \bm{A} \right\rvert &= a_{11} a_{22} - a_{12} a_{21} \\ &= -a_{11}^2 - a_{21}^2 \\ &= -1 \end{aligned}

从而同样存在 θ[0,2π)\theta \in \left[ 0, 2\pi \right),使得

{a11=cosθa21=sinθ\left\lbrace\begin{aligned} a_{11} &= \cos \theta \\ a_{21} &= \sin \theta \end{aligned}\right.

A=[cosθsinθsinθcosθ]\bm{A} = \begin{bmatrix} \cos \theta & \sin \theta \\ \sin \theta & -\cos \theta \end{bmatrix}

A=[cosθsinθsinθcosθ][1001]\bm{A} = \begin{bmatrix} \cos \theta & - \sin \theta \\ \sin \theta & \cos \theta \end{bmatrix} \begin{bmatrix} 1 & 0 \\ 0 & -1 \end{bmatrix}

A\bm{A} 代表的线性变换为先关于 xx 轴对称(反射),再绕原点逆时针旋转 θ\theta

旋转矩阵似乎会改变全部向量的方向,也就是说,似乎找不到特征向量。实际上其特征值为复数,特征向量为复向量。

实际上有

{λ1=eiθλ2=eiθ\left\lbrace\begin{aligned} \lambda_1 &= \e ^{\i \theta} \\ \lambda_2 &= \e ^{-\i \theta} \end{aligned}\right.

A\bm{A}nn 阶正交矩阵(正交矩阵要求为实矩阵),λ\lambdaA\bm{A} 的特征值,α,βRn\bm{\alpha},\, \bm{\beta} \in \R^n,则

  1. A,Ak,A1,A\bm{A}^\intercal,\, \bm{A}^k,\, \bm{A}^{-1},\, \bm{A}^{*} 也是正交矩阵
  2. A2=1\left\lvert \bm{A} \right\rvert^2 = 1(即 A\bm{A} 的行列式为 ±1\pm 1
  3. (Aα,Aβ)=(α,β)(\bm{A} \bm{\alpha}, \bm{A} \bm{\beta}) = (\bm{\alpha}, \bm{\beta})(即 A\bm{A} 保持内积不变,若取 α=β\bm{\alpha} = \bm{\beta},则 A\bm{A} 保持长度不变,即 Aα=α\| \bm{A} \bm{\alpha} \| = \| \bm{\alpha} \|
  4. λ=1\left\lvert \lambda \right\rvert = 1(即 λ=eiθ\lambda = \e ^{\i \theta},实特征值只能为 ±1\pm 1

  1. 证明:

A,A1,A\bm{A}^\intercal,\, \bm{A}^{-1},\, \bm{A}^{*} 显然。

对于 Ak\bm{A}^k,有

(Ak)Ak=(AAAk)AAAk=AA(AkAA(AA)E(AA)AAA)AAk=E\begin{aligned} (\bm{A}^k)^\intercal \bm{A}^k &= (\overbrace{\bm{A} \bm{A} \cdots \bm{A}}^{k})^\intercal \overbrace{\bm{A} \bm{A} \cdots \bm{A}}^{k} \\ &= \rlap{$\overbrace{\phantom{\bm{A}^\intercal \bm{A}^\intercal \cdots (\bm{A}^\intercal}}^{k}$} \bm{A}^\intercal \bm{A}^\intercal \cdots \rlap{$\underbrace{(\bm{A}^\intercal \bm{A})}_{\bm{E}}$} \phantom{(\bm{A}^\intercal \bm{A})} \bm{A} \cdots \bm{A} \llap{$\overbrace{\phantom{\bm{A}) \bm{A} \cdots \bm{A}}}^{k}$}\\ &\kern{0.5em}\vdots \\ &= \bm{E} \end{aligned}

  1. 证明:

AA=E\bm{A}^\intercal \bm{A} = \bm{E}

A2=AA=E=1\left\lvert \bm{A} \right\rvert^2 = \lvert \bm{A}^\intercal \rvert \left\lvert \bm{A} \right\rvert = \left\lvert \bm{E} \right\rvert = 1

  1. 证明:

(Aα,Aβ)=(Aα)Aβ=αAAβ=αEβ=αβ=(α,β)\begin{aligned} (\bm{A} \bm{\alpha}, \bm{A} \bm{\beta}) &= (\bm{A} \bm{\alpha})^\intercal \bm{A} \bm{\beta} \\ &= \bm{\alpha}^\intercal \bm{A}^\intercal \bm{A} \bm{\beta} \\ &= \bm{\alpha}^\intercal \bm{E} \bm{\beta} \\ &= \bm{\alpha}^\intercal \bm{\beta} \\ &= (\bm{\alpha}, \bm{\beta}) \end{aligned}

那么同时还有

Aα,Aβ=arccos(Aα,Aβ)AαAβ=arccos(α,β)αβ=α,β\begin{aligned} \langle \bm{A} \bm{\alpha}, \bm{A} \bm{\beta} \rangle &= \arccos \dfrac{(\bm{A} \bm{\alpha}, \bm{A} \bm{\beta})}{\| \bm{A} \bm{\alpha} \| \| \bm{A} \bm{\beta} \|} \\ &= \arccos \dfrac{(\bm{\alpha}, \bm{\beta})}{\| \bm{\alpha} \| \| \bm{\beta} \|} \\ &= \langle \bm{\alpha}, \bm{\beta} \rangle \end{aligned}

  1. 证明:

Aη=λη\bm{A} \bm{\eta} = \lambda \bm{\eta},其中 ηCn\{θ}\bm{\eta} \in \C^n \backslash \left\lbrace \bm{\theta} \right\rbrace,则

(Aη,Aη)C=(Aη)Aη=ηAAη=ηAAηˉ=ηηˉ=η2\begin{aligned} \left( \bm{A} \bm{\eta}, \bm{A} \bm{\eta} \right)_{\C} &= \left( \bm{A} \bm{\eta} \right) ^\intercal \overline{\bm{A} \bm{\eta}} \\ &= \bm{\eta}^\intercal \bm{A}^\intercal \overline{\bm{A} \bm{\eta}} \\ &= \bm{\eta}^\intercal \bm{A}^\intercal \bm{A} \bar{\bm{\eta}} \\ &= \bm{\eta}^\intercal \bar{\bm{\eta}} \\ &= \| \bm{\eta} \| ^2 \end{aligned}

(λη,λη)C=(λη)λη=λλˉηηˉ=λ2η2\begin{aligned} \left( \lambda \bm{\eta}, \lambda \bm{\eta} \right)_{\C} &= (\lambda \bm{\eta})^\intercal \overline{\lambda \bm{\eta}} \\ &= \lambda \bar{\lambda} \bm{\eta}^\intercal \bar{\bm{\eta}} \\ &= \left\lvert \lambda \right\rvert^2 \| \bm{\eta} \| ^2 \end{aligned}

从而 λ2=1\left\lvert \lambda \right\rvert^2 = 1

考虑 Rn\R^n 空间中两个法正交组

A ⁣:α1,α2,,αn\bm{A}\colon \bm{\alpha}_1 ,\, \bm{\alpha}_2 ,\, \cdots ,\, \bm{\alpha}_n

B ⁣:β1,β2,,βn\bm{B}\colon\bm{\beta}_1 ,\, \bm{\beta}_2 ,\, \cdots ,\, \bm{\beta}_n

则存在矩阵 C\bm{C},使得

B=AC[β1β2βn]=[α1α2αn][c11c12c1nc21c22c2ncn1cn2cnn]\begin{aligned} \bm{B} &= \bm{A} \bm{C}\\ \begin{bmatrix} \bm{\beta}_1 & \bm{\beta}_2 & \cdots & \bm{\beta}_n \end{bmatrix} &= \begin{bmatrix} \bm{\alpha}_1 & \bm{\alpha}_2 & \cdots & \bm{\alpha}_n \end{bmatrix} \begin{bmatrix} c_{11} & c_{12} & \cdots & c_{1n} \\ c_{21} & c_{22} & \cdots & c_{2n} \\ \vdots & \vdots & \ddots & \vdots \\ c_{n1} & c_{n2} & \cdots & c_{nn} \end{bmatrix} \end{aligned}

那么

C=A1B\bm{C} = \bm{A}^{-1} \bm{B}

C\bm{C} 也是正交矩阵。

实对称矩阵的对角化

A\bm{A}nn 阶实方阵,且 A=A\bm{A}^\intercal = \bm{A},则称 A\bm{A}实对称矩阵

实对称矩阵的特征值一定为实数


证明:

对实对称矩阵 A\bm{A},设其一特征值为 λ\lambda,则存在特征向量 ηCn\{θ}\bm{\eta} \in \C^n\backslash \left\lbrace \bm{\theta} \right\rbrace,即 Aη=λη\bm{A} \bm{\eta} = \lambda \bm{\eta},等价于证明 λˉ=λ\bar{\lambda} = \lambda

{Aη=λη(Aη)=(λη)    {Aηˉ=λˉηˉηA=λη\left\lbrace\begin{aligned} \overline{\bm{A}\bm{\eta}} &= \overline{\lambda \bm{\eta}}\\ (\bm{A} \bm{\eta})^\intercal &= (\lambda \bm{\eta})^\intercal \end{aligned}\right. \implies \left\lbrace\begin{aligned} \bm{A} \bar{\bm{\eta}} &= \bar{\lambda} \bar{\bm{\eta}}\\ \bm{\eta}^\intercal \bm{A} &= \lambda \bm{\eta}^\intercal \end{aligned}\right.

两式分别左乘 η\bm{\eta}^\intercal 和右乘 ηˉ\bar{\bm{\eta}},得

{ηAηˉ=λˉηηˉ=λˉη2ηAηˉ=ληηˉ=λη2\left\lbrace\begin{aligned} \bm{\eta}^\intercal \bm{A} \bar{\bm{\eta}} &= \bar{\lambda} \bm{\eta}^\intercal \bar{\bm{\eta}} &= \bar{\lambda} \| \bm{\eta} \|^2 \\ \bm{\eta}^\intercal \bm{A} \bar{\bm{\eta}} &= \lambda \bm{\eta}^\intercal \bar{\bm{\eta}} &= \lambda \| \bm{\eta} \|^2 \end{aligned}\right.

从而

λˉη2=λη2\bar{\lambda} \| \bm{\eta} \|^2 = \lambda \| \bm{\eta} \|^2

也即

λˉ=λ\bar{\lambda} = \lambda

对于 nn 阶实对称矩阵 A\bm{A} 与任意向量 α,βRn\bm{\alpha},\, \bm{\beta} \in \R^n,有

(Aα,β)=(α,Aβ)(\bm{A} \bm{\alpha}, \bm{\beta}) = (\bm{\alpha}, \bm{A} \bm{\beta})


证明:

(Aα,β)=(Aα)β=αAβ=αAβ=(α,Aβ)\begin{aligned} (\bm{A} \bm{\alpha}, \bm{\beta}) &= (\bm{A} \bm{\alpha})^\intercal \bm{\beta} \\ &= \bm{\alpha}^\intercal \bm{A}^\intercal \bm{\beta} \\ &= \bm{\alpha}^\intercal \bm{A} \bm{\beta} \\ &= (\bm{\alpha}, \bm{A} \bm{\beta}) \end{aligned}

实对称矩阵属于不同特征值的特征向量必正交。

即对实对称矩阵 A\bm{A},若 λ1,λ2\lambda_1,\, \lambda_2 为不同特征值,η1,η2\bm{\eta}_1,\, \bm{\eta}_2 为对应的特征向量,则 (η1,η2)=0(\bm{\eta}_1, \bm{\eta}_2) = 0


由上面有 (Aη1,η2)=(η1,Aη2)(\bm{A} \bm{\eta}_1, \bm{\eta}_2) = (\bm{\eta}_1, \bm{A} \bm{\eta}_2)

从而 (λ1η1,η2)=(η1,λ2η2)(\lambda_1 \bm{\eta}_1, \bm{\eta}_2) = (\bm{\eta}_1, \lambda_2 \bm{\eta}_2),即 (λ1η1,η2)=(λ2η1,η2)(\lambda_1 \bm{\eta}_1, \bm{\eta}_2) = (\lambda_2 \bm{\eta}_1, \bm{\eta}_2),从而 (λ1λ2)(η1,η2)=0(\lambda_1 - \lambda_2) (\bm{\eta}_1, \bm{\eta}_2) = 0

λ1λ2\lambda_1 \neq \lambda_2,从而 (η1,η2)=0(\bm{\eta}_1, \bm{\eta}_2) = 0

实对称矩阵必可对角化。

对于实对称矩阵 AMn(R)\bm{A} \in M_n(\R),存在正交矩阵 P\bm{P},使得 P1AP=Λ\bm{P}^{-1} \bm{A} \bm{P} = \bm{\Lambda},其中 Λ\bm{\Lambda} 为对角矩阵。


证明:

A\bm{A} 有特征值 λ1,λ2,,λnR\lambda_1,\, \lambda_2,\, \cdots,\, \lambda_n \in \R

λ1\lambda_1,存在单位向量 α1Rn\{θ}\bm{\alpha}_1 \in \R^n\backslash\left\lbrace \bm{\theta} \right\rbrace 使得 Aα1=λ1α1,α1=1\bm{A} \bm{\alpha}_1 = \lambda_1 \bm{\alpha}_1,\, \|\bm{\alpha}_1\| = 1

总能将其扩充为 Rn\R^n 的一组法正交组 P1 ⁣:α1,α2,,αn\bm{P}_1\colon\bm{\alpha}_1,\, \bm{\alpha}_2,\, \cdots,\, \bm{\alpha}_n(总能先扩充为一个向量组⟦上面谈过证明。由于提及的锚点问题,不设置锚点跳转,因为设置也跳转不对⟧,再使用施密特正交化,并单位化)。

P1=[α1α2αn]\bm{P}_1 = \begin{bmatrix} \bm{\alpha}_1 & \bm{\alpha}_2 & \cdots & \bm{\alpha}_n \end{bmatrix} 为正交矩阵。

从而有

AP1=A[α1α2αn]=[Aα1Aα2Aαn]=[λ1α1Aα2Aαn]=[α1α2αn][λ1a2an00]=P1[λ1a2an00A1]=P1Q1\begin{aligned} \bm{A} \bm{P}_1 &= \bm{A} \begin{bmatrix} \bm{\alpha}_1 & \bm{\alpha}_2 & \cdots & \bm{\alpha}_n \end{bmatrix} \\ &= \begin{bmatrix} \bm{A} \bm{\alpha}_1 & \bm{A} \bm{\alpha}_2 & \cdots & \bm{A} \bm{\alpha}_n \end{bmatrix} \\ &= \begin{bmatrix} \lambda_1 \bm{\alpha}_1 & \bm{A} \bm{\alpha}_2 & \cdots & \bm{A} \bm{\alpha}_n \end{bmatrix} \\ &= \begin{bmatrix} \bm{\alpha}_1 & \bm{\alpha}_2 & \cdots & \bm{\alpha}_n \end{bmatrix} \begin{bmatrix} \lambda_1 & a_2 & \cdots & a_n \\ 0 & * & \cdots & * \\ \vdots & \vdots & \ddots & \vdots \\ 0 & * & \cdots & * \end{bmatrix} \\ &= \bm{P}_1 \begin{bmatrix} \lambda_1 & \begin{matrix} a_2 & \cdots & a_n \end{matrix}\\ \begin{matrix} 0 \\ \vdots \\ 0 \end{matrix} & \LARGE\bm{A}_1 \end{bmatrix}\\ &= \bm{P}_1 \bm{Q}_1 \end{aligned}

Q1=P11AP1=P1AP1\begin{aligned} \bm{Q}_1 &= \bm{P}_1^{-1} \bm{A} \bm{P}_1 \\ &= \bm{P}_1^\intercal \bm{A} \bm{P}_1 \end{aligned}

Q1=(P1AP1)=P1AP1=P1AP1=Q1\begin{aligned} \bm{Q}_1^\intercal &= \left( \bm{P}_1^\intercal \bm{A} \bm{P}_1 \right)^\intercal \\ &= \bm{P}_1^\intercal \bm{A}^\intercal \bm{P}_1 \\ &= \bm{P}_1^\intercal \bm{A} \bm{P}_1 \\ &= \bm{Q}_1 \end{aligned}

Q1\bm{Q}_1 也为实对称矩阵,从而 a2=a3==an=0a_2 = a_3 = \cdots = a_n = 0,且 A1\bm{A}_1n1n - 1 阶实对称矩阵,特征值为 λ2,λ3,,λn\lambda_2,\, \lambda_3,\, \cdots,\, \lambda_n

以此类推,存在正交矩阵 P1,P2,,Pn\bm{P}_1,\, \bm{P}_2,\, \cdots,\, \bm{P}_n 使

PnP2P1AP1P2Pn=Λ\bm{P}_n^\intercal \cdots \bm{P}_2^\intercal \bm{P}_1^\intercal \bm{A} \bm{P}_1 \bm{P}_2 \cdots \bm{P}_n = \bm{\Lambda}

P=P1P2Pn\bm{P} = \bm{P}_1 \bm{P}_2 \cdots \bm{P}_n

显然 P\bm{P} 也是正交矩阵。则

PAP=Λ\bm{P}^\intercal \bm{A} \bm{P} = \bm{\Lambda}

也即

P1AP=Λ\bm{P}^{-1} \bm{A} \bm{P} = \bm{\Lambda}