矩阵计算选做

正定矩阵Gauss消元

设$A$是对称正定矩阵,经过Gauss消去法一步后,$A$约化为$\left(\begin{array}{cc}{a}_{11}& a_1^T\\ 0& A_2\end{array}\right)$,其中
$A=(a_{ij}{)}_n$, $A_2=(a_{ij}^{(2)}{)}_{n-1}$.
(1) $A$的对角元素$a_{ii}>0(i=1,2,\dots ,n)$;
(2) $A_2$是对称正定矩阵;
(3) $a_{ii}^{(2)}⩽a_{ii}(i=2,3,\cdots ,n)$;
(4) $A$的绝对值最大的元素必在对角线上;
(5) $\underset{2⩽i,j⩽n}{\max }|a_{ij}^{(2)}|⩽\underset{2⩽i,j⩽n}{\max }|a_{ij}|$;
(6) 从(2),(3),(5)推出,如果$|a_{ij}|<1$,则对于所有$k$,$|a_{ij}^{(k)}|<1$.

注意记号: 如果上标(k)则指明是“第k-1次Gauss消元之后”“第k次Gauss消元之前”的矩阵/向量/元素.
例如上标(2)指明是“第1次Gauss消元之后”“第2次Gauss消元之前”的矩阵/向量/元素.
除了题目中定义的符号, 如果不加上标则默认上标是(1), “第1次Gauss消元之前”即的矩阵/向量/元素.

$A=A^T$, $A\succcurlyeq 0$.
$A^{(1)}=A$
$A^{(2)}=G^{(1)}{A}^{(1)}$

$A^{(1)}=\left(\begin{array}{cc}{a}_{11}^{(1)}& {a_1^{(1)}}^T\\ a_1^{(1)}& A_2^{(1)}\end{array}\right)$
$G^{(1)}=\left(\begin{array}{ccccc}1& 0& 0& & 0\\ -\frac{a_{21}}{a_{11}}& 1& 0& & 0\\ -\frac{a_{31}}{a_{11}}& 0& 1& & 0\\ ⋮& & & \ddots & \\ -\frac{a_{n1}}{a_{11}}& 0& 0& & 1\end{array}\right)=\left(\begin{array}{cc}1& 0^T\\ g^{(1)}& I\end{array}\right)$

(1) $A$的对角元素$a_{ii}>0(i=1,2,\dots ,n)$;

取基向量$e_i$, $(e_i{)}_i=1$, $(e_i{)}_j=0(j\ne i)$.
由正定性$A_{ii}=e_i^{T}A{e}_i>0$.

$A_{ii}=e_i^{T}A{e}_i$是矩阵乘法的展开.
$e_i^{T}A{e}_i>0$是正定性的推论.

(2) $A_2$是对称正定矩阵;

$\left(\begin{array}{cc}{a}_{11}& a_1^T\\ 0& A_2\end{array}\right)=A^{(2)}=G^{(1)}{A}^{(1)}=\left(\begin{array}{cc}1& 0^T\\ g^{(1)}& I\end{array}\right)\left(\begin{array}{cc}{a}_{11}^{(1)}& {a_1^{(1)}}^T\\ a_1^{(1)}& A_2^{(1)}\end{array}\right)$
$A_2=A_2^{(1)}+g^{(1)}{a_1^{(1)}}^T=A_2^{(1)}-\frac{1}{a_{11}^{(1)}}{a}_1^{(1)}{a_1^{(1)}}^T$

对称性
由于$A^{(1)}$对称, 所以$A_2^{(1)}$对称.
显然$a_1^{(1)}{a_1^{(1)}}^T$对称.
由于对称矩阵的线性组合仍是对称矩阵, 所以$A_2=A_2^{(1)}-\frac{1}{a_{11}^{(1)}}{a}_1^{(1)}{a_1^{(1)}}^T$是对称矩阵.

正定性
要证明$A_2\succ 0$, 只需证明对于任意的向量$v\ne 0$有$v^T{A}_{2}v>0$.
$v^T{A}_{2}v=v^T{A}_2^{(1)}v-\frac{1}{a_{11}^{(1)}}{\left(v^T{a}_1^{(1)}\right)}^2$
构造$\tilde{v}=\left(\begin{array}{c}-\frac{1}{a_{11}^{(1)}}\left(v^T{a}_1^{(1)}\right)\\ v\end{array}\right)$.
${\tilde{v}}^T{A}^{(1)}\tilde{v}=a_{11}^{(1)}{\left(-\frac{1}{a_{11}^{(1)}}\left(v^T{a}_1^{(1)}\right)\right)}^2-2\frac{1}{a_{11}^{(1)}}\left(v^T{a}_1^{(1)}\right)v^T{a}_1^{(1)}+v^T{A}_2^{(1)}v=v^T{A}_2^{(1)}v-\frac{1}{a_{11}^{(1)}}{\left(v^T{a}_1^{(1)}\right)}^2=v^T{A}_{2}v$
由于$A^{(1)}\succ 0$, 所以${\tilde{v}}^T{A}^{(1)}\tilde{v}>0$, 亦即$v^T{A}_{2}v>0$.
注意到$v\ne 0$是任取的, 总可以通过构造$\tilde{v}$来证明$v^T{A}_{2}v>0$, 所以$A_2\succ 0$.

注意: $v\in {\mathbb{R}}^{n-1}$, $\tilde{v}\in {\mathbb{R}}^n$.
将n元二次型归纳为n-1元二次型: ${\left(\begin{array}{c}x\\ \mathbf{y}\end{array}\right)}^T\left(\begin{array}{cc}\mu & {\mathbf{m}}^T\\ \mathbf{m}& \mathbf{M}\end{array}\right)\left(\begin{array}{c}x\\ \mathbf{y}\end{array}\right)=\mu x^2+2x{\mathbf{m}}^T\mathbf{y}+{\mathbf{y}}^T\mathbf{M}\mathbf{y}$
$x=-\frac{1}{a_{11}^{(1)}}\left(v^T{a}_1^{(1)}\right)$
取 $\mathbf{y}=v$
取 $\mu =a_{11}^{(1)}$
取 $\mathbf{m}=a_1^{(1)}$
取 $\mathbf{M}=A_2^{(1)}$
取 $\left(\begin{array}{c}x\\ \mathbf{y}\end{array}\right)=\tilde{v}$
取 $\left(\begin{array}{cc}\mu & {\mathbf{m}}^T\\ \mathbf{m}& \mathbf{M}\end{array}\right)=A^{(1)}$

(3) $a_{ii}^{(2)}⩽a_{ii}(i=2,3,\cdots ,n)$;

(2)中已经证明$A_2=A_2^{(1)}-\frac{1}{a_{11}^{(1)}}{a}_1^{(1)}{a_1^{(1)}}^T$.
注意到$a_{ii}^{(2)}={\left(A_2\right)}_{ii}={\left(A_2^{(1)}\right)}_{ii}-\frac{1}{a_{11}^{(1)}}{\left(a_1^{(1)}{a_1^{(1)}}^T\right)}_{ii}=a_{ii}-\frac{1}{a_{11}^{(1)}}{a}_{i1}^2$.
(1)中已经证明$a_{11}^{(1)}>0$, 所以$a_{ii}^{(2)}=a_{ii}-\frac{1}{a_{11}^{(1)}}{a}_{i1}^2⩽a_{ii}$

(4) $A$的绝对值最大的元素必在对角线上;

由于对称矩阵是Hermite矩阵, 因此根据Hermite矩阵的性质知$A$①可对角化②特征值均为实数③特征向量正交, 进而有$A=Q\Lambda Q^T$, 其中$\Lambda =\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}({\lambda }_1,\cdots ,{\lambda }_n)$, $Q^{T}Q=Q{Q}^T=I$.
由于正定性, $\Lambda \succ 0$.

$\begin{array}{rl}A& =Q\Lambda Q^T\\ & =\left(\begin{array}{ccc}\uparrow & & \uparrow \\ q_1& \cdots & q_n\\ \downarrow & & \downarrow \end{array}\right)\left(\begin{array}{ccc}{\lambda }_1& & \\ & \ddots & \\ & & {\lambda }_n\end{array}\right)\left(\begin{array}{ccc}\leftarrow & q_1^T& \to \\ & ⋮& \\ \leftarrow & q_n^T& \to \end{array}\right)\\ & =\left(\begin{array}{ccc}\uparrow & & \uparrow \\ q_1& \cdots & q_n\\ \downarrow & & \downarrow \end{array}\right)\left(\begin{array}{ccc}\uparrow & & \uparrow \\ {\lambda }_1{e}_1& \cdots & {\lambda }_n{e}_n\\ \downarrow & & \downarrow \end{array}\right)\left(\begin{array}{ccc}\leftarrow & q_1^T& \to \\ & ⋮& \\ \leftarrow & q_n^T& \to \end{array}\right)\\ & =\left(\begin{array}{ccc}\uparrow & & \uparrow \\ q_1& \cdots & q_n\\ \downarrow & & \downarrow \end{array}\right)\left(\sum _{i=1}^n{\lambda }_i{e}_i{q}_i^T\right)\\ & =\left(\begin{array}{ccc}\uparrow & & \uparrow \\ q_1& \cdots & q_n\\ \downarrow & & \downarrow \end{array}\right)\left(\begin{array}{ccc}\leftarrow & {\lambda }_1{q}_1^T& \to \\ & ⋮& \\ \leftarrow & {\lambda }_n{q}_n^T& \to \end{array}\right)\end{array}$

将A的元素重写成 $A_{ij}=\sum _{k=1}^n{\lambda }_k{q}_{ki}{q}_{kj}$, ${\lambda }_k>0$.

根据柯西不等式, ${\left(\sum _{i=1}^n{\lambda }_k{q}_{ki}{q}_{kj}\right)}^2⩽\left(\sum _{i=1}^n{\left(\sqrt{{\lambda }_k}{q}_{ki}\right)}^2\right)\left(\sum _{i=1}^n{\left(\sqrt{{\lambda }_k}{q}_{kj}\right)}^2\right)=\left(\sum _{i=1}^n{\lambda }_k{q}_{ki}^2\right)\left(\sum _{i=1}^n{\lambda }_k{q}_{kj}^2\right)$, 即$A_{ij}^2⩽A_{ii}{A}_{jj}$.
(1)中已经证明$A_{ii},A_{jj}>0$, 因此 max ⁡ { ∣ A i i ∣ , ∣ A j j ∣ } = max ⁡ { A i i , A j j } = ( max ⁡ { A i i , A j j } ) 2 ⩾ A i i A j j = A i j 2 = ∣ A i j ∣ \max\{|A_{ii}|,|A_{jj}|\} = \max\{A_{ii},A_{jj}\} = \sqrt{\left(\max\{A_{ii},A_{jj}\}\right)^2} \geqslant \sqrt{A_{ii}A_{jj}} = \sqrt{A_{ij}^2} = |A_{ij}| max{∣Aii​∣,∣Ajj​∣}=max{Aii​,Ajj​}=(max{Aii​,Ajj​})2 ​⩾Aii​Ajj​ ​=Aij2​ ​=∣Aij​∣.
即 ∣ A i j ∣ ⩽ max ⁡ { ∣ A i i ∣ , ∣ A j j ∣ } |A_{ij}| \leqslant \max\{|A_{ii}|,|A_{jj}|\} ∣Aij​∣⩽max{∣Aii​∣,∣Ajj​∣}, 这样就证明了$A$的绝对值最大的元素必在对角线上.

[半]正定矩阵(semi positive definite matrix)主子式(principal minor)必然[半]正定.
以上证明过程可以简化为: 由于A正定, 因此二阶主子式$\left(\begin{array}{cc}{A}_{ii}& A_{ij}\\ A_{ij}& A_{jj}\end{array}\right)$正定, 故其行列式$A_{ii}{A}_{jj}-A_{ij}^2>0$.

(5) $\underset{2⩽i,j⩽n}{\max }|a_{ij}^{(2)}|⩽\underset{2⩽i,j⩽n}{\max }|a_{ij}|$;

由于$A$是正定矩阵, 因此主子式$A_2^{(1)}$是正定对称矩阵.
(2)中已经证明$A_2$是正定对称矩阵.
(4)中已经证明正定对称的绝对值最大的元素必在对角线上.

因此不妨假设$\underset{2⩽i,j⩽n}{\max }|a_{ij}^{(2)}|=\underset{2⩽i⩽n}{\max }{a}_{ii}^{(2)}=a_{xx}^{(2)}$, $\underset{2⩽i,j⩽n}{\max }|a_{ij}|=\underset{2⩽i⩽n}{\max }{a}_{ii}=a_{yy}$, $2⩽x,y⩽n$.

(3)中已经证明$a_{ii}^{(2)}⩽a_{ii}(i=2,3,\cdots ,n)$.
因此$\underset{2⩽i,j⩽n}{\max }|a_{ij}^{(2)}|=a_{xx}^{(2)}⩽a_{xx}⩽\underset{2⩽i⩽n}{\max }{a}_{ii}=\underset{2⩽i,j⩽n}{\max }|a_{ij}|$.

(6) 从(2),(3),(5)推出,如果$|a_{ij}|<1$,则对于所有$k$,$|a_{ij}^{(k)}|<1$.

根据(5)知$\underset{2⩽i,j⩽n}{\max }|a_{ij}^{(2)}|⩽\underset{2⩽i,j⩽n}{\max }|a_{ij}^{(1)}|⩽\underset{1⩽i,j⩽n}{\max }|a_{ij}^{(1)}|$推广到$\underset{k⩽i,j⩽n}{\max }|a_{ij}^{(k)}|⩽\underset{k⩽i,j⩽n}{\max }|a_{ij}^{(k-1)}|⩽\underset{k-1⩽i,j⩽n}{\max }|a_{ij}^{(k-1)}|$.

对于$k⩽i,j⩽n$, $|a_{ij}^{(k)}|⩽\underset{k⩽i,j⩽n}{\max }|a_{ij}^{(k)}|⩽\underset{k-1⩽i,j⩽n}{\max }|a_{ij}^{(k-1)}|⩽\cdots ⩽\underset{1⩽i,j⩽n}{\max }|a_{ij}^{(1)}|=\underset{1⩽i,j⩽n}{\max }|a_{ij}|<1$.
对于$i<k\vee j<k$, 注意到 a i j ( k ) = a i j ( k − 1 ) = ⋯ = a i j ( min ⁡ { i , j } ) a_{ij}^{(k)} = a_{ij}^{(k-1)} = \cdots = a_{ij}^{(\min\{i,j\})} aij(k)​=aij(k−1)​=⋯=aij(min{i,j})​.
∣ a i j ( k ) ∣ = ∣ a i j ( k − 1 ) ∣ = ⋯ = ∣ a i j ( min ⁡ { i , j } ) ∣ ⩽ max ⁡ min ⁡ { i , j } ⩽ i , j ⩽ n ∣ a i j ( min ⁡ { i , j } ) ∣ ⩽ max ⁡ min ⁡ { i , j } − 1 ⩽ i , j ⩽ n ∣ a i j ( min ⁡ { i , j } − 1 ) ∣ ⩽ ⋯ ⩽ max ⁡ 1 ⩽ i , j ⩽ n ∣ a i j ( 1 ) ∣ = max ⁡ 1 ⩽ i , j ⩽ n ∣ a i j ∣ < 1 |a_{ij}^{(k)}| = |a_{ij}^{(k-1)}| = \cdots = |a_{ij}^{(\min\{i,j\})}| \leqslant \max\limits_{\min\{i,j\} \leqslant i,j \leqslant n} |a_{ij}^{(\min\{i,j\})}| \leqslant \max\limits_{\min\{i,j\}-1 \leqslant i,j \leqslant n} |a_{ij}^{(\min\{i,j\}-1)}| \leqslant \cdots \leqslant \max\limits_{1 \leqslant i,j \leqslant n} |a_{ij}^{(1)}| = \max\limits_{1 \leqslant i,j \leqslant n} |a_{ij}| < 1 ∣aij(k)​∣=∣aij(k−1)​∣=⋯=∣aij(min{i,j})​∣⩽min{i,j}⩽i,j⩽nmax​∣aij(min{i,j})​∣⩽min{i,j}−1⩽i,j⩽nmax​∣aij(min{i,j}−1)​∣⩽⋯⩽1⩽i,j⩽nmax​∣aij(1)​∣=1⩽i,j⩽nmax​∣aij​∣<1.

综上所述, 对于$1⩽i,j⩽n$, $|a_{ij}^{(k)}|<1$.

追赶法

$\left(\begin{array}{ccccc}{b}_1& c_1& & & \\ a_2& b_2& c_2& & \\ & \ddots & \ddots & \ddots & \\ & & a_{n-1}& b_{n-1}& c_{n-1}\\ & & & a_n& b_n\end{array}\right)=\begin{array}{cc}& \left(\begin{array}{ccccc}1& {\beta }_1& & & \\ & 1& {\beta }_2& & \\ & & \ddots & \ddots & \\ & & & 1& {\beta }_{n-1}\\ & & & & 1\end{array}\right)\\ \left(\begin{array}{ccccc}{\alpha }_1& & & & \\ {\gamma }_2& {\alpha }_2& & & \\ & \ddots & \ddots & & \\ & & {\gamma }_{n-1}& {\alpha }_{n-1}& \\ & & & {\gamma }_n& {\alpha }_n\end{array}\right)& \end{array}$

$A=LU⟹\{\begin{array}{l}{a}_i={\gamma }_i\\ b_i={\alpha }_i+{\beta }_{i-1}{\gamma }_i\\ c_i={\alpha }_i{\beta }_i\end{array}$

$Ly=f⟹f_i={\alpha }_i{y}_i+{\gamma }_i{y}_{i-1}$

$Ux=y⟹y_i=x_i+{\beta }_i{x}_{i+1}$

$a_1=a_{n+1}=b_0=b_{n+1}=c_0=c_n={\alpha }_0={\alpha }_{n+1}={\beta }_0={\beta }_n={\gamma }_1={\gamma }_{n+1}=0$

追循环(第1轮)
${\alpha }_1=b_1$
${\alpha }_i=b_i-{\beta }_{i-1}{a}_i$ ($2⩽i⩽n$)
${\beta }_i=c_i/{\alpha }_i$ ($1⩽i⩽n-1$)
追循环(第2轮)
$y_1=f_1/{\alpha }_1$
$y_i=\left(f_i-a_i{y}_{i-1}\right)/{\alpha }_i$ ($2⩽i⩽n$)
赶循环
$x_n=y_n$
$x_i=y_i-{\beta }_i{x}_{i+1}$ ($1⩽i⩽n-1$)

显然$\gamma$只是用于推导的符号, 不具备实际意义(不具备自由度).
如果要节省内存空间, 可以不储存$\alpha$.

A = sym([
  2 -1  0  0  0;
 -1  2 -1  0  0;
  0 -1  2 -1  0;
  0  0 -1  2 -1;
  0  0  0 -1  2;
]);
f = sym([
  1;
  0;
  0;
  0;
  0;
])
n = 5;

a = @(i) A(i,i-1);
b = @(i) A(i,i);
c = @(i) A(i,i+1);

alpha = sym(zeros(n,1));
beta = sym(zeros(n-1,1));

y = sym(zeros(n,1));
x = sym(zeros(n,1));

% alpha, beta, 1 -> n
alpha(1) = b(1);
for i = 2:n
    beta(i-1) = c(i-1) ./ alpha(i-1);
    alpha(i) = b(i) - beta(i-1) .* a(i);
end

% y 1 -> n
y(1) = f(1) ./ alpha(1);
for i = 2:n
    y(i) = (f(i) - a(i) .* y(i-1)) ./ alpha(i);
end

% x n -> 1
x(n) = y(n);
for j = 1:n-1
    i = n - j;
    x(i) = y(i) - beta(i) .* x(i+1);
end

disp(x);

矩阵算子范数

L1算子范数证明:
$\Vert Ax{\Vert }_1=\sum _{i=1}^n\left|\sum _{j=1}^n{a}_{ij}{x}_j\right|⩽\sum _{j=1}^n|x_j|\sum _{i=1}^n|a_{ij}|⩽\sum _{j=1}^n|x_j|\left(\underset{j}{\max }\sum _{i=1}^n|a_{ij}|\right)=\Vert x{\Vert }_1\left(\underset{j}{\max }\sum _{i=1}^n|a_{ij}|\right)$
取等条件 $x_{j_0}=\pm \Vert x{\Vert }_1$, $x_{\mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}}=0$, $j_0=\arg \underset{j}{\max }\sum _{i=1}^n|a_{ij}|$.

L2算子范数证明:
见下一部分.

L∞算子范数证明:
$\Vert Ax{\Vert }_{+\infty }=\underset{i}{\max }\left|\sum _{j=1}^n{a}_{ij}{x}_j\right|⩽\underset{i}{\max }\sum _{j=1}^n|a_{ij}||x_j|⩽\underset{j}{\max }|x_j|\underset{i}{\max }\sum _{j=1}^n|a_{ij}|=\Vert x{\Vert }_{+\infty }\underset{i}{\max }\sum _{j=1}^n|a_{ij}|$
取等条件 $x_i=\mathrm{s}\mathrm{i}\mathrm{g}\mathrm{n}(a_{i_{0}j})\Vert x{\Vert }_{+\infty }$ 或 $x_i=-\mathrm{s}\mathrm{i}\mathrm{g}\mathrm{n}(a_{i_{0}j})\Vert x{\Vert }_{+\infty }$, $i_0=\arg \underset{i}{\max }\sum _{j=1}^n|a_{ij}|$.

$\Vert A{\Vert }_F=\sum _{i=1}^n\sum _{j=1}^n{a}_{ij}^2$
$\Vert A{\Vert }_1=\underset{\Vert x{\Vert }_1=1}{\sup }\Vert Ax{\Vert }_1=\underset{j}{\max }\sum _{i=1}^n|a_{ij}|$
$\Vert A{\Vert }_2=\underset{\Vert x{\Vert }_2=1}{\sup }\Vert Ax{\Vert }_2=|\sigma {|}_{\max }(A)$
$\Vert A{\Vert }_{+\infty }=\underset{\Vert x{\Vert }_{+\infty }=1}{\sup }\Vert Ax{\Vert }_{+\infty }=\underset{i}{\max }\sum _{j=1}^n|a_{ij}|$

% F
nF = sqrt(sum(A(:).^2))
assert(nF == norm(A,'fro'));

% 1
n1 = max(sum(abs(A)))
assert(n1 == norm(A,1));

% inf
nInf = max(sum(abs(A')))
assert(nInf == norm(A,inf));

% 2
n2 = max(svd(A))
assert(n2 == norm(A,2));

向量矩阵范数等价性

证明

1. $\Vert x{\Vert }_{+\infty }⩽\Vert x{\Vert }_1⩽n\Vert x{\Vert }_{+\infty }$
2. $\frac{1}{\sqrt{n}}\Vert A{\Vert }_F⩽\Vert A{\Vert }_2⩽\Vert A{\Vert }_F$

$\Vert x{\Vert }_{+\infty }=\underset{1⩽i⩽n}{\max }|a_i|⩽\sum _{1⩽i⩽n}|a_i|=\Vert x{\Vert }_1⩽\sum _{1⩽i⩽n}\left(\underset{1⩽i⩽n}{\max }|a_i|\right)=n\underset{1⩽i⩽n}{\max }|a_i|=n\Vert x{\Vert }_{+\infty }$

$A=U\Sigma V^H$

为什么选用奇异值分解而非特征值分解?

1. 不是所有矩阵的特征值分解都可以写成$A=Q\Lambda Q^{-1}$(考虑Jordan标准型, 只有代数重数等于几何重数的矩阵才可对角化)
2. 不是所有可对角化矩阵的特征值分解都可以写成$A=Q\Lambda Q^T$(只有归正矩阵$A{A}^H=A^{H}A$才能保证特征向量正交)

$\Vert A{\Vert }_F=\sqrt{\mathrm{t}\mathrm{r}(A^{H}A)}=\sqrt{\mathrm{t}\mathrm{r}(V{\Sigma }^2{V}^H)}=\sqrt{\mathrm{t}\mathrm{r}({\Sigma }^{2}V{V}^H)}=\sqrt{\mathrm{t}\mathrm{r}({\Sigma }^{2}I)}=\sqrt{\mathrm{t}\mathrm{r}({\Sigma }^2)}=\sqrt{\sum _{1⩽i⩽n}{\sigma }_i(A^{H}A)}=\sqrt{\sum _{1⩽i⩽n}{\sigma }_i^2(A)}$
$\Vert A{\Vert }_2=\underset{\Vert x{\Vert }_2=1}{\sup }\Vert Ax{\Vert }_2=\underset{\Vert x{\Vert }_2=1}{\sup }\Vert U\Sigma V^{H}x{\Vert }_2\stackrel{z=V^{H}x}{=}\underset{\Vert z{\Vert }_2=1}{\sup }\Vert U\Sigma z{\Vert }_2=\underset{\Vert z{\Vert }_2=1}{\sup }\sqrt{\sum _{i=1}^n{\sigma }_i^2(A)⟨u_i,z{⟩}^2}=\sqrt{{\sigma }_{\max }^2(A)}=|\sigma {|}_{\max }(A)$

其中$u_i$是$U$的第$i$行, 由于$U$是酉矩阵, 因此 { u i } i = 1 n \{u_i\}_{i=1}^{n} {ui​}i=1n​也构成一组正交基.
其中$⟨x,y⟩$是两个$n$维向量内积, 忽略行向量和列向量的格式差异.
由于 { u i } i = 1 n \{u_i\}_{i=1}^{n} {ui​}i=1n​构成一组正交基且$\Vert z{\Vert }_2=1$, 由Parseval定理知$\sum _{i=1}^n⟨u_i,z{⟩}^2=\Vert z{\Vert }_2^2=1$.

$\frac{1}{\sqrt{n}}\Vert A{\Vert }_F=\sqrt{\frac{1}{n}\sum _{1⩽i⩽n}{\sigma }_i^2(A)}⩽\sqrt{\frac{1}{n}\sum _{1⩽i⩽n}{\sigma }_{\max }^2(A)}=\sqrt{{\sigma }_{\max }^2(A)}=\Vert A{\Vert }_2⩽\sqrt{\sum _{1⩽i⩽n}{\sigma }_i^2(A)}=\Vert A{\Vert }_F$

矩阵算子范数下的条件数

设A,B∈ℝ^n×n且∥·∥为ℝ^n×n上矩阵的算子范数,证明cond(AB)⩽cond(A)cond(B).

假设$A$,$B$均非奇异,
$\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{d}(A)\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{d}(B)=\Vert A\Vert \Vert A^{-1}\Vert \Vert B\Vert \Vert B^{-1}\Vert$
$\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{d}(AB)=\Vert AB\Vert \Vert B^{-1}{A}^{-1}\Vert$
即证$\Vert AB\Vert \Vert B^{-1}{A}^{-1}\Vert ⩽(\Vert A\Vert \Vert B\Vert )(\Vert B^{-1}\Vert \Vert A^{-1}\Vert )$

我们证明如下更强的结论, 假设$X$,$Y$均非奇异, $\Vert XY\Vert ⩽\Vert X\Vert \Vert Y\Vert$
$\Vert XY\Vert =\underset{w\ne 0}{\sup }\frac{\Vert XYw\Vert }{\Vert w\Vert }=\underset{w\ne 0}{\sup }\frac{\Vert XYw\Vert }{\Vert Yw\Vert }\frac{\Vert Yw\Vert }{\Vert w\Vert }⩽\underset{w\ne 0}{\sup }\frac{\Vert XYw\Vert }{\Vert Yw\Vert }\underset{v\ne 0}{\sup }\frac{\Vert Yv\Vert }{\Vert v\Vert }\stackrel{u=Yw}{=}\underset{v\ne 0}{\sup }\frac{\Vert Yv\Vert }{\Vert v\Vert }\underset{u\ne 0}{\sup }\frac{\Vert Xu\Vert }{\Vert u\Vert }=\Vert Y\Vert \Vert X\Vert$

注意$w\ne 0$当且仅当$u=Yw\ne 0$, 原因是$Y$非奇异.