矩阵范数学习

矩阵范数

定义

设 A ∈ C m × n A\in \mathbb{C}^{m\times n} A∈Cm×n,按某一法则在 C m × n \mathbb{C}^{m\times n} Cm×n上规定 A A A的一个实值函数,记作 ∥ A ∥ \Vert A \Vert ∥A∥,它满足下面4个条件:
(1)非负性:如果 A ≠ 0 A\neq 0 A​=0,则 ∥ A ∥ > 0 \Vert A \Vert>0 ∥A∥>0;如果 A = 0 A=0 A=0,则 ∥ A ∥ = 0 \Vert A \Vert=0 ∥A∥=0
(2)齐次性:对于任意的 k ∈ C , ∥ k A ∥ = ∣ k ∣ ∥ A ∥ k \in \mathbb{C}, \Vert kA \Vert=\left| k \right| \Vert A \Vert k∈C,∥kA∥=∣k∣∥A∥
(3)三角不等式: ∀ A , B ∈ C m × n , ∥ A + B ∥ ≤ ∥ A ∥ ∥ B ∥ \forall A,B \in \mathbb{C}^{m\times n}, \Vert A+B \Vert \le \Vert A \Vert \Vert B \Vert ∀A,B∈Cm×n,∥A+B∥≤∥A∥∥B∥
(4)次乘性:当矩阵乘积 A B AB AB有意义时,若有
∥ A B ∥ ≤ ∥ A ∥ ∥ B ∥ \Vert AB \Vert \le \Vert A \Vert \Vert B \Vert ∥AB∥≤∥A∥∥B∥

则称 ∥ A ∥ \Vert A \Vert ∥A∥为矩阵范数

(如果次乘性的不等号反向,则幂等矩阵的矩阵范数为0,与非负性矛盾;
次乘性保证了矩阵幂级数的敛散性的“合理性”)

常用的矩阵范数

设 A ∈ C m × n A\in C^{m\times n} A∈Cm×n
∥ A ∥ m 1 = ∑ i = 1 m ∑ i = 1 n ∣ a i j ∣ ∥ A ∥ m ∞ = n ⋅ max ⁡ i , j ∣ a i j ∣ ∥ A ∥ F = ∥ A ∥ m 2 = ( ∑ i = 1 m ∑ i = 1 n ∣ a i j 2 ∣ 2 ) 1 2 \Vert A \Vert_{m_1}=\sum_{i=1}^{m}\sum_{i=1}^{n}\left|a_{ij}\right|\\ \Vert A \Vert_{m_\infty}=n\cdot \max \limits_{i,j}\left|a_{ij}\right|\\ \Vert A \Vert_F=\Vert A \Vert_{m_2}=(\sum_{i=1}^{m}\sum_{i=1}^{n}\left| a_{ij}^2\right|^2)^{\frac{1}{2}} ∥A∥m1​​=i=1∑m​i=1∑n​∣aij​∣∥A∥m∞​​=n⋅i,jmax​∣aij​∣∥A∥F​=∥A∥m2​​=(i=1∑m​i=1∑n​∣∣​aij2​∣∣​2)21​

等价

设 A ∈ C m × n A\in \mathbb{C}^{m\times n} A∈Cm×n, ∥ A ∥ \Vert A \Vert ∥A∥是 C m × n \mathbb{C}^{m\times n} Cm×n上的矩阵范数,则 C m × n \mathbb{C}^{m\times n} Cm×n上的任意两个矩阵范数等价

相容

设 A ∈ C m × n , x ∈ C n A\in \mathbb{C}^{m\times n},x \in \mathbb{C}^{n} A∈Cm×n,x∈Cn,如果取定的向量范数 ∥ x ∥ \Vert x \Vert ∥x∥和矩阵范数 ∥ A ∥ \Vert A\Vert ∥A∥满足
∥ A x ∥ ≤ ∥ A ∥ ∥ x ∥ \Vert Ax \Vert \le \Vert A \Vert\Vert x \Vert ∥Ax∥≤∥A∥∥x∥
则称矩阵范数 ∥ A ∥ \Vert A \Vert ∥A∥与向量范数 ∥ x ∥ \Vert x \Vert ∥x∥是相容

算子范数

设 A ∈ C m × n , x = ( x 1 , ⋯   , x n ) T ∈ C n A\in \mathbb{C}^{m\times n},x=(x_1,\cdots,x_n)^T \in \mathbb{C}^{n} A∈Cm×n,x=(x1​,⋯,xn​)T∈Cn,且在 C n \mathbb{C}^{n} Cn中已规定了向量的范数(即 C n \mathbb{C}^{n} Cn是 n n n维赋范线性空间),定义
∥ A ∥ = sup ⁡ ∥ x ∥ ≠ 0 ∥ A x ∥ ∥ x ∥ = max ⁡ ∥ x ∥ = 1 ∥ A x ∥ \Vert A \Vert = \sup \limits_{\Vert x \Vert \neq 0} \frac{\Vert Ax \Vert}{\Vert x \Vert}=\max \limits_{\Vert x \Vert =1}\Vert Ax \Vert ∥A∥=∥x∥​=0sup​∥x∥∥Ax∥​=∥x∥=1max​∥Ax∥
则上式定义了一个与向量范数 ∥ ⋅ ∥ \Vert \cdot \Vert ∥⋅∥相容的矩阵范数,称为向量范数 ∥ ⋅ ∥ \Vert \cdot \Vert ∥⋅∥诱导的矩阵范数或算子范数

证明:
需要证明这个矩阵范数满足4条性质以及相溶性

相溶性:
设 y ≠ 0 , x = 1 ∥ y ∥ y , ∥ x ∥ = 1 y\neq 0,x=\frac{1}{\Vert y \Vert} y,\Vert x \Vert =1 y​=0,x=∥y∥1​y,∥x∥=1
∥ A y ∥ = ∥ A ( ∥ y ∥ ) x ∥ = ∥ y ∥ ∥ A x ∥ ≤ ∥ y ∥ ∥ A ∥ = ∥ A ∥ ∥ y ∥ \begin{aligned} &\quad \Vert Ay \Vert \\ &= \Vert A(\Vert y \Vert) x \Vert\\ &= \Vert y \Vert \Vert Ax \Vert\\ &\le \Vert y \Vert \Vert A \Vert\\ &= \Vert A \Vert \Vert y \Vert \end{aligned} ​∥Ay∥=∥A(∥y∥)x∥=∥y∥∥Ax∥≤∥y∥∥A∥=∥A∥∥y∥​
非负性:
若 A ≠ 0 A\neq 0 A​=0,则可以找到 ∥ x ∥ = 1 \Vert x \Vert=1 ∥x∥=1的向量 x x x,使得 A x ≠ 0 Ax \neq 0 Ax​=0,从而 ∥ A x ∥ ≠ 0 \Vert Ax \Vert \neq 0 ∥Ax∥​=0
所以 ∥ A ∥ = max ⁡ ∥ x ∥ = 1 ∥ A x ∥ > 0 \Vert A \Vert=\max \limits_{\Vert x \Vert=1} \Vert Ax \Vert>0 ∥A∥=∥x∥=1max​∥Ax∥>0
当 A = 0 A=0 A=0,一定有 ∥ A ∥ = max ⁡ ∥ x ∥ = 1 ∥ 0 x ∥ = 0 \Vert A \Vert=\max \limits_{\Vert x \Vert=1} \Vert 0x \Vert=0 ∥A∥=∥x∥=1max​∥0x∥=0

齐次性:
对于 ∀ k ∈ C \forall k \in \mathbb{C} ∀k∈C,有
∥ k A ∥ = max ⁡ ∥ x ∥ = 1 ∥ k A x ∥ = ∣ k ∣ max ⁡ ∥ x ∥ = 1 ∥ A x ∥ = ∣ k ∣ ∥ A ∥ \Vert kA \Vert=\max \limits_{\Vert x \Vert=1} \Vert kAx \Vert = \left|k \right| \max \limits_{\Vert x \Vert=1} \Vert Ax \Vert=\left|k \right| \Vert A \Vert ∥kA∥=∥x∥=1max​∥kAx∥=∣k∣∥x∥=1max​∥Ax∥=∣k∣∥A∥

三角不等式:
对于矩阵 A + B A+B A+B,可以找到向量 x 0 x_0 x0​,使得
∥ A + B ∥ = ∥ ( A + B ) x 0 ∥ ( ∥ x 0 ∥ = 1 ) \Vert A+B \Vert= \Vert (A+B)x_0 \Vert \quad (\Vert x_0 \Vert=1) ∥A+B∥=∥(A+B)x0​∥(∥x0​∥=1)
于是
∥ A + B ∥ = ∥ ( A + B ) x 0 ∥ = ∥ A x 0 + B x 0 ∥ ≤ ∥ A x 0 ∥ + ∥ B x 0 ∥ ≤ ∥ A ∥ ∥ x 0 ∥ + ∥ B ∥ ∥ x 0 ∥ = ∥ A ∥ + ∥ B ∥ \begin{aligned} &\quad \Vert A+B \Vert\\ &= \Vert (A+B)x_0 \Vert\\ &=\Vert Ax_0+Bx_0 \Vert\\ &\le \Vert Ax_0 \Vert+ \Vert Bx_0 \Vert\\ &\le \Vert A \Vert \Vert x_0 \Vert+ \Vert B \Vert \Vert x_0 \Vert\\ &= \Vert A \Vert + \Vert B \Vert \end{aligned} ​∥A+B∥=∥(A+B)x0​∥=∥Ax0​+Bx0​∥≤∥Ax0​∥+∥Bx0​∥≤∥A∥∥x0​∥+∥B∥∥x0​∥=∥A∥+∥B∥​

次乘性:
对于矩阵 A B AB AB,可以找到向量 x 0 x_0 x0​,使得
∥ A B x 0 ∥ = ∥ A B ∥ ( ∥ x 0 ∥ = 1 ) \Vert ABx_0 \Vert = \Vert AB \Vert \quad (\Vert x_0 \Vert=1) ∥ABx0​∥=∥AB∥(∥x0​∥=1)
于是
∥ A B ∥ = ∥ A B x 0 ∥ = ∥ A ( B x 0 ) ∥ ≤ ∥ A ∥ ∥ B x 0 ∥ ≤ ∥ A ∥ ∥ B ∥ ∥ x 0 ∥ = ∥ A ∥ ∥ B ∥ \begin{aligned} &\quad \Vert AB \Vert\\ &= \Vert ABx_0 \Vert\\ &=\Vert A(Bx_0) \Vert\\ &\le \Vert A \Vert \Vert Bx_0 \Vert\\ &\le \Vert A \Vert \Vert B \Vert \Vert x_0\Vert\\ &= \Vert A \Vert \Vert B \Vert \end{aligned} ​∥AB∥=∥ABx0​∥=∥A(Bx0​)∥≤∥A∥∥Bx0​∥≤∥A∥∥B∥∥x0​∥=∥A∥∥B∥​
证毕

常见的算子范数

设 A ∈ C m × n , x ∈ C n A\in \mathbb{C}^{m\times n},x\in \mathbb{C}^{n} A∈Cm×n,x∈Cn,则从属于向量 x x x的三种范数 ∥ x ∥ 1 , ∥ x ∥ 2 , ∥ x ∥ ∞ \Vert x \Vert_1,\Vert x \Vert_2 , \Vert x \Vert_\infty ∥x∥1​,∥x∥2​,∥x∥∞​的算子范数依次是
(1)
∥ A ∥ 1 = max ⁡ j ∑ i = 1 m ∣ a i j ∣ \Vert A \Vert_1=\max \limits_{j} \sum_{i=1}^{m} \left| a_{ij}\right| ∥A∥1​=jmax​i=1∑m​∣aij​∣
称为列范数
(2)
∥ A ∥ 2 = λ max ⁡ ( A H A ) \Vert A \Vert_2 =\sqrt{\lambda_{\max}(A^HA)} ∥A∥2​=λmax​(AHA)
称为谱范数
(3)
∥ A ∥ ∞ = max ⁡ i ∑ j = 1 n ∣ a i j ∣ \Vert A \Vert_\infty =\max \limits_{i} \sum_{j=1}^{n}\left|a_{ij}\right| ∥A∥∞​=imax​j=1∑n​∣aij​∣
称为行范数

证明:
(1)
对于任何非零向量 x x x,设 ∥ x ∥ 1 = 1 \Vert x \Vert_1 =1 ∥x∥1​=1,则
∥ A x ∥ 1 = ∑ i = 1 m ∑ j = 1 n ∣ a i j ∣ ∣ x j ∣ = ∑ j = 1 n ∑ i = 1 m ∣ a i j ∣ ∣ x j ∣ = ∑ j = 1 n ( ∑ i = 1 m ∣ a i j ∣ ) ∣ x j ∣ ≤ max ⁡ j ∑ i = 1 m ∣ a i j ∣ ∑ j = 1 n ∣ x j ∣ = max ⁡ j ∑ i = 1 m ∣ a i j ∣ \begin{aligned} &\quad \Vert Ax \Vert_1\\ &=\sum_{i=1}^{m}\sum_{j=1}^{n}\left| a_{ij}\right|\left| x_j\right|\\ &=\sum_{j=1}^{n}\sum_{i=1}^{m}\left| a_{ij}\right|\left| x_j\right|\\ &=\sum_{j=1}^{n}(\sum_{i=1}^{m}\left| a_{ij}\right|)\left| x_j\right|\\ &\le \max\limits_{j}\sum_{i=1}^{m}\left| a_{ij}\right| \sum_{j=1}^{n}\left| x_j\right|\\ &=\max\limits_{j}\sum_{i=1}^{m}\left| a_{ij}\right| \end{aligned} ​∥Ax∥1​=i=1∑m​j=1∑n​∣aij​∣∣xj​∣=j=1∑n​i=1∑m​∣aij​∣∣xj​∣=j=1∑n​(i=1∑m​∣aij​∣)∣xj​∣≤jmax​i=1∑m​∣aij​∣j=1∑n​∣xj​∣=jmax​i=1∑m​∣aij​∣​
所以
∥ A x ∥ 1 ≤ max ⁡ j ∑ i = 1 m ∣ a i j ∣ \quad \Vert Ax \Vert_1\le \max\limits_{j}\sum_{i=1}^{m}\left| a_{ij}\right| ∥Ax∥1​≤jmax​i=1∑m​∣aij​∣
设在 j = j 0 j=j_0 j=j0​时, ∑ i = 1 m ∣ a i j ∣ \sum_{i=1}^{m}\left|a_{ij}\right| ∑i=1m​∣aij​∣达到最大值,即
∑ i = 1 m ∣ a i j 0 ∣ = max ⁡ 1 ≤ j ≤ n ∑ i = 1 m ∣ a i j ∣ \sum_{i=1}^{m}\left|a_{ij_0}\right|=\max\limits_{1\le j \le n}\sum_{i=1}^{m}\left| a_{ij}\right| i=1∑m​∣aij0​​∣=1≤j≤nmax​i=1∑m​∣aij​∣
去向量 x 0 = ( 0 , ⋯   , 0 , 1 , 0 , ⋯ 0 ) T x_0=(0,\cdots,0,1,0,\cdots 0)^T x0​=(0,⋯,0,1,0,⋯0)T
其中第 j 0 j_0 j0​个分量为 1 1 1,其余为 0 0 0,显然 ∥ x ∥ 1 = 1 \Vert x \Vert_1 =1 ∥x∥1​=1
∥ A x 0 ∥ 1 = ∑ i = 1 m ∣ ∑ j = 1 n a i j x j ∣ = ∑ i = 1 m ∣ a i j 0 ∣ = max ⁡ j ∑ i = 1 m ∣ a i j ∣ \Vert Ax_0 \Vert_1=\sum_{i=1}^{m}\left|\sum_{j=1}^{n}a_{ij}x_j\right|=\sum_{i=1}^{m}\left|a_{ij_0}\right|=\max\limits_{j}\sum_{i=1}^{m}\left|a_{ij}\right| ∥Ax0​∥1​=i=1∑m​∣∣∣∣∣​j=1∑n​aij​xj​∣∣∣∣∣​=i=1∑m​∣aij0​​∣=jmax​i=1∑m​∣aij​∣
于是
∥ A ∥ 1 = max ⁡ ∥ x ∥ 1 = 1 ∥ A x ∥ 1 = max ⁡ j ∑ i = 1 m ∣ a i j ∣ \Vert A \Vert_1=\max\limits_{\Vert x \Vert_1=1} \Vert Ax \Vert_1=\max\limits_{j} \sum_{i=1}^{m}\left| a_{ij} \right| ∥A∥1​=∥x∥1​=1max​∥Ax∥1​=jmax​i=1∑m​∣aij​∣

(2)
∥ A ∥ 2 = max ⁡ ∥ x ∥ 2 = 1 ∥ A x ∥ 2 \Vert A \Vert_2=\max\limits_{\Vert x \Vert_2 =1} \Vert Ax \Vert_2 ∥A∥2​=∥x∥2​=1max​∥Ax∥2​
因为
∥ A x ∥ 2 2 = ( A x , A x ) = ( x , A H A x ) \Vert Ax \Vert_2^2 =(Ax ,Ax)=(x,A^H Ax) ∥Ax∥22​=(Ax,Ax)=(x,AHAx)
显然,矩阵 A H A A^HA AHA是埃尔米特矩阵(复数版实对称矩阵),且非负,从而他的特征值都是非负实数
设 λ 1 ≥ λ 2 ≥ ⋯ ≥ λ n ≥ 0 \lambda_1\ge \lambda_2 \ge \cdots \ge \lambda_n \ge 0 λ1​≥λ2​≥⋯≥λn​≥0为 A H A A^HA AHA的特征值,
而 x 1 , x 2 , ⋯   , x n x_1,x_2,\cdots,x_n x1​,x2​,⋯,xn​为这些特征值对应的一组标准正交特征向量,任何一个范数为 1 1 1的向量 x x x都可以表示为
x = a 1 x 1 + ⋯ a n x n x=a_1x_1+\cdots a_nx_n x=a1​x1​+⋯an​xn​

( x , x ) = ∣ a 1 ∣ 2 + ⋯ ∣ a n ∣ 2 = 1 (x,x)=\left|a_1\right|^2+\cdots \left|a_n\right|^2=1 (x,x)=∣a1​∣2+⋯∣an​∣2=1
又因为
∥ A x ∥ 2 2 = ( x , A H A x ) = ( a 1 x 1 + ⋯ a n x n , λ 1 a 1 x 1 + ⋯ λ n a n x n ) = λ 1 ∣ a 1 ∣ 2 + ⋯ λ n ∣ a n ∣ 2 ≤ λ 1 ( ∣ a 1 ∣ 2 + ⋯ ∣ a n ∣ 2 ) = λ 1 = λ max ⁡ ( A H A ) \begin{aligned} &\quad \Vert Ax \Vert_2^2 \\ &=(x,A^H Ax)\\ &=(a_1x_1+\cdots a_n x_n,\lambda_1 a_1 x_1+\cdots \lambda_n a_n x_n)\\ &=\lambda_1\left|a_1\right|^2+\cdots \lambda_n \left|a_n\right|^2\\ &\le \lambda_1(\left|a_1\right|^2+\cdots \left|a_n\right|^2)\\ &=\lambda_1\\ &=\lambda_{\max}(A^HA) \end{aligned} ​∥Ax∥22​=(x,AHAx)=(a1​x1​+⋯an​xn​,λ1​a1​x1​+⋯λn​an​xn​)=λ1​∣a1​∣2+⋯λn​∣an​∣2≤λ1​(∣a1​∣2+⋯∣an​∣2)=λ1​=λmax​(AHA)​

取向量 x = x 1 x=x_1 x=x1​,有
∥ A x 1 ∥ 2 2 = ( x 1 , A H A x 1 ) = ( x 1 λ 1 x 1 ) = λ 1 ‾ ( x 1 , x 1 ) = λ 1 = λ max ⁡ ( A H A ) \begin{aligned} &\quad \Vert Ax_1 \Vert_2^2\\ &=(x_1,A^H Ax_1)\\ &=(x_1 \lambda_1 x_1)\\ &=\overline{\lambda_1}(x_1,x_1)\\ &=\lambda_1\\ &=\lambda_{\max}(A^HA) \end{aligned} ​∥Ax1​∥22​=(x1​,AHAx1​)=(x1​λ1​x1​)=λ1​​(x1​,x1​)=λ1​=λmax​(AHA)​
所以
∥ A ∥ 2 = max ⁡ ∥ x ∥ 2 = 1 ∥ A x ∥ 2 = λ max ⁡ ( A H A ) \Vert A \Vert_2= \max\limits_{\Vert x \Vert_2=1}\Vert Ax \Vert_2=\sqrt{\lambda_{\max}(A^HA)} ∥A∥2​=∥x∥2​=1max​∥Ax∥2​=λmax​(AHA)

(3)设 ∥ x ∥ ∞ = 1 \Vert x \Vert_\infty=1 ∥x∥∞​=1,则
∥ A x ∥ ∞ = max ⁡ i ∣ ∑ j = 1 n a i j x j ∣ ≤ max ⁡ j ∑ j = 1 n ∣ a i j ∣ ∣ x j ∣ ≤ max ⁡ j ∑ j = 1 n ∣ a i j ∣ \begin{aligned} &\quad \Vert Ax \Vert_\infty\\ &=\max\limits_{i}\left|\sum_{j=1}^{n}a_{ij}x_j\right|\\ &\le \max\limits_{j}\sum_{j=1}^{n}\left|a_{ij}\right|\left|x_j\right|\\ &\le \max\limits_{j}\sum_{j=1}^{n}\left|a_{ij}\right| \end{aligned} ​∥Ax∥∞​=imax​∣∣∣∣∣​j=1∑n​aij​xj​∣∣∣∣∣​≤jmax​j=1∑n​∣aij​∣∣xj​∣≤jmax​j=1∑n​∣aij​∣​
所以
max ⁡ ∥ x ∥ ∞ = 1 ∥ A x ∥ ∞ ≤ max ⁡ i ∑ j = 1 n ∣ a i j ∣ \max\limits_{\Vert x \Vert_\infty=1}\Vert Ax \Vert_\infty \le \max\limits_{i}\sum_{j=1}^{n}\left|a_{ij}\right| ∥x∥∞​=1max​∥Ax∥∞​≤imax​j=1∑n​∣aij​∣

设 ∑ j = 1 n ∣ a i j ∣ \sum_{j=1}^{n}\left|a_{ij}\right| ∑j=1n​∣aij​∣在 i = i 0 i=i_0 i=i0​是取到最大值,取向量
x 0 = ( x 1 , ⋯   , x n ) T x_0=(x_1,\cdots,x_n)^T x0​=(x1​,⋯,xn​)T
其中 x j = { ∣ a i 0 j ∣ a i 0 j , a i 0 j = 0 1 , a i 0 j = 0 x_j=\begin{cases} \frac{\left|a_{i_0 j}\right|}{a_{i_0 j}},a_{i_0 j}=0\\ 1,a_{i_0 j} =0 \end{cases} xj​=⎩⎨⎧​ai0​j​∣ai0​j​∣​,ai0​j​=01,ai0​j​=0​
易知
∥ x 0 ∥ ∞ = 1 \Vert x_0 \Vert_\infty=1 ∥x0​∥∞​=1
且当 i = i 0 i=i_0 i=i0​时
∣ ∑ j = 1 n a i j x j ∣ = max ⁡ i ∑ j = 1 n ∣ a i j ∣ \left|\sum_{j=1}^{n}a_{ij}x_j\right|=\max\limits_{i}\sum_{j=1}^{n}\left|a_{ij}\right| ∣∣∣∣∣​j=1∑n​aij​xj​∣∣∣∣∣​=imax​j=1∑n​∣aij​∣
从而
∥ A x 0 ∥ ∞ = max ⁡ i ∑ j = 1 n ∣ a i j ∣ \Vert Ax_0 \Vert_\infty=\max\limits_{i}\sum_{j=1}^{n}\left| a_{ij}\right| ∥Ax0​∥∞​=imax​j=1∑n​∣aij​∣
所以
∥ A ∥ ∞ = max ⁡ ∥ x ∥ ∞ = 1 ∥ A x ∥ ∞ = max ⁡ i ∑ j = 1 n ∣ a i j ∣ \Vert A \Vert_\infty=\max\limits_{\Vert x \Vert_\infty=1} \Vert Ax \Vert_\infty=\max\limits_{i}\sum_{j=1}^{n}\left| a_{ij}\right| ∥A∥∞​=∥x∥∞​=1max​∥Ax∥∞​=imax​j=1∑n​∣aij​∣

F范数性质

F范数又叫做费罗贝尼乌斯(Frobenius)范数
设 A ∈ C m × n A\in\mathbb{C}^{m\times n} A∈Cm×n,而 U ∈ C m × m , V ∈ C n × n U\in \mathbb{C}^{m\times m} ,V \in \mathbb{C}^{n\times n} U∈Cm×m,V∈Cn×n都是酉矩阵

∥ U A ∥ F = ∥ A ∥ F = ∥ A V ∥ F \Vert UA \Vert_F= \Vert A \Vert_F= \Vert AV \Vert_F ∥UA∥F​=∥A∥F​=∥AV∥F​

证明:
设 A = ( α 1 , ⋯   , α n ) A=(\alpha_1,\cdots,\alpha_n) A=(α1​,⋯,αn​),则
∥ U A ∥ F 2 = ∥ U ( α 1 , ⋯   , α n ) ∥ F 2 = ∑ i = 1 n ∥ U α i ∥ 2 2 = ∑ i = 1 n ∥ α i ∥ 2 2 = ∥ A ∥ F 2 \begin{aligned} &\quad \Vert UA \Vert_F^2\\ &= \Vert U(\alpha_1,\cdots ,\alpha_n) \Vert_F^2\\ &=\sum_{i=1}^{n}\Vert U\alpha_i\Vert_2^2\\ &=\sum_{i=1}^{n}\Vert\alpha_i \Vert_2^2\\ &=\Vert A \Vert_F^2 \end{aligned} ​∥UA∥F2​=∥U(α1​,⋯,αn​)∥F2​=i=1∑n​∥Uαi​∥22​=i=1∑n​∥αi​∥22​=∥A∥F2​​
于是
∥ U A ∥ F = ∥ A ∥ F \Vert UA \Vert_F= \Vert A \Vert_F ∥UA∥F​=∥A∥F​

∥ A V ∥ F = ∥ ( A V ) H ∥ F = ∥ V H A H ∥ F = ∥ A H ∥ F = ∥ A ∥ F \Vert AV \Vert_F = \Vert (AV)^H \Vert_F= \Vert V^H A^H \Vert_F=\Vert A^H \Vert_F = \Vert A \Vert_F ∥AV∥F​=∥(AV)H∥F​=∥VHAH∥F​=∥AH∥F​=∥A∥F​

推论

与 A A A酉相似的矩阵的F范数相同
即 B = U H A U B=U^HAU B=UHAU,则 ∥ B ∥ F = ∥ A ∥ F \Vert B \Vert_F= \Vert A \Vert_F ∥B∥F​=∥A∥F​,其中 U U U是酉矩阵

谱范数的性质和谱半径

定理1

设 A ∈ C m × n A\in \mathbb{C}^{m\times n} A∈Cm×n,则
(1) ∥ A ∥ 2 = max ⁡ ∥ x ∥ 2 = ∥ y ∥ 2 = 1 ∣ y H A x ∣ , x ∈ C n , y ∈ C m \Vert A \Vert_2 =\max\limits_{\Vert x \Vert_2=\Vert y \Vert_2=1} \left| y^H Ax\right|,x \in \mathbb{C}^{n},y\in \mathbb{C}^{m} ∥A∥2​=∥x∥2​=∥y∥2​=1max​∣∣​yHAx∣∣​,x∈Cn,y∈Cm
(2) ∥ A H ∥ 2 = ∥ A ∥ 2 \Vert A^H \Vert_2= \Vert A \Vert_2 ∥AH∥2​=∥A∥2​
(3) ∥ A H A ∥ 2 = ∥ A ∥ F 2 \Vert A^H A \Vert_2 = \Vert A\Vert_F^2 ∥AHA∥2​=∥A∥F2​
证明:
(1)对满足 ∥ x ∥ 2 = ∥ y ∥ 2 = 1 \Vert x \Vert_2=\Vert y \Vert_2=1 ∥x∥2​=∥y∥2​=1的 x , y x,y x,y,有
∣ y H A x ∣ ≤ ∥ y ∥ 2 ∥ A x ∥ 2 ≤ ∥ A ∥ 2 \left| y^H Ax \right| \le \Vert y \Vert_2 \Vert Ax \Vert_2 \le \Vert A \Vert_2 ∣∣​yHAx∣∣​≤∥y∥2​∥Ax∥2​≤∥A∥2​
设有 ∥ x ∥ 2 = 1 \Vert x \Vert_2=1 ∥x∥2​=1,使得 ∥ A x ∥ 2 = ∥ A ∥ 2 ≠ 0 \Vert Ax \Vert_2= \Vert A \Vert_2 \neq 0 ∥Ax∥2​=∥A∥2​​=0
令 y = A x ∥ A x ∥ 2 y=\frac{Ax}{\Vert Ax \Vert_2} y=∥Ax∥2​Ax​,就有
∣ y H A x ∣ = ∥ A x ∥ 2 2 ∥ A x ∥ 2 = ∥ A x ∥ 2 = ∥ A ∥ 2 \left|y^H Ax\right|=\frac{\Vert Ax \Vert_2^2}{\Vert Ax \Vert_2}= \Vert Ax \Vert_2=\Vert A \Vert_2 ∣∣​yHAx∣∣​=∥Ax∥2​∥Ax∥22​​=∥Ax∥2​=∥A∥2​
从而
max ⁡ ∥ x ∥ 2 = ∥ y ∥ 2 = 1 ∣ y H A x ∣ = ∥ A ∥ 2 \max\limits_{\Vert x \Vert_2=\Vert y \Vert_2=1}\left| y^H Ax \right|= \Vert A \Vert_2 ∥x∥2​=∥y∥2​=1max​∣∣​yHAx∣∣​=∥A∥2​
(2)
∥ A ∥ 2 = max ⁡ ∥ x ∥ 2 = ∥ y ∥ 2 = 1 ∣ y H A x ∣ = max ⁡ ∥ x ∥ 2 = ∥ y ∥ 2 = 1 ∣ x H A H y ∣ = ∥ A H ∥ 2 \begin{aligned} &\quad \Vert A \Vert_2\\ &=\max\limits_{\Vert x \Vert_2=\Vert y \Vert_2=1} \left| y^H Ax \right|\\ &=\max\limits_{\Vert x \Vert_2=\Vert y \Vert_2=1}\left| x^H A^H y \right|\\ &=\Vert A^H \Vert_2 \end{aligned} ​∥A∥2​=∥x∥2​=∥y∥2​=1max​∣∣​yHAx∣∣​=∥x∥2​=∥y∥2​=1max​∣∣​xHAHy∣∣​=∥AH∥2​​
(3)
由 ∥ A H A ∥ 2 ≤ ∥ A H ∥ 2 ∥ A ∥ 2 , ∥ A H ∥ 2 = ∥ A ∥ 2 \Vert A^H A\Vert_2 \le \Vert A^H \Vert_2 \Vert A \Vert_2,\Vert A^H \Vert_2= \Vert A \Vert_2 ∥AHA∥2​≤∥AH∥2​∥A∥2​,∥AH∥2​=∥A∥2​,有
∥ A H A ∥ 2 ≤ ∥ A ∥ 2 2 \Vert A^HA \Vert_2 \le \Vert A \Vert_2^2 ∥AHA∥2​≤∥A∥22​
令 ∥ x ∥ 2 = 1 \Vert x \Vert_2=1 ∥x∥2​=1,使得 ∥ A x ∥ 2 = ∥ A ∥ 2 \Vert Ax \Vert_2= \Vert A \Vert_2 ∥Ax∥2​=∥A∥2​,于是
∥ A H A ∥ 2 ≥ max ⁡ ∥ x ∥ 2 = 1 ∣ x H A H A x ∣ = max ⁡ ∥ x ∥ 2 = 1 ∥ A x ∥ 2 2 = ∥ A ∥ 2 2 \begin{aligned} &\quad \Vert A^HA \Vert_2\\ &\ge\max\limits_{\Vert x\Vert_2=1}\left|x^HA^H Ax\right|\\ &=\max\limits_{\Vert x \Vert_2=1} \Vert Ax \Vert_2^2\\ &= \Vert A \Vert_2^2 \end{aligned} ​∥AHA∥2​≥∥x∥2​=1max​∣∣​xHAHAx∣∣​=∥x∥2​=1max​∥Ax∥22​=∥A∥22​​

定理2

设 A ∈ C m × n , U ∈ C m × n , V ∈ C n × n A\in \mathbb{C}^{m\times n},U\in \mathbb{C}^{m\times n},V \in \mathbb{C}^{n\times n} A∈Cm×n,U∈Cm×n,V∈Cn×n,且 U H U = I m , V H V = I n U^HU=I_m,V^HV=I_n UHU=Im​,VHV=In​,则
∥ U A V ∥ 2 = ∥ A ∥ 2 \Vert UAV \Vert_2 = \Vert A \Vert_2 ∥UAV∥2​=∥A∥2​

证明:
令 v = V H x , u = U y v=V^Hx,u=Uy v=VHx,u=Uy,则
∥ x ∥ 2 = 1 ⇔ ∥ v ∥ 2 = 1 \Vert x \Vert_2=1 \Leftrightarrow \Vert v \Vert_2=1 ∥x∥2​=1⇔∥v∥2​=1
∥ y ∥ 2 = 1 ⇔ ∥ u ∥ 2 = 1 \Vert y \Vert_2 =1 \Leftrightarrow \Vert u \Vert_2=1 ∥y∥2​=1⇔∥u∥2​=1
于是
∥ A ∥ 2 = max ⁡ ∥ x ∥ 2 = ∥ y ∥ 2 = 1 ∣ y H A x ∣ = max ⁡ ∥ v ∥ 2 = ∥ u ∥ 2 = 1 ∣ u H U A V v ∣ = ∥ U A V ∥ 2 \begin{aligned} &\quad \Vert A \Vert_2\\ &=\max\limits_{\Vert x \Vert_2=\Vert y \Vert_2=1}\left| y^H Ax \right|\\ &=\max \limits_{\Vert v \Vert_2 =\Vert u \Vert_2=1}\left|u^HUAVv\right|\\ &=\Vert UAV \Vert_2 \end{aligned} ​∥A∥2​=∥x∥2​=∥y∥2​=1max​∣∣​yHAx∣∣​=∥v∥2​=∥u∥2​=1max​∣∣​uHUAVv∣∣​=∥UAV∥2​​

定理3

设 A ∈ C n × n A \in \mathbb{C}^{n\times n} A∈Cn×n,若 ∥ A ∥ < 1 \Vert A \Vert<1 ∥A∥<1,则 I − A I-A I−A为非奇异矩阵,且
∥ ( I − A ) − 1 ∥ ≤ ( 1 − ∥ A ∥ ) − 1 \Vert (I-A)^{-1} \Vert\le (1- \Vert A \Vert)^{-1} ∥(I−A)−1∥≤(1−∥A∥)−1
证明:
设 x x x为任一非零向量,则
∥ ( I − A ) x ∥ = ∥ x − A x ∥ ≥ ∥ x ∥ − ∥ A x ∥ ≥ ∥ x ∥ − ∥ A ∥ ∥ x ∥ = ( 1 − ∥ A ∥ ) ∥ x ∥ > 0 \begin{aligned} &\quad \Vert (I-A)x \Vert\\ &= \Vert x-Ax \Vert\\ &\ge \Vert x \Vert- \Vert Ax \Vert\\ &\ge \Vert x \Vert- \Vert A \Vert \Vert x \Vert\\ &=(1-\Vert A \Vert)\Vert x \Vert\\ &>0 \end{aligned} ​∥(I−A)x∥=∥x−Ax∥≥∥x∥−∥Ax∥≥∥x∥−∥A∥∥x∥=(1−∥A∥)∥x∥>0​
所以,若 x ≠ 0 x\neq 0 x​=0,则 ( I − A ) x ≠ 0 (I-A)x \neq 0 (I−A)x​=0
从而方程
( I − A ) x = 0 (I-A)x=0 (I−A)x=0
无非零解,故 I − A I-A I−A非奇异
( I − A ) − 1 = ( ( I − A ) + A ) ( I − A ) − 1 = I + A ( I − A ) − 1 \begin{aligned} (I-A)^{-1}&=((I-A)+A)(I-A)^{-1}\\ &=I+A(I-A)^{-1} \end{aligned} (I−A)−1​=((I−A)+A)(I−A)−1=I+A(I−A)−1​
从而
∥ ( I − A ) − 1 ∥ = ∥ I + A ( I − A ) − 1 ∥ ≤ ∥ I ∥ + ∥ A ∥ ∥ ( I − A ) − 1 ∥ = 1 + ∥ A ∥ ∥ ( I − A ) − 1 ∥ \begin{aligned} &\quad \Vert (I-A)^{-1} \Vert\\ &=\Vert I+A(I-A)^{-1} \Vert\\ &\le \Vert I \Vert + \Vert A \Vert \Vert (I-A)^{-1} \Vert\\ &=1+\Vert A \Vert \Vert (I-A)^{-1} \Vert \end{aligned} ​∥(I−A)−1∥=∥I+A(I−A)−1∥≤∥I∥+∥A∥∥(I−A)−1∥=1+∥A∥∥(I−A)−1∥​
观察首尾,得到
∥ ( I − A ) − 1 ∥ ≤ ( 1 − ∥ A ∥ ) − 1 \Vert (I-A)^{-1} \Vert\le (1- \Vert A \Vert)^{-1} ∥(I−A)−1∥≤(1−∥A∥)−1

谱半径

设 A ∈ C n × n A\in \mathbb{C}^{n\times n} A∈Cn×n, λ 1 , ⋯   , λ n \lambda_1,\cdots, \lambda_n λ1​,⋯,λn​为 A A A的特征值,我们称
ρ ( A ) = max ⁡ i ∣ λ i ∣ \rho(A)=\max \limits_{i}\left|\lambda_i\right| ρ(A)=imax​∣λi​∣
为 A A A的谱半径

特征值上界

对于任意矩阵 A ∈ C n × n A\in \mathbb{C}^{n\times n} A∈Cn×n,总有
ρ ( A ) ≤ ∥ A ∥ \rho(A)\le \Vert A \Vert ρ(A)≤∥A∥
证明:
设 λ \lambda λ是 A A A的任一特征值, x x x为对应的特征向量,则有 A x λ x Ax\lambda x Axλx
根据相容性
∣ λ ∣ ∥ x ∥ = ∥ λ x ∥ ≤ ∥ A ∥ ∥ x ∥ \left|\lambda \right| \Vert x \Vert = \Vert \lambda x \Vert \le \Vert A \Vert \Vert x \Vert ∣λ∣∥x∥=∥λx∥≤∥A∥∥x∥
于是
∣ λ ∣ ≤ ∥ A ∥ \left|\lambda \right| \le \Vert A \Vert ∣λ∣≤∥A∥

ρ ( A ) ≤ ∥ A ∥ \rho(A) \le \Vert A \Vert ρ(A)≤∥A∥

定理4

如果 A ∈ C n × n A\in \mathbb{C}^{n\times n} A∈Cn×n,且 A A A是正规矩阵(包括实对称矩阵),则
ρ ( A ) = ∥ A ∥ 2 \rho(A)=\Vert A \Vert_2 ρ(A)=∥A∥2​
证明:
因为是正规矩阵,存在酉矩阵 U U U,使得
U H A U = d i a g ( λ 1 , ⋯   , λ n ) = A U^H AU =diag(\lambda_1,\cdots,\lambda_n)=A UHAU=diag(λ1​,⋯,λn​)=A
于是
∥ A ∥ 2 = ∥ U H A U ∥ = ∥ d i a g ( λ 1 , ⋯   , λ n ) ∥ = λ max ⁡ ( A H A ) = max ⁡ i ( λ ‾ i λ i ) = max ⁡ i ∣ λ i ∣ 2 = ρ ( A ) \begin{aligned} &\quad \Vert A \Vert_2\\ &= \Vert U^H AU \Vert\\ &=\Vert diag(\lambda_1,\cdots , \lambda_n)\Vert\\ &=\sqrt{\lambda_{\max}(A^HA)}\\ &=\sqrt{\max\limits_{i}(\overline{\lambda}_i\lambda_i)}\\ &=\sqrt{\max\limits_{i}\left|\lambda_i\right|^2}\\ &=\rho(A) \end{aligned} ​∥A∥2​=∥UHAU∥=∥diag(λ1​,⋯,λn​)∥=λmax​(AHA) ​=imax​(λi​λi​) ​=imax​∣λi​∣2 ​=ρ(A)​

定理5

对于任意非奇异矩阵 A ∈ C n × n A\in \mathbb{C}^{n\times n} A∈Cn×n, A A A的谱范数为
∥ A ∥ 2 = ρ ( A H A ) = ρ ( A A H ) \Vert A \Vert_2 = \sqrt{\rho(A^HA)}=\sqrt{\rho(AA^H)} ∥A∥2​=ρ(AHA) ​=ρ(AAH)
证明:
∥ A ∥ 2 = λ max ⁡ ( A H A ) = ρ ( A H A ) \begin{aligned} &\quad \Vert A \Vert_2\\ &=\sqrt{\lambda_{\max}(A^HA)}\\ &=\sqrt{\rho(A^HA)} \end{aligned} ​∥A∥2​=λmax​(AHA) ​=ρ(AHA) ​​
因为 A A H = A ( A H A ) A − 1 AA^H=A(A^HA)A^{-1} AAH=A(AHA)A−1,所以 A A H ∼ A H A AA^H\sim A^HA AAH∼AHA,特征值相同,从而
∥ A ∥ 2 = ρ ( A H A ) = ρ ( A A H ) \Vert A \Vert_2= \sqrt{\rho(A^HA)}=\sqrt{\rho(AA^H)} ∥A∥2​=ρ(AHA) ​=ρ(AAH)

上一篇:Noip模拟76 2021.10.12


下一篇:「tricks」整体DP