本文主要内容如下:

1. 矩阵范数

定义 如果矩阵 A ∈ R m × n \bold A\in\mathbb{R}^{m\times n} ARm×n 的某个非负实值函数 N ( A ) = ∣ ∣ A ∣ ∣ N(\bold A)=||\bold A|| N(A)=∣∣A∣∣,满足如下条件:

(1) ∣ ∣ A ∣ ∣ ≥ 0 ,当且仅当 A = 0 时取等号(正定性) ||\bold A||\ge0,当且仅当\bold A=0时取等号(正定性) ∣∣A∣∣0,当且仅当A=0时取等号(正定性)

(2) ∣ ∣ c A ∣ ∣ = ∣ c ∣ ⋅ ∣ ∣ A ∣ ∣ , c ∈ R (齐次性) ||c\bold{A}||=|c|\cdot||\bold{A}||,c\in\mathbb{R}(齐次性) ∣∣cA∣∣=c∣∣A∣∣cR(齐次性)

(3) ∣ ∣ A + B ∣ ∣ ≤ ∣ ∣ A ∣ ∣ + ∣ ∣ B ∣ ∣ (三角不等式) ||\bold{A+B}||\le||\bold{A}||+||\bold{B}||(三角不等式) ∣∣A+B∣∣∣∣A∣∣+∣∣B∣∣(三角不等式)

(4) ∣ ∣ A ⋅ B ∣ ∣ m × s ≤ ∣ ∣ A ∣ ∣ m × n ⋅ ∣ ∣ B ∣ ∣ n × s (相容性 / 次乘性) ||\bold{A\cdot B}||_{m\times s}\le||\bold A||_{m\times n}\cdot||\bold{B}||_{n\times s}(相容性/次乘性) ∣∣ABm×s∣∣Am×n∣∣Bn×s(相容性/次乘性)

则称 ∣ ∣ A ∣ ∣ ||\bold A|| ∣∣A∣∣ 为定义在 R n × n \mathbb{R}^{n\times n} Rn×n 上的一个矩阵范数。若该非负实值函数不满足条件 (4) 仅满足条件 (1~3),则称为广义矩阵范数

根据矩阵范数的三角不等式有:
∣ ∣ A − B ∣ ∣ + ∣ ∣ B ∣ ∣ ≥ ∣ ∣ A ∣ ∣ ⟹ ∣ ∣ A − B ∣ ∣ ≥ ∣ ∣ A ∣ ∣ − ∣ ∣ B ∣ ∣ ∣ ∣ A − B ∣ ∣ + ∣ ∣ A ∣ ∣ = ∣ ∣ B − A ∣ ∣ + ∣ ∣ A ∣ ∣ ≥ ∣ ∣ B ∣ ∣ ⟹ ∣ ∣ A − B ∣ ∣ ≥ ∣ ∣ B ∣ ∣ − ∣ ∣ A ∣ ∣ \begin{aligned} &||\bold A-\bold B||+||\bold B||\ge||\bold A||\Longrightarrow ||\bold A-\bold B||\ge||\bold A||-||\bold B|| \\\\ &||\bold A-\bold B||+||\bold A||=||\bold B-\bold A||+||\bold A||\ge||\bold B||\Longrightarrow ||\bold A-\bold B||\ge||\bold B||-||\bold A|| \end{aligned} ∣∣AB∣∣+∣∣B∣∣∣∣A∣∣∣∣AB∣∣∣∣A∣∣∣∣B∣∣∣∣AB∣∣+∣∣A∣∣=∣∣BA∣∣+∣∣A∣∣∣∣B∣∣∣∣AB∣∣∣∣B∣∣∣∣A∣∣
结合上述两式与三角不等式,便有:
∣   ∣ ∣ A ∣ ∣ − ∣ ∣ B ∣ ∣   ∣ ≤ ∣ ∣ A − B ∣ ∣ ≤ ∣ ∣ A ∣ ∣ + ∣ ∣ B ∣ ∣ |\ ||\bold A||-||\bold B||\ |\le||\bold A-\bold B||\le||\bold A||+||\bold B||  ∣∣A∣∣∣∣B∣∣ ∣∣AB∣∣∣∣A∣∣+∣∣B∣∣

由常用的向量范数:1-范数,欧式范数,最大范数,可以定义如下矩阵的实值函数:
( 1 ) ∣ ∣ A ∣ ∣ M 1 ≜ ∑ i = 1 m ∑ j = 1 n ∣ A ∣ i j ; ( 2 ) ∣ ∣ A ∣ ∣ F ≜ ∑ i = 1 m ∑ j = 1 n a i j 2 = t r ( A T A ) ( F r o b e n i u s 范数); ( 3 ) ∣ ∣ A ∣ ∣ M ∞ ′ = max ⁡ 1 ≤ i ≤ m , 1 ≤ j ≤ n ∣ a i j ∣ (一种广义矩阵范数) \begin{aligned} &(1)\quad ||\bold{A}||_{M_1}\triangleq\sum_{i=1}^m\sum_{j=1}^n|A|_{ij};\\\\ &(2)\quad ||\bold{A}||_{F}\triangleq\sqrt{\sum_{i=1}^m\sum_{j=1}^na_{ij}^2}=\sqrt{tr(\bold A^T\bold A)}(Frobenius范数);\\\\ &(3)\quad ||\bold{A}||_{M'_\infty}=\max_{1\le i\le m,1\le j\le n}|a_{ij}|(一种广义矩阵范数) \end{aligned} (1)∣∣AM1i=1mj=1nAij(2)∣∣AFi=1mj=1naij2 =tr(ATA) Frobenius范数);(3)∣∣AM=1im,1jnmaxaij(一种广义矩阵范数)
它们直接由向量范数推广而来,显然是满足条件(1~3)的,但是否满足相容性条件(是否可以被称为矩阵范数)仍需要验证:
∣ ∣ A m × n B n × s ∣ ∣ M 1 = ∑ i = 1 m ∑ j = 1 s ( ∣ a i 1 b 1 j + a i 2 b 2 j + ⋯ + a i n b n j ∣ ) ≤ ∑ i = 1 m ∑ j = 1 s ( ∣ a i 1 ∣ ∣ b 1 j ∣ + ∣ a i 2 ∣ ∣ b 2 j ∣ + ⋯ + ∣ a i n ∣ ∣ b n j ∣ ) ≤ ∑ i = 1 m ∑ j = 1 s ( ∣ a i 1 ∣ + ∣ a i 2 ∣ + ⋯ + ∣ a i n ∣ ) ( ∣ b 1 j ∣ + ∣ b 2 j ∣ + ⋯ + ∣ b n j ∣ ) ≤ ∑ i = 1 m ( ∣ a i 1 ∣ + ∣ a i 2 ∣ + ⋯ + ∣ a i n ∣ ) ⋅ ∑ j = 1 s ( ∣ b 1 j ∣ + ∣ b 2 j ∣ + ⋯ + ∣ b n j ∣ ) = ∑ i = 1 m ∑ j = 1 n ∣ a i j ∣   ∑ i = 1 n ∑ j = 1 s ∣ b i j ∣ = ∣ ∣ A ∣ ∣ M 1 ⋅ ∣ ∣ B ∣ ∣ M 1 ∣ ∣ A m × n B n × s ∣ ∣ F = ∑ i = 1 m ∑ j = 1 s ( a i 1 b 1 j + a i 2 b 2 j + ⋯ + a i n b n j ) 2 ≤ ∑ i = 1 m ∑ j = 1 s ( a i 1 2 + a i 2 2 + ⋯ + a i n 2 ) ⋅ ( b 1 j 2 + b 2 j 2 + ⋯ + b n j 2 ) ≤ ∑ i = 1 m ( a i 1 2 + a i 2 2 + ⋯ + a i n 2 ) ∑ j = 1 s ( b 1 j 2 + b 2 j 2 + ⋯ + b n j 2 ) = ∣ ∣ A ∣ ∣ F ⋅ ∣ ∣ B ∣ ∣ F \begin{aligned} &||\bold{A_{m\times n}B_{n\times s}}||_{M_1}=\sum_{i=1}^m\sum_{j=1}^s(|a_{i1}b_{1j}+a_{i2}b_{2j}+\dots+a_{in}b_{nj}|)\\\\ &\qquad\qquad\qquad\quad\le\sum_{i=1}^m\sum_{j=1}^s(|a_{i1}||b_{1j}|+|a_{i2}||b_{2j}|+\dots+|a_{in}||b_{nj}|)\\\\ &\qquad\qquad\qquad\quad\le\sum_{i=1}^m\sum_{j=1}^s(|a_{i1}|+|a_{i2}|+\dots+|a_{in}|)(|b_{1j}|+|b_{2j}|+\dots+|b_{nj}|)\\\\ &\qquad\qquad\qquad\quad\le\sum_{i=1}^m(|a_{i1}|+|a_{i2}|+\dots+|a_{in}|)\cdot\sum_{j=1}^s(|b_{1j}|+|b_{2j}|+\dots+|b_{nj}|)\\\\ &\qquad\qquad\qquad\quad=\sum_{i=1}^m\sum_{j=1}^n|a_{ij}|\ \sum_{i=1}^n\sum_{j=1}^s|b_{ij}|=||\bold{A}||_{M_1}\cdot||\bold{B}||_{M_1}\\\\ &\\\\ &||\bold{A_{m\times n}B_{n\times s}}||_{F}=\sqrt{\sum_{i=1}^m\sum_{j=1}^s(a_{i1}b_{1j}+a_{i2}b_{2j}+\dots+a_{in}b_{nj})^2}\\\\ &\qquad\qquad\qquad\quad\le\sqrt{\sum_{i=1}^m\sum_{j=1}^s(a_{i1}^2+a_{i2}^2+\dots+a_{in}^2)\cdot(b_{1j}^2+b_{2j}^2+\dots+b_{nj}^2)}\\\\ &\qquad\qquad\qquad\quad\le\sqrt{\sum_{i=1}^m(a_{i1}^2+a_{i2}^2+\dots+a_{in}^2)}\sqrt{\sum_{j=1}^s(b_{1j}^2+b_{2j}^2+\dots+b_{nj}^2)}\\\\ &\qquad\qquad\qquad\quad=||\bold{A}||_F\cdot||\bold B||_F \end{aligned} ∣∣Am×nBn×sM1=i=1mj=1s(ai1b1j+ai2b2j++ainbnj)i=1mj=1s(ai1∣∣b1j+ai2∣∣b2j++ain∣∣bnj)i=1mj=1s(ai1+ai2++ain)(b1j+b2j++bnj)i=1m(ai1+ai2++ain)j=1s(b1j+b2j++bnj)=i=1mj=1naij i=1nj=1sbij=∣∣AM1∣∣BM1∣∣Am×nBn×sF=i=1mj=1s(ai1b1j+ai2b2j++ainbnj)2 i=1mj=1s(ai12+ai22++ain2)(b1j2+b2j2++bnj2) i=1m(ai12+ai22++ain2) j=1s(b1j2+b2j2++bnj2) =∣∣AF∣∣BF
对于(3)定义的矩阵实值函数而言,其并不满足相容性条件,通过举反例可以说明。对(3)重新进行定义便可符合矩阵范数的定义:
∣ ∣ A m × n ∣ ∣ M ∞ ≜ m n ( max ⁡ 1 ≤ i ≤ m , 1 ≤ j ≤ n ∣ a i j ∣ ) ||\bold A_{m\times n}||_{M_\infty}\triangleq mn\left(\max_{1\le i\le m,1\le j\le n}|a_{ij}|\right) ∣∣Am×nMmn(1im,1jnmaxaij)
证明,其满足相容性条件:
∣ ∣ A m × n B n × s ∣ ∣ M ∞ = m s [ max ⁡ 1 ≤ i ≤ m , 1 ≤ j ≤ s ∣ a i 1 b 1 j + a i 2 b 2 j + ⋯ + a i n b n j ∣ ]    ≤ m s [ max ⁡ 1 ≤ i ≤ m , 1 ≤ j ≤ s ( ∣ a i 1 ∣ ∣ b 1 j ∣ + ∣ a i 2 ∣ ∣ b 2 j ∣ + ⋯ + ∣ a i n ∣ ∣ b n j ∣ ) ]    ≤ m s [ max ⁡ 1 ≤ i ≤ m , 1 ≤ j ≤ s ( ∣ a i 1 ∣ + ∣ a i 2 ∣ + ⋯ + ∣ a i n ∣ ) ( ∣ b 1 j ∣ + ∣ b 2 j ∣ + ⋯ + ∣ b n j ∣ ) ]    ≤ m s [ n ( max ⁡ 1 ≤ i ≤ m , 1 ≤ j ≤ n ∣ a i j ∣ ) n ( max ⁡ 1 ≤ i ≤ n , 1 ≤ j ≤ s ∣ b i j ∣ ∣ ) ]    = m n ( max ⁡ 1 ≤ i ≤ m , 1 ≤ j ≤ n ∣ a i j ∣ ) ⋅ n s ( max ⁡ 1 ≤ i ≤ n , 1 ≤ j ≤ s ∣ b i j ∣ ∣ )    = ∣ ∣ A m × n ∣ ∣ M ∞ ⋅ ∣ ∣ B n × s ∣ ∣ M ∞ (证毕) \begin{aligned} &||\bold A_{m\times n}\bold B_{n\times s}||_{M_\infty}=ms\left[\max_{1\le i\le m,1\le j\le s}|a_{i1}b_{1j}+a_{i2}b_{2j}+\dots+a_{in}b_{nj}|\right]\\\\ &\qquad\qquad\qquad\ \ \le ms\left[\max_{1\le i\le m,1\le j\le s}(|a_{i1}||b_{1j}|+|a_{i2}||b_{2j}|+\dots+|a_{in}||b_{nj}|)\right]\\\\ &\qquad\qquad\qquad\ \ \le ms\left[\max_{1\le i\le m,1\le j\le s}(|a_{i1}|+|a_{i2}|+\dots+|a_{in}|)(|b_{1j}|+|b_{2j}|+\dots+|b_{nj}|)\right]\\\\ &\qquad\qquad\qquad\ \ \le ms\left[n\left(\max_{1\le i\le m,1\le j\le n}|a_{ij}|\right)n\left(\max_{1\le i\le n,1\le j\le s}|b_{ij}||\right)\right]\\\\ &\qquad\qquad\qquad\ \ = mn\left(\max_{1\le i\le m,1\le j\le n}|a_{ij}|\right)\cdot ns\left(\max_{1\le i\le n,1\le j\le s}|b_{ij}||\right)\\\\ &\qquad\qquad\qquad\ \ =||\bold A_{m\times n}||_{M_\infty}\cdot||\bold B_{n\times s}||_{M_\infty}(证毕) \end{aligned} ∣∣Am×nBn×sM=ms[1im,1jsmaxai1b1j+ai2b2j++ainbnj]  ms[1im,1jsmax(ai1∣∣b1j+ai2∣∣b2j++ain∣∣bnj)]  ms[1im,1jsmax(ai1+ai2++ain)(b1j+b2j++bnj)]  ms[n(1im,1jnmaxaij)n(1in,1jsmaxbij∣∣)]  =mn(1im,1jnmaxaij)ns(1in,1jsmaxbij∣∣)  =∣∣Am×nM∣∣Bn×sM(证毕)

2. 矩阵的算子范数

定义 x ⃗ ∈ R n , A ∈ R m × n \vec{x}\in\mathbb{R}^n,\bold{A}\in\mathbb{R}^{m\times n} x Rn,ARm×n,则 A x ⃗ ∈ R m \bold{A}\vec{x}\in\mathbb{R}^{m} Ax Rm,设已定义了向量空间 R m \mathbb{R}^{m} Rm R n \mathbb{R}^{n} Rn 上的同种向量范数 ∣ ∣ ∙ ∣ ∣ {||\bullet||} ∣∣∣∣ ,相应地可以定义一个矩阵的非负实值函数:
∣ ∣ A ∣ ∣ v ≜ max ⁡ ∣ ∣ x ⃗ ∣ ∣ v ≠ 0 ∣ ∣ A x ⃗ ∣ ∣ ∣ ∣ x ⃗ ∣ ∣ = max ⁡ ∣ ∣ x ⃗ ∣ ∣ v = 1 ∣ ∣ A x ⃗ ∣ ∣ ( ∗ ) ||\bold A||_{v}\triangleq\max_{||\vec{x}||_v\ne0}\frac{||\bold{A}\vec{x}||}{||\vec{x}||}=\max_{||\vec{x}||_v=1}||\bold{A}\vec{x}||\quad(*) ∣∣Av∣∣x v=0max∣∣x ∣∣∣∣Ax ∣∣=∣∣x v=1max∣∣Ax ∣∣()
若该实函数为矩阵范数,将其称作算子范数,这意味着可以通过向量的函数来定义出一种矩阵范数,故通常也将其称之为从属于向量范数的矩阵范数(从属范数)

证明算子范数的定义式满足矩阵范数的定义:

(1) 根据向量范数的非负性知 ( ∗ ) (*) () 的非负性, ∣ ∣ A ∣ ∣ v = 0 ||\bold A||_{v}=0 ∣∣Av=0 当且仅当
∀ x ⃗ ≠ 0 , ∣ ∣ A x ⃗ ∣ ∣ ≡ 0 \forall \vec{x}\ne0,||\bold{A}\vec{x}||\equiv0 x =0∣∣Ax ∣∣0
不妨取
x ⃗ = e ⃗ i   ( i = 1 , 2 , … , n ) \vec{x}=\vec{e}_i\ (i=1,2,\dots,n) x =e i (i=1,2,,n)
则可证得, A ≡ [ 0 ] \bold A\equiv[0] A[0]。其中, e ⃗ i \vec{e}_i e i 为n维向量空间的单位向量。

(2) 对于任意实数 c,
∣ ∣ c A ∣ ∣ v = max ⁡ ∣ ∣ x ⃗ ∣ ∣ v ≠ 0 ∣ ∣ c A x ⃗ ∣ ∣ ∣ ∣ x ⃗ ∣ ∣ = ∣ c ∣ ( max ⁡ ∣ ∣ x ⃗ ∣ ∣ v ≠ 0 ∣ ∣ A x ⃗ ∣ ∣ ∣ ∣ x ⃗ ∣ ∣ ) = ∣ c ∣ ⋅ ∣ ∣ A ∣ ∣ v ||c\bold{A}||_v=\max_{||\vec{x}||_v\ne0}\frac{||c\bold{A}\vec{x}||}{||\vec{x}||}=|c|\left(\max_{||\vec{x}||_v\ne0}\frac{||\bold{A}\vec{x}||}{||\vec{x}||}\right)=|c|\cdot||\bold A||_v ∣∣cAv=∣∣x v=0max∣∣x ∣∣∣∣cAx ∣∣=c(∣∣x v=0max∣∣x ∣∣∣∣Ax ∣∣)=c∣∣Av
(3)
∣ ∣ A + B ∣ ∣ v = max ⁡ ∣ ∣ x ⃗ ∣ ∣ v ≠ 0 ∣ ∣ ( A + B ) x ⃗ ∣ ∣ ∣ ∣ x ⃗ ∣ ∣ = max ⁡ ∣ ∣ x ⃗ ∣ ∣ v ≠ 0 ∣ ∣ A x ⃗ + B x ⃗ ∣ ∣ ∣ ∣ x ⃗ ∣ ∣   ≤ max ⁡ ∣ ∣ x ⃗ ∣ ∣ v ≠ 0 ∣ ∣ A x ⃗ ∣ ∣ ∣ ∣ x ⃗ ∣ ∣ + max ⁡ ∣ ∣ x ⃗ ∣ ∣ v ≠ 0 ∣ ∣ B x ⃗ ∣ ∣ ∣ ∣ x ⃗ ∣ ∣ = ∣ ∣ A ∣ ∣ v + ∣ ∣ B ∣ ∣ v ||\bold{A+B}||_v=\max_{||\vec{x}||_v\ne0}\frac{||\bold{(A+B)}\vec{x}||}{||\vec{x}||}=\max_{||\vec{x}||_v\ne0}\frac{||\bold{A}\vec{x}+\bold{B}\vec{x}||}{||\vec{x}||}\\\ \\ \le\max_{||\vec{x}||_v\ne0}\frac{||\bold{A}\vec{x}||}{||\vec{x}||}+\max_{||\vec{x}||_v\ne0}\frac{||\bold{B}\vec{x}||}{||\vec{x}||}=||\bold A||_v+||\bold B||_v ∣∣A+Bv=∣∣x v=0max∣∣x ∣∣∣∣(A+B)x ∣∣=∣∣x v=0max∣∣x ∣∣∣∣Ax +Bx ∣∣ ∣∣x v=0max∣∣x ∣∣∣∣Ax ∣∣+∣∣x v=0max∣∣x ∣∣∣∣Bx ∣∣=∣∣Av+∣∣Bv
(4) 由 ( ∗ ) (*) () 式:
∣ ∣ A x ⃗ ∣ ∣ ≤ ∣ ∣ A ∣ ∣ v ⋅ ∣ ∣ x ⃗ ∣ ∣ ( 称为:算子范数与向量范数的相容性条件 ) ||\bold{A}\vec{x}||\le||\bold{A}||_v\cdot||\vec{x}||\qquad(称为:算子范数与向量范数的相容性条件) ∣∣Ax ∣∣∣∣Av∣∣x ∣∣(称为:算子范数与向量范数的相容性条件)

∣ ∣ A ( B x ⃗ ) ∣ ∣ ≤ ∣ ∣ A ∣ ∣ v ⋅ ∣ ∣ B x ⃗ ∣ ∣ ≤ ∣ ∣ A ∣ ∣ v ⋅ ∣ ∣ B ∣ ∣ v ⋅ ∣ ∣ x ⃗ ∣ ∣ ||\bold{A(B}\vec{x})||\le||\bold{A}||_v\cdot||\bold{B}\vec{x}||\le||\bold{A}||_v\cdot||\bold{B}||_v\cdot||\vec{x}|| ∣∣A(Bx )∣∣∣∣Av∣∣Bx ∣∣∣∣Av∣∣Bv∣∣x ∣∣
x ⃗ ≠ 0 \vec{x}\ne0 x =0 时,对任意 x ⃗ \vec{x} x 均有:
∣ ∣ A B x ⃗ ∣ ∣ ∣ ∣ x ⃗ ∣ ∣ ≤ ∣ ∣ A ∣ ∣ v ⋅ ∣ ∣ B ∣ ∣ v \frac{||\bold{AB}\vec{x}||}{||\vec{x}||}\le||\bold{A}||_v\cdot||\bold{B}||_v ∣∣x ∣∣∣∣ABx ∣∣∣∣Av∣∣Bv
成立,故
∣ ∣ A B ∣ ∣ v = max ⁡ ∣ ∣ x ⃗ ∣ ∣ ≠ 0 ∣ ∣ A B x ⃗ ∣ ∣ ∣ ∣ x ⃗ ∣ ∣ ≤ ∣ ∣ A ∣ ∣ v ⋅ ∣ ∣ B ∣ ∣ v ( 证毕 ) ||\bold{AB}||_v=\max_{||\vec{x}||\ne0}\frac{||\bold{AB}\vec{x}||}{||\vec{x}||}\le||\bold{A}||_v\cdot||\bold{B}||_v\qquad(证毕) ∣∣ABv=∣∣x ∣∣=0max∣∣x ∣∣∣∣ABx ∣∣∣∣Av∣∣Bv(证毕)

依据算子范数的定义可以得到从属于 ∣ ∣ x ⃗ ∣ ∣ 1 ||\vec{x}||_1 ∣∣x 1 ∣ ∣ x ⃗ ∣ ∣ 2 ||\vec{x}||_2 ∣∣x 2 ∣ ∣ x ⃗ ∣ ∣ ∞ ||\vec{x}||_{\infty} ∣∣x 的矩阵 A m × n \bold A_{m\times n} Am×n的算子范数
( 1 ) ∣ ∣ A ∣ ∣ 1 = max ⁡ ∣ ∣ x ⃗ ∣ ∣ ≠ 0 ∣ ∣ A x ⃗ ∣ ∣ 1 ∣ ∣ x ⃗ ∣ ∣ 1 = max ⁡ j ( ∑ i = 1 m ∣ a i j ∣ ) (列范数) ( 2 ) ∣ ∣ A ∣ ∣ ∞ = max ⁡ ∣ ∣ x ⃗ ∣ ∣ ≠ 0 ∣ ∣ A x ⃗ ∣ ∣ ∞ ∣ ∣ x ⃗ ∣ ∣ ∞ = max ⁡ i ( ∑ j = 1 n ∣ a i j ∣ ) (行范数) ( 3 ) ∣ ∣ A ∣ ∣ 2 = max ⁡ ∣ ∣ x ⃗ ∣ ∣ ≠ 0 ∣ ∣ A x ⃗ ∣ ∣ 2 ∣ ∣ x ⃗ ∣ ∣ 2 = λ m a x ( A T A ) ( 2 − 范数, λ m a x ( A T A ) 表示 A T A 的最大特征值) \begin{aligned} &(1)\quad||\bold A||_1=\max_{||\vec{x}||\ne0}\frac{||\bold{A}\vec{x}||_1}{||\vec{x}||_1}=\max_{j}\left(\sum_{i=1}^m|a_{ij}|\right) (列范数)\\\\ &(2)\quad||\bold A||_{\infty}=\max_{||\vec{x}||\ne0}\frac{||\bold{A}\vec{x}||_{\infty}}{||\vec{x}||_{\infty}}=\max_{i}\left(\sum_{j=1}^n|a_{ij}|\right) (行范数)\\\\ &(3)\quad||\bold A||_2=\max_{||\vec{x}||\ne0}\frac{||\bold{A}\vec{x}||_2}{||\vec{x}||_2}=\sqrt{\lambda_{max}(\bold{A^TA})}(2-范数,\lambda_{max}(\bold{A^TA})表示\bold{A^TA}的最大特征值) \end{aligned} (1)∣∣A1=∣∣x ∣∣=0max∣∣x 1∣∣Ax 1=jmax(i=1maij)(列范数)(2)∣∣A=∣∣x ∣∣=0max∣∣x ∣∣Ax =imax(j=1naij)(行范数)(3)∣∣A2=∣∣x ∣∣=0max∣∣x 2∣∣Ax 2=λmax(ATA) 2范数,λmax(ATA)表示ATA的最大特征值)
证明:

(1) 由于
∣ ∣ A x ⃗ ∣ ∣ 1 = ∑ i = 1 m ∣ a i 1 x 1 + a i 2 x 2 + ⋯ + a i n x n ∣ ≤ ∑ i = 1 m ( ∣ a i 1 ∣ ∣ x 1 ∣ + ∣ a i 2 ∣ ∣ x 2 ∣ + ⋯ + ∣ a i n ∣ ∣ x n ∣ ) = ∑ i = 1 m ∣ a i 1 ∣ ∣ x 1 ∣ + ⋯ + ∑ i = 1 m ∣ a i n ∣ ∣ x n ∣ ≤ max ⁡ j ( ∑ i = 1 m ∣ a i j ∣ ) ( ∣ x 1 ∣ + ∣ x 2 ∣ + ⋯ + ∣ x n ∣ ) \begin{aligned} &||\bold{A}\vec{x}||_1=\sum_{i=1}^m|a_{i1}x_1+a_{i2}x_2+\dots+a_{in}x_n|\\\\ &\qquad\quad\le\sum_{i=1}^m(|a_{i1}||x_1|+|a_{i2}||x_2|+\dots+|a_{in}||x_n|)\\\\ &\qquad\quad=\sum_{i=1}^m|a_{i1}||x_1|+\dots+\sum_{i=1}^m|a_{in}||x_n|\\\\ &\qquad\quad\le\max_{j}\left(\sum_{i=1}^m|a_{ij}|\right)(|x_1|+|x_2|+\dots+|x_n|) \end{aligned} ∣∣Ax 1=i=1mai1x1+ai2x2++ainxni=1m(ai1∣∣x1+ai2∣∣x2++ain∣∣xn)=i=1mai1∣∣x1++i=1main∣∣xnjmax(i=1maij)(x1+x2++xn)
那么,
∀ x ⃗ ≠ 0 , ∣ ∣ A x ⃗ ∣ ∣ 1 ∣ ∣ x ⃗ ∣ ∣ 1 ≤ max ⁡ j ( ∑ i = 1 m ∣ a i j ∣ ) \forall \vec{x}\ne0,\frac{||\bold{A}\vec{x}||_1}{||\vec{x}||_1}\le\max_{j}\left(\sum_{i=1}^m|a_{ij}|\right) x =0∣∣x 1∣∣Ax 1jmax(i=1maij)
A \bold A A k k k 列的元素绝对值之和最大,并取 x ⃗ 0 = [ 0 , … , 0 , 1 k , 0 , … , 0 ] \vec{x}_0=[0,\dots,0,\underset{k}1,0,\dots,0] x 0=[0,,0,k1,0,,0], 则有:
max ⁡ j ( ∑ i = 1 m ∣ a i j ∣ ) = ∑ i = 1 m ∣ a i k ∣ = ∣ ∣ A x ⃗ 0 ∣ ∣ 1 , ∣ ∣ x ⃗ 0 ∣ ∣ ∞ = 1 \max_{j}\left(\sum_{i=1}^m|a_{ij}|\right)=\sum_{i=1}^m|a_{ik}|=||\bold A\vec{x}_0||_1,||\vec{x}_0||_{\infty}=1 jmax(i=1maij)=i=1maik=∣∣Ax 01∣∣x 0=1
这说明:
max ⁡ ∣ ∣ x ⃗ ∣ ∣ ≠ 0 ( ∣ ∣ A x ⃗ ∣ ∣ 1 ∣ ∣ x ⃗ ∣ ∣ 1 ) = max ⁡ j ( ∑ i = 1 m ∣ a i j ∣ ) ( 证毕 ) \max_{||\vec{x}||\ne0}\left(\frac{||\bold{A}\vec{x}||_1}{||\vec{x}||_1}\right)=\max_{j}\left(\sum_{i=1}^m|a_{ij}|\right)\quad(证毕) ∣∣x ∣∣=0max(∣∣x 1∣∣Ax 1)=jmax(i=1maij)(证毕)
(2) 由于
∣ ∣ A x ⃗ ∣ ∣ ∞ = max ⁡ 1 ≤ i ≤ m ∣ a i 1 x 1 + a i 2 x 2 + ⋯ + a i n x n ∣ ≤ max ⁡ 1 ≤ i ≤ m ( ∣ a i 1 ∣ ∣ x 1 ∣ + ∣ a i 2 ∣ ∣ x 2 ∣ + ⋯ + ∣ a i n ∣ ∣ x n ∣ ) ≤ max ⁡ 1 ≤ i ≤ m ( ∣ a i 1 ∣ + ∣ a i 2 ∣ + ⋯ + ∣ a i n ∣ ) ⋅ ∣ ∣ x ⃗ ∣ ∣ ∞ = max ⁡ i ( ∑ j = 1 n ∣ a i j ∣ ) ⋅ ∣ ∣ x ⃗ ∣ ∣ ∞ \begin{aligned} &||\bold{A}\vec{x}||_{\infty}=\max_{1\le i\le m}|a_{i1}x_1+a_{i2}x_2+\dots+a_{in}x_n|\\\\ &\qquad\quad\le\max_{1\le i\le m}(|a_{i1}||x_1|+|a_{i2}||x_2|+\dots+|a_{in}||x_n|)\\\\ &\qquad\quad\le\max_{1\le i\le m}(|a_{i1}|+|a_{i2}|+\dots+|a_{in}|)\cdot||\vec{x}||_{\infty}\\\\ &\qquad\quad=\max_{i}\left(\sum_{j=1}^n|a_{ij}|\right)\cdot||\vec{x}||_{\infty} \end{aligned} ∣∣Ax =1immaxai1x1+ai2x2++ainxn1immax(ai1∣∣x1+ai2∣∣x2++ain∣∣xn)1immax(ai1+ai2++ain)∣∣x =imax(j=1naij)∣∣x
故,
∀ x ⃗ ≠ 0 , ∣ ∣ A x ⃗ ∣ ∣ ∞ ∣ ∣ x ⃗ ∣ ∣ ∞ ≤ max ⁡ i ( ∑ j = 1 n ∣ a i j ∣ ) \forall \vec{x}\ne0,\frac{||\bold{A}\vec{x}||_{\infty}}{||\vec{x}||_{\infty}}\le\max_{i}\left(\sum_{j=1}^n|a_{ij}|\right) x =0∣∣x ∣∣Ax imax(j=1naij)
A \bold A A k k k 行的元素绝对值之和最大,并取 x ⃗ 0 = [ s g n ( x 1 ) , s g n ( x 2 ) , … , s g n ( x n ) ] \vec{x}_0=[sgn(x_1),sgn(x_2),\dots,sgn(x_n)] x 0=[sgn(x1),sgn(x2),,sgn(xn)], 则有:
max ⁡ i ( ∑ j = 1 n ∣ a i j ∣ ) = ∣ ∣ A x ⃗ 0 ∣ ∣ ∞ , ∣ ∣ x ⃗ 0 ∣ ∣ ∞ = 1 \max_{i}\left(\sum_{j=1}^n|a_{ij}|\right)=||\bold A\vec{x}_0||_{\infty},||\vec{x}_0||_{\infty}=1 imax(j=1naij)=∣∣Ax 0∣∣x 0=1
这说明:
max ⁡ ∣ ∣ x ⃗ ∣ ∣ ≠ 0 ( ∣ ∣ A x ⃗ ∣ ∣ ∞ ∣ ∣ x ⃗ ∣ ∣ ∞ ) = max ⁡ i ( ∑ j = 1 n ∣ a i j ∣ ) ( 证毕 ) \max_{||\vec{x}||\ne0}\left(\frac{||\bold{A}\vec{x}||_{\infty}}{||\vec{x}||_{\infty}}\right)=\max_{i}\left(\sum_{j=1}^n|a_{ij}|\right)\quad(证毕) ∣∣x ∣∣=0max(∣∣x ∣∣Ax )=imax(j=1naij)(证毕)
(3) 由于矩阵 A T A \bold{A^TA} ATA 满足:
{ ( A T A ) T = A T A x ⃗ T A T A x ⃗ = ( A x ⃗ , A x ⃗ ) ≥ 0 \begin{cases} (\bold{A}^T\bold{A})^T=\bold{A}^T\bold{A}\\\\ \vec{x}^T\bold{A}^T\bold{A}\vec{x}=(\bold{A}\vec{x},\bold{A}\vec{x})\ge0 \end{cases} (ATA)T=ATAx TATAx =(Ax ,Ax )0
故矩阵 A T A \bold{A^TA} ATA 为对称半正定矩阵,设其有特征值:
λ 1 ≥ λ 2 ⋯ ≥ λ n ≥ 0 \lambda_1\ge\lambda_2\dots\ge\lambda_n\ge0 λ1λ2λn0
分别对应有单位特征向量 u ⃗ i   ( i = 1 , 2 , … , n ) \vec{u}_i\ (i=1,2,\dots,n) u i (i=1,2,,n),且满足
u ⃗ i ⋅ u ⃗ j = δ i j \vec{u}_i\cdot\vec{u}_j=\delta_{ij} u iu j=δij
对于任意 x ⃗ = ∑ i = 1 n c i u ⃗ i ≠ 0 \vec{x}=\sum\limits_{i=1}^nc_i\vec{u}_i\ne0 x =i=1nciu i=0 有:
∣ ∣ A x ⃗ ∣ ∣ 2 ∣ ∣ x ⃗ ∣ ∣ 2 = ( x ⃗ , A T A x ⃗ ) ( x ⃗ , x ⃗ ) = ∑ i = 1 n c i 2 λ i ∑ i = 1 n c i 2 ≤ λ 1 \frac{||\bold{A}\vec{x}||_2}{||\vec{x}||_2}=\frac{(\vec{x},\bold{A^TA}\vec{x})}{(\vec{x},\vec{x})}=\sqrt{\frac{\sum\limits_{i=1}^nc_i^2\lambda_i}{\sum\limits_{i=1}^nc_i^2}}\le\sqrt{\lambda_1} ∣∣x 2∣∣Ax 2=(x ,x )(x ,ATAx )=i=1nci2i=1nci2λi λ1
x ⃗ 0 = u ⃗ 1 \vec{x}_0=\vec{u}_1 x 0=u 1 则有:
∣ ∣ A x ⃗ 0 ∣ ∣ 2 ∣ ∣ x ⃗ 0 ∣ ∣ 2 = λ 1 \frac{||\bold{A}\vec{x}_0||_2}{||\vec{x}_0||_2}=\sqrt{\lambda_1} ∣∣x 02∣∣Ax 02=λ1
说明:
max ⁡ ∣ ∣ x ⃗ ∣ ∣ ≠ 0 ∣ ∣ A x ⃗ ∣ ∣ 2 ∣ ∣ x ⃗ ∣ ∣ 2 = λ 1 ( 证毕 ) \max_{||\vec{x}||\ne0}\frac{||\bold{A}\vec{x}||_2}{||\vec{x}||_2}=\sqrt{\lambda_1}\quad(证毕) ∣∣x ∣∣=0max∣∣x 2∣∣Ax 2=λ1 (证毕)

定理 ∣ ∣ B ∣ ∣ v < 1 ||\bold B||_v<1 ∣∣Bv<1 E ± B \bold{E\pm B} E±B 为可逆矩阵,且
∣ ∣ E ± B − 1 ∣ ∣ v ≤ 1 1 − ∣ ∣ B ∣ ∣ v ||\bold{E\pm B}^{-1}||_v\le\frac{1}{1-||\bold B||_v} ∣∣E±B1v1∣∣Bv1

证明:若 E ± B \bold{E\pm B} E±B 为不可逆,则
( E ± B ) x ⃗ = 0 (\bold{E\pm B})\vec{x}=0 (E±B)x =0
存在非零解,即存在 x ⃗ 0 ≠ 0 \vec{x}_0\ne0 x 0=0使得
B x ⃗ 0 = ∓ x ⃗ 0 ⟹ ∣ ∣ B ∣ ∣ v = ∣ ∣ B x ⃗ 0 ∣ ∣ ∣ ∣ x ⃗ 0 ∣ ∣ = 1 ( 矛盾 ) \bold{B}\vec{x}_0=\mp\vec{x}_0\Longrightarrow||\bold B||_v=\frac{||\bold{B}\vec{x}_0||}{||\vec{x}_0||}=1\quad(矛盾) Bx 0=x 0∣∣Bv=∣∣x 0∣∣∣∣Bx 0∣∣=1(矛盾)
又由于
( E ± B ) − 1 ( E ± B ) = E ⟹ ( E ± B ) − 1 = E ∓ B ( E ± B ) − 1 (\bold{E\pm B})^{-1}(\bold{E\pm B})=\bold{E}\Longrightarrow(\bold{E\pm B})^{-1}=\bold{E}\mp\bold{B(E\pm B)^{-1}} (E±B)1(E±B)=E(E±B)1=EB(E±B)1

∣ ∣ ( E ± B ) − 1 ∣ ∣ v ≤ ∣ ∣ E ∣ ∣ v + ∣ ∣ B ∣ ∣ v ⋅ ∣ ∣ ( E ± B ) − 1 ∣ ∣ v = 1 + ∣ ∣ B ∣ ∣ v ⋅ ∣ ∣ ( E ± B ) − 1 ∣ ∣ v ||(\bold{E\pm B})^{-1}||_v\le||\bold{E}||_v+||\bold{B}||_v\cdot||\bold{(E\pm B)^{-1}}||_v=1+||\bold{B}||_v\cdot||\bold{(E\pm B)^{-1}}||_v ∣∣(E±B)1v∣∣Ev+∣∣Bv∣∣(E±B)1v=1+∣∣Bv∣∣(E±B)1v

∣ ∣ ( E ± B ) − 1 ∣ ∣ v ≤ 1 1 − ∣ ∣ B ∣ ∣ v ( 证毕 ) ||(\bold{E\pm B})^{-1}||_v\le\frac{1}{1-||\bold B||_v}\quad(证毕) ∣∣(E±B)1v1∣∣Bv1(证毕)

Logo

开放原子开发者工作坊旨在鼓励更多人参与开源活动,与志同道合的开发者们相互交流开发经验、分享开发心得、获取前沿技术趋势。工作坊有多种形式的开发者活动,如meetup、训练营等,主打技术交流,干货满满,真诚地邀请各位开发者共同参与!

更多推荐