跳转至

向量和矩阵范数

向量范数的定义

\(\mathbb{R}^n\) 上的一个函数 \(\left\Vert\cdot\right\Vert\) 为范数,若满足:

  1. 非负性\(\left\Vert\mathbf{x}\right\Vert\ge0\), \(\left\Vert\mathbf{x}\right\Vert=0\)当且仅当\(\mathbf{x}=\mathbf{0}\)

  2. 齐次性\(\left\Vert\alpha\mathbf{x}\right\Vert=\left\vert \alpha \right\vert\left\Vert\mathbf{x}\right\Vert, \alpha\in\mathbb{R}\)

  3. 三角不等式\(\left\Vert\mathbf{x}+\mathbf{y}\right\Vert\le\left\Vert\mathbf{x}\right\Vert+\left\Vert\mathbf{y}\right\Vert\)

常见的向量范数有:

  • 1-范数\(\left\Vert\mathbf{x}\right\Vert_{1}=\sum_{i=1}^{n}\left\vert x_{i} \right\vert\)

  • 2-范数\(\left\Vert\mathbf{x}\right\Vert_{2}=\left(\sum_{i=1}^{n}\left\vert x_{i} \right\vert^{2}\right)^{\frac 1{2}}\)

  • \(\infty\)-范数\(\left\Vert\mathbf{x}\right\Vert_{\infty}=\max\limits_{1\le i\le n}\left\vert x_{i} \right\vert\)

向量范数的性质

向量范数的等价性

对于 \(\mathbb{R}^n\) 中的任意两种范数 \(\left\Vert\mathbf{x}\right\Vert_{\alpha}\)\(\left\Vert\mathbf{x}\right\Vert_{\beta}\),都存在两个正数 \(c_1,c_2\),使得对任意 \(\mathbf{x}\in\mathbb{R}^n\) 都有:

\[ c_1\left\Vert\mathbf{x}\right\Vert_{\alpha}\le\left\Vert\mathbf{x}\right\Vert_{\beta}\le c_2\left\Vert\mathbf{x}\right\Vert_{\alpha} \]

Proof.

\(f(\mathbf{x})=\left\Vert\mathbf{x}\right\Vert_{\beta}\)\(\mathbf{S}=\{\mathbf{x}\in\mathbb{R}^n: \left\Vert\mathbf{x}\right\Vert_{\alpha}= 1\}\) 为一个有界闭集,则连续函数 \(f(\mathbf{x})\)\(\mathbf{S}\) 上存在最小值 \(c_1=\min\limits_{\mathbf{x}\in\mathbf{S}}f(\mathbf{x})\) 和最大值 \(c_2=\max\limits_{\mathbf{x}\in\mathbf{S}}f(\mathbf{x})\)

对于 \(\mathbf{x}\ne \mathbf{0}\)\(\frac{\mathbf{x}}{\left\Vert\mathbf{x}\right\Vert_{\alpha}}\in \mathbf{S}\),则有:

\[ c_1\le f\left(\frac{\mathbf{x}}{\left\Vert\mathbf{x}\right\Vert_{\alpha}}\right)=\frac{\left\Vert\mathbf{x}\right\Vert_{\beta}}{\left\Vert\mathbf{x}\right\Vert_{\alpha}}\le c_2 \]

即:

\[ c_1\left\Vert\mathbf{x}\right\Vert_{\alpha}\le\left\Vert\mathbf{x}\right\Vert_{\beta}\le c_2\left\Vert\mathbf{x}\right\Vert_{\alpha} \]

\(\square\)

对于常用的向量范数,有如下关系:

\[ \begin{aligned} \left\Vert\mathbf{x}\right\Vert_2\le\left\Vert\mathbf{x}\right\Vert_1\le\sqrt{n}\left\Vert\mathbf{x}\right\Vert_2 \\ \left\Vert\mathbf{x}\right\Vert_{\infty}\le\left\Vert\mathbf{x}\right\Vert_1\le n\left\Vert\mathbf{x}\right\Vert_{\infty} \\ \left\Vert\mathbf{x}\right\Vert_{\infty}\le\left\Vert\mathbf{x}\right\Vert_2\le\sqrt{n}\left\Vert\mathbf{x}\right\Vert_{\infty} \end{aligned} \]

向量序列的收敛性

在空间 \(\mathbb{R}^n\) 中,向量序列 \(\{\mathbf{x}^{(k)}\}\) 收敛于向量 \(\mathbf{x}^*\) 的充要条件是存在范数 \(\left\Vert\cdot\right\Vert\) 使得:

\[ \lim_{k\to\infty}\left\Vert\mathbf{x}^{(k)}-\mathbf{x}^*\right\Vert=0 \]

压缩映射

设有非空集合 \(\mathbf{D}\subset\mathbb{R}^n\),对于映射 \(\mathbf{f}:\mathbf{D}\to\mathbf{D}\),若存在范数\(\left\Vert\cdot\right\Vert\) 和常数 \(q\in[0,1)\) 使得对任意 \(\mathbf{x},\mathbf{y}\in\mathbf{D}\) 都有

\[ \left\Vert \mathbf{f}(\mathbf{x})-\mathbf{f}(\mathbf{y})\right\Vert\le q\left\Vert \mathbf{x}-\mathbf{y}\right\Vert \]

则称 \(\mathbf{f}\)\(\mathbf{D}\) 上的压缩映射。

Banach压缩映射原理

\(\mathbf{D}\subset\mathbb{R}^n\) 为闭集,映射 \(\mathbf{f}\)\(\mathbf{D}\) 上的压缩映射,则 \(\mathbf{f}\)\(\mathbf{D}\) 上有唯一不动点 \(\mathbf{x}\),使得 \(\mathbf{f}(\mathbf{x})=\mathbf{x}\)

Proof.

\(\mathbf{x}^{(0)}\in\mathbf{D}\),构造序列\(\{\mathbf{x}^{(k)}\}\) 如下:

\[ \mathbf{x}^{(k+1)}=\mathbf{f}(\mathbf{x}^{(k)}),\quad k=0,1,2,\ldots \]

则对任意 \(k\ge0\),有

\[ \left\Vert\mathbf{x}^{(k+1)}-\mathbf{x}^{(k)}\right\Vert=\left\Vert\mathbf{f}(\mathbf{x}^{(k)})-\mathbf{f}(\mathbf{x}^{(k-1)})\right\Vert\le q\left\Vert\mathbf{x}^{(k)}-\mathbf{x}^{(k-1)}\right\Vert \]

由此可得

\[ \left\Vert\mathbf{x}^{(k+1)}-\mathbf{x}^{(k)}\right\Vert\le q^k\left\Vert\mathbf{x}^{(1)}-\mathbf{x}^{(0)}\right\Vert \]

对于 \(m>n\ge0\),有

\[ \left\Vert\mathbf{x}^{(m)}-\mathbf{x}^{(n)}\right\Vert\le\sum_{k=n}^{m-1}\left\Vert\mathbf{x}^{(k+1)}-\mathbf{x}^{(k)}\right\Vert\le\sum_{k=n}^{m-1}q^k\left\Vert\mathbf{x}^{(1)}-\mathbf{x}^{(0)}\right\Vert=\frac{q^n-q^m}{1-q}\left\Vert\mathbf{x}^{(1)}-\mathbf{x}^{(0)}\right\Vert \]

由于 \(q\in[0,1)\),当 \(n\to\infty\)

\[ \lim_{n\to\infty}\left\Vert\mathbf{x}^{(m)}-\mathbf{x}^{(n)}\right\Vert=0 \]

即序列 \(\{\mathbf{x}^{(k)}\}\) 为 Cauchy 序列,故在 \(\mathbb{R}^n\) 中收敛,设其极限为 \(\mathbf{x}^*\in\mathbf{D}\),则由映射的连续性可得

\[ \mathbf{x}^*=\lim_{k\to\infty}\mathbf{x}^{(k+1)}=\lim_{k\to\infty}\mathbf{f}(\mathbf{x}^{(k)})=\mathbf{f}(\lim_{k\to\infty}\mathbf{x}^{(k)})=\mathbf{f}(\mathbf{x}^*) \]

\(\mathbf{x}^*\)\(\mathbf{f}\) 的不动点。 设存在不动点 \(\mathbf{x}_1,\mathbf{x}_2\),则有:

\[ \left\Vert\mathbf{x}_1-\mathbf{x}_2\right\Vert=\left\Vert\mathbf{f}(\mathbf{x}_1)-\mathbf{f}(\mathbf{x}_2)\right\Vert\le q\left\Vert\mathbf{x}_1-\mathbf{x}_2\right\Vert \]

由于 \(q\in[0,1)\),上式只在 \(\mathbf{x}_1=\mathbf{x}_2\) 时成立,故不动点唯一。 \(\square\)

矩阵范数的定义

\(\mathbb{R}^{n\times n}\) 上的一个函数 \(\left\Vert\cdot\right\Vert\) 为范数,若满足:

  1. 非负性\(\left\Vert\mathbf{A}\right\Vert\ge0\), \(\left\Vert\mathbf{A}\right\Vert=0\) 当且仅当 \(\mathbf{A}=\mathbf{0}\)

  2. 齐次性\(\left\Vert\alpha\mathbf{A}\right\Vert=\left\vert \alpha \right\vert\left\Vert\mathbf{A}\right\Vert, \alpha\in\mathbb{R}\)

  3. 三角不等式\(\left\Vert\mathbf{A}+\mathbf{B}\right\Vert\le\left\Vert\mathbf{A}\right\Vert+\left\Vert\mathbf{B}\right\Vert\)

  4. 矩阵乘法不等式\(\left\Vert\mathbf{A}\mathbf{B}\right\Vert\le\left\Vert\mathbf{A}\right\Vert\cdot\left\Vert\mathbf{B}\right\Vert\)

常见的矩阵范数有:

  • 列范数\(\left\Vert\mathbf{A}\right\Vert_{1}=\max\limits_{1\le j\le n}\sum_{i=1}^{n}\left\vert a_{ij} \right\vert\)

  • 行范数\(\left\Vert\mathbf{A}\right\Vert_{\infty}=\max\limits_{1\le i\le n}\sum_{j=1}^{n}\left\vert a_{ij} \right\vert\)

  • 谱范数\(\left\Vert\mathbf{A}\right\Vert_{2}=\sqrt{\lambda_{\max}(\mathbf{A}^T\mathbf{A})}\)

  • F-范数\(\left\Vert\mathbf{A}\right\Vert_{F}=\sqrt{\sum_{i=1}^{n}\sum_{j=1}^{n}\left\vert a_{ij} \right\vert^2}\)

矩阵范数的性质

算子范数

称向量范数导出的矩阵范数为算子范数,定义如下:

\[ \left\Vert\mathbf{A}\right\Vert=\max_{\mathbf{x}\ne \mathbf{0}}\frac{\left\Vert\mathbf{A}\mathbf{x}\right\Vert}{\left\Vert\mathbf{x}\right\Vert}=\max_{\left\Vert\mathbf{x}\right\Vert=1}\left\Vert\mathbf{A}\mathbf{x}\right\Vert \]

由定义可知算子范数是矩阵范数,且与向量范数相容:

\[ \left\Vert\mathbf{A}\mathbf{x}\right\Vert\le\left\Vert\mathbf{A}\right\Vert\cdot\left\Vert\mathbf{x}\right\Vert \]

矩阵范数的等价性

对于 \(\mathbb{R}^{n\times n}\) 中的任意两种范数 \(\left\Vert\mathbf{A}\right\Vert_{\alpha}\)\(\left\Vert\mathbf{A}\right\Vert_{\beta}\),都存在两个正数\(c_1,c_2\), 使得对任意 \(\mathbf{A}\in\mathbb{R}^{n\times n}\) 都有

\[ c_1\left\Vert\mathbf{A}\right\Vert_{\alpha}\le\left\Vert\mathbf{A}\right\Vert_{\beta}\le c_2\left\Vert\mathbf{A}\right\Vert_{\alpha} \]

对于常用的矩阵范数,有如下关系:

  1. \(\frac 1n\left\Vert\mathbf{A}\right\Vert_{\infty}\le\left\Vert\mathbf{A}\right\Vert_{1}\le n\left\Vert\mathbf{A}\right\Vert_{\infty}\)

Proof.

对于任意\(a_{ij}\in\mathbf{A}\),有

\[ \left\vert a_{ij} \right\vert\le\left\Vert\mathbf{A}\right\Vert_{\infty} \]

则有

\[ \sum_{i=1}^{n}\left\vert a_{ij} \right\vert\le n\left\Vert\mathbf{A}\right\Vert_{\infty} \]

取最大值可得

\[ \left\Vert\mathbf{A}\right\Vert_{1}=\max_{1\le j\le n}\sum_{i=1}^{n}\left\vert a_{ij} \right\vert\le n\left\Vert\mathbf{A}\right\Vert_{\infty} \]

同理可得

$$ \left\Vert\mathbf{A}\right\Vert_{\infty}=\max_{1\le i\le n}\sum_{j=1}^{n}\left\vert a_{ij} \right\vert\le n\left\Vert\mathbf{A}\right\Vert_{1} $$ \(\square\)

  1. \(\frac{1}{\sqrt{n}}\left\Vert\mathbf{A}\right\Vert_{\infty}\le\left\Vert\mathbf{A}\right\Vert_{2}\le\sqrt{n}\left\Vert\mathbf{A}\right\Vert_{\infty}\)

Proof.

先证明 \(\left\Vert\mathbf{x}\right\Vert_{\infty}\le\left\Vert\mathbf{x}\right\Vert_{2}\le\sqrt{n}\left\Vert\mathbf{x}\right\Vert_{\infty}\)

对于任意 \(x_i\in\mathbf{x}\),有 \(\left\vert x_i \right\vert\le\left\Vert\mathbf{x}\right\Vert_{2}\),取最大值可得 \(\left\Vert\mathbf{x}\right\Vert_{\infty}\le\left\Vert\mathbf{x}\right\Vert_{2}\)

对于任意 \(x_i\in\mathbf{x}\),有 \(\left\vert x_i \right\vert\le\left\Vert\mathbf{x}\right\Vert_{\infty}\),则有 \(\left\vert x_i \right\vert^2\le\left\Vert\mathbf{x}\right\Vert_{\infty}^2\),对所有 \(i\) 求和可得 \(\left\Vert\mathbf{x}\right\Vert_{2}^2\le n\left\Vert\mathbf{x}\right\Vert_{\infty}^2\),即 \(\left\Vert\mathbf{x}\right\Vert_{2}\le\sqrt{n}\left\Vert\mathbf{x}\right\Vert_{\infty}\)

由算子范数定义可得

\[ \left\Vert\mathbf{A}\right\Vert_{2}=\max_{\mathbf{x}\ne \mathbf{0}}\frac{\left\Vert\mathbf{A}\mathbf{x}\right\Vert_{2}}{\left\Vert\mathbf{x}\right\Vert_{2}}, \; \left\Vert\mathbf{A}\right\Vert_{\infty}=\max_{\mathbf{x}\ne \mathbf{0}}\frac{\left\Vert\mathbf{A}\mathbf{x}\right\Vert_\infty}{\left\Vert\mathbf{x}\right\Vert_{\infty}} \]

对于任意 \(\mathbf{x}\ne \mathbf{0}\),有

\[ \frac{\left\Vert\mathbf{A}\mathbf{x}\right\Vert_2}{\left\Vert\mathbf{x}\right\Vert_2}\le\frac{\sqrt{n}\left\Vert\mathbf{A}\mathbf{x}\right\Vert_\infty}{\left\Vert\mathbf{x}\right\Vert_2}\le\frac{\sqrt{n}\left\Vert\mathbf{A}\mathbf{x}\right\Vert_\infty}{\left\Vert\mathbf{x}\right\Vert_\infty}\le\sqrt{n}\left\Vert\mathbf{A}\right\Vert_{\infty} \]

取最大值可得

\[ \left\Vert\mathbf{A}\right\Vert_{2}\le\sqrt{n}\left\Vert\mathbf{A}\right\Vert_{\infty} \]

对于任意 \(\mathbf{x}\ne \mathbf{0}\),有

\[ \frac{\left\Vert\mathbf{A}\mathbf{x}\right\Vert_\infty}{\left\Vert\mathbf{x}\right\Vert_\infty}\le\frac{\left\Vert\mathbf{A}\mathbf{x}\right\Vert_2}{\left\Vert\mathbf{x}\right\Vert_\infty}\le\sqrt{n}\frac{\left\Vert\mathbf{A}\mathbf{x}\right\Vert_2}{\left\Vert\mathbf{x}\right\Vert_2}\le\sqrt{n}\left\Vert\mathbf{A}\right\Vert_{2} \]

取最大值可得

\[ \left\Vert\mathbf{A}\right\Vert_{\infty}\le\sqrt{n}\left\Vert\mathbf{A}\right\Vert_{2} \]

\(\square\)