谱半径学习

谱半径

设${\lambda }_1,\cdots ,{\lambda }_n$为$\mathbf{A}\in {\mathbb{C}}^{n\times n}$的特征值，则谱半径定义为
ρ ( A ) = max ⁡ { ∣ λ 1 ∣ , ⋯   , ∣ λ n ∣ } \rho\left(\mathbf{A}\right)=\max\left\{\left|\lambda_1\right|,\cdots, \left|\lambda_n\right|\right\} ρ(A)=max{∣λ1​∣,⋯,∣λn​∣}

注意$\Vert \mathbf{Av}\Vert \le \rho \left(\mathbf{A}\right)\Vert \mathbf{v}\Vert$不一定成立
$${\mathbf{C}}_r=\left(\begin{array}{cc}0& \frac{1}{r}\\ r& 0\end{array}\right)$$
其中$r>1$
则${\mathbf{C}}_r$的特征值为$\pm 1$
${\mathbf{C}}_r{\mathbf{e}}_1=r{\mathbf{e}}_2$
$$\Vert {\mathbf{C}}_r{\mathbf{e}}_1\Vert =r>1=\rho \left({\mathbf{C}}_r\right)\Vert {\mathbf{e}}_1\Vert$$
如果$\mathbf{A}$是Hermitian矩阵的话则$\Vert \mathbf{Av}\Vert \le \rho \left(\mathbf{A}\right)\Vert \mathbf{v}\Vert$

性质

性质1

$$\rho \left(\mathbf{A}\right)\le \Vert \mathbf{A}\Vert$$
其中$\Vert \cdot \Vert$是任意算子范数

证明：
设$\lambda$是$\mathbf{A}$的特征值，$\mathbf{x}$是对应的特征向量
$$\left|\lambda \right|\Vert \mathbf{x}\Vert =\Vert \lambda \mathbf{x}\Vert =\Vert \mathbf{Ax}\Vert \le \Vert \mathbf{A}\Vert \Vert \mathbf{x}\Vert \Rightarrow \rho \left(\mathbf{A}\right)\le \Vert \mathbf{A}\Vert$$

性质2

对于$ϵ>0$，存在某种矩阵范数$\Vert \cdot \Vert$,使得
$$\Vert \mathbf{A}\Vert \le \rho \left(\mathbf{A}\right)+ϵ$$

证明：
由若尔当分解定理
$$\mathbf{A}=\mathbf{S}\left(\begin{array}{cccc}{\mathbf{J}}_{n_1}\left({\lambda }_1\right)& 0& \cdots & 0\\ 0& {\mathbf{J}}_{n_2}\left({\lambda }_2\right)& \ddots & ⋮\\ ⋮& \ddots & \ddots & 0\\ 0& \cdots & 0& {\mathbf{J}}_{n_k}\left({\lambda }_k\right)\end{array}\right){\mathbf{S}}^{-1}$$
其中$\mathbf{S}$是可逆矩阵，${\lambda }_1,\cdots ,{\lambda }_k$是$\mathbf{A}$的特征值，$n_1+\cdots +n_k=n$

设
$$\mathbf{D}\left(\eta \right)=\left(\begin{array}{cccc}{\mathbf{D}}_{n_1}\left(\eta \right)& 0& \cdots & 0\\ 0& {\mathbf{D}}_{n_2}\left(\eta \right)& \ddots & ⋮\\ ⋮& \ddots & \ddots & 0\\ 0& \cdots & 0& {\mathbf{D}}_{n_k}\left(\eta \right)\end{array}\right)$$
其中
$${\mathbf{D}}_m\left(\eta \right)=\left(\begin{array}{cccc}\eta & 0& \cdots & 0\\ 0& {\eta }^2& \ddots & ⋮\\ ⋮& \ddots & \ddots & 0\\ 0& \cdots & 0& {\eta }^m\end{array}\right)$$

$$\mathbf{D}\left(\frac{1}{ϵ}\right){\mathbf{S}}^{-1}\mathbf{ASD}\left(ϵ\right)=\left(\begin{array}{cccc}{\mathbf{B}}_{n_1}\left({\lambda }_1,ϵ\right)& 0& \cdots & 0\\ 0& {\mathbf{B}}_{n_2}\left({\lambda }_2,ϵ\right)& \ddots & ⋮\\ ⋮& \ddots & \ddots & 0\\ 0& \cdots & 0& {\mathbf{B}}_{n_k}\left({\lambda }_k,ϵ\right)\end{array}\right)$$
其中
$${\mathbf{B}}_m\left(\lambda ,ϵ\right)={\mathbf{D}}_m\left(\frac{1}{ϵ}\right){\mathbf{J}}_m\left(\lambda \right){\mathbf{D}}_m\left(ϵ\right)=\left(\begin{array}{ccccc}\lambda & ϵ& 0& \cdots & 0\\ 0& \lambda & ϵ& 0& ⋮\\ 0& \ddots & \ddots & \ddots & 0\\ ⋮& \ddots & \ddots & \lambda & ϵ\\ 0& \cdots & 0& 0& \lambda \end{array}\right)$$

定义矩阵范数
$$\Vert \mathbf{A}\Vert =\Vert \mathbf{D}\left(\frac{1}{ϵ}\right){\mathbf{S}}^{-1}\mathbf{ASD}\left(ϵ\right){\Vert }_1$$
接下来验证$\Vert \cdot \Vert$是矩阵范数

容易验证非负性和正齐次性

三角不等式：
$$\begin{array}{rl}\Vert \mathbf{A}+\mathbf{B}\Vert & =\Vert \mathbf{D}\left(\frac{1}{ϵ}\right){\mathbf{S}}^{-1}\left(\mathbf{A}+\mathbf{B}\right)\mathbf{SD}\left(ϵ\right){\Vert }_1\\ & =\Vert \mathbf{D}\left(\frac{1}{ϵ}\right){\mathbf{S}}^{-1}\mathbf{ASD}\left(ϵ\right)+\mathbf{D}\left(\frac{1}{ϵ}\right){\mathbf{S}}^{-1}\mathbf{BSD}\left(ϵ\right){\Vert }_1\\ & \le \Vert \mathbf{D}\left(\frac{1}{ϵ}\right){\mathbf{S}}^{-1}\mathbf{ASD}\left(ϵ\right){\Vert }_1+\Vert \mathbf{D}\left(\frac{1}{ϵ}\right){\mathbf{S}}^{-1}\mathbf{BSD}\left(ϵ\right){\Vert }_1\\ & =\Vert \mathbf{A}\Vert +\Vert \mathbf{B}\Vert \end{array}$$

次乘性：

因为$\mathbf{D}\left(\frac{1}{ϵ}\right)\mathbf{D}\left(ϵ\right)=\mathbf{I}$
$$\begin{array}{rl}\Vert \mathbf{AB}\Vert & =\Vert \mathbf{D}\left(\frac{1}{ϵ}\right){\mathbf{S}}^{-1}\mathbf{AB}\mathbf{SD}\left(ϵ\right){\Vert }_1\\ & =\Vert \mathbf{D}\left(\frac{1}{ϵ}\right){\mathbf{S}}^{-1}\mathbf{ASD}\left(ϵ\right)\mathbf{D}\left(\frac{1}{ϵ}\right){\mathbf{S}}^{-1}\mathbf{BSD}\left(ϵ\right){\Vert }_1\\ & \le \Vert \mathbf{D}\left(\frac{1}{ϵ}\right){\mathbf{S}}^{-1}\mathbf{ASD}\left(ϵ\right){\Vert }_1\Vert \mathbf{D}\left(\frac{1}{ϵ}\right){\mathbf{S}}^{-1}\mathbf{BSD}\left(ϵ\right){\Vert }_1\\ & =\Vert \mathbf{A}\Vert \Vert \mathbf{B}\Vert \end{array}$$
所以$\Vert \cdot \Vert$是矩阵范数

于是
∥ A ∥ = max ⁡ i ∈ { 1 , 2 , ⋯   , n } ( ∣ λ i ∣ + ϵ ) = ρ ( A ) + ϵ \|\mathbf{A}\|=\max_{i\in\left\{1,2,\cdots,n\right\}}\left(\left|\lambda_i\right|+\epsilon\right)=\rho\left(\mathbf{A}\right)+\epsilon ∥A∥=i∈{1,2,⋯,n}max​(∣λi​∣+ϵ)=ρ(A)+ϵ

性质3

$\rho \left(\mathbf{A}\right)<1\iff \underset{k\to \infty }{\lim }{\mathbf{A}}^k=0$

证明：
设$\mathbf{A}$的特征值为$\lambda$,对应的特征向量为$\mathbf{v}$
假设$\underset{k\to \infty }{\lim }{\mathbf{A}}^k=0$
$$\begin{array}{rl}0& =\left(\underset{k\to \infty }{\lim }{\mathbf{A}}^k\right)\mathbf{v}\\ & =\underset{k\to \infty }{\lim }\left({\mathbf{A}}^k\mathbf{v}\right)\\ & =\underset{k\to \infty }{\lim }\left({\lambda }^k\mathbf{v}\right)\\ & =\mathbf{v}\underset{k\to \infty }{\lim }{\lambda }^k\end{array}$$
因为$\mathbf{v}\ne 0$,
$$\underset{k\to \infty }{\lim }{\lambda }^k=0\Rightarrow \left|\lambda \right|<1\Rightarrow \rho \left(\mathbf{A}\right)<1$$

假设$\rho \left(\mathbf{A}\right)<1$
由若尔当分解定理
$$\mathbf{A}=\mathbf{S}\left(\begin{array}{cccc}{\mathbf{J}}_{n_1}\left({\lambda }_1\right)& 0& \cdots & 0\\ 0& {\mathbf{J}}_{n_2}\left({\lambda }_2\right)& \ddots & ⋮\\ ⋮& \ddots & \ddots & 0\\ 0& \cdots & 0& {\mathbf{J}}_{n_k}\left({\lambda }_k\right)\end{array}\right){\mathbf{S}}^{-1}$$
其中$\mathbf{S}$是可逆矩阵，${\lambda }_1,\cdots ,{\lambda }_k$是$\mathbf{A}$的特征值，$n_1+\cdots +n_k=n$
由若尔当块的性质,对于充分大的$k$,有
$${\mathbf{J}}_{n_i}^k\left({\lambda }_i\right)=\left[\begin{array}{ccccc}{\lambda }_i^k& \left(\begin{array}{c}k\\ 1\end{array}\right){\lambda }_i^{k-1}& \left(\begin{array}{c}k\\ 2\end{array}\right){\lambda }_i^{k-2}& \cdots & \left(\begin{array}{c}k\\ n_i-1\end{array}\right){\lambda }_i^{k-n_i+1}\\ 0& {\lambda }_i^k& \left(\begin{array}{c}k\\ 1\end{array}\right){\lambda }_i^{k-1}& \cdots & \left(\begin{array}{c}k\\ n_i-2\end{array}\right){\lambda }_i^{k-n_i+2}\\ ⋮& ⋮& \ddots & \ddots & ⋮\\ 0& 0& \dots & {\lambda }_i^k& \left(\begin{array}{c}k\\ 1\end{array}\right){\lambda }_i^{k-1}\\ 0& 0& \cdots & 0& {\lambda }_i^k\end{array}\right]$$
所以
$$\underset{k\to \infty }{\lim }{\mathbf{J}}_{n_i}^k=0\Rightarrow \underset{k\to \infty }{\lim }{\mathbf{J}}^k=0\Rightarrow \underset{k\to \infty }{\lim }{\mathbf{A}}^k=\underset{k\to \infty }{\lim }\mathbf{S}{\mathbf{J}}^k{\mathbf{S}}^{-1}=\mathbf{S}\left(\underset{k\to \infty }{\lim }{\mathbf{J}}^k\right){\mathbf{S}}^{-1}=0$$

Gelfand定理

$$\rho \left(\mathbf{A}\right)=\underset{k\to \infty }{\lim }\Vert {\mathbf{A}}^k{\Vert }^{\frac{1}{k}}$$

证明：$k\ge 0$
$$\rho {\left(\mathbf{A}\right)}^k=\rho \left({\mathbf{A}}^k\right)=\Vert {\mathbf{A}}^k\Vert \Rightarrow \rho \left(\mathbf{A}\right)\le \Vert {\mathbf{A}}^k{\Vert }^{\frac{1}{k}}\Rightarrow \rho \left(\mathbf{A}\right)\le \underset{k\to \infty }{\lim }\Vert {\mathbf{A}}^k{\Vert }^{\frac{1}{k}}$$

存在范数$\Vert \cdot {\Vert }_M$,使得$\Vert \mathbf{A}{\Vert }_M\le \rho \left(\mathbf{A}\right)+ϵ$
($\Vert \cdot {\Vert }_M$的$M$主要是为了和上面的范数区分)
由范数的等价性
$$\exists C>0,s.t.\Vert \cdot \Vert \le C\Vert \cdot {\Vert }_M$$
所以
$$\begin{gathered} \Vert {\mathbf{A}}^k\Vert \le C\Vert {\mathbf{A}}^k{\Vert }_M\le C\Vert \mathbf{A}{\Vert }_M^k\le C{\left(\rho \left(\mathbf{A}\right)+ϵ\right)}^k \\ \Vert {\mathbf{A}}^k{\Vert }^{\frac{1}{k}}\le C^{\frac{1}{k}}\left(\rho \left(\mathbf{A}\right)+ϵ\right) \\ \underset{k\to \infty }{\lim }\Vert {\mathbf{A}}^k{\Vert }^{\frac{1}{k}}\le \rho \left(\mathbf{A}\right)+ϵ \end{gathered}$$

参考：
https://www.math.drexel.edu/~foucart/TeachingFiles/F12/M504Lect6.pdf
https://en.wikipedia.org/wiki/Spectral_radius