Weyl不等式

预备知识

定义（Hermite矩阵定义）：复矩阵$A=[a_{ij}]\in M_n$称为Hermite矩阵，是指$A=A^{\ast }$，其中$A^{\ast }\equiv {\bar{A}}^{\top }=[{\bar{a}}_{ji}]$。如果$A=-A^{\ast }$，则称之为斜Hermite矩阵。

定义（酉矩阵定义）：设矩阵$U\in M_n$，若$U^{\ast }U=I$，就称$U$为酉矩阵。

定理（Courant-Fischer）：设$A\in M_n$是具有特征值${\lambda }_1\le {\lambda }_2\le \cdots \le {\lambda }_n$的Hermite矩阵，$k$是给定的整数，$1\le k\le n$，那么$$\underset{w_1,w_2,\cdots ,w_{n-k}\in {\mathbb{C}}^n}{\min }\underset{\stackrel{x\ne 0,x\in {\mathbb{C}}^n}{x\perp w_1,w_2,\cdots ,w_{n-k}}}{\max }\frac{x^{\ast }Ax}{x^{\ast }x}={\lambda }_k$$$$\underset{w_1,w_2,\cdots ,w_{k-1}\in {\mathbb{C}}^n}{\max }\underset{\stackrel{x\ne 0,x\in {\mathbb{C}}^n}{x\perp w_1,w_2,\cdots ,w_{k-1}}}{\min }\frac{x^{\ast }Ax}{x^{\ast }x}={\lambda }_k$$

证明：把$A$写成$A=U\Lambda U^{\ast }$，其中$U$是酉矩阵，而 Λ = d i a g { λ 1 , λ 2 , ⋯   , λ n } \Lambda=\mathrm{diag}\{\lambda_1,\lambda_2,\cdots,\lambda_n\} Λ=diag{λ1​,λ2​,⋯,λn​}，且设$1<k\le n$。如果$x\ne 0$，则$$\frac{x^{\ast }Ax}{x^{\ast }x}=\frac{(U^{\ast }x{)}^{\ast }\Lambda (U^{\ast }x)}{x^{\ast }x}=\frac{(U^{\ast }x{)}^{\ast }\Lambda (U^{\ast }x)}{(U^{\ast }x{)}^{\ast }(U^{\ast }x)}$$且 { U ∗ x : x ∈ C n , x ≠ 0 } = { y ∈ C n , y ≠ 0 } \{U^{*}x:x\in \mathbb{C}^n,x\ne 0\}=\{y \in \mathbb{C}^n,y\ne 0\} {U∗x:x∈Cn,x​=0}={y∈Cn,y​=0}。因此给定$w_1,w_2,\cdots ,w_{n-k}\in {\mathbb{C}}^n$，便有$$\begin{array}{rl}\underset{\stackrel{x\ne 0}{x\perp w_1,\cdots ,w_{n-k}}}{\sup }\frac{x^{\ast }Ax}{x^{\ast }x}& =\underset{\stackrel{y\ne 0}{y\perp U^{\ast }{w}_1,\cdots ,U^{\ast }{w}_{n-k}}}{\sup }\frac{y^{\ast }\Lambda y}{y^{\ast }y}\\ & =\underset{\stackrel{y^{\ast }y=1}{y\perp U^{\ast }{w}_1,\cdots ,U^{\ast }{w}_{n-k}}}{\sup }\sum _{i=1}^n{\lambda }_i|y_i{|}^2\\ & \ge \underset{\stackrel{\stackrel{y^{\ast }y=1}{y\perp U^{\ast }{w}_1,\cdots ,U^{\ast }{w}_{n-k}}}{y_1=y_2=\cdots y_{k-1}=0}}{\sup }\sum _{i=1}^n{\lambda }_i|y_i{|}^2\\ & =\underset{\stackrel{|y_k{|}^2+|y_{k+1}{|}^2+\cdots +|y_n{|}^2=1}{y\perp U^{\ast }{w}_1,\cdots ,U^{\ast }{w}_{n-k}}}{\sup }\sum _{i=k}^n{\lambda }_i|y_i{|}^2\ge {\lambda }_k\end{array}$$这说明，对于任意$n-k$个向量$w_1,w_2,\cdots ,w_{n-k}$ $$\underset{\stackrel{x\ne 0}{x\perp w_1,\cdots ,w_{n-k}}}{\sup }\frac{x^{\ast }Ax}{x^{\ast }x}\ge {\lambda }_k$$为了使等式成立，存在向量$w_i=u_{n-i+1}$，其中$U=[u_1,\cdots ,u_n]$，则有$$\underset{w_1,\cdots ,w_{n-k}}{\inf }\underset{\stackrel{x\ne 0}{x\perp w_1,\cdots ,w_{n-k}}}{\sup }\frac{x^{\ast }Ax}{x^{\ast }x}={\lambda }_k$$又因为极值是可以达到的，所以用$\min$和$\max$替换$\inf$和$\sup$。

Weyl不等式

定理（Weyl不等式）：设$A$,$B\in M_n$是Hermite矩阵，又设各个特征值${\lambda }_i(A),{\lambda }_i(B)$以及${\lambda }_i(A+B)$均按照递增顺序排列，则对每个$k=1,2,\cdots ,n$有$${\lambda }_k(A)+{\lambda }_1(B)\le {\lambda }_k(A+B)\le {\lambda }_k(A)+{\lambda }_n(B)$$

证明：对于任意复数$x\in {\mathbb{C}}^n$，有不等式$${\lambda }_1(B)\le \frac{x^{\ast }Bx}{x^{\ast }x}\le {\lambda }_n(B)$$因而对任意$k=1,2,\cdots ,n$有$$\begin{array}{rl}{\lambda }_k(A+B)& =\underset{w_1,\cdots ,w_n-k\in {\mathbb{C}}^n}{\min }\underset{\stackrel{x\ne 0}{x\perp w_1,\cdots ,w_{n-k}}}{\max }\frac{x^{\ast }(A+B)x}{x^{\ast }x}\\ & =\underset{w_1,\cdots ,w_n-k\in {\mathbb{C}}^n}{\min }\underset{\stackrel{x\ne 0}{x\perp w_1,\cdots ,w_{n-k}}}{\max }\left[\frac{x^{\ast }Ax}{x^{\ast }x}+\frac{x^{\ast }Bx}{x^{\ast }x}\right]\\ & \ge \underset{w_1,\cdots ,w_n-k\in {\mathbb{C}}^n}{\min }\underset{\stackrel{x\ne 0}{x\perp w_1,\cdots ,w_{n-k}}}{\max }\left[\frac{x^{\ast }Ax}{x^{\ast }x}+{\lambda }_1(B)\right]={\lambda }_k(A)+{\lambda }_1(B)\end{array}$$

推论（Weyl不等式）：设$A$,$B\in M_n$是Hermite矩阵。假定$B$是半正定矩阵，且$A$和$A+B$均按照递增顺序排列，则对每个$k=1,2,\cdots ,n$有$${\lambda }_k(A)\le {\lambda }_k(A+B)$$