论文: Eigenvectors from eigenvalues 提出并简要证明了一种根据矩阵和子矩阵的特征值判断Hermitian矩阵的normed特征向量的取值(约束条件见下文中的“方法适用约束条件”部分)的方法,即下文中的Lemma 2。本文给出部分命题的详细证明过程。
对于任意一个 n × n n \times n n×n 的Hermitian矩阵 A A A, 假设特征值为 λ i ( A ) \lambda _ {i} {(A)} λi(A), 对应的线性无关的 normed特征向量为 v i v_{i} vi, i ∈ { 1 , … , n } i \in \{1, \dots, n \} i∈{1,…,n}。假设 v i v_{i} vi的各元素为 v i , j v_{i, j} vi,j, j ∈ { 1 , … , n } j \in \{1, \dots, n \} j∈{1,…,n}。
假如 A A A有一个特征值为 0 0 0,不失一般性,假如 λ n ( A ) = 0 \lambda _ {n} {(A)} = 0 λn(A)=0。
∏ i = 1 n − 1 λ i ( A ) ∣ det ( B , v n ) ∣ 2 = det ( B ∗ A B ) (1) \prod \limits_{i = 1} ^{n - 1} \lambda _ {i} {(A)} \left \vert \det \left ( \begin{matrix} B, v_{n} \end{matrix} \right ) \right \vert ^{2}=\det (B^{*}AB) \tag {1} i=1∏n−1λi(A)∣∣det(B,vn)∣∣2=det(B∗AB)(1)
其中 B B B 是任意一个 n × ( n − 1 ) n \times (n - 1) n×(n−1) 的矩阵。 ( B , v n ) \begin{pmatrix} B, v_{n} \end{pmatrix} (B,vn) 是由 B B B和 v n v_{n} vn 按列拼接起来的 n × n n \times n n×n 矩阵。
由于 A A A 是 Hermitian矩阵,令矩阵 V = ( v 1 , ⋯ , v n ) V = \begin{pmatrix} v_{1}, \cdots , v_{n} \end{pmatrix} V=(v1,⋯,vn), D = ( λ 1 ( A ) ⋱ λ n − 1 ( A ) 0 ) = ( Λ 0 ) D = \begin{pmatrix} \lambda _ {1} {(A)} & & & \\ & \ddots & \\ & & \lambda _ {n - 1} {(A)} & \\ & & & 0 \end{pmatrix} = \begin{pmatrix} \Lambda & \\ & 0 \end{pmatrix} D=⎝⎜⎜⎛λ1(A)⋱λn−1(A)0⎠⎟⎟⎞=(Λ0),则 A = V D V ∗ A = VDV ^{*} A=VDV∗,且 V V V是酉矩阵。
令 V ∗ B = B 1 = ( B ′ X ) V^{*} B = B_{1} = \begin{pmatrix} B' \\ X \end{pmatrix} V∗B=B1=(B′X),其中 B ′ B' B′是一个 ( n − 1 ) × ( n − 1 ) (n - 1) \times (n - 1) (n−1)×(n−1) 的矩阵, X X X是一个 1 × ( n − 1 ) 1 \times (n - 1) 1×(n−1) 的矩阵。
易知 V e n = v n Ve_{n}=v _{n} Ven=vn,则 e n = V ∗ v n e_{n}=V^{*} v _{n} en=V∗vn
则 V ∗ ( B , v n ) = ( V ∗ B , V ∗ v n ) = ( B 1 , e n ) = ( B ′ 0 1 × ( n − 1 ) X 1 ) V^{*} \begin{pmatrix} B, v_{n} \end{pmatrix} = \begin{pmatrix} V^{*} B, V^{*} v_{n} \end{pmatrix} = \begin{pmatrix} B_{1}, e_{n} \end{pmatrix} = \begin{pmatrix} B' & 0_{1 \times (n - 1)} \\ X & 1 \end{pmatrix} V∗(B,vn)=(V∗B,V∗vn)=(B1,en)=(B′X01×(n−1)1)
因此 det ( B , v n ) = det ( V ( B ′ 0 1 × ( n − 1 ) X 1 ) ) = det ( V ) det ( B ′ ) = ± 1 det ( B ′ ) \det \left ( \begin{matrix} B, v_{n} \end{matrix} \right ) = \det \left ( V \begin{pmatrix} B' & 0_{1 \times (n - 1)} \\ X & 1 \end{pmatrix} \right ) = \det (V) \det (B') = \pm1 \det (B') det(B,vn)=det(V(B′X01×(n−1)1))=det(V)det(B′)=±1det(B′)
于是 ∏ i = 1 n − 1 λ i ( A ) ∣ det ( B , v n ) ∣ 2 = ∏ i = 1 n − 1 λ i ( A ) ∣ det ( B ′ ) ∣ 2 \prod \limits_{i = 1} ^{n - 1} \lambda _ {i} {(A)} \left \vert \det \left ( \begin{matrix} B, v_{n} \end{matrix} \right ) \right \vert ^{2}=\prod \limits_{i = 1} ^{n - 1} \lambda _ {i} {(A)} \left \vert \det \left ( \begin{matrix} B' \end{matrix} \right ) \right \vert ^{2} i=1∏n−1λi(A)∣∣det(B,vn)∣∣2=i=1∏n−1λi(A)∣∣det(B′)∣∣2
又 B ∗ A B = B ∗ V D V ∗ B = B 1 ∗ D B 1 = ( ( B ′ ) ∗ X ∗ ) ( Λ 0 ) ( B ′ X ) = ( B ′ ) ∗ Λ B ′ B^{*}AB =B^{*}VDV^{*}B = { B_{1} } ^{*} D B_{1} = \begin{pmatrix} {(B')} ^{*} & X ^{*} \end{pmatrix} \begin{pmatrix} \Lambda & \\ & 0 \end{pmatrix} \begin{pmatrix} B' \\ X \end{pmatrix} = {(B')} ^{*} \Lambda B' B∗AB=B∗VDV∗B=B1∗DB1=((B′)∗X∗)(Λ0)(B′X)=(B′)∗ΛB′
因此 det ( B ∗ A B ) = ∏ i = 1 n − 1 λ i ( A ) ∣ det ( B ′ ) ∣ 2 \det (B^{*}AB) =\prod \limits_{i = 1} ^{n - 1} \lambda _ {i} {(A)} \left \vert \det \left ( \begin{matrix} B' \end{matrix} \right ) \right \vert ^{2} det(B∗AB)=i=1∏n−1λi(A)∣∣det(B′)∣∣2
因此命题成立。
特征向量的每个元素的绝对值的平方与矩阵和子矩阵的特征值的关系:
∣ v i , j ∣ 2 ∏ k = 1 , k ≠ i n [ λ i ( A ) − λ k ( A ) ] = ∏ k = 1 n − 1 [ λ i ( A ) − λ k ( M j ) ] (2) \left \vert v_{i, j} \right \vert ^{2} \prod \limits_{k = 1, k \neq i} ^{n} \left [ \lambda _ {i} {(A)} - \lambda _{k} {(A)} \right ] = \prod \limits_{k = 1} ^{n - 1} \left [ \lambda _{i} {(A)} - \lambda _{k} {(M_{j})} \right ] \tag {2} ∣vi,j∣2k=1,k=i∏n[λi(A)−λk(A)]=k=1∏n−1[λi(A)−λk(Mj)](2)
不失一般性,只证明 i = n , j = 1 i = n, j = 1 i=n,j=1的情况。
令 A 1 = ( a 11 ⋯ ⋮ M 1 ′ ) = A − λ n I A_1 = \begin{pmatrix} a_{11} & \cdots \\ \vdots & M_{1}' \end{pmatrix} = A - \lambda _{n} I A1=(a11⋮⋯M1′)=A−λnI,则 λ n ( A 1 ) = 0 \lambda _{n} {(A_{1})} = 0 λn(A1)=0
等式 (2)可化为:
∣ v n , 1 ∣ 2 ∏ k = 1 n − 1 λ k ( A 1 ) = ∏ k = 1 n − 1 λ k ( M 1 ′ ) (3) \left \vert v_{n, 1} \right \vert ^{2} \prod \limits_{k = 1} ^{n - 1} \lambda _{k} {(A_{1})} = \prod \limits_{k = 1} ^{n - 1} \lambda _{k} {(M_{1}')} \tag {3} ∣vn,1∣2k=1∏n−1λk(A1)=k=1∏n−1λk(M1′)(3)
在 Lemma 1中令 B = ( 0 1 , n − 1 I n − 1 ) B = \begin{pmatrix} 0_{1, n-1} \\ I_{n - 1} \end{pmatrix} B=(01,n−1In−1),则
det ( B , v n ) = v n , 1 \det \left ( \begin{matrix} B, v_{n} \end{matrix} \right ) = v_{n, 1} det(B,vn)=vn,1
B ∗ A 1 B = ( 0 n − 1 , 1 I n − 1 ) ( a 11 ⋯ ⋮ M 1 ′ ) ( 0 1 , n − 1 I n − 1 ) = M 1 ′ B^{*}A _{1} B = \begin{pmatrix} 0_{n-1, 1} & I_{n - 1} \end{pmatrix} \begin{pmatrix} a_{11} & \cdots \\ \vdots & M_{1}' \end{pmatrix} \begin{pmatrix} 0_{1, n-1} \\ I_{n - 1} \end{pmatrix} = M_{1}' B∗A1B=(0n−1,1In−1)(a11⋮⋯M1′)(01,n−1In−1)=M1′
于是
∣ v n , 1 ∣ 2 ∏ k = 1 n − 1 λ k ( A 1 ) = ∏ i = 1 n − 1 λ i ( A 1 ) ∣ v n , 1 ∣ 2 = ∏ i = 1 n − 1 λ i ( A 1 ) ∣ det ( B , v n ) ∣ 2 = det ( B ∗ A 1 B ) = det ( M 1 ′ ) = ∏ k = 1 n − 1 λ k ( M 1 ′ ) \left \vert v_{n, 1} \right \vert ^{2} \prod \limits_{k = 1} ^{n - 1} \lambda _{k} {(A_{1})} = \prod \limits_{i = 1} ^{n - 1} \lambda _{i} {(A_{1})} \left \vert v_{n, 1} \right \vert ^{2} = \prod \limits_{i = 1} ^{n - 1} \lambda _ {i} {(A_{1})} \left \vert \det \left ( \begin{matrix} B, v_{n} \end{matrix} \right ) \right \vert ^{2} = \det (B^{*}A_{1} B) = \det(M_{1}') = \prod \limits_{k = 1} ^{n - 1} \lambda _{k} {(M_{1}')} ∣vn,1∣2k=1∏n−1λk(A1)=i=1∏n−1λi(A1)∣vn,1∣2=i=1∏n−1λi(A1)∣∣det(B,vn)∣∣2=det(B∗A1B)=det(M1′)=k=1∏n−1λk(M1′)
因此命题成立。
https://terrytao.wordpress.com/2019/08/13/eigenvectors-from-eigenvalues/
https://arxiv.org/abs/1908.03795