矩阵求逆引理(Matrix inversion lemma):
现有矩阵 A A 可以写为如下分块矩阵形式:
A=[A11A21A12A22](m+n)×(m+n) A = [ A 11 A 12 A 21 A 22 ] ( m + n ) × ( m + n )
矩阵 A A 为 (m+n) ( m + n ) 阶方阵,其中 A11 A 11 为 n n 阶非奇异方阵, A22 A 22 为 m m 阶非奇异方阵。那么可以得到: (A11−A12A−122A21) ( A 11 − A 12 A 22 − 1 A 21 ) 和 (A22−A21A−111A12) ( A 22 − A 21 A 11 − 1 A 12 ) 都是非奇异矩阵。
引理结论:
A−1=[A−111+A−111A12(A22−A21A−111A12)−1A21A−111−(A22−A21A−111A12)−1A21A−111−A−111A12(A22−A21A−111A12)−1(A22−A21A−111A12)−1] A − 1 = [ A 11 − 1 + A 11 − 1 A 12 ( A 22 − A 21 A 11 − 1 A 12 ) − 1 A 21 A 11 − 1 − A 11 − 1 A 12 ( A 22 − A 21 A 11 − 1 A 12 ) − 1 − ( A 22 − A 21 A 11 − 1 A 12 ) − 1 A 21 A 11 − 1 ( A 22 − A 21 A 11 − 1 A 12 ) − 1 ]
A−1=[(A11−A12A−122A21)−1−A−122A21(A11−A12A−122A21)−1−(A11−A12A−122A21)−1A12A−122A−122+A−122A21(A11−A12A−122A21)−1A12A−122] A − 1 = [ ( A 11 − A 12 A 22 − 1 A 21 ) − 1 − ( A 11 − A 12 A 22 − 1 A 21 ) − 1 A 12 A 22 − 1 − A 22 − 1 A 21 ( A 11 − A 12 A 22 − 1 A 21 ) − 1 A 22 − 1 + A 22 − 1 A 21 ( A 11 − A 12 A 22 − 1 A 21 ) − 1 A 12 A 22 − 1 ]
以上为矩阵 A−1 A − 1 的两种等价结果,其中对应位置的分块矩阵分别相等,由此可以分别得到4个推广等式,其中对角线上的两组等式又是相互等价的,最终得到2个等式:
(A11−A12A−122A21)−1=A−111+A−111A12(A22−A21A−111A12)−1A21A−111 ( A 11 − A 12 A 22 − 1 A 21 ) − 1 = A 11 − 1 + A 11 − 1 A 12 ( A 22 − A 21 A 11 − 1 A 12 ) − 1 A 21 A 11 − 1
A−111A12(A22−A21A−111A12)−1=(A11−A12A−122A21)−1A12A−122 A 11 − 1 A 12 ( A 22 − A 21 A 11 − 1 A 12 ) − 1 = ( A 11 − A 12 A 22 − 1 A 21 ) − 1 A 12 A 22 − 1
证明:
对矩阵 A A 以块为单位进行LU分解。得到以下两种分解结果:
A=[InA21A−1110Im][A110A12A22−A21A−111A12] A = [ I n 0 A 21 A 11 − 1 I m ] [ A 11 A 12 0 A 22 − A 21 A 11 − 1 A 12 ]
A=[In0A12A−122Im][A11−A12A−122A21A210A22] A = [ I n A 12 A 22 − 1 0 I m ] [ A 11 − A 12 A 22 − 1 A 21 0 A 21 A 22 ]
对以上分解得到的四个矩阵分别求逆(分块矩阵求逆方法):
[InA21A−1110Im]−1=[In−A21A−1110Im] [ I n 0 A 21 A 11 − 1 I m ] − 1 = [ I n 0 − A 21 A 11 − 1 I m ]
[A110A12A22−A21A−111A12]−1=[A110−A−111A12(A22−A21A−111A12)−1(A22−A21A−111A12)−1] [ A 11 A 12 0 A 22 − A 21 A 11 − 1 A 12 ] − 1 = [ A 11 − A 11 − 1 A 12 ( A 22 − A 21 A 11 − 1 A 12 ) − 1 0 ( A 22 − A 21 A 11 − 1 A 12 ) − 1 ]
[In0A12A−122Im]−1=[In0−A12A−122Im] [ I n A 12 A 22 − 1 0 I m ] − 1 = [ I n − A 12 A 22 − 1 0 I m ]
[A11−A12A−122A21A210A22]−1=[(A11−A12A−122A21)−1A−122A21(A11−A12A−122A21)−10A−122] [ A 11 − A 12 A 22 − 1 A 21 0 A 21 A 22 ] − 1 = [ ( A 11 − A 12 A 22 − 1 A 21 ) − 1 0 A 22 − 1 A 21 ( A 11 − A 12 A 22 − 1 A 21 ) − 1 A 22 − 1 ]
由LU分解的结果,矩阵 A−1 A − 1 可以写为:
A−1=[A110A12A22−A21A−111A12]−1[InA21A−1110Im]−1 A − 1 = [ A 11 A 12 0 A 22 − A 21 A 11 − 1 A 12 ] − 1 [ I n 0 A 21 A 11 − 1 I m ] − 1
A−1=[A11−A12A−122A21A210A22]−1[In0A12A−122Im]−1 A − 1 = [ A 11 − A 12 A 22 − 1 A 21 0 A 21 A 22 ] − 1 [ I n A 12 A 22 − 1 0 I m ] − 1
带入分块矩阵求逆的结果得到:
A−1=[A−111+A−111A12(A22−A21A−111A12)−1A21A−111−(A22−A21A−111A12)−1A21A−111−A−111A12(A22−A21A−111A12)−1(A22−A21A−111A12)−1] A − 1 = [ A 11 − 1 + A 11 − 1 A 12 ( A 22 − A 21 A 11 − 1 A 12 ) − 1 A 21 A 11 − 1 − A 11 − 1 A 12 ( A 22 − A 21 A 11 − 1 A 12 ) − 1 − ( A 22 − A 21 A 11 − 1 A 12 ) − 1 A 21 A 11 − 1 ( A 22 − A 21 A 11 − 1 A 12 ) − 1 ]
A−1=[(A11−A12A−122A21)−1−A−122A21(A11−A12A−122A21)−1−(A11−A12A−122A21)−1A12A−122A−122+A−122A21(A11−A12A−122A21)−1A12A−122] A − 1 = [ ( A 11 − A 12 A 22 − 1 A 21 ) − 1 − ( A 11 − A 12 A 22 − 1 A 21 ) − 1 A 12 A 22 − 1 − A 22 − 1 A 21 ( A 11 − A 12 A 22 − 1 A 21 ) − 1 A 22 − 1 + A 22 − 1 A 21 ( A 11 − A 12 A 22 − 1 A 21 ) − 1 A 12 A 22 − 1 ]
根据两个矩阵对应位置的分块矩阵相同,得到结论:
(A11−A12A−122A21)−1=A−111+A−111A12(A22−A21A−111A12)−1A21A−111 ( A 11 − A 12 A 22 − 1 A 21 ) − 1 = A 11 − 1 + A 11 − 1 A 12 ( A 22 − A 21 A 11 − 1 A 12 ) − 1 A 21 A 11 − 1
A−111A12(A22−A21A−111A12)−1=(A11−A12A−122A21)−1A12A−122 A 11 − 1 A 12 ( A 22 − A 21 A 11 − 1 A 12 ) − 1 = ( A 11 − A 12 A 22 − 1 A 21 ) − 1 A 12 A 22 − 1