Gram 矩阵
假设 A A A 是一个 m × n m\times n m×n 阶矩阵,
- 列向量 Gram 矩阵
A A A 由列向量 α i \mathbf{\alpha}_i αi 表示, 即
A = [ α 1 α 2 ⋯ α n ] A=\begin{bmatrix}\mathbf{\alpha}_1 & \mathbf{\alpha}_2 &\cdots & \mathbf{\alpha}_n \end{bmatrix} A=[α1α2⋯αn]
则
G = A T A = [ α 1 T α 2 T ⋮ α n T ] [ α 1 α 2 ⋯ α n ] = [ α 1 T α 1 α 1 T α 2 ⋯ α 1 T α n α 2 T α 1 α 2 T α 2 ⋯ α 2 T α n ⋮ ⋮ ⋮ α n T α 1 α n T α 2 ⋯ α n T α n ] \begin{aligned} G &= \, A^{\mathsf T}A \\[3pt] &= \begin{bmatrix} \mathbf{\alpha}_1^{\mathsf T} \\ \mathbf{\alpha}_2^{\mathsf T} \\ \vdots \\ \mathbf{\alpha}_n^{\mathsf T} \end{bmatrix} \begin{bmatrix} \mathbf{\alpha}_1 & \mathbf{\alpha}_2 & \cdots & \mathbf{\alpha}_n \end{bmatrix} \\[3pt] & = \begin{bmatrix} \mathbf{\alpha}_1^{\mathsf T}\mathbf{\alpha}_1 & \mathbf{\alpha}_1^{\mathsf T}\mathbf{\alpha}_2 & \cdots & \mathbf{\alpha}_1^{\mathsf T}\mathbf{\alpha}_n \\ \mathbf{\alpha}_2^{\mathsf T}\mathbf{\alpha}_1 & \mathbf{\alpha}_2^{\mathsf T}\mathbf{\alpha}_2 & \cdots &\mathbf{\alpha}_2^{\mathsf T}\mathbf{\alpha}_n \\ \vdots & \vdots & & \vdots \\ \mathbf{\alpha}_n^{\mathsf T}\mathbf{\alpha}_1 & \mathbf{\alpha}_n^{\mathsf T}\mathbf{\alpha}_2 & \cdots & \mathbf{\alpha}_n^{\mathsf T}\mathbf{\alpha}_n \end{bmatrix} \end{aligned} G=ATA=⎣⎢⎢⎢⎡α1Tα2T⋮αnT⎦⎥⎥⎥⎤[α1α2⋯αn]=⎣⎢⎢⎢⎡α1Tα1α2Tα1⋮αnTα1α1Tα2α2Tα2⋮αnTα2⋯⋯⋯α1Tαnα2Tαn⋮αnTαn⎦⎥⎥⎥⎤
- 行向量 Gram 矩阵
A A A 由行向量 β i T \mathbf{\beta}_i^{\mathsf T} βiT 表示, 即
A = [ β 1 T β 2 T ⋮ β m T ] A=\begin{bmatrix}\mathbf{\beta}_1^{\mathsf T} \\ \mathbf{\beta}_2^{\mathsf T} \\ \vdots \\ \mathbf{\beta}_m^{\mathsf T} \end{bmatrix} A=⎣⎢⎢⎢⎡β1Tβ2T⋮βmT⎦⎥⎥⎥⎤
则
G = A A T = [ β 1 T β 2 T ⋮ β m T ] [ β 1 β 2 ⋯ β m ] = [ β 1 T β 1 β 1 T β 2 ⋯ β 1 T β m β 2 T β 1 β 2 T β 2 ⋯ β 2 T β m ⋮ ⋮ ⋮ β m T β 1 β m T β 2 ⋯ β m T β m ] \begin{aligned} G &= \, AA^{\mathsf T} \\[3pt] &= \begin{bmatrix}\mathbf{\beta}_1^{\mathsf T} \\ \mathbf{\beta}_2^{\mathsf T} \\ \vdots \\ \mathbf{\beta}_m^{\mathsf T} \end{bmatrix} \begin{bmatrix} \mathbf{\beta}_1 & \mathbf{\beta}_2 & \cdots & \mathbf{\beta}_m \end{bmatrix} \\[3pt] & = \begin{bmatrix} \mathbf{\beta}_1^{\mathsf T}\mathbf{\beta}_1 & \mathbf{\beta}_1^{\mathsf T}\mathbf{\beta}_2 & \cdots & \mathbf{\beta}_1^{\mathsf T}\mathbf{\beta}_m \\ \mathbf{\beta}_2^{\mathsf T}\mathbf{\beta}_1 & \mathbf{\beta}_2^{\mathsf T}\mathbf{\beta}_2 & \cdots &\mathbf{\beta}_2^{\mathsf T}\mathbf{\beta}_m \\ \vdots & \vdots & & \vdots \\ \mathbf{\beta}_m^{\mathsf T}\mathbf{\beta}_1 & \mathbf{\beta}_m^{\mathsf T}\mathbf{\beta}_2 & \cdots & \mathbf{\beta}_m^{\mathsf T}\mathbf{\beta}_m \end{bmatrix} \end{aligned} G=AAT=⎣⎢⎢⎢⎡β1Tβ2T⋮βmT⎦⎥⎥⎥⎤[β1β2⋯βm]=⎣⎢⎢⎢⎡β1Tβ1β2Tβ1⋮βmTβ1β1Tβ2β2Tβ2⋮βmTβ2⋯⋯⋯β1Tβmβ2Tβm⋮βmTβm⎦⎥⎥⎥⎤
6 大性质
下面只考虑列向量 Gram 矩阵
(1) G = A T A G = \, A^{\mathsf T}A G=ATA 是对称矩阵
G T = ( A T A ) T = A T A = G G^{\mathsf T } = \, (A^{\mathsf T}A)^{\mathsf T} = \, A^{\mathsf T}A = G GT=(ATA)T=ATA=G
(2) 对于实矩阵 A A A r a n k ( A T A ) = r a n k ( A ) \mathrm{rank} (A^{\mathsf T}A) = \mathrm{rank} (A) rank(ATA)=rank(A)
证明 { A x = 0 A T A x = 0 \begin{cases} A\mathsf{x} = 0 \\ A^{\mathsf T}A\mathbf{x} = 0 \end{cases} {Ax=0ATAx=0 同解即可.
证明过程详见经典例题(第3小问)
(3) 若 A T A = 0 A^{\mathsf T}A=0 ATA=0, 则 A = 0 A = 0 A=0
由上面性质
r
a
n
k
(
A
T
A
)
=
r
a
n
k
(
A
)
=
r
a
n
k
(
0
)
=
0
\begin{aligned} \mathrm{rank} (A^{\mathsf T}A) &= \mathrm{rank} (A) \\ &= \mathrm{rank} \ (0) = 0 \end{aligned}
rank(ATA)=rank(A)=rank (0)=0
(4) 对于实矩阵 A A A, 则 A T A A^{\mathsf T}A ATA 是半正定矩阵
x T A T A x = ( A x ) T A x ≥ 0 \mathbf{x}^{\mathsf T}A^{\mathsf T}A\mathbf{x} = (A\mathbf{x})^{\mathsf T}A\mathbf{x} \geq 0 xTATAx=(Ax)TAx≥0
(5) 对于任意 n n n 阶实对称半正定矩阵 M M M, 存在矩阵 A A A 使得 M = A T A M=A^{\mathsf T}A M=ATA 成立.
因为矩阵
M
M
M 实对称, 所以
M
M
M 可以正交对角化
, 即
M
=
Q
Λ
Q
T
M = Q\Lambda Q^{\mathsf T}
M=QΛQT 又因为矩阵
M
M
M 半正定, 所以其特征值 $\lambda_i \geq 0 $, 所以可记
Λ
1
/
2
=
d
i
a
g
(
λ
1
,
…
,
λ
n
)
\Lambda^{1/2} = \mathrm{diag} ( \sqrt{\lambda_1}, \dots, \sqrt{\lambda_n})
Λ1/2=diag(λ1,…,λn) 且 KaTeX parse error: Expected 'EOF', got '}' at position 29: …2}Q^\{\mathsf T}̲ 则可得
M
=
Q
Λ
Q
T
=
(
Λ
1
/
2
Q
T
)
T
Λ
1
/
2
Q
T
=
A
T
A
\begin{aligned} M &= Q\Lambda Q^{\mathsf T} \\ &= (\Lambda^{1/2}Q^{\mathsf T})^{\mathsf T}\Lambda^{1/2}Q^{\mathsf T} \\ &= A^{\mathsf T}A \end{aligned}
M=QΛQT=(Λ1/2QT)TΛ1/2QT=ATA
(6) 若 A = [ α 1 α 2 ⋯ α n ] A=\begin{bmatrix}\mathbf{\alpha}_1 & \mathbf{\alpha}_2 &\cdots & \mathbf{\alpha}_n \end{bmatrix} A=[α1α2⋯αn] 列满秩, 则 A T A A^{\mathsf T}A ATA 正定
- 由性质 (2), 知 r a n k ( A T A ) = r a n k ( A ) = n \mathrm{rank} (A^{\mathsf T}A) = \mathrm{rank} (A) = n rank(ATA)=rank(A)=n
- 因为
A
x
=
0
A\mathbf{x}=0
Ax=0 只有零解, 结合性质 (4), 对于非零
x
∈
R
n
\mathbf{x}\in \mathbb{R}^n
x∈Rn
x T A T A x = ( A x ) T A x > 0 \mathbf{x}^{\mathsf T}A^{\mathsf T}A\mathbf{x} = (A\mathbf{x})^{\mathsf T}A\mathbf{x} > 0 xTATAx=(Ax)TAx>0
原文链接
[1] matnoble.me/posts/gram
[2] 关注我吧