大写字母表示矩阵(如 A A A),小写粗斜体表示向量(如 x \boldsymbol{x} x),正常字体表示标量或坐标点等(如 x 1 x_1 x1)。
向量
a
×
(
b
×
c
)
=
a
×
b
×
c
=
(
a
×
b
×
)
c
=
a
×
(
b
×
c
)
≠
(
a
×
b
)
×
c
=
a
×
b
×
c
=
(
a
×
b
)
×
c
\begin{aligned} & a\times(b\times c)=a^\times b^\times c=(a^\times b^\times)c =a^\times(b^\times c) \\ \neq& (a\times b)\times c= a\times b\times c=(a^\times b)^\times c \\ \end{aligned}
=a×(b×c)=a×b×c=(a×b×)c=a×(b×c)(a×b)×c=a×b×c=(a×b)×c
a
×
b
×
c
≠
a
×
b
×
c
(
R
a
)
×
b
≠
R
(
a
×
b
)
(
a
×
R
)
b
=
a
×
(
R
b
)
a\times b\times c\neq a^\times b^\times c \\ (Ra)\times b\neq R(a\times b) \\ (a\times R)b= a\times (Rb) \\
a×b×c=a×b×c(Ra)×b=R(a×b)(a×R)b=a×(Rb)
(
a
×
b
)
×
=
b
a
T
−
a
b
T
a
×
b
×
c
=
(
a
T
c
)
b
−
(
a
T
b
)
c
(a^\times b)^\times=ba^T-ab^T \\ a^\times b^\times c=(a^Tc)b-(a^Tb)c
(a×b)×=baT−abTa×b×c=(aTc)b−(aTb)c
方阵的迹
定义
tr
(
A
)
=
∑
i
=
1
n
(
a
i
i
)
\text{tr}(A)=\sum_{i=1}^n(a_{ii})
tr(A)=i=1∑n(aii)
性质
tr
(
A
)
=
tr
(
A
T
)
tr
(
A
+
B
)
=
tr
(
A
)
+
tr
(
B
)
tr
(
α
A
)
=
α
tr
(
A
)
tr
(
A
)
=
∑
i
=
1
n
λ
i
tr
(
A
B
)
=
tr
(
B
A
)
tr
(
A
T
A
)
=
∑
i
=
1
n
∑
j
=
1
n
a
i
j
2
\begin{aligned} & \text{tr}(A)=\text{tr}(A^{\text{T}}) \\ & \text{tr}(A+B)=\text{tr}(A)+\text{tr}(B) \\ & \text{tr}(\alpha A)=\alpha\text{tr}(A) \\ & \text{tr}(A)=\sum_{i=1}^n\lambda_i \\ & \text{tr}(AB)=\text{tr}(BA) \\ & \text{tr}(A^{\text{T}} A)=\sum_{i=1}^n\sum_{j=1}^n a_{ij}^2 \end{aligned}
tr(A)=tr(AT)tr(A+B)=tr(A)+tr(B)tr(αA)=αtr(A)tr(A)=i=1∑nλitr(AB)=tr(BA)tr(ATA)=i=1∑nj=1∑naij2
矩阵的秩
- A和B等价或相似或合同 ⇒ r ( A ) = r ( B ) \Rightarrow r(A)=r(B) ⇒r(A)=r(B)
- r ( A ) = r ( A T ) = r ( A T A ) = r ( A A T ) = r ( k A ) r(A)=r(A^{\text{T}})=r(A^{\text{T}}A)=r(AA^{\text{T}})=r(kA) r(A)=r(AT)=r(ATA)=r(AAT)=r(kA)
- 若 A A A列满秩则 r ( A B ) = r ( B ) r(AB)=r(B) r(AB)=r(B),若 A A A行满秩则 r ( B A ) = r ( B ) r(BA)=r(B) r(BA)=r(B)
- r ( A ± B ) ≤ r ( A ) + r ( B ) r(A\pm B)\le r(A)+r(B) r(A±B)≤r(A)+r(B)
- r ( A ) + r ( B ) − n ≤ r ( A B ) ≤ min { r ( A ) , r ( B ) } ≤ max { r ( A ) , r ( B ) } ≤ r ( A ⋮ B ) ≤ r ( A ) + r ( B ) r(A)+r(B)-n\le r(AB)\le\min\{r(A), r(B)\}\le\max\{r(A), r(B)\}\le r(A\vdots B)\le r(A)+r(B) r(A)+r(B)−n≤r(AB)≤min{r(A),r(B)}≤max{r(A),r(B)}≤r(A⋮B)≤r(A)+r(B)
- A B = O ⇒ r ( A ) + r ( B ) ≤ n AB=O\Rightarrow r(A)+r(B)\le n AB=O⇒r(A)+r(B)≤n
- 幂零阵: A k = O ⇒ r ( A ) < n A^k=O\Rightarrow r(A)<n Ak=O⇒r(A)<n
- 幂等阵: A 2 = A ⇒ r ( A ) + r ( I − A ) = n A^2=A\Rightarrow r(A)+r(I-A)=n A2=A⇒r(A)+r(I−A)=n
- 幂幺阵: A k = I ⇒ r ( A ) = n A^k=I\Rightarrow r(A)=n Ak=I⇒r(A)=n
- r ( A O O B ) = r ( A ) + r ( B ) ≤ r ( A 0 C B ) ≤ r ( A ) + r ( B ) + r ( C ) r\left(\begin{array}{ll} A & O \\ O & B \end{array}\right)=r(A)+r(B) \leq r\left(\begin{array}{ll} A & 0 \\ C & B \end{array}\right) \leq r(A)+r(B)+r(C) r(AOOB)=r(A)+r(B)≤r(AC0B)≤r(A)+r(B)+r(C)
- r ( A ∗ ) = { n ⇔ r ( A ) = n 1 ⇔ r ( A ) = n − 1 0 ⇔ r ( A ) < n − 1 r(A*)=\begin{cases} n \Leftrightarrow r(A)=n \\ 1 \Leftrightarrow r(A)=n -1 \\ 0 \Leftrightarrow r(A)<n -1 \\ \end{cases} r(A∗)=⎩ ⎨ ⎧n⇔r(A)=n1⇔r(A)=n−10⇔r(A)<n−1
- 若A可逆,则 r ( A B ) = r ( B A ) = r ( B ) r(AB)=r(BA)=r(B) r(AB)=r(BA)=r(B)
- 若 A A A是n阶方阵则 r ( A ) ≥ μ ( A ) r(A)\ge\mu(A) r(A)≥μ(A), μ ( A ) \mu(A) μ(A)表示矩阵 A A A的非零特征值的个数;若 A A A可相似对角化则 r ( A ) = μ ( A ) r(A)=\mu(A) r(A)=μ(A)
- 若 A A A是正定或正交矩阵则 r ( A ) = n r(A)=n r(A)=n
- m×n 矩阵 A 的秩为 r 的充分与必要条件为:存在 m 阶满秩矩阵 P 与 n 阶满秩矩阵 Q 使得
P A Q = Λ = ( I r O O O ) PAQ=\Lambda=\left(\begin{matrix} I_r & O \\ O & O \end{matrix}\right) PAQ=Λ=(IrOOO)
若 r ( A ) = 1 r(A)=1 r(A)=1,则
- ∣ λ I − A ∣ = λ n − tr ( A ) λ n − 1 |\lambda I-A|=\lambda^n-\text{tr}(A)\lambda^{n-1} ∣λI−A∣=λn−tr(A)λn−1
- λ 1 = tr ( A ) \lambda_1=\text{tr}(A) λ1=tr(A), λ 2 = ⋯ = λ n = 0 \lambda_2=\cdots=\lambda_n=0 λ2=⋯=λn=0
- A 2 = tr ( A ) A A^2=\text{tr}(A)A A2=tr(A)A, A n = ( tr ( A ) ) n − 1 A A^n=(\text{tr}(A))^{n-1}A An=(tr(A))n−1A
- (充要条件)存在非零向量 α \alpha α、 β \beta β使得 A = α β A=\alpha\beta A=αβ
特征值与特征向量
- 特征值都是单值的矩阵一定可以对角化。
- n阶矩阵有n个线性无关的特征向量是可以相似对角化的充分必要条件。
- 实对称矩阵的不同特征值对应的特征向量两两正交。
- 实对称矩阵可对角化,即必有n个线性无关的特征向量。
- 不同特征值对应的特征向量线性无关。
- 特征值相同的两个矩阵不一定相似。
- λ 1 + λ 2 + ⋯ + λ n = tr ( A ) \lambda_1+\lambda_2+\cdots+\lambda_n=\text{tr}(A) λ1+λ2+⋯+λn=tr(A)
- λ 1 λ 2 ⋯ λ n = ∣ A ∣ \lambda_1\lambda_2\cdots\lambda_n=|A| λ1λ2⋯λn=∣A∣
设 A A A是 n 阶矩阵, λ 0 \lambda_0 λ0是 A A A的 k k k阶特征值,则
- 若 k = 1 k=1 k=1,则属于特征值 λ 0 \lambda_0 λ0的线性无关的特征向量只有一个;
- 若 k > 1 k>1 k>1,则属于特征值 λ 0 \lambda_0 λ0的线性无关的特征向量不超过 k k k个。
分块矩阵
设矩阵
A
A
A是一个
n
+
m
n+m
n+m阶方阵,它具有分块三角阵的结构,即
A
=
[
A
11
A
12
0
A
22
]
或
A
=
[
A
11
0
A
21
A
22
]
A=\left[\begin{array}{cc} A_{11} & A_{12} \\ 0 & A_{22} \end{array}\right]\text{或} A=\left[\begin{array}{cc} A_{11} & 0 \\ A_{21} & A_{22} \end{array}\right]
A=[A110A12A22]或A=[A11A210A22]
其中
A
11
A_{11}
A11和
A
22
A_{22}
A22分别是n阶和m阶可逆方阵,这意味着 A 是可逆阵。利用
A
−
1
A
=
A
A
−
1
=
I
n
+
m
A^{-1}A=AA^{-1}=I_{n+m}
A−1A=AA−1=In+m
可以推得
A
−
1
=
[
A
11
−
1
−
A
11
−
1
A
12
A
22
−
1
0
A
22
−
1
]
A
−
1
=
[
A
11
−
1
0
−
A
22
−
1
A
21
A
11
−
1
A
22
−
1
]
A^{-1}=\left[\begin{array}{cc} A_{11}^{-1} & -A_{11}^{-1}A_{12}A_{22}^{-1} \\ 0 & A_{22}^{-1} \end{array}\right] \\ A^{-1}=\left[\begin{array}{cc} A_{11}^{-1} & 0 \\ -A_{22}^{-1}A_{21}A_{11}^{-1} & A_{22}^{-1} \end{array}\right]
A−1=[A11−10−A11−1A12A22−1A22−1]A−1=[A11−1−A22−1A21A11−10A22−1]
一般地,若n+m阶方阵A可以写成分块形式
A
=
[
A
11
A
12
A
21
A
11
−
1
A
22
]
A=\left[\begin{array}{cc} A_{11} & A_{12} \\ A_{21}A_{11}^{-1} & A_{22} \end{array}\right]
A=[A11A21A11−1A12A22]
其中
A
11
A_{11}
A11和
A
22
A_{22}
A22具有与前相同的性质,那么利用矩阵分解关系式
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=\left[\begin{array}{cc} I_n & 0 \\ A_{21}A_{11}^{-1} & I_m \end{array}\right]\left[\begin{array}{cc} A_{11} & A_{12} \\ 0 & A_{22}-A_{21}A_{11}^{-1}A_{12} \end{array}\right]
A=[InA21A11−10Im][A110A12A22−A21A11−1A12]
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
]
A=\left[\begin{array}{cc} I_n & A_{12}A_{22}^{-1} \\ 0 & I_m \end{array}\right]\left[\begin{array}{cc} A_{11}-A_{12}A_{22}^{-1}A_{21} & 0 \\ A_{21} & A_{22} \end{array}\right]
A=[In0A12A22−1Im][A11−A12A22−1A21A210A22]
和前面关于三角阵的求逆结果,可以推得矩阵 A 的分块求逆公式如下:
A
−
1
=
[
A
11
−
1
+
A
11
−
1
A
12
A
~
22
−
1
A
21
A
11
−
1
−
A
11
−
1
A
12
A
~
22
−
1
−
A
~
22
−
1
A
21
A
11
−
1
A
~
22
−
1
]
A^{-1}=\left[\begin{array}{cc} A_{11}^{-1}+A_{11}^{-1}A_{12}\tilde{A}_{22}^{-1}A_{21}A_{11}^{-1} & -A_{11}^{-1}A_{12}\tilde{A}_{22}^{-1} \\ -\tilde{A}_{22}^{-1}A_{21}A_{11}^{-1} & \tilde{A}_{22}^{-1} \end{array}\right]
A−1=[A11−1+A11−1A12A~22−1A21A11−1−A~22−1A21A11−1−A11−1A12A~22−1A~22−1]
A
−
1
=
[
A
~
11
−
1
−
A
~
11
−
1
A
12
A
22
−
1
−
A
22
−
1
A
21
A
~
11
−
1
A
22
−
1
+
A
22
−
1
A
21
A
~
11
−
1
A
12
A
22
−
1
]
A^{-1}=\left[\begin{array}{cc} \tilde{A}_{11}^{-1} & -\tilde{A}_{11}^{-1}A_{12}A_{22}^{-1} \\ -A_{22}^{-1}A_{21}\tilde{A}_{11}^{-1} & A_{22}^{-1}+A_{22}^{-1}A_{21}\tilde{A}_{11}^{-1}A_{12}A_{22}^{-1} \end{array}\right]
A−1=[A~11−1−A22−1A21A~11−1−A~11−1A12A22−1A22−1+A22−1A21A~11−1A12A22−1]
其中
A
~
11
=
A
11
−
A
12
A
22
−
1
A
21
A
~
22
=
A
22
−
A
21
A
11
−
1
A
12
\tilde{A}_{11}=A_{11}-A_{12}A_{22}^{-1}A_{21} \\ \tilde{A}_{22}=A_{22}-A_{21}A_{11}^{-1}A_{12}
A~11=A11−A12A22−1A21A~22=A22−A21A11−1A12
假定矩阵A是可逆矩阵,因而
A
~
11
−
1
\tilde{A}_{11}^{-1}
A~11−1和
A
~
22
−
1
\tilde{A}_{22}^{-1}
A~22−1总是存在的。根据逆矩阵的唯一性,对比两式立即得到(这附近的推导可能有问题)
(
A
11
−
A
12
A
22
−
1
A
21
)
−
1
=
A
11
−
1
+
A
11
−
1
A
12
A
~
22
−
1
A
21
A
11
−
1
(A_{11}-A_{12}A_{22}^{-1}A_{21})^{-1}=A_{11}^{-1}+A_{11}^{-1}A_{12}\tilde{A}_{22}^{-1}A_{21}A_{11}^{-1}
(A11−A12A22−1A21)−1=A11−1+A11−1A12A~22−1A21A11−1
A
22
−
A
21
A
11
−
1
A
12
=
A
22
−
1
+
A
22
−
1
A
21
A
~
11
−
1
A
12
A
22
−
1
A_{22}-A_{21}A_{11}^{-1}A_{12}=A_{22}^{-1}+A_{22}^{-1}A_{21}\tilde{A}_{11}^{-1}A_{12}A_{22}^{-1}
A22−A21A11−1A12=A22−1+A22−1A21A~11−1A12A22−1
这是两个非常重要的矩阵恒等式,在矩阵变换中经常用到。其中第一式习惯上称为矩阵反馈公式。
- 若A和B是方阵,则
∣ A B ∣ = ∣ A ∣ ⋅ ∣ B ∣ |AB|=|A|\cdot|B| ∣AB∣=∣A∣⋅∣B∣
上式可用初等变换方法证明。 - 若A和D是方阵,则
∣ A B 0 D ∣ = ∣ A ∣ ⋅ ∣ D ∣ \left|\begin{matrix}A & B \\ 0 & D\end{matrix}\right|=|A|\cdot|D| A0BD =∣A∣⋅∣D∣
QR分解法证明:
∣ A 0 0 D ∣ = ∣ A 0 0 I ∣ ∣ I 0 0 D ∣ = ∣ A ∣ ⋅ ∣ D ∣ ∣ A B 0 D ∣ = ∣ Q A 0 0 Q D ∣ ⋅ ∣ R A Q A T B 0 R D ∣ = ∣ Q A ∣ ⋅ ∣ Q D ∣ ⋅ ∣ R A ∣ ⋅ ∣ R D ∣ \begin{aligned} &\left|\begin{matrix}A & 0 \\ 0 & D\end{matrix}\right| =\left|\begin{matrix}A & 0 \\ 0 & I\end{matrix}\right| \left|\begin{matrix}I & 0 \\ 0 & D\end{matrix}\right| =|A|\cdot|D| \\ &\left|\begin{matrix}A & B \\ 0 & D\end{matrix}\right| =\left|\begin{matrix}Q_A & 0 \\ 0 & Q_D\end{matrix}\right| \cdot\left|\begin{matrix}R_A & Q_A^{\text{T}}B \\ 0 & R_D\end{matrix}\right| =|Q_A|\cdot|Q_D|\cdot|R_A|\cdot|R_D| \end{aligned} A00D = A00I I00D =∣A∣⋅∣D∣ A0BD = QA00QD ⋅ RA0QATBRD =∣QA∣⋅∣QD∣⋅∣RA∣⋅∣RD∣ - 若A和D是方阵且A可逆,则
∣ A B C D ∣ = ∣ A ∣ ⋅ ∣ D − C A − 1 B ∣ \left|\begin{matrix}A & B \\ C & D\end{matrix}\right| =|A|\cdot|D-CA^{-1}B| ACBD =∣A∣⋅∣D−CA−1B∣
若D可逆则
∣ A B C D ∣ = ∣ D ∣ ⋅ ∣ A − B D − 1 C ∣ \left|\begin{matrix}A & B \\ C & D\end{matrix}\right| =|D|\cdot|A-BD^{-1}C| ACBD =∣D∣⋅∣A−BD−1C∣
证明
∣ A B C D ∣ = ∣ I 0 C A − 1 I ∣ ∣ A B 0 D − C A − 1 B ∣ = ∣ I B D − 1 0 I ∣ ∣ A − B D − 1 C 0 C D ∣ \begin{aligned} \left|\begin{matrix}A & B \\ C & D\end{matrix}\right| =& \left|\begin{matrix}I & 0 \\ CA^{-1} & I\end{matrix}\right| \left|\begin{matrix}A & B \\ 0 & D-CA^{-1}B\end{matrix}\right|\\ =& \left|\begin{matrix}I & BD^{-1} \\ 0 & I\end{matrix}\right| \left|\begin{matrix}A-BD^{-1}C & 0 \\ C & D\end{matrix}\right| \end{aligned} ACBD == ICA−10I A0BD−CA−1B I0BD−1I A−BD−1CC0D
矩阵的正定与负定
设矩阵
A
A
A为n阶对称阵,如果对于所有n维列向量
x
\boldsymbol{x}
x,二次型
x
T
A
x
\boldsymbol{x}^{\text{T}} A\boldsymbol{x}
xTAx均为非负,则称矩阵
A
A
A为非负定矩阵,并用
A
≥
0
A\ge 0
A≥0来表示。进一步,如果矩阵
A
A
A为非负定矩阵,且对所有非零向量
x
\boldsymbol{x}
x,二次型
x
T
A
x
\boldsymbol{x}^{\text{T}}A\boldsymbol{x}
xTAx总大于零,则称矩阵
A
A
A为正定阵,并且用
A
>
0
A>0
A>0来表示。对称矩阵
A
A
A当且仅当其所有特征值非负时才是非负定阵;当且仅当特征值均为正时才是正定阵。显然,若矩阵
A
A
A为正定阵,则其逆矩阵存在且也为正定阵。
如果
D
D
D是任意
n
×
m
n\times m
n×m阶矩阵,则
A
=
D
D
′
A=DD'
A=DD′是非负定阵;当且仅当
D
D
D行满秩时,
A
=
D
D
′
A=DD'
A=DD′才是正定阵。
如果
A
A
A和
B
B
B是同阶非负定阵,
α
\alpha
α和
β
\beta
β为非负常数,则
α
A
+
β
B
\alpha A+\beta B
αA+βB为非负定阵;若
A
A
A、
B
B
B两者之一是正定阵而另一个为非负定阵且
α
\alpha
α和
β
\beta
β均大于零,则
α
A
+
β
B
\alpha A+\beta B
αA+βB是正定阵。设
A
A
A和
B
B
B分别为非负定阵和正定阵,称-A和-
B
B
B分别是非正定阵和负定阵。非正定阵和负定阵分别与非负定阵和正定阵具有相反而类似的性质。
几个名词:
- 标准形
- 规范形
- 矩阵等价
- 矩阵相似
- 矩阵合同
二次型的标准形的系数中正负系数的个数保持不变,分别称为二次型的正负惯性指数。
n
n
n阶实对称矩阵
A
A
A和
B
B
B的正负惯性指数相同是
A
A
A和
B
B
B合同的充分必要条件。
记二次型
f
(
x
)
=
x
T
A
x
f(\boldsymbol{x})=\boldsymbol{x}^{\text{T}} A\boldsymbol{x}
f(x)=xTAx,
A
A
A为
n
n
n阶实对称矩阵,下列命题相互等价:
- f f f是正定(负定)二次型
- f f f的正(负)惯性指数为 n n n
- A A A的特征值全为正(负)
- A A A合同于 + I +I +I( − I -I −I)
- A A A的顺序主子式都大于0( A A A的奇数阶顺序主子式为负,偶数阶顺序主子式为正)
- 存在可逆矩阵 B B B使得 A = B T B A=B^TB A=BTB
- a i i > 0 a_{ii}>0 aii>0, i = 1 , 2 , ⋯ , n i=1,2,\cdots,n i=1,2,⋯,n
内积与范数
向量范数:
- 非负性:对所有 x ⃗ \vec{x} x均有 ∣ ∣ x ⃗ ∣ ∣ ≥ 0 ||\vec{x}||\ge 0 ∣∣x∣∣≥0;当且仅当 x ⃗ = 0 \vec{x}=0 x=0 时,才有 ∣ ∣ x ⃗ ∣ ∣ = 0 ||\vec{x}||=0 ∣∣x∣∣=0;
- 齐次性: ∣ ∣ k x ⃗ ∣ ∣ = ∣ k ∣ ∣ ∣ x ⃗ ∣ ∣ ||k\vec{x}||=|k| \ ||\vec{x}|| ∣∣kx∣∣=∣k∣ ∣∣x∣∣;
- 三角不等式 ∣ ∣ A + B ∣ ∣ ≤ ∣ ∣ A ∣ ∣ + ∣ ∣ B ∣ ∣ ||A+B||\le||A|| + ||B|| ∣∣A+B∣∣≤∣∣A∣∣+∣∣B∣∣。
矩阵范数除此以外还有次乘性:
- 次乘性: ∣ ∣ A B ∣ ∣ ≤ ∣ ∣ A ∣ ∣ ∣ ∣ B ∣ ∣ ||AB||\le||A||\ ||B|| ∣∣AB∣∣≤∣∣A∣∣ ∣∣B∣∣。
如果向量范数 ∣ ∣ x ∣ ∣ ||x|| ∣∣x∣∣和矩阵范数 ∣ ∣ A ∣ ∣ ||A|| ∣∣A∣∣满足 ∣ ∣ A x ∣ ∣ ≤ ∣ ∣ A ∣ ∣ ∣ ∣ x ∣ ∣ ||Ax||\le||A||\ ||x|| ∣∣Ax∣∣≤∣∣A∣∣ ∣∣x∣∣,则称相容。
矩阵的微分运算
矩阵微分运算有几种不同的情况。
矩阵函数对标量的导数
设n×m阶矩阵
A
A
A、
B
B
B和m×1阶矩阵
C
C
C的元素都是实变数t的函数,
λ
=
λ
(
t
)
\lambda=\lambda(t)
λ=λ(t)是
t
t
t的标量实值函数。定义矩阵
A
A
A对
t
t
t的导数等于
A
A
A的每个元素
a
i
j
(
t
)
a_{ij}(t)
aij(t)对
t
t
t分别求导所构成的n×m阶矩阵,即
d
A
d
t
=
[
d
a
i
j
(
t
)
d
t
]
\frac{\text{d}A}{\text{d}t}=\left[\frac{\text{d}a_{ij}(t)}{\text{d}t}\right]
dtdA=[dtdaij(t)]
比如,对于n维列向量
x
=
[
x
1
(
t
)
x
2
(
t
)
⋯
x
n
(
t
)
]
T
\boldsymbol{x}=[x_1(t)\ x_2(t)\cdots x_n(t)]^{\text{T}}
x=[x1(t) x2(t)⋯xn(t)]T,按定义就有
d
x
d
t
=
[
d
x
1
(
t
)
d
t
d
x
2
(
t
)
d
t
⋯
d
x
n
(
t
)
d
t
]
T
\frac{\text{d}\boldsymbol{x}}{\text{d}t}=\left[\frac{\text{d}x_1(t)}{\text{d}t}\frac{\text{d}x_2(t)}{\text{d}t}\cdots\frac{\text{d}x_n(t)}{\text{d}t}\right]^{\text{T}}
dtdx=[dtdx1(t)dtdx2(t)⋯dtdxn(t)]T
关于矩阵函数对标量的导数,根据上述定义容易验证如下运算规则;
d
(
A
+
B
)
d
t
=
d
A
d
t
+
d
B
d
t
d
(
λ
A
)
d
t
=
d
λ
d
t
A
+
λ
d
A
d
t
d
(
A
C
)
d
t
=
d
A
d
t
C
+
A
d
C
d
t
\begin{aligned} &\frac{\text{d}(A+B)}{\text{d}t}=\frac{\text{d}A}{\text{d}t}+\frac{\text{d}B}{\text{d}t} \\ &\frac{\text{d}(\lambda A)}{\text{d}t}=\frac{\text{d}\lambda}{\text{d}t}A+\lambda\frac{\text{d}A}{\text{d}t} \\ &\frac{\text{d}(AC)}{\text{d}t}=\frac{\text{d}A}{\text{d}t}C+A\frac{\text{d}C}{\text{d}t} \end{aligned}
dtd(A+B)=dtdA+dtdBdtd(λA)=dtdλA+λdtdAdtd(AC)=dtdAC+AdtdC
标量函数对矩阵的导数
设
f
=
f
(
A
)
f=f(A)
f=f(A)、
g
=
g
(
A
)
g=g(A)
g=g(A)是以矩阵
A
A
A的n×m个元素为自变量的标量定值函数。定义
f
f
f对
A
A
A的导数为如下
n
×
m
n\times m
n×m阶矩阵
d
f
d
A
≜
[
∂
f
∂
a
i
j
]
\frac{\text{d}f}{\text{d}A}\triangleq\left[\frac{\partial f}{\partial a_{ij}}\right]
dAdf≜[∂aij∂f]
对于上述这类微分运算,显然有
d
(
f
+
g
)
d
A
=
d
f
d
A
+
d
g
d
A
d
(
f
g
)
d
A
=
d
f
d
A
g
+
f
d
g
d
A
\begin{aligned} &\frac{\text{d}(f+g)}{\text{d}A}=\frac{\text{d}f}{\text{d}A}+\frac{\text{d}g}{\text{d}A} \\ &\frac{\text{d}(fg)}{\text{d}A}=\frac{\text{d}f}{\text{d}A}g+f\frac{\text{d}g}{\text{d}A} \end{aligned}
dAd(f+g)=dAdf+dAdgdAd(fg)=dAdfg+fdAdg
矩阵函数对向量的导数
设
F
(
x
)
F(\boldsymbol{x})
F(x)是n维列向量
x
\boldsymbol{x}
x的
m
×
l
m\times l
m×l阶矩阵函数,即
F
(
x
)
=
(
f
i
j
(
x
)
)
m
×
l
F(\boldsymbol{x})=(f_{ij}(\boldsymbol{x}))_{m×l}
F(x)=(fij(x))m×l,而
x
=
[
x
1
x
2
⋯
x
n
]
T
\boldsymbol{x}=[x_1\ x_2\cdots x_n]^{\text{T}}
x=[x1 x2⋯xn]T。定义
F
(
x
)
F(\boldsymbol{x})
F(x)对
x
\boldsymbol{x}
x的导数为如下nm×l阶矩阵:
d
F
d
x
≜
[
∂
F
(
x
)
∂
x
1
⋮
∂
F
(
x
)
∂
x
2
⋯
⋮
∂
F
(
x
)
∂
x
n
]
T
\frac{\text{d}F}{\text{d}\boldsymbol{x}}\triangleq[\frac{\partial F(\boldsymbol{x})}{\partial x_1}\vdots\frac{\partial F(\boldsymbol{x})}{\partial x_2}\cdots\vdots\frac{\partial F(\boldsymbol{x})}{\partial x_n}]^{\text{T}}
dxdF≜[∂x1∂F(x)⋮∂x2∂F(x)⋯⋮∂xn∂F(x)]T
其中
∂
F
(
x
)
∂
x
k
≜
∂
f
i
j
(
x
)
∂
x
k
\frac{\partial F(\boldsymbol{x})}{\partial x_k}\triangleq\frac{\partial f_ij(\boldsymbol{x})}{\partial x_k}
∂xk∂F(x)≜∂xk∂fij(x)
对于这类运算,我们有
d
F
(
x
)
+
G
(
x
)
d
x
=
d
F
(
x
)
d
x
+
d
G
(
x
)
d
x
d
(
F
T
(
x
)
G
(
x
)
)
d
x
=
d
F
T
(
x
)
d
x
G
(
x
)
+
d
G
T
(
x
)
d
x
F
(
x
)
\begin{aligned} &\frac{\text{d}F(\boldsymbol{x})+G(\boldsymbol{x})}{\text{d}\boldsymbol{x}} =\frac{\text{d}F(\boldsymbol{x})}{\text{d}\boldsymbol{x}}+\frac{\text{d}G(\boldsymbol{x})}{\text{d}\boldsymbol{x}} \\ &\frac{\text{d}(F^{\text{T}}(\boldsymbol{x})G(\boldsymbol{x}))}{\text{d}\boldsymbol{x}} =\frac{\text{d}F^{\text{T}}(\boldsymbol{x})}{\text{d}\boldsymbol{x}}G(\boldsymbol{x})+\frac{\text{d}G^{\text{T}}(\boldsymbol{x})}{\text{d}\boldsymbol{x}}F(\boldsymbol{x}) \end{aligned}
dxdF(x)+G(x)=dxdF(x)+dxdG(x)dxd(FT(x)G(x))=dxdFT(x)G(x)+dxdGT(x)F(x)
几个常用的矩阵微分公式
根据前面的定义,不难验证以下矩阵微分公式:
- 设
f
=
f
(
x
)
f=f(\boldsymbol{x})
f=f(x)是n维列向量
x
\boldsymbol{x}
x的标量定值函数,则有
d f d t = [ d f d x ] T d x d t \frac{\text{d}f}{\text{d}t}=\left[\frac{\text{d}f}{\text{d}\boldsymbol{x}}\right]^{\text{T}} \frac{\text{d}\boldsymbol{x}}{\text{d}t} dtdf=[dxdf]Tdtdx
式中t为实变数。 - 设
x
\boldsymbol{x}
x为n维列向量,
a
\boldsymbol{a}
a和
B
B
B分别为与
x
\boldsymbol{x}
x无关的m维列向量和m×n阶矩阵,f为
x
\boldsymbol{x}
x的一个二次型,且
f = ( a + B x ) T ( a + B x ) f=(\boldsymbol{a}+B\boldsymbol{x})^{\text{T}}(\boldsymbol{a}+B\boldsymbol{x}) f=(a+Bx)T(a+Bx)
则有
d ( a + B x ) d x = B T d f d x = 2 B T ( a − B x ) \begin{aligned} &\frac{\text{d}(\boldsymbol{a}+B\boldsymbol{x})}{\text{d}\boldsymbol{x}}=B^{\text{T}} \\ &\frac{\text{d}f}{\text{d}\boldsymbol{x}}=2B^{\text{T}}(\boldsymbol{a}-B\boldsymbol{x}) \end{aligned} dxd(a+Bx)=BTdxdf=2BT(a−Bx) - 设
A
A
A为n阶方阵,其元素是实变数t的函数,且对所有的t,
A
−
1
A^{-1}
A−1存在,则有
d A − 1 d t = − A − 1 d A d t A − 1 \frac{\text{d}A^{-1}}{\text{d}t}=-A^{-1}\frac{\text{d}A}{\text{d}t}A^{-1} dtdA−1=−A−1dtdAA−1
此式可通过恒等式
d I n d t = d A A − 1 d t = 0 \frac{\text{d}I_n}{\text{d}t}=\frac{\text{d}AA^{-1}}{\text{d}t}=0 dtdIn=dtdAA−1=0
导出。 - 设,则有的解为矩阵 A A A的伪迹。式中星号*表示转置兼取复数共轭。
- 其它常用公式
d x T d x = d x d x T = I n d ( x T A x ) d x = ( A + A T ) x d ( x T A x ) d A = x x T d J d X = d ( α T X β ) d X = α β T e x p \begin{aligned} &\frac{\text{d}\boldsymbol{x}^{\text{T}}}{\text{d}\boldsymbol{x}} =\frac{\text{d}\boldsymbol{x}}{\text{d}\boldsymbol{x}^{\text{T}}}=I_n \\ & \frac{\text{d}(\boldsymbol{x}^{\text{T}}\mathbf{A}\boldsymbol{x})}{\text{d}\boldsymbol{x}} =(\mathbf{A}+\mathbf{A}^{\text{T}})\boldsymbol{x} \\ &\frac{\text{d}(\boldsymbol{x}^{\text{T}}\mathbf{A}\boldsymbol{x})}{\text{d}\mathbf{A}} =\boldsymbol{x}\boldsymbol{x}^{\text{T}} \\ &\frac{\text{d}J}{\text{d}\mathbf{X}} =\frac{\text{d}(\boldsymbol{\alpha}^{\text{T}}\mathbf{X}\boldsymbol{\beta})}{\text{d}\mathbf{X}} =\boldsymbol{\alpha}\boldsymbol{\beta}^{\text{T}}exp \\ \end{aligned} dxdxT=dxTdx=Indxd(xTAx)=(A+AT)xdAd(xTAx)=xxTdXdJ=dXd(αTXβ)=αβTexp
矩阵函数及其导数的一些性质
定义
e
A
=
I
+
A
t
+
A
2
2
!
+
A
3
3
!
+
⋯
sin
A
=
A
−
A
3
3
!
+
A
5
5
!
−
⋯
cos
A
=
1
−
A
2
2
!
+
A
4
4
!
+
⋯
ln
(
1
+
A
)
=
A
−
A
2
2
+
A
3
3
−
⋯
(
1
)
\begin{aligned} & \text{e}^{A}=I+At+\frac{A^2}{2!}+\frac{A^3}{3!}+\cdots \\ & \sin A=A-\frac{A^3}{3!}+\frac{A^5}{5!}-\cdots \\ & \cos A=1-\frac{A^2}{2!}+\frac{A^4}{4!}+\cdots\\ & \ln(1+A)=A-\frac{A^2}{2}+\frac{A^3}{3}-\cdots\qquad(1)\\ \end{aligned}
eA=I+At+2!A2+3!A3+⋯sinA=A−3!A3+5!A5−⋯cosA=1−2!A2+4!A4+⋯ln(1+A)=A−2A2+3A3−⋯(1)
性质
A
e
A
=
e
A
A
e
A
⋅
e
B
=
e
B
⋅
e
A
=
e
A
+
B
(
2
)
(
e
A
)
−
1
=
e
−
A
e
j
A
=
cos
A
+
j
sin
A
e
−
A
t
=
(
e
A
t
)
−
1
e
P
−
1
A
P
t
=
P
−
1
e
A
t
P
ln
(
e
A
)
=
A
(
3
)
ln
(
A
B
)
=
ln
A
+
ln
B
(
4
)
d
d
t
e
A
t
=
A
e
A
t
=
e
A
t
A
d
d
t
(
A
(
t
)
B
(
t
)
)
=
A
′
B
+
A
B
′
d
d
t
(
A
2
(
t
)
)
=
A
′
A
+
A
A
′
≠
2
A
A
′
d
d
t
(
e
A
(
t
)
)
≠
A
′
(
t
)
e
A
(
t
)
≠
e
A
(
t
)
A
′
(
t
)
(
5
)
\begin{aligned} & A\text{e}^{A}=\text{e}^{A}A \\ & \text{e}^{A}\cdot\text{e}^{B}=\text{e}^{B}\cdot\text{e}^{A}=\text{e}^{A+B}\qquad(2) \\ & (\text{e}^{A})^{-1}=\text{e}^{-A} \\ & \text{e}^{jA}=\cos A+j\sin A \\ & \text{e}^{-At}=(\text{e}^{At})^{-1} \\ & \text{e}^{P^{-1}APt}=P^{-1}\text{e}^{At}P \\ & \ln(\text{e}^A)=A\qquad(3) \\ & \ln(AB)=\ln A+\ln B\qquad(4) \\ & \frac{\text{d}}{\text{d}t}\text{e}^{At}=A\text{e}^{At}=\text{e}^{At}A \\ & \frac{\text{d}}{\text{d}t}(A(t)B(t))=A'B+AB' \\ & \frac{\text{d}}{\text{d}t}(A^2(t))=A'A+AA'\neq2AA' \\ & \frac{\text{d}}{\text{d}t}(\text{e}^{A(t)})\neq A'(t)\text{e}^{A(t)}\neq\text{e}^{A(t)}A'(t)\qquad(5) \\ \end{aligned}
AeA=eAAeA⋅eB=eB⋅eA=eA+B(2)(eA)−1=e−AejA=cosA+jsinAe−At=(eAt)−1eP−1APt=P−1eAtPln(eA)=A(3)ln(AB)=lnA+lnB(4)dtdeAt=AeAt=eAtAdtd(A(t)B(t))=A′B+AB′dtd(A2(t))=A′A+AA′=2AA′dtd(eA(t))=A′(t)eA(t)=eA(t)A′(t)(5)
注:
(1)需满足收敛域条件
∣
A
∣
<
1
|A|<1
∣A∣<1
(2)需满足
A
B
=
B
A
AB=BA
AB=BA。
(3)需满足
∣
e
A
−
I
∣
<
1
|\text{e}^A-I|<1
∣eA−I∣<1
(4)需满足
∣
A
B
−
I
∣
<
1
|AB-I|<1
∣AB−I∣<1和
A
B
=
B
A
AB=BA
AB=BA。
(5)仅当
A
′
A
=
A
A
′
A'A=AA'
A′A=AA′时才能为等号但一般不等。
常系数齐次/非齐次线性微分方程组
x
′
(
t
)
=
A
x
(
t
)
x
(
t
)
=
e
A
(
t
−
t
0
)
x
(
t
0
)
x
′
(
t
)
=
A
x
(
t
)
+
f
(
t
)
x
(
t
)
=
e
A
(
t
−
t
0
)
x
(
t
0
)
+
∫
t
0
t
e
A
(
t
−
τ
)
f
(
τ
)
d
τ
\begin{aligned} & \boldsymbol{x}'(t)=A\boldsymbol{x}(t) \\ & \boldsymbol{x}(t)=\text{e}^{A(t-t_0)}\boldsymbol{x}(t_0) \\ & \boldsymbol{x}'(t)=A\boldsymbol{x}(t)+\boldsymbol{f}(t) \\ & \boldsymbol{x}(t)=\text{e}^{A(t-t_0)}\boldsymbol{x}(t_0) +\int_{t_0}^t\text{e}^{A(t-\tau)}\boldsymbol{f}(\tau)\text{d}\tau \\ \end{aligned}
x′(t)=Ax(t)x(t)=eA(t−t0)x(t0)x′(t)=Ax(t)+f(t)x(t)=eA(t−t0)x(t0)+∫t0teA(t−τ)f(τ)dτ
证明
x
′
−
A
x
=
f
e
−
A
t
(
x
′
−
A
x
)
=
(
e
−
A
t
x
)
′
=
e
−
A
t
f
e
−
A
t
x
(
t
)
=
∫
t
0
t
e
−
A
τ
f
(
τ
)
d
τ
+
C
C
=
e
−
A
t
0
x
(
t
0
)
x
(
t
)
=
e
A
(
t
−
t
0
)
x
(
t
0
)
+
∫
t
0
t
e
A
(
t
−
τ
)
f
(
τ
)
d
τ
\begin{aligned} & x'-Ax=f \\ & \text{e}^{-At}(x'-Ax)=(\text{e}^{-At}x)'=\text{e}^{-At}f \\ & \text{e}^{-At}x(t)=\int_{t_0}^t\text{e}^{-A\tau}f(\tau)\text{d}\tau+C \\ & C=\text{e}^{-At_0}x(t_0) \\ & x(t)=\text{e}^{A(t-t_0)}x(t_0)+\int_{t_0}^t\text{e}^{A(t-\tau)}f(\tau)\text{d}\tau \\ \end{aligned}
x′−Ax=fe−At(x′−Ax)=(e−Atx)′=e−Atfe−Atx(t)=∫t0te−Aτf(τ)dτ+CC=e−At0x(t0)x(t)=eA(t−t0)x(t0)+∫t0teA(t−τ)f(τ)dτ
矩阵的伪逆
设A为n×m阶矩阵,其元素可以是复数。我们称基于A形成的矩阵代数方程组
Jordan标准型
A
=
[
3
1
−
1
1
2
−
1
2
1
0
]
A=\left[\begin{matrix} 3 & 1 & -1 \\ 1 & 2 & -1 \\ 2 & 1 & 0 \end{matrix}\right]
A=
312121−1−10
方法一:
求矩阵的特征值和特征向量
f
A
(
λ
)
=
λ
I
−
A
=
(
λ
−
1
)
(
λ
−
2
)
2
f_A(\lambda)=\lambda I-A=(\lambda-1)(\lambda-2)^2
fA(λ)=λI−A=(λ−1)(λ−2)2,
对于矩阵
A
A
A的特征值
λ
0
\lambda_0
λ0,
λ
0
\lambda_0
λ0的代数重数指
λ
0
\lambda_0
λ0作为特征多项式
f
A
(
λ
)
f_A(\lambda)
fA(λ)的根的重数。
λ
0
\lambda_0
λ0的几何重数指
λ
0
\lambda_0
λ0对应的特征向量的个数。
任一特征值的代数重数不小于它的几何重数。
λ
=
1
\lambda=1
λ=1对应的特征向量为
[
0
,
1
,
1
]
T
[0,1,1]^\text{T}
[0,1,1]T,代数重数1,几何重数1。
λ
=
2
\lambda=2
λ=2对应的特征向量为
[
1
,
0
,
1
]
T
[1,0,1]^\text{T}
[1,0,1]T,代数重数2,几何重数1。
将Jordan标准型写成
A
P
=
P
J
AP=PJ
AP=PJ的形式
J
=
[
1
0
0
0
2
1
0
0
2
]
[
3
1
−
1
1
2
−
1
2
1
0
]
[
0
1
1
1
0
1
1
1
1
]
=
[
0
1
1
1
0
1
1
1
1
]
[
1
0
0
0
2
1
0
0
2
]
\begin{aligned} J=&\left[\begin{matrix} 1 & 0 & 0 \\ 0 & 2 & 1 \\ 0 & 0 & 2 \end{matrix}\right] \\ \left[\begin{matrix} 3 & 1 & -1 \\ 1 & 2 & -1 \\ 2 & 1 & 0 \end{matrix}\right] \left[\begin{matrix} 0 & 1 & 1 \\ 1 & 0 & 1 \\ 1 & 1 & 1 \end{matrix}\right] =&\left[\begin{matrix} 0 & 1 & 1 \\ 1 & 0 & 1 \\ 1 & 1 & 1 \end{matrix}\right] \left[\begin{matrix} 1 & 0 & 0 \\ 0 & 2 & 1 \\ 0 & 0 & 2 \end{matrix}\right] \end{aligned}
J=
312121−1−10
011101111
=
100020012
011101111
100020012
其中
P
P
P由
A
A
A的特征向量和广义特征向量组成。
方法二:
行列式因子
D
k
(
λ
)
D_k(\lambda)
Dk(λ):所有子式的最大公约数(3阶矩阵有4+4+1个2阶子式)。
不变因子
d
k
(
λ
)
d_k(\lambda)
dk(λ):n阶与n-1阶行列式因子的商。
初等因子:所有
(
λ
−
a
)
b
(\lambda-a)^b
(λ−a)b 形式的项。
几个例子
行列式
D
1
=
1
,
D
2
=
λ
−
1
,
D
3
=
(
λ
−
1
)
3
D_1=1,D_2=\lambda-1,D_3=(\lambda-1)^3
D1=1,D2=λ−1,D3=(λ−1)3
不变
d
1
=
1
,
d
2
=
λ
−
1
,
d
3
=
(
λ
−
1
)
2
d_1=1,d_2=\lambda-1,d_3=(\lambda-1)^2
d1=1,d2=λ−1,d3=(λ−1)2
初等
λ
−
1
,
(
λ
−
1
)
2
\lambda-1,(\lambda-1)^2
λ−1,(λ−1)2
行列式
D
1
=
1
,
D
2
=
1
,
D
3
=
(
λ
−
1
)
(
λ
−
2
)
2
D_1=1,D_2=1,D_3=(\lambda-1)(\lambda-2)^2
D1=1,D2=1,D3=(λ−1)(λ−2)2
不变
d
1
=
1
,
d
2
=
1
,
d
3
=
(
λ
−
1
)
(
λ
−
2
)
2
d_1=1,d_2=1,d_3=(\lambda-1)(\lambda-2)^2
d1=1,d2=1,d3=(λ−1)(λ−2)2
初等
λ
−
1
,
(
λ
−
2
)
2
\lambda-1,(\lambda-2)^2
λ−1,(λ−2)2
λ
\lambda
λ 矩阵的初等变换之间等价,Smith标准型中的各个非零元素为不变因子。
矩阵分解
LU分解
[ 1 0 0 0 1 0 0 c 1 ] [ 1 0 0 a 1 0 b 0 1 ] A = U A = [ 1 0 0 a 1 0 b 0 1 ] − 1 [ 1 0 0 0 1 0 0 c 1 ] − 1 U = [ 1 0 0 − a 1 0 − b 0 1 ] [ 1 0 0 0 1 0 0 − c 1 ] U = [ 1 0 0 − a 1 0 − b − c 1 ] U \begin{aligned} &\left[\begin{matrix} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & c & 1 \\ \end{matrix}\right] \left[\begin{matrix} 1 & 0 & 0 \\ a & 1 & 0 \\ b & 0 & 1 \\ \end{matrix}\right]A=U \\ A=&\left[\begin{matrix} 1 & 0 & 0 \\ a & 1 & 0 \\ b & 0 & 1 \\ \end{matrix}\right]^{-1} \left[\begin{matrix} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & c & 1 \\ \end{matrix}\right]^{-1}U \\ =&\left[\begin{matrix} 1 & 0 & 0 \\ -a & 1 & 0 \\ -b & 0 & 1 \\ \end{matrix}\right] \left[\begin{matrix} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & -c & 1 \\ \end{matrix}\right]U =\left[\begin{matrix} 1 & 0 & 0 \\ -a & 1 & 0 \\ -b & -c & 1 \\ \end{matrix}\right]U \end{aligned} A== 10001c001 1ab010001 A=U 1ab010001 −1 10001c001 −1U 1−a−b010001 10001−c001 U= 1−a−b01−c001 U
QR分解
A
=
[
2
−
3
1
1
1
−
6
2
4
2
]
H
=
I
−
2
ω
ω
T
\begin{aligned} A=&\left[\begin{matrix} 2 & -3 & 1 \\ 1 & 1 & -6 \\ 2 & 4 & 2 \\ \end{matrix}\right] \\ H=&I-2\omega\omega^\text{T} \end{aligned}
A=H=
212−3141−62
I−2ωωT
将第一列向量
[
2
,
1
,
2
]
T
[2,1,2]^\text{T}
[2,1,2]T变换到x轴上并保持长度不变,即
[
3
,
0
,
0
]
T
[3,0,0]^\text{T}
[3,0,0]T,
ω
T
=
[
2
,
1
,
2
]
−
[
3
,
0
,
0
]
=
[
−
1
,
1
,
2
]
\omega^\text{T}=[2,1,2]-[3,0,0]=[-1,1,2]
ωT=[2,1,2]−[3,0,0]=[−1,1,2]后再取单位向量,代入
H
H
H,则。
H
=
[
2
3
1
3
2
3
1
3
2
3
−
2
3
2
3
−
2
3
−
1
3
]
[
2
1
2
]
=
[
3
0
0
]
H=\left[\begin{matrix} \displaystyle\frac{2}{3} & \displaystyle\frac{1}{3} & \displaystyle\frac{2}{3} \\ \displaystyle\frac{1}{3} & \displaystyle\frac{2}{3} & -\displaystyle\frac{2}{3} \\ \displaystyle\frac{2}{3} & -\displaystyle\frac{2}{3} & -\displaystyle\frac{1}{3} \\ \end{matrix}\right] \left[\begin{matrix} 2 \\ 1 \\ 2 \\ \end{matrix}\right] =\left[\begin{matrix} 3 \\ 0 \\ 0 \\ \end{matrix}\right]
H=
3231323132−3232−32−31
212
=
300
满秩分解
A = [ F ⋮ G ] , L F = I L A = C = [ I ⋮ L G ] , B = F B C = F [ I ⋮ L G ] = [ F ⋮ F L G ] = [ F ⋮ L F G ] = A A=[F\vdots G], LF=I \\ LA=C=[I\vdots LG], B=F \\ BC=F[I\vdots LG]=[F\vdots FLG]=[F\vdots LFG]=A A=[F⋮G],LF=ILA=C=[I⋮LG],B=FBC=F[I⋮LG]=[F⋮FLG]=[F⋮LFG]=A
奇异值分解
A = U D V T V 1 = A T U 1 D − 1 U 1 = A V 1 D − 1 A=UDV^\text{T} \\ V_1=A^\text{T}U_1D^{-1} \\ U_1=AV_1D^{-1} \\ A=UDVTV1=ATU1D−1U1=AV1D−1
广义逆矩阵
方程组有解称作相容,否则称为不相容或矛盾方程组。相容方程组的通解为
x
=
A
−
b
+
(
I
−
A
−
A
)
z
x=A^-b+(I-A^-A)z
x=A−b+(I−A−A)z
其中
z
z
z是与
x
x
x同维的任意向量,
A
−
A^-
A−是
A
A
A的任一减号逆,满足
A
A
−
A
=
A
AA^-A=A
AA−A=A
如果还另外满足
A
−
A
A
−
=
A
−
A^-AA^-=A^-
A−AA−=A−,则称
A
−
A^-
A−是
A
A
A的自反减号逆,记作
A
r
−
A_r^-
Ar−。证明:
A
x
=
A
A
−
b
+
A
(
I
−
A
−
A
)
z
=
A
A
−
b
Ax = AA^-b+A(I-A^-A)z = AA^-b
Ax=AA−b+A(I−A−A)z=AA−b
矩阵
A
A
A的左逆和右逆分别满足 $ A_L{-1}A=I$和$AA_R{-1}=I
,当
,当
,当A$分别列满秩和行满秩时,分别存在左逆和右逆
A
L
−
1
=
(
A
T
A
)
−
1
A
T
A
R
−
1
=
A
T
(
A
A
T
)
−
1
A_L^{-1} = (A^{\text{T}}A)^{-1}A^\text{T} \\ A_R^{-1} = A^\text{T}(AA^{\text{T}})^{-1}
AL−1=(ATA)−1ATAR−1=AT(AAT)−1
且分别是列满秩和行满秩矩阵
A
A
A的自反减号逆
A
r
−
A_r^-
Ar−,同时也是最小范数广义逆
A
m
−
A_m^-
Am−、最小二乘广义逆
A
l
−
A_l^-
Al−、加号逆
A
+
A^+
A+。
举例,方程组
{
x
1
+
2
x
2
+
x
3
=
1
−
x
2
+
2
x
3
=
2
,
A
x
=
b
\begin{cases} x_1+2x_2+x_3 &= 1 \\ -x_2+2x_3 &= 2 \end{cases}, Ax=b
{x1+2x2+x3−x2+2x3=1=2,Ax=b
行满秩,可求其右逆
A
R
−
1
A_R^{-1}
AR−1为减号逆
A
−
A^-
A−,并求出通解。
A
A
A的加号逆为同时满足下面4个式子的矩阵,且唯一。
A
G
A
=
A
G
A
G
=
G
(
A
G
)
T
=
A
G
(
G
A
)
T
=
G
A
\begin{aligned} &AGA=A \\ &GAG=G \\ &(AG)^\text{T}=AG \\ &(GA)^\text{T}=GA \\ \end{aligned}
AGA=AGAG=G(AG)T=AG(GA)T=GA
当
其他
- 列满秩矩阵的列向量线性无关,一般为竖形状。列范数 ∣ ∣ A ∣ ∣ 1 ||A||_1 ∣∣A∣∣1指对每一列元素(每个列向量)的绝对值求和的最大值。
- ∣ ∣ A ∣ ∣ m 1 ||A||_{m_1} ∣∣A∣∣m1范数是所有元素绝对值的和; ∣ ∣ A ∣ ∣ m 2 / ∣ ∣ A ∣ ∣ F ||A||_{m_2}/||A||_F ∣∣A∣∣m2/∣∣A∣∣F范数是所有元素平方和开根号; ∣ ∣ A ∣ ∣ m ∞ ||A||_{m_\infty} ∣∣A∣∣m∞范数是所有元素绝对值最大值×阶数。
(
A
+
B
C
D
)
−
1
=
A
−
1
−
A
−
1
B
(
D
A
−
1
B
+
C
−
1
)
−
1
D
A
−
1
(
A
+
u
⃗
u
⃗
T
)
−
1
=
A
−
1
−
A
−
1
u
⃗
u
⃗
T
A
−
1
1
+
u
⃗
T
A
−
1
u
⃗
\begin{aligned} (A+BCD)^{-1} =& A^{-1}-A^{-1} B (DA^{-1}B+C^{-1})^{-1}DA^{-1} \\ (A+\vec{u}\vec{u}^\text{T})^{-1} =& A^{-1} -\frac{A^{-1}\vec{u}\vec{u}^\text{T} A^{-1}} {1+\vec{u}^\text{T}A^{-1}\vec{u}} \end{aligned}
(A+BCD)−1=(A+uuT)−1=A−1−A−1B(DA−1B+C−1)−1DA−1A−1−1+uTA−1uA−1uuTA−1
a
⃗
×
(
b
⃗
×
c
⃗
)
=
(
a
⃗
⋅
c
⃗
)
b
⃗
−
(
a
⃗
⋅
b
⃗
)
c
⃗
\vec a\times(\vec b\times\vec c) =(\vec a\cdot\vec c)\vec b -(\vec a\cdot\vec b)\vec c
a×(b×c)=(a⋅c)b−(a⋅b)c
(
a
⃗
×
b
⃗
)
×
c
⃗
=
(
a
⃗
⋅
c
⃗
)
b
⃗
−
a
⃗
(
b
⃗
⋅
c
⃗
)
(\vec a\times\vec b)\times\vec c =(\vec a\cdot\vec c)\vec b -\vec a(\vec b\cdot\vec c)
(a×b)×c=(a⋅c)b−a(b⋅c)