目录
0. 顺序主子式
1. 三角分解(为方便Guass消去法-化上阶梯矩阵的过程)
1.1 LU分解
- 目的: 将 n \boldsymbol{n} n 阶矩阵 A A A分解为一个下三角矩阵L和一个上三角矩阵U的乘积
- 形式: A = L U A = LU A=LU
- 推论:设
A
\boldsymbol{A}
A 是
n
\boldsymbol{n}
n 阶非奇异矩阵,
A
\boldsymbol{A}
A 有三角分解
A
=
L
U
\boldsymbol{A}=\boldsymbol{L} \boldsymbol{U}
A=LU,
的充要条件是 A \mathrm{A} A 的顺序主子式 Δ k ≠ 0 k = 1 , 2 , ⋯ , n \Delta_{k} \neq 0 \quad k=1,2, \cdots, n Δk=0k=1,2,⋯,n
1.2 LDU分解
- 目的: 将 n \boldsymbol{n} n 阶矩阵 A A A分解为一个单位下三角矩阵L、对角矩阵D和一个单位上三角矩阵U的乘积
- 形式: A = L D U A=LDU A=LDU
- 定理:
p:矩阵 A = ( a i j ) n × n A=\left(a_{i j}\right)_{n \times n} A=(aij)n×n 的分解式唯一 ⇔ \Leftrightarrow ⇔ q:A的顺序主子式 Δ k ≠ 0 \Delta_{k} \neq 0 Δk=0
且这个唯一的分解式 A = L D U A=LDU A=LDU中, D = diag ( d 1 , d 2 , ⋯ , d n ) , d k = Δ k Δ k − 1 , k = 1 , 2 , ⋯ , n ( Δ 0 = 1 ) D=\operatorname{diag}\left(d_{1}, d_{2}, \cdots, d_{n}\right), d_{k}=\frac{\Delta_{k}}{\Delta_{k-1}}, k=1,2, \cdots, n \quad\left(\Delta_{0}=1\right) D=diag(d1,d2,⋯,dn),dk=Δk−1Δk,k=1,2,⋯,n(Δ0=1)
1.3 LDU分解步骤
-
目的: A = L D U = L ~ U \boldsymbol{A}=\boldsymbol{L} \boldsymbol{D} \boldsymbol{U}=\tilde{\boldsymbol{L}} \boldsymbol{U} A=LDU=L~U,求其中的L和U, L = L ~ = [ l 11 l 21 l 22 ⋮ ⋮ ⋱ l n 1 l n 2 ⋯ l n n ] , U = [ 1 u 12 ⋯ u 1 n 1 ⋯ u 2 n ⋱ ⋮ 1 ] \boldsymbol{L}=\tilde{\boldsymbol{L}}=\left[\begin{array}{cccc}l_{11} & & & \\ l_{21} & l_{22} & & \\ \vdots & \vdots & \ddots & \\ l_{n 1} & l_{n 2} & \cdots & l_{n n}\end{array}\right], \boldsymbol{U}=\left[\begin{array}{cccc}1 & u_{12} & \cdots & u_{1 n} \\ & 1 & \cdots & u_{2 n} \\ & & \ddots & \vdots \\ & & & 1\end{array}\right] L=L~=⎣⎢⎢⎢⎡l11l21⋮ln1l22⋮ln2⋱⋯lnn⎦⎥⎥⎥⎤,U=⎣⎢⎢⎢⎡1u121⋯⋯⋱u1nu2n⋮1⎦⎥⎥⎥⎤,使得 [ a 11 a 12 ⋯ a 1 n a 21 a 22 a 2 n ⋮ ⋮ ⋱ ⋮ a n 1 a n 2 ⋯ a n n ] = [ l 11 l 21 l 22 ⋮ ⋮ ⋱ l n 1 l n 2 ⋯ l n n ] [ 1 u 12 ⋯ u 1 n 1 ⋯ u 2 n ⋱ ⋮ 1 ] \left[\begin{array}{cccc}a_{11} & a_{12} & \cdots & a_{1 n} \\ a_{21} & a_{22} & & a_{2 n} \\ \vdots & \vdots & \ddots & \vdots \\ a_{n 1} & a_{n 2} & \cdots & a_{n n}\end{array}\right]=\left[\begin{array}{ccc}l_{11} & & \\ l_{21} & l_{22} & & \\ \vdots & \vdots & \ddots & \\ l_{n 1} & l_{n 2} & \cdots & l_{n n}\end{array}\right]\left[\begin{array}{ccc}1 & u_{12} & \cdots & u_{1 n} \\ & 1 & \cdots & u_{2 n} \\ & & \ddots & \vdots \\ & & & 1\end{array}\right] ⎣⎢⎢⎢⎡a11a21⋮an1a12a22⋮an2⋯⋱⋯a1na2n⋮ann⎦⎥⎥⎥⎤=⎣⎢⎢⎢⎡l11l21⋮ln1l22⋮ln2⋱⋯lnn⎦⎥⎥⎥⎤⎣⎢⎢⎢⎡1u121⋯⋯⋱u1nu2n⋮1⎦⎥⎥⎥⎤
-
步骤:
-
例题:
2. QR分解
- 目的:将矩阵分解为正交矩阵与上三角矩阵之积
- 定理: A n × n A_{n \times n} An×n 可逆 ⇒ ∃ \Rightarrow \exists ⇒∃ 正交矩阵 Q \boldsymbol{Q} Q, 可逆上三角矩阵 R \boldsymbol{R} R, 使得 A = Q R \boldsymbol{A}=\boldsymbol{Q} \boldsymbol{R} A=QR 。
2.0 施密特正交化法进行QR分解
- 方法:Schmidt正交化方法
( a 1 , a 2 , ⋯ , a n ) = ( b 1 , b 2 , ⋯ , b n ) K \left(a_{1}, a_{2}, \cdots, a_{n}\right)=\left(b_{1}, b_{2}, \cdots, b_{n}\right) K (a1,a2,⋯,an)=(b1,b2,⋯,bn)K
其中, { b 1 = a 1 b 2 = a 2 − k 21 b 1 ⋯ ⋯ b n = a n − k n , n − 1 b n − 1 − ⋯ − k n 1 b 1 \left\{\begin{aligned} b_{1}=& a_{1} \\ b_{2}=& a_{2}-k_{21} b_{1} \\ & \cdots \cdots \\ b_{n}=& a_{n}-k_{n, n-1} b_{n-1}-\cdots-k_{n 1} b_{1} \end{aligned}\right. ⎩⎪⎪⎪⎪⎨⎪⎪⎪⎪⎧b1=b2=bn=a1a2−k21b1⋯⋯an−kn,n−1bn−1−⋯−kn1b1, K = [ 1 k 21 ⋯ k n 1 1 ⋯ k n 2 ⋱ ⋮ 1 ] K=\left[\begin{array}{cccc} 1 & k_{21} & \cdots & k_{n 1} \\ & 1 & \cdots & k_{n 2} \\ & & \ddots & \vdots \\ & & & 1 \end{array}\right] K=⎣⎢⎢⎢⎡1k211⋯⋯⋱kn1kn2⋮1⎦⎥⎥⎥⎤
令 Q = ( q 1 , q 2 , ⋯ , q n ) , R = [ ∣ b 1 ∣ ∣ b 2 ∣ ⋱ ∣ b n ∣ ] [ 1 k 21 ⋯ k n 1 1 ⋯ k n 2 ⋱ ⋮ 1 ] Q=\left(q_{1}, q_{2}, \cdots, q_{n}\right), R=\left[\begin{array}{llll}\left|b_{1}\right| & & & \\ & \left|b_{2}\right| & & \\ & & \ddots & \\ & & & \left|b_{n}\right|\end{array}\right]\left[\begin{array}{cccc}1 & k_{21} & \cdots & k_{n 1} \\ & 1 & \cdots & k_{n 2} \\ & & \ddots & \vdots \\ & & & & 1\end{array}\right] Q=(q1,q2,⋯,qn),R=⎣⎢⎢⎡∣b1∣∣b2∣⋱∣bn∣⎦⎥⎥⎤⎣⎢⎢⎢⎡1k211⋯⋯⋱kn1kn2⋮1⎦⎥⎥⎥⎤ - 例题:
2.1 Givens变换(由Givens矩阵确定的变换)
Givens矩阵性质:
- T i j T T i j = T i j ( c , − s ) = [ I c − s I s c I ] , ∣ T i j ∣ = 1 T_{i j}^{T} T_{i j}=T_{i j}(c,-s)=\left[\begin{array}{ccccc} I & & & & \\ & c & & -s & \\ & & I & & \\ & s & & c & \\ & & & & I \end{array}\right], \quad\left|T_{i j}\right|=1 TijTTij=Tij(c,−s)=⎣⎢⎢⎢⎢⎡IcsI−scI⎦⎥⎥⎥⎥⎤,∣Tij∣=1
-
x
=
[
ξ
1
⋮
ξ
n
]
,
T
i
j
x
=
[
η
1
⋮
η
n
]
⇒
{
η
i
=
c
ξ
i
+
s
ξ
j
η
j
=
−
s
ξ
i
+
c
ξ
j
η
k
=
ξ
k
(
k
≠
i
,
j
)
x=\left[\begin{array}{c}\xi_{1} \\ \vdots \\ \xi_{n}\end{array}\right], \quad T_{i j} x=\left[\begin{array}{c}\eta_{1} \\ \vdots \\ \eta_{n}\end{array}\right] \Rightarrow\left\{\begin{array}{l}\eta_{i}=c \xi_{i}+s \xi_{j} \\ \eta_{j}=-s \xi_{i}+c \xi_{j} \\ \eta_{k}=\xi_{k}(k \neq i, j)\end{array}\right.
x=⎣⎢⎡ξ1⋮ξn⎦⎥⎤,Tijx=⎣⎢⎡η1⋮ηn⎦⎥⎤⇒⎩⎨⎧ηi=cξi+sξjηj=−sξi+cξjηk=ξk(k=i,j)
若 ξ i 2 + ξ j 2 ≠ 0 \xi_{i}^{2}+\xi_{j}^{2} \neq 0 ξi2+ξj2=0, , 取 c = ξ i ξ i 2 + ξ j 2 , s = ξ j ξ i 2 + ξ j 2 c=\frac{\xi_{i}}{\sqrt{\xi_{i}^{2}+\xi_{j}^{2}}}, s=\frac{\xi_{j}}{\sqrt{\xi_{i}^{2}+\xi_{j}^{2}}} c=ξi2+ξj2ξi,s=ξi2+ξj2ξj
则 η i = ξ i 2 + ξ j 2 > 0 , η j = 0 \eta_{i}=\sqrt{\xi_{i}^{2}+\xi_{j}^{2}}>0, \quad \eta_{j}=0 ηi=ξi2+ξj2>0,ηj=0
- 例题:
2.2 HouseHolder变换
- 定义:
设单位列向量 u ∈ R n u \in R^{n} u∈Rn, 称 H = I − 2 u u T H=I-2 u u^{T} H=I−2uuT为Householder矩阵(初等反射矩阵), 由 H \mathbf{H} H 矩阵确定的线性变换称为Householder变换。
注意: e i e i T 即 在 第 i 行 i 列 为 1 , 其 余 都 为 0 的 矩 阵 e_{i} e_{i}^{T}即在第i行i列为1,其余都为0的矩阵 eieiT即在第i行i列为1,其余都为0的矩阵
-
性质:
H u = I n − 2 u u T ( u ∈ R n \boldsymbol{H}_{u}=\boldsymbol{I}_{n}-2 u u^{T} \quad\left(u \in R^{n}\right. Hu=In−2uuT(u∈Rn 是单位列向量 ) ) )
(1) H = H T \boldsymbol{H}=\boldsymbol{H}^{T} H=HT 对称
(2) H T H = I \boldsymbol{H}^{T} \boldsymbol{H}=\boldsymbol{I} HTH=I 正交
(3) H 2 = I \boldsymbol{H}^{2}=\boldsymbol{I} H2=I 对合
(4) H − 1 = H \boldsymbol{H}^{-1}=\boldsymbol{H} H−1=H 自逆
(5) det H = − 1 \operatorname{det} \boldsymbol{H}=-\mathbf{1} detH=−1 自逆 -
例题:
u = x − ∣ x ∣ z ∣ x − ∣ x ∣ z ∣ u=\frac{x-|x| z}{|x-| x|z|} u=∣x−∣x∣z∣x−∣x∣z
2.3 G矩阵和H矩阵关系
- G矩阵 T i j ( c , s ) ⇒ ∃ H , T_{i j}(c, s) \Rightarrow \exists \mathrm{H}, Tij(c,s)⇒∃H,矩阵 H u H_{u} Hu 与 H v H_{v} Hv, st T i j = H u H v T_{i j}=H_{u} H_{v} Tij=HuHv
T i j ( c , s ) = [ I cos θ sin θ I − sin θ cos θ I ] T_{i j}(c, s)=\left[\begin{array}{cccc}I & & & \\ & \cos \theta & & \sin \theta & \\ & & I & & \\ & -\sin \theta & & \cos \theta & \\ & & & &I\end{array}\right] Tij(c,s)=⎣⎢⎢⎢⎢⎡Icosθ−sinθIsinθcosθI⎦⎥⎥⎥⎥⎤
ν = [ 0 ⋯ 0 sin θ 4 0 ⋯ 0 cos θ 4 0 ⋯ 0 ] T \nu=\left[\begin{array}{lllllllll}0 & \cdots & 0 & \sin \frac{\theta}{4} & 0 & \cdots & 0 & \cos \frac{\theta}{4} & 0 & \cdots & 0\end{array}\right]^{T} ν=[0⋯0sin4θ0⋯0cos4θ0⋯0]T
u = [ 0 ⋯ 0 sin 3 θ 4 0 ⋯ 0 cos 3 θ 4 0 ⋯ 0 ] T u=\left[\begin{array}{lllllllll}0 & \cdots & 0 & \sin \frac{3 \theta}{4} & 0 & \cdots & 0 & \cos \frac{3 \theta}{4} & 0 & \cdots & 0\end{array}\right]^{T} u=[0⋯0sin43θ0⋯0cos43θ0⋯0]T
- H-矩阵不能由若干个 G矩阵的乘积来表示
det H = − 1 \operatorname{det} H=-1 detH=−1, 而 det G = 1 \operatorname{det} G=1 detG=1
2.4 G变化法进行QR分解
2.5 H变换法进行QR分解
3. 矩阵的满秩分解
- 目的:对 A ∈ C r m × n ( n ≥ 1 ) A \in C_{r}^{m \times n}(n \geq 1) A∈Crm×n(n≥1), 求 F ∈ C r m × r F \in C_{r}^{m \times r} F∈Crm×r, 及 G ∈ C r r × n G \in C_{r}^{r \times n} G∈Crr×n 使 A m × n = F m × r G r × n A_{m×n}=F_{m×r} G_{r×n} Am×n=Fm×rGr×n
- 例题:
A = P − 1 B = ( F m × r ∣ S m × ( m − r ) ) ( G O ) = F G A=P^{-1} B=\left(F_{m \times r} \mid S_{m \times(m-r)}\right)\left(\begin{array}{l}G \\ O\end{array}\right)=F G A=P−1B=(Fm×r∣Sm×(m−r))(GO)=FG
4. 奇异值分解SVD
4.0 Hermite(半)正定矩阵和酉矩阵
矩阵A是Hermite矩阵,则 A H = A A^{H}=A AH=A
设 A \mathrm{A} A 为 n \mathrm{n} n 阶Hermite矩阵,如果对任意 n n n 维向量 X \mathrm{X} X 都有 x H A x ≥ 0 x^{H} A x \geq 0 xHAx≥0 ,则称 A \mathrm{A} A 为半正定(非 负定) 矩阵; 如果对任意 n n n 维向量X都有 x H A x > 0 x^{H} A x>0 xHAx>0 ,则称 A A A 为正定矩阵。
酉矩阵: Q H Q = I Q^{H} Q=I QHQ=I
4.1 奇异值分解
目的:
A
=
U
D
V
H
A=U D V^{H}
A=UDVH
定理 16:
A
m
×
n
∈
C
r
m
×
n
(
r
≥
1
)
,
Σ
r
=
[
σ
1
⋱
σ
r
]
A_{m \times n} \in C_{r}^{m \times n}(r \geq 1), \Sigma_{r}=\left[\begin{array}{lll}\sigma_{1} & & \\ & \ddots & \\ & & \sigma_{r}\end{array}\right]
Am×n∈Crm×n(r≥1),Σr=⎣⎡σ1⋱σr⎦⎤为奇异值矩阵,
存在酉矩阵
U
m
×
m
U_{m \times m}
Um×m 及
V
n
×
n
V_{n \times n}
Vn×n, 使得
U
H
A
V
=
[
Σ
r
0
0
0
]
m
×
n
≜
D
U^{H} A V=\left[\begin{array}{cc}\Sigma_{r} & 0 \\ 0 & 0\end{array}\right]_{m \times n} \triangleq D
UHAV=[Σr000]m×n≜D
定理17:
A
∈
C
r
m
×
n
(
r
≥
0
)
A \in C_{r}^{m \times n}(r \geq 0)
A∈Crm×n(r≥0) 的奇异值分解
A
=
U
[
Σ
0
0
0
]
V
H
A=U\left[\begin{array}{ll}\Sigma & 0 \\ 0 & 0\end{array}\right] V^{H}
A=U[Σ000]VH 中, 划分
U
=
(
u
1
,
u
2
,
⋯
,
u
m
)
,
V
=
(
v
1
,
v
2
,
⋯
,
v
n
)
U=\left(u_{1}, u_{2}, \cdots, u_{m}\right), V=\left(v_{1}, v_{2}, \cdots, v_{n}\right)
U=(u1,u2,⋯,um),V=(v1,v2,⋯,vn), 则有
(1)
N
(
A
)
=
span
{
v
r
+
1
,
v
r
+
2
,
⋯
,
v
n
}
N(A)=\operatorname{span}\left\{v_{r+1}, v_{r+2}, \cdots, v_{n}\right\}
N(A)=span{vr+1,vr+2,⋯,vn};
(2)
R
(
A
)
=
span
{
u
1
,
u
2
,
⋯
,
u
r
}
R(A)=\operatorname{span}\left\{u_{1}, u_{2}, \cdots, u_{r}\right\}
R(A)=span{u1,u2,⋯,ur};
(3)
A
=
σ
1
u
1
v
1
H
+
σ
2
u
2
v
2
H
+
⋯
+
σ
r
u
r
v
r
H
A=\sigma_{1} u_{1} v_{1}^{\mathrm{H}}+\sigma_{2} u_{2} v_{2}^{\mathrm{H}}+\cdots+\sigma_{r} u_{r} v_{r}^{\mathrm{H}}
A=σ1u1v1H+σ2u2v2H+⋯+σrurvrH
A
m
×
n
,
B
m
×
n
A_{m \times n}, B_{m \times n}
Am×n,Bm×n, 若有酉矩阵
U
m
×
m
U_{m \times m}
Um×m 及
V
n
×
n
V_{n \times n}
Vn×n, 使
U
H
A
V
=
B
U^{H} A V=B
UHAV=B 称
A
\boldsymbol{A}
A 与
B
\boldsymbol{B}
B 正交相抵。
定理18:
A
与
B
正交相抵
⇒
σ
A
=
σ
B
证明:
B
=
U
H
A
V
⇒
B
H
B
=
⋯
=
V
−
1
(
A
H
A
)
V
⇒
λ
B
H
B
=
λ
A
H
A
≥
0
⇒
σ
A
=
σ
B
\begin{aligned} A \text { 与 } B \text { 正交相抵 } & \Rightarrow \sigma_{A}=\sigma_{B} \\ \text { 证明: } \quad B=U^{H} A V & \Rightarrow B^{H} B=\cdots=V^{-1}\left(A^{H} A\right) V \\ & \Rightarrow \lambda_{B^{H} B}=\lambda_{A^{H} A} \geq 0 \\ & \Rightarrow \sigma_{A}=\sigma_{B} \end{aligned}
A 与 B 正交相抵 证明: B=UHAV⇒σA=σB⇒BHB=⋯=V−1(AHA)V⇒λBHB=λAHA≥0⇒σA=σB