Householder变换 https://www.cnblogs.com/reasno/p/9606224.html
Householder transformation https://en.wikipedia.org/wiki/Householder_transformation
[矩阵的QR分解系列三] 豪斯霍尔德(Householder)变换 https://blog.csdn.net/honyniu/article/details/110351391
北化 矩阵论
Householder矩阵
Householder矩阵:
H
=
I
−
2
w
w
H
H = I - 2ww^H
H=I−2wwH,
w
w
w为单位列向量
Householder矩阵性质:厄米特矩阵、正交矩阵、对合矩阵、反射矩阵
注:如果一个对合矩阵也是厄米特矩阵,那么它也是正交矩阵
Householder变换
Householder变换(初等反射阵或镜象变换阵):将向量
x
x
x通过Householder矩阵
H
H
H变换为
H
x
Hx
Hx
x
x
x与
H
x
Hx
Hx关于一个超平面对称,该超平面与
w
w
w垂直。
H
x
Hx
Hx是
x
x
x关于与
w
w
w垂直的超平面的镜面反射。
H
x
=
x
−
2
w
w
H
x
=
x
−
2
<
x
,
w
>
w
Hx = x - 2ww^H x = x - 2<x,w>w
Hx=x−2wwHx=x−2<x,w>w
x
x
x减去
x
x
x在
w
w
w上的投影分量,便是位于超平面内的
u
u
u
u
u
u减去
x
x
x在
w
w
w上的投影分量,便是
x
x
x的反射向量
H
x
Hx
Hx
x
x
x和其反射向量
H
x
Hx
Hx可以分解为
u
u
u与
w
w
w
x
=
u
+
α
w
x = u + \alpha w
x=u+αw
H
x
=
u
−
α
w
Hx = u - \alpha w
Hx=u−αw
α
\alpha
α为一个常数
H x = x − 2 w w H x Hx = x - 2ww^H x Hx=x−2wwHx 可写作: H x = y Hx = y Hx=y
定理1
定理:H为householder矩阵,则 [ I r 0 0 H ] \begin{bmatrix} I_r & 0 \\ 0 & H\end{bmatrix} [Ir00H]也是为householder矩阵
证明:
[
I
r
0
0
H
]
=
[
I
r
0
0
I
n
−
2
w
w
H
]
=
I
n
+
r
−
[
0
0
0
2
w
w
H
]
=
I
n
+
r
−
2
[
0
u
]
[
0
H
u
H
]
=
I
n
+
r
−
2
u
1
u
1
H
\begin{bmatrix} I_r & 0 \\ 0 & H\end{bmatrix} = \begin{bmatrix} I_r & 0 \\ 0 & I_n - 2ww^H \end{bmatrix} = I_{n+r} - \begin{bmatrix} 0 & 0 \\ 0 & 2ww^H \end{bmatrix} = I_{n+r} - 2\begin{bmatrix} 0 \\ u \end{bmatrix} \begin{bmatrix} 0^H & u^H \end{bmatrix} = I_{n+r} - 2u_1 u_1^H
[Ir00H]=[Ir00In−2wwH]=In+r−[0002wwH]=In+r−2[0u][0HuH]=In+r−2u1u1H
定理2
H
=
I
−
2
w
w
H
H = I - 2ww^H
H=I−2wwH
定理:必存在
H
x
=
±
∣
∣
x
∣
∣
z
,
∣
∣
z
∣
∣
=
1
Hx = \pm ||x|| z, ||z||=1
Hx=±∣∣x∣∣z,∣∣z∣∣=1
证明:
x
=
0
x=0
x=0,任取
w
w
w,则
H
x
=
±
∣
∣
x
∣
∣
z
Hx=\pm ||x|| z
Hx=±∣∣x∣∣z,定理在这种情况下成立
x
=
±
∣
∣
x
∣
∣
z
x = \pm ||x|| z
x=±∣∣x∣∣z,取与
x
x
x正交的
w
w
w,则
H
x
=
x
=
±
∣
∣
x
∣
∣
z
Hx=x=\pm ||x|| z
Hx=x=±∣∣x∣∣z,定理在这种情况下成立
x
≠
±
∣
∣
x
∣
∣
z
x \neq \pm ||x|| z
x=±∣∣x∣∣z,取
w
=
x
−
α
z
∣
∣
x
−
α
z
∣
∣
w = \frac{x- \alpha z}{||x - \alpha z||}
w=∣∣x−αz∣∣x−αz,则
H
x
=
x
−
2
(
x
−
α
z
)
(
x
−
α
z
)
H
∣
∣
x
−
α
z
∣
∣
2
x
=
α
z
=
±
∣
∣
x
∣
∣
z
Hx = x - 2\frac{(x- \alpha z)(x- \alpha z)^H}{||x - \alpha z||^2}x = \alpha z = \pm ||x|| z
Hx=x−2∣∣x−αz∣∣2(x−αz)(x−αz)Hx=αz=±∣∣x∣∣z,定理在这种情况下成立
推论: x x x与 e 1 e_1 e1不共线,取 w = x − α e 1 ∣ ∣ x − α e 1 ∣ ∣ w = \frac{x- \alpha e_1}{||x - \alpha e_1||} w=∣∣x−αe1∣∣x−αe1,则 H x = ± ∣ ∣ x ∣ ∣ e 1 Hx = \pm ||x|| e_1 Hx=±∣∣x∣∣e1
Householder矩阵的特征对
H
H
H为正交矩阵,则其有
n
n
n个特征对
如果
x
x
x与
w
w
w垂直,则
H
x
=
x
Hx = x
Hx=x
如果
x
=
w
x=w
x=w垂直,则
H
w
=
−
w
Hw = -w
Hw=−w
n
n
n维空间存在
n
−
1
n-1
n−1个向量与
w
w
w垂直
因此可以看出
H
H
H的n个特征对:
n
−
1
n-1
n−1个
(
1
,
x
i
)
(1,x_i)
(1,xi)和1个
(
−
1
,
w
)
(-1,w)
(−1,w)
因此
H
H
H的行列式为-1
利用Householder变换实现正交分解
矩阵
A
=
[
a
1
,
.
.
.
,
a
n
]
A = [a_1, ..., a_n]
A=[a1,...,an],实现
A
=
Q
R
A = QR
A=QR
利用
n
−
1
n-1
n−1个Householder矩阵
H
i
H_i
Hi对
A
A
A进行变换,使得
H
n
−
1
.
.
.
H
2
H
1
A
=
R
H_{n-1}...H_2 H_1 A = R
Hn−1...H2H1A=R,
Q
=
H
1
H
2
.
.
.
H
n
−
1
Q=H_1H_2...H_{n-1}
Q=H1H2...Hn−1
如果
a
1
≠
0
a_1 \neq 0
a1=0,有
H
1
a
1
=
α
1
e
1
H_1 a_1 = \alpha_1 e_1
H1a1=α1e1
H
1
A
n
=
[
α
1
∗
0
B
n
−
1
]
H_1 A_n = \begin{bmatrix} \alpha_1 & * \\ 0 & B_{n-1} \end{bmatrix}
H1An=[α10∗Bn−1]
如果
a
1
=
0
a_1 = 0
a1=0,则不需要进行householder变换,也可以说
H
1
=
I
H_1=I
H1=I
B
n
−
1
=
[
b
1
,
.
.
.
,
b
n
−
1
]
B_{n-1} = [b_1, ..., b_{n-1}]
Bn−1=[b1,...,bn−1]
如果
b
1
≠
0
b_1 \neq 0
b1=0,有
H
2
~
b
1
=
α
2
e
1
~
\tilde{H_2} b_1 = \alpha_2 \tilde{e_1}
H2~b1=α2e1~
H
2
~
B
n
−
1
=
[
α
2
∗
0
C
n
−
2
]
\tilde{H_2} B_{n-1} = \begin{bmatrix} \alpha_2 & * \\ 0 & C_{n-2} \end{bmatrix}
H2~Bn−1=[α20∗Cn−2]
如果
b
1
=
0
b_1 = 0
b1=0,则
H
2
~
=
I
\tilde{H_2}=I
H2~=I
H
2
=
[
1
0
T
0
H
2
~
]
H_2 = \begin{bmatrix} 1 & 0^T \\ 0 & \tilde{H_2} \end{bmatrix}
H2=[100TH2~]
也就是说对于n维矩阵,如果其列向量都不为0,用 n − 1 n-1 n−1个Householder矩阵就能实现正交分解
Householder变换的另一个作用便是实现 A A A的正交化,将 A A A变为 Q Q Q
例题
例1:通过Householder变换,使得 x = [ − 3 , 0 , 0 , 4 ] T x = [-3,0,0,4]^T x=[−3,0,0,4]T与 e 1 e_1 e1同向
解1:
α
=
∣
∣
x
∣
∣
=
5
\alpha = ||x|| = 5
α=∣∣x∣∣=5
w
=
x
−
5
e
1
∣
∣
x
−
5
e
1
∣
∣
=
1
80
[
−
8
,
0
,
0
,
4
]
T
w = \frac{x- 5e_1}{||x - 5 e_1||} = \frac{1}{\sqrt{80}}[-8,0,0,4]^T
w=∣∣x−5e1∣∣x−5e1=801[−8,0,0,4]T
H
x
=
x
−
2
w
w
T
x
=
5
e
1
Hx = x - 2ww^Tx = 5e_1
Hx=x−2wwTx=5e1
解2:
α
=
∣
∣
x
∣
∣
=
−
5
\alpha = ||x|| = -5
α=∣∣x∣∣=−5
w
=
x
+
5
e
1
∣
∣
x
+
5
e
1
∣
∣
=
1
20
[
2
,
0
,
0
,
4
]
T
w = \frac{x + 5e_1}{||x + 5 e_1||} = \frac{1}{\sqrt{20}}[2,0,0,4]^T
w=∣∣x+5e1∣∣x+5e1=201[2,0,0,4]T
H
x
=
x
−
2
w
w
T
x
=
−
5
e
1
Hx = x - 2ww^Tx = -5e_1
Hx=x−2wwTx=−5e1
例2:通过Householder变换,使得 A = [ 0 3 1 0 4 − 2 2 1 2 ] A = \begin{bmatrix} 0 & 3 & 1 \\ 0 & 4 & -2 \\ 2 & 1 & 2 \end{bmatrix} A=⎣⎡0023411−22⎦⎤实现正交分解
解:
a
1
=
[
0
0
2
]
a_1 = \begin{bmatrix} 0 \\ 0 \\2 \end{bmatrix}
a1=⎣⎡002⎦⎤,
α
1
=
∣
∣
a
1
∣
∣
=
2
\alpha_1 = ||a_1|| = 2
α1=∣∣a1∣∣=2,
w
=
a
1
−
2
e
1
∣
∣
a
1
−
2
e
1
∣
∣
=
1
2
[
−
1
0
1
]
w = \frac{a_1 - 2e_1}{||a_1 - 2 e_1||} = \frac{1}{\sqrt{2}}\begin{bmatrix} -1 \\ 0 \\1 \end{bmatrix}
w=∣∣a1−2e1∣∣a1−2e1=21⎣⎡−101⎦⎤
H
1
A
=
[
2
1
2
0
4
−
2
0
3
1
]
H_1 A = \begin{bmatrix} 2 & 1 & 2 \\ 0 & 4 & -2 \\ 0 & 3 & 1 \end{bmatrix}
H1A=⎣⎡2001432−21⎦⎤
b
1
=
[
4
3
]
b_1 = \begin{bmatrix} 4 \\ 3 \end{bmatrix}
b1=[43],
α
2
=
∣
∣
b
1
∣
∣
=
5
\alpha_2 = ||b_1|| = 5
α2=∣∣b1∣∣=5,
w
=
b
1
−
5
e
~
1
∣
∣
b
1
−
5
e
~
1
∣
∣
=
1
10
[
−
1
3
]
w = \frac{b_1 - 5 \tilde e_1}{||b_1 - 5 \tilde e_1||} = \frac{1}{\sqrt{10}}\begin{bmatrix} -1 \\ 3 \end{bmatrix}
w=∣∣b1−5e~1∣∣b1−5e~1=101[−13]
H
2
H
1
A
=
[
2
1
2
0
5
−
1
0
0
−
2
]
H_2H_1 A = \begin{bmatrix} 2 & 1 & 2 \\ 0 & 5 & -1 \\ 0 & 0 & -2 \end{bmatrix}
H2H1A=⎣⎡2001502−1−2⎦⎤
A = Q [ 2 1 2 0 5 − 1 0 0 − 2 ] , Q = H 1 H 2 A = Q \begin{bmatrix} 2 & 1 & 2 \\ 0 & 5 & -1 \\ 0 & 0 & -2 \end{bmatrix}, Q = H_1H_2 A=Q⎣⎡2001502−1−2⎦⎤,Q=H1H2
补充知识—正交补
https://zh.wikipedia.org/wiki/正交补
W
W
W为内积空间
V
V
V的子空间,则
W
W
W的正交补
W
⊥
W^{\perp}
W⊥定义为:
W
⊥
=
{
x
∈
V
:
∀
y
∈
W
,
<
x
,
y
>
=
0
}
W^{\perp} = \{ x \in V: \forall y \in W, <x,y>=0 \}
W⊥={x∈V:∀y∈W,<x,y>=0}
注:
W
⊥
W^{\perp}
W⊥也是
V
V
V的子空间
W
⊥
⊥
=
W
ˉ
W^{\perp \perp} = \bar{W}
W⊥⊥=Wˉ
疑问:线性空间
W
W
W与线性空间的闭包
W
ˉ
\bar{W}
Wˉ有什么区别
A
A
A为一个矩阵,
R
o
w
A
,
C
o
l
A
,
N
u
l
A
RowA, ColA, NulA
RowA,ColA,NulA分别为其行空间,列空间,零空间
(
R
o
w
A
)
⊥
=
N
u
l
A
(RowA)^{\perp} = NulA
(RowA)⊥=NulA
(
C
o
l
A
)
⊥
=
N
u
l
A
T
(ColA)^{\perp} = NulA^T
(ColA)⊥=NulAT
注:
A
A
A的右零空间是由
A
x
=
0
Ax=0
Ax=0的解张成的空间
A
A
A的左零空间是由
A
T
x
=
0
A^Tx=0
ATx=0的解张成的空间
W
W
W为巴拿赫空间
V
V
V的子空间,则
W
W
W的正交补
W
⊥
W^{\perp}
W⊥定义为:
W
⊥
=
{
x
∈
V
∗
:
∀
y
∈
W
,
x
(
y
)
=
0
}
W^{\perp} = \{ x \in V^*: \forall y \in W, x(y)=0 \}
W⊥={x∈V∗:∀y∈W,x(y)=0}
注:
W
⊥
W^{\perp}
W⊥是
V
V
V的对偶空间的子空间
对偶空间是行向量与列向量关系的抽象化
疑问:
V
∗
∗
V^{**}
V∗∗与
V
V
V的关系?
设 V V V为 在域 F F F上的向量空间,定义其对偶空间 V ∗ V^∗ V∗ 为由 V → F V \rightarrow F V→F 的所有线性函数的集合
V V V的元素称为反变或逆变contravariant向量; V ∗ V^* V∗的元素称为共变或协变covariant向量,也称余向量或同向量co-vectors,一形或线性型one-form