線性代數 中,廣義奇異值分解 (GSVD)是基於奇異值 (SVD)的兩種不同算法的統稱。其區別在於,一個是分解兩個矩陣(類似於高階或張量SVD ),另一種使用施加於單矩陣SVD奇異向量上的約束。
版本1:雙矩陣分解
廣義奇異值分解 (GSVD)是對矩陣對的矩陣分解 ,將奇異值分解 推廣到兩個矩陣的情形。它由Van Loan [ 1] 於1976年提出,後來由Paige與Saunders完善,[ 2] 也就是本節描述的版本。與SVD相對,GSVD可以同時分解具有相同列數的矩陣對。SVD、GSVD及SVD的其他一些推廣[ 3] [ 4] [ 5] 被廣泛用於研究線性系統在二次半範數 方面的條件調節 與正則化 。下面設
F
=
R
{\displaystyle \mathbb {F} =\mathbb {R} }
,或
F
=
C
{\displaystyle \mathbb {F} =\mathbb {C} }
。
定義
A
1
∈
F
m
1
×
n
{\displaystyle A_{1}\in \mathbb {F} ^{m_{1}\times n}}
與
A
2
∈
F
m
2
×
n
{\displaystyle A_{2}\in \mathbb {F} ^{m_{2}\times n}}
的廣義奇異值分解 為
A
1
=
U
1
Σ
1
[
W
∗
D
,
0
D
]
Q
∗
,
A
2
=
U
2
Σ
2
[
W
∗
D
,
0
D
]
Q
∗
,
{\displaystyle {\begin{aligned}A_{1}&=U_{1}\Sigma _{1}[W^{*}D,0_{D}]Q^{*},\\A_{2}&=U_{2}\Sigma _{2}[W^{*}D,0_{D}]Q^{*},\end{aligned}}}
,其中
U
1
∈
F
m
1
×
m
1
{\displaystyle U_{1}\in \mathbb {F} ^{m_{1}\times m_{1}}}
為酉矩陣 ;
U
2
∈
F
m
2
×
m
2
{\displaystyle U_{2}\in \mathbb {F} ^{m_{2}\times m_{2}}}
為酉矩陣;
Q
∈
F
n
×
n
{\displaystyle Q\in \mathbb {F} ^{n\times n}}
為酉矩陣;
W
∈
F
k
×
k
{\displaystyle W\in \mathbb {F} ^{k\times k}}
為酉矩陣;
D
∈
R
k
×
k
{\displaystyle D\in \mathbb {R} ^{k\times k}}
對角線元素為正實數,包含
C
=
[
A
1
A
2
]
{\displaystyle C={\begin{bmatrix}A_{1}\\A_{2}\end{bmatrix}}}
的非零奇異值的降序排列,
0
D
=
0
∈
R
k
×
(
n
−
k
)
{\displaystyle 0_{D}=0\in \mathbb {R} ^{k\times (n-k)}}
,
Σ
1
=
⌈
I
A
,
S
1
,
0
A
⌋
∈
R
m
1
×
k
{\displaystyle \Sigma _{1}=\lceil I_{A},S_{1},0_{A}\rfloor \in \mathbb {R} ^{m_{1}\times k}}
是非負實數分塊對角陣 ,其中
S
1
=
⌈
α
r
+
1
,
…
,
α
r
+
s
⌋
{\displaystyle S_{1}=\lceil \alpha _{r+1},\dots ,\alpha _{r+s}\rfloor }
,其中
1
>
α
r
+
1
≥
⋯
≥
α
r
+
s
>
0
{\displaystyle 1>\alpha _{r+1}\geq \cdots \geq \alpha _{r+s}>0}
,
I
A
=
I
r
{\displaystyle I_{A}=I_{r}}
,且
0
A
=
0
∈
R
(
m
1
−
r
−
s
)
×
(
k
−
r
−
s
)
{\displaystyle 0_{A}=0\in \mathbb {R} ^{(m_{1}-r-s)\times (k-r-s)}}
;
Σ
2
=
⌈
0
B
,
S
2
,
I
B
⌋
∈
R
m
2
×
k
{\displaystyle \Sigma _{2}=\lceil 0_{B},S_{2},I_{B}\rfloor \in \mathbb {R} ^{m_{2}\times k}}
是非負實數分塊對角陣,其中
S
2
=
⌈
β
r
+
1
,
…
,
β
r
+
s
⌋
{\displaystyle S_{2}=\lceil \beta _{r+1},\dots ,\beta _{r+s}\rfloor }
,其中
0
<
β
r
+
1
≤
⋯
≤
β
r
+
s
<
1
{\displaystyle 0<\beta _{r+1}\leq \cdots \leq \beta _{r+s}<1}
,
I
B
=
I
k
−
r
−
s
{\displaystyle I_{B}=I_{k-r-s}}
,且
0
B
=
0
∈
R
(
m
2
−
k
+
r
)
×
r
{\displaystyle 0_{B}=0\in \mathbb {R} ^{(m_{2}-k+r)\times r}}
;
Σ
1
∗
Σ
1
=
⌈
α
1
2
,
…
,
α
k
2
⌋
{\displaystyle \Sigma _{1}^{*}\Sigma _{1}=\lceil \alpha _{1}^{2},\dots ,\alpha _{k}^{2}\rfloor }
,
Σ
2
∗
Σ
2
=
⌈
β
1
2
,
…
,
β
k
2
⌋
{\displaystyle \Sigma _{2}^{*}\Sigma _{2}=\lceil \beta _{1}^{2},\dots ,\beta _{k}^{2}\rfloor }
,
Σ
1
∗
Σ
1
+
Σ
2
∗
Σ
2
=
I
k
{\displaystyle \Sigma _{1}^{*}\Sigma _{1}+\Sigma _{2}^{*}\Sigma _{2}=I_{k}}
,
k
=
rank
(
C
)
{\displaystyle k={\textrm {rank}}(C)}
.
記
α
1
=
⋯
=
α
r
=
1
,
α
r
+
s
+
1
=
⋯
=
α
k
=
0
,
β
1
=
⋯
=
β
r
=
0
,
β
r
+
s
+
1
=
⋯
=
β
k
=
1
{\displaystyle \alpha _{1}=\cdots =\alpha _{r}=1,\ \alpha _{r+s+1}=\cdots =\alpha _{k}=0,\ \beta _{1}=\cdots =\beta _{r}=0,\ \beta _{r+s+1}=\cdots =\beta _{k}=1}
。而
Σ
1
{\displaystyle \Sigma _{1}}
是對角陣,
Σ
2
{\displaystyle \Sigma _{2}}
不總是對角陣,因為前導矩形零矩陣;相反,
Σ
2
{\displaystyle \Sigma _{2}}
是「副對角陣」。
變體
GSVD有許多變體,與這樣一個事實有關:
Q
∗
{\displaystyle Q^{*}}
總可以左乘
E
E
∗
=
I
<
(
E
∈
F
n
×
n
)
{\displaystyle EE^{*}=I<(E\in \mathbb {F} ^{n\times n})}
是任意酉矩陣。記
X
=
(
[
W
∗
D
,
0
D
]
Q
∗
)
∗
{\displaystyle X=([W^{*}D,0_{D}]Q^{*})^{*}}
X
∗
=
[
0
,
R
]
Q
^
∗
{\displaystyle X^{*}=[0,R]{\hat {Q}}^{*}}
,其中
R
∈
F
k
×
k
{\displaystyle R\in \mathbb {F} ^{k\times k}}
是上三角可逆陣;
Q
^
∈
F
n
×
n
{\displaystyle {\hat {Q}}\in \mathbb {F} ^{n\times n}}
是酉矩陣。QR分解 總可以得到這樣的矩陣。
Y
=
W
∗
D
{\displaystyle Y=W^{*}D}
,那麼
Y
{\displaystyle Y}
可逆。
下面是GSVD的一些變體:
MATLAB (gsvd):
A
1
=
U
1
Σ
1
X
∗
,
A
2
=
U
2
Σ
2
X
∗
.
{\displaystyle {\begin{aligned}A_{1}&=U_{1}\Sigma _{1}X^{*},\\A_{2}&=U_{2}\Sigma _{2}X^{*}.\end{aligned}}}
LAPACK (LA_GGSVD):
A
1
=
U
1
Σ
1
[
0
,
R
]
Q
^
∗
,
A
2
=
U
2
Σ
2
[
0
,
R
]
Q
^
∗
.
{\displaystyle {\begin{aligned}A_{1}&=U_{1}\Sigma _{1}[0,R]{\hat {Q}}^{*},\\A_{2}&=U_{2}\Sigma _{2}[0,R]{\hat {Q}}^{*}.\end{aligned}}}
簡化:
A
1
=
U
1
Σ
1
[
Y
,
0
D
]
Q
∗
,
A
2
=
U
2
Σ
2
[
Y
,
0
D
]
Q
∗
.
{\displaystyle {\begin{aligned}A_{1}&=U_{1}\Sigma _{1}[Y,0_{D}]Q^{*},\\A_{2}&=U_{2}\Sigma _{2}[Y,0_{D}]Q^{*}.\end{aligned}}}
廣義奇異值
A
1
{\displaystyle A_{1}}
與
A
2
{\displaystyle A_{2}}
的廣義奇異值 是一對
(
a
,
b
)
∈
R
2
{\displaystyle (a,b)\in \mathbb {R} ^{2}}
使得
lim
δ
→
0
det
(
b
2
A
1
∗
A
1
−
a
2
A
2
∗
A
2
+
δ
I
n
)
/
det
(
δ
I
n
−
k
)
=
0
,
a
2
+
b
2
=
1
,
a
,
b
≥
0.
{\displaystyle {\begin{aligned}\lim _{\delta \to 0}\det(b^{2}A_{1}^{*}A_{1}-a^{2}A_{2}^{*}A_{2}+\delta I_{n})/\det(\delta I_{n-k})&=0,\\a^{2}+b^{2}&=1,\\a,b&\geq 0.\end{aligned}}}
我們有
A
i
A
j
∗
=
U
i
Σ
i
Y
Y
∗
Σ
j
∗
U
j
∗
{\displaystyle A_{i}A_{j}^{*}=U_{i}\Sigma _{i}YY^{*}\Sigma _{j}^{*}U_{j}^{*}}
A
i
∗
A
j
=
Q
[
Y
∗
Σ
i
∗
Σ
j
Y
0
0
0
]
Q
∗
=
Q
1
Y
∗
Σ
i
∗
Σ
j
Y
Q
1
∗
{\displaystyle A_{i}^{*}A_{j}=Q{\begin{bmatrix}Y^{*}\Sigma _{i}^{*}\Sigma _{j}Y&0\\0&0\end{bmatrix}}Q^{*}=Q_{1}Y^{*}\Sigma _{i}^{*}\Sigma _{j}YQ_{1}^{*}}
根據這些性質,可以證明廣義奇異值正是成對的
(
α
i
,
β
i
)
{\displaystyle (\alpha _{i},\beta _{i})}
。有
det
(
b
2
A
1
∗
A
1
−
a
2
A
2
∗
A
2
+
δ
I
n
)
=
det
(
b
2
A
1
∗
A
1
−
a
2
A
2
∗
A
2
+
δ
Q
Q
∗
)
=
det
(
Q
[
Y
∗
(
b
2
Σ
1
∗
Σ
1
−
a
2
Σ
2
∗
Σ
2
)
Y
+
δ
I
k
0
0
δ
I
n
−
k
]
Q
∗
)
=
det
(
δ
I
n
−
k
)
det
(
Y
∗
(
b
2
Σ
1
∗
Σ
1
−
a
2
Σ
2
∗
Σ
2
)
Y
+
δ
I
k
)
.
{\displaystyle {\begin{aligned}&\det(b^{2}A_{1}^{*}A_{1}-a^{2}A_{2}^{*}A_{2}+\delta I_{n})\\=&\det(b^{2}A_{1}^{*}A_{1}-a^{2}A_{2}^{*}A_{2}+\delta QQ^{*})\\=&\det \left(Q{\begin{bmatrix}Y^{*}(b^{2}\Sigma _{1}^{*}\Sigma _{1}-a^{2}\Sigma _{2}^{*}\Sigma _{2})Y+\delta I_{k}&0\\0&\delta I_{n-k}\end{bmatrix}}Q^{*}\right)\\=&\det(\delta I_{n-k})\det(Y^{*}(b^{2}\Sigma _{1}^{*}\Sigma _{1}-a^{2}\Sigma _{2}^{*}\Sigma _{2})Y+\delta I_{k}).\end{aligned}}}
因此
lim
δ
→
0
det
(
b
2
A
1
∗
A
1
−
a
2
A
2
∗
A
2
+
δ
I
n
)
/
det
(
δ
I
n
−
k
)
=
lim
δ
→
0
det
(
Y
∗
(
b
2
Σ
1
∗
Σ
1
−
a
2
Σ
2
∗
Σ
2
)
Y
+
δ
I
k
)
=
det
(
Y
∗
(
b
2
Σ
1
∗
Σ
1
−
a
2
Σ
2
∗
Σ
2
)
Y
)
=
|
det
(
Y
)
|
2
∏
i
=
1
k
(
b
2
α
i
2
−
a
2
β
i
2
)
.
{\displaystyle {\begin{aligned}{}&\lim _{\delta \to 0}\det(b^{2}A_{1}^{*}A_{1}-a^{2}A_{2}^{*}A_{2}+\delta I_{n})/\det(\delta I_{n-k})\\=&\lim _{\delta \to 0}\det(Y^{*}(b^{2}\Sigma _{1}^{*}\Sigma _{1}-a^{2}\Sigma _{2}^{*}\Sigma _{2})Y+\delta I_{k})\\=&\det(Y^{*}(b^{2}\Sigma _{1}^{*}\Sigma _{1}-a^{2}\Sigma _{2}^{*}\Sigma _{2})Y)\\=&|\det(Y)|^{2}\prod _{i=1}^{k}(b^{2}\alpha _{i}^{2}-a^{2}\beta _{i}^{2}).\end{aligned}}}
對某個
i
{\displaystyle i}
,當
a
=
α
i
,
b
=
β
i
{\displaystyle a=\alpha _{i},\ b=\beta _{i}}
時,表達式恰為零。
在[ 2] 中,廣義奇異值被認為是求解
det
(
b
2
A
1
∗
A
1
−
a
2
A
2
∗
A
2
)
=
0
{\displaystyle \det(b^{2}A_{1}^{*}A_{1}-a^{2}A_{2}^{*}A_{2})=0}
的奇異值。然而,這只有當
k
=
n
{\displaystyle k=n}
時才成立,否則行列式對每對
(
a
,
b
)
∈
R
2
{\displaystyle (a,b)\in \mathbb {R} ^{2}}
都將是0;這可通過替換上面的
δ
=
0
{\displaystyle \delta =0}
得到。
廣義逆
對任意可逆陣
E
∈
F
n
×
n
{\displaystyle E\in \mathbb {F} ^{n\times n}}
,令
E
+
=
E
−
1
{\displaystyle E^{+}=E^{-1}}
,對任意零矩陣
0
∈
F
m
×
n
{\displaystyle 0\in \mathbb {F} ^{m\times n}}
,令
0
+
=
0
∗
{\displaystyle 0^{+}=0^{*}}
,對任意分塊對角陣令
⌈
E
1
,
E
2
⌋
+
=
⌈
E
1
+
,
E
2
+
⌋
{\displaystyle \left\lceil E_{1},E_{2}\right\rfloor ^{+}=\left\lceil E_{1}^{+},E_{2}^{+}\right\rfloor }
。定義
A
i
+
=
Q
[
Y
−
1
0
]
Σ
i
+
U
i
∗
{\displaystyle A_{i}^{+}=Q{\begin{bmatrix}Y^{-1}\\0\end{bmatrix}}\Sigma _{i}^{+}U_{i}^{*}}
可以證明這裡定義的
A
i
+
{\displaystyle A_{i}^{+}}
是
A
i
{\displaystyle A_{i}}
的廣義逆陣 ;特別是
A
i
{\displaystyle A_{i}}
的
{
1
,
2
,
3
}
{\displaystyle \{1,2,3\}}
逆。由於它一般不滿足
(
A
i
+
A
i
)
∗
=
A
i
+
A
i
{\displaystyle (A_{i}^{+}A_{i})^{*}=A_{i}^{+}A_{i}}
,所以不是摩爾-彭若斯廣義逆 ;否則可以得出,對任意所選矩陣都有
(
A
B
)
+
=
B
+
A
+
{\displaystyle (AB)^{+}=B^{+}A^{+}}
,這隻對特定類型的矩陣成立。
設
Q
=
[
Q
1
Q
2
]
{\displaystyle Q={\begin{bmatrix}Q_{1}&Q_{2}\end{bmatrix}}}
,其中
Q
1
∈
F
n
×
k
,
Q
2
∈
F
n
×
(
n
−
k
)
{\displaystyle Q_{1}\in \mathbb {F} ^{n\times k},\ Q_{2}\in \mathbb {F} ^{n\times (n-k)}}
。這個廣義逆具有如下性質:
Σ
1
+
=
⌈
I
A
,
S
1
−
1
,
0
A
T
⌋
{\displaystyle \Sigma _{1}^{+}=\lceil I_{A},S_{1}^{-1},0_{A}^{T}\rfloor }
Σ
2
+
=
⌈
0
B
T
,
S
2
−
1
,
I
B
⌋
{\displaystyle \Sigma _{2}^{+}=\lceil 0_{B}^{T},S_{2}^{-1},I_{B}\rfloor }
Σ
1
Σ
1
+
=
⌈
I
,
I
,
0
⌋
{\displaystyle \Sigma _{1}\Sigma _{1}^{+}=\lceil I,I,0\rfloor }
Σ
2
Σ
2
+
=
⌈
0
,
I
,
I
⌋
{\displaystyle \Sigma _{2}\Sigma _{2}^{+}=\lceil 0,I,I\rfloor }
Σ
1
Σ
2
+
=
⌈
0
,
S
1
S
2
−
1
,
0
⌋
{\displaystyle \Sigma _{1}\Sigma _{2}^{+}=\lceil 0,S_{1}S_{2}^{-1},0\rfloor }
Σ
1
+
Σ
2
=
⌈
0
,
S
1
−
1
S
2
,
0
⌋
{\displaystyle \Sigma _{1}^{+}\Sigma _{2}=\lceil 0,S_{1}^{-1}S_{2},0\rfloor }
A
i
A
j
+
=
U
i
Σ
i
Σ
j
+
U
j
∗
{\displaystyle A_{i}A_{j}^{+}=U_{i}\Sigma _{i}\Sigma _{j}^{+}U_{j}^{*}}
A
i
+
A
j
=
Q
[
Y
−
1
Σ
i
+
Σ
j
Y
0
0
0
]
Q
∗
=
Q
1
Y
−
1
Σ
i
+
Σ
j
Y
Q
1
∗
{\displaystyle A_{i}^{+}A_{j}=Q{\begin{bmatrix}Y^{-1}\Sigma _{i}^{+}\Sigma _{j}Y&0\\0&0\end{bmatrix}}Q^{*}=Q_{1}Y^{-1}\Sigma _{i}^{+}\Sigma _{j}YQ_{1}^{*}}
商SVD
'
A
1
{\displaystyle A_{1}}
與
A
2
{\displaystyle A_{2}}
的'廣義奇異比是
σ
i
=
α
i
β
i
+
{\displaystyle \sigma _{i}=\alpha _{i}\beta _{i}^{+}}
。由以上性質,
A
1
A
2
+
=
U
1
Σ
1
Σ
2
+
U
2
∗
{\displaystyle A_{1}A_{2}^{+}=U_{1}\Sigma _{1}\Sigma _{2}^{+}U_{2}^{*}}
。注意
Σ
1
Σ
2
+
=
⌈
0
,
S
1
S
2
−
1
,
0
⌋
{\displaystyle \Sigma _{1}\Sigma _{2}^{+}=\lceil 0,S_{1}S_{2}^{-1},0\rfloor }
是對角陣,忽略前導零矩陣,按降序包含着奇異比。若
A
2
{\displaystyle A_{2}}
可逆,則
Σ
1
Σ
2
+
{\displaystyle \Sigma _{1}\Sigma _{2}^{+}}
沒有前導零,廣義奇異比就是奇異值,
U
1
{\displaystyle U_{1}}
與
U
2
{\displaystyle U_{2}}
則是
A
1
A
2
+
=
A
1
A
2
−
1
{\displaystyle A_{1}A_{2}^{+}=A_{1}A_{2}^{-1}}
的奇異向量矩陣。事實上計算
A
1
A
2
−
1
{\displaystyle A_{1}A_{2}^{-1}}
的SVD是GSVD的動機之一,因為「形成
A
B
−
1
{\displaystyle AB^{-1}}
並求SVD,當
B
{\displaystyle B}
的方程解條件不佳時,可能產生不必要、較大的數值誤差」。[ 2] 因此有時也被稱為「商GSVD」,雖然這並不是使用GSVD的唯一原因。若
A
2
{\displaystyle A_{2}}
不可逆,並放寬奇異值降序排列的要求,則
U
1
Σ
1
Σ
2
+
U
2
∗
{\displaystyle U_{1}\Sigma _{1}\Sigma _{2}^{+}U_{2}^{*}}
仍是
A
1
A
2
+
{\displaystyle A_{1}A_{2}^{+}}
的SVD。或者,把前導零移到後面,也可以找到降序SVD:
U
1
Σ
1
Σ
2
+
U
2
∗
=
(
U
1
P
1
)
P
1
∗
Σ
1
Σ
2
+
P
2
(
P
2
∗
U
2
∗
)
{\displaystyle U_{1}\Sigma _{1}\Sigma _{2}^{+}U_{2}^{*}=(U_{1}P_{1})P_{1}^{*}\Sigma _{1}\Sigma _{2}^{+}P_{2}(P_{2}^{*}U_{2}^{*})}
,其中
P
1
{\displaystyle P_{1}}
與
P
2
{\displaystyle P_{2}}
是適當的置換矩陣。由於秩等於非零奇異值的個數,所以
r
a
n
k
(
A
1
A
2
+
)
=
s
{\displaystyle \mathrm {rank} (A_{1}A_{2}^{+})=s}
。
構造
令
C
=
P
⌈
D
,
0
⌋
Q
∗
{\displaystyle C=P\lceil D,0\rfloor Q^{*}}
為
C
=
[
A
1
A
2
]
{\displaystyle C={\begin{bmatrix}A_{1}\\A_{2}\end{bmatrix}}}
的SVD,其中
P
∈
F
(
m
1
+
m
2
)
×
(
m
1
×
m
2
)
{\displaystyle P\in \mathbb {F} ^{(m_{1}+m_{2})\times (m_{1}\times m_{2})}}
是酉矩陣,
Q
{\displaystyle Q}
與
D
{\displaystyle D}
如上所述;
P
=
[
P
1
,
P
2
]
{\displaystyle P=[P_{1},P_{2}]}
,其中
P
1
∈
F
(
m
1
+
m
2
)
×
k
{\displaystyle P_{1}\in \mathbb {F} ^{(m_{1}+m_{2})\times k}}
與
P
2
∈
F
(
m
1
+
m
2
)
×
(
n
−
k
)
{\displaystyle P_{2}\in \mathbb {F} ^{(m_{1}+m_{2})\times (n-k)}}
;
P
1
=
[
P
11
P
21
]
{\displaystyle P_{1}={\begin{bmatrix}P_{11}\\P_{21}\end{bmatrix}}}
,其中
P
11
∈
F
m
1
×
k
{\displaystyle P_{11}\in \mathbb {F} ^{m_{1}\times k}}
與
P
21
∈
F
m
2
×
k
{\displaystyle P_{21}\in \mathbb {F} ^{m_{2}\times k}}
;
P
11
=
U
1
Σ
1
W
∗
{\displaystyle P_{11}=U_{1}\Sigma _{1}W^{*}}
通過
P
11
{\displaystyle P_{11}}
的SVD得到,其中
U
1
{\displaystyle U_{1}}
、
Σ
1
{\displaystyle \Sigma _{1}}
與
W
{\displaystyle W}
如上所述,
P
21
W
=
U
2
Σ
2
{\displaystyle P_{21}W=U_{2}\Sigma _{2}}
經過類似於QR分解 的分解,其中
U
2
{\displaystyle U_{2}}
與
Σ
2
{\displaystyle \Sigma _{2}}
如上所述。
那麼,
C
=
P
⌈
D
,
0
⌋
Q
∗
=
[
P
1
D
,
0
]
Q
∗
=
[
U
1
Σ
1
W
∗
D
0
U
2
Σ
2
W
∗
D
0
]
Q
∗
=
[
U
1
Σ
1
[
W
∗
D
,
0
]
Q
∗
U
2
Σ
2
[
W
∗
D
,
0
]
Q
∗
]
.
{\displaystyle {\begin{aligned}C&=P\lceil D,0\rfloor Q^{*}\\{}&=[P_{1}D,0]Q^{*}\\{}&={\begin{bmatrix}U_{1}\Sigma _{1}W^{*}D&0\\U_{2}\Sigma _{2}W^{*}D&0\end{bmatrix}}Q^{*}\\{}&={\begin{bmatrix}U_{1}\Sigma _{1}[W^{*}D,0]Q^{*}\\U_{2}\Sigma _{2}[W^{*}D,0]Q^{*}\end{bmatrix}}.\end{aligned}}}
還有
[
U
1
∗
0
0
U
2
∗
]
P
1
W
=
[
Σ
1
Σ
2
]
.
{\displaystyle {\begin{bmatrix}U_{1}^{*}&0\\0&U_{2}^{*}\end{bmatrix}}P_{1}W={\begin{bmatrix}\Sigma _{1}\\\Sigma _{2}\end{bmatrix}}.}
因此
Σ
1
∗
Σ
1
+
Σ
2
∗
Σ
2
=
[
Σ
1
Σ
2
]
∗
[
Σ
1
Σ
2
]
=
W
∗
P
1
∗
[
U
1
0
0
U
2
]
[
U
1
∗
0
0
U
2
∗
]
P
1
W
=
I
.
{\displaystyle \Sigma _{1}^{*}\Sigma _{1}+\Sigma _{2}^{*}\Sigma _{2}={\begin{bmatrix}\Sigma _{1}\\\Sigma _{2}\end{bmatrix}}^{*}{\begin{bmatrix}\Sigma _{1}\\\Sigma _{2}\end{bmatrix}}=W^{*}P_{1}^{*}{\begin{bmatrix}U_{1}&0\\0&U_{2}\end{bmatrix}}{\begin{bmatrix}U_{1}^{*}&0\\0&U_{2}^{*}\end{bmatrix}}P_{1}W=I.}
由於
P
1
{\displaystyle P_{1}}
的列歸一正交,
|
|
P
1
|
|
2
≤
1
{\displaystyle ||P_{1}||_{2}\leq 1}
,因此
|
|
Σ
1
|
|
2
=
|
|
U
1
∗
P
1
W
|
|
2
=
|
|
P
1
|
|
2
≤
1.
{\displaystyle ||\Sigma _{1}||_{2}=||U_{1}^{*}P_{1}W||_{2}=||P_{1}||_{2}\leq 1.}
對每個
x
∈
R
k
{\displaystyle x\in \mathbb {R} ^{k}}
,有
|
|
x
|
|
2
=
1
{\displaystyle ||x||_{2}=1}
,使得
|
|
P
21
x
|
|
2
2
≤
|
|
P
11
x
|
|
2
2
+
|
|
P
21
x
|
|
2
2
=
|
|
P
1
x
|
|
2
2
≤
1.
{\displaystyle ||P_{21}x||_{2}^{2}\leq ||P_{11}x||_{2}^{2}+||P_{21}x||_{2}^{2}=||P_{1}x||_{2}^{2}\leq 1.}
因此
|
|
P
21
|
|
2
≤
1
{\displaystyle ||P_{21}||_{2}\leq 1}
;
|
|
Σ
2
|
|
2
=
|
|
U
2
∗
P
21
W
|
|
2
=
|
|
P
21
|
|
2
≤
1.
{\displaystyle ||\Sigma _{2}||_{2}=||U_{2}^{*}P_{21}W||_{2}=||P_{21}||_{2}\leq 1.}
應用
張量GSVD是比較譜分解的一種,是SVD在多張量上的推廣,提出動機是同時識別其中的相似與不相似數據,並從任何數量和維度的任意數據類型中得到單一相干模型。
GSVD是一種比較譜分解,[ 6] 已成功應用於信號處理和數據科學,如基因組信號處理。[ 7] [ 8] [ 9]
這些應用啟發了其他幾種比較譜分解,即高階GSVD(HO GSVD)[ 10] 與張量GSVD。[ 11] [ 12]
當特徵函數以線性模型(即再生核希爾伯特空間 )為參數時,它同樣適於估計線性運算的譜分解。[ 13]
版本2:加權單矩陣分解
廣義奇異值分解 (GSVD)的加權情形是一種有約束矩陣分解 ,約束施加在奇異向量上。[ 14] [ 15] [ 16] 這種GSVD 是SVD 的推廣。給定m×n 實或複數矩陣M 的SVD 分解
M
=
U
Σ
V
∗
{\displaystyle M=U\Sigma V^{*}\,}
,其中
U
∗
W
u
U
=
V
∗
W
v
V
=
I
.
{\displaystyle U^{*}W_{u}U=V^{*}W_{v}V=I.}
其中I 是單位矩陣 ;
U
{\displaystyle U}
與
V
{\displaystyle V}
在約束條件下(
W
u
{\displaystyle W_{u}}
;
W
v
{\displaystyle W_{v}}
)是標準正交矩陣。另外,
W
u
{\displaystyle W_{u}}
、
W
v
{\displaystyle W_{v}}
是正定矩陣(通常是權的對角矩陣)。這種形式的GSVD 是某些算法的核心,如廣義主成分分析和對應分析 。
加權形式的GSVD 之所以被稱為加權形式,是因為在正確取權時,可以推出許多算法(如多維標度 與線性判別分析 )。[ 17]
參考文獻
^ Van Loan CF. Generalizing the Singular Value Decomposition. SIAM J. Numer. Anal. 1976, 13 (1): 76–83. Bibcode:1976SJNA...13...76V . doi:10.1137/0713009 .
^ 2.0 2.1 2.2 Paige CC, Saunders MA. Towards a Generalized Singular Value Decomposition. SIAM J. Numer. Anal. 1981, 18 (3): 398–405. Bibcode:1981SJNA...18..398P . doi:10.1137/0718026 .
^ Hansen PC. Rank-Deficient and Discrete Ill-Posed Problems: Numerical Aspects of Linear Inversion. SIAM Monographs on Mathematical Modeling and Computation. 1997. ISBN 0-89871-403-6 .
^ de Moor BL, Golub GH. Generalized Singular Value Decompositions A Proposal for a Standard Nomenclauture (PDF) . 1989 [2023-09-25 ] . (原始內容存檔 (PDF) 於2023-07-23).
^ de Moor BL, Zha H. A tree of generalizations of the ordinary singular value decomposition . Linear Algebra and Its Applications. 1991, 147 : 469–500. doi:10.1016/0024-3795(91)90243-P .
^ Alter O, Brown PO, Botstein D. Generalized singular value decomposition for comparative analysis of genome-scale expression data sets of two different organisms . Proceedings of the National Academy of Sciences of the United States of America. March 2003, 100 (6): 3351–6. Bibcode:2003PNAS..100.3351A . PMC 152296 . PMID 12631705 . doi:10.1073/pnas.0530258100 .
^ Lee CH, Alpert BO, Sankaranarayanan P, Alter O. GSVD comparison of patient-matched normal and tumor aCGH profiles reveals global copy-number alterations predicting glioblastoma multiforme survival . PLOS ONE. January 2012, 7 (1): e30098. Bibcode:2012PLoSO...730098L . PMC 3264559 . PMID 22291905 . doi:10.1371/journal.pone.0030098 .
^ Aiello KA, Ponnapalli SP, Alter O. Mathematically universal and biologically consistent astrocytoma genotype encodes for transformation and predicts survival phenotype . APL Bioengineering. September 2018, 2 (3): 031909. PMC 6215493 . PMID 30397684 . doi:10.1063/1.5037882 .
^ Ponnapalli SP, Bradley MW, Devine K, Bowen J, Coppens SE, Leraas KM, Milash BA, Li F, Luo H, Qiu S, Wu K, Yang H, Wittwer CT, Palmer CA, Jensen RL, Gastier-Foster JM, Hanson HA, Barnholtz-Sloan JS , Alter O. Retrospective Clinical Trial Experimentally Validates Glioblastoma Genome-Wide Pattern of DNA Copy-Number Alterations Predictor of Survival . APL Bioengineering. May 2020, 4 (2): 026106. PMC 7229984 . PMID 32478280 . doi:10.1063/1.5142559 . Press Release .
^ Ponnapalli SP, Saunders MA, Van Loan CF, Alter O. A higher-order generalized singular value decomposition for comparison of global mRNA expression from multiple organisms . PLOS ONE. December 2011, 6 (12): e28072. Bibcode:2011PLoSO...628072P . PMC 3245232 . PMID 22216090 . doi:10.1371/journal.pone.0028072 .
^ Sankaranarayanan P, Schomay TE, Aiello KA, Alter O. Tensor GSVD of patient- and platform-matched tumor and normal DNA copy-number profiles uncovers chromosome arm-wide patterns of tumor-exclusive platform-consistent alterations encoding for cell transformation and predicting ovarian cancer survival . PLOS ONE. April 2015, 10 (4): e0121396. Bibcode:2015PLoSO..1021396S . PMC 4398562 . PMID 25875127 . doi:10.1371/journal.pone.0121396 .
^ Bradley MW, Aiello KA, Ponnapalli SP, Hanson HA, Alter O. GSVD- and tensor GSVD-uncovered patterns of DNA copy-number alterations predict adenocarcinomas survival in general and in response to platinum . APL Bioengineering. September 2019, 3 (3): 036104. PMC 6701977 . PMID 31463421 . doi:10.1063/1.5099268 . Supplementary Material .
^ Cabannes, Vivien; Pillaud-Vivien, Loucas; Bach, Francis; Rudi, Alessandro. Overcoming the curse of dimensionality with Laplacian regularization in semi-supervised learning. 2021. arXiv:2009.04324 [stat.ML ].
^ Jolliffe IT. Principal Component Analysis . Springer Series in Statistics 2nd. NY: Springer. 2002. ISBN 978-0-387-95442-4 .
^ Greenacre M. Theory and Applications of Correspondence Analysis. London: Academic Press. 1983. ISBN 978-0-12-299050-2 .
^ Abdi H, Williams LJ. Principal component analysis.. Wiley Interdisciplinary Reviews: Computational Statistics. 2010, 2 (4): 433–459. doi:10.1002/wics.101 .
^ Abdi H. Singular Value Decomposition (SVD) and Generalized Singular Value Decomposition (GSVD).. Salkind NJ (編). Encyclopedia of Measurement and Statistics. . Thousand Oaks (CA): Sage. 2007: 907 –912.
閱讀更多