1. 矩阵微分的常用法则
令
A
,
B
A,B
A,B为常数矩阵,并且
U
,
V
,
W
,
X
U,V,W,X
U,V,W,X为矩阵函数。下面汇总了矩阵微分的常用法则
(1)常数矩阵的微分矩阵为零矩阵,即
d
A
=
O
dA=O
dA=O
(2)常数
α
\alpha
α与矩阵函数
U
U
U的乘积的微分矩阵,即
d
(
α
U
)
=
α
d
U
d(\alpha U)=\alpha dU
d(αU)=αdU
(3)矩阵转置的微分矩阵等于原矩阵的微分矩阵的转置,即
d
(
U
T
)
=
(
d
U
)
T
d(U^T)=(dU)^T
d(UT)=(dU)T
(4)两个矩阵函数的和(差)的微分矩阵为
d
(
U
+
V
)
=
d
U
±
d
V
d(U+V)=dU±dV
d(U+V)=dU±dV
(5)常数矩阵与矩阵的数乘积的微分矩阵为
d
(
A
X
B
)
=
A
(
d
X
)
B
d(AXB)=A(dX)B
d(AXB)=A(dX)B
(6)矩阵函数乘积的微分矩阵为
d
(
U
V
)
=
(
d
U
)
V
+
U
(
d
V
)
d(UV)=(dU)V+U(dV)
d(UV)=(dU)V+U(dV)
d
(
U
V
W
)
=
(
d
U
)
V
W
+
U
(
d
V
)
W
+
U
V
(
d
W
)
d(UVW)=(dU)VW+U(dV)W+UV(dW)
d(UVW)=(dU)VW+U(dV)W+UV(dW)
特别地,若
A
A
A为常数矩阵,则
d
(
X
A
X
T
)
=
(
d
X
)
A
X
T
+
X
A
(
d
X
)
T
d(XAX^T)=(dX)AX^T+XA(dX)^T
d(XAXT)=(dX)AXT+XA(dX)T
d
(
X
T
A
X
)
=
(
d
X
)
T
A
X
+
X
T
A
d
X
d(X^TAX)=(dX)^TAX+X^TAdX
d(XTAX)=(dX)TAX+XTAdX
(7)矩阵函数的Kronecker积的微分矩阵为
d
(
U
⊗
V
)
=
(
d
U
)
⊗
V
+
U
⊗
d
V
d(U\otimes V)=(dU)\otimes V+U\otimes dV
d(U⊗V)=(dU)⊗V+U⊗dV
(8)矩阵函数的Hadamard积的微分矩阵,为
d
(
U
⊙
V
)
=
(
d
U
)
⊙
V
+
U
⊙
d
V
d(U\odot V)=(dU)\odot V+U\odot dV
d(U⊙V)=(dU)⊙V+U⊙dV
(9)向量化函数
v
e
c
(
U
)
vec(U)
vec(U)的微分矩阵等于
U
U
U的微分矩阵的向量化函数,即
d
(
v
e
c
(
U
)
)
=
v
e
c
(
d
U
)
d(vec(U))= vec(dU)
d(vec(U))=vec(dU)
(10)行列式的微分为
d
∣
X
∣
=
∣
X
∣
t
r
(
X
−
1
d
X
)
d|X|=|X|tr(X^{-1}dX)
d∣X∣=∣X∣tr(X−1dX)
(11)矩阵
U
U
U的迹的微分
d
(
t
r
U
)
d(trU)
d(trU)等于微分矩阵
d
U
dU
dU的迹
t
r
(
d
U
)
tr(dU)
tr(dU),即有
d
(
t
r
(
U
)
)
=
t
r
(
d
U
)
d(tr(U))=tr(dU)
d(tr(U))=tr(dU)
d
(
t
r
(
X
T
X
)
)
=
2
t
r
(
X
T
d
X
)
d(tr(X^TX))=2tr(X^TdX)
d(tr(XTX))=2tr(XTdX)
(12)逆矩阵的微分矩阵为
d
(
X
−
1
)
=
−
X
−
1
(
d
X
)
X
−
1
d(X^{-1})=-X^{-1}(dX)X^{-1}
d(X−1)=−X−1(dX)X−1
(13)Moore-Penrose逆矩阵的微分矩阵为
d
(
X
+
)
=
−
X
+
(
d
X
)
X
+
+
X
+
(
X
+
)
T
(
d
X
T
)
(
I
−
X
X
+
)
+
(
I
−
X
+
X
)
(
d
X
T
)
(
X
+
)
T
X
+
d(X^+)=-X^+(dX)X^++X^+(X^+)T(dX^T)(I-XX^+)+(I-X^+X)(dX^T)(X^+)^TX^+
d(X+)=−X+(dX)X++X+(X+)T(dXT)(I−XX+)+(I−X+X)(dXT)(X+)TX+
d
(
X
+
X
)
=
X
+
(
d
X
)
(
I
−
X
+
X
)
+
(
X
+
(
d
X
)
(
I
−
X
+
X
)
)
T
d(X^+X)=X^+(dX)(I-X^+X)+(X^+(dX)(I-X^+X))^T
d(X+X)=X+(dX)(I−X+X)+(X+(dX)(I−X+X))T
d
(
X
X
+
)
=
(
I
−
X
X
+
)
(
d
X
)
X
+
+
(
(
I
−
X
X
+
)
(
d
X
)
X
+
)
T
d(XX^+)=(I-XX^+)(dX)X^++((I-XX^+)(dX)X^+)^T
d(XX+)=(I−XX+)(dX)X++((I−XX+)(dX)X+)T
(14)矩阵对数的微分矩阵为
d
l
o
g
X
=
X
−
1
d
X
dlogX=X^{-1}dX
dlogX=X−1dX
d
l
o
g
(
X
T
A
X
)
=
X
T
A
X
−
1
[
(
d
X
)
T
A
X
+
X
T
A
d
X
]
dlog(X^TAX)=X^TAX^{-1}[(dX)^TAX+X^TAdX]
dlog(XTAX)=XTAX−1[(dX)TAX+XTAdX]
例5.3.10
考虑例5.1.7的CDMA系统中,仍然共有K个用户在通信,但每个用户的扩频波形向量变成复向量
s
k
(
t
)
s_k(t)
sk(t),接收信号向量
y
y
y也为复向量。此时,设计多用户检测器
M
M
M的目标函数变为
J
(
M
)
=
E
{
∣
∣
b
−
M
y
∣
∣
2
2
}
J(M)=E\{||b-My||_2^2\}
J(M)=E{∣∣b−My∣∣22}
=
t
r
(
c
o
v
(
b
−
M
y
)
)
=tr(cov(b-My))
=tr(cov(b−My))
=
t
r
(
I
)
+
t
r
(
M
(
R
A
2
R
+
σ
2
R
)
M
H
)
−
t
r
(
A
R
M
H
)
−
t
r
(
M
R
A
)
=tr(I)+tr(M(RA^2R+\sigma^2R)M^H)-tr(ARM^H)-tr(MRA)
=tr(I)+tr(M(RA2R+σ2R)MH)−tr(ARMH)−tr(MRA)
易求出
∂
J
(
M
)
∂
M
∗
=
M
(
R
A
2
R
+
σ
2
R
)
−
A
R
\frac{\partial J(M)}{\partial M^*}=M(RA^2R+\sigma^2R)-AR
∂M∗∂J(M)=M(RA2R+σ2R)−AR
令其等于零,并假定
R
R
R非奇异,立即有
M
=
A
(
R
A
2
+
σ
2
I
)
−
1
M=A(RA^2+\sigma^2I)^{-1}
M=A(RA2+σ2I)−1
这里的检测器为复矩阵。
2. 矩阵的奇异值分解
令
A
∈
R
m
×
n
A\in R_{m\times n}
A∈Rm×n(或
C
m
×
n
C_{m\times n}
Cm×n),则存在正交(或西)矩阵
U
∈
R
m
×
m
U\in R_{m\times m}
U∈Rm×m(或
C
m
×
m
C_{m\times m}
Cm×m)和
V
∈
R
n
×
n
V\in R_{n\times n}
V∈Rn×n(或
C
n
×
n
C_{n\times n}
Cn×n)使得
A
=
U
Σ
V
T
(
或
U
Σ
V
H
)
A=U\Sigma V^T(或U\Sigma V^H)
A=UΣVT(或UΣVH)
式中
Σ
=
[
Σ
1
O
O
O
]
\Sigma= \begin{bmatrix} \Sigma_1&O\\ O&O \end{bmatrix}
Σ=[Σ1OOO]
且
Σ
1
=
d
i
a
g
(
σ
1
,
σ
2
,
…
,
σ
r
)
\Sigma_1=diag(\sigma_1,\sigma_2,…,\sigma_r)
Σ1=diag(σ1,σ2,…,σr),其对角元素按照顺序
σ
1
≥
σ
2
≥
…
≥
σ
r
>
0
,
r
=
r
a
n
k
(
A
)
\sigma_1≥\sigma_2≥…≥\sigma_r > 0, r=rank(A)
σ1≥σ2≥…≥σr>0,r=rank(A)
排列。