本文对信号处理中一些常用的求导作以记录,如有表述不当之处欢迎批评指正。欢迎任何形式的转载,但请务必注明出处。
1. 实值函数相对于复变量的梯度
a. 定义 f ( w ) f(w) f(w) 是复变量 w w w 的实函数;
b. 其中 w w w 和 w ∗ w^{*} w∗ 分别表示为: w = x + j y w=x+jy w=x+jy w ∗ = x − j y w^{*}=x-jy w∗=x−jy
c. 那么: ∂ f ∂ w = 1 2 ( ∂ f ∂ x − j ∂ f ∂ y ) \frac{\partial{f}}{\partial{w}}=\frac{1}{2}(\frac{\partial{f}}{\partial{x}}-j\frac{\partial{f}}{\partial{y}}) ∂w∂f=21(∂x∂f−j∂y∂f) ∂ f ∂ w ∗ = 1 2 ( ∂ f ∂ x + j ∂ f ∂ y ) \frac{\partial{f}}{\partial{w^{*}}}=\frac{1}{2}(\frac{\partial{f}}{\partial{x}}+j\frac{\partial{f}}{\partial{y}}) ∂w∗∂f=21(∂x∂f+j∂y∂f) ∂ w ∂ w = 1 \frac{\partial{w}}{\partial{w}}=1 ∂w∂w=1 ∂ w ∂ w ∗ = 0 \frac{\partial{w}}{\partial{w^{*}}}=0 ∂w∗∂w=0
2. 标量函数相对于复向量的梯度
a. 定义 f ( w ⃗ ) f(\vec{\bm{w}}) f(w) 是复向量 w ⃗ \vec{\bm{w}} w 的标量函数;
b. 其中 w ⃗ = [ w 0 , . . . , w M − 1 ] T \vec{\bm{w}}=[w_0,...,w_{M-1}]^{T} w=[w0,...,wM−1]T w n = x n + j y n w_n=x_n+jy_n wn=xn+jyn
c. 那么
∂
f
∂
w
⃗
=
∇
w
⃗
f
(
w
⃗
)
=
[
∂
f
∂
w
0
,
.
.
.
,
∂
f
∂
w
M
−
1
]
T
\frac{\partial{f}}{\partial{\vec{\bm{w}}}}=\nabla_{\vec{\bm{w}}}f(\vec{\bm{w}})=[\frac{\partial{f}}{\partial{w_0}},...,\frac{\partial{f}}{\partial{w_{M-1}}}]^{T}
∂w∂f=∇wf(w)=[∂w0∂f,...,∂wM−1∂f]T
∂
f
∂
w
∗
⃗
=
∇
w
∗
⃗
f
(
w
⃗
)
=
[
∂
f
∂
w
0
∗
,
.
.
.
,
∂
f
∂
w
M
−
1
∗
]
T
\frac{\partial{f}}{\partial{\vec{\bm{w^{*}}}}}=\nabla_{\vec{\bm{w^{*}}}}f(\vec{\bm{w}})=[\frac{\partial{f}}{\partial{w_0^{*}}},...,\frac{\partial{f}}{\partial{w_{M-1}^{*}}}]^{T}
∂w∗∂f=∇w∗f(w)=[∂w0∗∂f,...,∂wM−1∗∂f]T
其中, T T T 表示转置操作。
3. 向量函数相对于复向量的梯度
a. 定义 f ⃗ \vec{\bm{f}} f 是复向量 w ⃗ \vec{\bm{w}} w 的向量函数
b. 其中
f
⃗
=
[
f
0
(
w
⃗
)
,
.
.
.
,
f
M
−
1
(
w
⃗
)
]
T
\vec{\bm{f}}=[f_0(\vec{\bm{w}}),...,f_{M-1}(\vec{\bm{w}})]^{T}
f=[f0(w),...,fM−1(w)]T
w
⃗
=
[
w
0
,
.
.
.
,
w
M
−
1
]
T
\vec{\bm{w}}=[w_0,...,w_{M-1}]^{T}
w=[w0,...,wM−1]T
w
n
=
x
n
+
j
y
n
w_n = x_n + jy_n
wn=xn+jyn
c. 那么
∂
f
⃗
∂
w
⃗
=
[
∂
f
0
∂
w
0
,
.
.
.
,
∂
f
0
∂
w
M
−
1
∂
f
1
∂
w
0
,
.
.
.
,
∂
f
1
∂
w
M
−
1
.
.
.
.
.
.
.
.
.
∂
f
M
−
1
∂
w
0
,
.
.
.
,
∂
f
M
−
1
∂
w
M
−
1
]
\frac{\partial{\vec{\bm{f}}}}{\partial{\vec{\bm{w}}}}=\begin{bmatrix} \frac{\partial{f_0}}{\partial{w_0}},...,\frac{\partial{f_0}}{\partial{w_{M-1}}} \\ \frac{\partial{f_1}}{\partial{w_0}},...,\frac{\partial{f_1}}{\partial{w_{M-1}}} \\ ......... \\ \frac{\partial{f_{M-1}}}{\partial{w_0}},...,\frac{\partial{f_{M-1}}}{\partial{w_{M-1}}} \end{bmatrix}
∂w∂f=
∂w0∂f0,...,∂wM−1∂f0∂w0∂f1,...,∂wM−1∂f1.........∂w0∂fM−1,...,∂wM−1∂fM−1
d. 结合上述概念可得到 ∂ f ⃗ ∂ w ∗ ⃗ \frac{\partial{\vec{\bm{f}}}}{\partial{\vec{\bm{w^{*}}}}} ∂w∗∂f
4. 关于复变量/向量的一些结论
a. 定义 ∇ = ( ∂ ∂ z 0 ∗ , . . . , ∂ ∂ z N − 1 ∗ ) T \nabla=(\frac{\partial}{\partial{z_0^{*}}},...,\frac{\partial}{\partial{z_{N-1}^{*}}})^{T} ∇=(∂z0∗∂,...,∂zN−1∗∂)T z n = x n + j y n z_n=x_n+jy_n zn=xn+jyn n = 0 , . . . , N − 1 n=0,...,N-1 n=0,...,N−1
b. ∂ ∂ z n ∗ = 1 2 ( ∂ ∂ x n + j ∂ ∂ y n ) \frac{\partial}{\partial{z_n^{*}}}=\frac{1}{2}(\frac{\partial}{\partial{x_n}}+j\frac{\partial}{\partial{y_n}}) ∂zn∗∂=21(∂xn∂+j∂yn∂) z n = x n + j y n z_n=x_n+jy_n zn=xn+jyn n = 0 , . . . , N − 1 n=0,...,N-1 n=0,...,N−1
c.
{
∂
z
∂
z
∗
=
0
∂
z
∂
z
=
1
∂
z
∗
∂
z
∗
=
1
\begin{cases} \frac{\partial{z}}{\partial{z^{*}}}=0\\ \frac{\partial{z}}{\partial{z}}=1\\ \frac{\partial{z^{*}}}{\partial{z^{*}}}=1 \end{cases}
⎩
⎨
⎧∂z∗∂z=0∂z∂z=1∂z∗∂z∗=1
d.
{
∇
(
a
⃗
H
z
⃗
)
=
0
∇
(
z
⃗
H
a
⃗
)
=
a
⃗
∇
(
z
⃗
H
R
⃗
z
⃗
)
=
R
⃗
z
⃗
\begin{cases} \nabla(\vec{\bm{a}}^{H}\vec{\bm{z}}) = 0\\ \nabla(\vec{\bm{z}}^{H}\vec{\bm{a}}) = \vec{\bm{a}}\\ \nabla(\vec{\bm{z}}^{H}\vec{\bm{R}}\vec{\bm{z}}) = \vec{\bm{R}}\vec{\bm{z}}\\ \end{cases}
⎩
⎨
⎧∇(aHz)=0∇(zHa)=a∇(zHRz)=Rz
其中, a ⃗ \vec{\bm{a}} a 和 z ⃗ \vec{\bm{z}} z 均为列向量, R ⃗ \vec{\bm{R}} R 为矩阵, H H H 表示共轭转置
e.
f
(
w
⃗
)
=
p
⃗
H
w
⃗
⇒
{
∂
f
∂
w
⃗
=
p
⃗
∗
∂
f
∂
w
⃗
∗
=
0
f(\vec{\bm{w}})=\vec{\bm{p}}^{H}\vec{\bm{w}}\Rightarrow \begin{cases} \frac{\partial{f}}{\partial{\vec{\bm{w}}}}=\vec{\bm{p}}^{*}\\ \frac{\partial{f}}{\partial{\vec{\bm{w}}^{*}}}=0 \end{cases}
f(w)=pHw⇒{∂w∂f=p∗∂w∗∂f=0
其中, p ⃗ \vec{\bm{p}} p 和 w ⃗ \vec{\bm{w}} w 均为列向量, f ( w ⃗ ) f(\vec{\bm{w}}) f(w) 为标量
f.
f
(
w
⃗
)
=
w
⃗
H
p
⃗
⇒
{
∂
f
∂
w
⃗
=
0
∂
f
∂
w
⃗
∗
=
p
⃗
f(\vec{\bm{w}})=\vec{\bm{w}}^{H}\vec{\bm{p}}\Rightarrow \begin{cases} \frac{\partial{f}}{\partial{\vec{\bm{w}}}}=0\\ \frac{\partial{f}}{\partial{\vec{\bm{w}}^{*}}}=\vec{\bm{p}} \end{cases}
f(w)=wHp⇒{∂w∂f=0∂w∗∂f=p
其中, p ⃗ \vec{\bm{p}} p 和 w ⃗ \vec{\bm{w}} w 均为列向量, f ( w ⃗ ) f(\vec{\bm{w}}) f(w) 为标量
g.
f
(
w
⃗
)
=
w
⃗
H
A
⃗
w
⃗
⇒
{
∂
f
∂
w
⃗
=
A
⃗
T
w
⃗
∗
∂
f
∂
w
⃗
∗
=
A
⃗
w
⃗
f(\vec{\bm{w}})=\vec{\bm{w}}^{H}\vec{\bm{A}}\vec{\bm{w}}\Rightarrow \begin{cases} \frac{\partial{f}}{\partial{\vec{\bm{w}}}}=\vec{\bm{A}}^{T}\vec{\bm{w}}^{*}\\ \frac{\partial{f}}{\partial{\vec{\bm{w}}^{*}}}=\vec{\bm{A}}\vec{\bm{w}} \end{cases}
f(w)=wHAw⇒{∂w∂f=ATw∗∂w∗∂f=Aw
其中, w ⃗ \vec{\bm{w}} w 为列向量, A ⃗ \vec{\bm{A}} A 为矩阵, f ( w ⃗ ) f(\vec{\bm{w}}) f(w) 为标量
5. 关于实变量/向量的一些结论
a. ∂ a ⃗ T b ⃗ ∂ b ⃗ = ∂ b ⃗ T a ⃗ ∂ b ⃗ = a ⃗ \frac{\partial{\vec{\bm{a}}}^{T}{\vec{\bm{b}}}}{\partial{\vec{\bm{b}}}}=\frac{\partial{\vec{\bm{b}}}^{T}{\vec{\bm{a}}}}{\partial{\vec{\bm{b}}}}=\vec{\bm{a}} ∂b∂aTb=∂b∂bTa=a
b. ∂ b ⃗ T A ⃗ b ⃗ ∂ b ⃗ = 2 A ⃗ b ⃗ (or = 2 b ⃗ T A ⃗ \frac{\partial{\vec{\bm{b}}}^{T}{\vec{\bm{A}}}{\vec{\bm{b}}}}{\partial{\vec{\bm{b}}}}=2\vec{\bm{A}}\vec{\bm{b}}(\text{or} \; =2\vec{\bm{b}}^{T}\vec{\bm{A}} ∂b∂bTAb=2Ab(or=2bTA
其中, a ⃗ \vec{\bm{a}} a 和 b ⃗ \vec{\bm{b}} b 均为列向量, A ⃗ \vec{\bm{A}} A 为对称矩阵
6. 参考文献
[1] 张贤达.矩阵分析与应用[M].北京:高等教育出版社,2004.