Matrix Cookbook
Matrix Calculus
| 列向量求导 A ∈ R N × m , x ∈ R m , b ∈ R n A \in \mathbb{R}^{N \times m}, x \in \mathbb{R}^{m}, b \in \mathbb{R}^{n} A∈RN×m,x∈Rm,b∈Rn | 矩阵求导 |
|---|---|
| ∂ x T D x ∂ x = ( D + D T ) x \frac{\partial x^{T} D x}{\partial x} =\left(D+D^{T}\right) x ∂x∂xTDx=(D+DT)x | ∂ tr ( X H A X ) ∂ X = ( A + A H ) X \frac{\partial \operatorname{tr}\left({\boldsymbol{X}}^{H} \boldsymbol{A} {\boldsymbol{X}}\right)} {\partial \boldsymbol{X}} =\left(\boldsymbol{A} +\boldsymbol{A} ^{H}\right) \boldsymbol{X} ∂X∂tr(XHAX)=(A+AH)X |
| ∂ D T x ∂ x = D \frac{\partial D^{T} x}{\partial x} =D ∂x∂DTx=D | ∂ tr ( A H X ) ∂ X = A \frac{\partial \operatorname{tr}\left(\boldsymbol{A}^{H} {\boldsymbol{X}}\right)}{\partial \boldsymbol{X}} =\boldsymbol{A} ∂X∂tr(AHX)=A |
| ∂ x T D ∂ x = D \frac{\partial x^{T} D}{\partial x} =D ∂x∂xTD=D | ∂ tr ( X H A ) ∂ X = A \frac{\partial \operatorname{tr}\left({\boldsymbol{X}}^{H} {\boldsymbol{A}}\right)}{\partial \boldsymbol{X}} =\boldsymbol{A} ∂X∂tr(XHA)=A |
| ∂ x T x ∂ x = 2 x \frac{\partial x^{T} x}{\partial x} =2x ∂x∂xTx=2x， ∥ x ∥ 2 2 = t r ( x H x ) \Vert{x}\Vert_{2}^2 =tr(x^H x) ∥x∥22=tr(xHx) | ∂ tr ( X H X ) ∂ X = 2 X \frac{\partial \operatorname{tr}\left({\boldsymbol{X}}^{H} {\boldsymbol{X}}\right)} {\partial \boldsymbol{X}} =2 \boldsymbol{X} ∂X∂tr(XHX)=2X， ∥ X ∥ F 2 = t r ( X H X ) \Vert{\boldsymbol{X}}\Vert_{F}^2 =tr({\boldsymbol{X}}^H{\boldsymbol{X}}) ∥X∥F2=tr(XHX) |
| ∂ ( 1 2 ∥ A x − b ∥ 2 2 ) ∂ x = A H A x − A H b \frac{\partial (\frac{1}{2}\|A x-b\|_{2}^{2})}{\partial x}=A^{H} A x-A^{H} b ∂x∂(21∥Ax−b∥22)=AHAx−AHb | ∂ ( 1 2 ∥ Y − A X ∥ F 2 ) ∂ x i j = [ A H A X − A H Y ] i j \frac{\partial\left(\frac{1}{2}\|Y-A X\|_{F}^{2}\right)}{\partial x_{i j}} = [A^{H} A X-A^{H} Y]_{ij} ∂xij∂(21∥Y−AX∥F2)=[AHAX−AHY]ij |
| ∂ ( 1 2 ∥ Y − A X ∥ F 2 ) ∂ a i j = [ A X X H − Y X H ] i j \frac{\partial\left(\frac{1}{2}\|Y-A X\|_{F}^{2}\right)}{\partial a_{i j}} = [A X X^{H}-Y X^{H}]_{ij} ∂aij∂(21∥Y−AX∥F2)=[AXXH−YXH]ij | |
向量求导
∂
x
T
D
x
∂
x
=
(
D
+
D
T
)
x
\frac{\partial x^{T} D x}{\partial x} =\left(D+D^{T}\right) x
∂x∂xTDx=(D+DT)x
∂
D
T
x
∂
x
=
D
\frac{\partial D^{T} x}{\partial x} =D
∂x∂DTx=D
∂
x
T
D
∂
x
=
D
\frac{\partial x^{T} D}{\partial x} =D
∂x∂xTD=D
向量2范数的求导
A
∈
R
N
×
m
,
x
∈
R
m
,
b
∈
R
n
A \in \mathbb{R}^{N \times m}, x \in \mathbb{R}^{m}, b \in \mathbb{R}^{n}
A∈RN×m,x∈Rm,b∈Rn
∂
(
1
2
∥
A
x
−
b
∥
2
2
)
∂
x
=
A
T
(
A
x
−
b
)
=
A
T
A
x
−
A
T
b
\frac{\partial (\frac{1}{2}\|A x-b\|_{2}^{2})}{\partial x}=A^{T}(A x-b)=A^{T} A x-A^{T} b
∂x∂(21∥Ax−b∥22)=AT(Ax−b)=ATAx−ATb
矩阵求导
矩阵理论学习指导88页
∂
tr
(
X
H
A
X
)
∂
X
=
(
A
+
A
H
)
X
\frac{\partial \operatorname{tr}\left({\boldsymbol{X}}^{H} \boldsymbol{A} {\boldsymbol{X}}\right)} {\partial \boldsymbol{X}} =\left(\boldsymbol{A} +\boldsymbol{A} ^{H}\right) \boldsymbol{X}
∂X∂tr(XHAX)=(A+AH)X 108
∂
tr
(
X
A
X
H
)
∂
X
=
X
(
A
+
A
H
)
\frac{\partial \operatorname{tr}\left({\boldsymbol{X}} \boldsymbol{A} {\boldsymbol{X}}^{H}\right)} {\partial \boldsymbol{X}} =\boldsymbol{X} \left(\boldsymbol{A} +\boldsymbol{A} ^{H}\right)
∂X∂tr(XAXH)=X(A+AH) 111
∂
tr
(
A
H
X
)
∂
X
=
A
\frac{\partial \operatorname{tr}\left(\boldsymbol{A}^{H} {\boldsymbol{X}}\right)} {\partial \boldsymbol{X}} =\boldsymbol{A}
∂X∂tr(AHX)=A
∂
tr
(
X
H
A
)
∂
X
=
A
\frac{\partial \operatorname{tr}\left({\boldsymbol{X}}^{H} {\boldsymbol{A}}\right)} {\partial \boldsymbol{X}} =\boldsymbol{A}
∂X∂tr(XHA)=A
矩阵理论学习指导95页 矩阵m2范数的求导
∥
X
∥
F
2
=
t
r
(
X
H
X
)
\Vert{\boldsymbol{X}}\Vert_{F}^2 =tr({\boldsymbol{X}}^H{\boldsymbol{X}})
∥X∥F2=tr(XHX)
∂
tr
(
X
H
X
)
∂
X
=
2
X
\frac{\partial \operatorname{tr}\left({\boldsymbol{X}}^{H} {\boldsymbol{X}}\right)} {\partial \boldsymbol{X}} =2 \boldsymbol{X}
∂X∂tr(XHX)=2X
矩阵m2范数的求导
矩阵m2范数的求导
∂
(
1
2
∥
Y
−
A
X
∥
F
2
)
∂
a
i
j
=
[
A
X
X
T
−
Y
X
T
]
i
j
\frac{\partial\left(\frac{1}{2}\|Y-A X\|_{F}^{2}\right)}{\partial a_{i j}} = [A X X^{T}-Y X^{T}]_{ij}
∂aij∂(21∥Y−AX∥F2)=[AXXT−YXT]ij
∂
(
1
2
∥
Y
−
A
X
∥
F
2
)
∂
x
i
j
=
[
A
T
A
X
−
A
T
Y
]
i
j
\frac{\partial\left(\frac{1}{2}\|Y-A X\|_{F}^{2}\right)}{\partial x_{i j}} = [A^{T} A X-A^{T} Y]_{ij}
∂xij∂(21∥Y−AX∥F2)=[ATAX−ATY]ij
迹
- tr(A) = tr(A’)
- tr(kA) = k*tr(A) ,
- tr(A+B) = tr(A) + tr(B)
- tr(AB)=tr(BA)
求导
关于矩阵X的函数f(X)对矩阵X求导 = 由(∂f(X)/∂xij)组成的矩阵
3050
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
