F ( θ ) = ∫ ( ∂ ∂ θ l o g q ( x ; θ ) ) ( ∂ ∂ θ T l o g q ( x ; θ ) ) q ( x ; θ ) d x F(\theta)=\int\bigg(\frac{\partial}{\partial\theta}logq(x;\theta)\bigg)\bigg(\frac{\partial}{\partial\theta^T}logq(x;\theta)\bigg)q(x;\theta)dx F(θ)=∫(∂θ∂logq(x;θ))(∂θT∂logq(x;θ))q(x;θ)dx ⇕ \bigg\Updownarrow ⇓‖‖⇑ F ( θ ) = − ∫ ( ∂ 2 ∂ θ ∂ θ T l o g q ( x ; θ ) ) q ( x ; θ ) d x F(\theta)=-\int\bigg(\frac{\partial^2}{\partial\theta\partial\theta^T}logq(x;\theta)\bigg)q(x;\theta)dx F(θ)=−∫(∂θ∂θT∂2logq(x;θ))q(x;θ)dx
证明过程:
我
们
可
以
从
F
(
θ
2
)
后
面
那
一
堆
入
手
:
我们可以从F(\theta_2)后面那一堆入手:
我们可以从F(θ2)后面那一堆入手:
∫
(
∂
2
∂
θ
∂
θ
T
l
o
g
q
(
x
;
θ
)
)
q
(
x
;
θ
)
d
x
\int\bigg(\frac{\partial^2}{\partial\theta\partial\theta^T}logq(x;\theta)\bigg)q(x;\theta)dx
∫(∂θ∂θT∂2logq(x;θ))q(x;θ)dx
首
先
,
(
∂
2
∂
θ
∂
θ
T
l
o
g
q
(
x
;
θ
)
)
可
以
写
成
:
(
∂
∂
θ
[
∂
∂
θ
T
l
o
g
q
(
x
;
θ
)
]
)
首先,\bigg(\frac{\partial^2}{\partial\theta\partial\theta^T}logq(x;\theta)\bigg)可以写成:\bigg(\frac{\partial}{\partial\theta}\big[\frac{\partial}{\partial\theta^T}logq(x;\theta)\big]\bigg)
首先,(∂θ∂θT∂2logq(x;θ))可以写成:(∂θ∂[∂θT∂logq(x;θ)])
对
中
括
号
里
面
的
东
西
求
导
:
[
1
q
(
x
;
θ
)
⋅
∂
q
(
x
;
θ
)
∂
θ
T
]
对中括号里面的东西求导: \big[\frac{1}{q(x;\theta)} · \frac{\partial q(x;\theta)}{\partial\theta^T}\big]
对中括号里面的东西求导:[q(x;θ)1⋅∂θT∂q(x;θ)]
此
时
上
式
就
变
成
了
:
(
∂
∂
θ
[
1
q
(
x
;
θ
)
⋅
∂
q
(
x
;
θ
)
∂
θ
T
]
)
此时上式就变成了: \bigg(\frac{\partial}{\partial\theta}\big[\frac{1}{q(x;\theta)} · \frac{\partial q(x;\theta)}{\partial\theta^T}\big]\bigg)
此时上式就变成了:(∂θ∂[q(x;θ)1⋅∂θT∂q(x;θ)])
继
续
对
圆
括
号
里
的
内
容
求
导
:
继续对圆括号里的内容求导:
继续对圆括号里的内容求导:
(
−
∂
q
(
x
;
θ
)
∂
θ
q
(
x
;
θ
)
2
⋅
∂
q
(
x
;
θ
)
∂
θ
T
+
∂
∂
θ
⋅
∂
q
(
x
;
θ
)
∂
θ
T
⋅
1
q
(
x
;
θ
)
)
\bigg(-\frac{\frac{\partial q(x;\theta)}{\partial\theta}}{q(x;\theta)^2}·\frac{\partial q(x;\theta)}{\partial\theta^T}+\frac{\partial}{\partial\theta}·\frac{\partial q(x;\theta)}{\partial\theta^T}·\frac{1}{q(x;\theta)}\bigg)
(−q(x;θ)2∂θ∂q(x;θ)⋅∂θT∂q(x;θ)+∂θ∂⋅∂θT∂q(x;θ)⋅q(x;θ)1)
带
到
原
式
子
里
面
就
变
成
了
:
带到原式子里面就变成了:
带到原式子里面就变成了:
∫
(
−
∂
q
(
x
;
θ
)
∂
θ
q
(
x
;
θ
)
2
⋅
∂
q
(
x
;
θ
)
∂
θ
T
+
∂
∂
θ
⋅
∂
q
(
x
;
θ
)
∂
θ
T
⋅
1
q
(
x
;
θ
)
)
q
(
x
;
θ
)
d
x
\int\bigg(-\frac{\frac{\partial q(x;\theta)}{\partial\theta}}{q(x;\theta)^2}·\frac{\partial q(x;\theta)}{\partial\theta^T}+\frac{\partial}{\partial\theta}·\frac{\partial q(x;\theta)}{\partial\theta^T}·\frac{1}{q(x;\theta)}\bigg)q(x;\theta)dx
∫(−q(x;θ)2∂θ∂q(x;θ)⋅∂θT∂q(x;θ)+∂θ∂⋅∂θT∂q(x;θ)⋅q(x;θ)1)q(x;θ)dx
然
后
将
这
个
式
子
拆
开
:
然后将这个式子拆开:
然后将这个式子拆开:
∫
∂
2
∂
θ
∂
T
q
(
x
;
θ
)
q
(
x
;
θ
)
q
(
x
;
θ
)
d
x
−
∫
∂
q
(
x
;
θ
)
∂
θ
⋅
∂
q
(
x
;
θ
)
∂
θ
T
q
(
x
;
θ
2
)
q
(
x
;
θ
)
d
x
\int\frac{\frac{\partial^2}{\partial\theta\partial^T}q(x;\theta)}{q(x;\theta)}q(x;\theta)dx-\int\frac{\frac{\partial q(x;\theta)}{\partial\theta}·\frac{\partial q(x;\theta)}{\partial\theta^T}}{q(x;\theta^2)}q(x;\theta)dx
∫q(x;θ)∂θ∂T∂2q(x;θ)q(x;θ)dx−∫q(x;θ2)∂θ∂q(x;θ)⋅∂θT∂q(x;θ)q(x;θ)dx
∫
∂
2
∂
θ
∂
θ
T
q
(
x
;
θ
)
d
x
−
∫
(
∂
∂
θ
l
o
g
q
(
x
;
θ
)
)
(
∂
∂
θ
T
l
o
g
q
(
x
;
θ
)
)
q
(
x
;
θ
)
d
x
⏟
注
意
这
里
就
是
开
篇
的
第
一
个
式
子
\int\frac{\partial^2}{\partial\theta\partial\theta^T}q(x;\theta)dx-\underbrace{\int\bigg(\frac{\partial}{\partial\theta}logq(x;\theta)\bigg)\bigg(\frac{\partial}{\partial\theta^T}logq(x;\theta)\bigg)q(x;\theta)dx}_{注意这里就是开篇的第一个式子}
∫∂θ∂θT∂2q(x;θ)dx−注意这里就是开篇的第一个式子
∫(∂θ∂logq(x;θ))(∂θT∂logq(x;θ))q(x;θ)dx
∂
2
∂
θ
∂
θ
T
∫
q
(
x
;
θ
)
d
x
−
F
(
θ
)
\frac{\partial^2}{\partial\theta\partial\theta^T}\int q(x;\theta)dx-F(\theta)
∂θ∂θT∂2∫q(x;θ)dx−F(θ)
∂
2
∂
θ
∂
θ
T
⋅
1
−
F
(
θ
)
=
−
F
(
θ
)
\frac{\partial^2}{\partial\theta\partial\theta^T}·1-F(\theta)=-F(\theta)
∂θ∂θT∂2⋅1−F(θ)=−F(θ)
完整过程:
F
(
θ
1
)
=
∫
(
∂
∂
θ
l
o
g
q
(
x
;
θ
)
)
(
∂
∂
θ
T
l
o
g
q
(
x
;
θ
)
)
q
(
x
;
θ
)
d
x
F(\theta_1)=\int\bigg(\frac{\partial}{\partial\theta}logq(x;\theta)\bigg)\bigg(\frac{\partial}{\partial\theta^T}logq(x;\theta)\bigg)q(x;\theta)dx
F(θ1)=∫(∂θ∂logq(x;θ))(∂θT∂logq(x;θ))q(x;θ)dx
F
(
θ
2
)
=
−
∫
(
∂
2
∂
θ
∂
θ
T
l
o
g
q
(
x
;
θ
)
)
q
(
x
;
θ
)
d
x
F(\theta_2)=-\int\bigg(\frac{\partial^2}{\partial\theta\partial\theta^T}logq(x;\theta)\bigg)q(x;\theta)dx
F(θ2)=−∫(∂θ∂θT∂2logq(x;θ))q(x;θ)dx
∫
(
∂
2
∂
θ
∂
θ
T
l
o
g
q
(
x
;
θ
)
)
q
(
x
;
θ
)
d
x
\int\bigg(\frac{\partial^2}{\partial\theta\partial\theta^T}logq(x;\theta)\bigg)q(x;\theta)dx
∫(∂θ∂θT∂2logq(x;θ))q(x;θ)dx
=
∫
(
∂
∂
θ
[
∂
∂
θ
T
l
o
g
q
(
x
;
θ
)
]
)
q
(
x
;
θ
)
d
x
=\int\bigg(\frac{\partial}{\partial\theta}\big[\frac{\partial}{\partial\theta^T}logq(x;\theta)\big]\bigg)q(x;\theta)dx
=∫(∂θ∂[∂θT∂logq(x;θ)])q(x;θ)dx
=
∫
(
∂
∂
θ
[
1
q
(
x
;
θ
)
⋅
∂
q
(
x
;
θ
)
∂
θ
T
]
)
q
(
x
;
θ
)
d
x
=\int\bigg(\frac{\partial}{\partial\theta}\big[\frac{1}{q(x;\theta)} · \frac{\partial q(x;\theta)}{\partial\theta^T}\big]\bigg)q(x;\theta)dx
=∫(∂θ∂[q(x;θ)1⋅∂θT∂q(x;θ)])q(x;θ)dx
=
∫
(
−
∂
q
(
x
;
θ
)
∂
θ
q
(
x
;
θ
)
2
⋅
∂
q
(
x
;
θ
)
∂
θ
T
+
∂
∂
θ
⋅
∂
q
(
x
;
θ
)
∂
θ
T
⋅
1
q
(
x
;
θ
)
)
q
(
x
;
θ
)
d
x
=\int\bigg(-\frac{\frac{\partial q(x;\theta)}{\partial\theta}}{q(x;\theta)^2}·\frac{\partial q(x;\theta)}{\partial\theta^T}+\frac{\partial}{\partial\theta}·\frac{\partial q(x;\theta)}{\partial\theta^T}·\frac{1}{q(x;\theta)}\bigg)q(x;\theta)dx
=∫(−q(x;θ)2∂θ∂q(x;θ)⋅∂θT∂q(x;θ)+∂θ∂⋅∂θT∂q(x;θ)⋅q(x;θ)1)q(x;θ)dx
=
∫
∂
2
∂
θ
∂
T
q
(
x
;
θ
)
q
(
x
;
θ
)
q
(
x
;
θ
)
d
x
−
∫
∂
q
(
x
;
θ
)
∂
θ
⋅
∂
q
(
x
;
θ
)
∂
θ
T
q
(
x
;
θ
2
)
q
(
x
;
θ
)
d
x
=\int\frac{\frac{\partial^2}{\partial\theta\partial^T}q(x;\theta)}{q(x;\theta)}q(x;\theta)dx-\int\frac{\frac{\partial q(x;\theta)}{\partial\theta}·\frac{\partial q(x;\theta)}{\partial\theta^T}}{q(x;\theta^2)}q(x;\theta)dx
=∫q(x;θ)∂θ∂T∂2q(x;θ)q(x;θ)dx−∫q(x;θ2)∂θ∂q(x;θ)⋅∂θT∂q(x;θ)q(x;θ)dx
=
∫
∂
2
∂
θ
∂
θ
T
q
(
x
;
θ
)
d
x
−
∫
(
∂
∂
θ
l
o
g
q
(
x
;
θ
)
)
(
∂
∂
θ
T
l
o
g
q
(
x
;
θ
)
)
q
(
x
;
θ
)
d
x
⏟
F
(
θ
1
)
=\int\frac{\partial^2}{\partial\theta\partial\theta^T}q(x;\theta)dx-\underbrace{\int\bigg(\frac{\partial}{\partial\theta}logq(x;\theta)\bigg)\bigg(\frac{\partial}{\partial\theta^T}logq(x;\theta)\bigg)q(x;\theta)dx}_{F(\theta_1)}
=∫∂θ∂θT∂2q(x;θ)dx−F(θ1)
∫(∂θ∂logq(x;θ))(∂θT∂logq(x;θ))q(x;θ)dx
=
∂
2
∂
θ
∂
θ
T
∫
q
(
x
;
θ
)
d
x
−
F
(
θ
1
)
=\frac{\partial^2}{\partial\theta\partial\theta^T}\int q(x;\theta)dx-F(\theta_1)
=∂θ∂θT∂2∫q(x;θ)dx−F(θ1)
=
∂
2
∂
θ
∂
θ
T
⋅
1
⏟
0
−
F
(
θ
1
)
=
−
F
(
θ
1
)
=\underbrace{\frac{\partial^2}{\partial\theta\partial\theta^T}·1}_0-F(\theta_1)=-F(\theta_1)
=0
∂θ∂θT∂2⋅1−F(θ1)=−F(θ1)
即
F
(
θ
2
)
=
−
(
−
F
(
θ
1
)
)
=
F
(
θ
1
)
即F(\theta_2)=-\big(-F(\theta_1)\big)=F(\theta_1)
即F(θ2)=−(−F(θ1))=F(θ1)
证
毕
证毕
证毕