假设
n
n
n 维高斯随机变量
X
∼
N
(
μ
,
Σ
)
X \sim N(\mu, \Sigma)
X∼N(μ,Σ)其概率密度函数为:
f
X
(
x
)
=
α
e
−
1
2
[
(
x
−
μ
)
T
Σ
−
1
(
x
−
μ
)
]
(1)
\tag 1 f_X(x) =\alpha e^{-\frac 1 2 [ ( x - \mu) ^ T \Sigma^{-1} (x - \mu)]}
fX(x)=αe−21[(x−μ)TΣ−1(x−μ)](1)其中
α
\alpha
α是归一化因子。
根据Wikipedia对Fisher Information的定义,
[
I
(
θ
)
]
i
,
j
=
−
E
X
[
∂
2
∂
θ
i
∂
θ
j
l
o
g
f
X
(
x
;
θ
)
∣
θ
]
(2)
\tag 2 [\Iota(\theta)]_{i,j} = - E_X[\frac {\partial^2} {\partial {\theta_i} \partial \theta_j}log f_X(x;\theta) | \theta]
[I(θ)]i,j=−EX[∂θi∂θj∂2logfX(x;θ)∣θ](2)对照式(1)和式(2),视式(1)中的
μ
\mu
μ为式(2)中的参数
θ
\theta
θ,并假设随机变量
X
X
X的协方差矩阵
Σ
\Sigma
Σ以及
f
X
f_X
fX中的归一化因子
α
\alpha
α与
μ
\mu
μ无关;则有:
∂
2
∂
μ
i
∂
μ
j
l
n
f
X
(
x
;
μ
)
=
∂
2
μ
i
∂
μ
j
l
n
(
α
e
−
1
2
[
(
x
−
μ
)
T
Σ
−
1
(
x
−
μ
)
]
)
=
∂
2
∂
μ
i
∂
μ
j
(
l
n
α
−
1
2
[
(
x
−
μ
)
T
Σ
−
1
(
x
−
μ
)
]
)
=
−
1
2
∂
2
∂
μ
i
∂
μ
j
[
(
x
−
μ
)
T
Σ
−
1
(
x
−
μ
)
]
=
−
[
Σ
−
1
]
i
,
j
(3)
\begin{aligned} \tag 3 \frac {\partial^2} {\partial {\mu_i} \partial \mu_j}ln f_X(x;\mu) &= \frac {\partial^2} {\ {\mu_i} \partial \mu_j}ln (\alpha e^{-\frac 1 2 [ ( x - \mu) ^ T \Sigma^{-1} (x - \mu)]}) \\ &= \frac {\partial^2} {\partial {\mu_i} \partial \mu_j}(ln \alpha -\frac 1 2 [ (x - \mu) ^ T \Sigma^{-1} (x - \mu)])\\ &= -\frac 1 2 \frac {\partial^2} {\partial {\mu_i} \partial \mu_j}[( x - \mu) ^ T \Sigma^{-1} (x - \mu)]\\ &=-[\Sigma^{-1}]_{i,j} \end{aligned}
∂μi∂μj∂2lnfX(x;μ)= μi∂μj∂2ln(αe−21[(x−μ)TΣ−1(x−μ)])=∂μi∂μj∂2(lnα−21[(x−μ)TΣ−1(x−μ)])=−21∂μi∂μj∂2[(x−μ)TΣ−1(x−μ)]=−[Σ−1]i,j(3)
于是:
[
I
(
μ
)
]
i
j
=
−
E
X
[
−
[
Σ
−
1
]
i
,
j
]
(3)
\tag {3} [\Iota(\mu)]_{ij}=-E_X[-[\Sigma^{-1}]_{i,j}]
[I(μ)]ij=−EX[−[Σ−1]i,j](3)
由于协方差矩阵的逆
Σ
−
1
\Sigma^{-1}
Σ−1是与
μ
\mu
μ无关的常数,因此
[
I
(
μ
)
]
i
j
=
E
X
[
[
Σ
−
1
]
i
,
j
]
=
∫
[
Σ
−
1
]
i
,
j
f
X
(
x
)
d
x
=
[
Σ
−
1
]
i
,
j
∫
f
X
(
x
)
d
x
=
[
Σ
−
1
]
i
,
j
∗
1
=
[
Σ
−
1
]
i
,
j
(4)
\begin{aligned} \tag {4} [\Iota(\mu)]_{ij}&=E_X[[\Sigma^{-1}]_{i,j}]\\ &=\intop [\Sigma^{-1}]_{i,j} f_X(x)dx \\ &=[\Sigma^{-1}]_{i,j}\intop f_X(x)dx \\ &=[\Sigma^{-1}]_{i,j} * 1 \\ &=[\Sigma^{-1}]_{i,j} \end{aligned}
[I(μ)]ij=EX[[Σ−1]i,j]=∫[Σ−1]i,jfX(x)dx=[Σ−1]i,j∫fX(x)dx=[Σ−1]i,j∗1=[Σ−1]i,j(4)
因此信息矩阵
Λ
=
I
(
μ
)
=
Σ
−
1
(5)
\tag {5} \Lambda = \Iota(\mu)=\Sigma^{-1}
Λ=I(μ)=Σ−1(5)
即信息矩阵等于协方差矩阵的逆