设真值为
θ
θ
θ,
x
1
,
x
2
⋯
,
x
n
x_1,x_2⋯,x_n
x1,x2⋯,xn为 n 次独立测量值, 每次测量的误差为
e
i
=
x
i
–
θ
e_i=x_i–θ
ei=xi–θ,假设误差
e
i
e_i
ei的密度函数为
f
(
e
)
f(e)
f(e), 则测量值的联合概率为
n
n
n个误差的联合概率,记为:
L
(
θ
)
=
L
(
x
1
,
x
2
,
⋯
,
x
n
:
θ
)
=
∏
i
=
1
n
f
(
e
i
)
=
∏
i
=
1
n
f
(
x
i
−
θ
)
\begin{aligned} L(θ)&=L(x_1,x_2,⋯,x_n:\theta) \\ &= \prod_{i=1}^{n}f(e_i) \\ &= \prod_{i=1}^{n}f(x_i−θ) \end{aligned}
L(θ)=L(x1,x2,⋯,xn:θ)=i=1∏nf(ei)=i=1∏nf(xi−θ)
对上式两边取对数:
l
n
L
(
θ
)
=
∑
i
=
1
n
l
n
f
(
x
i
−
θ
)
lnL(\theta) = \sum_{i=1}^{n}lnf(x_i-\theta)
lnL(θ)=i=1∑nlnf(xi−θ)
对上式两边求导:
d
l
n
L
(
θ
)
d
θ
=
∑
i
=
1
n
d
l
n
f
(
x
i
−
θ
)
d
θ
=
∑
i
=
1
n
f
′
(
x
i
−
θ
)
f
(
x
i
−
θ
)
\begin{aligned} \frac {dlnL(θ)}{dθ} &= \sum_{i=1}^{n} \frac{dlnf(x_i-\theta)}{d\theta} \\ &= \sum_{i=1}^{n} \frac{f'(x_i-\theta)}{f(x_i-\theta)} \end{aligned}
dθdlnL(θ)=i=1∑ndθdlnf(xi−θ)=i=1∑nf(xi−θ)f′(xi−θ)
为求极大似然估计,令
d
l
n
L
(
θ
)
d
θ
=
0
\frac {dlnL(θ)}{dθ}=0
dθdlnL(θ)=0
记
g
(
x
)
=
f
′
(
x
)
f
(
x
)
g(x)=\frac{f'(x)}{f(x)}
g(x)=f(x)f′(x),则有:
∑
i
=
1
n
g
(
x
i
−
θ
)
=
0
\sum_{i=1}^{n} g(x_i-\theta) = 0
i=1∑ng(xi−θ)=0
由于高斯假设极大似然估计的解就是算术平均
x
‾
\overline{x}
x,把解代入上式,可以得到:
∑
i
=
1
n
g
(
x
i
−
x
‾
)
=
0
(1)
\sum_{i=1}^{n} g(x_i-\overline{x}) = 0 \tag 1
i=1∑ng(xi−x)=0(1)
(
1
)
(1)
(1) 式中取
n
=
2
n=2
n=2, 有:
g
(
x
1
−
x
‾
)
+
g
(
x
2
−
x
‾
)
=
0
g(x_1−\overline{x})+g(x_2−\overline{x})=0
g(x1−x)+g(x2−x)=0
由于此时有
x
1
−
x
‾
=
−
(
x
2
−
x
‾
)
x_1−\overline{x}=−(x_2−\overline{x})
x1−x=−(x2−x), 并且
x
1
,
x
2
x_1,x_2
x1,x2是任意的,由此得到:
g
(
−
x
)
=
−
g
(
x
)
g(−x)=−g(x)
g(−x)=−g(x)
(
1
)
(1)
(1) 式中再取
n
=
m
+
1
n=m+1
n=m+1, 并且取
x
1
=
x
2
⋯
=
x
m
=
−
x
;
x
m
+
1
=
m
x
x_1=x_2⋯=x_m=−x; x_{m+1}=mx
x1=x2⋯=xm=−x;xm+1=mx, 则有
x
‾
=
0
\overline{x}=0
x=0, 并且:
∑
i
=
1
n
g
(
x
i
−
x
‾
)
=
m
g
(
−
x
)
+
g
(
m
x
)
=
0
\sum_{i=1}^{n}g(x_i-\overline{x}) = mg(-x)+g(mx) = 0
i=1∑ng(xi−x)=mg(−x)+g(mx)=0
即有:
g
(
m
x
)
=
m
g
(
x
)
g(mx)=mg(x)
g(mx)=mg(x)
而满足上式的唯一的连续函数就是
g
(
x
)
=
c
x
g(x)=cx
g(x)=cx, 从而进一步可以求解出:
f
(
x
)
=
M
e
c
x
2
f(x)=Me^{cx2}
f(x)=Mecx2
由于
f
(
x
)
f(x)
f(x)是概率密度函数,把
f
(
x
)
f(x)
f(x)正规化一下就得到均值为0的正态分布密度函数
N
(
0
,
σ
2
)
N(0,σ^2)
N(0,σ2)。