极大似然估计
问题背景
以高斯分布引出问题,高斯分布的重要性体现于:
- 1.根据中心极限定理,当样本量足够大的时候,任意分布的均值都趋近于一个高斯分布,高斯分布具有工程应用的普适性;
- 2.高斯分布是许多模型的基础,比如线性高斯模型(卡尔曼滤波器),高斯过程等;
假设有一组观测到的样本数据
X
=
(
x
1
,
x
2
,
.
.
.
,
x
N
)
X=(x_{1},x_{2},...,x_{N})
X=(x1,x2,...,xN),他们服从参数
θ
=
(
μ
,
σ
2
)
\theta=(\mu,\sigma^{2})
θ=(μ,σ2)的一元高斯分布,可以使用极大似然估计得到高斯分布的参数,首先回顾一元高斯分布概率密度函数的表达:
p
(
x
)
=
1
2
π
σ
2
e
x
p
(
−
(
x
−
μ
)
2
2
σ
2
)
p(x)=\frac{1}{\sqrt{2\pi\sigma^{2}}}exp(-\frac{(x-\mu)^{2}}{2\sigma^{2}})
p(x)=2πσ21exp(−2σ2(x−μ)2)
极大似然估计(Maximum Likelihood Estimation,简称mle)的本质是估计参数
θ
\theta
θ,使得所观测样本
X
X
X出现的概率最大;此处需要熟悉一种数学格式
p
(
x
∣
θ
)
p(x|\theta)
p(x∣θ),指的是明确了参数
θ
\theta
θ情况下,服从高斯分布的样本
x
x
x出现的概率,事实上,这个格式的写法和概率密度一致:
p
(
x
∣
θ
)
=
1
2
π
σ
2
e
x
p
(
−
(
x
−
μ
)
2
2
σ
2
)
p(x|\theta)=\frac{1}{\sqrt{2\pi\sigma^{2}}}exp(-\frac{(x-\mu)^{2}}{2\sigma^{2}})
p(x∣θ)=2πσ21exp(−2σ2(x−μ)2)
极大似然估计的计算方法
假设样本
X
=
(
x
1
,
x
2
,
.
.
.
,
x
N
)
X=(x_{1},x_{2},...,x_{N})
X=(x1,x2,...,xN)中每个样本
x
i
x_{i}
xi都是独立同分布的,即满足同一个高斯分布,且彼此间相互独立,则极大似然的优化目标可以写为:
m
a
x
θ
p
(
X
∣
θ
)
=
m
a
x
θ
∏
i
=
1
N
p
(
x
i
∣
θ
)
max_{\theta}p(X|\theta)=max_{\theta}\prod_{i=1}^{N}p(x_{i}|\theta)
maxθp(X∣θ)=maxθi=1∏Np(xi∣θ)
取对数可以将乘积转为求和:
l
o
g
(
p
(
X
∣
θ
)
)
=
∑
i
=
1
N
l
o
g
(
p
(
x
i
∣
θ
)
)
=
∑
i
=
1
N
l
o
g
(
1
2
π
σ
2
e
x
p
(
−
(
x
i
−
μ
)
2
2
σ
2
)
)
=
∑
i
=
1
N
[
l
o
g
1
2
π
+
l
o
g
1
σ
−
(
x
i
−
μ
)
2
2
σ
2
]
log(p(X|\theta))=\sum_{i=1}^{N}log(p(x_{i}|\theta))=\sum_{i=1}^{N}log(\frac{1}{\sqrt{2\pi\sigma^{2}}}exp(-\frac{(x_{i}-\mu)^{2}}{2\sigma^{2}}))=\sum_{i=1}^{N}[log\frac{1}{\sqrt{2\pi}}+log\frac{1}{\sigma}-\frac{(x_{i}-\mu)^{2}}{2\sigma^{2}}]
log(p(X∣θ))=i=1∑Nlog(p(xi∣θ))=i=1∑Nlog(2πσ21exp(−2σ2(xi−μ)2))=i=1∑N[log2π1+logσ1−2σ2(xi−μ)2]
对上式求偏导,解出偏导为0的根即得到参数的取值:
μ
m
l
e
=
1
N
∑
i
=
1
N
x
i
\mu_{mle}=\frac{1}{N}\sum_{i=1}^{N}x_{i}
μmle=N1i=1∑Nxi
可以看出,样本的均值就是高斯分布参数
μ
\mu
μ的极大似然估计值;
同样的方式得到:
σ
m
l
e
2
=
1
N
∑
i
=
1
N
(
x
i
−
μ
m
l
e
)
2
\sigma^{2}_{mle}=\frac{1}{N}\sum_{i=1}^{N}(x_{i}-\mu_{mle})^{2}
σmle2=N1i=1∑N(xi−μmle)2
参数估计的有偏性和无偏性
通过极大似然估计的参数是否与模型参数真实值存在差距,如何衡量它们是一个问题,所以引入无偏估计的概念:如果估计量的期望等于被估计量的真实值,则称估计值满足无偏性;
对于上面的高斯分布参数估计,进行无偏性的检验,先从均值的估计值考虑,计算
μ
m
l
e
\mu_{mle}
μmle的期望:
E
[
μ
m
l
e
]
=
E
[
1
N
∑
i
=
1
N
x
i
]
=
1
N
∑
i
=
1
N
E
[
x
i
]
=
1
N
∑
i
=
1
N
μ
=
μ
E[\mu_{mle}]=E[\frac{1}{N}\sum_{i=1}^{N}x_{i}]=\frac{1}{N}\sum_{i=1}^{N}E[x_{i}]=\frac{1}{N}\sum_{i=1}^{N}\mu=\mu
E[μmle]=E[N1i=1∑Nxi]=N1i=1∑NE[xi]=N1i=1∑Nμ=μ
可以得到,
E
[
μ
m
l
e
]
=
μ
E[\mu_{mle}]=\mu
E[μmle]=μ,即估计值的期望等于模型参数的真实值,因此,均值的极大似然估计
μ
m
l
e
\mu_{mle}
μmle是无偏估计;
然后,检验方差估计的无偏性,首先对表达式变形:
σ
m
l
e
2
=
1
N
∑
i
=
1
N
(
x
i
−
μ
m
l
e
)
2
=
1
N
∑
i
=
1
N
(
x
i
2
−
2
x
i
μ
m
l
e
+
μ
m
l
e
2
)
=
1
N
∑
i
=
1
N
x
i
2
+
1
N
∑
i
=
1
N
μ
m
l
e
2
−
2
μ
m
l
e
1
N
∑
i
=
1
N
x
i
\sigma^{2}_{mle}=\frac{1}{N}\sum_{i=1}^{N}(x_{i}-\mu_{mle})^{2}=\frac{1}{N}\sum_{i=1}^{N}(x_{i}^{2}-2x_{i}\mu_{mle}+\mu_{mle}^{2})=\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}+\frac{1}{N}\sum_{i=1}^{N}\mu_{mle}^{2}-2\mu_{mle}\frac{1}{N}\sum_{i=1}^{N}x_{i}
σmle2=N1i=1∑N(xi−μmle)2=N1i=1∑N(xi2−2xiμmle+μmle2)=N1i=1∑Nxi2+N1i=1∑Nμmle2−2μmleN1i=1∑Nxi
发现
1
N
∑
i
=
1
N
x
i
\frac{1}{N}\sum_{i=1}^{N}x_{i}
N1∑i=1Nxi就是
μ
m
l
e
\mu_{mle}
μmle,所以进行替换:
σ
m
l
e
2
=
1
N
∑
i
=
1
N
x
i
2
+
1
N
∑
i
=
1
N
μ
m
l
e
2
−
2
μ
m
l
e
1
N
∑
i
=
1
N
x
i
=
1
N
∑
i
=
1
N
x
i
2
+
μ
m
l
e
2
−
2
μ
m
l
e
2
=
1
N
∑
i
=
1
N
x
i
2
−
μ
m
l
e
2
\sigma_{mle}^{2}=\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}+\frac{1}{N}\sum_{i=1}^{N}\mu_{mle}^{2}-2\mu_{mle}\frac{1}{N}\sum_{i=1}^{N}x_{i}=\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}+\mu_{mle}^{2}-2\mu_{mle}^{2}=\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}-\mu_{mle}^{2}
σmle2=N1i=1∑Nxi2+N1i=1∑Nμmle2−2μmleN1i=1∑Nxi=N1i=1∑Nxi2+μmle2−2μmle2=N1i=1∑Nxi2−μmle2
因此,得到:
E
[
σ
m
l
e
2
]
=
E
[
1
N
∑
i
=
1
N
x
i
2
−
μ
m
l
e
2
]
=
E
[
1
N
∑
i
=
1
N
x
i
2
−
μ
2
−
(
μ
m
l
e
2
−
μ
2
)
]
=
E
[
1
N
∑
i
=
1
N
x
i
2
−
μ
2
]
−
E
[
μ
m
l
e
2
−
μ
2
]
E[\sigma_{mle}^{2}]=E[\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}-\mu_{mle}^{2}]=E[\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}-\mu^{2}-(\mu_{mle}^{2}-\mu^{2})]=E[\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}-\mu^{2}]-E[\mu_{mle}^{2}-\mu^{2}]
E[σmle2]=E[N1i=1∑Nxi2−μmle2]=E[N1i=1∑Nxi2−μ2−(μmle2−μ2)]=E[N1i=1∑Nxi2−μ2]−E[μmle2−μ2]
对于第一项
E
[
1
N
∑
i
=
1
N
x
i
2
−
μ
2
]
E[\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}-\mu^{2}]
E[N1∑i=1Nxi2−μ2]:
E
[
1
N
∑
i
=
1
N
x
i
2
−
μ
2
]
=
E
[
1
N
∑
i
=
1
N
x
i
2
−
1
N
∑
i
=
1
N
μ
2
]
=
1
N
E
[
∑
i
=
1
N
(
x
i
2
−
μ
2
)
]
=
1
N
∑
i
=
1
N
E
[
(
x
i
2
−
μ
2
)
]
E[\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}-\mu^{2}]=E[\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}-\frac{1}{N}\sum_{i=1}^{N}\mu^{2}]=\frac{1}{N}E[\sum_{i=1}^{N}(x_{i}^{2}-\mu^{2})]=\frac{1}{N}\sum_{i=1}^{N}E[(x_{i}^{2}-\mu^{2})]
E[N1i=1∑Nxi2−μ2]=E[N1i=1∑Nxi2−N1i=1∑Nμ2]=N1E[i=1∑N(xi2−μ2)]=N1i=1∑NE[(xi2−μ2)]
注意到:
E
[
(
x
i
2
−
μ
2
)
]
=
E
[
x
i
2
]
−
μ
2
=
E
[
x
i
2
]
−
E
[
x
i
]
2
=
v
a
r
(
x
i
)
=
σ
2
E[(x_{i}^{2}-\mu^{2})]=E[x_{i}^{2}]-\mu^{2}=E[x_{i}^{2}]-E[x_{i}]^{2}=var(x_{i})=\sigma^{2}
E[(xi2−μ2)]=E[xi2]−μ2=E[xi2]−E[xi]2=var(xi)=σ2
所以:
E
[
1
N
∑
i
=
1
N
x
i
2
−
μ
2
]
=
1
N
∑
i
=
1
N
E
[
(
x
i
2
−
μ
2
)
]
=
σ
2
E[\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}-\mu^{2}]=\frac{1}{N}\sum_{i=1}^{N}E[(x_{i}^{2}-\mu^{2})]=\sigma^{2}
E[N1i=1∑Nxi2−μ2]=N1i=1∑NE[(xi2−μ2)]=σ2
处理第二项
E
[
μ
m
l
e
2
−
μ
2
]
E[\mu_{mle}^{2}-\mu^{2}]
E[μmle2−μ2]:
E
[
μ
m
l
e
2
−
μ
2
]
=
E
[
μ
m
l
e
2
]
−
E
[
μ
2
]
=
E
[
μ
m
l
e
2
]
−
μ
2
E[\mu_{mle}^{2}-\mu^{2}]=E[\mu_{mle}^{2}]-E[\mu^{2}]=E[\mu_{mle}^{2}]-\mu^{2}
E[μmle2−μ2]=E[μmle2]−E[μ2]=E[μmle2]−μ2
之前已经证明,均值的极大似然估计是无偏的,因此
μ
=
E
[
μ
m
l
e
]
\mu=E[\mu_{mle}]
μ=E[μmle],因此可以替换得到:
E
[
μ
m
l
e
2
]
−
μ
2
=
E
[
μ
m
l
e
2
]
−
E
[
μ
m
l
e
]
2
=
v
a
r
(
μ
m
l
e
)
=
v
a
r
(
1
N
∑
i
=
1
N
x
i
)
=
1
N
σ
2
E[\mu_{mle}^{2}]-\mu^{2}=E[\mu_{mle}^{2}]-E[\mu_{mle}]^{2}=var(\mu_{mle})=var(\frac{1}{N}\sum_{i=1}^{N}x_{i})=\frac{1}{N}\sigma^{2}
E[μmle2]−μ2=E[μmle2]−E[μmle]2=var(μmle)=var(N1i=1∑Nxi)=N1σ2
合并结果:
E
[
σ
m
l
e
2
]
=
σ
2
−
1
N
σ
2
=
N
−
1
N
σ
2
E[\sigma_{mle}^{2}]=\sigma^{2}-\frac{1}{N}\sigma^{2}=\frac{N-1}{N}\sigma^{2}
E[σmle2]=σ2−N1σ2=NN−1σ2
方差的极大似然估计值的期望不等于真实值,所以是有偏的,为了变成无偏,需要进行修正:
σ
^
2
=
N
N
−
1
σ
m
l
e
2
=
1
N
−
1
∑
i
=
1
N
(
x
i
−
μ
m
l
e
)
2
\widehat{\sigma}^{2}=\frac{N}{N-1}\sigma_{mle}^{2}=\frac{1}{N-1}\sum_{i=1}^{N}(x_{i}-\mu_{mle})^{2}
σ
2=N−1Nσmle2=N−11i=1∑N(xi−μmle)2