贝叶斯估计
一、基础知识
1.1 常用分布函数
- X ~
B
(
α
,
β
)
B(\alpha, \beta)
B(α,β)：
f
(
x
)
=
y
α
−
1
(
1
−
y
)
β
−
1
f(x) = y^{\alpha - 1}(1-y)^{\beta - 1}
f(x)=yα−1(1−y)β−1
E ( x ) = α α + β E(x) = \frac{\alpha}{\alpha + \beta} E(x)=α+βα - X ~
Γ
(
α
,
β
)
\Gamma(\alpha, \beta)
Γ(α,β)：
f
(
x
)
=
β
α
Γ
(
α
)
x
α
−
1
e
−
β
x
,
x
>
0
,
α
>
0
,
β
>
0
f(x)=\frac{\beta^\alpha}{\Gamma(\alpha)}x^{\alpha-1}e^{-\beta x},x>0,\alpha >0, \beta >0
f(x)=Γ(α)βαxα−1e−βx,x>0,α>0,β>0
E ( x ) = α β , D ( x ) = α β 2 , E ( x 2 ) = α 2 + α β 2 E(x) = \frac{\alpha}{\beta}, D(x) = \frac{\alpha}{\beta^2}, E(x^2) = \frac{\alpha^2 + \alpha}{\beta^2} E(x)=βα,D(x)=β2α,E(x2)=β2α2+α - X ~
I
Γ
(
α
,
β
)
I \Gamma(\alpha, \beta)
IΓ(α,β)：
f
(
x
)
=
β
α
Γ
(
α
)
x
α
+
1
e
−
β
x
,
x
>
0
,
α
>
0
,
β
>
0
f(x)=\frac{\beta^\alpha}{\Gamma(\alpha)}x^{\alpha+1}e^{-\beta x},x>0,\alpha >0, \beta >0
f(x)=Γ(α)βαxα+1e−βx,x>0,α>0,β>0
E ( 1 x ) = α β , D ( 1 x ) = α β 2 , E ( 1 x 2 ) = α 2 + α β 2 E(\frac{1}{x}) = \frac{\alpha}{\beta}, D(\frac{1}{x}) = \frac{\alpha}{\beta^2}, E(\frac{1}{x^2}) = \frac{\alpha^2 + \alpha}{\beta^2} E(x1)=βα,D(x1)=β2α,E(x21)=β2α2+α
1.2 相关定义
-
损失函数 L ( θ , a ) L(\theta, a) L(θ,a) ：表示用 a 去估计 θ \theta θ 时所造成的的损失。（平方差损失函数、平方相对差损失函数、加权平方差损失函数、绝对差损失函数）
-
决策函数 d ( ξ 1 , ξ 2 , ⋅ ⋅ ⋅ , ξ n ) d(\xi_1, \xi_2, ··· , \xi_n) d(ξ1,ξ2,⋅⋅⋅,ξn) ：决策 a，它是一个随机变量，所对应的损失为 L [ θ , d ( ξ 1 , ξ 2 , ⋅ ⋅ ⋅ , ξ n ) ] L[\theta, d(\xi_1, \xi_2, ··· , \xi_n)] L[θ,d(ξ1,ξ2,⋅⋅⋅,ξn)]
-
风险函数 R ( θ , d ) R(\theta, d) R(θ,d)：是损失函数在参数为 θ \theta θ 时的数学期望，代表了使用 d d d 估计 θ \theta θ 时所造成的平均损失
R ( θ , d ) = E θ { L [ θ , d ( ξ 1 , ξ 2 , ⋅ ⋅ ⋅ , ξ n ) ] } R(\theta, d) = E_\theta \{ L[\theta, d(\xi_1, \xi_2, ··· , \xi_n)] \} R(θ,d)=Eθ{L[θ,d(ξ1,ξ2,⋅⋅⋅,ξn)]} -
最大风险最小化估计量、极小极大估计量 ：P66页，例2.5.1
-
贝叶斯风险函数 B ( d ) B(d) B(d)： B ( d ) B(d) B(d) 为决策函数 d ( ξ 1 , ξ 2 , ⋅ ⋅ ⋅ , ξ n ) d(\xi_1, \xi_2, ··· , \xi_n) d(ξ1,ξ2,⋅⋅⋅,ξn) 的贝叶斯风险函数
B ( d ) = E { L [ θ , d ( ξ 1 , ξ 2 , ⋅ ⋅ ⋅ , ξ n ) ] } = E { E ( L [ θ , d ( ξ 1 , ξ 2 , ⋅ ⋅ ⋅ , ξ n ) ] ∣ θ ) } B(d) = E \{ L[\theta, d(\xi_1, \xi_2, ··· , \xi_n)] \} = E \{ E(L[\theta, d(\xi_1, \xi_2, ··· , \xi_n)]| \theta) \} B(d)=E{L[θ,d(ξ1,ξ2,⋅⋅⋅,ξn)]}=E{E(L[θ,d(ξ1,ξ2,⋅⋅⋅,ξn)]∣θ)}
= E { E ( L [ θ , d ( ξ 1 , ξ 2 , ⋅ ⋅ ⋅ , ξ n ) ] ∣ ξ 1 , ξ 2 , ⋅ ⋅ ⋅ , ξ n ) } = E \{ E(L[\theta, d(\xi_1, \xi_2, ··· , \xi_n)]| \xi_1, \xi_2, ··· , \xi_n) \} =E{E(L[θ,d(ξ1,ξ2,⋅⋅⋅,ξn)]∣ξ1,ξ2,⋅⋅⋅,ξn)} -
贝叶斯估计量、贝叶斯解 d ~ ( ξ 1 , ξ 2 , ⋅ ⋅ ⋅ , ξ n ) \tilde{d}(\xi_1, \xi_2, ··· , \xi_n) d~(ξ1,ξ2,⋅⋅⋅,ξn)： B ( d ~ ) = m i n d ∈ G { B ( d ) } B(\tilde{d}) = min_{d \in G} \{B(d)\} B(d~)=mind∈G{B(d)}
-
核：如果函数 φ ( x ) \varphi(x) φ(x) 与函数 f ( x ) f(x) f(x) 只相差一个常熟因子，则称 φ ( x ) \varphi(x) φ(x) 为 f ( x ) f(x) f(x) 的核，记为
f ( x ) ∝ φ ( x ) f (x) \propto \varphi(x) f(x)∝φ(x)
二、所用公式
h ( y ∣ x 1 , ⋅ ⋅ ⋅ , x n ) = π ( y ) f ( x 1 , ⋅ ⋅ ⋅ , x n ∣ y ) ∫ − ∞ ∞ f ( x 1 , ⋅ ⋅ ⋅ , x n ∣ y ) d F θ ( y ) ∝ π ( y ) f ( x 1 , ⋅ ⋅ ⋅ , x n ∣ y ) h(y|x_1, ···, x_n) = \frac{\pi(y)f(x_1, ···, x_n | y )}{\int_{-\infty}^{\infty}f(x_1, ···, x_n | y)dF_\theta(y)} \propto \pi(y)f(x_1, ···, x_n | y ) h(y∣x1,⋅⋅⋅,xn)=∫−∞∞f(x1,⋅⋅⋅,xn∣y)dFθ(y)π(y)f(x1,⋅⋅⋅,xn∣y)∝π(y)f(x1,⋅⋅⋅,xn∣y)
1万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
