贝叶斯课后习题（一）先验分布与后验分布

先验分布与后验分布

1. 先验概率可以写为：$\pi (\theta )\{\begin{array}{l}0.1,0.7\\ 0.2,0.3\end{array}$

   x服从二项分布$B(8,\theta )$

   即$p(x|\theta )=C_8^x{\theta }^x(1-\theta {)}^{n-x}=\{\begin{array}{l}{C}_8^{x}0.1^x(1-0.1{)}^{8-x}\\ C_8^{x}0.2^x(1-0.2{)}^{8-x}\end{array}$

   可有联合分布函数为$h(x,\theta )=p(x|\theta )\pi (\theta )=\{\begin{array}{l}{C}_8^{x}0.1^x(1-0.1{)}^{8-x}0.7\\ C_8^{x}0.2^x(1-0.2{)}^{8-x}0.3\end{array}$

   可有x的边缘分布为$m(x)=C_8^{x}0.1^x(1-0.1{)}^{8-x}0.7+C_8^{x}0.2^x(1-0.2{)}^{8-x}0.3$

   可有后验分布为$\pi (\theta |x)=\frac{h(x,\theta )}{m(x)}=\{\begin{array}{l}\frac{C_8^{x}0.1^x(1-0.1{)}^{8-x}0.7}{C_8^{x}0.1^x(1-0.1{)}^{8-x}0.7+C_8^{x}0.2^x(1-0.2{)}^{8-x}0.3}\\ \frac{C_8^{x}0.2^x(1-0.2{)}^{8-x}0.3}{C_8^{x}0.1^x(1-0.1{)}^{8-x}0.7+C_8^{x}0.2^x(1-0.2{)}^{8-x}0.3}\end{array}$

   带入x=3可求得$\pi (\theta |x)=\{\begin{array}{l}0.1,0.5418274451\\ 0.2,0.4581725549\end{array}$

2. 先验分布概率可以有：$\pi (\theta )=\{\begin{array}{l}1.0,0.4\\ 1.5,0.6\end{array}$

   x服从泊松分布$p(x|\theta )=\frac{{\theta }^x{e}^{-\theta }}{\Gamma (x+1)}=\{\begin{array}{l}\frac{e^{-1}}{\Gamma (x+1)}\\ \frac{1.5^x{e}^{-1.5}}{\Gamma (x+1)}\end{array}$

   有联合概率函数$h(x,\theta )=p(x|\theta )\pi (\theta )=\{\begin{array}{l}\frac{e^{-1}}{\Gamma (x+1)}0.4\\ \frac{1.5^x{e}^{-1.5}}{\Gamma (x+1)}0.6\end{array}$

   有x的边缘分布$m(x)=\frac{e^{-1}}{\Gamma (x+1)}0.4+\frac{1.5^x{e}^{-1.5}}{\Gamma (x+1)}0.6$

   可有后验分布$\pi (\theta |x)=\frac{h(x,\theta )}{m(x)}=\{\begin{array}{l}\frac{\frac{e^{-1}}{\Gamma (x+1)}0.4}{\frac{e^{-1}}{\Gamma (x+1)}0.4+\frac{1.5^x{e}^{-1.5}}{\Gamma (x+1)}0.6}\\ \frac{\frac{1.5^x{e}^{-1.5}}{\Gamma (x+1)}0.6}{\frac{e^{-1}}{\Gamma (x+1)}0.4+\frac{1.5^x{e}^{-1.5}}{\Gamma (x+1)}0.6}\end{array}$

   带入x=3，可有$\pi (\theta |x)=\{\begin{array}{l}0.2456663555\\ 0.7543336445\end{array}$

3. 先求先验分布

   1. $\pi (\theta )=\{\begin{array}{ll}1& 0<\theta <1\\ 0& other\end{array}$

      x是满足二项分布的，即$p(x|\theta )=C_8^x{\theta }^x(1-\theta {)}^{8-x}$

      可有联合密度函数$h(x,\theta )=p(x|\theta )\pi (\theta )=C_8^x{\theta }^x(1-\theta {)}^{n-x},0<\theta <1$

      可有x的边缘分布$m(x)={\int }_0^{1}h(x,\theta )d\theta =C_8^x{\int }_0^1{\theta }^x(1-\theta {)}^{n-x}d\theta$

      可有 后验分布 $\pi (\theta |x)=\frac{h(x|\theta )}{m(x)}=\frac{{\theta }^x(1-\theta {)}^{n-x}}{{\int }_0^1{\theta }^x(1-\theta {)}^{n-x}d\theta }=\frac{{\theta }^x(1-\theta {)}^{n-x}}{\frac{\Gamma (x+1)\Gamma (8-x+1)}{\Gamma (x+1+8-x+1)}}=\frac{\Gamma (10){\theta }^x(1-\theta {)}^{8-x}}{\Gamma (x+1)\Gamma (9-x)}$

      带入x=3可有$\pi (\theta |x=3)=\frac{{\theta }^3(1-\theta {)}^{8-3}}{{\int }_0^1{\theta }^3(1-\theta {)}^{8-3}d\theta }=\frac{\Gamma (10){\theta }^3(1-\theta {)}^{8-3}}{\Gamma (3+1)\Gamma (9-3)}=504{\theta }^3(1-\theta {)}^5$

   2. $\pi (\theta )=\{\begin{array}{ll}2(1-\theta )& 0<\theta <1\\ 0& other\end{array}$

      x是满足二项分布的，即$p(x|\theta )=C_8^x{\theta }^x(1-\theta {)}^{8-x}$

      可有联合密度函数$h(x,\theta )=p(x|\theta )\pi (\theta )=C_8^x{\theta }^x(1-\theta {)}^{n-x}2(1-\theta ),0<\theta <1$

      可有x的边缘分布$m(x)={\int }_0^{1}h(x,\theta )d\theta =2C_8^x{\int }_0^1{\theta }^x(1-\theta {)}^{n-x+1}d\theta$

      可有后验分布$\pi (\theta |x)=\frac{h(x|\theta )}{m(x)}=\frac{{\theta }^x(1-\theta {)}^{n-x+1}}{{\int }_0^1{\theta }^x(1-\theta {)}^{n-x+1}d\theta }=\frac{{\theta }^x(1-\theta {)}^{n-x+1}}{\frac{\Gamma (x+1)\Gamma (8-x+1+1)}{\Gamma (x+1+8-x+1+1)}}=\frac{\Gamma (11){\theta }^x(1-\theta {)}^{9-x}}{\Gamma (x+1)\Gamma (10-x)}$

      带入x=3可有$\pi (\theta |x=3)=\frac{{\theta }^3(1-\theta {)}^{n-3+1}}{\frac{\Gamma (3+1)\Gamma (8-3+1+1)}{\Gamma (3+1+8-3+1+1)}}=\frac{\Gamma (11){\theta }^3(1-\theta {)}^{9-3}}{\Gamma (3+1)\Gamma (10-3)}=840{\theta }^3(1-\theta {)}^6$

4. 由$\pi (\theta |x)\propto p(\theta |x)\pi (\theta )$显然易得

5. 首先求先验分布$\pi (\theta )=\{\begin{array}{ll}\frac{1}{10}& (10,20)\\ 0& other\end{array}$

   1. x满足$p(x|\theta )=\{\begin{array}{ll}1& (\theta -\frac{1}{2},\theta +\frac{1}{2})\\ 0& other\end{array}$

      可有联合概率函数$h(x,\theta )=p(x|\theta )\pi (\theta )=\frac{1}{10},11.5<\theta <12.5$

      可有x的边缘分布$m(x)={\int }_{11.5}^{12.5}\frac{1}{10}d\theta =0.1$

      可有后验分布$\pi (\theta |x)=1,11.5<\theta <12.5$

   2. x满足$p(x|\theta )=\{\begin{array}{ll}1& (\theta -\frac{1}{2},\theta +\frac{1}{2})\\ 0& other\end{array}$

      可有联合概率函数$h(x,\theta )=p(x|\theta )\pi (\theta )=\frac{1}{10},11.5<\theta <11.6$。这里是$\theta$的范围取交集，即下限的上限，上限的下限

      可有x的边缘分布$m(x)={\int }_{11.5}^{11.6}\frac{1}{10}d\theta =0.01$

      可有后验分布$\pi (\theta |x)=10,11.5<\theta <11.6$

6. 先验分布$\pi (\lambda )\propto {\lambda }^{\alpha -1}{e}^{-\beta \lambda }$

   x的概率函数$p(x|\lambda )\propto {\lambda }^x{e}^{-\lambda }$

   $\lambda$的后验分布$\pi (\lambda |x)\propto p(x|\lambda )\pi (\lambda )\propto {\lambda }^x{e}^{-\lambda }{\lambda }^{\alpha -1}{e}^{-\beta \lambda }\propto {\lambda }^{x+\alpha -1}{e}^{-(\beta +1)\lambda }$

   可以看到，这是一个伽马分布$Ga(x+\alpha ,\beta +1)$

7. 先求先验

   1. $\pi (\theta )=\{\begin{array}{ll}1& (0,1)\\ 0& other\end{array}$

      再求x的概率函数$p(x|\theta )=\frac{2x}{{\theta }^2}$

      可有联合概率函数$h(x,\theta )=p(x|\theta )\pi (\theta )=\frac{2x}{{\theta }^2},x<\theta <1$

      可有x的边缘分布$m(x)={\int }_x^{1}h(x,\theta )d\theta ={\int }_x^1\frac{2}{{\theta }^2}d\theta =2(1-x)$

      可有后验分布$\pi (\theta |x)=\frac{h(x,\theta )}{m(x)}=\frac{x}{1-x}{\theta }^{-2},x<\theta <1$

   2. $\pi (\theta )=\{\begin{array}{ll}3{\theta }^2& 0<\theta <1\\ 0& other\end{array}$

      再求x的概率函数$p(x|\theta )=\frac{2x}{{\theta }^2}$

      可有联合概率函数$h(x,\theta )=p(x|\theta )\pi (\theta )=\frac{2x}{{\theta }^2}3{\theta }^2=6x,x<\theta <1$

      可有x的边缘分布$m(x)={\int }_x^{1}6xd\theta =6x(1-x)$

      可有后验分布$\pi (\theta |x)=\frac{h(x,\theta )}{m(x)}=\frac{6x}{6x(1-x)}=\frac{1}{1-x},x<\theta <1$

8. 先求先验分布$\pi (\theta )=\frac{\Gamma (202)}{\Gamma (2)\Gamma (200)}{\theta }^1(1-\theta {)}^{199}$

   再求x的概率函数，为二项分布$B(100,\theta ),p(x|\theta )=C_{100}^x{\theta }^x(1-\theta {)}^{100-x}$

   可有联合概率函数$h(x,\theta )=p(x|\theta )\pi (\theta )=C_{100}^x{\theta }^x(1-\theta {)}^{100-x}\frac{\Gamma (202)}{\Gamma (2)\Gamma (200)}{\theta }^1(1-\theta {)}^{199},0<\theta <1$

   可有x的边缘分布$$\begin{gathered} m(x)={\int }_0^{1}h(x,\theta )d\theta ={\int }_0^1{C}_{100}^x{\theta }^x(1-\theta {)}^{100-x}\frac{\Gamma (202)}{\Gamma (2)\Gamma (200)}{\theta }^1(1-\theta {)}^{199} \\ =C_{100}^x\frac{\Gamma (202)}{\Gamma (2)\Gamma (200)}{\int }_0^1{\theta }^x(1-\theta {)}^{100-x}{\theta }^1(1-\theta {)}^{199}d\theta \end{gathered}$$

   可有后验分布$$\begin{gathered} \pi (\theta |x)=\frac{h(x,\theta )}{m(x)}=\frac{C_{100}^x\frac{\Gamma (202)}{\Gamma (2)\Gamma (200)}{\theta }^x(1-\theta {)}^{100-x}{\theta }^1(1-\theta {)}^{199}}{C_{100}^x\frac{\Gamma (202)}{\Gamma (2)\Gamma (200)}{\int }_0^1{\theta }^x(1-\theta {)}^{100-x}{\theta }^1(1-\theta {)}^{199}d\theta } \\ =\frac{{\theta }^x(1-\theta {)}^{100-x}{\theta }^1(1-\theta {)}^{199}}{{\int }_0^1{\theta }^x(1-\theta {)}^{100-x}{\theta }^1(1-\theta {)}^{199}d\theta }=\frac{{\theta }^x(1-\theta {)}^{100-x}{\theta }^1(1-\theta {)}^{199}}{\frac{\Gamma (x+1+1)\Gamma (299-x+1)}{\Gamma (302)}} \end{gathered}$$

   带入x=3有后验分布$\pi (\theta |x=3)=\frac{{\theta }^4(1-\theta {)}^{296}}{\frac{4!296!}{302!}}$即$Be(5,297)$

9. 先求先验分布$N(172.72,2.54)$

   x的分布为$N(\theta ,5^2)$

   1. 有$\overline{x}-N(\theta ,\frac{5^2}{10})$

      有 p ( x ‾ ∣ θ ) ∝ ( 1 2 π 25 10 ) e x p { ( x ‾ − θ ) 2 5 } p(\overline{x}|\theta)\varpropto(\frac1{\sqrt{2\pi\frac{25}{10}}})exp\{\frac{(\overline{x}-\theta)^2}{5}\} p(x∣θ)∝(2π1025​ ​1​)exp{5(x−θ)2​}

      可有 π ( θ ) ∝ ( 1 2 π 2.54 ) e x p { ( θ − 172.72 ) 2 2 ∗ 2.54 } \pi(\theta)\varpropto(\frac1{\sqrt{2\pi2.54}})exp\{\frac{(\theta-172.72)^2}{2*2.54}\} π(θ)∝(2π2.54 ​1​)exp{2∗2.54(θ−172.72)2​}

      可有 π ( θ ∣ x ) ∝ p ( x ∣ θ ) π ( θ ) ∝ ( 1 2 π 2.5 ∗ 2.54 ) e x p { ( 176.53 − θ ) 2 5 + ( θ − 172.72 ) 2 2 ∗ 2.54 } \pi(\theta|x)\varpropto p(x|\theta)\pi(\theta)\\\varpropto(\frac1{\sqrt{2\pi2.5*2.54}})exp\{\frac{(176.53-\theta)^2}{5}+\frac{(\theta-172.72)^2}{2*2.54}\} π(θ∣x)∝p(x∣θ)π(θ)∝(2π2.5∗2.54 ​1​)exp{5(176.53−θ)2​+2∗2.54(θ−172.72)2​}

      可有$\theta |\overline{x}-N(174.64,1.26)$

   2. 想要在区间上的概率最大，就需要让这个区间的重点与分布的均值重合，所以区间是$[173.39,175.89]$

   3. 求积分即可，$p(\theta \in \Theta )={\int }_{173.39}^{175.89}\pi (\theta |x)d\theta$

10. 首先，有$\overline{x}-N(\theta ,\frac{1}{25})$

    后有$\pi (\theta )-N(\mu ,{\sigma }^2)$

    可有 π ( θ ) ∝ ( 1 σ ) e x p { ( θ − μ ) 2 2 σ 2 } \pi(\theta)\varpropto(\frac1{\sigma})exp\{\frac{(\theta-\mu)^2}{2\sigma^2}\} π(θ)∝(σ1​)exp{2σ2(θ−μ)2​}

    可有 p ( x ‾ ∣ θ ) = e x p { ( x ‾ − θ ) 2 2 / 25 } p(\overline{x}|\theta)=exp\{\frac{(\overline{x}-\theta)^2}{2/25}\} p(x∣θ)=exp{2/25(x−θ)2​}

    即 π ( θ ∣ x ‾ ) ∝ e x p { ( θ − μ ) 2 2 σ 2 + 25 ( x ‾ − θ ) 2 2 } ∝ e x p { ( θ − μ ) 2 + 25 σ 2 ( x ‾ − θ ) 2 2 σ 2 } \pi(\theta|\overline x)\varpropto exp\{\frac{(\theta-\mu)^2}{2\sigma^2}+\frac{25(\overline{x}-\theta)^2}{2}\}\\\varpropto exp\{\frac{(\theta-\mu)^2+25\sigma^2(\overline x-\theta)^2}{2\sigma^2}\} π(θ∣x)∝exp{2σ2(θ−μ)2​+225(x−θ)2​}∝exp{2σ2(θ−μ)2+25σ2(x−θ)2​}

    转为正态分布的形式，可有 π ( θ ∣ x ‾ ) ∝ e x p { ( 25 σ 2 + 1 ) θ 2 − ( 50 σ 2 x ‾ + 2 μ ) θ + μ 2 + 25 σ 2 x ‾ 2 2 σ 2 } \pi(\theta|\overline x)\varpropto exp\{\frac{(25\sigma^2+1)\theta^2-(50\sigma^2\overline{x}+2\mu)\theta+\mu^2+25\sigma^2\overline x^2}{2\sigma^2}\} π(θ∣x)∝exp{2σ2(25σ2+1)θ2−(50σ2x+2μ)θ+μ2+25σ2x2​}

    可有，后验分布的方差为$\frac{{\sigma }^2}{25{\sigma }^2+1}$，可化简为$\frac{1}{25+1/{\sigma }^2}$，其中${\sigma }^2$是先验分布的方差

    可以看到，后验方差存在上界$\frac{1}{25}$，故后验标准差必然存在上界$\frac{1}{5}$

11. 先求先验分布$\pi (\theta )=\{\begin{array}{ll}192/{\theta }^4& \theta \ge 4\\ 0& \theta <0\end{array}$

    根据样本可有$p(x|\theta )=\frac{1}{{\theta }^3},0<x<\theta$

    可有联合概率函数$h(x,\theta )=192/{\theta }^7,0<x<\theta ,\theta >8$

    可有x的边缘分布$m(x)={\int }_8^{\infty }192/{\theta }^{7}d\theta$

    可有后验分布$\pi (\theta |x)=\frac{h(x,\theta )}{m(x)}=\frac{8192}{{\theta }^7}$

12. 首先有先验分布$\pi (\theta )\propto \frac{\alpha {\theta }_0^{\alpha }}{{\theta }^{\alpha -1}},\theta >{\theta }_0$

    后有x的概率函数$p(x|\theta )\propto \frac{1}{\theta }$

    $\pi (\theta |x)\propto p(x|\theta )\pi (\theta )\propto \frac{\alpha {\theta }_0^{\alpha }}{{\theta }^{\alpha -1}}\frac{1}{\theta },\theta >{\theta }_0\propto \frac{\alpha {\theta }_0^{\alpha }}{{\theta }^{\alpha }}$

    可以看到，后验分布仍然是pareto分布

13. 根据均值和方差可以求出，$\frac{\alpha }{\alpha +\beta }=1/3,\frac{\alpha \beta }{(\alpha +\beta {)}^2(\alpha +\beta +1)}=1/45$

    可有$\alpha =3,\beta =6$

    此时先验分布为$Be(3,6)$

14. 有样本的分布函数为$p(x|\lambda )={\lambda }^n{e}^{-\lambda {\sum }_{i=1}^n{x}_i}$

    1. 先验分布有$\pi (\lambda )\propto {\lambda }^{\alpha -1}{e}^{-\beta \lambda }$

       可有后验分布$$\begin{gathered} \pi (\lambda |x)\propto p(x|\lambda )\pi (\lambda )\propto {\lambda }^n{e}^{-\lambda {\sum }_{i=1}^n{x}_i}{\lambda }^{\alpha -1}{e}^{-\beta \lambda } \\ \propto {\lambda }^{n+\alpha -1}{e}^{-\lambda ({\sum }_{i=1}^n{x}_i+\beta )} \end{gathered}$$

       可以看到，仍然是伽马分布的形式

    2. 根据伽马分布可有$\frac{\alpha }{\beta }=0.0002,\frac{\alpha }{{\beta }^2}=0.0001$

       可解得$\alpha =0.0004,\beta =2$

15. 有$x-N({\theta }_1,\frac{1}{2{\theta }_2})$
    p ( x ∣ θ ) = ( 1 2 π ∗ 1 2 θ 2 ) e x p { − ( x − θ 1 ) 2 1 θ 2 } p(x|\theta)=(\frac1{\sqrt{2\pi*\frac1{2\theta_2}}})exp\{-\frac{(x-\theta_1)^2}{\frac1{\theta2}}\} p(x∣θ)=(2π∗2θ2​1​ ​1​)exp{−θ21​(x−θ1​)2​}
    可有 p ( x ‾ ∣ θ 1 , θ 2 ) ∝ θ 2 n 2 e x p { − θ 2 ∑ i = 1 n ( x i − θ 1 ) 2 } p(\overline{x}|\theta_1,\theta_2)\varpropto \theta_2^{\frac n2}exp\{-\theta_2\sum_{i=1}^n(x_i-\theta_1)^2\} p(x∣θ1​,θ2​)∝θ22n​​exp{−θ2​∑i=1n​(xi​−θ1​)2}
    有${\theta }_1|{\theta }_2-N(0,\frac{1}{2{\theta }_2})$
    可有 π ( θ 1 , θ 2 ) = π ( θ 1 ∣ θ 2 ) π ( θ 2 ) ∝ ( 1 2 π 1 2 θ 2 ) e x p { − ( θ 1 ) 2 1 θ 2 } λ α Γ ( α ) θ 2 α − 1 e − λ θ 2 \pi(\theta_1,\theta_2)=\pi(\theta_1|\theta_2)\pi(\theta_2)\varpropto (\frac1{\sqrt{2\pi\frac1{2\theta_2}}})exp\{-\frac{(\theta_1)^2}{\frac1{\theta_2}}\}\frac{\lambda^{\alpha}}{\Gamma(\alpha)}\theta_2^{\alpha-1}e^{-\lambda\theta_2} π(θ1​,θ2​)=π(θ1​∣θ2​)π(θ2​)∝(2π2θ2​1​ ​1​)exp{−θ2​1​(θ1​)2​}Γ(α)λα​θ2α−1​e−λθ2​
    即 π ( θ 1 , θ 2 ) ∝ λ α Γ ( α ) θ 2 1 2 + α − 1 e x p { − θ 2 ( θ 1 2 + λ ) } \pi(\theta_1,\theta_2)\varpropto \frac{\lambda^{\alpha}}{\Gamma(\alpha)}\theta_2^{\frac12+\alpha-1}exp\{-\theta_2(\theta_1^2+\lambda)\} π(θ1​,θ2​)∝Γ(α)λα​θ221​+α−1​exp{−θ2​(θ12​+λ)}
    可有 π ( θ 1 , θ 2 ∣ x ) ∝ p ( x ‾ ∣ θ 1 , θ 2 ) π ( θ 1 , θ 2 ) ∝ θ 2 n 2 e x p { − θ 2 ∑ i = 1 n ( x i − θ 1 ) 2 } λ α Γ ( α ) θ 2 1 2 + α − 1 e x p { − θ 2 ( θ 1 2 + λ ) } ∝ θ 2 n + 1 2 + α − 1 e x p { − θ 2 ( ∑ i = 1 n x i 2 − 2 θ 1 ∑ i = 1 n x i + ( n + 1 ) θ 1 2 + λ ) } \pi(\theta_1,\theta_2|x)\varpropto p(\overline{x}|\theta_1,\theta_2)\pi(\theta_1,\theta_2)\\ \varpropto \theta_2^{\frac n2}exp\{-\theta_2\sum_{i=1}^n(x_i-\theta_1)^2\}\frac{\lambda^{\alpha}}{\Gamma(\alpha)}\theta_2^{\frac12+\alpha-1}exp\{-\theta_2(\theta_1^2+\lambda)\}\\\varpropto \theta_2^{\frac{n+1}2+\alpha-1}exp\{-\theta_2(\sum_{i=1}^nx_i^2-2\theta_1\sum_{i=1}^nx_i+(n+1)\theta_1^2+\lambda)\} π(θ1​,θ2​∣x)∝p(x∣θ1​,θ2​)π(θ1​,θ2​)∝θ22n​​exp{−θ2​∑i=1n​(xi​−θ1​)2}Γ(α)λα​θ221​+α−1​exp{−θ2​(θ12​+λ)}∝θ22n+1​+α−1​exp{−θ2​(∑i=1n​xi2​−2θ1​∑i=1n​xi​+(n+1)θ12​+λ)}

16. 比较麻烦，用核比较好
    有 p ( X ∣ θ ) = ∏ i = 1 n p ( x i ∣ θ ) = c n ( θ ) e x p { ϕ ( θ ) ∑ i = 1 n T ( x i ) } ∏ i = 1 n h ( x i ) p(X|\theta)=\prod_{i=1}^np(x_i|\theta)=c^n(\theta)exp\{\phi(\theta)\sum_{i=1}^nT(x_i)\}\prod_{i=1}^nh(x_i) p(X∣θ)=∏i=1n​p(xi​∣θ)=cn(θ)exp{ϕ(θ)∑i=1n​T(xi​)}∏i=1n​h(xi​)
    可有 π ( θ ∣ X ) ∝ p ( X ∣ θ ) π ( θ ) ∝ c n ( θ ) e x p { ϕ ( θ ) ∑ i = 1 n T ( x i ) } ∏ i = 1 n h ( x i ) c η ( θ ) e x p { ϕ ( θ ) ν } ∝ c n + η ( θ ) e x p { ( ∑ i = 1 n T ( x i ) ∏ i = 1 n h ( x i ) + ν ) ϕ ( θ ) } \pi(\theta|X)\varpropto p(X|\theta)\pi(\theta) \\\varpropto c^n(\theta)exp\{\phi(\theta)\sum_{i=1}^nT(x_i)\}\prod_{i=1}^nh(x_i)c^{\eta}(\theta)exp\{\phi(\theta)\nu\}\\\varpropto c^{n+\eta}(\theta)exp\{(\sum_{i=1}^nT(x_i)\prod_{i=1}^nh(x_i)+\nu)\phi(\theta)\} π(θ∣X)∝p(X∣θ)π(θ)∝cn(θ)exp{ϕ(θ)∑i=1n​T(xi​)}∏i=1n​h(xi​)cη(θ)exp{ϕ(θ)ν}∝cn+η(θ)exp{(∑i=1n​T(xi​)∏i=1n​h(xi​)+ν)ϕ(θ)}
    可以看到，这和先验分布是同形式的

17. 超纲啦。。。

18. 假设$\lambda$的先验分布为$\pi (\lambda )$
    令$X-P(\lambda )$
    即$p(x|\lambda )=\frac{{\lambda }^x{e}^{-\lambda }}{\Gamma (x+1)}$
    可有样本的分布函数为$p(\overline{x}|\lambda )={\prod }_{i=1}^{n}p(x_i|\lambda )=\frac{{\lambda }^{{\sum }_{i=1}^n{x}_i}{e}^{-n\lambda }}{{\prod }_{i=1}^n\Gamma (x_i)}$
    从而有$\pi (\lambda |\overline{x})\propto p(x|\lambda )\pi (\lambda ){\lambda }^{{\sum }_{i=1}^n{x}_i}{e}^{-n\lambda }\pi (\lambda )$
    可以看到，后验分布只与样本的和有关，故可以用样本的和${\sum }_{i=1}^n{x}_i$来代替样本，故可证为充分统计量