4.Random Variable__随机变量（3）混合分布

2.3 混合分布

2.3.1 高斯与伽马混合

$$X\sim \mathcal{N}\left(x|\mu ,\frac{{\sigma }^2}{\tau }\right),r\sim Ga\left(r|\frac{\tau }{2},\frac{\tau }{2}\right)$$
$$p\left(x|\mu ,\frac{{\sigma }^2}{r}\right)=\frac{r^{\frac{1}{2}}}{\sqrt{2\pi }\sigma }exp\left[-\frac{1}{2}\frac{r{\left(x-\mu \right)}^2}{{\sigma }^2}\right],p\left(r|\frac{\tau }{2},\frac{\tau }{2}\right)=\frac{{\left(\frac{\tau }{2}\right)}^{\frac{\tau }{2}}}{\Gamma \left(\frac{\tau }{2}\right)}{r}^{\frac{\tau }{2}-1}exp\left(-\frac{\tau }{2}r\right)$$
$$\begin{array}{rl}& {\int }_0^{+\infty }p\left(x|\mu ,\frac{{\sigma }^2}{r}\right)p\left(r|\frac{\tau }{2},\frac{\tau }{2}\right)dr\\ =& {\int }_0^{+\infty }\frac{{\left(\frac{\tau }{2}\right)}^{\frac{\tau }{2}}{r}^{\frac{\tau -1}{2}}}{\sqrt{2\pi }\Gamma \left(\frac{\tau }{2}\right)\sigma }exp\left\{-\frac{r}{2}\left[\frac{{\left(x-\mu \right)}^2}{{\sigma }^2}+\tau \right]\right\}dr\\ & \text{Let}B=\frac{{\left(\frac{\tau }{2}\right)}^{\frac{\tau }{2}}}{\sqrt{2\pi }\Gamma \left(\frac{\tau }{2}\right)\sigma },A=\frac{{\left(x-\mu \right)}^2}{{\sigma }^2}+\tau ,u=\frac{A}{2}r\Rightarrow dr=\frac{2}{A}dr\\ =& B{\left(\frac{2}{A}\right)}^{\frac{\tau +1}{2}}{\int }_0^{+\infty }{u}^{\frac{\tau -1}{2}}exp(-u)du\\ =& \frac{\Gamma \left(\frac{\tau +1}{2}\right)}{\Gamma \left(\frac{\tau }{2}\right)\sqrt{\tau \pi }\sigma }{\left[1+\frac{{\left(x-\mu \right)}^2}{\tau {\sigma }^2}\right]}^{-\left(\frac{\tau +1}{2}\right)}\end{array}$$
显然服从$T$分布，即正态分布与伽马分布的混合分布可得$T$分布，可将$Ga\left(r|\frac{\tau }{2},\frac{\tau }{2}\right)$视为$\vec{w}$（权重矢量）。

2.3.2 高斯与指数混合

$$X\sim \mathcal{N}\left(x|\mu ,r\right),r\sim Exp\left(r|2{\sigma }^2\right)$$
$$p(x|\mu ,r)=\frac{r^{\frac{1}{2}}}{\sqrt{2\pi }}exp\left[-\frac{1}{2}\frac{{\left(x-\mu \right)}^2}{r}\right],p\left(r|s{\sigma }^2\right)=\frac{1}{2{\sigma }^2}exp\left(-\frac{r}{2{\sigma }^2}\right)$$
$${\int }_0^{+\infty }p(x|\mu ,r)p\left(r|2{\sigma }^2\right)dr={\int }_0^{+\infty }\frac{r^{\frac{1}{2}}}{2\sqrt{2\pi }{\sigma }^2}exp\left\{-\frac{1}{2}\left[\frac{{\left(x-\mu \right)}^2}{r}+\frac{r}{{\sigma }^2}\right]\right\}dr$$
显然积分号内分布函数符合广义逆高斯分布，令$\alpha =1/{\sigma }^2$，$\beta ={\left(x-\mu \right)}^2$，则
$$\begin{array}{rl}& \{\begin{array}{l}\frac{\alpha }{\beta }=\frac{1}{{\sigma }^2{\left(x-\mu \right)}^2}\\ \alpha \beta =\frac{{\left(x-\mu \right)}^2}{{\sigma }^2}\end{array},p=\frac{1}{2}\Rightarrow \{\begin{array}{l}{\left(\frac{\alpha }{\beta }\right)}^{\frac{p}{2}}=\frac{1}{\sqrt{\sigma (x-\mu )}}\\ 2K_p\left(\sqrt{\alpha \beta }\right)=\sqrt{\frac{2\pi \sigma }{x-\mu }}exp\left(-\frac{|x-\mu |}{\sigma }\right)\end{array}\\ \Rightarrow & \frac{{\left(\frac{\alpha }{\beta }\right)}^{\frac{p}{2}}}{2K_p\left(\sqrt{\alpha \beta }\right)}=\frac{1}{\sqrt{2\pi }\sigma }exp\left(-\frac{|x-\mu |}{\sigma }\right)\triangleq A\\ & {\int }_0^{+\infty }p(x|\mu ,r)p\left(r|2{\sigma }^2\right)dr=\frac{1}{2\sigma }exp\left(-\frac{|x-\mu |}{\sigma }\right)\end{array}$$
可见高斯分布与指数分布混合服从拉普拉斯分布。

2.3.3 泊松与伽马分布

$$\begin{array}{rl}& X\sim Poisson(\lambda ),\lambda \sim Gamma\left(r,\frac{1-p}{p}\right),0<p<1,r>0\\ & P(X=k)=\frac{{\lambda }^k}{k!}exp(\lambda )=\frac{{\lambda }^k}{\Gamma (k+1)}exp(\lambda )\\ & p\left(\lambda |r,\frac{1-p}{p}\right)=\frac{(1-p{)}^r}{\Gamma (r)p^r}{\lambda }^{r-1}exp\left(\frac{p-1}{p}\lambda \right)\\ & {\int }_0^{+\infty }P(X=k)P\left(\lambda |r,\frac{1-p}{p}\right)d\lambda \\ =& \frac{(1-p{)}^r}{\Gamma (k+1)\Gamma (r)p^r}{\int }_0^{+\infty }{\lambda }^{k+r-1}exp\left(-\frac{\lambda }{p}\right)d\lambda \\ =& \frac{\Gamma (k+r)}{\Gamma (k)\Gamma (r)}{p}^k(1-p{)}^r\\ \sim & NB(k,p)\end{array}$$
可见，泊松分布与伽马分布混合可得负二项分布。