不同方法推导Gamma分布可加性产生的矛盾
Gamma分布的概率密度函数表示如下:
\[X \backsim G(\alpha,\beta): f(x)=\frac{\beta^\alpha}{\Gamma(\alpha)}x^{\alpha-1}e^{-\beta x}
\]
其对应的矩母函数为
\[{\rm M}_x(t)=(1+\beta t)^{-\alpha}
\]
显然,若\(X_1 \backsim G(\alpha_1,\beta),X_2 \backsim G(\alpha_2,\beta)\),则\({\rm M}_{x_1}(t)=(1+t/\beta)^{-\alpha_1},{\rm M}_{x_2}(t)=(1+t/\beta)^{-\alpha_2}\),所以\({\rm M}_{x_1}(t){\rm M}_{x_2}(t)=(1+t/\beta)^{-(\alpha_1+\alpha_2)}\),所以\(X_1+X_2 \backsim G(\alpha_1+\alpha_2,\beta)\),这即为Gamma分布的可加性原则。
从矩母函数的角度出发,Gamma分布可加性原则是显而易见的。然而,本人在利用其它方法推导这一性质时,却出现了矛盾,一时难以发现端倪。现将推导过程展示如下(针对\(\alpha_1,\alpha_2\)为大于等于1的整数的情况):
由概率论的知识我们知道,两个随机变量的和的概率密度函数是各自概率密度函数的卷积,因此若\(X=X_1+X_2\),则
\[\begin{equation}
\begin{aligned}
f_x(x)&=f_{x_1}(x)*f_{x_2}(x)=\int_0^x f_{x_1}(y)f_{x_2}(x-y)dy\\
&=\frac{\beta^{\alpha_1+\alpha_2}}{\Gamma(\alpha_1)\Gamma(\alpha_2)}\int_0^x y^{\alpha_1-1}e^{-\beta y} (x-y)^{\alpha_2-1}e^{-\beta(x-y)}dy\\
&=\frac{\beta^{\alpha_1+\alpha_2}}{\Gamma(\alpha_1)\Gamma(\alpha_2)}e^{-\beta x}\int_0^x y^{\alpha_1-1} (x-y)^{\alpha_2-1}dy
\end{aligned}
\end{equation}\tag{1}
\]
针对(1)中最后一个等式的计算,采用不同方法出现了不同的结果:
方法1:用二项展开可以得到\((x-y)^{\alpha_2-1}=\sum_{i=0}^{\alpha_2-1} (-1)^i\begin{pmatrix}\alpha_2-1\\i\end{pmatrix}x^{\alpha_2-1-i} y^i\),将其代入(1)中可以得到
\[\begin{equation}
\begin{aligned}
f_x(x)&=\frac{\beta^{\alpha_1+\alpha_2}}{\Gamma(\alpha_1)\Gamma(\alpha_2)}e^{-\beta x}\sum_{i=0}^{\alpha_2-1}(-1)^i\begin{pmatrix}\alpha_2-1\\i \end{pmatrix}x^{\alpha_2-1-i} \int_0^x y^{\alpha_1+i-1}dy\\
&=\frac{\beta^{\alpha_1+\alpha_2}}{\Gamma(\alpha_1)\Gamma(\alpha_2)}[\sum_{i=0}^{\alpha_2-1}(-1)^i\begin{pmatrix}\alpha_2-1\\i \end{pmatrix}/(\alpha_1+i)]x^{\alpha_1+\alpha_2-1}e^{-\beta x}
\end{aligned}
\end{equation}\tag{2}
\]
显然,要想使命题得证,需要有\(\sum_{i=0}^{\alpha_2-1}(-1)^i\begin{pmatrix}\alpha_2-1\\i \end{pmatrix}/(\alpha_1+i)=\Beta(\alpha_1,\alpha_2)\),(其中\(B(\alpha_1,\alpha_2)\)表示beta函数)但是这个等式似乎并不成立(可以简单地用数值验证)。因此利用上述方法无法使命题得证。
方法2:令\(y=tx\),将(1)转化为对\(t\)的积分,可以得到
\[\begin{equation}
\begin{aligned}
f_x(x)&=\frac{\beta^{\alpha_1+\alpha_2}}{\Gamma(\alpha_1)\Gamma(\alpha_2)}[\int_0^1 t^{\alpha_1-1}(1-t)^{\alpha_2-1}dt]x^{\alpha_1+\alpha_2-1}e^{-\beta x}\\
&=\frac{\beta^{\alpha_1+\alpha_2}}{\Gamma(\alpha_1)\Gamma(\alpha_2)}\Beta(\alpha_1,\alpha_2)x^{\alpha_1+\alpha_2-1}e^{-\beta x}\\
&=\frac{\beta^{\alpha_1+\alpha_2}}{\Gamma(\alpha_1+\alpha_2)}x^{\alpha_1+\alpha_2-1}e^{-\beta x}
\end{aligned}
\end{equation}\tag{3}
\]
显然,命题得证。
暂时没发现为什么方法1没有能够使命题得证!!!
说明: 经过证明,确定上面两种推导方法得到的结果是一致的。具体地,我们要证明\(\sum_{i=0}^{\alpha_2-1}(-1)^i\begin{pmatrix}\alpha_2-1\\i \end{pmatrix}/(\alpha_1+i)=\Beta(\alpha_1,\alpha_2)\)这个等式是成立的,证明过程如下:
证明过程需要用到一个积分等式(该等式可从参考文献[1]中找到):
\[\int_0^\infty(1-e^{-x/\beta})^{\alpha-1}e^{-\mu x}dx=\beta \Beta(\beta\mu,\alpha),({\rm Re\beta>0,{\rm Re}\alpha>0,{\rm Re}\mu>0})\tag{4}
\]
根据(4),我们令\(\beta=1\),则有
\[\begin{equation}
\begin{aligned}
\int_0^\infty(1-e^{-x})^{\alpha-1}e^{-\mu x}dx&=\int_0^\infty \sum_{i=0}^{\alpha-1}(-1)^i \begin{pmatrix}\alpha-1\\i\end{pmatrix}e^{-(i+\mu)x}dx\\
&=\sum_{i=0}^{\alpha-1}(-1)^i \begin{pmatrix}\alpha-1\\i\end{pmatrix}\int_0^\infty e^{-(i+\mu)x}dx\\
&=\sum_{i=0}^{\alpha-1}(-1)^i \begin{pmatrix}\alpha-1\\i\end{pmatrix}/(i+\mu)\\
&=\Beta(\mu,\alpha)
\end{aligned}
\end{equation}\tag{5}
\]
至此,得证。
参考文献
[1] Gradshteyn, I. S. and Ryzhik, L.M. "Table of Integrals, Series, and Products (6th ed)", New York: Academic Press, 2000, pp. 331 and 899.