t分布k阶矩推导

一大盘洋芋

于 2023-05-19 16:25:08 发布

阅读量355

点赞数

文章标签：概率论

本文链接：https://blog.csdn.net/weixin_51362945/article/details/130764810

版权

$T=\frac{X}{Y/n}$

其中， $X\sim N(0,1)$ ， $Y\sim \chi^{2}(n)\sim Ga(\frac{n}{2},\frac{1}{2})$

$f(x)=\frac{1}{\sqrt{2\pi}}e^{-\frac{x^{2}}{2}},x\in(-\infty,\infty)$

$f(y)=\frac{\lambda^{\alpha}}{\Gamma(\alpha)}y^{\alpha-1}e^{-\lambda y},y> 0$

转换的雅可比式： $\left | \frac{\varphi (x,y)}{\varphi (t,y)} \right |=\begin{vmatrix} \sqrt{y/n} &\,\,\,\,\, \frac{T}{2\sqrt{yn}}\\ 0 & 1 \end{vmatrix}=\left |\sqrt{y/n} \right |$

$\begin{aligned}f_{T}(t)&=\int_{0}^{\infty}\sqrt{y/n}f_{X}(t\sqrt{y/n})f_{Y}(y)dy\\ &=\int_{0}^{\infty}\frac{1}{\sqrt{2n\pi}\Gamma(\frac{n}{2})2^{\frac{n}{2}}}y^{\frac{n-1}{2}}e^{-\frac{y}{2}(1+\frac{t^{2}}{n})}dy\\&u=\frac{y}{2}(1+\frac{t^{2}}{n}) \\&= \int_{0}^{\infty}\frac{1}{\sqrt{2n\pi}\Gamma(\frac{n}{2})2^{\frac{n}{2}}}u^{\frac{n+1}{2}-1}e^{-u}(\frac{1}{2}(1+\frac{t^{2}}{n})^{-\frac{n+1}{2}}du\\ &=\frac{\Gamma(\frac{n+1}{2})}{\Gamma(\frac{n}{2})\sqrt{n\pi}} (1+\frac{t^{2}}{n})^{-\frac{n+1}{2}}\end{aligned}$

T分布密度函数是偶函数，则奇数次阶矩为0，偶数次阶乘为：

$E(T^{2k})=2\int_{0}^{\infty}t^{2k}\frac{\Gamma(\frac{n+1}{2})}{\Gamma(\frac{n}{2})\sqrt{n\pi}} (1+\frac{t^{2}}{n})^{-\frac{n+1}{2}}dt$

上式中积分变换难以直接看出怎么解，先利用t分布和其他分布关系来求：

$E(T^{2k})=E\left(\frac{X^{2}}{Y/n} \right )^{k}=n^{k}E(X^{2k})E(Y^{-k})$

$\begin{aligned} E(X^{2k})&=2\int_{0}^{\infty}x^{2k}\frac{1}{\sqrt{2\pi}}e^{-\frac{x^{2}}{2}}dx\\ &t=\frac{x^{2}}{2}\\ &=\int_{0}^{\infty}\frac{2^{k}}{\sqrt{\pi}}t^{k-\frac{1}{2}}e^{-t}dt\\ &=\frac{2^{k}\Gamma(k+\frac{1}{2})}{\sqrt{\pi}}\\ &=\frac{2^{k}(k-\frac{1}{2})(k-\frac{3}{2})...\frac{1}{2}\Gamma(\frac{1}{2})}{\Gamma(\frac{1}{2})}\\ &=(2k-1)(2k-3)...1\\ &=(2k-1)!! \end{aligned}$

$\begin{aligned} E(Y^{-k})&=\int_{0}^{\infty}\frac{1}{\Gamma(\frac{n}{2})2^{\frac{n}{2}}}y^{\frac{n}{2}-k-1}e^{-\frac{1}{2}y}dy\\ &=\int_{0}^{\infty}\frac{2^{-k}}{\Gamma(\frac{n}{2})}\left(\frac{y}{2}\right)^{\frac{n}{2}-k-1}e^{-\frac{1}{2}y}d\frac{y}{2}\\ &=\frac{2^{-k}\Gamma(\frac{n}{2}-k)}{\Gamma(\frac{n}{2})} \end{aligned}$

所以，

$\begin{aligned} E(T^{2k})&=n^{k}\frac{2^{k}\Gamma(k+\frac{1}{2})}{\sqrt{\pi}}\frac{2^{-k}\Gamma(\frac{n}{2}-k)}{\Gamma(\frac{n}{2})}\\ &=\frac{n^{k}\Gamma(k+\frac{1}{2})\Gamma(\frac{n}{2}-k)}{\sqrt{\pi}\Gamma(\frac{n}{2})} \end{aligned}$

对比上式与密度函数积分的差别：

$\frac{\Gamma(k+\frac{1}{2})\Gamma(\frac{n}{2}-k)}{\Gamma(\frac{n}{2})}=\frac{\Gamma(\frac{n+1}{2})}{\Gamma(\frac{n}{2})}\cdot B(k+\frac{1}{2},\frac{n}{2}-k)$

所以证明下式成立即可：

$2\int_{0}^{\infty}t^{2k}\frac{n^{-k}}{\sqrt{n}} (1+\frac{t^{2}}{n})^{-\frac{n+1}{2}}dt=B(k+\frac{1}{2},\frac{n}{2}-k)$

寻找左边与贝塔函数相似的地方，

设 $1-m=1+\frac{t^{2}}{n}$ ，则 $m=-\frac{t^{2}}{n}$ ，m范围不是(0,1)

设 $\frac{1}{1-m}=1+\frac{t^{2}}{n}$ ，则 $m=\frac{t^{2}}{n+t^{2}}$ ，m的范围是(0,1)，尝试换元是否正确：

$\begin{aligned} &2\int_{0}^{\infty}t^{2k}\frac{n^{-k}}{\sqrt{n}} (1+\frac{t^{2}}{n})^{-\frac{n+1}{2}}dt\\ &t^{2}=n\frac{m}{1-m},n+t^{2}=\frac{n}{1-m},t=\sqrt{n(\frac{1}{1-m}-1}) \\ =&2\int_{0}^{1}(n\frac{m}{1-m})^{k}\frac{n^{-k}}{\sqrt{n}} (\frac{1}{1-m})^{-\frac{n+1}{2}}\frac{1}{2\sqrt{n\frac{m}{1-m}}}\frac{n}{(1-m)^{2}}dm\\ =&\int_{0}^{1}m^{k+\frac{1}{2}-1}(1-m)^{\frac{n}{2}-k-1}dm\\ =&B(k+\frac{1}{2},\frac{n}{2}-k) \end{aligned}$