卡方分布上侧α分位数的近似公式及其证明

命题: 当 $n>45$ 时有 ${\chi }_{\alpha }^2(n)\approx \frac{1}{2}(u_{\alpha }+\sqrt{2n-1}{)}^2$, 其中 ${\chi }_{\alpha }^2(n)$ 和 $u_{\alpha }$ 分别为 ${\chi }^2(n)$ 和 $N(0,1)$ 的上侧 $\alpha$ 分位数.

引理1: $\Gamma (s+1)=s\Gamma (s)$.

$proof$: $LHS={\int }_0^{\infty }{x}^s{e}^{-x}dx=-{\int }_0^{\infty }{x}^{s}d(e^{-x})=-x^s{e}^{-x}{\mid }_0^{\infty }+s{\int }_0^{\infty }{x}^{s-1}{e}^{-x}dx=RHS.$

引理2: $\ln \Gamma (s)$ 严格上凸.

$proof$: $\forall \lambda \in (0,1)$, 有
$$\begin{array}{rl}\Gamma (\lambda x+(1-\lambda )y)& ={\int }_0^{\infty }{t}^{\lambda x+(1-\lambda )y-1}{e}^{-t}dt\\ & ={\int }_0^{\infty }[t^{x-1}{e}^{-t}{]}^{\lambda }[t^{y-1}{e}^{-t}{]}^{1-\lambda }dt\\ & \stackrel{Holder}{\le }{\left[{\int }_0^{\infty }{t}^{x-1}{e}^{-t}dt\right]}^{\lambda }{\left[{\int }_0^{\infty }{t}^{y-1}{e}^{-t}dt\right]}^{1-\lambda }\\ & ={\Gamma }^{\lambda }(x){\Gamma }^{1-\lambda }(y)\end{array}$$ 由Holder不等式取等条件可知不等号严格成立, 进一步对不等式两边取对数即可知
$$\ln \Gamma (\lambda x+(1-\lambda )y)<\lambda \ln \Gamma (x)+(1-\lambda )\ln \Gamma (y)$$

Holder不等式及其证明详见 Holder不等式（赫尔德不等式）

引理3: $\underset{x\to +\infty }{\lim }\frac{\Gamma (x+s)}{(x+\delta {)}^s\Gamma (x)}=1,\forall \delta \in (-1,1)$.

$proof$: 由引理1, 仅需考虑$s\in (0,1]$的情况, 则由引理2可知
$$\Gamma (x+s)=\Gamma ((1-s)x+s(x+1))\le {\Gamma }^{1-s}(x){\Gamma }^s(x+1)=x^s\Gamma (x)$$ 可知$\frac{\Gamma (x+s)}{(x+\delta {)}^s\Gamma (x)}\le {\left(\frac{x}{s+\delta }\right)}^s\to 1,x\to \infty$, 同理有
$$\begin{array}{rl}x\Gamma (x)& =\Gamma (x+1)=\Gamma (s(x+s)+(1-s)(x+s+1))\\ & \le {\Gamma }^s(x+s){\Gamma }^{1-s}(x+s+1)=(x+s{)}^{1-s}\Gamma (x+s)\end{array}$$ 可知 $\frac{\Gamma (x+s)}{(x+\delta {)}^s\Gamma (x)}\ge {\left(\frac{x}{x+\delta }\right)}^s\cdot {\left(\frac{x}{x+s}\right)}^{1-s}\to 1,x\to \infty$. 由夹逼原理即可知引理3成立.

下面我们回到原命题的证明,

$Proof$: 设 $X\sim {\chi }^2(n)$, 则 $X$ 的概率密度函数为 $$p(x)=\frac{1}{2^{n/2}\Gamma (n/2)}{x}^{\frac{n}{2}-1}{e}^{-\frac{x}{2}},x>0$$
其中 $\Gamma (z)={\int }_0^{\infty }{x}^{z-1}{e}^{-x}dx$, 考虑$Y=\sqrt{2X}$, 则 $Y$ 的概率密度函数为 $$q(y)=p\left(\frac{y^2}{2}\right)y=\frac{1}{\Gamma (\frac{n}{2})}{\left(\frac{y}{2}\right)}^{n-1}{e}^{-(\frac{y}{2}{)}^2},y>0$$ 则
$$\begin{array}{rl}EY& ={\int }_0^{\infty }yq(y)dy\\ & ={\int }_0^{\infty }\frac{y}{\Gamma (\frac{n}{2})}{\left(\frac{y}{2}\right)}^{n-1}{e}^{-(\frac{y}{2}{)}^2}dy\\ & \stackrel{t=(\frac{y}{2}{)}^2}{=}\frac{2}{\Gamma (\frac{n}{2})}{\int }_0^{\infty }{t}^{\frac{n-1}{2}}{e}^{-t}dt\\ & =\frac{2\Gamma (\frac{n+1}{2})}{\Gamma (\frac{n}{2})}\\ & \underset{取\delta =-\frac{1}{4}}{\stackrel{引理3}{\to }}2(\frac{n}{2}-\frac{1}{4}{)}^{\frac{1}{2}}=\sqrt{2n-1},a.s.n\to \infty \end{array}$$ 同理
$$E{Y}^2=\frac{4\Gamma (\frac{n}{2}+1)}{\Gamma (\frac{n}{2})}=2n$$ 则 $\mathrm{Var}Y=E{Y}^2-(EY{)}^2=1$. 由 Lindeberg-Levy CLT可知
$$\frac{Y-\sqrt{2n-1}}{1}\stackrel{d}{\to }N(0,1),a.s.n\to \infty$$ 即有
$$\frac{\sqrt{2{\chi }^2(\alpha )}-\sqrt{2n-1}}{1}\approx u_{\alpha },a.s.n\to \infty (一般为n>45)$$
可知 $n$ 较大时有 ${\chi }_{\alpha }^2(n)\approx \frac{1}{2}(u_{\alpha }+\sqrt{2n-1}{)}^2$, 证毕!

证明过程全为本人所写，大家若对上述证明过程有疑问或者发现纰漏欢迎在评论区指正！