Box–Muller Transform

0. Abstract

这几天看了那么多采样, 回过头来, 仍然不知道最基本的 Normal Distribution 怎么采样. 于是查看 Normal distribution - Wikipedia, 有很多种采样方法, 不过多数还是要 approximate, 或者使用 rejection method, 似乎只有 Box–Muller method 是精确的解析式方法.

但 Wikipedia 对其的介绍很简洁, 看不出个所以然来. 即使是 Box–Muller transform - Wikipedia, 也讲的一头雾水. 最终看到 B 站视频 <Box Muller 算法推导>, 才豁然开朗, 这里记录其推导过程.

1. Box Muller 算法推导

Box Muller 算法: 设 $U_1,U_2\sim Uniform(0,1)$, 且相互独立, 令 $\{\begin{array}{c}X=\sqrt{-2ln{U}_1}cos(2\pi U_2)\\ Y=\sqrt{-2ln{U}_1}sin(2\pi U_2)\end{array}$, 则 $X,Y\sim Normal(0,1)$, 且相互独立.

推导:
$$\begin{array}{rl}\{\begin{array}{c}X=\sqrt{-2ln{U}_1}cos(2\pi U_2)=Rcos\Theta \\ Y=\sqrt{-2ln{U}_1}sin(2\pi U_2)=Rsin\Theta \end{array}⟹& \{\begin{array}{c}R=\sqrt{-2ln{U}_1}\\ \Theta =2\pi U_2\end{array}\\ R^2=X^2+Y^2& \end{array}$$ 求 $R$ 的累积分布函数: $$\begin{array}{rl}{F}_R(r)& =P(R\le r)=P(\sqrt{-2ln{U}_1}\le r)=P(-2ln{U}_1\le r^2)=P(U_1\ge exp(-\frac{r^2}{2}))\\ & =1-exp(-\frac{r^2}{2})\\ f_R(r)& =r\ast exp(-\frac{r^2}{2})\end{array}$$ 由于 $U_1$ 和 $U_2$ 相互独立, 所以 $R$ 和 $\Theta$ 相互独立, 求两者联合概率分布: $$\begin{array}{r}{f}_{R,\Theta }(r,\theta )=f_R(r)f_{\Theta }(\theta )=\frac{r}{2\pi }exp(-\frac{r^2}{2})\end{array}$$ 继续求 $(X,Y)$ 的联合概率分布: $$\begin{array}{rl}{f}_{X,Y}(x,y)=& f_{R,\Theta }(r,\theta )\left|\frac{\partial (r,\theta )}{\partial (x,y)}\right|\\ =& \frac{r}{2\pi }exp(-\frac{r^2}{2})/\left|\frac{\partial (x,y)}{\partial (r,\theta )}\right|\\ =& \frac{r}{2\pi }exp(-\frac{r^2}{2})/\left|\begin{array}{cc}cos\theta & -rsin\theta \\ sin\theta & rcos\theta \end{array}\right|\\ =& \frac{1}{2\pi }exp(-\frac{r^2}{2})\\ =& \frac{1}{2\pi }exp(-\frac{x^2+y^2}{2})\\ =& \left[\frac{1}{\sqrt{2\pi }}exp(-\frac{x^2}{2})\right]\left[\frac{1}{\sqrt{2\pi }}exp(-\frac{y^2}{2})\right]\end{array}$$ 很明显(或者可以积分求边缘分布证明): $$\begin{array}{r}{f}_X(x)=\frac{1}{\sqrt{2\pi }}exp(-\frac{x^2}{2})\\ f_Y(y)=\frac{1}{\sqrt{2\pi }}exp(-\frac{y^2}{2})\end{array}$$ $X,Y\sim Normal(0,1)$, 且相互独立.