3

I am thinking about this classical problem in statistical simulations. We want to apply the Rejection sampling to simulate a random vector $Y=(Y_1, Y_2)$ of a uniform distribution from a unit disc $D = \{ (X_1 , X_2) \in R^2: \sqrt{x^2_1 + x^2_2} ≤ 1\}$ such that $X = (X_1 , X_ 2)$ is random vector of a uniform distribution in the square $S = [−1, 1]^2$ and the joint density $$ f(y_1,y_2) = \frac{1}{\pi} \mathbb{1}_{D(y_1,y_2)}.$$

Many answers presented a very simple solution like the answer of Haitao Du and page 26 but I can't understand how did they applied the rejection sampling method.

In the rejection method, we accept a sample generally if $f(x) \leq C \times g(x)$. In the linked answers they simply do 

Repeat generate independent V1, V2  which distribute U(−1,1)  
Until (V1)^2 + (V2)^2 <= 1
Return (V1,V2).

So the code 

x=runif(1e4,-1,1)
y=runif(1e4,-1,1)

d=data.frame(x=x,y=y)
disc_sample=d[d$x^2+d$y^2<1,]
plot(disc_sample)

What I am can't understand how did they use the definition of Rejection sampling. Where did the define the constant $C$, the function $g(x)$ or even found the maximum of $f(x)$. I understand what did they do but I don't see how it is applying the definition of the Rejection sampling. Thanks for help

Sophie Allan
  • 31
  • 3

1 Answers1

0

The constant is irrelevant because all distributions are uniform. If you take the formal set-up, you can quickly see that you can simplify it to the procedure you describe.

The maximum of $f(x)$ is the same as the value throughout, because of the uniformity. Therefore, we don't need to do the comparison.

By the way, there is a mistake in the code. d$x^2+d$y^2<1should be (d$x^2+d$y^2)<1.

Kees Mulder
  • 1,414
  • 1
  • 10
  • 10