Suppose I want to simulate the value of the integral $\int_{0}^{1} \int_{2}^{3} 2xy \ dx dy$ using Monte Carlo methods.

So, now, I draw a random sample from $U_1,U_2,...,U_n$ from $U(0,1)$ and for each $i \in [1,n]$, I take $X_i =U_i+2$. So, now, I get a random sample of $X_i$'s from $U(2,3)$.

Again I take another sample $Y_1,Y_2,...,Y_n$ from $U(0,1)$.

So, by WLLN, I calculate the mean of $2X_iY_i$ defined over the space $[0,1] \times [2,3] $ to get the estimate.

My question is do we always need to draw another sample for $Y_1,Y_2,..,Y_n$?

That is, if I had taken $Y_i=U_i$ , i.e. the initial drawn sample what would go wrong?

I get it that if I don't draw a fresh sample, then $X_i, Y_i$'s are not independent. But do we at all require independence of $X_i,Y_i$ here?


First, let me point out [somewhat pedantly!] that one does not simulate the value of the integral since this is a fixed number.

Second, if using the same $U_i$'s for simulating $X_i$ and $Y_i$, as e.g. $Y_i=U_i$ and $X_i=U_i+1$, the functional dependence between these two rvs modifies the value of the integral $$\int_0^1\int_2^3 2xy\text{d}x\text{d}y=x^2|_0^1 y^2/2|_2^3=5/2$$ versus $$\int_0^1 2x(x+2)\text{d}x=2x^2/3+2x^2|^1_0=5/2=8/3$$ as can be checked by simulation

  > y=runif(1e7)
  > x=y+2
  > print(mean(2*x*y))
  [1] 2.667504 #8/3
  > x=runif(1e7)
  > print(mean(2*x*y))
  [1] 2.500612 #5/2

The simulation of the points $(X_i,Y_i)$'s in $(0,1)\times(2,3)$ need mimick a Uniform generation over the unit square. They may be dependent in $i$ but not between $X_i$ and $Y_i$ since this modifies the joint distribution. An alternative is to simulate directly $2X_iY_i$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.