Suppose $X_1 \sim f_{X_1}(x_1)$, $X_2 \sim f_{X_2}(x_2)$ are random variables with known probability density function.

Is there any way to compute the probability density function of a bivariate function $g(x_1,x_2)$, assuming on specific finite support $x_1 \in [a_1,b_1]$ and $x_2 \in [a_2,b_2]$ without Monte Carlo sampling or any other sampling method? $X_1$ and $X_2$ are independent, and $g $ is "smooth".

I'm looking for a way to possibly implement the computation numerically for an arbitrary $g$.

  • $\begingroup$ what do you know of the relationship between $X_1,X_2$? E.g. if they are independent, $g(x_1, x_2) = f_{X_1}(x_1) \times f_{X_2}(x_2)$... $\endgroup$ – gt6989b Apr 21 '16 at 17:34
  • $\begingroup$ yes, assuming they are independent. thanks for pointing it out! $\endgroup$ – kensaii Apr 21 '16 at 17:34
  • $\begingroup$ why not condition on $X_2$? $\endgroup$ – gt6989b Apr 21 '16 at 17:40
  • $\begingroup$ I'm not so familiar with Bayesian rules conditioning... can you please elaborate? :) $\endgroup$ – kensaii Apr 21 '16 at 17:42


Let's rename your variables $X,Y$ for easier notation. So you are defining $$ Z = g(X,Y) $$ and asking what is the probability density function of $Z$. For simplicity, let's assume a particular (very simple) $g$, let's say $Z = X+Y$.

$Z=z$ means $X+Y=z$, so $X=z-Y$; let's condition on $Y=y$, then you end up with $$ f_Z(z) = \int_\mathbb{R} f_X(z-y) f_Y(y) dy. $$

So if $g(x,y)$ has a nice form allowing to solve $z = g(x,y)$ for one of the variables in terms of the other, you can use this technique to get the pdf...

  • $\begingroup$ thank you! this is also known as the convolution formula, right? what if $g$ is not a one-to-one transformation? In this case the Jacobian is 1 so it becomes quite straightforward... $\endgroup$ – kensaii Apr 22 '16 at 13:50
  • $\begingroup$ @kensaii indeed this is a convolution. If you cannot solve for one in terms of the other, this is not really usable. $\endgroup$ – gt6989b Apr 22 '16 at 16:17

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.