Let $X \sim U(0,\theta_1) \text{ and } Y \sim U(0,\theta_2)$ be independent RV's. We're interested in $\mathbb{P}[X-Y\le z]$.

I've done the convolution (or double integrals), denote $Z=X-Y$ \begin{align} \mathbb{P}[Z\le z] &= \int_{[x-y<z]} dF_X dF_Y \\ &=\int_0^{\theta_2} \int_0^{z+y} \frac{1}{\theta_1} dx \frac{1}{\theta_2} dy\\ &=\frac{z}{\theta_1}+\frac{\theta_2}{2\theta_1} \end{align}

is this correct?

for $\theta_1=\theta_2 = 1$ we're getting $\mathbb{P}[X-Y\le z] = z+\frac12 $, which intuitively makes sense.

I'm also curious about the joint pdf. We get the pdf by differentiating the cdf w.r.t. $z$. So $f_Z(z) = \frac{1}{\theta_1}$. Had we chosen the convolution the other way around $$\int_0^{\theta_1} \int_0^{z+x} \frac{1}{\theta_2} dy \frac{1}{\theta_2} dx $$ we would have obtained the joint pdf $f_Z(z) = \frac{1}{\theta_2}$. Isnt' that weird?


Assuming that $X$ and $Y$ are independent, the joint distribution is

$f(x,y) = \frac{1}{\theta_1 \theta_2}$ if $0 \le x \le \theta_1, 0 \le y \le \theta_2$ and $0$ otherwise.

The correct way to do this is to integrate $f(x,y)$ over the region $x-y\le z$. The easiest way to compute this integral is to do it geometrically. Compute the area of the intersection of the region $x-y\le z$ and the rectangle $[0, \theta_1] \times [0, \theta_2]$ and divide it by $\theta_1 \theta_2$.

You can also use the convolution to compute the distribution of $Z = X - Y$ and find $P\{Z \le z\}$ by integrating $\int_0^z f_Z(s) ds$.

  • $\begingroup$ I've just edited the question. $\endgroup$ – shimee May 29 '13 at 20:57
  • $\begingroup$ Both your integrals are incorrect because $z+y$ and $z+x$ can potentially be negative in your integration range. (Z can assume values between $-\theta_2$ and $\theta_1$) You have to split into cases depending on the value of $z$. $\endgroup$ – Narut Sereewattanawoot May 29 '13 at 21:08

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.