Let $X,Y$ be two independent random variables, uniformly distributed on $(0,1)$. I have to compute $E(X\mid\min(X,Y)=t)$, but in very educational way, I mean, I have to explain it to my students and I would rather prefer do not use very advanced methods. Is it possible? I was trying to compute common distribution function and then take its density, but I think that density does not exist. So, is there any easy, FORMAL way?

  • 1
    $\begingroup$ What do your students know? It's pretty easy to calculate the joint distribution $(X,\min (X,Y))$ if your students know about general measures. It doesn't have density, though, as you say. $\endgroup$ – WoolierThanThou Feb 19 '20 at 10:09
  • $\begingroup$ Okay, assume that I computed joint distribution and it has no density. How to compute conditional expectation? $\endgroup$ – mwrooo Feb 19 '20 at 12:31
  • $\begingroup$ Well... you don't need a density, you just need a Markov kernel. If $(X,Y)\sim \mu$ and $\mu(A\times B)=\int_B \nu_y(A)\textrm{d} \gamma(y)$ for some appropriate measure $\gamma$ and family of measure $\nu_y$, then $\mathbb{E}(X|Y)=\int_{\mathbb{R}} x\textrm{d}\nu_Y$ by exactly the same argument that works in the density case. $\endgroup$ – WoolierThanThou Feb 19 '20 at 14:30
  • $\begingroup$ In your case, if we denote by $\lambda^2$ the 2-dimensional Lesbegue measure and $\lambda_{\Delta}$ the Lesbegue measure on the diagonal lifted from $\mathbb{R}$. Then your measure is $\mu=1_{x\geq y} 1_{0\leq x\leq 1} \textrm{d}\lambda^2+(1-x)\textrm{d}\lambda$, which does have a Markov kernel decomposition with $\nu_y= \frac{1}{2}\delta_y+\frac{1}{2(1-y)}\cdot 1_{[y,1]}\textrm{d}\lambda $. $\endgroup$ – WoolierThanThou Feb 19 '20 at 14:45
  • $\begingroup$ Okay, I agree, but the problem is how to find that $\gamma$ measure and the family of $\nu_y$? $\endgroup$ – mwrooo Feb 19 '20 at 14:46

You could try drawing a square something like this as the joint distribution is uniform on the square

enter image description here

The condition $\min(X,Y)=t$ corresponds to points on the orange and red lines (orange for $t=X \le Y$ and red for $t=Y \le X$

Conditioned on this, the probability $P(X=t \mid \min(X,Y)=t) =\frac12$ by symmetry, and if not then $X$ is uniformly distributed between $t$ and $1$ so with density $f(x)=\frac{1}{2(1-t)}\mathbf 1_{t \lt x \lt 1}$

This makes the conditional expected value $\mathbb E[X \mid \min(X,Y)=t ]=\frac12 t + \frac12\left(\frac{t +1}{2}\right) = \frac 34 t +\frac14$

This turns out to be formal, with the corresponding conditional CDF $$\mathbb P(X \le x \mid \min(X,Y)=t )=\left\{ \begin{matrix}0 & \text{ for } x\lt t \\\frac12 & \text{ for } x= t \\ \frac{x+1-2t}{2-2t} & \text{ for } t \lt x\le 1 \\ 1 & \text{ for } 1 \lt x \end{matrix} \right.$$

  • 1
    $\begingroup$ Very clear and transparent explanation! Thank you! $\endgroup$ – NCh Feb 19 '20 at 10:15
  • $\begingroup$ Thank You very much! $\endgroup$ – mwrooo Feb 19 '20 at 10:42
  • 1
    $\begingroup$ @StubbornAtom $P(X\le t \mid \min(X,Y)=t) =\frac12$ is correct, but since $P(X\lt t \mid \min(X,Y)=t) =0$ that would suggest to me $P(X= t \mid \min(X,Y)=t) =\frac12$ too $\endgroup$ – Henry Feb 19 '20 at 14:43
  • 1
    $\begingroup$ @StubbornAtom On the density, you have to stretch a probability of $\frac12$ uniformly over an interval of length $1-t$, which I would have thought would make the density $\frac{1/2}{1-t} =\frac{1}{2(1-t)}$ $\endgroup$ – Henry Feb 19 '20 at 14:46
  • 1
    $\begingroup$ @mwrooo - I would guess $\mathbb E[\min(X,Y)\mid \max(X,Y) =s] = \int\limits_{t=0}^s \frac{t \lambda e^{-t\lambda}}{1-e^{-s\lambda}}\, dt = \frac{1-(1+s\lambda)e^{-s \lambda}}{\lambda\left(1-e^{-s \lambda}\right)}$ $\endgroup$ – Henry Feb 20 '20 at 23:20

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.