As the title says, I'm looking for the distribution of $X$ given that $X>Y$.


Assume that $X$ and $Y$ are independent $\mathrm{N}(0,1)$ rv's. Denote by $\phi$ and $\Phi$ the corresponding pdf and df, respectively. By definition, $$ P(X\leq x\mid X>Y) = \frac{P(X\leq x,X>Y)}{P(X>Y)} \ \, . $$ Since $X$ and $Y$ are iid, by symmetry we have $P(X>Y)=P(X<Y)$. But we know that $P(X>Y)+P(X<Y)=1$ (the probability of a tie is zero). Hence, $P(X>Y)=1/2$.

Now, $$ \begin{align*} P(X\leq x,X>Y) &= \mathrm{E}[P(X\leq x,X>Y\mid Y)] \quad \text{(total expectation)} \\ &= \int_{-\infty}^\infty P(X\leq x,X>Y\mid Y=y)\,\phi(y)\,dy \\ &= \int_{-\infty}^\infty P(X\leq x,X>y\mid Y=y)\,\phi(y)\,dy \quad \text{(use what you know)} \\ &= \int_{-\infty}^\infty P(y<X\leq x)\,\phi(y)\,dy \quad \text{(independence)} \\ &= \int_{-\infty}^\infty (\Phi(x)-\Phi(y)) I_{(-\infty,x]}(y)\,\phi(y)\,dy \\ &= \Phi^2(x) - \int_{-\infty}^x \Phi(y)\,\phi(y)\,dy \, . \\ \end{align*} $$ Therefore, $$ P(X\leq x\mid X>Y) = 2\left(\Phi^2(x) - \int_{-\infty}^x \Phi(y)\,\phi(y)\,dy\right) \, . $$ Integrating by parts, we have $$ \int_{-\infty}^x \Phi(y)\,\phi(y)\,dy = \left[\Phi(y)\Phi(y)\right]_{-\infty}^x - \int_{-\infty}^x \Phi(y)\,\phi(y)\,dy \, , $$ yielding $$ \int_{-\infty}^x \Phi(y)\,\phi(y)\,dy = \frac{\Phi^2(x)}{2} \, , $$ and finally $$ P(X\leq x\mid X>Y) = \Phi^2(x) \, . $$

  • 4
    $\begingroup$ Note that $\Pr(X\le x|X\gt Y)$ = $\Pr(\max(X,Y)\le x)$ = $\Pr(X\le x\text{ and }Y\le x)$ = $\Phi(x)^2$. This agrees with your result after you perform the final integration using the substitution $z=\Phi(y)$ (no Monte-Carlo needed). $\endgroup$ – whuber Apr 8 '14 at 18:54
  • $\begingroup$ Thanks! This is much simpler. How would you prove the (intuitively clear) first equality? $\endgroup$ – Zen Apr 8 '14 at 20:01
  • 2
    $\begingroup$ You need (a) exchangeability of $X$ and $Y$ and (b) that $\Pr(X=Y)=0$. It follows immediately that $\Pr(X\gt Y)=\Pr(Y\gt X)=1/2$. A total probability law then gives $\Pr(\max(X,Y)\le x) \\= \Pr(X\le x|X\gt Y)\Pr(X\gt Y)+\Pr(Y\le x|Y\gt X)\Pr(Y\gt X)+\Pr(X\le x|X=Y)\Pr(X=Y)\\ = \Pr(X\le x|X\gt Y)(1/2+1/2)+0\\ = \Pr(X\le x|X\gt Y).\\$ All of this is essentially already in the first few lines of your answer. The remainder (about eight lines) therefore seems to be just a particularly cumbersome way to derive the distribution of the maximum of two continuous iid variables. $\endgroup$ – whuber Apr 8 '14 at 20:11

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.