Let $X$ and $Y$ be two independent standard Normal variables. Let $M := \max(X, Y)$ and $L := \min(X, Y)$. It is given that the covariance between $M$ and $L$ is given by $\text{Cov}(M, L) = 1 / \pi$ How do I find the variance of $M$ and $L$?

We first note that $M - L = |X - Y|$ and $M + L = X + Y$. Since $X$ and $Y$ are independent we have $X + Y \sim \mathcal{N}(0, 2)$. Therefore $\text{Var}(X + Y) = 2 = \text{Var}(M) + \text{Var}(L) + 2 \text{Cov}(M, L)$. Hence $\text{Var}(M) + \text{Var}(L) + \frac{2}{\pi} = 2$.

I could not find another equation involving $\text{Var}(M)$ and $\text{Var}(L)$ using $M - L = |X - Y|$, as I could not find the variance of $|X - Y|$. How do I continue from this point? Also,are there any other approaches which use the symmetry property of the standard normals?

  • 1
    $\begingroup$ Note that $$ \Pr(M\leq m)=\Pr(X\leq m,Y\leq m) $$ $\endgroup$ – Chamberlain Foncha May 1 '18 at 7:42
  • $\begingroup$ As $X-Y\sim\mathcal N(0,2)$, $|X-Y|$ has a half-normal density. It's mean and variance is easily found. $\endgroup$ – StubbornAtom May 1 '18 at 10:23
  • $\begingroup$ Since $M+L=X+Y$, the variance of $M+L$ is $2.$ By symmetry, $M$ and $L$ have the same variance (after all, the distribution one is the distribution of the negative of the other) and it is now easily found by solving the equation $\operatorname{Var}(M+L)=\operatorname{Var}(M)+\operatorname{Var}(L)+2\operatorname{Cov}(M,L).$ $\endgroup$ – whuber May 1 '18 at 17:44

As edited from the earlier version of the question, we have that $$\mathbb{E}[M]=\underbrace{\mathbb{E}[X]}_{=0}+\underbrace{\mathbb{E}[(Y-X)\mathbb{I}_{Y-X\ge 0}]}_{Y-X\sim{\cal N}(0,2)}=\frac{1}{\sqrt\pi}$$ implies $$\text{cov}(M,L)=\mathbb{E}[ML]-\mathbb{E}[M]\mathbb{E}[L]=\underbrace{\mathbb{E}[XY]}_{=0}+\underbrace{\mathbb{E}[M]^2}_{L\sim-M}=\frac{1}{\pi}$$ where $L\sim -M$ means that $L$ and $-M$ have the same (marginal) distribution since $$M\sim\max(X,Y)\sim\max(-X,-Y)=-\min(X,Y)\sim -L$$ Now, the derivation proposed in the second paragraph is that \begin{align} \text{var}(M+L)&=\text{var}(X+Y)\\&=\underbrace{\text{var}(M)+\text{var}(L)}_{\text{identical: }L\sim -M}+2\text{cov}(M,L)\\&=2\left\{\text{var}(M)+\text{cov}(M,L)\right\}\\ &=2\left\{\text{var}(M)+\frac{1}{\pi}\right\}=2\\ \end{align} Therefore $$\text{var}(M)=\dfrac{\pi-1}{\pi}$$ which somewhat surprisingly implies$$\mathbb{E}[M^2]=1=\mathbb{E}[X^2]$$ Except that$$2\mathbb{E}[M^2]=\mathbb{E}[M^2+L^2]=\mathbb{E}[X^2+Y^2]=2\mathbb{E}[X^2]=2$$explains the identity.

  • 2
    $\begingroup$ I think the part of this proof that is missing is that $var(M) = var(L)$. You may want to elaborate that point. $\endgroup$ – Greenparker May 1 '18 at 8:57

I don't think one needs to know what $\mathrm{Cov}(M,L)$ is to find the variances.

As $X$ and $Y$ are independent, $X-Y\sim\mathcal{N}(0,2)$, so that $E|X-Y|=\sqrt\frac{2}{\pi}\cdot \sqrt{2}=\frac{2}{\sqrt{\pi}}$.

So we have $E(M)=E(\frac{1}{2}(X+Y+|X-Y|))=\frac{1}{2}E(|X-Y|)=\frac{1}{\sqrt{\pi}}$.

Again, $2M=X+Y+|X-Y|=U+|V|$, say, where $U=X+Y,V=X-Y$

Then $E(UV)=E(X^2-Y^2)=1-1=0$, which implies $\mathrm{Cov}(U,V)=0$.

This in turn means that $U$ and $V$ are independent as $(U,V)$ is jointly normal.

So, $4E(M^2)=E(U+|V|)^2=E(U^2)+E(V^2)+2E(U)E(|V|)$


Finally we have $E(M^2)=1$ which means $\mathrm{Var}(M)=1-\frac{1}{\pi}$.

Try doing something similar to find $\mathrm{Var}(L)$.


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.