9
$\begingroup$

Assume that $X$ and $Y$ follow the bivariate normal distribution with correlation coeffcient $\rho > 0$, zero means and scale parameters equal to one. I am looking for an elegant way to compute

$$E\left[ \Phi \left(X \right) \Phi \left(Y \right) \right]$$

I can see that the result is

$$E\left[ \Phi \left(X \right) \Phi \left(Y \right) \right] = \frac{1}{4} + \frac{\arcsin\left(\rho/2\right)} {\left(2\pi\right)}$$

but it's not obvious how one gets there. The CDFs are still uniformly distibuted RVs but the dependence complicates things. Had it not been for the dependence we would only have the $\frac{1}{4}$ factor on the RHS but now the shape of the distribution has to be taken into account.

I have experimented with the Law of Iterated Expectations but haven't gone very far. I don't think a trigonometric transformation would be helpful either. I would therefore appreciate it if someone could give some hints on how to approach this.

Thank you.

Improve this question
$\endgroup$
3
  • 2
    $\begingroup$ Defining $\Phi$ is a good place to start ... $\endgroup$
    – wolfies
    Commented Jan 24, 2016 at 19:45
  • 2
    $\begingroup$ @wolfies I thought it was the universal symbol for the CDF of a standard normal variable... $\endgroup$
    – JohnK
    Commented Jan 24, 2016 at 19:46
  • 3
    $\begingroup$ Have to dash, but can't help but notice that your solution is almost identical to the standardised bivariate Normal orthant probability $$P(X<0,Y<0) = \frac{1}{4} + \frac{\arcsin\left(\rho\right)} {\left(2\pi\right)}$$ ... which may (or may not) help, as related integrals may be involved. $\endgroup$
    – wolfies
    Commented Jan 24, 2016 at 20:06

1 Answer 1

Reset to default
8
$\begingroup$

That it relates to the standardised bivariate Normal orthant probability pointed out by wolfies is due to the fact that $$\begin{align*} \mathbb{E}^{X,Y}\left[ \Phi \left(X \right) \Phi \left(Y \right) \right] &= \mathbb{E}^{X,Y} \overbrace{\left[\mathbb{E}^Z\{\mathbb{I}(Z\le X)\}\mathbb{E}^W\{\mathbb{I}(W\le Y)\}\right]}^{Z,W\sim\mathcal{N}(0,1)}\\ &=\mathbb{E}^{X,Y,Z,W} \left[\mathbb{I}(Z\le X) \mathbb{I}(W\le Y)\right]\\ &= \mathbb{E}^{X,Y,Z,W} \left[\mathbb{I}(Z-X\le 0) \mathbb{I}(W-Y\le 0)\right]\\ &= \mathbb{E}^{X_1,Y_1} \left[\mathbb{I}(X_1\le 0) \mathbb{I}(Y_1\le 0)\right]\\&= \mathbb{P}^{X_1,Y_1} \left[X_1\le 0,Y_1\le 0\right]\\&= \mathbb{P}^{X_1,Y_1} \left[X_1/\sqrt2\le 0,Y_1/\sqrt2\le 0\right]\\\end{align*} $$ where $(X_1,Y_1)$ is now a bivariate normal vector with correlation $\rho/2$: $$\mathbb{E} \left[X_1Y_1\right]=\mathbb{E} \left[(Z-X)(W-Y)\right]=\mathbb{E} \left[XY \right]=\rho$$ and $$\mathrm{var}(X_1)=\mathrm{var}(Y_1)=\mathrm{var}(Z)+\mathrm{var}(X)=2$$

Improve this answer
$\endgroup$
4
  • $\begingroup$ I think I got it. You move the expectations with respect to $Z$ and $W$ in front, because they are independent, yes? $\endgroup$
    – JohnK
    Commented Jan 24, 2016 at 20:55
  • $\begingroup$ Not really, I simply use a single expectation for all rv's involved, whether they are dependent or independent. The expectation is against the joint distribution over the four rv's, $X,Y,Z,W$. $\endgroup$
    – Xi'an
    Commented Jan 24, 2016 at 21:12
  • $\begingroup$ Sorry for being slow but isn't that like claiming that $E\left(X\right) E \left(Y\right) = E\left(XY\right)$? $\endgroup$
    – JohnK
    Commented Jan 24, 2016 at 21:22
  • $\begingroup$ No, I am using$$\mathbb{E}^X[\mathbb{E}^Z[h(X,Z)]]=\mathbb{E}^{X,Z}[h(X,Z)]$$ $\endgroup$
    – Xi'an
    Commented Jan 24, 2016 at 21:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.