3
$\begingroup$

What I have done is use Cholesky decompostion to get $X = \rho X^{1} + \sqrt{1 - \rho^2}X^{2}$ and $Y = X^{1}$ where $X^{1}$ and $X^{2}$ are independent standard normal. Then $$ \begin{aligned} P(X>Y>0) &= P(X>Y,Y>0) \\ &= P(\rho X^{1} + \sqrt{1 - \rho^2}X^{2} > X^{1},X^{1} > 0)\\ &= P(\sqrt{1 - \rho^2}X^{2} > (1 - \rho)X^{1},X^{1}>0)\\ &= P(X^{2} > \sqrt{\frac{1-\rho}{1+\rho}}X^{1},X^{1} > 0) \quad \quad \quad \text{ if } \rho \neq 1,-1 \end{aligned} $$

Thus $P(X > Y >0) \in [\frac{1}{8},\frac{1}{4})$ when $\rho \in [0,1)$ and $P(X > Y >0) \in(0,\frac{1}{8})$ when $\rho \in (-1,0)$ and $P(X > Y >0) = 0 $ for $\rho = 1,-1$. Does that right? Could anyone use the graph of the coordinates that are not vertical to figure out this question?

$\endgroup$
7
  • 2
    $\begingroup$ I would like to suggest that simply looking at the plots in my post at stats.stackexchange.com/a/71303/919 will enable you to write down a correct simple formula. All you need to is reverse the transformations to find where the wedge $x\gt y\gt 0$ originated; the probability will equal the angle divided by $2\pi.$ $\endgroup$
    – whuber
    Commented Sep 13, 2023 at 17:07
  • $\begingroup$ Thank you! I will read it now. By the way, What do you think about what I have done using Cholesky? $\endgroup$
    – Zhihao Xu
    Commented Sep 13, 2023 at 17:34
  • 6
    $\begingroup$ According to Abramowitz and Stegun 26.3.19, p 937, $$P(X>0, Y>0) - \frac 14 + \frac{\arcsin \rho}{2\pi}$$ and so, because of the symmetry, $P(X>Y>0)$ should be half that number. $\endgroup$ Commented Sep 13, 2023 at 18:24
  • 4
    $\begingroup$ @whuber I guess Dilip means $P(X > 0, Y > 0) = P(X > 0, Y > 0, X > Y) + P(X > 0, Y > 0, Y > X)$, and the probabilities on the right hand side are identical to $P(X > Y > 0)$. $\endgroup$
    – Zhanxiong
    Commented Sep 13, 2023 at 20:53
  • 2
    $\begingroup$ @whuber Zhanxiong has already answered your question regarding symmetry. I apologize for the typo in my displayed equation; that $-$ should have been a $=$. Oh well, comments can’t be edited this late. $\endgroup$ Commented Sep 13, 2023 at 22:35

1 Answer 1

2
$\begingroup$

Given that the joint density of $(X, Y)$ is $f(x, y) = \frac{1}{2\pi\sqrt{1 - \rho^2}}\exp\left(-\frac{x^2 + y^2 - 2\rho xy}{2(1 - \rho^2)}\right)$, the probability of interest is \begin{align} P(X > Y > 0) = \iint_{(x, y): x > y > 0}\frac{1}{2\pi\sqrt{1 - \rho^2}}\exp\left(-\frac{x^2 + y^2 - 2\rho xy}{2(1 - \rho^2)}\right)dxdy. \tag{1}\label{1} \end{align} A natural way of simplifying the integral $\eqref{1}$ is through the polar coordinates $x = \rho\cos\theta, y = r\sin\theta$, under which $\eqref{1}$ becomes \begin{align} P(X > Y > 0) = \frac{1}{2\pi\sqrt{1 - \rho^2}}\int_0^{\pi/4}\left[\int_0^\infty r\exp\left(-\frac{1 - \rho\sin(2\theta)}{2(1 - \rho^2)}r^2\right)dr\right]d\theta. \tag{2}\label{2} \end{align} It is easy to show that inner integral of $\eqref{2}$ evaluates to $\frac{1 - \rho^2}{1 - \rho\sin(2\theta)}$, whence it suffices to evaluate the integral \begin{align} I = \int_0^{\pi/4}\frac{1}{1 - \rho\sin(2\theta)}d\theta. \tag{3}\label{3} \end{align} One way to evaluate $\eqref{3}$ is considering the transformation $\theta = \arctan(u)$ (inspired by the tangent half-angle formula), which gives \begin{align} I = \int_0^1 \frac{1}{1 + u^2 - 2\rho u}du = \frac{1}{\sqrt{1 - \rho^2}}\arctan\left(\sqrt{\frac{1 + \rho}{1 - \rho}}\right). \end{align} Therefore, \begin{align} P(X > Y > 0) = \frac{1}{2\pi}\arctan\left(\sqrt{\frac{1 + \rho}{1 - \rho}}\right). \end{align} This agrees with the answer given in @Dilip Sarwate's comment (to verify, try matching derivatives of these two expressions).


An alternative approach makes use of the representation in your post. Assume $\rho \in (-1, 1)$ (the case $\rho = \pm 1$ is trivial), from the expression (I rewrote $X^i$ as $X_i$ to avoid confusion with exponents of $X$, and denote $\sqrt{\frac{1 - \rho}{1 + \rho}}$ by $\tau$ to save some typing) $P(X > Y > 0) = P\left(X_2 > \tau X_1, X_1 > 0\right)$ that you have correctly derived and the condition that $X_1, X_2 \text{ i.i.d.} \sim N(0, 1)$, it follows that \begin{align} & P(X > Y > 0) = \int_0^\infty P\left(X_2 > \tau x\right)\phi(x)dx = \int_0^\infty \left(1 - \Phi\left(\tau x\right)\right)\phi(x)dx \\ =& \frac{1}{2} - \int_0^\infty\int_{-\infty}^{\tau x}\phi(y)\phi(x) dydx \\ =& \frac{1}{2} - \frac{1}{2\pi}\int_0^\infty\int_{-\infty}^{\tau x}\exp\left(-\frac{1}{2}(x^2 + y^2)\right) dydx, \tag{4}\label{4} \end{align} where $\phi(x) = \frac{1}{\sqrt{2\pi}}e^{-x^2/2}$ and $\Phi(x) = \int_{-\infty}^x \phi(y)dy$.

It turns out that applying polar coordinates $x = r\cos\theta, y = r\sin\theta$ to $\eqref{4}$ would make the integration more straightforward (than to $\eqref{1}$). Indeed, by observing that the integration region $\{(x, y): x > 0, y < \tau x\}$ in the Cartesian coordinate system is the integration region $\{(r, \theta): r > 0, \cos\theta > 0, \sin\theta < \tau\cos\theta\} = \{(r, \theta): r > 0, \theta \in (0, \arctan\tau) \cup (\frac{3\pi}{2}, 2\pi)\}$ in the polar coordinate system, the integral in $\eqref{4}$ under polar coordinates becomes: \begin{align} \int_0^{\arctan\tau}\int_0^\infty re^{-\frac{1}{2}r^2}drd\theta + \int_{3\pi/2}^{2\pi}\int_0^\infty re^{-\frac{1}{2}r^2}drd\theta = \arctan\tau + \frac{1}{2}\pi. \tag{5}\label{5} \end{align} Substituting $\eqref{5}$ back to $\eqref{4}$ yields \begin{align} P(X > Y > 0) = \frac{1}{4} - \frac{1}{2\pi}\arctan\tau = \frac{1}{4} - \frac{1}{2\pi}\arctan\left(\sqrt{\frac{1 - \rho}{1 + \rho}}\right). \tag{6}\label{6} \end{align} Since $\arctan(x^{-1}) = \frac{\pi}{2} - \arctan(x)$ when $x > 0$, $\eqref{6}$ agrees with the answer found in the first approach.

$\endgroup$
5
  • $\begingroup$ Thank you! What about $\rho = 1,-1$? $\endgroup$
    – Zhihao Xu
    Commented Sep 13, 2023 at 21:37
  • $\begingroup$ @ZhihaoXu with $\rho=-1$ you will get $0$ as $X=-Y$ with probability $1$. $\rho=1$ is a special case as $X=Y$ with probability $1$, so $P(X \ge Y \ge 0)=\frac12$ while $P(X > Y > 0)=0$ $\endgroup$
    – Henry
    Commented Sep 13, 2023 at 21:42
  • $\begingroup$ By the way, what I have done is plot $x^{1}$ is perpendicular to $x^{2}$ since they are independent in my work. Then see how is the region $X>Y>0$ change and compute the region's angle then divided by 2 $\pi$ . $\endgroup$
    – Zhihao Xu
    Commented Sep 13, 2023 at 21:42
  • $\begingroup$ Thank you guys! Appreciate that for helping me! $\endgroup$
    – Zhihao Xu
    Commented Sep 13, 2023 at 21:44
  • 3
    $\begingroup$ Note that $\arctan\left(\sqrt{\frac{1 + \rho}{1 - \rho}}\right) =\frac12 \arcsin(\rho) +\frac\pi4$ $\endgroup$
    – Henry
    Commented Sep 13, 2023 at 21:44

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.