8

The question is mainly in the title:

Given two standard normal random variables with correlation $\rho$, what is the distribution of sign of their product?

I understand that when $\rho=0$, we have two iid standard normal random variables and therefore, they take positive and negative values independently with probability $\frac12$.

But I don't know what to do if $\rho\ne0$. We can take $\rho>0$ without loss of generality (because if $X$ and $Y$ are std normal with correlation $\rho$, then $-X$ and $Y$ are std normal with correlation $-\rho$). But I could not proceed further.

Martund
  • 431
  • 2
  • 8

3 Answers3

14

As long as $(X,Y)$ is standard bivariate normal with correlation $\rho$, the probability that $XY$ is positive or negative can be found using the well-known result for the positive quadrant probability $$P(X>0,Y>0)=\frac14+\frac1{2\pi}\sin^{-1}\rho \tag{1}$$

(This is likely discussed here before but I cannot quite find the question.)

You have

\begin{align} P(XY>0)&=P(X>0,Y>0)+P(X<0,Y<0) \\&=P(X>0,Y>0)+P(-X>0,-Y>0) \end{align}

Because $(-X,-Y)$ has the same distribution as $(X,Y)$, this probability is just $$P(XY>0)=2P(X>0,Y>0)$$

Similarly,

\begin{align} P(XY<0)&=P(X>0,Y<0)+P(X<0,Y>0) \\&=P(X>0,-Y>0)+P(-X>0,Y>0) \end{align}

Again, $(X,-Y)$ and $(-X,Y)$ have the same distribution, so

$$P(XY<0)=2P(X>0,-Y>0)$$

And since $(X,-Y)$ is bivariate normal with correlation $-\rho$, we have from $(1)$ that

$$P(X>0,-Y>0)=\frac14-\frac1{2\pi}\sin^{-1}\rho$$

StubbornAtom
  • 8,662
  • 1
  • 21
  • 67
  • 6
    The initial formula follows immediately from the analysis and illustrations at https://stats.stackexchange.com/a/71303/919. The first quadrant corresponds to the sector $X^\prime \gt 0, Y^\prime \gt -\rho/\sqrt{1-\rho}^2$ where $(X^\prime,Y^\prime)$ has a standard Normal distribution. Because the latter is rotationally symmetric, the chance of this sector is simply its angle divided by $2\pi.$ For positive $\rho$ this sector consists of the entire first quadrant plus the wedge $X\gt 0, -\rho/\sqrt{1-\rho^2}\lt Y \le 0,$ whose angle (by elementary geometry) is $\sin^{-1}\rho.$ – whuber Oct 12 '21 at 14:46
  • 1
    There are equivalent formulas, btw. One of them is that the expectation of $\operatorname{sgn}(XY)$ equals $2\arctan(\rho/\sqrt{1-\rho^2})/\pi.$ (Of interest is the appearance of the slope $\rho/\sqrt{1-\rho^2}$ of the boundary line, expressed as an angle.) I believe this form of the result has appeared elsewhere on CV. – whuber Oct 12 '21 at 14:48
  • 1
    @whuber's first comment is also nicely illustrated in [this](https://math.stackexchange.com/a/1688568/321264) answer on Math.SE. – StubbornAtom Oct 12 '21 at 15:21
9

Consider $X,Y\sim N(0,1)$ with correlation $\rho$. Then (Nadarajah & Pogány, 2016; Gaunt, 2018) their product is variance-gamma distributed: $$ XY \sim \text{VG}(1,\rho,\sqrt{1-\rho^2},0). $$

Its PDF is $$ f_{XY}(x) = \frac{1}{\pi\sqrt{1-\rho^2}}\exp\left(\frac{\rho x}{1-\rho^2}\right) K_0\left(\frac{|x|}{1-\rho^2}\right), $$ where $K_0$ is the modified Bessel function of the second kind of order $0$.

Thus $$ P(\text{sign}(XY) = -1) = \int_{-\infty}^0 f_{XY}(x)\,dx. $$

You may be able to evaluate this using your favorite computer algebra system. Unfortunately, this exceeds the standard computation time for WolframAlpha. COOLSerdash notes that the integral evaluates to a nice round $\frac{\arccos\rho}{\pi}$.

Alternatively, you could map the above parameterization to the one employed in the VarianceGamma package for R and use the functions in there, if all you are interested in is numerical results.

Stephan Kolassa
  • 95,027
  • 13
  • 197
  • 357
5

The correlation alone is not enough to be able to derive the probability distribution for the sign.

See the example below for a case where $\rho = 2/\pi$ and the sign of $XY$ is always positive.

example

Sextus Empiricus
  • 43,080
  • 1
  • 72
  • 161
  • 1
    Normality is given in the question. – Martund Oct 13 '21 at 12:28
  • @Martund The variables $x$ and $y$ are following a standard normal distribution here. – Sextus Empiricus Oct 13 '21 at 12:29
  • It is impossible. You claim that product of two distinct standard normal random variables takes positive values with probability $1$???!!!!!!!!!!!!!!! – Martund Oct 13 '21 at 12:32
  • 1
    @Martund I created the pairs of variables $X_i, Y_i$ from the answer in the following way. Let $X_i \sim N(0,1)$, let $Z_i \sim N(0,1)$ and let $Y_i = \text{abs}(Z_i) \times \text{sign}(X_i)$. Then $X_i$ and $Y_i$ each follow a standard normal distribution. Of course, they won't be independent and that is how you get that their product is always positive. – Sextus Empiricus Oct 13 '21 at 12:35
  • In the same way: If $\rho = 0$, then this does not need to mean that the variables are independent. For $\rho = 0$ you can also have different probabilities for the sign of the product $XY$. – Sextus Empiricus Oct 13 '21 at 12:41
  • 2
    Martund, please see https://stats.stackexchange.com/a/30205/919. It will erase all the exclamation points in your comment. – whuber Oct 13 '21 at 14:06
  • Thank You @whuber, I got the point now.. – Martund Oct 14 '21 at 13:13