Suppose $X_1$ and $X_2$ are independent $N(0,1)$ variables.

Define $$Y_1=X_1\,\text{sign}(X_2)\quad,\quad Y_2=X_2\,\text{sign}(X_1)$$

I have to show that $(Y_1,Y_2)$ is not bivariate normal and find the correlation between $Y_1$ and $Y_2$.

It is easy to see that both $Y_1$ and $Y_2$ are themselves univariate normal.

And the joint distribution function of $(Y_1,Y_2)$ can be derived to show that it is not jointly normal. But I am wondering if there is an alternate way to see this result, i.e. without explicitly finding the distribution of $(Y_1,Y_2)$ can I justify that the distribution is not jointly normal?

I am asking this because I think I do not require the joint distribution of $(Y_1,Y_2)$ to find the correlation. Since

$$Y_1=\begin{cases}X_1&,\text{ if }X_2>0\\-X_1&,\text{ if }X_2<0\end{cases}\quad\text{ and }\quad Y_2=\begin{cases}X_2&,\text{ if }X_1>0\\-X_2&,\text{ if }X_1<0\end{cases}$$

I have $$Y_1Y_2=\begin{cases}X_1X_2&,\text{ if }X_1X_2>0\\-X_1X_2&,\text{ if }X_1X_2<0\end{cases}$$

so that $$E(Y_1Y_2)=E(|X_1X_2|)=E(|X_1|)E(|X_2|)=\frac{2}{\pi},$$

thus implying that the correlation is $2/\pi$.

  • $\begingroup$ Run a few samples of $X_1$ and $X_2$ pairs and plot $Y_1$ against $Y_2$ to see that a zero correlation is not the correct answer $\endgroup$ – JimB Dec 25 '18 at 7:16
  • $\begingroup$ @JimB Let me check then. I thought $E(X_1X_2\mid X_1X_2>0)$ and $E(X_1X_2\mid X_1X_2<0)$ are equal. $\endgroup$ – StubbornAtom Dec 25 '18 at 7:18
  • 1
    $\begingroup$ Equal in magnitude certainly. But you’re subtracting a negative value so they don’t cancel out. $\endgroup$ – JimB Dec 25 '18 at 7:26
  • 3
    $\begingroup$ @StubbornAtom (edited) Showing they don't have 0 correlation doesn't show that they're not bivariate normal. $\endgroup$ – Glen_b Dec 25 '18 at 7:28
  • $\begingroup$ @Glen_b That wasn't what I was indicating at. My point was that since I could find the correlation without the joint distribution of $(Y_1,Y_2)$ then maybe I could somehow say that the joint is not bivariate normal without the actual joint distribution. That is possibly less work for the problem at hand. $\endgroup$ – StubbornAtom Dec 25 '18 at 7:37

without explicitly finding the distribution of $(Y_1,Y_2)$ can I justify that the distribution is not jointly normal?

One obvious way would be to see that $Y_1$ and $Y_2$ cannot be opposite in sign, and therefore cannot be bivariate normal.

Equivalently, note that $Y_1Y_2=\text{sign}(X_1)X_1\,\text{sign}(X_2)X_2$ $=|X_1|\cdot |X_2|$ must be non-negative.

  • 1
    $\begingroup$ In fact, the joint density must be given by $$p(x,y) = \begin{cases} 2p(x)p(y), & \text{if}~ x > 0, y > 0, \text{or}~ x < 0, y < 0,\\ 0,& \text{otherwise}.\end{cases}$$ as depicted in the third picture in this answer by Moderator cardinal. See this answer for the source. $\endgroup$ – Dilip Sarwate Dec 25 '18 at 15:02

To see what happens, let's explicitly find the distribution.

You could see it as a transformation from the entire plane to the first and third quadrants.

  • Transform the first quadrant ($X_1>0, X_2>0$) to itself $Y_1,Y_2 = X_1,X_2$.
  • Mirror the third quadrant ($X_1<0, X_2<0$) to the first trough the origin $Y_1,Y_2 = -X_1,-X_2$.
  • Mirror the second quadrant ($X_1<0, X_2>0$) along the x-axis to the third quadrant $Y_1,Y_2 = X_1,-X_2$
  • Mirror the fourth quadrant ($X_1>0,X_2 <0$) along the y-axis to the third quadrant $Y_1,Y_2 = -X_1,X_2$.

Overall you map all four quadrants to only the first and third quadrant. This relates to the answer of Glen_b.


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.