Let $(X_1, X_2)$ be a pair of uniform distributed random coordinates on $S:=([-1,0]\times[-1,0])\;\cup\; [0,1]\times[0,1])$. Are $X_1, X_2$ independent? Are $X_1, X_2$ uncorrelated?

If you draw the picture you can see they're obviously dependent are positively correlated. If I know $X_1$ is negative, then by all means $X_2$ is negative as well, and vice versa. Also, they occur in the same "direction". So positively correlated. The intuition is quite obvious. Here is my formal approach:

Show that $f_{X,Y}(x,y)\not= f_X(x)f_Y(y)$.

We know that $$f_{X,Y}(x,y)=\frac{1}{2}\left(\mathbf{1}_{[-1,0]^2}(x,y)+\mathbf{1}_{[0,1]^2}(x,y)\right)$$ and we use $1/2$ because when we integrate the above it needs to equal 1. To obtain $f_X(x)$ we integrate over $y$ and analogous for $f_Y(y)$. $$f_X(x)=\int_{\mathbb{R}}f_{X,Y}(x,y)dy=\int_{\mathbb{R}}\frac{1}{2}\left(\mathbf{1}_{[-1,0]^2}(x,y)+\mathbf{1}_{[0,1]^2}(x,y)\right)dy$$ $$=\frac{1}{2}\int_{\mathbb{R}}\mathbf{1}_{[-1,0]^2}(x,y)dy+\frac{1}{2}\int_{\mathbb{R}}\mathbf{1}_{[0,1]^2}(x,y)dy$$ $$=\frac{1}{2}\mathbf{1}_{[-1,0]}\int_{0}^1dy+\frac{1}{2}\mathbf{1}_{[0,1]}\int_{0}^1dy$$ $$=\frac{1}{2}\left(\mathbf{1}_{[-1,0]}+\mathbf{1}_{[0,1]}\right)$$

For $f_Y$ I will obtain the exact same thing. So what remains to be shown is $$\frac{1}{2}\left(\mathbf{1}_{[-1,0]^2}(x,y)+\mathbf{1}_{[0,1]^2}(x,y)\right)\not=\frac{1}{2}\left(\mathbf{1}_{[-1,0]}+\mathbf{1}_{[0,1]}\right)\cdot\frac{1}{2}\left(\mathbf{1}_{[-1,0]}+\mathbf{1}_{[0,1]}\right)$$

I tried this by integrating over $x$ and $y$ but I will be left with an identity saying $1=1$. But I know they are not independent. Where is my error?

  • $\begingroup$ Why are you integrating at the end? Two functions can be different but have the same integral... $\endgroup$ – Eric Wofsey Dec 5 '19 at 21:31
  • $\begingroup$ I thought this way I can show they are not equal. Using the pdf. $\endgroup$ – Marc Dec 5 '19 at 21:39
  • $\begingroup$ If you integrate, you are just computing the total probability of the whole space with respect to each density function, which of course is going to be $1$ for both of them. $\endgroup$ – Eric Wofsey Dec 5 '19 at 21:40
  • $\begingroup$ Does this approach make sense at all? Because now looking at my last equation, on the left side I have my indicator variable with two arguments and on the right side they are just constants.. I'm really confused on this. $\endgroup$ – Marc Dec 5 '19 at 21:47

Two things to note about that last equation you wrote. First, observe that the indicator functions on the right should carry variable inputs with them -- i.e. $\mathbf 1_{[-1, 0]}(x)$, etc. Also, note that $$\mathbf 1_{[-1, 0]}(x) + \mathbf 1_{[0, 1]}(x) = \mathbf 1_{[-1, 1]}(x).$$ (This isn't actually quite right because the function should evaluate to $2$ at $x = 0$, but that's just a single point and we can safely ignore it.)

Hence, that last equation should actually be: $$\frac{1}{2}\left(\mathbf{1}_{[-1,0]^2}(x,y)+\mathbf{1}_{[0,1]^2}(x,y)\right)\not=\frac{1}{2}\left(\mathbf{1}_{[-1,1]}(x)\right)\cdot\frac{1}{2}\left(\mathbf{1}_{[-1,1]}(y)\right)$$ and your task is now to find a region on the plane with positive measure for which this equation does not hold. Can you take it from here?

  • 1
    $\begingroup$ Well, for the same reason that you can ignore $x=0$, it's also not enough to find a single pair $(x,y)$ where the two sides are not equal. Rather you need a set of positive Lebesgue measure of such points. $\endgroup$ – Eric Wofsey Dec 5 '19 at 21:50
  • $\begingroup$ @EricWofsey Ah, good point. $\endgroup$ – Aaron Montgomery Dec 5 '19 at 21:51
  • $\begingroup$ Okay, I'm not sure about the Lebesgue measure. I tried reading into it, but it seems a little too technical for me. What I'm doing now is: For all $x,y \in [0.5,1]$ it holds that the right side is $1/4$. For the left side it's $0$ because the indicator variable is only 1 when $x,y \in [-1,0]$. So the whole left part of the left equation leaves everything with 0. Is that correct? $\endgroup$ – Marc Dec 5 '19 at 22:17
  • 1
    $\begingroup$ I didn't even check the right indicator of the left equation, the left indicator of the left equation is always 0, so that works :-). And that's what I meant by "So the whole left part of the left equation leaves everything with 0". Thanks Aaron (and Eric)! Now I have to make this work for the correlation, but I'm getting more comfortable with this.. should be fine. $\endgroup$ – Marc Dec 5 '19 at 22:35
  • 1
    $\begingroup$ I just noticed. I don't have product there, that's why you asked again. It's a sum.. but still it holds (luckily) since $1/2 \not = 1/4$. $\endgroup$ – Marc Dec 5 '19 at 23:36

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.