Let $X, Y, Z$ be three random variables such that $Z$ and $X$ are independent, and $Z$ and $Y$ are independent. Let $\rho(X,Y)$ be the Pearson correlation coefficient between $X$ and $Y$. It seems to me that $\rho(XZ, YZ) \geq \rho(X, Y)$, but I can't prove it. I tried using the identity $\rho(X,Y) = \frac{Cov(X,Y)}{\sqrt{V(X)V(Y)}}$ and grinding through the algebra, to no avail.

  • 1
    $\begingroup$ Have you taken some sample X, Y, and Z and seen whether it's true for those variables? I think you will find it's not true, especially if they have mean zero. $\endgroup$ – Acccumulation Nov 8 '17 at 0:44
  • $\begingroup$ @Accumulation It looks to me like the claim will be true if X and Y both have mean 0 (it would be equality in that case); the algebra is simple in that situation $\endgroup$ – Glen_b Nov 8 '17 at 0:52
  • $\begingroup$ However, here's a clear counterexample in R by giving one of X and Y a zero mean and the other a large mean (relative to its sd), while Z has a moderate mean relative to its standard deviation: xi=rnorm(10000); y=.8*xi+.6*rnorm(10000); x=xi+10; z=rnorm(10000,1,1); cor(x,y); cor(x*z,y*z) $\endgroup$ – Glen_b Nov 8 '17 at 1:10
  • $\begingroup$ As for the algebra, you'll need to make use of the result $\text{Var}(UV) = \text{Var}(U)\text{Var}(V)+E(U)^2 \text{Var}(V)+\text{Var}(U)E(V)^2$ for independent U and V. If you take all the variables as having variance 1 and Y as having mean 0, the calculations are considerably simplified. Taking the mean of Z to be 1 simplifies things further which makes it easier to see how to obtain counterexamples. $\endgroup$ – Glen_b Nov 8 '17 at 1:22
  • $\begingroup$ The claim cannot generally be true for the simple reason that the inequality is reversed when either of $X,Y$ (but not both) is negated. But you will find that if you try to patch this up by considering the absolute value of the correlation, the inequality still doesn't hold. $\endgroup$ – whuber Nov 8 '17 at 15:07

The claim is generally not true. Thank you @Glen_b for the following counterexample (in R):

xi <- rnorm(10000)
y <- .8*xi+.6*rnorm(10000)
x <- xi+10
z <- rnorm(10000,1,1)
cor(x*z,y*z) - cor(x,y) > 0 # returns FALSE

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.