Suppose $X,Y$ are positive random variables with $\operatorname{Cov}(X,Y)=0$. Define $Z= \log Y$. Does it necessarily follow that $\operatorname{Cov}(X,Z) = 0$? I know it's true for linear transformations of $Y$, but now I have something more complicated. Conceptually I feel like it should be true, however there could be a counterexample I haven't thought of.

  • $\begingroup$ I edited your title so that it summarises your question, instead of just being the first line of it. Feel free to revert if you think it's worse this way. $\endgroup$ – Ben Millwood Jun 10 '15 at 22:35
  • $\begingroup$ I think your edit is better. Thank you :) $\endgroup$ – Brenton Jun 10 '15 at 22:36
  • $\begingroup$ If $X$ and $Y$ are independent random variables, then $\iint (X-\bar X)(Y-\bar Y)\,\rho(x,y)dxdy=0$ and $\iint (X-\bar X)(\log Y-\bar \log Y)\,\rho(x,y)dxdy=\int (X-\bar X)\rho_X(x)dx\int (\log Y-\bar \log Y)\rho_Y(x)dy=0$ also. If $X$ and $Y$ are not independent, then one cannot conclude generally that $\text{Cov}(X,Y)=0\implies \text{Cov}(X,\log Y)=0$. $\endgroup$ – Mark Viola Jun 10 '15 at 22:50
  • $\begingroup$ My answer isn't any good. Will try again. $\endgroup$ – Clarinetist Jun 10 '15 at 23:21

A counterexample requires that $X$ and $Y$ not be independent. Let $(X,Y)$ be uniform on the three points $(1,a)$, $(3,b)$, and $(4,c)$. Then $E(X)=8/3$, $E(Y)=(a+b+c)/3$, and $E(XY)=(a+3b+4c)/3$. To have $\mathrm{Cov}(X,Y)=0$ means $E(XY)=E(X)E(Y)$, which means $$3(a+3b+4c)=8a+8b+8c\quad\Longleftrightarrow\quad b+4c=5a\;.$$ You can find positive $(a,b,c)$ to satisfy the above, but then (in general) $(\log a, \log b, \log c)$ won't satisfy the same equation, which would be necessary if $\mathrm{Cov}(X,\log Y)=0$, since $(X,\log Y)$ is uniform on the three points $(1,\log a)$, $(3,\log b)$, $(4, \log c)$.

Example: $a=6$, $b=10$, $c=5$ wil do. But $\log b + 4\log c-5\log a=\log(bc^4/a^5)$ is not zero.

  • $\begingroup$ Why does $\log Y$ have uniform distribution? $\endgroup$ – Sina Jun 11 '15 at 0:04
  • $\begingroup$ @Sina $\log Y$ is uniform on the three points $\log a$, $\log b$, $\log c$, since $Y$ is uniform on the three points $a$, $b$, $c$. $\endgroup$ – grand_chat Jun 11 '15 at 0:06
  • $\begingroup$ (Never mind, my previous comment was wrong.) $\endgroup$ – Clarinetist Jun 11 '15 at 0:35
  • $\begingroup$ @Clarinetist True. But if the pair $(X,Y)$ has a discrete uniform distribution, then the answer is yes, in general (barring duplicate values for $XY$). In the above example $XY$ takes three values: $a$, $3b$, $4c$ each with prob $1/3$. $\endgroup$ – grand_chat Jun 11 '15 at 0:38
  • $\begingroup$ Great counterexample. +1 $\endgroup$ – Clarinetist Jun 11 '15 at 0:40

$\newcommand{\Cov}{\operatorname{Cov}} \newcommand{\E}{\mathbb{E}}$

The answer to you question is NO, in general.

Let $g(Y)$ be a function of $Y$. Then, $$\Cov[X,g(Y)] = \E[(X-\mu_X)(Y-\mu_{g(Y)})] %= \E[Xg(Y)-\mu_{g(Y)}X-\mu_Xg(Y)+\mu_X\mu_{g(Y)}] = \E[Xg(Y)]-\mu_X\mu_{g(Y)}.$$

Having $\Cov[X,g(Y)]=0$ only means $$\E[Xg(Y)]=\mu_X\mu_{g(Y)},$$ which does not hold in general (and your specific $g(Y)=\operatorname{log}{Y}$).

Remark: If $X$ and $g(Y)$ are independent, then $\Cov[X,g(Y)]=0$. This is definitely correct when $X$ and $Y$ are independent.

  • $\begingroup$ My question would be, then, do you have an example of uncorrelated but dependent positive random variables? $\endgroup$ – Clarinetist Jun 10 '15 at 23:35
  • $\begingroup$ I do not actually, because I realized in my work, I made a small mistake: $Y$ is positive but $X$ does not have to be (in fact, $X$ is a normal r.v.). I believe I might be able to show that $Cov(X, logY) = 0$, but was hoping that would tell me something about $Cov(X,Y)$ $\endgroup$ – Brenton Jun 10 '15 at 23:37
  • $\begingroup$ You need a counter example to validate your conclusion. $\endgroup$ – Mark Viola Jun 11 '15 at 0:02
  • $\begingroup$ I couldn't come up with a counterexample. This book might be helpful: Counterexamples in probability and statistics Author: Joseph P Romano; Andrew F Siegel $\endgroup$ – Bob Jun 11 '15 at 0:28

Let's see if we can make an example really simple. Let $$ Y = \begin{cases} 1 \\ 2 \\ 3 \end{cases} $$ each with probability $1/3$, and let $X=(Y-2)^2+1$. Then $\operatorname{cov}(X,Y)=0$ and $\operatorname{cov}(X,\log Y)\ne0$. The difference happens because the simple arithmetic progression $1,2,3$ is replaced by $\log1=0, \log2, \log3$, which is not an arithmetic progression: the amount you add to the first term to get the second is not the same as the amount you add to the second to get the third.


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.