Under what conditions does the below statement hold:

X and Y are uncorrelated if and only if X and Y are independent.

I totally understand that this statement does not always hold, but I would like to know under what specific conditions (for what distribution of X and Y) it will be true.

  • 1
    $\begingroup$ Do not forget self study tag $\endgroup$ – zlon Feb 11 '17 at 23:34
  • 1
    $\begingroup$ It's difficult to determine what you asking. What do you mean by a "special case"? What do you mean by "always holds"? You might be interested in stats.stackexchange.com/questions/4364. $\endgroup$ – whuber Feb 11 '17 at 23:35
  • $\begingroup$ Sorry for that, however, I was looking for a distribution under which this statement will always hold. $\endgroup$ – Saba Feb 12 '17 at 0:20

The statement that you are asking about has two parts:

  1. If $X$ and $Y$ are independent, then $X$ and $Y$ are uncorrelated.

  2. If $X$ and $Y$ are uncorrelated, then $X$ and $Y$ are independent.

Statement 1 is always true and imposes no additional constraints on $X$ and $Y$ other than what already has been assumed, viz. that they are independent random variables. Statement 2 does not hold in general, but it does hold if we constrain $X$ and $Y$ to be jointly Gaussian random variables. That is,

2'. If jointly Gaussian random variables $X$ and $Y$ are uncorrelated, then $X$ and $Y$ are independent.

is a true statement, and so

Jointly Gaussian random variables $X$ and $Y$ are uncorrelated if and only if they are independent

is a true statement but

"Random variables $X$ and $Y$ are uncorrelated if and only if they are independent"

does not hold in general. Nor is

"Gaussian random variables $X$ and $Y$ are uncorrelated if and only if they are independent"

a true statement. (Note that in contrast to 2'. the word jointly is missing from the statement). For example, suppose that $X\sim N(0,1)$ and $Z$, independent of $X$ is a Bernoulli random variable with parameter $\frac 12$. Set $Y = (-1)^ZX = \pm X$ and note that $Y \sim N(0,1)$, just like $X$. But, $$E[XY] = E[(-1)^Z X^2] = E[(-1)^Z]E[X^2] = 0 = E[X]E[Y]$$ showing that $X$ and $Y$ are (marginally) Gaussian random variables that are uncorrelated. That they are not independent is easily see because conditioned on the event that $X = x_0$, $Y$ takes on values $x_0$ and $-x_0$ and is thus a discrete random variable instead of continuing to enjoy the standard Gaussian density as it would have if only $X$ and $Y$ were independent random variables. Note that $X$ and $Y$ do not have a jointly Gaussian density.

Finally, if $X$ and $Y$ are Bernoulli random variables or more generally, discrete random variables that take on only two different values, then the statement

Bernoulli random variables (more generally, dichotomous random variables) $X$ and $Y$ are uncorrelated if and only if they are independent

is a true statement. See this question and its answers for some details.

  • 2
    $\begingroup$ @Saba If this answers your question please mark this answer as accepted (by clicking on a green tick to the left of it). $\endgroup$ – amoeba Feb 13 '17 at 10:12

For a joint distribution function (CDF) constructed as follows

$$H_{X,Y}(x,y)=F_X(x)G_Y(y)\left[1+\alpha\big(1-F_X(x)\big)\big(1-G_Y(y)\big)\right],\;\;\; \alpha >1$$

where $F_X(x)$ and $G_Y(y)$ are any two marginal CDF's,

uncorrelatedness (zero covariance) is equivalent to independence.

This is the "Farlie-Gumbel-Morgenstern" family of joint distributions. For an analysis of the correlation structure, see

Schucany, W. R., Parr, W. C., & Boyer, J. E. (1978). Correlation structure in farlie-gumbel-morgenstern distributions. Biometrika, 65(3), 650-653.


The result is only guaranteed to hold when X and Y form a bivariate normal distribution. You will find this in most multivariate analysis texts as well as on some threads on this site.

  • 1
    $\begingroup$ Thank you very much, I was in specific looking for a distribution that let this statement always holds. $\endgroup$ – Saba Feb 12 '17 at 0:20
  • $\begingroup$ All I meant to say by my answer was that the bivariate normal is the only case where two univariate random variables X and Y are guaranteed to be independent when they are uncorrelated. Dilip Sawate subsequently added more detail to that. I made a minor edit to my answer to clarify that. $\endgroup$ – Michael R. Chernick Mar 27 '17 at 12:16

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.