Let $\big\{X_1, X_2, ..., X_n \big\}$ be $n$ jointly exchangeable Bernoulli random variables, i.e., exchanging the order of these random variables does not change the joint distribution. If we know that $$\sum_{i=1}^{n} X_i \leq m < n$$ holds for sure, does this imply that any $X_i$ and $X_j$ are negatively correlated? This is quite intuitive to me because these random variables are symmetric and their sum is bounded from above...

  • 4
    $\begingroup$ Consider the case when with probability $1/2$ random $m$ of them are $1$ and with probability $1/2$ random $n-m$ of them are $1$ where $m=n-1$, say. Then it looks to me like the correlations are rather positive for large $n$. $\endgroup$
    – fedja
    Nov 18, 2018 at 21:42
  • 1
    $\begingroup$ @IlyaBogdanov On the contrary. If $X_1=1$, it makes it almost certain that we are in case $1$, so the probability that $X_2=1$ gets close to $1$. More precisely, $P(X_1=1,X_2=1)=P(X_1=0,X_2=0)=\frac 12\frac{(n-1)(n-2)}{n^2}$, which is almost full correlation. $\endgroup$
    – fedja
    Nov 19, 2018 at 1:07
  • $\begingroup$ @fedja : Very nice example. In fact, the random variables are even more correlated: $P(X_1=X_2=1)=\frac12\,\frac{n-2}n$ for $n\ge2$, I think. $\endgroup$ Nov 19, 2018 at 2:29
  • $\begingroup$ @fedja: Ah yes, you are right! $\endgroup$ Nov 19, 2018 at 5:05
  • $\begingroup$ @fedja Nice example! Thanks so much! $\endgroup$
    – zxzx179
    Nov 19, 2018 at 5:34


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Browse other questions tagged or ask your own question.