1
$\begingroup$

If you have $n$ identically distributed Bernoulli trials whose sum is binomially distributed random variable, does it then follow that the $n$ Bernoulli trials are independent?

$\endgroup$
  • 1
    $\begingroup$ I think this question would be more appropriate for math.stackexchange.com. The answer is no because you could take $X_1,\ldots,X_n$ to be iid Bernoulli and define $Y_1,\ldots,Y_n$ to be the increasing rearrangement of those random variables. The sum has the same distribution. $\endgroup$ – Anthony Quas Feb 22 at 15:25
  • $\begingroup$ @AnthonyQuas: I might misunderstand your reasoning, but these $Y_i$ are, I think, not identically distributed? $\endgroup$ – Steve Feb 22 at 15:31
  • $\begingroup$ @Steve: Ah yes... $\endgroup$ – Anthony Quas Feb 22 at 15:38
6
$\begingroup$

That's false even for $n=3$.

Denote by $B(p)$ the Bernoulli distribution which has probability $p$ of being 1 and $(1-p)$ to be 0.

Define the three Bernoulli variables $(X_1, X_2, X_3)$ by $ X_1 \sim B(0.5) \\ X_2|(X_1=1) \sim B(0.25);~ X_2|(X_1=0) \sim B(0.75) \\ X_3|(X_1+X_2=2) \sim B(1);~ X_3|(X_1+X_2=0) \sim B(0);~ X_3|(X_1+X_2=1) \sim B(0.5) $

Quick calculation shows they're all $B(0.5)$ distributed (basically by symmetry) and that the sum has the same distribution as the sum of three i.i.d. trials.

$\endgroup$
  • $\begingroup$ Okay, that was actually a quite easy counterexample. Thanks. $\endgroup$ – Kristian Feb 22 at 16:19
  • $\begingroup$ Another way of writing the same counterexample: Let $Y_1, Y_2, Y_3$ be independent Bernouli($1/2$). Let $X_i = Y_i$, except in the following cases: If $(Y_1, Y_2, Y_3) = (0,0,1)$ then take $X_3 = 0$, and either $(X_1,X_2) = (0,1)$ or $(1,0)$, with equal probabilities. If $(Y_1,Y_2,Y_3) = (1,1,0)$ then take $X_3 = 1$, and either $(X_1,X_2) = (0,1)$ or $(1,0)$, with equal probabilities. $\endgroup$ – Robert Israel Feb 22 at 18:23
1
$\begingroup$

The answer is no for large enough $n$, because the independence of the $n$ Bernoulli random variables (r.v.'s) is given by about $2^n$ equations, whereas to describe the individual distributions of the Bernoulli r.v.'s and their sum one needs only $O(n)$ equations.

$\endgroup$
  • $\begingroup$ It's good intuition but not exactly a proof, since there are a further $2^n$ inequalities to be satisfied: that the probability of every outcome must be nonnegative. $\endgroup$ – Nate Eldredge Feb 22 at 15:43
  • 1
    $\begingroup$ @NateEldredge : Of course, you are right. However, with this intuition, one immediately knows there should be a counterexample, which is then easy to find, by using e.g. the Mathematica command FindInstance[]. $\endgroup$ – Iosif Pinelis Feb 22 at 18:02

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.