Do there exist random variables that force transitivity of dependence? In general, statistical dependence is not transitive. If $Y$ and $X_{1}$ are dependent, and $Y$ and $X_{2}$ are dependent, then $X_{1}$ and $X_{2}$ are NOT necessarily dependent.
However, in some cases they CAN be dependent. A question which occurred to me earlier this week on which I have not been able to make some headway, is: are there any random variables $Y$ such that any variables dependent on $Y$ must be dependent on each other?
Related is the idea of decomposability (whether the random variable can be expressed as the sum of independent random variables), but I was unable to find anything on this more general problem.
The simplest case which occurs to me is a Bernoulli random variable, but I have neither been able to prove it nor offer a counterexample. If someone could shed light on the issue, it would be much appreciated.
 A: Here's a counterexample.
Consider a probability space with $4$ outcomes, and the following probabilities:
$$ \matrix{ Y & X_1 & X_2 & \text{probability} \cr
            0 & 0   & 0   &    (1-s)^2 \cr
            1 & 0   & 1   &    s - s^2 \cr
            1 & 1   & 0   &    s - s^2 \cr
            1 & 1   & 1   &    s^2 \cr} $$
where $0 < s < 1$.  $Y$ is Bernoulli with parameter $p = 1 - (1-s)^2$,
which can be anything in $(0,1)$.
$X_1$ and $X_2$ are independent Bernoulli random variables
with parameter $s$.  However, $X_i$ and $Y$ are dependent, since 
${\mathbb P}(X_i = 0 | Y = 0) = 1 \ne {\mathbb P}(X_i = 0)$.
EDIT: This can be generalized.  Let $Y$ be any random variable that is not a.s. constant.  Take $A$ so that $0 < p = {\mathbb P}(Y \in A) < 1$, and take 
$s = 1 - \sqrt{1-p}$ so that $p = 1 - (1-s)^2$.  Let $U$ be independent of $Y$ with ${\mathbb P}(U=1) = {\mathbb P}(U=2) = (s -s^2)/p$, ${\mathbb P}(U=3) = s^2/p$. Let 
$$(X_1,X_2) = \cases{(0,0) & if $Y \notin A$\cr
(0,1) & if $Y \in A$ and $U=1$\cr
(1,0) & if $Y \in A$ and $U= 2$\cr
(1,1) & if $Y \in A$ and $U = 3$\cr }$$
Then $X_1$ and $X_2$ are independent Bernoulli with parameter $s$, but $X_i$ and $Y$ are dependent.
