5
$\begingroup$

Suppose $B, \{A_i: i \in \omega\}$ are i.i.d. random variables with uniform distributions on $[0,1]$. If $f$ is a map such that $\{f(A_i, B): i \in \omega\}$ are independent, must $\{f(A_i, B): i \in \omega\}$ also be independent of $B$?

$\endgroup$
  • 1
    $\begingroup$ I can prove it for discrete uniform distributions. I wonder if that helps for continuous uniform distributions. $\endgroup$ – Brendan McKay Feb 23 '15 at 14:16
  • $\begingroup$ Do you think there is a counterexample for general (i.e. non-uniform) distributions? $\endgroup$ – usul Feb 23 '15 at 17:22
  • $\begingroup$ No, I think any distributions at all are ok, even a different one for each of the arguments. See the stuff I added to my answer. $\endgroup$ – Brendan McKay Feb 24 '15 at 1:19
5
$\begingroup$

It is true.

Let $\chi_Z(x,y)$ be the indicator function of the set $\lbrace (x,y) \,|\, f(x,y)\in Z\rbrace$ for some set $Z$.

The condition that the events $f(A_1,B)\in Z$ and $f(A_2,B)\in Z$ are independent is that $$ \int_{[0,1]^3} \chi_Z(x,y)\chi_Z(x'y)\,dxdx'dy = \int_{[0,1]^2} \chi_Z(x,y) \,dxdy \ \ \int_{[0,1]^2} \chi_Z(x',y') \,dx'dy'.$$ Define $g(y) = \int_0^1 \chi_Z(x,y)\,dx$. Then the above condition can be rearranged into $$ \int_{[0,1]^2} g(y) (g(y)-g(y'))\, dy dy' = 0.$$ That's just a change of variables from $$ \int_{[0,1]^2} g(y') (g(y)-g(y'))\, dy dy' = 0,$$ so we can subtract the two to get $$ \int_{[0,1]^2} (g(y)-g(y'))^2 \,dydy' = 0$$ which by positivity means $g(y)$ is constant a.e.. Harking back to the definition of $\chi_Z$ and noting that the set $Z$ was arbitrary, we see that $f(A_1,B)$ is independent of $B$.

I have only considered pairwise independence, but once you find that $g(y)$ is independent of $y$ it makes the whole lot of them independent. At least, it seems so at 3am...

ADDED after waking up: At the moment I can't see how uniform distributions are needed for this argument. Just consider any vertical measure $\mu$ and horizontal measure $\nu$ and replace $dx$ by $d\mu$, $dy$ by $d\nu$, etc, in the above argument. Still works, doesn't it?

MORE ADDED: Nate's question leads me to note the following, which surely must be well-known in the theory of exchangeable random variables (but I didn't know it).

Theorem. Let $X,Y$ be random variables such that $(X,Y)$ and $(Y,X)$ have the same distribution. Then $X$ and $Y$ are independent iff for all measurable sets $S$, $P(X\in S \wedge Y\in S)=P(X\in S)^2$.

Proof. Let $S,T$ be measurable sets. First assume $S$ and $T$ are disjoint. Since by the exchangeability assumption $P(X\in T\wedge Y\in S)=P(X\in S\wedge Y\in T)$, we have \begin{align} 2 P(X\in S\wedge Y\in T)&=P(X,Y\in S\cup T) - P(X,Y\in S) - P(X,Y\in T) \\ &= P(X\in S\cup T)^2 - P(X\in S)^2 - P(X\in T)^2 \\ &= (P(X\in S) + P(X\in T))^2 - P(X\in S)^2 - P(X\in T)^2 \\ &= 2 P(X\in S) P(X\in T) \\ &= 2 P(X\in S) P(Y\in T). \end{align} If $S$ and $T$ are not disjoint, break $P(X\in S\wedge Y\in T)$ into four disjoint cases like $P(X\in S\wedge Y\in T\setminus S)$ and apply the above to each. It works.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ In general, to show two random variables are independent, you have to show $P(X \in U, Y \in V) = P(X \in U) P(Y \in V)$ for $U,V$ ranging over an appropriately rich collection of sets. Why does it suffice to prove the special case where $U=V=Z$? Sets of the form $Z \times Z$ don't generate the product $\sigma$-algebra, do they? $\endgroup$ – Nate Eldredge Feb 24 '15 at 1:52
  • $\begingroup$ @Nate : What you say is correct, but I don't think I'm making that mistake. Independence of $X$ and $Y$ implies that the events $X\in Z$ and $Y\in Z$ are independent; the converse is false but I don't use it. Once we know $g(y)$ is a.e. constant, the distributions of $f(A,B)$ conditional on any two non-null $B$-events are the same. But let me know if you still think there's a problem. $\endgroup$ – Brendan McKay Feb 24 '15 at 2:27
  • $\begingroup$ @Nate : To my surprise, the converse is true in this case, see my addition. $\endgroup$ – Brendan McKay Feb 24 '15 at 4:08

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.