3
$\begingroup$

As noted in the recent answer by Yuval Peres, the sum of independent uniformly distributed random variables (r.v.'s) cannot have a normal distribution.

The question is, what happens without the independence condition? More specifically:

Do there exist positive real numbers $a_1,a_2,\dots$ and r.v.'s $U_1,U_2,\dots$ such that for each natural $i$ the r.v. $U_i$ is uniformly distributed on the interval $[-a_i,a_i]$ and the sum $U_1+U_2+\cdots$ has the standard normal distribution?

$\endgroup$
6
$\begingroup$

Yes. Let the random variables $V_i$ be independent, with $V_i$ uniform on $[-\frac 1i,\frac 1i]$. Let $N$ be an independent standard normal. Inductively define $Z_i\in \{\pm 1\}$ by $$ Z_i= \begin{cases} 1&\text{if $\text{sgn} (N-\sum_{j<i}Z_jV_j)=\text{sgn}(V_i)$}\\ -1&\text{otherwise.} \end{cases} $$ Notice that $\sum Z_iV_i=N$ almost surely - since the $|V_i|$ approach 0, but are almost surely not summable, the partial sums overshoot successively in one direction then the other, converging to $N$.

An observation that we need is the following: the involution that negates all of the $V$’s and $N$ preserves probabilities; and the values of the $Z$ variables are unchanged. In particular for any $a$ and $i$, we have $\mathbb P(V_i<a,\, Z_i=-1)=\mathbb P(V_i>-a,\, Z_i=-1)$.

Now we claim that $U_i=Z_iV_i$ has the same distribution as $V_i$. To see this, notice that $\mathbb P(U_i<a)=\mathbb P(V_i<a,\,Z_i=1) + \mathbb P(V_i>-a,\, Z_i=-1)$. But by the above observation, $\mathbb P(V_i>-a,\,Z_i=-1)=\mathbb P(V_i<a,\,Z_i=-1)$, so that $\mathbb P(U_i<a)=\mathbb P(V_i<a)$ as required.

added comment: @Michael Hardy pointed out the use of conditional convergence in this answer. It turns out this is essential: there does not exist a collection of dependent uniform random variables where the sum is (non-trivially) normal and the convergence is absolute. To see this, note that if $\sum a_i<\infty$, then the sum is compactly supported, so cannot be normal. If $\sum a_i=\infty$, then $\mathbb E\sum_{i=1}^M|U_i|=\sum_{i=1}^M a_i/2$ by linearity of expectation. Since $\sum_{i =1}^M |U_i|\le \sum_{i=1}^M a_i$, we deduce $\mathbb P(\sum_{i=1}^M |U_i|\ge \sum_{i=1}^M a_i/4)\ge 1/4$. In particular, the limit superior of these sets has measure at least $1/4$, but on this set the sum of the absolute values diverges.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Can you clarify — what is the pdf for the sequence $u_1, u_2, \ldots$? $\endgroup$ – Matt F. Oct 14 '19 at 9:05
  • $\begingroup$ Very nice! However, it may seem unclear when you write "by symmetry (under replacing each random variable by its negation)" but actually (of course) do not negate the $Z_i$'s. Can you detail this symmetry consideration? $\endgroup$ – Iosif Pinelis Oct 14 '19 at 15:05
  • $\begingroup$ @MattF. : Sorry. I have no idea what the joint pdf looks like. (I do know the marginal PDFs though!) $\endgroup$ – Anthony Quas Oct 14 '19 at 18:26
  • $\begingroup$ @IosifPinelis : I tried to expand the explanation in the way that you suggested. $\endgroup$ – Anthony Quas Oct 14 '19 at 18:27
  • 1
    $\begingroup$ @MichaelHardy: I haven't seen conditoinal convergence in probability before. On the other hand, what I'm doing here is a version of my favourite party trick: essentially using extra randomness (in this case the $N$ which is the target random variable) to build something doing what you want (Notice that the $N$ once built in to the $Z$'s is obtained from the $Z$'s rather than the other way around). I have generally done things a bit like this in dynamical systems, where I call the technique "Coupling and Splicing". The symmetry is a bonus making things work out nicely. $\endgroup$ – Anthony Quas Oct 15 '19 at 6:21

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.