5
$\begingroup$

Let $\{X_i\}$ be i.i.d random variables uniform on a measurable, symmetric set $A$ contained in $[-1,1]$. Let $g_{n}$ be density of $X_1+\ldots + X_n$.

Question (general): Is there any non-trivial bound from below on $g_n(0)$ in terms of the Lebesgue measure of $A$.

Let us consider an example, say sets with $|A|=1/4$. It is easy to find an example of $A$ such that $g_1(0)=g_3(0)=0$. One checks also that $g_2(0)=1/|A|$.

Question (specific): What about $g_4(0)$?

$\endgroup$
1
  • $\begingroup$ local limit theorems don't care much about the details of your distribution, but they say for large n the one that maximizes the variance minimizes the density. $\endgroup$
    – user83457
    Jun 16 '16 at 9:38
2
+50
$\begingroup$

I'll only discuss even $n$, I don't have any ideas for odd $n$, plus, as your comments indicate, there may not be any lower bounds for odd $n$.

Let's write $f(t)=|A|^{-1}\widehat{\chi}_A(t)=(1/|A|)\int_A e^{2\pi itx}\, dx$. Then $$ g_n(0)=\int \widehat{g_n}(x)\, dx = \int f^n(x)\, dx . $$

First attempt: To obtain a lower bound, I'll just focus on the positive contribution coming from $x$ close to zero. Notice that $f(0)=1$, $f'(0)=0$ (because $A$ is symmetric), $|f'(x)|, |f''(x)|\le C$ (the constants could be worked out here). So $h=f^n$ satisfies $h(0)=1$, $h'(0)=0$, $|h''(x)|\lesssim n^2$, thus $$ |f^n(x)|\ge 1-cn^2x^2 . $$ It follows that $$ g_n(0)\ge C/n $$ for even $n$.

This was clearly not a very sophisticated estimate and there's probably room for improvement; for example, for $n=2$, we're off by a factor of $|A|$ for small $|A|$.

Second attempt: A perhaps more useful bound that tries more seriously to keep track of the dependence on $A$ is obtained if we use Hölder as follows: $$ \frac{1}{|A|}=\int f^2 = \int f^{a+2-a} \le \left(\int f^n\right)^{a/n}\left( \int |f|^{(2-a)n/(n-a)}\right)^{(n-a)/n} $$ If I could take $a=n/(n-1)$ here, this would (formally) give that $$ g_n(0)\ge |A|^{-1/(n-1)} \left( \int |f|\right)^{-(n-2)/(n-1)^2} . $$ That was not correct because $f\notin L^1$, but if $f\in L^{1+\epsilon}$ (this is an assumption on $A$, it's not true automatically), I can take a slightly smaller $a$ to produce a bound of the form $g_n(0)\gtrsim |A|^{-1/(n-1)+\delta}$.

$\endgroup$
1
  • $\begingroup$ Thanks! These are nice arguments. I see some hopes for the second argument. Note that in order to get a good bound one needs to take into account that the A is in the unit interval (I can construct bad examples when this is not valid). $\endgroup$ Jun 23 '16 at 15:12

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.