2
$\begingroup$

Suppose we have a sequence of random variables, $\{X_{n}\}_{n\geq 1}$ satisfying:

$\mathbb{P}(X_{j} = 2^{j}) = \mathbb{P}(X_{j} = -2^{j}) = \frac{1}{2}$

Then is it true that the CLT holds? Or sufficiently, does the following Lindeberg condition hold?

$\lim\limits_{n\rightarrow\infty}\dfrac{1}{s_{n}^{2}}\displaystyle\sum\limits_{j=1}^{n}E[(X_{j}-E[X_{j}])^2\chi_{\{|X_{j}|>\epsilon s_{n}\}}] = 0 \;\; \forall \epsilon>0$

where $s_{n}^{2} = \displaystyle\sum\limits_{j=1}^{n}Var[X_{j}]$. Specifically, we have $E[X_{j}] = 0$ and $s_{n}^{2} = \displaystyle\sum\limits_{j=1}^{n}4^{j}$. So we need to show:

$\lim\limits_{n\rightarrow\infty}\frac{1}{\sum\limits_{j=1}^{n}4^{j}}\displaystyle\sum\limits_{j=1}^{n}E[X_{j}^2\chi_{\{|X_{j}|>\epsilon s_{n}\}}] = 0 \;\; \forall \epsilon>0$.

I'm not sure how to deal with the set $\{|X_{j}| > \epsilon s_{n}\}$.

What happens if Lindeberg condition does not hold? I don't think the Lindeberg condition is necessary since $\{X_{n}\}_{n\geq 1}$ does not satisfy Feller's criterion:

$\lim\limits_{n\rightarrow\infty}\max\limits_{1\leq j\leq n}\dfrac{Var[X_{j}]}{s_{n}^{2}} = \lim\limits_{n\rightarrow\infty}\dfrac{4^{n}}{\displaystyle\sum\limits_{j=1}^{n}4^{j}} = \dfrac{3}{4} \neq 0$

$\endgroup$
0
$\begingroup$

Back for a second try. Running a couple of simulations suggests that the normalized sum $S_n/s_n$ indeed converges, but strangely the histogram of $S_n/s_n$ seems more rectangular than bell-shaped. To understand what the limiting distribution is, calculate the characteristic function of the partial sum $S_n=X_1+\cdots+X_n$: $$ \begin{align} E\exp(itS_n) & = \prod_{j=1}^nE \exp(itX_j)\\ &= \prod_{j=1}^n\left( {\exp(it2^j)+\exp(-it2^j)\over 2}\right)\\ &= \prod_{j=1}^n\cos(2^jt)\\ &= \prod_{j=1}^n\cos\left({2^{n+1}t\over 2^j}\right)\\ &= {\sin(2^{n+1}t)\over 2^n\sin(2t)} \end{align} $$ where in the last step we use the identity $$ \prod_{j=1}^n\cos\left({x\over2^j}\right)={\sin x\over 2^n\sin(x/2^n)}\;. $$ Now normalize by $s_n=c_n2^n$, where $c_n$ tends to a constant $c=\sqrt{4/3}$ as $n\to\infty$. We have $$ E\exp(it(S_n/s_n)) = {\sin(2^{n+1}t/c_n2^n)\over 2^n\sin(2t/c_n2^n)} = {\sin(2t/c_n)\over 2^n\sin(2t/c_n2^n)}. $$ As $n\to\infty$, this last tends to $\displaystyle{\sin(2t/c)\over 2t/c}$ , which is the CF of a uniform$[-2/c,2/c]$ random variable, i.e. $S_n/s_n$ converges to a uniform$[-\sqrt3,\sqrt3]$ distribution. (Please check my algebra.)

$\endgroup$
  • $\begingroup$ Wow, this is great, thank you! Can I ask how you arrived at the equality $cos(2^{j}t) = cos(\frac{2^{n+1}t}{2^{j}})$? $\endgroup$ – Quicksilver Apr 24 '15 at 22:17
  • $\begingroup$ Apply the change of index $j\mapsto n+1-j$ (in preparation for applying the identity). We are just iterating in reverse order. $\endgroup$ – grand_chat Apr 24 '15 at 22:19
1
$\begingroup$

If you want to show the CLT holds, you'll need a criterion other than the Lindeberg condition. Note that $s_n^2< 4^{n+1}$. Use this inequality to bound the Lindeberg expression away from zero: $$\sum\limits_{j=1}^{n}E[X_{j}^2\chi_{\{|X_{j}|>\epsilon s_{n}\}}]= \sum_{j=1}^n 4^jP(|X_j|>\epsilon s_n)\ge 4^nP(|X_n|>\epsilon s_n) \ge 4^nP(|X_n|>\epsilon 2^{n+1})\;. $$ For every small $\epsilon$ the last probability equals $1$, so $$\frac{1}{s_n^2}\displaystyle\sum\limits_{j=1}^{n}E[X_{j}^2\chi_{\{|X_{j}|>\epsilon s_{n}\}}]\ge{4^n\over4^{n+1}}=\frac14\;, $$ i.e., the Lindeberg condition doesn't hold for your sequence.

$\endgroup$
  • $\begingroup$ Thank you. How did you arrive at the conclusion that: $s_{n}^{2} = \sum\limits_{j=1}^{n}4^{j} < 4^{n+1}$? Perhaps it should be $s_{n}^{2} < n4^{n}$? $\endgroup$ – Quicksilver Apr 24 '15 at 21:27
  • $\begingroup$ Sorry, it seems like a simple induction argument to prove $s_{n}^{2} < 4^{n+1}$. Thank you again. $\endgroup$ – Quicksilver Apr 24 '15 at 21:35
  • $\begingroup$ I used $s_n^2=\frac43(4^n-1)$ and tried to find a nice upper bound . Your inequality is also valid, but it makes it harder to argue that $P(|X_n|>\epsilon s_n)=1$ $\endgroup$ – grand_chat Apr 24 '15 at 21:37

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.