The Central Limit Theorem tells us that for an iid sequence of random variables $(X_n)_{n\geq 0}$ of finite variance $\sigma^2$ and zero mean

$$\lim_{n\to\infty}\frac{S_n}{\sqrt{n}}=^d N(0,\sigma^2)$$

where $S_n=X_1+\cdots+X_n$.

Suppose we have a similar sequence, except now we suppose that $X_n$ has infinite variance. Then is it possible for the sequence $\frac{S_n}{\sqrt{n}}$ to converge in distribution? Is there always some $\alpha$ such that $n^\alpha S_n$ converges to a non-constant distribution?

(It seems to me that the answer to the first question should be no, but I'm having trouble showing this.)

Thank you.

  • $\begingroup$ There is the Cauchy distribution for instance. But there are many other distributions with infinite variance. You'll need to specify more to have a definite limiting distribution. $\endgroup$ – Raskolnikov Oct 5 '11 at 17:37
  • 1
    $\begingroup$ Suppose the random variables have symmetric density $f(x)$ and $f(x)$ is asymptotic to $|x|^{-(1+\alpha)}$ as $|x| \rightarrow \infty$. If $\alpha >2$, then the central limit theorem applies. If $0 < \alpha <2$, then $\frac{S_n}{n^{1/\alpha}}$ converges in distribution to a symmetric $\alpha$-stable distribution. See en.wikipedia.org/wiki/Stable_distribution $\endgroup$ – ShawnD Oct 5 '11 at 18:02

The generalized central limit theorem states (see this for a summary), that if $X_i$ are i.i.d. such that its density function has left tail power-law asymptotic $\mathbb{P}(X < -x) \sim d x^{-\mu}$ and right tail asymptotic $\mathbb{P}(X > x) \sim 1- c x^{-\mu}$ as $x \to +\infty$, then there exist sequences $a_n$ and $b_n$ such that the random variate $Z_n = ((\sum_{i=1}^n X_i) - a_n )/b_n$ converges in probability to a stable distribution with stability index $\alpha = \min(\mu, 2)$ and asymmetry parameter $\beta = \frac{c-d}{c+d}$.

Details on the constructive choice of sequences $a_n$ and $b_n$ are given in the table found at the link above. Also see page 62 of Zolotarev and Uchaikin on Google books.


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.