3
$\begingroup$

Let $\{ X_i \}_{i=1}^n$ be sequence of i.i.d. mean zero ($\mathbb{E} X_i = 0$) random variables with bounded second moment ($\mathbb{E} X_i^2 < \infty$) . Now what would be the limiting behavior of

$$S^a_n := \frac{\sum_{i=1}^n X_i}{n^a},$$

as $n \to \infty$ where $a \in [0,1]$ or even $\mathbb{R}$.

When $a = 1$, we have the Strong Law of Large Numbers which say that $S^1_n$ almost surely converges to $0$.

When $a = 1/2$, we have the Central Limit Theorem which say that $S^{1/2}_n$ converges to a normal distribution ($\mathcal{N} (0, \mathbb{E}X_i^2 )$) in distribution.

But what happens for other choices of $a$ and how can one prove it converges or diverges in those scenarios?

This is Terence Tao's comment on it (which I didn't fully understand) - https://terrytao.wordpress.com/2008/06/18/the-strong-law-of-large-numbers/

Assuming the random variables have bounded second moment, the limit will be almost surely zero for a > 1/2, converge (in a distributional sense) to a normal distribution for a=1/2, and diverge to infinity almost surely for a < 1/2, all thanks to the (strong) law of large numbers. For heavy-tailed random variables with infinite second moment, the situation is going to be more complicated, but can be worked out for any specific distribution by a variety of tools (e.g. Fourier analysis)"

$\endgroup$

1 Answer 1

Reset to default
4
$\begingroup$

Using Slutsky's lemma, it is immediate that if $a\gt 1/2$, then $S_n^a=S_n^{1/2}n^{1/2 -a}\to 0$ in probability. More can be said thank s to the law of the iterated logarithms. Indeed, since $\limsup_n S_n^{1/2} /\sqrt{2\log\log n}$ and $\liminf_n S_n^{1/2} /\sqrt{2\log\log n}$ are almost surely finite, we derive that $S_n^a\to 0$ almost everywhere. This can also be derived from Marcinkiewicz law of large numbers.

When $a\lt 1/2$, we cannot even expect a convergence in distribution. If $\left(Y_n\right)_{ n\geqslant 1}$ is a non-negative sequence of random variables which converges in distribution to a non-degenerated random variable $Y$, then for any sequence $\left(R_n\right)_{n\geqslant 1}$ going to infinity, the sequence $\left(R_n Y_n\right)_{n\geqslant 1}$ cannot be convergent in distribution.

$\endgroup$
3
  • $\begingroup$ Thank you for the answer, @Davide Giraudo. I looked at Marcinkiewicz law of large numbers, it gives exactly what we want for $a \in (1/2, \infty)$. Just to make it clear I understand by Slutsky's lemma we can say that $S^a_n \overset{d}{\to} 0$ and convergence in probability follows from the fact that $X_n \overset{d}{\to} c \implies X_n \overset{p}{\to} c$, where $c$ is a constant. $\endgroup$ Commented Dec 5, 2016 at 0:05
  • $\begingroup$ But I have two follow up questions. (1) By law of iterated logarithms I understand that $S_n^{1/2}/\sqrt{2 \log \log n}$ finite a.s, but how does it imply (along with the "in probability" convergence proved by you) that $S^a_n = 0$ a.s. for $a > 1/2$? $\endgroup$ Commented Dec 5, 2016 at 0:06
  • $\begingroup$ (2) I didn't understand why $(R_n Y_n)_{n \geq 1}$ does not converge and why does it imply that $S_n^a$ doesn't converge for any $a \in (1, 1/2)$, could you give me some hints please. $\endgroup$ Commented Dec 5, 2016 at 0:06

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.