3
$\begingroup$

Here is an exercise in the book of author Achim Klenke. Let $(X_n)$ be iid non-negative random variables. By using Borel-Cantelli lemma, show that: $$ \limsup_n \dfrac{X_n}{n} = 0 \text{ a.s} $$ if $\mathbb{E}(X_1) < \infty$. Otherwise, show that $\limsup_n \dfrac{X_n}{n} = \infty$ a.s

As suggested by the problem, I tried to express the event $\{\limsup_n X_n/n = 0\}$ as the limsup of events as follows: $$ \begin{align*} \left(\limsup_n \dfrac{X_n}{n} = 0 \right) &= \bigcap_{n \in \mathbb{N}} \bigcup_{m \in \mathbb{N}} \bigcap_{k \ge m} \left(\dfrac{X_k}{k} \le \dfrac{1}{n}\right) \end{align*} $$ Thus, $$ \mathbb{P}\left(\limsup_n \dfrac{X_n}{n} = 0\right) = 1 - \lim_{n \rightarrow \infty} \mathbb{P}\left(\limsup_m \left[\dfrac{X_m}{m} > \dfrac{1}{n}\right]\right) $$

From here, I want to show that $$ \sum_{m = 1}^\infty \mathbb{P}\left(\dfrac{X_m}{m} > \dfrac{1}{n}\right) < \infty \ \forall n $$ However, I can only prove this if we add the condition $\mathbb{E}(X_1^2) < \infty$, then $$ \sum_{m = 1}^\infty \mathbb{P}\left(\dfrac{X_m}{m} > \dfrac{1}{n}\right) = \sum_{m = 1}^\infty \mathbb{P}\left(\dfrac{X_m^2}{m^2} > \dfrac{1}{n^2}\right) \le n^2 \mathbb{E}(X_1^2)\sum_{m = 1}^\infty \dfrac{1}{m^2} < \infty $$ Without the assumption $\mathbb{E}(X_1^2) < \infty$, I'm pretty much stuck, so any hints for other ways are appreciated. Thanks

$\endgroup$

1 Answer 1

4
$\begingroup$

You have shown that the goal is to prove for any $\epsilon > 0$ (which is equivalent to $n^{-1}$ in your post), it holds that \begin{align} \sum_{m = 1}^\infty P(X_m > m\epsilon) < \infty. \end{align}

By the i.i.d. assumption (in fact, just "identically distributed" condition would suffice), this is equivalent to prove \begin{align} \sum_{m = 1}^\infty P(X_1 > m\epsilon) < \infty. \tag{1} \end{align}

Now the inequality (the first equality below uses $X_1$ is nonnegative) \begin{align} E[X_1] &= \int_0^\infty P[X_1 > t]dt = \sum_{m = 1}^\infty \int_{(m - 1)\epsilon}^{m\epsilon}P[X_1 > t]dt \geq \sum_{m = 1}^\infty \epsilon P[X_1 > m\epsilon] \end{align} and $E[X_1] < \infty$ imply that $(1)$ holds. This completes the proof.

The same trick can be used to prove $\limsup_n n^{-1}X_n = \infty$ almost surely given $E[X_1] = \infty$. But for this to hold, now we need the independent assumption as well (as opposed to the $E[X_1] < \infty$ case which only needs $X_1, X_2, \ldots$ are identically distributed) to deploy the second Borel-Cantelli lemma.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.