I'm trying to solve the following exercise.

Show that if $(X_n)_{n =1}^\infty$ is a sequence of independent real-valued random variables with mean zero and uniformly bounded fourth moment, $n^{-1}\sum_{j=1}^n X_j \to 0$ almost surely. (Hint: Think about a fourth moment tail bound.)

Here's what I did. Without loss, by rescaling $X_i$, we may assume that $\mathbf{E} X_i^4 \leq 1$. Note that this means $\mathbf{E} X_i^2 = \|X_i\|_2^2 \leq \|X_i\|_4^2 = (\mathbf{E} X_i^4)^{1/2} \leq 1$. Put $S_n = \sum_{j=1}^n X_j$, observe that $\mathbf{E}S_n^4 \leq n + (n^2 - n) = n^2$, simply because the other cross terms vanish due to independence.

Following the hint, notice that by Markov's inequality, $$ \mathbf{P}(|n^{-1}S_n|\geq n^{-1/8}) = \mathbf{P}(S_n^4 \geq n^{3.5}) \leq n^{-3.5} \mathbf{E}S_n^4 \leq n^{-1.5}. $$ The upshot of this is that with $A_n = \{|n^{-1}S_n| \geq n^{-1/8}\}$, $\mathbf{P}(A_n)$ is summable, and hence $\mathbf{P}(A_n~\mathrm{i.o.}) = 0$ by Borel-Cantelli Lemma 1.

If $\omega$ is such that $|n^{-1}S_n(\omega)| \geq n^{-1/8}$ for finitely many terms, then $\limsup_n |n^{-1}S_n(\omega)| = 0$, and hence $n^{-1}S_n(\omega) \to 0$. But the set of such $\omega$ is almost sure (indeed, the complement of $A_n~\mathrm{i.o.}$), proving the claim.


This is almost correct, except for the crucial part: it is not necessarily true that the expected value of $S_n^4$ will not be greater than $n$. You fail to take into account the $\sim n^2$ terms $\mathbb{E}[X_i^2X_j^2]$.

  • $\begingroup$ Ah, @Mindlack, good point. I think, however since $u\mapsto u^4$ is convex, we have that $(1/n \sum_{j=1}^n X_j(\omega))^4 \leq 1/n \sum_{j=1}^n X_j(\omega)^4$, in which case we actually have that $\mathbf{E} (S_n/n)^4 \leq 1$. $\endgroup$ – Drew Brady Dec 12 '18 at 0:22
  • $\begingroup$ Ah, but this bound is not strong enough to carry forth the argument. $\endgroup$ – Drew Brady Dec 12 '18 at 0:25
  • $\begingroup$ @Drew Brady: Yes indeed. However, the terms mentioned in my answer are the only ones that matter, so the expected value is less than $n^2$. $\endgroup$ – Mindlack Dec 12 '18 at 0:26
  • $\begingroup$ Thanks for the edits. Can you take a look at my revised argument. Tell me if you think the details look right now. I forgot that $L^p$ is monotone on probability spaces! $\endgroup$ – Drew Brady Dec 12 '18 at 0:48
  • $\begingroup$ The edited solution works now! $\endgroup$ – Mindlack Dec 12 '18 at 12:00

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.