I've been trying to prove that given a sequence of independent random variables with identical distribution $\{X_n\}_{n \in \mathbb{N}}$ such that $P(X_1 \neq 0)>0$, so also $P(X_i \neq 0) >0 \ \ \forall i$, the series $$\sum_{n \in \mathbb{N}} X_n \ \ \text{is diverent almost surely}.$$

That is, I need to prove that $P(\omega| \ \exists \varepsilon>0 : \forall N \in \mathbb{N} \exists m, n \ge N : |S_m(\omega) - S_n(\omega)| > \varepsilon)$.

We are dealing with a series of random variables, so I thought I could use Borel-Cantelli lemma. It implies that the series $\sum_{n \in \mathbb{N}} P(X_m \neq 0) $ is divergent.

This is because $\forall n \in \mathbb{N} : P(X_n \neq 0) = c >0$, because the variables have identical distributions.

But I don't think it is helpful, because by Markov's inequality we also have for $a \ge 0$:

$$\mathbb P (|X| \ge a) \le \frac{\mathbb E(|X|)}{a}.$$

So it would seem that the series is divergent in $L^1.$ But why is it divergent almost surely?

Could I use the law of large numbers? If so, how? The variables satisfy all necessary conditions for the following to hold:

$$\frac{S_n}{n} = \frac{X_1 + ... + X_n}{n} \to \mathbb{E}X_1 \ \ \text{a.s.}$$

But does that imply that the series is divergent?

Could you help me finish that?

  • 2
    $\begingroup$ Your LLN approach is good if you assume $E[X_1] \neq 0$. Then $\frac{1}{n} \sum_{i=1}^n X_i \rightarrow E[X_1]$ almost surely, so for any $\epsilon$ you can (with prob 1) find an $N$ such that $E[X_1] + \epsilon \geq \frac{1}{n}\sum_{i=1}^n X_i \geq E[X_1] - \epsilon$ for all $n \geq N$. $\endgroup$
    – Michael
    May 1 '15 at 11:40
  • 3
    $\begingroup$ So if $E[X_1]$ is a real number such that $E[X_1] \neq 0$, you can in fact show that $\sum_{i=1}^n X_i$ diverges to either $\infty$ or $-\infty$ (depending on the sign of $E[X_1]$). If you just want to show the limit of $\sum_{i=1}^n X_i$ does not exist, just consider that the difference between successive terms is $X_{n+1}$, which is not converging to 0. $\endgroup$
    – Michael
    May 1 '15 at 11:49
  • 1
    $\begingroup$ @Michael Thank you. So $E[X_1] + \epsilon \geq \frac{1}{n}\sum_{i=1}^n X_i \geq E[X_1] - \epsilon$ means that the arithmetic mean is from a certain moment between $\mathbb{E}X_1 - \varepsilon $ and $\mathbb{E}X_1 + \varepsilon$. But how can I extract the sum itself from that? How can I conclude that the series is divergent? $\endgroup$
    – Bruce
    May 1 '15 at 12:12
  • 2
    $\begingroup$ If $E[X_1]=c>0$ then choose $\epsilon =c/2$ to conclude there is an $N>0$ such that $\frac{1}{n}\sum_{i=1}^n X_i \geq c/2$ for all $n \geq N$. So $\sum_{i=1}^n X_i \geq cn/2$ for all $n \geq N$. Take limit as $n\rightarrow\infty$. $\endgroup$
    – Michael
    May 1 '15 at 22:02


  1. Since $\mathbb{P}(X_1 \neq 0)>0$, there exists $\epsilon>0$ such that $\mathbb{P}(|X_1|>\epsilon)>0$.
  2. Conclude from $$\sum_{n \geq 1} \mathbb{P}(|X_n|>\epsilon)= \sum_{n \geq 1} \mathbb{P}(|X_1|>\epsilon) =\infty$$ and the Borel-Cantelli lemma that $|X_n(\omega)|>\epsilon$ happens infinitely often for almost all $\omega \in \Omega$.
  3. Deduce that $$\sum_{n \geq 1} X_n $$ is divergent almost surely.

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.