I have a solution to the following exercise and I want to know if there is another way to solve it:

Let $X_1,X_2,\dots$ be i.i.d. real r.v. with $\operatorname{Var}(X_1)=1,\ E(X_1)=0$ and let $S_n=X_1+\cdots + X_n,\quad n\in\mathbb{N}$

Deduce $\lim\limits_{n\to \infty}E(|S_n|)=\infty$

Solution: if $N\sim\mathcal{N}(0,1)$ is a r.v. then $p:=P(N\geq 1)>0$ and the CLT gives $$P\left(\left|\frac{S_n}{\sqrt{n}}\right|\geq 1\right)\xrightarrow{n\to\infty}P(|N|\geq 1)=2p$$

With such $p>0$ and $n_0\in\mathbb{N}$ by Markov's inequality $$\forall n\geq n_0,\quad E(|S_n|)\geq \sqrt{n}P(|S_n|\geq \sqrt{n})\geq p\sqrt{n}$$ hence $\lim\limits_{n\to \infty}E(|S_n|)=\infty$

For me this $p$ just looks like something arbitrary and not really intuitively. It can be with Markov chains or martingales or something like that.

  • $\begingroup$ Are you sure that the sum diverges? As a random walk it is symmetrical, so it would stay centered. $\endgroup$ – Cehhΐro Jan 16 '17 at 16:20
  • $\begingroup$ @O.VonSeckendorff Unless the solution is wrong. But I'm pretty confident that it is not...? $\endgroup$ – MarcE Jan 16 '17 at 16:37
  • $\begingroup$ Are the $X_n$ always normally distributed? $\endgroup$ – Cehhΐro Jan 16 '17 at 16:37
  • $\begingroup$ @O.VonSeckendorff We don't know the distribution of the $X_n$ besides of the expected value and the variance $\endgroup$ – MarcE Jan 16 '17 at 16:39

With a martingale argument, you can drop the assumption that the variances exist.

The process $(S_n)_{n\in\mathbb{N}}$ is a martingale and so $(|S_n|)_{n\in\mathbb{N}}$ is a submartingale. This means that the sequence $\mathbb{E}(|S_n|)$ is non-decreasing, and so it has a limit (possibly $+\infty$). If the limit is finite, then $\sup_n \mathbb{E}(|S_n|)<\infty$ so that $(S_n)_{n\in\mathbb{N}}$ is an $L^1$-bounded martingale.

By the martingale convergence theorem, $S_n$ converges almost surely to some random variable $Y$. On the other hand, the process $S_n-X_1=X_2+\cdots + X_n$ is an identically distributed martingale which converges to a limit $Z$, where $Z$ is independent of $X_1$ and has the same distribution as $Y$.

This gives $Z+X_1\stackrel{D}{=}Z$ where $Z$ is independent of $X_1$. Taking characteristic functions gives $\varphi_Z(t)\varphi_{X_1}(t)=\varphi_Z(t)$ for all $t$. Since $\varphi_Z(t)$ is continuous and takes the value 1 at $t=0$, we can divide out to obtain $\varphi_{X_1}(t)=1$ for all sufficiently small $t$. You can now show that $\mathbb{P}(X_1=0)=1$, and so $S_n\equiv 0$ for all $n$.

This is the only one of these random walks that is $L^1$-bounded.

| cite | improve this answer | |
  • $\begingroup$ Thank you very much for your answer! $\endgroup$ – MarcE Jan 22 '17 at 12:57
  • $\begingroup$ Glad to help out. $\endgroup$ – user940 Jan 22 '17 at 14:12

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.