# Durret problem 4.8.3 - Random walk and optional sampling theorem application

Exercise 4.8.3 Durret states

Let $$S_n= \xi_1 + \ldots \xi_n$$, $$(\xi_i)_i$$ independent with $$0$$ mean and $$Var(\xi_i)=\sigma ^2$$. Then $$S_n^2 - n \sigma^2$$ is a martingale. Let consider the stopping time $$T= \inf \{ n: |S_n|> a \}$$.

Show $$E[T]> \frac{a^2}{\sigma^2}$$.

I need to know if my argument is okay, especially the application of the optional stopping theorem!

Of course the filtration is the canonical one $$F_n = \sigma(\xi_i, i \leq n)$$ and is easy to show that $$S_n^2 - n \sigma^2$$ is a $$(F_n)_n$$-adapted martingale.

Morever, $$T$$ is a stopping time since $$\{T= n \}$$ can be written as \begin{align} ( \cap_{i=1}^{N-1} \{ |S_i|a \} \end{align} and by construction this belongs to $$F_n$$

To show the claim, I need to use the optional stopping theorem applied to the given martingale. The standard argument is that $$T$$ is not bounded, so I consider the stopped martingale $$S_{n ∧ T} - (n ∧ T) \sigma^2$$.

For every $$n$$, $$n∧T$$ is bounded and $$(n∧T) \rightarrow_n T$$ Here I should show that $$P(T< +\infty)=1$$ a.s, but don't know how

Hence, by OST applied to stopping times $$T ∧ n$$ and $$0$$:

\begin{align} E[S_T - T] = E[\lim_n S_{T ∧ n} - (T ∧ n) \sigma^2] = \lim_n E[S_{T ∧ n} - (T ∧ n) \sigma ^2] = E[S_0]=0 \end{align}

where I used DCT.

Hence $$E[T] \sigma^2= E[S_T^2]$$, but when I stop at $$T$$ I have by definition: $$|S_T|>a$$ and hence follows $$E[T] \sigma^2 = E[S_T^2]>E[a^2]=a^2$$ and hence

\begin{align} E[T]> \frac{a^2}{\sigma^2} \end{align}

EDIT

To apply DCT, I note that $$S_{ n \wedge T}^2 \leq (a+1)^2$$ and hence, by OST, $$E[T \wedge n]=E[S_{ n \wedge T}^2] \leq (a+1)^2$$ and hence I can take the limit in $$E[S_{ n \wedge T}^2 - T \wedge n]=0$$ and obtain the thesis

• Actually, I have noticed something: what is the dominating function that allows you to use DCT? Commented Nov 11, 2019 at 16:40
• @TheoreticalEconomist I thought $S_{T ∧ N} - (T ∧ n) \sigma^2 < a^2$
– VoB
Commented Nov 11, 2019 at 16:43
• I don't see why that's true. Bear in mind that what you actually need is that $\vert S_{T\wedge n} - (T\wedge n)\sigma^2 \vert \le X$ where $X$ is some integrable random variable. $X=a^2$ would do, if the inequality were true, but I just don't see it. Commented Nov 11, 2019 at 16:46
• @TheoreticalEconomist you're right. Could you confirm the following? I think it should work: $|S_{n \wedge T}| \leq |S_{n \wedge T-1} | + |S_{ n \wedge T -1} - S_{ n \wedge T} | < a +1$ since the steps of the r.w. are of size 1, and so I have a uniform bound
– VoB
Commented Nov 11, 2019 at 16:56
• The idea is good, but I think your last inequality should actually be $\le a + \vert \xi_{T\wedge n}\vert$, which is fine because this is integrable. Note in particular that the inequality is weak, not strict. Commented Nov 11, 2019 at 17:10

I just want to summarise the discussion in the comments (as some of them have been deleted), give an argument that allows you to interchange taking limits and expectations, and make a remark about your application of the DCT.

First, you don't need to show that $$P(T<\infty)=1$$. You are able to apply the optional sampling theorem since $$T\wedge n \le n$$, which guarantees that $$T\wedge n$$ is a bounded stopping time.

We will now justify the step $$E\left[\lim_{n\to \infty} S_{T\wedge n}^2 -(T\wedge n)\sigma^2\right] = \lim_{n\to \infty} E\left[ S_{T\wedge n}^2 -(T\wedge n)\sigma^2\right], \tag{*}\label{*}$$ in two separate steps using the dominated and monotone convergence theorems respectively.

As you argue in the comments, we observe first that $$\vert S_{T\wedge n} \vert \le \vert S_{(T\wedge n) -1} \vert + \vert S_{T\wedge n} - S_{(T\wedge n) -1}\vert \le a + \vert \xi_{T\wedge n} \vert.$$ Thus, $$\vert S_{T\wedge n} \vert^2 = S_{T\wedge n}^2 \le \left(a + \left\vert \xi_{T\wedge n} \right\vert\right)^2,$$ where we know that the right-hand side is integrable. This means that $$E\left[\lim_{n\to \infty} S_{T\wedge n}^2 \right] = \lim_{n\to \infty} E\left[ S_{T\wedge n}^2 \right].$$

Now, observe that $$T\wedge n$$ is non-negative, and non-decreasing in $$n$$. Thus, we can use the monotone convergence theorem to conclude that $$E\left[\lim_{n\to \infty}(T\wedge n)\sigma^2\right] = \lim_{n\to \infty} E\left[(T\wedge n)\sigma^2\right].$$

Putting these two steps together gives us \eqref{*}.

Note that the argument you make in your edit is not correct, because it does not allow you to apply the DCT. The DCT requires that you find an integrable random variable $$Z$$ such that $$T\wedge n \le Z$$. What you've done is instead found an upper bound for $$E[T\wedge n]$$, which is not sufficient to apply the DCT.

The rest of your post looks fine.

• Many thanks, you're right, I was thinking about a RW!
– VoB
Commented Nov 12, 2019 at 11:49