26 — Optional Stopping for Random Walks, Expected Hitting Times

We study a biased random walk and compute the expected hitting time of a level using martingales and OST (Optional Stopping Theorem).


Setup: Biased Random Walk

Let

  • $ {Z_k}_{k\ge 1} $ be iid with
    \(P(Z_k = +1) = p,\qquad P(Z_k=-1)=q,\qquad p>q,\; p+q=1.\)
  • $S_0=0$, $S_n=\sum_{k=1}^n Z_k$.

Then by LLN:
\(\frac{S_n}{n} \xrightarrow{a.s.} p-q>0.\)

Define the standard martingale: \(\psi(S_n)=\left(\frac{q}{p}\right)^{S_n},\qquad M_n=\psi(S_n).\) Since $E[(q/p)^{Z_{n+1}}]=1$, this is a MG.

Let \(T=T_a\wedge T_b,\qquad a<0<b\in\mathbb Z,\) where
\(T_x=\inf\{n\ge 0:S_n=x\}.\)

We know $T_b<\infty$ a.s. because the drift is positive.

The random variables $\psi(S_{T\wedge n})$ are bounded by $(p/q)^{-a}$, hence UI, which allows OST.


Hitting Probabilities (review from Lecture 26)

Using OST on $M_{T\wedge n}$:

\[E(M_{T\wedge n}) = M_0 = 1.\]

As $n\to\infty$: \(M_T = \begin{cases} \left(\frac{q}{p}\right)^a & \text{if }T_a<T_b,\\)6pt] \left(\frac{q}{p}\right)^b & \text{if }T_b<T_a. \end{cases} $$

Solving the linear system gives:

\[P(T_a<T_b) = \frac{\varphi(b)-\varphi(0)}{\varphi(b)-\varphi(a)},\qquad \varphi(x)=\left(\frac{q}{p}\right)^x.\]

Goal: Compute $E(T_b)$

We use the martingale: \(S_n - (p-q)n.\)

Indeed: \(E(Z_k-(p-q))=0 \quad\Rightarrow\quad \{S_n-(p-q)n\} \text{ is a MG}.\)

Apply OST to $T_b\wedge n$:

\[E\left[S_{T_b\wedge n} - (p-q)(T_b\wedge n)\right]=0.\]

Since $S_{T_b\wedge n}=b$ when $T_b\le n$, and ≤$b$ otherwise:

\[E(S_{T_b\wedge n}) = (p-q)E(T_b\wedge n).\]

Letting $n\to\infty$ (using MCT):

\[E(S_{T_b}) = b = (p-q)E(T_b).\]

Thus:

\[\boxed{E(T_b)=\frac{b}{p-q}}.\]

Why is ${S_{T_b\wedge n}}$ UI?

We need UI or DCT to exchange limit and expectation.

We know \(S_{T_b\wedge n} \ge \min_{k\ge 0} S_k.\)

And from page 2 of the notes (:contentReference[oaicite:1]{index=1}):

\[E\left(\min_{k\ge 0} S_k\right) = -\sum_{k=1}^\infty P(T_{-k}<\infty) = -\sum_{k=1}^\infty \left(\frac{q}{p}\right)^k > -\infty.\]

Since $0<q/p<1$, the geometric series converges.

Thus:

  • ${S_{T_b\wedge n}}$ is bounded above by $b$,
  • bounded below in $L^1$ by a finite constant,

so the family is UI.

Therefore OST applies and the limit–expectation exchange is valid.


Summary of Result

\[\boxed{ E(T_b)=\frac{b}{p-q} < \infty }\]

This matches the intuition: the drift is $p-q>0$, and it takes on average $b/(p-q)$ steps to climb from 0 to $b$.


Positive Supermartingales (from pages 3–4)

Let ${X_n}$ be a positive supermartingale. For any stopping time $T$:

\[E(X_0)\ge E(X_T)\ge E(X_\infty).\]

Key example:
\(X_n = e^{S_n - n/2},\qquad Z_k\sim N(0,1).\)

Here: \(E(e^{Z_k - 1/2}) = 1,\) so $X_n$ is a MG.

But $X_n\to 0$ a.s. and is not UI (page 4). OST still works using Fatou’s lemma because the MG is positive.


Final OST Statement for Positive Supermartingales

If $X_n\ge 0$ is a supermartingale and $T$ is a stopping time, then

\[E(X_0)\ge E(X_T).\]

If $X_n\to X_\infty$ a.s., then also:

\[E(X_T)\ge E(X_\infty).\]

This uses Fatou’s lemma and the positivity of $X_n$.


Comments