The Optional Stopping Theorem: When You Can't Beat a Fair Game

Precise statement: martingales, stopping times, and the three hypotheses

Fix a probability space (Ω, ℱ, ℙ) with a filtration (ℱₙ)ₙ≥₀ — an increasing chain ℱ₀ ⊂ ℱ₁ ⊂ … of σ-algebras encoding information up to time n. A process (Xₙ) is a martingale if each Xₙ is ℱₙ-measurable, integrable (𝔼|Xₙ| < ∞), and satisfies the fair-game identity 𝔼[Xₙ₊₁ | ℱₙ] = Xₙ. A stopping time τ: Ω → {0,1,…,∞} is a random time with {τ ≤ n} ∈ ℱₙ for all n — you decide to stop using only information available so far, no peeking into the future.

The Optional Stopping (a.k.a. Optional Sampling) Theorem asserts 𝔼[X_τ] = 𝔼[X₀] provided at least one of the following holds: (A) τ is bounded, τ ≤ N a.s.; (B) τ < ∞ a.s. and the stopped process is dominated, |X_{τ∧n}| ≤ Y with 𝔼[Y] < ∞; or (C) 𝔼[τ] < ∞ and the conditional increments are bounded, 𝔼[|Xₙ − Xₙ₋₁| | ℱₙ₋₁] ≤ c. For a submartingale (𝔼[Xₙ₊₁|ℱₙ] ≥ Xₙ) the equality becomes 𝔼[X_τ] ≥ 𝔼[X₀]; for a supermartingale, ≤.

The picture: a fair game stays fair, whenever you quit

Imagine your fortune Xₙ playing a fair game — each round adds a mean-zero, unpredictable increment. The martingale property says that on average, from where you stand now, you neither gain nor lose next round: 𝔼[Xₙ₊₁ | ℱₙ] = Xₙ. The seductive idea is that a smart exit rule could tilt the odds: quit while ahead, keep playing while behind. Optional stopping says this is an illusion — the exit rule τ, because it cannot see the future, is powerless to bias the expectation.

Geometrically, think of the martingale as a random path whose conditional slope is always zero. Stopping is choosing when to freeze the path, and the freezing decision is measurable with respect to the past. The stopped process Xₙ^τ := X_{τ∧n} is still a martingale — freezing preserves fairness at every finite horizon. The only subtlety is the limit n → ∞: does 𝔼[X_{τ∧n}] → 𝔼[X_τ]? The three hypotheses are exactly the conditions that license this passage to the limit; without one, mass can 'escape to infinity' and the fair game appears beatable.

Key idea of the proof: stop early, then take the limit safely

The mechanism has two moves. Move 1 (finite horizon is free). For any stopping time τ, the stopped process Xₙ^τ = X_{τ∧n} is a martingale. The one-line reason: X_{τ∧n} − X_{τ∧(n−1)} = 1_{τ ≥ n}(Xₙ − Xₙ₋₁), and the indicator 1_{τ ≥ n} = 1 − 1_{τ ≤ n−1} is ℱₙ₋₁-measurable (since {τ ≤ n−1} ∈ ℱₙ₋₁). So 𝔼[X_{τ∧n} − X_{τ∧(n−1)} | ℱₙ₋₁] = 1_{τ ≥ n} · 𝔼[Xₙ − Xₙ₋₁ | ℱₙ₋₁] = 0. Telescoping gives 𝔼[X_{τ∧n}] = 𝔼[X₀] for every finite n — this needs no extra hypothesis.

Move 2 (pass to the limit). Since τ < ∞ a.s., X_{τ∧n} → X_τ pointwise. We need 𝔼[X_{τ∧n}] → 𝔼[X_τ]. Under (A) the sequence is finite (τ∧n = τ for n ≥ N). Under (B), Dominated Convergence applies via the envelope Y. Under (C), a telescoping bound |X_{τ∧n} − X₀| ≤ Σ_{k=1}^{τ} |Dₖ| gives an integrable majorant (its expectation is ≤ c·𝔼[τ] by a conditional Wald bound), so DCT again applies. The unifying condition is uniform integrability of {X_{τ∧n}}, which is precisely what makes L¹-convergence follow from a.s. convergence.

Canonical example: gambler's ruin and Wald's identity

Let (ξᵢ) be i.i.d. with ℙ(ξ = +1) = ℙ(ξ = −1) = ½, and Sₙ = ξ₁ + … + ξₙ the simple symmetric random walk, S₀ = 0. Then Sₙ is a martingale. For 0 < a and start at position k with 0 < k < a, let τ = inf{n : Sₙ = 0 or Sₙ = a} be the exit time of the interval. Here τ has 𝔼[τ] < ∞ and increments are bounded (|Dₙ| = 1), so hypothesis (C) holds. Optional stopping gives 𝔼[S_τ] = S₀ = k. Since S_τ ∈ {0, a}, writing p = ℙ(hit a first): 0·(1−p) + a·p = k, so ℙ(hit a before 0) = k/a — the classic gambler's-ruin formula, extracted in one line.

To get the expected duration, use the companion martingale Mₙ = Sₙ² − n (a martingale because 𝔼[Sₙ₊₁² | ℱₙ] = Sₙ² + 1). Optional stopping yields 𝔼[S_τ²] − 𝔼[τ] = k², and since 𝔼[S_τ²] = a²·(k/a) = ak, we get 𝔼[τ] = k(a − k). The same two martingales, Sₙ and Sₙ² − n, deliver both hitting probabilities and expected hitting times.

Why the hypotheses matter: the doubling counterexample

Drop every integrability hypothesis and the theorem is false. Take Sₙ the simple symmetric random walk from 0 and τ = inf{n : Sₙ = 1}, the first time you are one unit ahead. Because the walk is recurrent, τ < ∞ almost surely, and X_τ = 1 identically. Hence 𝔼[X_τ] = 1 ≠ 0 = 𝔼[X₀] — the fair game looks beaten. What fails? Here 𝔼[τ] = ∞, so (C) is out; the stopped values X_{τ∧n} are unbounded below (you can dip arbitrarily negative before hitting +1), so (A) and (B) fail too; and {X_{τ∧n}} is not uniformly integrable. The martingale-doubling betting scheme is the same phenomenon dressed as a casino strategy: it 'guarantees' a $1 profit but requires an unbounded bankroll and unbounded time.

The moral connects to uniform integrability and Doob's convergence theorems: 𝔼[X_τ] = 𝔼[X₀] holds iff the family {X_{τ∧n}} is uniformly integrable, the sharp necessary-and-sufficient refinement. This ties optional stopping to the Dominated Convergence Theorem and to martingale L¹/Lᵖ convergence.

Why it matters: from random walks to the pricing of options

Optional stopping is the workhorse behind a startling range of results. In probability it yields Wald's identities (𝔼[S_τ] = 𝔼[τ]·𝔼[ξ] for i.i.d. sums), exit-time and exit-position laws for random walks and Brownian motion, and hitting-probability formulas that solve discrete Dirichlet problems. Its continuous-time analogue (Doob's optional sampling for right-continuous martingales, with τ a stopping time and X_{τ∧t} uniformly integrable) underlies the strong Markov property and Itô-calculus arguments.

In mathematical finance, the fundamental theorem of asset pricing says a market is arbitrage-free iff a discounted-price martingale measure exists; optional stopping is precisely why no stopping-based trading strategy — including early exercise of an American option below its fair value — can produce risk-free profit. In analysis and PDE, martingale representations of harmonic functions use optional stopping to derive mean-value and maximum principles. In algorithms and statistics it certifies sequential-test validity (Wald's SPRT) and powers concentration bounds. It is, quite literally, the theorem that formalizes 'you can't beat a fair game.'

Sufficient hypotheses for 𝔼[X_τ] = 𝔼[X₀], and what each rules out. Here X is a martingale, τ a stopping time, and increments are Dₙ = Xₙ − Xₙ₋₁.

Hypothesis	Conclusion	What it forbids / typical use
(A) τ ≤ N almost surely (bounded stopping time)	𝔼[X_τ] = 𝔼[X₀] always	No hypothesis on X needed; the cleanest case. Rules out 'wait arbitrarily long'.
(B) τ < ∞ a.s. and |X_{τ∧n}| ≤ Y with 𝔼[Y] < ∞ (dominated)	𝔼[X_τ] = 𝔼[X₀]	Bounded game payoffs. Rules out unbounded fortunes like the doubling martingale.
(C) 𝔼[τ] < ∞ and 𝔼[|Dₙ| | ℱₙ₋₁] ≤ c a.s. (bounded increments)	𝔼[X_τ] = 𝔼[X₀]	Random walks with fixed step size; underlies Wald's identity and gambler's ruin.
(D) X_{τ∧n} uniformly integrable and τ < ∞ a.s.	𝔼[X_τ] = 𝔼[X₀]	The sharpest single sufficient condition; (A)–(C) are special cases of it.
None (only τ < ∞ a.s.)	Can FAIL: 𝔼[X_τ] > 𝔼[X₀]	Doubling strategy / simple random walk hitting +1: 𝔼[X_τ] = 1 ≠ 0 = X₀.