Martingales: The Mathematics of Fair Games
Set the static theory in motion: a process whose best forecast of tomorrow is today, with respect to the observer's own information. The definition is three lines of Chapter 8.1; the theorems are the most applied results in the course — and the negative examples are the most expensive to ignore.
In this subchapter
8.3.1The definition, and the reading
Given a filtration F₁ ⊆ F₂ ⊆ ⋯ (4.3 §5 — information arriving over time) and a process M₁, M₂, … adapted to it: M is a martingale if each Mn is integrable and
The conditional expectation — the machinery of 8.1, the variable defined by the two-property specification — of the next value, given everything known now, is the current value. A supermartingale replaces = by ≤ (drifts down on average: a gambler being bled by house edge); a submartingale by ≥ (drifts up).
Fair coin flips: Mn = Σi≤nYi with Yi = ±1 at probability 1/2 each — the running fortune of a gambler betting one unit per flip. Check the definition honestly: E[Mn+1|Fn] = E[Mn + Yn+1|Fn] = Mn + E[Yn+1] — the taking-out-what-is-known property of 8.1 Theorem 3 pulls Mn through (it is Fn-measurable, known), and the independence property kills the second term: E[Yn+1] = 0. The entire modern theory of fair pricing runs on exactly this two-line maneuver. Other members of the family: a random walk with iid mean-zero increments of any law; the product Mn = exp(ΣLi)/E[exp(ΣLi)] of likelihood ratios (the exponential martingale, behind pricing and large deviations); the fortune process under any betting strategy on a fair game — that is §2's theorem, not an accident.
Coin flips with P(H) = 0.49: the same fortune process is a supermartingale — E[Mn+1|Fn] = Mn − 0.02. The house edge is a supermartingale statement, quantified: 0.02 per flip per unit bet, in expectation, conditional on everything. The martingale framework converts "the casino wins in the long run" from folklore to a theorem with a rate — and every "near-fair" strategy analysis is a supermartingale estimate with the drift named. (Submartingale mirror: the casino's own book, drifting up — the same mathematics, the other side of the table.)
8.3.2Betting systems: the no-free-lunch theorem
Let M be a martingale adapted to (Fn), and let Hn be a predictable process: Hn is Fn−1-measurable (your stake for round n is decided before the n-th outcome — the formalization of "a strategy"). If HnMn is integrable, then the strategy's fortune
is again a martingale. No choice of stakes — however clever, adaptive, or data-driven — turns a fair game unfair in your favor.
Integrability: assumed where flagged. The conditional expectation: E[Wn+1 | Fn] = Wn + E[Hn+1(Mn+1 − Mn) | Fn]. The stake Hn+1 is Fn-measurable (predictable!) — take out what is known (8.1 Theorem 3): = Wn + Hn+1·E[Mn+1 − Mn | Fn] = Wn + Hn+1·0 — the martingale property kills the increment's forecast. Note the proof's honest crux: predictability is where it enters. If Hn+1 could depend on the (n+1)-th outcome itself, the taking-out step would be unlicensed and the theorem false — this is not a technicality; it is the mathematical boundary between strategy and clairvoyance. ∎
Read M as a fair (risk-neutral-priced) asset and H as a position: any dynamic, self-financed position in a fair asset is fair — the basis of every derivative-pricing argument ("the replicating portfolio is a martingale, hence its initial cost is the unique consistent price"). Read M as a market under the efficient-markets measure and the theorem is the weak-form efficiency statement: no signal computed from past prices (necessarily Fn−1-measurable — predictable) extracts expected value. The theorem's honesty is its hypothesis ledger: predictability (no lookahead), integrability (no explosions — §4's museum shows what dodges it), and the martingale property itself (the fairness you must prove, not assume).
8.3.3Optional stopping: the fine print
For a martingale M and a stopping time T (an integer-valued random time with {T ≤ n} ∈ Fn for every n — "the decision to stop by time n is made from information available at n"): E[MT] = E[M0] holds under any of the sufficient conditions: (a) T bounded; (b) T finite a.s. and M's increments uniformly bounded with E[T] < ∞; (c) MT dominated by an integrable Z. Without some such condition, the conclusion can fail — §4.
Define the stopped process MTn = Mn∧T. It is a martingale: MTn+1 − MTn = 1{T>n}·(Mn+1 − Mn), and the indicator is Fn-measurable (complement of {T ≤ n}, a stopping-time property) — predictable, so Theorem 1's argument runs verbatim (stop betting after T: that is a strategy). Then E[MTn] = E[M0] for every n, and passing n → ∞ to the limit — the step that needs exactly the hypotheses of the theorem, an MCT/DCT exercise on the case in hand — gives E[MT] = E[M0]. The bounded-T case needs no limit passage: MTN = MT for N ≥ T, done. ∎-sketch with the limit-step flagged, as always.
8.3.4The failure museum: doubling, and the ruin it hides
The classic system: bet 1; on loss, double; on first win, stop. On a fair coin, with unlimited credit and no table limits, the game is fair — Theorem 1 — and T (first heads) is finite a.s. (6.3 Example 3's runs logic). The seduction: MT = +1 always — profit 1 with probability 1, yet by optional stopping... the check of the hypotheses: the time is fine — T geometric with p = 1/2 has E[T] = 2 < ∞ — but the fortune is the issue: the drawdown before T reaches 2T−1, and E[2T−1] = Σn (2n−1)·2−n = ∞ (each term is ≥ 1/2). So the fortune process is not uniformly integrable, condition (b) fails on the increment-side, and the honest accounting is: you gain 1 with probability 1, financed by a tail of astronomical losses with vanishing-but-not-null probability and infinite expectation. Every hypothesis of optional stopping fails where the money is. This is not a loophole — it is the theorem working: the "guaranteed profit" is exactly the ∞−∞ that 5.1 §5's discipline refuses to sign. Casinos close the same loophole with table limits — the constraint is not commercial, it is the missing integrability, priced.
The symmetric ±1 random walk on ℤ, started at 0, stopped at T = first hit of +1. T is finite a.s. (recurrence — flag: provable via Borel–Cantelli on refresh times), MT = 1 — a.s. profit 1 from a fair game, with no doubling at all. The fine print: E[T] = ∞, and the honest proof is optional stopping applied to the martingale Nn = M²n − n (check the martingale property: increments of M² decompose as cross-term (mean zero by independence) plus 1) at the bounded time T ∧ N: E[M²T∧N] = E[T∧N]. The left side is ≥ P(T ≤ N)·1 → 1 as N → ∞ (T finite a.s.), so E[T] ≥ limN E[T∧N] = ∞ by monotone convergence. Again the hypothesis ledger — condition (b)'s E[T] < ∞ — fails, and the a.s.-profit is a genuine counterexample to the hypothesis-free statement. The two exhibits of the museum share one lesson: martingale conclusions are exactly as strong as their integrability fine print, and the fine print lives in the tails of the stopping time.
8.3.5Convergence: martingales settle
A martingale bounded in L² converges a.s. and in L² (Doob); more generally, an L¹-bounded martingale converges a.s., and the limit is genuine in L¹ exactly when the process is uniformly integrable — the exact concept the failure museum was circling. Behind the proof: the upcrossing inequality — the expected number of upcrossings of any interval [a,b] by a martingale is bounded by E[(Mn − a)−]/(b−a); a bounded-oscillation argument (countably many rational intervals) then forces a.s. convergence. The upcrossing inequality is the last great ε-management argument of the course, and its pattern — control crossings of a fixed gap to control oscillation — is worth learning as a reusable move. We state the theorem, prove the upcrossing inequality's core step, and flag the assembly as an exercise in everything the course has trained.
Every "learning settles, markets converge, filtering stabilizes" claim in applied work is a martingale-convergence argument in costume: Bayesian posterior updating is a martingale (in the conditional probabilities) under the true-measure's filtration — so posterior beliefs converge a.s.; the Kalman filter's error is a supermartingale under detectability; self-financed trading against a fair market, with bounded risk, converges a.s. to a limiting wealth — and the limit is what the no-free-lunch theorem says it must be (a fair price, not a profit machine). The a.s. mode matters here — 7.1's hierarchy — because pathwise settling is exactly the practical claim.
8.3.6What the theory powers
- Pricing: fundamental theorems of asset pricing — under a "no free lunch" condition (precisely: no-arbitrage plus a technical closure), prices are expectations under an equivalent martingale measure: a Radon–Nikodym change of measure (8.2!) applied to a martingale process. Chapters 8.1–8.3 in one sentence.
- Stochastic calculus: continuous-time martingales (Itô calculus) extend this subchapter word for word — optional stopping, convergence, and no-arbitrage conclusions — with integrals of Chapter 5's type against Brownian paths. The course's machinery transfers without conceptual change.
- Statistics and ML: sequential testing (stop when the likelihood-ratio martingale crosses a threshold — optional stopping, with the same fine print), online learning (regret bounds as supermartingale estimates), and bandit algorithms all run on the theorems of this subchapter.
- Risk: the failure museum is the risk textbook — every leverage strategy that "cannot lose" is a doubling system wearing a suit, and the theorem identifies exactly which hypothesis (integrability, capital constraint, table limit) its survival depends on.
A fair game stays fair under every strategy you can compute from the past — and every system that claims otherwise is spending integrability it does not have.
The theory is built, end to end: from σ-algebras to measure to integration to independence to convergence to conditioning to motion. The final subchapter closes the course: where each piece of the machinery points (measure-theoretic probability's graduate curriculum, stochastic calculus, large deviations, ergodic theory), and how to keep the discipline you now hold — naming the space, the measure, and the theorem's fine print — when the mathematics gets sold to you secondhand.