Martingale

Why martingale matters

• Financial models. Risk-neutral pricing reduces an option's value to a martingale expectation; Black-Scholes is a martingale measure on geometric Brownian motion.
• Stochastic calculus. Every Itô integral against Brownian motion is a martingale; the Itô isometry, Girsanov's theorem, and martingale representation underpin all of continuous-time finance and stochastic control.
• Gambling theory. The optional stopping theorem proves no clever strategy beats a fair coin — including the historical "martingale" doubling strategy, which technically loses in expectation when bankroll or time is bounded.
• Hypothesis testing. Sequential tests use likelihood-ratio martingales; Ville's inequality gives anytime-valid p-values that don't suffer from peeking.
• Branching processes. Galton–Watson trees: dividing population by E[offspring]^n yields a martingale. Convergence theorem gives the extinction probability.
• Pólya urns and Bayesian posteriors. Posterior probabilities form a martingale under the prior — convergence is automatic.
• Concentration inequalities. Azuma–Hoeffding bounds for martingales with bounded increments are the foundation of probabilistic combinatorics and online learning.

Formal setup

Fix a filtered probability space (Ω, F, (Fₙ)_{n≥0}, P). A sequence (Xₙ) of random variables is a martingale with respect to (Fₙ) if:

(1) Xₙ is Fₙ-measurable           (adapted)
(2) E[|Xₙ|] < ∞                    (integrable)
(3) E[Xₙ₊₁ | Fₙ] = Xₙ a.s.        (martingale property)

The natural filtration is Fₙ = σ(X₀, X₁, ..., Xₙ) — the σ-algebra generated by the process up to time n. Replacing equality with ≥ gives a submartingale; with ≤, a supermartingale.

Canonical examples

• Simple random walk. Sₙ = X₁ + ... + Xₙ where Xᵢ are i.i.d. with E[Xᵢ] = 0. Then Sₙ is a martingale; Sₙ² − n is also a martingale.
• Wealth in a fair game. Bet $1 on a fair coin; your wealth Xₙ is a martingale.
• Pólya urn. Start with one red and one blue ball; draw, return with another of the same color. The fraction of red balls is a bounded martingale — converges a.s. to a Beta(1,1) = Uniform(0,1) limit.
• Likelihood ratio. Lₙ = ∏ q(Xᵢ)/p(Xᵢ) under P-distribution p is a martingale — basis for Wald's sequential probability ratio test.
• Conditional expectation tower. For any L¹ random variable Y, Mₙ = E[Y | Fₙ] is automatically a martingale (Doob's martingale).

Optional stopping theorem

A stopping time τ is a random time such that the event {τ ≤ n} is Fₙ-measurable — you can decide whether to stop using only information available so far, no peeking ahead. Doob's optional stopping theorem says that under any of these regularity conditions:

• τ is bounded (τ ≤ N a.s. for some constant N), or
• Xₙ is uniformly bounded and τ < ∞ a.s., or
• E[τ] < ∞ and the increments |Xₙ₊₁ − Xₙ| are bounded,

then E[X_τ] = E[X₀]. The simple-random-walk-hits-+1 example violates the third condition (E[τ] = ∞) — and the conclusion fails: E[X_τ] = 1, not 0.

Martingale convergence theorem

If Xₙ is a martingale with sup_n E[|Xₙ|] < ∞, then Xₙ → X_∞ almost surely for some X_∞ with E[|X_∞|] < ∞. The proof uses Doob's upcrossing inequality: the expected number of times the path crosses any interval [a, b] is bounded — so it can only oscillate finitely many times, hence must converge.

Caveat: a.s. convergence does not imply L¹ convergence. The doubling strategy on a fair coin is a martingale that converges a.s. to a finite limit but has E[Xₙ] = E[X₀] for all n — yet E[X_∞] can be different. L¹ convergence requires uniform integrability.

Doob's maximal inequality

For a non-negative submartingale Xₙ and λ > 0:

P(max_{0 ≤ n ≤ N} Xₙ ≥ λ) ≤ E[X_N] / λ

This is the martingale analog of Markov's inequality — but it bounds the running maximum, not just a single time. The L^p version (Doob's L^p inequality) gives ‖max Xₙ‖_p ≤ (p/(p−1)) ‖X_N‖_p for p > 1. These bounds underpin Burkholder–Davis–Gundy inequalities and most of modern probabilistic analysis.

Connection to mathematical finance

The fundamental theorem of asset pricing (Harrison–Kreps 1979, Delbaen–Schachermayer 1994): a discrete-time market is arbitrage-free if and only if there exists an equivalent martingale measure Q under which discounted asset prices are martingales. Pricing a derivative reduces to computing E_Q[discounted payoff].

For geometric Brownian motion dS_t = μ S_t dt + σ S_t dW_t, a Girsanov change of measure absorbs the drift μ into a new Brownian motion, leaving the discounted price as a martingale. The expectation E_Q[e^(-rT)(S_T − K)⁺] then evaluates to the Black–Scholes formula — and that's where the closed form comes from.

Common misconceptions

• "A martingale always converges." Only if it's bounded in L¹. The simple random walk is a martingale that does not converge.
• "No strategy beats a martingale." True for stopping times satisfying optional-stopping regularity. False if you allow unbounded τ — random walks hit any level eventually.
• "Fair game means equal probability of winning." Fair game means conditional expectation of next state equals current — no edge in expectation, regardless of variance or skew.
• "Martingales are about gambling." Originally yes, but the modern theory is the language of stochastic processes — Itô calculus, filtering theory, sequential analysis, and machine learning concentration bounds all rest on it.
• "Submartingales drift down." Reverse: submartingales drift up. The naming follows convex-function analogy: super = above the diagonal, sub = below.
• "The doubling strategy works." It's a supermartingale once you account for finite bankroll; expected loss is bounded but eventual loss is certain.

Azuma–Hoeffding inequality

If (Xₙ) is a martingale with bounded increments |Xₙ − Xₙ₋₁| ≤ cₙ, then for any t > 0:

P(|X_N − X₀| ≥ t) ≤ 2 exp(−t² / (2 ∑ cₙ²))

This Gaussian-tail bound is the workhorse of probabilistic combinatorics — random graph properties, randomized algorithms, online learning regret bounds — wherever you can express a quantity as a martingale with bounded differences.