Martingale Convergence Theorems

Core Idea

The martingale convergence theorems characterize when a martingale (or submartingale) converges as time goes to infinity. Doob's forward convergence theorem states that any L¹-bounded martingale converges almost surely. However, a.s. convergence alone does not guarantee L¹ convergence — for that, the martingale must be uniformly integrable. The L² martingale convergence theorem gives a cleaner sufficient condition: if sup E[M_n²] < ∞, then M_n converges both a.s. and in L². These results are the backbone of asymptotic theory for stochastic processes and underpin applications from Bayesian updating to the optional stopping theorem in unbounded time.

Explainer

The martingale convergence theorems answer the fundamental question: when does a martingale have a well-defined long-run limit? The answer turns on a hierarchy of integrability conditions, each giving stronger conclusions. These results, primarily due to Joseph Doob, form the asymptotic foundation of martingale theory and are essential wherever one needs to take limits of conditional expectations — Bayesian inference, branching processes, ergodic theory, and mathematical finance.

Doob's forward convergence theorem is the starting point. If {M_n, ℱ_n} is a martingale (or submartingale) with sup_n E[|M_n|] < ∞, then M_n converges to a finite limit M_∞ almost surely. The proof uses Doob's upcrossing inequality: the expected number of times {M_n} crosses upward through any interval [a,b] is bounded by E[(M_N - a)⁺]/(b-a), which is finite under the L¹-bound. Finite expected upcrossings for every rational interval [a,b] forces the sequence to converge on a set of full measure — a path that oscillates infinitely between two values would generate infinitely many upcrossings, contradicting the bound.

However, almost sure convergence is weaker than one might hope. The classic counterexample is M_n = ∏_{k=1}^n (2X_k) where X_k are i.i.d. Bernoulli(1/2) — this satisfies E[M_n] = 1 for all n but converges a.s. to 0, losing all its mass. The missing property is uniform integrability (UI): a family {M_n} is UI if sup_n E[|M_n| · 1_{|M_n| > K}] → 0 as K → ∞. Uniform integrability is necessary and sufficient for upgrading a.s. convergence to L¹ convergence, and it exactly characterizes the martingales that are "closed" — those of the form M_n = E[X | ℱ_n] for some integrable terminal variable X. In the closed case, M_∞ = E[X | ℱ_∞] and E[M_∞] = E[M_0], so no mass is lost.

The L² martingale convergence theorem provides the cleanest sufficient condition: if sup_n E[M_n²] < ∞, then M_n converges a.s. and in L². The proof exploits the orthogonality of martingale increments: E[(M_j - M_{j-1})(M_k - M_{k-1})] = 0 for j ≠ k, so Var(M_n) = Σ_{k=1}^n E[(ΔM_k)²]. The L²-bound forces this series to converge, making {M_n} Cauchy in L², and L² completeness gives the limit. L²-boundedness implies uniform integrability (by the de la Vallée-Poussin criterion), so the L¹ conclusions follow automatically. This is the workhorse version of the theorem in applications — it applies to square-integrable martingales, which include most naturally arising examples.

These convergence results have immediate consequences throughout probability. In Bayesian statistics, the posterior given n observations is a martingale that converges to the true parameter (on the support of the prior). In branching processes, the normalized population size W_n = Z_n/μ^n is a martingale whose convergence (or lack thereof) determines whether the population grows exponentially or dies out. In the theory of stochastic processes, martingale convergence is the main tool for proving laws of large numbers, 0-1 laws (Lévy's version: E[X | ℱ_n] → E[X | ℱ_∞] a.s. and in L¹), and for justifying the passage to continuous time via discrete approximations.