Martingale Convergence Theorems

The martingale convergence theorems guarantee that a martingale which stays bounded in an appropriate sense settles down to a limiting random variable, providing a versatile route to almost-sure convergence.

Definition

The martingale convergence theorems are results stating that a martingale bounded in L1 converges almost surely and that a uniformly integrable martingale converges almost surely and in L1 to a random variable that closes the martingale as a conditional expectation.

Scope

This topic covers Doob's upcrossing inequality and maximal inequalities, the almost-sure convergence of L1-bounded martingales, convergence in mean for uniformly integrable martingales and the notion of a closing variable, Lp-bounded martingale convergence, and the backward martingale convergence theorem with its applications to the strong law of large numbers.

Core questions

• How does the upcrossing inequality force a bounded martingale to converge?
• What is the difference between almost-sure and mean convergence for martingales?
• What does uniform integrability add, and what is a closing variable?
• How do backward martingales yield the strong law of large numbers?

Key theories

Doob's upcrossing inequality and L1-bounded convergence

Bounding the expected number of times a martingale crosses any interval shows it cannot oscillate indefinitely, so an L1-bounded martingale converges almost surely to a finite limit.

Uniform integrability and L1 convergence

A uniformly integrable martingale converges in L1 as well as almost surely, and equals the conditional expectations of its limit, so it is closed by a single integrable random variable, the form needed for many applications.

Clinical relevance

Martingale convergence underlies proofs of the strong law of large numbers, the convergence of Bayesian posterior beliefs as data accumulate, Levy's zero-one law, and the almost-sure limits of branching-process population sizes, making it a recurring engine for almost-sure asymptotics.

History

Doob established the convergence theorem and the upcrossing argument in the 1940s and presented them in his 1953 treatise, and the uniformly integrable and backward versions, together with Levy's downward and upward theorems, became standard parts of the graduate probability curriculum.

Key figures

• Joseph Doob
• Paul Levy
• David Williams

Seminal works

• williams1991

Frequently asked questions

When does a martingale converge?

If it stays bounded in L1, meaning its expected absolute value is bounded over time, it converges almost surely; uniform integrability additionally gives convergence in mean to a closing variable.

What is an upcrossing?

An upcrossing of an interval is an occasion when the martingale moves from below the lower endpoint to above the upper endpoint; bounding the expected number of these crossings proves convergence.