Martingale

A martingale is a stochastic process for which the conditional expected value at a future time, given the information currently available, equals its current value. It formalizes the idea of a fair game: knowing the past provides no expected gain or loss in the process itself. The definition concerns conditional averages, not individual trajectories, which may fluctuate substantially. Martingales are defined relative to both a probability measure and a specified flow of information. (math.dartmouth.edu)

Mathematical definition

Let $(\Omega,\mathcal F,\mathbb P)$ be a probability space, and let $(\mathcal F_n)_{n\geq0}$ be a filtration: an increasing sequence of sigma-algebras representing the information available at successive times. A sequence of real-valued random variables $(M_n)_{n\geq0}$ is a martingale with respect to this filtration if:

1. Adaptedness: $M_n$ is $\mathcal F_n$-measurable, so its current value is determined by current information.

2. Integrability: $\mathbb E[|M_n|]<\infty$ for every $n$.

3. Martingale property:

   $$\mathbb E[M_{n+1}\mid\mathcal F_n]=M_n \quad\text{almost surely}.$$

Here $\mathbb E[\cdot\mid\mathcal F_n]$ denotes conditional expectation, and “almost surely” permits exceptions on an event of probability zero. (math.cmu.edu)

The tower property of conditional expectation extends the one-step condition to every pair $m\geq n$:

Consequently, the expected value is constant:

The converse is false: constant unconditional expectation does not ensure the conditional martingale property. (tropp.caltech.edu)

In continuous time, an adapted, integrable process $(M_t)_{t\geq0}$ is a martingale when

Many continuous-time results additionally assume a right-continuous, complete filtration and paths that are right-continuous with left limits. These regularity assumptions are distinct from the conditional-expectation definition. (web.stanford.edu)

Interpretation and examples

Centered sums

Suppose $X_1,X_2,\ldots$ are integrable, mutually independent random variables with $\mathbb E[X_k]=0$. Then

is a martingale for $\mathcal F_n=\sigma(X_1,\ldots,X_n)$. Independence makes the conditional mean of the next increment zero. A symmetric random walk, whose increments are $+1$ and $-1$ with equal probabilities, is a basic example. (n.ethz.ch)

More generally, independent increments are unnecessary. It is enough that the increments $D_n=M_n-M_{n-1}$ satisfy

Such increments form a martingale difference sequence. (math.cmu.edu)

Successive predictions of one quantity

For an integrable random variable $Y$, define

This is a Doob martingale, also called a Lévy–Doob martingale. As information accumulates, the prediction of $Y$ changes, but its next revision has conditional mean zero. This construction is useful when a complicated random object is revealed one component at a time. (tropp.caltech.edu)

Brownian examples

Standard Brownian motion $(B_t)$ is a martingale with respect to its natural filtration. So are

for each fixed real $\theta$. The compensating terms remove the predictable increase in the square or exponential. These examples illustrate that nonlinear transformations generally need correction terms to preserve the martingale property. (math.uchicago.edu)

Related classes and transformations

A submartingale replaces equality by

a supermartingale uses the reverse inequality. Both retain adaptedness and integrability. Their values therefore have, respectively, nonnegative or nonpositive conditional drift; their sample paths need not be monotone. A process is a martingale precisely when it belongs to both classes. (math.cmu.edu)

If $\varphi$ is a convex function and $\varphi(M_n)$ is integrable, conditional Jensen’s inequality shows that $\varphi(M_n)$ is a submartingale. Examples include $|M_n|$ and, for square-integrable martingales, $M_n^2$. (tropp.caltech.edu)

In discrete time, a submartingale $X_n$ has the Doob decomposition

where $N_0=A_0=0$, $N_n$ is a martingale, and $A_n$ is predictable and increasing. Its increments are

This separates unpredictable fluctuations from accumulated conditional drift. (web.stanford.edu)

A martingale transform takes the form

where $H_k$ is determined by $\mathcal F_{k-1}$, before the next increment is observed. If the coefficients are bounded, $G_n$ is a martingale. In the gambling interpretation, adjusting bounded stakes using past observations does not create a positive expected gain from a fair game. (n.ethz.ch)

Stopping times and optional stopping

A stopping time $T$, taking values in the nonnegative integers or infinity, satisfies

Thus, whether stopping has occurred by time $n$ can be decided without future information. The first time a process reaches a prescribed level is a typical example. (math.dartmouth.edu)

The optional stopping theorem gives conditions under which

For a discrete-time martingale, any one of the following is sufficient:

• $T$ is bounded by a deterministic integer.
• $T$ is almost surely finite and $|M_n|\leq C$ for all $n$, for a deterministic constant $C$.
• $\mathbb E[T]<\infty$ and $|M_n-M_{n-1}|\leq C$ for all $n$.

These are alternative sufficient conditions, not interchangeable descriptions of the same assumption. (math.dartmouth.edu)

Almost-sure finiteness alone is insufficient. A symmetric random walk starting at zero reaches $+1$ almost surely, so stopping at its first visit gives $M_T=1$, although $\mathbb E[M_0]=0$. The stopping time has infinite expectation, preventing the bounded-increment version from applying. Similar failures explain why the gambling strategy of doubling a stake after every loss does not contradict optional stopping: it allows unbounded stakes and potentially enormous intermediate losses. (n.ethz.ch)

Maximal inequalities and concentration

Martingale inequalities control an entire trajectory, rather than only its value at one fixed time. For a nonnegative submartingale $X_n$, Doob’s maximal inequality states

For a martingale with $\mathbb E|M_n|^p<\infty$, $p>1$, the $L^p$ version gives

These estimates connect terminal moments with the largest earlier fluctuation. (math.cmu.edu)

The Azuma–Hoeffding inequality provides exponential concentration. If

almost surely for deterministic constants $c_k$, then, for $a>0$,

Unlike bounds restricted to sums of independent variables, this permits dependence compatible with the martingale property. Exposure martingales apply it to random graphs and randomized algorithms. (cl.cam.ac.uk)

Convergence and integrability

Martingales need not converge without further assumptions. Doob’s martingale convergence theorem states that a discrete-time martingale satisfying

converges almost surely to a finite, integrable limit $M_\infty$. In particular, every nonnegative martingale has such a limit, since its expectations are constant. Almost-sure convergence alone does not ensure preservation of expectation. (web.stanford.edu)

The stronger condition of uniform integrability,

also gives convergence in $L^1$:

It follows that

Thus uniformly integrable martingales are precisely martingales representable as successive conditional expectations of an integrable terminal variable. (tropp.caltech.edu)

The distinction is essential: a martingale bounded in $L^1$ may converge almost surely while losing expectation in the limit because rare, increasingly large values continue to contribute to its finite-time means. Bounds

instead yield both almost-sure and $L^p$ convergence. (web.stanford.edu)

Continuous-time theory and applications

A local martingale behaves as a martingale after stopping at each member of a sequence of stopping times increasing almost surely to infinity. A local martingale need not be a true martingale: localization does not by itself supply the global integrability needed to preserve conditional expectations. Nonnegative local martingales are supermartingales. (math.uchicago.edu)

Martingales underlie stochastic calculus. For a predictable integrand $H$, the Brownian stochastic integral

is a square-integrable martingale when

on every finite horizon. Under weaker local assumptions it may be only a local martingale. (math.uchicago.edu)

In statistics and probability, martingale methods handle sequential information and dependent observations; maximal and concentration inequalities quantify their fluctuations. In mathematical finance, discounted asset prices are modeled as martingales or local martingales under an appropriate equivalent pricing measure, subject to the assumptions of the model. This is not a claim that actual prices have zero expected return under the real-world probability measure. (cl.cam.ac.uk)

Historical development

Martingale theory grew from attempts to formalize fair games and the limitations of gambling systems. Earlier work by Sergei Bernstein, Paul Lévy, and Andrey Kolmogorov anticipated aspects of the theory. Jean Ville’s 1939 book used martingale methods and a maximal inequality in studying randomness. Joseph L. Doob subsequently formulated the measure-theoretic framework, developed stopping and convergence results, and established martingales as a central class of stochastic processes in his 1953 book Stochastic Processes. In a later interview, Doob described Ville’s work as an important stimulus while explicitly acknowledging the earlier contributions. (chance.dartmouth.edu)

References

1. The Martingale Stopping Theoremmath.dartmouth.edu
2. Probability Theory & Stochastic Processes / Caltech CMS 117tropp.caltech.edu
3. 3. The Convergence of Martingalesweb.stanford.edu
4. Lecture 7: Martingales and Concentrationcl.cam.ac.uk
5. Stochastic Calculus and Applicationsmath.uchicago.edu
6. A Conversation with Joe Doobchance.dartmouth.edu