10.3 Martingale convergence theorems

Martingale Convergence Theorems

Martingale convergence theorems tell you when a martingale sequence settles down to a well-defined limit as time goes on. These results are central to understanding the long-term behavior of random processes, with direct applications in mathematical finance, sequential analysis, and ergodic theory. This section covers the main convergence results (Doob's theorems), the role of $L^p$ boundedness and uniform integrability, key applications, and important counterexamples.

Martingale Definition and Properties

A martingale is a stochastic process where your best prediction of the future value, given everything you know so far, is the current value. Formally, a sequence $(X_n)$ adapted to a filtration $(\mathcal{F}_n)$ is a martingale if $E[|X_n|] < \infty$ for all $n$ and

$E[X_{n+1} \mid \mathcal{F}_n] = X_n \quad \text{for all } n.$

This "fair game" property means the process has no systematic drift upward or downward.

Submartingale vs Supermartingale

• A submartingale satisfies $E[X_{n+1} \mid \mathcal{F}_n] \geq X_n$. On average, it tends upward (or stays flat). Convex functions of martingales are submartingales by Jensen's inequality, so $|X_n|$ and $X_n^2$ are submartingales when $(X_n)$ is a martingale.
• A supermartingale satisfies $E[X_{n+1} \mid \mathcal{F}_n] \leq X_n$. On average, it tends downward (or stays flat). A gambler playing a subfair game has wealth that forms a supermartingale.

Most convergence theorems are stated for submartingales or supermartingales, and the martingale case follows as a special case of both.

Martingale Transformations

A martingale transform constructs a new process from an existing one using a predictable process $(H_n)$ (meaning $H_n$ is $\mathcal{F}_{n-1}$-measurable):

$(H \cdot X)_n = \sum_{k=1}^{n} H_k (X_k - X_{k-1}).$

Think of $H_n$ as a betting strategy: you decide how much to wager based on information available before round $n$. A key fact is that a martingale transform of a martingale is again a martingale (provided integrability holds). This is the discrete analogue of the stochastic integral and underpins the connection between martingale theory and mathematical finance.

Stopping Times for Martingales

A stopping time $T$ with respect to $(\mathcal{F}_n)$ is a random variable taking values in $\{0, 1, 2, \ldots\} \cup \{\infty\}$ such that $\{T = n\} \in \mathcal{F}_n$ for all $n$. The decision to stop at time $n$ depends only on information available at time $n$.

The Optional Stopping Theorem states: if $(X_n)$ is a martingale and $T$ is a bounded stopping time (i.e., $T \leq N$ for some constant $N$), then

$E[X_T] = E[X_0].$

For unbounded stopping times, you need additional conditions (e.g., uniform integrability of $(X_{T \wedge n})$, or dominated convergence conditions). Careless application of optional stopping to unbounded times is a common source of errors.

Martingale Convergence Theorem Types

The convergence theorems answer a natural question: if a martingale runs forever, does it settle down? The answer depends on what boundedness conditions you impose and what mode of convergence you want.

Discrete-Time Martingale Convergence

The foundational result is Doob's upcrossing inequality, which controls how many times a submartingale can cross an interval $[a, b]$. If the expected number of upcrossings is finite, the process can't oscillate forever, forcing convergence.

Doob's Forward Convergence Theorem (the most important result in this unit): If $(X_n)$ is a submartingale satisfying $\sup_n E[X_n^+] < \infty$, then $X_n \to X_\infty$ almost surely, where $X_\infty$ is a finite random variable with $E[|X_\infty|] < \infty$.

For a martingale, $\sup_n E[X_n^+] < \infty$ is equivalent to $\sup_n E[|X_n|] < \infty$ (i.e., $L^1$-boundedness). So an $L^1$-bounded martingale converges almost surely.

Continuous-Time Martingale Convergence

The continuous-time analogue holds for right-continuous martingales (càdlàg paths). If $(X_t)_{t \geq 0}$ is a right-continuous submartingale with $\sup_t E[X_t^+] < \infty$, then $X_t \to X_\infty$ almost surely as $t \to \infty$.

The proof strategy is similar: control upcrossings over rational times, then use right-continuity to extend to all real times. This result is essential in stochastic calculus and the theory of continuous-time financial models.

$L^p$ Bounded Martingale Convergence

For $p > 1$, $L^p$-boundedness gives you both almost sure and $L^p$ convergence:

If $(X_n)$ is a martingale with $\sup_n E[|X_n|^p] < \infty$ for some $p > 1$, then:

1. $X_n \to X_\infty$ almost surely.
2. $X_n \to X_\infty$ in $L^p$ (meaning $E[|X_n - X_\infty|^p] \to 0$).

The reason $p > 1$ works so much better than $p = 1$ is that $L^p$-boundedness for $p > 1$ implies uniform integrability (by de la Vallée-Poussin's criterion), which is exactly the missing ingredient for norm convergence. The case $p = 1$ is the boundary case where this implication fails.

Doob's Martingale Convergence Theorem

Doob's Forward Convergence Theorem

This is the cornerstone result. Here's the precise statement and the logic behind it:

Theorem: Let $(X_n, \mathcal{F}_n)_{n \geq 0}$ be a submartingale with $\sup_n E[X_n^+] < \infty$. Then there exists a random variable $X_\infty$ with $E[|X_\infty|] < \infty$ such that $X_n \to X_\infty$ almost surely.

Proof sketch (via upcrossings):

1. For any $a < b$, let $U_n[a,b]$ count the number of upcrossings of the interval $[a,b]$ by $X_0, X_1, \ldots, X_n$.

2. Doob's upcrossing inequality gives $E[U_n[a,b]] \leq \frac{E[(X_n - a)^+]}{b - a}$.

3. Since $\sup_n E[X_n^+] < \infty$, the right side is bounded uniformly in $n$, so $E[U_\infty[a,b]] < \infty$.

4. This means $U_\infty[a,b] < \infty$ a.s. for each rational pair $a < b$.

5. On the event where $\liminf X_n < \limsup X_n$, there would exist rationals $a < b$ with infinitely many upcrossings, which has probability zero.

6. Therefore $\liminf X_n = \limsup X_n$ a.s., so the limit exists.

The upcrossing argument is elegant because it converts an analytic convergence question into a combinatorial counting problem.

Doob's Backward Convergence Theorem

Backward (or reverse) martingale convergence works in the opposite direction. A reverse martingale (or backward martingale) is a sequence $(X_n, \mathcal{F}_n)_{n \leq 0}$ (indexed by negative integers, or equivalently a decreasing filtration) satisfying the martingale property.

Theorem: If $(X_n)_{n \geq 1}$ is a reverse martingale with respect to a decreasing filtration $\mathcal{F}_1 \supset \mathcal{F}_2 \supset \cdots$, then $X_n \to X_\infty$ almost surely and in $L^1$, where $X_\infty = E[X_1 \mid \mathcal{F}_\infty]$ and $\mathcal{F}_\infty = \bigcap_n \mathcal{F}_n$.

A striking feature: reverse martingales always converge in $L^1$ without any extra conditions. This is because reverse martingales are automatically uniformly integrable (each $X_n = E[X_1 \mid \mathcal{F}_n]$, and conditional expectations of a fixed $L^1$ random variable are uniformly integrable).

A classic application: the strong law of large numbers can be proved using backward martingale convergence.

Doob's $L^p$ Convergence Theorem

Theorem: If $(X_n)$ is a martingale bounded in $L^p$ for some $p > 1$, then $X_n \to X_\infty$ almost surely and in $L^p$.

The key steps:

1. $L^p$-boundedness with $p > 1$ implies $L^1$-boundedness, so Doob's forward theorem gives a.s. convergence to some $X_\infty$.
2. $L^p$-boundedness with $p > 1$ implies uniform integrability.
3. Almost sure convergence plus uniform integrability upgrades to $L^p$ convergence.

Martingale Convergence in $L^1$ and $L^2$

Uniform Integrability and $L^1$ Convergence

The correct condition for $L^1$ convergence is uniform integrability (UI), not just $L^1$-boundedness. A family of random variables $(X_n)$ is uniformly integrable if

$\lim_{K \to \infty} \sup_n E[|X_n| \mathbf{1}_{|X_n| > K}] = 0.$

Theorem: For a martingale $(X_n)$, the following are equivalent:

1. $(X_n)$ is uniformly integrable.
2. $X_n \to X_\infty$ in $L^1$.
3. $X_n = E[X_\infty \mid \mathcal{F}_n]$ for all $n$ (the martingale is "closed").

This equivalence is sometimes called the martingale closure theorem. Condition (3) is especially useful: it says a UI martingale is always the sequence of conditional expectations of its own limit.

Krickeberg's Decomposition

Krickeberg's decomposition provides a structural result for $L^1$-bounded martingales. Any $L^1$-bounded martingale $(X_n)$ can be written as

$X_n = M_n - N_n,$

where $(M_n)$ and $(N_n)$ are both non-negative submartingales (in fact, non-negative supermartingales that are also martingales). This decomposition is the martingale analogue of writing a function as the difference of its positive and negative parts.

The decomposition is useful for reducing questions about general $L^1$-bounded martingales to questions about non-negative supermartingales, which are better behaved.

Martingale Convergence in $L^2$

For $L^2$ convergence, there's a clean characterization using the predictable quadratic variation. A martingale $(X_n)$ with $X_0 = 0$ converges in $L^2$ if and only if

$\sum_{k=1}^{\infty} E[(X_k - X_{k-1})^2 \mid \mathcal{F}_{k-1}] < \infty \quad \text{a.s.}$

This connects convergence to the total accumulated variance of the increments. If the increments keep adding substantial variance, the martingale won't settle down in $L^2$.

Relation Between $L^1$ and $L^2$ Convergence

The hierarchy of convergence modes for martingales:

• $L^2$ convergence $\Rightarrow$ $L^1$ convergence $\Rightarrow$ almost sure convergence (given $L^1$-boundedness for the last implication).
• None of these arrows reverse in general.

$L^2$ convergence is strictly stronger because $E[|X|] \leq (E[X^2])^{1/2}$ by Jensen's inequality (or Cauchy-Schwarz). A martingale can converge in $L^1$ without being $L^2$-bounded, so $L^1$ convergence does not imply $L^2$ convergence.

Applications of Martingale Convergence

Gambler's Ruin Problem

Consider a gambler who starts with $k$ dollars and bets $1$ dollar on each round of a fair game (probability $1/2$ of winning or losing each bet). The gambler's wealth $(X_n)$ is a martingale. The gambler stops upon reaching $N$ dollars or going broke.

Since the stopping time is bounded (the game must end in at most $N$ absorbing states), the optional stopping theorem gives:

$k = E[X_0] = E[X_T] = N \cdot P(\text{reach } N) + 0 \cdot P(\text{ruin}),$

so $P(\text{ruin}) = 1 - k/N$. As $N \to \infty$, the ruin probability approaches 1, illustrating that a fair-game martingale with an absorbing barrier at 0 will eventually be absorbed.

Pólya's Urn Model

Start with an urn containing one red and one blue ball. At each step, draw a ball uniformly at random, then return it along with one additional ball of the same color.

Let $X_n$ be the proportion of red balls after $n$ draws. Then $(X_n)$ is a martingale (you can verify the conditional expectation property directly). Since $0 \leq X_n \leq 1$, the martingale is bounded, so Doob's theorem guarantees $X_n \to X_\infty$ almost surely. The limit $X_\infty$ turns out to be uniformly distributed on $[0,1]$, which is a beautiful and non-obvious result.

Martingale Convergence in Mathematical Finance

The Fundamental Theorem of Asset Pricing connects arbitrage-free markets to martingales: a market is arbitrage-free if and only if there exists an equivalent probability measure (the risk-neutral measure) under which discounted asset prices are martingales.

Martingale convergence results then have direct financial interpretations:

• Convergence of discounted price processes relates to long-run asset behavior.
• The martingale representation theorem (closely tied to convergence theory) underpins the replication of contingent claims.
• The Black-Scholes formula arises from computing conditional expectations under the risk-neutral measure, where the discounted stock price is a continuous-time martingale.

Counterexamples and Limitations

Martingales Without Almost Sure Convergence

The simple symmetric random walk $X_n = \sum_{k=1}^n Y_k$, where $Y_k$ are i.i.d. with $P(Y_k = +1) = P(Y_k = -1) = 1/2$, is a martingale. It does not converge almost surely: it oscillates between $+\infty$ and $-\infty$.

Why doesn't Doob's theorem apply? Because $E[|X_n|] = E[|X_n|] \sim \sqrt{2n/\pi} \to \infty$, so the martingale is not $L^1$-bounded. The boundedness condition in Doob's theorem is genuinely necessary, not just a technicality.

$L^1$-Bounded Martingales Without $L^1$ Convergence

A classic example: let $X_n = 2^n \mathbf{1}_{A_n}$, where $A_0 \supset A_1 \supset \cdots$ with $P(A_n) = 2^{-n}$, and each $A_{n+1}$ is obtained by choosing one of two equal halves of $A_n$. Then $E[X_n] = 1$ for all $n$, so the martingale is $L^1$-bounded. It converges almost surely to $X_\infty = 0$ (since $P(A_n) \to 0$). But $E[X_\infty] = 0 \neq 1 = E[X_n]$, so convergence in $L^1$ fails.

This martingale is not uniformly integrable: all the mass concentrates on an ever-smaller set. This is the standard example showing that $L^1$-boundedness gives a.s. convergence but not $L^1$ convergence.

Non-Uniqueness of Martingale Limits

If $(X_n)$ is a martingale that converges almost surely, the limit $X_\infty$ is unique (as a random variable, up to a.s. equivalence). There is no non-uniqueness issue for the limit itself.

What can happen is that different filtrations or different starting random variables produce different martingales with different limits. For instance, $E[Y \mid \mathcal{F}_n]$ converges to $E[Y \mid \mathcal{F}_\infty]$ when $Y \in L^1$, and the limit depends on both $Y$ and the filtration. But given a fixed martingale, its almost sure limit (when it exists) is determined.