Martingale central limit theorem

The martingale central limit theorem is a family of results in probability theory describing when normalized sums of martingale differences converge in distribution to a normal distribution. It extends the classical central limit theorem to dependent random variables whose conditional expectations vanish relative to an increasing sequence of information sets.

The theorem replaces the independence assumption of classical limit theory with conditions on conditional variance and conditionally large increments. Its standard triangular-array form applies when the number and distribution of the increments vary with the row of the array. The limiting variance may be deterministic, or it may be random when convergence is formulated through stable convergence.

Mathematical setting

Let

[ {X_{n,k},\mathcal F_{n,k}:1\leq k\leq k_n} ]

be a triangular array adapted to filtrations

[ \mathcal F_{n,0}\subseteq\mathcal F_{n,1}\subseteq\cdots \subseteq\mathcal F_{n,k_n}. ]

Each (X_{n,k}) is assumed to be integrable and (\mathcal F_{n,k})-measurable. The martingale-difference condition is

[ \operatorname E[X_{n,k}\mid\mathcal F_{n,k-1}]=0 ]

almost surely. Consequently, the partial sums

[ S_{n,j}=\sum_{k=1}^{j}X_{n,k} ]

form a martingale with respect to the corresponding row filtration.

The relevant variance quantity is the predictable quadratic variation

[ V_n=\sum_{k=1}^{k_n} \operatorname E[X_{n,k}^{2}\mid\mathcal F_{n,k-1}]. ]

Unlike an ordinary sum of marginal variances, (V_n) records the variance visible immediately before each increment occurs. This distinction permits the theorem to accommodate dependence generated by the filtration.

Standard triangular-array theorem

Suppose that

[ V_n\xrightarrow{\mathrm P}\sigma^2 ]

for a constant (\sigma^2\geq 0). Suppose additionally that, for every (\varepsilon>0),

[ \sum_{k=1}^{k_n} \operatorname E!\left[ X_{n,k}^{2} \mathbf 1_{{|X_{n,k}|>\varepsilon}} \mid\mathcal F_{n,k-1} \right] \xrightarrow{\mathrm P}0. ]

The second condition is the conditional Lindeberg condition. Under these assumptions,

[ S_{n,k_n} =\sum_{k=1}^{k_n}X_{n,k} \xrightarrow{\mathrm d}N(0,\sigma^2). ]

When (\sigma^2=0), the conclusion means that the terminal sum converges in probability to zero. When (\sigma^2>0), division by (\sigma) produces convergence to the standard normal law.

The conditional Lindeberg condition prevents a non-negligible fraction of the total quadratic variation from being contributed by increments exceeding a fixed threshold. It is therefore the martingale counterpart of the condition used in the Lindeberg–Feller theorem. The conditional variance hypothesis determines the scale of the limiting distribution.

A frequently used sufficient condition is the conditional Lyapunov relation

[ \sum_{k=1}^{k_n} \operatorname E!\left[ |X_{n,k}|^{2+\delta} \mid\mathcal F_{n,k-1} \right] \xrightarrow{\mathrm P}0 ]

for a fixed (\delta>0), together with convergence of (V_n). Applying the conditional form of Markov's inequality shows that this relation implies the conditional Lindeberg condition.

Interpretation through quadratic variation

For each row, the sum of squared increments is

[ [S_n]{k_n}=\sum{k=1}^{k_n}X_{n,k}^{2}, ]

whereas its predictable counterpart is (V_n). Under suitable integrability and negligibility conditions, the difference

[ [S_n]_{k_n}-V_n ]

is itself a martingale fluctuation that converges to zero in probability. The theorem can consequently be stated using convergence of actual quadratic variation rather than predictable quadratic variation, provided the accompanying assumptions control this difference.

This formulation clarifies the relation with continuous-time martingale theory. A continuous local martingale whose quadratic variation approaches a deterministic clock behaves asymptotically like Brownian motion evaluated at that clock. In discrete arrays, the terminal martingale sum has the corresponding one-time normal limit.

Random limiting variance

The conditional variance sum need not converge to a constant. If

[ V_n\xrightarrow{\mathrm P}V, ]

where (V) is a nonnegative random variable measurable with respect to an appropriate limiting information field, the limiting law is conditionally Gaussian. In a stable-convergence formulation,

[ S_{n,k_n}\xrightarrow{\mathrm{st}} \sqrt V,Z, ]

where (Z) has the standard normal distribution and is independent of the limiting information field.

The unconditional distribution of (\sqrt V,Z) is a normal variance mixture. It is generally not normal unless (V) is almost surely constant. This distinction explains why convergence of the predictable quadratic variation merely in distribution is insufficient for the ordinary deterministic-variance conclusion.

Historical development

The theorem emerged from the interaction between classical central limit theory and the systematic theory of martingales. Paul Lévy connected conditional expectations, sums of dependent variables, and Gaussian limits during the early development of modern probability. Joseph L. Doob subsequently placed discrete and continuous martingales within a unified measure-theoretic framework.

During the early 1970s, You Watanabe formulated a triangular-array version in which convergence of the predictable conditional variance was paired directly with a conditional Lindeberg condition. Watanabe's treatment separated the Gaussian approximation from assumptions of unconditional independence and identified the filtration as the relevant structure governing the variance calculation.

Elsewhere in the same period, B. M. Brown established martingale central limit criteria based on conditional variances and asymptotic negligibility. Donald L. McLeish developed related array theorems that allowed dependence to be controlled through martingale approximation. These formulations became standard sources for later treatments of the subject.

Subsequent work organized the results around predictable quadratic variation, stopping-time arguments, and stable limits. The resulting framework supports both scalar limit theorems and functional versions in spaces of sample paths.

Outline of the convergence argument

The proof compares the conditional characteristic function of the martingale sum with the characteristic function of a Gaussian variable. For a fixed real number (t), the exponential associated with one increment admits the expansion

[ e^{itX_{n,k}}

1+itX_{n,k} -\frac{t^2}{2}X_{n,k}^{2} +r_{n,k}(t). ]

Conditional expectation removes the linear term because (X_{n,k}) is a martingale difference. The accumulated quadratic terms approach

[ -\frac{t^2}{2}\sigma^2 ]

through convergence of (V_n). The conditional Lindeberg condition controls the total contribution of the remainders by separating increments below a fixed threshold from increments exceeding it.

Iteration over the filtration then yields

[ \operatorname E[e^{itS_{n,k_n}}] \longrightarrow e^{-t^2\sigma^2/2}. ]

The limiting function is the characteristic function of (N(0,\sigma^2)), so Lévy's continuity theorem gives convergence in distribution. Proofs using truncation or martingale approximation implement the same division between quadratic accumulation and negligible exceptional increments.

Functional form

A functional central limit theorem considers the partial-sum process

[ S_n(t)= \sum_{k\leq \lfloor k_nt\rfloor}X_{n,k}, \qquad 0\leq t\leq 1. ]

If the predictable quadratic-variation process converges to a deterministic continuous function (v(t)), and a process-level conditional Lindeberg condition holds, then (S_n) converges weakly in an appropriate Skorokhod space to a time-changed Brownian motion

[ B(v(t)). ]

When (v(t)=t), the limit is standard Brownian motion. Tightness requires control of fluctuations over short time intervals in addition to convergence of terminal quadratic variation. The functional theorem therefore contains information not supplied by the one-dimensional terminal-sum result.

Relation to martingale approximation

A dependent partial sum may be decomposed as

[ T_n=M_n+R_n, ]

where (M_n) is a martingale sum and (R_n) is a remainder. If the normalized remainder converges to zero in probability, the asymptotic distribution of (T_n) agrees with that of (M_n) by Slutsky's theorem. This mechanism transfers martingale central limit theorems to dependent processes that are not themselves martingales.

The substantive requirement is that the approximation preserve the asymptotic quadratic variation. A small remainder at the terminal time may be adequate for a scalar limit, while a functional limit requires uniform control of the remainder across the entire time interval.