Functional Central Limit Theorem (FCLT)

• FCLT is a generalization of the classical CLT that establishes convergence of normalized partial-sum processes to Brownian motion in function spaces.
• It quantifies convergence using metrics like Wasserstein-1 and Lévy–Prokhorov, showing a sharp rate of n⁻¹ᐟ²√(ln n) for stationary Gaussian sequences.
• The theorem underpins applications in stochastic differential equations and Gaussian processes, highlighting precise error bounds via coupling and duality methods.

The Functional Central Limit Theorem (FCLT) is a fundamental result in probability theory and stochastic process theory, extending the classical central limit theorem from finite-dimensional distributions of sums of random variables to convergence in distribution of suitably normalized stochastic processes to limiting Gaussian processes, typically Brownian motion. In the Gaussian case, the FCLT becomes nontrivial in the topology of function spaces (such as the space of continuous functions), and the rate of convergence and the choice of probability metrics play a crucial role in quantifying this convergence.

1. Model Framework and Process Construction

Let $\{X_k : k \in \mathbb{Z}\}$ be a doubly infinite, strictly stationary Gaussian sequence with $\mathbb{E}[X_k]=0$, $\operatorname{Var}(X_k)=1$, and $\operatorname{Cov}(X_k, X_{k+n}) = \rho(n)$, where the covariance function $\rho$ is summable and normalized such that $\sum_{n=-\infty}^{\infty}\rho(n)=1$. For each $n \geq 1$, the normalized partial-sum process $S_n$ is defined by

$$S_{n}(t) = n^{-1/2} \sum_{k=1}^{\lfloor nt \rfloor} X_k + (nt - \lfloor nt \rfloor) n^{-1/2} X_{\lfloor nt \rfloor + 1}, \quad 0 \leq t \leq 1.$$

This is a continuous, piecewise-linear interpolation of normalized partial sums of the $X_k$, known as the Donsker–Prokhorov chain. As $\mathbb{E}[X_k]=0$0, $\mathbb{E}[X_k]=0$1 converges in distribution in $\mathbb{E}[X_k]=0$2 to a standard Brownian motion $\mathbb{E}[X_k]=0$3.

2. Topologies and Probability Metrics

Probability measures on $\mathbb{E}[X_k]=0$4 can be compared using different metrics. Two primary distances in this context are:

• Wasserstein-1 (Kantorovich–Rubinstein) distance: For $\mathbb{E}[X_k]=0$5 probability measures on $\mathbb{E}[X_k]=0$6,

$\mathbb{E}[X_k]=0$7

where $\mathbb{E}[X_k]=0$8 is the set of all couplings of $\mathbb{E}[X_k]=0$9.

• Lévy–Prokhorov metric $\operatorname{Var}(X_k)=1$0: Defined by

$\operatorname{Var}(X_k)=1$1

with $\operatorname{Var}(X_k)=1$2.

In general, $\operatorname{Var}(X_k)=1$3, and convergence in $\operatorname{Var}(X_k)=1$4 is strictly stronger than weak convergence (equivalently, convergence in $\operatorname{Var}(X_k)=1$5).

3. Main Theorem: Sharp FCLT in Wasserstein-1

The central quantitative result for Gaussian sequences is: There exist positive constants $\operatorname{Var}(X_k)=1$6 (depending only on $\operatorname{Var}(X_k)=1$7) such that for all $\operatorname{Var}(X_k)=1$8,

$\operatorname{Var}(X_k)=1$9

This shows that in probability law, in the $\operatorname{Cov}(X_k, X_{k+n}) = \rho(n)$0 metric, the convergence of the normalized partial-sum process to Brownian motion occurs at rate $\operatorname{Cov}(X_k, X_{k+n}) = \rho(n)$1. The exponent $\operatorname{Cov}(X_k, X_{k+n}) = \rho(n)$2 is the leading power, but the $\operatorname{Cov}(X_k, X_{k+n}) = \rho(n)$3 reflects a logarithmic correction inherent to the Gaussian setting.

4. Proof Sketch: Two-Sided Bounds

Upper Bound:

• By constructing an explicit coupling, $\operatorname{Cov}(X_k, X_{k+n}) = \rho(n)$4 is decomposed as increments of $\operatorname{Cov}(X_k, X_{k+n}) = \rho(n)$5 plus a negligible boundary correction. In the Gaussian–Markov case (e.g., $\operatorname{Cov}(X_k, X_{k+n}) = \rho(n)$6), there is an exact representation $\operatorname{Cov}(X_k, X_{k+n}) = \rho(n)$7, where $\operatorname{Cov}(X_k, X_{k+n}) = \rho(n)$8 is a mean-zero Gaussian process with supremum of order $\operatorname{Cov}(X_k, X_{k+n}) = \rho(n)$9.
• Using Kantorovich duality, $\rho$0.
• Borell–TIS inequality yields that the expected supremum of a mean-zero Gaussian process with variance $\rho$1 grows as $\rho$2.

Lower Bound:

• For the lower bound, a 1-Lipschitz functional $\rho$3 with a carefully chosen $\rho$4 is used to detect the size of oscillations.
• Computations establish $\rho$5, which by duality implies the same lower bound for $\rho$6.

The presence of the $\rho$7 factor, and matching upper and lower bounds, demonstrate the sharpness of the result.

5. Relationship to Lévy–Prokhorov Metric and Interpretation

By Skorokhod embedding or Komlós–Major–Tusnády (KMT) techniques, the convergence in the Lévy–Prokhorov metric is

$\rho$8

and this is also sharp. Thus, in Wasserstein-1, the convergence is quantitatively faster, with $\rho$9 as $\sum_{n=-\infty}^{\infty}\rho(n)=1$0. This phenomenon is attributed to the fact that for a Gaussian perturbation of order $\sum_{n=-\infty}^{\infty}\rho(n)=1$1, the $\sum_{n=-\infty}^{\infty}\rho(n)=1$2 distance to zero is $\sum_{n=-\infty}^{\infty}\rho(n)=1$3, while the $\sum_{n=-\infty}^{\infty}\rho(n)=1$4 distance is $\sum_{n=-\infty}^{\infty}\rho(n)=1$5.

6. Extensions and Corollaries

• Continuous-time Gaussian Markov processes: For $\sum_{n=-\infty}^{\infty}\rho(n)=1$6 a stationary Gaussian Markov process, the normalized continuous-time integral $\sum_{n=-\infty}^{\infty}\rho(n)=1$7 satisfies the same two-sided rate in $\sum_{n=-\infty}^{\infty}\rho(n)=1$8:

$\sum_{n=-\infty}^{\infty}\rho(n)=1$9

• Applications to stochastic differential equations (SDEs): By the continuous mapping theorem and Lipschitz property of the Itô map, error bounds in $n \geq 1$0 for the driving Brownian motion directly yield corresponding bounds for the solutions to linear or nonlinear SDEs with additive Gaussian noise.
• Limitations of sharpness: The $n \geq 1$1 correction is specific to the Gaussian setting. For non-Gaussian or heavy-tailed sequences, or for sequences with only mixing conditions rather than full Gaussianity, rates may differ and sharpness may be lost.

7. Broader Context and Summary

In the setting of stationary Gaussian and Markov sequences, the rate $n \geq 1$2 in the $n \geq 1$3 metric for the functional CLT in $n \geq 1$4 is strictly faster (by a logarithmic factor) than the classical Lévy–Prokhorov rate. The $n \geq 1$5 correction is an unavoidable feature arising from the behavior of suprema of Gaussian processes. This result exhibits the maximally rapid convergence permitted in law for the Donsker–Prokhorov chain with Gaussian input when measured by the Wasserstein-1 distance. The analytic framework integrates coupling, Gaussian process inequalities, and duality arguments to provide non-asymptotic quantitative control of functional convergence rates in strong topologies (Lototsky, 2022).