Functional Central Limit Theorems (FCLTs)

• FCLTs are foundational principles that establish the functional convergence of normalized stochastic processes to a limiting process in function spaces.
• They are applied in various domains such as queueing theory, random graphs, and statistical inference, rigorously capturing entire process trajectories.
• Key methodologies include martingale difference arrays, spectral analysis, and entropy conditions to verify tightness and finite-dimensional convergence.

A functional central limit theorem (FCLT) provides the weak convergence of a sequence of stochastic processes to a limiting process in a function space, typically under suitable normalization and centering. This concept generalizes classical central limit theorems from finite-dimensional distributions to the convergence of random functions, thereby enabling a rigorous description of the asymptotic behavior of entire sample paths. FCLTs are foundational in the study of stochastic processes, empirical processes, random walks, interacting particle systems, queueing, random graphs, network models, survey designs, and many other domains. The modern theory encompasses both classical Gaussian limits and non-Gaussian (e.g., stable, self-similar, or Volterra-type) processes as limits.

1. Foundational Structures and General Framework

In the canonical setting, an FCLT addresses sequences of processes $\{X_n(t),\, t \in [0,1]\}$ (or a suitable index set) in a metric function space $D[0,1]$ (the Skorokhod space of c`adl`ag functions, or $C[0,1]$ for continuous limits). The goal is to establish weak convergence

$$X_n(\cdot) \xRightarrow{d} X(\cdot) \quad \text{in} \quad D[0,1],$$

where “$\xRightarrow{d}$” denotes convergence in law. Generally, the classical invariance principle (Donsker’s theorem) provides that the properly normalized sum of i.i.d. mean-zero random variables with finite variance converges to standard Brownian motion in $C[0,1]$: $$\frac{1}{\sqrt{n}}\, S_{\lfloor nt \rfloor} \xRightarrow{d} \sigma\,B(t),$$ where $B(t)$ is a standard Brownian motion and $\sigma^2$ is the variance of the increments (Lo et al., 2018).

This paradigm generalizes to non-i.i.d., dependent, heavy-tailed, or non-Gaussian settings, requiring more sophisticated scaling and normalization and leading to varied limit processes (e.g., symmetric $\alpha$-stable, fractional Brownian motion, time-changed Brownian motion, Gaussian Volterra processes, etc.).

2. FCLTs in Classical, Heavy-Tailed, and Dependent Regimes

Classical i.i.d. Increments: Under independence, centering, and finite variance, Donsker’s theorem yields functional convergence to Brownian motion in multi-dimensional settings, with covariance inherited from the covariance matrix of the increments. Tightness is verified by Arzelà–Ascoli-type modulus of continuity arguments and Etemadi-style maximal inequalities (Lo et al., 2018).

Heavy-Tailed and Infinitely Divisible Processes: For stationary, symmetric, infinitely divisible processes with regularly varying Lévy measures, with no Gaussian component and satisfying an integral representation via a conservative, measure-preserving transformation $D[0,1]$0, the FCLT yields convergence to a new class of self-similar, symmetric $D[0,1]$1-stable processes $D[0,1]$2 with stationary increments (Owada et al., 2012). The normalization is dictated by both the tail index $D[0,1]$3 of the Lévy measure and an ergodic-theoretical memory parameter $D[0,1]$4 derived from the flow generating the process, with explicit dependence on the wandering rate and the Darling–Kac (return time) normalization. The limiting process exhibits $D[0,1]$5-self-similarity, with $D[0,1]$6.

Discrete Memory and Martingale Cases: For finite-memory dependent models (e.g., $D[0,1]$7-sum dependence, reinforced random walks such as the minimal RW and elephant RW), martingale decomposition techniques are employed. The martingale difference array admits explicit uniform fourth-moment bounds, and the predictable quadratic variation leads to a Brownian process limit with explicit variance scaling (Guevara et al., 13 Jun 2025). Tightness and finite-dimensional convergence are achieved using Rebolledo–Hall–Heyde martingale FCLT criteria.

Sub-linear Expectation and Nonlinear Probability: FCLTs extend to sub-linear expectation (capacity) frameworks, where central limit theorems for martingale-like and independent arrays yield convergence to $D[0,1]$8-Brownian motion (the law-solving a sublinear PDE), with time-changed covariance structure determined by quadratic variation limits (Zhang, 2019).

3. Non-Gaussian and Non-classical Limits

Stable and Self-Similar Limits: Beyond Gaussianity, limits can be non-Gaussian, such as stable Lévy processes where scaling is revised (e.g., $D[0,1]$9 normalization for stable laws), and the limit is a symmetric $C[0,1]$0-stable process (Kosiński, 2010, Owada et al., 2012).

Fractional and Bi-Fractional Brownian Motions: In infinite-urn and occupancy models with regularly varying urn probabilities, functional CLTs yield fractional Brownian motion (FBM) with Hurst index $C[0,1]$1 (for occupancy or missing-mass processes) or bi-fractional Brownian motion for certain odd-occupancy processes (Chebunin et al., 2019). The Gaussian limits possess explicit covariance kernels derived via Poissonization and Riemann–Stieltjes analysis.

Volterra-Type and Long-Range Dependence: In models with long-range memory (e.g., Hawkes processes with heavy-dispersion kernels), FCLTs yield Gaussian Volterra processes with covariance involving fractional powers (Horst et al., 2024). The same type of memory arises in interacting particle systems (random graphs, reinforced walks), superprocesses with small eigenvalue gaps, and critical percolation-type limits.

Infinite-Dimensional and Functional-Infinite Limits: For inhomogeneous random graphs, limits are conditionally Gaussian processes in weighted $C[0,1]$2-spaces indexed by component-type vectors, or functionals of these infinite-dimensional processes yielding macroscopic CLTs for features like giant component size and the minimal spanning tree weight (Bhamidi et al., 2024).

4. FCLTs in Statistical, Survey, and Nonparametric/Machine Learning Applications

Survey Design-Based Empirical Processes: Under single-stage and conditional Poisson sampling designs, Horvitz-Thompson (HT) and Hájek empirical processes, both when centered by finite-population or superpopulation means, satisfy FCLTs in $C[0,1]$3 (or $C[0,1]$4 for index sets of test functions) under mild moment and design-correlation assumptions (Boistard et al., 2015, Pasquazzi, 2019). Covariances incorporate higher-order correlations in the inclusion indicators and combinations of model-based and design-based uncertainty. These FCLTs are essential for Hadamard-differentiable statistical functionals, enabling uniform inference for survey quantile processes and continuity-based functionals (e.g., poverty rates).

Nonparametric Bayesian Priors and Random Measures: For stick-breaking priors (Dirichlet process and generalizations), FCLTs are established using method-of-moments arguments, yielding convergence of the empirical process of the random measure to a Brownian bridge process in appropriate function spaces (Hu et al., 2020). Precise conditions on weights are required (control of higher moments, vanishing cumulants), which guarantee that only pairwise (Wick) contractions persist, leading to a Gaussian limiting process with covariance structure matching that of the base law.

Topological Data Analysis: In persistent homology and Betti number processes on geometric (e.g., Poisson) networks, FCLTs are proved for the empirical Betti trajectories under stabilization and moment criteria, yielding tight Gaussian fields in Skorokhod spaces $C[0,1]$5 for multidimensional parameter grids (Krebs et al., 2020).

5. FCLTs in Complex Systems: Queues, Random Graphs, Particle Systems, and Beyond

Markov-Modulated Queues: In Markov-modulated infinite-server systems, the FCLT yields convergence to an inhomogeneous Ornstein–Uhlenbeck (OU) process. The dichotomy between “fast” and “slow” background modulation (via scaling of arrival/service rates and background transition rates) determines the normalization, limit SDE, and the emergence of long-memory fluctuations (via the deviation matrix) (Blom et al., 2016).

Epidemic Dynamics and Volterra-Type Limits: In epidemic models with structured infectivity and waning immunity, aggregate fluctuation processes for susceptibility and infectivity converge in law to the unique solution of stochastic Volterra integral equations, with driving Gaussian noise reflecting random individual-level infectivity paths (Zotsa-Ngoufack, 2023).

Interacting Reinforced Random Walks and Synchronization: In mean-field reinforced random walks, synchronization rates and global fluctuation scaling are captured by FCLTs for the mean and difference processes, with time-changed Brownian motion limits exhibiting distinct rates (synchronization is faster than simple convergence) (Crimaldi et al., 2016).

Supercritical Superprocesses: For supercritical superprocesses with spatially varying branching mechanisms, FCLTs describe the process-level convergence (in Skorokhod space) of critical, subcritical, and supercritical components to independent Gaussian processes, with variances and covariances determined by spectral properties of the Feynman–Kac semigroup and the structure of the underlying motion and branching mechanism (Ren et al., 2014).

6. Methodologies and Proof Techniques

• Series/Poisson Integral Representations: Used for infinitely divisible processes, especially in the heavy-tailed case, to facilitate the truncation of contributions from large and small jumps and connect the law of partial sums to stable or self-similar limits (Owada et al., 2012).
• Martingale Difference Arrays and Predictable Quadratic Variation: Central to establishing FCLTs for models with dependence or memory, by verifying quadratic variation convergence and Lindeberg-Feller conditions under bounded increments (Guevara et al., 13 Jun 2025, Zhang, 2019).
• Weak Convergence in Function Spaces: Tightness is established via oscillation modulus or functional entropy conditions, using entropy inequalities (e.g., for survey sampling or nonparametric tests (Pasquazzi, 2019, Boistard et al., 2015)), and finite-dimensional convergence is obtained via standard CLT arguments and continuous mapping theorems.
• Spectral, Ergodic, or ODE-based Reductions: For particle systems, superprocesses, and inhomogeneous random graphs, explicit spectral decompositions, or law-of-large-numbers for ODE systems, reduce the process dynamics to infinite-dimensional stochastic differential or integral equations, whose solutions characterize the Gaussian limit (Ren et al., 2014, Bhamidi et al., 2024).

7. Exemplary Limit Processes, Key Formulas, and Classes

The following summarizes representative limit process forms in FCLTs:

Case	Limiting Process	Key Normalization
i.i.d. finite variance (Donsker)	Brownian motion $C[0,1]$6	$C[0,1]$7
i.i.d. heavy-tailed ($C[0,1]$8-stable)	$C[0,1]$9-stable Lévy motion	$$X_n(\cdot) \xRightarrow{d} X(\cdot) \quad \text{in} \quad D[0,1],$$0
Infinite Markov-modulated queue	Ornstein–Uhlenbeck process	$$X_n(\cdot) \xRightarrow{d} X(\cdot) \quad \text{in} \quad D[0,1],$$1 or $$X_n(\cdot) \xRightarrow{d} X(\cdot) \quad \text{in} \quad D[0,1],$$2
Reinforced finite-memory dependent models	Brownian motion with explicit variance	$$X_n(\cdot) \xRightarrow{d} X(\cdot) \quad \text{in} \quad D[0,1],$$3
Occupancy/missing-mass infinite urn with RV tails	fBM/bi-fBM, Hurst $$X_n(\cdot) \xRightarrow{d} X(\cdot) \quad \text{in} \quad D[0,1],$$4	$$X_n(\cdot) \xRightarrow{d} X(\cdot) \quad \text{in} \quad D[0,1],$$5
Supercritical superprocesses, spatial branching	Gaussian process with spectral kernel	$$X_n(\cdot) \xRightarrow{d} X(\cdot) \quad \text{in} \quad D[0,1],$$6
Stick-breaking Bayesian priors	Brownian bridge $$X_n(\cdot) \xRightarrow{d} X(\cdot) \quad \text{in} \quad D[0,1],$$7	$$X_n(\cdot) \xRightarrow{d} X(\cdot) \quad \text{in} \quad D[0,1],$$8 (prior parameter)
Topological/plenoptic Betti numbers	Gaussian field (multi-parameter)	$$X_n(\cdot) \xRightarrow{d} X(\cdot) \quad \text{in} \quad D[0,1],$$9
Sub-linear expectation (nonlinear probability)	$\xRightarrow{d}$0-Brownian motion	Variance by quadratic form

The precise kernel/covariance structure is model-dependent and often constructed using the spectral, combinatorial, ergodic, or empirical process framework, as detailed in the cited works.

---

By integrating ergodic theory, probability, functional analysis, and combinatorics, FCLTs provide a unified mathematical apparatus for quantifying fluctuations of stochastic systems at the functional/process level, yielding robust asymptotic approximations valid in a wide range of complex probabilistic models. This suggests that further generalizations may be possible by extending the ergodic, martingale, and sub-linear frameworks to even broader classes of stochastic systems.