Regularity Assumptions & Empirical Process Conditions

• Regularity assumptions and empirical process conditions are vital criteria that ensure convergence and tightness in stochastic models.
• They integrate smoothness, mixing, and entropy measures to overcome challenges in high- and infinite-dimensional settings.
• These frameworks underpin applications such as empirical copula analysis, multidimensional diffusion studies, and advanced bootstrap techniques.

Regularity assumptions and empirical process conditions form the foundational framework underlying the convergence and tightness properties of modern empirical process theory. These conditions govern the behavior of stochastic processes indexed by function classes or sets, allowing precise determination of limiting Gaussian laws, rates of convergence, and applicability of bootstrap and other asymptotic techniques. Central themes include smoothness, mixing, tail behavior, complexity measures (bracketing/covering entropy), and metric regularity, each of which addresses a distinct obstacle to uniform convergence in high- or infinite-dimensional probability.

1. Smoothness and Functional Regularity Conditions

Empirical process theory typically requires the indexing functions to obey quantifiable smoothness or regularity constraints, which directly determine the feasibility of classical CLT or functional convergence.

• Empirical Copula Processes: Weak convergence holds under the minimal condition that the first-order partial derivatives $\dot{C}_j(u)$ of the copula exist and are continuous on open faces $V_{d,j}=\{u\in[0,1]^d:0<u_j<1\}$, with appropriate one-sided boundary extensions. Such continuity ensures existence and path regularity of the limiting Gaussian process, and, crucially, is minimal for validity of multiplier bootstrap resampling (Segers, 2010).
• Multidimensional Diffusions: The generator coefficients $a_{ij}(x)$ and $b_i(x)$ are imposed to be in Hölder balls $C^\beta(L)$, typically with $\beta>d/2$ for $d\geq3$ to permit undersmoothing and optimal rates. Uniform ellipticity and at-most-linear growth are also necessary. The invariant Lebesgue density $\pi$ must be strictly positive and smooth enough to support equivalence between $L^2(\pi)$ and spatial $L^2$ norms (Rohde et al., 2010).
• Empirical Measures and Regular Test Functions: For empirical measures, if the class of test functions $V_{d,j}=\{u\in[0,1]^d:0<u_j<1\}$0 enjoys Hӧlder or Sobolev regularity with $V_{d,j}=\{u\in[0,1]^d:0<u_j<1\}$1, the convergence in dual norms is of canonical Monte Carlo order $V_{d,j}=\{u\in[0,1]^d:0<u_j<1\}$2, completely countering the curse of dimensionality. Conversely, for $V_{d,j}=\{u\in[0,1]^d:0<u_j<1\}$3, the rate degrades to $V_{d,j}=\{u\in[0,1]^d:0<u_j<1\}$4 (Kloeckner, 2018).
• Critical Point Convergence: For functions $V_{d,j}=\{u\in[0,1]^d:0<u_j<1\}$5 converging to $V_{d,j}=\{u\in[0,1]^d:0<u_j<1\}$6 on a manifold $V_{d,j}=\{u\in[0,1]^d:0<u_j<1\}$7, $V_{d,j}=\{u\in[0,1]^d:0<u_j<1\}$8-convergence plus positive minimal separation between critical points ensures stability of the maxima, minima, saddle counts, with $V_{d,j}=\{u\in[0,1]^d:0<u_j<1\}$9 convergence to a Morse function bolstering this to convergence of individual Morse indices. Empirical verification involves uniform convergence in $a_{ij}(x)$0 and bracketing (entropy) control (Maullin-Sapey et al., 2 Jul 2025).

2. Dependence Structure and Mixing Conditions

Temporal or spatial dependence among observations necessitates quantitative control beyond independence.

• $a_{ij}(x)$1-, $a_{ij}(x)$2-, and $a_{ij}(x)$3-mixing: For stationary sequences, $a_{ij}(x)$4 (absolute regularity) and $a_{ij}(x)$5 (maximal correlation) are quantified via total variation and $a_{ij}(x)$6 correlations of sigma-algebras separated by lag $a_{ij}(x)$7. The key regularity threshold is a polynomial decay $a_{ij}(x)$8, with “phase transitions” in empirical process rates at $a_{ij}(x)$9 (short-range) vs. $b_i(x)$0 (long-range). Optimal rates can still be attained for complex enough function classes, even under persistent dependence (Deb et al., 2024).
• Indicator-based Coefficient Conditions: Instead of classical mixing, strong approximation can be established if the coefficients $b_i(x)$1, defined via indicator function covariances, decay strictly faster than $b_i(x)$2, i.e., $b_i(x)$3 for some $b_i(x)$4. This is essentially optimal for uniform strong approximation in dependent time series (Dedecker et al., 2013).
• Geometric Ergodicity via Drift Conditions: For Markov chain models, geometric Foster-Lyapunov drift (i.e., $b_i(x)$5) together with minorization on a “small” set and regular growth on observables guarantees geometric ergodicity and mixing, which, when appropriately matched with regular variation, ensures weak convergence of tail empirical processes (Kulik et al., 2015).
• Control for Nonstationary Time Series: In nonstationary arrays, sequential empirical process convergence is handled via uniform moment bounds for partial sums and tail control, which rely on mixing coefficients and moment functionals tailored to the process structure (Scholze et al., 2024).

3. Complexity and Entropy Constraints on Function Classes

The function class complexity, measured via covering and bracketing numbers, is critical in controlling oscillations and ensuring tightness.

Complexity Metric	Definition/Requirement	Empirical-Process Target
$b_i(x)$6 bracketing	$b_i(x)$7	Donsker class, classical CLT
$b_i(x)$8 envelope	$b_i(x)$9, $C^\beta(L)$0	New maximal inequalities
VC-, polynomial entropy	$C^\beta(L)$1	Sharp influence on rates
Weighted covering/entropy	$C^\beta(L)$2 with $C^\beta(L)$3	Extension to heavy tails

• Empirical Copula: Weak convergence, validity of bootstrap, and almost-sure rates hinge on $C^\beta(L)$4-covering number bounds for the relevant functionals, combined with smoothness of the copula derivatives (Segers, 2010).
• Diffusions and Empirical Measures: Bracketing number integrals in $C^\beta(L)$5 or $C^\beta(L)$6 must be finite or polynomially bounded in exponent strictly below 2 for the Donsker property. For smooth observable classes, this is equivalent to $C^\beta(L)$7 Hӧlder/Sobolev regularity (Rohde et al., 2010, Kloeckner, 2018).
• Heavy-Tailed Data: For trimmed empirical processes, the Donsker condition shifts from global $C^\beta(L)$8 bracketing to the trimmed law $C^\beta(L)$9, dramatically relaxing complexity requirements and restoring Gaussian limits even when the original moment structure fails (Camara et al., 8 Dec 2025).
• Non-Donsker and Heavy-Tailed Classes: Dudley-type maximal inequalities can be extended to envelope functions in $\beta>d/2$0 or even $\beta>d/2$1 norm, with expected (possibly data-dependent) covering entropies instead of uniform $\beta>d/2$2 bounds. This allows empirical process theory for ReLU networks and robust estimators under infinite-variance noise (Ding et al., 19 Nov 2025).

4. Moment and Tail Conditions

Moment, envelope, and tail restrictions are crucial, especially for divergence of classical variances.

• Minimal Moment for Donsker/CLT: Uniformly $\beta>d/2$3 integrable indexing class (or envelope) is typical except in trimmed or weighted contexts, where only truncated moments are needed (Camara et al., 8 Dec 2025).
• Tail-negligibility and Trimming: When moments diverge globally, as in Pareto or Cauchy laws with $\beta>d/2$4, trimming a symmetric fraction and focusing on the resulting law $\beta>d/2$5 neutralizes the pathological contribution of the tails, producing $\beta>d/2$6-Donsker convergence for all function classes with finite $\beta>d/2$7 moment under $\beta>d/2$8 (Camara et al., 8 Dec 2025).
• Envelope and Weighted Integrability: For new Dudley-type bounds, only the envelope and weight inverses need to be integrable to the specified exponent; uniform boundedness is unnecessary (Ding et al., 19 Nov 2025).

5. Metric and Oscillation Control

Fine control of regularity via metrics and oscillation modulus is essential to ensure equicontinuity and control the modulus of increments.

• Generic-Chaining and Gaussian Surrogate Metrics: Pre-Gaussianity and chaining bounds underpin tightness in Skorokhod or uniform topology, with the empirical process viewed as a random element in $\beta>d/2$9 or $d\geq3$0. Surrogate Gaussian processes and distributional transforms are invoked as in the L-condition of Kuelbs–Kurtz–Zinn (Kuelbs et al., 2010).
• Besov Regularity: The exact path-regularity of the empirical process is characterized in Besov spaces $d\geq3$1, and the sample path falls almost surely in the nonseparable space, not its “little” separable subspace, encoding sharp modulus-of-continuity and rate properties (Lo et al., 2012).
• Maximal Inequalities: Modern maximal inequalities established under minimal envelope or moment conditions are now central; e.g., outer expected $d\geq3$2 covering entropies, weighted $d\geq3$3 metrics, and expected random entropy integrals serve as the backbone for high-probability error bounds, even for non-measurable, non-Donsker, or highly complex classes (Ding et al., 19 Nov 2025, Scholze et al., 2024).

6. Bootstrap, Smoothed Processes, and Nonstationarity

• Multiplier Bootstrap for Copulas: Valid under consistency and uniform boundedness of partial derivative estimators, as long as weak smoothness conditions (continuity in open faces) are met and maximal inequalities on estimator classes are available (Segers, 2010).
• Smoothed and Sequential Empirical Processes: For dependent or nonstationary arrays, weak convergence is established via $d\geq3$4-Lipschitz regularity and bracketing entropy of the indexing (set, function) class; the key is moment control, bounded modulus, and complexity expressed through integrals like $d\geq3$5 (Scholze et al., 2024).
• Diffusion Smoothing (Undersmoothing, Exponential Regimes): In diffusions, undersmoothing bandwidths are admissible if diffusion and drift are smooth ($d\geq3$6), exploitation of maximal occupation time bounds allows exponentially small bandwidths even in $d\geq3$7, and path regularity is guaranteed by exponential-type inequalities and generic chaining (Rohde et al., 2010).

7. Synthesis and Sharpness

Across regimes, the minimality or sharpness of the regularity assumption is a recurrent theme:

• For copulas and Markov structures, smoothness away from boundaries or geometric drift conditions suffice and are optimal in that further weakening precludes uniform CLT (e.g., presence of atoms or slow dependence accelerations) (Segers, 2010, Kulik et al., 2015, Dedecker et al., 2013).
• The balance between dependence (mixing decay) and function class complexity (entropy growth) precisely describes when i.i.d.-level learning rates are recoverable, including in settings with strong long-range dependence (Deb et al., 2024).
• Empirical process limit theorems accommodate robustification (e.g., via trimming) and deep learning enrichments by lifting uniform boundedness and Donsker integrability conditions, provided moment/entropy integrals and envelope control are preserved (Camara et al., 8 Dec 2025, Ding et al., 19 Nov 2025).

• "Asymptotics of empirical copula processes under non-restrictive smoothness assumptions" (Segers, 2010)
• "On the Weak Convergence of the Function-Indexed Sequential Empirical Process and its Smoothed Analogue under Nonstationarity" (Scholze et al., 2024)
• "Uniform Central Limit Theorems for Multidimensional Diffusions" (Rohde et al., 2010)
• "Besov regularity of the uniform empirical process" (Lo et al., 2012)
• "Asymptotic theory and statistical inference for the samples problems with heavy-tailed data using the functional empirical process" (Camara et al., 8 Dec 2025)
• "Strong approximation results for the empirical process of stationary sequences" (Dedecker et al., 2013)
• "New Empirical Process Tools and Their Applications to Robust Deep ReLU Networks and Phase Transitions for Nonparametric Regression" (Ding et al., 19 Nov 2025)
• "Empirical measures: regularity is a counter-curse to dimensionality" (Kloeckner, 2018)
• "The tail empirical process of regularly varying functions of geometrically ergodic Markov chains" (Kulik et al., 2015)
• "Regularity Conditions for Critical Point Convergence" (Maullin-Sapey et al., 2 Jul 2025)
• "An Empirical Process Central Limit Theorem for Multidimensional Dependent Data" (Durieu et al., 2011)
• "A CLT for empirical processes involving time-dependent data" (Kuelbs et al., 2010)
• "Optimal use of auxiliary information : information geometry and empirical process" (Arradi-Alaoui, 2021)
• "Trade-off Between Dependence and Complexity for Nonparametric Learning -- an Empirical Process Approach" (Deb et al., 2024)