Analytic Regularity Bounds Overview

• Analytic regularity bounds are explicit factorial-type estimates for weighted high-order derivatives, ensuring that solutions inherit analytic structure even near singularities.
• They integrate methods from PDEs, harmonic analysis, stochastic analysis, and algebraic geometry to guide optimal numerical schemes and quantify approximation barriers.
• This framework underpins exponential convergence in finite element and spectral methods and constrains approximation rates in neural network models.

Analytic regularity bounds quantify how solutions to differential equations and related problems inherit analytic (or sub-analytic) structure from data, often in the presence of geometric singularities, nonlinearities, or other obstructions. In many contexts, this is formalized as explicit factorial-type bounds for weighted high-order derivatives, indicating that the solution, when appropriately reweighted, behaves essentially as an analytic function. These bounds play a critical role in regularity theory, enabling optimal numerical approximation rates and underpinning approximation-theoretic barriers for model classes, such as neural networks constrained by analytic data. The theory integrates methods from PDEs, harmonic analysis, stochastic analysis, and algebraic geometry, and has widespread influence in numerical analysis, uncertainty quantification, and algebraic geometry.

1. Weighted Analytic Regularity in Domains with Singular Geometries

Analytic regularity in non-smooth domains, such as polyhedra or polygons, requires weighted function spaces whose norms account for the singular behavior near corners, edges, and faces. For the Dirichlet problem for the integral fractional Laplacian on a polyhedral Lipschitz domain $\Omega \subset \mathbb{R}^3$,

$$(-\Delta)^s u = f \text{ in } \Omega, \quad u=0 \text{ in } \mathbb{R}^3\setminus\overline{\Omega},$$

where $0 < s < 1$ and analytic data $f$, analytic regularity is captured by bounds of the form

$$\|r_v^{a_v} r_e^{a_e} r_f^{a_f} \partial_x^{\alpha}u\|_{L^2(\Omega)} \leq C R^{|\alpha|} |\alpha|!,$$

where $r_v(x), r_e(x), r_f(x)$ denote the distances to vertices, edges, and faces respectively, and the exponents $a_v, a_e, a_f$ are chosen to compensate exactly for the local singularity profiles of the derivatives of $u$.

The construction utilizes the Caffarelli–Silvestre extension for localization, geometric decompositions near singular strata (vertices, edges, faces, and their intersections), and a bootstrapping argument involving careful use of weighted Caccioppoli inequalities on dyadic decompositions. This leads to the analytic control of all weighted derivatives, showing that the singularities are fully compensated by explicit distance weights (Faustmann et al., 2023).

This weighted analytic regularity structure generalizes the classical (unweighted) analytic regularity available in smooth domains, and it provides the analytic foundation for high-order numerical methods, including the derivation of exponential convergence rates for hp-finite-element methods on appropriately graded meshes.

2. Analytic Regularity and Approximation Barriers in Neural Network Classes

Analytic regularity bounds are crucial in understanding the limitations of approximation capabilities of coefficient-constrained shallow neural networks with analytic (or Gevrey-regular) activation functions. For networks of the form

$$g(x) = \sum_{k=1}^m \lambda_k \varphi(\alpha_k x),\qquad \sum_{k=1}^m |\lambda_k| \le B_m,\ |\alpha_k| \le L_m,$$

with analytic activation $\varphi$, all outputs $$(-\Delta)^s u = f \text{ in } \Omega, \quad u=0 \text{ in } \mathbb{R}^3\setminus\overline{\Omega},$$0 are holomorphic in a Bernstein ellipse and satisfy

$$(-\Delta)^s u = f \text{ in } \Omega, \quad u=0 \text{ in } \mathbb{R}^3\setminus\overline{\Omega},$$1

for explicit constants $$(-\Delta)^s u = f \text{ in } \Omega, \quad u=0 \text{ in } \mathbb{R}^3\setminus\overline{\Omega},$$2 set by the parameter bounds and the analyticity domain of $$(-\Delta)^s u = f \text{ in } \Omega, \quad u=0 \text{ in } \mathbb{R}^3\setminus\overline{\Omega},$$3. Uniform analyticity of the network outputs means their best polynomial approximation error decays exponentially, and a comparison argument yields a rigidity result: for any target $$(-\Delta)^s u = f \text{ in } \Omega, \quad u=0 \text{ in } \mathbb{R}^3\setminus\overline{\Omega},$$4,

$$(-\Delta)^s u = f \text{ in } \Omega, \quad u=0 \text{ in } \mathbb{R}^3\setminus\overline{\Omega},$$5

where $$(-\Delta)^s u = f \text{ in } \Omega, \quad u=0 \text{ in } \mathbb{R}^3\setminus\overline{\Omega},$$6 is the polynomial best-approximation error of degree $$(-\Delta)^s u = f \text{ in } \Omega, \quad u=0 \text{ in } \mathbb{R}^3\setminus\overline{\Omega},$$7 (Attali, 8 Jan 2026). As a consequence, such networks cannot outperform polynomial approximation rates on non-analytic targets, establishing a sharp approximation barrier tied directly to analyticity properties enforced by the coefficient constraints.

If the activation is only Gevrey-regular ($$(-\Delta)^s u = f \text{ in } \Omega, \quad u=0 \text{ in } \mathbb{R}^3\setminus\overline{\Omega},$$8, $$(-\Delta)^s u = f \text{ in } \Omega, \quad u=0 \text{ in } \mathbb{R}^3\setminus\overline{\Omega},$$9), the synthesizable functions admit sub-exponential regularity bounds, and the approximation barrier becomes sub-exponential in the width $0 < s < 1$0. Extensions to multidimensional inputs use polyelliptic tube analyticity and multivariate Bernstein inequalities.

3. Analytic Regularity in Nonlinear and Parametric PDEs

In high-dimensional parametric and nonlinear PDEs, such as semilinear reaction-diffusion systems with analytic or Gevrey-regular parametric data,

$0 < s < 1$1

analytic and Gevrey regularity for the coefficient maps is inherited by the parametric solution map $0 < s < 1$2. The analytic regularity is formalized for each multi-index $0 < s < 1$3 as

$0 < s < 1$4

with explicit control on the constants in both analytic ($0 < s < 1$5) and Gevrey ($0 < s < 1$6) cases. The proof utilizes alternative-to-factorial combinatorics, which improves over naïve brute-force partitions and provides factorial-type estimates while controlling combinatorial constants (Chernov et al., 2023).

Quantitative regularity leads to exponential or root-exponential convergence rates for quadrature schemes, such as Gaussian or quasi-Monte Carlo integration, and the analytic radius deteriorates with increased nonlinearity.

4. Weighted Analytic Regularity for Linear Elliptic Systems and the Navier-Stokes Equations

For classical elliptic systems in domains with corners/edges, analytic regularity estimates are obtained in weighted Sobolev or analytic classes: $0 < s < 1$7 in 2D polygons, and in 3D polyhedra,

$0 < s < 1$8

for explicit weights related to the geometry (vertex and edge distances, dihedral angles) (Costabel et al., 2010). The proof combines dyadic partitioning near corners with nested open-set estimates along edges, yielding global analytic control compatible with the anisotropy of singularities.

For the stationary incompressible Navier-Stokes equations in 2D polygons, similar corner-weighted analytic regularity bounds are established under small analytic data: $0 < s < 1$9 for velocity and pressure, with the weights determined by the local spectral gaps at each corner (Marcati et al., 2020). This validates exponential decay of $f$0-widths and exponential convergence rates for high-order finite element and spectral discretizations.

5. Analytic Regularity for Stochastic Processes and Algebraic Contexts

Regularity bounds also arise in stochastic analysis and homological algebra. For regularity properties of SDE densities, pointwise Fourier analytic bounds enable the derivation of spatial Hölder regularity for local densities $f$1 of SDE solutions under minimal assumptions on the coefficients. Explicit decay of the Fourier transform,

$f$2

for any $f$3, translates into $f$4 spatial regularity for the density itself. The approach leverages Lamperti transforms, Euler approximations, and careful pathwise error control, circumventing the use of Malliavin calculus. Joint continuity in $f$5 follows directly (Ellinger, 30 Apr 2025).

In the algebraic field, the Castelnuovo-Mumford regularity of complexes and their homology admits precise bounds via an Auslander-Buchsbaum-type equality: $f$6 under explicit homological depth and dimension conditions. This refines classical Tor/Ext regularity bounds and gives methods to compute sharp regularity for modules, complexes, and their quotient/extension objects in commutative algebra (Nguyen, 2013).

6. Implications for Numerical Methods and Computational Complexity

Analytic regularity bounds imply that, after compensating for singularities, the regularized solution admits factorial-type control of its derivatives, a hallmark of analytic functions. This structure ensures that hp-finite element methods, spectral discretizations, and reduced-basis approaches achieve exponential-rate convergence, provided the mesh or basis is adapted to the location and strength of singularities. For example, in the context of the fractional Laplacian on polyhedra, the error satisfies

$f$7

where $f$8 is the number of degrees of freedom, and the exponent reflects the analytic regularity radius and combinatorial structure of singular sets (Faustmann et al., 2023).

In approximation theory, analytic regularity rigidifies the function class, placing hard lower bounds on best-approximation errors by neural networks or other bases, unless the target function is analytic on a suitable domain (Attali, 8 Jan 2026).

These analytic regularity bounds thus underpin both theoretical limits and practical performance in high-accuracy simulation, uncertainty quantification, machine learning approximation theory, and computational algebra.