Spectral Barron Space

• Spectral Barron Space is a Banach space characterized by an L¹-Fourier norm with polynomial, subexponential, or exponential weights ensuring integrability.
• It underpins dimension‐free neural network approximations and offers rigorous tools for PDE regularity, spectral theory, and inverse problem analysis.
• Its robust interpolation, embedding properties, and extensions to operator-valued and quantum settings make it vital for high-dimensional functional analysis.

A spectral Barron space is a Banach space of functions on Euclidean domains, compact groups, or generalized settings (e.g., operator-algebras or vector-valued functions) characterized by a Fourier-side integrability condition: specifically, integrability of the modulus of the Fourier transform of the function against a polynomial or, in some cases, subexponential or exponential weight. This L¹-based Fourier norm defines a function class that is simultaneously apt for dimension-free approximation by neural networks and for rigorous functional-analytic and PDE-theoretic analyses—including duality, compact embeddings, boundary-value problems, spectral theory, and operator learning—without suffering from the curse of dimensionality. Spectral Barron spaces provide the mathematical backbone for recent results on the regularity of high-dimensional PDEs (especially Schrödinger and Hamilton-Jacobi-Bellman equations), generalization and universal approximation with neural networks, and the analysis of inverse and interpolation problems in function spaces governed by Fourier decay.

1. Formal Definition and Fundamental Properties

Let $f\colon\mathbb{R}^d\to\mathbb{C}$ be a tempered distribution with Fourier transform $\hat f$. For $s\in\mathbb{R}$, the (scalar-valued) spectral Barron space is defined as

$$\mathcal{B}^s(\mathbb{R}^d) := \left\{f\in\mathcal{S}'(\mathbb{R}^d)\; :\; \|f\|_{\mathcal{B}^s} := \int_{\mathbb{R}^d} (1+|\xi|^2)^{s/2} |\hat f(\xi)|\,d\xi < \infty\right\}$$

with norm induced by the weighted $L^1$-norm of the Fourier transform (Chen et al., 2022, Lu et al., 6 Feb 2025). For $s=1$, this coincides with Barron's original definition (Chen et al., 2022). Equivalent formulations appear in (Yserentant, 25 Feb 2025, Liao et al., 2023, Feng et al., 24 Mar 2025) for various domains and group structures.

Key properties include:

• Banach structure: Completeness of the weighted $L^1$-norm establishes that $\mathcal{B}^s$ is a Banach space, not a Hilbert space (Chen et al., 2022, Choulli et al., 9 Jul 2025, Mensah et al., 13 Dec 2025).
• Monotonic embeddings: For $s'\geq s$, $\mathcal{B}^{s'}\hookrightarrow\mathcal{B}^s$ and $\|f\|_{\mathcal{B}^s}\leq\|f\|_{\mathcal{B}^{s'}}$ (Choulli et al., 9 Jul 2025).
• Pointwise control and algebra closure: $\mathcal{B}^0\subset L^\infty$ with $\|f\|_{L^\infty}\leq\|f\|_{\mathcal{B}^0}$. For $f,g\in\mathcal{B}^s$, $fg\in\mathcal{B}^s$ and $$\|\;fg\|_{\mathcal{B}^s}\le\|f\|_{\mathcal{B}^s}\|g\|_{\mathcal{B}^s}$$ (Feng et al., 24 Mar 2025). $\mathcal{B}^s$ is a Banach algebra for $s>\frac{d}{2}$ (Choulli et al., 9 Jul 2025).
• Approximation expressivity: Finite atomic sums $u(x)=\sum_j \alpha_j e^{i\omega_j\cdot x}$ with $\sum_j|\alpha_j|(1+|\omega_j|^2)^{s/2}<\infty$ are dense in $\mathcal{B}^s$ (Yserentant, 25 Feb 2025).

Extensions to vector-valued and operator-valued settings (e.g., on compact groups or quantum systems) replace the Euclidean Fourier transform with the appropriate harmonic analysis machinery, and the $L^1$-control is imposed on the coefficients or quantum-Fourier transforms, with suitable Schatten or trace-class norms (Mensah et al., 13 Dec 2025, Mensah, 18 Sep 2025).

2. Embedding Relations, Interpolation, and Comparison with Other Spaces

Spectral Barron spaces admit precise embedding relations with Sobolev and Besov scales, supporting interpolation and scaling laws:

• Besov and Sobolev Embeddings: There exist sharp, dimension-independent continuous embeddings

$$B^{s+\frac{d}{2}}_{2,1}(\mathbb{R}^d)\hookrightarrow\mathcal{B}^s(\mathbb{R}^d)\hookrightarrow B^s_{\infty,1}(\mathbb{R}^d)$$

with explicit dimension-free constants (Liao et al., 2023).

• Interpolation Inequalities: Real interpolation in the scale generated by $L=I-\Delta$ yields, for $r<s<t$, the inequality

$$\|u\|_{B^s}\leq C\|u\|_{B^r}^\theta \|u\|_{B^t}^{1-\theta}, \quad \theta = \frac{t-s}{t-r}$$

(Lu et al., 6 Feb 2025).

• Barron-type vs Spectral Barron Spaces: The probabilistic (path-norm) Barron space $\mathcal{B}_s(\Omega)$ and the spectral Barron space $\mathcal{F}_s(\Omega)$ (where the norm involves the $L^1$-based Fourier moment of any extension) are related by tight continuous embeddings:

$$\delta \|f\|_{\mathcal{F}_{s-\delta}(\Omega)} \lesssim_s \|f\|_{\mathcal{B}_s(\Omega)} \lesssim_s \|f\|_{\mathcal{F}_{s+1}(\Omega)}$$

with constants independent of $d$ (Wu, 2023). These embeddings are sharp; one cannot remove the exponent shift $+1$ or let $\delta\to 0$. This equivalence (up to a small loss of smoothness) unifies the spectral, neural, and probabilistic perspectives.

• Real Interpolation and Hölder Embedding: For $k\in\mathbb{N}_0$ and $\theta\in(0,1)$, $B^{k+\theta}\hookrightarrow C_b^{k,\gamma}$ for any $0<\gamma<\theta$ with explicit quantitative estimates for Hölder coefficients (Choulli et al., 9 Jul 2025).

3. Regularity in PDEs and Applications to Schrödinger and HJB Equations

Spectral Barron spaces provide an optimal Banach scale for PDE regularity theory, especially in the context of equations admitting high-dimensional or singular data:

• Schrödinger Equation Regularity: For the elliptic PDE $-\Delta u + V(x)u = f(x)$ with $V(x)=\alpha+W(x)$, $\alpha>0$ and $W\in\mathcal{B}^s$, if $f\in\mathcal{B}^s$ then the solution $u\in\mathcal{B}^{s+2}$ and $\|u\|_{B^{s+2}}\leq C(V,d,s)\|f\|_{B^s}$—matching the regularity shift in Sobolev theory but in the $L^1$-Fourier framework (Chen et al., 2022, Choulli et al., 9 Jul 2025).
• Electronic Schrödinger Equation: For the $N$-electron problem under Coulomb potential, bound-state solutions $u$ satisfy $u\in\mathcal{B}^s(\mathbb{R}^{3N})$ for all $s<1$, and this exponent is shown sharp via the hydrogen atom ground state (Yserentant, 25 Feb 2025). Similar results hold for many-particle systems with general Fourier-Lebesgue regularity in the potential (Ming et al., 25 Aug 2025).
• Hamilton-Jacobi-Bellman (HJB) Equations: A discount-thresholded existence/uniqueness theory holds for HJB with spectral Barron data, and the solution can be constructed as the locally uniform limit of bounded spectral Barron functions (Feng et al., 24 Mar 2025).
• Boundary Value and Spectral Problems: On bounded domains, spectral Barron spaces admit quotient and interpolation structures compatible with classical Dirichlet and Neumann theories; spectral analysis of associated linear operators (e.g., $(1-\Delta+V)$) is governed by compactness and point spectrum, with eigenfunctions in all $\mathcal{B}^s$ and hence in $C^\infty$ (Choulli et al., 9 Jul 2025).

4. Neural Network Approximation and Curse-Free Expressivity

Spectral Barron spaces are precisely the function spaces for which next-generation approximation theorems for neural networks—especially shallow and moderately deep two-layer ReLU/cosine networks—admit dimension-free, sharp approximation rates:

• Shallow Networks: For $f\in\mathcal{B}^s$ with $s\in(0,1/2]$ (with current best theory at $s=1/2$), every $N$-width 1-hidden-layer network can achieve $L^p$- or $L^\infty$-error $\mathcal{O}(N^{-1/2})$, uniformly in $d$ (Liao et al., 9 Jul 2025).
• Depth Effect: For $L$-layer, $N$-width networks, $f\in\mathcal{B}^s$ with $sL\leq1/2$, the optimal rate is $\mathcal{O}(N^{-sL})$. This shows that depth amplifies the effect of regularity, and the lower bounds confirm sharpness up to logarithmic factors (Liao et al., 9 Jul 2025, Liao et al., 2023).
• Approximation by Cosine Dictionaries: The integral Fourier representation of $f\in\mathcal{B}^s$ can be interpreted as an infinitely wide cosine network. Monte Carlo truncation preserves the dimension-free rate (Lu et al., 6 Feb 2025, Abdeljawad et al., 2024).
• Operator Learning: For linear operators $$T:u\mapsto\mathcal{F}^{-1}[m(\cdot)\mathcal{F}[u](\cdot)]$$ with $m$ in an exponential spectral Barron class, neural operator approximation algorithms can achieve subexponential error in the Fréchet topology (Abdeljawad et al., 2024).

5. Extensions: Group, Quantum, and Vector-Valued Settings

Spectral Barron spaces generalize beyond $\mathbb{R}^d$:

• Compact Groups: The spectral Barron space $\mathfrak{B}^s_\gamma(G;A)$ consists of $A$-valued functions $f$ on a compact group $G$ whose group-Fourier coefficients $\widehat f(\pi)$ satisfy

$$\sum_{\pi\in\widehat G} d_\pi (1+\gamma(\pi)^2)^{s/2} \sum_{i,j}\|\widehat f(\pi)(\xi_j^\pi,\xi_i^\pi)\|_A < \infty$$

for a suitable weight $\gamma$, with Banach structure and continuous embeddings into Sobolev and $L^\infty$ spaces (Mensah et al., 13 Dec 2025).

• Quantum Harmonic Analysis: For a locally compact abelian group $G$ and projective representation $\rho$, the quantum Fourier transform leads to a spectral Barron space of trace-class operators defined by weighted $L^1$-summability of the Fourier images (Mensah, 18 Sep 2025).
• Vector-Valued and Operator-Valued Functions: Both the quantum setting and vector-valued Barron spaces (involving Schatten norms) preserve the main interpolation, completeness, and embedding properties (Mensah, 18 Sep 2025, Mensah et al., 13 Dec 2025).

6. Interpolation, Inverse Problems, and Regularization

Spectral Barron spaces play a central role in the analysis of inverse problems and regularization:

• Moment Inequalities and Scaling: Real interpolation yields conditional stability exponents that translate regularity across scales and relate noise levels to reconstruction accuracy (Lu et al., 6 Feb 2025).
• Link Conditions for Inverse Operators: For elliptic pseudo-differential operators, the static Schrödinger operator, and the Radon transform, mapping properties between Barron scales induce bi-Lipschitz “link conditions” governing stability and error (Lu et al., 6 Feb 2025).
• Tikhonov Regularization: Penalizing by a higher-order Barron norm in Tikhonov functionals yields rate-optimal recovery guarantees, and empirical neural-network approximants can achieve these rates dimension-independently (Lu et al., 6 Feb 2025).
• Universal Approximation: Classical Monte Carlo arguments adapted to the spectral Barron setting show that two-layer ReLU/RePU networks of width $n$ can achieve $L^2$-error $\mathcal{O}(n^{-1/2})$ for all $f\in\mathcal{B}^{s+1}$, with rates independent of $d$ (Lu et al., 6 Feb 2025).

7. Exponential and Non-Polynomial Weights, Analyticity, and Open Directions

Several developments extend the spectral Barron paradigm to symbol classes and analytic functions:

• Exponential Spectral Barron Spaces: For $f$ defined on $U\subset\mathbb{R}^d$, the exponential spectral Barron space

$$B_{\beta,c}[U] := \left\{f : \exists f_e\text{ with }\int_{\mathbb{R}^d} e^{c|\xi|^{\beta}}|\widehat{f_e}(\xi)|\,d\xi<\infty\right\}$$

captures rapid Fourier decay. Neural network dictionaries then achieve subexponential error in Sobolev/Fréchet metrics, with complexity scaling as $N=O((|\log\epsilon|)^{d/\beta})$ (Abdeljawad et al., 2024).

• Analytic Regularity: Such spaces are related to Gelfand–Shilov classes, and uniform control on derivatives yields sufficient conditions for inclusion in $B_{\beta,c}$ (Abdeljawad et al., 2024).
• Open Problems: Major unresolved areas include deterministic, curse-free neural approximations for spectral Barron functions of high (e.g., $s\gg1$) regularity; extension of the policy-iteration approach for HJB equations to weakly discounted or non-quadratic formulations; boundary behaviors and spectral properties on non-Euclidean or singular domains; and explicit generalization bounds in the sup-norm for deep networks (Liao et al., 2023, Liao et al., 9 Jul 2025, Feng et al., 24 Mar 2025).

Summary Table: Main Variants of Spectral Barron Spaces

Domain/Setting	Fourier Notion	Space Definition and Norm
$\mathbb{R}^d$ (scalar)	Classical FT	$$\|f\|_{\mathcal{B}^s} = \int (1+|\xi|^2)^{s/2}|\hat f|$$
Compact group $G$	Group FT ($\widehat G$)	$$\sum_{\pi\in\widehat G} d_\pi (1+\gamma(\pi)^2)^{s/2} \|\widehat f(\pi)\|$$
Operator-valued/quantum	Quantum FT	$$\int_{\widehat G}(1+\gamma(\xi)^2)^{s/2}|\mathcal{F}_U(T)|\,d\xi$$
Exponential weights	$\mathbb{R}^d$ FT	$\int e^{c|\xi|^\beta}|\hat f|\,d\xi$

Spectral Barron spaces thus provide a robust, unifying functional analytic framework underpinning the analysis of both high-dimensional approximation with neural networks and regularity theory for PDEs. Their dimension-free properties, natural interpolation structure, and connections to both probabilistic and harmonic analytic regimes make them central to current research in analysis, operator learning, and high-dimensional inverse problems (Chen et al., 2022, Liao et al., 2023, Abdeljawad et al., 2024, Choulli et al., 9 Jul 2025, Yserentant, 25 Feb 2025, Feng et al., 24 Mar 2025, Liao et al., 9 Jul 2025, Lu et al., 6 Feb 2025, Mensah et al., 13 Dec 2025, Mensah, 18 Sep 2025, Ming et al., 25 Aug 2025, Wu, 2023).