PaperlayerView exact revision on arXiv

Source-linked AI summary

Operator Barron Regularity for Elliptic Equations and Application to Neural Operator Approximation

Shuhao Cao, Ruchi Guo, Daniel Zhengyu Huang, Na Liu

arXiv:2610.37101v1math.NAcs.NA

TL;DR

The paper addresses how regularity of elliptic solution operators can quantify neural-operator approximation. It develops operator spectral Barron bounds and applies them to an FNO-type network, obtaining arbitrary-order regularity and stretched-exponential width decay for fixed input truncation.

  • Problem

    Neural-operator approximation depends on regularity of the underlying solution maps, while operator-level Barron regularity has been comparatively less explored.

    Source: §1 · p. 1, §8 · p. 23

  • Method

    The paper combines finite-dimensional input truncations with Hilbert-valued Fourier representations to establish operator Barron regularity and derive FNO-type approximation bounds.

    Source: §3 · p. 8, §8 · p. 23

  • Results

    The elliptic solution operator has arbitrary-order Barron regularity, and optimizing the approximation order yields stretched-exponential error decay in network width for fixed input truncation.

    Source: §3 · pp. 7–8, §8 · p. 23

  • Takeaways & Limitations

    The results provide a quantitative link between elliptic PDE structure, operator regularity, and neural-operator approximation.

    Source: §8 · p. 23

    The analysis assumes uniform boundedness and coercivity of the admissible coefficient family, without requiring a smallness condition on the coefficient amplitudes.

    Source: §3 · p. 7

Abstract

from arXiv · show

We develop a spectral Barron regularity framework for nonlinear operators between infinite-dimensional function spaces. Using Hilbert-valued Fourier representations of finite-dimensional input truncations, we establish bounds that are uniform in the input truncation dimension and prove arbitrary-order Barron regularity for a class of elliptic solution operators. Second, we use this operator Barron regularity to derive approximation error bounds for an FNO-type neural operator, with explicit dependence on the input truncation dimension, output reconstruction dimension, and network width. For a fixed input truncation dimension, optimizing the neural approximation bound over the Barron regularity order yields stretched-exponential decay of the approximation error in the network width. We also show some numerical experiments to support the predicted rate.

1. Introduction.

The paper studies regularity of elliptic coefficient-to-solution maps as a basis for quantitative neural-operator approximation. It introduces operator spectral Barron regularity and uses it to establish FNO-type approximation bounds.

1. Introduction.

  • Neural operators can act as surrogates for repeated PDE solves, but their approximation properties depend on regularity of the underlying solution maps.The paper considers an abstract elliptic problem with stochastic coefficient inputs.Source: §1 · p. 1

1.1. Literature review.

  • Prior operator-level Barron regularity is less explored than function-level Barron regularity, motivating a framework for quantifying solution-map dependence on inputs.Much existing work on Barron regularity for PDEs focuses on spatial regularity, while operator regularity concerns smooth dependence on the input function.Source: §1.1 · pp. 2–3

1.2. Contributions.

  • The paper defines an operator spectral Barron norm and proves arbitrary-order regularity for elliptic solution operators, then derives FNO-type approximation bounds.The norm combines finite-dimensional input truncations with Hilbert-valued Fourier representations; its stated bounds depend explicitly on truncation dimensions and network width.Source: §1.2 · p. 3

1.3. Notations.

  • N and M denote the input and output truncation dimensions, respectively.The notation overview also reserves τ, s, k, and q for regularity or approximation orders.Source: §1.3 · p. 4

2. Barron regularity of functions and operators.

The paper extends spectral Barron norms from scalar functions to Hilbert-valued maps and then to operators on a parameterized input family. This construction supports basis-independent bounds on output coefficients for neural reconstruction.

2. Barron regularity of functions and operators.

  • The Hilbert-valued spectral Barron norm uses Fourier representations in the parameter variables and integrable extensions into the complexified Hilbert space.The scalar spectral Barron definition is motivated by Barron’s Fourier-moment condition.Source: §2 · pp. 4–5
  • The operator spectral Barron norm is defined through input-truncated parameterizations of continuous maps from the coefficient family into a Hilbert space.The input family is parameterized by a weighted sequence of modes, with summability ensuring absolute convergence in L∞.Source: §2 · pp. 4–5
  • For any orthonormal output basis, the squared scalar Barron norms of the output coefficients are controlled by the operator Barron norm.This bound enables neural approximation of retained coefficients uniformly over the choice of output basis; truncation error for discarded coefficients depends on the chosen basis.Source: §2 · pp. 5–6

3. The target approximation problem and the main result.

The section sets up approximation of a centered coefficient-to-solution operator for a uniformly well-posed family of elliptic problems. It gives a truncation-uniform Barron regularity result and FNO-type approximation bounds that expose input, output, and width dependence.

3. The target approximation problem and the main result.

  • Uniform boundedness and coercivity, together with the stated summability and complex-extension assumptions, define the coefficient family used for the well-posedness and regularity results.The summability condition permits splitting the parameter sequence into a finite part and a sufficiently small tail.Source: §3 · pp. 6–7
  • The centered operator maps coefficient perturbations g to the solution for coefficient ā + g, aligning its domain with the perturbation parameterization.The input and output truncations are used to study finite-dimensional parameter-to-solution maps and their approximation.Source: §3 · pp. 6–7
  • Under Assumption 3.1, the solution operator is continuous and has spectral Barron bounds for every τ, with constants independent of input truncation N and output dimension M.The bounds depend on fixed problem data and the prescribed data b, but not on τ, M, or N.Source: §3 · p. 7
  • An FNO-type spectral neural operator achieves the stated approximation bound with width max{4N, MK, 4M}, while the error accounts for input truncation and output reconstruction.The input truncation error tends to zero as N increases; the reconstruction error tends to zero as M increases for each fixed N.Source: §3 · p. 8
  • For fixed input truncation N, choosing the Barron regularity index in terms of width K yields nearly exponential decay of the neural approximation error.The section describes this decay as stretched-exponential in the paper context.Source: §3 · p. 8
  • The output reconstruction rate improves when the solution is smoother; the stated M-dependent estimate also uses an additional Weyl asymptotic assumption on background eigenvalues.The eigenvalue condition is λ_m ∼ c_W m^(2/d).Source: §3 · p. 9

4. Basic results in harmonic analysis.

This section develops Fourier-moment tools for Hilbert-valued holomorphic functions and monomial extensions. Dimension-independent monomial moment bounds and product estimates support the paper’s uniform Barron-regularity argument.

4. Basic results in harmonic analysis.

  • A product of Hilbert-valued and scalar functions in disjoint variables has a Barron bound controlled by the Fourier moments of both factors.The lemma assumes integrable extensions with finite zeroth and τ-order moments.Source: §4 · pp. 9–10
  • Holomorphic Hilbert-valued functions admit compactly supported extensions with Fourier-moment bounds simultaneously for all orders 0 ≤ s ≤ τ.The extension estimates’ constants depend on dimension and the neighborhood parameters.Source: §4 · p. 10
  • Monomial extensions have Fourier moment bounds independent of dimension, although the constant-function extension’s L1 norm may depend on dimension and regularity order.The bound is stated for the Fourier moment; the dimension dependence of the constant-function extension’s L1 norm is a separate qualification.Source: §4 · p. 11
  • The regularity proof combines parametric holomorphy with Fourier-moment estimates for holomorphic functions and dimension-independent estimates for monomials.These estimates are used to establish Barron regularity for the parameter-to-solution map with bounds uniform in the input truncation dimension.Source: §4 · pp. 10, 12

5. Barron regularity under uniform coercivity.

The section proves arbitrary-order spectral Barron regularity for elliptic solution maps under uniform coercivity. It extends the parameter-to-solution map holomorphically, controls a partial Taylor expansion in tail variables, and obtains a Barron bound uniform in the input truncation dimension.

5.1. Holomorphic extension near the real cube.

  • Uniform coercivity permits a holomorphic extension of the parameter-to-solution map to a complex neighborhood, where the associated operator remains invertible.The extension is given by uN(zN) = A(zN)−1f, with bounds stated for the map and inverse.Source: §5.1 · p. 13

5.2. Partial Taylor expansion and coefficient estimates.

  • The tail split is chosen using coefficient summability, and only the solution is expanded in the tail variables.This distinguishes the procedure from the cited framework in [14, Section 3].Source: §5.1 · p. 13, §5.2 · p. 14
  • Splitting parameter indices into a finite set and a tail yields a partial Taylor expansion that converges absolutely and uniformly in the V_C norm.The Taylor coefficients are holomorphic on the finite-variable neighborhood and satisfy the bound in (5.12).Source: §5.2 · pp. 14–15

5.3. Proof of Theorem 3.3.

  • Combining the partial Taylor expansion with Fourier-moment estimates yields an operator Barron norm bound uniform in the input truncation dimension.The resulting estimate is transferred from V to H using the embedding V → H.Source: §5.3 · pp. 15–16

6. Approximation error estimates.

The approximation construction combines input extraction, approximation of a truncated coefficient-to-solution map, and spatial reconstruction, separating the total error into three components. Its FNO-type network realizes extraction and reconstruction exactly while approximating the finite-dimensional map.

6. Approximation error estimates.

  • The network combines input parameter extraction, approximation of the truncated coefficient-to-solution map, and final spatial reconstruction, yielding three components in the total error.The construction realizes the input encoder and output decoder exactly and approximates the coefficient map.Source: §6 · p. 16, §6.1 · p. 17

6.1. Network construction.

  • The input encoder extracts the first N parameters exactly with an FNO-type network of width 4N.The construction uses Fourier layers and activations to implement the required inner products and multiplication by Γ⁻¹.Source: §6.1 · pp. 17–18
  • Cosine networks approximate the retained coefficient-to-solution map, while two Fourier layers and an output projection realize spatial reconstruction exactly.The reconstruction uses fixed basis features and has maximum channel width 4M.Source: §6.1 · p. 19

6.2. Proof of Theorem 3.5.

  • The Theorem 3.5 proof bounds input truncation, output reconstruction, and neural approximation separately; the constructed width is max{4N, MK, 4M} and is independent of n.The estimates are combined and the supremum is taken over n ≥ N.Source: §6.2 · pp. 20–21

6.3. Proof of Theorem 3.6.

  • Theorem 3.6 combines the neural approximation bound with input and output truncation terms to give the total error estimate.The section’s construction uses arbitrary-order Barron regularity to optimize the order with respect to network width, yielding nearly exponential convergence.Source: §6 · p. 16, §6.3 · p. 21

7. Numerical Experiments.

The experiments vary input truncation dimension, output modes, and network width for a one-dimensional elliptic problem. The observed error trends broadly match the truncation predictions, and the width experiments are consistent with the predicted stretched-exponential rate.

Experimental setup

  • The experiments use a one-dimensional parametric elliptic problem with coefficient decay exponents r ∈ {1.5, 2, 3}, and evaluate L2 test error on 8192 test pairs.For the truncation experiments, the FNO has four layers and 16 Fourier modes per layer and is trained for 600 epochs.Source: §7 · p. 21, Figure 7.1 · p. 22

Error versus truncation dimensions and width

  • The observed N- and M-dependent convergence rates are broadly consistent with theory, while the width-dependent rate is slightly faster than K^-1/2.The curves level off when other error contributions become dominant.Source: §7 · p. 22, Figure 7.1 · p. 22

Stretched-exponential width rate

  • For r = 2 and N = 1 or N = 2, the observed width-dependent decay is consistent with the fitted rate AK^-1/2 exp(-cK^(1/(2N))).These tests use finite-parameter input families, so no input-truncation error arises relative to the target.Source: §7 · p. 22, Figure 7.2 · p. 22

8. Conclusion.

The paper develops a spectral Barron framework for nonlinear operators and establishes arbitrary-order regularity for elliptic solution operators. It uses this regularity to derive FNO-type approximation bounds and reports experiments consistent with predicted rates.

Conclusion

  • The paper links elliptic solution-operator regularity to FNO-type approximation, with experiments illustrating predicted roles of input regularity, truncation dimensions, and network width.Optimizing the Barron regularity order yields a nearly exponential approximation rate with respect to network width.Source: §8 · p. 23

Appendix A. Representation of Barron spaces.

Appendix A establishes closure of scalar Barron spaces under products and bounds the Barron norm of functions represented by weighted trigonometric coefficient sequences. The proof uses a cutoff extension and Fourier estimates.

Product closure

  • For p > 0, the product of two functions in Bp(Rd) also belongs to Bp(Rd).The proof uses the Fourier product formula and a weight inequality.Source: Appendix A · p. 23

Coefficient representation

  • The trigonometric-series map Φ sends weighted coefficient sequences into Bp(Ω), with norm bounded by CΩ,p times the weighted sequence norm.The series converges absolutely and uniformly on QL.Source: Appendix A · p. 23

Extension argument

  • A cutoff function yields an admissible extension of Φ(α) to Bp(Rd), enabling the Barron norm estimate on Ω.The proof defines the extension by multiplying the coefficient series by the cutoff and extending it by zero outside QL.Source: Appendix A · pp. 23–24

Appendix B. Barron regularity of cutoff functions.

The appendix constructs smooth compactly supported cutoff functions and establishes their Barron regularity. It also estimates derivatives of the cutoff construction.

Appendix B. Barron regularity of cutoff functions.

  • For every p > 0, a cutoff satisfying (B.1) belongs to Bp(Rd).The proof uses rapid Fourier-transform decay obtained by repeated integration by parts.Source: Appendix B · p. 24
  • For q ∈ N and δ > 0, the construction gives a Cq cutoff equal to 1 on [−1, 1] and supported in [−1 − δ, 1 + δ].It is built by integrating a scaled cardinal B-spline.Source: Appendix B · pp. 24–25
  • The appendix estimates derivatives of the cutoff using convolution inequalities, including Young’s inequality for functions and finite signed measures.The derivatives vanish where the cutoff is constant.Source: Appendix B · p. 25

Appendix C. Explicit Barron approximation by cosine networks.

This appendix proves an explicit cosine-network approximation rate for Barron functions in probability L2. It improves direct random sampling by correcting its error with a polynomial projection.

Appendix C. Explicit Barron approximation by cosine networks.

  • A real single-hidden-layer cosine network of width at most K approximates every real-valued g ∈ Bτ(UN) with error O(K^-1/2-τ/N).The appendix proves this rate in probability L2 with an explicit constant depending only on the Barron regularity order.Source: Appendix C · pp. 25–26
  • For p ≥ τ, the proof projects the sampling error onto the orthogonal complement of polynomials of degree at most p − 1, then approximates its polynomial component by another cosine network.The construction reserves pN hidden units for the polynomial correction and uses the remaining units for samples.Source: Appendix C · pp. 26–27
  • For p < τ, the proof instead uses the averaging argument without polynomial projection, and combines the two cases to prove the stated rate.The polynomial projection estimate relies on approximating cosine atoms by total-degree polynomials.Source: Appendix C · pp. 25–27

Appendix D. Input truncation errors.

The appendix bounds the error from truncating the PDE coefficient under decay and frequency-ordering assumptions. Its estimate is uniform over input parameters and yields an L2 error rate in the truncation dimension.

Appendix D. Input truncation errors.

  • The truncation estimate compares solutions for the full and truncated coefficients, with the truncated solution viewed as a function on the larger input space by appending zeros.The constant Ctruc depends only on Ω and Creg.Source: Appendix D · p. 28
  • If |bj| ≤ Cσ(j + 1)^-r, the modes are ordered by nondecreasing frequency, and r + 1/d > 1, then truncation error is bounded by Ctruc Cσ cd,L (r + 1/d − 1) ∥f∥L2(Ω) N^(1−r−1/d).This bound holds for N ≥ Nd,L; the constants cd,L and Nd,L depend only on d and L.Source: Appendix D · p. 28
  • The proof obtains a bound uniform in the input parameter, then integrates over its probability measure and takes a supremum over larger truncation levels.The argument uses integration by parts, a trace estimate, the product rule, and Cauchy–Schwarz.Source: Appendix D · p. 28
Loading 2610.37101v1…