Pith. sign in

REVIEW 2 major objections 5 minor 37 references

On Generalized Barron Spaces for Shallow Neural Networks

T0 review · 2 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A generalized Barron space $B^\varphi_\sigma$ is a Banach space whose $\varphi$-weighted parameter norm controls Sobolev derivatives up to order $k$ when the activation and $\varphi$ satisfy a matching growth condition.

desk verdict A genuinely useful phi-weighted Barron space framework with a clean Sobolev embedding theorem, but the necessity argument in Remark 2.4 has a divergence error and the higher-rate approximation theorem is stated without proof. read the letter →

arxiv 2608.06843 v1 pith:VXSKWLLL submitted 2026-08-07 math.NA cs.NA

Shuai Lu , Leqi Zhang This is my paper · ORCID
classification math.NAcs.NA MSC 46E3541A2568T07
keywords generalizedBarronspacesshallowneuralnetworksSobolevembeddingnormfunctionactivationapproximationratesTikhonovregularizationspectral
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Classical Barron spaces describe functions that shallow neural networks can approximate well, but they are tied to particular activations such as ReLU or RePU and to parameter norms that rely on homogeneity. This paper proposes a generalized Barron space $B^\varphi_\sigma$ in which a non-decreasing norm function $\varphi$ is paired with an arbitrary activation $\sigma$, and the norm is the infimum of $\int \varphi(\|w\|_1+|b|)\,d|\rho|$ over all parameter measures representing the function. The central claim is that this space is a Banach space continuously embedded in $L^\infty$, and that when the activation's derivatives satisfy a matching growth bound against $\varphi$, every function in $B^\varphi_\sigma$ has bounded Sobolev derivatives up to order $k$ with norm controlled by the generalized Barron norm. If true, the choice of $\varphi$ becomes a dial for smoothness: faster-growing $\varphi$ yields higher-order Sobolev regularity while keeping the same activation function, and the framework unifies spectral and (extended) Barron spaces as special cases. The paper also derives dimension-independent approximation rates and uses the $\varphi$-norm as a Tikhonov regularizer, with numerical tests showing stable high-order derivative recovery.

What carries the argument

The central object is the generalized Barron space $B^\varphi_\sigma$ defined by the integral representation and the weighted norm in Definition 2.2, together with the norm function $\varphi$ from Definition 2.1. The machinery that carries the argument is the compatibility condition (5) in Theorem 2.3: each derivative $\partial^m\sigma$ is dominated by $C\varphi(|x|)/(1+\mathrm{ReLU}(|x|))^m$ with $\varphi_k$ non-decreasing. This inequality is what permits differentiating under the integral to obtain the weak-derivative formula (6) and then to bound $|\partial^\alpha f(x)|$ by $C\|f\|_{B^\varphi_\sigma}$; without it, the $\varphi$-norm does not control high-order derivatives, as the sine activation with $\varphi=1+\mathrm{ReLU}$ shows. A second device, the finite signed complex kernel of Definition 2.3 and Lemma 2.8, carries the embedding relations between spaces with different $\sigma$ and $\varphi$ pairs, including the Taylor and Fourier representations used in Propositions 2.5, 2.6, and 2.11.

What would settle it

For $\sigma(x)=\sin(x)$ and $\varphi(x)=1+\mathrm{ReLU}(x)$, the paper's own Remark 2.4 constructs a function with finite $B^\varphi_\sigma$-norm whose derivative is a nowhere-differentiable trigonometric series, so $B^\varphi_{\sin}$ is not embedded in $W^{2,\infty}$ exactly when condition (5) fails for $m=2$. To test the sufficiency side, repeat the construction with $\sigma(x)=\sin(x)$ and $\varphi(x)=\exp(x)$, where (5) holds for every $k$: the theorem predicts that the $k$-th derivative of every finite-$\varphi$-norm function is bounded by a constant times the norm, so numerical differentiation of the same series should show bounded high-order derivatives whose bounds scale with the $\varphi$-norm; any divergence of these derivatives while the $\varphi$-mass stays bounded would refute the embedding.

Watch

Extended reading notes

Core claim

In the author's own terms, functions represented as $f(x)=\int \sigma(w\cdot x+b)\,d\rho(w,b)$ form a normed Banach space $B^\varphi_\sigma$ when the parameter measure is weighted by $\varphi(\|w\|_1+|b|)$, and this space is continuously embedded in $L^\infty$. Under the matching condition $|\partial^m \sigma(x)|\le C\varphi_m(|x|)$ with $\varphi_m=\varphi/(1+\mathrm{ReLU})^m$ and $\varphi_k$ positive and non-decreasing, the same $\varphi$-norm controls all weak derivatives up to order $k$, giving the embedding $B^\varphi_\sigma\hookrightarrow W^{k,\infty}(\Omega)$. Consequently the generalized Barron norm simultaneously measures parametric complexity and classical smoothness, and the framework recovers the extended Barron spaces for $\mathrm{ReLU}^s$ activations (with $\varphi=1+\mathrm{ReLU}^t$, $t\ge s$) and the spectral Barron spaces for $\sigma=\exp(i2\pi x)$ (with $\varphi=(1+\mathrm{ReLU})^s$) with equivalent norms.

Load-bearing premise

The load-bearing premise is the matching condition (5): for each $m\le k$, the $m$-th derivative of the activation $\sigma$ must be bounded by $C\varphi(|x|)/(1+\mathrm{ReLU}(|x|))^m$ with $\varphi_k$ positive and non-decreasing; if $\varphi$ grows too slowly relative to the activation's derivatives, the $\varphi$-norm no longer guarantees Sobolev regularity, as the sine example shows.

Editorial extensions

If this is right

  • If $f$ has finite $\varphi$-norm and condition (5) holds, then the weak derivatives of $f$ up to order $k$ are bounded by a constant times its Barron norm, so the $\varphi$-norm acts as a control on smoothness during training.
  • Choosing a faster-growing $\varphi$ raises the guaranteed Sobolev order for the same activation; for example, an exponential $\varphi$ gives $W^{k,\infty}$ embeddings for every $k$ for activations such as sine whose derivatives are globally bounded.
  • The norm equivalences with spectral and extended Barron spaces mean that existing approximation and regularity results for those spaces transfer to the generalized spaces with their $\varphi$-weights.
  • A type-p sampling argument yields dimension-free approximation rates of order $O(n^{1/p-1})$ and faster rates under tail-decay and smoothness conditions, extending known rates to non-homogeneous activations.
  • Tikhonov regularization penalized by the discrete $\varphi$-weighted cost gives $W^{m,p}$ error bounds of order $(\delta+r_n)^{(k-m)/k}$ under the parameter choice $\delta+r_n\asymp\lambda^{1/p}$.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the embedding is sharp, one could select $\varphi$ from data to prescribe the Sobolev order of the function class for a fixed activation, making regularity an explicit architectural knob rather than a property of the activation.
  • The $\varphi$-weighted regularizer is an activation-aware alternative to Sobolev-norm penalties for derivative recovery and PDE problems; the paper's own 5D experiments suggest a milder growth rate may be preferable in high dimensions, a trade-off it does not fully resolve.
  • The kernel-based embedding theorem provides a general recipe: any change of activation representable as an integral kernel with controlled $\varphi$-mass yields an embedding between generalized Barron spaces, so wavelet or other atomic representations could generate new embedding chains beyond the Taylor and Fourier cases.
  • A direct testable extension is to check numerically whether the predicted $W^{k,\infty}$ bounds hold with reasonable constants for oscillatory activations with exponential $\varphi$ on random finite networks, or whether the constant in condition (5) makes the bound vacuous in practice.
Share X Bluesky LinkedIn Reddit HN

Formalized claims in Lean

  1. Claim #1: In the author's own terms, functions represented as $f(x)=\int \sigma(w\cdot x+b)\,d\rho(w,b)$ form a normed Banach space $B^\varphi_\sigma$ when the parameter measure is weighted by $\varphi(\|w\|_1+|b|)$, and this space is continuously embedded in $L^\infty$. Under the matching condition $|\partial^m \sigma(x)|\le C\varphi_m(|x|)$ with $\varphi_m=\varphi/(1+\mathrm{ReLU})^m$ and $\varphi_k$ posi

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The manuscript introduces a generalized Barron space B^phi_sigma for shallow neural networks with a generic activation sigma and a norm function phi, defined by phi-weighted integral norms over parameter measures. It proves that B^phi_sigma is a Banach space continuously embedded in L^infinity (Proposition 2.2), and that under a matching condition between phi and the derivatives of sigma it embeds continuously into W^{k,infinity} (Theorem 2.3). It then relates these spaces to extended Barron spaces and spectral Barron spaces via norm equivalences (Propositions 2.5 and 2.6), establishes an embedding criterion between generalized spaces (Theorem 2.9), derives approximation rates in type-p spaces (Theorems 3.1 and 3.2), and gives error bounds for Tikhonov regularization with the generalized Barron norm (Proposition 3.3 and Theorem 3.4). Numerical experiments in one and five dimensions illustrate the effect of the norm function on derivative approximation.

Significance. If correct, the framework provides a unified and flexible way to associate a norm with any sufficiently regular activation function, and the Sobolev embedding theorem (Theorem 2.3) cleanly explains how the choice of phi controls the regularity of the realized functions. The proof of the Banach-space property and the main embedding are self-contained and appear sound, and the norm equivalences with classical spaces are plausible and largely verified. The regularization application is a useful addition. However, the paper currently contains an invalid necessity example (Remark 2.4) and states a central approximation-rate theorem (Theorem 3.2) without proof; these issues must be fixed before the claims can be fully accepted.

major comments (2)
  1. [Remark 2.4 (Section 2.1)] For f(x)=sum_{i=1}^infty 2^{6-i} pi^{-1} sin(13^i pi x) in Remark 2.4, the bound ||f||_{B^{1+ReLU}_{sin}} <= sum_i 2^{6-i} pi^{-1}(1+13^i pi) is incorrect because the summand behaves like 2^6 (13/2)^i and the series diverges. Moreover, the displayed derivative f'(x)=sum 2^{-i} cos(13^i pi x) is not the derivative of the displayed f, whose derivative coefficients are 2^6 (13/2)^i. Therefore the example does not show B^{1+ReLU}_{sin} not subset of W^{2,infinity}, and the claimed necessity of condition (5) is not established. The authors should correct the coefficients (e.g., use amplitudes 2^{6-i}(13^i pi)^{-1}) or remove the necessity claim and state that condition (5) is sufficient.
  2. [Theorem 3.2 (Section 3.1)] The higher approximation rate in Theorem 3.2 is stated without proof; the text says 'For brevity, we omit the detailed proof.' This is a central application advertised in the abstract, and the result involves new elements (tail decay theta(R), the Lip^infty(s,X) condition on P_sigma/phi, and the exponent alpha/(s+alpha) * s/(d+1) + 1/p - 1) that are not immediate consequences of [32]. The paper should either supply a complete proof in an appendix or clearly mark the theorem as conditional on a full adaptation of [32].
minor comments (5)
  1. [Proposition 2.5 (Section 2.2)] The proof of Proposition 2.5 contains garbled notation in the definition of rho_g (e.g., 'R+1_D' and missing exponents) and the choice g=||w||_1+|b| is undefined where ||w||_1+|b|=0. Rewriting the change of variables with a positive measurable g and explicitly treating the zero set would make the proof readable.
  2. [Sections 2.2-2.3] The same symbol B^s is used for different spaces: in Proposition 2.5 it denotes the extended Barron space defined via ReLU^s parameter norms, while in Proposition 2.6 and Remark 2.7 it denotes the spectral Barron space defined via weighted Fourier norms. This overloading makes claims such as the equivalence in Remark 2.7 ambiguous and should be fixed with distinct notations.
  3. [Section 4.2, Tables 1-4] The discussion emphasizes variance across random seeds, but the tables report only mean relative errors; reporting standard deviations or a variance measure would support the stability claims made in the text.
  4. [Theorem 2.3 (Section 2.1)] The theorem assumes phi_k is non-decreasing and positive, while the proof uses non-decreasing phi_m for all m <= k. This follows because phi_m = phi_k (1+x)^{k-m}, but stating this implication explicitly would improve readability.
  5. [Theorem 3.2 (Section 3.1)] The approximation exponent is printed without parentheses as n^{-alpha/(s+alpha) * s/d+1 + 1/p - 1}, which is ambiguous; the intended expression should be written as n^{- (alpha/(s+alpha))*(s/(d+1)) + 1/p - 1}.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the generalized Barron embedding is proved from stated growth conditions; the only flaw (Remark 2.4) is a non-circular correctness gap.

full rationale

The paper's central chain is self-contained rather than circular. Definition 2.2 fixes B^φ_σ via a φ-weighted parameter-measure norm; Proposition 2.2 derives the Banach and L∞ properties directly from Definition 2.1's growth and monotonicity assumptions, not from the desired conclusion. Theorem 2.3 is a conditional embedding: given the derivative-growth condition (5) on σ, the proof differentiates the integral representation and obtains each Sobolev derivative bound by ∫φ(||w||_1+|b|)d|ρ|; the conclusion is not part of the definition and the condition concerns σ, not the target f. Propositions 2.5 and 2.6 prove norm equivalences with independently defined extended and spectral Barron spaces by explicit rescaling and Fourier/push-forward arguments, so the 'unification' claim is not a renaming of known results. Theorem 2.9's transfer machinery rests on a standard kernel lemma and is then applied through Taylor, Fourier, and Gevrey estimates; Proposition 2.11 is a concrete estimate, not an assumed embedding. Self-citations ([7], [19], [21]) supply background definitions and a previously proved embedding B^k↪W^{k,∞} that is used only as motivation before Theorem 2.3; the main proofs do not reduce to them. The only notable flaw is Remark 2.4's necessity claim: the displayed bound ∑2^{6-i}π^{-1}(1+13^iπ) diverges like ∑(13/2)^i, so the constructed f is not shown to lie in B^{1+ReLU}_{sin}, and the remark therefore does not establish non-embedding into W^{2,∞}. That is a correctness gap in a motivating remark, not a circularity, so it does not raise the circularity score.

Assumptions & free parameters 2 free parameters · 7 assumptions · 1 invented entities

The central claim rests on standard measure theory and on explicit compatibility assumptions between sigma and phi. No fitted parameters are used in the theory; the two listed free parameters are experimental tuning choices. The only invented entity is the new function space itself, which is the paper's contribution rather than an ad hoc explanatory device.

free parameters (2)
  • lambda_c (regularization scaling) = 0.1
    Hand-chosen in the numerical experiments (Section 4.1) to set the regularization strength; it is an experimental tuning constant, not part of the theoretical derivations.
  • noise level delta = 0.2
    Chosen relative noise in the numerical experiments (Equation 17); it is an experimental setting that affects the reported error tables but is not fitted to the theory.
assumptions (7)
  • standard math Fubini's theorem and measure-theoretic tools for finite signed complex Radon measures, including total variation and pushforward measures.
    Used throughout the definitions and proofs, especially in Proposition 2.2, Theorem 2.3, and Lemma 2.8.
  • domain assumption The activation sigma is Borel measurable; for Sobolev results, sigma is in W^{k,1}_loc and satisfies the derivative growth condition (5) with phi_k positive and non-decreasing.
    These are the hypotheses of Definition 2.2 and Theorem 2.3; the Sobolev embedding conclusion is conditional on them.
  • domain assumption The norm function phi is positive, non-decreasing, and dominates |sigma| (Definition 2.1).
    This is the defining compatibility between the activation and the weight, and it is required for the Banach space and L^infinity embedding in Proposition 2.2.
  • domain assumption The target space X is a type-p Banach space for p in (1,2] in the approximation theorem.
    Theorem 3.1 relies on the Maurey-Barron sampling argument, which requires the type-p inequality (12).
  • domain assumption In Theorem 2.9, sigma2 admits a finite signed complex kernel representation in terms of sigma1 with a uniform weighted bound (8).
    This is the key structural assumption enabling embeddings between spaces with different activations; it is verified only for examples such as Taylor and Fourier representations.
  • standard math Gevrey-class cutoff functions with exponential Fourier decay exist, as stated in Theorem 1.6.1 of [27].
    Used in Proposition 2.11 to construct the compactly supported cutoff chi whose Fourier transform decays like exp(-epsilon'||xi||_1^q).
  • standard math Sobolev interpolation between L^p(Omega) and W^{k,p}(Omega) holds for 0 <= m <= k.
    Used in Theorem 3.4 to convert L^p error bounds into W^{m,p} error bounds.
invented entities (1)
  • Generalized Barron space B^phi_sigma with norm function phi
    purpose: To provide a unified normed function space for shallow neural networks in which the weight phi can be tailored to the activation sigma and controls Sobolev regularity.
    This is a new mathematical object introduced in Definition 2.2. Its properties are established by the paper's proofs, but it is not a physical entity with an external falsifiable handle; 'independent evidence' in the empirical sense does not apply.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On Generalized Barron Spaces for Shallow Neural Networks." pith.science (2026). https://pith.science/paper/VXSKWLLL

@misc{pith2026260806843,
  author       = {Pith},
  title        = {Pith review of: On Generalized Barron Spaces for Shallow Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VXSKWLLL}},
  note         = {Machine review of arXiv:2608.06843}
}
abstract

Classical Barron spaces are function spaces specifically designed for shallow neural networks mostly with ReLU, $\mathrm{ReLU}^k$ (RePU) or Lipschitz continuous activation functions. In the present work, we introduce a generalized Barron space \( B_{\sigma}^{\varphi} \) for shallow neural networks with a generic activation function possessing certain smoothness properties. The subscript \( \sigma \) denotes the activation function, while the superscript \( \varphi \) controls the smoothness of the generalized Barron spaces, defined via a \( \varphi \)-weighted integral norm imposed on neural network parameter measures. Under certain assumptions on \( \varphi \) and \( \sigma \), we show that \( B_{\sigma}^{\varphi} \) can be continuously embedded into Sobolev spaces. We also explore the relationships among various function spaces for shallow neural networks, demonstrating that our definition encompasses most conventional ones. As applications of the proposed generalized Barron spaces, we derive approximation rates within these spaces and establish error bounds for numerical differentiation with regularization penalized by the newly introduced generalized Barron norm. Numerical examples confirm that the proposed spaces allow the construction of neural networks with varying degrees of smoothness while using the same activation function.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 31 canonical work pages

  1. [32]

    Siegel and Jinchao Xu

    Jonathan W. Siegel and Jinchao Xu. Sharp bounds on the approximation rates, metric entropy, andn-widths of shallow neural networks.Found. Comput. Math., 24(2):481–537, 2024

    work page 2024

  2. [1]

    Weighted Sobolev approximation rates for neural networks on unbounded domains.arXiv preprint arXiv:2411.04108, 2024

    Ahmed Abdeljawad and Thomas Dittrich. Weighted Sobolev approximation rates for neural networks on unbounded domains.arXiv preprint arXiv:2411.04108, 2024

    arXiv 2024

  3. [2]

    Approximations with deep neural networks in Sobolev time-space.Anal

    Ahmed Abdeljawad and Philipp Grohs. Approximations with deep neural networks in Sobolev time-space.Anal. Appl. (Singap.), 20(3):499–541, 2022

    work page 2022

  4. [3]

    Breaking the curse of dimensionality with convex neutral networks.J

    Francis Bach. Breaking the curse of dimensionality with convex neutral networks.J. Mach. Learn. Res., 18:Paper No. 19, 53, 2017

    work page 2017

  5. [4]

    Andrew R. Barron. Universal approximation bounds for superpositions of a sigmoidal function. IEEE Trans. Inform. Theory, 39(3):930–945, 1993

    work page 1993

  6. [5]

    Analysis of Tikhonov regularization for function approximation by neural networks.Neural Networks, 16(1):79–90, 2003

    Martin Burger and Andreas Neubauer. Analysis of Tikhonov regularization for function approximation by neural networks.Neural Networks, 16(1):79–90, 2003

    work page 2003

  7. [6]

    A regularity theory for static Schrödinger equations onRd in spectral Barron spaces.SIAM J

    Ziang Chen, Jianfeng Lu, Yulong Lu, and Shengxuan Zhou. A regularity theory for static Schrödinger equations onRd in spectral Barron spaces.SIAM J. Math. Anal., 55(1):557–570, 2023

    work page 2023

  8. [7]

    Functional analysis and partial differential equations in spectral Barron spaces.arXiv:2507.06778; to appear at SIAM J

    Mourad Choulli, Shuai Lu, and Hiroshi Takase. Functional analysis and partial differential equations in spectral Barron spaces.arXiv:2507.06778; to appear at SIAM J. Math. Anal., 2025

    arXiv 2025

Show all 37 references
  1. [8]

    Cohn.Measure theory

    Donald L. Cohn.Measure theory. Birkhäuser Advanced Texts: Basler Lehrbücher. [Birkhäuser Advanced Texts: Basel Textbooks]. Birkhäuser/Springer, New York, second edition, 2013. 28 SHUAI LU AND LEQI ZHANG

    2013

  2. [9]

    Approximation by superpositions of a sigmoidal function.Math

    George Cybenko. Approximation by superpositions of a sigmoidal function.Math. Ccontrol. Signal., 2(4):303–314, 1989

    1989

  3. [10]

    Neural network approximation.Acta Numer., 30:327–444, 2021

    Ronald DeVore, Boris Hanin, and Guergana Petrova. Neural network approximation.Acta Numer., 30:327–444, 2021

    2021

  4. [11]

    The Barron space and the flow-induced function spaces for neural network models.Constr

    Weinan E, Chao Ma, and Lei Wu. The Barron space and the flow-induced function spaces for neural network models.Constr. Approx., 55(1):369–406, 2022

    2022

  5. [12]

    Engl, Martin Hanke, and Andreas Neubauer.Regularization of inverse problems

    Heinz W. Engl, Martin Hanke, and Andreas Neubauer.Regularization of inverse problems. Kluwer Academic Publishers Group, Dordrecht, 1996

    1996

  6. [13]

    Approximation spaces of deep neural networks.Constr

    Rémi Gribonval, Gitta Kutyniok, Morten Nielsen, and Felix Voigtlaender. Approximation spaces of deep neural networks.Constr. Approx., 55(1):259–367, 2022

    2022

  7. [14]

    Schwenninger, and Christoph Brune

    Tjeerd Jan Heeringa, Len Spek, Felix L. Schwenninger, and Christoph Brune. Embeddings between Barron spaces with higher-order activation functions.Appl. Comput. Harmon. Anal., 73:Paper No. 101691, 21, 2024

    2024

  8. [15]

    Approximation capabilities of multilayer feedforward networks.Neural Networks, 4(2):251–257, 1991

    Kurt Hornik. Approximation capabilities of multilayer feedforward networks.Neural Networks, 4(2):251–257, 1991

    1991

  9. [16]

    Deep neural networks with ReLU-sine-exponential activations break curse of dimensionality in approximation on Hölder class.SIAM J

    Yuling Jiao, Yanming Lai, Xiliang Lu, Fengru Wang, Jerry Zhijian Yang, and Yuanyuan Yang. Deep neural networks with ReLU-sine-exponential activations break curse of dimensionality in approximation on Hölder class.SIAM J. Math. Anal., 55(4):3635–3649, 2023

    2023

  10. [17]

    Klusowski and Andrew R

    Jason M. Klusowski and Andrew R. Barron. Approximation by combinations of ReLU and squared ReLU ridge functions withℓ1 and ℓ0 controls.IEEE Trans. Inform. Theory, 64(12):7649–7656, 2018

    2018

  11. [18]

    Deep learning.Nature, 521(7553):436–444, May 2015

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning.Nature, 521(7553):436–444, May 2015

    2015

  12. [19]

    Pereverzev

    Yuanyuan Li, Shuai Lu, Peter Mathé, and Sergei V. Pereverzev. Two-layer networks with the ReLUk activation function: Barron spaces and derivative approximation.Numer. Math., 156(1):319–344, 2024

    2024

  13. [20]

    Spectral Barron space for deep neural network approximation

    Yulei Liao and Pingbing Ming. Spectral Barron space for deep neural network approximation. SIAM J. Math. Data Sci., 7(3):1053–1076, 2025

    2025

  14. [21]

    Interpolation and inverse problems in spectral Barron spaces

    Shuai Lu and Peter Mathé. Interpolation and inverse problems in spectral Barron spaces. arXiv preprint arXiv:2502.03819, to appear SIAM J. Appl. Math., 2025

    2025

  15. [22]

    Pereverzev.Regularization theory for ill-posed problems, volume 58

    Shuai Lu and Sergei V. Pereverzev.Regularization theory for ill-posed problems, volume 58. De Gruyter, Berlin, 2013. Selected topics

    2013

  16. [23]

    A logical calculus of the ideas immanent in nervous activity (1943)

    Warren McCulloch and Walter Pitts. A logical calculus of the ideas immanent in nervous activity (1943). InIdeas that created the future—classic papers of computer science, pages 79–88. MIT Press, Cambridge, MA, [2021]©2021. Reprint of [0010388]

    1943

  17. [24]

    A new function space from Barron class and application to neural network approximation.Commun

    Yan Meng and Pingbing Ming. A new function space from Barron class and application to neural network approximation.Commun. Comput. Phys., 32(5):1361–1400, 2022

    2022

  18. [25]

    H. N. Mhaskar. Neural networks for optimal approximation of smooth and analytic functions. Neural Computation, 8(1):164–177, 1996

    1996

  19. [26]

    Rahul Parhi and Robert D. Nowak. Banach space representer theorems for neural networks and ridge splines.J. Mach. Learn. Res., 22:Paper No. 43, 40, 2021

    2021

  20. [27]

    Longman Scientific & Technical, Harlow, 1993

    Luigi Rodino.Linear Partial Differential Operators in Gevrey Spaces, volume 258 ofPitman Research Notes in Mathematics Series. Longman Scientific & Technical, Harlow, 1993

    1993

  21. [28]

    The perceptron: a probabilistic model for information storage and orga- nization (1958)

    Frank Rosenblatt. The perceptron: a probabilistic model for information storage and orga- nization (1958). InIdeas that created the future—classic papers of computer science, pages 183–190. MIT Press, Cambridge, MA, [2021]©2021. Reprinted from Psychol. Rev.65(1958), no. 6, 386–408

    1958

  22. [29]

    Siegel and Jinchao Xu

    Jonathan W. Siegel and Jinchao Xu. Approximation rates for neural networks with general activation functions.Neural Networks, 128:313–321, 2020

    2020

  23. [30]

    Siegel and Jinchao Xu

    Jonathan W. Siegel and Jinchao Xu. High-order approximation rates for shallow neural networks with cosine andReLUk activation functions.Appl. Comput. Harmon. Anal., 58:1–26, 2022

    2022

  24. [31]

    Siegel and Jinchao Xu

    Jonathan W. Siegel and Jinchao Xu. Characterization of the variation spaces corresponding to shallow neural networks.Constr. Approx., 57(3):1109–1132, 2023

    2023

  25. [33]

    Implicit neural representations with periodic activation functions

    Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In H. Larochelle, M. Ranzato, ON GENERALIZED BARRON SPACES FOR SHALLOW NEURAL NETWORKS 29 R. Hadsell, M.F. Balcan, and H...

    2020

  26. [34]

    Embedding inequalities for Barron-type spaces.J

    Lei Wu. Embedding inequalities for Barron-type spaces.J. Mach. Learn., 2(4):259–270, 2023

    2023

  27. [35]

    Finite neuron method and convergence analysis.Commun

    Jinchao Xu. Finite neuron method and convergence analysis.Commun. Comput. Phys., 28(5):1707–1745, 2020

    2020

  28. [36]

    Error bounds for approximations with deep ReLU networks.Neural Networks, 94:103–114, 2017

    Dmitry Yarotsky. Error bounds for approximations with deep ReLU networks.Neural Networks, 94:103–114, 2017

    2017

  29. [37]

    Universality of deep convolutional neural networks.Appl

    Ding-Xuan Zhou. Universality of deep convolutional neural networks.Appl. Comput. Harmon. Anal., 48(2):787–794, 2020. School of Mathematical Sciences, Fudan University, 220 Handan Road, Shanghai 200433, China. Corresponding author: S. Lu Email address:slu@fudan.edu.cn (S. Lu); ...

    2020

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.