Approximation Rates for Metaplectic Neural Networks

Ahmed Abdeljawad; Elena Cordero; Marcello Carioni

REVIEW 1 major objections 4 minor 54 references

Approximation Rates for Metaplectic Neural Networks

T0 review · 1 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Metaplectic transforms define a Barron-type space whose functions are approximated by chirped ridge networks at dimension-free $N^{-1/2}$ Sobolev rates.

desk verdict A real extension of Barron theory with a promising metaplectic dictionary, but the printed inversion lemma doesn't match the representation used in the main theorem — a repairable error a referee should catch before acceptance. read the letter →

arxiv 2608.08872 v1 pith:LZCHFCGS submitted 2026-08-09 cs.LG math.FA

Ahmed Abdeljawad , Marcello Carioni , Elena Cordero This is my paper · ORCID

classification cs.LGmath.FA MSC 41A2541A4641A3041A6546E3568T0762M4568T05

keywords approximationratesneuralnetworksBarronspacescurseofdimensionalitymetaplectictransformSobolevphysics-informedSchrödingerequation

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

This paper establishes that functions whose frequency content is controlled by a metaplectic transform—a symplectic generalization of the Fourier transform—can be approximated by shallow networks of chirped ridge neurons at the same $N^{-1/2}$ rate as classical Barron spaces, with a constant independent of the input dimension. The approximation holds in Sobolev norms on bounded domains and in weighted Sobolev norms on unbounded domains, so derivatives of the target are controlled as well as its values. The proof represents the target as a finite-mass average of dictionary atoms, shows the total variation of the representing measure is bounded by the metaplectic Barron norm, and then samples $N$ atoms to form the approximation. The paper closes with a deep network built from metaplectic atoms that reaches a lower training loss than a standard physics-informed network on harmonic-oscillator Schrödinger solutions, with the gap growing as the number of eigenmodes increases.

What carries the argument

The carrying object is the metaplectic transform, $\hat{S}f(x) = |\det B|^{-1/2}\int e^{2\pi i W_S(x,\xi)}f(\xi)\,d\xi$ with quadratic phase $W_S$, which replaces the Fourier transform in the definition of a Barron norm and yields the metaplectic Barron spaces $B^S_s$. The approximation machinery is the empirical sampling method for sparse dictionary approximation: after writing $f$ as a finite-mass mixture of weighted dictionary atoms via the variation-space embedding in Eq. (4.14), sampling $N$ atoms independently turns the mixture into an $N$-term sum with expected Sobolev error of order $N^{-1/2}$. The dictionary's chirp factor, $\cos(\pi x\cdot(B^{-1}A)x - \theta)$, adapts the atoms to the symplectic phase of the problem; when $S=J$ the factor is constant and the dictionary reduces to the classical ridge dictionary.

What would settle it

Compute the $N$-term Sobolev error for a target with explicitly known metaplectic transform, e.g. a chirped Gaussian, for increasing $N$; the theorem predicts a line of slope $-1/2$ on a log-log plot with the same slope across dimensions, so observing a shallower slope or a constant that grows with dimension would falsify the dimension-free rate. A companion check is to take an activation decaying like $1/(1+|t|)$ and verify numerically that the variation bound in Eq. (4.14) diverges, as the theory predicts.

Watch

Extended reading notes

Core claim

The central result is Theorem 19: for any free symplectic matrix $S$ with invertible block $B$, any real-valued $f$ in the metaplectic Barron space $B^S_{n+1}(\mathbb{R}^d)$, and any $N$, there is an $N$-term combination of atoms from the neural metaplectic dictionary such that $\|f - f_N\|_{W^{n,r}(\Omega)} \le C N^{-1/2}\|f\|_{B^S_{n+1}(\mathbb{R}^d)}$. The atoms are chirped ridges, $\sigma(\omega\cdot B^{-1}x+b)\cos(\pi x\cdot(B^{-1}A)x - \theta)$, and the proof represents $f$ exactly as a Bochner integral over these atoms using the inversion formula for the phase-free metaplectic transform, then applies the dictionary sampling argument in the type-2 Sobolev space $W^{n,r}(\Omega)$. The companion Theorem 20 extends the same rate to weighted Sobolev spaces on unbounded domains, and Theorem 16 gives the Sobolev-versus-Barron embedding that makes derivative control possible.

Load-bearing premise

The rate depends on being able to write $f$ as a finite-mass mixture of dictionary atoms; that requires the activation to have nonzero frequency content at some scale and to decay strictly faster than $1/(1+|t|)$, so if the activation's Fourier transform vanished identically or its tails decayed too slowly, the representing measure would have infinite total variation and the $N^{-1/2}$ bound would not follow from this argument.

Editorial extensions

If this is right

Any function in $B^S_{n+1}(\mathbb{R}^d)$ can be approximated in $W^{n,r}(\Omega)$ with $N$ chirped-ridge neurons at error $O(N^{-1/2})$, with constants independent of the dimension, so the curse of dimensionality is avoided for this symplectically structured class.
Setting $S=J$ recovers the classical Fourier Barron space and the standard ridge-function dictionary, so the classical $N^{-1/2}$ approximation result and its Sobolev versions are special cases.
The embedding theorem controls $\|f\|_{W^{n,r}(\Omega)}$ by metaplectic Barron norms of polynomially weighted combinations of $f$, which is the kind of derivative control needed for neural PDE solvers.
On unbounded domains with polynomial weight $v_{-u}$, the same $N^{-1/2}$ rate holds for $f\in B^S_{m+s}$, extending the guarantee beyond compact sets.
In the Schrödinger benchmark, the metaplectic architecture reaches a lower final training loss than a same-size standard network, and the advantage grows with the oscillator mode, reaching about one order of magnitude at $n=20$.

Reading between the lines

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

Beyond the paper, the same argument should transfer to any oscillatory integral dictionary attached to a quadratic phase with symmetric kernel, including fractional Fourier and Fresnel dictionaries used in optics, since those are metaplectic transforms for specific free symplectic matrices.
Beyond the paper, the numerical experiment measures training loss only, so it does not by itself establish better generalization; a matched test-error comparison on unseen space-time points and on conserved quantities beyond mass would separate representation benefit from optimization benefit.
Beyond the paper, because the constants contain $|\det B|$ and the weighted dilation factor $D_s(B^{-1})$, highly anisotropic symplectic matrices should erode the rate in practice; this predicts measurable performance loss as the condition number of $B$ grows, and suggests preconditioning $B$ as a design choice.
Beyond the paper, the requirement $s>1$ in the finite-variation estimate suggests a sharp boundary: activations with exactly algebraic decay $v_s^{-1}$ with $s=1$ should break the rate or turn it logarithmic, which is a concrete testable limit of the theory.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Two things you should know. The paper builds a genuine new object — metaplectic Barron spaces and a chirped-ridge dictionary — and proves a dimension-independent N^{-1/2} Sobolev approximation rate for functions in those spaces. That is real, meaningful work, likely useful for symmetry-adapted networks in PDE and signal processing. And there is a concrete error in the central proof as printed. Lemma 18's inversion formula has B^{-T} in the phase; Eq. (4.3), which Theorem 19 actually uses, has B^{-1}. For a free symplectic matrix with non-symmetric B, these are not the same, so a reader who follows Lemma 18 cannot derive (4.3). The stress-test note is right: inverting (2.18) gives B^{-1}, so the lemma is the typo and the theorem is recoverable, but it needs to be fixed. Lemma 11 has a matching sign slip (Q1 = -B^{-1}A vs p_{B^{-1}A}) that should be corrected in the same pass.

The good parts first: the definitions are natural, the Sobolev embedding results (Prop. 12, Thm. 16) are useful extensions, and the two Monte-Carlo theorems are proved in detail. The machinery is standard Maurey/variation-space technology, and several key bounds are imported from the authors' prior papers [2,3] rather than re-derived. That is not disqualifying, but it does mean the novelty is incremental — the new objects do the work, not the proof techniques. The numerical section is a reasonable proof-of-concept: the metaplectic-inspired PINN beats a matched-parameter tanh network on the harmonic-oscillator Schrödinger equation, with the gap growing with mode number. It is a single benchmark, no code, and the loss includes snapshot terms from the exact solution, so I would not lean heavily on it.

The weakest spot, beyond the typo, is transparency: the paper should spell out which prior lemmas are used where, since Theorem 16 delegates to [2, Prop. 12] and a few steps in Theorem 19 cite [3] and [4]. The variation-norm bound (4.14) checks out under s>1 and nonzero σ-hat; the reader's concern there does not land as a flaw, just a scope condition.

Who it's for: anyone working on Barron spaces, approximation of PDE solutions, or architecture design informed by symplectic structure. I'd send it to a serious referee, with a request to correct the inversion formula and its sign matches before acceptance. Fix that, and the paper deserves to stand.

Referee Report

1 major / 4 minor

Summary. The paper introduces metaplectic Barron spaces B^S_s(R^d), defined by integrability of the metaplectic transform $\hat bS f$ against the weight $(1+|\xi|)^s$ for a free symplectic matrix S, and studies approximation of their elements by finite linear combinations of chirped ridge atoms $\sigma(\omega\cdot B^{-1}x+b)\cos(\pi x\cdot(B^{-1}A)x-\theta)$. After deriving local Sobolev control of metaplectic Barron functions (Theorem 16) and uniform Sobolev bounds for a weighted dictionary (Lemma 21), the authors prove Monte-Carlo-type estimates: for real-valued $f\in B^S_{n+1}(\mathbb{R}^d)$, the $W^{n,r}(\Omega)$ error of an N-term dictionary approximation is $O(N^{-1/2})\|f\|_{B^S_{n+1}}$ (Theorem 19), with a weighted unbounded-domain analogue (Theorem 20). A physics-informed numerical experiment for the one-dimensional harmonic-oscillator Schrödinger equation compares a metaplectic-inspired network with a plain tanh network.

Significance. If the central theorem holds, the paper gives a genuine generalization of Barron's dimension-independent $N^{-1/2}$ approximation rate to a dictionary adapted to symplectic geometry, and it does so in Sobolev norms rather than merely in $L^2$. The proof strategy is explicit and standard: a representation via the inverse metaplectic transform, construction of a finite-total-variation coefficient measure, identification of the variation space, and an application of Maurey's inequality. The constants are stated explicitly and are not fitted to data, and the assertions are falsifiable. The paper also contains a careful treatment of strong measurability of the dictionary-valued maps, which is often omitted in this literature. The numerical comparison is suggestive but not decisive, because the training loss includes data-assisted terms and no code is provided.

major comments (1)

[§2.3–§4, Eqs. (2.24), (3.20), (4.3)] The inverse formula used for the central representation has a transpose error. Lemma 18 states the inverse plane wave as $e^{2\pi i x\cdot B^{-T}\xi}$, but the inverse of the phase-free metaplectic operator (2.23) has $e^{2\pi i x\cdot B^{-1}\xi}$ with the same quadratic phase factors. The error originates in (2.24): the linear term in $W_{S^{-1}}$ should be $x\cdot B^{-1}\xi$, not $(B^{-1}x)\cdot\xi$, since the (1,2) block of $S^{-1}$ is $-B^T$, not $B$. As printed, Eq. (4.3) does not follow from Lemma 18. Because (4.3) is the starting point for the measure $\lambda_f$ in (4.9) and hence for the variation-space bound (4.14), this is a load-bearing gap in the proof of the main rate (4.1). The remainder of the proof is consistent with the corrected formula (for example, the change of variables $\eta=zL\xi$ in (4.6)), so I expect a local correction to suffice, but (2.24), Lemma 18, any downstream formulas affected by (2.25), and the surrounding text must be updated before the theorem can be considered verified.

minor comments (4)

[§3, Lemma 11 and Proposition 12] The factorization is written with $Q_1=-B^{-1}A$, contradicting (2.18), which has $p_{B^{-1}A}$. The sign cancels in the $L^r$ norm estimates because $|p_{Q_1}|=1$, so the stated bounds remain true, but the proofs as printed are not self-consistent and the polynomials $P_{\alpha-\beta}$ in Proposition 12 are defined with the wrong matrix. Please correct the sign to $Q_1=B^{-1}A$ throughout.
[§5 and Abstract] The numerical experiment is data-assisted: the loss (5.4) includes $L_{\mathrm{snap}}$ and $L_{\mathrm{mass}}$, both of which use the exact solution. The abstract's phrase 'demonstrating better performance compared to classical physics informed neural networks architectures' should therefore be qualified. In addition, no code repository is provided and the choices of the fixed matrix $B$, the trainable matrices $M_\ell$, the initialization, and the hyperparameters are not fully specified, which limits reproducibility.
[Abstract] The word 'phyisics' in the abstract should be 'physics'.
[Theorem 19, Step 4] The application of Proposition 5 to the parameter-space representation requires a Borel measure on the dictionary $\widetilde D_S$, not only strong measurability of the parameter map. A short sentence explaining that the measure is the pushforward of $\mu_f$ under $(\xi,b,\theta)\mapsto \widetilde\varrho_S(\cdot,\xi,b,\theta)$ would make this step fully rigorous.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Theorems 19–20 derive the N^{-1/2} rates from the metaplectic inversion formula, an explicit coefficient measure, and Maurey's inequality; constants are analytic rather than fitted, and the numerical experiment does not feed back into the proof. The Lemma 18 / Eq. (4.3) phase mismatch is a correctness issue, not a circular reduction.

full rationale

The central derivation is self-contained. Theorem 19 starts from the definition of the metaplectic Barron space B^S_{n+1}, applies the inversion formula (Lemma 18) to write f as an absolutely convergent oscillatory integral involving bS f, represents the plane wave by translates of the activation sigma with bsigma(z) != 0, constructs the explicit complex measure lambda_f, proves the total-variation bound ||lambda_f|| <= C ||f||_{B^S_{n+1}}, invokes Proposition 5 to embed B^S into the variation space of the weighted dictionary, and then applies Maurey's estimate (Proposition 4). No step assumes the target approximation rate, no constant is fitted to data, and the numerical experiment in Section 5 is a separate validation that does not influence the constants in (4.1) or (4.15). The self-citations to [2], [3], and [4] are used for standard technical bounds (mollifier limits, type-2 Sobolev spaces) that are either reproduced in the proof or are elementary and not equivalent to the main theorem; they are therefore not load-bearing in a circular sense. The only anomaly is the note before Eq. (4.2), 'The matrix L is introduced solely to reconcile the corrected inverse plane wave with the unchanged ridge variable in D_S,' together with the mismatch between B^{-T} in Lemma 18, Eq. (3.20), and B^{-1} in Eq. (4.3). If the printed Lemma 18 phase is wrong, the theorem's representation step is not derivable as written; however, this is a proof-correctness defect, not a circularity. The approximation statement does not reduce to its inputs by construction, and no fitted parameter is renamed as a prediction. Therefore the circularity score is 0.

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

The theoretical results are derived from standard tools (Maurey, variation spaces, metaplectic analysis) with explicit hypotheses; no constants are fitted to data in the theorems. The free parameters listed are the assumptions on the weight exponent s and the Fourier evaluation point z, which affect the constants but are not estimated from data.

free parameters (2)

s (weight exponent) = > 1
Required for the b-marginal integrability of the weight omega(xi,b) in Eq. (4.8); s <= 1 would make the total variation of lambda_f infinite.
z (Fourier evaluation point) = some z != 0 with nonzero Fourier transform of sigma at z
The plane-wave representation (4.4) divides by the Fourier transform of sigma at z; the rate constant scales with the reciprocal of that value.

assumptions (7)

standard math Metaplectic representation theory and generating functions for free symplectic matrices
Used throughout Section 2 and in the inversion formula Lemma 18; referenced to de Gosson [26,27,28].
standard math Maurey's approximation theorem in type-2 Banach spaces (Proposition 4)
The N^{-1/2} rate is inherited directly from this result [42,47].
standard math Variation space characterization via representing measures (Proposition 5)
Used to bound the variation norm by the measure in Step 4 of Theorem 19 and Step 3 of Theorem 20; from [46, Lemma 3].
standard math L^p boundedness of free metaplectic operators (Theorems 13-14)
Used in Lemma 18 and Theorem 16; from Giacchi [25].
standard math Weighted Young inequality and submultiplicativity of the weights v_s
Used in Lemma 10, Lemma 11, and the total variation estimates.
domain assumption Activation condition: sigma in W^(k,infinity)(v_s) and the Fourier transform of sigma is nonzero at some point z
Required for the plane-wave representation (4.4) and the uniform atom bound (Lemma 21).
domain assumption Target functions are real-valued
Theorems 19 and 20 are stated for real-valued f; the complex case is not treated, though it would follow by splitting into real and imaginary parts.

invented entities (2)

Metaplectic Barron spaces B^S_s(R^d)
purpose: Function class of targets efficiently approximable by the metaplectic dictionary; extends classical Barron spaces.
New mathematical object defined in Definition 9; its properties are established inside the paper, so it does not have a falsifiable handle outside the paper.
Neural metaplectic dictionary D_S
purpose: Set of chirped ridge atoms used in Theorems 19 and 20 and as inspiration for the deep architecture in Section 5.
New dictionary defined in Eq. (1.3); its usefulness is shown by the theorems and the numerical experiment, which are internal to the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Approximation Rates for Metaplectic Neural Networks." pith.science (2026). https://pith.science/paper/LZCHFCGS

@misc{pith2026260808872,
  author       = {Pith},
  title        = {Pith review of: Approximation Rates for Metaplectic Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LZCHFCGS}},
  note         = {Machine review of arXiv:2608.08872}
}

read the original abstract

In this paper we develop quantitative approximation results for shallow neural networks constructed using a dictionary based on metaplectic operators. First, we extend the concept of Barron spaces by considering a symplectically motivated extension of the Fourier transform, known as the metaplectic transform. Then, after establishing embedding between metaplectic Barron spaces and Sobolev spaces we consider a neural metaplectic dictionary and we prove Monte-Carlo approximation bounds for metaplectic Barron functions using finite linear combinations of atoms of the dictionary. Finally, we validate the introduction of the neural metaplectic dictionary by devising a deep neural network architecture that uses as building blocks the atoms of the dictionary. We test it to approximate solutions of time-dependent Schr\"odinger equations, demonstrating better performance compared to classical phyisics informed neural networks architectures.

Figures

Figures reproduced from arXiv: 2608.08872 by the authors.

**Figure 1.** Terminal training objective as a function of the harmonic-oscillator mode n. For each run, the reported value is the median total loss over the final 5% of training. The curves and error bars show the median and interquartile range over five matched seeds, and the transparent points show the individual runs. 10 0 10 1 10 2 10 3 10 4 10 5 Epoch 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 20 Eigenmode Matched-seed advantage… view at source ↗

**Figure 2.** Matched-seed advantage during training. The color at mode n and epoch k is medians log10(Lplain/Lmeta). Positive values indicate a smaller total loss for the metaplectic-inspired model, while negative values indicate a smaller loss for the plain model. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_2.png] view at source ↗

**Figure 3.** Training objective for each harmonic-oscillator mode. Solid curves show the median over five seeds, and shaded regions show the 25–75% range. No temporal smoothing is applied. Acknowledgments A. Abdeljawad acknowledges the support of the Austrian Science Fund (FWF) through project PAT4788625 (Grant-DOI: 10.55776/PAT4788625). M. Carioni acknowledges the support of NWO through the Vidi grant SPARGO: Exploring and Exp… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 25 canonical work pages

[3]

Abdeljawad and T

A. Abdeljawad and T. Dittrich,Space-Time Approximation with Shallow Neural Net- works in Fourier Lebesgue Spaces, arXiv:2312.08461 [cs], Dec. 2023

arXiv 2023
[4]

Abdeljawad and T

A. Abdeljawad and T. Dittrich,Weighted Sobolev Approximation Rates for Neural Networks on Unbounded Domains, Version Number: 1, 2024.doi:10.48550/ARXIV. 2411.04108

arXiv doi:10.48550/arxiv.2411.04108 2024
[1]

Uniform approximation with quadratic neural networks,

A. Abdeljawad, “Uniform approximation with quadratic neural networks,”Neural Net- works, vol. 192, p. 107742, Dec. 2025.doi:10.1016/j.neunet.2025.107742

work page arXiv 2025
[2]

Time-frequency analysis for neuralnetworks,

A. Abdeljawadand E. Cordero, “Time-frequency analysis for neuralnetworks,”Applied and Computational Harmonic Analysis, vol. 84, p. 101883, 2026.doi:https://doi. org/10.1016/j.acha.2026.101883

work page arXiv 2026
[5]

Approximations with deep neural networks in Sobolev time-space,

A. Abdeljawad and P. Grohs, “Approximations with deep neural networks in Sobolev time-space,”Analysis and Applications, vol. 20, no. 03, pp. 499–541, May 2022.doi: 10.1142/S0219530522500014

doi:10.1142/s0219530522500014 2022
[6]

The fractional fourier transform and time-frequency representations,

L. B. Almeida, “The fractional fourier transform and time-frequency representations,” IEEE Transactions on Signal Processing, vol. 42, no. 11, pp. 3084–3091, 1994

work page 1994
[7]

Breaking the Curse of Dimensionality With Convex Neural Networks,

F. Bach, “Breaking the Curse of Dimensionality With Convex Neural Networks,”Jour- nal of Machine Learning Research, vol. 18, no. 19, pp. 1–53, 2017.doi:10.48550/ arXiv.1412.8690

arXiv doi:10.48550/arxiv.1412.8690 2017
[8]

Universal approximation bounds for superpositions of a sigmoidal func- tion,

A. Barron, “Universal approximation bounds for superpositions of a sigmoidal func- tion,”IEEE Transactions on Information Theory, vol. 39, no. 3, pp. 930–945, May 1993.doi:10.1109/18.256500

doi:10.1109/18.256500 1993

Show all 54 references

[9]

Approximation and learning by greedy algorithms,

A. R. Barron, A. Cohen, W. Dahmen, and R. A. DeVore, “Approximation and learning by greedy algorithms,”The Annals of Statistics, vol. 36, no. 1, pp. 64–94, 2008.doi: 10.1214/009053607000000631

2008 doi
[10]

V. I. Bogachev,Measure Theory. Berlin: Springer, 2007, vol. 2.doi:10.1007/978-3- 540-34514-5 37

2007 doi
[11]

Stochastic partial differential equations in M-type 2 Banach spaces,

Z. Brzeźniak, “Stochastic partial differential equations in M-type 2 Banach spaces,” Potential Analysis, vol. 4, no. 1, pp. 1–45, Feb. 1995.doi:10.1007/BF01048965

1995 doi
[12]

The discrete fractional fourier trans- form,

C. Candan, M. A. Kutay, and H. M. Ozaktas, “The discrete fractional fourier trans- form,”IEEE Transactions on Signal Processing, vol. 48, no. 5, pp. 1329–1337, 2000

2000
[13]

Universal approximation to nonlinear operators by neural networkswitharbitraryactivationfunctionsanditsapplicationtodynamicalsystems,

T. Chen and H. Chen, “Universal approximation to nonlinear operators by neural networkswitharbitraryactivationfunctionsanditsapplicationtodynamicalsystems,” IEEE transactions on neural networks, vol. 6, no. 4, pp. 911–917, 1995

1995
[14]

A Regularity Theory for Static Schrödinger Equa- tions on{R} d in Spectral Barron Spaces,

Z. Chen, J. Lu, Y. Lu, and S. Zhou, “A Regularity Theory for Static Schrödinger Equa- tions on{R} d in Spectral Barron Spaces,”SIAM Journal on Mathematical Analysis, vol. 55, no. 1, pp. 557–570, Feb. 2023.doi:10.1137/22M1478719

2023 doi
[15]

Cordero and L

E. Cordero and L. Rodino,Time-Frequency Analysis of Operators. De Gruyter, Sep. 2020.doi:10.1515/9783110532456

2020 doi
[16]

Approximation by Superpositions of a Sigmoidal Function,

G. Cybenko, “Approximation by Superpositions of a Sigmoidal Function,”Mathe- matics of Control, Signals, and Systems, vol. 2, no. 4, pp. 303–314, Dec. 1989.doi: 10.1007/BF02551274

1989 doi
[17]

On the approximation of functions by tanh neural networks,

T. De Ryck, S. Lanthaler, and S. Mishra, “On the approximation of functions by tanh neural networks,”Neural Networks, vol. 143, pp. 732–750, 2021

2021
[18]

Nonlinear Approximation,

R. A. DeVore, “Nonlinear Approximation,”Acta Numerica, vol. 7, pp. 51–150, Jan. 1998.doi:10.1017/S0962492900002816

1998 doi
[19]

A metaplectic perspective of uncertainty principles in the linear canonical transform domain,

N. C. Dias, M. de Gosson, and J. N. Prata, “A metaplectic perspective of uncertainty principles in the linear canonical transform domain,”Journal of Functional Analysis, vol. 287, p. 110494, 2024.doi:10.1016/j.jfa.2024.110494

2024
[20]

C. F. Dunkl and Y. Xu,Orthogonal Polynomials of Several Variables, 2nd ed. Cam- bridge University Press, Aug. 21, 2014.doi:10.1017/CBO9781107786134

2014 doi
[21]

The Barron Space and the Flow-Induced Function Spaces for Neural Network Models,

W. E, C. Ma, and L. Wu, “The Barron Space and the Flow-Induced Function Spaces for Neural Network Models,”Constructive Approximation, vol. 55, no. 1, pp. 369–406, Feb. 2022.doi:10.1007/s00365-021-09549-y

2022 doi
[22]

Deep Neural Network Approximation Theory,

D. Elbrachter, D. Perekrestenko, P. Grohs, and H. Bolcskei, “Deep Neural Network Approximation Theory,”IEEE Transactions on Information Theory, vol. 67, no. 5, pp. 2581–2623, May 2021.doi:10.1109/TIT.2021.3062161

2021
[23]

G. B. Folland,Harmonic Analysis in Phase Space(Annals of Mathematics Studies). Princeton University Press, 1989, vol. 122

1989
[24]

The metaplectic action on modulation spaces,

H. Führ and I. Shafkulovska, “The metaplectic action on modulation spaces,”Applied and Computational Harmonic Analysis, vol. 68, p. 101604, 2024.doi:https://doi. org/10.1016/j.acha.2023.101604

2024
[25]

Boundedness of metaplectic operators withinL p spaces, applications to pseudodifferential calculus, and time–frequency representations,

G. Giacchi, “Boundedness of metaplectic operators withinL p spaces, applications to pseudodifferential calculus, and time–frequency representations,”Journal of Fourier Analysis and Applications, vol. 30, p. 69, 2024.doi:10.1007/s00041-024-10129-4

2024 doi
[26]

M. A. de Gosson,Maslov Classes, Metaplectic Representation and Lagrangian Quan- tization(Research Notes in Mathematics). Berlin: Wiley–VCH, 1997, vol. 95. 38

1997
[27]

M. A. de Gosson,Symplectic Geometry and Quantum Mechanics(Operator Theory: Advances and Applications). Basel: Birkhäuser, 2006, vol. 166, Subseries: Advances in Partial Differential Equations

2006
[28]

M. A. de Gosson,Symplectic Methods in Harmonic Analysis and in Mathematical Physics(Pseudo-Differential Operators). Birkhäuser, 2011, vol. 7.doi:10.1007/978- 3-7643-9992-4

2011 doi
[29]

D. J. Griffiths and D. F. Schroeter,Introduction to Quantum Mechanics, 3rd ed. Cam- bridge University Press, 2018

2018
[30]

Gröchenig,Foundations of Time-Frequency Analysis(Applied and Numerical Har- monic Analysis), J

K. Gröchenig,Foundations of Time-Frequency Analysis(Applied and Numerical Har- monic Analysis), J. J. Benedetto, Ed. Boston, MA: Birkhäuser Boston, 2001.doi: 10.1007/978-1-4612-0003-1

2001 doi
[31]

Approximation capabilities of multilayer feedforward networks,

K. Hornik, “Approximation capabilities of multilayer feedforward networks,”Neural networks, vol. 4, no. 2, pp. 251–257, 1991

1991
[32]

Multilayer feedforward networks are uni- versal approximators,

K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are uni- versal approximators,”Neural networks, vol. 2, no. 5, pp. 359–366, 1989

1989
[33]

A simple lemma on greedy approximation in Hilbert space and conver- gence rates for projection pursuit regression and neural network training,

L. K. Jones, “A simple lemma on greedy approximation in Hilbert space and conver- gence rates for projection pursuit regression and neural network training,”The Annals of Statistics, vol. 20, no. 1, pp. 608–613, 1992.doi:10.1214/aos/1176348546

1992
[34]

On universal approximation and error bounds for fourier neural operators,

N. Kovachki, S. Lanthaler, and S. Mishra, “On universal approximation and error bounds for fourier neural operators,”Journal of Machine Learning Research, vol. 22, no. 290, pp. 1–76, 2021

2021
[35]

A Theoretical Analysis of Deep Neural Networks and Parametric PDEs,

G. Kutyniok, P. Petersen, M. Raslan, and R. Schneider, “A Theoretical Analysis of Deep Neural Networks and Parametric PDEs,”Constructive Approximation, vol. 55, no. 1, pp. 73–125, Feb. 2022.doi:10.1007/s00365-021-09551-4

2022 doi
[36]

Error estimates for deeponets: A deep learning framework in infinite dimensions,

S. Lanthaler, S. Mishra, and G. E. Karniadakis, “Error estimates for deeponets: A deep learning framework in infinite dimensions,”Transactions of Mathematics and its Applications, vol. 6, no. 1, tnac001, 2022

2022
[37]

The expressive power of neural networks: A view from the width,

Z. Lu, H. Pu, F. Wang, Z. Hu, and L. Wang, “The expressive power of neural networks: A view from the width,”Advances in neural information processing systems, vol. 30, 2017

2017
[38]

The fractional order fourier transform and its application to quantum me- chanics,

V. Namias, “The fractional order fourier transform and its application to quantum me- chanics,”Journal of the Institute of Mathematics and its Applications, vol. 25, pp. 241– 265, 1980

1980
[39]

Fractional fourier transforms and their optical implementation,

H. M. Ozaktas and D. Mendlovic, “Fractional fourier transforms and their optical implementation,”Journal of the Optical Society of America A,vol.10,no.12,pp.2522– 2531, 1993

1993
[40]

H. M. Ozaktas, Z. Zalevsky, and M. A. Kutay,The Fractional Fourier Transform: With Applications in Optics and Signal Processing. Wiley, 2001. 39

2001
[41]

Modulation Spaces and the Curse of Dimensionality,

R. Parhi and M. Unser, “Modulation Spaces and the Curse of Dimensionality,” in 2023 International Conference on Sampling Theory and Applications (SampTA), New Haven, CT, USA: IEEE, Jul. 10, 2023, pp. 1–5.doi:10.1109/SampTA59647.2023. 10301395

2023
[42]

Remarques sur un résultat non publié de b. maurey,

G. Pisier, “Remarques sur un résultat non publié de b. maurey,”Séminaire d’Analyse Fonctionnelle, pp. 1–12, 1981

1981
[43]

Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

M. Raissi, P. Perdikaris, and G. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,”Journal of Computational Physics, vol. 378, pp. 686– 707, Feb. 2019.doi:10.10...

2019 doi
[44]

Rudin,Real and Complex Analysis, 3rd ed

W. Rudin,Real and Complex Analysis, 3rd ed. McGraw–Hill, 1987

1987
[45]

Approximation Rates for Neural Networks With General Activation Functions,

J. W. Siegel and J. Xu, “Approximation Rates for Neural Networks With General Activation Functions,”Neural Networks, vol. 128, pp. 313–321, Aug. 2020.doi:10. 1016/j.neunet.2020.05.019

2020
[46]

Characterization of the Variation Spaces Corresponding to Shallow Neural Networks,

J. W. Siegel and J. Xu, “Characterization of the Variation Spaces Corresponding to Shallow Neural Networks,”Constructive Approximation, Feb. 2023.doi:10 . 1007 / s00365-023-09626-4

2023
[47]

Sharp Bounds on the Approximation Rates, Metric Entropy, and n-Widths of Shallow Neural Networks,

J. W. Siegel and J. Xu, “Sharp Bounds on the Approximation Rates, Metric Entropy, and n-Widths of Shallow Neural Networks,”Foundations of Computational Mathemat- ics, vol. 24, no. 2, pp. 481–537, Apr. 2024.doi:10.1007/s10208-022-09595-3

2024 doi
[48]

Tartar,An Introduction to Sobolev Spaces and Interpolation Spaces(Lecture Notes of the Unione Matematica Italiana)

L. Tartar,An Introduction to Sobolev Spaces and Interpolation Spaces(Lecture Notes of the Unione Matematica Italiana). Berlin, Heidelberg: Springer Berlin Heidelberg, 2007, vol. 3.doi:10.1007/978-3-540-71483-5

2007 doi
[49]

V. N. Temlyakov,Greedy Approximation(Cambridge Monographs on Applied and Computational Mathematics). Cambridge: Cambridge University Press, 2011, vol. 20

2011
[50]

Terdik,Multivariate Statistical Methods: Going Beyond the Linear(Frontiers in Probability and the Statistical Sciences)

G. Terdik,Multivariate Statistical Methods: Going Beyond the Linear(Frontiers in Probability and the Statistical Sciences). Cham: Springer International Publishing, 2021.doi:10.1007/978-3-030-81392-5

2021 doi
[51]

American Mathematical Soci- ety, 2014, vol

G.Teschl,Mathematical Methods in Quantum Mechanics: With Applications to Schrödinger Operators(Graduate Studies in Mathematics), 2nd ed. American Mathematical Soci- ety, 2014, vol. 157

2014
[52]

Thangavelu,Lectures on Hermite and Laguerre Expansions(Mathematical Notes)

S. Thangavelu,Lectures on Hermite and Laguerre Expansions(Mathematical Notes). Princeton University Press, 1993, vol. 42

1993
[53]

Lp Sampling Numbers for the Fourier-Analytic Barron Space,

F. Voigtlaender, “Lp Sampling Numbers for the Fourier-Analytic Barron Space,”arXiv preprint arXiv:2208.07605, 2022, Publisher: arXiv tex.version: 1

2022 arXiv
[54]

Some observations on high-dimensional partial differ- ential equations with barron data,

E. Weinan and S. Wojtowytsch, “Some observations on high-dimensional partial differ- ential equations with barron data,” inMathematical and Scientific Machine Learning, PMLR, 2022, pp. 253–269. 40

2022

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [3]

Abdeljawad and T

A. Abdeljawad and T. Dittrich,Space-Time Approximation with Shallow Neural Net- works in Fourier Lebesgue Spaces, arXiv:2312.08461 [cs], Dec. 2023

arXiv 2023

[2] [4]

Abdeljawad and T

A. Abdeljawad and T. Dittrich,Weighted Sobolev Approximation Rates for Neural Networks on Unbounded Domains, Version Number: 1, 2024.doi:10.48550/ARXIV. 2411.04108

arXiv doi:10.48550/arxiv.2411.04108 2024

[3] [1]

Uniform approximation with quadratic neural networks,

A. Abdeljawad, “Uniform approximation with quadratic neural networks,”Neural Net- works, vol. 192, p. 107742, Dec. 2025.doi:10.1016/j.neunet.2025.107742

work page arXiv 2025

[4] [2]

Time-frequency analysis for neuralnetworks,

A. Abdeljawadand E. Cordero, “Time-frequency analysis for neuralnetworks,”Applied and Computational Harmonic Analysis, vol. 84, p. 101883, 2026.doi:https://doi. org/10.1016/j.acha.2026.101883

work page arXiv 2026

[5] [5]

Approximations with deep neural networks in Sobolev time-space,

A. Abdeljawad and P. Grohs, “Approximations with deep neural networks in Sobolev time-space,”Analysis and Applications, vol. 20, no. 03, pp. 499–541, May 2022.doi: 10.1142/S0219530522500014

doi:10.1142/s0219530522500014 2022

[6] [6]

The fractional fourier transform and time-frequency representations,

L. B. Almeida, “The fractional fourier transform and time-frequency representations,” IEEE Transactions on Signal Processing, vol. 42, no. 11, pp. 3084–3091, 1994

work page 1994

[7] [7]

Breaking the Curse of Dimensionality With Convex Neural Networks,

F. Bach, “Breaking the Curse of Dimensionality With Convex Neural Networks,”Jour- nal of Machine Learning Research, vol. 18, no. 19, pp. 1–53, 2017.doi:10.48550/ arXiv.1412.8690

arXiv doi:10.48550/arxiv.1412.8690 2017

[8] [8]

Universal approximation bounds for superpositions of a sigmoidal func- tion,

A. Barron, “Universal approximation bounds for superpositions of a sigmoidal func- tion,”IEEE Transactions on Information Theory, vol. 39, no. 3, pp. 930–945, May 1993.doi:10.1109/18.256500

doi:10.1109/18.256500 1993