Pith. sign in

REVIEW 2 major objections 4 minor 2 cited by

This paper argues that the sample complexity of learning an operator from point evaluations is governed by the operator's regularity, proving that holomorphic operators admit algebraic error rates whereas Lipschitz or C^k operators cannot.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A survey of operator learning theory showing holomorphy gives fast sample-complexity rates, general smoothness gives a polylogarithmic barrier, and FNO-approximable classes cap out at n^{-1/2}.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A transparent survey of operator learning sample-complexity bounds; the FNO exponent ceiling in Thm. 5 rests on an unverified lemma transfer. the 2 major comments →

arxiv 2603.00819 v2 pith:PY5VZ5XN submitted 2026-02-28 math.NA cs.LGcs.NAmath.STstat.TH

A short tour of operator learning theory: Convergence rates, statistical limits, and open questions

Simone Brugiapaglia , Nicola Rares Franco , Nicholas H. Nelsen This is my paper
classification math.NA cs.LGcs.NAmath.STstat.TH MSC 41A4647A5862C2068T07
keywords operator learningempirical risk minimizationminimax ratesholomorphic operatorsneural networksFourier Neural Operatorsn-widthssample complexity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 survey of operator learning theory sets out to clarify how many samples are needed to approximate a nonlinear map between function spaces from point evaluations. Drawing together recent upper bounds for empirical risk minimization and lower bounds from minimax analysis, it proposes that the smoothness of the target operator—not the expressivity of the neural network—is the decisive factor. The paper establishes that holomorphic operators can be learned at error rates arbitrarily close to the Monte Carlo rate n^{-1/2} (or faster), while merely Lipschitz or C^k operators suffer a 'curse of sample complexity' with only polylogarithmic decay. Fourier Neural Operator–approximable classes sit in between, with optimal algebraic rate exponents between 1/(2+16/α) and 1/2. If these results hold, they tell practitioners where to place their bets: analytic regularity justifies optimism, while Lipschitz-only problems are fundamentally hard.

Core claim

On the paper's own terms, the central discovery is a hierarchy of minimax learnability for operator classes. For the unit ball in the Lipschitz or C^k spaces, the best achievable worst-case error over all methods using n point evaluations decays only polylogarithmically, so no algebraic rate is possible (Thm. 3). For holomorphic operators with l^p-summable parameter sequences, the minimax error is n^{-(1/p-1/2)} up to log factors, a rate that can be made arbitrarily close to n^{-1/2} as p→0 and is attained by an empirical risk minimizer based on compressed sensing (Thms. 2 and 4). For operators that are efficiently approximated by Fourier Neural Operators at rate α, the optimal algebraic min

What carries the argument

The central objects are the nonlinear sampling n-width s_n(K)_X, which records the best worst-case error achievable with n point evaluations over a class K, and the regularity classes themselves: the holomorphic classes H(b) built from Bernstein polyellipses with l^p-summability, and the FNO-approximable classes K_alpha^FNO. The upper bounds are obtained by balancing approximation error from finite-dimensional encoding/decoding with statistical error from finite samples, using empirical process theory (for ReLU networks) and compressed sensing (for handcrafted tanh networks). The lower bounds combine n-width arguments with concrete constructions of hard operator families. The key mechanism i

Load-bearing premise

The load-bearing premise is that the empirical risk minimization problem (3) has a solution and that training can find it: the theorems are existential for global minimizers, and no algorithm is proven to reach them, so the sample-complexity rates apply only to an idealized optimization.

What would settle it

Find a concrete learning method—neural or otherwise—that achieves an algebraic error rate for the unit Lipschitz ball from n point evaluations, contradicting Theorem 3; or construct a holomorphic class with l^p-summable parameters whose minimax rate is strictly slower than n^{-(1/p-1/2)}, contradicting Theorem 4.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • For holomorphic operators with l^p-summable parameter sequences and negligible encoder/decoder error, empirical risk minimization achieves minimax-optimal error rates n^{-(1/p - 1/2)} in the absence of noise, which are algebraic and can be made arbitrarily fast as p decreases (Thms. 2 and 4).
  • No learning method can achieve algebraic sample complexity for the unit Lipschitz or C^k ball from point evaluations; the best possible worst-case error decays only polylogarithmically (Thm. 3).
  • For operators that are well approximated by Fourier Neural Operators at rate α, the optimal algebraic minimax rate exponent is between 1/(2+16/α) and 1/2, so algebraic rates are possible but capped at n^{-1/2} (Thm. 5).
  • The minimax errors for the holomorphic and FNO classes are both bounded above by the minimax error for the Lipschitz class, giving a clean separation between regularity classes (paper's §3 discussion).

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A practical consequence of the separation is that estimating the analytic regularity of a solution map (e.g., from parameter-to-solution maps of PDEs) before choosing an architecture would be more valuable than architecture search; holomorphy justifies expecting fast convergence.
  • The lower bounds assume point-evaluation measurements; if linear functionals (e.g., Fourier coefficients) are allowed as measurements, the Lipschitz lower bound may not apply, and the sample complexity could change—this is not explored in the survey.
  • The theorem results are existential for global minimizers of a nonconvex problem; without proofs that gradient-based training can reach these minimizers, the theoretical rates may not be realized in practice, an open problem the paper acknowledges.
  • One could test the FNO upper bound by training FNOs on holomorphic operators: if the observed error exponent exceeds 1/2 for large α, the true minimax exponent might be higher than the survey's guess.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. This survey paper presents a unified, notationally consistent overview of recent results in operator learning theory with emphasis on sample complexity. Section 2 reviews two empirical-risk-minimization bounds for holomorphic operators: Theorem 1, adapted from Reinhardt–Wang–Zech, gives an expectation error bound with a rate approaching n^{-1/2} for deep ReLU networks; Theorem 2, adapted from Adcock–Dexter–Moraga, gives a faster-than-Monte-Carlo algebraic rate for a handcrafted tanh-network class under polynomial truncation-error decay and bounded noise. Section 3 surveys minimax results: Theorem 3 gives a polylogarithmic lower bound for C^k and Lipschitz operator classes from point evaluations; Theorem 4 gives lower bounds for holomorphic classes showing rates of order n^{-(1/p-1/2)}; Theorem 5 gives two-sided bounds for FNO-approximable classes, with optimal algebraic exponent between roughly 1/2 and 1/2; and Theorem 6 extends the story to a noisy sampling setting. The paper closes with several open problems. The authors are transparent that they are re-presenting results from cited papers in a unified notation, with a partial loss of generality.

Significance. If the results are correct as presented, the survey provides a valuable synthesis: holomorphy yields arbitrarily fast algebraic minimax rates, Lipschitz/C^k regularity yields only polylogarithmic rates, and FNO-approximability yields algebraic but no-better-than-n^{-1/2} rates. This gives a clear organizing principle, essentially 'regularity governs sample complexity.' The explicit attribution of each theorem to a cited source is a strength, as is the careful separation of upper and lower bounds and the candid discussion of open problems. The paper does not prove new theorems, but a survey does not need to; its contribution is the coherent framing. The weakest point is a single load-bearing verification in Theorem 5, discussed below; if that is repaired, the survey would be a useful reference for the operator-learning community.

major comments (2)
  1. [§3, Theorem 5 and the proof paragraph after Eq. (10)] The upper bound β*≤1/2 is load-bearing for the FNO half of the paper's central conclusion, yet the proof rests on the sentence that 'perusal' of the proof of [16, Lem. 3.22] shows that [16, Eqn. (3.30)] remains valid with K^α_FNO(K) in place of U^{α,∞}_{ℓ,NO}. This is not a formality: the class K^α_FNO(K) defined in Eq. (9) involves a specific FNO architecture, parameter norm bound exp(m), and sup_{m} m^α e_K(f,N_m^{FNO}) approximation error, while the class in [16] may measure approximation differently. The paper gives no argument that the metric-entropy lower bound or the relevant construction transfers. Please supply a proof or a precise reduction—for example, show that U^{α,∞}_{ℓ,NO} embeds into K^α_FNO(K) with explicit parameter choices, or state and prove a lemma establishing the lower bound directly for K^α_FNO(K). Without this, Eq. (10)'s upper bound is unsupported and the 'no fa
  2. [§3, Theorem 4 and the paragraph following it] The displayed theorem contains only lower bounds on s_n(H(b)∘ι), but the text states that the rate is 'optimal, up to log factors.' This optimality conclusion is not a consequence of the theorem as stated; it requires a matching upper bound, which is only indirectly available through Theorem 2 under extra hypotheses (e.g., Eq. (6), σ=0). Please state the matching upper bound explicitly, or rephrase the optimality claim to identify precisely which theorem supplies the matching upper bound and under which assumptions. This would prevent a reader from attributing to Theorem 4 an assertion it does not contain.
minor comments (4)
  1. [§2, Theorems 1–2] Both theorems are stated for global minimizers of the nonconvex ERM objective (3). Existence is asserted or guaranteed with high probability, but no algorithm is claimed to find such minimizers. The paper acknowledges this only in the §4 open-problem discussion. Since the abstract promises convergence rates for empirical risk minimization, add a remark in §2 clarifying the oracle nature of these results and referring the reader to the open problems.
  2. [§3, Eq. (9)] The notation e_K(f,𝒩) in the definition of Γ^α_FNO is ambiguous because 𝒩 is not defined at that point; it should be e_K(f,𝒩_m^{FNO}). Also, the rendering of the map ι as U→R^N should be U→ℝ^ℕ, and the inclusion H_{0,1}=[−1,1]^ℕ⊆R(b) is meant in ℓ^∞, not ℓ^2; clarify the ambient space.
  3. [§3, Theorem 6, Eq. (12)] The displayed chain contains three inequalities, but the following text refers only to 'the first' and 'the second' of them. Label the inequalities (e.g., (12a), (12b), (12c)) to make the explanation unambiguous.
  4. [§2, Theorem 1, Eq. (4)] The additive τ in the right-hand side means that for fixed τ>0 the displayed bound does not tend to zero as n→∞. The text's phrase 'approximate Monte Carlo rate' is intended to address this, but the dependence of c on τ is not stated. Please write c_τ or add a sentence clarifying that c depends on τ and other fixed parameters, and that the τ term is to be sent to zero in the limiting-rate statement.

Circularity Check

0 steps flagged

No significant circularity: the load-bearing theorems are imported from external groups and are not equivalent to the survey's inputs.

full rationale

The paper is a survey. Its main theorems are quoted from external groups: Thm. 1 from Reinhardt-Wang-Zech [35], Thm. 2 from Adcock-Dexter-Moraga [4], Thms. 3 and 5 from Kovachki-Lanthaler-Mhaskar [20] and Grohs-Lanthaler-Trautner [16], and Thm. 6 from Adcock-Maier-Parhi [6]. The survey adds no derivation in which an output quantity is defined in terms of, or fitted to, the quantity it later predicts. The ERM upper bounds and minimax lower bounds are independent statements; matching them in Thm. 4 and the Discussion is a synthesis, not a construction. The class K^alpha_FNO is defined by FNO approximation rates, not by sample complexity, so the bound beta* <= 1/2 is a genuine transfer from metric-entropy estimates in [16], not a tautology. The authors' self-citations ([13], [15], [24], [30], [31]) appear in background, context, and one illustrative example; none supplies a central conclusion. In particular, [13] is cited only as an explicit measure example satisfying hypotheses already provided by [16, Prop. 3.5], so it is not load-bearing. The only load-bearing concern is the unshown transfer in Theorem 5: the paper says 'perusal of the proof of [16, Lem. 3.22] show that [16, Eqn. (3.30)] remains valid with K^alpha_FNO in place of U^{alpha,infty}_{ell,NO}', which is a missing verification at a joint, not a circular reduction. It should be weighed as correctness risk, not circularity. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' own prior work, and no ansatz is smuggled in via self-citation. The derivation chain is not equivalent to its inputs, so the circularity burden is low.

Axiom & Free-Parameter Ledger

2 free parameters · 7 axioms · 0 invented entities

This paper is a survey, so the central claims are not derived here; they inherit assumptions from the cited sources. No data-fitted parameters appear; the only hand-chosen slack parameters are τ and δ in Thms. 1–2, used to absorb log factors and width constants. The main structural assumptions are holomorphy/l^p summability, spectral decay of the input measure, specific noise models, global attainability of the ERM minimizer, and the point-evaluation-only minimax information model.

free parameters (2)
  • tau in Theorem 1 (slack exponent) = arbitrarily small > 0 (not estimated from data)
    Introduced to suppress log factors in the ERM bound (4); the paper calls it 'merely an algebraic artifact.' It is hand-chosen and appears in the stated convergence-rate exponent.
  • delta in Theorem 2 (width slack) = arbitrarily small > 0 (not estimated from data)
    Appears in the network-width bound width <= c0 * ntilde^(1+delta); hand-chosen to close the compressed-sensing network construction, and does not affect the final rate exponent.
axioms (7)
  • domain assumption Assumption 1: existence of boundedly invertible linear maps E_infty: U -> l2 and D_infty: l2 -> V, with truncated encoder-decoder architecture D_q ∘ g ∘ E_d.
    Underlies both ERM analyses in §2 and restricts applicability to Riesz-basis-type representations; the paper notes its presentation is less general than [35]'s frame setting.
  • domain assumption Holomorphy: G admits a holomorphic extension over O containing E_infty^{-1}(H_{r,R}) (Thm. 1), or G∘E_infty^{-1} ∈ H(b) with b ∈ l^p_M on Bernstein polyellipses (Thm. 2).
    The fast algebraic rates in Thms. 1–2 and 4 are conditional on holomorphy and l^p summability; this is the central regularity assumption inherited from the cited works.
  • domain assumption Spectral decay: supp(ρ) ⊆ E_infty^{-1}(H_{r,R}) and sup norm condition in Thm. 1; truncation errors decay as s^{-γ} and s^{-ν+1/2} in (6a)–(6b); covariance eigenvalues satisfy inf_j j^{2ϑ} λ_j > 0 or λ_j ≍ j^{-2ϑ} in Thms. 3 and 6.
    All sample-complexity rates are governed by these decay/regularity parameters; if spectra decay slowly, the bounds degrade (e.g., polylog lower bounds).
  • domain assumption Measure conditions: (E_infty)_#ρ is quasi-uniform on H_{0,1} and (R_s∘E_infty)_#ρ is absolutely continuous w.r.t. uniform measure (Thm. 2); ρ is Gaussian in Thms. 3 and 6; ρ is supported on a compact convex infinite-dimensional K in Thm. 5.
    The specific minimax and ERM results are proven under these measure assumptions; they are not universal over all input distributions.
  • domain assumption Noise models: i.i.d. subgaussian noise (Thm. 1), bounded noise (Thm. 2), Gaussian noise (Thm. 6).
    Rates in §2 depend on the noise model: Thm. 1 is in expectation, Thm. 2 is high-probability with a bounded-noise threshold σ ≤ c ntilde^{-(1/p−1/2)}.
  • ad hoc to paper Existence and attainability of a global minimizer of the ERM objective (3).
    Thms. 1–2 state bounds for minimizers of a nonconvex ERM problem; no algorithm is shown to find such a minimizer, and Thm. 2's network class has handcrafted early-layer weights. This is the paper's main practical-existence caveat, acknowledged in §4.
  • domain assumption Minimax information model: the learner is restricted to n point evaluations of the target operator with arbitrary decoding, as encoded in Map_n and the nonlinear sampling n-width (8).
    Defines the minimax setting of §3; the hardness results do not automatically apply if other information types (e.g., derivative evaluations or general linear measurements) are allowed.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A short tour of operator learning theory: Convergence rates, statistical limits, and open questions." pith.science (2026). https://pith.science/paper/PY5VZ5XN

@misc{pith2026260300819,
  author       = {Pith},
  title        = {Pith review of: A short tour of operator learning theory: Convergence rates, statistical limits, and open questions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PY5VZ5XN}},
  note         = {Machine review of arXiv:2603.00819}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper surveys recent developments at the intersection of operator learning, statistical learning theory, and approximation theory. First, it reviews error bounds for empirical risk minimization with a focus on holomorphic operators and neural network approximations. Next, it illustrates fundamental performance limits in terms of sample size by adopting a minimax perspective and considering various notions of regularity beyond holomorphy. The paper ends with a discussion on the interplay between these two perspectives and related open questions.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. One Operator for Many Densities: Amortized Approximation of Conditioning by Neural Operators

    stat.ML 2026-05 unverdicted novelty 7.0

    A single neural operator can approximate the map from arbitrary joint densities to their conditionals, backed by new continuity results and illustrated on Gaussian mixtures.

  2. One Operator for Many Densities: Amortized Approximation of Conditioning by Neural Operators

    stat.ML 2026-05 unverdicted novelty 6.0

    A single neural operator can approximate the map from joint densities to conditional densities to arbitrary accuracy, with a proof based on continuity of the conditioning operator and a demonstration on Gaussian mixtures.

Reference graph

Works this paper leans on

40 extracted references · 6 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Adcock, S

    B. Adcock, S. Brugiapaglia, N. Dexter, and S. Moraga. Learning smooth functions in high dimensions: From sparse polynomials to deep neural networks. In S. Mishra and A. Townsend, editors,Handb. Numer. Anal., volume 25, pages 1–52. Elsevier, 2024

    2024

  2. [2]

    Adcock, S

    B. Adcock, S. Brugiapaglia, N. Dexter, and S. Moraga. Near-optimal learning of Banach-valued, high-dimensional functions via deep neural networks.Neural Netw., 181, 2025

    2025

  3. [3]

    Adcock, S

    B. Adcock, S. Brugiapaglia, and C. G. Webster.Sparse Polynomial Approximation of High- Dimensional Functions, volume 25. SIAM, 2022

    2022

  4. [4]

    Adcock, N

    B. Adcock, N. Dexter, and S. Moraga. Optimal deep learning of holomorphic operators between Banach spaces. InAdv. Neural Inf. Process. Syst., volume 37, pages 27725–27789, 2024

    2024

  5. [5]

    Adcock, M

    B. Adcock, M. Griebel, and G. Maier. The sample complexity of learning Lipschitz operators with respect to Gaussian measures.preprint arXiv:2410.23440, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Adcock, G

    B. Adcock, G. Maier, and R. Parhi. Towards sharp minimax risk bounds for operator learning. preprint arXiv:2512.17805, 2025

    arXiv 2025

  7. [7]

    Bhattacharya, B

    K. Bhattacharya, B. Hosseini, N. B. Kovachki, and A. M. Stuart. Model reduction and neural networks for parametric PDEs.The SMAI J. Comput. Math., 7:121–157, 2021

    2021

  8. [8]

    Boull ´e and A

    N. Boull ´e and A. Townsend. A mathematical guide to operator learning. In S. Mishra and A. Townsend, editors,Handb. Numer. Anal., volume 25, pages 83–125. Elsevier, 2024

    2024

  9. [9]

    Chen and H

    T. Chen and H. Chen. Universal approximation to nonlinear operators by neural networks with arbitrary activation functions and its application to dynamical systems.IEEE Trans. Neural Netw., 6(4):911–917, 1995

    1995

  10. [10]

    Cohen and R

    A. Cohen and R. DeVore. Approximation of high-dimensional parametric PDEs.Acta Numer., 24:1–159, 2015

    2015

  11. [11]

    Cohen, C

    A. Cohen, C. Schwab, and J. Zech. Shape holomorphy of the stationary Navier–Stokes equations. SIAM J. Math. Anal., 50(2):1720–1752, 2018

    2018

  12. [12]

    G. Cybenko. Approximation by superpositions of a sigmoidal function.Math. Control. Signals Syst., 2(4):303–314, 1989

    1989

  13. [13]

    M. V. de Hoop, N. B. Kovachki, M. Lassas, and N. H. Nelsen. Extension and neural operator approximation of the electrical impedance tomography inverse map.preprint arXiv:2511.20361, 2025

    arXiv 2025

  14. [14]

    DeVore, R

    R. DeVore, R. D. Nowak, R. Parhi, G. Petrova, and J. W. Siegel. Optimal recovery meets minimax estimation.preprint arXiv:2502.17671, 2025. 12 Simone Brugiapaglia, Nicola Rares Franco, Nicholas H. Nelsen

    Pith/arXiv arXiv 2025

  15. [15]

    N. R. Franco and S. Brugiapaglia. A practical existence theorem for reduced order models based on convolutional autoencoders.Found. Data Sci., 7(1):72–98, 2025

    2025

  16. [16]

    Grohs, S

    P. Grohs, S. Lanthaler, and M. Trautner. Theory-to-practice gap for neural networks and neural operators.preprint arXiv:2503.18219, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    G¨ uhring and M

    I. G¨ uhring and M. Raslan. Approximation rates for neural networks with encodable weights in smoothness spaces.Neural Netw., 134:107–130, 2021

    2021

  18. [18]

    Gy ¨orfi, M

    L. Gy ¨orfi, M. Kohler, A. Krzy˙ zak, and H. Walk.A distribution-free theory of nonparametric regression. Springer New York, 2002

    2002

  19. [19]

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

    1991

  20. [20]

    N. B. Kovachki, S. Lanthaler, and H. Mhaskar. Data complexity estimates for operator learning. preprint arXiv:2405.15992, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    N. B. Kovachki, S. Lanthaler, and A. M. Stuart. Operator learning: Algorithms and analysis. In S. Mishra and A. Townsend, editors,Handb. Numer. Anal., volume 25, pages 419–467. Elsevier, 2024

    2024

  22. [22]

    Krieg and M

    D. Krieg and M. Ullrich. Approximation of functions: Optimal sampling and complexity.preprint arXiv:2602.02066, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    Lanthaler, S

    S. Lanthaler, S. Mishra, and G. E. Karniadakis. Error estimates for DeepONets: A deep learning framework in infinite dimensions.Trans. Math. Appl., 6(1), 2022

    2022

  24. [24]

    Lanthaler and N

    S. Lanthaler and N. H. Nelsen. Error bounds for learning with vector-valued random features. In Adv. Neural Inf. Process. Syst., volume 36, pages 71834–71861, 2023

    2023

  25. [25]

    Z. Li, N. B. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. M. Stuart, and A. Anand- kumar. Fourier neural operator for parametric partial differential equations.Int. Conf. Learn. Represent., 2021

    2021

  26. [26]

    H. Liu, H. Yang, M. Chen, T. Zhao, and W. Liao. Deep nonparametric estimation of operators between infinite dimensional spaces.J. Mach. Learn. Res., 25(24):1–67, 2024

    2024

  27. [27]

    L. Lu, P. Jin, G. Pang, Z. Zhang, and G. E. Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators.Nature Mach. Intell., 3(3):218–229, 2021

    2021

  28. [28]

    Marcati, J

    C. Marcati, J. A. Opschoor, P. C. Petersen, and C. Schwab. Exponential ReLU neural network approximation rates for point and edge singularities.Found. Comput. Math., 23(3):1043–1127, 2023

    2023

  29. [29]

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

    1996

  30. [30]

    N. H. Nelsen and A. M. Stuart. Operator learning using random features: A tool for scientific computing.SIAM Rev., 66(3):535–571, 2024

    2024

  31. [31]

    N. H. Nelsen and Y. Yang. Operator learning meets inverse problems: A probabilistic perspective. preprint arXiv:2508.20207, 2025

    arXiv 2025

  32. [32]

    Novak and H

    E. Novak and H. Wo´ zniakowski.Tractability of Multivariate Problems. Volume I: Linear Infor- mation, volume 6 ofEMS Tracts Math.EMS, Z¨ urich, Switzerland, 2008

    2008

  33. [33]

    Parhi and B

    R. Parhi and B. Adcock. Upper bounds on averaged sampling numbers for general model classes. In2025 Int. Conf. Sampl. Theory Appl., pages 1–5. IEEE, 2025

    2025

  34. [34]

    Petersen and F

    P. Petersen and F. Voigtlaender. Optimal approximation of piecewise smooth functions using deep ReLU neural networks.Neural Netw., 108:296–330, 2018

    2018

  35. [35]

    Reinhardt, S

    N. Reinhardt, S. Wang, and J. Zech. Statistical learning theory for neural operators.preprint arXiv:2412.17582, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Schwab and J

    C. Schwab and J. Zech. Deep learning in high dimension: Neural network expression rates for analytic functions in𝑙 2(R𝑑,𝛾 𝑑).SIAM/ASA J. Uncertain. Quantif., 11(1):199–234, 2023

    2023

  37. [37]

    J. W. Siegel. Optimal approximation rates for deep ReLU neural networks on Sobolev and Besov spaces.J. Mach. Learn. Res., 24(357):1–52, 2023

    2023

  38. [38]

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

    2024

  39. [39]

    A. v. d. Vaart and J. A. Wellner. Empirical processes. InWeak Convergence and Empirical Processes: With Applications to Statistics, pages 127–384. Springer, 2023

    2023

  40. [40]

    Yarotsky

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

    2017

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.