Pith. sign in

REVIEW 3 major objections 4 minor 33 references

Discretization and Statistical Consistency of Functional Flow Matching

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper proves that finite-rank reconstructions of random functions give conditional velocity targets that converge strongly in L2 to the infinite-dimensional flow-matching target, even when sensor sigma-algebras are not nested.

desk verdict Nonnested sensor-limit convergence for functional flow matching is proved cleanly; the end-to-end rate leans on an explicitly verified stability assumption, and the paper deserves refereeing. read the letter →

arxiv 2608.04531 v1 pith:LSOEHLNW submitted 2026-08-05 cs.LG math.PRstat.ML

Lennon J. Shikhman This is my paper
classification cs.LGmath.PRstat.ML
keywords functionalflowmatchingneuraloperatorsHilbert-spaceprobabilityflowsfinitesensorsdiscretizationconsistencyconditionalexpectationWassersteindistance
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

Functional flow matching is meant to learn a time-dependent velocity field that transports a distribution of functions to another distribution, but every implementation sees only finitely many coefficients or point values and solves a finite-dimensional ODE. The open problem was whether the finite conditional velocity targets converge to the infinite-dimensional target as sensors are refined, especially when scattered or adaptive meshes make the observation sigma-algebras nonnested, so martingale arguments do not apply. The paper closes this gap: for every uniformly bounded sequence of finite-rank reconstructions that converges strongly to the identity, the finite conditional targets converge in L2 to the continuum target, with quantitative bounds for orthogonal projections and a point-sensor version through a regularity space. It then propagates this convergence through learning, giving an end-to-end Wasserstein bound whose error terms are separated and individually certified. A reader should care because without such a sensor limit, 'resolution invariance' of functional flow matching is only an assumption; with it, arbitrary mesh refinement during training is theoretically sound.

What carries the argument

The carrying identity is $v^*_m(\tau,A_m X_\tau)=A_m E[U_\tau\mid G_m]=A_m M_m$, where $M_m$ is the conditional expectation of the continuum target onto the finite observation $\sigma$-algebra; because bounded operators commute with Bochner conditional expectations, the sensor limit reduces to proving $M_m\to V$ in $L^2$. That convergence comes from strong convergence of $A_m$ plus density of bounded continuous functions, with no nesting required. For the end-to-end result, the working object is the superposition measure of the finite reconstructed continuity equation, which exists without ODE uniqueness; a comparison of the learned flow against a path drawn from that measure gives Wasserstein control with an exponential factor governed only by the learned field's Lipschitz constant. The operator-realization certificate runs on a layerwise recurrence for the normalized quadrature neural operator, combining a magnitude bound with a spatial regularity and cubature estimate that makes the discretization error depend on input regularity rather than an unverified quadrature consistency.

What would settle it

Compute the finite conditional targets for the paper's noncommuting Gaussian model at increasing $m$ and verify that the integrated squared error decays at the predicted $\Theta(m^{1-2\alpha})$ rate; a persistent nonvanishing plateau would contradict the main consistency theorem. More directly, any strongly consistent uniformly bounded finite-rank reconstruction and square-integrable interpolation for which the $L^2$ difference of finite and continuum conditional targets does not tend to zero would refute the paper's central claim.

Watch

Extended reading notes

Core claim

In the paper's own terms: for an absolutely continuous stochastic interpolation $X_t$ with velocity $U_t$, the continuum target is $v^*(t,X_t)=E[U_t\mid X_t]$ and the level-$m$ target is $v^*_m(t,A_m X_t)=E[A_m U_t\mid A_m X_t]$, with $A_m$ a finite-rank reconstruction. The paper proves that if $A_m$ is uniformly bounded and converges strongly to the identity, then $v^*_m(\tau,A_m X_\tau)\to v^*(\tau,X_\tau)$ in $L^2(dt\otimes P;H)$, with no requirement that the observation $\sigma$-algebras be nested. The proof uses the commutation of bounded linear operators with conditional expectation and a bounded-continuous approximation argument. Quantitative versions give tail bounds for orthogonal projections, and a regularity-space variant extends the result to literal point sensors. For learned fields, the paper couples the learned ODE to a population superposition path and obtains Wasserstein convergence of generated laws from excess risk plus a learned-field Lipschitz bound, without assuming uniqueness of the population finite-dimensional ODE. A noncommuting Gaussian example shows that projected restriction of the continuum field is not the same as conditioning (boundary multiplier $0$ versus $0.72$), yet the global $L^2$ error still vanishes, and an exactly realizable clipped Gaussian scaling model yields an explicit $O((\log n)/\sqrt{n})$ end-to-end rate.

Load-bearing premise

The learned velocity field must have a Lipschitz constant whose integral over time stays bounded independent of the sensor level $m$; this is a separate hypothesis about optimization and architecture, not a consequence of the excess-risk bound.

Editorial extensions

If this is right

  • Functional flow matching can be trained on scattered, remeshed, or adaptively refined grids: as long as the reconstruction operators are uniformly bounded and strongly consistent, the finite velocity targets approach the continuum target even though the observation sigma-algebras are not nested.
  • The end-to-end Wasserstein error separates into endpoint truncation, sensor and reconstruction error, operator realization error, function-class approximation error, excess risk, solver error, and learned-field stability; each term has its own certificate and can be shrunk independently.
  • A normalized quadrature neural operator with ReLU, GeLU, or SiLU activations admits sensor-independent output and parameter constants, so refining the mesh does not inflate the statistical estimation bound.
  • In the exactly realizable clipped Gaussian scaling model, the generated law converges to the Gaussian endpoint at rate $O((\log n)/\sqrt{n})$ with high probability, including clipping, sensing, optimization, solver, and statistical contributions.
  • Projecting a continuum-trained field is not equivalent to conditioning on finite observations; the paper's Gaussian example shows the boundary multiplier can be $0$ under projection and $0.72$ under exact conditioning, while global $L^2$ convergence still holds because the affected direction lies in the trace-class tail.

Reading between the lines

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

  • The paper does not develop an adaptive training algorithm, but its nonnested consistency theorem implies that changing the sensor set between training iterations has a well-defined continuum limit; a practical next step is an adaptive refinement schedule that shrinks reconstruction error while training.
  • The paper's Gaussian check only tests the conditioning mechanism; a learned-operator benchmark that varies the quadrature rule and measures the realization term $d_m$ would test whether the certified constants are tight in practice.
  • The learned-field Lipschitz assumption could be checked empirically by tracking the integrated Lipschitz constant of the trained field as $m$ grows; the paper leaves this as a separate hypothesis about optimization and architecture.
  • The point-sensor extension suggests that fill-distance rates from scattered-data reconstruction translate directly into target-consistency rates; plugging standard reconstruction bounds into the quantitative target bounds would give explicit $L^2$ convergence rates for specific sensor layouts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper studies the discretization limit of functional flow matching. For an infinite-dimensional stochastic interpolation X_t with velocity U_t, the population target is v*(t,X_t)=E[U_t|X_t], while implementations use finite-rank observations A_m X_t and targets v*_m(τ,A_m X_τ)=E[A_m U_τ|σ(τ,A_m X_τ)]. The main theorem (Theorem 7) proves strong L^2 convergence of v*_m to v* under only strong consistency and uniform boundedness of A_m, without requiring the observation σ-algebras to be nested. The paper supplies quantitative orthogonal-projection bounds, a point-sensor extension through a regularity space, a population flow convergence theorem under a mesh-uniform stability assumption, and an end-to-end Wasserstein bound for learned fields that avoids uniqueness of the population finite-dimensional ODE. It also verifies sensor-independent constants for a normalized quadrature neural operator, analyzes a noncommuting trace-class Gaussian example, and closes the paper with an exactly realizable clipped-Gaussian specialization yielding an explicit O((log n)/√n) rate.

Significance. The central sensor-limit result is the key novel contribution: it removes the long-standing gap between infinite-dimensional flow matching objectives and finite-sensor implementations under scattered or adaptive refinement, where the conditioning σ-algebras are not nested and martingale arguments do not apply. The proof of Theorem 7 is explicit and self-contained, and it uses no fitted constants or circular reasoning. The paper is remarkably honest about its assumptions: Theorem 13 shows that mesh-uniform population ODE stability is not automatic, Remark 12 warns that Lipschitz regularity of v* does not transfer to v*_m, Remark 24 separates learned-field stability from excess risk, and Section 10 lists the remaining limitations. Theorem 23 gives a clean error decomposition into sensing, approximation, statistical, numerical, and endpoint terms, which is a substantial conceptual advance. The paper also ships a concrete certificate for a quadrature neural operator (Theorem 17), a noncommuting Gaussian example with the sharp boundary multiplier 0 versus 0.72 (Section 9.2, Table 2), and an exactly realizable clipped-Gaussian scaling model with an explicit rate (Corollary 33).

major comments (3)
  1. [5.1, Theorem 7] The proof relies on the density of bounded continuous H-valued functions in L^2(Law(ξ);H), but the justification is compressed into a parenthetical. Since this is a load-bearing approximation step and H is infinite-dimensional, please provide a short proof or a precise reference (for example, the standard measure-theoretic approximation of L^p by bounded continuous functions on a Polish space, used together with the vector-valued nature of the integrand). The step is correct, but as written it is too terse for the main theorem of the paper.
  2. [8, Assumption 22 and Theorem 26] The end-to-end Wasserstein bound is conditional on a learned-field Lipschitz envelope that is not delivered by the excess-risk theory. The manuscript is explicit about this in Remark 24 and verifies the condition for its quadrature architecture through Theorem 17, so this is not a mathematical error. However, the abstract and the statement of Theorem 26 should state more visibly that the general bound applies to realized fields satisfying Assumption 22, not to arbitrary learned fields. This would prevent a reader from over-generalizing the end-to-end claim.
  3. [9.3, Corollary 33] The explicit rate in Corollary 33 is derived in the proof under the additional certificate ε_sol,m(h) ≤ C_sol r h^β with C_sol independent of m and r, and then h_n = n^{-1/(2β)} is chosen. This uniformity in m is not listed among the hypotheses of Corollary 33, even though m_n grows with n and a dimension-dependent solver constant could change the rate. Please add this uniformity condition to the statement of the corollary, or argue that it follows from the preceding assumptions.
minor comments (4)
  1. [3.3] The symbol R for the reconstruction map clashes with the radius R used later in Assumption 16 and in the statistical rates. Consider renaming the reconstruction map (e.g., Rec_m or R_m) to avoid notational confusion.
  2. [5.3] In the finite-sensor obstruction argument, the sentence about the same bound holding over the L^2-unit ball in C_c^∞(D) would benefit from one sentence explaining the bump construction: choose a smooth function supported away from every sensor point with L^2 norm 1. This makes the independence of the reconstruction scheme immediate.
  3. [9.2, Table 2] Table 2 would be more informative with a 95% confidence interval or standard error in addition to the sample standard deviation; the RMSE column already provides useful information, but the additional column would quantify the Monte Carlo error more directly.
  4. [7.2, Remark 19] The notation O-tilde(n^{-1}) is used without a definition. Please define the tilde-O convention or replace it with an explicit O((log n)/n) statement to avoid ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 7 and the end-to-end bounds are proved from stated assumptions with external, independent mathematical ingredients.

full rationale

The paper's central claim, Theorem 7, is derived from standard measure-theoretic ingredients: strong convergence of uniformly bounded finite-rank operators, density of bounded continuous H-valued functions in L2 over the Polish space [0,T]×H, and the commutation of bounded linear maps with Bochner conditional expectation. The proof does not assume the conclusion, and the non-nesting claim is genuinely used: the argument goes through the almost-sure convergence of ξm to ξ and bounded convergence rather than martingale nesting. Corollary 10 likewise uses the external Lusin–Souslin theorem and a V-valued approximation argument, not the target statement itself. The learning bounds in Theorem 18 are standard Bernstein/net arguments with envelope constants B and G that are assumed rather than fitted to the data; Assumption 16 explicitly calls B and G 'the common output–target magnitude envelope and parameter sensitivity', and these constants are subsequently verified for the quadrature neural operator in Theorem 17 rather than being calibrated to match the conclusion. Assumption 22 (stable learned field) is exactly what the paper says it is: a separate stability hypothesis about the realized learned field. The paper is transparent that the general theorem requires it, and Theorem 13 and Remark 12 explicitly delimit why it is not derived from continuum regularity. For the paper's concrete architecture, Theorem 17 supplies the uniform Lipschitz and envelope estimates, so Assumption 22 is not merely imported; for general fields it is honestly stated as an assumption. The clipped-Gaussian specialization is described as 'an exactly realizable clipped Gaussian scaling specialization' and Remark 34 explicitly warns that 'It is not an independent-endpoint result', so the resulting rate is a composition certificate rather than a hidden fit presented as a generic prediction. The Gaussian examples in Section 9 are numerical/analytical stress tests used to illustrate the distinction between projection and conditioning; Table 2 is explicitly a check of the conditioning mechanism and not a learned-operator benchmark, and the paper says so. Citations to superposition principles and scattered-data approximation are external standard results (Stepanov and Trevisan 2017, Wendland 2005, Narcowich et al. 2005), and no load-bearing claim is justified solely by a self-citation by the author of this paper.

Assumptions & free parameters 0 free parameters · 8 assumptions · 0 invented entities

No numbers are fitted to data in the central results; all constants such as B, G, R, L, Lambda, and c_mass/C_mass enter as assumption bounds, not as posterior fits. The illustrative Gaussian parameters and clipping radius schedules are examples, not free parameters of the theorems. No new entities are postulated beyond the existing mathematical objects.

assumptions (8)
  • standard math Factorizability: [0,T] x H is a standard Borel space, so Bochner conditional expectations admit Borel field representatives.
    Invoked in Section 3.1 to define v* via the factorization lemma; standard for Polish spaces.
  • domain assumption Assumption 1: X has almost surely absolutely continuous paths with Bochner derivative U and finite second moments for X0 and U in L2(dt tensor P; H).
    Underlies all flow identities and the L2 conditional-target framework; stated in Section 3.1.
  • domain assumption Reconstruction condition: operators A_m are finite rank, uniformly bounded, and converge strongly to I; for point sensors, V embeds in H and the same-information condition sigma(tau, S_m X_tau) = sigma(tau, A_m X_tau) holds.
    The exact hypothesis of Theorem 7 and Corollary 10; point evaluation is not continuous on L2 without a regularity space.
  • domain assumption Mesh-uniform stability: either Assumption 11 for population finite targets or Assumption 22 for the learned field has an integrable Lipschitz constant uniformly in m.
    Needed for the strong flow convergence theorem and for the end-to-end Wasserstein bounds. The paper's Theorem 13 shows it is not implied by continuum regularity.
  • domain assumption Statistical envelope Assumption 16: outputs, targets, and parameter differences are bounded by B, G, R on a single probability-one event.
    Used in Theorem 18 for the Bernstein oracle bound; verified for the normalized quadrature operator in Theorem 17.
  • domain assumption Solver coupling Assumption 25: the numerical endpoint is coupled to the exact learned endpoint within epsilon_sol,m(h).
    Entered explicitly in Theorem 26; the paper states the guarantee as an assumption rather than hiding mesh-dependent constants.
  • standard math Superposition theorems of Stepanov-Trevisan and Zhang-Scott giving a Borel measure on characteristic curves.
    Used in Corollary 3 and Proposition 6 to obtain measure-valued flows without uniqueness of the population ODE.
  • standard math Gaussian concentration inequality for Hilbert-space Gaussian measures.
    Used in Proposition 31 to bound the clipping cost in the end-to-end rate; standard reference is Bogachev.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Discretization and Statistical Consistency of Functional Flow Matching." pith.science (2026). https://pith.science/paper/LSOEHLNW

@misc{pith2026260804531,
  author       = {Pith},
  title        = {Pith review of: Discretization and Statistical Consistency of Functional Flow Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LSOEHLNW}},
  note         = {Machine review of arXiv:2608.04531}
}
abstract

Functional flow matching is posed on distributions of functions but implemented from finitely many coefficients or point values. Under scattered or adaptive refinement, the resulting conditioning sigma-algebras need not be nested, so martingale convergence does not justify the sensor limit. We prove strong $L^2$ convergence of finite conditional velocity targets for every strongly consistent sequence of finite-rank reconstructions, with quantitative bounds for orthogonal projections and a point-sensor extension through a regularity space. For learned flows, coupling directly to a population superposition path yields an end-to-end Wasserstein bound without assuming uniqueness of the population finite-dimensional ODE. We verify sensor-independent constants for a normalized quadrature neural operator, including globally Lipschitz activations through an explicit magnitude recurrence. A noncommuting trace-class Gaussian example gives boundary multiplier $0$ under projected restriction and $0.72$ under exact conditioning. A spatial regularity--cubature certificate closes the operator-realization term, a Bernstein argument gives a $\widetilde{O}(n^{-1})$ excess-risk term for fixed model dimension and envelopes, and an exactly realizable clipped Gaussian scaling specialization yields an explicit end-to-end rate.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 27 canonical work pages

  1. [1]

    Adams and John J

    Robert A. Adams and John J. F. Fournier.Sobolev Spaces. Academic Press, Amsterdam, second edition, 2003

    work page 2003

  2. [2]

    Albergo, Nicholas M

    Michael S. Albergo, Nicholas M. Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.Journal of Machine Learning Research, 26(209):1–80, 2025. URLhttps://www.jmlr.org/papers/v26/23-1605.html

    work page 2025

  3. [3]

    Birkh¨ auser, Basel, second edition, 2008

    Luigi Ambrosio, Nicola Gigli, and Giuseppe Savar´ e.Gradient Flows in Metric Spaces and in the Space of Probability Measures. Birkh¨ auser, Basel, second edition, 2008

    work page 2008

  4. [4]

    Error bounds for flow match- ing methods.Transactions on Machine Learning Research, 2024

    Joe Benton, George Deligiannidis, and Arnaud Doucet. Error bounds for flow match- ing methods.Transactions on Machine Learning Research, 2024. URL https: //openreview.net/forum?id=uqQPyWFDhY

    work page 2024

  5. [5]

    Bogachev.Gaussian Measures

    Vladimir I. Bogachev.Gaussian Measures. American Mathematical Society, Providence, Rhode Island, 1998

    work page 1998

  6. [6]

    Oxford University Press, Oxford, 2013

    St´ ephane Boucheron, G´ abor Lugosi, and Pascal Massart.Concentration Inequalities: A Nonasymptotic Theory of Independence. Oxford University Press, Oxford, 2013

    work page 2013

  7. [7]

    Scale-adaptive generative flows for multiscale scientific data.arXiv preprint arXiv:2509.02971, 2025

    Yifan Chen and Eric Vanden-Eijnden. Scale-adaptive generative flows for multiscale scientific data.arXiv preprint arXiv:2509.02971, 2025. doi: 10.48550/arXiv.2509.02971

    arXiv doi:10.48550/arxiv.2509.02971 2025

  8. [8]

    Cambridge University Press, Cambridge, second edition, 2014

    Giuseppe Da Prato and Jerzy Zabczyk.Stochastic Equations in Infinite Dimensions. Cambridge University Press, Cambridge, second edition, 2014

    work page 2014

Show all 33 references
  1. [9]

    Flow matching achieves almost minimax optimal convergence

    Kenji Fukumizu, Taiji Suzuki, Noboru Isobe, Kazusato Oko, and Masanori Koyama. Flow matching achieves almost minimax optimal convergence. InThe Thirteenth International Conference on Learning Representations, 2025. URL https://arxiv. org/abs/2405.20879

    2025 arXiv

  2. [10]

    Multilevel diffusion: Infinite dimensional score-based diffusion models for image generation.SIAM Journal on Mathematics of Data Science, 7(3):1337–1366,

    Paul Hagemann, Sophie Mildenberger, Lars Ruthotto, Gabriele Steidl, and Nicole Tian- jiao Yang. Multilevel diffusion: Infinite dimensional score-based diffusion models for image generation.SIAM Journal on Mathematics of Data Science, 7(3):1337–1366,

  3. [11]

    Springer, Cham, third edition,

    Olav Kallenberg.Foundations of Modern Probability. Springer, Cham, third edition,

  4. [12]

    Functional flow matching

    Gavin Kerrigan, Giosue Migliorini, and Padhraic Smyth. Functional flow matching. In Sanjoy Dasgupta, Stephan Mandt, and Yingzhen Li, editors,Proceedings of the 27th International Conference on Artificial Intelligence and Statistics, volume 238 ofProceedings of Machine Learning...

    2024

  5. [13]

    On universal approximation and error bounds for Fourier neural operators.Journal of Machine Learning Research, 22(290):1–76, 2021

    Nikola Kovachki, Samuel Lanthaler, and Siddhartha Mishra. On universal approximation and error bounds for Fourier neural operators.Journal of Machine Learning Research, 22(290):1–76, 2021. URLhttps://www.jmlr.org/papers/v22/21-0806.html. 29 Shikhman

    2021

  6. [14]

    Neural operator: Learning maps between function spaces with applications to PDEs.Journal of Machine Learning Re- search, 24(89):1–97, 2023

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhat- tacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to PDEs.Journal of Machine Learning Re- search, 24(89):1–97, 2023. URLhttps:/...

    2023

  7. [15]

    Error estimates for DeepONets: A deep learning framework in infinite dimensions.Transactions of Mathematics and Its Applications, 6(1):tnac001, 2022

    Samuel Lanthaler, Siddhartha Mishra, and George Em Karniadakis. Error estimates for DeepONets: A deep learning framework in infinite dimensions.Transactions of Mathematics and Its Applications, 6(1):tnac001, 2022. doi: 10.1093/imatrm/tnac001

    2022 doi

  8. [16]

    Optimal-transport-guided functional flow matching for turbulent field generation in Hilbert space.arXiv preprint arXiv:2604.05700, 2026

    Kunpeng Li, Chenguang Wan, Zhisong Qu, Kyungtak Lim, Virginie Grandgirard, Xavier Garbet, Hua Yu, and Yew Soon Ong. Optimal-transport-guided functional flow matching for turbulent field generation in Hilbert space.arXiv preprint arXiv:2604.05700, 2026. doi: 10.48550/arXiv.2604.05700

    2026 doi arXiv

  9. [17]

    Functional mean flow in hilbert space

    Zhiqi Li, Yuchen Sun, Greg Turk, and Bo Zhu. Functional mean flow in hilbert space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1928–1938, June 2026

    1928

  10. [18]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Representations, 2023. URLhttps://openreview.net/forum?id=PqvMRDCJT9t

    2023

  11. [19]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InThe Eleventh International Conference on Learning Representations, 2023. URLhttps://openreview.net/forum? id=XVjTT1nw5z

    2023

  12. [20]

    Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3:218–229, 2021

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3:218–229, 2021. doi: 10.1038/ s42256-021-00302-5

    2021

  13. [21]

    Probability-flow ode in infinite-dimensional function spaces

    Kunwoo Na, Junghyun Lee, Se-Young Yun, and Sungbin Lim. Probability-flow ode in infinite-dimensional function spaces. 2025. URL https://arxiv.org/abs/2503.10219

    2025 arXiv

  14. [22]

    Narcowich, Joseph D

    Francis J. Narcowich, Joseph D. Ward, and Holger Wendland. Sobolev bounds on func- tions with scattered zeros, with applications to radial basis function surface fitting.Math- ematics of Computation, 74(250):743–763, 2005. doi: 10.1090/S0025-5718-04-01708-9

    2005 doi

  15. [23]

    Infinite- dimensional diffusion models.Journal of Machine Learning Research, 25(414):1–52,

    Jakiw Pidstrigach, Youssef Marzouk, Sebastian Reich, and Sven Wang. Infinite- dimensional diffusion models.Journal of Machine Learning Research, 25(414):1–52,

  16. [24]

    Statistical learning theory for neural operators.Journal of Machine Learning Research, 27(102):1–82, 2026

    Niklas Reinhardt, Sven Wang, and Jakob Zech. Statistical learning theory for neural operators.Journal of Machine Learning Research, 27(102):1–82, 2026. URL https: //jmlr.org/beta/papers/v27/25-0543.html

    2026

  17. [25]

    Ross, Domniki Asimaki, and Kamyar Aziz- zadenesheli

    Yaozhong Shi, Zachary E. Ross, Domniki Asimaki, and Kamyar Aziz- zadenesheli. Stochastic process learning via operator flow matching. In 30 Statistical Consistency of Functional Flow Matching Advances in Neural Information Processing Systems, volume 38, 2025. URL https://proce...

    2025

  18. [26]

    Three superposition principles: Currents, continu- ity equations and curves of measures.Journal of Functional Analysis, 272(3):1044–1103,

    Eugene Stepanov and Dario Trevisan. Three superposition principles: Currents, continu- ity equations and curves of measures.Journal of Functional Analysis, 272(3):1044–1103,

  19. [27]

    Cambridge University Press, Cam- bridge, 2005

    Holger Wendland.Scattered Data Approximation. Cambridge University Press, Cam- bridge, 2005

    2005

  20. [28]

    Flow straight and fast in hilbert space: Functional rectified flow

    Jianxin Zhang and Clayton Scott. Flow straight and fast in hilbert space: Functional rectified flow. InThe Fourteenth International Conference on Learning Representations,

  21. [2017]

    doi: 10.1016/j.jfa.2016.10.025

    2016 doi

  22. [2021]

    doi: 10.1007/978-3-030-61871-1

    doi

  23. [2024]

    URLhttps://www.jmlr.org/papers/v25/23-1271.html

  24. [2025]

    doi: 10.1137/23M1614092

    doi

  25. [2026]

    URLhttps://openreview.net/forum?id=GWK8fm1r9y. 31

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.