PIKS: Universal Physics-Informed Kernel Methods

Giacomo Meanti; Joachim Bona-Pellissier; Lorenzo Rosasco; Matteo Santacesaria

REVIEW 2 major objections 6 minor 12 references

Kernel ridge regression with linear differential constraints is universally consistent even when the target lies outside the RKHS.

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 ↗

T0 review · grok-4.5

2026-07-30 12:18 UTC pith:TYH3O45F

load-bearing objection Solid misspecified consistency theorem for physics-informed kernels; the real delta is joint L2 recovery of u and Du without u* in H. the 2 major comments →

arxiv 2607.27062 v1 pith:TYH3O45F submitted 2026-07-29 stat.ML cs.LG

PIKS: Universal Physics-Informed Kernel Methods

Joachim Bona-Pellissier , Giacomo Meanti , Matteo Santacesaria , Lorenzo Rosasco This is my paper

classification stat.ML cs.LG MSC 68T0562G0865N3546E22

keywords physics-informed machine learningkernel methodsuniversal consistencymisspecified settingdifferential operatorsreproducing kernel Hilbert spacesource conditionsPDE collocation

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.

Physics-informed learning adds knowledge of differential operators to ordinary regression, but most theory either assumes the unknown function already lives in the model space or relies on neural networks whose optimization is hard to analyze. This paper introduces PIKS, a kernel method that simply appends a physics residual term to the usual squared loss and solves the resulting linear system in closed form. The central claim is that, for universal kernels such as Gaussians or Matérns, the estimator simultaneously recovers the target values and the action of the differential operator as both the ordinary data and the physics collocation samples grow, even when the true function is too rough to belong to the native RKHS. Finite-sample rates are also given under source conditions, and small experiments show the method remaining competitive with PINNs and finite-element solvers under misspecification. A sympathetic reader cares because the result supplies the first universal-consistency guarantee for a tractable physics-informed learner and shows that classical operator analysis extends cleanly once both data sources are allowed to grow.

Core claim

Under mild boundedness and embedding assumptions, any regularizing sequence that tends to zero slowly enough makes the PIKS estimator a universal learner: almost surely both the L2 error on function values and the L2 error on the differential residual vanish, even when the target lies outside the RKHS of a universal kernel.

What carries the argument

The composite sampling operator A ho that maps a candidate function to the pair (function values under ho X, differential residual under ho Z), together with its associated integral operator L = A ho A ho*; consistency and rates are obtained by extending the classical bias-variance decomposition of kernel ridge regression to this joint operator.

Load-bearing premise

Both the ordinary sample size and the physics collocation sample size must tend to infinity together; if only one of them grows, the joint consistency claim does not hold.

What would settle it

Fix a universal kernel and a target outside its RKHS; keep the number of physics collocation points fixed while letting ordinary samples grow, and check whether the differential residual of the estimator converges to zero in L2—if it does not, the necessity of joint growth is confirmed.

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

If this is right

Universal kernels can be used for physics-informed tasks without first verifying that the unknown solution is smooth enough to lie in the RKHS.
Closed-form kernel solvers become theoretically justified competitors to PINNs for linear differential constraints.
Elliptic regularity can be imported after the fact: L2 consistency of values and residual automatically upgrades to stronger Sobolev convergence on compact subdomains.
Source-condition rates supply concrete guidance for choosing the regularization parameter once both sample sizes are known.

Where Pith is reading between the lines

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

The same operator framework should extend, with only technical changes, to systems of linear PDEs or to constraints that mix several differential operators of different orders.
Introducing an explicit weight that vanishes with the ordinary sample size would decouple the two convergences and recover ordinary kernel ridge regression as a special case, at the price of losing asymptotic physical consistency.
Scalable approximations already developed for derivative-augmented kernel matrices can be dropped in immediately, turning the cubic cost into a practical large-scale method while preserving the consistency guarantees.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Solid misspecified consistency theorem for physics-informed kernels; the real delta is joint L2 recovery of u and Du without u* in H.

read the letter

The thing worth knowing is Theorem 1: for universal kernels, under standard boundedness and density assumptions, the block-kernel ridge estimator jointly recovers both the target values and the linear differential residual in L2, almost surely, even when u* lies outside the RKHS. That is the gap the paper actually fills. Prior Hermite–Birkhoff, RBF collocation, GP-PDE, and Doumèche-style physics kernels mostly stayed in the well-specified or deterministic fill-distance regime; the operator-theoretic misspecified argument here is the new piece.

They do the work cleanly. Closed-form estimator via the usual block Gram matrix of K and KD, error split into approximation plus estimation, integral/covariance operators on the product space, Hoeffding plus Borel–Cantelli for the a.s. limit. The Sobolev/Cs0-universality lemmas and the elliptic upgrades (H2 interior, H1/2 global under boundary sampling) sit on standard PDE facts and are stated carefully. Source-condition rates are the expected extension; the periodic Laplacian example makes the source condition readable. Related work is honest about Kimeldorf–Wahba, Fasshauer, Wendland, and Doumèche.

Soft spots are real but proportionate. Both n and m must go to infinity together; the convergences are not decoupled without an extra vanishing weight (they flag this). Boundedness of targets and noise is load-bearing for the concentration they use. Scope is linear operators only. Experiments are small-scale and competitive with PINNs/FEM/PIKL on a few PDEs, including a misspecification sweep with Matérn targets; they illustrate rather than stress-test. No code shipped. None of that breaks the central claim.

This is for people who care about learning theory for scientific ML or meshless methods and want a justification that does not pretend the target lives in the native RKHS. It deserves a serious referee. I would engage with it and expect to cite the consistency result.

Referee Report

2 major / 6 minor

Summary. The paper introduces Physics-Informed Kernel methodS (PIKS): kernel ridge regression with an additive empirical residual for a known linear operator D, yielding a closed-form estimator via a block Gram matrix built from K and the derivative representers K^D. The central theoretical claim (Theorem 1) is universal consistency in the misspecified regime: under Assumptions 1–5, for Cs0-universal kernels and any λ_{n,m}→0 with log N/(λ³_{n,m} N)→0 (N=min(n,m)), one has almost-sure joint convergence ∥û_λ−u*∥_{L²(ρ_X)}→0 and ∥Dû_λ−Du*∥_{L²(ρ_Z)}→0, even when u*∉H. The argument extends classical operator-theoretic KRR (sampling/covariance operators, estimation–approximation split, Hoeffding/Borel–Cantelli) to the product space via A_ρ. Specializations cover differential operators on Sobolev spaces, in-domain and boundary sampling (PDE BVP), elliptic upgrades to H²_loc / H^{1/2}, and finite-sample rates under a physics-informed source condition (Theorem 2), illustrated for the periodic Laplacian. Small-scale experiments compare PIKS to PINNs, PIKL, classical solvers, and FEM, including a misspecification study with Matérn targets of varying smoothness.

Significance. The misspecified consistency result fills a genuine gap: prior kernel/RBF and physics-informed kernel analyses largely treat the well-specified case u*∈H or deterministic fill-distance settings, while PINN theory remains limited by nonconvex optimization. Extending the integral-operator toolkit to joint learning of values and linear differential residuals, with explicit Cs0-universality and trace/elliptic upgrades, is a solid contribution to statistical learning theory for scientific ML. Strengths include a closed-form estimator, fully spelled-out assumptions, complete appendix proofs, interpretable source conditions in a model example, and empirical evidence that PIKS remains competitive under misspecification and against PINNs/FEM. Limitations (linear D only; coupled (n,m)→∞; cubic cost) are stated and do not erase the main advance.

major comments (2)

[§4.2, Theorem 1; §4.5] §4.2, Remark after Theorem 1 and the rate statements (Theorem 2 / Cor. 5): joint consistency and rates require both n and m → ∞ (N=min(n,m)). In the PDE BVP regime of §4.4 this is natural, but many practical settings have abundant collocation points and scarce boundary/measurement data (or vice versa). The paper flags decoupling via a vanishing physics weight but does not analyze it. A short proposition or remark giving conditions under which one residual still converges when the other sample size is fixed (or grows much more slowly), even if only in a weaker topology, would make the main theorem more usable and would clarify when the physics term is statistically necessary versus optional regularization.
[§4.5, Assumption 6, Proposition 6] §4.5, Assumption 6 and Example 1: finite-sample rates are stated in terms of ran(L^r) for the composite operator L=A_ρ A_ρ*, which is hard to check outside the periodic Laplacian. The paper correctly leaves general interpretation to future work, but the claim that rates are obtained “under suitable source conditions” is only operationally meaningful in that example. Either (i) add one more concrete case (e.g., 1D elliptic operator on an interval with explicit Green’s function / spectral picture), or (ii) temper the abstract/intro wording to stress that rates are conditional on an abstract source condition whose Sobolev meaning is verified only for the periodic Laplacian.

minor comments (6)

[§4.5, Appendix C] Notation overload: A_ρ is used both for the map F→L²(ρ_X)×L²(ρ_Z) (Eq. 22) and for A_ρ∘i on H (Eq. 23); the same symbol appears again in the appendix with slightly different roles. A consistent distinction (e.g., A_ρ vs A_ρ,H) would help.
[§3, §4] Lemma/Proposition numbering in the main text sometimes refers to “Proposition 1” for what is labeled Lemma 1 (representers), and appendix restatements (Prop. 8, 29, …) are easy to lose track of. Align labels or add a short “main text ↔ appendix” map.
[§5.2] Figure 2 caption and Table 2: state explicitly the kernel (RBF), length-scale selection, and λ schedule used for PIKS so that the comparison to PIKL/PINN is reproducible from the main text alone.
[§2.2] Related work: the distinction from Doumèche et al. (continuous physics penalty / well-specified rates) and from RBF collocation (fill-distance, noiseless, often well-specified) is good; a single sentence on how the present random-design + noise + misspecification guarantees relate to Schräder–Wendland-type fill-distance bounds would round out §2.2.
[§1, §6] Typos/style: “Wemakethisgeneralresultconcrete” (contribution list); occasional missing spaces after periods in the contribution bullets; “theresulting” etc. in the conclusions. A proofreading pass is enough.
[§4.1, Assumption 4] Assumption 4 (bounded targets and noise) is standard but strong for PDE data. A brief forward pointer that Bernstein-type extensions are possible, as already alluded to in §4.1, would set expectations for practitioners with heavy-tailed measurement noise.

Circularity Check

0 steps flagged

No significant circularity: consistency is a theorem from stated assumptions, not a fit or self-definitional loop.

full rationale

The central claim (Theorem 1 / Prop. 29) is an asymptotic consistency result derived from Assumptions 1–5 via a standard operator-theoretic KRR argument extended to the product space A_ρ: density of H in F plus boundedness of A_ρ put (h,q) in the closure of ran A_ρ (Prop. 28); approximation (Prop. 16) and estimation via Hoeffding/Borel–Cantelli (Props. 19, 26) then yield the joint L² limits. The estimator is defined as the minimizer of a regularized empirical risk and is not defined to equal the target. Source conditions and λ schedules are classical devices, not circular redefinitions of u*. Self-citations (if any) to prior kernel work by overlapping authors are foundational KRR facts, not load-bearing uniqueness theorems that force the physics-informed conclusion. Empirical tables compare external baselines (PINNs, FEM, PIKL) without fitting a parameter and relabeling it as a prediction. No step reduces Eq. X to Eq. Y by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 8 axioms · 1 invented entities

The theory rests on standard RKHS/PDE background plus several domain assumptions (bounded features/noise, dense embedding in the Sobolev space via Cs₀-universality, sampling measures compatible with traces). No new physical entities. Free parameters appear only in experiments (kernel length-scale, λ, point counts), not inside the consistency theorem itself.

free parameters (3)

Regularization λ_{n,m} = problem-dependent schedule / tuned
Must satisfy λ→0 and log N/(λ³ N)→0 for consistency; for rates, explicit schedules λ=N^{-1/2} or N^{-1/(2r+1)}. Chosen by theory or tuning in experiments, not identified from data as a physical constant.
Kernel hyperparameters (length-scale, Matérn ν) = tuned per experiment (not fully tabulated)
Affect finite-sample performance and whether Cs₀-universality/smoothness assumptions hold; experiments use RBF/Gaussian with small-scale tuning.
Sample sizes n, m and collocation design = experiment-specific
Both must grow for the joint guarantee; empirical tables fix specific counts (e.g. 100 boundary points, ~22k total for wave).

axioms (8)

domain assumption Assumption 1: u ↦ Du(x) is a bounded linear functional on H for each x (Riesz representers K^D_x exist).
Needed for the physics feature maps and closed-form estimator; verified for differential operators when K∈C^{2s}.
domain assumption Assumption 2: F↪L²(ρ_X) and G↪L²(ρ_Z) are well-defined and bounded (including via trace when ρ_X lives on ∂Ω).
Makes the risk and sampling of u*(X), Du*(Z) meaningful for Sobolev targets.
domain assumption Assumption 3 (Universality): H is densely embedded in F (stronger than L² universality; Cs₀-universality for Sobolev).
Load-bearing for misspecified consistency; Gaussian/Matérn with sufficient smoothness are cited as examples.
domain assumption Assumption 4: u* and Du* bounded in L∞ under the sampling measures; noise ϵ,η a.s. bounded.
Enables Hoeffding-type concentration; authors note moment relaxations are possible but out of scope.
domain assumption Assumption 5: Bounded kernel features K(x,x)≤κ² and K^D(x,x)≤κ_D².
Standard for operator-norm control of covariance estimators on compact X.
domain assumption Assumption 6 (rates): physics-informed source condition (h,q)∈ran(L^r) for the composite integral operator L=A_ρ A_ρ*.
Needed only for finite-sample rates, not bare consistency; interpreted for periodic Laplacian.
standard math Classical RKHS reproducing properties for derivatives (Zhou 2008) and elliptic regularity / trace theorems (Evans; Lions–Magenes).
Used to verify assumptions and upgrade L² residual convergence to local H² or H^{1/2} convergence.
ad hoc to paper D is a known linear differential operator; analysis does not cover nonlinear PDEs.
Scope restriction stated throughout; central theorems are for linear D.

invented entities (1)

PIKS estimator (physics-informed kernel ridge with block Gram matrix of K and K^D) independent evidence
purpose: Name the concrete regularized ERM and its closed form used for analysis and experiments.
Algorithmically continuous with Kimeldorf–Wahba / RBF collocation / GP-PDE block systems; the new content is the misspecified learning theory around it, not a new physical object.

pith-pipeline@v1.2.0-daily-grok45 · 53322 in / 4014 out tokens · 69981 ms · 2026-07-30T12:18:49.903177+00:00 · methodology

0 comments

read the original abstract

Physics-informed machine learning incorporates physical principles --often expressed via differential operators-- into data-driven models. While physics-informed neural networks (PINNs) dominate empirical applications, the complexity of neural network architectures and optimization landscapes hinders the development of a corresponding learning theory. In turn, kernel methods offer an appealing alternative with closed-form solutions and analytical tractability, yet existing guarantees primarily cover the well-specified setting where the target belongs to the native Reproducing Kernel Hilbert Space (RKHS). This imposes unrealistic regularity assumptions that physical targets often fail to satisfy. In this paper, we introduce and analyze Physics-Informed Kernel methodS (PIKS). We establish the universal consistency of PIKS for linear differential constraints, proving that for universal kernels (such as Gaussian or Mat\'ern), the estimator asymptotically learns the target while satisfying physical constraints. We further derive finite-sample bounds under suitable source conditions. Our analysis is based on extending classical operator-theoretic analysis of kernel methods to physics-informed machine learning. Numerical experiments demonstrate that PIKS can be competitive with PINNs and traditional finite element methods.

Figures

Figures reproduced from arXiv: 2607.27062 by Giacomo Meanti, Joachim Bona-Pellissier, Lorenzo Rosasco, Matteo Santacesaria.

**Figure 1.** Figure 1: Improved learning accuracy with derivative data. A dataset of function values (in red) is augmented by gradient (in blue) information to improve accuracy [PITH_FULL_IMAGE:figures/full_fig_p017_1.png] view at source ↗

**Figure 3.** Figure 3: FEM vs. PIKS on noiseless data of different smoothness. All FEM results apart from ν = 0.5 overlap at the top of the plot. PIKS with ν = ∞ plateaus at numerical precision. 10−4 10−2 100 Noise standard deviation 10−3 10−2 10−1 100 RMSE FEM PIKS [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗

**Figure 5.** Figure 5: Four Matérn Laplacians with increasing smoothness. [PITH_FULL_IMAGE:figures/full_fig_p054_5.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 2 canonical work pages

[11]

Joel A Tropp

doi: 10.1007/s44379-025-00015-1. Joel A Tropp. User-friendly tools for random matrices: An introduction,

doi:10.1007/s44379-025-00015-1
[12]

doi: https: //doi.org/10.1016/j.jcp.2021.110768

ISSN 0021-9991. doi: https: //doi.org/10.1016/j.jcp.2021.110768. Wenjia Wang and Bing-Yi Jing. Gaussian process regression: Optimality, robustness, and relation- ship with kernel ridge regression.Journal of Machine Learning Research, 23(193):1–67,

arXiv 2021
[1990]

Alfio Quarteroni, Paola Gervasio, and Francesco Regazzoni

doi: 10.1126/science.247.4945.978. Alfio Quarteroni, Paola Gervasio, and Francesco Regazzoni. Combining physics-based and data- driven models: advancing the frontiers of research with scientific machine learning.Math- ematical Models and Methods in Applied Sciences, 35(04):905–1071,

work page doi:10.1126/science.247.4945.978
[1994]

optimum bounds for the distributions of martingales in banach spaces

Iosif Pinelis. Correction:“optimum bounds for the distributions of martingales in banach spaces”[ann. probab. 22 (1994), no. 4, 1679–1706; mr 96b: 60010].The Annals of Probability, 27(4):2119–2119,

1994
[1998]

Bharath K Sriperumbudur, Kenji Fukumizu, and Gert RG Lanckriet

doi: https: //doi.org/10.1016/S0893-6080(98)00032-X. Bharath K Sriperumbudur, Kenji Fukumizu, and Gert RG Lanckriet. Universality, characteristic kernels and rkhs embedding of measures.Journal of Machine Learning Research, 12(7),

work page doi:10.1016/s0893-6080(98)00032-x
[2010]

Solvingpartialdifferentialequationsbycollocationwithradialbasisfunctions

GregoryEFasshauer. Solvingpartialdifferentialequationsbycollocationwithradialbasisfunctions. InProceedings of Chamonix, volume 1997, pages 1–8,

1997
[2018]

Analysis of p-laplacian regularization in semisupervised learn- ing.SIAM Journal on Mathematical Analysis, 51(3):2085–2120,

Dejan Slepcev and Matthew Thorpe. Analysis of p-laplacian regularization in semisupervised learn- ing.SIAM Journal on Mathematical Analysis, 51(3):2085–2120,

2085
[2020]

Carl-Johann Simon-Gabriel and Bernhard Schölkopf

doi: 10.4208/cicp.oa-2020-0193. Carl-Johann Simon-Gabriel and Bernhard Schölkopf. Kernel distribution embeddings: Universal kernels, characteristic kernels and kernel metrics on distributions.Journal of Machine Learning Research, 19(44):1–29,

doi:10.4208/cicp.oa-2020-0193 2020
[2022]

Challenges in training pinns: A loss landscape perspective.arXiv preprint arXiv:2402.01868,

Pratik Rathore, Weimu Lei, Zachary Frangella, Lu Lu, and Madeleine Udell. Challenges in training pinns: A loss landscape perspective.arXiv preprint arXiv:2402.01868,

Pith/arXiv arXiv
[2023]

Pau Batlle, Yifan Chen, Bamdad Hosseini, Houman Owhadi, and Andrew M Stuart

doi: 10.5281/zenodo.10447666. Pau Batlle, Yifan Chen, Bamdad Hosseini, Houman Owhadi, and Andrew M Stuart. Error analysis of kernel/gp methods for nonlinear and parametric pdes.Journal of Computational Physics, 520: 113488,

doi:10.5281/zenodo.10447666
[2024]

Fast kernel methods: Sobolev, physics-informed, and additive models.arXiv preprint arXiv:2509.02649, 2025a

55 Nathan Doumèche, Francis Bach, Gérard Biau, and Claire Boyer. Fast kernel methods: Sobolev, physics-informed, and additive models.arXiv preprint arXiv:2509.02649, 2025a. Nathan Doumèche, Francis Bach, Gérard Biau, and Claire Boyer. Physics-informed kernel learning. Journal of Machine Learning Research, 26(124):1–39, 2025b. Nathan Doumèche, Gérard Biau,...

Pith/arXiv arXiv
[2025]

Christopher Rackauckas, Yingbo Ma, Julius Martensen, Collin Warner, Kirill Zubov, Rohit Supekar, Dominic Skinner, Ali Ramadhan, and Alan Edelman

doi: 10.1142/ S0218202525500125. Christopher Rackauckas, Yingbo Ma, Julius Martensen, Collin Warner, Kirill Zubov, Rohit Supekar, Dominic Skinner, Ali Ramadhan, and Alan Edelman. Universal differential equations for scientific machine learning.arXiv preprint arXiv:2001.04385,

Pith/arXiv arXiv 2001

[1] [11]

Joel A Tropp

doi: 10.1007/s44379-025-00015-1. Joel A Tropp. User-friendly tools for random matrices: An introduction,

doi:10.1007/s44379-025-00015-1

[2] [12]

doi: https: //doi.org/10.1016/j.jcp.2021.110768

ISSN 0021-9991. doi: https: //doi.org/10.1016/j.jcp.2021.110768. Wenjia Wang and Bing-Yi Jing. Gaussian process regression: Optimality, robustness, and relation- ship with kernel ridge regression.Journal of Machine Learning Research, 23(193):1–67,

arXiv 2021

[3] [1990]

Alfio Quarteroni, Paola Gervasio, and Francesco Regazzoni

doi: 10.1126/science.247.4945.978. Alfio Quarteroni, Paola Gervasio, and Francesco Regazzoni. Combining physics-based and data- driven models: advancing the frontiers of research with scientific machine learning.Math- ematical Models and Methods in Applied Sciences, 35(04):905–1071,

work page doi:10.1126/science.247.4945.978

[4] [1994]

optimum bounds for the distributions of martingales in banach spaces

Iosif Pinelis. Correction:“optimum bounds for the distributions of martingales in banach spaces”[ann. probab. 22 (1994), no. 4, 1679–1706; mr 96b: 60010].The Annals of Probability, 27(4):2119–2119,

1994

[5] [1998]

Bharath K Sriperumbudur, Kenji Fukumizu, and Gert RG Lanckriet

doi: https: //doi.org/10.1016/S0893-6080(98)00032-X. Bharath K Sriperumbudur, Kenji Fukumizu, and Gert RG Lanckriet. Universality, characteristic kernels and rkhs embedding of measures.Journal of Machine Learning Research, 12(7),

work page doi:10.1016/s0893-6080(98)00032-x

[6] [2010]

Solvingpartialdifferentialequationsbycollocationwithradialbasisfunctions

GregoryEFasshauer. Solvingpartialdifferentialequationsbycollocationwithradialbasisfunctions. InProceedings of Chamonix, volume 1997, pages 1–8,

1997

[7] [2018]

Analysis of p-laplacian regularization in semisupervised learn- ing.SIAM Journal on Mathematical Analysis, 51(3):2085–2120,

Dejan Slepcev and Matthew Thorpe. Analysis of p-laplacian regularization in semisupervised learn- ing.SIAM Journal on Mathematical Analysis, 51(3):2085–2120,

2085

[8] [2020]

Carl-Johann Simon-Gabriel and Bernhard Schölkopf

doi: 10.4208/cicp.oa-2020-0193. Carl-Johann Simon-Gabriel and Bernhard Schölkopf. Kernel distribution embeddings: Universal kernels, characteristic kernels and kernel metrics on distributions.Journal of Machine Learning Research, 19(44):1–29,

doi:10.4208/cicp.oa-2020-0193 2020

[9] [2022]

Challenges in training pinns: A loss landscape perspective.arXiv preprint arXiv:2402.01868,

Pratik Rathore, Weimu Lei, Zachary Frangella, Lu Lu, and Madeleine Udell. Challenges in training pinns: A loss landscape perspective.arXiv preprint arXiv:2402.01868,

Pith/arXiv arXiv

[10] [2023]

Pau Batlle, Yifan Chen, Bamdad Hosseini, Houman Owhadi, and Andrew M Stuart

doi: 10.5281/zenodo.10447666. Pau Batlle, Yifan Chen, Bamdad Hosseini, Houman Owhadi, and Andrew M Stuart. Error analysis of kernel/gp methods for nonlinear and parametric pdes.Journal of Computational Physics, 520: 113488,

doi:10.5281/zenodo.10447666

[11] [2024]

Fast kernel methods: Sobolev, physics-informed, and additive models.arXiv preprint arXiv:2509.02649, 2025a

55 Nathan Doumèche, Francis Bach, Gérard Biau, and Claire Boyer. Fast kernel methods: Sobolev, physics-informed, and additive models.arXiv preprint arXiv:2509.02649, 2025a. Nathan Doumèche, Francis Bach, Gérard Biau, and Claire Boyer. Physics-informed kernel learning. Journal of Machine Learning Research, 26(124):1–39, 2025b. Nathan Doumèche, Gérard Biau,...

Pith/arXiv arXiv

[12] [2025]

Christopher Rackauckas, Yingbo Ma, Julius Martensen, Collin Warner, Kirill Zubov, Rohit Supekar, Dominic Skinner, Ali Ramadhan, and Alan Edelman

doi: 10.1142/ S0218202525500125. Christopher Rackauckas, Yingbo Ma, Julius Martensen, Collin Warner, Kirill Zubov, Rohit Supekar, Dominic Skinner, Ali Ramadhan, and Alan Edelman. Universal differential equations for scientific machine learning.arXiv preprint arXiv:2001.04385,

Pith/arXiv arXiv 2001