Pith. sign in

REVIEW 2 major objections 6 minor 16 references

On Symmetric Kernel Collocation for Nonlinear PDEs

T0 review · 2 major / 6 minor · reviewed 2026-07-08 · glm-5.2

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

Pith's one-line read Kernel collocation for nonlinear PDEs converges without uniqueness

desk verdict Solid convergence theory for nonlinear kernel collocation with greedy point selection; one narrow gap in the greedy well-definedness argument. read the letter →

arxiv 2607.06276 v1 pith:CA7PTU72 submitted 2026-07-07 math.NA cs.NA

Milan Bacchetta , Tobias Ehring , Bernard Haasdonk This is my paper · ORCID
classification math.NAcs.NA
keywords nonlinearpointproblemcollocationconvergencegeneralizedresidual-greedyrkhs
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 extends symmetric kernel collocation—a meshless method for approximating PDE solutions using reproducing kernel Hilbert spaces (RKHSs)—from linear to nonlinear boundary value problems. The key reformulation casts the collocation problem as an optimal-recovery or minimum-norm generalized interpolation problem: find the function of smallest RKHS norm that satisfies the nonlinear PDE and boundary conditions at selected collocation points. The central analytical contribution is a convergence framework that operates directly in the RKHS norm and, crucially, does not require the PDE to have a unique solution. When the solution is nonunique, the sequence of interpolants is shown to converge to the set of minimum-norm RKHS solutions. The paper establishes this convergence for two collocation point selection strategies: first, for point sets whose fill distance (a measure of how densely points cover the domain and boundary) vanishes; second, for a novel residual-greedy strategy that adaptively selects the next collocation point where the current PDE residual is largest. The greedy convergence proof (Theorem 4.5) uses a compactness-and-separation argument: if residuals fail to vanish, the greedy rule forces selected points to be separated by a fixed distance, which contradicts compactness of the domain. Numerical experiments on a stationary nonlinear heat equation show the residual-greedy approach achieves markedly smaller PDE residuals than geometric point selection, and remains effective even when the true solution lies outside the chosen RKHS.

What carries the argument

The central objects are: (1) the minimum-norm generalized interpolation problem (4), which minimizes the RKHS norm subject to nonlinear PDE and boundary constraints at collocation points; (2) the set S_k of minimum-norm RKHS solutions, defined as the argmin of the RKHS norm over all solutions; (3) the alpha-weak greedy sequence (Definition 4.4), which selects each new collocation point where the current residual exceeds alpha times the supremal residual; (4) Proposition 4.1, the abstract convergence tool showing that pointwise residual convergence implies RKHS convergence to S_k; and (5) the effective fill distance (Definition 4.2), controlling point density on both domain interior and.

What would settle it

If one can exhibit a nonlinear PDE satisfying Assumption 3.1 for which the finite-dimensional problem (5) has no minimizer at some collocation configuration, the greedy sequence is not well-defined and Theorem 4.5 does not apply.

Watch

Extended reading notes

Core claim

The paper's central result is that a residual-greedy kernel collocation scheme for nonlinear PDEs produces a sequence of minimum-norm generalized interpolants that converge to the set of minimum-norm RKHS solutions of the PDE, without requiring solution uniqueness. The mechanism is an optimal-recovery formulation—minimizing the RKHS norm subject to nonlinear pointwise PDE constraints—combined with an alpha-weak greedy selection rule that picks collocation points where the residual is largest. The convergence proof hinges on showing that if the residual fails to vanish, the greedy rule generates a sequence of points that are separated by a fixed distance, which is impossible on a compact set.

Load-bearing premise

The convergence proofs assume that the nonlinear optimal recovery problem (4) attains a minimizer at each step of the greedy iteration. This existence is built into the definition of the greedy sequence rather than proven from the PDE structure. If the nonlinear optimization fails to have a solution at some step, the entire greedy construction breaks down.

Editorial extensions

If this is right

  • The convergence framework removes the uniqueness assumption that prior nonlinear kernel collocation analyses required, broadening the class of PDEs to which optimal-recovery-based kernel methods can be rigorously applied.
  • The residual-greedy strategy provides a principled, target-dependent alternative to fill-distance-based point selection for nonlinear PDEs, with convergence guaranteed by Theorem 4.5.
  • The minimum-norm selection principle means the method preferentially finds the smoothest (in the RKHS sense) solution among possibly many, which could be advantageous for regularization and ill-posed problems.
  • Numerical evidence of effectiveness for solutions outside the RKHS suggests the method may have practical utility beyond the theoretical setting, motivating analysis in the spirit of escape-from-native-space results.
  • The framework extends naturally to integral operators, as noted in the outlook, potentially broadening applicability to nonlocal equations.

Reading between the lines

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

  • The minimum-norm selection criterion implicitly ties the method to Tikhonov-style regularization: the RKHS norm acts as a smoothness penalty, and the kernel choice determines which solutions are preferred. This connection is not made explicit but could unify the method with regularized inverse-problem theory.
  • The alpha-weak greedy convergence proof does not yield rates, and the numerical experiments do not clearly suggest rate-type behavior. This is consistent with the linear theory where greedy rates require additional assumptions (e.g., Kolmogorov n-width decay); analogous nonlinear conditions might be identifiable.
  • The assumption that a minimizer u_n exists at each greedy step (built into Definition 4.4) is a nontrivial requirement for nonlinear optimization. The finite-dimensional reduction in Proposition 3.2 reduces the problem to minimizing z^T K_X^{-1} z subject to F(z) = y, but existence of a solution to this nonconvex system is not guaranteed in general and depends on the structure of the nonlinear ope
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 / 6 minor

Summary. This paper studies symmetric kernel collocation for nonlinear PDEs, formulating the problem as an optimal-recovery (minimum-norm generalized interpolation) problem in an RKHS with nonlinear functional constraints. The authors prove convergence of the resulting interpolants to the set of minimum-norm RKHS solutions, without assuming uniqueness of the PDE solution. Two regimes are analyzed: (i) sequences of collocation points with vanishing effective fill distance, and (ii) a novel residual-greedy, target-dependent point selection strategy. Numerical experiments on a stationary nonlinear heat equation compare the residual-greedy approach against geometric greedy collocation, showing improved residual reduction. The analysis builds on prior work on Gaussian-process methods for nonlinear PDEs [4] and optimal recovery in RKHSs [5].

Significance. The paper makes a solid contribution to the kernel collocation literature by extending the optimal-recovery framework to nonlinear PDEs and proving RKHS convergence without a uniqueness assumption. The convergence proofs are parameter-free derivations based on standard functional analysis tools (boundedness via comparison with a feasible point, weak compactness, equicontinuity, and contradiction arguments). The reduction to a finite-dimensional problem (Proposition 3.2, adapted from [5]) is a clean structural result. The residual-greedy scheme and its convergence proof (Theorem 4.5) constitute the most novel contribution. The numerical experiments are illustrative and include a case where the exact solution lies outside the RKHS, which is a useful stress test. The framework is honestly presented, including acknowledgment of the gap between theory and implementation.

major comments (2)
  1. Definition 4.4 and Theorem 4.5: The α-weak greedy sequence is defined to include the existence of a minimizer u_n at each step as part of the definition. The authors note that 'this existence requirement is included in the definition.' However, Proposition 3.2 establishes existence of a minimizer for problem (4) whenever the collocation functionals in X(X) are linearly independent (Assumption 3.1). The gap is that Definition 4.4 selects x_n greedily from Ω̄ without verifying that the augmented functional set X(X_n) remains linearly independent at every step. If at some step the newly selected point produces a linearly dependent functional, Assumption 3.1 fails for X_n, Proposition 3.2 no longer applies, and the minimizer u_n may not exist — breaking the greedy sequence and the proof of Theorem 4.5. The authors should either (a) prove that greedy selection preserves linear independence of
  2. Definition 4.4 and Theorem 4.5 (continued): the collocation functionals (at least for standard kernels such as Gaussian or Wendland with distinct points), or (b) explicitly state this as a conditional result and discuss how likely the condition is to hold in practice. As written, Theorem 4.5 is conditional on the greedy sequence being well-defined, but this is not clearly stated in the theorem itself. This is a load-bearing issue for the central claim of Theorem 4.5.
minor comments (6)
  1. Section 5: The numerical experiments use a single shape parameter γ=5 for the Gaussian kernel. No sensitivity analysis or justification for this choice is provided. A brief discussion of how γ was selected and whether the results are robust to this choice would strengthen the experimental contribution.
  2. Figure 1 and Figure 2 appear to be swapped relative to their captions: Figure 1 is labeled as showing the interior residual but the y-axis is labeled ∥u−u_n∥∞, while Figure 2 is labeled as showing the approximation error but the y-axis is labeled ∥R_P[u_n]∥∞. Please verify the numbering.
  3. Section 5: The implementation differs from the theoretical framework in two respects (alternating interior/boundary selection, discrete grid maximization). These deviations are acknowledged but their impact on the applicability of Theorem 4.5 is not discussed. A sentence clarifying that the experiments are illustrative rather than a direct validation of Theorem 4.5 would help.
  4. The reference to the master's thesis [2] is cited for 'more detailed experiments and further discussion.' Since this is not publicly accessible (master's thesis, University of Stuttgart, 2025), key supporting content should be summarized in the paper itself.
  5. Definition 4.2: The effective fill distance is defined as max{h^Ω_{X_Ω}, h^{∂Ω}_{X_{∂Ω}}}. For domains where ∂Ω is a (d−1)-dimensional manifold, the fill distance h^{∂Ω}_{X_{∂Ω}} is computed in the subspace topology of ∂Ω. This should be stated explicitly to avoid ambiguity about the metric used on the boundary.
  6. The proof of Theorem 4.5 states 'Since Ω̄, and hence also ∂Ω, is compact.' The compactness of ∂Ω follows from Ω being bounded and open (so Ω̄ is compact), but the implication 'Ω̄ compact, hence ∂Ω compact' should be stated more precisely (e.g., '∂Ω is a closed subset of the compact set Ω̄').

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and the constructive report. The referee's single major comment concerns the relationship between Definition 4.4 (α-weak greedy sequence) and the linear independence condition in Assumption 3.1, and whether Theorem 4.5 should more explicitly state its conditional nature. We agree that this point needs to be addressed and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: Definition 4.4 and Theorem 4.5: The α-weak greedy sequence is defined to include the existence of a minimizer u_n at each step as part of the definition. However, Proposition 3.2 establishes existence of a minimizer only when the collocation functionals in X(X) are linearly independent (Assumption 3.1). The gap is that Definition 4.4 selects x_n greedily from Ω̄ without verifying that the augmented functional set X(X_n) remains linearly independent at every step. If linear dependence occurs, Assumption 3.1 fails, Proposition 3.2 no longer applies, and the minimizer u_n may not exist — breaking the greedy sequence and the proof of Theorem 4.5. The authors should either (a) prove that greedy selection preserves linear independence of the collocation functionals, or (b) explicitly state this as a conditional result and discuss how likely the condition is to hold in practice.

    Authors: We thank the referee for identifying this issue, which is a legitimate gap in the presentation. We agree that the conditional nature of Theorem 4.5 is not sufficiently clearly stated and will revise the manuscript to address this. We will take the following steps: First, we will add a remark after Definition 4.4 clarifying that the definition implicitly requires that the augmented set of collocation functionals X(X_n) satisfies the linear independence condition of Assumption 3.1 at every step n, and that the existence of u_n (included in the definition) depends on this condition via Proposition 3.2. Second, we will add an explicit hypothesis to Theorem 4.5 stating that the α-weak greedy sequence is well-defined, meaning that linear independence of X(X_n) is preserved at each step. This makes the conditional nature of the result transparent. Third, we will add a remark discussing when this condition holds in practice. For strictly positive definite kernels such as the Gaussian kernel (used in our experiments) and Wendland kernels, the collocation matrix K_X is positive definite whenever the collocation functionals correspond to evaluation of distinct linear differential operators at distinct points. This is a standard result in the RBF literature (see, e.g., Wendland, Scattered Data Approximation, Theorem 16.5 and surrounding discussion): for strictly positive definite kernels, the generalized collocation matrix involving derivative functionals at distinct points is positive definite, hence the functionals are linearly independent. Since the greedy algorithm in Definition 4.4 selects distinct points (x_n ∈ Ω̄ ∖ {x_1, ..., x_{n-1}}), and the differential operators L^P_q and L^B_r are fixed and finite in number, linear independence is preserved at each step for thesekernel revision: no

  2. Referee: As written, Theorem 4.5 is conditional on the greedy sequence being well-defined, but this is not clearly stated in the theorem itself. This is a load-bearing issue for the central claim of Theorem 4.5.

    Authors: We agree. The theorem statement will be revised to explicitly include the hypothesis that the α-weak greedy sequence is well-defined (i.e., that linear independence of X(X_n) is maintained at each step, ensuring the existence of u_n via Proposition 3.2). This is already implicit in the definition, but we agree it should be stated in the theorem itself for clarity. We believe this revision fully addresses the referee's concern without changing the substance of the result: the convergence proof itself is unaffected, as it only uses properties of the sequence (u_n) once existence is established. revision: no

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: convergence proofs are parameter-free derivations; the one self-citation (Prop. 3.2 from [5]) is structural, not a fitted input renamed as prediction.

full rationale

The paper's central results (Theorem 4.3 and Theorem 4.5) are convergence proofs that proceed from Assumption 3.1 through Proposition 4.1 without fitting any parameter to data and then claiming the fit as a prediction. The proofs use standard functional-analytic arguments: boundedness of (u_n) from optimality, weak compactness, equicontinuity (Lemma A.2), and a contradiction argument for the greedy case. Proposition 3.2 is adapted from [5, Thm. 3.6] (co-authored by Ehring), but it establishes a structural fact — that the infinite-dimensional problem (4) reduces to the finite-dimensional problem (5) and attains a minimum under linear independence of collocation functionals. This is not a fitted quantity renamed as a prediction; it is a reduction theorem whose hypotheses (linear independence, compactness, existence of a PDE solution) are stated independently of the conclusion. The numerical experiments use manufactured solutions with known exact answers, which is standard validation practice, not circular fitting. Definition 4.4 includes minimizer existence in the definition of a greedy sequence rather than proving it, but this is an assumption made explicit, not a circular derivation. No step in the chain reduces to its inputs by construction.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities or postulated objects. The free parameter gamma is a standard kernel hyperparameter. The axioms are standard existence and regularity assumptions for PDEs, except for the minimizer existence at each greedy step, which is folded into the definition of the greedy sequence rather than proven.

free parameters (1)
  • Gaussian kernel shape parameter gamma = 5
    Set to gamma=5 in all numerical experiments (Section 5). Not fitted to data but chosen by hand; sensitivity to this choice is not reported.
assumptions (4)
  • domain assumption The PDE (3) admits at least one solution in H_k(Omega)
    Assumption 3.1. This is a standard existence assumption but is nontrivial for nonlinear PDEs and restricts the class of problems to which the theory applies.
  • domain assumption Linear independence of collocation functionals at each step
    Assumption 3.1 requires X(X) functionals to be linearly independent. For the greedy scheme, this must hold at every iteration, which is not automatically guaranteed.
  • ad hoc to paper Existence of a minimizer u_n for the nonlinear problem (4) at each greedy step
    Definition 4.4 includes this as part of the definition of an alpha-weak greedy sequence rather than deriving it from Assumption 3.1. The paper notes that 'in certain RKHS settings, such existence can be guaranteed' but does not verify this in general.
  • domain assumption Continuity of P and B (nonlinear functions composing the differential operators)
    Stated in Section 3. Used for equicontinuity in Lemma A.2 and for passing weak limits in Proposition 4.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On Symmetric Kernel Collocation for Nonlinear PDEs." pith.science (2026). https://pith.science/paper/CA7PTU72

@misc{pith2026260706276,
  author       = {Pith},
  title        = {Pith review of: On Symmetric Kernel Collocation for Nonlinear PDEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CA7PTU72}},
  note         = {Machine review of arXiv:2607.06276}
}
read the original abstract

This paper considers kernel-based approximation methods for nonlinear partial differential equations. To this end, the problem is formulated as an optimal-recovery generalized interpolation problem, that is, as an optimization problem in an RKHS with nonlinear functional constraints. This formulation provides the basis for a convergence analysis carried out directly in the RKHS and extends existing results by relaxing the uniqueness assumption on the PDE solution. In the nonunique case, the limiting object is characterized as a minimum-norm solution. Furthermore, a residual-greedy strategy for adaptive collocation point selection is proposed, and convergence of the resulting sequence of generalized interpolants is established. Numerical experiments for a stationary nonlinear heat equation illustrate the method and indicate that residual-greedy point selection can lead to markedly smaller PDE residuals than point sets selected according to fill-distance criteria.

Figures

Figures reproduced from arXiv: 2607.06276 by the authors.

Figure 1
Figure 1. Moreover, the benefit of adding further precomputed points appears to [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Comparison of the ∞-norm approximation error for uH using residual￾greedy and geometric greedy collocation points. 0 100 200 300 400 500 10−12 10−10 10−8 10−6 10−4 10−2 100 102 Number of collocation points ∥R P [u n]∥∞ Residual-greedy Geometric greedy [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 1
Figure 1. Comparison of the ∞-norm of the interior residual for uH using residual￾greedy and geometric greedy collocation points. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: Distribution of the collocation points after residual-greedy iterations for [PITH_FULL_IMAGE:figures/full_fig_p015_3.png]
Figure 4
Figure 4. Figure 4: Comparison of the interior residual for residual-greedy and geometric [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

  1. [5]

    Ehring, B

    T. Ehring, B. Azmi, and B. Haasdonk.Recovery of the Optimal Control Value Function in Reproducing Kernel Hilbert Spaces from Verification Con- ditions. 2025.URL:https://arxiv.org/abs/2512.07477

    work page arXiv 2025

  2. [4]

    Solving and Learning Nonlinear PDEs with Gaussian Processes

    Y . Chen, B. Hosseini, H. Owhadi, and A. M. Stuart. “Solving and Learning Nonlinear PDEs with Gaussian Processes”. In:J. Comput. Phys.447 (2021), p. 110668

    work page 2021

  3. [1]

    Theory of Reproducing Kernels

    N. Aronszajn. “Theory of Reproducing Kernels”. In:Trans. Amer. Math. Soc.68 (1950), pp. 337–404

    work page 1950

  4. [2]

    Approximating Solutions to Nonlinear PDEs with Kernel Methods

    M. Bacchetta. “Approximating Solutions to Nonlinear PDEs with Kernel Methods”. Master’s thesis. Stuttgart, Germany: University of Stuttgart, 2025

    work page 2025

  5. [3]

    Error Analy- sis of Kernel/GP Methods for Nonlinear and Parametric PDEs

    P. Batlle, Y . Chen, B. Hosseini, H. Owhadi, and A. M. Stuart. “Error Analy- sis of Kernel/GP Methods for Nonlinear and Parametric PDEs”. In:J. Com- put. Phys.520 (2025), p. 113488

    work page 2025

  6. [6]

    Ehring, M.-P

    T. Ehring, M.-P. V ogel, and B. Haasdonk.Escaping the Native Space of Sobolev Kernels by Interpolation. 2025.URL:https://arxiv.org/ abs/2512.07262

    work page arXiv 2025

  7. [7]

    Solving Partial Differential Equations by Col- location Using Radial Basis Functions

    C. Franke and R. Schaback. “Solving Partial Differential Equations by Col- location Using Radial Basis Functions”. In:Appl. Math. Comput.93 (1998), pp. 73–82

    work page 1998

  8. [8]

    Kernel-based Greedy Approximation of Parametric Elliptic Boundary Value Problems

    B. Haasdonk, G. Santin, and T. Wenzel.Kernel-Based Greedy Approxima- tion of Parametric Elliptic Boundary Value Problems. 2025.URL:https: //arxiv.org/abs/2507.06731

    work page Pith review arXiv 2025

Show all 16 references
  1. [9]

    Refined Rates of Con- vergence for Target-Data Dependent Greedy Generalized Interpolation with Sobolev Kernels

    B. Haasdonk, G. Santin, T. Wenzel, and D. Winkle. “Refined Rates of Con- vergence for Target-Data Dependent Greedy Generalized Interpolation with Sobolev Kernels”. In:Appl. Math. Lett.181 (2026), p. 110005

    2026

  2. [10]

    Near-Optimal Data-Independent Point Locations for Radial Basis Function Interpolation

    S. De Marchi, R. Schaback, and H. Wendland. “Near-Optimal Data-Independent Point Locations for Radial Basis Function Interpolation”. In:Adv. Comput. Math.23 (2005), pp. 317–330. 17

    2005

  3. [11]

    Convergence Rate of the Data-Independent P-Greedy Algorithm in Kernel-Based Approximation

    G. Santin and B. Haasdonk. “Convergence Rate of the Data-Independent P-Greedy Algorithm in Kernel-Based Approximation”. In:Dolomites Res. Notes Approx.10 (2017), pp. 68–78

    2017

  4. [12]

    Adaptive Greedy Techniques for Approx- imate Solution of Large RBF Systems

    R. Schaback and H. Wendland. “Adaptive Greedy Techniques for Approx- imate Solution of Large RBF Systems”. In:Numer. Algorithms24 (2000), pp. 239–254

    2000

  5. [13]

    Steinwart and A

    I. Steinwart and A. Christmann.Support Vector Machines. Information Sci- ence and Statistics. New York, NY: Springer, 2008

    2008

  6. [14]

    Piecewise Polynomial, Positive Definite and Compactly Sup- ported Radial Functions of Minimal Degree

    H. Wendland. “Piecewise Polynomial, Positive Definite and Compactly Sup- ported Radial Functions of Minimal Degree”. In:Adv. Comput. Math.4 (1995), pp. 389–396

    1995

  7. [15]

    Wendland.Scattered Data Approximation

    H. Wendland.Scattered Data Approximation. V ol. 17. Cambridge Mono- graphs on Applied and Computational Mathematics. Cambridge: Cambridge University Press, 2010

    2010

  8. [16]

    Adaptive Meshfree Approximation for Linear Elliptic Partial Differential Equations with PDE- Greedy Kernel Methods

    T. Wenzel, D. Winkle, G. Santin, and B. Haasdonk. “Adaptive Meshfree Approximation for Linear Elliptic Partial Differential Equations with PDE- Greedy Kernel Methods”. In:BIT Numer. Math.65 (2025), p. 11. A Auxiliary results Lemma A.1.Assume the setting of Assumption 3.1. Let ...

    2025

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.