Pith. sign in

REVIEW 2 major objections 4 minor 20 references

Nonlinear Optimal Recovery in Hilbert Spaces

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proves that the solution of a nonlinear optimal recovery problem in Hilbert space converges strongly to the true solution of the underlying nonlinear equation as the number of measurements grows, and gives sufficient conditions…

desk verdict The convergence framework is a real contribution, but Proposition 3's optimality condition is false without a constraint qualification, and the finite-dimensional representation results rely on it. read the letter →

arxiv 2506.00704 v2 pith:PK3HTDF5 submitted 2025-05-31 math.NA cs.NA

Daozhe Lin , Qiang Du This is my paper
classification math.NAcs.NA MSC 65J2041A6546E22
keywords optimalrecoverynonlinearproblemsfinitemeasurementsreproducingkernelHilbertspacefinite-dimensionalrepresentationconvergenceweakcontinuityrepresentertheorem
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 introduces a framework, called nonlinear optimal recovery, for solving nonlinear equations in infinite-dimensional Hilbert spaces when only finitely many measurements (linear functionals of the solution) are available. The central result is that, under two conditions—the test functions have dense span and the nonlinear map is weakly continuous—the minimal-norm solution of the constrained problem exists for each N and converges strongly to the true solution as N tends to infinity. Because the solution is generally infinite-dimensional, the paper gives a sufficient 'finite dimensionality condition' under which the solution lies in a finite span and can be computed by a finite-dimensional optimization. For cases where that condition fails, it proposes a relaxed version with inequality constraints and proves that its solutions converge to the original recovery solution and then to the true solution. These results matter because they offer a rigorous, data-limited route to solving nonlinear PDEs with convergence guarantees.

What carries the argument

The load-bearing object is the nonlinear optimal recovery problem (12): minimize the Hilbert-space norm over u satisfying N exact measurement equations. Its optimality condition (14) expresses the solution as a combination of Riesz representatives psi(u_N, phi_n) of the Frechet derivatives of the constraint functionals, which encode the sensitivity of the measurements. The Finite Dimensionality Condition (16) postulates that each such psi is a linear combination, with u-dependent coefficients, of finitely many fixed functions psi_l(phi_n), turning the problem into a finite-dimensional minimization. For test functions that are not eligible, the relaxed formulation (18) replaces the equality constraints by inequalities with an approximation error tolerance and uses a dense subset of test functions, such as point evaluations in a reproducing kernel Banach space, preserving finite-dimensionality and convergence.

What would settle it

Take U = $R^{2}$ with the Euclidean norm, B = R, one measurement with test function phi_1 = 1, f = 0, and F(x,y) = (x-1)^3 - $y^{2}$. The nonlinear optimal recovery problem min ($x^{2}$+$y^{2}$) subject to (x-1)^3 - $y^{2}$ = 0 has solution u_N = (1,0). But the Frechet derivative of F at (1,0) is zero, so the optimality condition would force u_N = $\lambda$ * 0 = 0, contradicting u_N = (1,0); this shows the claimed representation fails without a constraint qualification.

Watch

Extended reading notes

Core claim

The paper proves that the nonlinear optimal recovery problem min ||u|| subject to [F(u), phi_n] = [f, phi_n] has a minimizer u_N for each N and that u_N converges strongly in U to the true solution u* of the abstract nonlinear problem as N grows, assuming the test functions span the dual space densely and F is weak-to-weak continuous (Theorem 1). It further shows that the minimizer satisfies the optimality condition u_N = sum_n lambda_n psi(u_N, phi_n), where psi is the Riesz representative of the Frechet derivative DuF(u). Under the finite dimensionality condition psi(u, phi_n) = sum_l c_l(u, phi_n) psi_l(phi_n), the solution collapses into a finite-dimensional span and the problem reduces to a finite-dimensional optimization (Propositions 3 and 4). When the test functions are not in the favorable subset, the relaxed nonlinear optimal recovery, which uses inequality constraints and approximations by point evaluations, yields finite-dimensional solutions and converges to u_N and then to u* (Theorem 6). An example with a reaction-diffusion equation illustrates the regularity trade-offs and the utility of the linear-nonlinear decomposition.

Load-bearing premise

The finite-dimensional representation of the solution rests on the Lagrangian optimality condition u_N = sum_n lambda_n psi(u_N, phi_n) holding at the minimizer, which requires a constraint qualification, such as surjectivity of the derivative of the constraint map, that the paper neither states nor proves.

Editorial extensions

If this is right

  • If the theorem is correct, finite measurements, not full data, are enough to recover solutions of nonlinear PDEs, with strong convergence as the number of measurements grows.
  • For nonlinear operators satisfying the finite dimensionality condition, the optimal recovery solution is computable by a finite-dimensional optimization, making the method numerically tractable without discretizing the whole domain.
  • The relaxed formulation extends finite-dimensional representability to arbitrary test functions that can be approximated by a dense subset, so the method applies to general nonlinear equations in reproducing kernel Banach spaces.
  • The convergence results cover regularization variants, linear-nonlinear decompositions, and multi-domain settings, so the framework can handle boundary conditions and lower-regularity solutions.
  • The sequential optimal recovery variant suggests a unified view of Galerkin and Petrov-Galerkin methods as special cases of a measurement-driven minimal-norm approach.

Reading between the lines

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

  • A missing constraint qualification means the finite-dimensional representation could fail for nonconvex constraints without a surjectivity condition on the derivative of the constraint map; adding such a condition would make the representer result robust.
  • The point-evaluation relaxed formulation connects this framework to kernel and Gaussian-process methods for PDEs, suggesting that those existing solvers may be viewed as instances of the same minimal-norm principle.
  • The paper's qualitative convergence does not give rates; deriving error estimates would require stronger regularity assumptions and could lead to practical stopping criteria for choosing N and M.
  • The multi-domain generalization could be extended to interface conditions and domain decomposition, potentially linking the framework to substructuring methods.
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

2 major / 4 minor

Summary. This paper proposes a general framework, termed nonlinear optimal recovery, for solving nonlinear equations in Hilbert spaces when only finitely many measurements of the data are available. The authors formulate the problem as a minimum-norm problem with nonlinear equality constraints, prove existence of solutions and their strong convergence to the true solution as the number of measurements tends to infinity (Theorem 1), and analyze a regularized variant and a relaxed inequality-constrained variant with analogous convergence properties (Theorems 2 and 6). They also introduce a sufficient condition for the recovered solution to admit a finite-dimensional representation, specialize the theory to reproducing kernel Banach spaces with point-evaluation measurements, and illustrate the framework on a reaction-diffusion PDE.

Significance. If the results held, the paper would provide a useful theoretical foundation for solving nonlinear PDEs from limited measurements, extending classical linear optimal recovery and connecting to operator-learning formulations. The paper is well structured, states its assumptions explicitly, and supplies proofs of the main convergence theorems in an appendix, together with a concrete PDE example. The convergence results and the abstract formulation are genuine contributions. However, the finite-dimensional representation, which is a central advertised contribution, rests on an optimality condition that is false without a constraint qualification, and the paper's proof does not supply one. This issue is load-bearing for the finite-dimensionality claims and must be repaired before the results can be accepted as stated.

major comments (2)
  1. [Section 3.1, Eq. (14)] The optimality condition (14) is derived by differentiating a Lagrangian for the nonconvex equality-constrained problem (12), which is valid only under a constraint qualification (e.g., surjectivity of the derivative of the constraint map at the minimizer). No such condition is stated or proved. The condition can fail even when all other assumptions of the paper hold. For example, take U=R^2, B=R, phi_1=1, F(u)=((u_1^2-1)^2+u_2^2), and f=0. The feasible set is {u_1=±1, u_2=0}; u_N=(1,0) is a minimizer for N=1. The derivative of F at u_N is zero, so psi(u_N,phi_1)=0 and (14) would require (1,0)=0. This example satisfies Assumptions 1 and 2 and the finite-dimensionality condition (16) with L=2, so it lies within the stated scope. Consequently, Proposition 3 is false as stated, and the reduction in Section 3.2 together with Propositions 4, 7, and 8, all of which invoke (14), are not justified by the arguments given. A projection/invariance argument using condition (16) might repair the conclusion, but that is not what the paper presents.
  2. [Appendix 8.3 (proof of Theorem 6)] In the proof of the second convergence claim of Theorem 6, the bound |[F(˜u), Σ_m c_mh φ_mh]| ≤ ε/(4 C_H) is asserted directly from the feasibility constraint, but the constraint controls [F(u^M_N), Σ_m c_mh φ_mh], not [F(˜u), Σ_m c_mh φ_mh]. Passing from one to the other requires an additional step using the weak continuity of F and the convergence of the approximating functionals, which is not supplied in the manuscript. This is a gap in the proof of a main convergence theorem, though it appears repairable.
minor comments (4)
  1. [Section 3.3.2, Proposition 5 and Lemma 1] Proposition 5 states 'Under the assumption 1' but the proof uses the weak-continuity assumption, which is Assumption 2; in the proof of Lemma 1, 'assumption 5' should refer to Assumption 4. These cross-referencing errors should be corrected.
  2. [Section 3.1, Corollary 1] Corollary 1 attributes to the regularization problem (13) the same optimality condition (15) as the constrained problem and writes u_N rather than u^µ_N. The stationarity condition for (13) necessarily contains the fidelity residual terms, so the corollary as stated is incorrect; it should be rederived or removed.
  3. [Section 6.1.1] The claim that continuous embedding of U into H^1_0(Ω) yields compact embedding into L^4(Ω) requires Ω to be bounded or at least of finite measure; this hypothesis should be stated explicitly in the example.
  4. [Appendix 8.3] The proof of Theorem 6 contains notation inconsistencies (δnM versus ϵ_nM, cmn versus cnm, and references to 'the true solution u' where u_N is meant). A careful proofreading pass is needed for this section.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the derivation chain is self-contained and rests on explicit assumptions, not on fitted data or self-citations.

full rationale

The paper contains no fitted parameters, no generated predictions, and no self-citation chain. Theorem 1, Theorem 2, and Theorem 6 are proved from explicit assumptions (dense span, weak continuity, uniqueness, approximation by B*_0) by standard weak-compactness and weak-lower-semicontinuity arguments; the constant choice ϵ_nM = C(∥u*∥) ε̄_nM in Lemma 1 is a bound derived from the boundedness of F on the ball of radius ∥u*∥, not a fitted value. The finite-dimensionality results are conditional on the explicitly stated sufficient condition (16), and for finite-measurement nonlinear operators Lemma 2 derives (16) by differentiating the defining representation; this is a direct verification, not a renaming of the conclusion. The paper's citations are to external prior work (e.g., Chen et al. 2021, Owhadi and Scovel 2019, Wang and Xu 2021), not to the present authors, so there is no self-referential loop. One genuine caveat is that Proposition 3 obtains the optimality condition by differentiating a Lagrangian without stating or proving a constraint qualification, so the representation (14) can fail at abnormal minimizers; however, a missing hypothesis is a correctness risk rather than a circularity, because the conclusion is not shown to be equivalent to the input by construction.

Assumptions & free parameters 1 free parameters · 7 assumptions · 0 invented entities

The framework explicitly assumes dense test functions, weak continuity, uniqueness, and an approximation property. The main hidden premise is the validity of the Lagrange multiplier rule in Proposition 3, which is asserted without a constraint qualification. The relaxation tolerance epsilon_nM is a free parameter whose theoretical choice depends on the unknown solution.

free parameters (1)
  • relaxation tolerance epsilon_nM = C(||u*||) epsilon_bar_nM (depends on unknown true solution)
    The convergence proofs for the relaxed nonlinear optimal recovery require the tolerance to be set using the norm of the unknown solution u*; no implementable selection rule is provided.
assumptions (7)
  • domain assumption Assumption 1: weak*-dense span of test functions
    Required for passing from finitely many measurements to the full equation in the limit N to infinity; stated in Section 2.2.
  • domain assumption Assumption 2: weak continuity of F
    Used in Theorems 1, 2, 6 and Corollaries 2, 3 to pass limits through the nonlinear operator; verified for finite measurement operators under Assumption 5.
  • domain assumption Assumption 3: uniqueness of u_N
    Needed to identify weak limits with the optimal recovery solution in convergence proofs; stated in Section 2.3.
  • domain assumption Assumption 4: strong approximation of B* by B*_0
    Used in the relaxed nonlinear optimal recovery to guarantee u* and u_N are feasible; automatic for point evaluations in reflexive reproducing kernel Banach spaces by Lemma 4.
  • domain assumption Assumption 5: continuity of F_phi
    Ensures weak continuity of finite measurement nonlinear operators via Lemma 3.
  • domain assumption Assumption 6: weak continuity of L
    Used in the linear-nonlinear decomposition convergence, Corollary 2.
  • ad hoc to paper Existence of Lagrange multipliers for problem (12) without constraint qualification
    Proposition 3 differentiates the Lagrangian and concludes the optimality condition u_N = sum lambda_n psi(u_N, phi_n) without stating any constraint qualification; this can fail for nonconvex equality-constrained problems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Nonlinear Optimal Recovery in Hilbert Spaces." pith.science (2026). https://pith.science/paper/PK3HTDF5

@misc{pith2026250600704,
  author       = {Pith},
  title        = {Pith review of: Nonlinear Optimal Recovery in Hilbert Spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PK3HTDF5}},
  note         = {Machine review of arXiv:2506.00704}
}
read the original abstract

This paper investigates solution strategies for nonlinear problems in Hilbert spaces, such as nonlinear partial differential equations (PDEs) in Sobolev spaces, when only finite measurements are available. We formulate this as a nonlinear optimal recovery problem, establishing its well-posedness and proving its convergence to the true solution as the number of measurements increases. However, the resulting formulation might not have a finite-dimensional solution in general. We thus present a sufficient condition for the finite dimensionality of the solution, applicable to problems with well-defined point evaluation measurements. To address the broader setting, we introduce a relaxed nonlinear optimal recovery and provide a detailed convergence analysis. An illustrative example is given to demonstrate that our formulations and theoretical findings offer a comprehensive framework for solving nonlinear problems in infinite-dimensional spaces with limited data.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 18 canonical work pages

  1. [1]

    Journal of Computational Physics 447, 110668 (2021)

    Chen, Y., Hosseini, B., Owhadi, H., Stuart, A.M.: Solving and learning nonlinear pdes with gaussian processes. Journal of Computational Physics 447, 110668 (2021)

    work page 2021

  2. [2]

    Owhadi, H., Scovel, C.: Operator-Adapted Wavelets, Fast Solvers, and Numerical Homogenization: From a Game Theoretic Approach to Numerical Approximation and Algorithm Design vol. 35. Cambridge University Press, Cambridge (2019)

    work page 2019

  3. [3]

    Solving PDEs with Incomplete Information

    Binev, P., Bonito, A., Cohen, A., Dahmen, W., Devore, R., Petrova, G.: Solving pdes with incomplete information. arXiv preprint arXiv:2301.05540 (2023)

    work page Pith review arXiv 2023

  4. [4]

    arXiv preprint arXiv:2305.04962 (2023)

    Batlle, P., Chen, Y., Hosseini, B., Owhadi, H., Stuart, A.M.: Error analy- sis of kernel/gp methods for nonlinear and parametric pdes. arXiv preprint arXiv:2305.04962 (2023)

    arXiv 2023

  5. [5]

    Nature machine intelligence 3(3), 218–229 (2021)

    Lu, L., Jin, P., Pang, G., Zhang, Z., Karniadakis, G.E.: Learning nonlinear oper- ators via deeponet based on the universal approximation theorem of operators. Nature machine intelligence 3(3), 218–229 (2021)

    work page 2021

  6. [6]

    The SMAI journal of computational mathematics 7, 121–157 (2021)

    Bhattacharya, K., Hosseini, B., Kovachki, N.B., Stuart, A.M.: Model reduction and neural networks for parametric pdes. The SMAI journal of computational mathematics 7, 121–157 (2021)

    work page 2021

  7. [7]

    Advances in Neural Information Processing Systems 35, 5601–5613 (2022)

    Seidman, J., Kissas, G., Perdikaris, P., Pappas, G.J.: Nomad: Nonlinear mani- fold decoders for operator learning. Advances in Neural Information Processing Systems 35, 5601–5613 (2022)

    work page 2022

  8. [8]

    arXiv preprint arXiv:2306.15924 (2023)

    Lanthaler, S., Stuart, A.M.: The parametric complexity of operator learning. arXiv preprint arXiv:2306.15924 (2023)

    arXiv 2023

Show all 20 references
  1. [9]

    Optimal estimation in approximation theory, 1–54 (1977)

    Micchelli, C.A., Rivlin, T.J.: A survey of optimal recovery. Optimal estimation in approximation theory, 1–54 (1977)

    1977

  2. [10]

    Wiley Interdisciplinary Reviews: Computa- tional Statistics 1(1), 93–100 (2009)

    McDonald, G.C.: Ridge regression. Wiley Interdisciplinary Reviews: Computa- tional Statistics 1(1), 93–100 (2009)

    2009

  3. [11]

    Journal of the Royal Statistical Society Series B: Statistical Methodology 58(1), 267–288 (1996)

    Tibshirani, R.: Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology 58(1), 267–288 (1996)

    1996

  4. [12]

    In: 38 Helmbold, D., Williamson, B

    Sch¨ olkopf, B., Herbrich, R., Smola, A.J.: A generalized representer theorem. In: 38 Helmbold, D., Williamson, B. (eds.) Computational Learning Theory, pp. 416–

  5. [13]

    Advances in neural information processing systems 25 (2012)

    Dinuzzo, F., Sch¨ olkopf, B.: The representer theorem for hilbert spaces: a necessary and sufficient condition. Advances in neural information processing systems 25 (2012)

    2012

  6. [14]

    Foundations of Computational Mathematics 21(4), 941–960 (2021)

    Unser, M.: A unifying representer theorem for inverse problems and machine learning. Foundations of Computational Mathematics 21(4), 941–960 (2021)

    2021

  7. [15]

    Journal of Machine Learning Research 22(225), 1–65 (2021)

    Wang, R., Xu, Y.: Representer theorems in banach spaces: Minimum norm inter- polation, regularized learning and semi-discrete inverse problems. Journal of Machine Learning Research 22(225), 1–65 (2021)

    2021

  8. [16]

    Springer, New York (2008)

    Steinwart, I., Christmann, A.: Support Vector Machines. Springer, New York (2008)

    2008

  9. [17]

    Journal of Machine Learning Research 10(12) (2009)

    Zhang, H., Xu, Y., Zhang, J.: Reproducing kernel banach spaces for machine learning. Journal of Machine Learning Research 10(12) (2009)

    2009

  10. [18]

    Journal of Machine Learning Research 25(93), 1–45 (2024)

    Wang, R., Xu, Y., Yan, M.: Sparse representer theorems for learning in repro- ducing kernel banach spaces. Journal of Machine Learning Research 25(93), 1–45 (2024)

    2024

  11. [19]

    Springer (2011) 39

    Brezis, H.: Functional Analysis, Sobolev Spaces and Partial Differential Equations. Springer (2011) 39

    2011

  12. [426]

    Springer, Berlin, Heidelberg (2001)

    2001

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.