pith. sign in

arxiv: 2606.27285 · v1 · pith:GPKUPKAJnew · submitted 2026-06-25 · 💻 cs.LG · cs.IT· math.CA· math.DS· math.IT

Recovering Governing Equations from Solution Data: Identifiability Bounds for Linear and Nonlinear ODEs

Yang Pan , Helmut B\"olcskei This is my paper

Pith reviewed 2026-06-26 05:11 UTC · model grok-4.3

classification 💻 cs.LG cs.ITmath.CAmath.DSmath.IT
keywords identifiability boundsgoverning equationsordinary differential equationsHausdorff distancesample complexitysolution datalinear and nonlinear ODEsvector fields
0
0 comments X

The pith

The Hausdorff distance between solution sets determines when two governing ODEs can be distinguished from data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces the Hausdorff distance on solution sets to compare differential equations and derives identifiability bounds that tell exactly when two distinct equations produce distinguishable trajectories. It covers linear ODEs and nonlinear ones whose vector fields are Lipschitz or Hölder continuous, then uses the metric to obtain entropy estimates and sample-complexity results. A reader cares because these bounds supply the first quantitative account of how many solution observations are required to recover the governing equation in the worst case over initial conditions.

Core claim

We introduce the Hausdorff distance on solution sets as the natural metric for comparing differential equations because it captures the worst-case separation over all admissible initial conditions. Using this metric we establish identifiability bounds for linear ODEs and for nonlinear ODEs with Lipschitz or Hölder-continuous vector fields, characterizing precisely when two distinct equations can be told apart from solution data. The same metric yields metric-entropy estimates for the relevant classes and produces sample-complexity bounds that quantify the number of solution observations needed to recover the governing equation.

What carries the argument

The Hausdorff distance on solution sets, which quantifies the largest separation between any pair of trajectories generated by two different equations over all possible initial conditions.

If this is right

  • For linear ODEs the identifiability threshold is controlled by the separation of their coefficient matrices in the induced Hausdorff metric.
  • For nonlinear ODEs with Lipschitz vector fields the same threshold depends on the Lipschitz constant and the diameter of the domain of initial conditions.
  • The derived sample-complexity bounds scale with the metric entropy of the ODE class, giving explicit rates for both linear and nonlinear families.
  • Once the Hausdorff distance exceeds the identifiability threshold, finitely many solution trajectories suffice to certify which equation generated the data.

Where Pith is reading between the lines

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

  • The same metric could be used to compare the stability of learned versus true equations under small perturbations of initial conditions.
  • Sample-complexity results might guide the minimal number of experiments needed when designing data-collection protocols for physical systems.
  • Extensions to time-varying or stochastic ODEs would require only a suitable enlargement of the solution-set metric.

Load-bearing premise

The Hausdorff distance on solution sets is the right metric because it encodes the worst-case separation over all admissible initial conditions.

What would settle it

A pair of linear ODEs whose solution sets have positive Hausdorff distance yet produce identical trajectories for every initial condition in a dense set, or a pair with zero Hausdorff distance that can still be distinguished from finitely many observed solutions.

Figures

Figures reproduced from arXiv: 2606.27285 by Helmut B\"olcskei, Yang Pan.

Figure 1
Figure 1. Figure 1: Lipschitz functions f and ˜f We have [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Red points represent the randomly generated samples. Yellow and blue [PITH_FULL_IMAGE:figures/full_fig_p025_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Red points represent the randomly generated samples. Yellow and blue [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗
read the original abstract

Learning governing equations from observed solution data is a fundamental challenge in scientific machine learning \cite{bruntonDiscoveringGoverningEquations2016,kovachkiNeuralOperatorLearning2023,longPDENetLearningPDEs2018,rudyDatadrivenDiscoveryPartial2017,raonicConvolutionalNeuralOperators2023}, yet the theoretical conditions under which a ground-truth ODE can be uniquely and stably identified from multiple solution observations remain largely undeveloped, and no quantitative analysis of the sample complexity of such learning tasks exists in the literature. To address this gap, we introduce the Hausdorff distance on solution sets as the natural metric for comparing differential equations, since it captures the worst-case separation between two equations over all admissible initial conditions and thus encodes the minimax structure of the identification problem. We establish identifiability bounds for governing ODEs across a wide class of structure equations--ranging from linear ODEs to nonlinear classes with Lipschitz (H\"older)-continuous vector fields--characterizing precisely when two distinct equations can be distinguished from solution data. Using this metric, we derive metric entropy estimates for the relevant ODE classes and analyze sample complexity bounds, quantifying how many solution observations are needed to reliably recover the governing equation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The paper introduces the Hausdorff distance on solution sets as the metric for comparing ODEs, since it encodes the worst-case separation over admissible initial conditions. It derives identifiability bounds for linear ODEs and for nonlinear ODEs whose vector fields are Lipschitz or Hölder continuous, characterizes when two distinct equations produce distinguishable solution trajectories, obtains metric-entropy estimates for the resulting function classes, and supplies sample-complexity bounds on the number of solution observations needed to recover the governing equation.

Significance. If the derivations hold, the work supplies the first quantitative identifiability and sample-complexity theory for data-driven recovery of ODEs, directly addressing an acknowledged gap in the scientific machine-learning literature. The Hausdorff construction yields a clean minimax formulation, the extension from linear to Hölder classes is technically substantive, and the entropy and sample-complexity results are of immediate practical value. The absence of free parameters or ad-hoc constants in the stated program is a strength.

minor comments (3)
  1. [§2] §2 (or wherever the Hausdorff metric is first defined): the precise statement of the admissible initial-condition set and the time horizon should be stated explicitly before the metric is introduced, to make the subsequent entropy calculations fully reproducible.
  2. [Abstract] The abstract claims that 'no quantitative analysis of the sample complexity … exists in the literature.' A short paragraph contrasting the new bounds with existing results on parameter identifiability for linear systems or on Lipschitz ODEs would strengthen this claim.
  3. [Theorem 3.2] Theorem statements that invoke covering numbers should include a brief reminder of the dependence on the Hölder exponent α and the dimension d, even if the full proof is deferred.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary and recommendation of minor revision. The report accurately reflects the paper's contributions on identifiability bounds, Hausdorff distance, metric entropy, and sample complexity for ODE recovery.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper introduces the Hausdorff distance on solution sets as a metric for ODE comparison and derives identifiability bounds, metric entropy, and sample complexity results directly from the properties of this metric space for linear and Lipschitz/Hölder classes. This is a self-contained mathematical construction: the distance is defined to encode worst-case separation over initial conditions, and the bounds follow from standard covering-number arguments in the induced metric. No load-bearing step reduces to a self-definition, fitted parameter renamed as prediction, or self-citation chain. External citations (e.g., Brunton et al.) are contextual and not invoked to justify the core uniqueness or metric properties. The derivation stands independently against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated or can be extracted.

pith-pipeline@v0.9.1-grok · 5759 in / 1033 out tokens · 25426 ms · 2026-06-26T05:11:31.343349+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

36 extracted references · 5 canonical work pages · 1 internal anchor

  1. [1]

    Lecture Notes Mathematics of Information

    Helmut Bolcskei. “Lecture Notes Mathematics of Information”. In: (2020)

    2020

  2. [2]

    Discovering governing equations from data: Sparse identification of nonlinear dynamical systems

    Steven L. Brunton, Joshua L. Proctor, and J. Nathan Kutz. “Discovering Gov- erning Equations from Data: Sparse Identification of Nonlinear Dynamical Sys- tems”. In:Proceedings of the National Academy of Sciences113.15 (Apr. 2016), pp. 3932–3937.issn: 0027-8424, 1091-6490.doi:10.1073/pnas.1517384113. arXiv:1509.03580 [math]

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1073/pnas.1517384113 2016

  3. [3]

    Finite sample properties of system identi- fication methods

    Marco C Campi and Erik Weyer. “Finite sample properties of system identi- fication methods”. In:IEEE Transactions on Automatic Control47.8 (2002), pp. 1329–1334

    2002

  4. [4]

    Physics-informed learning of governing equations from scarce data

    Zhao Chen, Yang Liu, and Hao Sun. “Physics-informed learning of governing equations from scarce data”. In:Nature communications12.1 (2021), p. 6136

    2021

  5. [5]

    Identifiability of linear and nonlinear dynamical systems

    M Grewal and Keith Glover. “Identifiability of linear and nonlinear dynamical systems”. In:IEEE Transactions on automatic control21.6 (2003), pp. 833– 837

    2003

  6. [6]

    Martin Holler and Erion Morina.On Uniqueness in Structured Model Learning. Oct. 2024. arXiv:2410.22009

    arXiv 2024

  7. [7]

    Metric entropy limits on recurrent neural network learning of linear dynamical systems

    Clemens Hutter, Recep G¨ ul, and Helmut B¨ olcskei. “Metric entropy limits on recurrent neural network learning of linear dynamical systems”. In:Applied and Computational Harmonic Analysis59 (2022), pp. 198–223

    2022

  8. [8]

    Springer, 2006

    Victor Isakov.Inverse problems for partial differential equations. Springer, 2006

    2006

  9. [9]

    Entropy per Unit Time as a Metric Invariant of Automor- phism

    AN Kolgomorov. “Entropy per Unit Time as a Metric Invariant of Automor- phism”. In:Doklady of Russian Academy of Sciences: Moscow, Russia124 (1959), pp. 754–755

    1959

  10. [10]

    Neural Operator: Learning Maps Between Function Spaces With Applications to PDEs

    Nikola Kovachki et al. “Neural Operator: Learning Maps Between Function Spaces With Applications to PDEs”. In:Journal of Machine Learning Re- search24.89 (2023), pp. 1–97.issn: 1533-7928

    2023

  11. [11]

    Neural operator: Learning maps between function spaces with applications to pdes

    Nikola Kovachki et al. “Neural operator: Learning maps between function spaces with applications to pdes”. In:Journal of Machine Learning Research 24.89 (2023), pp. 1–97

    2023

  12. [12]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li et al. “Fourier neural operator for parametric partial differential equations”. In:arXiv preprint arXiv:2010.08895(2020)

    Pith/arXiv arXiv 2010

  13. [13]

    System identification

    Lennart Ljung. “System identification”. In:Signal analysis and prediction. Springer, 1998, pp. 163–173

    1998

  14. [14]

    On global identifiability for arbitrary model parametrizations

    Lennart Ljung and Torkel Glad. “On global identifiability for arbitrary model parametrizations”. In:automatica30.2 (1994), pp. 265–276

    1994

  15. [15]

    PDE-Net: Learning PDEs from Data

    Zichao Long et al. “PDE-Net: Learning PDEs from Data”. In:Proceedings of the 35th International Conference on Machine Learning. PMLR, July 2018, pp. 3208–3216. 55

    2018

  16. [16]

    Extrapolation and learning equa- tions

    Georg Martius and Christoph H Lampert. “Extrapolation and learning equa- tions”. In:arXiv preprint arXiv:1610.02995(2016)

    Pith/arXiv arXiv 2016

  17. [17]

    EMS press Berlin, 2023

    Richard Nickl.Bayesian non-linear statistical inverse problems. EMS press Berlin, 2023

    2023

  18. [18]

    Non-asymptotic identification of lti sys- tems from a single trajectory

    Samet Oymak and Necmiye Ozay. “Non-asymptotic identification of lti sys- tems from a single trajectory”. In:2019 American control conference (ACC). IEEE. 2019, pp. 5655–5661

    2019

  19. [19]

    Baburao G Pachpatte.Integral and finite difference inequalities and applica- tions. Vol. 205. Elsevier, 2006

    2006

  20. [20]

    Metric-Entropy Limits on Nonlinear Dynamical System Learning

    Yang Pan, Clemens Hutter, and Helmut B¨ olcskei. “Metric-Entropy Limits on Nonlinear Dynamical System Learning”. In:arXiv preprint arXiv:2407.01250 (2024)

    arXiv 2024

  21. [21]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations”. In:Journal of Computational physics378 (2019), pp. 686–707

    2019

  22. [22]

    Convolutional Neural Operators for Robust and Ac- curate Learning of PDEs

    Bogdan Raonic et al. “Convolutional Neural Operators for Robust and Ac- curate Learning of PDEs”. In:Advances in Neural Information Processing Systems36 (Dec. 2023), pp. 77187–77200

    2023

  23. [23]

    Convolutional neural operators for robust and accurate learning of PDEs

    Bogdan Raonic et al. “Convolutional neural operators for robust and accurate learning of PDEs”. In:Advances in Neural Information Processing Systems36 (2024)

    2024

  24. [24]

    ESPRIT-estimation of signal parameters via rotational invariance techniques

    Richard Roy and Thomas Kailath. “ESPRIT-estimation of signal parameters via rotational invariance techniques”. In:IEEE Transactions on acoustics, speech, and signal processing37.7 (2002), pp. 984–995

    2002

  25. [25]

    ScienceAdvances9,eadf8537

    Samuel H. Rudy et al. “Data-Driven Discovery of Partial Differential Equa- tions”. In:Science Advances3.4 (Apr. 2017), e1602614.doi:10.1126/sciadv. 1602614

    work page doi:10.1126/sciadv 2017

  26. [26]

    Learning equa- tions for extrapolation and control

    Subham Sahoo, Christoph Lampert, and Georg Martius. “Learning equa- tions for extrapolation and control”. In:International Conference on Machine Learning. Pmlr. 2018, pp. 4442–4450

    2018

  27. [27]

    Enhancing and Adversarial: Improve ASR with Speaker Labels

    Philipp Scholl et al. “The Uniqueness Problem of Physical Law Learning”. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Rhodes Island, Greece: IEEE, June 2023, pp. 1– 5.isbn: 978-1-72816-327-7.doi:10.1109/ICASSP49357.2023.10095017

    work page doi:10.1109/icassp49357.2023.10095017 2023

  28. [28]

    Philipp Scholl et al.Well-Definedness of Physical Law Learning: The Unique- ness Problem. Jan. 2023. arXiv:2210.08342 [math-ph]

    arXiv 2023

  29. [29]

    On the notion of entropy of a dynamical system

    Yakov G Sinai. “On the notion of entropy of a dynamical system”. In:Doklady of Russian Academy of Sciences. Vol. 124. 3. 1959, pp. 768–771

    1959

  30. [30]

    Inverse problems: a Bayesian perspective

    Andrew M Stuart. “Inverse problems: a Bayesian perspective”. In:Acta nu- merica19 (2010), pp. 451–559. 56 [31]System identification - Wikipedia — en.wikipedia.org. Aug. 2025.url:https: //en.wikipedia.org/wiki/System_identification

    2010

  31. [31]

    Akoma-Ntoso for Legal Documents

    V. M. Tikhomirov. “ε-Entropy andε-Capacity of Sets In Functional Spaces”. In:Selected Works of A. N. Kolmogorov: Volume III: Information Theory and the Theory of Algorithms. Ed. by A. N. Shiryayev. Dordrecht: Springer Netherlands, 1993, pp. 86–170.isbn: 978-94-017-2973-4.doi:10.1007/978- 94-017-2973-4_7

    work page doi:10.1007/978- 1993

  32. [32]

    AI Feynman: A physics-inspired method for symbolic regression

    Silviu-Marian Udrescu and Max Tegmark. “AI Feynman: A physics-inspired method for symbolic regression”. In:Science advances6.16 (2020), eaay2631

    2020

  33. [33]

    Martin J Wainwright.High-dimensional statistics: A non-asymptotic view- point. Vol. 48. Cambridge university press, 2019

    2019

  34. [34]

    Wolfgang Walter.Ordinary Differential Equations. Vol. 182. Graduate Texts in Mathematics. New York, NY: Springer, 1998.doi:10.1007/978-1-4612- 0601-9

    work page doi:10.1007/978-1-4612- 1998

  35. [35]

    On the metric complexity of causal linear systems: Estimates ofε- entropy andε-dimension

    G Zames. “On the metric complexity of causal linear systems: Estimates ofε- entropy andε-dimension”. In:1977 IEEE Conference on Decision and Control including the 16th Symposium on Adaptive Processes and A Special Symposium on Fuzzy Set Theory and Applications. IEEE. 1977, pp. 807–810

    1977

  36. [36]

    Classes of ODE solutions: smoothness, cov- ering numbers, implications for noisy function fitting, and the curse of smooth- ness phenomenon

    Ying Zhu and Mozhgan Mirzaei. “Classes of ODE solutions: smoothness, cov- ering numbers, implications for noisy function fitting, and the curse of smooth- ness phenomenon”. In:arXiv preprint arXiv:2011.11371(2020). 57

    arXiv 2011