Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Counterfactual Operator Relevance for PDE Discovery: Screening, Pruning, and Identifiability

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Counterfactual deletion of a PDE term, not residual magnitude, separates functional necessity from mere fit; under sparsity and incoherence, exact support recovery follows.

desk verdict The core recovery theorem is self-contradictory (it sets the measurements to zero and then proves a nonzero sparse solution), and the body doesn't deliver the abstract's six theorems, though the orthogonality examples make a valid point about residual-based screening. read the letter →

arxiv 2506.20181 v2 pith:WDNEGTEI submitted 2025-06-25 cs.LG cs.NAmath.NA

Ronald Katende This is my paper · ORCID
classification cs.LGcs.NAmath.NA MSC 35R3035Q9368T0762M4535A0265M32
keywords PDEdiscoverycounterfactualoperatorinterventionscausalrelevanceidentifiabilitysparserecoveryphysics-informedneuralnetworksmutualcoherenceSensitivityIndex
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 argues that a PDE term can lower the residual of a learned model without being functionally necessary, and that the correct test of necessity is counterfactual: delete or perturb the candidate operator, re-solve the dynamics, and measure how far the trajectory moves. The theory is organized around six reusable results, including a residual–counterfactual gap theorem, a certified decision rule for relevance and abstention under surrogate error, an aliasing theorem tied to the null space of the operator-evaluation design, a constraint-manifold obstruction, a pruning-consistency theorem, and an observable-level adjoint theorem. It further claims that when the operator-evaluation matrix satisfies a restricted isometry or mutual-coherence bound, the sparsity-regularized residual program recovers the true operator support exactly. Synthetic PDEs and public climate and ocean fields are used to show where residual-only methods mislead, with real-data results framed as operator-surrogate diagnostics rather than unconditional physical laws.

What carries the argument

The central object is the counterfactual PDE: the system obtained by setting a candidate coefficient αj to zero (or perturbing it) and re-solving for a new trajectory ucf. Causal relevance is quantified by the L2 deviation ||u−ucf||, which the paper treats as the ground-truth measure of functional necessity. The companion machinery for support recovery is the operator-evaluation design matrix Aij = Tj[u](xi,ti) together with the ℓ1-regularized program min ||Aα||₂² + λ||α||₁, whose exact-recovery guarantee is imported from compressive-sensing theory under restricted isometry or mutual coherence. The paper also introduces the Causal Sensitivity Index (CSI), a normalized residual-based score used as a screening proxy; the paper shows CSI can be fooled by orthogonal terms and must be checked against counterfactual deviation.

What would settle it

Numerically solve a Fisher–KPP equation $\partial_t u = D\partial_{xx}u + u - u^2$ on a periodic domain, build the library $\{\partial_{xx}u, u, u^2\}$, form $A$ from a sampled trajectory, and compute the mutual coherence $\mu$. With $s=3$ the theorem requires $\mu < 1/(2s-1) = 1/5$; if a dataset satisfying that bound still fails to recover $\{\partial_{xx}u, u, u^2\}$ from $\min\|A\alpha\|_2^2 + \lambda\|\alpha\|_1$, the identifiability claim is false, and if $\mu \geq 1/5$ the theorem's conditions simply do not apply.

Watch

Extended reading notes

Core claim

The paper's central claim is that operator relevance should be defined by counterfactual trajectory deviation, not by residual contribution. An operator Tj is functionally necessary if setting its coefficient αj to zero and re-solving the PDE moves the solution by more than a specified tolerance; terms that pass residual checks but fail this counterfactual test are classified as non-identifiable or spurious. The paper gives explicit examples where a term contributes a large residual norm yet zero trajectory deviation, and where a residual-based sensitivity index assigns high influence to a structurally inert term. On the recovery side, the paper claims that for an s-sparse true PDE and a design matrix A with entries Tj[u](xi,ti) satisfying restricted isometry of order 2s or mutual coherence below 1/(2s−1), the minimizer of ||Aα||₂² + λ||α||₁ recovers the exact support. Real-data outputs are scoped as surrogate diagnostics relative to a specified library, experiment class, norm, and tolerance.

Load-bearing premise

The recovery guarantee depends on the candidate operators being nearly uncorrelated when evaluated on the data, and on the true equation having only a few terms; for smooth PDEs these evaluations are usually strongly correlated, so the condition is unverifiable and often false.

Editorial extensions

If this is right

  • Library screening becomes a causal operation: after counterfactual pruning, the retained operators are exactly those whose deletion changes the trajectory, not merely those that improve the loss.
  • Non-identifiability becomes diagnosable: operators lying in the null space of the operator-evaluation design, or vanishing on invariant constraint classes, are flagged as experiment-dependent rather than falsely recovered.
  • Under the restricted isometry or mutual-coherence bound, the sparsity-regularized residual program provably retrieves the true sparse support, so screening followed by counterfactual deletion recovers the functionally relevant terms.
  • Relevance testing extends to observables via the adjoint theorem, so scientific quantities of interest can be tested without always reconstructing the full state.
  • Real-data discoveries carry an explicit scope: the climate and ocean results are operator-surrogate diagnostics relative to a chosen library, experiment class, norm, and tolerance, not unconditional physical laws.

Reading between the lines

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

  • The null-space characterization of aliasing suggests an active-measurement scheme: adding observations at times and locations that reduce the mutual coherence of A should resolve non-identifiability; the paper does not develop this selection rule.
  • The residual–counterfactual gap theorem implies a cheap two-stage protocol—rank all candidates with the CSI proxy, then run counterfactual deletion only on the top-ranked terms—but the paper does not analyze the statistical error of this sequential decision rule.
  • The certified decision theorem could be inverted into a principled choice of tolerance: given a measured surrogate error, the relevance threshold ε could be set so that abstentions occur exactly when the surrogate cannot distinguish deletion effects; the paper leaves ε as a user input.
  • The constraint-manifold result transfers to reduced-order modeling: if observed trajectories lie on a low-dimensional invariant manifold, operators vanishing on that manifold are invisible, so identifying them would require manifold-breaking perturbations or richer data.
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

4 major / 5 minor

Summary. The paper proposes a framework for PDE discovery that distinguishes residual-based term selection from counterfactual operator relevance. It formalizes a PDE as a structural causal model, defines causal sensitivity indices and counterfactual deviations, and claims a theoretical guarantee (Theorem 3.9, restated as Theorem 4.1) that l1-regularized residual minimization exactly recovers the support of a sparse true operator library under RIP or mutual coherence. It also presents a residual error bound, an aliasing/orthogonality analysis, and experiments on synthetic and real-world data (tumor imaging, ERA5 climate fields, HYCOM ocean currents). The abstract states results as operator-surrogate diagnostics, while the full-text abstract and conclusion claim exact recovery and superiority over PINNs and DeepONets.

Significance. If the central recovery theorem were valid, the paper would provide a useful bridge between sparse PDE discovery and causal inference, and the explicit orthogonality examples would be a helpful caution. The paper honestly identifies cases where residual-based influence measures fail (Section 3.6, Examples 3.14-3.16). However, the main theoretical result is invalid because the homogeneous measurement model makes the zero vector a trivial global solution, so the claimed foundational guarantee is not established. The remaining diagnostics are heuristic, and the real-data evaluation does not test predictive or counterfactual validity. Credit is due for the paper's clear statement of the distinction it targets and for the instructive counterexamples; those parts are not the problem.

major comments (4)
  1. [Section 3.5, Theorem 3.9; Section 4, Theorem 4.1] The recovery program is internally inconsistent. In the proof of Theorem 3.9, the measurement vector is defined as y := Aalpha*, but the governing equation sum_j alpha*_j T_j[u] = 0 gives y = 0. The minimization objective displayed in the theorem statement is ||Aalpha||_2^2 + lambda||alpha||_1 (with the data term omitted), whose value at alpha = 0 is 0, whereas at alpha = alpha* it is lambda||alpha*||_1 > 0 for every lambda > 0. Hence the minimizer is alpha = 0, not alpha*. Moreover, with rank(A) = m, the equation Aalpha* = 0 forces alpha* = 0, contradicting the premise ||alpha*||_0 = s >= 1. The RIP and mutual coherence conditions do not repair this because they cannot change the fact that the objective is minimized at zero. Theorem 4.1 repeats the same error. Since the abstract and Section 4 advertise exact support recovery as the core theoretical contribution, this invalidates the central claim.
  2. [Section 3.3, Theorem 3.5] The residual error bound assumes that the weak form a(w,v) = <N[w;gamma'], v> is continuous and coercive on H^1(Omega) with coercivity constant alpha, but no coercivity or continuity is verified for any operator considered in the paper, including the nonlinear reaction-diffusion and Navier-Stokes examples. For nonlinear N, the pairing <N[w;gamma'], v> is not the standard H^{-1}-H^1 product, and for neural surrogates v_theta it is unclear in what space the residual should be interpreted. Additionally, the bound is stated in H^1(Omega), whereas the quantities actually used in the framework, such as the counterfactual deviation in Definition 3.6 and the delta_j in Section 3.8, are L^2(Omega x (0,T)) norms of space-time deviations. The theorem therefore does not establish the claimed rigorous link between residual norm and the solution deviations used throughout the paper.
  3. [Section 3.7 and Section 3.8 (Definitions 3.7, 3.12; Algorithm 1)] The counterfactual relevance diagnostic is self-referential. The residual influence metric C_j[v_theta] and the CSI_j are computed by removing a term from the residual of the same surrogate u_theta that was trained to minimize that residual, and the counterfactual deviation delta_j is obtained by re-optimizing the same loss with alpha_j = 0. Deleting a term in this way does not implement the structural intervention of solving the counterfactual PDE N[ucf;gamma'] = 0 unless one proves that the retrained surrogate converges to the true counterfactual solution, which would require a result of the type of Theorem 3.5 under hypotheses that are not verified. The paper's own Example 3.14 shows that residual-based CSI can misattribute relevance to Delta u; the proposed counterfactual re-optimization is not shown to avoid that failure, because the retrained ucf_theta is again a minimizer of a residual-based loss.
  4. [Section 6, Tables 3 and 6] The real-data evaluation reports only in-sample reconstruction metrics (MAE, RMSE, support counts) for the tumor imaging, ERA5 climate, and HYCOM ocean datasets. No held-out spatial or temporal prediction is evaluated, no uncertainty quantification is provided, and the 'true' causal support for these real datasets is asserted by the authors rather than independently known. Consequently, the 'Support Accuracy 100%' entries in Table 6 and the conclusion's claim that the method recovers the true generative law are not supported by the experimental design. The abstract's own qualification that real-data results are 'operator-surrogate diagnostics, not as unconditional recovery of physical laws' is more accurate than the claims made in the body of the paper.
minor comments (5)
  1. [Title and abstract] The full-text title ('Causal Operator Discovery in Partial Differential Equations via Counterfactual Physics-Informed Neural Networks') differs from the arXiv title ('Counterfactual Operator Relevance for PDE Discovery: Screening, Pruning, and Identifiability'), and the two abstracts present different levels of claim; these should be harmonized.
  2. [Section 5.1] The text references 'Eq. XX' as a placeholder for the total loss; the equation number should be filled in.
  3. [Section 3.5, proof of Theorem 3.9] The theorem statement defines A_ij := T_j[u](x_i,t_i), but the proof redefines A_ij := T_j[phi_i](x_i,t_i) without ever specifying phi_i; this is confusing and should be corrected.
  4. [Figures] The text refers to Figure 2a, Figure 2b, and Figure 3 with descriptions, but no figures appear in the manuscript; the figure captions and in-text references are inconsistent.
  5. [Table 6] The 'Sparsity (Recovered/True)' column contains entries such as '6/3' and '7/4', which appear to be counts rather than ratios; the column heading and the reported values should be clarified.

Circularity Check

1 steps flagged · score 8.0 of 10

The central support-recovery theorem (Thm 3.9 / Thm 4.1) is self-defeating: its own premise makes y=Aα*=0, so the ℓ1-regularized minimizer is α=0, not the claimed α*.

  1. self definitional [Theorem 3.9, Section 3.5 (Operator Identifiability via Sparse Recovery); restated as Theorem 4.1, Section 4]
    "Let y := Aα⋆ ∈ RN denote the residual vector. The recovery program is ˆα := arg min α∈Rm ∥Aα − y∥2 2 + λ∥α∥1. ... Since u† satisfies the PDE exactly, we have y := Aα† = 0. ... L(α) := ∥Aα∥2 2 + λ∥α∥1, and study its minimizer."

    The theorem defines the measurement vector as y = Aα*, but its own premise Σ α*_j T_j[u] = 0 forces every entry y_i = Σ α*_j T_j[u](x_i,t_i) = 0. So the recovery objective ||Aα − y||^2 + λ||α||_1 is, by construction, identical to ||Aα||^2 + λ||α||_1. At α = 0 this objective is 0; at α = α* it equals λ||α*||_1 > 0 for any λ > 0, and ||Aα*||^2 = 0, so even the quadratic term cannot distinguish α* from 0. Under the theorem's additional assumption rank(A) = m, the equation Aα = 0 has only the solution α = 0. Hence the minimizer is α = 0, not α*. The claimed exact-support recovery thus reduces, by the theorem's own definitions, to asserting supp(0) = supp(α*), which is false for nonzero sparse α*.

full rationale

The paper's main theoretical pillar, the operator-identifiability guarantee, collapses under its own definitions. Both Theorem 3.9 and Theorem 4.1 define the design matrix A from evaluations of the candidate operators on the true solution and then set y = Aα*, where α* is the true sparse coefficient vector. But the governing equation Σ α*_j T_j[u] = 0 means Aα* = 0 row by row. The proof of Theorem 4.1 explicitly acknowledges 'we have y := Aα† = 0', then studies the minimizer of ||Aα||^2 + λ||α||_1. That minimizer is zero, not α*, since the objective is zero at α=0 and strictly positive at α=α*. Therefore the central claim that the causal PINN loss recovers supp(α*) is not merely unproved; it is contradicted by the theorem's own construction. This is a load-bearing reduction-by-definition, not a mere technical gap, because the abstract and introduction present this recovery guarantee as the rigorous basis for separating counterfactual operator relevance from residual usefulness. I also considered the surrogate-based counterfactual diagnostics (CSI, δ_j), which are indeed self-referential in that the same fitted PINN generates both factual and counterfactual states; however, the paper explicitly frames real-data results as 'operator-surrogate diagnostics' and not unconditional recovery, and this approximate-computation issue is not a formal circular derivation of the same kind. The fatal circular/degenerate step is the identifiability theorem, which forces score 8: the central result is, in effect, determined by definition of y = Aα* = 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The framework's conclusions rest on a set of unverified mathematical assumptions about the PDE operator (coercivity, invertibility of the linearized map), about the operator library (linear independence, RIP/coherence), and about the data-generating process (exact sparse PDE, noiseless residual). The thresholds and regularization weights are free parameters that directly determine the reported causal support. No new physical entities are introduced.

free parameters (5)
  • relevance threshold epsilon (ε) = user-defined; not reported
    Definition 3.6 and the real-data sections declare Tj causally relevant only if ∥u − ucf_j∥ ≥ ε; the value of ε is chosen by the user and directly controls which operators are deemed relevant.
  • counterfactual deviation threshold eta (η) = user-defined; not reported
    In Section 6.2, an operator is deemed relevant only if both CSI_j > ε and δ_j > η; η is chosen post hoc and affects the reported support.
  • regularization weights lambda_r, lambda_s = (1.0, 0.01)
    These hyperparameters control the balance between residual fit, data fidelity, and sparsity; the results depend on them but no sensitivity analysis is given.
  • stability constant delta = δ > 0, value not specified
    Introduced in the CSI definition to avoid division by zero; its value is not specified and affects CSI magnitudes.
  • coercivity constant alpha = not estimated
    The residual error bound Theorem 3.5 depends on the coercivity constant α of the weak form, which is assumed but never computed or validated for the examples.
assumptions (6)
  • domain assumption The weak form a(w,v) = ⟨N[w;γ′], v⟩ is continuous and coercive on H1 with constant α > 0
    Invoked in Theorem 3.5 to bound the counterfactual surrogate error; fails for many PDEs such as first-order hyperbolic systems.
  • domain assumption The Fréchet derivative ∂R/∂uθ is invertible at the solution
    Invoked in Eq. (5) for the causal derivative formula and in the adjoint derivation; invertibility is not guaranteed for degenerate or constrained PDEs.
  • domain assumption The candidate operator library {Tj} is linearly independent on the relevant function space
    Used in Theorem 3.9 and Prop 3.8; linear independence of differential operators on the solution's trajectory is hard to verify and often fails for overcomplete libraries.
  • domain assumption The true PDE is sparse with coefficient vector α* satisfying ∥α*∥_0 = s << m
    This sparsity assumption is the foundation of the ℓ1 recovery guarantees and is taken as given without evidence.
  • domain assumption The design matrix A = [Tj[u](x_i,t_i)] satisfies the RIP of order 2s or mutual coherence µ < 1/(2s−1)
    Needed for Theorem 3.9 and Theorem 4.1; unverifiable and unlikely for smooth differential operators.
  • standard math The neural surrogate uθ lies in a dense, differentiable subspace of H1 and can represent the true solution
    Standard universal approximation assumption, stated in Section 4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Counterfactual Operator Relevance for PDE Discovery: Screening, Pruning, and Identifiability." pith.science (2026). https://pith.science/paper/WDNEGTEI

@misc{pith2026250620181,
  author       = {Pith},
  title        = {Pith review of: Counterfactual Operator Relevance for PDE Discovery: Screening, Pruning, and Identifiability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WDNEGTEI}},
  note         = {Machine review of arXiv:2506.20181}
}
read the original abstract

We study operator relevance in data-driven partial differential equation (PDE) discovery. Sparse residual methods can select terms that improve residual fit, but residual contribution is not the same as functional necessity. We formalize this distinction through counterfactual operator interventions, where a candidate term is deleted or perturbed and the factual and intervened trajectories, or observables, are compared. The resulting theory gives six reusable results. A residual--counterfactual gap theorem shows that deletion effects are governed by the inverse linearized PDE map, not by residual magnitude alone. A certified decision theorem gives error margins for relevance, irrelevance, and abstention under neural or numerical surrogate error. An aliasing theorem characterizes experiment-dependent non-identifiability through the null space of the operator-evaluation design. A constraint-manifold theorem shows that operators vanishing on invariant constraint classes cannot be identified from trajectories restricted to those classes. A pruning-consistency theorem proves that sparse screening followed by counterfactual deletion recovers the functionally relevant support under a recall and margin condition. An observable-level adjoint theorem extends relevance testing from full-state deviations to scientific quantities of interest. Validation experiments test these mechanisms on synthetic PDEs with known support and on public geophysical fields from atmospheric reanalysis and NOAA OISST. The real-data results are reported as operator-surrogate diagnostics, not as unconditional recovery of physical laws. The framework provides a rigorous diagnostic layer for distinguishing residual usefulness from counterfactual operator relevance within a specified library, experiment class, norm, and tolerance.

Figures

Figures reproduced from arXiv: 2506.20181 by the authors.

Figure 3
Figure 3. Benchmark Comparison across Discovery Frameworks. [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figure 4
Figure 4. Causal PDE recovery from medical imaging data using Causal PINNs. The approach [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗
Figure 5
Figure 5. Discovered causal operators and predicted field over Uganda using sparse ERA5 data. [PITH_FULL_IMAGE:figures/full_fig_p025_5.png] view at source ↗
Figures from the paper (1 more)
Figure 6
Figure 6. Figure 6: Casual Discovery of Ocean Flow Dynamics in the Mozambique Channel [PITH_FULL_IMAGE:figures/full_fig_p026_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 29 canonical work pages

  1. [1]

    The Mathematical Theory of Finite Element Methods

    Susanne C Brenner and L Ridgway Scott. The Mathematical Theory of Finite Element Methods. Springer, 2008

    work page 2008

  2. [2]

    Chebyshev and Fourier Spectral Methods

    John P Boyd. Chebyshev and Fourier Spectral Methods . Dover Publications, 2001

    work page 2001

  3. [3]

    Runge–kutta discontinuous galerkin methods for convection-dominated problems

    Bernardo Cockburn and Chi-Wang Shu. Runge–kutta discontinuous galerkin methods for convection-dominated problems. Journal of Scientific Computing , 16(3):173–261, 2001

    work page 2001

  4. [4]

    A Multigrid Tutorial

    William L Briggs, Van Emden Henson, and Steve F McCormick. A Multigrid Tutorial . SIAM, 2000

    work page 2000

  5. [5]

    Convergence and error analysis of physics-informed neural networks

    Shirong Dong, George Em Karniadakis, and Zhicheng Huang. Convergence and error analysis of physics-informed neural networks. Journal of Computational Physics , 495:112527, 2023

    work page 2023

  6. [6]

    Numerical Solution of Time-dependent Advection- diffusion-reaction Equations

    Willem Hundsdorfer and Jan G Verwer. Numerical Solution of Time-dependent Advection- diffusion-reaction Equations. Springer, 2003

    work page 2003

  7. [7]

    A matrix-free finite element method for solid mechanics

    Jed Brown, Matthew G Knepley, and Barry F Smith. A matrix-free finite element method for solid mechanics. SIAM Journal on Scientific Computing , 34(1):A479–A504, 2012

    work page 2012

  8. [8]

    Automated Solution of Differential Equa- tions by the Finite Element Method

    Anders Logg, Kent-Andre Mardal, and Garth Wells. Automated Solution of Differential Equa- tions by the Finite Element Method . Springer, 2012

    work page 2012

Show all 31 references
  1. [9]

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

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear pdes. Journal of Computational Physics , 378:686–707, 2019

    2019

  2. [10]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew M Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. ICLR, 2021. 28

    2021

  3. [11]

    Neural operators: Learning maps between function spaces

    Nikola B Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew M Stuart, and Anima Anandkumar. Neural operators: Learning maps between function spaces. Nature Machine Intelligence , 5:684–696, 2023

    2023

  4. [12]

    Learning nonlinear operators via deep- onet based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, and George Em Karniadakis. Learning nonlinear operators via deep- onet based on the universal approximation theorem of operators. Nature Machine Intelligence , 3(3):218–229, 2021

    2021

  5. [13]

    Marrying causal representation learning with dynamical systems

    Dingling Yao, Caroline M¨ uller, Francesco Locatello, and Bernhard Sch¨ olkopf. Marrying causal representation learning with dynamical systems. In Advances in Neural Information Processing Systems, volume 37, pages 16425–16438, 2024

    2024

  6. [14]

    Robust discovery of partial differential equations in complex situations

    Dong Zhang, Xuedong Lu, Paris Perdikaris, and George Em Karniadakis. Robust discovery of partial differential equations in complex situations. Physical Review Research, 3(3):033270, 2021

    2021

  7. [15]

    A causal graph ode for continuous-time treatment effect modeling

    Zijie Huang, Yizhou Sun, Wei Wang, and Shen Gao. A causal graph ode for continuous-time treatment effect modeling. In Proceedings of The Web Conference, page 11, 2024

    2024

  8. [16]

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems

    Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. PNAS, 113(15):3932–3937, 2016

    2016

  9. [17]

    Data-driven dis- covery of partial differential equations

    Samuel H Rudy, Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Data-driven dis- covery of partial differential equations. Science Advances, 3(4):e1602614, 2017

    2017

  10. [18]

    Discovering physics from data: Universal pde identifiability with sparse regression and neural networks

    Hyunghoon Kim, Eugene Choi, Houman Owhadi, and Marin Soljaˇ ci´ c. Discovering physics from data: Universal pde identifiability with sparse regression and neural networks. arXiv preprint arXiv:2111.12335, 2021

    2021 arXiv

  11. [19]

    Ensemble learning for robust pde discovery

    Ulysse Fasel, Luca Dede, Alfio Quarteroni, and Petros Koumoutsakos. Ensemble learning for robust pde discovery. PNAS, 119(36):e2200925119, 2022

    2022

  12. [20]

    Learning deep koopman representations for data-efficient control

    Enoch Yeung, Saurabh Kundu, and Sebastian Peitz. Learning deep koopman representations for data-efficient control. Nature Communications, 12(1):7333, 2021

    2021

  13. [21]

    Hybrid ² neural ode causal modeling

    Jian Zou, Ziyang Li, Xinyue Zhang, Zheng Li, and Jie Wang. Hybrid ² neural ode causal modeling. In Proceedings of the 41st International Conference on Machine Learning , pages 1123–1134, 2024

    2024

  14. [22]

    Deep koopman operator framework for causal discovery in nonlinear dynamical systems

    John Nathaniel, Carla Roesch, Jatan Buch, Derek DeSantis, Adam Rupe, Kara Lamb, and Pierre Gentine. Deep koopman operator framework for causal discovery in nonlinear dynamical systems. arXiv preprint arXiv:2501.01234 , 2025

    2025 arXiv

  15. [23]

    Cf-gode: Continuous- time causal inference for multi-agent dynamical systems

    Song Jiang, Zijie Huang, Xiao Luo, Yizhou Sun, and Bernhard Sch¨ olkopf. Cf-gode: Continuous- time causal inference for multi-agent dynamical systems. arXiv preprint arXiv:2308.05241 , 2023

    2023 arXiv

  16. [24]

    Causality: Models, Reasoning, and Inference

    Judea Pearl. Causality: Models, Reasoning, and Inference . Cambridge University Press, 2009

    2009

  17. [25]

    Toward causal representation learning.PNAS, 118(38):e2103171118, 2021

    Bernhard Sch¨ olkopf. Toward causal representation learning.PNAS, 118(38):e2103171118, 2021

    2021

  18. [26]

    Djuri´ c

    Kurt Butler, Daniel Waxman, and Petar M. Djuri´ c. Tangent space causal inference: Vector fields for causal discovery in dynamical systems. arXiv preprint arXiv:2404.07321 , 2024

    2024 arXiv

  19. [27]

    Is l2 loss always suitable for pinn training? In Advances in Neural Information Processing Systems 35 , pages 4021–4033, 2022

    Chuwei Wang, Shanda Li, Di He, and Liwei Wang. Is l2 loss always suitable for pinn training? In Advances in Neural Information Processing Systems 35 , pages 4021–4033, 2022. 29

    2022

  20. [28]

    Physpde: A physical hypothesis selection benchmark for pde discovery

    Mingquan Feng, Yixin Huang, Yizhou Liu, Qixuan Chen, and Xiaolin Zhong. Physpde: A physical hypothesis selection benchmark for pde discovery. In International Conference on Learning Representations, 2025

    2025

  21. [29]

    Cand` es, Justin Romberg, and Terence Tao

    Emmanuel J. Cand` es, Justin Romberg, and Terence Tao. Stable signal recovery from incomplete and inaccurate measurements. Communications on Pure and Applied Mathematics, 59(8):1207– 1223, 2006

    2006

  22. [30]

    Joel A. Tropp. Just relax: Convex programming methods for identifying sparse signals in noise. IEEE Transactions on Information Theory , 52(3):1030–1051, 2006

    2006

  23. [31]

    A Mathematical Introduction to Compressive Sensing

    Simon Foucart and Holger Rauhut. A Mathematical Introduction to Compressive Sensing . Birkh¨ auser, New York, 2013. 30

    2013

Pith tools

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