Pith. sign in

REVIEW 3 major objections 4 minor 13 references

By placing a Gaussian process prior on the Hamiltonian functional instead of on the discretized state, the paper learns the tension and mass of a nonlinear wave equation while keeping the port-Hamiltonian structure intact.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-08-03 16:29 UTC pith:RI5CNT7Z

load-bearing objection PFEM-GP-dPHS is a genuinely new late-lumping way to put GP priors on distributed port-Hamiltonian systems, but the control term in the key equations is missing M^{-1} and the validation is in-sample. the 3 major comments →

arxiv 2512.13163 v2 pith:RI5CNT7Z submitted 2025-12-15 math.AP

PFEM-GP-dPHS : a finite element framework for combining Gaussian processes and infinite-dimensional port-Hamiltonian systems

Florian Courteville (ILL , KTH) , Iain Henderson , Denis Matignon , Sylvain Dubreuil This is my paper
classification math.AP MSC 65M6060G1562F1535L05
keywords Gaussian processport-Hamiltonian systemspartitioned finite element methodsystem identificationnonlinear wave equationlate lumpingquadratic mean prior
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that the right way to learn an unknown distributed port-Hamiltonian system from data is to place a Gaussian process prior on the infinite-dimensional Hamiltonian functional, not on the discretized state vector, and to discretize only afterwards with a structure-preserving finite element method. This 'late lumping' keeps the hyperparameters as continuous functions, so their discretization is chosen independently from the discretization of the state, avoiding the hundreds of hyperparameters that an early-lumping approach would need. The authors give the prior a quadratic mean, so the squared-exponential kernel only has to represent the non-quadratic part of the Hamiltonian. In a nonlinear 1D wave equation with unknown tension and linear mass, the method reproduces the true trajectory and, when the hyperparameters are piecewise-linear functions, the optimized quadratic coefficients match the physical parameters. If this holds beyond the example, it offers a way to learn physics-based, passivity-preserving distributed-parameter models with built-in uncertainty.

Core claim

The paper's central claim is that the partitioned finite element method (PFEM) can be read as an affine transformation that takes a Gaussian process prior over the Hamiltonian functional \hat H to a Gaussian process prior over the discrete time derivative \dot α. Explicitly, \dot α = M^{-1}(J-R)M^{-1}∫ Φ(x)(δ_α \hat H)(Φ^T α)(x)dx + Gu, where the mass matrices M_q, M_p, D, R_{22} and control matrices B_L, B_R come from PFEM and guarantee the discrete power balance \dot H_d = -e_p^T R_{22} e_p + u^T y. Because the hyperparameters of \hat H are functions of space, they can be discretized on their own mesh—the late-lumping idea. Choosing a quadratic mean m(α)=∫(m_q α_q^2 + m_p α_p^2) makes the

What carries the argument

The load-bearing object is the affine-transformation identity \dot α = M^{-1}(J-R)M^{-1}∫ Φ(x)(δ_α \hat H)(Φ^T α)(x)dx + Gu, together with the PFEM mass and interconnection matrices that make the discrete system preserve the continuous power balance exactly. This identity is what turns the infinite-dimensional GP over \hat H into a finite-dimensional GP over \dot α without losing the port-Hamiltonian structure. The functional hyperparameters (m_q, m_p, ℓ_q^{-1}, ℓ_p^{-1}) are then introduced directly in the prior mean and kernel, and their independent discretization is what keeps the number of tunable parameters small.

Load-bearing premise

The method assumes the true Hamiltonian is a spatially varying quadratic form plus a smooth, bounded residual that the squared-exponential kernel can represent, and that the hyperparameter discretization is fine enough to actually capture m_q and m_p; if either fails, the kernel compensates for the misspecified model and the recovered parameters are not the physical ones.

What would settle it

Repeat the same identification experiment with the hyperparameters discretized on meshes of, say, 5, 11, and 21 linear elements; if the optimized m_q and m_p do not converge to the true tension T and inverse linear mass ρ^{-1}, then the identification reported for the 11-point mesh is an artifact of that particular parametrization rather than a consequence of the quadratic-mean prior.

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

Share X LinkedIn Reddit HN

If this is right

  • The discrete GP prior inherits the continuous system's passivity: at any mesh refinement the learned model obeys the exact power balance \dot H_d = -e_p^T R_{22} e_p + u^T y, so energy cannot be created artificially.
  • The number of hyperparameters is decoupled from the number of state degrees of freedom; the wave example needs 46 or 18 tunable parameters instead of 166 for an early-lumping approach.
  • With a quadratic mean, physically meaningful parameters (tension, linear mass) can be read off directly from the optimized m_q and m_p, rather than being hidden inside the kernel.
  • The construction transfers to any structure-preserving discretization that is an affine transformation of the continuous Hamiltonian, so other PDEs in port-Hamiltonian form can be learned the same way.
  • The posterior GP gives uncertainty on the vector field \dot α, so downstream control or optimization can use variance information, not just the point estimate.

Where Pith is reading between the lines

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

  • The paper fixes the interconnection and damping operators (J and R) and learns only the Hamiltonian; a natural extension is to treat those operators as functional hyperparameters too, though the current affine-transformation derivation would need to be reworked.
  • The quadratic-mean separation is only as clean as the hyperparameter discretization: the polynomial case in the paper shows the kernel silently compensating for a misspecified mean, which implies that physical interpretation of m_q and m_p as 'the' parameters depends on a discretization choice the method does not yet justify.
  • A testable extension is to apply the method to a Hamiltonian with a discontinuous or non-smooth coefficient; the squared-exponential kernel would likely fail, revealing whether the kernel is truly learning nonlinearity or just absorbing model error.
  • Because the method yields a GP over the vector field, one could feed the posterior variance into an active-learning loop that chooses which sensor locations or input signals to query next, reducing the amount of trajectory data needed.

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

3 major / 4 minor

Summary. The paper proposes PFEM-GP-dPHS, a late-lumping framework that puts a Gaussian-process prior directly on the Hamiltonian functional of a distributed port-Hamiltonian system and then discretizes the resulting stochastic dynamics with the partitioned finite element method. The GP prior uses a quadratic mean plus an SE kernel whose hyperparameters are continuous functions, allowing the hyperparameter discretization to be chosen independently of the state discretization. The method is illustrated on a nonlinear 1D wave equation with unknown tension and linear mass, using a single simulated trajectory for hyperparameter fitting and posterior simulation. The paper also provides a public code repository.

Significance. If correct, the late-lumping idea is valuable: it avoids the explosion of hyperparameters that occurs when a GP prior is placed on a high-dimensional discretized Hamiltonian, and the quadratic-mean prior is a sensible way to separate dominant linear energetics from nonlinear corrections. The passivity/structure-preservation goal is well motivated, and the paper makes the derivation reasonably explicit. However, the central algebraic derivation currently contains concrete errors in the control input map and in the discrete covariance factorization, both of which affect the validity of the structure-preservation claim and the numerical estimates. These are fixable, but they are load-bearing rather than cosmetic.

major comments (3)
  1. [§3.3, Eq. (15); §4.2, Eq. (23)] The control term is missing a factor M^{-1}. From Eq. (10), M dot_alpha = (J-R)e + G u, so dot_alpha = M^{-1}(J-R)e + M^{-1}G u. Eq. (15) and the mean in Eq. (23) instead write the control term as G u. This breaks the discrete power balance Eq. (9), since with the published input map dH_d/dt = e^T(J-R)e + e^T M G u rather than e^T(J-R)e + u^T y. Section 5 uses u_L(t)=sin(pi t), so the incorrect input map biases the GP prior and the hyperparameter estimates. Please correct Eq. (15), Eq. (23), and the implementation, and rerun the numerical experiments.
  2. [§4.2, Eqs. (21)-(22)] The discrete mean and covariance are inconsistent with Eq. (14). In Eq. (21), the explicit mean is written as [M_{m,q} alpha_q; M_{m,p} alpha_p], but the preceding definition me = M^{-1} integral Phi me dx requires an additional M^{-1}. For the covariance, Eq. (22) with M_k = M^{-1} Diag(M_{k,q}, M_{k,p}) gives a first term proportional to M^{-1} M_{k,q}^2 M^{-1}, whereas Eqs. (14) and (20) imply M^{-1} M_{k,q} M^{-1}. This changes the scaling of the prior covariance with the length-scale hyperparameters and affects every subsequent hyperparameter fit. The covariance formula should be re-derived and corrected.
  3. [§5 and Conclusion] The numerical validation is in-sample: training and evaluation use the same trajectory and the same input signal, and no quantitative error or uncertainty metrics are reported. More importantly, Figure 2(b) shows that with the degree-3 hyperparameter discretization the quadratic mean does not recover T and rho^{-1}; the kernel compensates for the misspecified mean. Thus the claim that the quadratic mean enables identification of physical parameters depends on the particular P1 hyperparameter discretization and is not robust. Provide out-of-sample or multi-trajectory validation and quantitative errors, or substantially soften the identification claim.
minor comments (4)
  1. [§4.1, Eq. (18)] The positive semi-definiteness of the kernel with spatially varying length-scale functions should be stated or proved, and the conditions on m_q, m_p, ell_q, ell_p that guarantee a bounded-below Hamiltonian should be discussed.
  2. [§2.2, Eq. (5)] The notation for the boundary port term is compressed; the distinction between the distributed dissipation integral and the boundary pairing should be made explicit when the operators act on infinite-dimensional function spaces.
  3. [General] Some formulas use 'Schwarz's lemma' where the intended statement is the equality of mixed partials; the wording can be corrected. Also, Section 3.3 uses both X and O for the spatial domain; unify the notation.
  4. [§5] The figures are described only visually; report numerical values such as relative L2 errors between the true and posterior trajectories and between m_q, m_p and their physical counterparts.

Circularity Check

0 steps flagged

No significant circularity: the GP-prior construction and PFEM discretization are derived from stated assumptions, with hyperparameters fitted to data rather than defined in terms of the target quantities.

full rationale

The paper's derivation chain is self-contained. It defines a function-valued GP prior over the Hamiltonian with a quadratic mean (Eq. 17) and an SE kernel (Eq. 18), then applies the variational derivative, the PFEM affine transformation, and matrix operations to obtain a GP prior over the discrete right-hand side f_dot_alpha (Eqs. 19-23). None of these steps define the hyperparameters m_q, m_p, l_q, l_p, or sigma_f in terms of the quantities to be identified (T and rho^{-1}); the hyperparameters are fitted from simulated training data by minimizing the negative log marginal likelihood. The claim that m_q and m_p identify T and rho^{-1} is an empirical observation on the fitted values, not a result forced by construction. The PFEM structure-preservation property is invoked from prior work (Cardoso-Ribeiro et al.), but the discrete equations and power balance are restated and used directly in the present paper, so the self-citation is not load-bearing. The validation is in-sample, and the possible missing M^{-1} in Eq. (15)/(23) is a correctness concern, not circularity. No step reduces by definition or by self-citation to the result it claims to predict.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The paper introduces no new physical entities. Its contributed structure consists of the functional hyperparameters of the GP prior (quadratic mean coefficient functions and lengthscale functions), all fitted to the data. The main axioms are standard GP/pHs theory plus the specific structural prior choices; the weakest link is the ad hoc quadratic-mean-plus-SE-kernel decomposition of the Hamiltonian.

free parameters (4)
  • m_q(x), m_p(x) (quadratic mean functional hyperparameters) = Piecewise-linear functions on an 11-point mesh (case a) or degree-3 polynomials (case b); values shown in Fig. 2
    These functions set the quadratic part of the GP prior over the Hamiltonian. They are optimized via NLML and, in case (a), recover T and ρ^{-1}; in case (b) they do not.
  • ℓ_q^{-2}(x), ℓ_p^{-2}(x) (inverse squared lengthscale functions) = Piecewise-linear or polynomial functions; values shown in Fig. 3
    These control the kernel's spatial correlation lengths and are fitted to the data. Their shape is interpreted as identifying the nonlinear spatial modulator c(x).
  • σ_f (kernel amplitude) = Scalar hyperparameter, optimized by NLML
    Overall scale of the SE kernel on the Hamiltonian; fitted to the single trajectory.
  • Noise variance (σ_n²) = Scalar hyperparameter, included in the 46/18 hyperparameter count
    Observation noise in the training data; fitted with NLML.
axioms (6)
  • standard math Gaussian process regression theory (mean, covariance, conditioning, derivative processes)
    Used throughout Section 2.1 and Section 4.2; standard results from Rasmussen & Williams (2006) and Alvarez et al. (2012).
  • standard math The SE kernel with weighted L2 norm is positive definite on the Hilbert space of state functions
    Required for Eq. (18) to define a valid GP prior over the Hamiltonian functional; standard for RBF kernels on Hilbert spaces of negative type.
  • domain assumption The 1D wave equation can be written in distributed port-Hamiltonian form with operators J and R as in Eq. (8)
    This is the physical modeling step, relying on van der Schaft & Maschke (2002) and the chosen strain/momentum variables.
  • domain assumption PFEM preserves the discrete power balance and the structure of the dPHS (Cardoso-Ribeiro et al. 2021)
    The whole method relies on PFEM being structure-preserving; the paper cites the prior work for this property rather than re-proving it.
  • domain assumption The variational derivative of the GP prior is a well-defined GP and conditioning commutes with the linear PFEM map
    Implicit in Section 2.3 and Section 4.2: the posterior mean of the derivative process is taken as the derivative of the posterior mean, which is needed for the passivity claim.
  • ad hoc to paper The Hamiltonian is well approximated by a quadratic mean plus a zero-mean SE-kernel residual
    Eqs. (17)-(18) are a modeling choice, not derived from first principles. The paper's own case (b) shows this assumption can fail in practice.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of PFEM-GP-dPHS : a finite element framework for combining Gaussian processes and infinite-dimensional port-Hamiltonian systems." pith.science (2026). https://pith.science/paper/RI5CNT7Z

@misc{pith2026251213163,
  author       = {Pith},
  title        = {Pith review of: PFEM-GP-dPHS : a finite element framework for combining Gaussian processes and infinite-dimensional port-Hamiltonian systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RI5CNT7Z}},
  note         = {Machine review of arXiv:2512.13163}
}
Share X LinkedIn Reddit HN
read the original abstract

In order to learn distributed port-Hamiltonian systems (dPHS) using Gaussian processes (GPs), the partitioned finite element method (PFEM) is combined with the Gp-dPHS method. By following a late lumping approach, the discretization of the functional hyperparameters of the GP prior over the Hamiltonian functional is chosen independently from the discretization of the dPHS, thus reducing the numerical complexity of our method. We next model the mean of the GP prior of the Hamiltonian as a quadratic form, enabling the GP kernel to focus on the nonlinear part of a given dPHS. We illustrate our method on a nonlinear one dimensional wave equation with unknown physical parameters (tension and linear mass).

Figures

Figures reproduced from arXiv: 2512.13163 by Denis Matignon, Florian Courteville (ILL, Iain Henderson, KTH), Sylvain Dubreuil.

Figure 1
Figure 1. Figure 1: True αp trajectory obtained with SCRIMP plotted with the estimated αp trajectory obtained using the posterior mean of the PFEM-Gp-dpHs and discretizing hyperparameters with (a) finite element (b) polynomial decomposition. Results obtained after successful regression for both hyperparameter discretizations are presented in Figures 1, 2 and 3. The true (SCRIMP output) and approximate trajectory of αp can be … view at source ↗
Figure 2
Figure 2. Figure 2: Quadratic hyperparameters of the GP mean plotted along their respective true Hamiltonian [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) Inverse squared correlation lengths ℓ −2 q and ℓ −2 p plotted with the spatial modulator c for hyperparameters discretize with (a) finite element (b) polynomial decomposition. Correlation lengths and c are plotted with separate axes. (“dimensions x”) where its nonlinearity is mild, i.e. c(x) ≃ 0. Hence we expect that the tuned kernel k should not contribute at such locations x. Ensuring this latter pro… view at source ↗

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

13 extracted references · 1 canonical work pages

  1. [1]

    Alvarez, Lorenzo Rosasco, and Neil D

    Mauricio A. Alvarez, Lorenzo Rosasco, and Neil D. Lawrence. Kernels for vector-valued functions: A review. Foundations and Trends in Machine Learning, pages vol. 4, no. 3, pp. 195–266, 2012

    2012

  2. [2]

    Kernel methods are competitive for operator learning

    Pau Batlle, Matthieu Darcy, Bamdad Hosseini, and Houman Owhadi. Kernel methods are competitive for operator learning. J. Comput. Phys., 496: 0 Paper No. 112549, 24, 2024. ISSN 0021-9991. doi:10.1016/j.jcp.2023.112549. URL https://doi.org/10.1016/j.jcp.2023.112549

    arXiv 2024

  3. [3]

    Gaussian process port- H amiltonian systems : Bayesian learning with physics prior

    Thomas Beckers. Gaussian process port- H amiltonian systems : Bayesian learning with physics prior. IEEE 61st Conference on Decision and Control (CDC), 2022

    2022

  4. [4]

    Partitioned finite element method for power-preserving discretization of open systems of conservation laws

    Flávio Luiz Cardoso-Ribeiro, Denis Matignon, and Laurent Lefèvre. Partitioned finite element method for power-preserving discretization of open systems of conservation laws. IMA Journal of Mathematical Control and Information, 38: 0 493–533, 2021

    2021

  5. [5]

    Simulation and control of interactions in multi-physics, a python package for port-hamiltonian systems

    Giuseppe Ferraro, Michel Fournié, and Ghislain Haine. Simulation and control of interactions in multi-physics, a python package for port-hamiltonian systems. IFAC-PapersOnLine, 58 0 (6): 0 119--124, 2024. doi:https://doi.org/10.1016/j.ifacol.2024.08.267

    work page doi:10.1016/j.ifacol.2024.08.267 2024

  6. [6]

    NAPI-MPC : Neural accelerated physics-informed MPC for nonlinear PDE systems

    Peilun Li, Kaiyuan Tan, and Thomas Beckers. NAPI-MPC : Neural accelerated physics-informed MPC for nonlinear PDE systems. Proceedings of Machine Learning Research, 283: 0 1--13, 2025

    2025

  7. [7]

    Early-and late-lumping observer designs for long hydraulic pipelines: Application to pumped-storage power plants

    L Marko, J-F Mennemann, L Jadachowski, W Kemmetm \"u ller, and A Kugi. Early-and late-lumping observer designs for long hydraulic pipelines: Application to pumped-storage power plants. International Journal of Robust and Nonlinear Control, 28 0 (7): 0 2759--2779, 2018

    2018

  8. [8]

    Physics-guided transfer learning for B ayesian optimization of chemical port- H amiltonian systems

    Mahboubi Negareh, Xie Junyao, and Huang Biao. Physics-guided transfer learning for B ayesian optimization of chemical port- H amiltonian systems. Computers and Chemical Engineering, page 109331, 2025. ISSN 0098-1354. doi:10.1016/j.compchemeng.2025.109331

    arXiv 2025

  9. [9]

    Gaussian Processes for Machine Learning

    Carl Edward Rasmussen and Christopher Williams. Gaussian Processes for Machine Learning. the MIT Press, 2006. ISBN 026218253X

    2006

  10. [10]

    Metamodeling with G aussian processes

    Olivier Roustant. Metamodeling with G aussian processes. HAL Id: hal-04509167, 2024. doi:https://hal.science/hal-04509167v2

    2024

  11. [11]

    Physics-constrained learning of PDE systems with uncertainty quantified port- H amiltonian models

    Kaiyuan Tan, Peilun Li, and Thomas Beckers. Physics-constrained learning of PDE systems with uncertainty quantified port- H amiltonian models. In Alessandro Abate, Mark Cannon, Kostas Margellos, and Antonis Papachristodoulou, editors, Proceedings of the 6th Annual Learning for Dynamics amp; Control Conference, volume 242 of Proceedings of Machine Learning...

    2024

  12. [12]

    Plug-and-play physics-informed learning using uncertainty quantified port- H amiltonian models

    Kaiyuan Tan, Peilun Li, and Jun Wang. Plug-and-play physics-informed learning using uncertainty quantified port- H amiltonian models. arXiv:2504.17966v1, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Hamiltonian formulation of distributed-parameter systems with boundary energy flow

    Arjan van der Schaft and Bernhard Maschke. Hamiltonian formulation of distributed-parameter systems with boundary energy flow. Journal of Geometry and Physics, 42: 0 166--194, 2002

    2002

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.