{"id":"b02fff69-08dc-4189-9db2-027adf2096f3","arxiv_id":"2512.13163","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A late-lumping GP prior over the Hamiltonian, discretized independently with PFEM, learns a 1D nonlinear wave equation and identifies its quadratic parameters.","lead":"This paper builds a Gaussian-process framework for learning distributed port-Hamiltonian systems by combining them with a structure-preserving finite element discretization. It shows the method on a nonlinear 1D wave equation with unknown parameters, using fewer hyperparameters than earlier approaches.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (15)/(23) omit M^{-1} from the control term, breaking the discrete power balance and biasing the GP prior under the nonzero input used in the experiments.","rationale":"The reader's stated weakest_assumption was the representability of the Hamiltonian as a quadratic mean plus SE-kernel residual and the dependence of parameter identification on the hyperparameter discretization. That is a real limitation, and the paper's own case (b) shows the kernel can compensate for a misspecified mean. However, the control-term error in Eqs. (15) and (23) is a more fundamental, concrete inconsistency: it directly invalidates the derivation of the GP prior and the passivity-preserving claim, and it can bias the numerical results because the experiments use a nonzero control. This error is load-bearing because the central claim depends on the correctness of Eq. (15) as the link between the dpHs and the GP prior. The reader did mention the control-gain formula as an item to fix, so there is partial agreement, but the reader did not make it the weakest assumption. The concern is fixable; if the equations are corrected (and the code checked), the method may still be valid, so I do not move the verdict beyond CONDITIONAL. I therefore keep verdict_should_be UNCHANGED, while emphasizing that the control-term correction is a prerequisite for the paper's claims.","tokens_in":9807,"tokens_out":7636,"duration_ms":62868,"concrete_test":"Re-derive Eq. (15) step by step from Eqs. (10) and (14); the control term must be M^{-1}G u, not G u. To settle whether the error is consequential, run the provided code on a simple linear wave equation (c(x)=0, known T, ρ) with a nonzero control; if the code implements the published G u term, the predicted trajectory under the posterior mean will deviate from the true trajectory by an amount attributable to the missing M^{-1}, and the discrete power balance will fail. If the code actually uses M^{-1}G u, the paper's equations are a typo and the numerical results may stand, but the derivation still needs correction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central derivation of the GP prior for f_dot_alpha contains a concrete algebraic error. From Eq. (10), M * dot_alpha = (J - R) e + G u, so dot_alpha = M^{-1}(J - R)e + M^{-1}G u. Using Eq. (14), e = M^{-1} ∫ Φ (δ_α H)(Φ^T α) dx, the correct expression is dot_alpha = M^{-1}(J - R)M^{-1} ∫ Φ (δ_α H) dx + M^{-1}G u. However, Eq. (15) writes f_dot_alpha = M^{-1}(J - R)M^{-1} ∫ Φ (δ_α H) dx + G u, and Eq. (23) repeats this with mean M^{-1}(J - R) me(α) + G u. The control term is missing a factor M^{-1}. This is not a notational quibble: the discrete power balance (9) is dH_d/dt = -ep^T R22 ep + u^T y, which holds only when the input enters through M^{-1}G u. With the published control term, the passivity-preserving property is not established for the GP prior. Moreover, the numerical experiments use a nonzero control u_L(t) = sin(πt); a wrong input map in the prior can bias the posterior mean and the estimated hyperparameters m_q, m_p, undermining the claimed identification of T and ρ^{-1}. This is a load-bearing concern because it affects both the theoretical structure-preservation claim and the numerical validation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":10253,"tokens_out":12444,"duration_ms":111251,"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":[{"comment":"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.","section":"§3.3, Eq. (15); §4.2, Eq. (23)"},{"comment":"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.","section":"§4.2, Eqs. (21)-(22)"},{"comment":"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.","section":"§5 and Conclusion"}],"minor_comments":[{"comment":"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.","section":"§4.1, Eq. (18)"},{"comment":"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.","section":"§2.2, Eq. (5)"},{"comment":"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.","section":"General"},{"comment":"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.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The algebraic errors in the control map and covariance factorization are significant enough that the theoretical claims should not be accepted as written, but they appear fixable within the paper's scope. The code availability is a strength and should be preserved. If the authors correct the formulas and add out-of-sample validation, the paper could become a useful contribution to physics-informed learning for distributed port-Hamiltonian systems."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading if you work on structure-preserving system identification or physics-informed learning. The genuinely new piece is the late-lumping strategy: they put a function-valued GP prior on the Hamiltonian functional, then discretize with PFEM afterwards, so the number of hyperparameters is controlled by a separate, coarser discretization rather than by the PFEM mesh. The quadratic-mean prior is also a sensible modification—it lets the kernel focus on the nonlinear part of the energy. I was glad to see they also ship code and show the passivity argument in the PFEM setting; that is real work.\n\nThat said, there is a concrete algebraic error that the stress-test note correctly identifies. In Eq. (10) the dynamics is M \\dot{alpha} = (J-R)e + G u, so the control enters the state equation as M^{-1}G u. But Eqs. (15) and (23) write +G u, missing the M^{-1}. This is not a notational quibble: the discrete power balance only holds with the M^{-1} in place, and since the experiments use a nonzero u_L, the prior mean is biased and the hyperparameter estimates for m_q and m_p can be skewed. The authors should fix this and verify the power balance numerically. Everything else in the GP derivation checks out.\n\nOn validation, the paper is thinner. The experiments use a single simulated trajectory, evaluate the posterior mean on the same trajectory they trained on, and give no error bars or comparisons against the earlier GP-dPHS methods (Tan et al., Li et al.). The identification of T and ρ^{-1} is only shown for one hyperparameter discretization; case (b) in the paper itself shows the kernel silently compensating for a misspecified quadratic mean, so the clean separation between 'quadratic mean identified' and 'kernel handles nonlinearity' depends on an arbitrary discretization choice. That is a honest admission, but it complicates the claim that the method identifies physical parameters.\n\nOverall: the central idea is plausible and the paper deserves a serious referee, but as written it needs a revision on the control term and a stronger numerical study—out-of-sample testing, at least one baseline, and uncertainty bounds. If those are addressed, this could be a useful contribution. If you are refereeing it, I would accept with major revision; the error is fixable, not fatal.","headline":"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.","tokens_in":10732,"tokens_out":3631,"would_cite":true,"duration_ms":31041,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M60","60G15","62F15","35L05"],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["Gaussian process","port-Hamiltonian systems","partitioned finite element method","system identification","nonlinear wave equation","late lumping","quadratic mean prior"],"falsifier":"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 $\\rho^{-1}$, 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.","tokens_in":9714,"feed_emoji":"🌊","tokens_out":10442,"duration_ms":91822,"temperature":0.7,"texified_at":"2026-08-05T20:43:33.141163+00:00","pith_summary":"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.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":5418,"prompt_tokens":849,"completion_tokens":4569,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":849,"completion_tokens_details":{"reasoning_tokens":3769}},"feed_headline":"Learn a wave equation's tension and mass with one GP prior","feed_subtitle":"Placing the Gaussian process on the Hamiltonian separates linear from nonlinear physics and preserves passivity.","key_machinery":"The load-bearing object is the affine-transformation identity $\\dot \\alpha = M^{-1}(J-R)M^{-1}\\int \\Phi(x)(\\delta_\\alpha \\hat H)(\\Phi^T \\alpha)(x)\\,dx + Gu$, 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 $\\dot \\alpha$ without losing the port-Hamiltonian structure. The functional hyperparameters $(m_q, m_p, \\ell_q^{-1}, \\ell_p^{-1})$ are then introduced directly in the prior mean and kernel, and their independent discretization is what keeps the number of tunable parameters small.","core_discovery":"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 $\\dot \\alpha$. Explicitly, $\\dot \\alpha = M^{-1}(J-R)M^{-1}\\int \\Phi(x)(\\delta_\\alpha \\hat H)(\\Phi^T \\alpha)(x)\\,dx + Gu$, 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 $m(\\alpha)=\\int (m_q \\alpha_q^2 + m_p \\alpha_p^2)$ makes the","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Late lumping: one GP prior learns tension and mass","GP prior on Hamiltonian learns nonlinear wave parameters","PFEM meets GP: learn wave tension and mass","One Gaussian process learns a wave equation's physics"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Late lumping: one GP prior learns tension and mass","GP prior on Hamiltonian learns nonlinear wave parameters","PFEM meets GP: learn wave tension and mass","One Gaussian process learns a wave equation's physics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1217,"prompt_tokens":709,"completion_tokens":508,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":447}},"tokens_in":453,"tokens_out":508,"duration_ms":5627,"temperature":1.0,"reasoning_tokens":447,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T16:29:43.922603+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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 $\\rho^{-1}$, 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.","supporting_citations":[],"review_version":1}