{"id":"9f71c4d9-52a3-4321-b0f1-0b1d4295e279","arxiv_id":"2504.13396","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A closed-form structure-preserving kernel ridge regression estimator learns the Hamiltonian of a Poisson system up to a Casimir function, with PAC bounds, demonstrated on rigid body, underwater vehicle, and two-vortex dynamics.","lead":"This paper presents a kernel-based machine learning method that recovers Hamiltonian functions of Poisson systems on curved manifolds from noisy vector-field data, with convergence guarantees. It extends structure-preserving kernel regression from flat Euclidean spaces to manifolds such as spheres and Lie algebras, which matters for physics-informed learning of mechanical and fluid systems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.13 derives the closed-form coefficients from an equality of kernel-represented functions without an injectivity argument; in degenerate Poisson settings this step is invalid as written and the estimator formula needs a repaired proof.","rationale":"The reader's verdict is CONDITIONAL, and the reader's rationale already flags that the proof of Theorem 3.13 glosses over an injectivity step. My stress-test sharpens this into the single most load-bearing concern: the closed-form estimator is the method's computational backbone, and the contested step is not merely a missing detail but an inference that fails in permitted degenerate Poisson settings, such as the zero Poisson bracket. I have not found evidence that the final formula is false; in fact, a repair via the identity Q_N g_N(c, X_K·(Z_N)) = (1/N) g_N(G_N c, X_K·(Z_N)) plus invertibility of Q_N + \\lambda I appears to close the gap. I therefore do not recommend moving the verdict to reject or to unverified: the paper should be accepted, if at all, only after the proof of Theorem 3.13 is corrected or the missing injectivity argument is supplied. The other weaknesses noted by the reader, the absence of noisy numerical experiments and the failure of the distance regularity assumption in Proposition 4.10 on S^2×S^2, are real but secondary: they affect validation and a downstream flow-approximation statement, whereas the representer-theorem gap directly concerns the formula used in every experiment and in the implementation of the method. The agreement is 'partial' because the reader's formal weakest assumption was Assumption 3.7, not the Theorem 3.13 proof gap, although the same gap appears in the reader's rationale.","tokens_in":42655,"tokens_out":29095,"duration_ms":271290,"concrete_test":"Re-derive Theorem 3.13 without the contested step: show that the operator identity Q_N g_N(c, X_K·(Z_N)) = (1/N) g_N(G_N c, X_K·(Z_N)) holds, define c_0 = (G_N + \\lambda N I)^{-1}X_{\\sigma^2,N}, and verify that h_0 = g_N(c_0, X_K·(Z_N)) satisfies (Q_N + \\lambda I)h_0 = (1/N) g_N(X_{\\sigma^2,N}, X_K·(Z_N)). If this check succeeds, the formula (3.19) follows without injectivity. A complementary numerical check is to run the rigid-body example with N = 50 using both the closed-form estimator (3.16)-(3.19) and the operator estimator (3.14), and compare the predicted Hamiltonian vector fields; agreement of the vector fields to machine precision would indicate the formula is correct even though the printed proof step is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central computational object of the paper is the closed-form estimator (3.16)-(3.19), and the proof of Theorem 3.13 is the only derivation of that formula. The proof represents the minimizer as h = g_N(c, X_K·(Z_N)) for some c in T_{Z_N}P (equation (3.17)), then applies Q_N + λI to both sides and obtains the function identity (3.18). From this identity the paper immediately concludes the vector equality G_N c + λN c = X_{\\sigma^2,N} (the line before (3.19)). This step is not valid without injectivity of the coefficient map c \\mapsto g_N(c, X_K·(Z_N)) into H_K. In a degenerate Poisson structure the kernel sections X_{K_z} can be dependent: for the zero Poisson bracket, X_h = 0 for every h, so the coefficient map is the zero map and the conclusion from function equality to coefficient equality is simply false. The same structural issue can arise whenever the span of the kernel sections has nontrivial linear dependencies. The theorem may still be true, and the gap is likely repairable: define c_0 = (G_N + \\lambda N I)^{-1}X_{\\sigma^2,N}, set h_0 = g_N(c_0, X_K·(Z_N)), prove the operator identity Q_N g_N(c, X_K·(Z_N)) = (1/N) g_N(G_N c, X_K·(Z_N)), and use invertibility of Q_N + \\lambda I on the finite-dimensional invariant subspace. But that argument is not what appears in the paper. Since every numerical experiment and the claimed representer theorem depend on the unproved vector equation (3.19), this is a load-bearing gap in the central claim as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a structure-preserving kernel ridge regression estimator for Hamiltonian functions on Poisson manifolds from noisy observations of Hamiltonian vector fields. The estimator is given in closed form as a finite linear combination of kernel sections (Theorem 3.13), with an operator representation and a detailed error analysis yielding PAC-type rates under a source condition (Theorems 4.6 and 4.8) and a flow-approximation statement (Proposition 4.10). The method is designed to handle the non-identifiability caused by Casimir functions through Tikhonov regularization, and a symmetry preservation result is stated in terms of invariant kernels and Noether's theorem. Numerical experiments cover Lie-Poisson systems (rigid body, underwater vehicle), exact recovery in a Gaussian RKHS on R^3, and Hamiltonian systems on S^2 x S^2 (spherical 3-norm and two-vortex dynamics).","tokens_in":42996,"tokens_out":20237,"duration_ms":190793,"significance":"If the technical gaps identified below are repaired, this would be a valuable contribution to structure-preserving learning: it provides a global, coordinate-free estimator on Poisson manifolds, a closed-form representer-type formula, explicit handling of Casimir degeneracy via regularization, and convergence rates in the RKHS norm under a source condition. The paper also gives detailed proofs and nontrivial numerical examples, including singular vector fields. The claimed results are plausible, but because the current proof gaps affect the derivation of the estimator and the main convergence and flow-approximation statements, the significance is conditional on those repairs.","major_comments":[{"comment":"The passage from the function identity (3.18) to the coefficient equation (3.19) is valid only if the map c in T_{Z_N}P |-> g_N(c, X_{K·}(Z_N)) in H_K is injective. This injectivity is neither stated nor proved, and it fails in degenerate Poisson structures: for the zero Poisson bracket, X_{K_z}=0 for every z, so (3.18) reduces to 0=0 and cannot imply G_N c + lambda N c = X_{sigma^2,N}. The theorem statement is likely still correct, and the gap is repairable: define c_0=(G_N+lambda N I)^{-1}X_{sigma^2,N}, set h_0=g_N(c_0,X_{K·}(Z_N)), and prove directly that (Q_N+lambda I)h_0=(1/sqrt N)A_N^*X_{sigma^2,N} using Q_N g_N(c,X_{K·}(Z_N))=(1/N)g_N(G_N c,X_{K·}(Z_N)) and the definition of G_N. Please replace the present inference with this argument or add an explicit injectivity assumption.","section":"Section 3.4, Theorem 3.13, Eqs. (3.18)-(3.19)"},{"comment":"The first displayed bound in Theorem 4.8 has a prefactor C(gamma,delta,kappa)=max{||B^{-gamma}H||, 8 sqrt(4 log(8/delta)) d^{3/2} kappa^3 ||H||} that is independent of the noise variance sigma^2 and of the constant c in the choice lambda=cN^{-alpha}. This is inconsistent with Lemma 4.2, whose noisy-sampling bound contains sigma kappa lambda^{-1} N^{-1/2}; substituting lambda=cN^{-alpha} gives a contribution with prefactor sigma kappa/c multiplied by N^{alpha-1/2}, so this factor must appear in C. As written, the theorem claims a rate that cannot hold uniformly for arbitrary sigma>0. In addition, 'B^{-gamma}H' is undefined in the statement; the source condition (4.2) uses Q, so the notation should be Q^{-gamma}H or the operator should be defined.","section":"Section 4, Theorem 4.8"},{"comment":"Lemma 4.9 is false as stated: X_h[f]=df(X_h) is not the same as the quantity Dh(y) with y=(z,X_f(z)) that is used in the proof; the latter is X_f[h]. In a Euclidean symplectic chart on a compact manifold, take h=1/2 sin^2(x_1) and f=M sin(x_2) on the two-torus with the standard symplectic form. Then X_h[f]=M cos(x_2) sin(x_1) cos(x_1), whose Lipschitz constant grows linearly in M, while ||h||_{C^2_b} is independent of M. The claimed bound with constant ||h||_{C^2_b} therefore fails for large M. The proof also chooses the curve v(t) in TP to be parallel along the geodesic, which does not in general interpolate X_f(z_1) and X_f(z_2). Since Proposition 4.10 uses this lemma directly, the flow-approximation statement needs a corrected statement with additional assumptions on the test functions f and a repaired proof.","section":"Section 4, Lemma 4.9 and Proposition 4.10"},{"comment":"The numerical section does not validate the noisy-data setting advertised in the abstract and used in the theory. The rigid-body experiment explicitly states sigma=0; the underwater vehicle, exact-recovery, S^2 x S^2, and two-vortex experiments do not state a noise level either, and none of the reported tests appears to inject observation noise. Since the main formal contribution concerns recovery from noisy vector-field observations, please add experiments with nonzero sigma^2 and report the estimator and vector-field errors as sigma varies.","section":"Section 5, numerical experiments"},{"comment":"The definition of argumentwise invariance is K(g·x,g'·x')=K(x,x') for all g,g'. From this condition it does not follow immediately that K(g·x,·)=K(x,·) for each g, which is the property needed to conclude that every f in H_K is G-invariant via f(gx)=<f,K(gx,·)>_{H_K}. The assertion that the RKHS consists exclusively of G-invariant functions therefore needs either a proof from the stated definition or an alignment of the definition with the property cited from [Gins 12, Property 3.13] (for example, first-argument invariance K(gx,y)=K(x,y)).","section":"Section 3.5, Definition 3.15 and Proposition 3.16"}],"minor_comments":[{"comment":"The phrase 'with probability as least 1-delta' should read 'with probability at least 1-delta'.","section":"Section 4, Theorem 4.8"},{"comment":"The analysis paragraph refers to 'Theorem 3.2 (iii)'; the intended reference appears to be Theorem 3.3 (iii).","section":"Section 5.2.4"},{"comment":"In the local-coordinate expression following Eq. (3.21), the index ordering in G_{i,j}^N=B(z^{(i)}) partial_{1,2}K(z^{(i)},z^{(j)}) B^T(z^{(j)}) g^{(j)} should be checked for consistency with the definition (3.15), particularly the order of the two points in the kernel derivative, since the text alternates between (z^{(i)},z^{(j)}) and (z^{(j)},z^{(i)}).","section":"Section 3.4, local coordinate formula"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the manuscript leans substantially on the authors' previous work [Hu 24] for the operator framework and the Euclidean proof skeleton. This is not circular, since the Poisson-manifold differential reproducing property, the representer formula, and the Casimir discussion are new, but the novelty statement could be sharpened. The main technical gaps are localized and appear repairable within the scope of the paper, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2504.13396. The paper extends the authors' earlier Euclidean structure-preserving kernel method to Poisson manifolds, and the extension is genuine: the differential reproducing property on Riemannian manifolds, the Poisson version of the representer theorem, and the explicit treatment of Casimir non-identifiability via Tikhonov regularization are all new relative to the cited literature. The error analysis with source conditions and adaptive regularization is a serious piece of work, and the Lie-Poisson and S^2×S^2 examples show the method works in settings beyond the Euclidean case.\n\nNow the soft spots. The proof of Theorem 3.13, which is the foundation of the closed-form estimator, has a real gap: from the function equality (3.18) the paper concludes the coefficient equality G_N c + λN c = X_{σ^2,N}, but this requires injectivity of the map c ↦ g_N(c, X_K·(Z_N)) into H_K. That map can be non-injective in degenerate Poisson settings (trivially for the zero bracket). The result is very likely true—there's a standard repair via (G_N + λNI)⁻¹ X_{σ^2,N} and an invariant-subspace argument—but the proof as written doesn't deliver it. The authors should fix this before publication.\n\nSecond, the numerical experiments never inject noise, even though the problem statement and error analysis are explicitly about noisy observations. That's a mismatch that needs to be addressed by adding at least one noisy-data experiment.\n\nThird, the rigid body demonstration uses a manually fitted Casimir correction (a·||Π||² with a=1.85 chosen by trial and error). The learned Hamiltonian alone is far from the truth; the success of the example depends on the correction. That's honest but weakens the demonstration. The 'exact recovery' example with a Hamiltonian in the RKHS is more convincing.\n\nFourth, Proposition 4.10 on flow approximation assumes the geodesic distance d(·, y) is in C¹_b with a uniform bound, which fails on the spheres used in the paper (distance to a point is not smooth at the antipode and cut locus). The proposition should be scoped more carefully.\n\nThese are addressable issues. The core estimator appears sound, the writing is clear, and the authors are careful about what they claim. This paper is for researchers in structure-preserving learning and geometric mechanics. It deserves a serious peer review, and I'd send it out—with the expectation that the representer theorem proof and the numerical validation are fixed.","headline":"A genuinely new manifold extension of a structure-preserving kernel method with a repairable proof gap and uneven numerical validation.","tokens_in":43585,"tokens_out":3868,"would_cite":true,"duration_ms":32096,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65P10","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A closed-form kernel regression recovers Hamiltonian functions on Poisson manifolds from noisy vector-field data.","keywords":["Poisson manifolds","Hamiltonian systems","kernel ridge regression","reproducing kernel Hilbert spaces","structure-preserving learning","Casimir functions","Noether's theorem","convergence rates"],"falsifier":"On a Poisson manifold with nontrivial Casimir directions, take a target Hamiltonian $H$ that is a known kernel section in the source space (as in Section 5.1.3), generate noisy vector-field samples with positive variance, estimate with $\\lambda=cN^{-\\alpha}$, and measure the empirical RKHS error over many independent trials. If, for a compatible structure that violates Assumption 3.7 on the sampling support, the error still decays at the claimed rate with high probability, then the boundedness assumption is not necessary; if the decay breaks, it is confirmed as the load-bearing condition the paper states it to be.","tokens_in":42394,"feed_emoji":"🧮","tokens_out":10601,"duration_ms":105602,"temperature":0.7,"pith_summary":"The paper tries to establish that the Hamiltonian function $H$ of a Poisson system can be recovered from noisy samples of its Hamiltonian vector field by a single convex kernel ridge regression, with no iterative training and no local coordinate charts. The learned function lies in a reproducing kernel Hilbert space on the manifold and is a Hamiltonian of the same Poisson structure by construction, so the geometry is preserved exactly rather than approximately. The authors prove high-probability convergence in the RKHS norm, at rate $N^{-\\min\\{\\alpha\\gamma,(1-3\\alpha)/2\\}}$ under an adaptive regularization $\\lambda\\propto N^{-\\alpha}$ when the target lies in a source space, and show the flow of the learned Hamiltonian tracks the true flow. A reader should care because many physical systems—rigid bodies, underwater vehicles, point vortices—are Poisson systems, and this method turns their inverse problem into a closed-form regression with certified error.","feed_headline":"Closed-form kernel method learns Poisson systems from noisy data","feed_subtitle":"One regression recovers a global, chart-free Hamiltonian with convergence guarantees and Casimir ambiguity resolved","key_machinery":"The object that carries the argument is the compatible structure $J=B^{\\sharp}\\circ g^{\\flat}$, the bundle map over a Riemannian metric $g$ that writes every Hamiltonian vector field as $X_h=J\\nabla h$; locally $J=-Bg$. On it rest three tools: the differential reproducing property on Riemannian manifolds, $D^k f(y)\\cdot v=\\langle D^{(k,0)}K(y,\\cdot)\\cdot v,f\\rangle_{H_K}$, which converts differentials into RKHS inner products; the generalized differential Gram matrix $G_N$, whose symmetry and positive semidefiniteness make the ridge problem well posed; and the operator $Q=A^*A$, whose spectral decomposition together with the source condition $H=Q^\\gamma\\psi$ produces the convergence rates. Group-invariant kernels feed into the same machinery, so Noether's theorem carries momentum conservation to the estimated flow.","core_discovery":"On the paper's own terms, the central discovery is that learning $H$ from $X_H=B^{\\sharp}(dH)=J\\nabla H$ reduces to a non-standard kernel ridge regression whose minimizer is a differential version of the Representer Theorem on Poisson manifolds: $$\\hat h_{\\$\\lambda$,N}=g_N\\!\\left(\\hat c,\\,X_{K_\\cdot}(Z_N)\\right),\\qquad \\hat c=(G_N+\\$\\lambda$ N I)^{-1}X_{\\$sigma^{2}$,N},$$ where $G_N$ is the generalized differential Gram matrix built from the compatible structure $J$ and the kernel sections. The estimator is globally defined and chart-independent; it is unique for $\\lambda>0$, and it automatically lies in the orthogonal complement of the kernel of the operator $h\\mapsto J\\nabla h$, which is what resolves the ambiguity caused by Casimir functions. Around this formula the paper builds an operator framework ($Q=A^*A$, $Q_N=A_N^*A_N$) that yields high-probability estimation and approximation error bounds, and a flow-approximation statement under a boundedness assumption on $J$.","pith_inferences":["Extension the paper leaves implicit: the same operator argument should transfer to any inverse problem whose observed quantity is a bundle map applied to the differential of an unknown function, such as learning a potential from measured forces in a gradient or controlled system; the rates in this paper constrain the data-observation model more than the Hamiltonian character of the dynamics.","Balancing the two exponents in $N^{-\\min\\{\\alpha\\gamma,(1-3\\alpha)/2\\}}$—an optimization the paper does not perform—yields an optimal regularization choice $\\alpha=1/(2\\gamma+3)$ with rate $\\gamma/(2\\gamma+3)$, and under coercivity $\\alpha=1/(2\\gamma+2)$ with rate $\\gamma/(2\\gamma+2)$.","The two-vortex experiment, where the true Hamiltonian is singular and outside the RKHS, suggests the practical method is wider than the theorem: qualitative structure is still learned, while error concentrates near the singular set—an inviting test bed for singular-aware or locally refined kernels."],"forward_implications":["Hamiltonian discovery on Poisson manifolds becomes a one-step convex problem: compute one $dN$-dimensional linear system from $N$ noisy vector-field samples, and the output is a global Hamiltonian, not a patchwise model.","The learned vector field is Hamiltonian by construction, so the manifold constraint and the conservation of Casimirs are built in exactly rather than enforced as soft losses.","Ridge regularization gives a unique estimator even though adding a Casimir function would leave the observed vector field unchanged; the estimator lies in $H^{\\perp}_{\\mathrm{null}}$.","If the kernel is invariant under a canonical group action, the estimated Hamiltonian is invariant and every momentum map of that action is exactly conserved by the estimated flow.","Under the source condition and bounded-compatible-structure assumption, finite-sample high-probability error bounds hold, and the continuous-time flow of the learned system stays within order of the RKHS error of the true flow over finite time intervals."],"supporting_citations":[{"why":"Supplies the Euclidean symplectic estimator and error decomposition that the manifold/Poisson version extends.","marker":"[Hu 24]"},{"why":"Source of the derivative reproducing property generalized to Riemannian manifolds in Theorem 3.3.","marker":"[Zhou 08]"},{"why":"Provides the estimation-versus-approximation error analysis pattern reused in Section 4.","marker":"[Feng 23]"},{"why":"The Hanson-Wright inequality used to control the noisy sampling error in Lemma 4.2.","marker":"[Rude 13]"},{"why":"Establishes that argumentwise invariant kernels have invariant RKHS functions, used in the Noether conservation result.","marker":"[Gins 12]"},{"why":"Supplies the standard Poisson mechanics and Noether theorem background, including Casimirs and momentum maps.","marker":"[Mars 99]"},{"why":"Used in Section 5.1.3 to identify a Gaussian-kernel section that lies in the RKHS and satisfies the source condition.","marker":"[Minh 10]"}],"fun_headline_variants":["Kernel regression recovers Hamiltonian from noisy Poisson dynamics","One regression solves Casimir ambiguity in Poisson Hamiltonian learning","Chart-free kernel recovery of Hamiltonians from noisy data","Differential representer theorem yields convergent Hamiltonian recovery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the skew structure of the system never amplifies a gradient into arbitrarily long vector fields at the sampled points, in a quadratic sense bounded by an $L^1(\\mu_Z)$ function (Assumption 3.7); without this, the estimator can still be computed, but the convergence and flow-approximation theorems do not apply.","fun_headline_variants_meta":{"raw":{"variants":["Kernel regression recovers Hamiltonian from noisy Poisson dynamics","One regression solves Casimir ambiguity in Poisson Hamiltonian learning","Chart-free kernel recovery of Hamiltonians from noisy data","Differential representer theorem yields convergent Hamiltonian recovery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000935,"raw_usage":{"total_tokens":3976,"prompt_tokens":896,"completion_tokens":3080,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":3018}},"tokens_in":512,"tokens_out":3080,"duration_ms":24949,"temperature":1.0,"reasoning_tokens":3018,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:10:40.891853+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a Poisson manifold with nontrivial Casimir directions, take a target Hamiltonian $H$ that is a known kernel section in the source space (as in Section 5.1.3), generate noisy vector-field samples with positive variance, estimate with $\\lambda=cN^{-\\alpha}$, and measure the empirical RKHS error over many independent trials. If, for a compatible structure that violates Assumption 3.7 on the sampling support, the error still decays at the claimed rate with high probability, then the boundedness assumption is not necessary; if the decay breaks, it is confirmed as the load-bearing condition the paper states it to be.","supporting_citations":[],"review_version":1}