{"id":"b3c41101-68c9-426f-9c8c-a756ee6bfef9","arxiv_id":"2606.01596","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Randomized Jacobian matching implicitly enforces second-order consistency in learned chaotic vector fields at O(d^2) cost without forming the full Hessian.","lead":"The paper introduces model-constrained randomized Jacobian matching to enforce second-order geometric supervision when learning chaotic dynamical systems from data. This could help models maintain correct long-term attractor geometry and invariant statistics better than first-order methods alone.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest assumption is the accuracy of the Taylor expansion for a chosen variance. Because the tested systems make the expansion exact, that assumption carries no approximation risk and does not threaten the central claim. The derivation is standard multivariable calculus and is realized without remainder on the reported examples.","tokens_in":1794,"tokens_out":300,"duration_ms":37602,"concrete_test":"On a quadratic system, compute the Monte-Carlo estimate of the randomized Jacobian loss for several noise variances and verify that it equals the sum of the nominal Jacobian term and variance \times ||Hess_f − Hess_g||_F^{2} to machine precision; any systematic deviation would indicate an implementation or sampling issue in the decomposition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on a Taylor expansion of the Jacobian under random perturbations. For the quadratic vector fields in the reported experiments (Lorenz 63, Lorenz 96), third- and higher-order derivatives vanish identically, so the expansion J(x + ε) = J(x) + Hess·ε is exact. The expected randomized Jacobian loss therefore equals the nominal Jacobian mismatch plus a term exactly proportional to the squared Hessian mismatch (scaled by perturbation variance) with zero remainder. The second modeling assumption (querying the learned model at perturbed inputs) holds by construction for the neural-network architectures used.","agreement_with_reader":"disagree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes model-constrained randomized Jacobian matching as a way to enforce second-order geometric consistency when learning chaotic vector fields from data. A Taylor expansion of the Jacobian under random perturbations is used to show that the expected loss decomposes into the nominal first-order Jacobian mismatch plus a term proportional to the squared Hessian mismatch scaled by perturbation variance, achieving this supervision at O(d^2) cost without explicitly forming the O(d^3) Hessian. Experiments on Lorenz 63 and Lorenz 96 demonstrate that the approach eliminates failures of first-order methods (catastrophic Lyapunov outliers under minimal supervision; loss of invariant measure under out-of-distribution forcing) while matching the performance of explicit Hessian matching at lower cost.","tokens_in":1902,"tokens_out":510,"duration_ms":24279,"significance":"If the central claim holds, the work supplies a practical, scalable route to second-order supervision for data-driven modeling of chaotic dynamics, where preservation of attractor geometry and long-time statistics is essential. The explicit Taylor decomposition, the exactness result for quadratic fields, and the reproducible numerical comparisons on standard benchmarks are strengths that strengthen the contribution.","major_comments":[{"comment":"Abstract and derivation of the randomized loss: the Taylor expansion is exact (remainder zero) only when third- and higher-order derivatives vanish, as occurs for the quadratic Lorenz systems in the experiments; for general C^3 fields the remainder term must be bounded or shown negligible for the chosen perturbation variance, otherwise the implicit second-order enforcement is not guaranteed.","section":"Abstract"},{"comment":"Numerical experiments on Lorenz 96 (out-of-distribution forcing sweep): the separation of methods for F>18 is reported for the invariant measure and Lyapunov spectrum, but the manuscript must specify the number of Monte Carlo samples used to estimate the randomized loss, the perturbation variance schedule, and whether the reported differences are statistically significant across independent trials.","section":"Numerical experiments"}],"minor_comments":[{"comment":"The notation for the perturbation variance (denoted ε or σ in different places) should be unified and its scaling with dimension d discussed explicitly.","section":null},{"comment":"A short paragraph on the computational complexity of Jacobian-vector products versus full Jacobian evaluations would clarify the claimed O(d^2) advantage.","section":null}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the positive assessment and constructive comments. We address each major comment below.","responses":[{"response":"We agree that the Taylor expansion of the Jacobian is exact (remainder zero) only for quadratic fields. The derivation isolates the leading Hessian term in the expectation under random perturbations, which holds exactly for the quadratic Lorenz systems used throughout the experiments. For general C^3 vector fields a remainder involving third derivatives scaled by the cube of the perturbation size appears. We will revise the manuscript to state this limitation explicitly, note that sufficiently small perturbation variance renders the remainder negligible relative to the O(σ^{2}) Hessian contribution, and add a brief remark on how the variance is chosen in practice to maintain this regime.","revision_made":"yes","referee_comment":"[Abstract] Abstract and derivation of the randomized loss: the Taylor expansion is exact (remainder zero) only when third- and higher-order derivatives vanish, as occurs for the quadratic Lorenz systems in the experiments; for general C^3 fields the remainder term must be bounded or shown negligible for the chosen perturbation variance, otherwise the implicit second-order enforcement is not guaranteed."},{"response":"We will revise the numerical-experiments section to report the exact number of Monte Carlo samples used to estimate the randomized Jacobian loss, the perturbation-variance schedule, and the outcome of statistical-significance tests (e.g., standard errors or p-values) computed across independent random seeds to substantiate the separation observed for F>18.","revision_made":"yes","referee_comment":"[Numerical experiments] Numerical experiments on Lorenz 96 (out-of-distribution forcing sweep): the separation of methods for F>18 is reported for the invariant measure and Lyapunov spectrum, but the manuscript must specify the number of Monte Carlo samples used to estimate the randomized loss, the perturbation variance schedule, and whether the reported differences are statistically significant across independent trials."}],"tokens_in":1478,"tokens_out":417,"duration_ms":23401,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core contribution is a randomized perturbation scheme that matches Jacobians at noisy inputs to implicitly penalize Hessian mismatch. A Taylor expansion of the vector field shows the expected loss splits into the usual first-order term plus a Hessian term scaled by perturbation variance, all without ever forming the full third-order tensor.\n\nThis is new relative to standard trajectory or Jacobian matching in the literature. The experiments back the claim on concrete systems. For Lorenz 63 with sparse temporal data, first-order models produce outlier Lyapunov exponents while the second-order version stays stable and recovers the correct attractor. On Lorenz 96, an out-of-distribution forcing sweep shows first-order methods diverge past F=18 on the invariant measure and spectrum, whereas randomized Jacobian matching holds up and matches the cost of explicit Hessian supervision at far lower expense.\n\nThe math is clean for the reported cases because the Lorenz vector fields are quadratic, so the expansion is exact with zero remainder. The method also scales naturally to higher dimensions since it only needs Jacobian evaluations.\n\nA minor limitation is that the benefit is demonstrated only on polynomial systems where higher derivatives vanish; the approximation quality for general nonlinear fields will depend on perturbation size and could leave residual higher-order error. The paper would be stronger with at least one non-quadratic test case and more detail on variance selection.\n\nThis is for people working on data-driven modeling of chaotic dynamics who already use Jacobian supervision and want better long-term statistics without cubic cost. The derivation is direct, the experiments are targeted, and there is no circularity in the loss. It deserves peer review.","headline":"Randomized Jacobian matching adds cheap second-order supervision for chaotic vector fields via Taylor, with clear gains on Lorenz attractor stats over first-order baselines.","tokens_in":2358,"tokens_out":390,"would_cite":false,"duration_ms":15853,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Randomized Jacobian matching at perturbed inputs implicitly enforces second-order consistency for learned chaotic vector fields.","keywords":["chaotic dynamics","Jacobian matching","second-order supervision","randomized perturbations","attractor preservation","Lyapunov exponents","Lorenz systems","numerical ODE methods"],"falsifier":"A numerical verification that the difference in randomized Jacobian loss between two models exactly matches the predicted Hessian-mismatch term when Jacobian mismatch is held fixed and perturbation variance is varied.","tokens_in":2706,"feed_emoji":"🌀","tokens_out":641,"duration_ms":21560,"temperature":0.7,"pith_summary":"Learning chaotic dynamical systems requires models that preserve attractor geometry and invariant statistics over long times, not just short-term accuracy. Trajectory and Jacobian matching ensure correct values and tangents at supervised points but leave the curvature of the vector field unconstrained, allowing drift to incorrect attractors. The paper establishes that second-order consistency fixes this failure mode. It introduces model-constrained randomized Jacobian matching, which compares Jacobians of the true and learned fields at randomly perturbed states. A Taylor expansion shows the expected loss decomposes into the first-order mismatch plus a scaled second-order term, delivering the required supervision at quadratic cost.","feed_headline":"Randomized Jacobians add second-order supervision to chaotic models","feed_subtitle":"The loss decomposition recovers correct attractors and statistics in Lorenz systems where first-order matching fails, at quadratic rather th","key_machinery":"model-constrained randomized Jacobian matching, which compares Jacobians of the true and learned vector fields at randomly perturbed inputs to capture curvature via the Taylor decomposition.","core_discovery":"The expected randomized Jacobian loss decomposes into the nominal Jacobian mismatch plus a Hessian mismatch scaled by the noise variance, implicitly enforcing second-order consistency at O(d^2) cost without forming the O(d^3) Hessian tensor.","pith_inferences":["The same randomized-loss construction could be tested on other high-dimensional systems where explicit curvature information is unavailable but model evaluations remain cheap.","If the perturbation variance must be tuned per system, the method may require an additional validation step not described in the core argument.","The approach assumes the learned model remains queryable at arbitrary off-trajectory points, which may restrict its use when the model is defined only on observed data."],"forward_implications":["In Lorenz 63, second-order supervision eliminates catastrophic Lyapunov-exponent outliers that appear under first-order methods with minimal temporal data.","In coupled Lorenz 96, second-order methods preserve the invariant measure and Lyapunov spectrum at forcing values beyond F=18 where first-order methods diverge.","The randomized approach achieves performance comparable to explicit Hessian matching while remaining feasible in high dimensions.","Only Jacobian evaluations are required, avoiding the prohibitive cost of forming the full Hessian tensor."],"fun_headline_variants":["Second-order consistency at O(d^2) cost via randomized Jacobian matching","Randomized Jacobians enforce second-order chaotic dynamics without Hessian","Loss decomposition captures Hessian mismatch through randomized Jacobians","Implicit second-order supervision scales chaotic learning beyond first-order","Jacobian randomization recovers attractor geometry in chaotic dynamics"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The Taylor expansion of the vector field around the nominal point accurately captures the second-order contribution for the chosen perturbation variance.","fun_headline_variants_meta":{"raw":{"variants":["Second-order consistency at O(d^2) cost via randomized Jacobian matching","Randomized Jacobians enforce second-order chaotic dynamics without Hessian","Loss decomposition captures Hessian mismatch through randomized Jacobians","Implicit second-order supervision scales chaotic learning beyond first-order","Jacobian randomization recovers attractor geometry in chaotic dynamics"]},"model":"grok-4.3","cost_usd":0.004992,"raw_usage":{"total_tokens":2462,"prompt_tokens":715,"num_sources_used":0,"completion_tokens":77,"cost_in_usd_ticks":49924500,"prompt_tokens_details":{"text_tokens":715,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1670,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":715,"tokens_out":77,"duration_ms":12490,"temperature":1.0,"reasoning_tokens":1670,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T14:04:47.336024+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A numerical verification that the difference in randomized Jacobian loss between two models exactly matches the predicted Hessian-mismatch term when Jacobian mismatch is held fixed and perturbation variance is varied.","supporting_citations":[],"review_version":1}