{"id":"eaa59cd8-3473-4e8d-9b03-d777bb60c112","arxiv_id":"2608.09404","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A trainable input-dependent contraction metric plus an incremental ISS bound for neural ODEs, presented as a theoretical and protocol contribution without experimental validation.","lead":"A method for training neural differential equations with a learned, input-dependent contraction metric is proposed for controlled systems with unknown disturbances. The paper derives an incremental input-to-state bound and specifies reproducible Duffing and PMSM evaluation protocols, but all numerical results are synthetic placeholders because no experiments were run.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof is incomplete: for distinct inputs, (10) is a linearization, not an identity for x1-x2, and the metric's total derivative requires a virtual-flow path; the theorem is not established as written.","rationale":"The reader's verdict is CONDITIONAL, and that is the right severity, but the reason should be sharpened. The weakest point is not only that the sampled loss cannot verify the contraction premise; it is that the proof of Theorem 1, as written, does not add up even under the ideal premise. Equation (10) is introduced as 'variational dynamics' but is then used as a statement about δx = x1 - x2. The difference of two nonlinear vector fields with different inputs is not the Jacobian expression; a remainder appears that no assumption in the theorem bounds. The standard way to make contraction arguments with inputs work is to consider the virtual system with the input linearly interpolated in s; then δ = ∂x/∂s satisfies the variational equation exactly, and the metric derivative in Eq. (5) is well-defined along that virtual trajectory. The proof should be rewritten this way, and the premise must then be quantified over all s and all input derivatives in the convex hull. This is a repair, not a rejection: the theorem is plausible and likely recoverable. However, because the repair changes the statement of the premise and the definition of the path, the central claim is not established in the submitted text. The paper's transparency about placeholders and its explicit Remark 1 lower the risk of overclaiming, but the mathematical gap and the lack of any executed certificate keep the paper at CONDITIONAL. If the proof repair is done and a verification procedure (e.g., SOS or interval bounding) is applied to the learned metric, the contribution could become ACCEPT; without it, the theorem should not be relied upon.","tokens_in":1011,"tokens_out":1450,"duration_ms":147691,"concrete_test":"Independently complete the proof using the virtual system x_dot(s,t) = f(x(s,t), u_s(t)), u_s = (1-s)u1 + s u2, with δ = ∂x/∂s and V_s = δ^T M(x,u_s)δ. Check whether differentiating V_s and using the pointwise inequality at (x,u_s) yields (14) for every s and then integrating over s gives (12). If yes, Theorem 1 is true but needs a rewritten statement with the virtual-flow premise and a bound on u_s derivatives; if no, produce a two-dimensional counterexample with M depending on u to show the missing term. Also recompute b_u as sup over the convex hull of U, not just at training points.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution is Theorem 1's incremental ISS bound, and the proof is not valid as written. It sets V = δx^T M δx and quotes the 'variational dynamics' δx_dot = Aθ δx + Bθ δu + Eδw. For two finite trajectories with different inputs, δx = x1 - x2 does not satisfy this equation: the exact difference f(x1,u1)-f(x2,u2) is a path integral of Aθ and Bθ over the connecting segment, and the pointwise inequality M_dot + Aθ^T M + M Aθ <= -2λM does not control the omitted remainder. If δx is instead a tangent vector along a path, then the conclusion must be a bound on the path integral of sqrt(V), and the proof must define the path as a virtual flow x_dot(s,t) = f(x(s,t), u_s(t)) with u_s(t) = (1-s)u1(t) + s u2(t). Along such a path Eq. (5) is well-defined, but the theorem's premise must be imposed for every s and for the interpolated input derivative (1-s)u1_dot + s u2_dot. The manuscript neither states this construction nor verifies the strengthened premise. Because u_dot enters the inequality and the training loss (7) samples it only at collocation points, the theorem's premise is not certified for arbitrary admissible inputs. The bound (12) therefore remains conditional on a proof repair and on a domain/input-rate specification that is absent.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Adaptive Stability-Constrained Neural Differential Equations (AS-NDE), a continuous-time controlled neural model that jointly learns a vector field f_theta(x,u) and an input-dependent Riemannian metric M_phi(x,u). Positive definiteness of the metric is enforced by a Cholesky-type construction with a floor term, and a sampled differential inequality penalizes violations of a prescribed contraction rate. The main theoretical claim is Theorem 1, an incremental input-to-state bound stating that the Riemannian distance between two trajectories of the learned model decays exponentially up to gains proportional to the differences in their controls and disturbances. An experimental protocol is specified for a forced Duffing oscillator and a PMSM model, with all numerical tables and figures explicitly labeled as illustrative placeholders pending actual experiments.","tokens_in":1584,"tokens_out":4106,"duration_ms":106050,"significance":"If Theorem 1 is established rigorously, the paper would provide a useful quantitative sensitivity bound for learned controlled neural ODEs, explicitly accounting for the total derivative of an input-dependent metric, a term often omitted in heuristic stability regularizers. The paper is unusually transparent: it separates the method from the unexecuted experiments, labels all placeholder data as such, and includes a thoughtful threats-to-validity section. These are strengths. However, the central proof as written is incomplete, and the premise of the theorem is never certified for the learned model; therefore the contribution is currently a plausible methodology with a conditional guarantee, not a proven result. The reproducibility protocol and the honest treatment of limitations are commendable and should be preserved in a revised version.","major_comments":[{"comment":"The proof states that for two solutions with different inputs (u1,w1) and (u2,w2), the variational dynamics are delta x_dot = A_theta delta x + B_theta delta u + E delta w, and then differentiates V = delta x^T M delta x. For delta x = x1 - x2, this equation is not exact: the difference f(x1,u1) - f(x2,u2) is a path integral of A_theta and B_theta over the connecting segment plus a remainder that is not controlled by the pointwise contraction inequality. To make the argument rigorous, one must define a virtual flow x_s(t) satisfying x_s_dot = f(x_s,u_s) + E w_s with u_s = (1-s)u1 + s u2 and w_s = (1-s)w1 + s w2, and set delta x_s = partial x_s / partial s; only then is the linearized equation exact. The proof as written is a linearization argument and does not establish the claimed bound for finite trajectory differences. This is the paper's central contribution, so the proof must be repaired and the virtual-flow construction stated explicitly.","section":"Section V, Eq. (10) and proof of Theorem 1"},{"comment":"The contraction inequality is assumed to hold throughout a forward-invariant domain, but because the metric depends on u, its total derivative in Eq. (5) contains the term partial M / partial u times u_dot. When comparing two different input trajectories, the connecting path must specify the input derivative u_s_dot = (1-s) u1_dot + s u2_dot, and the inequality must hold for all s in [0,1] along the virtual flow. The manuscript does not state this strengthened premise, and the training loss (7) only samples the inequality at a finite set of collocation points. Consequently, Theorem 1's bound is conditional on a property that is never certified for the learned model. The paper's own Remark 1 concedes this point, but the theorem as stated is not established for arbitrary admissible inputs unless the premise is explicitly extended to the virtual paths and verified.","section":"Section V, Eq. (5) and Theorem 1 premise"},{"comment":"Because the contraction loss L_ctr directly penalizes lambda_max(S), any violation rate reported for AS-NDE is a fitted quantity and cannot serve as independent evidence that the theorem's premise holds. The certificate audit described in Section VI-E is the appropriate check, but it must report the full distribution of lambda_max(S) on dense held-out grids and exterior shells, together with the metric condition numbers, rather than a single violation percentage. The placeholder numbers in Table I are correctly labeled as illustrative, but the relationship between the penalty and the reported violation metric should be made explicit so that readers do not interpret low violation rates as a verification of the theoretical condition.","section":"Section VI-E and Table I"}],"minor_comments":[{"comment":"The second line of the Duffing equation contains a stray 'quad' that should be removed.","section":"Section VI-A, Eq. (15)"},{"comment":"The notation 'D+ sqrt(V)' is used without definition; please define the upper Dini derivative and spell out the integration step from the differential inequality to Eq. (12).","section":"Section V, proof of Theorem 1"},{"comment":"The Riemannian distance d_M(t) is computed between points (x1(t),u1(t)) and (x2(t),u2(t)) using a metric that depends on both arguments; the paper should define this distance explicitly as the infimum over smooth paths in the joint (x,u) space to avoid ambiguity.","section":"Section V, Theorem 1 statement"},{"comment":"The phrase 'four-layer multilayer perceptron' is redundant; replace it with 'four-layer perceptron' or 'four-layer MLP' for clarity.","section":"Section VI-D"}],"recommendation":"major_revision","confidential_remarks":"The paper is a draft with explicitly labeled placeholder results, which is unusual but honest. The main issue is the proof of Theorem 1: it requires a virtual-flow construction and a strengthened, explicitly stated premise about the input-dependent metric along interpolated paths. This is a standard repair that should be within the authors' capability. The paper's scope and topic are suitable for the journal, and the transparency about missing experiments is commendable. I recommend major revision rather than reject, because the central idea is plausible and the proof gap is fixable without changing the method's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take: the paper's central theorem is not proved as written. Equation (10) is a linearization along a tangent path; it isn't the exact evolution of the finite difference x1 - x2 for two different inputs. The metric's total derivative in (5) also requires a well-defined input path, and the paper never defines the interpolated input between u1 and u2. So Theorem 1's bound is a conjecture unless the proof is repaired with a virtual-flow construction.\n\nThe paper does several things well. It is transparent to the point of labeling every table entry and curve as a placeholder, and it lays out a detailed protocol with ablations, threats to validity, and a certificate audit. That is the right way to write a methods draft. The idea of an input-dependent contraction metric that keeps the total derivative in training is genuinely new relative to ICODE, CSODE, and control contraction metrics. The paper also correctly distinguishes a sampled penalty from a formal certificate.\n\nThe soft spots, in order: the proof gap is the biggest; it may be fixable but must be stated. Second, no experiments, so nothing validates the training procedure actually satisfies the inequality on a meaningful set. Third, the u_dot term is underspecified; the theorem needs a domain condition on input rate, not just a collocation penalty.\n\nThis is a paper for people working on learned contraction metrics and stability-constrained NODEs. I'd take it as a starting point for a discussion, not as an established result. Send it to review? Yes, because the core idea is useful and the gap is repairable, but the review should be strict about the proof.\n\nRecommendation: engage, but don't cite the bound as fact until the proof is fixed and experiments are run.","headline":"The theorem isn't proven as written, but the transparent protocol and the input-conditioned metric idea make it worth a rigorous referee.","tokens_in":9183,"tokens_out":3004,"would_cite":false,"duration_ms":30631,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["34D23","93D25","93B30","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"An input-dependent contraction metric gives controlled neural ODEs a quantitative bound on prediction sensitivity to differing controls and disturbances.","keywords":["neural ordinary differential equations","contraction analysis","incremental stability","input-to-state stability","nonlinear system identification","permanent-magnet synchronous motor","input-dependent metric","unknown disturbances"],"falsifier":"Train the model on the forced Duffing benchmark exactly as the protocol specifies, then from the same initial condition run two rollouts under different test command sequences and disturbances and compare the measured Riemannian distance to the right side of (12); a persistent overshoot of the bound would show that the contraction premise, or the way it is enforced, does not hold.","tokens_in":8107,"feed_emoji":"⚙️","tokens_out":10378,"duration_ms":101965,"temperature":0.7,"pith_summary":"This paper proposes an adaptive stability-constrained neural differential equation (AS-NDE) that jointly learns a controlled vector field and a state- and input-dependent Riemannian metric, so that predicted trajectories stay close even when the inputs used at rollout differ from those seen in training. Its central analytical claim is Theorem 1, an incremental input-to-state contraction bound: the metric distance between two trajectories decays exponentially at rate $\\lambda$, up to additive gains from differences in the measured controls and the unknown disturbances. The paper's distinctive move is to keep the total derivative of the input-dependent metric, including the $\\partial M/\\partial u$ term, inside the training residual, a term it says heuristic stability regularizers omit. The draft is explicit that its numerical tables and curves are illustrative synthetic placeholders and that no measured data or executed training runs were available, so the contribution is the method and the proof, not empirical evidence. A sympathetic reader would care because, if the theorem's premise can be made to hold, neural models of nonlinear systems would come with a quantitative prediction-sensitivity guarantee instead of an ad hoc regularizer.","feed_headline":"Learned stability metric bounds rollout error under unseen inputs","feed_subtitle":"A neural differential equation with a learned Riemannian metric keeps predictions close when controls and disturbances differ.","key_machinery":"The load-bearing object is the input-conditioned Riemannian metric $M_\\phi(x,u) = L_\\phi(x,u) L_\\phi(x,u)^\\top + \\epsilon_M I$, where $L_\\phi$ is a lower-triangular network output whose diagonal uses softplus plus a floor, so positive definiteness is guaranteed by construction. Alongside it, the contraction residual $S_\\Theta = \\dot M + A_\\theta^\\top M + M A_\\theta + 2\\lambda M$ is evaluated at collocation points, with the total derivative $\\dot M = \\sum_i \\partial M/\\partial x_i f_{\\theta,i} + \\sum_j \\partial M/\\partial u_j \\dot u_j$ retained. Training penalizes $\\mathrm{softplus}(\\lambda_{\\max}(S_\\Theta)/\\tau)\\tau$, and Theorem 1's premise is exactly the negative semidefiniteness of $S_\\Theta$ on the whole domain. The metric's job is to define a distance in which the variational dynamics are contractive; the residual's job is to encode the differential inequality that makes the exponential bound hold.","core_discovery":"The core discovery is Theorem 1. If Assumption 1 holds and the contraction inequality $\\dot M + A_\\theta^\\top M + M A_\\theta \\preceq -2\\lambda M$ is satisfied throughout a forward-invariant domain, then the Riemannian distance between two trajectories of the learned model obeys $d_M(t) \\le e^{-\\lambda t} d_M(0) + \\int_0^t e^{-\\lambda(t-\\tau)} [b_u \\|u_1-u_2\\| + b_w \\|w_1-w_2\\|]\\,d\\tau$, with gains $b_u = \\sup \\|M^{1/2} B_\\theta\\|_2$ and $b_w = \\sup \\|M^{1/2} E\\|_2$. The Euclidean bound (13) follows from metric equivalence. The deliberate inclusion of the total derivative of the metric, especially the $\\partial M/\\partial u$ term, is what makes the inequality correct for an input-dependent metric; the paper identifies this as easily omitted in heuristic regularizers. The stated object of the certificate is the learned model, not the unknown plant; Remark 1 says that a plant-level claim would need a verified bound on the modeling error $f^\\star - f_\\theta$ or independent validation.","pith_inferences":["Not stated in the paper, but the same proof could be re-run with a time-varying rate $\\lambda(t)$ or with metric-dependent gains; the gain structure $b_w/\\lambda$ already suggests a concrete trade-off between disturbance attenuation and stiffness.","Not stated in the paper, but the theorem's premise could be verified ex post on the Duffing and PMSM benchmarks using interval bound propagation, turning the sampled penalty into a real certificate; the paper names this tool in its certificate audit.","Not stated in the paper, but the bound (13) applies only to the learned model; combining it with an independently validated bound on $f^\\star - f_\\theta$ would yield a plant-level prediction-sensitivity guarantee, which is the result an end user would ultimately need.","Not stated in the paper, but the proposed ablations would settle a key ambiguity of the method: whether the adaptive metric itself buys the improvement, or whether the same gain comes from the contraction penalty acting on a fixed identity metric."],"forward_implications":["If the contraction premise can be verified, rollout error under new inputs is no longer an open-ended risk: the Euclidean form (13) gives a closed-form bound in terms of initial mismatch, input mismatch, disturbance mismatch, the rate $\\lambda$, and the metric condition number.","The certificate is for the learned model, not the real plant; the paper explicitly says that claiming robustness of the plant requires an additional verified bound on $f^\\star - f_\\theta$.","The sampled contraction penalty is an empirical regularizer, and a zero violation rate on finite samples is not a proof; the prescribed certificate audit on a dense held-out grid plus an exterior shell is needed before calling the bound a guarantee.","For piecewise-constant commands, the $\\partial M/\\partial u$ term forces a hybrid treatment at jumps, and the paper proposes interval-by-interval checking rather than a single smooth certificate.","A negative result is meaningful under the paper's protocol: if residual violations drop without an improvement in held-out rollout error, the honest conclusion is that the certificate regularizes sensitivity without demonstrated predictive benefit."],"supporting_citations":[{"why":"Supplies the contraction-analysis framework that defines distance contraction between neighboring trajectories.","marker":"[15]"},{"why":"Provides the differential Lyapunov theory that justifies treating M as a state-dependent Lyapunov function for variations.","marker":"[16]"},{"why":"Introduces control contraction metrics, the direct precedent for an input-dependent metric in controlled systems.","marker":"[17]"},{"why":"Defines incremental stability, the pair-of-trajectories notion that Theorem 1 quantifies.","marker":"[18]"},{"why":"Defines input-to-state stability, whose gain structure the disturbance and control terms in (12) follow.","marker":"[19]"},{"why":"Defines the neural ODE class whose rollout sensitivity is the problem being addressed.","marker":"[1]"},{"why":"Provides the SINDy sparse-regression baseline used in the evaluation protocol.","marker":"[8]"},{"why":"An input-concomitant NODE with contraction conditions, the immediate construction whose input interface AS-NDE extends.","marker":"[35]"},{"why":"Shows a scalable NODE can satisfy linear inequalities guaranteeing convergence, a comparative approach for a guaranteed-stability NODE.","marker":"[36]"},{"why":"Provides the continuous-time PMSM differential-neural-network baseline for the application protocol.","marker":"[37]"}],"fun_headline_variants":["Learned metric enforces contraction in neural ODEs under unknown inputs","Adaptive stability-constrained neural ODEs robust to unknown inputs","Neural ODEs with adaptive metric stay close under input shifts","Metric learning gives neural ODEs stability under unknown inputs","Learned metric keeps neural ODE rollout error bounded under unknown inputs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the contraction inequality $\\dot M + A^\\top M + M A \\preceq -2\\lambda M$ holds throughout the forward-invariant domain, including along paths connecting two different input trajectories, yet the training loss only samples this condition at finitely many collocation points and the metric's total derivative is not well-defined for two differing inputs.","fun_headline_variants_meta":{"raw":{"variants":["Learned metric enforces contraction in neural ODEs under unknown inputs","Adaptive stability-constrained neural ODEs robust to unknown inputs","Neural ODEs with adaptive metric stay close under input shifts","Metric learning gives neural ODEs stability under unknown inputs","Learned metric keeps neural ODE rollout error bounded under unknown inputs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00328,"raw_usage":{"total_tokens":12420,"prompt_tokens":1034,"completion_tokens":11386,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":11296}},"tokens_in":650,"tokens_out":11386,"duration_ms":78707,"temperature":1.0,"reasoning_tokens":11296,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:52:53.337303+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the model on the forced Duffing benchmark exactly as the protocol specifies, then from the same initial condition run two rollouts under different test command sequences and disturbances and compare the measured Riemannian distance to the right side of (12); a persistent overshoot of the bound would show that the contraction premise, or the way it is enforced, does not hold.","supporting_citations":[{"cited_title":"A differential Lyapunov framework for contraction analysis,","cited_arxiv_id":null,"evidence_quote":"Provides the differential Lyapunov theory that justifies treating M as a state-dependent Lyapunov function for variations."},{"cited_title":"Control contraction metrics: Convex and intrinsic criteria for nonlinear feedback design,","cited_arxiv_id":null,"evidence_quote":"Introduces control contraction metrics, the direct precedent for an input-dependent metric in controlled systems."},{"cited_title":"Smooth stabilization implies coprime factorization,","cited_arxiv_id":null,"evidence_quote":"Defines input-to-state stability, whose gain structure the disturbance and control terms in (12) follow."},{"cited_title":"Neural ordinary differential equations,","cited_arxiv_id":null,"evidence_quote":"Defines the neural ODE class whose rollout sensitivity is the problem being addressed."},{"cited_title":"Discovering governing equations from data by sparse identification of nonlinear dynamical systems,","cited_arxiv_id":null,"evidence_quote":"Provides the SINDy sparse-regression baseline used in the evaluation protocol."},{"cited_title":"ICODE: Modeling dynamical systems with extrinsic input information,","cited_arxiv_id":null,"evidence_quote":"An input-concomitant NODE with contraction conditions, the immediate construction whose input interface AS-NDE extends."},{"cited_title":"ControlSynth neural ODEs: Modeling dynamical systems with guaranteed convergence,","cited_arxiv_id":null,"evidence_quote":"Shows a scalable NODE can satisfy linear inequalities guaranteeing convergence, a comparative approach for a guaranteed-stability NODE."}],"review_version":1}