{"id":"b4ab3129-c17d-4e5f-93d8-7a945d01fa7c","arxiv_id":"2505.16511","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Contraction-based LMI conditions on learned neural ODE models guarantee convergence of unknown nonlinear systems to a neighborhood of their equilibrium.","lead":"This paper presents a method for controlling nonlinear systems whose dynamics are unknown, by first learning a neural ODE model and then using contraction analysis to certify and enforce stability. It proves the controlled system converges close to the unknown equilibrium, with the error tied to how accurately the neural model matches the real system.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The convergence guarantee in Prop. 3.8 rests on a uniform bound on the approximation error over the forward-invariant set C, but the paper only estimates this error on the finite training set D, so the claimed neighborhood size is unverified.","rationale":"I read the paper in good faith. The theoretical core—contraction of the NODE system plus an ISS-type argument around the unknown equilibrium—is coherent and does not appear internally inconsistent under its stated assumptions. The reader's weakest_assumption identifies the right soft spot: the uniform approximation-error bound on C is load-bearing but is only estimated on D. I would not move the verdict to REJECT because the theory is conditional and the conditional is made explicit; however, the paper currently presents simulations as evidence of the guarantee, and the unregularized optimization in Ex. 4.2 returns µ≈8e-15 and a claimed bound of 3.6e-16, which is far below any plausible resolution of the reported approximation error. This reinforces the need for a certification step. Therefore the CONDITIONAL verdict stands: accept conditional on certifying sup_{x∈C}||ε(x)|| (and ideally forward invariance) and on regularizing or independently validating the LMI-based neighborhood computation.","tokens_in":22417,"tokens_out":19065,"duration_ms":156737,"concrete_test":"For each example, evaluate ε(x)=f(x)-Ax-Znn(x) on a fine grid over C (e.g., 10^6 points in [-2,2]^2 for Ex. 4.1 and [-1,1]^2 for Ex. 4.2) and compute the sup norm; compare with the reported ε values 0.002302 and 0.008532. Then simulate the closed-loop system from many grid initial conditions and record the maximum steady-state distance to the observed equilibrium, comparing with the claimed bounds 0.0239 and 3.56e-16. If either the grid sup error or the simulated steady-state distance exceeds the claimed value by a significant margin, the practical guarantee is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central result, Prop. 3.8, asserts that trajectories of the perturbed system (32) satisfy (37): ||x(t)-x*|| ≤ sqrt(p/p)||x(0)-x*||e^{-γt/2} + (4ε/γ)sqrt(p/p). This requires a known uniform bound sup_{x∈C}||ε(x)|| ≤ ε, with ε(x)=f(x)-Ax-Znn(x). The paper's only mechanism for obtaining ε is the empirical fit (2) over a finite dataset D; there is no validation on a dense grid, no Lipschitz or generalization certificate, and no argument that the reported values 0.002302 (Ex. 4.1) and 0.008532 (Ex. 4.2) dominate the error on the actual operating set C. The same assumption block also requires forward invariance of C for the perturbed closed-loop system, which is not checked. Because the terminal neighborhood in (37), Algorithm 1, and Algorithm 2 scales linearly in ε, an underestimated ε makes the headline 'convergence to a neighborhood determined by the approximation error' vacuous. If ε were certified, the Lyapunov argument would be a standard ISS-type bound; the gap is the unverified step from training loss to worst-case error.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies NN-based control of unknown affine nonlinear systems using contraction analysis. A NODE model is fitted to data, LMI-type sufficient conditions are proposed under which the NODE system is contractive, and it is shown that, if the approximation error is uniformly bounded by a known constant on a forward invariant set, trajectories of the original system converge to a neighborhood of the unknown equilibrium whose size is proportional to that bound. For NODE models that do not satisfy the contraction conditions, the paper proposes a two-part controller with a linear gain and an NN cancellation term, again with LMI reformulations. Two numerical examples, a pendulum and a wheeled-vehicle path-following model, illustrate the procedure.","tokens_in":22660,"tokens_out":12498,"duration_ms":102061,"significance":"If the stated assumptions were fully verified, the paper would provide a useful design workflow for unknown nonlinear systems when the equilibrium is not known a priori. The LMI conditions for contractivity of multi-layer NODEs with incremental sector bounds are a practical contribution, and the convergence-to-neighborhood result is a clean ISS-type Lyapunov argument. The examples are simple but demonstrate the overall architecture. On the other hand, the paper does not provide code or machine-checked proofs, and the central quantity used in the main guarantee, the uniform approximation-error bound, is only estimated on a finite training set rather than certified.","major_comments":[{"comment":"The main convergence guarantee requires a known uniform bound sup_{x in C} ||epsilon(x)|| <= epsilon, but the only numerical mechanism in the paper is the training loss (2) over the finite data set D. The values 0.002302 (Example 4.1) and 0.008532 (Example 4.2) are reported as 'upper bounds' after 5000 epochs, yet no dense verification, Lipschitz/generalization argument, or other certificate is provided that they dominate the error on the forward invariant set C. Because the terminal neighborhood in Eq. (37), Algorithm 1, and Algorithm 2 scales linearly in epsilon, an underestimated epsilon invalidates the headline guarantee. The forward invariance of C for the perturbed closed-loop system is also assumed, not checked. Please either certify the bound on C, restrict all claims to a known epsilon and remove 'yielding an upper bound epsilon' from Algorithm 1 Step 1, or explicitly state that the result is conditional on an unverified uniform error bound.","section":"Section III, Proposition 3.7-3.8, Algorithms 1-2"},{"comment":"The displayed Lyapunov derivative in the proof of Theorem 3.2 contains W^1 in the cross terms, but the derivative of V along the single-layer NODE system (12) contains W^2, because Znn(x) = W^2 phi(W^1 x + b^1) + b^2. The correct cross term appears implicitly in Eq. (15) through L = P W^2 + (alpha1+beta1) W^{1T}. As printed, the proof does not establish the claimed bound; replacing the two W^1 factors by W^2 makes the argument consistent. Please correct the proof.","section":"Theorem 3.2, Eq. (17)"},{"comment":"The reported incremental sector bound 0.9978 <= (phi1(x)-phi1(y))/(x-y) <= 1 for all x,y in [-2,2] is false for tanh, whose derivative at x=2 is sech^2(2) approximately 0.07065. The bound can only be true on a much smaller interval, presumably the range of the pre-activation values W^1 x + b^1, but the text states x in [-2,2]. In addition, the coefficient -102850 in the reported matrix bar(R) is not consistent with the coefficient -2 alpha beta/(beta-alpha)^2 from Theorem 3.12, which for alpha=0.9978, beta=1 is approximately -4.12e5. Please reconcile the numerical values with the formulas in the paper.","section":"Example 4.1, Eq. (68)"},{"comment":"The LMI status of the multilayer controller design is unclear. The term -2 s / lambda S in bar(X) of Eq. (56) is bilinear in the decision variables if s is a decision variable, while the theorem states 'there exist a fixed positive constant s' and the optimization (66) appears to optimize over s. The approximation bar(R) = (AS+gY)^T + AS + gY in Example 4.2, made 'provided that lambda is sufficiently small,' is not justified and the value of lambda is not reported. Since the feasibility of the LMI is the basis for the controller guarantee in the multilayer case, this needs to be clarified and made rigorous.","section":"Theorem 3.15 and Example 4.2, Eq. (56), (70), (72)"}],"minor_comments":[{"comment":"The sentence 'The symbols R, R+ denote the set of real nonnegative real numbers, respectively' should read that R denotes the real numbers and R+ denotes the nonnegative real numbers.","section":"Section II, Notation"},{"comment":"The symbol D is used both for the finite training data set and for the domain on which the incremental sector bound and approximation error are defined. Please introduce a separate symbol for the domain, for example Omega, and state explicitly how C relates to it.","section":"Section II, Definition 2.1 and Eq. (11)"},{"comment":"The notation p and p (underline and overline) is easy to confuse; in Eq. (34) the coefficient sqrt(p) should be the upper bound overline(p) from Eq. (14). Please use a more distinct notation, for example p_min and p_max.","section":"Proposition 3.7, Eq. (34)"},{"comment":"Step 3 in both algorithms solves the unconstrained least-squares problem min ||W^{k+1} + g W_u^{k+1}||; the closed-form solution should be stated, and the algorithm should clarify what happens if g is not injective.","section":"Algorithms 1 and 2"},{"comment":"Please state the numerical value of lambda used in Eq. (70) and explain why the approximation bar(R) ≈ (AS+gY)^T + AS + gY is valid for that value.","section":"Example 4.2"},{"comment":"There are several typographical and grammatical errors; for example, 'the author analyze' should be 'the authors analyze' and 'non-convex conditions can complicate the search' should be 'non-convex conditions can complicate the search for feasible solutions'. A careful proofreading pass is recommended.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The core conditional result is plausible and the Lyapunov argument is standard once the uniform approximation-error bound is granted. However, the gap between the empirical training loss and the certified worst-case error on the forward invariant set is central to the claimed guarantees, and the numerical inconsistencies in the examples would need to be resolved before the paper can be accepted. I recommend asking the authors to address the certification issue head-on, either by adding a genuine certificate or by reformulating the contribution as conditional on a known epsilon, and to clean up the proof typo and the LMI details in Theorem 3.15."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my read on 2505.16511. The genuinely new piece is Proposition 3.8: if the learned NODE is contractive on a forward invariant set C, then trajectories of the original perturbed system converge to a neighborhood of an unknown equilibrium, with radius proportional to the approximation error. That is a real improvement over prior work that assumes a known equilibrium or a Hurwitz linearization. The LMI reformulations for both single- and multi-layer networks are competent, and the controller that enforces contractivity when the NODE fails the conditions is a reasonable extension of approximate nonlinearity cancellation.\n\nThe proof of Prop 3.8 is a standard Lyapunov argument and, as far as I can tell from the garbled equation rendering, it holds under the stated assumptions. The same goes for the sector-bound / LMI derivation: the quadratic constraint in Lemma 2.3 is correct, and the contraction conditions are sensible.\n\nWhere the paper is soft: first, the approximation error bound epsilon is only estimated from the training set via (2), but the convergence guarantee in (37) needs a uniform bound on C, and forward invariance of C for the perturbed system is assumed, not verified. The stress-test note lands: if epsilon is underestimated, the 'neighborhood determined by the approximation error' is vacuous. This is fixable with a denser grid or a Lipschitz-based certificate, but the paper should say how epsilon is computed and what it actually dominates. Second, Example 4.2 reports mu ~ 8e-15 and an ultimate bound of 3.6e-16; that is machine-precision nonsense and it weakens confidence in the simulation. The optimization is clearly degenerating. Third, the LMI feasibility is never guaranteed: the paper gives sufficient conditions, but if those conditions are infeasible for a given learned model, the design simply does not apply. That is acceptable, but it should be stated as a limitation rather than buried.\n\nWho this is for: researchers in contraction-based learning control or NN verification. It deserves a serious referee; with the epsilon certification fixed and the simulations cleaned up, it would be a solid contribution. My recommendation: send it to review, with the error-certification gap as the main concern.","headline":"Useful contraction-based framework for NN control of unknown systems; main guarantee depends on an approximation-error bound that is only estimated on training data, and one simulation reports an implausible machine-precision bound.","tokens_in":23206,"tokens_out":4218,"would_cite":true,"duration_ms":35783,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C10","93D05","93D30","68T07","90C22"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that a learned neural-ODE model, if certifiably contractive, lets a controller drive an unknown nonlinear system into a neighborhood of its unseen equilibrium, with the neighborhood size set by the approximation error.","keywords":["neural ordinary differential equations","contraction analysis","incremental sector bounds","linear matrix inequalities","unknown nonlinear systems","Hopfield neural network","nonlinearity cancellation","unknown equilibrium convergence"],"falsifier":"Train the neural network on data from a region $D$, compute the claimed error bound $\\varepsilon$, and then evaluate $\\sup_{x \\in C\\setminus D} \\|\\varepsilon(x)\\|$ on a larger forward-invariant set $C$; if the empirical supremum exceeds the claimed $\\varepsilon$, then Proposition 3.8's bound $\\|x(t)-x^*\\| \\le (4\\varepsilon/\\gamma)\\sqrt{\\bar{p}/\\underline{p}}$ fails because the derivation uses $\\varepsilon = \\sup_{C}\\|\\varepsilon(x)\\|$. Alternatively, simulate a system whose drift lies outside the training region and check whether trajectories exit the predicted neighborhood of the unseen equilibrium.","tokens_in":22130,"feed_emoji":"🎯","tokens_out":4944,"duration_ms":39350,"temperature":0.7,"pith_summary":"The paper proposes a way to design controllers for nonlinear systems whose drift dynamics are unknown, using contraction analysis instead of knowing the equilibrium. The authors first train a feedforward neural network to approximate the unknown dynamics, producing a Neural ODE model. They impose incremental sector bounds on the activation functions and linear-matrix-inequality (LMI) conditions on the network weights that certify the NODE model is contractive, meaning all its trajectories converge exponentially toward each other. If the NODE model is contractive, they prove that trajectories of the original system converge to a neighborhood of the unknown equilibrium, with the neighborhood's radius proportional to the approximation error. When the NODE model fails the contraction conditions, they design a controller combining a linear gain with an NN compensation term that enforces contractivity and again guarantees convergence to such a neighborhood.","feed_headline":"Neural control reaches a neighborhood of an unseen equilibrium","feed_subtitle":"Contraction certificates let a learned network steer unknown systems without knowing where the equilibrium sits.","key_machinery":"The central object is the pair of incremental sector bounds $[K^i_1, K^i_2]$ on the activation functions, combined with a quadratic differential Lyapunov function $V = (x_1-x_2)^\\top P(x_1-x_2)$. Lemma 2.3 turns the sector bounds into a matrix inequality on the network weights, and Theorems 3.2 and 3.4 convert that into LMI conditions certifying contraction. The Schur complement reformulations in Theorems 3.3 and 3.6 make the conditions convex and hence solvable by standard semidefinite programming. Proposition 3.7 then propagates the approximation error through the contractive comparison argument, and Proposition 3.8 uses the unknown equilibrium as a trajectory of an augmented NODE to derive the explicit neighborhood bound.","core_discovery":"The central claim is that contractivity of the learned NODE system is a workable certificate for controlling unknown nonlinear affine systems without knowing their equilibria. Concretely, if the NODE system $\\dot{x} = Ax + Z_{\\mathrm{nn}}(x)$ is contractive on a forward invariant set $C$, then the original system $\\dot{x} = Ax + Z_{\\mathrm{nn}}(x) + \\varepsilon(x)$ has every trajectory converge to a ball around the unknown equilibrium $x^*$ whose radius is $(4\\varepsilon/\\gamma)\\sqrt{\\bar{p}/\\underline{p}}$, where $\\varepsilon$ is the uniform bound on the approximation error and $\\gamma$ is the contraction rate. This removes the need to know $x^*$ explicitly. For a single-hidden-layer network the NODE becomes a continuous-time Hopfield network, so the same certificate applies there. If the NODE is not contractive, the proposed controller $u(x) = Hx + W_{u,\\min} w_k + b_u$ adds a linear state feedback and a neural-network compensation term that cancels most nonlinearities and restores contractivity, again guaranteeing convergence to a neighborhood of the unknown equilibrium.","pith_inferences":["The same incremental-sector-plus-LMI machinery could plausibly be adapted to discrete-time neural network dynamics or to stochastic disturbance settings, where contraction notions already exist, though the paper does not pursue those extensions.","The neighborhood bound $4\\varepsilon/\\gamma \\sqrt{\\bar{p}/\\underline{p}}$ is global on the set $C$; a testable refinement would use localized error bounds on smaller regions of the state space to obtain tighter radii than a uniform supremum.","Because the equilibrium is never identified, the method offers a form of equilibrium-free stabilization that could be combined with online model adaptation, a direction the authors leave for future work.","A practical deployment step suggested by the results is to run the LMI certificate on a trained network before use, turning the data-fitting step into a verifiable contractivity guarantee."],"forward_implications":["If the NODE model is contractive, the original autonomous system's trajectories converge to a neighborhood of the unknown equilibrium, and the radius shrinks as the approximation error $\\varepsilon$ goes to zero or the contraction rate $\\gamma$ grows.","The contractivity conditions are convex LMIs, so a trained network can be checked for contractivity, and the controller gain can be optimized via a convex program that minimizes the guaranteed neighborhood size.","When the NODE fails the contractivity test, the two-part controller (linear gain plus NN compensation) restores contractivity and still gives convergence to a specified neighborhood of the unknown equilibrium.","For single-layer networks the framework specializes to continuous-time Hopfield networks, giving a direct contractivity and convergence certificate for that popular recurrent architecture.","The bias term of the NN controller is free to choose, so the designer can shift the closed-loop equilibrium, for instance to the origin, without destroying the contraction guarantee."],"supporting_citations":[{"why":"Supplies the contraction-analysis framework used to define and certify contractivity of the NODE system.","marker":"[25]"},{"why":"States the universal approximation theorem that justifies modeling the unknown drift with a feedforward neural network.","marker":"[19]"},{"why":"Analyzes contractivity of continuous-time neural networks, providing the Hopfield-network connection used in the single-layer case.","marker":"[12]"},{"why":"Provides the approximate-nonlinearity-cancellation method that motivates the neural-network controller design.","marker":"[13]"},{"why":"Introduces the quadratic-constraint and incremental-sector-bound technique that Lemma 2.3 extends to vector activations with layer-dependent slopes.","marker":"[15]"},{"why":"Presents the single-hidden-layer ReLU stability analysis and the wheeled-vehicle example used as a comparison and simulation benchmark.","marker":"[29]"}],"fun_headline_variants":["Contraction certificates let NNs control unknown systems","Learn contractivity to reach unseen equilibria","NN tracking without knowing the target equilibrium","Contractive neural nets guarantee safe convergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approximation error $\\varepsilon(x)$ must be uniformly bounded by a known $\\varepsilon$ on the forward-invariant operating set $C$, even though it is only measured on the training set $D$; if the neural network's error is larger on the actual operating region than the bound used, the predicted neighborhood radius is not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Contraction certificates let NNs control unknown systems","Learn contractivity to reach unseen equilibria","NN tracking without knowing the target equilibrium","Contractive neural nets guarantee safe convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1590,"prompt_tokens":1015,"completion_tokens":575,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":521}},"tokens_in":631,"tokens_out":575,"duration_ms":5310,"temperature":1.0,"reasoning_tokens":521,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:59:20.691586+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the neural network on data from a region $D$, compute the claimed error bound $\\varepsilon$, and then evaluate $\\sup_{x \\in C\\setminus D} \\|\\varepsilon(x)\\|$ on a larger forward-invariant set $C$; if the empirical supremum exceeds the claimed $\\varepsilon$, then Proposition 3.8's bound $\\|x(t)-x^*\\| \\le (4\\varepsilon/\\gamma)\\sqrt{\\bar{p}/\\underline{p}}$ fails because the derivation uses $\\varepsilon = \\sup_{C}\\|\\varepsilon(x)\\|$. Alternatively, simulate a system whose drift lies outside the training region and check whether trajectories exit the predicted neighborhood of the unseen equilibrium.","supporting_citations":[{"cited_title":"Multilayer feedforwa rd networks are universal approximators[J]","cited_arxiv_id":null,"evidence_quote":"Supplies the contraction-analysis framework used to define and certify contractivity of the NODE system."},{"cited_title":"Learning controllers from data via ap- proximate nonlinearity cancellation[J]","cited_arxiv_id":null,"evidence_quote":"States the universal approximation theorem that justifies modeling the unknown drift with a feedforward neural network."},{"cited_title":"Modeling and contractivi ty of neural- synaptic networks with Hebbian learning[J]","cited_arxiv_id":null,"evidence_quote":"Analyzes contractivity of continuous-time neural networks, providing the Hopfield-network connection used in the single-layer case."},{"cited_title":"Euclidean cont ractivity of neural networks with symmetric weights[J]","cited_arxiv_id":null,"evidence_quote":"Provides the approximate-nonlinearity-cancellation method that motivates the neural-network controller design."},{"cited_title":"Neural ordina ry differential equations[J]","cited_arxiv_id":null,"evidence_quote":"Introduces the quadratic-constraint and incremental-sector-bound technique that Lemma 2.3 extends to vector activations with layer-dependent slopes."},{"cited_title":"Stable neural ode with lyapu nov-stable equilibrium points for defending against adversarial atta cks[J]","cited_arxiv_id":null,"evidence_quote":"Presents the single-hidden-layer ReLU stability analysis and the wheeled-vehicle example used as a comparison and simulation benchmark."}],"review_version":1}