Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Learning Stable Controlled Dynamical Systems via Input-Contraction Neural Differential Models

T0 review · 3 major / 5 minor · reviewed 2026-07-11 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A neural model can learn controlled continuous-time dynamics that remain incrementally stable under time-varying inputs by jointly fitting the vector field and a Riemannian contraction metric.

desk verdict Solid methods paper that packages input-aware Neural ODEs with a learnable contraction metric and shows real long-horizon gains on forced oscillators and a physical PMSM; the soft-to-hard gap on Lc is real but does not sink the work. read the letter →

arxiv 2607.05718 v1 pith:5F5B4HRV submitted 2026-07-07 eess.SY cs.SY

Syed Pouladi This is my paper · ORCID
classification eess.SYcs.SY
keywords NeuralOrdinaryDifferentialEquationsContractionTheorySystemIdentificationPermanentMagnetSynchronousMotorPhysics-InformedMachineLearningInput-to-StateControlledDynamicalSystems
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper shows how to train continuous-time neural models of systems driven by external controls so that the learned dynamics stay stable over long prediction horizons. Prior neural ODEs either treat inputs without stability guarantees or only guarantee contraction when inputs are constant or vanishing. The Input-Contraction Neural Differential Model embeds an input encoder, parameterizes the non-autonomous vector field, and simultaneously learns a state-dependent positive-definite metric that enforces contraction even while the input changes. Theory proves that when the contraction condition holds, trajectories driven by the same input converge exponentially, and state mismatch under different bounded inputs remains controlled by an input-to-state bound. Experiments on chaotic oscillators and a physical permanent-magnet motor drive show order-of-magnitude lower long-horizon rollout error and slower error growth under input noise than existing neural differential baselines. Anyone who needs reliable long-horizon simulators for motors, robots or vehicles has a concrete reason to care: the model no longer drifts or explodes when controls keep changing.

What carries the argument

The input-dependent contraction matrix Ψ(x,u) assembled from the state Jacobian of the neural field, the learned metric M_ω(x) (kept positive-definite by a Cholesky network) and the metric's time derivative; a soft spectral barrier loss drives the largest eigenvalue of Ψ+βM below zero during training, supplying the sufficient condition used by the two theorems.

What would settle it

After training, evaluate the largest eigenvalue of Ψ(x,u)+βM(x) on a dense grid of states and inputs drawn from the claimed domain; if it is positive on a positive-measure set, or if 200-step rollout MSE on held-out Duffing, Van der Pol or PMSM trajectories is not substantially lower than unconstrained Neural ODEs, the central claim fails.

Watch

Extended reading notes

Core claim

The Input-Contraction Neural Differential Model learns both a non-autonomous neural vector field and a generalized Riemannian metric so that the closed system is input-to-state contractive: under identical inputs any two trajectories converge exponentially at a rate set by the metric, and under bounded input differences the state mismatch is globally bounded by a decaying initial-error term plus a term proportional to the supremum input mismatch.

Load-bearing premise

Training must drive the soft spectral penalty all the way to zero so the hard contraction inequality actually holds at every state and input in the operating domain.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes the Input-Contraction Neural Differential Model (ICNDM), a Neural ODE architecture that jointly learns a non-autonomous vector field f_ heta(x, E_ heta(u)) and a state-dependent Riemannian metric M_ heta(x) (via Cholesky factorization) so that the generalized contraction matrix Ψ(x,u) satisfies a soft spectral barrier. Under the hypothesis that the barrier residual vanishes, Theorem 1 recovers standard incremental exponential stability for trajectories sharing the same input, and Theorem 2 sketches an input-to-state contraction bound under bounded input mismatch. Empirical comparisons on forced Duffing and Van der Pol oscillators and on a physical PMSM drive dataset show 1–2 orders of magnitude lower 200-step rollout MSE and slower error growth under input noise relative to Standard NODE, Controlled NODE, ICODE, ControlSynth, and DNN baselines.

Significance. If the soft barrier is driven sufficiently close to zero that the hard inequality of Theorem 1 holds on the operating domain, the work supplies a practical route to input-aware, incrementally stable neural models for controlled systems—an important gap between autonomous contraction Neural ODEs and purely heuristic controlled Neural CDEs. The PMSM experiment and the noise-robustness curves are concrete, falsifiable demonstrations that would be useful to the data-driven control community. The theoretical contribution itself is largely a standard Riemannian contraction argument plus a sketched ISC extension; its value therefore hinges on whether the learned networks actually satisfy the premise that is claimed to explain the empirical gains.

major comments (3)
  1. The central formal claim (Theorems 1–2) is conditioned on L_c = 0, i.e., Ψ(x,u) ⪰ -β M_ heta(x) everywhere on the operating set. The only enforcement mechanism is the soft spectral penalty (Eq. 9). Nowhere in §VI or the tables is the post-training residual of λ_max(Ψ + β M) (or any proxy such as max residual on a dense grid, fraction of violating samples, or histogram of eigenvalues) reported. Without that certificate it is unknown whether the deployed networks satisfy the hypothesis of the theorems or merely receive a mild regularizing effect; the causal link between “input-contraction” and the reported long-horizon gains therefore remains unproven.
  2. Theorem 2’s proof is only sketched (“leads directly to the input-to-state contraction bound”). The differential inequality for the Riemannian distance under simultaneous state and input mismatch is not written out, nor is the precise role of the input-Lipschitz constant L_u made rigorous. A complete derivation (or an explicit reference to a standard ISC lemma that applies verbatim) is needed before the bound can be treated as established.
  3. Table I and Fig. 1 compare ICNDM only against unconstrained or differently regularized Neural ODEs. There is no ablation that isolates the contribution of the contraction penalty (e.g., ICNDM with λ_c = 0 versus λ_c > 0, or versus a fixed Euclidean metric). Consequently it is impossible to tell whether the 1–2-order rollout improvement is produced by the contraction regularizer, by the input encoder, by capacity differences, or by hyper-parameter tuning.
minor comments (5)
  1. §IV-B heading contains a duplicated word: “Joint Neural Metric Metric Formulation”.
  2. Assumption 1 and the metric bounds α_1, β appear only after the architecture is defined; a forward reference or earlier statement would improve readability.
  3. The encoder latent dimension k, network widths, and the concrete values of β, λ_c, λ_u, ε used for each experiment are never listed; reproducibility is therefore limited.
  4. Fig. 1 caption claims “500-step” rollout while the surrounding text and Table I speak of 200-step horizons; the discrepancy should be resolved.
  5. Several self-citations ([1]–[3]) are to very recent or concurrent works; a short paragraph clarifying the precise technical delta relative to ControlSynth and ICODE would help the reader.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity; theorems are conditional on a soft regularizer that is not forced to zero by architecture, and empirical gains are measured against external baselines and physical data.

full rationale

The derivation chain is self-contained and non-circular. Theorem 1 states that if the spectral loss Lc equals zero (i.e., Ψ(x,u) ⪯ −β Mω(x) holds), then incremental exponential convergence follows from the standard Riemannian Lyapunov argument (V = δxᵀ M δx, V̇ ≤ −β V). Theorem 2 extends this to an input-to-state bound under a Lipschitz assumption on the input channel. Both results are classical contraction theory applied under an explicit hypothesis; they do not redefine the hypothesis in terms of the conclusion. The architecture (input encoder + Cholesky metric network) does not enforce the hypothesis by construction; Lc is only a soft barrier inside the multi-objective loss (Eq. 11). The paper never claims that the trained networks satisfy Lc = 0, nor does it present any post-training residual of λmax(Ψ + βM) as a “prediction.” Empirical claims (lower 200-step MSE on Duffing/Van der Pol/PMSM, slower growth under input noise) are obtained by direct comparison with five external baselines on held-out trajectories and a physical dataset; they are not algebraically forced by any fitted constant. Citations [1–3] are to distinct prior frameworks used only as baselines, not as uniqueness theorems or hidden premises that close a self-referential loop. No ansatz is smuggled, no known empirical pattern is merely renamed, and no quantity is fitted then re-labeled a first-principles prediction. The only residual concern is that the formal guarantees may not attach to the deployed networks (because Lc is never certified), but that is a verification gap, not circularity. Score 1 reflects the minor presentational over-statement that regularization “ensures” the property, while the mathematics itself remains non-circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central stability claims rest on classical contraction theory plus the unproven practical attainment of a soft spectral constraint, uniform metric bounds, and Lipschitzness in the input. Free hyperparameters (β, λ_c, λ_u, ε, architectures) are chosen by the authors and directly shape whether the theorems apply. No new physical entity is postulated; the invented objects are architectural.

free parameters (4)
  • target contraction rate β
    Hand-chosen positive scalar that defines the desired exponential rate in Lc and Theorems 1–2; not derived from data or first principles.
  • regularization weights λ_c, λ_u
    Trade-off coefficients in the multi-objective loss (Eq. 11); fitted/tuned to balance prediction vs. contraction vs. input sensitivity.
  • metric floor ε
    Fixed positive constant in the Cholesky factorization M=LL^T+εI ensuring positive-definiteness; chosen by hand.
  • encoder latent dimension k and network widths
    Architectural sizes of E_ϕ, f_θ, and L_ω are free design choices that affect capacity and the attainable contraction residual; not reported numerically.
assumptions (4)
  • standard math Classical differential contraction: if Ψ ⪯ −βM with α1 I ⪯ M ⪯ α2 I then trajectories converge exponentially under a common input (Lohmiller–Slotine).
    Invoked as the backbone of Theorem 1; standard in the field.
  • domain assumption f_θ is globally Lipschitz continuous in u and C1 in x on the domains of interest.
    Stated in §III-A and used for Theorem 2’s Lu bound; not verified for the trained networks.
  • domain assumption The learned metric network remains uniformly bounded α1 I ⪯ M_ω(x) ⪯ α2 I on X (Assumption 1).
    Required for converting Riemannian decay into Euclidean exponential bounds; only structurally encouraged by εI, not proved for the trained L_ω.
  • ad hoc to paper Soft spectral penalty Lc≈0 after training is sufficient for the hard inequality of Theorem 1 to hold on the operating set.
    The theory assumes Lc=0; training only minimizes a soft max(0,·)^2 expectation. No residual certificate is given.
invented entities (2)
  • Input-Contraction Neural Differential Model (ICNDM)
    purpose: Joint architecture of input encoder E_ϕ, neural field f_θ, and Cholesky metric network M_ω trained under Ψ-regularization.
    The named model is the paper’s primary construct; independent evidence is limited to the reported numerical/PMSM experiments.
  • Input-dependent generalized contraction matrix Ψ(x,u)
    purpose: Combines state Jacobian of the controlled field with the metric and its Lie derivative along the flow to encode non-autonomous contraction.
    Standard CCM-style object specialized to the neural setting; not a new physical quantity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Stable Controlled Dynamical Systems via Input-Contraction Neural Differential Models." pith.science (2026). https://pith.science/paper/5F5B4HRV

@misc{pith2026260705718,
  author       = {Pith},
  title        = {Pith review of: Learning Stable Controlled Dynamical Systems via Input-Contraction Neural Differential Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5F5B4HRV}},
  note         = {Machine review of arXiv:2607.05718}
}
read the original abstract

Learning continuous-time representations of dynamical systems from observation data has emerged as a cornerstone of data-driven control and scientific machine learning. However, existing neural differential equations either treat external control inputs heuristically without providing strict structural guarantees, or enforce stability properties under the restrictive assumption of constant or vanishing inputs. This paper proposes the Input-Contraction Neural Differential Model (ICNDM), a novel deep learning framework that seamlessly incorporates time-varying control inputs while ensuring incremental exponential convergence via input-dependent contraction regularization. By leveraging an embedded input encoder and a parameterized metric network, the proposed architecture learns both the non-autonomous neural vector fields and a generalized Riemannian contraction metric simultaneously. We derive sufficient conditions for input-dependent contraction and formally establish an input-to-state contraction property under bounded external excitations. Extensive numerical evaluations on highly nonlinear chaotic oscillators and experimental data from a Permanent Magnet Synchronous Motor (PMSM) drive system demonstrate that ICNDM yields substantial reductions in long-horizon rollout errors and exhibits superior structural robustness against input perturbations compared with state-of-the-art neural differential benchmarks.

Figures

Figures reproduced from arXiv: 2607.05718 by the authors.

Figure 1
Figure 1. Long-horizon trajectory rollout comparison on the PMSM drive system [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Tracking error evolution under increasing input noise standard [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages

  1. [1]

    ICODE: Modeling dynamical systems with extrinsic input information,

    Z. Li, W. Mei, K. Yu, Y . Bai, and S. Li, “ICODE: Modeling dynamical systems with extrinsic input information,”IEEE Trans. Autom. Sci. Eng., 2025, doi:10.1109/TASE.2025.3559462

    work page doi:10.1109/tase.2025.3559462 2025

  2. [2]

    ControlSynth neural ODEs: Modeling dynamical systems with guaranteed convergence,

    W. Mei, D. Zheng, and S. Li, “ControlSynth neural ODEs: Modeling dynamical systems with guaranteed convergence,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 37, 2024

    work page 2024

  3. [3]

    Learning and current prediction of PMSM drive via differential neural networks,

    W. Mei, X. Wang, Y . Lu, K. Yu, and S. Li, “Learning and current prediction of PMSM drive via differential neural networks,”IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 72, no. 3, pp. 489–493, 2025

    work page 2025

  4. [4]

    Design of interval observers for non- autonomous systems under input disturbances,

    D. Efimov and A. Fradkov, “Design of interval observers for non- autonomous systems under input disturbances,”IEEE Transactions on Automatic Control, vol. 64, no. 1, pp. 240–247, 2019

    work page 2019

  5. [5]

    J.-J. E. Slotine and W. Li,Applied Nonlinear Control. Englewood Cliffs, NJ: Prentice Hall, 1991

    work page 1991

  6. [6]

    Neu- ral ordinary differential equations,

    R. T. Q. Chen, Y . Rubanova, J. Bettencourt, and D. K. Duvenaud, “Neu- ral ordinary differential equations,” inAdvances in Neural Information Processing Systems, 2018, pp. 6571–6583

    work page 2018

  7. [7]

    Neural controlled differential equations for irregular time series,

    P. Kidger, J. Morrill, J. Foster, and T. Lyons, “Neural controlled differential equations for irregular time series,” inAdvances in Neural Information Processing Systems, 2020, pp. 6696–6707

    work page 2020

  8. [8]

    Universal Differential Equations for Scientific Machine Learning

    C. Rackauckas, Y . Ma, J. Martensen, C. Warner, K. Nie, R. Dixit, and A. Edelman, “Universal differential equations for scientific machine learning,”arXiv preprint arXiv:2001.04385, 2020

    work page Pith review arXiv 2001

Show all 31 references
  1. [9]

    On contraction analysis for non-linear systems,

    W. Lohmiller and J.-J. E. Slotine, “On contraction analysis for non-linear systems,”Automatica, vol. 34, no. 6, pp. 683–700, 1998

    1998

  2. [10]

    Control contracting metrics: Gen- eralizing output regulation and Lyapunov stability,

    I. R. Manchester and J.-J. E. Slotine, “Control contracting metrics: Gen- eralizing output regulation and Lyapunov stability,”IEEE Transactions on Automatic Control, vol. 62, no. 11, pp. 5473–5488, 2017

    2017

  3. [11]

    Learning contracting vector fields for stable system identification,

    S. Singh, B. Anisetti, V . Sindhwani, and J.-J. E. Slotine, “Learning contracting vector fields for stable system identification,”IEEE Control Systems Letters, vol. 5, no. 4, pp. 1213–1218, 2020

    2020

  4. [12]

    Sufficient conditions for input-to-state contraction of non-autonomous systems,

    J. Tsinias and C. Papadimitriou, “Sufficient conditions for input-to-state contraction of non-autonomous systems,”Systems & Control Letters, vol. 143, p. 104742, 2020

    2020

  5. [13]

    Contractive neural ordinary differential equations,

    M. Tobia, G. A. De La Torre, and M. Egerstedt, “Contractive neural ordinary differential equations,”IEEE Control Systems Letters, vol. 6, pp. 2342–2347, 2022

    2022

  6. [14]

    Hamiltonian neural net- works,

    S. Greydanus, M. Dzamba, and J. Yosinski, “Hamiltonian neural net- works,” inAdvances in Neural Information Processing Systems, 2019, pp. 15379–15389

    2019

  7. [15]

    Invertible residual networks for stable modeling,

    J. Behrmann, W. Grathwohl, R. T. Q. Chen, D. Duvenaud, and J.-H. Ja- cobsen, “Invertible residual networks for stable modeling,”International Conference on Machine Learning, 2019, pp. 573–582

    2019

  8. [16]

    Deep Lagrangian networks: Using physics priors for safe learning,

    M. Lutter, C. Ritter, and J. Peters, “Deep Lagrangian networks: Using physics priors for safe learning,” inIEEE International Conference on Robotics and Automation (ICRA), 2019, pp. 8518–8524

    2019

  9. [17]

    Bose,Modern Power Electronics and AC Drives

    B. Bose,Modern Power Electronics and AC Drives. Standards Press, 2015

    2015

  10. [18]

    A Lyapunov approach to incremental stability properties,

    D. Angeli, “A Lyapunov approach to incremental stability properties,” IEEE Transactions on Automatic Control, vol. 47, no. 3, pp. 410–421, 2002

    2002

  11. [19]

    Ergodic theory of chaos and strange attractors,

    D. Ruelle and J.-P. Eckmann, “Ergodic theory of chaos and strange attractors,”Reviews of Modern Physics, vol. 57, no. 3, p. 617, 1985

    1985

  12. [20]

    H. K. Khalil,Nonlinear Systems, 3rd ed. Upper Saddle River, NJ: Prentice Hall, 2002

    2002

  13. [21]

    Sastry,Nonlinear Systems: Analysis, Stability, and Control

    S. Sastry,Nonlinear Systems: Analysis, Stability, and Control. Springer Science & Business Media, 2013

    2013

  14. [22]

    Duffing,Erzwungene Schwingungen bei ver ¨anderlicher Eigenfre- quenz und ihre technische Bedeutung

    G. Duffing,Erzwungene Schwingungen bei ver ¨anderlicher Eigenfre- quenz und ihre technische Bedeutung. Vieweg, Braunschweig, 1918

    1918

  15. [23]

    On relaxation-oscillations,

    B. van der Pol, “On relaxation-oscillations,”The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, vol. 2, no. 11, pp. 978–992, 1926

    1926

  16. [24]

    Deterministic nonperiodic flow,

    E. N. Lorenz, “Deterministic nonperiodic flow,”Journal of the Atmo- spheric Sciences, vol. 20, no. 2, pp. 130–141, 1963

    1963

  17. [25]

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems,

    S. L. Brunton, J. L. Proctor, and J. N. Kutz, “Discovering governing equations from data by sparse identification of nonlinear dynamical systems,”Proceedings of the National Academy of Sciences, vol. 113, no. 15, pp. 3932–3937, 2016

    2016

  18. [26]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems,

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems,”Journal of Computational Physics, vol. 378, pp. 686– 707, 2019

    2019

  19. [27]

    M. S. Pinsker,Information and Information Stability of Random Vari- ables and Processes. Holden-Day, 1964

    1964

  20. [28]

    B. T. Polyak,Introduction to Optimization. Optimization Software, New York, 1987

    1987

  21. [29]

    Smooth stabilization implies coprime factorization,

    E. D. Sontag, “Smooth stabilization implies coprime factorization,”IEEE Transactions on Automatic Control, vol. 34, no. 4, pp. 435–443, 1989

    1989

  22. [30]

    S. Boyd, L. El Ghaoui, E. Feron, and V . Balakrishnan,Linear Matrix Inequalities in System and Control Theory. SIAM, 1994

    1994

  23. [31]

    R. A. Horn and C. R. Johnson,Matrix Analysis. Cambridge University Press, 2012

    2012

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.