Pith. sign in

REVIEW 2 major objections 4 minor 63 references

React to Surprises: Stable-by-Design Neural Feedback Control and the Youla-REN

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proves that a nonlinear Youla parameterization built on contracting recurrent equilibrium networks gives stable-by-design policies for partially observed nonlinear systems, guaranteeing d-tube contraction and Lipschitzness under…

desk verdict Genuinely new stability notion and a solid if patchy theory; the robustness experiments claim more than the theorems certify. read the letter →

arxiv 2506.01226 v3 pith:2N5LXQ33 submitted 2025-06-02 eess.SY cs.LGcs.SY

Nicholas H. Barbara , Ruigang Wang , Alexandre Megretski , Ian R. Manchester This is my paper
classification eess.SYcs.LGcs.SY MSC 93B5293D0593D0968T07
keywords Youlaparameterizationrecurrentequilibriumnetworkscontractionanalysisincrementalstabilityd-tubeneuralfeedbackcontroldeepreinforcementlearningintegralquadraticconstraints
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 tries to establish that a Youla-Kučera-style parameterization of feedback policies, in which a learned Youla parameter reacts to observer innovations, can serve as a stable-by-design policy class for learning-based control of nonlinear, partially observed systems. For a contracting and Lipschitz Youla parameter, the closed loop is shown to be d-tube contracting and Lipschitz: trajectories respond smoothly to known reference inputs, while responses to unknown disturbances converge to a tube around the disturbance-free trajectory whose radius scales with the disturbance size. The paper also gives converse results showing that, under certainty-equivalence observers, every controller that achieves a contracting and Lipschitz closed loop can be written in this Youla form. Combined with a direct parameterization of contracting recurrent equilibrium networks, this yields unconstrained, differentiable policies that keep stability guarantees during deep-reinforcement-learning training. The point of the result is to decouple closed-loop stability from the reward function, training horizon, and model uncertainty, so that gradients can be optimized freely without destabilizing the system.

What carries the argument

The central object is the innovations-feedback form of the nonlinear Youla parameterization: an observer $O$ produces an output estimate $\hat{y}_t$, the difference $\tilde{y}_t = y_t - \hat{y}_t$ (the 'surprises') drives a dynamical system $Q$, and $Q$'s output $\tilde{u}_t$ augments a base controller $K_b$. Theorem 4's proof converts the plant dynamics into observer dynamics plus a disturbance term (Appendix B, equation (47)), applies the observer's contraction and Lipschitz bounds to show $|\tilde{x}_t|$ is bounded independently of $\tilde{u}$, and composes this bound with the contraction of the base closed loop to obtain the d-tube estimate. The recurrent equilibrium network (REN) provides a direct, smooth, surjective parameterization of contracting and Lipschitz maps $\theta \mapsto Q_\theta$, so the whole pipeline $\theta \mapsto Q \mapsto K_Q$ is unconstrained and differentiable and can be searched with first-order optimization methods.

What would settle it

Run the scalar system (21)–(23) from Example 1 with $w \equiv 2$ from two different initial conditions: if the two state trajectories converge to each other, then the claimed loss of contraction does not occur; if they diverge beyond a bound linear in $\|w\|$, then the claimed d-tube bound fails.

Watch

Extended reading notes

Core claim

The central claim is Theorem 4: under assumptions A1–A4 and Lipschitz regularity, for any contracting and Lipschitz Youla parameter $Q$, the closed-loop system (2)–(13) is d-tube contracting and Lipschitz, meaning trajectories converge to a tube around the disturbance-free path with radius proportional to $\|d\|$. The proof shows that the observer error and innovations are constrained within a tube whose radius depends only on the disturbance $d$, independent of the control augmentation $\tilde{u}$, which keeps the feedback loop from being destabilized by high-gain $Q$. A counterexample (Example 1) shows that full contraction and Lipschitzness are lost in general when disturbances are present, because the innovations $\tilde{y}$ couple back into the observer through the control input. When the innovations are generated by a contracting and Lipschitz system decoupled from $\tilde{u}$ (Proposition 1, Theorem 5), the Youla parameterization recovers full contraction and Lipschitzness. Conversely, Theorem 6 and Corollary 3 show that for certainty-equivalence observers, every controller yielding a contracting and Lipschitz closed loop can be realized in Youla form with a contracting and Lipschitz $Q$.

Load-bearing premise

The argument assumes the observer is exactly model-correct (A3): with zero disturbances and matching initial state, the observer reproduces the plant one-step dynamics exactly, so any model mismatch acts as an unbounded state-dependent disturbance and would break the tube bound.

Editorial extensions

If this is right

  • If $\tilde{y}$ and $\tilde{u}$ are decoupled (LTI plants, perfect state feedback, or zero disturbances), the Youla parameterization gives closed loops that are contracting and Lipschitz in $(\eta, d)$, so small changes to references or noise cannot cause large output variations.
  • For partially observed nonlinear plants with disturbances, only d-tube contraction holds in general: two trajectories under the same disturbance converge to a common tube whose radius is proportional to the disturbance norm, and the closed loop remains Lipschitz with respect to known inputs.
  • With certainty-equivalence observers, the parameterization is complete: it contains every controller that stabilizes the observer and yields a contracting and Lipschitz closed loop, so restricting to Youla-REN policies does not sacrifice any contracting and Lipschitz controller.
  • Because the parameterization is direct and unconstrained, stability does not depend on the reward function, the training horizon, or the model parameters encountered during training, so any first-order gradient method can be used.
  • Numerically, Youla-REN policies retain stability and near-optimal cost on LQG control with short training horizons and uncertain pole mass, whereas residual and black-box recurrent policies diverge outside the training horizon.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the d-tube result suggests a design principle for model-based reinforcement learning: train a contracting and Lipschitz observer, then treat every residual prediction error as a disturbance; the tube radius becomes a certificate of the worst-case effect of model error.
  • Because the tube radius scales linearly with the disturbance norm and with the observer's contraction and Lipschitz constants, the framework points to a concrete objective: shrink the observer's contraction rate and Lipschitz bound to tighten the tube, providing a measurable proxy for robustness.
  • The converse under certainty-equivalence observers indicates that, for systems admitting such observers, stability-by-design need not trade away performance: the policy class is complete, so in principle an optimal contracting closed loop is representable.
  • A natural testable extension is to use the frequency-weighted loop shaping of Proposition 2 to shape the disturbance-to-innovation gain in nonlinear systems, tuning the filter $W_2$ to trade robustness against performance in a data-driven way.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes a nonlinear Youla parameterization for feedback control of partially-observed nonlinear systems with disturbances, implemented with recurrent equilibrium networks (RENs) as the Youla parameter. The central theoretical contribution is that, under assumptions A1--A4, any contracting and Lipschitz Youla parameter Q yields a closed loop that is d-tube contracting and Lipschitz (Theorem 4), meaning trajectories converge to a tube around the disturbance-free path whose radius scales with the disturbance norm. The paper also shows that full contraction and Lipschitzness are recovered when the innovations are decoupled from the Youla augmentation (Theorem 5), and provides partial converse results (Theorem 6, Corollaries 2--3). Numerical experiments address economic rewards, short training horizons, and model uncertainty.

Significance. If the theoretical results hold, the paper makes a useful contribution to learning-based control by extending Youla parameterization to nonlinear partial observation with incremental stability guarantees, introducing the weaker but meaningful notion of d-tube contraction, and providing an unconstrained, differentiable policy class with built-in stability certificates. The paper is honest about the need for a good observer and provides reproducible code. The main gap is that the model-uncertainty experiment is presented as a demonstration of 'stable-by-design' behavior, but the rigorous theorem (Theorem 4) does not apply to that setting because the observer-correctness assumption A3 fails off-nominal; the actual stability evidence there is heuristic.

major comments (2)
  1. [Sec. VII-D and V-D, Prop. 2] The claim that Youla-REN policies maintain closed-loop stability under model parameter uncertainty is not supported by the theorems. Theorem 4 assumes A3 (observer correctness), but in the cart-pole experiment the observer is designed at the nominal pole mass m_p^*=0.2 while m_p ranges over [0.14,0.35], so A3 fails at almost every mass in the evaluation set. The text invokes Prop. 2 and a frequency-weighted small-gain argument, but the verification in Fig. 9 is a Bode-magnitude heuristic ('serves as a good heuristic'), not a rigorous incremental IQC/SOS certificate over the uncertainty set. Consequently, the statement in Fig. 10c that Youla-REN 'ensures stability' across the m_p range is an empirical observation, not a certified property. Please either add a rigorous certificate for the uncertain family, e.g. by verifying the incremental IQC condition of Prop. 2 over the parameter range, or explicitly restrict the stability guarantees to the exact-model case and present the uncertainty experiment as heuristic/empirical.
  2. [App. B, Eq. (47)] The proof of Theorem 4 uses A3 as a global pointwise identity f(x,eta,u)=f_o(x,eta,u,h(x)) for all x,u,eta, including trajectories with nonzero disturbance d. However, A3 is stated as a property that holds only when d is identically zero and x_0 equals the observer initial state. As written, the assumption is weaker than what the proof requires. Please restate A3 as a pointwise identity or prove that the zero-disturbance matching-condition statement implies the global identity given that arbitrary initial conditions and control inputs are admissible. This is a formal rigor issue in the central theorem.
minor comments (4)
  1. [Sec. IV-C, Theorem 3] The proof of Theorem 3 is omitted due to space constraints, but Section V states that 'All proofs are provided in Appendices B--F'. Please clarify the status of Theorem 3's proof, either by relocating the statement or by noting explicitly that this proof is omitted.
  2. [Sec. V-D, Prop. 2] The proof of Proposition 2 is deferred to the first author's thesis [63, Chapter 5]. Since this proposition is used as the theoretical basis for the uncertainty experiment, please include a short proof sketch in the paper or state the precise reference so the reader can verify the sufficient condition.
  3. [Sec. V-A, Example 1] The memoryless map Q(y)=min(2.5-5y,0) is described as 'contracting and Lipschitz'. As a static map it is Lipschitz but not strictly contracting in the sense of Definition 1; please clarify the wording, e.g. by treating memoryless maps as trivially contracting with rate 0.
  4. [Sec. II and Fig. 9] Minor typographical issues: in Fig. 9 the axis label '10!2' should likely be '10^{-2}', and in Eq. (20) the values x_{-1} and w_{-1} are used without a formal definition. Please correct these presentation issues.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central stability theorems are parameter-free sufficient conditions proved in the paper, and the self-citations are component-level rather than load-bearing.

full rationale

The main derivation chain is self-contained. Theorem 4 (Appendix B) begins from Assumptions A1-A4 and proves a tube bound on the observer error, Eq. (47)-(53), using Lemma 1; the d-tube contraction and Lipschitz conclusion then follows by composing this bound with the assumed contraction/Lipschitz properties of G_Kb and Q. The conclusion does not appear among the assumptions. Theorem 5 is a series-composition statement, and Theorem 6/Corollary 3 construct the Youla parameter Q from a given controller K as a feedback interconnection of K and the observer, verifying Q is contracting and Lipschitz by composition; no target representation is assumed. The paper does rely on prior work by the same authors for components: Theorem 1 is quoted from [44], but it is a special case and the authors state the proof is subsumed by Section V; the REN guarantee [42] and Lemma 1 [63] are supporting tools rather than the source of the main theorem. These are normal component citations, not circular. The paper explicitly acknowledges its limitations: Theorem 3's proof is omitted, Proposition 2's proof is deferred to [63, Chapter 5], and the cart-pole uncertainty experiment in Section VII-D uses a Bode-magnitude heuristic ('serves as a good heuristic', Fig. 9) rather than a rigorous incremental IQC certificate over the uncertain mass set. Those are verification/correctness gaps, not circular steps: the learned policies are tested empirically, and the stability claim for uncertain plants is not obtained by assuming the conclusion. Score 2 reflects the minor non-load-bearing self-citations and deferred proofs, not any reduction of the central result to its inputs.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The theoretical results rely on assumptions A1-A4, especially the existence of a contracting and correct observer, plus standard contraction and IQC results and one external perturbation lemma. The numeric experimental hyperparameters, Lipschitz bounds and filter gains, are not part of the theoretical claim. No new physical entities are introduced; d-tube contraction is a new mathematical property, not a postulated mechanism.

free parameters (2)
  • Lipschitz bound gamma of the Youla-REN module Q = 1.7 / 120 / 0.15 (Table II)
    Chosen via small-gain and H-infinity heuristics (Proposition 2) to certify closed-loop stability in the cart-pole experiments; not part of the main theoretical statement.
  • Filter parameter nu in W2(s) = 5e-4 (Youla-gammaREN), 1e-2 (Residual-gammaREN)
    Chosen to shape the loop gain and tuned empirically. Not part of the main theorems.
assumptions (6)
  • domain assumption Assumption A1: the base closed-loop system G_Kb is contracting and Lipschitz
    Required for Theorems 4 and 5; if the base controller does not make the loop contracting, the d-tube result does not follow.
  • domain assumption Assumption A2: the observer is contracting and the map (eta,u,y) maps to xhat is Lipschitz
    Needed to bound the observer error independently of the augmentation input, as done in Appendix B.
  • domain assumption Assumption A3: observer correctness under zero disturbances
    Used to express plant dynamics through observer dynamics in Appendix B, eq. (47); model error breaks the tube bound.
  • domain assumption Assumption A4: piecewise-differentiable dynamics and Lipschitz output maps
    Regularity condition used throughout the proofs, though it does not explicitly state that f_o is globally Lipschitz in y.
  • standard math Lemma 1: perturbation bound for contracting systems from [63]
    Used without proof in Appendix B to convert disturbance bounds into state-error bounds; it is a standard contraction theory result.
  • standard math Series composition properties of contracting and Lipschitz systems
    Used in the proofs of Theorems 5, 6, and Corollary 3; standard contraction analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of React to Surprises: Stable-by-Design Neural Feedback Control and the Youla-REN." pith.science (2026). https://pith.science/paper/2N5LXQ33

@misc{pith2026250601226,
  author       = {Pith},
  title        = {Pith review of: React to Surprises: Stable-by-Design Neural Feedback Control and the Youla-REN},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2N5LXQ33}},
  note         = {Machine review of arXiv:2506.01226}
}
read the original abstract

We study parameterizations of stabilizing nonlinear policies for learning-based control. We propose a structure based on a nonlinear version of the Youla-Kucera parameterization combined with robust neural networks such as the recurrent equilibrium network (REN). The resulting parameterizations are unconstrained, and hence can be searched over with first-order optimization methods, while always ensuring closed-loop stability by construction. We study the combination of (a) nonlinear dynamics, (b) partial observation, and (c) incremental closed-loop stability requirements (contraction and Lipschitzness). We find that for the combination of (c) with either (a) or (b), a contracting and Lipschitz Youla parameter always leads to contracting and Lipschitz closed loops. However, if all three hold, then incremental stability can be lost with exogenous disturbances. Instead, a weaker condition is maintained, which we call d-tube contraction and Lipschitzness. We further obtain converse results showing that the proposed parameterization covers all contracting and Lipschitz closed loops for certain classes of nonlinear systems. Numerical experiments illustrate the utility of our parameterization when learning controllers with built-in stability certificates for: (i) ``economic'' rewards without stabilizing effects; (ii) short training horizons; and (iii) uncertain systems.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 57 canonical work pages

  1. [1]

    Mastering the game of Go without human knowledge,

    D. Silveret al., “Mastering the game of Go without human knowledge,” Nature, vol. 550, pp. 354–359, 2017

    work page 2017

  2. [2]

    Deep reinforcement learning for robotics: A survey of real- world successes,

    C. Tang, B. Abbatematteo, J. Hu, R. Chandra, R. Mart ´ın-Mart´ın, and P. Stone, “Deep reinforcement learning for robotics: A survey of real- world successes,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 8, no. 1, pp. 153–188, 2025

    work page 2025

  3. [3]

    Magnetic control of tokamak plasmas through deep reinforcement learning,

    J. Degraveet al., “Magnetic control of tokamak plasmas through deep reinforcement learning,”Nature, vol. 602, pp. 414–419, 2022

    work page 2022

  4. [4]

    R. S. Sutton and A. G. Barto,Reinforcement Learning: An Introduction. The MIT Press, 2018

    work page 2018

  5. [5]

    Simple statistical gradient-following algorithms for connectionist reinforcement learning,

    R. J. Williams, “Simple statistical gradient-following algorithms for connectionist reinforcement learning,”Machine learning, vol. 8, no. 3, pp. 229–256, 1992

    1992

  6. [6]

    Meyn,Control systems and reinforcement learning

    S. Meyn,Control systems and reinforcement learning. Cambridge University Press, 2022

    work page 2022

  7. [7]

    Prox- imal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    arXiv 2017

  8. [8]

    Accelerated policy learning with parallel differentiable simulation,

    J. Xuet al., “Accelerated policy learning with parallel differentiable simulation,” inInternational Conference on Learning Representations (ICLR), 2022

    work page 2022

Show all 63 references
  1. [9]

    Do differentiable simulators give better policy gradients?

    H. J. Suh, M. Simchowitz, K. Zhang, and R. Tedrake, “Do differentiable simulators give better policy gradients?” inInternational Conference on Machine Learning (ICML), 2022

    2022

  2. [10]

    Brax – a differentiable physics engine for large scale rigid body simulation,

    C. D. Freeman, E. Frey, A. Raichuk, S. Girgin, I. Mordatch, and O. Bachem, “Brax – a differentiable physics engine for large scale rigid body simulation,” inAdvances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2021

    2021

  3. [11]

    Dojo: A differentiable physics engine for robotics,

    T. A. Howellet al., “Dojo: A differentiable physics engine for robotics,” arXiv preprint arXiv:2203.00806, 2022

    2022 arXiv

  4. [12]

    Neural networks in the loop: Learning with stability and robustness guarantees,

    I. R. Manchester, R. Wang, and N. H. Barbara, “Neural networks in the loop: Learning with stability and robustness guarantees,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 9, 2026

    2026

  5. [13]

    Residual policy learning,

    T. Silver, K. Allen, J. Tenenbaum, and L. Kaelbling, “Residual policy learning,”arXiv preprint arXiv:1812.06298, 2018

    2018 arXiv

  6. [14]

    Residual reinforcement learning for robot control,

    T. Johanninket al., “Residual reinforcement learning for robot control,” pp. 6023–6029, 2019

    2019

  7. [15]

    Residual policy learning for perceptive quadruped control using differentiable simulation,

    J. Y . Luo, Y . Song, V . Klemm, F. Shi, D. Scaramuzza, and M. Hut- ter, “Residual policy learning for perceptive quadruped control using differentiable simulation,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 1–8

    2025

  8. [16]

    Modern Wiener-Hopf design of optimal controllers — Part II: The multivariable case,

    D. C. Youla, J. J. Bongiorno, and H. A. Jabr, “Modern Wiener-Hopf design of optimal controllers — Part II: The multivariable case,”IEEE Transactions on Automatic Control, vol. 21, pp. 319–338, 1976

    1976

  9. [17]

    From Youla-Kucera to identification, adaptive and nonlinear control,

    B. D. Anderson, “From Youla-Kucera to identification, adaptive and nonlinear control,”Automatica, vol. 34, pp. 1485–1506, 12 1998

    1998

  10. [18]

    Advances in Youla- Kuˇcera parametrization: A review,

    I. Mahtout, F. Navas, V . Milanes, and F. Nashashibi, “Advances in Youla- Kuˇcera parametrization: A review,”Annual Reviews in Control, vol. 49, pp. 81–94, 2020

    2020

  11. [19]

    Stability of discrete linear feedback systems,

    V . Ku ˇcera, “Stability of discrete linear feedback systems,”IFAC Pro- ceedings Volumes, vol. 8, pp. 573–578, 8 1975

    1975

  12. [20]

    K. Zhou, J. C. Doyle, and K. Glover,Robust and optimal control. USA: Prentice-Hall, Inc., 1996

    1996

  13. [21]

    Feedback controller parameterizations for reinforcement learning,

    J. W. Roberts, I. R. Manchester, and R. Tedrake, “Feedback controller parameterizations for reinforcement learning,” inIEEE Symposium on Adaptive Dynamic Programming and Reinforcement Learning (ADPRL), 2011, pp. 310–317

    2011

  14. [22]

    Matrix interpolation theory and optimal control,

    J. C. Doyle, “Matrix interpolation theory and optimal control,” 1984

    1984

  15. [23]

    Feedback and optimal sensitivity: Model reference trans- formations, multiplicative seminorms, and approximate inverses,

    G. Zames, “Feedback and optimal sensitivity: Model reference trans- formations, multiplicative seminorms, and approximate inverses,”IEEE Transactions on automatic control, vol. 26, no. 2, pp. 301–320, 1981

    1981

  16. [24]

    Internal model control. a unifying review and some new results,

    C. E. Garcia and M. Morari, “Internal model control. a unifying review and some new results,”Industrial & Engineering Chemistry Process Design and Development, vol. 21, no. 2, pp. 308–323, 1982

    1982

  17. [25]

    Optimization over state feedback policies for robust control with constraints,

    P. J. Goulart, E. C. Kerrigan, and J. M. MacIejowski, “Optimization over state feedback policies for robust control with constraints,”Automatica, vol. 42, pp. 523–533, 4 2006

    2006

  18. [26]

    Improper learning for non- stochastic control,

    M. Simchowitz, K. Singh, and E. Hazan, “Improper learning for non- stochastic control,” inCOLT, 2020, pp. 3320–3436

    2020

  19. [27]

    Global parametrization of feedback systems with nonlinear plants,

    C. A. Desoer and R.-W. Liu, “Global parametrization of feedback systems with nonlinear plants,”System Control Letter, vol. 1, pp. 249– 251, 1982

    1982

  20. [28]

    Two-step compensation of nonlinear systems,

    C. A. Desoer and C.-A. Lin, “Two-step compensation of nonlinear systems,”Systems & Control Letters, vol. 3, no. 1, pp. 41–45, 1983

    1983

  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]

    Left coprime factorizations and a class of stabilizing controllers for nonlinear systems,

    T. T. Tay and J. B. Moore, “Left coprime factorizations and a class of stabilizing controllers for nonlinear systems,” in27th IEEE Conference on Decision and Control (CDC), 1988, pp. 449–454. 16 IEEE TRANSACTIONS AND JOURNALS TEMPLATE

    1988

  23. [31]

    On the Youla-Kucera parametrization for nonlinear systems,

    A. D. Paice and J. B. Moore, “On the Youla-Kucera parametrization for nonlinear systems,”Syst. Control Lett., vol. 14, pp. 121–129, 2 1990

    1990

  24. [32]

    The class of stabilizing nonlinear plant controller pairs,

    A. D. Paice and A. J. V . D. Schaft, “The class of stabilizing nonlinear plant controller pairs,”IEEE Transactions on Automatic Control, vol. 41, pp. 634–645, 1996

    1996

  25. [33]

    Characterization of all nonlinear stabiliz- ing controllers via observer-based kernel representations,

    K. Fujimoto and T. Sugie, “Characterization of all nonlinear stabiliz- ing controllers via observer-based kernel representations,”Automatica, vol. 36, pp. 1123–1135, 8 2000

    2000

  26. [34]

    A state-space approach to parameterization of stabilizing controllers for nonlinear systems,

    W. M. Lu, “A state-space approach to parameterization of stabilizing controllers for nonlinear systems,”IEEE Transactions on Automatic Control, vol. 40, pp. 1576–1588, 1995

    1995

  27. [35]

    Parametrization of all stabilizing con- trollers of nonlinear systems,

    J.-i. Imura and T. Yoshikawa, “Parametrization of all stabilizing con- trollers of nonlinear systems,”Systems & Control Letters, vol. 29, no. 4, pp. 207–213, 1997

    1997

  28. [36]

    On contraction analysis for non-linear systems,

    W. Lohmiller and J. J. E. Slotine, “On contraction analysis for non-linear systems,”Automatica, vol. 34, pp. 683–696, 6 1998

    1998

  29. [37]

    Direct parameterization of Lipschitz- bounded deep networks,

    R. Wang and I. Manchester, “Direct parameterization of Lipschitz- bounded deep networks,” inInternational Conference on Machine Learning (ICML), 2023

    2023

  30. [38]

    Convex optimization in identification of stable non-linear state space models,

    M. M. Tobenkin, I. R. Manchester, J. Wang, A. Megretski, and R. Tedrake, “Convex optimization in identification of stable non-linear state space models,” in49th IEEE Conference on Decision and Control (CDC), 2010, pp. 7232–7237

    2010

  31. [39]

    Convex parame- terizations and fidelity bounds for nonlinear identification and reduced- order modelling,

    M. M. Tobenkin, I. R. Manchester, and A. Megretski, “Convex parame- terizations and fidelity bounds for nonlinear identification and reduced- order modelling,”IEEE Transactions on Automatic Control, vol. 62, no. 7, pp. 3679–3686, 2017

    2017

  32. [40]

    A convex parameterization of robust recurrent neural networks,

    M. Revay, R. Wang, and I. R. Manchester, “A convex parameterization of robust recurrent neural networks,”IEEE Control Systems Letters, vol. 5, no. 4, pp. 1363–1368, 2020

    2020

  33. [41]

    Specialized interior-point algo- rithm for stable nonlinear system identification,

    J. Umenberger and I. R. Manchester, “Specialized interior-point algo- rithm for stable nonlinear system identification,”IEEE Transactions on Automatic Control, vol. 64, no. 6, pp. 2442–2456, 2018

    2018

  34. [42]

    Recurrent equilibrium networks: Flexible dynamic models with guaranteed stability and ro- bustness,

    M. Revay, R. Wang, and I. R. Manchester, “Recurrent equilibrium networks: Flexible dynamic models with guaranteed stability and ro- bustness,”IEEE Transactions on Automatic Control, vol. 69, no. 5, pp. 2855–2870, 2024

    2024

  35. [43]

    Youla-REN: Learning nonlinear feed- back policies with robust stability guarantees,

    R. Wang and I. R. Manchester, “Youla-REN: Learning nonlinear feed- back policies with robust stability guarantees,” in2022 American Control Conference (ACC). IEEE, 2022, pp. 2116–2123

    2022

  36. [44]

    Learning over all stabilizing nonlinear controllers for a partially-observed linear system,

    R. Wang, N. H. Barbara, M. Revay, and I. R. Manchester, “Learning over all stabilizing nonlinear controllers for a partially-observed linear system,”IEEE Control System Letter, vol. 7, pp. 91–96, 2022

    2022

  37. [45]

    Learning over contract- ing and Lipschitz closed-loops for partially-observed nonlinear systems,

    N. H. Barbara, R. Wang, and I. R. Manchester, “Learning over contract- ing and Lipschitz closed-loops for partially-observed nonlinear systems,” 62nd IEEE Conference on Decision and Control, pp. 1028–1033, 2023

    2023

  38. [46]

    Youla–ku ˇcera parameterization in contraction framework,

    Y . Kawano, A. J. van der Schaft, and J. M. Scherpen, “Youla–ku ˇcera parameterization in contraction framework,”IEEE Transactions on Au- tomatic Control, vol. 70, no. 3, pp. 1667–1682, 2025

    2025

  39. [47]

    Neural system level synthesis: Learning over all stabilizing policies for nonlinear systems,

    L. Furieri, C. L. Galimberti, and G. Ferrari-Trecate, “Neural system level synthesis: Learning over all stabilizing policies for nonlinear systems,” inIEEE Conference on Decision and Control, 2022, pp. 2765–2770

    2022

  40. [48]

    Learning to boost the performance of stable nonlinear systems,

    ——, “Learning to boost the performance of stable nonlinear systems,” IEEE Open Journal of Control Systems, vol. 3, pp. 342–357, 2024

    2024

  41. [49]

    Parametrizations of all stable closed-loop responses: From theory to neural network control design,

    C. L. Galimberti, L. Furieri, and G. Ferrari-Trecate, “Parametrizations of all stable closed-loop responses: From theory to neural network control design,”Annual Reviews in Control, vol. 60, p. 101012, 2025

    2025

  42. [50]

    System analysis via integral quadratic constraints,

    A. Megretski and A. Rantzer, “System analysis via integral quadratic constraints,”IEEE transactions on automatic control, vol. 42, no. 6, pp. 819–830, 1997

    1997

  43. [51]

    A new controller architecture for high per- formance, robust, and fault-tolerant control,

    K. Zhou and Z. Ren, “A new controller architecture for high per- formance, robust, and fault-tolerant control,”IEEE Transactions on Automatic Control, vol. 46, pp. 1613–1618, 2001

    2001

  44. [52]

    A system level approach to discrete-time nonlinear systems,

    D. Ho, “A system level approach to discrete-time nonlinear systems,” in IEEE American Control Conference (ACC), 2020, pp. 1625–1630

    2020

  45. [53]

    What matters for on-policy deep actor-critic methods? a large-scale study,

    M. Andrychowiczet al., “What matters for on-policy deep actor-critic methods? a large-scale study,” inInternational conference on learning representations (ICLR), 2021

    2021

  46. [54]

    Robustneuralnetworks.jl: a package for machine learning and data- driven control with certified robustness,

    N. H. Barbara, M. Revay, R. Wang, J. Cheng, and I. R. Manchester, “Robustneuralnetworks.jl: a package for machine learning and data- driven control with certified robustness,”Proc. JuliaCon Conf., vol. 7, no. 68, p. 163, 2025

    2025

  47. [55]

    Training efficient controllers via analytic policy gradient,

    N. Wiedemann, V . W ¨uest, A. Loquercio, M. M ¨uller, D. Floreano, and D. Scaramuzza, “Training efficient controllers via analytic policy gradient,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 1349–1356

    2023

  48. [56]

    R2DN: Scalable parameterization of contracting and Lipschitz recurrent deep networks,

    N. H. Barbara, R. Wang, and I. R. Manchester, “R2DN: Scalable parameterization of contracting and Lipschitz recurrent deep networks,” arXiv preprint arXiv:2504.01250, 2025

    2025 arXiv

  49. [57]

    Uniting two control Lyapunov functions for affine systems,

    V . Andrieu and C. Prieur, “Uniting two control Lyapunov functions for affine systems,”IEEE Transactions on Automatic Control, vol. 55, no. 8, pp. 1923–1927, 2010

    1923

  50. [58]

    Virtual control contraction metrics: Convex nonlinear feedback design via be- havioral embedding,

    R. Wang, R. T ´oth, P. J. W. Koelewijn, and I. R. Manchester, “Virtual control contraction metrics: Convex nonlinear feedback design via be- havioral embedding,”International Journal of Robust Nonlinear Control, vol. 34, pp. 7698–7721, 2024

    2024

  51. [59]

    Learning to walk in minutes using massively parallel deep reinforcement learning,

    N. Rudin, D. Hoeller, P. Reist, and M. Hutter, “Learning to walk in minutes using massively parallel deep reinforcement learning,” in Conference on Robot Learning (CoRL). PMLR, 2022, pp. 91–100

    2022

  52. [60]

    Guaranteed margins for LQG regulators,

    J. Doyle, “Guaranteed margins for LQG regulators,”IEEE Transactions on automatic Control, vol. 23, no. 4, pp. 756–757, 1978

    1978

  53. [61]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural Computation, vol. 9, pp. 1735–1780, 11 1997

    1997

  54. [62]

    Reinforcement learning with long short-term memory,

    B. Bakker, “Reinforcement learning with long short-term memory,” in Advances in Neural Information Processing Systems (NeurIPS), 2001

    2001

  55. [63]

    Parametrising neural feedback policies with stability and robustness guarantees,

    N. H. Barbara, “Parametrising neural feedback policies with stability and robustness guarantees,” Ph.D. dissertation, University of Sydney, 2025. Nicholas H. Barbarareceived the Ph.D. de- gree and the combined B.Sc. & B.E. (Hons 1, University Medal) degrees from the University...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.