{"id":"4de3652b-d63c-43d3-b629-8a713f9bdd02","arxiv_id":"2508.21684","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A data-driven robust controller, formed by adding a Lyapunov redesign term to a dual-EnKF optimal controller, stabilizes heat and Burgers PDEs under bounded disturbances in simulation.","lead":"This paper designs a robust controller for nonlinear systems and PDEs when the model is unknown but a simulator is available, adding a Lyapunov-based disturbance-rejection term on top of an optimal controller learned with the dual ensemble Kalman filter. Tests on the heat and Burgers equations show the robust term shrinks disturbance effects beyond the learned optimal controller alone.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stability guarantee requires full-rank B, but PDE experiments relax it; the Lyapunov inequality used in Props. 1/2 is invalid in the tested m << n regime.","rationale":"The reader's verdict is CONDITIONAL, and I find the same gap to be the critical one. The sign error in the proof of Prop. 1 is a typo: restoring the sign gives the standard LQ Lyapunov argument, so it is not load-bearing. The more serious issue is that the theorem's sufficient condition is structurally incompatible with the application: for PDE control with few actuators, rank(B)=n cannot hold, and the paper admits this. Once BBdagger is only a projection, the exact cancellation in the proof is replaced by a projection term, and no bound is supplied. The simulations with input-channel disturbances may still behave well, but the paper does not prove it, and no code or comparison baseline is provided to independently check the figures. I would keep the verdict CONDITIONAL rather than REJECT because the empirical evidence is plausible and the theoretical gap is explicitly acknowledged; the paper can be revised by proving (or numerically demonstrating) a practical-stability bound for matched disturbances in the low-rank case, and by adding a sensitivity study.","tokens_in":11671,"tokens_out":13609,"duration_ms":172615,"concrete_test":"Re-run the heat-equation test of Fig. 1 with the same learned Pbar(N), but add a state disturbance d(t)=eps v, where v is a unit vector in Null(B^T) (e.g., a high-frequency spatial mode not representable by the m=8 indicator controls). Compare ||z(T)||/||z(0)|| for eps=0 and eps>0 at the same lambda. If performance degrades or diverges with eps, the rank-deficient implementation violates the assumptions of Prop. 1; if it does not, the paper should replace Assumption 2(i) with a weaker condition that explains the observed robustness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 1's proof (Sec. II.B) uses Bud = -lambda Pbar x / |Pbar x|, which needs BBdagger = I, i.e., Assumption 2(i) rank(B)=n; Proposition 2 similarly needs rank(b(x))=n (Assumption 4(ii)). The heat and Burgers implementations have n=100, m=8/10 and n=128, m=10 (Sec. IV, App. II), so rank(B)=m < n. Remarks 1 and 4 say the condition is relaxed, but no replacement bound is proved. In this regime BBdagger is the orthogonal projection onto Range(B); the control contribution to Vdot is only -lambda |P_proj Pbar x|^2 / |Pbar x|, whereas the disturbance term can be as large as lambda|Pbar x| (or lambda|B^T Pbar x| for the matched input disturbances used in Sec. IV). Vdot can therefore be positive when Pbar x has a component in Null(B^T), so neither asymptotic nor practical stability follows from the given argument. Remark 2's 'practical stability' claim is asserted without proof. Thus the central stability guarantee does not cover the PDE experiments, and the reported robustness rests on the simulations alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven robust control design for nonlinear affine-in-control systems, combining an optimal feedback law computed via the dual ensemble Kalman filter (dual EnKF) with an additive Lyapunov-redesign term. Section II treats LTI systems, Section III extends the construction to nonlinear systems, and Section IV applies the method to discretized heat and Burgers PDEs, testing constant and sinusoidal disturbances. The main theoretical claims are Propositions 1 and 2, which assert asymptotic stability under the respective rank conditions rank(B)=n and rank(b(x))=n. The numerical experiments deliberately operate outside these conditions, as acknowledged in Remarks 1 and 4.","tokens_in":11962,"tokens_out":9768,"duration_ms":110876,"significance":"If the gap between the assumptions and the PDE experiments can be closed, the paper is a useful demonstration that a learning-based optimal controller can be robustified by a Lyapunov-redesign term computed with only simulator access. The dual EnKF component is prior work with its own convergence analysis, and the robustification theory is standard; the manuscript's contribution is the combination and its PDE application. The simulation study is reasonably broad: two PDEs, two viscosities, two disturbance types, and 100 random initial conditions, with robust control visibly outperforming optimal control alone. However, the main theoretical guarantee does not cover the demonstrated rank-deficient regime, and the claims of practical stability in Remarks 2 are asserted without proof. These are load-bearing issues for the paper's central claim, though they appear fixable by adding a matched-disturbance analysis or by explicitly repositioning the PDE results as empirical demonstrations.","major_comments":[{"comment":"The displayed Lyapunov derivative has a sign error. For the closed-loop matrix A_cl = A - B\\bar K, the true derivative is \\dot V = x^T \\bar P A_cl x + (Bu_d+d)^T\\bar P x = -\\tfrac12 x^T(Q + \\bar P B R^{-1}B^T\\bar P)x + (Bu_d+d)^T\\bar P x. The paper instead writes \\dot V = -x^T\\bar P(A-B\\bar K)x + ... and then uses the claim that this term is nonpositive. The conclusion of Proposition 1 is restorable with the standard ARE computation, but the proof as written is not valid and should be corrected.","section":"Section II.B, Proposition 1 proof"},{"comment":"Assumptions 2(i) and 4(ii) require rank(B)=n and rank(b(x))=n, but the PDE experiments violate these assumptions: heat equation has n=100, m=8/10, and Burgers full model has p=128, m=10. Remarks 1 and 4 say the conditions are relaxed, but no replacement stability bound is provided. When rank(B)<n, BB^\\dagger is the orthogonal projection onto Range(B), so u_d only cancels the component of \\bar P x in Range(B). For disturbances with a component in Null(B^T), the term d^T\\bar P x can make \\dot V positive. This is not merely a cosmetic gap: the asymptotic stability guarantee of Propositions 1 and 2 does not cover the systems on which the paper demonstrates the method.","section":"Section II.B/III.B vs. Section IV, App. II"},{"comment":"The claim that the regularized implementation gives 'practical stability' is asserted without proof. In the rank-deficient regime relevant to the PDE experiments, the control term is -\\lambda (BB^\\dagger \\bar P x)/|\\bar P x|; if \\bar P x is close to Null(B^T), this term is small while a disturbance in Null(B^T) contributes up to \\lambda |\\bar P x|. The derivative can then be positive for small |x|, so neither asymptotic nor practical stability follows from the given argument. A concrete matched/unmatched disturbance analysis or a modified controller is needed before the PDE claims can be supported theoretically.","section":"Remark 2, Section II.B"},{"comment":"In the unknown-B branch, the algorithm returns (\\bar u^{(N)})_i = (R^{-1}B^T\\bar P^{(N)}x)_i, which is the negative of the optimal control. Indeed, for H as defined in (4), minimizing in u gives \\bar u = -R^{-1}B^T\\bar P x. The B-known branch returns this negative value, so the two branches disagree. If a user implements the unknown-B branch, the learning control will have the wrong sign. This is a load-bearing implementation error for the model-free claims, although the reported PDE experiments may use the known-B branch.","section":"Appendix I-C, Algorithm 2"}],"minor_comments":[{"comment":"The heat equation section states m=10, while Appendix II-A states m=8. Please make these consistent.","section":"Section IV.A vs. Appendix II-A"},{"comment":"The disturbance notation is ambiguous: U(t)=u(t)+d(t) suggests d(t) is m-dimensional, but later d(t)=d0 sin(t) and d(t)=d0 are written as scalars. Clarify whether d is a scalar or a vector in R^m.","section":"Section IV, Eq. (8) discussion"},{"comment":"Typo: 'stailizing' should be 'stabilizing' in the sentence introducing the stabilizing control.","section":"Section II"},{"comment":"In the nonlinear case, the line using S(0,v) to minimize |b(x)v - \\nabla V/|\\nabla V|| is not generally correct, because S(0,v) = a(0)+b(0)v, not b(x)v. Please clarify how the simulator is used to evaluate b(x)v, e.g., via S(x,v)-S(x,0), or state the needed assumptions.","section":"Algorithm 1 and Section III.C"},{"comment":"The proof says 'For the first equality we recall Assumption 4 and Cauchy-Schwarz'; this should be 'first inequality'. The equality/inequality labels are currently inconsistent.","section":"Section III.B, proof of Proposition 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript builds legitimately on prior dual-EnKF work and Khalil's Lyapunov redesign, and I do not see a novelty or citation problem. The main issue is that the theorems do not cover the demonstrated PDE regime; the paper either needs a matched-disturbance practical-stability theorem for the rank-deficient case or a clear reframing of the PDE results as empirical. Given that the gap is identifiable and fixable, major revision rather than rejection seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a clear, well-organized step in the data-driven PDE control direction: it wraps the dual EnKF learned LQR with a Lyapunov redesign term and shows, on 100 random initial conditions, that the robustified controller beats the pure optimal controller on heat and Burgers, often by an order of magnitude in the final L2 norm. The writing is straightforward, the simulation study is careful, and the authors are transparent about where their assumptions are violated.\n\nThe genuinely new piece is the integration, not the theory. Propositions 1 and 2 are standard Khalil redesign with a pseudoinverse, and the proofs follow the textbook template. There is a sign error in Proposition 1's displayed Lyapunov derivative (the no-disturbance term should be negative, not preceded by a minus sign); it looks fixable without changing the conclusion. The larger soft spot is the rank condition. The theorems need rank(B)=n, but the heat and full nonlinear Burgers runs have m<<n, and the paper explicitly relaxes the assumption in Remarks 1 and 4. That means the asymptotic stability guarantee simply does not apply to the main experiments. Remark 2 asserts practical stability without proof, and in the underactuated case the redesign term only cancels disturbances in the range of B; Vdot can be positive when Pbar x has a null-space component. The stress-test note is right about this. The mismatch is not hidden, but the paper should either extend the analysis to the rank-deficient case or explicitly reframe the PDE results as empirical demonstrations, not consequences of the theorems.\n\nThere are smaller issues: the heat equation section says m=10 while the appendix says m=8; the nonlinear dual-EnKF uses a gradient approximation from [15] without restating its validity conditions; and there are no baseline comparisons against a model-based robust controller (e.g., an LQR with an H-infinity or disturbance-attenuation term). Still, the paper's central empirical claim is believable, and the honest discussion of assumptions counts in its favor.\n\nI would send this to peer review. A referee can reasonably ask for major revisions: fix the sign, either prove something for rank-deficient B or say plainly that the theory only covers the full-rank case, reconcile the m discrepancy, and add a baseline. The target audience is practitioners working on simulator-based control of PDEs who want a turnkey robustification for learned LQR policies; for them this is a useful data point even if not a definitive one.","headline":"Honest about its assumptions but the theory doesn't cover the PDE experiments; the simulations carry the paper.","tokens_in":12464,"tokens_out":3747,"would_cite":false,"duration_ms":47966,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C20","93C10","93E11"],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding a Lyapunov redesign term to a dual-EnKF learned control stabilizes nonlinear systems and PDE discretizations under bounded disturbances.","keywords":["dual ensemble Kalman filter","Lyapunov redesign","robust control","nonlinear affine systems","partial differential equations","heat equation","Burgers equation","data-driven control"],"falsifier":"For a minimal full-rank case (e.g., n=2, B=I, A with a nonzero unstable mode), compute the closed-loop V̇ exactly for a bounded disturbance with the controller of Proposition 1; if V̇ is not strictly negative for some state, the theorem is false. For the implemented method, run the heat-equation simulation with N=10000 EnKF particles and measure the steady-state L2 norm against the practical-stability radius r=0.002; norms far exceeding that radius would indicate that EnKF approximation error, not disturbance, limits performance.","tokens_in":11522,"feed_emoji":"🎛️","tokens_out":9938,"duration_ms":99135,"temperature":0.7,"pith_summary":"This paper seeks to show that a fully data-driven controller — no model identification, just a simulator — can robustly stabilize nonlinear control-affine systems, including discretized PDEs. The robust control is the sum of an optimal stabilizing control learned by the dual ensemble Kalman filter and a Lyapunov redesign term that cancels bounded disturbances. The authors prove asymptotic stability for both linear and nonlinear systems under a full-rank condition on the input map, then demonstrate on heat and Burgers equations that the robustified control drives the L2 norm of the state orders of magnitude closer to zero than the optimal control alone. The reader would care because the approach targets high-dimensional PDE control, where standard model-based design is impractical and data-driven methods must work with simulators only.","feed_headline":"Robust control drives heat and Burgers PDEs to zero under disturbances","feed_subtitle":"Adding a Lyapunov redesign term to a learned EnKF control cuts settling error by an order of magnitude.","key_machinery":"Two objects carry the argument. The first is the dual ensemble Kalman filter (dual EnKF), a controlled interacting-particle system running backward in time whose terminal empirical covariance (linear case) or empirical gradient (nonlinear case) approximates the value function of the optimal control problem without ever constructing the system matrices. The second is the Lyapunov redesign term u_d = −λ b†∇V / |∇V|, which uses the same learned value function to produce a control component that points along −∇V; under the rank condition on b (or B), b u_d equals −λ ∇V/|∇V|, exactly canceling the worst-case disturbance in the V̇ inequality.","core_discovery":"The paper's central claim is Proposition 2 (and its linear counterpart, Proposition 1): for the nonlinear affine system ẋ = a(x)+b(x)u+d, the feedback u = ū + u_d, with ū from the HJB-optimal problem and u_d = −λ b†∇V/|∇V|, renders the origin asymptotically stable for every disturbance bounded by |d| < λ. The proof shows V̇ ≤ (|d|−λ)|∇V| < 0, because the redesign term exactly cancels the disturbance once b has full rank. The dual EnKF supplies ū and V from simulator trajectories alone: backward-run interacting particles approximate the ARE solution or value-function gradient, and the control is recovered by minimizing the Hamiltonian H(x,u) = (∇V)ᵀS(x,u)+½L(x,u) using only simulator calls. E","pith_inferences":["Because the PDE implementations relax the full-rank assumption (n=100 with m=8 controls, n=128 with m=10), the demonstrated stability is practical, not asymptotic; a practitioner should tune λ and r explicitly rather than relying on Propositions 1–2 in underactuated regimes.","The Hamiltonian-minimization step is agnostic to the learning algorithm — any zero-order optimizer or policy-gradient method could supply the optimal-control component, so the Lyapunov redesign construction is a general plugin for robustifying learned controllers.","The framework extends naturally to stochastic disturbances (e.g., Gaussian white noise) and to output-feedback settings; the paper's filtering-based EnKF machinery gives a direct path, since value-function estimation could be carried out from sensor measurements.","For other dissipative PDEs such as Allen-Cahn or KdV, the same Lyapunov redesign mechanism should transfer, because the robustness term needs only a Lyapunov function and a known disturbance bound, not PDE-specific structure."],"forward_implications":["Under the full-rank input-map condition, the proposed controller guarantees asymptotic stability for any disturbance bounded by a known λ — not merely practical stability.","The controller can be implemented with only a disturbance-free simulator: both the optimal term and the robustifying term are computed through Hamiltonian evaluations S(x,u), with no need to identify a(x) and b(x).","Simulations on heat and Burgers equations show the robustified control reduces the terminal L2 norm by an order of magnitude compared with the optimal control alone under the same disturbances.","For the Burgers equation, the robustification term also masks model mismatch: the controller built on a reduced-order linear model attains settling performance close to that of the full nonlinear model.","Adding a small regularizing parameter r to u_d converts the asymptotic guarantee into practical stability in a ball; the experiments confirm this behavior at r = 0.002."],"supporting_citations":[{"why":"supplies the dual EnKF interacting-particle algorithm used to approximate the value function and the optimal control from simulator trajectories.","marker":"[15]"},{"why":"provides the Lyapunov redesign method (Chapter 14) that constructs the robustifying term ud against bounded disturbances.","marker":"[18]"},{"why":"gives the Riccati-equation theory and the stabilizing-gain theorem on which the linear robust control result rests.","marker":"[19]"},{"why":"provides the Hamilton-Jacobi-Bellman characterization of the optimal control used in the nonlinear case.","marker":"[22]"},{"why":"introduces dynamic mode decomposition with control (DMDc), used to build the linear reduced-order simulator for the Burgers equation.","marker":"[24]"},{"why":"supplies the PDE simulation environments (heat and Burgers equations) used in the numerical experiments.","marker":"[31]"},{"why":"the preceding DMDc-based PDE control approach that this paper extends, including the test setup and initial-condition sampling.","marker":"[32]"},{"why":"establishes the log-transform duality between optimal control and nonlinear filtering that motivates the dual EnKF sampling formulation.","marker":"[10]"}],"fun_headline_variants":["Dual EnKF plus Lyapunov redesign tames PDEs under disturbance","Robust control for nonlinear PDEs via dual ensemble Kalman filter","Lyapunov redesign teams with EnKF to stabilize heat and Burgers","Data-driven robust control drives PDEs to zero despite disturbances","Dual EnKF robustifies control for heat and Burgers equations"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The asymptotic-stability proofs require the input matrix b(x) to have rank equal to the state dimension, so the Lyapunov redesign term can exactly cancel the disturbance; the PDE experiments relax this condition, so the formal guarantee does not apply there and the reported behavior rests on the simulations.","fun_headline_variants_meta":{"raw":{"variants":["Dual EnKF plus Lyapunov redesign tames PDEs under disturbance","Robust control for nonlinear PDEs via dual ensemble Kalman filter","Lyapunov redesign teams with EnKF to stabilize heat and Burgers","Data-driven robust control drives PDEs to zero despite disturbances","Dual EnKF robustifies control for heat and Burgers equations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000584,"raw_usage":{"total_tokens":2552,"prompt_tokens":685,"completion_tokens":1867,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":429,"completion_tokens_details":{"reasoning_tokens":1789}},"tokens_in":429,"tokens_out":1867,"duration_ms":12521,"temperature":1.0,"reasoning_tokens":1789,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:03:33.117314+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a minimal full-rank case (e.g., n=2, B=I, A with a nonzero unstable mode), compute the closed-loop V̇ exactly for a bounded disturbance with the controller of Proposition 1; if V̇ is not strictly negative for some state, the theorem is false. For the implemented method, run the heat-equation simulation with N=10000 EnKF particles and measure the steady-state L2 norm against the practical-stability radius r=0.002; norms far exceeding that radius would indicate that EnKF approximation error, not disturbance, limits performance.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the dual EnKF interacting-particle algorithm used to approximate the value function and the optimal control from simulator trajectories."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the Lyapunov redesign method (Chapter 14) that constructs the robustifying term ud against bounded disturbances."},{"cited_title":"K WAKERNAAK AND R","cited_arxiv_id":null,"evidence_quote":"gives the Riccati-equation theory and the stabilizing-gain theorem on which the linear robust control result rests."},{"cited_title":"L IBERZON , Calculus of variations and optimal control theory , Princeton University Press, Princeton, NJ, 2012","cited_arxiv_id":null,"evidence_quote":"provides the Hamilton-Jacobi-Bellman characterization of the optimal control used in the nonlinear case."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces dynamic mode decomposition with control (DMDc), used to build the linear reduced-order simulator for the Burgers equation."},{"cited_title":"Controlgym: Large-Scale Control Environments for Benchmarking Reinforcement Learning Algorithms","cited_arxiv_id":"2311.18736","evidence_quote":"supplies the PDE simulation environments (heat and Burgers equations) used in the numerical experiments."},{"cited_title":"Policy Optimization for PDE Control with a Warm Start","cited_arxiv_id":"2403.01005","evidence_quote":"the preceding DMDc-based PDE control approach that this paper extends, including the test setup and initial-condition sampling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"establishes the log-transform duality between optimal control and nonlinear filtering that motivates the dual EnKF sampling formulation."}],"review_version":1}