{"id":"792e51a8-5844-41c4-8edd-7c81d8eb2533","arxiv_id":"2411.10096","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A pH-structured neural controller is proven to have a finite L2 gain for all parameters, but the claimed finite incremental L2 gain is not proven and fails in simple cases.","lead":"This paper designs port-Hamiltonian controllers whose finite L2 or incremental L2 gains are guaranteed for every parameter value, so neural policies can be trained without stability constraints. The L2 result is correct, but the incremental L2 theorem has a proof gap and is false as stated, making a central contribution unsafe as presented.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4's proof drops the term ½δξᵀ d/dt(∇²H(ξ)) δξ when differentiating Vδ; for H=0.1ξ²+log cosh(ξ), α=εδ, Λ=0, this term is positive along an unforced contracting trajectory, so the claimed all-parameter iL₂ bound fails.","rationale":"I re-derived the core step of Theorem 4 and confirmed the reader's weakest assumption: the dependence of ∇²H(ξ) on ξ is omitted when differentiating Vδ. This is not a cosmetic omission. For the paper's own log-cosh-plus-quadratic Hamiltonian in one dimension with the theorem's equality choice α=εδλ̄(GcGcᵀ), the missing term makes the derivative of Vδ exceed the required −εδ||δu||² by an amount proportional to δξ². The integrated dissipativity inequality therefore fails on arbitrarily small time intervals, so the conclusion of Theorem 4 is false as stated, not merely unproved. This is the load-bearing result of the paper: it supplies the incremental-L₂ unconstrained parametrization, underlies Remark 7 for discrete time, and motivates the DC microgrid experiment. The L₂ Theorem 3 remains valid because its proof uses V=Hc and only requires the standard pH dissipation inequality; the L₂ contribution is sound. No code or baseline comparisons are provided, but that is secondary; the mathematical counterexample is dispositive. The paper could be revised by replacing Theorem 4 with an incremental gain bound that depends on a trajectory-uniform metric-derivative condition or on a larger damping margin, but that would no longer be the advertised parameter-free a priori bound. Verdict: reject the central iL₂ claim; the L₂ part could be salvaged in a revision.","tokens_in":24486,"tokens_out":9739,"duration_ms":111387,"concrete_test":"Run the scalar counterexample simulation described above and compare the integrated dissipativity inequality at a few small times.","verdict_should_be":"REJECT","load_bearing_attack":"Appendix B differentiates Vδ(ξ,δξ)=½δξᵀ∇²H(ξ)δξ as though ∇²H(ξ) were constant along the flow. The full derivative is Vδdot=δξᵀ∇²H δξdot + ½δξᵀ(d/dt∇²H)δξ. Only the first term is used; the omitted metric-derivative term is not sign-definite and is not controlled by the assumptions (14). Because the theorem fixes α=εδλ̄(GcGcᵀ), there is no slack in the damping to absorb a positive ½δξᵀ(d/dt∇²H)δξ. A concrete failure: take scalar ξ, Gc=1, Λ=0, α=εδ, and Hc(ξ)=0.1ξ²+log cosh(ξ), which satisfies 0.2≤∇²H≤1.2. Consider input y=0 with two trajectories starting at ξ(0)=0.5 and ξ̃(0)=0. Writing P=0.2+sech²ξ and δξ=ξ, the variational output is δu=Pδξ, and Vδdot=½Pdotδξ²−αP²δξ². For α=εδ, the claimed εδ-output strict incremental passivity inequality requires Vδdot+εδδu²≤0, but Vδdot+εδδu²=½Pdotδξ², and Pdot=2α sech²ξ tanhξ(0.2ξ+tanhξ)>0 at ξ=0.5. Hence the inequality fails at t=0 for any εδ>0. Since Theorem 4 is the basis of the iL₂ novelty, Remark 7, and the DC microgrid experiment, the central all-parameters iL₂ guarantee is not established. Theorem 3's L₂ result is not affected, since its storage Hc has no variational metric derivative of this type.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a class of distributed port-Hamiltonian (pH) controllers whose storage functions can be parametrized by neural networks, and claims that this yields an unconstrained parametrization with a priori finite L2 or incremental L2 (iL2) gains. The L2 result (Theorem 3) is obtained by choosing the damping coefficient as α = ε λ̄(GcGcᵀ), making the controller ε-output strictly passive for all parameters. The iL2 result (Theorem 4) is obtained by applying the same damping choice to the variational dynamics and using a quadratic form in the Hessian as a differential storage function. The paper also presents a discrete-gradient discretization (Theorem 5) claimed to preserve the L2 passivity property, and two case studies: consensus of non-holonomic robots and voltage/power regulation in DC microgrids. The main iL2 claim is not established: the proof of Theorem 4 omits the time derivative of the Hessian in the storage function, and the theorem is in fact false as stated for a Hamiltonian satisfying its assumptions.","tokens_in":24802,"tokens_out":8108,"duration_ms":82300,"significance":"If Theorem 4 were correct, the paper would provide a genuinely attractive extension of unconstrained stable parametrizations to distributed nonlinear controllers with arbitrary strictly convex storage functions, going beyond quadratic-storage RENs. Theorem 3 and the discrete-gradient L2 preservation argument are valuable and appear sound, and the experimental section demonstrates that the proposed training pipeline can produce reasonable closed-loop behavior. However, the incremental-L2 contribution is the paper's central novelty in Section IV, underpins Remark 7 and the DC microgrid experiment, and the stated theorem has a concrete counterexample. Because the main iL2 guarantee is not merely unproved but false under the theorem's own hypotheses, the significance of the paper as a whole is substantially reduced. I also note that no code or data is provided, so the experimental results are not independently reproducible.","major_comments":[{"comment":"","section":"Appendix B, Theorem 4"},{"comment":"","section":"Remark 7 and Section VI-B"},{"comment":"","section":"Section V, Eq. (17)"}],"minor_comments":[{"comment":"","section":"Abstract"},{"comment":"","section":"Definitions 1 and 2"},{"comment":"","section":"Theorem 4 statement"},{"comment":"","section":"Section VI-A"},{"comment":"","section":"Appendix D"}],"recommendation":"reject","confidential_remarks":"The L2 part of the paper (Theorem 3 and Theorem 5) is sound and could form the basis of a useful separate contribution. The iL2 part, however, is load-bearing for the claimed novelty and is false as stated; the counterexample is elementary and uses the paper's own log-cosh-plus-quadratic Hamiltonian. Fixing this would require either a substantially stronger assumption (e.g., bounded third derivative with a matching damping slack) or a different storage function, and the experiments and Remark 7 would need to be revisited. Given the scope of the changes, I recommend rejection rather than major revision, while noting that the L2 results and the discrete-gradient preservation idea have merit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one for the L2 result, not the iL2 one. Theorem 3 is a clean, valid contribution: pH controllers with alpha = epsilon * lambda_max(Gc Gc^T) are epsilon-output strictly passive with L2 gain <= 1/epsilon for every parameter choice, and the sparsity pattern in Gc is preserved. That gives a genuinely useful unconstrained parametrization of distributed L2-stable neural controllers, more flexible than RENs in storage functions and better suited to distributed sparsity. The discrete-gradient discretization in Theorem 5 also looks sound and is a nice practical add-on.\n\nThe problem is Theorem 4. The proof of incremental passivity differentiates Vdelta = 1/2 delta_xi^T Hessian(H)(xi) delta_xi as if Hessian(H) were constant along the trajectory. It is not. The missed term 1/2 delta_xi^T d/dt(Hessian(H)) delta_xi is not sign-definite, and assumptions (14) do not control it. The stress-test counterexample is correct: with scalar xi, Gc = 1, Lambda = 0, alpha = epsilon_delta, Hc = 0.1 xi^2 + log cosh(xi), and two trajectories starting at 0.5 and 0, the claimed inequality fails at t = 0 for any epsilon_delta > 0. So the all-parameter iL2 guarantee is not established. Remark 7 inherits the gap for the discrete-time case, and the DC microgrid experiment is framed as relying on iL2 preservation, so that part of the abstract's promise does not hold.\n\nThe rest is thinner: no code or baselines in the experiments, and well-posedness of the implicit discrete scheme is deferred to an external reference. Those are minor relative to the Theorem 4 issue.\n\nFor peer review: yes, send it out, but ask the reviewer to focus on Theorem 4. The L2 portion is salvageable if the iL2 claim is removed or properly proved. I would not cite this paper as-is, but I would keep the L2 theorem in mind for future work.","headline":"Sound L2-gain parametrization, but the incremental-L2 Theorem 4 is not proven: the proof drops the metric-derivative term and the claimed all-parameter iL2 guarantee fails on a simple example.","tokens_in":25462,"tokens_out":1689,"would_cite":false,"duration_ms":17723,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C10","93A14","93D25"],"pacs":[],"model":"deepseek-v4-flash","headline":"A distributed controller class based on port-Hamiltonian dynamics is claimed to guarantee finite L2 and incremental L2 gains for every neural-network parameter value, making stability-preserving control an unconstrained optimization…","keywords":["port-Hamiltonian systems","distributed control","L2 gain","incremental L2 stability","unconstrained parametrization","neural network controllers","discrete gradient methods"],"falsifier":"Compute the variational energy $V_\\delta(\\xi,\\delta\\xi)=\\tfrac12\\delta\\xi^\\top\\nabla^2 H_c(\\xi)\\delta\\xi$ for $H_c(\\xi)=0.1\\xi^\\top\\xi+\\log\\cosh(\\xi)$ with $\\Lambda=0$ along a contracting trajectory of the controller; the time derivative contains the term $\\tfrac12\\delta\\xi^\\top \\frac{d}{dt}\\nabla^2H_c(\\xi)\\delta\\xi$, which is omitted in the proof of Theorem 4, and for this $H_c$ it can violate the required inequality, so the claimed bound $\\leq 1/\\epsilon_\\delta$ would not hold.","tokens_in":24143,"feed_emoji":"⚙️","tokens_out":10001,"duration_ms":93848,"temperature":0.7,"pith_summary":"This paper tries to establish that a distributed controller built from port-Hamiltonian dynamics can carry arbitrary neural-network parameters and still be stable by design, in the input-output sense. The main results are an L2-gain theorem and an incremental L2-gain theorem: with the damping coefficient set to a fixed function of the controller's input matrix, every choice of parameters produces a controller with a prescribed gain bound, so no projection or constraint is needed during training. A dissipation-preserving discretization theorem says the same passivity property survives time discretization when a discrete-gradient integrator is used. If these results hold, gradient-based training of distributed nonlinear controllers, with communication sparsity built into the matrices, becomes a standard unconstrained optimization problem, and the same parametrization could serve as a stable-by-design model class for identification.","feed_headline":"Port-Hamiltonian control guarantees L2 gain without constraints","feed_subtitle":"A port-Hamiltonian structure with fixed damping keeps neural controllers stable for every weight, so training needs no projections.","key_machinery":"The load-bearing object is the port-Hamiltonian controller $$\\dot\\xi = [J_c - (\\$\\alpha$ I + \\Lambda)]\\,\\nabla H_c(\\xi) + G_c y, \\quad u = G_c^\\top \\nabla H_c(\\xi),$$ where $J_c$ is a block-diagonal skew-symmetric interconnection matrix, $\\Lambda$ is a diagonal nonnegative damping matrix, $G_c$ carries the communication graph's sparsity pattern, and $H_c$ is a differentiable, radially unbounded energy (Hamiltonian) function that can be a neural network. Port-Hamiltonian means the drift splits into a lossless skew-symmetric part and a dissipative damping part, which forces energy to decrease along trajectories. The proof mechanism is the same in each theorem: use the energy $H_c$ (or its curvature $\\nabla^2 H_c$) as the storage function, then choose $\\alpha \\geq \\epsilon\\bar{\\lambda}(G_cG_c^\\top)$ so that the cross term involving $G_c$ is dominated and the required supply-rate inequality holds for every parameter value. This cancellation is what creates the unconstrained parametrization.","core_discovery":"The central claim is that the port-Hamiltonian controller (10) is dissipative for all parameters. Taking the Hamiltonian $H_c$ as the storage function and setting $\\alpha = \\epsilon \\bar{\\lambda}(G_cG_c^\\top)$ makes the inequality $\\dot H_c \\leq -\\epsilon\\|y\\|^2 + u^\\top y$ hold algebraically, which is exactly $\\epsilon$-output strict passivity and hence a finite $\\mathcal{L}_2$ gain at most $1/\\epsilon$ for any $J_c$, $\\Lambda$, $G_c$, and $H_c$ (Theorem 3). Applying the same idea to the variational system with the Hessian $\\nabla^2 H_c$ as the storage function, the paper claims that $0 < c_1 I \\leq \\nabla^2 H_c \\leq c_2 I$ yields $\\epsilon_\\delta$-output strictly incremental passivity and a finite incremental $\\mathcal{L}_2$ gain at most $1/\\epsilon_\\delta$ (Theorem 4). Discrete-gradient discretization of the same controller is claimed to preserve the $\\epsilon$-output strict passivity and therefore the $\\mathcal{L}_2$ bound (Theorem 5).","pith_inferences":["The incremental-gain proof implicitly requires a condition on how fast $\\nabla^2 H_c$ changes along trajectories; adding a Lipschitz bound on $\\nabla^2 H_c$ or a contraction analysis of the variational system would make the iL2 theorem robust, and such a condition is likely testable numerically during training.","Because the discrete-gradient update is implicit, deploying Theorem 5 in real time requires solving the fixed-point equation (17); guaranteeing existence and uniqueness is a separate numerical step, so the 'parameter-free' guarantee is only as strong as that solver.","In the robot consensus experiment the controller parameters are time-varying during the finite horizon and frozen afterward; one may infer that the a priori stability certificates fully cover the frozen regime, while collision-avoidance behavior during training is a learned property rather than a guaranteed one.","The same parametrization could be used for identification of distributed nonlinear systems with guaranteed $\\mathcal{L}_2$ stability, since the gain bound holds for all weights and the sparsity pattern is fixed."],"forward_implications":["Distributed controllers in this class can be trained end-to-end with standard gradient-based methods, because the stability and gain certificates hold at every iteration and for the final parameters.","Communication constraints enter only through $G_c$ and the block structure of $J_c$, so the guarantee is inherited by sparse, neighbor-based controllers without any added constraints.","If the plant also has a known $\\mathcal{L}_2$ or incremental gain, the small-gain theorem makes the closed-loop system stable with a computable bound, rather than only locally or after training.","The discrete-gradient discretization gives an implementable embedded controller whose $\\mathcal{L}_2$ gain is preserved, unlike forward Euler, which can destroy passivity."],"supporting_citations":[{"why":"Supplies the definitions of dissipativity, $\\epsilon$-output strict passivity, finite $\\mathcal{L}_2$ and incremental gains, the small-gain theorem, and the passivity-to-gain implications used in Theorems 3 and 4.","marker":"[22]"},{"why":"Provides the differential dissipativity and variational-dynamics framework, including the theorem used to pass from differential passivity to incremental passivity of the original controller.","marker":"[45]"},{"why":"Gives the discrete-gradient integrators whose exact energy balance is the mechanism for the passivity-preserving discretization in Theorem 5.","marker":"[38]"},{"why":"Frames the closed-loop trajectory simulation and backpropagation-through-time training used to solve the unconstrained optimal control problem in the experiments.","marker":"[55]"}],"fun_headline_variants":["Neural control stable for every weight, no constraints needed","Port-Hamiltonian guarantees L2 gain for any unconstrained policy","Stable distributed control via unconstrained parametrization","No parameter constraints: port-Hamiltonian nets stay passive","Unconstrained neural policies with guaranteed L2 gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that knowing the curvature of the controller's energy function lies between two positive constants is enough to bound the incremental gain, because the proof does not control how fast that curvature changes along a trajectory.","fun_headline_variants_meta":{"raw":{"variants":["Neural control stable for every weight, no constraints needed","Port-Hamiltonian guarantees L2 gain for any unconstrained policy","Stable distributed control via unconstrained parametrization","No parameter constraints: port-Hamiltonian nets stay passive","Unconstrained neural policies with guaranteed L2 gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000244,"raw_usage":{"total_tokens":1566,"prompt_tokens":1016,"completion_tokens":550,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":470}},"tokens_in":632,"tokens_out":550,"duration_ms":6820,"temperature":1.0,"reasoning_tokens":470,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:00:30.609723+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the variational energy $V_\\delta(\\xi,\\delta\\xi)=\\tfrac12\\delta\\xi^\\top\\nabla^2 H_c(\\xi)\\delta\\xi$ for $H_c(\\xi)=0.1\\xi^\\top\\xi+\\log\\cosh(\\xi)$ with $\\Lambda=0$ along a contracting trajectory of the controller; the time derivative contains the term $\\tfrac12\\delta\\xi^\\top \\frac{d}{dt}\\nabla^2H_c(\\xi)\\delta\\xi$, which is omitted in the proof of Theorem 4, and for this $H_c$ it can violate the required inequality, so the claimed bound $\\leq 1/\\epsilon_\\delta$ would not hold.","supporting_citations":[{"cited_title":"van der Schaft, L2-Gain and Passivity Techniques in Nonlinear Control, 3rd ed., ser","cited_arxiv_id":null,"evidence_quote":"Supplies the definitions of dissipativity, $\\epsilon$-output strict passivity, finite $\\mathcal{L}_2$ and incremental gains, the small-gain theorem, and the passivity-to-gain implications used in Theorems 3 and 4."},{"cited_title":"Convex in- cremental dissipativity analysis of nonlinear systems,","cited_arxiv_id":null,"evidence_quote":"Provides the differential dissipativity and variational-dynamics framework, including the theorem used to pass from differential passivity to incremental passivity of the original controller."},{"cited_title":"A geometric integration approach to smooth optimisation: Foundations of the discrete gradient method,","cited_arxiv_id":null,"evidence_quote":"Gives the discrete-gradient integrators whose exact energy balance is the mechanism for the passivity-preserving discretization in Theorem 5."},{"cited_title":"Neu- ral ordinary differential equations,","cited_arxiv_id":null,"evidence_quote":"Frames the closed-loop trajectory simulation and backpropagation-through-time training used to solve the unconstrained optimal control problem in the experiments."}],"review_version":1}