{"id":"b63e9249-08c9-4381-b3fe-07dbc069c814","arxiv_id":"1909.02702","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Neural network weights are trained as the state of a port-Hamiltonian ODE whose Hamiltonian is the loss, giving a continuous-time optimizer.","lead":"The paper replaces discrete neural-network weight updates with continuous-time dynamics: weights evolve like a damped physical system whose energy is the loss function. The authors claim this guarantees convergence to a loss minimum, and they demonstrate the idea on a linear classifier and a Duffing vector-field task.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The guarantee of convergence to a minimum is not established: passivity plus Lyapunov stability of minima does not rule out convergence to saddles or maxima.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: passivity implies monotone decrease of H, not convergence to a minimum. The paper's proof relies on 'any strict minimum of H is a Lyapunov stable equilibrium' and therefore concludes the state 'will eventually land in one minimum point'; this inference is invalid without additional invariance or strictness conditions. LaSalle's principle only yields convergence to the largest invariant set where damping vanishes, which includes saddles and maxima. The paper supplies no argument excluding non-minimum critical points from the omega-limit set, and one reported experiment even describes the state passing through a saddle point, which directly undercuts the saddle-avoidance claim. Since the central advertised guarantee is unsupported and the experiments are too minimal to substitute for the missing proof, the REJECT verdict stands without modification.","tokens_in":11919,"tokens_out":4593,"duration_ms":53913,"concrete_test":"Apply the claimed theorem to a one-dimensional loss J(ϑ) = -ϑ^2/2 with B > 0. For Eq. (9), (ϑ, dotϑ) = (0,0) is an equilibrium at the global maximum; starting there, the trajectory never converges to a minimum, so any theorem claiming guaranteed convergence to a minimum is false. For a non-degenerate version, take J(ϑ) = ϑ1^2 - ϑ2^2 and compute the largest invariant subset of {dotϑ = 0}; it contains the saddle (0,0), and LaSalle's principle gives no reason to exclude it. Either check settles that passivity alone cannot support the Section IV-A claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim appears in Section IV-A and the conclusion: PH passivity gives an implicit guarantee of convergence to a minimum and circumvents saddle points. For Eq. (9), with H = J* and R = diag(0, B), H-dot = -(partial_omega J*)^T B (partial_omega J*) = -dotϑ^T B dotϑ <= 0, so H is non-increasing. This only implies, via LaSalle's invariance principle, convergence to the largest invariant set contained in {dotϑ = 0}. Invariance inside that set requires partial_ϑ J* = 0, so the omega-limit set is a subset of the critical points of J* with zero velocity. Critical points include saddles and maxima. The paper asserts that because strict minima are Lyapunov stable, 'the state will eventually land in one minimum point'; that is a converse of Lyapunov's theorem and is false without additional assumptions, such as a strict Lyapunov function on the whole domain or a proof that non-minimum critical points cannot belong to the omega-limit set. The experimental passage through a saddle in Section V-B actually illustrates that the dynamics can approach a saddle, undercutting the claim that saddles are circumvented. This missing argument is load-bearing because the paper's advertised advantage over gradient-based methods is exactly this convergence certificate. In addition, Remark 4.7 concedes that sequential training converges to minima of individual-sample losses, not of the batch objective, so even the empirical setup does not realize the claimed guarantee.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a port-Hamiltonian (PH) framework for neural network training in which the network parameters evolve continuously in time as the state of an autonomous PH system whose Hamiltonian is the training loss. The authors claim that the passivity property of PH systems guarantees asymptotic convergence of the parameters to a minimum of the loss, thereby providing an implicit convergence certificate and avoiding the saddle-point problem of discrete gradient methods. The framework is instantiated with mechanical-like second-order dynamics (Eq. 9) in which the loss is augmented with a kinetic term, and two training schemes are described: sequential data training and batch training. Numerical experiments are reported on a linear classification task and on a Duffing oscillator vector-field approximation task.","tokens_in":12208,"tokens_out":6680,"duration_ms":72111,"significance":"If the central convergence claim were valid, the paper would offer a genuinely attractive alternative to discrete optimizers: a continuous-time training dynamics with an automatic guarantee of reaching a loss minimum, together with a principled way to inject damping and shape energy. The authors provide a clear derivation of the passivity inequality and make their code publicly available, which are strengths. However, the advertised guarantee is not established: the passivity calculation only gives non-increase of the Hamiltonian, and the leap to asymptotic convergence to a minimum is a substantial unsupported step. Since the paper's main contribution rests on this guarantee, the significance of the work in its present form is largely conditional on a missing analysis.","major_comments":[{"comment":"The central claim that the state 'will eventually land in one minimum point' is not justified by the preceding argument. The passivity property only implies H-dot = -(∂H)^T R ∂H ≤ 0, which for the mechanical PH system (9) gives equality only when ω = 0. Applying LaSalle's invariance principle, the omega-limit set is contained in the largest invariant set inside {ω = 0}; invariance there requires ∂ϑ J* = 0, so the limit set is a subset of the critical points of J*, which include saddle points and maxima, not just minima. That each strict minimum is Lyapunov stable does not imply that every trajectory converges to one of those minima; a converse Lyapunov statement of this type is false without additional assumptions (e.g., a strict Lyapunov function on the whole domain, or a proof that non-minimum critical points have no basin of attraction). This gap is load-bearing because the abstract, Section IV-B, and Section VI all rest on the assertion that PH passivity ensures convergence to a minimum.","section":"Section IV-A"},{"comment":"The numerical experiment in Section V-B reports that 'the state pas[s]es through a saddle point' of the batch loss. This observation directly undercuts the claim in Section VI that the framework 'circumvents the problem of getting stuck at saddle points.' While passing through a saddle is not the same as getting stuck at one, the paper provides no proof that a saddle cannot be an omega-limit point; indeed, the theoretical gap identified above allows exactly that possibility. The experimental passage through a saddle is therefore consistent with the weaker conclusion (convergence to critical points) and does not support the stronger convergence-to-minimum claim.","section":"Section V-B"},{"comment":"There is an inconsistency between the general definition of the augmented loss and the special case in Example 4.2. In Section IV-B, J*(û, ŷ, ξ) is defined as J(û, ŷ, ϑ) + ω^T M^{-1}(ϑ)ω, with no factor of 1/2 on either term. In Example 4.2, with M = I and ω = ϑ̇, the definition becomes J* = 1/2[α||ŷ - f||² + βϑ^Tϑ + ϑ̇^Tϑ̇], i.e., both the potential and kinetic terms are halved. Consequently, Example 4.2 is not a special case of the stated general Hamiltonian, and the gradient and equations of motion derived from it do not correspond to the PH system (9) with H as defined in the general formula. This makes the example internally inconsistent with the framework it is meant to illustrate.","section":"Example 4.2 and Section IV-B"},{"comment":"Assumption 4.1 only requires the Hamiltonian to be lower-bounded and the solutions to be forward-complete. These conditions are not sufficient for the claimed convergence to a minimum. Non-increase of H along trajectories does not by itself guarantee that trajectories are bounded or that their omega-limit sets are nonempty and compact; additional coercivity or compactness of sublevel sets is needed. Even if the omega-limit set is compact, LaSalle's principle only locates it inside the critical set, not at minima. The paper therefore omits a necessary part of the convergence analysis, regardless of how the saddle-point issue is resolved.","section":"Assumption 4.1 and convergence conditions"},{"comment":"Remark 4.7 concedes that sequential training converges to a minimum of the single-sample loss J*(û_ζ, ŷ_ζ, ξ) at the moment the procedure is stopped, which may not coincide with a minimum of the batch objective J*_batch. This is a serious limitation for the experimental validation: the classification experiment in Section V-A uses the sequential method, so its behavior is not governed by the convergence guarantee claimed for the batch objective. The paper would need to either provide a guarantee for the sequential/hybrid dynamics or clearly separate the theoretical claim (batch) from the experimental demonstration (sequential).","section":"Remark 4.7"}],"minor_comments":[{"comment":"The sentence 'The set R (R+) is the the set of real (non negative real) numbers' contains a duplicated article 'the'.","section":"Section III-A"},{"comment":"The caption of Figure 7 labels the loss axis as 'J*(t) [log]'; the plot in the text appears to show J* on a logarithmic scale, but the axis label should specify the base or the transformation to avoid ambiguity.","section":"Section V-A"},{"comment":"The conclusion states that the framework 'increases the probability of finding the global minimum of the objective function,' but no probability statement or mechanism is analyzed in the paper; this claim should be either supported or removed.","section":"Section VI"},{"comment":"Reference [29] is listed as 'Submitted to 11th IFAC Symposium on Nonlinear Control Systems (NOLCOS 2019)'; if it has since appeared, the published version should be cited, and if it remains unpublished, the dependence of Remark 4.4 on it should be clarified.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper's central theoretical claim—that port-Hamiltonian passivity ensures convergence to a minimum of the loss—is not justified and, as stated, is false in general. The missing argument is not a minor gap: it is the foundation of the paper's contribution. The authors would need to substantially weaken the claim (e.g., to convergence to critical points) or add nontrivial assumptions and a full proof, and the experimental program would need to align with the corrected claim. On that basis I recommend rejection, though a resubmission with a properly scoped theorem could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know up front: the paper's central claim is not established. It says the port-Hamiltonian dynamics guarantee convergence to a minimum of the loss and circumvent saddle points. The math only gives H-dot <= 0, which by LaSalle implies convergence to the largest invariant set where damping vanishes — that set includes saddles and maxima. The step from \"loss decreases\" to \"we land in a minimum\" needs strictness or extra invariance assumptions that are never supplied. Their own experiment in Section V-B even reports the trajectory passing near a saddle, which directly contradicts the saddle-escape claim. What's actually new is the packaging, not the dynamics. Equation (9) is the standard damped second-order gradient system, i.e., continuous-time momentum, and the paper honestly cites [22] and [23] for Hamiltonian dynamics in learning. The port-Hamiltonian framing is a legitimate notation, and the passivity identity is computed correctly. The writing is clear, and the code is available, which is more than many papers do. The soft spots are mostly consequences of the overclaim. The sequential training setup, by the paper's own Remark 4.7, converges to minima of individual sample losses, not the batch objective — so the experiment doesn't even realize the advertised guarantee. The experiments themselves are minimal: one linear boundary and one vector field, no baselines, no error bars, no comparison to standard optimizers. There is also a minor factor-of-1/2 inconsistency in Example 4.2, but that's cosmetic next to the convergence gap. For a reader, this is a useful cautionary example of why passivity alone doesn't give you convergence to a minimum. If the authors weakened the claim to \"convergence to critical points\" or added conditions that exclude saddles, there would be a publishable note here. As written, the main result is misleading. Who gets value from this? Control theorists curious how PH tools map onto optimization might skim it. But I wouldn't cite it, and I wouldn't bring it to reading group except as a teaching moment. It deserves a serious referee — the theoretical claim is substantive and the error is instructive — but the right outcome after review is rejection unless the convergence theorem is fixed.","headline":"The paper's advertised convergence-to-a-minimum guarantee does not follow from its passivity argument; the dynamics are continuous-time momentum in port-Hamiltonian clothing, and the experiments are too thin to rescue it.","tokens_in":727,"tokens_out":824,"would_cite":false,"duration_ms":35198,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","37N35","93D15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Training a neural network by treating its parameters as a port-Hamiltonian system yields a loss that falls continuously and, the paper argues, lands at a minimum.","keywords":["port-Hamiltonian systems","neural network training","continuous-time optimization","passivity","Lyapunov stability","damping injection","saddle-point escape"],"falsifier":"Construct a smooth lower-bounded loss with a strict saddle point, such as $J(x,y)=x^2-y^2+y^4$, initialize the port-Hamiltonian dynamics exactly at that saddle with zero velocity, and integrate. If the trajectory remains at the saddle instead of leaving it, the claim that the state will eventually land in a minimum fails without an extra assumption. A broader check is to run the dynamics on a loss with a continuum of critical points and observe whether the final point is always a local minimum.","tokens_in":11710,"feed_emoji":"📉","tokens_out":10602,"duration_ms":102116,"temperature":0.7,"pith_summary":"This paper proposes training neural networks by replacing discrete parameter updates with a continuous-time dynamical system. The parameters, together with a momentum variable, are treated as the state of a port-Hamiltonian system whose Hamiltonian is the loss function plus a kinetic-energy term. Because port-Hamiltonian systems dissipate energy, the augmented loss is non-increasing along every trajectory, and the authors argue that this forces the parameters to converge to a minimum of the original loss. The proposal therefore tries to supply what stochastic gradient descent lacks: an intrinsic convergence certificate for non-convex training. Two numerical experiments, a linear classification task and a nonlinear-oscillator vector-field approximation, illustrate the dynamics.","feed_headline":"Port-Hamiltonian weight dynamics converge to a loss minimum","feed_subtitle":"By treating weights as a physical system, the loss falls continuously to a minimum.","key_machinery":"The central object is a port-Hamiltonian system, a dynamical system of the form $\\dot{\\xi}=(J(\\xi)-R(\\xi))\\,\\partial H(\\xi)+g(\\xi)v$ with skew-symmetric $J$, dissipative $R\\succeq 0$, and Hamiltonian $H$; passivity gives $\\dot{H}\\le z^\\top v$, and in the autonomous case $\\dot{H}=-(\\partial H)^\\top R\\,\\partial H\\le 0$. Here the Hamiltonian is the augmented loss $H=J^*$, and the specific structure $F=\\begin{bmatrix}0&I\\\\-I&-B\\end{bmatrix}$ produces mechanical-like second-order parameter dynamics with damping $B$. This carries the argument because monotone decay of $J^*$, together with Lyapunov stability of its strict minima, is what the authors invoke to conclude convergence to a minimum of the loss.","core_discovery":"The central claim is that a neural network can be trained by making its parameters $\\vartheta$ the position variables of a dissipative mechanical system. The paper sets $\\xi=(\\vartheta,\\omega)$ with $\\omega = M(\\vartheta)\\dot{\\vartheta}$, defines the augmented Hamiltonian $J^*(\\hat{u},\\hat{y},\\xi)=J(\\hat{u},\\hat{y},\\vartheta)+\\frac{1}{2}\\omega^\\top M^{-1}\\omega$, and chooses the port-Hamiltonian structure $F=\\begin{bmatrix}0&I\\\\-I&-B\\end{bmatrix}$ with $B=B^\\top>0$. The resulting dynamics $\\dot{\\xi}=F\\,\\partial J^*$ make $J^*$ strictly decrease until the state reaches a minimizer of $J^*$, i.e. a minimum of $J$ with zero velocity. The paper states this as an implicit guarantee: passivity excludes getting stuck at saddle points, so the state will eventually land in one minimum point. This convergence claim is the load-bearing result, with the experiments serving as demonstrations rather than proofs.","pith_inferences":["The paper's inference from passivity to landing in a minimum is stronger than monotone energy decrease alone supports: without extra strictness or isolatedness assumptions, trajectories can in principle end at saddle points or other invariant sets where damping vanishes.","A concrete test would initialize the port-Hamiltonian dynamics exactly at a strict saddle of a lower-bounded loss and check whether the trajectory leaves; if it does not, the universal convergence claim needs qualifications.","The framework suggests viewing continuous-time optimizers as energy-shaping problems, where the augmented loss, damping, and injection term are designed to steer basins of attraction.","Because parameters evolve continuously, the training process defines a smooth path through weight space, which could make loss-landscape connectivity and model interpolation directly observable even though the paper does not explore that."],"forward_implications":["Loss monotonicity is built in: along any training run the augmented loss never increases, so interrupting the integration at any time gives a network whose training loss is no worse than at the start.","The dynamics are second-order and dissipative, so the scheme is a continuous-time analogue of momentum-based optimization, with the damping matrix $B$ playing the role of friction.","In batch training, the gradient of the average loss is the average of the gradients, so the port-Hamiltonian optimizer integrates a single ordinary differential equation rather than iterating discrete updates.","Damping injection $v=-k(t)\\dot{\\vartheta}$ provides a control port for reshaping basins of attraction, which the paper proposes as a path toward increasing the probability of reaching a global minimum.","Computational complexity remains linear in the number of parameters, with one extra momentum variable per weight, so the method is comparable to gradient descent in cost."],"supporting_citations":[{"why":"It introduces port-controlled Hamiltonian systems and their structural properties.","marker":"[12]"},{"why":"It supplies the passivity property of port-Hamiltonian systems that yields $\\dot{H}\\le z^\\top v$.","marker":"[14]"},{"why":"It provides the damping-injection stabilization result that the paper invokes to assert convergence to a minimum.","marker":"[15]"},{"why":"It defines continuous-depth networks that are still optimized with discrete methods, and this contrast motivates the continuous-parameter training approach.","marker":"[18]"},{"why":"It gives an earlier energy-based network with a Lyapunov convergence guarantee, which is the lineage the paper extends.","marker":"[25]"},{"why":"It defines the nonlinear oscillator whose vector field serves as the batch-training test problem.","marker":"[31]"}],"fun_headline_variants":["Weights as a mechanical system: training via port-Hamiltonian dynamics","Port-Hamiltonian weight dynamics: continuous descent to a loss minimum","Replacing gradient descent with continuous port-Hamiltonian evolution","From discrete steps to continuous flow: a port-Hamiltonian training rule","How port-Hamiltonian mechanics makes neural network training converge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an energy that never increases, together with stable minima, is enough to guarantee that training lands at a minimum rather than stalling at a saddle or another equilibrium.","fun_headline_variants_meta":{"raw":{"variants":["Weights as a mechanical system: training via port-Hamiltonian dynamics","Port-Hamiltonian weight dynamics: continuous descent to a loss minimum","Replacing gradient descent with continuous port-Hamiltonian evolution","From discrete steps to continuous flow: a port-Hamiltonian training rule","How port-Hamiltonian mechanics makes neural network training converge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000695,"raw_usage":{"total_tokens":3103,"prompt_tokens":867,"completion_tokens":2236,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":2147}},"tokens_in":483,"tokens_out":2236,"duration_ms":17864,"temperature":1.0,"reasoning_tokens":2147,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:42:12.484187+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a smooth lower-bounded loss with a strict saddle point, such as $J(x,y)=x^2-y^2+y^4$, initialize the port-Hamiltonian dynamics exactly at that saddle with zero velocity, and integrate. If the trajectory remains at the saddle instead of leaving it, the claim that the state will eventually land in a minimum fails without an extra assumption. A broader check is to run the dynamics on a loss with a continuum of critical points and observe whether the final point is always a local minimum.","supporting_citations":[{"cited_title":"Port-controlled hamiltonian systems: modelling origins and systemtheoretic properties","cited_arxiv_id":null,"evidence_quote":"It introduces port-controlled Hamiltonian systems and their structural properties."},{"cited_title":"Port-hamiltonian systems theory: An introductory overview","cited_arxiv_id":null,"evidence_quote":"It supplies the passivity property of port-Hamiltonian systems that yields $\\dot{H}\\le z^\\top v$."},{"cited_title":"Putting energy back in control","cited_arxiv_id":null,"evidence_quote":"It provides the damping-injection stabilization result that the paper invokes to assert convergence to a minimum."},{"cited_title":"Neural ordinary differential equations","cited_arxiv_id":null,"evidence_quote":"It defines continuous-depth networks that are still optimized with discrete methods, and this contrast motivates the continuous-parameter training approach."},{"cited_title":"Neural networks and physical systems with emer- gent collective computational abilities","cited_arxiv_id":null,"evidence_quote":"It gives an earlier energy-based network with a Lyapunov convergence guarantee, which is the lineage the paper extends."},{"cited_title":"The Duffing equation: nonlin- ear oscillators and their behaviour","cited_arxiv_id":null,"evidence_quote":"It defines the nonlinear oscillator whose vector field serves as the batch-training test problem."}],"review_version":1}