Learning the Brain's Dynamics as a Port-Hamiltonian System: A GNN-Surrogate Metriplectic Twin for Non-Equilibrium Cortical Dynamics and Closed-Loop Neuromodulation

Dibakar Sigdel

Motor cortex can be fit as a port-Hamiltonian system whose free-running stochastic dynamics recover near-critical avalanche branching from real EEG.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 11:43 UTC pith:T4KIMJ2A

load-bearing objection Body reports a real free-run branching result and mixed criticality; the abstract claims three rungs and closed-loop restoration that the results do not show. the 4 major comments →

arxiv 2607.10439 v2 pith:T4KIMJ2A submitted 2026-07-11 q-bio.NC cs.AI

Learning the Brain's Dynamics as a Port-Hamiltonian System: A GNN-Surrogate Metriplectic Twin for Non-Equilibrium Cortical Dynamics and Closed-Loop Neuromodulation

Dibakar Sigdel This is my paper

classification q-bio.NC cs.AI

keywords port-Hamiltonian systemsmetriplectic dynamicsEEGbrain-computer interfaceneural criticalitygraph neural networksneuromodulationphasor coordinates

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that waking motor cortex is better treated as a non-equilibrium physical system than as a black-box classifier. It represents multi-channel EEG as phasors on a torus, embeds them in a port-Hamiltonian structure with a conservative (skew-symmetric) connectome and a dissipative port, and upgrades the usual passivity picture to a metriplectic non-equilibrium steady state with a metabolic drive and fluctuation–dissipation noise. A graph-neural network learns the energy and the state-dependent coupling and dissipation from PhysioNet EEG motor-imagery recordings under a leakage-free by-subject split, reaching low kinematic test error. The free-running stochastic model reproduces near-critical avalanche branching; the authors present this as a structure-preserving substrate that can generate closed-loop neuromodulation signals and, with further upgrades, structure-preserving BCI decoders.

Core claim

A GNN-parameterised metriplectic port-Hamiltonian model, trained only on real EEG phasors from a wrist-extension / motor-imagery BCI task (1,109,250 training samples; three subjects held out), reconstructs held-out kinematics at test MSE 1.30×10⁻⁴ and, when free-run with fluctuation–dissipation noise, produces near-critical avalanche branching (σ≈1.00 versus measured ≈0.94). The same free runs do not yet match the real 1/f slope or DFA exponent; the paper treats the branching match as the cleared criticality rung and the spectral/temporal gaps as targets for excitation–inhibition and source-space upgrades.

What carries the argument

The Cortical GNN-pHNN: dynamics of the form ẋ = J(x)∇H + M∇S + Gu + b_met (metriplectic / port-Hamiltonian), with H a band-stratified graph energy, J exactly skew-symmetric and gated by measured phase-locking, R ⪰ 0 state-dependent dissipation, and a Fluctuation–Dissipation noise channel at an arousal temperature T.

Load-bearing premise

That alpha-band phase-locking values computed on average-referenced scalp EEG are a valid multiplicative prior for the reversible cortical connectome, rather than being inflated by volume conduction and the common reference.

What would settle it

Replace the scalp PLV prior with a volume-conduction-robust estimator (e.g. weighted phase-lag index) or move the state into source space with a structural-connectome mask; if free-run branching then collapses away from σ≈1 or the learned J becomes anatomically unsupported, the claim that the model captures cortical—not montage—dynamics fails. Concurrent TMS-EEG prediction of evoked responses and PCI (the paper’s own rung 5) would also decide the neuromodulation-port claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Structure-preserving BCI decoders can be built so that stimulation is an energy-shaping control input with passivity/orbital-stability guarantees rather than a black-box classifier output.
Closed-loop neuromodulation signals synthesised from the model can restore phase-locking in silico when applied to desynchronised inputs.
Pathology can be read as a failure of the metabolic–dissipative balance (runaway entropy or hypersynchronous energy growth) rather than as a mere classification error.
A single arousal temperature and a slow neuromodulatory compartment place rest and task conditions on a continuous axis instead of discrete labels.
Clearing the remaining spectral and DFA rungs via excitation–inhibition balance and source-space connectivity would make the free-run a testable digital twin of cortical criticality.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the PLV prior is the main load-bearing artefact, the same architecture with imaginary coherency or wPLI gating should improve the 1/f and DFA scores without changing the metriplectic integrator.
Energy-shaping (IDA-PBC) on the learned H and G offers a concrete route from this twin to adaptive tDCS/TMS protocols whose stability can be certified, not only simulated.
Subject-specific amortised inference over J₀, R₀, delays, and T would turn the population-level model into a personalised twin for closed-loop BCI calibration.
Seizure-onset modelling as dissipation failure is a direct, testable extension once the steady-state power balance is calibrated against haemodynamic data.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Body reports a real free-run branching result and mixed criticality; the abstract claims three rungs and closed-loop restoration that the results do not show.

read the letter

The thing to know first: Section V A is unusually straight. Free-run stochastic rollout on held-out subjects matches near-critical branching (σ≈1.00 vs 0.94) and fails the other two rungs they care about—1/f slope too steep (β≈1.96 vs 1.18) and DFA too persistent (α≈1.68 vs 0.68). That mixed score is useful. The abstract still says the model “passes three scale-free criticality rungs” and “generates closed-loop neuromodulation signals that restore phase-locking in silico.” Neither claim is supported by the free-run figures or by Section V I, which defers IDA-PBC and notes a reachability residual near 0.97 with the three anatomical ports.

What is actually new is the package, not any single ingredient. Port-Hamiltonian structure, metriplectic brackets, HNNs, and GNN energy models are prior art. Fitting a band-stratified GNN-pHNN to real EEGMMIDB phasors under a leakage-free by-subject split, then free-running with fluctuation–dissipation noise and scoring branching/spectrum/DFA like-for-like, is not in the cited literature. Test kinematic MSE of 1.30×10⁻⁴ is fine for what it is. The upgrade path (latent state + emission, volume-conduction-robust connectivity, E–I criticality control) is concrete and matches the failures they report.

Soft spots in proportion: the abstract–body mismatch is the load-bearing problem for anyone who only reads the front matter. Using average-referenced scalp PLV as a multiplicative gate on J(x) is a real but secondary issue—the paper itself later flags zero-lag PLV and field spread. No non-physics baseline, no shipped code. Closed-loop is aspirational, not demonstrated.

This is for people who care about structure-preserving BCI models and criticality diagnostics, not for someone hunting a finished neuromodulation controller. The math and citation pattern look solid; the honesty in the validation section is real. I would send it to peer review after the abstract is aligned with Section V A and the control claim is either shown or removed. Worth engaging once those are fixed; not worth citing as-is for the three-rung or closed-loop claims.

Referee Report

4 major / 5 minor

Summary. The paper proposes a Cortical GNN-pHNN that models motor-cortex EEG (PhysioNet EEGMMIDB, N=64) as a non-equilibrium port-Hamiltonian / metriplectic system on phasor coordinates x=[ϕ,ω]. A band-stratified GNN parameterises the Hamiltonian, a PLV-gated network produces a skew-symmetric connectome J(x), and state-dependent dissipation R(x) is constrained positive-semidefinite; an FDT-consistent noise channel yields stochastic free-runs. Fit under a leakage-free by-subject split (1,109,250 train samples; S010–S012 held out), the model reports test kinematic MSE 1.30×10⁻⁴. Free-run validation is scored against model-independent invariants: near-critical avalanche branching (σ≈1.00 vs measured 0.94) is reproduced, while aperiodic 1/f slope and DFA fail. An upgrade path (latent state, E–I criticality control, volume-conduction-robust connectivity, IDA-PBC neuromodulation) is outlined.

Significance. If the free-run branching result and the structure-preserving formulation hold under corrected reporting, the work is a useful step toward physics-informed BCI models that can, in principle, support energy-shaping control. Strengths that should be credited: (i) an unusually candid body that scores free-run criticality with the same estimators used on data (Section V A, Figs. 5–6); (ii) leakage-free by-subject protocol and multi-seed reporting; (iii) exact skew-symmetry and R⪰0 by construction; (iv) an explicit metriplectic / NESS reformulation with metabolic port and FDT noise rather than strict passivity. The closed-loop neuromodulation claim and the three-rung criticality claim, as currently written in the abstract, are not yet supported by the results and must be aligned with the body before the contribution can be fairly assessed.

major comments (4)

[Abstract; Section V A; Figs. 5–6] Abstract vs Section V A / Figs. 5–6: the abstract asserts that the model “passes three scale-free criticality rungs: near-critical branching ratio (σ≈1), 1/f power-law spectrum, and long-range DFA correlations.” The free-run results state the opposite for two of the three: β_model≈1.96 vs β_real≈1.18 (spectral fail) and α_DFA≈1.68 vs 0.68 (DFA fail); only branching passes (σ=1.00 vs 0.94). This is a load-bearing public claim. The abstract (and any parallel claims in the conclusion) must be rewritten to match the scored free-run outcomes exactly.
[Abstract; Section V I; Eq. (18)] Abstract and closing claims of closed-loop neuromodulation: the abstract states that the model “generates closed-loop neuromodulation signals that restore phase-locking in silico when applied to de-synchronised inputs.” Section V I formulates IDA-PBC as future work, reports a structural reachability residual ≈0.97 with the three anatomical ports of Eq. (18), and defers a closed-loop convergence result until after source-space fit. No in-silico restoration experiment, figure, or quantitative metric is provided. Either supply the experiment with controls or remove/qualify the claim.
[Section III C; Eqs. (7), (16), (19); Section V D] Eqs. (7), (16) and the PLV prior: alpha-band PLV on average-referenced scalp EEG is used both as a multiplicative gate on J(x) and in the L_PLV term of the composite loss (Eq. 19). Section V D itself notes that zero-lag PLV is strongly inflated by volume conduction (controlled example: PLV=0.88 vs wPLI=0.015 for pure field spread). Free-run branching is model-independent and therefore not circular, but any claim that the learned connectome or “functional connectivity” (rung 2) reflects cortical coupling rather than montage artefacts is under-supported until a volume-conduction-robust estimator (imaginary coherency / wPLI) is used, or the limitation is stated as binding on present results rather than only as future work.
[Section III A; Eq. (14); Section V A] Primary coordinate choice: Section III A states that the canonical state x(t)∈R^{128} is formed from the alpha-band phase and frequency only, while the Hamiltonian is band-stratified into five sub-energies and a PAC term (Eq. 14). It is unclear how multi-band H and θ–γ PAC dynamics are evolved when the free-run state is alpha-only, and whether free-run criticality metrics are computed on alpha phasors or on reconstructed multi-band signals. Clarify the state dimension, the emission of non-alpha bands, and which signal enters the avalanche / spectrum / DFA estimators.

minor comments (5)

[Abstract] Abstract placeholders \FitTrainN and \FitTestMSE should be replaced by the concrete numbers used in the body (1,109,250; 1.30×10^{-4}).
[Title; Abstract; Section III A] Title and abstract emphasise a “wrist-extension BCI task,” but EEGMMIDB conditions are eyes-open/closed rest and left/right hand and hands/feet motor imagery; align the task description with the actual dataset.
[Fig. 6; Section V A] Fig. 6 reports real β≈1.4 and model β≈2.1 in the caption, while the text and Fig. 5 use 1.18 and 1.96; reconcile the reported slopes.
[Table II; Section V A] Table II lists six validation rungs; only rungs 1 and 4 are scored on free-run output. A short explicit “not scored” note for rungs 2, 3, 5, 6 would help readers.
[Table I; Section V A] Notation: both the DFA exponent and the alpha band use α; consider α_DFA vs α-band to avoid collision in Section V A and Table I.

Circularity Check

1 steps flagged

Mild PLV prior/loss reuse for connectivity; free-run criticality metrics are independent of the training objective and not circular.

specific steps

fitted input called prediction [Eqs. 16, 19; Section V intro (rung 2)]
"J(x)←J(x)⊙PL V (α) ... LPL V=‖f|J| − ^PL V(α)‖2_F ... we report below the rungs we can score on open EEG data — spectral fidelity (rung 1), functional connectivity (rung 2, via the measured PLV prior), and criticality (rung 4)"

Measured alpha-band PLV is injected as a multiplicative gate on the reversible connectome and as an explicit training loss that pulls |J| toward the same PLV matrix. Scoring “functional connectivity” then partly via that measured prior reuses the fitted input rather than testing an independent free-run recovery of connectivity. Limited to rung 2; not load-bearing for free-run branching/spectrum/DFA.

full rationale

The load-bearing free-run validation (Section V A, Fig. 5) scores branching ratio, aperiodic 1/f slope, and DFA with the same model-independent estimators applied to recordings (Section III H). None of those three quantities appears in the five-term training loss (Eq. 19: kinematic, passivity, energy-balance, PLV, PAC). Branching therefore emerges from the stochastic metriplectic rollout rather than being fitted; spectrum and DFA fail in the body, which is the opposite of a forced success. Hard pH structure (skew-symmetric J by construction, Eq. 15; R ⪰ 0 via softplus, Eq. 17) is architectural constraint, not a circular prediction. The only mild circularity is that measured alpha PLV is both a multiplicative gate on J(x) (Eq. 16) and a coherence-loss target (L_PLV in Eq. 19), so any claim to recover functional connectivity is partly by construction—but the paper scores free-run connectivity only as “via the measured PLV prior” and does not report an independent free-run connectivity number in Fig. 5. No self-citation uniqueness chain; citations for metriplectic/GENERIC, HNN, and pH structure are external. Abstract overstatement (three criticality rungs pass; closed-loop restoration) versus body (only branching passes; IDA-PBC deferred) is a correctness/overclaim issue, not circular derivation. Overall circularity is low.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The central free-run criticality claim rests on a large fitted GNN surrogate, hand-chosen loss weights and temperature, and several domain modeling choices (Stuart-Landau phasors, scalp PLV as connectome prior, diagonal state-dependent R, three anatomical ports). The metriplectic degeneracy and FDT noise form are standard theory imported from GENERIC/FDT; the Cortical GNN-pHNN composite and the metabolic port as used here are paper-specific constructs without independent external measurement in this work.

free parameters (6)

GNN / MLP weights for H, J, R, g_PAC
All energy, connectome, dissipation, and PAC-gating networks are trained on 1,109,250 EEG samples; the free-run invariants depend on these fitted maps.
Composite loss weights λ_k, λ_p, λ_eb, λ_PLV, λ_PAC
Fixed by hand at 1.0, 0.5, 0.1, 0.05, 0.02 (Section III F); they determine the trade-off between kinematics, passivity, and coherence priors.
Fluctuation temperature T
Free-run criticality is scored at T=0.2 (Section V A); called a body-temperature / arousal proxy without physical calibration.
Per-channel baseline dissipation r0_j and softplus offsets
Learned regional decay rates (Eq. 17) that set the irreversible bracket linearisation.
Avalanche threshold 2.5 σ and spectral fit band 2–45 Hz
Estimator hyperparameters that define the criticality and 1/f targets used to score the model (Section III H).
Adam cosine schedule (η_max=5e-4, η_min=1e-5, clip=2.0, ≤400 epochs)
Training hyperparameters that affect the reported test MSE 1.30×10⁻⁴.

axioms (6)

domain assumption Near a Hopf bifurcation, band-limited cortical LFP is well described by Stuart-Landau oscillators, justifying phasor coordinates x=[ϕ,ω].
Section II A; motivates coordinates but is not used as a data generator.
standard math Metriplectic degeneracy J∇S=0 and M∇H=0, with R recovered as linearisation of the irreversible bracket.
Section II E; standard GENERIC structure (Morrison, Grmela, Öttinger).
domain assumption Fluctuation–dissipation: noise covariance σσᵀ = 2 T R(x).
Eq. 11; classical FDT form applied to cortical arousal temperature.
ad hoc to paper Alpha-band PLV on scalp EEG is a biologically valid multiplicative prior for skew-symmetric J(x).
Eqs. 15–16; later contradicted by the paper’s own volume-conduction critique in Section V D.
ad hoc to paper Dissipation is diagonal and state-dependent via softplus; ports G are fixed anatomical averages over 10–20 groups.
Eqs. 17–18; structural choices that limit controllability (~5% reachability, Section V I).
domain assumption Strict passivity is only a first approximation; the correct invariant is orbital NESS power balance with metabolic port.
Section II E; standard non-equilibrium refinement for self-sustained cortex.

invented entities (3)

Cortical GNN-pHNN (band-stratified graph energy + PLV-gated connectome + PAC path) no independent evidence
purpose: Parameterise H, J, R on 64-channel phasors while enforcing skew-symmetry and R⪰0 by construction.
Composite architecture assembled for this paper; no external independent measurement of its internal sub-energies beyond EEG fit residuals.
Explicit metabolic port b_met / P_met sustaining the resting limit cycle no independent evidence
purpose: Replace strict passivity with non-equilibrium steady-state power balance on the limit cycle.
Introduced in Section II E as the physical anchor for NESS; not calibrated to BOLD/fNIRS/CMRglc in this work.
Neuroanatomical three-port matrix G (frontal tDCS, P/O TMS, temporal DBS/tACS) no independent evidence
purpose: Provide structured actuation channels for closed-loop neuromodulation claims.
Anatomical averaging (Eq. 18); paper reports only ~5% reachability and defers closed-loop convergence.

pith-pipeline@v1.1.0-grok45 · 20500 in / 4583 out tokens · 54705 ms · 2026-07-14T11:43:07.560887+00:00 · methodology

0 comments

read the original abstract

We model human motor cortex, recorded during rest and motor-imagery BCI conditions, as a port-Hamiltonian system: a conservative interconnection (skew-symmetric coupling between band-limited neural phasors) together with a dissipative port whose state-dependent decay is set by a graph-neural-network surrogate. The Hamiltonian is resolved into five interpretable frequency sub-energies, and a phase-locking prior measured from the recordings gates the learned functional connectome so that coupling is admitted only where phase coherence is present. A metriplectic formulation places the resting cortex at a non-equilibrium steady state sustained by an explicit metabolic port, with a fluctuation-dissipation-consistent noise channel governed by a single arousal temperature. Fitting the model to 'FitTrainN' phasor samples from the PhysioNet EEG Motor Movement/Imagery database, under a leakage-free split with three subjects held out entirely, yields a held-out kinematic reconstruction error of 'FitTestMSE' that is stable across random seeds. We then score the free-running model against model-independent dynamical invariants it did not author: it reproduces near-critical avalanche branching ($\sigma\approx1$) but not yet the aperiodic $1/f$ spectral slope or the long-range temporal correlations of real cortex a concrete, falsifiable gap that we trace to specific, testable upgrades. The port-Hamiltonian structure supplies neuroanatomically grounded stimulation ports with stability guarantees, positioning the model as a physically principled, structure-preserving substrate for closed-loop neuromodulation.