△ Mechanism confirmed, baseline not beaten 2026

Hodge-Coercive Energy-Preserving Latent Dynamics

Implementation & benchmark of arXiv:2608.19360 — Hodge Coercivity and Global Dynamics in Two-Field Edge-Cochain Systems with MHD-Type Cancellation

Usefulness8/10

Difficulty6/10

Novelty6/10

Dynamics Stability Accuracy Graph nn Ssm Rnn Training dynamics Dynamical systems Control theory Geometry Linear algebra Spectral theory

Source paper: Hodge Coercivity and Global Dynamics in Two-Field Edge-Cochain Systems with MHD-Type Cancellation arXiv:2608.19360 ⓘ · analyzed Sep 1, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Replace an unconstrained graph-neural latent ODE or recurrent transition with two edge-cochain states whose linear drift is Hodge-Laplacian dissipation and whose quadratic coupling is generated by a skew-symmetric anticommutator. The coupling remains expressive while cancelling from the total energy, so the long-time envelope is determined by the Hodge spectral gap rather than uncontrolled nonlinear growth.

Formulas

$$J^\top=-J,\qquad \big(D(a)J+JD(a)\big)^\top=-\big(D(a)J+JD(a)\big).$$

$$B(a,b)=P\big(D(a)J+JD(a)\big)b,\qquad \langle x,N_x(x,y)\rangle+\langle y,N_y(x,y)\rangle=0.$$

$$E=\frac12(\|x\|^2+\|y\|^2),\qquad \dot E=-\nu\big(\langle x,Lx\rangle+\langle y,Ly\rangle\big)+\langle x,f_x\rangle+\langle y,f_y\rangle.$$

$$\ker L=\{0\}\Longleftrightarrow \lambda_1>0,\qquad \dot E\le -\nu\lambda_1E+\frac{\|f_x\|^2+\|f_y\|^2}{2\nu\lambda_1}.$$

Mathematical statement

Let $V$ be the divergence-free edge-cochain subspace, $P$ its orthogonal projector, $L=P\Delta_1P$ the restricted Hodge Laplacian, and $x,y\in V$ two latent fields. Let $D(a)$ denote diagonal multiplication by a cochain $a$, and let $J^\top=-J$ be a trainable skew-symmetric operator. The algebraic mechanism is that $D(a)J+JD(a)$ is skew-symmetric, so the projected quadratic map $B(a,b)=P(D(a)J+JD(a))b$ has the required trilinear antisymmetry and cancels in the total energy. For viscosity $\nu>0$, forcing $f_x,f_y$, and exchange terms $N_x,N_y$ satisfying $\langle x,N_x\rangle+\langle y,N_y\rangle=0$, the total energy $E=(\|x\|^2+\|y\|^2)/2$ obeys an energy inequality. If the harmonic space $\ker L$ is trivial, $\lambda_1=\lambda_{\min}(L)>0$ and the dynamics are globally dissipative.

Implementation notes

Use this as the transition function of a graph neural ODE, continuous-time state-space model, or recurrent edge layer. Construct the cell-complex incidence matrices once, form the edge Hodge Laplacian $\Delta_1$, compute the divergence-free projector $P$, and restrict latent edge channels to $V=\operatorname{range}(P)$. Maintain two states $x$ and $y$. For implementation, initialize an unconstrained matrix $A$ and set $J=A-A^\top$. Compute $L=P\Delta_1P$. At each step calculate $q_x=P(D(x)J+JD(x))y$ and construct $q_y$ with the corresponding two-field exchange signs so that $\langle x,q_x\rangle+\langle y,q_y\rangle=0$. Update $x\leftarrow x+\delta[-\nu Lx+q_x+f_x]$ and $y\leftarrow y+\delta[-\nu Ly+q_y+f_y]$. Prefer implicit midpoint or an energy-monitoring adaptive solver because forward Euler introduces discretization energy error. Compute $P,L,$ and $\lambda_1$ exactly on small complexes; estimate the gap with Lanczos on large graphs. The first cheap experiment should use a triangular-grid advection dataset and compare an unconstrained MLP transition, a Laplacian-only transition, and this model. With zero forcing, predict $E(t)\le E(0)e^{-\nu\lambda_1t}$; measured decay should approach the slope $-\nu\lambda_1$ independently of interaction amplitude. With forcing, the absorbing energy scale should be proportional to $(\|f_x\|^2+\|f_y\|^2)/(\nu^2\lambda_1^2)$.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Inconclusive

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 9/10

Built and ran a NumPy/SciPy MVP implementing the projected anticommutator interaction and two-field energy-neutral exchange dynamics on a small filled-square complex. The mechanism manifested: projector error was 3.45e-16, boundary identity error was 0, nonlinear energy-rate cancellation stayed below 2.2e-14 across interaction strengths, and the decay bound was never violated. Slow-mode decay matched -2 nu alpha lambda1 to relative error below 6e-12 across five gap scalings; the zero-gap control changed energy by only 2.2e-9. The unconstrained quadratic baseline exploded to 2.39e26 energy while the idea remained dissipative, although this was only a secondary toy comparison.

Agent confidence: 9/10

Baseline: Unconstrained random quadratic baseline: final/max energy 2.39e26 over T=8, versus initial energy approximately 3.33.
Idea: lambda1=2.0; maximum absolute cancellation rate 2.13e-14; decay-bound violations 0 for gamma=0..50; slow-mode slopes [-0.35,-0.70,-1.40,-2.80,-5.60] for alpha [0.25,0.5,1,2,4], matching theory; forced linear equilibrium energy/force^2=0.5647 across amplitudes 0.25..4; zero-gap relative energy change=2.16e-9.

Limitations: This is a deterministic low-dimensional numerical verification, not a trained graph-neural ODE or triangular-grid advection benchmark. The baseline is a deliberately random unconstrained quadratic transition; no implicit midpoint discretization, GPU implementation, or nonlinear forced absorbing-radius experiment was tested.

How to run: python3 hodge_energy_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.19360) Sep 2, 2026

No significant difference vs. the tuned baseline (+33.7%, p=0.4069, wins 3 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0034
Idea mean: 0.0045
Effect (Δ): 0.0011 (+33.7%; negative = idea better)
Wins: 3 / 8 paired seeds
p-value: 0.4069 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±79.5%
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 6 configs

Limitations:

Only the 400-sample built-in dynamics track was tested. Longer-horizon rollout metrics, implicit midpoint/adaptive integration, larger graph complexes, alternative latent dimensions, and additional datasets were not evaluated; the MVP uses a fixed identity Laplacian rather than a task-specific cell-complex Hodge Laplacian.

How to run: python3 bench_hodge_stage2.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Sep 1, 2026 latest Failed ✗

Agent confidence: 10/10

Implemented the Hodge-coercive two-field latent transition on the structurally matched built-in dynamics track with identical data, outer architecture, training loop, budgets, and shared hyperparameter grid. The trained-model signature confirmed near-zero interaction power, but the idea had worse test MSE than the tuned baseline, with no significant win.

Baseline: 8-seed test MSE 0.0033659555 ± 0.0008568638; best lr=0.01, weight_decay=0.0001.
Idea: 8-seed test MSE 0.0044999952 ± 0.0028251554; best lr=0.003, weight_decay=0.0001; observed interaction power 1.4863e-09, confirmed=true.

Benchmark result