✓✓ Beats tuned baseline 2026

Heterogeneity-Calibrated Hopf RNN

Implementation & benchmark of arXiv:2607.28275 — Synchronization, Kinematic Waves and Spike-Phase-Separation in Feedback Ising Neural Networks on Heterogeneous Graphs

Usefulness8/10

Difficulty5/10

Novelty7/10

Dynamics Stability Accuracy Rnn Graph nn Training dynamics Initialization Dynamical systems Bifurcations Control theory Statistical mechanics Graph theory

Source paper: Synchronization, Kinematic Waves and Spike-Phase-Separation in Feedback Ising Neural Networks on Heterogeneous Graphs arXiv:2607.28275 ⓘ · analyzed Aug 31, 2026

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 Not peer-reviewed.

Idea description

Construct a recurrent or graph-recurrent layer with a homeostatic feedback variable and set its recurrent gain using the graph degree-moment ratio $\alpha=\langle k^2\rangle/\langle k\rangle$. The feedback loop is deliberately placed below, near, or above a predicted Hopf boundary, allowing controlled persistent oscillations without unconstrained exploding states.

Formulas

$$\alpha=\frac{\langle k^2\rangle}{\langle k\rangle},\qquad \dot m=-m+\tanh\!\left[\beta(J\alpha m+h)\right],\qquad \dot h=-a h+c(m_0-m).$$

$$A_* = \begin{pmatrix}-1+\beta J\alpha q & \beta q\\-c & -a\end{pmatrix},\qquad q=\operatorname{sech}^2\!\left[\beta(J\alpha m_*+h_*)\right].$$

$$\text{Hopf boundary: }\operatorname{tr}(A_*)=-1-a+\beta J\alpha q=0,\qquad \det(A_*)=a(1-\beta J\alpha q)+\beta c q>0.$$

Mathematical statement

Use the mean-field Glauber-inspired system $\dot m=-m+\tanh[\beta(J\alpha m+h)]$ and $\dot h=-a h+c(m_0-m)$, where $m$ is global hidden-state activity, $h$ is a slow homeostatic field, $J$ is recurrent coupling, $\beta$ is activation sharpness, $a>0$ is feedback relaxation, $c>0$ is feedback strength, and $m_0$ is target activity. At fixed point $(m_*,h_*)$, define $q=\operatorname{sech}^2[\beta(J\alpha m_*+h_*)]$. The Jacobian is $A=[[-1+\beta J\alpha q,\beta q],[-c,-a]]$. A Hopf crossing occurs when $\operatorname{tr}(A)=0$ and $\det(A)>0$, yielding $\beta J\alpha q=1+a$. Thus graph heterogeneity directly shifts the oscillatory stability boundary.

Implementation notes

Integrate this mechanism directly into the recurrent update of a small RNN or message-passing GNN. For a graph, compute node states $x_i$, degrees $k_i$, and global activity $m=N^{-1}\sum_i x_i$. Maintain one scalar feedback state $h$ per layer or sequence. Estimate $\alpha$ once from the graph, or recompute it for dynamic graphs. A concrete Euler implementation is: compute neighbor input $r_i=J\sum_jA_{ij}x_j/k_i+h$; update $x_i\leftarrow x_i+\Delta t[-x_i+\tanh(\beta r_i)]$; then update $h\leftarrow h+\Delta t[-ah+c(m_0-m)]$. Parameterize $a,c,J,\beta$ as positive values. Log $\alpha$, the empirical fixed point, $q$, the Jacobian trace, determinant, and the largest hidden-state spectral growth rate. The graph degree moments and state statistics are computed directly; the fixed point and Jacobian quantities are estimated from long burn-in trajectories or automatic differentiation. First experiment: train this layer and a vanilla tanh RNN on synthetic graph-sequence prediction using graphs with matched mean degree but different $\alpha$. Sweep $J$ and $\beta$. The predicted signature is an oscillation onset when $\beta J\alpha q\approx1+a$, within 20 percent, followed by a narrow activity-spectrum peak; below onset, perturbations should decay exponentially.

Verification

Beats tuned baseline

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Significant win ✓

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 a self-contained Heterogeneity-Calibrated Hopf RNN toy verifier with degree-moment alpha, equilibrium/Jacobian diagnostics, trajectory simulation, and a vanilla graph-RNN comparison. Predicted Hopf boundaries matched numerical trace crossings within 5e-6 relative error; thresholds scaled inversely with alpha, with observed Jc ratio 1.562497 versus predicted 1.562500. Below the boundary perturbations decayed, while above it bounded oscillations emerged. The mechanism manifested clearly, although the untuned tiny prediction task was worse for the Hopf model, so there is no demonstrated practical accuracy win.

Agent confidence: 9/10

Baseline: Vanilla graph recurrence tiny-task MSE=0.2635055
Idea: Hopf graph recurrence tiny-task MSE=0.9331563; regular Jc=0.321402, heterogeneous Jc=0.205698; observed Jc ratio=1.562497 versus inverse-alpha prediction=1.562500; above-boundary amplitude at 1.2Jc≈0.6267

Limitations: Only a two-variable mean-field toy system and a tiny synthetic graph-memory task were tested; no trained multi-layer model, dynamic graphs, FLOP/speed analysis, long-horizon benchmark, or broad parameter/seed sweep was performed. The fixed-gain comparison was illustrative and the Hopf task parameters were not tuned for predictive accuracy.

How to run: python3 hopf_rnn_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.28275) Sep 2, 2026

Beats the tuned baseline by 25.3% (p=0.0081, wins 8 / 8 paired seeds; metric: lower is better).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0016
Idea mean: 0.0012
Effect (Δ): -0.0004 (−25.3%; negative = idea better)
Wins: 8 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±9.1%
Mechanism: Confirmed ✓
Practical effect: Helps
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in dynamics track was tested; tabular, vision, and sequence tracks were not run. The graph was fixed rather than dynamically varying, J/beta/a/c were not broadly swept, and the trained system remained below the predicted Hopf boundary, so persistent oscillation behavior was not validated at neural-network scale.

How to run: python3 hopf_stage2_bench.py

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

Stage 2 — Benchmark Aug 31, 2026 latest Worked ✓

Agent confidence: 8/10

Implemented a matched graph-recurrent dynamics benchmark using the standard bench training loop: vanilla graph-tanh recurrence versus the Euler homeostatic Hopf recurrence. Across 8 paired seeds, the idea achieved MSE 0.0011689134 versus baseline 0.0015645342, a 25.3% relative improvement, with 8/8 paired wins and permutation p=0.0081. The analytical Hopf boundary check passed, and the trained-model signature was measured from benchmark-model states; oscillation onset itself was not demonstrated in the trained model.

Baseline: MSE mean=0.0015645341773, std=0.0002957899302; best config lr=0.01, J=0.16
Idea: MSE mean=0.0011689133535, std=0.0001947871970; best config lr=0.01, J=0.16, beta=1.0; paired delta=-0.0003956208238; 8/8 wins; permutation p=0.0081; relative improvement=25.3%

Benchmark result