✗ Failed on benchmark 2026

Symmetry-Resolved Fourier Bifurcation Monitor

Implementation & benchmark of arXiv:2608.11101 — Local and Global Equivariant Bifurcation for Periodic Weyl and Riesz Fractional Equations

Usefulness7/10

Difficulty6/10

Novelty7/10

Dynamics Stability Accuracy Rnn Ssm Optimizer Training dynamics Dynamical systems Bifurcations Control theory Harmonic analysis Linear algebra

Source paper: Local and Global Equivariant Bifurcation for Periodic Weyl and Riesz Fractional Equations arXiv:2608.11101 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Decompose a periodic recurrent or state-space model into group-symmetry sectors and temporal Fourier modes, then monitor the restricted characteristic spectrum instead of only the full Jacobian. Use the first sector whose characteristic value approaches zero or whose winding number changes to reduce the learning rate, increase damping, or deliberately activate a new dynamical mode.

Formulas

hₜ₊₁ = Fθ(hₜ,xₜ), Mθ = ∂hₜ₊₁/∂h₀ = Πₜ₌₀ᵀ⁻¹ ∂Fθ/∂hₜ

H = ⊕σ ⊕k Hσ,k, χσ,k(μ) = det(μI − Jσ,k)

mσ,k = |1 − λσ,k|, candidate bifurcation when mσ,k ≤ ε

wσ,k = (1/(2πi)) ∮Γ [∂μχσ,k(μ)/χσ,k(μ)] dμ

Mathematical statement

Let G be a finite symmetry group acting on hidden states through a representation ρ(g), and let Mθ be the linearized one-period return map of the neural dynamical system at parameters θ. Decompose the hidden space into isotypical components Vσ indexed by irreducible representations σ of G, and periodic perturbations into temporal Fourier modes k. Let Jσ,k be the restriction of the linearized operator to the (σ,k) block and define the characteristic function χσ,k(μ) = det(μI − Jσ,k). A local bifurcation candidate occurs when χσ,k(1) = 0 for a discrete-time return map. For a complex symbol, a winding number wσ,k = (1/(2πi))∮Γ d log χσ,k(μ) detects robust zero crossings. In the neural adaptation, the blockwise spectral margin mσ,k = |1 − λmax(Jσ,k)| is used as a measurable stability signal; λmax denotes the eigenvalue closest to one in the relevant block.

Implementation notes

1. Integration point: use this monitor in a periodic RNN, linear state-space model, or recurrent memory module with rollout length T. Require a finite group action on hidden channels, such as cyclic channel permutations or a dihedral reflection-plus-rotation action. Precompute the representation projectors Pσ for each symmetry sector. 2. Pseudocode: periodically run a no-gradient rollout; estimate Jacobian-vector products for the one-period map Mθ; project vectors with Pσ; apply a temporal FFT to obtain approximate blocks Jσ,k; use power iteration or Arnoldi to estimate the eigenvalue λσ,k closest to one; compute mσ,k. If the minimum margin falls below ε, multiply the optimizer step by min(1,mmin/m0), or add damping −γPσh only to the critical sector. Optionally evaluate χσ,k on a small complex circle and unwrap its phase to estimate the winding number. 3. Computed from the paper's mechanism: symmetry decomposition, Fourier decomposition, restricted characteristic functions, and winding-style crossing indicators. Estimated empirically: Jacobian blocks, eigenvalues, truncation error, and the practical threshold ε. 4. First cheap experiment: train a cyclic-equivariant GRU on length-64 trajectories from a noisy Kuramoto ring. Compare global spectral clipping with blockwise monitoring over a grid of recurrent gain and learning rate. The prediction is that divergence begins when the first block eigenvalue reaches the unit circle, so the measured divergence boundary should agree with mσ,k = 0 within 20 percent. The critical spatial sector should also predict the Fourier pattern that dominates unstable rollouts.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

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 an exact cyclic-equivariant Fourier-sector monitor with restricted eigenvalue margins, characteristic-circle winding checks, temporal FFT utilities, and a reproducible toy sweep. All four mechanism predictions manifested: the unit-circle boundary was 1.075 versus predicted 1.0753 (0.025% error), the margin slope was -0.930 versus -0.930, and the predicted critical sector and winding transitions were recovered exactly. In the small optimization comparison, blockwise damping reached final loss 0.000181 versus 0.001430 for global damping, although this is a controlled linear toy rather than a trained GRU or Kuramoto task.

Agent confidence: 9/10

Baseline: Global margin damping: final MSE 0.00142994 after 80 steps.
Idea: Symmetry-resolved blockwise damping: final MSE 0.00018145 after 80 steps, approximately 7.9x lower; predicted/observed unit-circle boundary 1.07527/1.075, margin slope -0.93/-0.93, critical sector 3/3, winding inside/outside 1/0.

Limitations: The experiment uses an exactly known linear circulant return map, not a learned GRU, noisy Kuramoto trajectories, or estimated Jacobian-vector products. Temporal Fourier extraction is implemented as a utility but was not used in the optimization comparison, and no GPU or wall-clock benchmark was needed.

How to run: python3 symmetry_monitor.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.11101) Sep 2, 2026

Ahead of the baseline (−3.6%) but not statistically significant (p=0.7865, wins 4 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0097
Idea mean: 0.0093
Effect (Δ): -0.0004 (−3.6%; negative = idea better)
Wins: 4 / 8 paired seeds
p-value: 0.7865 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±30.3%
Mechanism: Not confirmed ✗
Practical effect: Inconclusive
Baseline tuning: swept over 9 configs

Limitations:

The benchmark was reduced to 400 training and 200 test examples and 5 epochs to fit the runtime budget, rather than the README example's 4000/1000 and 30 epochs. The built-in GRU is not cyclic-equivariant, so the hidden-channel Fourier decomposition is only a diagnostic proxy, not an exact symmetry restriction. No learned one-period Jacobian-vector-product or Kuramoto-ring experiment was run.

How to run: python3 bench_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: 9/10

Implemented Fourier-sector recurrent-gradient damping for the registered dynamics/rnn_small benchmark, matched against global gradient clipping on the same GRU and task. The idea slightly lowered mean MSE, but the paired permutation test was non-significant (p=0.78655), and the trained-model mechanism signature had 0/8 sector agreement. Therefore this transfer does not establish a real win.

Baseline: Best global clipping lr=0.01, clip=1.0: mean test MSE 0.00967748 over 8 seeds.
Idea: Best Fourier-sector damping lr=0.01, margin0=0.20: mean test MSE 0.00932742 over 8 seeds; paired delta -0.00035005, p=0.78655.

Benchmark result