△ Mechanism confirmed, baseline not beaten 2026

Grazing-Bifurcation Monitor for Recurrent Dynamics

Implementation & benchmark of arXiv:2607.27464 — Neutral Entry--Exit Cycles with Quadratic Grazing: Uniform Return Reduction and Local Two-Parameter Bifurcations

Usefulness8/10

Difficulty5/10

Novelty7/10

Dynamics Stability Accuracy Generalization Rnn Ssm Optimizer Training dynamics Regularization Dynamical systems Bifurcations Control theory Numerical analysis

Source paper: Neutral Entry--Exit Cycles with Quadratic Grazing: Uniform Return Reduction and Local Two-Parameter Bifurcations arXiv:2607.27464 ⓘ · analyzed Aug 31, 2026

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 Not peer-reviewed.

Idea description

Augment a recurrent or implicit neural layer with a local bifurcation monitor that estimates the scalar return-map coefficients A, B, c, and d near a latent fixed point. Penalize trajectories approaching the predicted fold or grazing curves, or deliberately target selected chambers when multistability is useful. The method converts local Jacobian and finite-difference measurements into a falsifiable prediction of when latent fixed points appear, disappear, or change stability.

Formulas

$$\Delta(q,p)=F_p(q)-q=S(q,p)+q_+^{3/2}K(\sqrt{q_+},q,p),\qquad q_+=\max(q,0).$$

$$\Delta(q,p)\approx A+Bq+cq^2+dq_+^{3/2},\qquad A=\Delta(0,p),\quad B=\partial_q\Delta(0,p),\quad c=\tfrac12\partial_{qq}S(0,p),\quad d=K(0,0,0).$$

$$q_*^-=-\frac{B}{2c},\qquad A_*^-(B)=\frac{B^2}{4c};\qquad q_*^+\approx\frac{4B^2}{9d^2},\qquad A_*^+(B)\approx\frac{4\,\operatorname{sgn}(d)|B|^3}{27d^2}.$$

$$\text{stable fixed point}\iff |F_p'(q_*)|=|1+\partial_q\Delta(q_*,p)|<1.$$

Mathematical statement

Let F_p be a recurrent return map and let q be a signed scalar latent coordinate relative to a switching surface, with q<0 nonpenetrating and q>0 penetrating. Define the displacement Delta(q,p)=F_p(q)-q. The paper's local normal form is Delta(q,p)=S(q,p)+q_+^{3/2}K(sqrt(q_+),q,p), where q_+=max(q,0), S is smooth, and K is regular. At the neutral organizer, S(0,0)=S_q(0,0)=0, c=S_qq(0,0)/2 is nonzero, and d=K(0,0,0) is nonzero. Use A=Delta(0,p) and B=Delta_q(0,p) as local parameter coordinates, giving the approximation Delta(q,p) approximately A+Bq+cq^2+dq_+^{3/2}. For q<0, the fold satisfies q_*^-=-B/(2c) and A_*^-(B)=B^2/(4c). For q>0, neglecting higher-order terms, the penetrating fold satisfies q_*^+ approximately 4B^2/(9d^2), requiring B/d<0, and A_*^+(B) approximately 4 sign(d)|B|^3/(27d^2). A fixed point is stable when |F_p'(q_*)|=|1+Delta_q(q_*,p)|<1. The measurable signatures are quadratic and cubic discriminant curves, a change in fixed-point count by two across each fold, and alternating attracting and repelling fixed points when the local derivative is monotone.

Implementation notes

First, integrate the method into a recurrent, state-space, or implicit neural layer. For a fixed input x0 and candidate fixed state h*, choose a unit direction v normal to a selected switching, saturation, or gating surface. Probe the one-dimensional section h=h*+qv and define F_theta(q)=v^T(G_theta(h*+qv,x0)-h*)+q, so Delta_theta(q)=F_theta(q)-q. Second, estimate A=Delta(0), B=[Delta(delta)-Delta(-delta)]/(2 delta), and c by fitting Delta(q)-A-Bq to c q^2 on negative q values. Estimate d from the positive-side residual [Delta(q)-A-Bq-cq^2]/q^(3/2). Compute the predicted nonpenetrating and penetrating fold curves, then add a penalty based on the squared soft distance from the current (A,B) to either curve. Also estimate the local multiplier by autodiff or finite differences and penalize max(0,|1+Delta_q|-rho)^2 for a target rho<1. Third, the normal-form relations and exponents are taken from the paper; coefficients, fixed points, residual quality, and multipliers are estimated from network probes. Disable the penalty when the fitted normal-form residual is too large. The first cheap experiment should use a one-dimensional hybrid tanh/ReLU RNN with two parameters controlling recurrent gain and bias. Sweep this parameter plane, solve h=G_theta(h,0), and compare empirical fold locations with predictions. The quantitative prediction is that the nonpenetrating fold obeys A proportional to B^2 and the penetrating fold obeys A proportional to |B|^3; log-log fits should yield slopes near 2 and 3 over a usable scaling range. Crossing each fold should change the number of fixed points by two, while stable and unstable multipliers should alternate. A second test can train the RNN on sequential MNIST or a synthetic long-horizon copy task and measure whether the monitored model reduces the fraction of updates with local multiplier magnitude above one and improves long-horizon error.

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 8/10

Built a self-contained grazing-bifurcation monitor prototype and numerical verification in experiment.py. The mechanism manifested strongly: observed log-log fold slopes were 2.000 for the nonpenetrating quadratic curve and 3.000 for the penetrating cubic curve, the fixed-point count changed by two across the fold, coefficient recovery was effectively exact, and stable/unstable multipliers alternated. The toy monitored update did not improve the secondary stability metric: its unstable fraction was 0.16495 versus 0.12465 for baseline, so this is evidence for the bifurcation mathematics, not a demonstrated training win.

Agent confidence: 8/10

Baseline: Toy recurrent comparison: mean local unstable-update fraction 0.12465 over 100 trials x 400 steps.
Idea: Mean local unstable-update fraction 0.16495; math checks: negative-fold slope 2.000 vs predicted 2, positive-fold slope 3.000 vs predicted 3, root-count change 2 vs predicted 2, coefficient absolute errors approximately [3.5e-17, 1.3e-15, 6.7e-16].

Limitations: The recurrent comparison is a small scalar synthetic dynamical system, not a trained RNN task or sequential benchmark. The monitor action is a simple gain reduction rather than a differentiable neural-layer penalty, GPU behavior and wall-clock/FLOP improvements were not tested, and the exact normal-form coefficients were used rather than estimated from a learned high-dimensional network.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.27464) Sep 2, 2026

Ahead of the baseline (−21.4%) but not statistically significant (p=0.4134, wins 4 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0132
Idea mean: 0.0104
Effect (Δ): -0.0028 (−21.4%; negative = idea better)
Wins: 4 / 8 paired seeds
p-value: 0.4134 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±47.4%
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

The run used a reduced smoke-scale workload of 400 training and 200 test samples for 5 epochs because the initial 15-epoch protocol exceeded the 40-minute budget. The fold-exponent check is analytical/toy, and the reported multiplier prediction is not an independent validation; no larger-scale or longer-training benchmark was completed.

How to run: python3 bench_experiment.py

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

Stage 2 — Benchmark Aug 31, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented a matched grazing-bifurcation monitor training loop for dynamics/rnn_small with paired seeds, baseline learning-rate sweep, and equal-sized idea sweep. The idea reduced mean test MSE from 0.01318146 to 0.01036343, but the paired permutation test was not significant (delta=-0.00281803, p=0.4134), so the required significant benchmark win was not observed. The toy fold-exponent check passed, but the trained multiplier signature is not independent because predicted and observed multipliers use the same finite-difference quantity.

Baseline: dynamics/rnn_small, best lr=0.003: mean test MSE 0.01318146 +/- 0.00641119 over 8 seeds
Idea: dynamics/rnn_small with monitor loss, best lr=0.006: mean test MSE 0.01036343 +/- 0.00647986 over 8 seeds; paired delta=-0.00281803, p=0.4134, 4/8 wins

Benchmark result