△ Mechanism confirmed, baseline not beaten 2026

Bifurcation-Aware Local Basin Regularizer

Implementation & benchmark of arXiv:2607.12189 — Stability and Bifurcations of Planar Switched Linear and Homogeneous Systems

Usefulness7/10

Difficulty5/10

Novelty8/10

Regularization Stability Generalization Rnn Mlp Regularization Training dynamics Dynamical systems Control theory Linear algebra Bifurcations

Source paper: Stability and Bifurcations of Planar Switched Linear and Homogeneous Systems arXiv:2607.12189 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Use the switched nonlinear extension to distinguish stability of the linearized modes from stability of the full neural dynamics. Stabilize worst-case linear products and limit the variation of each nonlinear Jacobian inside a specified radius, yielding an explicit local basin estimate and a penalty that prevents mode interactions from destroying attraction.

Formulas

Fₚ(z) = Mₚz + rₚ(z), Mₚ = DFₚ(0), rₚ(0) = 0, Drₚ(0) = 0

||Mₚₖ ... Mₚ₁|| ≤ C qᵏ, C > 0, 0 < q < 1

Lᴿ = sup₍ₚ, ||z||≤R₎ ||DFₚ(z) − Mₚ||

Cq + Lᴿ < 1 implies local contraction; Cq + Lᴿ = 1 predicts loss of the certified local basin

Mathematical statement

Let a switched neural update be Fₚ(z) = Mₚz + rₚ(z), where Mₚ = DFₚ(0), rₚ(0) = 0, and Drₚ(0) = 0. Assume all admissible linear products satisfy ||Mₚₖ ... Mₚ₁|| ≤ C qᵏ with C > 0 and 0 < q < 1. Define the nonlinear Jacobian variation over radius R as Lᴿ = sup over p and ||z||≤R of ||DFₚ(z) − Mₚ||. A sufficient local contraction condition is Cq + Lᴿ < 1, after expressing the dynamics in a compatible product norm. The predicted local stability boundary is Cq + Lᴿ = 1. The largest tested R satisfying this inequality is an empirical lower bound on the attraction basin radius.

Implementation notes

Attach the regularizer to a residual block or recurrent cell around a reference state, usually z = 0 after centering activations. Compute Mₚ = DFₚ(0) with automatic differentiation. Sample perturbations zⱼ on spheres with radii between zero and R, compute Jacobian-vector products for DFₚ(zⱼ) − Mₚ, and estimate each operator norm using power iteration. Estimate Cq from the worst-case product monitor above, or use a conservative maximum over sampled products. Add penalty λ times the squared positive part of Cq + Lᴿ − (1 − ε), with λ a tunable coefficient and ε = 0.05. Pseudocode: M[p]=jacobian(F[p],0); qC=worst_product_growth(M); L=max_jp operator_norm(jacobian(F[p],z[j])-M[p]); penalty=lambda*relu(qC+L-(1-eps))². The switched contraction argument is taken from the paper; Jacobian variation and the basin radius are estimated empirically. First test a two-mode two-dimensional nonlinear residual RNN on noisy sequence copying, comparing weight decay and per-mode spectral normalization. Increase a cubic nonlinearity coefficient or residual scale. The predicted signature is decay for Cq+Lᴿ below one, critical slowing near one, and growth above one; the measured certified radius should shrink toward zero as the boundary is approached.

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 a two-mode nonlinear switched residual map with exact Jacobian calculations, empirical basin-radius estimation, boundary sweeps, and a matched weight-decay baseline. The mechanism manifested quantitatively: L^R matched 3βR² to numerical precision, radius estimates matched sqrt((1−m)/(3β)) within 0.5%, and measured Jacobian norms crossed the predicted contraction boundary within about 0.05%. In the optimization toy, the basin regularizer slightly reduced the estimated gamma at R=0.45 (0.4934 vs 0.4964) but had slightly worse final loss (0.000255 vs 0.000245), so no task-performance win was observed.

Agent confidence: 9/10

Baseline: Weight decay: loss 0.0002449, gamma_R045 0.49644, m 0.48981, beta 0.01091
Idea: Basin penalty: loss 0.0002549, gamma_R045 0.49337, m 0.47631, beta 0.02809; mechanism checks: L relative error <4e-15, radius relative error <=0.5%, boundary Jacobian norms 0.96091/0.99993/1.04046 below/near/above

Limitations: The experiment uses an analytically tractable 2D cubic map rather than the proposed noisy sequence-copying RNN, does not test multiple random seeds or long-horizon switched products, and does not establish improved generalization or task loss. The empirical operator-norm sampling is replaced by exact Jacobians for this special toy map.

How to run: python3 verify_basin.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.12189) Sep 2, 2026

Ahead of the baseline (−13.1%) but not statistically significant (p=0.1963, wins 6 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.002
Idea mean: 0.0017
Effect (Δ): -0.0003 (−13.1%; negative = idea better)
Wins: 6 / 8 paired seeds
p-value: 0.1963 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±22.1%
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

Only the registered built-in dynamics track was tested. Jacobian variation used two random hidden-state directions during training and four during final signature measurement, with a fixed radius R=0.5 rather than exhaustive operator-norm estimation. The baseline and idea shared architecture, seeds, epochs, batch size, learning-rate grid, and weight-decay grid.

How to run: python3 bench_basin.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 the Bifurcation-Aware Local Basin Regularizer on the registered built-in dynamics track using the matched rnn_small architecture. The idea improved mean test MSE from 0.00197970 to 0.00172016, but the paired permutation p-value was 0.1963, so the improvement was not significant and worked=false. The trained-model mechanism signature confirmed a sub-unit local contraction indicator at radius 0.5, but this did not produce a statistically significant task-metric win.

Baseline: Tuned weight-decay rnn_small, best lr=0.006 and weight_decay=0.001; full 8-seed mean MSE 0.0019796995038632303.
Idea: Basin-regularized rnn_small, best lr=0.006 and weight_decay=0.001; full 8-seed mean MSE 0.0017201633454533294; paired delta -0.00025953615840990096; p=0.1963.

Benchmark result