✗ Failed on benchmark 2026

Bifurcation-Certified Piecewise-Linear Recurrent Cell

Implementation & benchmark of arXiv:2607.13519 — Noninvertibility and Bifurcation Phenomena in a Four-Partitions Piecewise Linear Map

Usefulness8/10

Difficulty5/10

Novelty6/10

Dynamics Stability Accuracy Scalability Rnn Ssm Training dynamics Regularization Dynamical systems Bifurcations Control theory Linear algebra

Source paper: Noninvertibility and Bifurcation Phenomena in a Four-Partitions Piecewise Linear Map arXiv:2607.13519 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Replace a standard recurrent update with a two-state absolute-value cell whose local dynamics are exactly piecewise affine. Train the coupling parameters while enforcing discrete-time Schur inequalities inside each activation quadrant, preventing exploding recurrent trajectories while retaining nonsmooth gating and richer dynamics than a globally contractive linear cell.

Formulas

$$z_{t+1}=T(z_t,u_t)=\begin{bmatrix}|x_t|+a y_t-1\\x_t+c|y_t|-1\end{bmatrix}+B u_t,$$

$$J_{s_x,s_y}=\begin{bmatrix}s_x&a\\1&cs_y\end{bmatrix},\qquad \tau=s_x+cs_y,\qquad \delta=s_xcs_y-a,$$

$$1-\tau+\delta>0,\qquad 1+\tau+\delta>0,\qquad 1-\delta>0,$$

$$\text{Flip boundary: }1+\tau+\delta=0\iff (1+s_x)(1+cs_y)-a=0.$$

Mathematical statement

The recurrent state is $z_t=(x_t,y_t)^\top$, and the autonomous core is $T(x,y)=(|x|+ay-1,\ x+c|y|-1)$. In a quadrant with $s_x=\operatorname{sign}(x)\in\{-1,+1\}$ and $s_y=\operatorname{sign}(y)\in\{-1,+1\}$, the Jacobian is $J_{s_x,s_y}=\begin{bmatrix}s_x&a\\1&cs_y\end{bmatrix}$. Its trace and determinant are $\tau=s_x+cs_y$ and $\delta=s_xcs_y-a$. Both eigenvalues lie strictly inside the unit disk if and only if the Jury conditions $1-\tau+\delta>0$, $1+\tau+\delta>0$, and $1-\delta>0$ hold. The Flip bifurcation surface, where an eigenvalue reaches $-1$, is $1+\tau+\delta=0$. A Border Collision occurs when a fixed point reaches $x=0$ or $y=0$, changing the active Jacobian.

Implementation notes

1. Integration point: use this cell as the recurrent transition in a small RNN or state-space model, with $z_t\in\mathbb{R}^{2d}$ represented as $d$ independent two-dimensional cells. Add input injection $Bu_t$ and an output projection $h_t=W_o z_t$. Generalize the fixed offset $-1$ to a learned bias, but keep the absolute values and shared couplings $a,c$ initially. 2. Pseudocode: for each batch and channel, compute sx=sign(x.detach()), sy=sign(y.detach()), and z_next=[abs(x)+a*y-1, x+c*abs(y)-1]+B*u. Compute tau=sx+c*sy and delta=sx*c*sy-a; add lambda * sum(relu(eps-(1-tau+delta)) + relu(eps-(1+tau+delta)) + relu(eps-(1-delta))) to the task loss. Optionally project $a,c$ after each optimizer step into a precomputed safe region. 3. Computed from the paper: the exact map, quadrant Jacobians, Jury inequalities, and Flip surface. Estimated empirically: active-quadrant frequencies, spectral radius of every observed Jacobian, and distance from the borders $x=0$ and $y=0$. Use a smooth absolute value only if gradients at zero are problematic, but evaluate stability with the exact nonsmooth map. 4. First cheap experiment: compare this cell against tanh-RNN and vanilla linear RNN on sequential MNIST and copy-memory sequences with equal hidden-state dimension. Sweep $a$ across the predicted Flip threshold while fixing $c$ and each quadrant sign, and measure hidden-state norm and Jacobian spectral radius. The quantitative prediction is that divergence or sustained period-2 oscillation begins near $1+\tau+\delta=0$, with the measured threshold within 10--20% of the analytic value; parameters satisfying all three Jury inequalities should show geometric perturbation decay away from borders.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: Significantly worse ✗

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 the certified piecewise-linear recurrent cell with quadrant Jacobians, Jury penalty, CPU/CUDA fallback, mathematical sweeps, and a fixed-seed toy sequence task. The mechanism manifested: Jury classification had 0/20,000 mismatches, the predicted stable interval for the (--), c=0.5 quadrant was observed as a=-0.45 through -0.05, and spectral-radius crossings occurred at the predicted boundaries a=-0.5 and a=0. Stable cases decayed while unstable cases grew strongly; the certified model also outperformed the tanh-RNN on this toy task, although this secondary result is not evidence of general superiority.

Agent confidence: 9/10

Baseline: tanh-RNN: mean final BCE 0.1698 and mean accuracy 94.6% across 3 seeds.
Idea: Certified PWL cell: mean final BCE 0.0651 and mean accuracy 99.7%; Jury test mismatches 0/20,000; perturbation norm ratios at t=15 were 0.554 and 0.123 for stable a=-0.4 and -0.2, versus 34.2 and 169.4 for unstable a=0.2 and 0.4.

Limitations: Only a small synthetic majority-bit sequence task was tested, not sequential MNIST or copy-memory. The perturbation sweep analyzes fixed active quadrants; full nonlinear border-collision statistics, FLOPs/speed, long-horizon robustness, and larger-scale generalization were not tested.

How to run: python3 experiment.py > results.json

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.13519) Aug 30, 2026

Significantly WORSE than the tuned baseline (+3823.8%, p=0.0081, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0001
Idea mean: 0.0025
Effect (Δ): 0.0024 (+3823.8%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±627.6%
Mechanism: Not confirmed ✗
Practical effect: Hurts
Baseline tuning: swept over 3 configs

Limitations:

Only the registered dynamics track was tested. The experiment used a 1000/300 train/test subset and 20 epochs; longer-horizon robustness, speed/FLOPs, additional Jury penalty weights, per-channel couplings, and sequential-MNIST or copy-memory tasks were not tested.

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 30, 2026 latest Failed ✗

Agent confidence: 10/10

The certified PWL recurrent cell was tested on the registered built-in dynamics track using bench.make_report(), a tuned three-learning-rate baseline sweep, and 8 paired seeds. The baseline achieved mean test MSE 0.0000627291 versus 0.00246135 for the idea; paired delta was +0.00239862 with permutation p=0.0081, yielding verdict idea worse (significant). The trained-model mechanism signature was not confirmed.

Baseline: mean MSE 6.27290992269991e-05, best lr 0.006, 8 seeds
Delta mean: 0.0023986215110198827
Idea: mean MSE 0.002461350610246882, best lr 0.006, 8 seeds
P value: 0.0081

Benchmark result