✗ Failed on benchmark 2026

Border-Collision Multi-Attractor Memory

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

Usefulness7/10

Difficulty7/10

Novelty7/10

Architecture Memory Accuracy Sample efficiency Rnn Ssm Memory World model Dynamical systems Bifurcations Linear algebra Control theory

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 Not peer-reviewed.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

Use the paper's stable periodic orbits and border-collision transitions as an intentional memory mechanism in a recurrent module. Different input-dependent parameter settings can place the same cell in fixed-point, period-2, or higher-period regimes, allowing a compact state to encode discrete modes without allocating one separate neural attractor per mode.

Formulas

$$T(x,y)=\bigl(|x|+ay-1,\ x+c|y|-1\bigr),\qquad T^p(z)=z,$$

$$M_p=\prod_{i=p-1}^{0}J_{q_i},\qquad J_{q_i}=\begin{bmatrix}s_{x,i}&a\\1&cs_{y,i}\end{bmatrix},\qquad \rho(M_p)<1\ \text{for a stable period-}p\text{ orbit},$$

$$\text{Border collision: }x_i=0\ \text{or}\ y_i=0\ \text{for some }i;\qquad \text{Flip: }1+\operatorname{tr}(J)+\det(J)=0,$$

$$\mathcal{L}_{\mathrm{orbit}}=\|T^p(z_0)-z_0\|^2+\beta\max\bigl(0,\rho(M_p)-1+\epsilon\bigr)^2.$$

Mathematical statement

The cell uses the piecewise-linear map $T(x,y)=(|x|+ay-1,x+c|y|-1)$. A period-$p$ attractor is a point $z$ satisfying $T^p(z)=z$ with minimal period $p$, whose product Jacobian $M_p=J_{q_{p-1}}\cdots J_{q_0}$ has spectral radius $\rho(M_p)<1$. Each $q_i$ is the quadrant visited at step $i$. A Border Collision occurs when one orbit point satisfies $x_i=0$ or $y_i=0$, changing one factor of $M_p$. A Fold Border Collision is detected when stable and saddle period-$p$ orbits coalesce at a border. A Flip transition from a fixed point to period 2 occurs when a fixed-point multiplier crosses $-1$, namely $1+\operatorname{tr}(J)+\det(J)=0$, provided the other multiplier is not on the unit circle.

Implementation notes

1. Integration point: create a bank of two-dimensional recurrent cells in an RNN, world model, or compact sequence classifier. Let an input-dependent hypernetwork select $(a_t,c_t)$ through bounded functions such as $a_t=a_{\max}\tanh(w_a^\top u_t)$ and $c_t=c_{\max}\tanh(w_c^\top u_t)$. The hidden state can then move among fixed-point and periodic attractor regimes. Decode the attractor mode with a linear head. 2. Pseudocode: during a warm-up phase, iterate z=cell(z,u) for $K$ steps; for candidate periods $p\in\{1,2,3,4\}$, compute residual=norm(T^p(z)-z)^2, record the quadrant sequence, form $M_p$ by multiplying the exact quadrant Jacobians, and add the orbit loss above when a target mode is requested. For classification, combine cross-entropy with a weak separation loss between attractor representatives. 3. Computed from the paper: the four-partition map, quadrant Jacobians, periodic-orbit stability criterion, and interpretation of border collisions as transitions between distinct cycles. Estimated empirically: period, basin membership, border distance, and $\rho(M_p)$ by multistart orbit iteration and finite differences. Do not assume closed-form period-3 or period-4 coordinates; locate them by multistart solves of $T^p(z)-z=0$. 4. First cheap experiment: train on a synthetic sequence task in which a mode token selects period 1, 2, 3, or 4 and the model must recall that mode after 100--1000 blank steps. Compare against same-parameter-count tanh-RNN and GRU. Sweep hypernetwork gain so trajectories cross $x=0$ or $y=0$. The predicted signature is an abrupt change in observed period and basin assignment at border crossings, stable modes satisfying $\rho(M_p)<1$, and a sharp recall-error increase when the measured multiplier crosses $\rho(M_p)=1$.

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 8/10

Built and ran a numerical MVP of the border-collision map with exact quadrant Jacobians, analytic fixed/2-cycle candidates, stability checks, border sweeps, and a small recall demo. The predicted flip boundary a=0 was observed exactly, fixed-point stability classification matched the analytic region on 98.9% of 832 grid points, and the predicted period-2 border c=(2-3a)/4 was observed exactly with the symbolic cycle changing from 34 to 24. The memory demo showed a promising but incomplete signal: 60% mode recall versus 25% for a shared contracting tanh baseline, but the selected parameters produced only periods 1 and 2 rather than four clean modes.

Agent confidence: 8/10

Baseline: Shared 2D stable tanh recurrence: 25% mode recall after 150 blank steps; all four states contracted to the same decoded prototype.
Idea: Border-collision map: 60% nearest-prototype mode recall; fixed-point stability sweep accuracy 98.9% (832 valid grid points); flip observed at a=0 with multiplier -1; period-2 border predicted and observed at c=-0.175 for a=0.9, with 34 below and 24 above the border; stable-point perturbation log slope -0.156 versus predicted log rho(J)=-0.131, though norm ratios oscillated.

Limitations: No trained hypernetwork RNN or GRU comparison was implemented; the memory demo used hand-selected parameter regimes and nearest-prototype decoding. It did not produce stable period-3/period-4 attractors, did not test long sequence datasets or FLOPs, and the contraction sweep crossed piecewise dynamics enough to make individual norm ratios non-monotonic.

How to run: python3 experiment.py

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

No significant difference vs. the tuned baseline (+58.2%, p=0.2034, wins 2 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: matched_tanh_vs_border_cells
Paired seeds: 8
Baseline mean: 0.0063
Idea mean: 0.0099
Effect (Δ): 0.0037 (+58.2%; negative = idea better)
Wins: 2 / 8 paired seeds
p-value: 0.2034 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±101.2%
Mechanism: Not confirmed ✗
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in dynamics pendulum rollout was tested; no custom long-delay recall track, period-3/period-4 attractor training, or larger hyperparameter sweep was run. The mechanism diagnostic used four-step products from trained final hidden states and was exploratory rather than a full basin/attractor census.

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 end-to-end dynamics benchmark: an exact input-conditioned border-collision recurrent system versus a same-width tanh RNN, using the canonical bench training path, shared learning-rate sweep, and 8 paired seeds. The idea was worse (MSE 0.00994385 vs 0.00628474; paired delta +0.00365911, permutation p=0.20345), so it did not produce a significant win. The trained-model mechanism signature was not confirmed: mean product spectral radius was 1.5247 and only 4.69% of measured states had rho<1.

Baseline: Test MSE mean 0.0062847389781381935, std 0.003016475610481406, best lr=0.01
Idea: Test MSE mean 0.009943852724973112, std 0.009340805162427421, best lr=0.01

Benchmark result