△ Mechanism confirmed, baseline not beaten 2026

Spectral Basin Allocation for Multimodal Neural Memories

Implementation & benchmark of arXiv:2609.02047 — Phase-delays shape multistability and basin sizes in Kuramoto networks: analytical estimates from network structure

Usefulness7/10

Difficulty7/10

Novelty8/10

Architecture Stability Accuracy Sample efficiency Rnn Ssm Training dynamics Embedding Dynamical systems Spectral theory Linear algebra Bifurcations

Source paper: Phase-delays shape multistability and basin sizes in Kuramoto networks: analytical estimates from network structure arXiv:2609.02047 ⓘ · analyzed Sep 3, 2026

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 Not peer-reviewed.

Idea description

Use several phase-locked states as distinct attractors of one recurrent network and shape their basin asymmetry through the phase-delay composite spectrum. This creates a controllable associative-memory architecture in which a desired memory receives a larger basin without adding a separate classifier or explicit nearest-neighbor lookup.

Formulas

$$C^{(q)}_{ij}=A_{ij}\cos(\psi_j^{(q)}-\psi_i^{(q)}-\alpha_{ij}),\qquad L^{(q)}=\operatorname{diag}(C^{(q)}\mathbf 1)-C^{(q)}$$

$$r_q=K\min_{k\ne\mathrm{gauge}}\operatorname{Re}\lambda_k\!\left(L^{(q)}\right),\qquad \|\delta\theta(t)\|\approx e^{-r_qt}\|\delta\theta(0)\|$$

$$\mathcal L_{\mathrm{basin}}=\sum_{q=1}^{M}[\gamma_q-r_q]_+^2+\lambda_b\sum_{q=1}^{M}(\widehat B_q-b_q)^2$$

Mathematical statement

Let $q\in\{1,\ldots,M\}$ index target phase-locked patterns $\psi^{(q)}$. For each pattern define $C^{(q)}_{ij}=A_{ij}\cos(\psi_j^{(q)}-\psi_i^{(q)}-\alpha_{ij})$ and $L^{(q)}=\operatorname{diag}(C^{(q)}\mathbf 1)-C^{(q)}$. The local contraction rate of attractor $q$ is $r_q=K\min_{k\ne\mathrm{gauge}}\operatorname{Re}\lambda_k(L^{(q)})$, and its slowest local mode has time constant approximately $1/r_q$. The paper's transferable mechanism is that this same composite spectrum organizes basin sizes, so spectral shaping can be used as a basin-control proxy. Define desired margins $\gamma_q$ and basin targets $b_q$, and optimize $\mathcal L=\sum_q[\gamma_q-r_q]_+^2+\lambda_b\sum_q(\widehat B_q-b_q)^2$, where $\widehat B_q$ is the measured fraction of initial states converging to attractor $q$.

Implementation notes

1. Integration point: implement an associative-memory neural ODE or recurrent layer with one shared graph $A$, shared delay matrix $\alpha$, and $M$ prototype phase patterns $\psi^{(q)}$. Decode the attractor index from the final phase state with a cosine-similarity head. 2. Pseudocode: initialize prototypes from class-conditioned latent embeddings; for every prototype compute $C^{(q)}$, $L^{(q)}$, and the smallest non-gauge real eigenvalue $r_q$; simulate short trajectories from noisy versions of each prototype; identify the converged attractor; estimate $\widehat B_q=N_q/N$ from uniformly sampled initial states; optimize task loss plus $\mathcal L_{\mathrm{basin}}$. To enlarge one target basin, increase its requested $\gamma_q$ or target $b_q$, while retaining positive margins for other memories. 3. Computed from the paper's mechanism: phase-delay composite spectra and their relationship to multistability and basin organization. Estimated empirically: global basin fractions, because the exact spectrum-to-volume mapping is network-dependent and should not be assumed universal. 4. First cheap experiment: use an 8-dimensional phase attractor memory with four prototype patterns and noisy-MNIST class embeddings. Compare unconstrained training with spectral basin allocation using 10,000 random initial states per model. The predicted signature is monotonic: increasing $r_q$ should increase the measured basin fraction $\widehat B_q$, and attractors with larger $r_q$ should show perturbation-recovery slopes near $-r_q$ on a semilog plot. Falsify the mechanism if basin fraction fails to correlate positively with $r_q$ across at least 20 independently initialized graphs.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: No measurable effect

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 7/10

Built a reproducible 12-oscillator delayed-Kuramoto associative-memory MVP with composite-Laplacian spectral rates, phase-locked prototypes, direct integration, and empirical basin classification. The local contraction claim was clearly verified: predicted rate 0.5070508 matched the Jacobian rate 0.5070508, and the measured semilog decay slope was -0.50834. Coupling reallocation increased the target q=1 basin from 0.0777 to 0.1677, demonstrating controllable basin asymmetry, but its local contraction rate decreased from 0.3660 to 0.2588, so the stronger universal rate-to-basin claim was not supported.

Agent confidence: 7/10

Baseline: weights [0.5, 0.5]; q=1 rate 0.3660; q=1 basin 0.0777; q=0 basin 0.9223
Idea: weights [0.9, 0.1]; q=1 rate 0.2588; q=1 basin 0.1677; q=0 basin 0.8323; math relative rate error 1.97e-15

Limitations: This MVP uses manually selected coupling weights rather than optimizing the proposed basin-aware objective, synthetic phase-gradient prototypes rather than MNIST embeddings, and no learned neural layer. The 20-graph correlation check is weak because only two attractors had positive rates in the sampled systems; no 10,000-sample or GPU experiment was performed.

How to run: python3 run_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2609.02047) Sep 3, 2026

No significant difference vs. the tuned baseline (0.0001 metric points, p=0.1228, wins 1 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0
Idea mean: 0
Effect (Δ): 0.0001 (n/a; negative = idea better)
Wins: 1 / 8 paired seeds
p-value: 0.1228 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Confirmed ✓
Practical effect: No effect
Baseline tuning: swept over 6 configs

Limitations:

Only the built-in dynamics track was tested; no custom multimodal associative-memory dataset or direct learned multi-attractor memory task was added. The benchmark uses 400 training and 400 test examples and 10 epochs, and the hidden-state phase embedding is a proxy rather than a full learned delayed-Kuramoto memory with explicit basin-volume estimation.

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 Sep 3, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented spectral basin allocation as a training-time composite-Laplacian penalty on phase coordinates of the shared rnn_small GRU, using the structurally matched dynamics track. The mechanism signature was confirmed on trained models: spectral rate increased from 0.37245 to 0.54203 and adjacent phase gap decreased from 0.27065 to 0.09760. However, test MSE worsened from 0.0034843 to 0.0035548, with paired delta +0.0000705 and permutation p=0.1228; therefore the idea did not produce a significant task-metric win.

Baseline: best cfg lr=0.006, weight_decay=0.0001, epochs=10; full 8-seed test MSE mean=0.0034843027, std=0.0013417531
Idea: best cfg lr=0.006, weight_decay=0.0, lam=0.02, epochs=10; full 8-seed test MSE mean=0.0035548201, std=0.0012805769
Mechanism signature: trained-model baseline rate=0.3724495, phase_gap=0.2706524; idea rate=0.5420256, phase_gap=0.0975972; confirmed=true
Paired delta: idea-baseline = +0.0000705174 MSE; idea wins 1/8 seeds; permutation p=0.1228

Benchmark result