✗ Failed on benchmark 2026

Topological Fluctuation Graph Layer

Implementation & benchmark of arXiv:2608.26055 — Topology of Fluctuation Bands in Chiral Active Matter

Usefulness7/10

Difficulty7/10

Novelty8/10

Architecture Stability Accuracy Generalization Graph nn Ssm Regularization Training dynamics Statistical mechanics Stochastic processes Spectral theory Differential geometry Dynamical systems

Source paper: Topology of Fluctuation Bands in Chiral Active Matter arXiv:2608.26055 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Replace a deterministic graph propagation layer by a stable stochastic linearized latent dynamics whose frequency-resolved covariance matrix defines spectral bands. Train or initialize the graph operator so that a selected covariance band has a nonzero Chern number and remains separated by a measurable spectral gap, producing representations that are robust to local perturbations and can support boundary-localized responses.

Formulas

$$\dot z_k(t)=A(k)z_k(t)+\xi_k(t),\qquad \operatorname{Re}\lambda_j(A(k))<0$$

$$S(\omega,k)=(-i\omega I-A(k))^{-1}D(k)(i\omega I-A(k)^\dagger)^{-1}$$

$$C_n(\omega)=\frac{1}{2\pi}\int_{\mathrm{BZ}} i\left(\langle\partial_{k_x}u_n|\partial_{k_y}u_n\rangle-\langle\partial_{k_y}u_n|\partial_{k_x}u_n\rangle\right)d^2k$$

$$g_n(\omega)=\min_k\left[s_n(\omega,k)-s_{n+1}(\omega,k)\right],\qquad g_n=0\ \text{is the predicted transition boundary}$$

Mathematical statement

Let $z_k(t)\in\mathbb{C}^d$ be the Fourier mode of a latent field on a two-dimensional periodic graph, $A(k)$ its learned drift matrix, and $\xi_k(t)$ zero-mean noise with covariance $D(k)$. The stable stochastic dynamics are $\dot z_k=A(k)z_k+\xi_k$, with $\operatorname{Re}\lambda_j(A(k))<0$. At angular frequency $\omega$, the fluctuation spectrum is $S(\omega,k)=G(\omega,k)D(k)G(\omega,k)^\dagger$, where $G(\omega,k)=(-i\omega I-A(k))^{-1}$. Let $u_n(\omega,k)$ be a normalized eigenvector of covariance band $n$, satisfying $S u_n=s_n u_n$. Its Chern number is $C_n=(2\pi)^{-1}\int_{\mathrm{BZ}}\Omega_n(k)\,d^2k$, with Berry curvature $\Omega_n=i(\langle\partial_{k_x}u_n|\partial_{k_y}u_n\rangle-\langle\partial_{k_y}u_n|\partial_{k_x}u_n\rangle)$. The topology can change only when the covariance-band gap $g_n=\min_k|s_n-s_{n+1}|$ closes.

Implementation notes

1. Integration point: implement a latent state-space graph layer on an $L_x\times L_y$ periodic grid or mesh. Store a small channel-space drift operator $A(k)$, obtained by Fourier-transforming a local graph or convolution stencil. Add antisymmetric chiral couplings between channels, while constraining the symmetric part to be negative definite, for example $A(k)=-(B(k)B(k)^\dagger+\epsilon I)+H(k)$, where $H(k)$ is skew-Hermitian. Inject Gaussian noise with covariance $D(k)$, simulate short trajectories, and pass filtered latent states to the predictor. 2. Pseudocode: update $z_{t+\Delta t}=z_t+\Delta t A z_t+\sqrt{\Delta t}L\varepsilon_t$, with $LL^\dagger=D$; estimate $S(\omega,k)$ by averaging FFT periodograms; eigendecompose each $S(\omega,k)$; compute the Fukui lattice Chern number from normalized neighboring eigenvector overlaps; and add $\lambda_g\max(0,g_0-g_n)^2$ to the task loss. Initially use the Chern number as a monitor rather than differentiating through its integer rounding, and optimize a smooth Berry-curvature or gap surrogate. 3. Computed from the mechanism: the resolvent spectrum, covariance eigenbands, band gap, Berry links, and Chern estimate. Estimated empirically: finite-sample spectra, useful frequency windows, and downstream accuracy. 4. First cheap experiment: use a two-channel graph convolution on a $16\times16$ periodic grid and train node classification on synthetic data whose labels depend on long-range wave propagation. Compare a deterministic GNN, a noisy GNN, and this covariance-band layer while sweeping chirality $q$, noise strength $\sigma^2$, and observation frequency $\omega$. The quantitative prediction is that the measured Chern number remains constant while $g_n(\omega)>0$, changes only near a frequency or noise value where $g_n$ approaches zero, and produces stronger boundary localization after opening the torus into a strip. Locate the transition by the minimum covariance gap and require agreement within 20 percent; quantify localization using inverse participation ratio or boundary-to-bulk energy ratio.

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 7/10

Built a readable toy topological fluctuation graph layer in topological_fluctuation.py, including stable chiral drift, resolvent covariance spectra, Fukui Chern estimation, Euler stability checking, covariance gaps, and open-strip localization. The mechanism manifested: the measured Euler threshold was 0.14752 versus the predicted 0.14752, Chern phases were 0,+1,+1,-1,-1,0 across the mass sweep, and chirality increased edge concentration from edge ratio/IPR 0.368/0.088 at q=0 to 1.000/1.000 at q=1. This is a promising toy verification, not evidence of downstream ML accuracy or a trained graph-layer win.

Agent confidence: 7/10

Baseline: Nonchiral q=0 strip: edge ratio 0.368 and IPR 0.088.
Idea: Chiral q=1 strip: edge ratio 1.000 and IPR 1.000; Chern sweep gave 0,+1,+1,-1,-1,0; stability-radius crossing was dt≈0.14752.

Limitations: No neural-network training, synthetic node-classification task, finite-sample stochastic trajectory/periodogram estimation, learned operator optimization, perturbation robustness, or systematic frequency/noise sweep was performed. The strip diagnostic is a small Hamiltonian proxy for boundary localization, and finite momentum grids make exact transition-gap verification sensitive to whether gap-closing momenta are sampled.

How to run: python3 topological_fluctuation.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.26055) Sep 2, 2026

Significantly WORSE than the tuned baseline (+31.2%, p=0.0081, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0019
Idea mean: 0.0025
Effect (Δ): 0.0006 (+31.2%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±17.9%
Mechanism: Not confirmed ✗
Practical effect: Hurts
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in dynamics track was tested; no vision, sequence, tabular, boundary-strip, finite-periodogram, learned Chern-number objective, or robustness sweep was run. The intervention uses a compact recurrent approximation of stable chiral latent dynamics rather than a spatial 2D graph Fourier operator, because the standard dynamics bench has no explicit periodic spatial graph. The mechanism signature measures trained-model perturbation response but does not establish covariance-band Chern topology at neural-network scale.

How to run: python3 stage2_bench.py

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

Stage 2 — Benchmark Sep 2, 2026 latest Failed ✗

Agent confidence: 10/10

Implemented a matched topological fluctuation recurrent system on the structurally appropriate dynamics track, using the canonical pendulum rollout benchmark and an equal-budget baseline learning-rate sweep. Across 8 paired seeds, the idea was significantly worse: test MSE 0.002484 versus baseline 0.001892, delta +0.000591, permutation p=0.0081. The trained-model perturbation signature was also not confirmed, with response ratio 1.0069 rather than the predicted <=1.0.

Baseline: Test MSE mean 0.0018923835, std 0.0005824769; best config lr=0.006, epochs=18, weight_decay=0.
Idea: Test MSE mean 0.0024836512, std 0.0007913987; best config lr=0.006, epochs=18, weight_decay=0., q=0.25, noise=0.02; paired delta +0.0005912678; permutation p=0.0081; 0/8 idea wins.

Benchmark result