△ Mechanism confirmed, baseline not beaten 2026

Contractive Floquet return map

Implementation & benchmark of arXiv:2608.10738 — Long-Time Trajectory Approximation via SA-NODEs: Model Predictive and Floquet Strategies

Usefulness7/10

Difficulty6/10

Novelty7/10

Regularization Stability Accuracy Sample efficiency Ssm World model Regularization Training dynamics Dynamical systems Control theory Geometry Numerical analysis

Source paper: Long-Time Trajectory Approximation via SA-NODEs: Model Predictive and Floquet Strategies arXiv:2608.10738 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

For systems with a repeating orbit, train a periodic neural dynamical model together with a return map whose transverse deviations contract after each period. Enforce and measure orbital contraction rather than requiring phase-aligned pointwise trajectories to remain close, allowing phase drift while suppressing divergence across many cycles.

Formulas

$$P(x)=\widehat{\Phi}_{\Theta}(\hat T;x),\qquad P(x^\star)=x^\star,$$

$$\|P(x)-P(y)\|\le q\|x-y\|,\qquad 0\le q<1,$$

$$e_{n+1}\le q e_n+\delta\quad\Longrightarrow\quad e_n\le q^n e_0+\delta\frac{1-q^n}{1-q},$$

$$d_{\mathrm{orb}}(x_n,\Gamma_\Theta)=\min_{\phi\in[0,2\pi)}\|x_n-\Gamma_\Theta(\phi)\|_2.$$

Mathematical statement

The transferable object is the stroboscopic return map $P$, defined by advancing a state by one period $\hat T$: $P(x)=\widehat\Phi_{\Theta}(\hat T;x)$. A contraction certificate on a neighborhood $U$ is $\|P(x)-P(y)\|\le q\|x-y\|$ for all $x,y\in U$, with $0\le q<1$. If $x^\star$ is a fixed point satisfying $P(x^\star)=x^\star$, then deviations contract as $\|P^n(x)-x^\star\|\le q^n\|x-x^\star\|$. With one-period defect $\delta$, the perturbed recursion is $e_{n+1}\le qe_n+\delta$, giving $e_n\le q^ne_0+\delta(1-q^n)/(1-q)$. For periodic oscillators, phase error should be removed by measuring distance to the learned orbit $\Gamma_\Theta$: $d_{orb}(x,\Gamma_\Theta)=\min_{\phi}\|x-\Gamma_\Theta(\phi)\|_2$. The paper's obstruction motivates this orbital rather than pointwise certificate.

Implementation notes

Use this in a periodic latent ODE or recurrent world model for locomotion, seasonal signals, rotating systems, or periodic control tasks. Parameterize the vector field with explicit period $\hat T$, feeding $(\sin(2\pi t/\hat T),\cos(2\pi t/\hat T))$ to the network. Learn a nominal periodic trajectory $\Gamma_\Theta(\phi)$ and define the one-period map $P_\Theta(x)=\operatorname{ODESolve}(f_\Theta,x,[0,\hat T])$. Add $L_{contract}=\max(0,\|P_\Theta(x)-P_\Theta(y)\|_2-q_{target}\|x-y\|_2)^2$ for nearby pairs sampled around the orbit, with $q_{target}<1$. Estimate the empirical contraction factor using $\hat q=\max_b\|P(x_b)-P(y_b)\|_2/\|x_b-y_b\|_2$, or estimate the largest local Jacobian singular value by power iteration through $J_P(x)$. Estimate the defect after phase alignment as $\hat\delta=\operatorname{Quantile}_q d_{orb}(P(x_b),\Gamma_\Theta)$. During inference, monitor $\hat q^ne_0+\hat\delta(1-\hat q^n)/(1-\hat q)$. The paper supplies the contraction-plus-defect recursion; the batch estimates, orbit matching, and penalty are engineering approximations. First experiment: train on noisy pendulum or Van der Pol cycles and compare an unconstrained periodic neural ODE against this model over 100--1000 periods. Evaluate phase-invariant orbital RMSE, pointwise RMSE, estimated $\hat q$, and the fraction of rollouts remaining inside the learned tube. Success means bounded orbital error with measured $\hat q<1$, even when pointwise phase error grows.

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 reproducible toy contractive Floquet return-map experiment in floquet_toy.py. All three mechanism predictions were confirmed: the q=1 stability transition and exponential rates matched exactly, the defect floor matched delta/(1-q) with relative error below 5e-16, and transverse orbital error contracted despite substantial phase-sensitive pointwise error. In the secondary noisy 1000-cycle comparison, the contractive map remained bounded while the q=1.02 baseline diverged, though this is not evidence from a learned neural ODE.

Agent confidence: 9/10

Baseline: q=1.02: median terminal orbital error 4.024e7, p95 5.286e7, fraction inside radius-0.05 tube 0.0, empirical qhat 1.02
Idea: q=0.80: median terminal orbital error 0.00428, p95 0.01293, fraction inside radius-0.05 tube 1.0, empirical qhat 0.80; q=0.85 phase-shear setting gave orbital error 3.39e-7 at 80 cycles versus pointwise error 1.747

Limitations: The experiment uses an analytically specified affine return map rather than training a neural ODE or recurrent world model on noisy oscillator data. It does not test optimization behavior, learned orbit quality, computational cost, multidimensional nonlinear Jacobians, or robustness beyond the sampled scalar transverse dynamics.

How to run: python3 floquet_toy.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.10738) Sep 2, 2026

Ahead of the baseline (−2.2%) but not statistically significant (p=0.328, wins 6 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0004
Idea mean: 0.0004
Effect (Δ): -0 (−2.2%; negative = idea better)
Wins: 6 / 8 paired seeds
p-value: 0.328 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

The built-in dynamics task predicts scalar future angle rather than a full one-period state return map, so contraction was approximated through nearby-window output sensitivity. Long-horizon rollout stability, phase-aligned orbital error, neural ODEs, and larger models were not tested.

How to run: python3 bench_floquet.py

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

Stage 2 — Benchmark Sep 1, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented contractive Floquet return-map regularization on the matched dynamics/rnn_small task using a nearby-perturbation contraction loss. The recursion sanity check passed, but the trained benchmark systems showed no significant task-metric win: paired delta was -8.7674e-06 with permutation p=0.328.

Baseline: mean test MSE 0.0003913901, best lr 0.006
Idea: mean test MSE 0.0003826227, best lr 0.006; 6/8 paired wins; delta -8.7674e-06; p 0.328

Benchmark result