# Periodic-Orbit Entropy Calibration

- ID: 1699
- Canonical URL: https://synthcore.org/idea/1699/periodic-orbit-entropy-calibration
- API JSON: https://synthcore.org/api/idea/1699.json
- API Markdown: https://synthcore.org/api/idea/1699.md
- Verification status: unverified
- Source: [arXiv:2608.02186](https://arxiv.org/abs/2608.02186)
- Category: regularization
- Solves: accuracy, generalization, stability
- ML areas: ssm, rnn, world-model, training-dynamics
- Math tags: dynamical-systems, statistics, statistical-mechanics
- Ratings: usefulness 6/10; difficulty 8/10; novelty 9/10

## Idea description

Use the paper's Margulis-type law as a structural constraint for neural continuous-time dynamics: the number of isolated periodic latent trajectories with period at most T should grow like exp(hT)/T in a positive-entropy regime. This provides a falsifiable test for orbit collapse, excessive chaos, or spurious recurrence in neural ODE world models, rather than relying only on one-step prediction loss.

## Mathematical statement

Let Pi(T,Delta) denote periodic orbits with periods in [T,T+Delta], Pi(T) denote cumulative periodic orbits up to T, and h>0 denote topological entropy on the relevant hyperbolic set. The paper establishes two-sided bounds C^{-1}exp(hT)/T <= #Pi(T,Delta) <= C exp(hT)/T and analogous cumulative bounds, with C>=1. For a learned flow, estimate periodic orbits by solving phi_T^theta(z)-z=0 from many initial seeds, remove phase-shift duplicates, and fit log N_theta(T)=h_theta T-log(T)+b_theta+error. The transferable signature is linearity of log[T N_theta(T)] in T with slope h_theta.

## Key formulas

- C^{-1}exp(hT)/T <= #Pi(T,Delta) <= C exp(hT)/T, and C^{-1}exp(hT)/T <= #Pi(T) <= C exp(hT)/T.
- log N_theta(T)=h_theta T-log(T)+b_theta+epsilon_T.
- L_orbit=sum_T w_T[log(T N_hat_theta(T))-h_hat T-b_hat]^2.
- h_hat=argmin_h min_b sum_T w_T[log(T N_hat_theta(T))-hT-b]^2.

## Implementation notes

1. Integration point: initially use this mechanism as a validation metric for neural ODEs, latent Hamiltonian-like flows, or recurrent world models with recurrent data. Only use it as a training regularizer when the observed system has a positive-entropy recurrence target; otherwise it can force artificial chaos. 2. Pseudocode: choose periods T on a logarithmic grid; initialize many latent states z0; solve the shooting residual r(z0,T)=phi_T^theta(z0)-z0; optimize each seed with Newton or L-BFGS; retain roots with norm(r)<tau; remove duplicates using trajectory Hausdorff distance and phase-shift matching; count the remaining solutions N_hat(T); fit h_hat and b_hat by regression; add lambda L_orbit to the objective or use it for checkpoint selection. 3. The exp(hT)/T law and -log(T) correction are taken from the paper. Root-finding, tolerance tau, clustering, finite-horizon counts, and entropy estimates are empirical, so report seed coverage and bootstrap confidence intervals. 4. First cheap experiment: train small neural ODEs on Lorenz-63 or a synthetic multi-cycle system, comparing one-step-loss training with orbit-calibrated training. Plot log N_hat(T) versus T and log[T N_hat(T)] versus T. The quantitative prediction is that the latter plot has an approximately linear slope h_hat and lower residual than a pure exp(hT) fit; orbit-collapse models should show subexponential counts, whereas over-chaotic models should show excess counts or failure of the -log(T) correction.

## Disclaimer

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