# Parabolic Torus Recurrent Core

- ID: 1539
- Canonical URL: https://synthcore.org/idea/1539/parabolic-torus-recurrent-core
- API JSON: https://synthcore.org/api/idea/1539.json
- API Markdown: https://synthcore.org/api/idea/1539.md
- Verification status: unverified
- Source: [arXiv:2607.28472](https://arxiv.org/abs/2607.28472)
- Category: dynamics
- Solves: stability, accuracy
- ML areas: rnn, ssm, world-model, training-dynamics
- Math tags: dynamical-systems, control-theory, geometry, bifurcations
- Ratings: usefulness 6/10; difficulty 5/10; novelty 8/10

## Idea description

Construct a recurrent state-space model with a neutral quasiperiodic phase variable and transverse amplitude variables whose non-autonomous coupling decays polynomially in inference time. The phase subsystem provides persistent torus-like memory, while the transverse subsystem receives only a vanishing perturbation, limiting long-horizon drift caused by continual corrections.

## Mathematical statement

Let \(\theta\in\mathbb{T}^{d}\) be the phase, \(z\in\mathbb{R}^{r}\) the transverse state, and \(t\geq1\) inference time. The paper studies a time-dependent Hamiltonian vector field \(X_{H_0+\varepsilon t^{-\alpha}H_1}\) near an isotropic invariant torus. Its invariance equation is \(\partial_tK(\varphi,t)+D_\varphi K(\varphi,t)\omega=X_H(K(\varphi,t),t)\), where \(K\) embeds the torus, \(\varphi\) is an internal angle, and \(\omega\) is the quasiperiodic frequency. The neural analogue is \(\dot\theta=\omega+\varepsilon t^{-\alpha}f_\theta(\theta,z,u_t)\), \(\dot z=A_0z+\varepsilon t^{-\alpha}f_z(\theta,z,u_t)\), with \(A_0\) parabolic or weakly damped. For bounded perturbations, cumulative forcing is bounded by \(C\int_1^Tt^{-\alpha}dt\), giving bounded drift for \(\alpha>1\), logarithmic drift for \(\alpha=1\), and power-law drift for \(0<\alpha<1\).

## Key formulas

- $$\partial_t K(\varphi,t)+D_\varphi K(\varphi,t)\,\omega=X_H(K(\varphi,t),t),\qquad H(\cdot,t)=H_0(\cdot)+\varepsilon t^{-\alpha}H_1(\cdot,t).$$
- $$\dot{\theta}_t=\omega+\varepsilon t^{-\alpha}f_\theta(\theta_t,z_t,u_t),\qquad \dot z_t=A_0z_t+\varepsilon t^{-\alpha}f_z(\theta_t,z_t,u_t).$$
- $$E(T):=\int_1^Tt^{-\alpha}dt=\begin{cases}(T^{1-\alpha}-1)/(1-\alpha),&0<\alpha<1,\\ \log T,&\alpha=1,\\ (1-T^{1-\alpha})/(\alpha-1),&\alpha>1.\end{cases}$$
- $$\|r(T)\|\lesssim\varepsilon E(T),\qquad r(T):=s(T)-K(\theta(T),T).$$

## Implementation notes

1. Integration point: replace the transition block of a small state-space model or GRU used for long-horizon forecasting. Split its hidden state into phase coordinates \(\theta\in\mathbb{R}^{d}\) and transverse coordinates \(z\in\mathbb{R}^{r}\), with \(\theta\) represented by sine and cosine features. Use a fixed or learned frequency vector \(\omega\), a zero or nilpotent upper-triangular parabolic matrix \(A_0\), and inject the input through the decaying perturbation. 2. Pseudocode: set \(t=1\); at each step compute \(g=t^{-\alpha}\), then \(f_\theta,f_z=\mathrm{MLP}([\sin\theta,\cos\theta,z,u_t])\). Update \(\theta\leftarrow\theta+\Delta t[\omega+\varepsilon g f_\theta]\) and \(z\leftarrow z+\Delta t[A_0z+\varepsilon g f_z]\), and predict from \([\sin\theta,\cos\theta,z]\). Train the output head and perturbation networks by teacher forcing, using spectral normalization or clipping to bound their Jacobians. 3. Computed from the mechanism: the factor \(t^{-\alpha}\), the phase transport term, and the cumulative bound \(E(T)\). Estimated empirically: the residual-to-torus error \(r(T)\), effective frequency, and perturbation norm. 4. First cheap experiment: train this model and a matched GRU and linear SSM on sums of two incommensurate sinusoids with transient inputs, then evaluate 1000-step free rollouts. Sweep \(\alpha\in\{0.5,0.75,1,1.25,1.5\}\). The prediction is residual drift proportional to \(T^{1-\alpha}\) for \(\alpha<1\), proportional to \(\log T\) at \(\alpha=1\), and saturation for \(\alpha>1\). Fitted log-log slopes should be approximately \(1-\alpha\) below the boundary, while the constant-coupling control should show linear drift.

## Disclaimer

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