# Log-Fourier Normal-Form Recurrent Cell

- ID: 1052
- Canonical URL: https://synthcore.org/idea/1052/log-fourier-normal-form-recurrent-cell
- API JSON: https://synthcore.org/api/idea/1052.json
- API Markdown: https://synthcore.org/api/idea/1052.md
- Verification status: unverified
- Source: [arXiv:2607.17068](https://arxiv.org/abs/2607.17068)
- Category: dynamics
- Solves: stability, accuracy
- ML areas: rnn, ssm, world-model, training-dynamics
- Math tags: dynamical-systems, bifurcations, harmonic-analysis, control-theory, numerical-analysis
- Ratings: usefulness 6/10; difficulty 7/10; novelty 8/10

## Idea description

Construct a second-order recurrent cell with an odd high-degree restoring force and lower-degree state-dependent velocity feedback, while representing time-dependent coefficients as a finite Fourier series. At each training or inference window, retain and normalize only Fourier modes below K = c_* log A, where A is the current hidden-state amplitude; apply bounded corrections to nonresonant low modes and leave the analytically small high-frequency tail untouched. The predicted benefit is bounded hidden trajectories and reduced long-horizon error when the restoring degree satisfies n >= 2(m+1).

## Mathematical statement

The source system is x'' + f(x, omega t)x' + g(x, omega t) = 0, with f(x,theta) = sum from j=0 to m of a_j(theta)x^(2j+1) and g(x,theta) = x^(2n+1) + sum from j=0 to n-1 of b_j(theta)x^(2j+1). Here x is the scalar state, theta = omega t is a torus phase, a_j and b_j are analytic even coefficient functions, m and n are nonnegative integers, and omega is Diophantine: |k dot omega| >= gamma |k|^(-tau) for every nonzero integer vector k, with gamma > 0 and tau > 0. The paper assumes n >= 2(m+1), uses the amplitude parameter A, and splits Fourier modes at K = c_* log A. Low modes are removed by a truncated homological equation on nonresonant action intervals. If the coefficient functions are analytic in a strip of width rho, the high-mode tail obeys ||h_>K|| <= C exp(-rho K) = C A^(-rho c_*), where C is a constant and rho > 0. The neural adaptation preserves the degree-gap condition and uses the same cutoff to constrain phase-dependent recurrent coefficients.

## Key formulas

- $$\ddot{x}+\left[\sum_{j=0}^{m}a_j(\omega t)x^{2j+1}\right]\dot{x}+\left[x^{2n+1}+\sum_{j=0}^{n-1}b_j(\omega t)x^{2j+1}\right]=0$$
- $$n\ge 2(m+1),\qquad K=c_*\log A$$
- $$|k\cdot\omega|\ge \gamma |k|^{-\tau}\quad (k\in\mathbb{Z}^d\setminus\{0\})$$
- $$\|h_{>K}\|_{\rho'}\le C e^{-\rho K}=C A^{-\rho c_*},\qquad 0<\rho'<\rho$$

## Implementation notes

1. Integration point: implement a second-order continuous-time RNN or neural-ODE cell. For hidden vector h, use oscillator states x and v with x_dot = v and v_dot = -x^(2n+1) - sum over j<n of b_j(theta)x^(2j+1) - [sum over j<=m of a_j(theta)x^(2j+1)]v + W u. Let theta(t) = omega t + theta_0, and parameterize each a_j and b_j by learned Fourier coefficients. Start with scalar or two-dimensional oscillator blocks before scaling to a full hidden state. 2. Pseudocode: compute A = 1 + norm(h), set K = ceil(c_* log(A)), zero Fourier coefficients whose integer frequency satisfies norm(k)_1 > K, estimate denominators k dot omega, and apply q_hat_k <- q_hat_k - r_hat_k divided by (i k dot omega) only when |k dot omega| >= gamma norm(k)^(-tau). Integrate the corrected cell with RK4 or a split integrator. 3. Computed from the paper: the degree condition, logarithmic cutoff, Diophantine test, and analytic-tail prediction. Estimated empirically: the analyticity constant, correction size, hidden-state growth rate, and practical resonance threshold. 4. First cheap experiment: compare a vanilla tanh GRU, an unfiltered oscillator RNN, and this cell on the adding task and Copy-Memory, followed by autonomous 1,000- to 10,000-step rollouts. Sweep c_*, m, and n, using n = 2(m+1) as the predicted boundary. Measure maximum hidden norm, Jacobian-product norm, and rollout MSE. The quantitative signature is a sharp reduction in state-growth probability for n >= 2(m+1). The discarded-mode residual should scale approximately as A^(-rho c_*); fitting log residual against log A should give slope approximately -rho c_*. Choosing K below c_* log A should increase resonance spikes and long-horizon divergence.

## Disclaimer

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