# Floquet-Sideband State-Space Layer

- ID: 1797
- Canonical URL: https://synthcore.org/idea/1797/floquet-sideband-state-space-layer
- API JSON: https://synthcore.org/api/idea/1797.json
- API Markdown: https://synthcore.org/api/idea/1797.md
- Verification status: unverified
- Source: [arXiv:2608.04558](https://arxiv.org/abs/2608.04558)
- Category: dynamics
- Solves: stability, accuracy, generalization
- ML areas: ssm, rnn, training-dynamics, regularization
- Math tags: dynamical-systems, control-theory, spectral-theory, stochastic-processes, linear-algebra
- Ratings: usefulness 6/10; difficulty 6/10; novelty 7/10

## Idea description

Replace a time-invariant linear state-space transition with a periodic transition whose coefficients have a learned period T. Constrain the product of one period to be contractive, and regularize its Fourier sidebands so that periodically driven modes do not accumulate unstable resonant energy. The architecture predicts an observable stability boundary through the spectral radius of its monodromy matrix and a measurable sideband occupation profile.

## Mathematical statement

Let the hidden state obey h_{k+1}=A_{k mod T}h_k+B_{k mod T}x_k, with period-T matrices A_r and B_r. The Floquet monodromy is M=A_{T-1}A_{T-2}\cdots A_0; homogeneous dynamics are asymptotically stable when the spectral radius rho(M)<1, and decay per period is approximately rho(M). In the continuous-time analogue, U(t)=P(t)exp(tH_F/(i hbar)), where P(t+2pi/Omega)=P(t) and H_F is the Floquet generator. Expand P(t)=sum_xi P_xi exp(i xi Omega t); a Floquet mode with quasienergy epsilon_j has sideband occupation n_j=sum_xi w_{jxi}f(epsilon_j+xi hbar Omega), where w_{jxi}=[P_xi^dagger P_xi]_{jj}, sum_xi w_{jxi}=1, and f(E)=(1+exp((E-mu)/(k_B T)))^{-1}. For the neural layer, use the same weighted sideband construction as a regularizer on singular-mode gains, but estimate gains from the discrete Fourier transform of the learned A_r rather than interpreting them as literal quantum occupations.

## Key formulas

- $$M=A_{T-1}A_{T-2}\cdots A_0,\qquad \rho(M)<1\ \Longrightarrow\ \|h_{qT}\|\lesssim \rho(M)^q\|h_0\|.$$
- $$U(t)=P(t)\exp\left(\frac{t}{i\hbar}H_F\right),\qquad P(t)=\sum_{\xi\in\mathbb Z}P_\xi e^{i\xi\Omega t}.$$
- $$n_j=\sum_{\xi}w_{j\xi}f(\varepsilon_j+\xi\hbar\Omega),\qquad w_{j\xi}=[P_\xi^\dagger P_\xi]_{jj},\qquad f(E)=\frac{1}{1+e^{(E-\mu)/(k_BT)}}.$$
- $$\mathcal L_{\mathrm{side}}=\sum_j\left[n_j^{\mathrm{emp}}-\sum_{|\xi|\le K}w_{j\xi}f(\lambda_j+\xi\Omega;\mu,\tau)\right]^2,\qquad \mathcal L=\mathcal L_{\mathrm{task}}+\alpha\mathcal L_{\mathrm{side}}+\beta\max(0,\rho(M)-(1-\delta))^2.$$

## Implementation notes

1. Integration point: implement a small periodic SSM/RNN block. For period T, store A_0,...,A_{T-1} and B_0,...,B_{T-1}; at sequence position k use A_{k mod T}. Initialize each A_r near a stable matrix, for example A_r=0.9 Q_r with orthogonal Q_r. During every optimizer step form M by multiplying the T transition matrices, estimate its largest eigenvalue with 10-20 power iterations, and add the monodromy penalty in the formula. Optionally rescale all A_r by ((1-delta)/max(rho(M),1-delta))^(1/T) after each update. 2. Pseudocode: compute h=0; for k in sequence: r=k%T; h=A[r]@h+B[r]@x[k]; y=C@h; then compute M, rho, and the stability penalty. Every N steps, FFT the sequence A_r across r, obtaining Fourier coefficients A_hat[xi]. Use their modal singular vectors to estimate sideband weights w[j,xi], truncate to |xi|<=K, and add L_side. 3. Computed from the paper's mechanism: Floquet decomposition, sideband shifts, weighted occupation form, and monodromy stability criterion. Estimated empirically: rho(M), modal gains, empirical mode energies n_emp from long-run hidden-state variance, and the effective temperature tau. 4. First cheap experiment: train T=8 and T=16 versions on sequential MNIST or permuted-MNIST, comparing a fixed-A SSM, unconstrained periodic SSM, and constrained Floquet model. Sweep a scalar multiplier s on all A_r. The quantitative prediction is a sharp long-horizon divergence transition near s_c=(1/rho(M_0))^(1/T); measured bounded/unbounded hidden-state behavior should switch within 10% of rho(M)=1. For stable models, hidden-state energy sampled at drive phase should show sideband peaks at frequencies xi/T, and the fitted occupation curve should agree with the weighted-sideband prediction with mean relative error below 20% in the locally linear, weak-noise regime.

## Disclaimer

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