# Task-Targeted Spectral Excitation for Dynamics Learning

- ID: 1227
- Canonical URL: https://synthcore.org/idea/1227/task-targeted-spectral-excitation-for-dynamics-learning
- API JSON: https://synthcore.org/api/idea/1227.json
- API Markdown: https://synthcore.org/api/idea/1227.md
- Verification status: unverified
- Source: [arXiv:2607.21280](https://arxiv.org/abs/2607.21280)
- Category: training
- Solves: sample-efficiency, accuracy, stability
- ML areas: world-model, ssm, rnn, data-augmentation
- Math tags: control-theory, dynamical-systems, harmonic-analysis, linear-algebra
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

When training a neural state-space model, SSM, or recurrent world model from trajectories, constrain the data-generation policy or augmentation process to satisfy both a Hankel-rank condition and a task-weighted frequency-coverage condition. The rank condition prevents unidentifiable dynamics, while the frequency condition concentrates samples at frequencies that affect the target prediction horizon, tracking objective, or closed-loop controller instead of merely producing broadband-looking data.

## Mathematical statement

For an input trajectory u_t ∈ R^{n_u}, define the depth-L Hankel matrix H_L(u) whose columns are [u_k^T,...,u_{k+L-1}^T]^T. Algebraic excitation requires rank(H_L(u)) = L n_u, or a numerically robust version σ_min(H_L(u)) ≥ ε_rank. Let U(e^{jω}) and Y(e^{jω}) be finite-length discrete Fourier transforms of input and output trajectories, and define W_ρ(e^{jω}) = [e^{-jωρ},...,e^{-jω}]^T. The paper's behavioral construction gives X(e^{jω}) = C_u(W_ρ(e^{jω}) ⊗ U(e^{jω})) + C_y(W_ρ(e^{jω}) ⊗ Y(e^{jω})), where X is the state spectrum and C_u,C_y are deterministic reconstruction matrices. For a neural predictor, replace unknown state sensitivity by an empirical task sensitivity q(ω) obtained by perturbing Fourier bins and measuring prediction-loss or closed-loop-loss change. Define spectral coverage E(ω)=|U(e^{jω})|^2+λ_y|Y(e^{jω})|^2 and require both σ_min(H_L(u))≥ε_rank and ∫q(ω)E(ω)dω≥ε_task. The mechanism predicts failure when rank is high but task-weighted spectral energy is low.

## Key formulas

- $$X(e^{j\omega})=\mathcal{C}_{u}\bigl(W_{\rho}(e^{j\omega})\otimes U(e^{j\omega})\bigr)+\mathcal{C}_{y}\bigl(W_{\rho}(e^{j\omega})\otimes Y(e^{j\omega})\bigr),\qquad W_{\rho}(e^{j\omega})=\begin{bmatrix}e^{-j\omega\rho}&\cdots&e^{-j\omega}\end{bmatrix}^{\top}.$$
- $$\operatorname{rank}\!\left(\mathcal{H}_{L}(u)\right)=Ln_u\quad\text{or numerically}\quad \sigma_{\min}\!\left(\mathcal{H}_{L}(u)\right)\geq\varepsilon_{\mathrm{rank}}.$$
- $$\mathcal{S}_{\mathrm{task}}=\int_{-\pi}^{\pi}q(\omega)E(\omega)\,d\omega,\qquad E(\omega)=\left\lVert U(e^{j\omega})\right\rVert_{2}^{2}+\lambda_y\left\lVert Y(e^{j\omega})\right\rVert_{2}^{2},\qquad \mathcal{S}_{\mathrm{task}}\geq\varepsilon_{\mathrm{task}}.$$
- $$u_{t+1}=u_t+\alpha\,\mathcal{F}^{-1}\!\left[q(\omega)\,\frac{U(e^{j\omega})}{\left\lVert U(e^{j\omega})\right\rVert_2+\delta}\right]_{t}.$$

## Implementation notes

1. Integration point: use this during data collection for a neural state-space model, recurrent world model, or nonlinear SSM, rather than as an ordinary minibatch loss. Begin with a controllable simulator such as a damped oscillator, CartPole, or MuJoCo pendulum. Generate trajectories using a nominal policy plus an additive colored excitation generator. 2. Pseudocode: collect a trajectory; construct H_L(u) and compute its smallest singular value; compute FFTs U and Y; estimate q(ω) by adding a small sinusoid at each of K frequency bins to validation rollouts and measuring the increase in H-step prediction loss or closed-loop tracking loss; update excitation amplitudes using E_k ← |U_k|²+λ_y|Y_k|² and a_k ← a_k+αq_k/(E_k+δ); inverse FFT the updated spectrum, clip actions to valid bounds, and accept the dataset only if σ_min(H_L(u))≥ε_rank and Σ_kq_kE_k≥ε_task. 3. Computed from the paper: the Hankel rank test, delayed Fourier representation, and the distinction between rank excitation and frequency-relevant excitation. Estimated empirically: q(ω), output spectra, thresholds, and the neural predictor. 4. First cheap experiment: identify a two-dimensional damped oscillator with a GRU or S4-style SSM from 50 trajectories, comparing white noise, rank-only colored noise, and proposed rank-plus-task excitation. Evaluate 50-step prediction and reference tracking. The quantitative prediction is that rank-only data can have σ_min(H_L(u)) above ε_rank but low Σ_kq_kE_k and poor performance; adding energy to high-q bins should reduce task loss. Plot performance against S_task and test whether the failure transition occurs near S_task=ε_task, with a stronger correlation between S_task and test error than between Hankel rank and test error.

## Disclaimer

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