# Finite-Excitation Latent Replay

- ID: 2914
- Canonical URL: https://synthcore.org/idea/2914/finite-excitation-latent-replay
- API JSON: https://synthcore.org/api/idea/2914.json
- API Markdown: https://synthcore.org/api/idea/2914.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.30977](https://arxiv.org/abs/2608.30977)
- Category: training
- Solves: stability, sample-efficiency, accuracy
- ML areas: optimizer, ssm, rnn, world-model
- Math tags: control-theory, linear-algebra, spectral-theory, dynamical-systems
- Ratings: usefulness 8/10; difficulty 5/10; novelty 7/10

## Idea description

Replace derivative-based latent-dynamics fitting with an integral regression and maintain a history stack selected by the smallest eigenvalue of its information matrix. The model should perform aggressive parameter updates only when the estimated latent regressors are sufficiently exciting, while a perturbation bound prevents false excitation caused by inaccurate hidden-state estimates.

## Mathematical statement

Consider a latent state model \(\dot{x}=Ax+Bu+\phi(x,u)+\Phi(x,u)\theta+Dd(t)\), where \(x\) is the latent state, \(u\) is the input, \(\theta\) is an unknown parameter vector, and \(d\) is a bounded disturbance. Over an interval \([t_i,t_i+T]\), integration gives \(x(t_i+T)-x(t_i)=\int_{t_i}^{t_i+T}[Ax+Bu+\phi(x,u)]dt+\Omega_i\theta+r_i\), where \(\Omega_i=\int_{t_i}^{t_i+T}\Phi(x(t),u(t))dt\) and \(r_i\) contains disturbance and modeling residuals. Using an estimated state produces \(\widehat{\Omega}_i=\Omega_i+\Delta_i\). The true information matrix is \(G=\sum_i\Omega_i^{\mathsf T}\Omega_i\), and the computable matrix is \(\widehat{G}=\sum_i\widehat{\Omega}_i^{\mathsf T}\widehat{\Omega}_i\). If \(\|\Delta_i\|\leq\varepsilon_i\), then \(\|\widehat{G}-G\|\leq\sum_i(2\|\Omega_i\|\varepsilon_i+\varepsilon_i^2)\). Thus \(\lambda_{\min}(G)\geq\lambda_{\min}(\widehat{G})-\|\widehat{G}-G\|\), so positive corrected excitation certifies informative data.

## Key formulas

- $$x(t_i+T)-x(t_i)=\int_{t_i}^{t_i+T}[Ax(t)+Bu(t)+\phi(x(t),u(t))]dt+\Omega_i\theta+r_i,\qquad \Omega_i=\int_{t_i}^{t_i+T}\Phi(x(t),u(t))dt.$$
- $$\widehat{G}=\sum_{i=1}^{M}\widehat{\Omega}_i^{\mathsf T}\widehat{\Omega}_i,\qquad G=\sum_{i=1}^{M}\Omega_i^{\mathsf T}\Omega_i,$$
- $$\|\widehat{G}-G\|\leq\sum_{i=1}^{M}\left(2\|\Omega_i\|\|\Delta_i\|+\|\Delta_i\|^2\right),\qquad \Delta_i=\widehat{\Omega}_i-\Omega_i,$$
- $$\lambda_{\min}(G)\geq\lambda_{\min}(\widehat{G})-\|\widehat{G}-G\|.$$

## Implementation notes

1. Integration point: use a latent ODE, state-space model, or recurrent model with a dynamics head that is linear in an adapter parameter vector \(\theta\); a neural feature map \(\Phi_\psi(z,u)\) can generate the regressors. Maintain a buffer of windows \((z(t_i),z(t_i+T),u_{i:i+T})\), where \(z\) is the encoder or recurrent hidden state. 2. Pseudocode: for every stored window, numerically integrate the feature matrix using \(\widehat{\Omega}_i=\sum_k\Phi_\psi(z_k,u_k)\Delta t\); form \(\widehat G=\sum_i\widehat\Omega_i^{\mathsf T}\widehat\Omega_i\); estimate latent uncertainty \(\varepsilon_i\) from an ensemble variance, dropout variance, or observer residual; compute \(q=\lambda_{\min}(\widehat G)-\sum_i(2\|\widehat\Omega_i\|\varepsilon_i+\varepsilon_i^2)\). If \(q>\gamma\), update \(\theta\) using the integral-regression residual; otherwise freeze or downweight the dynamics-adapter update. Select history windows greedily by maximizing the increase in \(\lambda_{\min}(\widehat G)\). 3. The integral regression and perturbation inequality are taken from the paper; the true matrix and exact state error are unavailable, so \(\varepsilon_i\) must be estimated empirically and inflated by a safety factor. 4. First cheap experiment: train a two- or four-dimensional latent state-space model on Lorenz-63 or a damped nonlinear oscillator, comparing Adam, ungated integral regression, and excitation-gated replay. Sweep observation noise and window length. The quantitative prediction is that parameter error decreases rapidly after \(q\) becomes positive, with the empirical transition from ill-conditioned to stable identification near \(q=0\). When true states are available for evaluation, measured \(\lambda_{\min}(G)\) should exceed the conservative lower bound \(q\), up to numerical integration error.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: harms
- Verdict: Built a NumPy MVP implementing integral regressors, Gramian excitation, perturbation correction, and greedy replay selection in finite_excitation_replay.py. The mechanism checks passed: 0/12 perturbation-bound violations, expected small-error O(epsilon) scaling, and a certificate threshold crossing at epsilon 0.240 observed versus 0.242 interpolated. Replay activated at step 4, but its final parameter error was worse than using all data (0.0541 versus 0.00697), so the mechanism is supported while this toy replay policy did not show an identification win.

### Mechanism check

- Verdict: Built a NumPy MVP implementing integral regressors, Gramian excitation, perturbation correction, and greedy replay selection in finite_excitation_replay.py. The mechanism checks passed: 0/12 perturbation-bound violations, expected small-error O(epsilon) scaling, and a certificate threshold crossing at epsilon 0.240 observed versus 0.242 interpolated. Replay activated at step 4, but its final parameter error was worse than using all data (0.0541 versus 0.00697), so the mechanism is supported while this toy replay policy did not show an identification win.
- Confidence: 8/10
- Limitations: This is a deterministic linear toy identification test, not a neural latent ODE or Lorenz experiment. It uses one random seed, synthetic matrix regressors, known perturbation magnitudes for validation, and a simple sequential greedy stack rather than a full online observer or ensemble uncertainty estimator; runtime, FLOPs, noisy latent-state estimation, and multi-seed statistical significance were not tested.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 0/8
- Benchmark verdict: idea worse (significant)

## Artifacts

- [bench_felr.py](https://synthcore.org/code/1077/bench_felr.py)
- [bench_report.json](https://synthcore.org/code/1077/bench_report.json)
- [finite_excitation_replay.py](https://synthcore.org/code/1077/finite_excitation_replay.py)
- [report.md](https://synthcore.org/code/1077/report.md)
- [report_bench_2026-09-01T235855.md](https://synthcore.org/code/1077/report_bench_2026-09-01T235855.md)
- [results.json](https://synthcore.org/code/1077/results.json)
- [Download all files as ZIP](https://synthcore.org/download/1077)

## Disclaimer

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