# State-Range Observer Gain Scheduler

- ID: 1419
- Canonical URL: https://synthcore.org/idea/1419/state-range-observer-gain-scheduler
- API JSON: https://synthcore.org/api/idea/1419.json
- API Markdown: https://synthcore.org/api/idea/1419.md
- Verification status: mechanism_failed
- Source: [arXiv:2607.25879](https://arxiv.org/abs/2607.25879)
- Category: dynamics
- Solves: stability, accuracy, generalization
- ML areas: ssm, rnn, optimizer, world-model
- Math tags: dynamical-systems, control-theory, linear-algebra, spectral-theory
- Ratings: usefulness 7/10; difficulty 7/10; novelty 7/10

## Idea description

Make the observation-injection gain state dependent, increasing it only when the projected unobserved dynamics approach the Hurwitz boundary. This creates a feedback controller for latent drift while avoiding the observation-noise amplification caused by using a globally oversized gain.

## Mathematical statement

Let $\Gamma$ be a compact region containing the latent trajectory and let $J_i$ be Jacobian samples or vertices representing the model over that region. For gain $G$, define $H_i(G)=U_K^{\mathsf T}(J_i-GC)U_K$, where $K=\ker C$ and $U_K$ is an orthonormal basis of $K$. Seek a common positive-definite matrix $P$ and margin $\alpha>0$ satisfying $H_i(G)^{\mathsf T}P+PH_i(G)\preceq-2\alpha P$ for every sampled vertex. During inference, use $G_t=\rho_tG_0$ and choose the smallest candidate multiplier $\rho_t$ for which the current restricted symmetric Jacobian has a negative Lyapunov margin. The resulting certificate predicts exponential decay $V(e_t)\leq e^{-2\alpha t}V(e_0)$ whenever the selected margin remains valid.

## Key formulas

- $$H_i(G)=U_K^{\mathsf T}\bigl(J_i-GC\bigr)U_K$$
- $$H_i(G)^{\mathsf T}P+PH_i(G)\preceq-2\alpha P,\qquad P\succ0,\ \alpha>0$$
- $$G_t=\rho_tG_0,\qquad \rho_t=\min\left\{\rho\in\mathcal{R}:\lambda_{\max}\left(PH_t(\rho)+H_t(\rho)^{\mathsf T}P\right)\leq-2\alpha\lambda_{\min}(P)\right\}$$
- $$V(e)=e^{\mathsf T}Pe,\qquad V(e_t)\leq e^{-2\alpha t}V(e_0)$$

## Implementation notes

1. Integration point: attach the scheduler to a neural ODE, continuous-time RNN, or learned world model receiving intermittent noisy measurements. Use a base gain $G_0$ and candidate multipliers $\mathcal{R}=\{1,1.25,1.5,2,3,4\}$. The scheduler changes only the residual-correction branch. 2. Pseudocode: before each observation update, compute Jacobian-vector products of $f_{\theta}$ at the current state; for each candidate $\rho$, form $H_t(\rho)=U_K^{\mathsf T}(J_t-\rho G_0C)U_K$; estimate the largest eigenvalue of $PH_t+H_t^{\mathsf T}P$ by power iteration; choose the smallest candidate satisfying the negative-margin condition; then update $z$ with $f_{\theta}(z)+\rho G_0(y-Cz)$. Add hysteresis by increasing $\rho$ immediately after a violation and decreasing it only after ten safe steps. 3. The Lyapunov inequality and gain-selection rule come from the observer mechanism; $J_t$, $P$, the state-range coverage, and the noise tradeoff are estimated. Solve for $P$ once using an SDP on stored Jacobians or learn it jointly. 4. First experiment: use a neural ODE on noisy partially observed Lorenz-63, comparing fixed low gain, fixed high gain, and scheduled gain. Sweep the estimated maximum projected eigenvalue. The predicted signature is negative reconstruction-error slope below the stability boundary and positive slope above it; the measured sign-change boundary should predict failure within 20 percent. During safe episodes, scheduling should reduce residual variance by at least 20 percent relative to fixed high gain while retaining its recovery rate after an unstable excursion.

## Verification

- Status: mechanism_failed
- Mechanism evidence: yes
- Mechanism confirmed: no
- Verdict: Built a reproducible two-state noisy observer with kernel projection, gain candidates, Lyapunov-margin selection, parameter sweeps, and fixed-gain comparisons. The core math check found C U_K=0 exactly, making U_K^T G C U_K identically zero for every gain; the observed projected stability boundary was gamma≈0 within the 0.01 grid resolution and did not move with rho. The scheduler selected rho=1 for stable cases and only fell back to rho=4 for unstable cases without fixing instability; on the stable noisy toy, scheduled RMSE matched fixed low gain (0.007893) and was better than fixed high gain (0.010423), but this is not evidence of the claimed mechanism.

### Mechanism check

- Verdict: Built a reproducible two-state noisy observer with kernel projection, gain candidates, Lyapunov-margin selection, parameter sweeps, and fixed-gain comparisons. The core math check found C U_K=0 exactly, making U_K^T G C U_K identically zero for every gain; the observed projected stability boundary was gamma≈0 within the 0.01 grid resolution and did not move with rho. The scheduler selected rho=1 for stable cases and only fell back to rho=4 for unstable cases without fixing instability; on the stable noisy toy, scheduled RMSE matched fixed low gain (0.007893) and was better than fixed high gain (0.010423), but this is not evidence of the claimed mechanism.
- Confidence: 10/10
- Limitations: This intentionally tests the formulas as written in a linear two-state observer rather than a neural ODE or Lorenz-63 system. It does not solve an SDP for P, estimate Jacobians with autodiff, test hysteresis, or evaluate a corrected observer formulation where injection acts outside ker(C).

## Artifacts

- [observer_scheduler_experiment.py](https://synthcore.org/code/557/observer_scheduler_experiment.py)
- [report.md](https://synthcore.org/code/557/report.md)
- [results.json](https://synthcore.org/code/557/results.json)
- [Download all files as ZIP](https://synthcore.org/download/557)

## Disclaimer

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