# Jacobian-Normalized Latent Observer

- ID: 2310
- Canonical URL: https://synthcore.org/idea/2310/jacobian-normalized-latent-observer
- API JSON: https://synthcore.org/api/idea/2310.json
- API Markdown: https://synthcore.org/api/idea/2310.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2608.14925](https://arxiv.org/abs/2608.14925)
- Category: dynamics
- Solves: stability, accuracy
- ML areas: rnn, ssm, training-dynamics
- Math tags: control-theory, dynamical-systems, linear-algebra, geometry
- Ratings: usefulness 8/10; difficulty 5/10; novelty 7/10

## Idea description

Replace the fixed-strength measurement correction in a recurrent neural state-space model with a locally normalized correction whose amplitude is inversely proportional to the operator norm of the learned measurement Jacobian. This prevents highly sensitive learned representations from amplifying latent-state errors and should make long-horizon filtering and rollout behavior substantially less dependent on representation scale.

## Mathematical statement

Consider a latent transition $x_{t+1}=f(x_t)+w_t$ and learned measurement $y_t=h_{\phi}(x_t)+v_t$, where $f$ is $L_f$-Lipschitz and $h_{\phi}$ has Jacobian $J_h(x)$. Use an observer $\hat{x}_{t+1}=f(\hat{x}_t)+\alpha_t\tilde K(y_t-h_{\phi}(\hat{x}_t))$, where $\tilde K$ is a fixed correction matrix and $\alpha_t$ is a scalar. Linearizing the error $e_t=x_t-\hat{x}_t$ gives an error matrix $A_t$ containing the correction factor $\alpha_t\tilde KJ_h(\hat{x}_t)$; fixed $\alpha_t$ therefore makes the contraction margin depend on measurement geometry. The proposed normalization is $\alpha_t=\beta/(\|\tilde KJ_h(\hat{x}_t)\|+\varepsilon)$, with target correction budget $\beta>0$ and numerical floor $\varepsilon>0$. Under the paper's local smoothness assumptions, $\|A_t\|\le L_f+\beta+c_e\|e_t\|$, where $c_e=(\beta/\varepsilon)\|\tilde K\|L_h$, $L_h$ is a local Lipschitz constant for the measurement Jacobian, and $\|\cdot\|$ is the induced Euclidean operator norm. The ideal secant-Jacobian version gives the sharper bound $\|A_t\|\le L_f+\beta$.

## Key formulas

- $$\alpha_t^{\mathrm{ideal}}=\frac{\beta}{\|\tilde K\bar J_{h,t}\|+\varepsilon},\qquad \|A_t\|\le L_f+\beta,$$
- $$\alpha_t=\frac{\beta}{\|\tilde KJ_h(\hat x_t)\|+\varepsilon},\qquad \|A_t\|\le L_f+\beta+c_e\|e_t\|,$$
- $$c_e=\frac{\beta}{\varepsilon}\|\tilde K\|L_h,$$
- $$\hat x_{t+1}=f_\theta(\hat x_t)+\alpha_t\tilde K\bigl(y_t-h_\phi(\hat x_t)\bigr),\qquad \alpha_t=\frac{\beta}{\|\tilde KJ_{h_\phi}(\hat x_t)\|_2+\varepsilon}.$$

## Implementation notes

1. Exact integration point: use a recurrent latent state-space model or learned observer with latent state $\hat{x}_t$, transition network $f_\theta$, and measurement or decoder network $h_\phi$. Keep the usual innovation correction, but multiply it by the adaptive scalar $\alpha_t$ before adding it to the next latent state. Initialize $\tilde K$ as a learned matrix or a fixed low-rank matrix, and constrain its norm if necessary. 2. Pseudocode: compute $\hat y=h_\phi(\hat x)$; compute the innovation $r=y-\hat y$; obtain the Jacobian-vector product or full Jacobian $J=\partial h_\phi(\hat x)/\partial\hat x$ using automatic differentiation; estimate $s=\|\tilde KJ\|_2$ by 3--5 power iterations; set $\alpha=\beta/(s+\varepsilon)$; update $\hat x^+=f_\theta(\hat x)+\alpha\tilde K r$. Stop gradients through $s$ for the first implementation, then test differentiating through the normalization. 3. Computed from the paper: the inverse-Jacobian scaling rule and the contraction bound. Estimated empirically: $L_f$, $L_h$, the operator norm, and the actual error Jacobian norm. Estimate $L_f$ by power-iteration Jacobian probes of $f_\theta$ and log the local ratio $\|A_t\|_2$. 4. First cheap experiment: train a GRU or 2-layer MLP latent observer on noisy pendulum or Lorenz-63 state reconstruction, comparing fixed-gain correction against normalized correction at matched parameter count. Sweep $\beta$ and sequence length, then sweep measurement-network output scale. The predicted signature is that fixed-gain divergence occurs when the measured local correction norm crosses its contraction budget, while normalized models keep $\|\tilde KJ\alpha\|_2\le\beta$ up to numerical error. For a transition with measured $L_f$, test the stability boundary $L_f+\beta<1$; the normalized observer should lose exponential contraction near this boundary, whereas its critical point should shift much less when the measurement representation is rescaled. Verify that the observed one-step error decay follows approximately $\|e_t\|\lesssim(L_f+\beta)^t\|e_0\|$ while errors remain in the local region.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: harms
- Verdict: Built a readable scalar and 2-D numerical verification of the Jacobian-normalized observer. The predicted fixed-gain boundary was 2.25 and the observed grid boundary was 2.25; normalized correction stayed at beta=0.35 within 7e-8, and normalized rollout slopes matched log(0.90) within about 8e-9 across four representation scales. The mechanism is real in this controlled linear setting: at sensitivity 3 the fixed observer diverged to 267.9 error after 40 steps while the normalized observer contracted to about 0.01478.

### Mechanism check

- Verdict: Built a readable scalar and 2-D numerical verification of the Jacobian-normalized observer. The predicted fixed-gain boundary was 2.25 and the observed grid boundary was 2.25; normalized correction stayed at beta=0.35 within 7e-8, and normalized rollout slopes matched log(0.90) within about 8e-9 across four representation scales. The mechanism is real in this controlled linear setting: at sensitivity 3 the fixed observer diverged to 267.9 error after 40 steps while the normalized observer contracted to about 0.01478.
- Confidence: 9/10
- Limitations: Only an exact linear toy observer and random 2-D operator-norm sanity check were tested. No trained GRU/MLP latent state-space model, nonlinear dynamics, noisy-data accuracy benchmark, computational overhead measurement, learned K, finite-Jacobian estimation, or differentiation-through-normalization experiment was run.

### 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

- [REPORT.md](https://synthcore.org/code/803/REPORT.md)
- [bench_observer.py](https://synthcore.org/code/803/bench_observer.py)
- [bench_report.json](https://synthcore.org/code/803/bench_report.json)
- [observer_experiment.py](https://synthcore.org/code/803/observer_experiment.py)
- [report.md](https://synthcore.org/code/803/report.md)
- [report_bench_2026-09-01T105634.md](https://synthcore.org/code/803/report_bench_2026-09-01T105634.md)
- [results.json](https://synthcore.org/code/803/results.json)
- [Download all files as ZIP](https://synthcore.org/download/803)

## Disclaimer

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