H-Infinity Disturbance-Attenuating Latent Observer
Source paper: Operator-Theoretic Stability and Observer Synthesis for Parameter-Dependent Vlasov--Maxwell Dynamics arXiv:2608.28349 ⓘ · analyzed Sep 2, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Turn a latent recurrent model into an observer that continuously corrects its hidden state from noisy or partial observations while certifying both estimation-error convergence and disturbance attenuation. The bounded-real operator inequality becomes a trainable regularizer for a neural correction gain, providing a principled alternative to unconstrained teacher forcing or ad hoc residual correction.
Formulas
Mathematical statement
Let the latent plant be $\dot z=A_\phi(\theta,t)z+B_w w$ with observations $y=C_\phi z+D_w w$. Use the observer $\dot{\hat z}=f_\phi(\hat z,\theta,t)+K_\eta(\theta,t)(y-C_\phi\hat z)$ and error $e=z-\hat z$. Its local error dynamics are $\dot e=A_c e+B_cw$, where $A_c=A_\phi-K_\eta C_\phi$ and $B_c=B_w-K_\eta D_w$. For performance output $q=C_e e+D_e w$, let $P(\theta,t)=P^\top\succ0$ be a storage metric. The bounded-real differential LMI is $\begin{bmatrix}\dot P+A_c^\top P+PA_c+C_e^\top C_e+2\alpha P & PB_c+C_e^\top D_e\\ B_c^\top P+D_e^\top C_e & D_e^\top D_e-\gamma^2I\end{bmatrix}\preceq0$. It implies $\dot V+2\alpha V+\|q\|^2-\gamma^2\|w\|^2\le0$ for $V=e^\top Pe$, giving decay rate $\alpha$ and induced disturbance gain at most $\gamma$.
Implementation notes
1. Integration point: attach an observer head to a neural state-space model, world model, or recurrent encoder. The predictor evolves $\hat z$ autonomously, while each observation produces innovation $r_k=y_k-C_\phi\hat z_k$ and correction $K_\eta r_k$. Use diagonal or low-rank $P$ for large latent dimensions. 2. Pseudocode: predict $\hat z_{k+1}=\hat z_k+h f_\phi(\hat z_k,\theta_k)$; calculate $r_k$ and correct with $hK_\eta r_k$; estimate $A_c$ using Jacobian-vector products; construct $B_c,C_e,D_e$ from the assumed noise and output maps; evaluate the largest eigenvalue of the symmetric bounded-real LMI matrix and add its positive part to the loss. Parameterize $P=LL^\top+p_{\min}I$ and optimize $K_\eta$ jointly with the predictor. 3. Computed from the mechanism: the storage inequality and target gain $\gamma$; estimated empirically: local Jacobians, observation-noise statistics, and minibatch averages over parameter/time points. 4. First experiment: use a partially observed damped oscillator and Lorenz-63, followed by a small video latent model. Compare an equal-parameter GRU with teacher forcing, a learned Kalman-style correction, and this bounded-real observer; ablate the LMI while retaining the correction gain. Evaluate open-loop rollouts under observation noise and impulsive disturbances. 5. Pre-registered signature: with zero disturbance, $V(e_k)$ should decay at least as $e^{-2\alpha kh}V(e_0)$ after a discretization transient; with disturbances, cumulative output energy should satisfy $\sum_k\|q_k\|^2\le\gamma^2\sum_k\|w_k\|^2+V_0$. Reject the transfer if measured induced gain exceeds $1.2\gamma$, or if the estimated LMI is feasible but held-out error decay fails.
Verification
Stage 1 · Toy mechanism gate: Passed ✓
Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗
Stage 2 · Practical benchmark result: No measurable effect
Stage 1 — Mechanism check agent confidence 8/10
Built a scalar H-infinity latent observer MVP with numerical storage-metric search, bounded-real LMI evaluation, exact disturbance-transfer verification, and time-domain dissipation checks. The feasible observer gain K=1.7804 had LMI maximum eigenvalue -0.00098, exact H-infinity gain 0.8166 below gamma=0.9, and zero-disturbance V decay rate 4.36 versus the required 0.4. The unconstrained high-gain baseline K=10 converged faster but was not LMI-feasible and exceeded the target gain with exact H-infinity gain 0.9615, demonstrating the claimed stability/attenuation tradeoff on this toy system.
- Agent confidence
- 8/10
- Baseline
- K=10.0; LMI max eigenvalue=0.07708 (infeasible); exact H-infinity gain=0.96154 > gamma=0.9; observed V decay rate=20.81
- Idea
- K=1.78043, P=0.50642; LMI max eigenvalue=-0.000981 (feasible); exact H-infinity gain=0.81655 < gamma=0.9; observed V decay rate=4.362 > 2*alpha=0.4; random-input energy ratio=0.01979
Limitations: Only a scalar linear continuous-time error system was tested; no neural predictor, learned gain, GRU comparison, Lorenz-63 system, partial-observation training, parameter/time-varying P, or large-scale computational overhead was evaluated. The gain and storage metric were selected by grid search rather than jointly learned.
How to run: python3 observer_experiment.py
Stage 2 — Benchmark vs. tuned baseline
No significant difference vs. the tuned baseline (−-0.0%, p=1, wins 0 / 8 paired seeds).
- Benchmark
- Dynamics forecasting
- Model
- rnn_small
- Paired seeds
- 8
- Baseline mean
- 0.0005
- Idea mean
- 0.0005
- Effect (Δ)
- 0 (−-0.0%; negative = idea better)
- Wins
- 0 / 8 paired seeds
- p-value
- 1 (permutation test, 20 000 shuffles)
- Smallest detectable effect
- ±0.0%
- Mechanism
- Not confirmed ✗
- Practical effect
- No effect
- Baseline tuning
- swept over 3 configs
Only the registered built-in controlled-pendulum dynamics track was tested. The intervention is a finite-dimensional Jacobian-gain penalty proxy rather than a full learned observer with explicit latent P, time-varying plant Jacobians, observation correction, or a formal differential LMI. Held-out impulsive-disturbance rollouts, Lorenz-63, larger models, and a true partial-observation observer head were not tested.
How to run: python3 stage2_bench.py
Verdict computed by deterministic test code from paired-seed statistics — not by the language model.
Stage 2 — Benchmark latest Failed ✗
Agent confidence: 10/10
The H-infinity-inspired Jacobian disturbance-gain penalty was implemented for the matched controlled-pendulum dynamics task using the same rnn_small GRU as baseline. The full paired benchmark showed no improvement: the idea exactly matched the tuned baseline, with paired delta 0 and permutation p=1.0. The trained-model sensitivity signature was also unchanged, so the proposed mechanism was not confirmed.
- Baseline
- Best lr=0.006; 8-seed test MSE mean=0.0005194859513721894, std=0.0001693891077925739.
- Idea
- Best shared lr=0.006; 8-seed test MSE mean=0.0005194859513721894, std=0.0001693891077925739; paired delta=0.0, p=1.0, wins=0/8.
Benchmark result
No significant difference vs. the tuned baseline (−-0.0%, p=1, wins 0 / 8 paired seeds).
- Benchmark
- Dynamics forecasting
- Model
- rnn_small
- Paired seeds
- 8
- Baseline mean
- 0.0005
- Idea mean
- 0.0005
- Effect (Δ)
- 0 (−-0.0%; negative = idea better)
- Wins
- 0 / 8 paired seeds
- p-value
- 1 (permutation test, 20 000 shuffles)
- Smallest detectable effect
- ±0.0%
- Mechanism
- Not confirmed ✗
- Practical effect
- No effect
- Baseline tuning
- swept over 3 configs
Only the registered built-in controlled-pendulum dynamics track was tested. The intervention is a finite-dimensional Jacobian-gain penalty proxy rather than a full learned observer with explicit latent P, time-varying plant Jacobians, observation correction, or a formal differential LMI. Held-out impulsive-disturbance rollouts, Lorenz-63, larger models, and a true partial-observation observer head were not tested.
How to run: python3 stage2_bench.py
Artifacts
- bench_report.json 2.1 KB View Raw JSON
- observer_experiment.py 4.3 KB View Raw JSON
- report.md 1.5 KB View
- report_bench_2026-09-02T122319.md 4.1 KB View
- results.json 1.5 KB View Raw JSON
- stage2_bench.py 6.9 KB View Raw JSON