# Low-Order Robust Functional Observer

- ID: 1384
- Canonical URL: https://synthcore.org/idea/1384/low-order-robust-functional-observer
- API JSON: https://synthcore.org/api/idea/1384.json
- API Markdown: https://synthcore.org/api/idea/1384.md
- Verification status: unverified
- Source: [arXiv:2607.25000](https://arxiv.org/abs/2607.25000)
- Category: architecture
- Solves: memory, stability, accuracy
- ML areas: ssm, rnn, world-model, regularization
- Math tags: control-theory, optimization, convex-analysis, dynamical-systems
- Ratings: usefulness 6/10; difficulty 6/10; novelty 5/10

## Idea description

Attach a small dynamical observer to a neural ODE, RNN, or state-space model and make it estimate only a task-relevant functional of the hidden state, such as logits, value features, or control-relevant projections. Use an incremental quadratic constraint and a bounded-real penalty to make the observer robust to hidden-state nonlinearities and input disturbances, instead of reconstructing the full latent state.

## Mathematical statement

The paper models a nonlinear state as E dot{x}=A x+B u+D v+bar{D}_xi xi+F g(K x,t), with measured output y=C x+G v and desired functional z=K_z x. Its filter can have order r less than or equal to p=dim(z), rather than reconstructing the entire n-dimensional state. The nonlinear map is constrained by an incremental quadratic constraint: for two states x_1 and x_2, define delta x=x_1-x_2 and delta g=g(K x_1,t)-g(K x_2,t), and require [delta x;delta g]^T Pi[delta x;delta g]>=0, where Pi=[[Q,S],[S^T,R]] is a multiplier matrix. The extracted example uses the one-sided-Lipschitz multiplier Pi=kappa [[0,-0.5],[-0.5,0]] with kappa=2 for g(x)=exp(-x_1). The neural adaptation uses a reduced observer dot{eta}=A_f eta+B_f y and z_hat=C_f eta+D_f y. For local augmented error dynamics dot{e}=A_e e+B_e w and performance error q=C_e e+D_e w, impose the bounded-real LMI with Lyapunov matrix P positive definite. Here e is the combined plant-observer error, w is disturbance or model mismatch, q is functional-estimation error, and gamma bounds the induced L2 gain.

## Key formulas

- $$E\dot{x}(t)=A x(t)+B u(t)+D v(t)+\bar{D}_{\xi}\xi(t)+F g(Kx,t),\qquad y(t)=C x(t)+Gv(t),\qquad z(t)=K_zx(t).$$
- $$\begin{bmatrix}\delta x\\\delta g\end{bmatrix}^{T}\Pi\begin{bmatrix}\delta x\\\delta g\end{bmatrix}\ge 0,\qquad \delta x=x_1-x_2,\quad \delta g=g(Kx_1,t)-g(Kx_2,t),\quad \Pi=\begin{bmatrix}Q&S\\S^T&R\end{bmatrix}.$$
- $$\dot{\eta}=A_f\eta+B_fy,\qquad \hat z=C_f\eta+D_fy,\qquad r=\dim(\eta)\le p=\dim(z).$$
- $$\begin{bmatrix}\operatorname{He}(PA_e)+C_e^TC_e&PB_e+C_e^TD_e\\B_e^TP+D_e^TC_e&D_e^TD_e-\gamma^2I\end{bmatrix}\prec0,\qquad P\succ0.$$

## Implementation notes

(1) Exact integration point: use a continuous-time state-space sequence model or neural ODE with hidden state x in R^n, measured feature y_t available during streaming inference, and task functional z_t=K_z x_t in R^p. Add an auxiliary observer state eta_t in R^r with r much smaller than n, preferably r less than or equal to p, and use z_hat_t=C_f eta_t+D_f y_t as an extra prediction head or as a replacement for the full-state readout. (2) Pseudocode: initialize eta=0; at each time step compute x_next=NeuralDynamics(x,u); compute eta_next=eta+Delta_t*(A_f*eta+B_f*y); compute z_hat=C_f*eta+D_f*y; minimize the task loss plus lambda_z times ||z_hat-K_z*x_teacher||^2 and lambda_r times max(0, estimated ||q||^2-gamma^2 estimated ||w||^2). (3) Quantities computed from the mathematics: estimate A_e,B_e,C_e,D_e by automatic differentiation or least-squares linearization over minibatch trajectories. Estimate the deltaQC multiplier using a conservative scalar one-sided-Lipschitz bound. For a scalar activation use Q=0, S=-kappa/2, R=0; for vector activations use block-diagonal multipliers. Solve the bounded-real LMI as a semidefinite program for P and, when the neural dynamics are fixed, for observer matrices. If joint optimization makes the constraints bilinear, alternate between neural-weight updates and SDP observer updates, or parameterize A_f=-L L^T-epsilon I to enforce stability. (4) First cheap experiment: use a GRU or 64-dimensional neural ODE on permuted-MNIST or an IMU time-series benchmark, expose only a 16-dimensional measurement y_t, and compare a full 64-dimensional readout against observers with r in {2,4,8,16}. Measure task accuracy, functional-estimation error, peak disturbance gain under additive input noise, and streaming memory. Success means equal accuracy with at least 4x smaller observer state, lower error under held-out perturbations, and empirical sum_t ||q_t||^2 divided by sum_t ||w_t||^2 below the chosen gamma^2.

## Disclaimer

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