# Schur-Constrained Neural Derivative Feedback

- ID: 1982
- Canonical URL: https://synthcore.org/idea/1982/schur-constrained-neural-derivative-feedback
- API JSON: https://synthcore.org/api/idea/1982.json
- API Markdown: https://synthcore.org/api/idea/1982.md
- Verification status: unverified
- Source: [arXiv:2608.08318](https://arxiv.org/abs/2608.08318)
- Category: dynamics
- Solves: stability, accuracy
- ML areas: rl, world-model, training-dynamics
- Math tags: control-theory, dynamical-systems, spectral-theory, linear-algebra
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Add a finite-difference derivative branch to a neural feedback policy, but constrain its gain using the sampled-system fast-mode criterion from the paper. The controller can retain derivative information while avoiding high-frequency instability caused by the stored previous observation, especially when the control loop is sampled rapidly.

## Mathematical statement

The plant is the linearization \(\dot{x}(t)=Ax(t)+Bu(t)\), with state \(x(t)\in\mathbb{R}^n\), control \(u(t)\in\mathbb{R}^m\), output \(y(t)=Cx(t)\in\mathbb{R}^p\), and matrices \(A,B,C\) of compatible dimensions. The ideal PID law is \(u(t)=-K_Py(t)-K_Iz(t)-K_D\dot y(t)\), with integral state \(\dot z(t)=y(t)\); \(K_D\in\mathbb{R}^{m\times p}\) is the derivative gain. Under backward differences, \(\dot y(t_k)\) is replaced by \((y_k-y_{k-1})/h\), where \(h\) is the sampling interval, so \(y_{k-1}\) becomes an additional state. The paper states that the fast eigenvalues of the exact lifted sampled model converge to \(-\operatorname{spec}(CBK_D)\), where \(\operatorname{spec}(M)\) is the set of eigenvalues of square matrix \(M\), and that away from boundary spectra stability is equivalent to ideal-loop stability plus Schur stability of \(CBK_D\). Here Schur stability means \(\rho(CBK_D)<1\), with spectral radius \(\rho(M)=\max_{\lambda\in\operatorname{spec}(M)}|\lambda|\). For a neural controller, use a margin \(\rho(CBK_D)\le 1-\varepsilon\) and penalize violations.

## Key formulas

- $$\dot{x}(t)=Ax(t)+Bu(t),\qquad y(t)=Cx(t)$$
- $$u(t)=-K_{P}y(t)-K_{I}z(t)-K_{D}\dot{y}(t),\qquad\dot{z}(t)=y(t)$$
- $$\operatorname{spec}_{\mathrm{fast}}(h)\longrightarrow-\operatorname{spec}(CBK_D),\qquad \rho(CBK_D)<1$$
- $$u_k=\pi_\theta(y_k)-K_D\frac{y_k-y_{k-1}}{h},\qquad \mathcal{L}_{\mathrm{stab}}=\left[\max\left(0,\rho(CBK_D)-(1-\varepsilon)\right)\right]^2$$

## Implementation notes

Integrate this at the action head of a neural feedback policy in an RL or model-based control system. Let the policy receive the current observation \(y_k\), maintain the previous observation \(y_{k-1}\) in its recurrent controller state, and output a base action \(\pi_\theta(y_k)\). Add a trainable derivative gain matrix \(K_D\) and form the action as \(u_k=\pi_\theta(y_k)-K_D(y_k-y_{k-1})/h\); optionally clip the final action to the actuator range. If a differentiable world model is available, estimate local plant Jacobians \(B_k=\partial u_{k+1}/\partial u_k\) and \(C_k=\partial y_{k+1}/\partial x_k\), or use a local linear system-identification model. At every training step compute \(M_k=C_kB_kK_D\). Estimate \(\rho(M_k)\) by power iteration: initialize a random vector \(v\), repeatedly set \(v\leftarrow M_kv/\|M_kv\|_2\), and use \(\hat\rho=\|M_kv\|_2\) after 5-10 iterations; for non-normal matrices, use several random initial vectors and take the maximum estimate. Add \(\lambda_{stab}[\max(0,\hat\rho-(1-\varepsilon))]^2\) to the policy loss, with \(\varepsilon\in[0.05,0.2]\). The paper supplies the fast-mode mechanism and Schur constraint; Jacobian estimation, the margin, and the differentiable penalty are engineering adaptations. First test on a linear or lightly nonlinear cart-pole or quadrotor benchmark with a two-layer MLP policy, comparing a finite-difference derivative branch with and without the spectral penalty across sampling intervals \(h\). Measure episode-return variance, action oscillation, instability rate, and performance after abruptly reducing \(h\). Success is fewer divergent rollouts and lower high-frequency action energy at equal return, while the unregularized controller should show failures correlated with \(\hat\rho(CBK_D)\ge 1\).

## Disclaimer

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