# Relative-Degree-Gated Passive Neural State Space

- ID: 2830
- Canonical URL: https://synthcore.org/idea/2830/relative-degree-gated-passive-neural-state-space
- API JSON: https://synthcore.org/api/idea/2830.json
- API Markdown: https://synthcore.org/api/idea/2830.md
- Verification status: unverified
- Source: [arXiv:2608.29474](https://arxiv.org/abs/2608.29474)
- Category: architecture
- Solves: stability, accuracy
- ML areas: ssm, rnn, training-dynamics, regularization
- Math tags: control-theory, dynamical-systems, linear-algebra, optimization
- Ratings: usefulness 6/10; difficulty 6/10; novelty 7/10

## Idea description

Construct a neural state-space model with an explicit first-order input-to-output path instead of forcing every output to depend only on deeply propagated hidden states. Penalize or reject learned linearizations whose transfer matrix has relative degree greater than one, then train a storage-function certificate for the remaining passive dynamics. This preserves the paper's relative-degree compatibility condition while allowing high-order internal memory.

## Mathematical statement

For a local neural state-space linearization \(\dot{x}=Ax+Bu\), \(y=Cx+Du\), the transfer matrix is \(G(s)=C(sI-A)^{-1}B+D\), where \(x\) is the hidden state, \(u\) the external input, \(y\) the output, and \(A,B,C,D\) the Jacobian blocks at an operating point. If \(D=0\), the relative degree is \(r=\min\{k\geq 1:CA^{k-1}B\neq 0\}\). Standard input-output passivity with supply rate \(u^{\mathsf T}y\) requires a nonzero passive rational transfer to have \(r\leq 1\); hence a model with \(CAB\neq 0\) but \(CB=0\) cannot satisfy the ordinary positive-real condition. A quadratic storage \(V(x)=\tfrac12x^{\mathsf T}Px\), with \(P=P^{\mathsf T}\succeq0\), is certified locally when the KYP matrix inequality \(\begin{bmatrix}A^{\mathsf T}P+PA & PB-C^{\mathsf T}\\ B^{\mathsf T}P-C & -(D+D^{\mathsf T})\end{bmatrix}\preceq0\) holds. The upper-left and off-diagonal blocks guarantee \(\dot V\leq u^{\mathsf T}y\), while the relative-degree gate prevents optimizing toward an impossible certificate.

## Key formulas

- $$G_i(s)={\rm diag}\left(-\frac{\Delta\theta_i}{\Delta P_i},-\frac{\Delta V_i}{\Delta Q_i}\right)$$
- $$G(s)=C(sI-A)^{-1}B+D,\qquad r=\min\{k\geq1:CA^{k-1}B\neq0\}\ \text{when }D=0$$
- $$\begin{bmatrix}A^{\mathsf T}P+PA & PB-C^{\mathsf T}\\ B^{\mathsf T}P-C & -(D+D^{\mathsf T})\end{bmatrix}\preceq0,\qquad V(x)=\tfrac12x^{\mathsf T}Px$$
- $$\mathcal{L}_{\rm rd}=\max\left(0,\tau-\|CB\|_F\right)^2+\gamma\|D\|_F^2\quad\text{for an intended first-order/direct pathway, with }\tau>0$$

## Implementation notes

1. Exact integration point: use a continuous-time neural state-space or recurrent model with \(\dot{x}=f_\theta(x,u)\), \(y=h_\theta(x,u)\). Implement \(h_\theta(x,u)=h_0(x)+D_\theta u\), or add a separate first-order pathway \(z=K_\theta(x)u\) to the output, while keeping long memory in \(x\). At sampled time steps, integrate the state with RK4 or an implicit solver. 2. Pseudocode: for each operating point \((x,u)\), compute \(A=\partial f/\partial x\), \(B=\partial f/\partial u\), \(C=\partial h/\partial x\), and \(D=\partial h/\partial u\) using automatic differentiation; add the relative-degree penalty; periodically solve the semidefinite feasibility problem for \(P\succeq0\) under the KYP inequality; if infeasible, increase the feedthrough or first-order pathway or reject the checkpoint. The training loss is task loss plus \(\lambda_{\rm rd}\mathcal{L}_{\rm rd}\) and, when a feasible \(P\) exists, a passivity residual based on the largest eigenvalue of the KYP matrix. 3. Computed versus estimated: the relative-degree quantities are computed from Jacobians, while passivity is only certified locally at sampled states; estimate global behavior by evaluating a grid or replay-buffer batch of operating points. 4. First cheap experiment: compare a vanilla GRU/neural-ODE, the same model with a learned direct pathway, and the gated model on damped second-order system identification and a small Lorenz or pendulum dataset. Sweep hidden-state dimension and pathway gain. The predicted signature is sharp: models with numerical \(\|CB\|_F\) near zero and nonzero \(CAB\) should have KYP infeasibility across all \(P\), while adding a pathway producing \(\|CB\|_F>\tau\) should create a feasible region. Measure the smallest KYP residual versus pathway gain and test whether the transition occurs within 20% of the gain where the estimated first Markov parameter \(CB\) rises above the numerical tolerance.

## Disclaimer

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