# Network-Respecting Controllability Regularizer

- ID: 2855
- Canonical URL: https://synthcore.org/idea/2855/network-respecting-controllability-regularizer
- API JSON: https://synthcore.org/api/idea/2855.json
- API Markdown: https://synthcore.org/api/idea/2855.md
- Verification status: unverified
- Source: [arXiv:2608.29871](https://arxiv.org/abs/2608.29871)
- Category: regularization
- Solves: stability, accuracy, generalization
- ML areas: ssm, graph-nn, regularization
- Math tags: representation-theory, linear-algebra, dynamical-systems, graph-theory
- Ratings: usefulness 6/10; difficulty 5/10; novelty 8/10

## Idea description

Apply the paper's quiver-semistability viewpoint to a graph-structured state-space layer, treating each node's latent state space as a quiver vertex and each message-passing or coupling matrix as an arrow. Penalize approximately invariant collections of node subspaces that receive little signal from the input, so the learned latent dynamics cannot hide useful information in unreachable subnetworks. A dual output-side penalty can prevent predictive information from becoming confined to unobservable subnetworks.

## Mathematical statement

The networked system is represented by vertex state spaces V_i, local linear maps A_i:V_i    o V_i, interconnection maps M_{ij}:V_j    o V_i, input maps B_i:U_i    o V_i, and output maps C_i:V_i    o Y_i. A subrepresentation is a collection of subspaces W_i\subseteq V_i satisfying A_iW_i\subseteq W_i and M_{ij}W_j\subseteq W_i for every local or interconnection arrow. King/Bader semistability tests such collections using a weight \sigma=(\sigma_i): the total weight must satisfy \sum_i\sigma_i\dim V_i=0, while every proper subrepresentation obeys \sum_i\sigma_i\dim W_i\le 0; the opposite inequality is used if the paper's sign convention is reversed. The paper's underlying controllability object is the standard finite-horizon matrix [B,AB,A^2B,\ldots]; for a structured system the analogous global matrix is \mathcal C_T=[B,\mathcal AB,\mathcal A^2B,\ldots,\mathcal A^{T-1}B], where \mathcal A is the block matrix assembled from A_i and M_{ij}. The ML surrogate searches for soft subspaces W_i=\operatorname{span}(Q_i), with Q_i\in\mathbb R^{d_i    imes r_i}, and penalizes leakage outside W_i, \|(I-Q_iQ_i^    op)A_iQ_i\|_F^2+\sum_{i,j}\|(I-Q_iQ_i^    op)M_{ij}Q_j\|_F^2, together with weak input reachability \|Q_i^    op B_i\|_F^2. This operationalizes the mathematical obstruction: an invariant subrepresentation receiving no input is a network-respecting uncontrollable component.

## Key formulas

- $$\dot{\mathbf{x}}(t)=A\mathbf{x}(t)+B\mathbf{u}(t),\qquad \mathbf{y}(t)=C\mathbf{x}(t),$$
- $$\mathcal C_T=\begin{bmatrix}B&AB&A^2B&\cdots&A^{T-1}B\end{bmatrix},$$
- $$\sum_i\sigma_i\dim V_i=0,\qquad \sum_i\sigma_i\dim W_i\le 0\quad\text{for every proper subrepresentation }W,$$
- $$\mathcal L_{\mathrm{NR-cont}}=\sum_i\left\|(I-Q_iQ_i^\top)A_iQ_i\right\|_F^2+\sum_{(j\to i)}\left\|(I-Q_iQ_i^\top)M_{ij}Q_j\right\|_F^2+\lambda_B\sum_i\left\|Q_i^\top B_i\right\|_F^2.$$

## Implementation notes

Integrate this into a graph-structured SSM or graph neural network whose hidden state is split into node tensors h_i\in\mathbb R^{d_i}; use learned linear maps A_i for per-node recurrence and M_{ij} for edge messages. Add the regularizer only during training, not inference. Choose a small rank r_i\ll d_i and parameterize Q_i by an unconstrained matrix followed by QR reorthogonalization, or use Q_i=\operatorname{orth}(R_i). At each optimization step, compute the local invariance residuals (I-Q_iQ_i^T)A_iQ_i and (I-Q_iQ_i^T)M_{ij}Q_j. Compute the controllability-side residual Q_i^TB_i; if inputs enter through a shared encoder, B_i is the explicit input projection into node i or its Jacobian with respect to the input. Add alpha times the invariance residuals plus beta times the input residual to the task loss. A more faithful but costlier variant estimates the finite-horizon structured controllability Gramian G_T=\sum_{k=0}^{T-1}\mathcal A^kBB^T(\mathcal A^T)^k and penalizes small eigenvalues of Q_i^TG_TQ_i. The paper supplies the invariant-subrepresentation and weight logic; rank r_i, alpha, beta, horizon T, and Q_i are engineering choices. Start on a 10-node synthetic graph and a graph forecasting benchmark with a 2-layer linear SSM, comparing no regularizer, ordinary global controllability-rank regularization, and this node-wise penalty. Measure validation error, long-horizon rollout stability, smallest controllability-Gramian eigenvalue, and whether useful gradients reach every node. Success is lower long-horizon error and fewer dead-node latent channels at equal parameter count, without materially increasing training time.

## Disclaimer

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