{
 "artifacts": null,
 "category": "dynamics",
 "description": "Cluster recurrent modules or MoE experts by the geometry of their observed finite-horizon input-output behaviors rather than by parameter distance. Train one shared optimizer/controller or low-rank adapter per cluster while retaining module-specific parameters and routing. This should reduce control and optimizer overhead without merging modules whose temporal responses are dynamically incompatible.",
 "formulas_latex": [
  "$$\\mathcal{G}_i=\\operatorname{im}(B_i),\\qquad B_i=\\operatorname{col}\\big(H_L(u_i),H_L(y_i)\\big),\\qquad P_i=Q_iQ_i^{\\mathsf T},$$",
  "$$\\operatorname{gap}^{\\mathcal{G}}_L(i,j)=\\|P_i-P_j\\|_2=\\max\\left\\{\\|P_j^{\\perp}P_i\\|_2,\\|P_i^{\\perp}P_j\\|_2\\right\\},\\qquad P_i^{\\perp}=I-P_i,$$",
  "$$\\ell_k=\\arg\\min_{i\\in\\mathcal{C}_k}\\operatorname{gap}^{\\mathcal{G}}_L(i,\\ell_k),\\qquad \\mathcal{C}_k=\\left\\{i:\\operatorname{gap}^{\\mathcal{G}}_L(i,\\ell_k)\\leq\\varepsilon\\right\\},$$",
  "$$\\mu_i=\\sigma_{\\min}\\!\\left([\\,Q_{\\mathcal{G}_i}\\;Q_{\\mathcal{G}(C_k)}\\,]\\right),\\qquad \\text{accept follower }i\\text{ only if }\\mu_i\\geq\\mu_{\\min}.$"
 ],
 "id": 3156,
 "implementation": "Use this at the level of recurrent cells, SSM layers, or MoE experts. During a calibration pass, feed every module the same batch of excitation sequences, record its hidden-state or output sequence, and build a block Hankel matrix with L consecutive time windows. Reset hidden states between sequences so that the measured response is comparable. For each module, compute B_i=col(H_L(u_i),H_L(y_i)), take a rank-r SVD B_i=U_i Sigma_i V_i^T, and set P_i=U_iU_i^T. Cluster modules greedily using ||P_i-P_leader||_2; share the leader's optimizer moments, controller, or low-rank adapter with followers. Recompute clusters every E epochs and use hysteresis so a module must violate epsilon for two recalculations before moving. The paper supplies the projector-gap construction; rank r, epsilon, mu_min, and finite-horizon responses must be estimated empirically. First experiment: train 8-16 small GRUs on sequential MNIST or permuted-MNIST with varied recurrent spectral radii. Compare independent Adam states against four behavior-gap clusters sharing optimizer moments or adapters. Plot validation loss, hidden-state norms, trajectory error, and mu_i versus gap. The predicted signature is a sharp increase in error and exploding hidden norms when gap exceeds epsilon or mu_i approaches zero. Repeat for L=16,32,64; the transition should move with the observation horizon.",
 "math_summary": "For module i, collect zero-state input-output trajectories over horizon L and construct a behavior matrix B_i=col(H_L(u_i),H_L(y_i)), where H_L is a block Hankel operator and u_i,y_i are input and output sequences. Let G_i=im(B_i) and P_i be the orthogonal projector onto G_i. The paper's similarity metric is gap_L^G(i,j)=||P_i-P_j||_2=max{||P_j^perp P_i||_2,||P_i^perp P_j||_2}. Small gap means that the systems have similar control-relevant trajectory subspaces. Assign module i to leader l_k only when this gap is at most epsilon. A practical compatibility margin is mu_i=sigma_min([Q_Gi Q_G(Ck)]), where Q_Gi and Q_G(Ck) are orthonormal bases for the module behavior graph and shared controller graph; require mu_i to remain above a chosen threshold.",
 "math_tags": [
  "control-theory",
  "linear-algebra",
  "spectral-theory",
  "geometry"
 ],
 "ml_areas": [
  "rnn",
  "ssm",
  "moe-routing",
  "optimizer"
 ],
 "paper": {
  "arxiv_id": "2609.03921",
  "arxiv_url": "https://arxiv.org/abs/2609.03921",
  "summary_what_math_gives_to_ml": "The paper provides a constructive mechanism for scaling control across many heterogeneous systems: represent each system by a finite-horizon input-output behavior subspace, cluster systems using the operator-norm gap between orthogonal projectors, and deploy one leader controller per cluster. The transferable asset is not merely clustering, but a control-relevant similarity metric with a geometric well-posedness interpretation: small graph-subspace gaps imply that a controller validated on a leader is less likely to become ill-posed on its followers. In neural networks, the same construction can cluster recurrent states, SSM layers, or MoE experts by their finite-horizon input-response subspaces and share an optimizer, adapter, or controller within each cluster. The key falsifiable prediction is that performance degradation and instability increase sharply when the within-cluster subspace gap crosses a measurable threshold.",
  "title": "Data-Based Clustering and Control of Similar Biological Systems",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 8,
  "usefulness": 8
 },
 "solves": [
  "stability",
  "scalability",
  "speedup"
 ],
 "title": "Behavior-Gap Clustered Neural Controllers",
 "url": "https://synthcore.org/idea/3156/behavior-gap-clustered-neural-controllers",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)",
  "verification_axes": {
   "benchmark_mechanism": {
    "confirmed": null,
    "tested": false
   },
   "practical_benchmark": {
    "beats_baseline": null,
    "tested": false
   },
   "toy_mechanism_gate": {
    "confirmed": null,
    "tested": false
   }
  }
 }
}