Where Should a Local Controller Look? State-Dependent Local Centrality for Truncated Control of Nonlinear Networks

Large physical networks such as power grids, building climate systems and neural circuits are increasingly controlled by many local controllers. Each one plans only over a small patch of the network around itself, because no single computer can solve the whole problem in real time. The key design choice is how big each patch should be. For linear systems, theory says a uniform patch radius works: a node's best action depends on far-away nodes with an influence that shrinks exponentially in graph distance. For nonlinear systems no such guarantee exists. How strongly two nodes couple depends on the current operating point, so a line that barely matters at light load can become the bottleneck that carries a disturbance across the grid when the system is stressed.

We propose that each local controller choose its own patch radius at each moment. The choice is based on a centrality score that it computes from its two-hop neighborhood. The score is not computed on the bare wiring diagram. It is computed on a "sensitivity graph" whose edge weights are the current local coupling strengths of the nonlinear dynamics. In two sentences: network science has learned to estimate which nodes are global bridges from small neighborhoods, and nonlinear control needs to know which nodes currently relay influence far. We test whether a bridge score computed locally and recomputed as the operating point moves can tell a controller how far it must look.

Core Claim

The first direction, localized centrality estimation, shows that global importance can often be read off small neighborhoods. Ego-betweenness counts shortest paths only inside a node's one-hop neighborhood. Efficient Top-k Ego-Betweenness Search reports that its top-$k$ nodes overlap with global betweenness by 60% to 80% on large real graphs. Latent Geometry-Driven Network Automata for Complex Network Dismantling uses common-neighbor dissimilarity, the idea that an edge whose endpoints share few neighbors spans a long latent distance. With it, the authors match global betweenness at dismantling networks across the 1,475-network ATLAS suite. Structural sparsity of complex networks explains why such local reads are trustworthy on sparse, clustered real-world graphs. All three papers are purely topological, and the field's own open problems note that their definitions ignore the state-dependent behavior of physical processes running on the graph.

The second direction, generalization to nonlinear systems, supplies the method we extend and the gap we target. Distributed Truncated Predictive Control for Networked Systems under Uncertainty has each node solve a predictive control problem restricted to its $\kappa$-hop neighborhood. It proves that the gap to centralized control (the regret) shrinks exponentially in $\kappa$. The proofs are for linear time-invariant systems only, the radius is uniform across nodes, and the authors leave even time-varying linear dynamics open.

Our claim has three parts. First, along a nonlinear trajectory, the spatial decay of a node's optimal-action sensitivity is governed by the Jacobian-weighted graph at the current state, not by the raw topology. Second, a bridge score computed on that weighted graph within two hops predicts the radius each node needs. Third, choosing radii from this score gives the regret of a large uniform radius at a fraction of the computation. What is new is the combination of three things: a per-node radius, a radius that changes over time with the state, and a radius set by a locally computed centrality. None of the cited works does all three.

Research Program

The central object is a weighted graph. At the current state $x$, an edge $(i,j)$ carries the weight

$$w_{ij}(x) = \left\| \frac{\partial f_i}{\partial x_j}(x) \right\| \Big/ \left\| \frac{\partial f_i}{\partial x_i}(x) \right\|,$$

where $f_i$ is node $i$'s dynamics. The weight is node $i$'s sensitivity to neighbor $j$ relative to its own damping. Each node can compute its incoming weights from its own model or from short local excitation data, so no global information is required.

On this graph we compute two local scores per node:

  • an ego-betweenness in which shortest paths use edge lengths $-\log w_{ij}$, so strong couplings count as short;
  • a weighted common-neighbor dissimilarity score.

Both need only two-hop communication. The radius rule maps the score $s_i(x)$ to

$$\kappa_i(x) = \kappa_{\min} + \left\lceil \frac{\log(1/\varepsilon)}{-\log \hat\rho_i(x)} \right\rceil.$$

Here $\varepsilon$ is a target sensitivity tolerance. The local decay-rate estimate $\hat\rho_i(x)$ comes from a calibrated monotone fit of measured decay against $s_i(x)$, learned on training graphs. The logarithmic form follows directly from assuming exponential spatial decay, $\text{sensitivity} \propto \rho^{d}$, as in the linear theory. Radii are recomputed whenever the score changes by more than a threshold. Hysteresis prevents a node's radius from chattering between values.

The theoretical work is deliberately modest. We linearize along the predicted trajectory to get a time-varying linear system. We then aim to prove spatial decay with distance measured on the weighted graph, under a contraction-type condition that the weighted out-degree stays below one. This would extend the linear argument of the truncated-control paper to the time-varying case it leaves open. The nonlinear error is handled as a bounded disturbance inside the existing input-to-state stability argument.

We do not claim a guarantee for regimes where nonlinear coupling creates genuinely nonlocal behavior, such as synchronization waves. We treat those regimes as stress tests.

Baselines

Every method uses the same solver: sequential convex programming warm-started from the previous step. Every method also gets the same prediction horizon, the same disturbance realizations, and the same wall-clock budget per control step. We plot regret against total neighborhood size and against compute, so no method can win by simply using more computation.

Reference point. Centralized nonlinear predictive control over the whole network defines zero regret. On the 26-area European network benchmark this reference is only feasible offline. A Benchmark for the Application of Distributed Control Techniques to the Electricity Network of the European Economic Area reports over 206 hours of compute to simulate 24 hours of centralized operation.

Naive baseline. Uniform-radius truncated control, applied unchanged to the nonlinear dynamics. We sweep $\kappa$ from 1 to 5. Beating it means reaching the regret of its best uniform $\kappa$ with at least 30% fewer total neighborhood nodes. Alternatively, it means achieving lower regret at equal total neighborhood size.

Strongest competitor. The a posteriori radius check. Each node solves at radius $\kappa$ and $\kappa+1$ and grows its radius while its first action still changes by more than $\varepsilon$. This directly measures what our score only predicts. Its weakness is cost: it needs at least two solves per node per step, and many more for bridge nodes. Our method beats it only if it matches its regret within 10% at lower compute per step. If the check is as cheap in practice, our method loses. No reported result was found on Lacuna for this check on any of these benchmarks, and none for uniform truncation either: the truncated-control paper shows regret decay only in figures, on its HVAC model.

Ablations. These isolate the key ingredient:

  • Static topological score. Ego-betweenness and dissimilarity computed on the unweighted graph. This tests whether state dependence matters.
  • Degree. The cheapest local proxy.
  • Oracle. Exact global betweenness on the weighted graph. This bounds what the two-hop restriction costs.

Why not just…?

Why not just use a larger uniform radius? Regret falls exponentially in $\kappa$, but each local problem grows with the number of nodes inside the radius. On scale-free or meshed grids that number grows very fast. The sweep against uniform $\kappa$ answers this question directly. If the regret-versus-neighborhood-size curve of adaptive radii does not dominate the uniform curve, the idea fails.

Why not just use degree or the static wiring diagram? Evidence from both literatures says topology alone is incomplete:

The static-score ablation is the decisive test. If static ego-betweenness predicts required radii as well as the weighted version across load levels, then the nonlinear part of our proposal is unnecessary. In that case the honest finding is "topological local centrality suffices."

Why not just measure the needed radius directly by trial solves? This is the real threat. Measurement does not have to predict anything. Our answer is cost: trial solves multiply computation exactly at the bridge nodes where problems are largest. A score computed in two hops costs almost nothing by comparison. We also test a hybrid in which the score sets the starting radius and a single check confirms it. If the hybrid wins, that is still a positive result for the score as a prior.

Why not just linearize once and run linear truncated control? Nonlinear Control of Networked Dynamical Systems shows that multistable networks must be steered across basins of attraction, which a fixed linearization cannot represent. We include a fixed-linearization baseline on the Hopfield and swing-equation tasks to measure this failure.

Proposed Experiments

Linear sanity check. On the synthetic HVAC model from the truncated-control paper, where the weighted graph is constant, the adaptive rule should reproduce the uniform-radius behavior. This shows the method does no harm in the linear case.

Nonlinear test beds. These carry the main evaluation:

  • Swing-equation (Kuramoto-type) dynamics on power-grid topologies drawn from ATLAS and on the IEEE 118-bus layout used in Grid2Op. Load is swept from light to near-collapse.
  • The European area benchmark, with nonlinear swing terms added to its piecewise-affine model.
  • 400-node Hopfield networks following the nonlinear networked-control paper. These test regimes where coupling saturates.

Decisive checks. There are three:

  1. A per-node rank correlation between the two-hop weighted score and the radius each node actually needs, measured by trial solves. This is compared against the static score across load levels. The prediction is that the gap widens with stress.
  2. Regret-versus-compute curves against all baselines.
  3. Constraint violations over 1,000 Monte Carlo disturbance runs. This mirrors the evaluation protocol of Dynamic Tube MPC for Nonlinear Systems, which reports zero tube exits.

We report where the predictor breaks down, namely in synchronization-dominated regimes and on dense expander-like graphs, as a result in its own right.