In large networked systems, such as the temperature zones of a building, the buses of a power grid, or a swarm of drones, each controller must decide how much of the network it needs to hear from. A known theoretical result, the spatial decay property, says that in many such systems the best action for one node depends on distant nodes less and less, exponentially so, as the number of network hops between them grows. Because of this, a controller that sees only its neighbors within κ hops can come exponentially close to a fully centralized one. Every method built on this result, however, picks a single radius κ for the whole network from worst-case constants. Those constants are usually unknown, and they are set by the hardest spot in the network.
We propose that each agent certify its own radius while it runs. An agent solves its local predictive-control problem over its current κ-hop window. It then computes, with one extra linear solve, how sensitive its chosen action is to the guessed values at the edge of that window. If the sensitivity is small, the edge does not matter and the agent stops. If it is large, the agent requests one more ring of neighbors. Agents in weakly coupled regions keep tiny windows, and agents near strong coupling or active constraints widen theirs. The rule turns spatial decay from an offline existence proof into an online stopping criterion. It also gives a direct way to measure whether decay survives the hard constraints that current theory excludes.
Core Claim
The known bounds say what happens in the worst case. The analysis of Localized Predictive Control in Decentralized Online Convex Optimization in Networked Systems bounds the error from a wrong boundary guess at κ hops by roughly $C\rho_S^{\kappa}$ times the size of the guess error. Here $\rho_S<1$ is a global spatial decay factor. The paper also warns that $\rho_S \to 1$ when neighbor coupling is strong relative to local costs. Distributed Truncated Predictive Control for Networked Systems under Uncertainty and Near-Optimal Distributed Linear-Quadratic Regulator for Networked Systems give the same kind of worst-case, uniform-κ guarantee.
Our claim is that the quantity these proofs bound can be computed locally, for the agent's actual current window. Let $u_i$ be agent $i$'s first-step action from its truncated problem, and let $z_{\partial}$ be the boundary values it assumes at hop κ. We define the local boundary sensitivity as
$$s_i(\kappa) = \Big\| \frac{\partial u_i}{\partial z_{\partial}} \Big\| \cdot \bar{e}_{\partial},$$
where $\bar e_\partial$ is a bound on how wrong the boundary guess can be. The Jacobian comes from differentiating the optimality conditions of the local convex program. It needs one adjoint solve using only data already in the window, with no extra communication.
The rule is simple: grow $\kappa_i$ until $s_i(\kappa_i) \le \varepsilon$. The hypothesis to prove is that, under the same convexity and graph-growth assumptions as the Localized Predictive Control analysis, per-agent certificates $s_i\le\varepsilon$ imply a global cost gap of order $N\varepsilon$ plus the usual temporal-horizon term. The theorem would replace the worst-case $\rho_S$ with the decay each agent actually measures.
The empirical claim is that, in networks with uneven coupling, adaptive radii reach the same regret as the best uniform κ with substantially fewer messages.
This differs from all existing uses of spatial decay, which fix one κ offline. It also differs from learned multi-agent methods such as GCBF+, whose sensing radius is a fixed, hand-tuned hyperparameter unrelated to any decay measurement.
Research Program
Theory. We first extend the perturbation argument of Localized Predictive Control from a uniform radius to heterogeneous, state-dependent radii. Each agent's window then differs, and windows overlap asymmetrically. We would show that the certified gap composes across agents.
Constraints. Next we handle hard constraints. Box limits on actuators and states, and linearized collision half-spaces for the swarm, keep each local problem convex. They change the local curvature, however, whenever a constraint becomes active. This is exactly the setting the research file flags as uncharacterized: existing decay theory, for example On the Optimal Control of Network LQR with Spatially-Exponential Decaying Structure, covers only unconstrained linear feedback. That paper also constrains our design, because it shows decay can fail without open-loop stability. The certificate must therefore detect non-decay, not assume decay. When $s_i$ stops shrinking as κ grows, the agent flags a region where locality does not help. We treat that as a result to report, not a failure to hide.
Implementation. A single distributed controller implements the rule. Each agent solves a receding-horizon problem over its κ-hop, k-step window, which reuses the Distributed Truncated Predictive Control and Localized Predictive Control formulations. It then computes $s_i$ by one extra solve of the optimality system and either acts or requests the next ring. To avoid radius chatter, the radius shrinks only after several consecutive steps below $\varepsilon/2$.
Accounting. Every message is counted: state exchanges, forecast exchanges, and the ring-expansion requests themselves.
Baselines
The naive reference is fully decentralized control with κ = 0 or 1. It bounds how much communication buys. The upper reference is the centralized receding-horizon controller with full information, which defines regret zero and total communication cost.
The strongest competitor is Distributed Truncated Predictive Control with a uniform radius. We sweep κ from 1 to 6 and tune it by grid search on the same training scenarios. It is the most direct rival because it carries the same stability and exponential-regret guarantees, and it differs from our method only in using a fixed radius. For the 100-zone building network and the IEEE 118-bus grid, Near-Optimal Distributed Linear-Quadratic Regulator for Networked Systems reports only that truncation error drops sharply as κ goes from 1 to 5. For the building network, the truncated-control papers report exponential regret decay without exact figures. So there is no reported number found on Lacuna for regret at a given κ on these benchmarks, and we will reproduce the fixed-κ curve ourselves. Beating this baseline means lying strictly below its Pareto front of regret against messages. Concretely, we would need at least 30% fewer messages at equal regret on heterogeneous-coupling instances, and no worse performance on uniform ones.
For the constrained setting we add Distributed and Localized Model Predictive Control via System Level Synthesis (Part II: Theoretical Guarantees). It is included because it handles constraints with feasibility guarantees, at a structural locality chosen in advance. That paper reports optimal cost differences on the order of $10^{-5}$ relative to unconstrained DLMPC, which is a cost-of-terminal-set measure rather than a regret-versus-communication number. We match its locality parameter to our median κ and compare regret and messages.
For the mobile swarm we compare against Distributed Differential Dynamic Programming Architectures for Large-Scale Multiagent Control. It reaches near-centralized quality through many rounds of neighbor consensus, and reports about 2 minutes against over 2 hours for centralized sequential quadratic programming on 1,024 cars. We also compare against GCBF+, which reports over 98% success at 256 agents.
All methods receive the same scenarios, the same horizon k, the same solver, and the same wall-clock budget per control step.
Two ablations isolate the key ingredient. The first replaces the measured sensitivity with the theoretical worst-case bound, which should recover uniform κ. The second is described under the proximity heuristic below.
Why not just…?
Why not just use a larger fixed κ everywhere? On uniformly coupled networks this is probably as good as our method, and we will say so. The proposal only pays off where coupling is uneven, which is why every benchmark includes both uniform and heterogeneous variants. If the gains vanish on heterogeneous instances too, the idea is wrong.
Why not just expand the radius whenever a neighbor is close or a constraint is active? This cheap event-triggered heuristic is the real threat. It captures the intuition that choke points need more information. We run it as the second ablation, tuned on the same budget. The sensitivity certificate should win in two situations. The first is strong coupling with no active constraint, such as tightly coupled building zones. The second is an active constraint that nonetheless does not propagate influence far. If the heuristic matches the certificate across both, we will report that the certificate's value is only its theoretical guarantee.
Why not just learn the radius or an attention mask, as graph-network controllers do? Learned policies such as GCBF+ and the infinite-horizon graph controller in Learning Distributed Safe Multi-Agent Navigation via Infinite-Horizon Optimal Graph Control scale impressively. The latter reports above 90% safety with 320 agents. Neither provides a regret bound, and benchmarking in POGEMA shows learned decentralized policies degrade on out-of-distribution maps. Our swarm experiment tests maps structurally different from the training ones. That is where a certificate computed from the model, rather than learned from data, should keep working.
Proposed Experiments
Building network. We construct the 100-zone building network and the IEEE 118-bus grid following the setups described in the spatial-decay papers. We then create heterogeneous variants in which one region has coupling five to ten times stronger, or actuator limits that become active under a load spike.
Swarm task. We add a mobile position-swap task through a narrow corridor using double-integrator agents with linearized collision constraints, at 64 to 1,024 agents. Graph distance changes over time in this setting, as noted in the research file's discussion of proximity graphs.
Metrics. We measure regret against the centralized controller, total and peak messages per step, the distribution of $\kappa_i$ over space and time, constraint violations, and swarm success rate.
Decisive checks. The first is whether the adaptive method's regret-versus-messages curve dominates the fixed-κ front on heterogeneous instances. The second is whether the measured $s_i(\kappa)$ curves decay geometrically near active constraints. This is the first empirical evidence on whether spatial decay survives active inequality constraints, and it is informative whichever way it comes out. The third is whether the certificate beats the proximity heuristic in the strongly-coupled-but-unconstrained regime. Failing the first check would falsify the proposal. Failing only the third would narrow it to its guarantee.
References
- Decentralized Online Convex Optimization in Networked Systems
- Distributed Truncated Predictive Control for Networked Systems under Uncertainty: Stability and Near-Optimality Guarantee
- Near-Optimal Distributed Linear-Quadratic Regulator for Networked Systems
- GCBF+: A Neural Graph Control Barrier Function Framework for Distributed Safe Multiagent Control
- On the Optimal Control of Network LQR with Spatially-Exponential Decaying Structure
- Distributed and Localized Model Predictive Control - Part II: Theoretical Guarantees
- Distributed Differential Dynamic Programming Architectures for Large-Scale Multiagent Control
- Learning Distributed Safe Multi-Agent Navigation via Infinite-Horizon Optimal Graph Control
- POGEMA: A Benchmark Platform for Cooperative Multi-Agent Navigation