Swarms That Speak Less, Track Better: The Mathematics of Sparse Coordination

In 2019, a fleet of autonomous delivery drones needed to follow a lead drone through a city. Each drone had limited radio bandwidth. Continuous communication would drain batteries in minutes. But sparse communication risked the swarm drifting apart or missing course corrections. The engineers faced a tradeoff that every designer of networked robots confronts: how do you guarantee precise coordinated behavior while minimizing the messages flying through the air?

A new control theory paper tackles exactly this problem—and solves it more elegantly than previous approaches. The result isn't just a smarter communication scheme. It's a mathematical guarantee that every drone will stay within a designer-specified performance envelope, even when transmissions are infrequent and asynchronous. The authors call it prescribed performance leader-following consensus under event-based broadcasting (Gkesoulis et al., 2026).

The Science

Multi-agent consensus is the problem of getting a network of agents—robots, drones, vehicles—to agree on a common state or trajectory. "Leader-following" means one agent sets the target and the others must track it. "Prescribed performance" means the tracking error must stay within a time-varying envelope that the designer specifies in advance. You might want the swarm to converge within 0.5 meters of the leader within 10 seconds, then stay within 0.1 meters thereafter.

The twist is communication. In theory, every agent could broadcast its state continuously. In practice, this wastes bandwidth and power. Event-based broadcasting offers a middle path: each agent transmits only when its current state differs too much from its last transmission. The threshold for "too much" is set locally by each transmitter, using only its own current state and its own last broadcast value. No agent needs to know the network topology, the leader's global position, or any global spectral information.

The researchers—based at the National Technical University of Athens, the University of Thessaly, and the University of Patras—consider a directed graph containing a spanning tree rooted at the leader. This is more general than many prior results, which required symmetric (undirected) communication links where information flows both ways equally. In many real scenarios—underwater acoustic networks, directional radio, adversarial environments—links are one-way. A drone might receive from the leader but not vice versa.

Their framework handles a time-varying leader with bounded velocity, which matters for real targets that accelerate and maneuver.

What They Found

The core result is Theorem 6, which establishes four properties simultaneously for the closed-loop system:

Well-posedness: The system admits unique, complete, non-Zeno solutions. "Non-Zeno" means no agent transmits an infinite number of messages in finite time—a practical necessity, since infinite chattering would defeat the purpose of sparse communication. This is mathematically guaranteed, not just observed in simulations.

Boundedness: All signals remain bounded. The follower states, their control inputs, their sampled disagreements with neighbors—nothing diverges.

Prescribed sampled performance: For each follower, the sampled neighborhood disagreement—the error computed from held incoming samples rather than live data—satisfies

$$|{\hat{e}}_i(t)|<{\rho }_i(t)+s_i(t)$$

where ${\rho }_i(t)$ is the prescribed performance function (a decaying exponential envelope) and $s_i(t)$ is a receiver-side modification term that temporarily expands the envelope.

Prescribed continuous performance: The true continuous-time error is similarly bounded:

$$|e_i(t)|<{\rho }_i(t)+{\Delta }_i+s_i(t)$$

where ${\Delta }_i$ captures the worst-case discrepancy between held samples and current states, depending on the communication thresholds of all incoming neighbors.

The leader-following error—the distance from each follower to the leader—gets its own explicit bound:

$$\Vert z(t)-1_N{z}_0(t){\Vert }_2\le \frac{1}{{\sigma }_{\min }(W)}\Vert \rho (t)+\Delta +\bar{s}{\Vert }_2$$

where ${\sigma }_{\min }(W)$ is the smallest singular value of the graph coupling matrix and $\bar{s}$ is the maximum modification term each follower can accumulate. This bound quantifies exactly how communication sparsity affects tracking precision.

The simulations—run with 5 followers tracking a sinusoidal leader over a directed graph—demonstrate these guarantees in action. Figure 2 shows the leader and follower trajectories converging and remaining close. Figure 3 shows the corresponding control inputs, which are bounded and eventually decay as the swarm settles.

Figure 4 is particularly compelling: it plots the maximum sampled and continuous normalized neighborhood-error magnitudes across all followers. Both stay below 1—the prescribed boundary—throughout the simulation. The envelope tightens over time as ${\rho }_i(t)$ decays toward its steady-state value ${\rho }_i^{\infty }$.

Why This Changes Things

Previous event-triggered prescribed performance control schemes required triggers that depended on edge dynamics, relative quantities between agents, neighboring control signals, or network-wide bounds. In other words, they still assumed agents could peek at their neighbors' states. The new framework needs none of this. Each transmitter decides when to broadcast using only its own current state and its own last transmission. This is truly decentralized.

The receiver-side performance modification term is the key conceptual innovation. When a follower receives new samples asynchronously, its computed error can jump discontinuously—potentially crossing the prescribed performance boundary. Rather than tightening control to prevent this (which would require faster, more aggressive responses), the researchers let the boundary expand temporarily. The term $s_i(t)$ absorbs the jump, then decays exponentially back toward zero. The performance envelope flexes to accommodate communication reality without compromising the guarantee.

Figure 5 illustrates this for follower 4: the continuous neighborhood error $e_4(t)$ stays within the envelope ${\rho }_4+{\Delta }_4+s_4$ at all times. When adverse receptions cause jumps, $s_4$ spikes and decays.

Figure 6 shows all four performance modification terms across the simulation. Each positive jump corresponds to an adverse reception—an update that increased the error magnitude. The decay rate ${\theta }_i$ controls how quickly the term returns to baseline. The designer can tune this tradeoff: faster decay means tighter performance but more aggressive correction cycles; slower decay means smoother envelopes but looser bounds.

The explicit topology-dependent bounds are especially valuable for design. The term ${\Delta }_i={\sum }_{j\in {\mathcal{Q}}_i}{a}_{ij}{c}_j+b_i{c}_0$ shows exactly how each incoming link's weight and threshold contributes to the worst-case error. A designer can allocate communication budget strategically: tighter thresholds on high-weight links yield larger performance gains.

What's Next

The paper focuses on single-integrator dynamics—agents whose state is position and whose control input is velocity. Extending to double-integrator or more complex dynamics would broaden applicability to physical systems with inertia. The assumption of continuous leader trajectory (Assumption 2) excludes scenarios where the leader teleports or resets; relaxing this would matter for target tracking with intermittent observations.

The framework currently treats each dimension independently (Remark 4), which works but may be conservative. A truly multi-dimensional analysis could tighten bounds for systems where errors in different directions are correlated.

Practical validation remains. The theoretical results assume ideal communication with no packet loss or delay. Real wireless networks introduce imperfect channels that could invalidate the error jump model. Adapting the receiver-side modification term to compensate for dropped packets is a natural next step.

For engineers building swarm systems, the paper offers something rare: a principled way to trade communication for performance without guessing. Set your thresholds based on available bandwidth. The mathematics tells you exactly how tracking precision degrades. Tighten thresholds where precision matters most. The swarm converges to the leader, stays within the envelope, and speaks only when necessary.

That's a small thing—until you scale to hundreds of drones running on batteries, or underwater vehicles where every transmitted bit costs energy you might not recoup.

Paper available on arXiv: Gkesoulis, Psillakis, Karras, and Bechlioulis (2026)