Abstract
This paper introduces a new lifting method for analyzing convergence of continuous-time distributed synchronization/consensus systems on the unit sphere. Points on the d-dimensional unit sphere are lifted to the (d+1)-dimensional Euclidean space. The consensus protocol on the unit sphere is the classical one, where agents move toward weighted averages of their neighbors in their respective tangent planes. Only local and relative state information is used. The directed interaction graph topologies are allowed to switch as a function of time. The dynamics of the lifted variables are governed by a nonlinear consensus protocol for which the weights contain ratios of the norms of state variables. We generalize previous convergence results for hemispheres. For a large class of consensus protocols defined for switching uniformly quasi-strongly connected time-varying graphs, we show that the consensus manifold is uniformly asymptotically stable relative to closed balls contained in a hemisphere. Compared to earlier projection based approaches used in this context such as the gnomonic projection, which is defined for hemispheres only, the lifting method applies globally. With that, the hope is that this method can be useful for future investigations on global convergence.
Introduction
In multi-agent systems—such as swarms of drones, satellite constellations, or biological flocks—"consensus" refers to agents reaching a common state. When this state represents a direction or heading, the agents evolve on a unit sphere ($S^{d-1}$). Analyzing synchronization on a sphere is inherently more complex than in Euclidean space due to the sphere's curvature and non-trivial topology. Standard linear consensus tools do not apply globally.
Historically, researchers have employed "projections" (e.g., gnomonic projections) to map the sphere onto a flat plane. However, these methods are typically restricted to a single hemisphere; if agents are distributed beyond this range, the mathematical framework fails. This paper addresses the need for a global analytical tool capable of handling diverse agent distributions and time-varying network topologies.
Methodology: The Lifting Method
The core innovation presented is the Lifting Method. Rather than projecting the $(d-1)$-dimensional sphere onto a lower-dimensional flat surface, the authors "lift" the states into a $d$-dimensional Euclidean space ($\mathbb{R}^d$).
The transformation associates an agent's position $x_i$ on the sphere with a variable $z_i$ in the ambient space. The authors derive a nonlinear consensus protocol in this higher-dimensional space that is mathematically equivalent to the standard protocol on the sphere, where agents move toward the weighted average of their neighbors in their respective tangent planes. In the lifted space, interaction weights become rational functions involving the norms (lengths) of the state variables. This framework accounts for "switching" networks—where communication links appear and disappear—provided the graph remains "uniformly quasi-strongly connected" over time.
Analysis and Results
The authors validate their theoretical framework through formal proofs and numerical simulations. A key finding is that the consensus manifold is uniformly asymptotically stable relative to any closed ball contained within an open hemisphere under switching topologies.
The lifting method offers several advantages over traditional projection-based approaches:
- Global Definition: Unlike local projection methods, the lifting method is defined globally on the sphere, except at the origin of the ambient space. This makes it a powerful tool for proving broader stability results.
- Well-Posedness: The authors demonstrate that if agents begin at valid positions, they will never reach the singular origin point, ensuring the system remains well-defined for all time $t > 0$.
- Equivalence: Trajectories in the lifted space, when projected back to the sphere, perfectly match the original spherical consensus protocol.
Discussion and Future Work
While the lifting method is globally defined, current proofs primarily guarantee convergence for agents starting within a hemisphere. True global convergence—where agents synchronize from any initial configuration—remains an open challenge, partly due to the non-differentiability of the origin in the lifted space.
Furthermore, the dimension of the sphere plays a role; the region of attraction for "bad" equilibria (where agents fail to synchronize) has a measure of zero for $d \ge 3$, but may be more significant for $d=2$ (the circle). The authors anticipate that this lifting framework will serve as a foundation for future proofs of almost-global convergence on the unit sphere, regardless of how scattered the agents are initially.
Summary updated