Dynamics on networks I. Combinatorial categories of modular continuous-time systems

We develop a new framework for the study of complex continuous time dynamical systems based on viewing them as collections of interacting control modules. This framework is inspired by and builds upon the groupoid formalism of Golubitsky, Stewart and their collaborators. Our approach uses the tools and — more importantly —the stance of category theory. This enables us to put the groupoid formalism in a coordinate-free setting and to extend it from ordinary differential equations to vector fields on manifolds. In particular, we construct combinatorial models for categories of modular continuous time dynamical systems. Each such model, as a category, is a fibration over an appropriate category of labeled directed graphs. This makes precise the relation between dynamical systems living on networks and the combinatorial structure of the underlying directed graphs, allowing us to exploit the relation in new and interesting ways.

💡 Research Summary

The paper presents a categorical framework for continuous‑time dynamical systems on networks, extending the groupoid symmetry approach of Golubitsky and Stewart to a coordinate‑free setting that works not only for ordinary differential equations but also for vector fields on arbitrary manifolds. The authors view a networked dynamical system as a collection of interacting control modules whose interconnections are encoded by a directed graph. For each vertex of a graph Γ they define an “input tree” I(a) consisting of all incoming edges; this tree determines a control system (a surjective submersion together with a smooth map into the tangent bundle of the target manifold).

A phase‑space functor P : Graphᵒᵖ → Man assigns to the vertex set of a graph the product of the manifolds attached to the vertices, thus turning the combinatorial data into a concrete state space. The collection of input trees together with their isomorphisms forms a groupoid G(Γ). This groupoid acts on the vector‑space valued functor Ctrl that assigns to each input tree the space of admissible control systems. The limit of this action, denoted V_Γ, is called the space of “virtual groupoid‑invariant vector fields.” Elements of V_Γ are not actual vector fields; rather, there is a linear map S_Γ : V_Γ → χ(P(Γ)) that sends a virtual invariant to a genuine vector field on the phase manifold P(Γ).

The crucial class of graph morphisms are the étale maps (those that preserve the input‑tree structure). Restricting to étale maps yields a subcategory Graphᵉᵗ. The authors define a contravariant functor V : (Graphᵉᵗ)ᵒᵖ → Vect extending the assignment Γ ↦ V_Γ. The main theorem states that V and the collection {S_Γ} are compatible: for any étale morphism φ : Γ → Γ′ and any virtual invariant w ∈ V(Γ′) there is an induced dynamical‑system morphism
Pφ : (P(Γ′), S_{Γ′}(w)) → (P(Γ), S_Γ(V(φ)w)).
Thus every étale graph map automatically lifts to a map between the corresponding dynamical systems, making the category V of pairs (Γ, w) a fibration over (Graphᵉᵗ)ᵒᵖ.

The paper further generalizes the construction by allowing “colored” graphs C that assign possibly different manifolds to different node types and different control systems to different edge types. In this setting the relevant category becomes the slice (Graph/C)ᵉᵗ, and the same machinery produces modular dynamical systems with heterogeneous components.

Sections 2–4 develop the theory first for Euclidean spaces (linear ODEs), then for linear vector fields, and finally for arbitrary manifolds, illustrating how the same categorical structures persist. Section 5 compares groupoid‑invariant vector fields with ordinary group‑invariant ones, proving that the former form a subspace of the latter when the underlying graph possesses non‑trivial group symmetries. Section 6 supplies technical results on categorical limits needed for the construction of the fibrations.

Overall, the authors provide a rigorous bridge between the combinatorial structure of directed graphs and the analytic structure of continuous‑time dynamics. By encoding control modules, input trees, and symmetry groupoids categorically, they obtain a clean, functorial description of how network topology dictates admissible dynamical systems and how morphisms of networks induce morphisms of dynamics. This framework promises to unify future work on discrete‑time, stochastic, and numerical models on networks under a single categorical umbrella.