A Non-Standard Semantics for Kahn Networks in Continuous Time

In a seminal article, Kahn has introduced the notion of process network and given a semantics for those using Scott domains whose elements are (possibly infinite) sequences of values. This model has since then become a standard tool for studying distributed asynchronous computations. From the beginning, process networks have been drawn as particular graphs, but this syntax is never formalized. We take the opportunity to clarify it by giving a precise definition of these graphs, that we call nets. The resulting category is shown to be a fixpoint category, i.e. a cartesian category which is traced wrt the monoidal structure given by the product, and interestingly this structure characterizes the category: we show that it is the free fixpoint category containing a given set of morphisms, thus providing a complete axiomatics that models of process networks should satisfy. We then use these tools to build a model of networks in which data vary over a continuous time, in order to elaborate on the idea that process networks should also be able to encompass computational models such as hybrid systems or electric circuits. We relate this model to Kahn’s semantics by introducing a third model of networks based on non-standard analysis, whose elements form an internal complete partial order for which many properties of standard domains can be reformulated. The use of hyperreals in this model allows it to formally consider the notion of infinitesimal, and thus to make a bridge between discrete and continuous time: time is “discrete”, but the duration between two instants is infinitesimal. Finally, we give some examples of uses of the model by describing some networks implementing common constructions in analysis.

💡 Research Summary

This paper revisits the foundational semantics of Kahn process networks and extends them to a continuous‑time setting by employing tools from non‑standard analysis. The authors first observe that, despite the widespread use of graphical drawings to represent process networks, a rigorous syntactic description of these drawings has been missing. To fill this gap they introduce nets, a formal structure built on a signature Σ = (Σ, σ, τ) where each symbol α ∈ Σ carries an arity σ(α) and a co‑arity τ(α). A net N = (P, O, λ, s, t) from m to n consists of a finite set P of ports, a finite set O of operators, a labeling function λ : O → Σ, an input‑port function s that maps each operator input to a port, and an injective output‑port function t that maps each operator output (and each external output of the whole net) to a port. The injectivity of t guarantees that no two wires share the same output port, while ports may be used as inputs for several wires, reflecting the FIFO nature of Kahn channels.

Having defined nets, the authors construct the category Net Σ whose objects are natural numbers (the numbers of external inputs and outputs) and whose morphisms are isomorphism classes of nets. They give explicit definitions of the categorical primitives:

• Identity net – a net with no operators and ports identified with the external interface.
• Composition – gluing the output ports of one net to the input ports of another, modulo the smallest equivalence relation that identifies the glued ports.
• Tensor product – juxtaposing two nets side‑by‑side, adding their input and output counts.
• Trace – a feedback operation that “closes” a set of ports, turning a net of type n₁ + n → n₂ + n into a net n₁ → n₂.

These constructions satisfy the axioms of a symmetric monoidal traced category (vanishing, superposing, yanking) as introduced by Joyal, Street and Verity. Consequently, Net Σ is a fixpoint category: every morphism f : A ⊗ X → B ⊗ X has a canonical least fixpoint, which models the semantics of feedback loops in Kahn networks. The authors prove that Net Σ is the free fixpoint category generated by the given signature Σ, i.e. any other fixpoint category equipped with interpretations of the symbols of Σ factors uniquely through Net Σ. This result provides a complete axiomatisation for any model of process networks.

To obtain a more compact representation, the paper introduces rewriting rules: sharing (identifying two operators with identical label and inputs) and erasing (removing an operator whose outputs are never used). The rewriting system is terminating and confluent; its normal forms correspond bijectively to shared nets, a variant where each operator appears at most once and every output is consumed. Quotienting Net Σ by the equivalence generated by these rules yields a category sNet Σ that possesses finite products (the tensor product coincides with the categorical product) while preserving the traced structure.

Having established the categorical backbone, the authors turn to continuous‑time semantics. Traditional Kahn semantics uses streams indexed by the natural numbers ℕ, which corresponds to discrete time steps. The paper proposes to replace ℕ by the non‑negative reals ℝ₊, but to retain a “discrete‑like” view by introducing an infinitesimal time step dt. In non‑standard analysis, ℝ is extended to the hyperreal field *ℝ, containing infinitesimals ε such that 0 < ε < 1/n for all standard n. Time is then modeled as the sequence 0, dt, 2·dt, 3·dt, …, where dt is a fixed positive infinitesimal.

Within this framework the authors define an internal complete partial order (internal cpo) of streams: functions from the hyperfinite index set {0,…,N} (with N a hypernatural) to a base domain D, ordered pointwise. Scott‑continuity lifts to internal Scott‑continuity, and the usual fixpoint theorem holds internally, guaranteeing least fixpoints for internal continuous operators. Consequently, every net can be interpreted as an internal Scott‑continuous function on internal streams, yielding a non‑standard semantics of Kahn networks that works over continuous time.

The paper then relates the non‑standard model to the classical one. There is a standard part map st : *ℝ → ℝ that sends a hyperreal to its nearest real number. Applying st pointwise to an internal stream produces a standard stream indexed by ℝ₊. The authors show that the interpretation of a net under the non‑standard semantics, followed by the standard part map, coincides with the traditional Kahn interpretation where time is discretized by ℕ and the limit dt → 0 is taken. Thus the two semantics are compatible: the non‑standard model refines the standard one by keeping track of infinitesimal temporal details.

Finally, the paper presents concrete examples illustrating the expressive power of the continuous‑time model:

• Differential operator – a net that, via a feedback loop, computes the derivative of a continuous‑time signal. The infinitesimal dt makes the finite‑difference approximation exact in the non‑standard setting.
• Electrical circuit components – resistors, inductors, and capacitors are modeled as primitive operators; Kirchhoff’s laws are expressed as equations between streams. The infinitesimal time step captures the instantaneous voltage‑current relationship.
• Hybrid systems – systems that combine continuous dynamics (e.g., a mass‑spring oscillator) with discrete events (e.g., a switch). The model cleanly separates continuous streams from event ports, yet allows them to interact through feedback.

These examples demonstrate that the non‑standard semantics can faithfully represent both purely continuous models and mixed discrete‑continuous (hybrid) systems, something that the original Kahn model could not do directly.

In summary, the paper makes three substantial contributions: (1) a rigorous categorical syntax for Kahn process networks, proved to be the free fixpoint category generated by a given signature; (2) a non‑standard analysis based semantics that extends Kahn’s model to continuous time while preserving the fixpoint property; and (3) a bridge between the discrete and continuous worlds via the standard part map, together with illustrative applications to analysis, circuit theory, and hybrid dynamics. This work deepens the theoretical foundations of concurrent computation and opens a pathway for applying process‑network ideas to domains traditionally modeled with differential equations and hybrid automata.