- The paper develops a multicategory-parameterized calculus for message-passing processes using polarized get/put operations and parallel let-bindings to model resource exchange.
- The rewrite system terminates and is confluent, giving every well-typed term a unique normal form and establishing deadlock freedom by eliminating internally resolvable let-bindings.
- The calculus forms a virtual double category with a sound functor into the free cornering, but faithfulness and support for branching protocols remain open problems.
Overview
The paper "A Simple Categorical Calculus of Interacting Processes" (2603.18321) by Nester and Voorneveld develops a term rewriting system for message-passing process interaction, parameterised by an arbitrary multicategory M whose morphisms are interpreted as non-interactive resource-transforming processes. The calculus augments these processes with a minimal interaction mechanism: processes running in parallel may exchange resources along left and right boundaries via polarised get/put constructors. The main technical contributions are threefold. First, the rewrite relation → on terms T(M) is proven confluent and terminating, so that terms possess unique normal forms. Second, terms modulo the induced convertibility relation ↔∗ form a virtual double category [M]. Third, there is a functor of virtual double categories from [M] into the underlying virtual double category of the free cornering of the free monoidal category F(M), yielding a sound denotational semantics.
The work is positioned relative to Cockett and Pastro's logic of message passing, Wadler's session-typed functional programming, and the line of work on the free cornering initiated by Nester. The stated motivation is to provide a more operational account of the free cornering: whereas the free cornering presents interaction statically, via equations between cells, the present calculus presents it dynamically, via directed reduction.
Background: the free cornering
The paper first recalls the free cornering ${\mathbb{A}$ of a strict monoidal category A, a single-object strict double category modelling two-party interaction. Vertical edges are elements of the monoid $#1{A_0} = (A_0 \times \{\circ,\bullet\})^*$ of exchanges: sequences of polarised resources read as simple protocols between a left and a right participant, where →0 requires the left participant to send an instance of →1 rightward and →2 the converse. Cells are generated by inclusions →3 of morphisms of →4, four corner cells (→5, →6, →7, →8) permitting resource exchange across boundaries, vertical identities →9, horizontal identities T(M)0, and the two composition operations. The defining equations include yanking laws asserting that exchanging a resource has no effect on it, e.g. T(M)1.
The relationship between multicategories and monoidal categories, and between virtual double categories and strict double categories, is given by free/forgetful adjunctions; both are instances of Leinster's generalised multicategory theory for cartesian monads. This underpins the choice of multicategories as the parameterising structure for the calculus.
The term calculus
Terms are presented as a sequent calculus with judgements T(M)2, where T(M)3 is a linear context (no repeated variables), T(M)4 is the return type, and T(M)5 are the left and right boundary protocols. The formation rules comprise:
- Neutral terms T(M)6 for morphisms T(M)7 of T(M)8, performing no interaction;
- Left/right-facing terms T(M)9, ↔∗0, ↔∗1, ↔∗2, which send or await a value along the corresponding boundary;
- Let-bindings ↔∗3, running subterms in parallel with boundaries concatenated sequentially, then feeding results to the continuation.
An admissible value-substitution rule resolves variables via composition in ↔∗4, defined by structural induction on terms. The rewrite relation is generated by eleven rules: [R0] substitutes neutral inputs into a let-binding; [R1] and [R2] perform internal interactions between adjacent matching put/get pairs — this is where computation happens; [R3–R6] hoist outward-facing gets and puts above let-bindings; and [R7–R10] commute left-facing past right-facing gets and puts at the top level. The authors observe that [R7–R10] contribute no computational work but are necessary for confluence, and note that the side conditions on [R8]/[R9] are not strictly required since typing already enforces them.
Two worked examples illustrate the mechanics: a sewing example in which a pattern-producing process sends its output to a sewing process within a let-binding, reducing to a normal form ↔∗5; and a coffee-machine example demonstrating how a ↔∗6 is hoisted above a let-binding so that interaction with the external environment occurs after internal message exchange completes.
The interpretation ↔∗7 maps each term to a cell of ↔∗8 by translating neutral terms to cell inclusions, let-bindings to horizontal composition followed by vertical composition, and each get/put to the corresponding corner cell composed appropriately. Coherence is established in two steps: substitution coherence (Lemma 4.x) and full coherence with respect to convertibility, i.e. ↔∗9 implies [M]0.
Termination and confluence
Termination follows from a size measure [M]1 under which every generating rewrite strictly decreases size; crucially, value substitution preserves size, so [R0] decreases the measure despite performing substitution. Confluence is obtained via Newman's lemma from local confluence, proven by resolving twelve classes of critical pairs grouped into three families: conflicts between internal interaction ([R1]/[R2]) and commutation rules ([R7–R10]); conflicts between left-hoisting ([R3]/[R4]) and right-hoisting ([R5]/[R6]); and conflicts between right-hoisting and commutation. The technically novel ingredient is the notion of term context [M]2 — a term with a hole appearing inside a let-binding sequence or beneath put/get binders — together with two lemmas: any effectful context reduces to an active one (one whose last bound subterm is right-facing), and a boundary-multipop joinability result. These supply the missing interactions needed to join critical pairs that cannot be resolved directly because an obstructing [M]3 intervenes.
A notable corollary of termination analysis is Lemma 4.x ("any let-binding reduces"), which the authors interpret as a cut-elimination theorem for the presenting sequent calculus and, semantically, as a deadlock-freedom property: normal forms contain no let-bindings, so every internally resolvable interaction is in fact resolved during reduction to normal form.
The rewrite relation is deliberately left-biased: when both a left-facing and right-facing resolution are available, the left is chosen first. The authors state plainly that a dual right-biased system would equally be confluent and terminating, and that the choice is arbitrary but unavoidable.
The virtual double category of terms
Modulo convertibility, the terms assemble into a single-object virtual double category [M]4: objects of [M]5 serve as horizontal morphisms, the exchange monoid [M]6 as the vertical edge monoid, derivations as cells, and composition is given by let-binding, with empty compositions handled by auxiliary terms [M]7 built from commuting get/put pairs that act as horizontal identities. Verifying associativity requires three technical lemmas showing that [M]8 behaves correctly: it commutes with let-binding, permits interaction across it, and permits hoisting across it. Associativity itself is proven by a simultaneous reduction argument indexed by total size of the innermost composites.
Finally, coherence of [M]9 plus preservation of identities and composition (using [M]0) yields the main theorem: [M]1 is a functor of virtual double categories. Read operationally, this is a sound denotational semantics: convertible programs denote equal cells of the free cornering.
Limitations and open questions
The authors are explicit about what remains open. The soundness functor is conjectured, but not proven, to be faithful; faithfulness would constitute adequacy of the denotational semantics relative to the operational one, a substantially stronger and more useful property than mere soundness. Expressiveness is also limited: protocols here are essentially linear session types without branching, so the calculus cannot express branching protocols such as those handled by the extended free cornering constructions of Nester (2024). Extending the calculus accordingly, with a view toward applications such as reasoning about cryptographic protocols, is proposed but not carried out. Additionally, the left-bias of the rewrite system, while harmless, means the normalisation strategy is a design artefact rather than a canonical feature.
Conclusion
The paper delivers a self-contained operational counterpart to the free cornering: a parameterised sequent calculus whose rewriting system is terminating and confluent, whose quotient forms a virtual double category, and which maps soundly into the free cornering of the free monoidal category on the parameterising multicategory. The critical-pair analysis mediated by term contexts is the principal proof technique, and the deadlock-freedom reading of let-reduction connects the metatheory to process-theoretic concerns. The remaining gap between soundness and adequacy, and the absence of branching protocols, delineate precisely what this calculus does not yet achieve.