An abstract view on syntax with sharing

The notion of term graph encodes a refinement of inductively generated syntax in which regard is paid to the the sharing and discard of subterms. Inductively generated syntax has an abstract expression in terms of initial algebras for certain endofunctors on the category of sets, which permits one to go beyond the set-based case, and speak of inductively generated syntax in other settings. In this paper we give a similar abstract expression to the notion of term graph. Aspects of the concrete theory are redeveloped in this setting, and applications beyond the realm of sets discussed.

💡 Research Summary

The paper “An abstract view on syntax with sharing” develops a categorical framework that generalizes the traditional inductive description of syntax to explicitly account for sharing and discarding of subterms, a phenomenon naturally captured by term graphs. Classical syntax is usually presented as the initial algebra μF of an endofunctor F on Set, which yields a tree‑like representation where each occurrence of a subterm is a distinct node. While this representation is mathematically convenient, it obscures the fact that many subterms are identical and can be reused, leading to redundancy in both theoretical models (e.g., rewriting) and practical implementations (e.g., compilers).

The authors propose to replace the tree‑based view with a graph‑based one, called a term graph, and to lift the whole construction from the concrete category of sets to an arbitrary base category C. The central technical contribution is the definition of an “F‑graph” as a span A ← X → B equipped with an algebraic structure compatible with the endofunctor F. Here X represents the shared subterms, while A and B encode the input and output contexts of a syntactic operation. By treating the span as a morphism in the slice category C/X, the authors are able to capture sharing as an explicit categorical object rather than as an implicit duplication in the underlying set.

A key insight is that an F‑graph can simultaneously be an initial F‑algebra and a final F‑coalgebra for a suitably defined comonad G on the category of spans. The paper constructs natural transformations η : Id ⇒ G and ε : G ⇒ Id satisfying the usual triangular identities, thereby showing that the category of F‑graphs is a G‑algebra and a G‑coalgebra at the same time. This “dual initial/final” property guarantees both the existence of a unique, well‑founded syntax (the initial algebra) and a canonical unfolding or observation mechanism (the final coalgebra) that respects sharing. In effect, the term graph becomes a fixed point of a combined algebra‑coalgebra equation, unifying induction and coinduction in a single structure.

The authors then demonstrate the versatility of the framework by instantiating it in several categories beyond Set. In a presheaf category, the span construction yields location‑dependent sharing, which is essential for modelling variable binding and context‑sensitive languages. In a domain‑theoretic setting (continuous maps between ω‑cpos), the same construction interacts nicely with least fixed‑point operators, allowing recursive definitions to be expressed without losing sharing information. In a presheaf‑like “finitary” category (often called a “free” or “finitary” category), the authors show how finite term graphs can be embedded into infinite ones, providing a bridge between finite program fragments and their potentially infinite unfolding. For each case, they verify that the required endofunctor F admits an initial algebra and that the associated comonad G admits a final coalgebra, thereby confirming that the abstract theory is robust under a wide range of semantic settings.

Beyond the theoretical development, the paper explores concrete applications. In graph rewriting systems, matching a left‑hand side pattern against a term graph must respect shared nodes; the categorical formulation guarantees that rewrite steps preserve the sharing structure, avoiding inadvertent duplication of subterms. In compiler construction, representing intermediate code as term graphs enables common subexpression elimination to be performed “for free” because shared subterms are already identified by the graph’s topology. Moreover, the dual algebra/coalgebra nature supports both forward evaluation (using the algebraic structure) and backward analysis or debugging (using the coalgebraic observation), offering a unified foundation for many program‑analysis tasks.

In conclusion, the paper provides a mathematically elegant and highly general abstraction of syntax with sharing. By recasting term graphs as initial algebras and final coalgebras in a suitable categorical environment, it unifies induction and coinduction, preserves sharing across transformations, and extends the reach of syntactic modelling far beyond the category of sets. This work opens the door to new research directions in categorical semantics, graph‑based rewriting, and the design of programming languages and tools that can exploit sharing at a foundational level.