What is a categorical model of the differential and the resource lambda-calculi?

In this paper we provide an abstract model theory for the untyped differential lambda-calculus and the resource calculus. In particular we propose a general definition of model of these calculi, namely the notion of linear reflexive object in a Cartesian closed differential category. Examples of models based on relations are provided.

💡 Research Summary

The paper develops a unified categorical model theory for two non‑standard λ‑calculi: the untyped differential λ‑calculus and the resource λ‑calculus. Both calculi extend the ordinary λ‑calculus with operations that make the use of variables explicit—differentiation in the former and controlled duplication/erasure in the latter. Traditional categorical semantics based on Cartesian closed categories (CCCs) are insufficient to capture these extra operations, because they lack a built‑in notion of linearity and a way to interpret a differential operator.

To address this gap, the authors first recall the structure of a differential category, a recent categorical abstraction introduced to model differential linear logic. A differential category is a CCC equipped with a differential combinator D that assigns to each object A a morphism D_A : A → A⊸A satisfying a set of axioms mirroring the algebraic properties of the ordinary derivative: linearity, the chain rule, Leibniz rule, and compatibility with the Cartesian product. This combinator makes it possible to speak about “linear” morphisms inside a generally non‑linear setting.

The central contribution of the paper is the definition of a linear reflexive object (LRO) within a differential category. An LRO consists of an object R together with two morphisms

• η : 1 → R (the “code” map), and
• ε : R⊸R → R (the “decode” map),

subject to two equations:

1. ε ∘ (η⊸id_R) = id_R, which is the usual reflexivity condition ensuring that R can be applied to itself; and
2. ε ∘ (Dη) = η, which forces the differential combinator to interact trivially with the code map.

The first equation guarantees that the object can interpret self‑application, a standard requirement for models of the untyped λ‑calculus. The second equation is novel: it says that differentiating the code does not change its meaning, thereby embedding the differential operator of the differential λ‑calculus into the categorical structure. Because ε expects a linear morphism (an element of R⊸R) and returns an element of R, the LRO automatically respects the linearity constraints that are essential for the resource λ‑calculus, where each variable’s usage count is tracked.

Having defined the abstract notion, the authors instantiate it in the concrete category Rel of sets and binary relations. Objects are sets, morphisms are relations, and the Cartesian closed structure is given by the usual product of sets and the set of all relations from A to B as the exponential A⊸B. The differential combinator D is defined by turning a relation R ⊆ A × B into a relation D(R) ⊆ A × (A⊸B) that pairs each a∈A with a linear function f such that (a, f(a))∈R. The code map η picks out the singleton relation linking the terminal object 1 to a distinguished element of R, while ε is simply relational composition. The authors verify in detail that the two LRO equations hold in this setting, thereby providing a concrete model that validates both the differential reduction rules and the resource calculus’s explicit copying/erasing rules.

The paper also situates its contribution within the broader literature. Earlier models of the differential λ‑calculus (e.g., those based on finiteness spaces or on the differential linear logic of Ehrhard and Regnier) treat differentiation but do not incorporate the linear reflexivity needed for the resource calculus. Conversely, models of the resource calculus often rely on linear logic semantics without a built‑in differential operator. By introducing the LRO, the authors present a single categorical framework that subsumes both. Moreover, because the definition only requires a Cartesian closed differential category, any such category—sets‑functions, coherence spaces, probabilistic coherence spaces, etc.—can potentially host an LRO, opening the door to a wide variety of concrete models.

In conclusion, the paper offers a clean, abstract semantics for two sophisticated extensions of the λ‑calculus. The linear reflexive object captures self‑application, linearity, and differentiation in a single structure, and the relational example demonstrates that the theory is not merely vacuous but can be instantiated concretely. Future work suggested by the authors includes extending the framework to typed settings, exploring connections with differential linear logic, and investigating computational interpretations such as abstract machines or compilation strategies that respect the categorical constraints.