On the discriminating power of tests in resource lambda-calculus

TL;DR

This paper investigates the discriminating power of tests in resource lambda-calculus models, providing a counter-example to full abstraction conjectures and exploring extensions with tests that achieve full abstraction.

Contribution

It presents a counter-example showing MRel is not fully abstract for resource lambda-calculus and analyzes extensions with tests that attain full abstraction.

Findings

MRel model is not fully abstract for resource lambda-calculus.

Counter-example disproves the full abstraction conjecture.

Extensions with tests can achieve full abstraction.

Abstract

Since its discovery, differential linear logic (DLL) inspired numerous domains. In denotational semantics, categorical models of DLL are now commune, and the simplest one is Rel, the category of sets and relations. In proof theory this naturally gave birth to differential proof nets that are full and complete for DLL. In turn, these tools can naturally be translated to their intuitionistic counterpart. By taking the co-Kleisly category associated to the ! comonad, Rel becomes MRel, a model of the \Lcalcul that contains a notion of differentiation. Proof nets can be used naturally to extend the \Lcalcul into the lambda calculus with resources, a calculus that contains notions of linearity and differentiations. Of course MRel is a model of the \Lcalcul with resources, and it has been proved adequate, but is it fully abstract? That was a strong conjecture of Bucciarelli, Carraro, Ehrhard and Manzonetto. However, in this paper we exhibit a counter-example. Moreover, to give more intuition on the essence of the counter-example and to look for more generality, we will use an extension of the resource \Lcalcul also introduced by Bucciarelli et al for which $\Minf$ is fully abstract, the tests.