Bisimulations for probabilistic linear lambda calculi

Abstract

We investigate a notion of probabilistic program equivalence under linear contexts. We show that both a statebased and a distribution-based bisimilarity are sound coinductive proof techniques for reasoning about higher-order probabilistic programs, but only the distribution-based one is complete for linear contextual equivalence. The completeness proof is novel and directly constructs linear contexts from transitions, rather than the traditional approach of characterizing bisimilarities as testing equivalences.