Tropical Mathematics and the Lambda-Calculus II: Tropical Geometry of Probabilistic Programming Languages

Abstract: In the last few years there has been a growing interest towards methods for statistical inference and learning based on ideas from computational algebraic geometry, and notably from tropical geometry, that is, the study of algebraic varieties over the min-plus semiring. At the same time, recent work has demonstrated the possibility of interpreting a higher-order probabilistic programming language in the framework of tropical mathematics, by exploiting the weighted relational semantics from linear logic. In this paper we try to put these two worlds in contact, by showing that actual methods from tropical geometry can indeed be exploited to perform statistical inference on higher order programs. For example, we show that the problem of describing the most-likely behavior of a probabilistic PCF program reduces to studying a tropical polynomial function associated with the program. We also design an intersection type system that captures such polynomials. As an application of our approach, we finally show that the tropical polynomial associated with a probabilistic protocol expressed in our language can be used to estimate its differential privacy.