A Convenient Fibration for Dependently-Typed Probability Theory

We describe semantic structures relevant for interpreting dependent types for statistical and probabilistic modelling. Our development extends the theory of quasi-Borel spaces (qbses) of Staton et. al, which support simply-typed, higher-order probability theory with continuous distributions. It is well-known that qbses can interpret a dependent-type theory supporting dependent function-spaces through the codomain fibration. We define an equivalent split fibration based on the family fibration, which we call quasi-Borel families (qbfs), characterise its structure, equip it with fibred monads of measures and probability, and use them to develop dependently-typed probability theory.

We characterise the structure of the qbf fibration that is relevant for dependently-typed probability theory in elementary form. Our characterisations include: context extension, dependent pairs, dependent functions, extensional identity types, fibred products and coproducts, subspaces, a universe of propositions, and straightforward internalisation and externalisation principles for discrete spaces. We use these concepts to define fibred distribution and probability monads, the semantic structure needed to interpret probability distributions under a dependent context. We show that this structure satisfies a fibred version of Kock's synthetic measure theory. We also use these concepts to develop a qbs counterpart to Kolmogorov's conditional expectation. Our main result is a version of the conditional expectation that, under standard regularity assumptions, is measurable in both the random variables we are conditioning, and the observation map we are conditioning by.