A Convenient Category for Higher-Order Probability Theory — Chris Heunen, Ohad Kammar, Sam Staton & Hongseok Yang (2017). arXiv:1701.02547 (v4, PDF); LICS 2017, doi:10.1109/LICS.2017.8005137.
Introduces quasi-Borel spaces, which replace the measurable sets of a measurable space by the random variables into it. They form a cartesian closed category, so probabilistic programs can sample functions, with a commutative probability monad that extends the Giry monad’s Kleisli category, a semantics for Bayesian regression over functions, and a de Finetti theorem.
Sources: the paper, arXiv:1701.02547v4, checked against the arXiv listing. Index: Papers.
Key definitions and results
- Proposition 6 (Aumann): the category of measurable spaces is not cartesian closed
- Definition 7: quasi-Borel space; Example 8: measurable spaces as quasi-Borel spaces
- Definitions 10–11: probability measures and morphisms; the category QBS
- Proposition 15: the adjunction between QBS and measurable spaces, full and faithful on standard Borel spaces
- Propositions 16–18: products, coproducts and function spaces (QBS is cartesian closed)
- Theorem 21, Proposition 22: the strong commutative probability monad P
- §VI: Bayesian linear regression with a prior on functions
- Theorem 29: de Finetti’s theorem for quasi-Borel spaces (weak form)