Bayesian inversion wants to be a dagger, a way of reversing every channel that squares to the identity, but on the nose it is not, because reversal depends on the prior and only holds almost surely. Quotienting a category of probability spaces by almost-sure equality repairs this, and what emerges is a strikingly rigid, self-dual category.

The quotient

Start from a category whose objects are probability spaces and whose morphisms are channels between them, and identify two morphisms whenever they agree almost surely. For the quotient to be a category the relation must be a categorical congruence. Such a relation is an equivalence relation compatible with composition and with identities, so that composing representatives is well defined. The result is the influence category , following Fritz.

What the quotient buys

Proposition on rigidity of the influence category

In every object is self-dual, so the category is dagger-compact; the monoidal unit coincides with both the initial and the terminal object; the dagger is Bayesian inversion, now well defined because almost-sure equality absorbs the prior-dependence and the null-set ambiguity; and the whole category is equivalent to a category of input–output coupling diagrams. Inference reversed twice is inference, up to almost-sure equality.

This is the setting in which the dagger question raised for a categorical statistics has a clean answer. Inversion is a genuine involution once one works up to null sets.

References

  • T. Fritz, A synthetic approach to Markov kernels, conditional independence and theorems on sufficient statistics (Advances in Mathematics, 2020)