A talk given in The Graduate Center on May 15, 2024.

Abstract: In previous talks at this Category seminar and at the Topology, Geometry and Physics seminar, Arthur Parzygnat showed how Bayesian inversion and its generalization to quantum mechanics may be interpreted as a functor on a suitable category of states which satisfies certain axioms. Such a functor is called a retrodiction and Parzygnat and collaborators conjectured that retrodiction is unique. In this talk, I will present a proof of this conjecture for the special case of classical probability theory on finite state spaces.

In this special case, the category in question has non-degenerate probability distributions on finite sets as its objects and stochastic matrices as its morphisms. After preliminary definitions and lemmas, the proof proceeds in three main steps.

In the first step, we focus on certain groups of automorphisms of certain objects. As a consequence of the axioms, it follows that these groups are preserved under any retrodiction functor and that the restriction of the functor to such a group is a certain kind of group automorphism. Since this group is isomorphic to a Lie group, it is easy to prove that the restriction of a retrodiction to such a group must equal Bayesian inversion if we assume continuity. If we do not make that assumption, we need to work harder and derive continuity "from scratch" starting from the positivity condition in the definition of stochastic matrix.

In the second step, we broaden our attention to the full automorphism groups of objects of our category corresponding to uniform distributions. We show that these groups are generated by the union of the subgroup consisting of permutation matrices and the subgroup considered in the first step. From this fact, it follows that the restriction of a retrodiction to this larger group must equal Bayesian inversion.

In the third step, we finally consider all the objects and morphisms of our category. As a consequence of what we have shown in the first two steps and some preliminary lemmas, it follows that retrodiction is given by matrix conjugation. Furthermore, Bayesian inversion is the special case where the conjugating matrices are diagonal matrices. Because the hom sets of our category are convex polytopes and a retrodiction functor is a continuous bijection of such sets, a retodiction must map polytope faces to faces. By an algebraic argument, this fact implies that the conjugating matrices are diagonal, answering the conjecture in the affirmative.