Picturing classical and quantum Bayesian inference

We introduce a graphical framework for Bayesian inference that is sufficiently general to accommodate not just the standard case but also recent proposals for a theory of quantum Bayesian inference wherein one considers density operators rather than probability distributions as representative of degrees of belief. The diagrammatic framework is stated in the graphical language of symmetric monoidal categories and of compact structures and Frobenius structures therein, in which Bayesian inversion boils down to transposition with respect to an appropriate compact structure. We characterize classical Bayesian inference in terms of a graphical property and demonstrate that our approach eliminates some purely conventional elements that appear in common representations thereof, such as whether degrees of belief are represented by probabilities or entropic quantities. We also introduce a quantum-like calculus wherein the Frobenius structure is noncommutative and show that it can accommodate Leifer’s calculus of `conditional density operators’. The notion of conditional independence is also generalized to our graphical setting and we make some preliminary connections to the theory of Bayesian networks. Finally, we demonstrate how to construct a graphical Bayesian calculus within any dagger compact category.

💡 Research Summary

The paper “Picturing classical and quantum Bayesian inference” develops a highly abstract, diagrammatic framework for Bayesian inference that works uniformly for classical probability theory, recent proposals for quantum Bayesian inference, and even more general probabilistic theories. The authors employ the language of symmetric monoidal categories (SMCs) equipped with compact structures and dagger Frobenius algebras, showing that the essential operations of Bayesian reasoning—conditioning, marginalisation, and Bayes’ rule—are captured by two categorical constructions: the Frobenius multiplication/comultiplication and the compact‑structure transpose.

The first part of the paper reviews the necessary categorical background. A compact structure on an object A provides a pair of morphisms (the “cup” η: I→A*⊗A and the “cap” ε: A⊗A*→I) satisfying the yanking equations, which give a canonical notion of transpose for any morphism. Adding a dagger functor yields a dagger compact category, where the dagger corresponds to flipping diagrams upside‑down. A dagger Frobenius algebra on an object A consists of a multiplication m: A⊗A→A, a unit u: I→A, and their adjoints, satisfying associativity, unit laws, and the Frobenius law. When the multiplication is commutative, the Frobenius algebra corresponds to an orthogonal basis in finite‑dimensional Hilbert spaces; when it is special (m∘δ = id) it corresponds to an orthonormal basis, and such structures have been called “classical structures”.

With this machinery the authors define a Bayesian graphical calculus. A state (belief) is a point I→A, a joint state is a point I→A⊗B, and a conditional state is a morphism A⊗B→I that satisfies a “normalisation” condition expressed via the counit. Conditioning is realised by the Frobenius comultiplication δ, while marginalisation uses the compact caps. Crucially, Bayesian inversion (the passage from p(A|B) to p(B|A)) is simply the transpose of the conditional morphism with respect to the appropriate compact structure. This makes Bayes’ rule a diagrammatic identity: the transpose of a conditional equals the conditional of the transposed joint.

The authors then specialise to classical Bayesian graphical calculi, characterised by the commutativity of the Frobenius multiplication. In this setting the induced compact structure is self‑dual and the diagrams reproduce the familiar algebra of probabilities. They show that any bijective re‑parameterisation of probabilities (e.g., using negative log‑probabilities) yields an equivalent graphical calculus, demonstrating that the choice of representation (probability vs. entropy) is a convention rather than a logical necessity.

Next, the paper introduces a quantum‑like calculus (the Q½‑calculus) where the Frobenius multiplication is non‑commutative. This captures Leifer’s proposal of “conditional density operators”: density operators play the role of states, reduced density operators are marginals, and conditional operators are positive operators that trace to the identity on the conditioned system. The non‑commutative Frobenius structure encodes operator multiplication, while its comultiplication implements a logical broadcasting operation (a map that copies a state’s marginals). The authors provide an explicit representation of this calculus in the category of finite‑dimensional Hilbert spaces, showing that the diagrammatic Bayes rule coincides with the approximate reversal channel of Barnum and Knill.

A substantial portion of the work is devoted to conditional independence. Using the Frobenius structure, the authors define a notion of independence that mirrors the classical definition: A and B are conditionally independent given C if a certain diagram involving δ_A, δ_B, and the conditional on C commutes. They prove that this definition satisfies the semi‑graphoid axioms (symmetry, decomposition, weak union, contraction), establishing a bridge to the theory of Bayesian networks. An example of “generalised pooling” demonstrates how, under conditional independence, multiple beliefs can be combined into a single joint belief using the graphical calculus.

Finally, the authors show how to construct a Bayesian graphical calculus in any dagger compact category. Starting from any object A, they define a non‑commutative Frobenius multiplication on A⊗A by using the category’s tensor product and composition, and then derive the associated comultiplication. This yields a generic model of quantum Bayesian inference that works not only in FdHilb but also in other operational theories (e.g., real Hilbert spaces, probabilistic relations). The construction clarifies that the essential ingredients of quantum Bayesian inference are categorical rather than Hilbert‑space specific.

Overall, the paper achieves a unifying perspective: Bayesian inference—classical, quantum, or more exotic—can be expressed entirely in terms of two graphical operations, transpose and Frobenius (co)multiplication, within a dagger compact category. This abstraction strips away conventional numerical representations, highlights the structural essence of Bayes’ rule, and opens the door to new applications such as quantum belief‑propagation algorithms, causal inference from quantum correlations, and systematic exploration of Bayesian inference in generalized probabilistic theories.