Bayesian inference and retrodiction for faithful states on von Neumann algebras

Abstract: Retrodiction is the act of inferring a cause from its effects, the most common example of which is Bayesian inference. Retrodiction can be defined by its structural process-theoretic properties, which are mathematically captured by category theory. This categorical definition of retrodiction has recently been shown to potentially isolate the Petz recovery map as a unique universal candidate for quantum Bayesian inference. This paper extends these results to the infinite-dimensional setting on von Neumann algebras. In the process, we provide a pedagogical review of the Petz recovery map in infinite dimensions and its relation to the more commonly used expression in the finite-dimensional setting. We formalize the open question as to whether these categorical axioms for retrodiction do in fact uniquely characterize the Petz recovery map. If such a characterization holds, this would show that Bayesian inversion and the Petz recovery map are structural necessities and not simply useful algorithms for classical and quantum inference.

• The paper proves that the Petz recovery map defines a retrodiction functor for faithful normal states on arbitrary von Neumann algebras, preserving identities, composition, tensor products, inverses, and involution.
• The authors establish representation-independent KMS inner products, construct unique normal completely positive adjoints, and resolve the technical challenges of tensoriality in infinite-dimensional settings.
• The framework reduces exactly to Bayes’ rule for commutative von Neumann algebras on standard Borel spaces, while leaving uniqueness beyond Petz retrodiction and extensions to nonfaithful states as open questions.

Overview and main result

This paper by Karmakar and Parzygnar establishes that the Petz recovery map, defined for arbitrary von Neumann algebras equipped with faithful normal states, satisfies the categorical axioms of a retrodiction functor. The setting is the category $\mathbf{Q}$ whose objects are pairs $(\mathcal{A},\omega)$ of a von Neumann algebra with a faithful normal state and whose morphisms are state-preserving normal completely positive unital (NCPU) maps. The main theorem states that the assignment sending a morphism $\mathcal{E}$ to its Petz recovery map $\mathcal{E}^{\star}$ defines a functor $\mathscr{R}:\mathbf{Q}\to\mathbf{Q}^{op}$ that is identity-preserving, compositional, tensorial, inverting (it sends isomorphisms to their inverses), and involutive. This extends to infinite-dimensional von Neumann algebras the results of Parzygnat and Buscemi for finite-dimensional matrix algebras (Bonini et al., 2022), and it provides a structural argument that the Petz recovery map is the correct quantum analogue of Bayes' rule rather than merely a convenient algorithm.

The paper also serves a pedagogical purpose: it carefully connects the finite-dimensional formula $$\mathcal{E}^{\star}(a)=\sigma^{-1/2}\,\mathcal{E}^{*}(\rho^{1/2}a\rho^{1/2})\,\sigma^{-1/2}$$, ubiquitous in quantum information theory, with the modular-theoretic definition of Petz [Pe84] valid in general. The authors emphasize that the finite-dimensional expression, written in terms of density operators and the modular operator as $\mathcal{L}_\rho\mathcal{R}_\rho^{-1}$, does not extend to infinite dimensions because the modular operator is generally unbounded; the Tomita--Takesaki modular operator acting on the GNS Hilbert space is the appropriate generalization.

The KMS inner product and its properties

The construction rests on the Connes selfpolar form, here called the KMS inner product. For a faithful normal state $\omega$ on $\mathcal{A}$ represented on a Hilbert space with cyclic and separating vector $\Omega$, the inner product is defined as

$(\mathcal{A},\omega)$0

where $(\mathcal{A},\omega)$1 is the modular conjugation. The authors prove that this form is positive and nondegenerate, and—crucially—that it is independent of the chosen faithful representation and cyclic separating vector, via a unitary intertwining argument that transfers the Tomita operators and their polar decompositions between representations. In the finite-dimensional case the inner product reduces to $(\mathcal{A},\omega)$2, connecting it to the standard quantum Fisher information geometry literature.

Two technical facts are established for later use: weak-operator-topology continuity of the inner product in the second argument (via normality of the representation and properties of $(\mathcal{A},\omega)$3), and multiplicativity under tensor products, $(\mathcal{A},\omega)$4. The latter underpins the tensoriality axiom. The authors also note a genuine infinite-dimensional obstruction: $(\mathcal{A},\omega)$5 is not complete under this inner product in general, and they give an explicit Cauchy sequence of bounded operators (with spectrum growing like $(\mathcal{A},\omega)$6 against eigenvalues $(\mathcal{A},\omega)$7) converging to an unbounded operator, which is not in $(\mathcal{A},\omega)$8. This incompleteness is a structural feature that any adjoint construction must confront.

Existence and uniqueness of the Petz recovery map

The central analytic result is the existence and uniqueness of the KMS-adjoint. For a state-preserving NCPU map $(\mathcal{A},\omega)$9, there is a unique NCPU map $\mathcal{E}$0 satisfying

$\mathcal{E}$1

for all $\mathcal{E}$2, $\mathcal{E}$3. The proof adapts Accardi and Cecchini's argument [AcCe82]: for $\mathcal{E}$4 in the commutant $\mathcal{E}$5, the functional $\mathcal{E}$6 is uniformly dominated by $\mathcal{E}$7, hence possesses a Radon--Nikodym derivative in $\mathcal{E}$8, which extends linearly to a normal completely positive map $\mathcal{E}$9. The Petz map is then recovered as $\mathcal{E}^{\star}$0. Normality of $\mathcal{E}^{\star}$1 holds whether or not $\mathcal{E}^{\star}$2 itself is normal—a noteworthy point in the proof.

A companion result shows that if $\mathcal{E}^{\star}$3 is an isomorphism in $\mathcal{E}^{\star}$4 (i.e., has an NCPU inverse), then $\mathcal{E}^{\star}$5 is necessarily a $\mathcal{E}^{\star}$6-isomorphism and $\mathcal{E}^{\star}$7. The proof uses the Kadison--Schwarz inequality and Choi's multiplicative domain theorem to establish the $\mathcal{E}^{\star}$8-homomorphism property, then shows the intertwiner $\mathcal{E}^{\star}$9 between GNS representations is unitary and satisfies $\mathscr{R}:\mathbf{Q}\to\mathbf{Q}^{op}$0, from which the adjoint identity yields the inverse. This is proved without invoking the modular automorphism group, using only modular operators and uniqueness of polar decomposition.

The retrodiction functor

The paper defines a retrodiction functor on $\mathscr{R}:\mathbf{Q}\to\mathbf{Q}^{op}$1 as an assignment $\mathscr{R}:\mathbf{Q}\to\mathbf{Q}^{op}$2 satisfying six axioms: the recovery property, identity-preservation, compositionality, tensoriality, extension of inversion, and involutivity. These are precisely the properties that make $\mathscr{R}:\mathbf{Q}\to\mathbf{Q}^{op}$3 a unitary dagger symmetric monoidal category. The main theorem then verifies all six for the Petz map:

• Compositionality follows from uniqueness of the adjoint applied to the chain $\mathscr{R}:\mathbf{Q}\to\mathbf{Q}^{op}$4, giving $\mathscr{R}:\mathbf{Q}\to\mathbf{Q}^{op}$5.
• Tensoriality is the most delicate part. The identity $\mathscr{R}:\mathbf{Q}\to\mathbf{Q}^{op}$6 is first established on algebraic tensor products using multiplicativity of the KMS inner product, then extended to the von Neumann algebra tensor products $\mathscr{R}:\mathbf{Q}\to\mathbf{Q}^{op}$7 via two successive approximation arguments using weak-operator convergence and normality of the adjoints. This step is where infinite-dimensionality genuinely requires new work beyond the finite-dimensional treatment.
• Involutivity $\mathscr{R}:\mathbf{Q}\to\mathbf{Q}^{op}$8 follows from conjugate-symmetry and nondegeneracy of the KMS inner product.

The consequence is that Petz retrodiction makes $\mathscr{R}:\mathbf{Q}\to\mathbf{Q}^{op}$9 a dagger category in which the dagger coincides with the categorical inverse on isomorphisms, generalizing the finite-dimensional picture of Parzygnat and Buscemi.

Reduction to classical Bayesian inversion

For commutative von Neumann algebras the Petz recovery map reproduces Bayes' rule exactly. On finite sets, with $$\mathcal{E}^{\star}(a)=\sigma^{-1/2}\,\mathcal{E}^{*}(\rho^{1/2}a\rho^{1/2})\,\sigma^{-1/2}$$0, $$\mathcal{E}^{\star}(a)=\sigma^{-1/2}\,\mathcal{E}^{*}(\rho^{1/2}a\rho^{1/2})\,\sigma^{-1/2}$$1, a stochastic matrix $$\mathcal{E}^{\star}(a)=\sigma^{-1/2}\,\mathcal{E}^{*}(\rho^{1/2}a\rho^{1/2})\,\sigma^{-1/2}$$2 and nowhere-vanishing distributions $$\mathcal{E}^{\star}(a)=\sigma^{-1/2}\,\mathcal{E}^{*}(\rho^{1/2}a\rho^{1/2})\,\sigma^{-1/2}$$3, $$\mathcal{E}^{\star}(a)=\sigma^{-1/2}\,\mathcal{E}^{*}(\rho^{1/2}a\rho^{1/2})\,\sigma^{-1/2}$$4, the adjointness condition forces $$\mathcal{E}^{\star}(a)=\sigma^{-1/2}\,\mathcal{E}^{*}(\rho^{1/2}a\rho^{1/2})\,\sigma^{-1/2}$$5, i.e., $$\mathcal{E}^{\star}(a)=\sigma^{-1/2}\,\mathcal{E}^{*}(\rho^{1/2}a\rho^{1/2})\,\sigma^{-1/2}$$6, which is Bayes' rule.

The substantial new content is the extension to standard Borel spaces. Given a Markov kernel $$\mathcal{E}^{\star}(a)=\sigma^{-1/2}\,\mathcal{E}^{*}(\rho^{1/2}a\rho^{1/2})\,\sigma^{-1/2}$$7 with $$\mathcal{E}^{\star}(a)=\sigma^{-1/2}\,\mathcal{E}^{*}(\rho^{1/2}a\rho^{1/2})\,\sigma^{-1/2}$$8 faithful, the authors form the joint measure $$\mathcal{E}^{\star}(a)=\sigma^{-1/2}\,\mathcal{E}^{*}(\rho^{1/2}a\rho^{1/2})\,\sigma^{-1/2}$$9 on $\mathcal{L}_\rho\mathcal{R}_\rho^{-1}$0, apply the disintegration theorem to obtain a regular conditional probability $\mathcal{L}_\rho\mathcal{R}_\rho^{-1}$1 consistent with the projection $\mathcal{L}_\rho\mathcal{R}_\rho^{-1}$2, and define the Bayesian inverse kernel $\mathcal{L}_\rho\mathcal{R}_\rho^{-1}$3. They prove that the induced NCPU map $\mathcal{L}_\rho\mathcal{R}_\rho^{-1}$4 is precisely the Petz recovery map of $\mathcal{L}_\rho\mathcal{R}_\rho^{-1}$5. Rewriting the adjointness identity with Radon--Nikodym derivatives yields the density-level Bayes' rule

$\mathcal{L}_\rho\mathcal{R}_\rho^{-1}$6

almost everywhere, i.e., $\mathcal{L}_\rho\mathcal{R}_\rho^{-1}$7. The reliance on the disintegration theorem restricts this classical identification to standard Borel spaces, where consistent disintegrations are guaranteed to exist and are unique up to $\mathcal{L}_\rho\mathcal{R}_\rho^{-1}$8-null sets.

Markov maps and the GNS adjoint

A morphism is a Markov map if it intertwines the modular automorphism groups, $\mathcal{L}_\rho\mathcal{R}_\rho^{-1}$9 (the Accardi--Cecchini condition). The authors show the Markov maps form a symmetric monoidal subcategory $\omega$0, and that on $\omega$1 the Petz recovery map coincides with the GNS adjoint, the map satisfying $\omega$2. Complete positivity of the GNS adjoint is equivalent to the Accardi--Cecchini condition, so the restriction of the retrodiction functor to $\omega$3 lands in $\omega$4. This connects the framework to conditional expectations (Takesaki's theorem), Davies maps, thermal operations, and perfect quantum error-correcting codes, all of which are Markov maps in this sense.

Limitations and open questions

The paper is candid that its central uniqueness claim remains a conjecture, not a theorem: if $\omega$5 is any retrodiction functor on $\omega$6, does it necessarily coincide with Petz retrodiction? Prior work showed that rotated Petz maps and their averaged variants fail the axioms, but no proof of uniqueness is given. If the conjecture holds, the square-root formula would have to emerge from the axioms alone; if not, the admissible alternatives are unclassified. Either resolution bears directly on whether Bayesian inversion is a structural necessity or one algorithm among many.

Several further questions are left open. The treatment is restricted to faithful states; the extension to semifinite weights, and whether a crossed-product transfer from type $\omega$7 to type $\omega$8 algebras yields a trace-based density formula for Petz maps useful in quantum field theory and gravity, are posed but not resolved. The existence conditions for non-GNS Bayesian inverses (e.g., those defined via the Jordan-product inner product) are known only in special cases such as unital qubit channels and the amplitude-damping channel. Finally, the other quantum Bayes' rules in the literature are not treated here; their categorical descriptions remain future work.

Conclusion

The paper establishes that the Petz recovery map on arbitrary von Neumann algebras with faithful normal states defines a retrodiction functor—functorial, monoidal, involutive, and inverting—thereby placing quantum Bayesian inference on the same categorical footing as classical Bayes' rule, to which it reduces on commutative algebras over standard Borel spaces. The technical contributions are the representation-independence of the KMS inner product, the tensoriality of the adjoint construction on von Neumann tensor products, and the measure-theoretic identification of Petz retrodiction with regular conditional probabilities. The open uniqueness conjecture, if true, would recast Bayes' rule and its quantum generalization as consequences of process-theoretic axioms rather than optimization principles.