A new arXiv preprint develops a formal way to run Bayesian inference backward in a setting that includes infinite-dimensional von Neumann algebras, mathematical structures that let the authors treat probability in a noncommutative setting. Its central result is that the Petz recovery map satisfies the proposed categorical rules for retrodiction, the paper’s term for reversing a state-preserving map. But the result stops short of answering the question the authors ultimately want to settle: whether those rules force the Petz map uniquely.
The work is a version-1 preprint dated 20 Aug 2026, rather than a report of an experiment or a study involving people, animals or laboratory measurements. Its evidence consists of mathematical definitions, lemmas, propositions, proofs and examples. There is no empirical sample or performance test to report.
A reverse map with familiar special cases
The authors study pairs made up of a von Neumann algebra and a faithful normal state. The arrows between these objects are state-preserving NCPU maps, the formal transformations used in the paper’s category, called Q. In ordinary language, the framework starts with a mathematical system and a state describing it, then asks how information can be transported backward while respecting that state.
The Petz map is built as the adjoint of the original map under a KMS inner product. An adjoint is the operation paired with another map through an inner-product rule. The authors show that this KMS inner product is well-defined and does not depend on the chosen representation or on the cyclic and separating vector used to describe the state.
That construction leads to a precise structural claim: every morphism in Q has a unique KMS-inner-product adjoint, and the adjoint is itself another morphism in Q. This gives the proposed reverse operation a well-defined place inside the same mathematical category, rather than treating it as an external recipe.
For full matrix algebras, the paper recovers the finite-dimensional Heisenberg-picture expression E⋆ = Adσ−1/2 ◦ E∗ ◦ Adρ1/2. The formula is the matrix form of the abstract adjoint construction and shows how the general framework specializes to a familiar density-matrix setting.
The framework also behaves as expected when the forward map is perfectly reversible. For an isomorphism in Q, Petz retrodiction is the categorical inverse. In that special case, the reverse operation agrees with the inverse already present in the category.
Where Bayes’ rule reappears
The clearest bridge to ordinary probability appears on finite commutative algebras. There, the Petz recovery map reproduces Bayes’ rule, and its reverse kernel represents the conditional probability of x given y. The result places the familiar classical calculation and the more general operator-algebraic construction in the same formal picture.
The authors also treat commutative settings built from standard Borel spaces. Starting from a Markov kernel, they construct a reverse kernel through disintegration, a way of building conditional or reverse probabilities from the relevant probability description. The NCPU map induced by that reverse kernel is identified with the Petz recovery map.
This correspondence connects the abstract map to a recognizable Bayesian operation without turning the paper into a practical prediction study. The supplied analysis describes the result as a structural extension of Bayes’ rule to infinite-dimensional and noncommutative settings.
A related match appears for modular-covariant Markov maps, a class of maps constrained by the state’s modular structure. On that Markov subcategory, the Petz retrodiction coincides with the GNS adjoint, another adjoint construction used in the mathematical treatment of states and maps.
The main claim remains unproved
The preprint’s headline mathematical question is still open. The authors state as Conjecture 7.1 that any retrodiction functor on Q coincides with the Petz retrodiction functor, but the paper does not establish that statement. In other words, it proves that the Petz construction has the proposed properties; it does not prove that no different construction could have them as well.
That distinction sets the limits of the result. Because the work concerns formal mathematical objects, it supplies no empirical validation, effect estimates, uncertainty intervals or comparison with alternative inference procedures. It also provides no direct evidence about human, animal, clinical or laboratory inference.
The main framework is restricted to von Neumann algebras with faithful normal states and state-preserving NCPU maps. The authors leave questions involving other Bayes rules and their categorical descriptions for future work, along with broader extensions beyond the setting developed in the paper.
The preprint therefore establishes a common construction and several structural correspondences while leaving its strongest uniqueness interpretation unresolved. The authors’ broader reading is conditional: if the uniqueness conjecture eventually holds, Bayesian inversion and the Petz recovery map would be structural necessities of the framework rather than merely useful algorithms.
A formal result, not a performance test
For readers outside mathematical quantum information, the takeaway is limited but clear. The preprint links ordinary finite Bayesian conditioning, reverse kernels on standard Borel spaces, categorical inverses and GNS adjoints to one Petz-based construction in the cases it studies. It does not show that the construction is more accurate, faster or superior to another inference method.
The acknowledgments report discussions with named individuals and hospitality from conference organizers and the University of Kentucky; no financial support is specified. The authors also declare using Microsoft 365 Copilot and Google Gemini for reference searching, BibTex generation, background learning, exploratory mathematical assistance and proofreading, while stating that the outputs were checked and the writing was done solely by the authors.
Paper data and sources
Original title: Bayesian inference and retrodiction for faithful states on von Neumann algebras
Authors: Pradyut Karmakar, Arthur J. Parzygnat
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text