arXiv:2610.20694
Quantum Partial Information Decompositions, an Operational Order on Ensembles, and Direct Quantifications of Objective Information
Alexander Nico-Katz
Abstract
The partial information decomposition (PID) framework is an extension of classical information theory that enables a systematic analysis of the structure of information in multipartite systems. We advance a simple extension of this framework enabling the construction of multipartite quantum PIDs (QPIDs); generalizing recent, exclusively bipartite, extensions. We construct two QPIDs of the Holevo quantity. First a naive QPID based on the classical minimum mutual information. Second a sophisticated QPID enabled by our development of an operational partial order on quantum ensembles in terms of ensemble (inter-)convertability. Finally, as a concrete example of how QPIDs can be applied, we investigate objectivity scenarios; using our Holevo QPIDs to directly quantify objective information in extended graph states. These findings ultimately establish a foundation upon which generic multipartite QPIDs can be systematically constructed, introduce an operational partial order on ensembles, advance QPIDs as interesting objects of study in their own right, and demonstrate their immediate applicability to problems in quantum information.
AI Summary
The paper asks how partial information decomposition (PID)—separating redundant, unique, and synergistic information—can be extended from bipartite to arbitrary multipartite quantum systems. It retains the classical antichain redundancy lattice and Möbius-inversion construction, placing the quantum-specific choices in the redundancy function. Two decompositions of the Holevo quantity are defined. The first assigns redundancy $\chi_\cap^{(1)}(\alpha)=\min_{P\in\alpha}\chi(\eta(P))$. The second maximizes $\chi(\mu)$ over ensembles obtainable from every input ensemble through a message-independent quantum channel and classical label garbling. Data processing establishes self-redundancy and the bound $\chi_\cap^{(2)}\leq\chi_\cap^{(1)}$. Ensemble convertibility is tested using alternating semidefinite optimizations. For binary qubit output ensembles, unitary equivalence reduces a discretized search from $O(\nu^7)$ to $O(\nu^4)$, or $O(\nu^3)$ with fixed probabilities. Applications use an anisotropic graph-state model with encoding errors, local rotations, and environmental scrambling. Bipartite decompositions display redundancy plateaus and redistribution into synergistic information under scrambling. Multipartite calculations identify the lowest lattice atom with information shared by all individual environment sites. For $N=8,10,12,14$, the difference between the two redundancy measures reaches approximately $0.04$ bits in the low-scrambling regime. Limitations include double-exponential growth of the full redundancy lattice, nonconvex optimization with possible local minima, and restriction of numerical output searches to binary qubit ensembles, generally giving lower bounds. The finite-size results are described as indicative. Holevo-based decompositions do not establish information accessibility or absence of quantum correlations; the application therefore addresses standard rather than strong quantum Darwinism.
Central claim may not hold: Independently garbling the target labels lets the proposed processing-based redundancy assign one bit to two independent perfectly readable bits, undermining its claimed distinction between shared information and equal information amounts. The stated classical-channel Kraus operators are generally not trace preserving, although separately indexing them by input and output labels repairs the definition and matches the later SDP. Processing-inferiority is a preorder on the stated ensembles, not a partial order, because distinct unitarily related ensembles are mutually convertible. The restricted, discretized output search with local feasibility optimization does not certify that the reported numerical redundancies equal the unrestricted maximum in the definition. For fixed finite dimensions the true minimum processing error is continuous, so discontinuities in see-saw results do not establish discontinuities of that minimum.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 18 pages, 7 figures, 2 tables