The Marginal Problem for Density Operators

Abstract: We study when local reduced density operators, viewed as quantum marginals, can be assembled into a global quantum state with a prescribed Markov structure. The starting point is a canonical logarithmic construction $T(\mathcal R)$, the noncommutative analogue of the junction-tree formula for decomposable graphical models. Unlike in the classical case, this formal construction may fail: noncommutativity can prevent it from being a normalized state with the prescribed marginals. We prove that this obstruction is captured exactly by a trace condition. For two overlapping marginals, and for clique marginals on a chordal graph, the condition $Tr(T(\mathcal R))=1$ is equivalent to the existence of a quantum Markov completion. When it exists, the completion is unique, equal to $T(\mathcal R)$, and selected by the maximum-entropy principle. In the two-clique case, we also give an equivalent conditional-reconstruction characterization: the two natural one-sided sandwich reconstructions agree if and only if the trace condition holds. We introduce the global quantum information $gI(\mathcal{G})_ρ$ associated with a chordal graph $\mathcal G$ and show that it is a relative-entropy discrepancy from $ρ$ to the logarithmic candidate, with a trace correction when the candidate is not normalized. We also prove an intersection property for strictly positive quantum conditional independence. Three-qubit Pauli examples show that the quantum obstructions are real: local consistency, feasibility, Markov feasibility, and maximum entropy can all separate.

• The paper establishes that, for consistent strictly positive marginals on two-clique or chordal graphs, a trace-one logarithmic candidate is equivalent to feasibility, quantum Markovity, marginal recovery, uniqueness, and maximum entropy.
• It develops noncommutative junction-tree formulas, divergence identities, and a global-information measure that quantify how noncommutativity obstructs normalization and classical-style marginal reconstruction.
• It proves the intersection property for strictly positive quantum conditional independence and uses Pauli examples to show that local consistency, feasibility, Markov feasibility, and maximum-entropy completion are distinct conditions.

Overview

The paper studies the quantum marginal problem for strictly positive density operators: given a family of reduced density operators (quantum marginals), when can they be assembled into a global state that satisfies a prescribed Markov structure? The authors, Lauritzen and Zwiernik, construct the noncommutative analogue of the classical junction-tree formula for decomposable graphical models and identify precisely when this construction succeeds. The central finding is that the entire classical picture survives in the quantum setting, but only under an additional trace-one condition on a canonical logarithmic operator $T(\mathcal R)$; without it, noncommutativity obstructs normalization, marginal recovery, and Markovity simultaneously.

The setting is finite-dimensional: Hilbert spaces $H_v$ indexed by a finite set $V$, with marginals given by partial traces. The paper works throughout with strictly positive density operators $\mathcal S_1^+(H)$, which permits use of matrix logarithms and the equality theory of quantum relative entropy. Two binary operations on positive operators play structural roles: $M\odot N=\exp(\log M+\log N)$, which is commutative and associative, and $M\star N=N^{1/2}MN^{1/2}$, which is neither but behaves well under partial trace and underlies conditional reconstruction.

The two-clique trace criterion

For a consistent pair of marginals $\rho_{A\cup C}$ and $\rho_{B\cup C}$ (agreeing on $C$), the canonical logarithmic candidate is

$$T(\mathcal R)=\exp\{\log\rho_{A\cup C}+\log\rho_{B\cup C}-\log\rho_C\},$$

the direct noncommutative analogue of the classical factorization $H_v$0. In general $H_v$1 need not be a density operator. The first main result establishes that $H_v$2 always holds — proved via Lieb's three-matrix inequality together with a pull-out property for partial traces — and that the following are equivalent: (i) $H_v$3; (ii) $H_v$4 has exactly the prescribed marginals; (iii) some completion exists satisfying quantum conditional independence $H_v$5. When these hold, the completion is unique and equals $H_v$6.

The proof rests on a divergence identity:

$H_v$7

where $H_v$8 by monotonicity of relative entropy under partial trace. Since both terms on the right are nonnegative, forcing them to vanish simultaneously pins down both feasibility and Markovity. This identity also yields uniqueness immediately: any Markov completion must coincide with $H_v$9 because Klein's inequality forces $V$0.

A complementary multiplicative characterization follows from Zhang's sandwich identities [1305.xxxx]: defining $V$1, the trace condition holds if and only if $V$2 is normal, in which case $V$3, i.e., the two one-sided conditional reconstructions $V$4 and $V$5 agree. Classically these reconstructions always agree; their agreement is thus exactly the noncommutative obstruction. The pair is then called Markov compatible over $V$6. Notably, the trace condition may hold even when the three operators do not commute, so commutativity is sufficient but not necessary.

Chordal graphs and the chordal trace criterion

The two-clique result extends to clique marginals on a chordal graph $V$7 with clique set $V$8 and separator set $V$9 with multiplicities $\mathcal S_1^+(H)$0. For a pairwise consistent family of strictly positive clique marginals, define

$\mathcal S_1^+(H)$1

Again $\mathcal S_1^+(H)$2, and the trace-one condition is equivalent to existence of a quantum Markov completion (a state whose global Markov property holds for $\mathcal S_1^+(H)$3, meaning separation in the graph implies vanishing conditional mutual information). When it exists, the completion is unique, equals $\mathcal S_1^+(H)$4, and is the unique maximum-entropy element among all completions, with entropy

$\mathcal S_1^+(H)$5

Two supporting results make this work. First, the paper proves a chordal analogue of strong subadditivity: every state satisfies the displayed entropy inequality, with equality if and only if it is quantum Markov on $\mathcal S_1^+(H)$6; the proof proceeds by induction along junction-tree decompositions using only the semi-graphoid axioms. Second, a general maximum-entropy theorem (not specific to chordal graphs) characterizes entropy maximizers over any feasible marginal family by a log-linear dual condition: $\mathcal S_1^+(H)$7 maximizes entropy over $\mathcal S_1^+(H)$8 if and only if $\mathcal S_1^+(H)$9 lies in the span of the local constraint spaces plus a scalar multiple of the identity. The proof uses a three-point Pythagorean identity for relative entropy, yielding the sharp gap formula $M\odot N=\exp(\log M+\log N)$0. An important caveat stated plainly: unlike the classical chordal case, where pairwise consistency always suffices, there is in general no simple criterion for $M\odot N=\exp(\log M+\log N)$1 to be nonempty unless the prescribed sets are cliques of a chordal graph.

Global information as divergence

The paper introduces the global quantum information

$M\odot N=\exp(\log M+\log N)$2

which interpolates between familiar quantities: it reduces to the conditional mutual information $M\odot N=\exp(\log M+\log N)$3 in the two-clique case and to Watanabe's total correlation (quantum multiinformation) for the empty graph. It is shown to be a relative-entropy discrepancy from $M\odot N=\exp(\log M+\log N)$4 to the logarithmic candidate built from its own clique marginals:

$M\odot N=\exp(\log M+\log N)$5

When the trace-one condition holds, the correction term vanishes and $M\odot N=\exp(\log M+\log N)$6, recovering exactly the classical form of connected-information decompositions organized through a chordal graph.

Intersection property

As a structural contribution needed for graphical arguments, the paper proves that for strictly positive density operators, entropic quantum conditional independence satisfies the intersection axiom ($M\odot N=\exp(\log M+\log N)$7 and $M\odot N=\exp(\log M+\log N)$8 imply $M\odot N=\exp(\log M+\log N)$9). This resolves affirmatively a question left open by Leifer and Poulin concerning whether the relation is a full graphoid. The proof is a variational argument based on the equality case of monotonicity of relative entropy under partial trace: each assumed independence forces the variational maximizers $M\star N=N^{1/2}MN^{1/2}$0 to be constant in one tensor coordinate, and constancy in both coordinates forces constancy in their product, giving the third independence. The appendix develops this via a weighted inner product on operator space, deriving both monotonicity and the Petz recovery formula for the partial trace in a self-contained finite-dimensional treatment.

Examples separating the classical equivalences

Three-qubit Pauli constructions demonstrate that four properties — local consistency, feasibility, Markov feasibility, and maximum-entropy completion — all separate. Prescribing $M\star N=N^{1/2}MN^{1/2}$1 and $M\star N=N^{1/2}MN^{1/2}$2, which are consistent on qubit 2:

• Local consistency does not imply feasibility: the marginals are feasible iff $M\star N=N^{1/2}MN^{1/2}$3, even though they are locally consistent for all $M\star N=N^{1/2}MN^{1/2}$4.
• Feasibility does not imply Markov feasibility: within the strictly feasible regime, a Markov completion exists iff $M\star N=N^{1/2}MN^{1/2}$5, since normality of $M\star N=N^{1/2}MN^{1/2}$6 forces the two anticommuting Pauli-supported marginals to commute.
• Maximum-entropy completions need not be Markov: the feasible completion $M\star N=N^{1/2}MN^{1/2}$7 is the unique maximum-entropy completion (its logarithm has the required log-linear form), yet fails to be Markov whenever $M\star N=N^{1/2}MN^{1/2}$8.
• Trace defect: explicitly, $M\star N=N^{1/2}MN^{1/2}$9 with $\rho_{A\cup C}$0 squared root, which equals 1 iff $\rho_{A\cup C}$1 — the trace defect quantifies the obstruction concretely.

A counterbalancing remark shows that Markov states need not have commuting overlapping marginals: product states $\rho_{A\cup C}$2 are Markov with noncommuting $\rho_{A\cup C}$3 and $\rho_{A\cup C}$4, so the obstruction in the example stems from the marginal problem rather than from Markovity per se.

Limitations and open questions

The results are confined to strictly positive density operators in finite dimensions; singular states, where logarithms and Petz-type reconstructions require approximation or support-restricted formulations, are outside scope. The trace criterion is established for pairs of overlapping marginals and for clique marginals of chordal graphs; no comparable characterization is given for general hypergraphs or non-chordal graphs, where even classically the existence of completions lacks a simple criterion. The paper also does not address computational aspects: deciding whether $\rho_{A\cup C}$5 for a given marginal family, or approximating the trace defect, remains open. Finally, whether the intersection property extends beyond strictly positive states, and how the graphoid structure interacts with other proposed definitions of quantum conditional independence, is left unresolved.

Conclusion

The paper gives a complete answer to the chordal quantum marginal problem for strictly positive states: the classical junction-tree construction has an exact noncommutative analogue, and a single scalar invariant — the trace of the logarithmic candidate — detects whether it is normalized, feasible, Markov, and maximum-entropic, all simultaneously. The accompanying divergence identity for $\rho_{A\cup C}$6 and the proof of the intersection axiom supply the graphical-model machinery needed to treat quantum states with chordal Markov structure, while the Pauli examples delineate sharply where the classical theory genuinely fails.