Foundations of quantum mechanics · Theory and numerical analysis

Solution of the Qubit Marginal Problem from Fermionic Representability

David A. Mazziotti

arXiv:2611.18503v1 ↗PDF ↗Submitted 13 November 2026

Question
Given few-qubit reduced states, is there a single global state, pure or mixed, that has them all as marginals? The question is QMA-complete in general, so practical answers are hierarchies of necessary conditions, and the issue is which conditions to impose and whether they converge. For fermions, the analogous \(N\)-representability problem of two-particle reduced density matrices already has a complete, constructive hierarchy. Does it carry over to qubits exactly?
Result
Yes. Encoding each qubit as one fermion in two orbitals, the representable \(p\)-qubit marginals are exactly the exposed face of the \(N\)-representable fermionic \(p\)-particle reduced density matrices on which every site-occupation variance \((\hat n_s-1)^2\) vanishes. The complete fermionic \((2,p)\)-positivity hierarchy therefore applies to qubits unchanged. In a Majorana basis, fermionic body order equals Pauli weight, and the conserved site occupations split each semidefinite program into polynomially many small blocks. On 12-qubit rings, the \((2,3)\) conditions are exact at the Majumdar–Ghosh point and within 0.075 J (0.5%) of the exact energy at the transverse-field Ising critical point.

The paper in detail

The quantum marginal problem asks whether a set of local reduced states is compatible with one global state. For qubits it underlies entanglement characterisation, the validation of data from quantum devices, and every simulation that targets few-body quantities instead of the full state. Since the general problem is QMA-complete, efficient methods are hierarchies of necessary conditions, and the questions are what conditions to use and whether the hierarchy becomes exact. Quantum chemistry faced the same problem for electrons, where one asks which two-particle reduced density matrices (2-RDMs) come from an \(N\)-electron state, and it developed a complete hierarchy: \((2,p)\)-positivity conditions, built from sums of squares of degree-\(p\) operators in which all terms above two-body cancel.

The paper places qubits inside that framework. Each qubit becomes one fermion shared between two orbitals on its site, and the Pauli operators become one-body operators confined to the site. These reproduce the qubit algebra exactly on the subspace where every site holds one fermion, and a Pauli string on \(p\) sites is a \(p\)-body operator. The theorem states that the representable \(p\)-qubit marginals form an exposed face of the convex set of \(N\)-representable fermionic \(p\)-RDMs, cut out by requiring the expectation of \(\sum_s(\hat n_s-1)^2\) to vanish. A complete description of the larger set therefore restricts to a complete description of the qubit set.

To use the hierarchy one needs to know which operator products generate high-body terms. In Majorana operators, four per site, the product of a site's four Majoranas acts as the identity on the singly occupied subspace, so a Pauli string of weight \(w\) is equivalent to a Majorana monomial of degree \(2w\). Cancelling three-body terms in a \((2,3)\) certificate is then the same as cancelling Pauli strings of weight three. The resulting semidefinite program maximises an energy shift \(E\) such that \(\hat H-E\) equals a two-body operator in the cone of certified squares, and its dual variables are the two-qubit marginals.

The fermionic generators temporarily empty or doubly fill a site, and each site's occupation is conserved by every qubit operator. Averaging over the phases \(e^{i\sum_s\theta_s\hat n_s}\) makes the certificate block diagonal by the pattern of occupation changes, and only changes of \(-1\), 0 or \(+1\) per site survive. For \(N=12\) the \((2,3)\) Gram matrix of dimension 15 712 splits into 24 blocks of size 68 and 1760 blocks of size 8, so the cost of a fixed level grows only polynomially with \(N\).

On the frustrated Heisenberg ring, \((2,2)\)-positivity underestimates the energy by about a third, while full \((2,3)\)-positivity errs by at most \(0.06\,J_1\) for \(J_2/J_1\le1/2\) and is exact at the Majumdar–Ghosh point, where the Hamiltonian is a sum of three-site positive terms. On the transverse-field Ising ring the critical-point errors are 0.40, 0.18 and 0.075 J for the \((2,2)\), partial \((2,3)\) and full \((2,3)\) levels. Open are the level needed for incommensurate spiral correlations and whether imposing these conditions on measured marginals actually mitigates hardware errors.

Transferable methods

  1. An exposed face cut out by a positive penalty operator

    The qubit states sit inside the fermionic states as exactly those annihilated by a positive two-body operator. A positive operator whose expectation is zero defines a supporting hyperplane, so the qubit marginals form a face of the fermionic set, and any complete description of the larger set restricts to the face.

    Each site occupation \(\hat n_s\) has eigenvalues 0, 1 and 2, so \(\hat Q_s=(\hat n_s-1)^2\succeq0\) vanishes exactly when site \(s\) is singly occupied. Because \(\hat Q=\sum_s\hat Q_s\) is two-body, its expectation is linear in the \(p\)-RDM for \(p\ge2\), and \[F^p_N=\bigl\{{}^pD\in P^p_N:\ \mathrm{Tr}(\hat Q\,{}^pD)=0\bigr\}\] is the intersection of \(P^p_N\) with a supporting hyperplane: an exposed face.

    If an \(N\)-fermion state \(\rho\) has \(\mathrm{Tr}(\hat Q\rho)=0\) with \(\hat Q\succeq0\), then \(\hat Q\rho=0\), so \(\rho\) lives on the singly occupied subspace and is an \(N\)-qubit state. On that subspace, a \(p\)-RDM element that changes any site occupation vanishes, and the others are combinations of Pauli strings of weight at most \(p\); conversely each such string is a \(w\)-body operator. So the face and the set of \(p\)-qubit marginals determine each other linearly.

    Editorial remark: the argument is a general lemma. For any convex set of expectation vectors and any positive operator whose expectation is a linear function of those vectors, the zero set is an exposed face equal to the image of the states supported on the operator's kernel, and the polar description of the face is the set of operators that are positive after compression to that kernel.

    How to reuse it

    Constrained Hilbert spaces (fixed occupation per site, hard-core constraints, code spaces) can be embedded in a larger representability problem with a known hierarchy and cut out by a positive penalty with zero expectation; the hierarchy then applies to the face without modification.

  2. Majorana compression and occupation-pattern block diagonalisation

    To use the hierarchy one must know which products create high-body terms, and the semidefinite programs must stay small. Compressing Majorana monomials to the qubit subspace turns body order into Pauli weight, and the conserved site occupations make the certificate block diagonal.

    With \(\hat\gamma_{2k-1}=\hat a_k+\hat a_k^\dagger\) and \(\hat\gamma_{2k}=-i(\hat a_k-\hat a_k^\dagger)\), the Majoranas anticommute and square to the identity. On the singly occupied subspace the site parity is fixed, so the product of the four Majoranas of a site is \(\pm1\) and \(\hat\sigma^z_s\equiv i\hat\gamma_{s1}\hat\gamma_{s2}\). A monomial compresses to zero if it has an odd number of Majoranas on some site, and otherwise to a Pauli string whose weight is half its degree. In a Gram certificate \(\sum_{IJ}X_{IJ}\hat O_I\hat O_J^\dagger\) every entry therefore feeds exactly one Pauli string, and two-body means that all strings of weight above two cancel.

    The rotations \(e^{i\sum_s\theta_s\hat n_s}\) leave \(\hat H\) invariant and map squares of degree-\(p\) polynomials to squares of the same degree, so \(X\) may be taken invariant, which zeroes every entry between generators with different occupation-change patterns. For \(p=3\) this leaves \[2N\ \text{blocks of size}\ 6N-4\quad\text{and}\quad8\binom N3\ \text{blocks of size}\ 8,\] which for \(N=12\) gives \(24\times68+1760\times8=15\,712\), the full Gram dimension (editorial recount).

    At the Majumdar–Ghosh point \(J_2=J_1/2\), \[\hat H=\frac{J_1}{4}\sum_i\bigl(\hat{\mathbf S}_i+\hat{\mathbf S}_{i+1}+\hat{\mathbf S}_{i+2}\bigr)^2-\frac{9}{16}J_1N,\] and each three-spin total spin squared is at least \(3/4\), so \(E\ge-\tfrac38J_1N\), attained by the dimer states. Positivity of three-qubit marginals, contained in the \((2,3)\) conditions, already enforces each term's bound, which is why that level is exact there. Editorial check: the operator identity holds exactly for \(N=12\), and exact diagonalisation gives \(-4.5\,J_1\).

    How to reuse it

    When a constrained problem carries extra conserved charges, symmetrise any sum-of-squares certificate under them; this replaces one large semidefinite block by many small ones. And a Hamiltonian that is a sum of positive \(k\)-site terms is solved exactly by any relaxation that includes positivity of \(k\)-site marginals.

Source: Theorem and proof; sections on Pauli strings as Majorana operators, site-occupation symmetry and results. Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. It connects the qubit marginal problem to a mature body of fermionic theory with a short exactness proof, so known hierarchies, solvers and structure (such as low-rank certificates) become available for qubits. The same conditions give rigorous lower bounds on spin-model energies, and because they hold for every state they can constrain noisy marginals measured on any qubit device.

Limitations. Exactness holds for the full hierarchy; any fixed level is a relaxation, and convergence may need levels that grow with the number of qubits, as QMA-completeness suggests. The numerics cover two 12-qubit rings, and in the spiral regime of the frustrated Heisenberg ring (\(J_2/J_1>1/2\)) the \((2,3)\) errors rise to 0.13–0.42 \(J_1\). The error-mitigation application is proposed, not demonstrated. The author states that Claude Opus 5.5 was used for scientific discussions and editing, and that all content was independently verified.

Description based on the full arXiv HTML of v1, read 2026-11-17: the whole Letter, including the theorem and proof, the Majorana dictionary, the site-occupation symmetry, the partial conditions, the results for both rings and the discussion; there is no supplement. Editorial checks: exact diagonalisation of the 12-qubit rings gives the Majumdar–Ghosh energy \(-4.5\,J_1\) and the critical Ising energy \(-15.3226\,J\) (so 0.075 J is 0.49%); the three-site decomposition of the Majumdar–Ghosh Hamiltonian holds exactly as an operator identity; the block count \(24\times68+1760\times8=15\,712\) was reproduced.