When can local unitaries generate a complete orthogonal basis?

The paper in detail

A pure state \(\ket\psi\) of \(n\) parties is locally encodable if there are \(D=\prod_jd_j\) local unitaries \(V_j=U^{(1)}_j\otimes\cdots\otimes U^{(n)}_j\) with \(\{V_j\ket\psi\}\) an orthonormal basis. Counting gives a heuristic: the \(D-1\) nontrivial encoders have \((D-1)\sum_j(d_j^2-1)\) real parameters and orthogonality imposes \(D(D-1)\) real constraints, so ULE should hold exactly when \[\sum_jd_j^2\ \ge\ \prod_jd_j+n.\] This always holds for two parties, and for equal local dimensions with three or more parties only for three qubits.

The group approach turns a per-state search into one structural check. If an abelian group \(G\) of \(D\) local unitaries has every element except the identity traceless, then for any state that is an equal-weight superposition of the common eigenbasis with arbitrary phases (a maximally coherent state), the orbit \(\{g\ket\psi\}\) is orthonormal. ULE then follows if every state is locally equivalent to such a state; for two parties this means every Schmidt vector must be reachable by choosing the phases.

For two qutrits the group generated by \(X\otimes X\) and \(Z\otimes Z^2\) has the generalised Bell basis as eigenbasis. The Schmidt spectrum is captured by \((\mathrm{Tr}B^2,\mathrm{Tr}B^3)\), with \(B=A^\dagger A\) and \(A\) the coefficient matrix, which ranges over a two-dimensional region with three boundary arcs: one eigenvalue zero, the two largest equal, the two smallest equal. Appendix C writes explicit phase families tracing each arc and glues them at \((1,0,0)\), \((\tfrac12,\tfrac12,0)\) and \((\tfrac13,\tfrac13,\tfrac13)\). A covering lemma then fills the interior.

Further constructions extend the picture. A lifting lemma combines encodings of \(d_A\times d_B\) and \(d_A'\times d_B\) (with \(d_A,d_A'\ge d_B\)) into \((d_A+d_A')\times d_B\). Hurwitz–Radon matrices give ULE whenever the Radon number of \(d_A\) is at least \(d_B\), and with care for convergence this extends to \(\infty\times\infty\). On the negative side, when the counting inequality fails the authors construct a one-uniform state with a discrete stabilizer, which by the Pauwels–Gühne argument cannot be encoded; three qutrits and four qubits are examples.

The operational result concerns estimating a Haar-random local unitary \(V\) from one copy of \(V\rho_0V^\dagger\) with an invariant loss. The optimal covariant POVM is the orbit of one rank-one seed \(\ket\eta\bra\eta\); if \(\ket\eta\) is locally encodable, the encoded basis is a projective measurement with the same loss. Conversely, for a non-encodable state and a small trace-distance tolerance, some POVM strictly beats every projective measurement. The open cases are \(5\times3\), general \(d\times d\) (conjectured via the Bell-basis group whenever \(\gcd(d_A,d_B)>1\)) and three qubits.

Transferable methods

1. Traceless abelian groups give orthonormal orbits

   Constructing \(D\) encoders for each state separately is hopeless in general. A group with traceless nonidentity elements makes orthogonality automatic for a whole family of states, leaving only the question of whether that family reaches every local-equivalence class.

   Let \(\{\ket{\psi_j}\}\) be the common eigenbasis of \(G\), with \(g\ket{\psi_k}=\lambda_k^{(g)}\ket{\psi_k}\), and \(\ket{\psi(\theta)}=D^{-1/2}\sum_je^{i\theta_j}\ket{\psi_j}\). Then \[\langle g_1\psi(\theta)\vert g_2\psi(\theta)\rangle=\frac1D\sum_k\lambda_k^{(g_1^{-1}g_2)}=\frac{\mathrm{Tr}(g_1^{-1}g_2)}{D}=\delta_{g_1,g_2},\] since the phases cancel on the diagonal and the trace vanishes unless \(g_1=g_2\).

   Editorial check: for \(d=3\) the nine operators \((X^iZ^j)\otimes(X^iZ^{2j})\) have trace moduli \(9,0,\dots,0\), all nine generalised Bell states are common eigenvectors, and for random phases the Gram matrix of the orbit differs from the identity by at most \(8\times10^{-16}\).

   How to reuse it

   When a basis must be generated from one vector by a restricted set of operations, look for an abelian group of allowed operations with vanishing characters off the identity; uniform superpositions of its eigenbasis then work for free, as in dense coding and unitary error bases.

2. Covering the Schmidt region by a boundary loop

   Showing that the phases reach every Schmidt vector would require solving polynomial equations for each target. It is enough to find a continuous loop of phases whose image runs once around the boundary of the region of possible spectra; a topological argument then forces every interior point to be hit.

   The region \(\mathcal T_3=\{(\sum\lambda_j^2,\sum\lambda_j^3)\}\) over probability vectors is a curved triangle with corners \((1,1)\), \((\tfrac12,\tfrac14)\) and \((\tfrac13,\tfrac19)\). Editorial reading of the covering step: the phase space \(\mathbb R^{3\times3}\) is contractible, so the loop of phases bounds a disc of phases. If the map to \(\mathcal T_3\) sent that disc into \(\mathcal T_3\) minus an interior point, the boundary loop, which winds once around that point, could be contracted in a punctured region, which is impossible.

   On the arc where the two smaller eigenvalues are equal, the construction gives \(B\) the repeated eigenvalue \((2\cos\phi_t+1)^2/27\), which decreases from \(\tfrac13\) to 0 as \(\phi_t\) runs from 0 to \(2\pi/3\).

   Editorial checks: the angles quoted in Appendix C.5 give Schmidt vectors \((1,0,0)\), \((\tfrac12,\tfrac12,0)\) and \((\tfrac13,\tfrac13,\tfrac13)\) to six digits; on the last arc at \(t=0.2,0.5,0.8\) the two smaller eigenvalues coincide with \((2\cos\phi_t+1)^2/27\); and 60 random Schmidt targets were each reached by optimising the nine phases, with largest squared residual \(9\times10^{-8}\).

   How to reuse it

   To show that a parametrised family reaches every point of a compact region, it can be easier to trace the boundary once with a contractible family of parameters than to invert the map; the same degree argument applies to reachable spectra, marginals or moment vectors.

Source: Sec. II.2, Eq. (1); Proposition IV.1, Proposition IV.2(e), Theorem IV.3; Appendix C; Theorem V.4; Theorem VI.1. 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. Every state in these systems can carry a dense-coding-style encoding by local operations. The equivalence with measurement optimality ties a geometric property of local-unitary orbits to a question about when general measurements beat projective ones.

Limitations. The \(5\times3\) case, general \(d\times d\) and three qubits remain open; the general bipartite case is stated as a conjecture backed by numerical searches. That the parameter count also suffices for ULE is unproven. The two-qutrit proof is an existence argument through a topological covering lemma and gives encoders only implicitly, through phases chosen for each state. The authors state that ChatGPT (GPT-5.6 Sol and GPT-6.1 Sol) assisted with the operational equivalence, the link between regularity and the stabilizer, parts of the constructions (including phases on one boundary arc of the two-qutrit construction and parts of the infinite-dimensional one), a counterexample, and some technical proof steps.

Description based on the full arXiv HTML of v1, read 2026-11-14: Sections I–VII, Appendix C (the explicit two-qutrit construction) and the AI statement; Appendices A, B, D–F were not read beyond statements cited in the main text. The group conditions, the three gluing points and one boundary arc of the two-qutrit path, and coverage of the Schmidt simplex were checked numerically.