Foundations of quantum mechanics · Theory and computer-assisted proof (AI-assisted)
No complete set of mutually unbiased bases in dimension ten contains the Fourier pair or three product bases
arXiv:2611.21188v1 ↗PDF ↗Submitted 16 November 2026
- Question
- Two bases of \(\mathbb{C}^d\) are mutually unbiased when every vector of one has overlap \(1/d\) with every vector of the other. At most \(d+1\) such bases fit, and complete sets exist in every prime-power dimension. After this autumn's computer-assisted proofs that dimension six allows only three, dimension ten is the smallest open case: three are known and eleven are not excluded. The two most natural seeds for a large set are product bases of \(\mathbb{C}^2\otimes\mathbb{C}^5\) and the pair formed by the standard and Fourier bases. Can either seed be extended to four mutually unbiased bases?
- Result
- Neither can. No vector is unbiased to all three bases of any triple of mutually unbiased product bases of \(\mathbb{C}^2\otimes\mathbb{C}^5\), so a set of four, and in particular a complete set, contains at most two product bases for a given factorisation. Exactly 3040 rays are unbiased to both the standard and the Fourier basis; each is unbiased to at most eight of the others, which rules out four bases containing the pair, and each of the at most 1136 possible third bases leaves no vector unbiased to all three. The first proof uses 910 polynomial infeasibility certificates checked modulo one prime; the second uses interval-arithmetic root enclosures and an exhaustive certified search.
The paper in detail
Mutually unbiased bases are the mathematical form of complementarity: preparing a state of one basis leaves every outcome of the other equally likely. They give optimal state tomography and underlie several key-distribution protocols. At most \(d+1\) bases of \(\mathbb{C}^d\) can be pairwise unbiased, and Ivanović, Wootters and Fields constructed complete sets in prime dimensions, later extended to prime powers. In every other dimension the maximum \(N(d)\) was unknown. For \(d=6\) several computer-assisted proofs of \(N(6)=3\) appeared in autumn 2026, one of them covered on this page on 5 November. That leaves \(d=10\) as the smallest open case, with \(3\le N(10)\le11\), and by Weiner's gap theorem \(N(10)\le9\) would already exclude eleven.
The first result concerns product bases of \(\mathbb{C}^2\otimes\mathbb{C}^5\), whose vectors all have the form \(a\otimes b\). McNulty, Pammer and Weigert showed that at most three such bases are pairwise unbiased and brought every such triple to a normal form: each basis pairs the two eigenvectors of one qubit Pauli operator with bases of \(\mathbb{C}^5\) from its complete set of six, labelled by the projective line over \(\mathbb{F}_5\). The 4080 admissible labellings fall into ten classes under the symmetries. For each class the authors prove that no unit vector is unbiased to all three bases. A fourth basis would consist of such vectors, so none exists, which is the statement McNulty, Pammer and Weigert anticipated in 2016.
The second result concerns the standard basis and the Fourier basis \(F_{10}\). A vector unbiased to both has entries of equal modulus and a discrete Fourier transform of equal modulus; Björck showed that these biunimodular sequences correspond to the unimodular roots of the cyclic 10-roots polynomial system. That system has 34,940 roots. The authors certify all of them in disjoint interval boxes, decide exactly which are unimodular (3040), and then prove by a separate exhaustive search that no biunimodular vector was missed. Enclosing all pairwise overlaps of the 3040 rays shows that 2240 are unbiased to none of the others and the rest to at most eight. A fourth basis would need ten rays unbiased to a given ray of the third, so none exists; the 1136 orthonormal ten-cliques bound the possible third bases, and none of them has a common unbiased neighbour.
Two remarks locate the obstruction. Under \(\mathbb{Z}_{10}\cong\mathbb{Z}_2\times\mathbb{Z}_5\) the Fourier pair is itself a pair of product bases, and the 40 product rays unbiased to it are Gaussian chirps, so the two results meet in one configuration. And the second-moment condition a complete set must satisfy, being a 2-design, has a fractional solution on the 3040 rays even without product rays: the obstruction is combinatorial, sitting in the sparse unbiasedness graph rather than in the design identity. Open next steps are extending the Fourier result to its affine families of Hadamard matrices, making explicit the neighbourhood of the Fourier pair in which no pair extends, and ultimately deciding \(N(10)\).
Transferable methods
Nullstellensatz certificates over a finite field
Unbiasedness is a condition on moduli, which is not polynomial in the vector's entries. Treating the complex conjugate as an independent unknown turns the question into whether a system of polynomial equations has a common zero. An identity \(1=\sum_vh_vg_v\) among the equations proves that it has none, and such an identity can be found and checked with exact integer arithmetic modulo a prime, provided reduction modulo that prime cannot lose a complex solution.
Let \(\psi\) be unbiased to \(U_Z,U_X,U_Y\) and put \(z=\sqrt{10}\,U_Z^*\psi\), so \(|z_k|^2=1\). Unbiasedness to row \(a\) of \(U_X\) or \(U_Y\) reads \(|\sum_kC_{ak}z_k|^2=10\), where every \(C_{ak}\) is a twentieth root of unity \(\zeta^{E_{ak}}\). The diagonal terms contribute \(\sum_k|z_k|^2=10\), so the off-diagonal part must vanish. Replacing \(\bar z\) by a second vector \(w\) gives the bihomogeneous system \[z_kw_k=z_0w_0\quad(1\le k\le9),\qquad f_a(z,w)=\sum_{k\ne l}\zeta^{E_{ak}-E_{al}}\,z_kw_l=0\quad(1\le a\le20),\] whose zero set \(\mathcal V\subset\mathbb P^9\times\mathbb P^9\) contains \(([z],[\bar z])\). The theorem follows once \(\mathcal V(\mathbb C)=\emptyset\) for the ten representatives.
Lemma 1 transfers emptiness from characteristic \(p=1073741741\equiv1\pmod{20}\), where \(\zeta\) maps to a primitive twentieth root of unity: a complex point lies over a number field; choose a prime above \(p\), scale each coordinate vector so that its entries are integral there and one is a unit, and reduce. The bihomogeneous equations survive, giving a nonzero point modulo \(p\). On \(\mathcal V\) either \(z_0w_0\ne0\) (one affine chart, scaled to \(z_0=w_0=1\)) or all \(z_kw_k=0\), and choosing \(z_i\ne0\), \(w_j\ne0\) with \(i\ne j\) gives 90 boundary charts. Each of the \(10\times91\) charts receives an identity \(1=\sum_vh_vg_v\) over \(\mathbb F_p\); for the affine charts, where Gröbner lifting ran out of memory, substituting \(w_k=z_k^{-1}\) and solving a sparse Macaulay-type linear system for Laurent cofactors found the identity in seconds. An independent program re-expands every product and checks that the sum is 1.
Editorial check: the twentieth-root exponents come from Gauss sums. For \(G_t(x,j)=\omega^{tx^2+jx}/\sqrt5\) with \(\omega=e^{2\pi i/5}\), completing the square gives \[\sqrt5\,(G_t^*G_s)_{jk}=\chi(s-t)\,\omega^{-(k-j)^2/(4(s-t))}\qquad(s\ne t),\] with \(\chi\) the quadratic character of \(\mathbb F_5\); a direct computation confirms this for all twenty ordered pairs. A numerical search over unit vectors for the triples of orbits 0, 2, 8 and 9 never brought the twenty unbiasedness residuals simultaneously below \(8\times10^{-3}\) in squared norm, while for any two of the bases the residual fell to rounding level. That is consistent with the certificates, though a search cannot prove absence.
How to reuse it
Questions of the form “is there a vector with prescribed overlaps” (extending sets of unbiased bases, SIC-POVM existence, unextendible product bases) become polynomial systems once the conjugate is an independent unknown. Certificates modulo a large prime are cheap to find and to check, and the projective reduction lemma makes them valid over \(\mathbb C\).
Certified roots of cyclic 10-roots and a Chinese-remainder search
To show that a finite set of vectors is the whole answer, one needs both a certificate that each listed solution exists and a proof that nothing else exists. Interval Newton tests give the first, an antiholomorphic involution decides exactly which certified roots are unimodular, and a low-dimensional exhaustive search, made possible by splitting \(\mathbb Z_{10}\) into \(\mathbb Z_2\times\mathbb Z_5\), gives completeness without relying on the known root count.
For unimodular \(z\) with \(z_0=1\), \(|\hat z_k|^2=\sum_sc_s\omega_{10}^{-ks}\) with the periodic autocorrelation \(c_s=\sum_nz_{n+s}\bar z_n\) and \(c_0=10\); all \(|\hat z_k|^2\) equal 10 exactly when \(c_s=0\) for \(s=1,\dots,9\). With \(x_n=z_{n+1}/z_n\) and \(\bar z_n=1/z_n\), \(c_s=\sum_nx_nx_{n+1}\cdots x_{n+s-1}\), which is the cyclic 10-roots system with \(\prod_nx_n=1\). Krawczyk's test certifies a root: if \[K(X)=\tilde x-YF(\tilde x)+\bigl(I-YJ(X)\bigr)(X-\tilde x)\subset\operatorname{int}X\] for an approximate root \(\tilde x\), an approximate inverse Jacobian \(Y\) and an interval Jacobian \(J(X)\), then \(X\) contains exactly one root. The map \(\sigma(x)=(1/\bar x_n)_n\) permutes the roots, and a root is unimodular exactly when it is fixed, which the boxes decide: 3040 of the 34,940 certified roots are fixed.
For completeness, write \(z_n=u_{n\bmod5}\) for even \(n\) and \(v_{n\bmod5}\) for odd \(n\), with five-point transforms \(U_k\) and \(V_k\). The ten-point transform at the frequency \((k_2,k_5)\) is \(U_{k_5}+(-1)^{k_2}V_{k_5}\), so biunimodularity requires \(|U_k+V_k|^2=|U_k-V_k|^2=10\). Subtracting and adding these gives \[\operatorname{Re}\bigl(U_k\overline{V_k}\bigr)=0,\qquad|U_k|^2+|V_k|^2=10,\] so \(V_k=\pm i\,U_k\,(10/|U_k|^2-1)^{1/2}\) whenever \(U_k\ne0\). Fixing \(u_0=1\) and the phases of \(u_1,\dots,u_4\) then determines \(v\) up to five signs, leaving five equations \(|v_m|^2=1\) in four real unknowns. An interval branch-and-bound over 256 four-dimensional subproblems excludes or certifies every box, \(3.1\times10^{10}\) boxes in total, and every certified zero lands in one of the 3040 known boxes.
Editorial check: solving \(|\hat z_k|^2=10\) by least squares from 4000 random phase vectors found 2042 distinct rays, and closing them under the cyclic shift, the modulation \(z_n\mapsto\omega_{10}^nz_n\), the dilations \(n\mapsto3n,7n,9n\), conjugation and the Fourier transform gave exactly 3040. Counting pairs with overlap \(1/10\) among them reproduces Theorem B(ii): 2240 rays have none and 800 have exactly eight, which is why no ray can sit in a basis with a fourth unbiased basis beside it.
How to reuse it
Whenever a set of solutions is expected to be finite, certified enclosures plus an involution that detects the real or unimodular locus give an exact list, and a group decomposition of the index set (here the Chinese remainder theorem) can turn completeness into a search of low enough dimension to finish. The same pattern applies to biunimodular vectors for other Hadamard matrices and to finite configurations in frame theory.
Context & limitations
Why it matters. The paper removes the two candidates that have shaped the search for a complete set in dimension ten, in the same way that the corresponding results for dimension six preceded its full resolution. It settles the product-basis case that McNulty, Pammer and Weigert expected but could not prove, and answers a problem from the recent McNulty–Weigert review for \(d=10\) in its strong form. The certification is exact or rigorous interval arithmetic throughout, with floating-point computation used only to find certificates.
Limitations. Neither result decides whether eleven mutually unbiased bases exist in dimension ten: a complete set might contain neither the Fourier pair nor three product bases, and other pairs built from complex Hadamard matrices come in continuous families with possibly infinitely many unbiased vectors. The product-basis theorem takes the classification of McNulty, Pammer and Weigert as input. The proofs are computer-assisted: the certificates (about 0.6 GB) are available on request until they are deposited, and two steps, a clique enumeration and the completeness search, have a single implementation, the latter audited independently. The authors state that Claude Opus 5.5 was used for coding and for speeding up the algorithms.