betaFeedback

WorksPhysics and Astronomy

No complete set of mutually unbiased bases in dimension ten contains the Fourier pair or three product bases

Abstract

With computer-assisted proofs that C6\mathbb{C}^6 admits at most three mutually unbiased bases (MUBs) now available, C10\mathbb{C}^{10} is the smallest dimension in which it is unknown whether a complete set of d+1=11d+1=11 MUBs exists; only three MUBs are known. We prove that every triple of mutually unbiased product bases of C2⊗C5\mathbb{C}^2\otimes\mathbb{C}^5 is strongly unextendible: no vector is unbiased to all three bases. Consequently no set of four MUBs in dimension ten, and in particular no complete set, contains three product bases. This settles the case d=2×5d=2\times5 anticipated by McNulty, Pammer and Weigert; it was known only for the Heisenberg-Weyl triple. The proof covers all ten equivalence classes with explicit Nullstellensatz certificates modulo one prime on 9191 charts of P9×P9\mathbb{P}^9\times\mathbb{P}^9 for each class (910910 in all), transferred to characteristic zero by a projective reduction lemma, together with an independent rank certificate for eight of the ten classes. We also prove that the pair formed by the standard and the Fourier basis of C10\mathbb{C}^{10} is contained in no set of four MUBs. Exactly 30403040 rays are unbiased to both bases (this confirms the count found numerically by Führ and Rzeszotnik), each of them is unbiased to at most eight of the others, and every triple formed by the pair and a third MUB is strongly unextendible. For this we certify 34,94034,940 cyclic 1010-roots in Krawczyk uniqueness boxes, decide unimodularity exactly through an involution, and prove completeness by a four-dimensional interval branch-and-bound obtained from the Chinese-remainder splitting Z10=Z2×Z5\mathbb{Z}_{10}=\mathbb{Z}_2\times\mathbb{Z}_5.

Introduction

Two orthonormal bases of Cd\mathbb{C}^{d} are mutually unbiased if every inner product between a vector of one and a vector of the other has modulus d−1/2d^{-1/2}. At most d+1d + 1 bases of Cd\mathbb{C}^{d} can be pairwise mutually unbiased, and complete sets of d+1d + 1 mutually unbiased bases (MUBs) exist whenever dd is a prime power [16, 42]. Let N(d)N(d) denote the maximal size of a set of MUBs in Cd\mathbb{C}^{d}. For every other dd the value of N(d)N(d) was open, and the smallest case, d=6d = 6, resisted for decades [8]. In the autumn of 2026 several computer-assisted proofs of N(6)=3N(6) = 3 appeared [31, 17, 7, 18]. Dimension ten is now the smallest dimension in which N(d)N(d) is unknown: 3≤N(10)≤113 \le N(10) \le11. By Weiner’s gap theorem [41], N(10)≤9N(10) \le9 is equivalent to the non-existence of eleven MUBs.

Two families of configurations have shaped the work on dimension six, and both have natural analogues in dimension ten.

  • Product bases. For C6=C2⊗C3\mathbb{C}^{6} = \mathbb{C}^{2} \otimes\mathbb{C}^{3}, McNulty and Weigert classified all mutually unbiased product bases, showed that no triple of them extends, and that a complete set contains at most one product basis [25, 26, 24]. For C2⊗C5\mathbb{C}^{2} \otimes\mathbb{C}^{5}, McNulty, Pammer and Weigert showed that there are at most three mutually unbiased product bases and brought every such triple into a normal form [28 Thm. 2, Cor. 3]. Recalling that in dimension six no vector is unbiased to three mutually unbiased product bases, they wrote: “We expect similar results to hold for larger product dimensions such as d=2×5d = 2 \times5, but we have not been able to generalize the proof for d=6d = 6” [28 after Cor. 6]; see also [40, 6]. For C10\mathbb{C}^{10}, Grassl, McNulty, Mišta and Paterek showed by computer algebra that the Heisenberg–Weyl triple {Z,X,XZ}\{Z, X, XZ\} is strongly unextendible [12] (see also [23]). The recent review by McNulty and Weigert asks whether the sets of mutually unbiased product bases giving the lower bound on N(d)N(d) in composite dimensions are unextendible, and records only d=6d = 6 as solved [30 Problem 10.8].

  • The Fourier pair. Grassl found that exactly 48 vectors are unbiased to both the standard basis II and the Fourier basis F6F_6, that they form 16 bases, and that no further vector is unbiased to II, F6F_6 and any of them [14], [30 Thm. 7.6]; hence no four MUBs contain (I,F6)(I,F_6) (see also [2], who note that Björck and coworkers had found the same vectors). Jaming, Matolcsi, Móra, Szöllősi and Weiner extended this to the two-parameter Fourier family [19]. The vectors unbiased to II and FdF_d are the biunimodular sequences of length dd, equivalently the unimodular cyclic dd-roots [3, 4, 15, 11]. For d=10d = 10 the cyclic system has 34,940 roots, a count obtained by Gröbner-basis computations modulo primes and by numerical homotopy methods [9, 10, 35, 37]. Führ and Rzeszotnik found numerically 3040 unimodular ones in ten symmetry orbits [11], and Magsino and Xu found the same number [22]. Grassl and Szöllősi announced a complete set of algebraic solutions to the cyclic-10 problem, noting that together with II and F10F_{10} they yield triples of MUBs [13]; the announcement does not address extensions to four MUBs, and to our knowledge the consequences for extending (I,F10)(I,F_{10}) have not appeared in print. Sun recently studied completions of Fourier-type pairs in dimension 2m2m [36].

We settle both questions in dimension ten. A set of bases is strongly unextendible if no unit vector is unbiased to all of its bases. We identify C10=C2⊗C5\mathbb{C}^{10} = \mathbb{C}^{2} \otimes\mathbb{C}^{5}; a product basis is an orthonormal basis consisting of product vectors a⊗ba \otimes b.

Theorem A (product triples). Every triple of pairwise mutually unbiased product bases of C2⊗C5\mathbb{C}^{2} \otimes\mathbb{C}^{5} is strongly unextendible.

In particular, Theorem A answers [30 Problem 10.8] for d=10d = 10, in the strong sense.

Theorem B (the Fourier pair). Let II be the standard basis and F10F_{10} the Fourier basis of C10\mathbb{C}^{10}. Then:

  • (i) exactly 3040 rays are unbiased to both II and F10F_{10};

  • (ii) each of these rays is unbiased to at most 8 of the others; 2240 of them are unbiased to none;

  • (iii) there are no four MUBs in C10\mathbb{C}^{10} containing II and F10F_{10};

  • (iv) there are at most 1136 bases unbiased to both II and F10F_{10}, and for every such basis BB the triple {I,F10,B}\{I,F_{10},B\} is strongly unextendible.

Theorem B(iv) is the ten-dimensional analogue of Grassl’s result for (I,F6)(I,F_6).

Corollary C. No set of four MUBs in C10\mathbb{C}^{10}, and in particular no complete set, contains (a) three bases that are product bases for one tensor factorisation C10≅C2⊗C5\mathbb{C}^{10} \cong\mathbb{C}^{2} \otimes\mathbb{C}^{5}, or (b) a pair of bases equivalent to (I,F10)(I,F_{10}).

Here two pairs of bases are equivalent if one common unitary or antiunitary map carries the first pair to the second, up to phases, the order of the vectors and the order of the two bases; such maps preserve all moduli of inner products. Corollary C(a) follows from Theorem A because a fourth MUB would consist of vectors unbiased to all three product bases, and (b) is Theorem B(iii). Together with [28 Thm. 2], (a) lowers the upper bound on the number of product bases (for one factorisation) in a set of four or more MUBs in C2⊗C5\mathbb{C}^{2} \otimes\mathbb{C}^{5} from three to two. Theorem A is the statement expected in [28 after Cor. 6].

The two results meet in one configuration. Under the Chinese-remainder identification Z10≅Z2×Z5\mathbb{Z}_{10} \cong\mathbb{Z}_{2} \times\mathbb{Z}_{5} the pair (I,F10)(I,F_{10}) is a pair of product bases, and the 40 product rays unbiased to it are the Gaussian chirps eπi(cn2+2bn)/10/10e^{\pi i(cn^{2}+2bn)/10}/\sqrt{10} with c∈{1,3,7,9}c \in\{1,3,7,9\}, b∈Z10b \in\mathbb{Z}_{10} (Remark 3). Theorem A excludes a fourth basis for every triple of product bases. Theorem B shows that for this particular pair every third basis, product or not, already blocks a fourth.

Neither result decides whether N(10)≤9N(10) \le9: a complete set need contain neither the Fourier pair nor three product bases. We return to this point in § 4.

Certification standard. Every computational step that the proofs use is exact (integer arithmetic, in particular modulo one prime, with identities that an independent program re-expands) or rigorous interval arithmetic with outward rounding. Non-rigorous floating-point computations were used only to find certificates and starting points, never to validate them. The root enclosures, the unimodularity decision and the overlap graphs were computed twice, with independent code and arithmetic, and the graphs agree bit for bit. Two steps have a single implementation: the clique enumeration on the conservative graphs (exact combinatorics) and the completeness step of Theorem B (§ 3.4). The logic of an earlier version of the latter was audited independently, and the audit’s blocking findings were fixed in the version used for all runs; we split it into 256 independent subproblems and record each completion. Theorem A uses as external input only the classification of [28], with the normalisation step from [5]; Theorem B uses none.

Product triples in C2⊗C5\mathbb{C}^{2} \otimes\mathbb{C}^{5}

Classification

Let ζ=e2πi/20\zeta= e^{2\pi i/20}, ω=ζ4\omega= \zeta^{4} and i=ζ5i = \zeta^{5}. The bases of C5\mathbb{C}^{5} that occur are G∞=I5G_{\infty}=I_{5} and GtG_t, t∈F5t \in\mathbb{F}_{5}, with entries Gt(x,j)=ωtx2+jx/5G_t(x,j)=\omega^{tx^{2}+jx}/\sqrt{5}; together they form the complete set of six MUBs of C5\mathbb{C}^{5}. The qubit vectors are the eigenvectors qb,0,qb,1q_{b,0},q_{b,1} of the Pauli operators b∈{Z,X,Y}b \in\{Z,X,Y\}. A label assignment ℓ=(ℓZ,0,ℓZ,1,ℓX,0,ℓX,1,ℓY,0,ℓY,1)\ell=(\ell_{Z,0},\ell_{Z,1},\ell_{X,0},\ell_{X,1},\ell_{Y,0},\ell_{Y,1}) with labels in P1(F5)={∞,0,1,2,3,4}\mathbb{P}^{1}(\mathbb{F}_{5})=\{\infty,0,1,2,3,4\} defines three product bases

Ub=[qb,0⊗Gℓb,0, qb,1⊗Gℓb,1],b∈{Z,X,Y}.(1)U_b=\left[q_{b,0}\otimes G_{\ell_{b,0}},\ q_{b,1}\otimes G_{\ell_{b,1}}\right],\qquad b\in\{Z,X,Y\}. \tag*{(1)}

They are pairwise mutually unbiased if and only if the three label sets {ℓb,0,ℓb,1}\{\ell_{b,0},\ell_{b,1}\} are pairwise disjoint; the two labels within one pair may coincide.

By the classification of McNulty, Pammer and Weigert [28 Cor. 3], every triple of mutually unbiased product bases of C2⊗C5\mathbb{C}^{2}\otimes\mathbb{C}^{5} is equivalent to one of the form (1). Their equivalences include local complex conjugations. On the normal forms these are realised by global maps: since Gt(i,j)=G_t(i,j)= G−t(:,−j)G_{-t}(:, -j) and M2Gt(:,j)=G−t(:,3j)M_{2}G_{t}(:,j)=G_{-t}(:,3j) with M2∣x⟩=∣2x⟩M_{2}|x\rangle=|2x\rangle, conjugation of the C5\mathbb{C}^{5} factor alone acts on the three bases as the unitary I2⊗M2I_{2}\otimes M_{2}, and conjugation of the qubit factor alone acts as the antiunitary (I2⊗M2)K(I_{2}\otimes M_{2})K, where KK is global complex conjugation. A simultaneous normalisation step uses the fact that in C5\mathbb{C}^{5} every pair of MUBs has a unique completion and the only rays unbiased to I5I_{5} and F5F_{5} are the 20 columns of G1,…,G4G_{1},\ldots,G_{4} [5 Sec. 4.1]. Hence strong unextendibility, which is invariant under common unitary or antiunitary maps, phases and permutations of vectors, needs to be checked only on the assignments (1).

There are ∑r=03(3r)(6)6−r=4080\sum_{r=0}^{3}\binom{3}{r}(6)_{6-r}=4080 admissible assignments, where rr is the number of bases whose two labels coincide and (n)k=n(n−1)⋯(n−k+1)(n)_{k}=n(n-1)\cdots(n-k+1). The qubit Clifford group (generated by HH and SS, acting on the six qubit rays), the action of PSL⁡(2,5)\operatorname{PSL}(2,5) on the labels (generated by F5F_{5}, D=diag⁡(ωx2)D=\operatorname{diag}(\omega^{x^{2}}) and M2M_{2}), and global conjugation act on them by explicit integer permutations. The 4080 assignments form ten orbits (Table 1); we checked this by enumerating all 666^{6} six-tuples with integer arithmetic.

orbitZZ labelsXX labelsYY labelssizesecond proof
0∞,∞\infty,\infty0,00,01,11,160—
1∞,∞\infty,\infty0,00,02,22,260—
2∞,∞\infty,\infty0,00,01,21,2720rank 975
3∞,∞\infty,\infty0,00,01,41,4180rank 975
4∞,∞\infty,\infty0,00,02,32,3180rank 975
5∞,∞\infty,\infty0,10,12,32,3720rank 975
6∞,∞\infty,\infty0,10,12,42,4720rank 975
7∞,∞\infty,\infty0,20,21,31,3720rank 975
8∞,0\infty,01,21,23,43,4480rank 987
9∞,0\infty,01,41,42,32,3240rank 987

Table 1. The ten classes of mutually unbiased product triples in C2⊗C5\mathbb{C}^{2}\otimes\mathbb{C}^{5}. Orbit 0 is the Heisenberg–Weyl triple of [12]. All ten are covered by the chart certificates of § 2.5; the last column refers to the independent rank certificate of § 2.6.

Exact coefficients

Write Tt,s(j,k)=5 (Gt∗Gs)jkT_{t,s}(j,k)=\sqrt{5}\,(G_{t}^{*}G_{s})_{jk} for t≠st\ne s and let χ\chi be the quadratic character of F5\mathbb{F}_{5}. Completing the square in the quadratic Gauss sum gives

Tt,∞(j,k)=ω−tk2−jk,T∞,s(j,k)=ωsj2+kj,Tt,s(j,k)=χ(s−t) ω−(k−j)2/(4(s−t))(s≠t∈F5).(2)\begin{aligned} T_{t,\infty}(j,k)&=\omega^{-tk^{2}-jk},\qquad T_{\infty,s}(j,k)=\omega^{sj^{2}+kj},\\ T_{t,s}(j,k)&=\chi(s-t)\,\omega^{-(k-j)^{2}/(4(s-t))}\qquad(s\ne t\in\mathbb{F}_{5}). \tag*{(2)} \end{aligned}

For an assignment put C=10[UY∗UZUX∗UZ]C=\sqrt{10}\begin{bmatrix}U_{Y}^{*}U_{Z}\\U_{X}^{*}U_{Z}\end{bmatrix}, a 20×1020\times10 matrix. By (2) each entry of CC is a twentieth root of unity ζEak\zeta^{E_{ak}} with an explicit integer exponent table EE. All computations below start from these integer tables; no floating-point quantity enters any certificate.

Polynomial formulation

Let ψ\psi be a unit vector unbiased to UZ,UX,UYU_{Z},U_{X},U_{Y} and put z=10UZ∗ψz=\sqrt{10}U_{Z}^{*}\psi. Unbiasedness to UZU_{Z} says ∣zk∣2=1|z_{k}|^{2}=1 for all kk; unbiasedness to the aa-th vector of UXU_{X} or UYU_{Y} says ∣∑kCakzk∣2=10\left|\sum_{k}C_{ak}z_{k}\right|^{2}=10. Replacing $\bar z by a second vector of unknowns ww, the diagonal terms contribute ∑kzkwk=10\sum_k z_kw_k = 10, so the conditions become

zkwk=z0w0(1≤k≤9),fa(z,w)=∑k≠lζEak−Ealzkwl=0(1≤a≤20).(3)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). \tag*{(3)}

These equations are bihomogeneous of bidegree (1,1)(1,1) and define a closed subscheme V⊂P9×P9\mathcal{V} \subset\mathbb{P}^9 \times\mathbb{P}^9 over Z[ζ,1/5]\mathbb{Z}[\zeta,1/5]. The point ([z],[zˉ])([z],[\bar z]) lies in V(C)\mathcal{V}(\mathbb{C}). Hence Theorem A follows once V(C)=∅\mathcal{V}(\mathbb{C}) = \varnothing for the ten representatives.

Reduction to one prime

Fix p=1073741741≡1(mod20)p = 1073741741 \equiv1 \pmod{20} and the residue map ζ↦a=95182621\zeta\mapsto a = 95182621, a primitive 20th root of unity modulo pp.

Lemma 1 (projective reduction). If V(F‾p)=∅\mathcal{V}(\overline{\mathbb{F}}_p) = \varnothing under ζ↦a\zeta\mapsto a, then V(C)=∅\mathcal{V}(\mathbb{C}) = \varnothing.

Proof. A complex point gives a point over a number field L⊇Q(ζ)L \supseteq\mathbb{Q}(\zeta). Choose a prime P\mathfrak{P} of LL above the prime (p,ζ−a)(p,\zeta-a). Scale each of the two homogeneous coordinate vectors so that all coordinates are P\mathfrak{P}-integral and one is a unit; the bihomogeneous equations are preserved. Reducing modulo P\mathfrak{P} gives two nonzero vectors over a finite field of characteristic pp satisfying the reduced system. □\square

Charts. On V\mathcal{V}, either z0w0≠0z_0w_0 \ne0, and we may scale to z0=w0=1z_0 = w_0 = 1 (the affine chart: zkwk=1z_kw_k = 1 and the twenty faf_a, in 18 variables), or z0w0=0z_0w_0 = 0, in which case all zkwk=0z_kw_k = 0; choosing zi≠0z_i \ne0 and wj≠0w_j \ne0 forces i≠ji \ne j, and we scale to zi=wj=1z_i = w_j = 1, wi=zj=0w_i = z_j = 0 (the 90 boundary charts). These 91 charts cover V\mathcal{V}. In each chart an identity 1=∑vhvgv1 = \sum_v h_vg_v over Fp\mathbb{F}_p among the chart equations gvg_v proves that the chart has no point over F‾p\overline{\mathbb{F}}_p.

The certificates

For each of the ten representatives and each of its 91 charts we produced such an identity, 910 in total.

  • The 900 boundary identities were obtained with the lift command of Singular [39] after msolve [29] and Singular had both found the reduced Gröbner basis {1}\{1\}.

  • For the ten affine charts Singular’s lift ran out of memory (at 6 GB per process; for orbit 0 also at 26 GB). We instead use the torus structure of the affine chart. Substituting wk=zk−1w_k = z_k^{-1} turns the generators into Laurent polynomials f~a=∑k≠lζEak−Ealzk/zl\widetilde{f}_a = \sum_{k\ne l}\zeta^{E_{ak}-E_{al}}z_k/z_l in z1,…,z9z_1,\ldots,z_9, and an identity 1=∑acaf~a1 = \sum_a c_a\widetilde{f}_a with Laurent cofactors of bounded degree is found by solving a sparse Macaulay-type linear system modulo pp, restricted to a symmetry-adapted subspace; this takes a few seconds and less than 0.1 GB per orbit (for orbits 0, 1 the system has 1789 rows). Each Laurent identity is then rewritten exactly as a polynomial identity 1=∑ktk(zkwk−1)+∑ahafa1 = \sum_k t_k(z_kw_k - 1) + \sum_a h_af_a in the 18 affine variables, of the same form as the Singular output.

The certificates were verified independently of how they were found. A short program expands every product hvgvh_vg_v with integer arithmetic modulo pp and checks that the sum is the constant 1. It reads the chart equations from inputs that a verification script regenerates from the exact exponent tables, and it requires that these inputs coincide byte for byte with the inputs given to the certificate search. All 910 identities pass (the largest affine identity has 479,691 cofactor terms). Modified certificates, such as swapped cofactors, a single altered coefficient, or the certificate of one orbit applied to another orbit, are rejected.

Proof of Theorem A. By the classification, a unit vector unbiased to a mutually unbiased product triple gives, after a common unitary or antiunitary map, a unit vector unbiased to one of the ten representatives, hence a complex point of the corresponding V\mathcal{V}. The 91 verified chart identities show V(Fp)=∅\mathcal{V}(\mathbb{F}_p)=\varnothing, and Lemma 1 gives V(C)=∅\mathcal{V}(\mathbb{C})=\varnothing, a contradiction.

An independent second proof for eight classes

Work in R=Q(ζ)[c,z,w]/(zkwk−c)R=\mathbb{Q}(\zeta)[c,z,w]/(z_k w_k-c), bigraded by deg⁡zk=(1,0)\deg z_k=(1,0), deg⁡wk=(0,1)\deg w_k=(0,1), deg⁡c=(1,1)\deg c=(1,1). The Weyl operators I2⊗X5I_2\otimes X_5 and I2⊗Z5I_2\otimes Z_5 preserve each of UZ,UX,UYU_Z,U_X,U_Y as sets of rays and therefore permute the faf_a up to phases; the resulting action of W≅Z52W\cong\mathbb{Z}_5^2 on the balanced components Rn,nR_{n,n} commutes with reduction modulo pp, because the stabiliser orders 1,5,251,5,25 are units. If the multiplication map ⨁Rn−1,n−1⟶Rn,nW\bigoplus R_{n-1,n-1}\longrightarrow R^W_{n,n} by the WW-projected generators has full row rank modulo pp, a nonzero maximal minor lifts to characteristic zero and cn∈⟨fa⟩c^n\in\langle f_a\rangle. Evaluating at z=10Uz∗ψz=\sqrt{10}U_z^*\psi, w=z‾w=\overline{z}, c=1c=1 gives 1=01=0. At n=3n=3 the matrices have full row rank 975 (orbits 2–7) and 987 (orbits 8, 9); square witnesses are part of the data. For orbits 0, 1 the rank at n=3n=3 is 972 < 975, so this route does not apply to them; they are covered by the chart certificates above.

Remark 2 (Galois conjugate classes). The chart equations depend on the exponent table only through the row differences Eak−EalE_{ak}-E_{al}. The automorphism ζ↦ζ17\zeta\mapsto\zeta^{17} of Q(ζ)\mathbb{Q}(\zeta) maps the multiset of rows of the table of orbit 0, up to a phase per row, onto that of orbit 1, and similarly orbit 3 onto orbit 4 (checked with exact integer arithmetic). The corresponding systems are Galois conjugate, so their complex points correspond. Either member of each pair would therefore have sufficed.

The Fourier pair

Let ω10=e2πi/10\omega_{10}=e^{2\pi i/10} and F10=(ω10jk/10)j,kF_{10}=(\omega_{10}^{jk}/\sqrt{10})_{j,k}.

Cyclic 10-roots

Let ψ\psi be unbiased to II and F10F_{10}. Then z=10ψz=\sqrt{10}\psi has unimodular entries, and we fix the phase by z0=1z_0=1. Unbiasedness to F10F_{10} means ∣z^k∣2=∣∑nznω10−kn∣2=10|\widehat{z}_k|^2=|\sum_n z_n\omega_{10}^{-kn}|^2=10 for all kk. Since ∣z^k∣2=∑scsω10−ks|\widehat{z}_k|^2=\sum_s c_s\omega_{10}^{-ks} with the periodic autocorrelation cs=∑nzn+szn‾c_s=\sum_n z_{n+s}\overline{z_n} and c0=10c_0=10, this holds if and only if cs=0c_s=0 for s=1,…,9s=1,\ldots,9. Put xn=zn+1/znx_n=z_{n+1}/z_n (indices modulo 10). As zn‾=1/zn\overline{z_n}=1/z_n,

cs=∑nxnxn+1⋯xn+s−1=:Cs(x),∏nxn=1.c_s=\sum_n x_nx_{n+1}\cdots x_{n+s-1}=:C_s(x),\qquad\prod_n x_n=1.

Conversely, every root of the cyclic 10-roots system

C1(x)=⋯=C9(x)=0,x0x1⋯x9=1(4)C_1(x)=\cdots=C_9(x)=0,\qquad x_0x_1\cdots x_9=1 \tag*{(4)}

with all ∣xn∣=1|x_n|=1 defines z0=1z_0=1, zn+1=xnznz_{n+1}=x_nz_n, consistently around the cycle. This is Björck’s correspondence [3]: rays unbiased to II and F10F_{10} are in bijection with unimodular roots of (4).

The number of roots

The cyclic 10-roots system is known computationally to have 34,940 roots: Gröbner-basis computations modulo the prime 32003 [35] and over a prime field (typically F65521\mathbb{F}_{65521}) [10] give a zero-dimensional ideal with quotient of dimension 34,940, cyclic 10-roots is a standard benchmark of the F5 algorithm [9], numerical irreducible decomposition finds 34,940 isolated solutions [37], a Gröbner basis over Q\mathbb{Q} has been computed by multi-modular methods [32], and Grassl and Szöllősi announced a complete set of algebraic solutions [13]. We do not use this count: below we prove that there are at least 34,940 distinct roots, and in § 3.4 we prove directly that the unimodular ones are exactly the certified 3040.

Certified enclosures of 34,940 roots

We start from the 1747 approximate representatives published with PHCpack [33] and expand them by the dihedral symmetry of (4) to 34,940 approximate roots; these serve only as starting points. For each, after a Newton polish, we apply Krawczyk’s test [21, 27, 34]: with x~\tilde{x} the approximate root, Y≈J(x~)−1Y \approx J(\tilde{x})^{-1} and a box X∋x~X \ni\tilde{x},

K(X)=x~−YF(x~)+(I−YJ(X))(X−x~)⊂int⁡XK(X)=\tilde{x}-YF(\tilde{x})+\left(I-YJ(X)\right)(X-\tilde{x})\subset\operatorname{int}X

implies that XX contains exactly one root, and that this root lies in K(X)K(X). Complex rectangular interval arithmetic on C10\mathbb{C}^{10} encloses the real interval Jacobian on R20\mathbb{R}^{20}, so the real theorem applies. We ran two independent implementations:

  1. a C++ program with double-precision intervals, in which every operation is rounded to nearest and then widened outward by one unit in the last place, compiled without floating-point contraction;

  1. a SageMath program [38] in which the system and its Jacobian are built symbolically over Q\mathbb{Q} and evaluated in Arb ball arithmetic [1] at 160 bits, with dyadic boxes whose endpoints are exact.

Both certify all 34,940 boxes (the C++ enclosures have width below 6×10−126 \times10^{-12}), and both find the enclosures pairwise disjoint. Hence (4) has at least 34,940 distinct roots. The C++ program checks its floating-point environment (binary64, round to nearest, gradual underflow) at run time and records the compiler; both programs exit with an error unless all 34,940 boxes are certified, every root is classified, the enclosures are disjoint and the degree bound below holds.

Exact unimodularity. The map σ(x)=(1/xn‾)n\sigma(x)=(1/\overline{x_n})_n permutes the roots of (4), since Ck(σ(x))=C10−k(x)‾/∏nxnC_k(\sigma(x))=\overline{C_{10-k}(x)}/\prod_n x_n and ∏nσ(x)n=1\prod_n\sigma(x)_n=1. A root is unimodular if and only if it is fixed by σ\sigma. If the interval image σ(K)\sigma(K) of an enclosure lies in the uniqueness box XX, the root is fixed, by uniqueness in XX; if some enclosure of ∣xn∣2|x_n|^2 excludes 1, it is not. This decides every certified root: exactly 3040 are unimodular and 31,900 are not, in both implementations.

All biunimodular vectors of length ten

We now show, without using the total number of roots, that the 3040 certified rays are all rays unbiased to II and F10F_{10}.

Reduction. Identify Z10=Z2×Z5\mathbb{Z}_{10}=\mathbb{Z}_2\times\mathbb{Z}_5 by n↦(n mod 2,n mod 5)n\mapsto(n\bmod2,n\bmod5) and write zn=un mod 5z_n=u_{n\bmod5} for even nn and zn=vn mod 5z_n=v_{n\bmod5} for odd nn. With ω5=e2πi/5\omega_5=e^{2\pi i/5}, Uk=∑mumω5−3kmU_k=\sum_m u_m\omega_5^{-3km} and Vk=∑mvmω5−3kmV_k=\sum_m v_m\omega_5^{-3km}, the discrete Fourier transform of zz at the frequency k↔(k2,k5)k\leftrightarrow(k_2,k_5) is Uk5+(−1)k2Vk5U_{k_5}+(-1)^{k_2}V_{k_5}. Hence zz is biunimodular if and only if

∣um∣=∣vm∣=1,ℜ(UkVk‾)=0,∣Uk∣2+∣Vk∣2=10(k∈Z5).(5)|u_m|=|v_m|=1,\qquad\mathop{\Re}(U_k\overline{V_k})=0,\qquad|U_k|^2+|V_k|^2=10\qquad(k\in\mathbb{Z}_5). \tag*{(5)}

If Uk≠0U_k \ne0 the last two conditions give Vk=εkiUk(10/∣Uk∣2−1)1/2V_k = \varepsilon_k i U_k (10/|U_k|^2 - 1)^{1/2} with εk=±1\varepsilon_k = \pm1 (boxes on which UkU_k may vanish are treated separately in the search below). After normalising u0=z0=1u_0 = z_0 = 1, a choice of u1,…,u4u_1,\ldots,u_4 on the unit circle and of signs ε\varepsilon determines vv, and biunimodularity becomes the five equations ∣vm∣2=1|v_m|^2 = 1 in four real unknowns.

Symmetry. A modulation zn↦e2πian/10znz_n \mapsto e^{2\pi ian/10}z_n fixes z0z_0 and maps (Uk,Vk)(U_k,V_k) to (Uk−a,(−1)aVk−a)(U_{k-a},(-1)^aV_{k-a}); the dilation zn↦z3nz_n \mapsto z_{3n} maps UkU_k to U2kU_{2k}. Since ∑k∣Uk∣2=25\sum_k |U_k|^2 = 25, every biunimodular ray is the image of one with ∣U0∣2=max⁡k∣Uk∣2|U_0|^2 = \max_k |U_k|^2, ∣U1∣2=max⁡k≥1∣Uk∣2|U_1|^2 = \max_{k \ge1}|U_k|^2 and ε0=+1\varepsilon_0 = +1. The certified set is closed under these maps; we check this by mapping the root enclosures and finding their images inside certified uniqueness boxes.

Search. Each umu_m, m=1,…,4m=1,\ldots,4, is parametrised by two overlapping rational charts u=±((1−s2)+2is)/(1+s2)u = \pm((1-s^2)+2is)/(1+s^2), s∈[−1716,1716]s \in[-\frac{17}{16},\frac{17}{16}]; with the 242^4 sign patterns (ε1,…,ε4)(\varepsilon_1,\ldots,\varepsilon_4) this gives 16⋅16=25616 \cdot16 = 256 four-dimensional subproblems. A box is excluded if an interval or mean-value enclosure (computed with interval automatic differentiation) of some residual misses 0, or if the symmetry inequalities fail on the whole box. Where the direction of UkU_k is not resolved, or ∣Vk∣|V_k| is small, VkV_k is computed instead from two of the linear equations that ∣vm∣2=1|v_m|^2 = 1 imposes on it, and the conditions ∣Vk∣2=10−∣Uk∣2|V_k|^2 = 10 - |U_k|^2 and ℜ(UkVk‾)=0\Re(U_k\overline{V_k}) = 0 are kept as residuals. A box is covered if a Krawczyk test on a slightly inflated box certifies a unique zero of four of the residuals, and that zero is either excluded by the remaining residuals or mapped by xn=zn+1/znx_n = z_{n+1}/z_n into a certified uniqueness box of a cyclic 10-root. Every other box is bisected. All 256 subproblems terminate with no unresolved box: 3.1×10103.1 \times10^{10} boxes in the certifying runs, about 2000 core-hours in total including duplicated runs (the subproblems are independent; they were run on a server, a cluster and a workstation with three compilers, all from one source file identified by its SHA-256 hash).

Overlap graphs

For the 3040 rays we enclose all overlaps g(r,s)=∣⟨ψr,ψs⟩∣2g(r,s) = |\langle\psi_r,\psi_s\rangle|^2. Let OO and UU be the graphs of genuinely orthogonal and genuinely unbiased pairs. Let O+O^+ join r≠sr \ne s unless the enclosure of g(r,s)g(r,s) excludes 0, and let U+U^+ join them unless it excludes 1/101/10. Then O⊆O+O \subseteq O^+ and U⊆U+U \subseteq U^+. The enclosures have width at most 1.1×10−111.1 \times10^{-11}, and the excluded pairs are far from the thresholds: every pair not in O+O^+ has g≥8.0×10−7g \ge8.0 \times10^{-7}, every pair not in U+U^+ has ∣g−1/10∣≥6.1×10−5|g - 1/10| \ge6.1 \times10^{-5}, and no pair lies in both graphs. O+O^+ has 57,680 edges and U+U^+ has 3200. The edge sets computed by the two implementations agree bit for bit.

A symmetry group of the pair, of order 1600 [11] (generated by the cyclic shift, the modulation zn↦ω10nznz_n \mapsto\omega_{10}^n z_n, the dilation zn↦z3nz_n \mapsto z_{3n}, complex conjugation and the Fourier transform) acts on the rays; numerically the 3040 rays form ten orbits of the sizes found in [11]. Table 2 lists the degrees in the conservative graphs along this orbit decomposition. The proof uses only the degree bounds, not the orbits.

orbit sizedegree in O+O^{+}degree in U+U^{+}remark
20, 201440the 40 product rays (Gaussian chirps)
200, 200440
200280
400, 400528the only rays that may have unbiased partners
400, 400280
800280

Table 2. The 3040 rays unbiased to II and F10F_{10}, grouped by (numerically identified) symmetry orbit, with their degrees in the conservative graphs O+⊇OO^{+} \supseteq O and U+⊇UU^{+} \supseteq U. A degree 0 in U+U^{+} is exact; a positive degree is an upper bound.

Proof of Theorem B. (i) follows from § 3: the 3040 certified unimodular roots are distinct, and by § 3.4 every biunimodular ray is one of them; (ii) is read off from U+⊇UU^+ \supseteq U. For (iii), suppose I,F10,B,DI,F_{10},B,D are MUBs. All vectors of BB and DD are among the 3040 rays, and for any r∈Br \in B the ten rays of DD are unbiased to rr, so rr has degree at least 10 in UU, hence in U+U^+, contradicting (ii). For (iv), a basis BB unbiased to II and F10F_{10} is a 10-clique of the orthogonality graph O⊆O+O \subseteq O^+. An exhaustive enumeration finds exactly 1136 ten-cliques in O+O^+, so there are at most 1136 such bases, and for each of these cliques the common U+U^+-neighbourhood of its ten vertices is empty.

A vector unbiased to II, F10F_{10} and BB would be one of the 3040 rays and would lie in that common neighbourhood.

Remark 3. Under n=5a+6xn=5a+6x (mod 10) the chirp eπi(cn2+2bn)/10e^{\pi i(cn^{2}+2bn)/10} factorises as (ic(−1)b)aω4cx2+3bx(i^{c}(-1)^{b})^{a}\omega_{4cx^{2}+3bx}, a product of a qubit vector unbiased to ZZ and XX and a column of some GtG_t. These 40 rays are exactly the product rays unbiased to the pair; their mutual overlaps lie in {0,1/5,1}\{0,1/5,1\}. Among the 1136 candidate third bases, 16 consist of product rays, and numerically the other 1120 each lie inside a single orbit. All 1136 have pairwise inner products of modulus below 1.1×10−151.1 \times10^{-15} numerically; we have not certified their exact orthonormality, and Theorem B(iv) does not need it. The candidate with ten rays from the eigenbasis of XZXZ recovers the Heisenberg–Weyl triple of [12].

Remark 4 (second moments do not suffice). A complete set is a complex projective 2-design [20], so the 90 rays of the other nine bases would satisfy ∑rwr(∣r⟩⟨r∣)⊗2=I+S−∑v∈I∪F10(∣v⟩⟨v∣)⊗2\sum_{r} w_r (|r\rangle\langle r|)^{\otimes2}=I+S-\sum_{v\in I\cup F_{10}}(|v\rangle\langle v|)^{\otimes2} with wr∈{0,1}w_r\in\{0,1\}. The linear relaxation 0≤wr≤10\leq w_r\leq1 over the 3040 rays is feasible numerically (residual 6×10−116\times10^{-11}), even with no weight on product rays. The obstruction of Theorem B is therefore not visible at the level of the 2-design identity; it comes from the sparse unbiasedness graph.

Discussion

Corollary C removes two natural candidates for a complete set in C10\mathbb{C}^{10}, but it does not show N(10)≤9N(10)\leq9. Since the set of pairs of bases that extend to four MUBs is closed (the set of quadruples of MUBs is compact), Theorem B also gives an open neighbourhood of (I,F10)(I,F_{10}) in which no pair extends to four MUBs; making this neighbourhood explicit would be of interest. A complete set might contain neither the Fourier pair nor three product bases. Other pairs of the form (I,H)(I,H) with HH a complex Hadamard matrix of order ten generally come in continuous families, whose unbiased vectors need not be finite in number. In dimension six the result for F6F_6 [14] was later extended to the whole Fourier family [19] and eventually to all triples [7, 18]. Extending Theorem B from F10F_{10} to its affine families is the natural next step in dimension ten.

The certification techniques are standard (rigorous numerical certification was used for d=6d=6 in [19] (discretisation with explicit error bounds) and [7] (interval and ball arithmetic), and Gröbner-basis arguments in [14, 23]); we do not claim novelty for them. Two details may still be useful. Laurent Nullstellensatz certificates on the torus chart replace a memory-bound Gröbner lift by a small sparse linear solve. Certified root enclosures, an involution that detects the unimodular locus exactly, and a low-dimensional interval branch-and-bound from a Chinese-remainder splitting give finite, checkable lists of vectors unbiased to a pair whose unbiased set is finite, without using the degree of the system.

Data availability

The programs that generate the inputs and verify all certificates, the lists of SHA-256 checksums of every certificate file, the logs of the verification runs, the logs of all branch-and-bound runs and the script that merges them (it checks that every one of the 256 subproblems finished without an unresolved box, under one source file and one input file identified by their SHA-256 hashes) are included as ancillary files of the arXiv version. The certificates themselves (910 chart identities, eight rank witnesses and the root enclosures; about 0.6 GB compressed) will be deposited in a public repository, with the DOI added in a revised version; until then they are available from the authors on request.

Use of AI

Claude Opus 5.5 was used for coding and for optimizing the speed of the algorithms.

References

  1. [1]F. Johansson, Arb: efficient arbitrary-precision midpoint-radius interval arithmetic, IEEE Trans. Comput. 66 (2017), 1281–1292.arxiv.org/abs/1611.02831
  2. [2]I. Bengtsson, W. Bruzda, Å. Ericsson, J.-Å. Larsson, W. Tadej and K. Życzkowski, Mutually unbiased bases and Hadamard matrices of order six, J. Math. Phys. 48 (2007), 052106.DOI
  3. [3]G. Björck, Functions of modulus 1 on Zₙ whose Fourier transforms have constant modulus, and “cyclic n-roots”, in Recent Advances in Fourier Analysis and its Applications, NATO Adv. Sci. Inst. Ser. C 315, Kluwer, 1990, 131–140.
  4. [4]G. Björck and R. Fröberg, A faster way to count the solutions of inhomogeneous systems of algebraic equations, with applications to cyclic n-roots, J. Symbolic Comput. 12 (1991), 329–336, doi:10.1016/S0747-7171(08)80153-8.
  5. [5]S. Brierley, S. Weigert and I. Bengtsson, All mutually unbiased bases in dimensions two to five, Quantum Inf. Comput. 10 (2010), 803–820.arxiv.org/abs/0907.4097
  6. [6]D. Cariello, Schmidt rank constraints in quantum information theory, Lett. Math. Phys. 111 (2021), 86, doi:10.1007/s11005-021-01434-w.DOI
  7. [7]M. Cárdenes Wuttig and J. Tindall, Using topology to reveal all mutually unbiased bases in dimension six, arXiv:2611.06917 [quant-ph], 2026.arxiv.org/abs/2611.06917
  8. [8]T. Durt, B.-G. Englert, I. Bengtsson and K. Życzkowski, On mutually unbiased bases, Int. J. Quantum Inf. 8 (2010), 535–640.DOI
  9. [9]J.-C. Faugère, A new efficient algorithm for computing Gröbner bases without reduction to zero (F5), in Proc. ISSAC 2002, ACM, 2002, 75–83, doi:10.1145/780506.780516.DOI
  10. [10]J.-C. Faugère and J. Svartz, Gröbner bases of ideals invariant under a commutative group: the non-modular case, in Proc. ISSAC 2013, ACM, 2013, 347–354, doi:10.1145/2465506.2465944.DOI
  11. [11]H. Führ and Z. Rzeszotnik, On binunimodular vectors for unitary matrices, Linear Algebra Appl. 484 (2015), 86–129, doi:10.1016/j.laa.2015.06.019.DOI
  12. [12]M. Grassl, D. McNulty, L. Mišta and T. Paterek, Small sets of complementary observables, Phys. Rev. A 95 (2017), 012118, doi:10.1103/PhysRevA.95.012118.DOI
  13. [13]M. Grassl (joint work with F. Szöllősi), Circulant complex Hadamard matrices of order 10, invited talk, Hadamard 2020+2 workshop, Kraków, 27 June 2022, book of abstracts, p. 7 (announcement).
  14. [14]M. Grassl, On SIC-POVMs and MUBs in dimension 6, in Proc. ERATO Conference on Quantum Information Science, Tokyo, 2004, 60–61; arXiv:quant-ph/0406175.
  15. [15]U. Haagerup, Cyclic p-roots of prime lengths p and related complex Hadamard matrices, arXiv:0803.2629, 2008.arxiv.org/abs/0803.2629
  16. [16]I. D. Ivanović, Geometrical description of quantal state determination, J. Phys. A: Math. Gen. 14 (1981), 3241–3245.
  17. [17]M. Jaswal, An independent certified computation excluding four mutually unbiased bases in dimension six, Zenodo preprint, 2026, https://doi.org/10.5281/zenodo.23533384.
  18. [18]M. Jaswal, Triplets of mutually unbiased bases in dimension six, Zenodo preprint, 2026, https://doi.org/10.5281/zenodo.23991331.
  19. [19]P. Jaming, M. Matolcsi, P. Móra, F. Szöllősi and M. Weiner, A generalized Pauli problem and an infinite family of MUB-triplets in dimension 6, J. Phys. A: Math. Theor. 42 (2009), 245305.DOI
  20. [20]A. Klappenecker and M. Rötteler, Mutually unbiased bases are complex projective 2-designs, in Proc. 2005 IEEE International Symposium on Information Theory, 1740–1744, 2005.
  21. [21]R. Krawczyk, Newton-Algorithmen zur Bestimmung von Nullstellen mit Fehlerschranken, Computing 4 (1969), 187–201.DOI
  22. [22]M. Magsino and Y. Xu, A nonlinear sum of squares search for CAZAC sequences, in 2023 IEEE Radar Conference (RadarConf23), 2023, 1–6, doi:10.1109/RadarConf2351548.2023.10149667.DOI
  23. [23]P. Mandayam, S. Bandyopadhyay, M. Grassl and W. K. Wootters, Unextendible mutually unbiased bases from Pauli classes, Quantum Inf. Comput. 14 (2014), 823–844, doi:10.26421/QIC14.9-10-8.DOI
  24. [24]D. McNulty and S. Weigert, On the impossibility to extend triples of mutually unbiased product bases in dimension six, Int. J. Quantum Inf. 10 (2012), 1250056, doi:10.1142/S0219749912500566.DOI
  25. [25]D. McNulty and S. Weigert, All mutually unbiased product bases in dimension 6, J. Phys. A: Math. Theor. 45 (2012), 135307.DOI
  26. [26]D. McNulty, B. Pammer and S. Weigert, The limited role of mutually unbiased product bases in dimension 6, J. Phys. A: Math. Theor. 45 (2012), 102001.DOI
  27. [27]R. E. Moore, A test for existence of solutions to nonlinear systems, SIAM J. Numer. Anal. 14 (1977), 611–615.DOI
  28. [28]D. McNulty, B. Pammer and S. Weigert, Mutually unbiased product bases for multiple qudits, J. Math. Phys. 57 (2016), 032202, doi:10.1063/1.4943301.DOI
  29. [29]J. Berthomieu, C. Eder and M. Safey El Din, msolve: a library for solving polynomial systems, in Proc. ISSAC 2021, ACM, 2021, 51–58.
  30. [30]D. McNulty and S. Weigert, Mutually unbiased bases in composite dimensions – a review, Quantum 10 (2026), 2051, doi:10.22331/q-2026-04-01-2051.DOI
  31. [31]OpenAI, The maximum number of mutually unbiased bases in dimension six, OpenAI Math Release preprint, 2026, https://github.com/openai/math/tree/main/preprints/The-maximum-number-of-mutually-unbiased-bases-in-dimension-six-September-24-2026.
  32. [32]B. Parisse, Computing huge Gröbner basis like cyclic10 over ℚ with Giac, arXiv:1903.12427, 2019.arxiv.org/abs/1903.12427
  33. [33]J. Verschelde, Algorithm 795: PHCPack: a general-purpose solver for polynomial systems by homotopy continuation, ACM Trans. Math. Software 25 (1999), 251–276; cyclic 10-roots data at https://homepages.math.uic.edu/~jan/Demo/cyclic10.html.DOI
  34. [34]S. M. Rump, Solving algebraic problems with high accuracy, in A New Approach to Scientific Computation, Academic Press, 1983, 51–120.
  35. [35]A. Steel, Allan Steel’s Gröbner basis timings page, https://magma.maths.usyd.edu.au/users/allan/gb/ (accessed 16 November 2026).
  36. [36]H.-Y. Sun, Mutually unbiased bases: common states, completion, and extension obstructions, arXiv:2610.12180, 2026.arxiv.org/abs/2610.12180
  37. [37]A. J. Sommese, J. Verschelde and C. W. Wampler, Numerical decomposition of the solution sets of polynomial systems into irreducible components, SIAM J. Numer. Anal. 38 (2001), 2022–2046, doi:10.1137/S0036142900372549.DOI
  38. [38]The Sage Developers, SageMath, the Sage Mathematics Software System, https://www.sagemath.org.
  39. [39]W. Decker, G.-M. Greuel, G. Pfister and H. Schönemann, SINGULAR — a computer algebra system for polynomial computations, https://www.singular.uni-kl.de.
  40. [40]M. Wiśniak, T. Pater ek and A. Zeilinger, Entanglement in mutually unbiased bases, New J. Phys. 13 (2011), 053047, doi:10.1088/1367-2630/13/5/053047.DOI
  41. [41]M. Weiner, A gap for the maximum number of mutually unbiased bases, Proc. Amer. Math. Soc. 141 (2013), 1963–1969.arxiv.org/abs/0902.0635
  42. [42]W. K. Wootters and B. D. Fields, Optimal state-determination by mutually unbiased measurements, Ann. Physics 191 (1989), 363–381.DOI

Paper details

vlnc:W_hn5xapde
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Claude Opus 5.5

View statement
Looking for related papers…
Contents