WorksPhysics and Astronomy
No complete set of mutually unbiased bases in C^6 without Fourier vanishing
Abstract
A recent exact certificate shows that C^6 admits no seven mutually unbiased bases, hence at most five (OpenAI, 2026). That argument has two parts. The first proves the Fourier-vanishing conjecture of Matolcsi, Ruzsa and Weiner for every complex Hadamard matrix of order six other than Tao's cubic matrix, and excludes the cubic matrix separately. The second feeds this vanishing into a positivity certificate on mixed moments of a hypothetical complete family. We show that the first part is not needed. We keep the same six families of test functions and the same linear identities, but we never set a moment to zero because of Fourier vanishing. New positive definite integer matrices still give the contradiction. The only inputs are orthonormality, unbiasedness, the second-moment (2-design) identity of a complete family, and the trivial bound |moment| <= 1. A standard-library program checks the certificate in exact integer and rational arithmetic in under a minute and a half. We also record computational observations that are not part of the proof: a systematic S_6-symmetric family of test functions appears to close the argument with relations of total order at most 7 instead of 11, and at order 9 two orbit types of test charges appear to suffice. In the prime-power dimensions 4,5,7,8, where complete families exist, the same pipeline finds no certificate, as it must.
Introduction
Two orthonormal bases of are mutually unbiased if every inner product between a vector of one and a vector of the other has modulus . At most bases of can be pairwise mutually unbiased, and complete families of such bases exist whenever is a prime power [2, 12]. Write for the maximal size of a mutually unbiased family in . Dimension six is the smallest case outside the prime-power constructions. Zauner conjectured that [13]. Weiner proved that mutually unbiased bases in always extend to a complete family [11]. Hence is equivalent to the non-existence of seven mutually unbiased bases in .
Several computer-assisted proofs addressing appeared in the autumn of 2026. OpenAI’s companion preprint [9] proves by a discretisation of the space of complex Hadamard matrices of order six. Its validity is stated under hypotheses on binary64 floating-point arithmetic and the compiler that ran the computation. Jaswal carried out an independent certified computation of that discretisation proof, again under a binary64-correctness hypothesis [3]. Cárdenes Wuttig and Tindall classify all triples of mutually unbiased bases in up to equivalence and show that none extends to a fourth basis; their proof is certified with interval arithmetic. They also give an analytic proof of the Fourier-vanishing conjecture discussed below [1]. Jaswal gives a further computer-assisted classification of triples and recovers that there are no four mutually unbiased bases [4].
The preprint [8] proves the weaker bound with a certificate that uses only integer and rational arithmetic and has no floating-point hypothesis at all. The bound is therefore not the news here. What is at stake is the structure of the exact certificate. That proof has two parts.
(A) Sections 3–5 of [8] prove the Fourier-vanishing conjecture of Matolcsi, Ruzsa and Weiner [7], which arose in the Fourier-analytic approach to mutually unbiased bases [6]: for and every coordinate permutation , for every complex Hadamard matrix of order six that is not equivalent to Tao’s cubic matrix [10]. The steps are an exact single-matrix identity [8 (10)], the dichotomy [8 Corollary 3.1], a classification of the cubic alternative [8 Section 4], and an exclusion of the cubic matrix from complete families [8 Proposition 5.2].
(B) Section 6 of [8] is a positivity certificate on mixed moments of a hypothetical complete family. It uses the vanishing as an input [8 (27)].
The point of this note is that part (A) is not needed for the bound. Using the framework of [8 Sections 6–7] with the vanishing switched off, the obstruction is already visible from identities that hold in every dimension for every complete family. These are orthogonality, unbiasedness and the second-moment identity. The single-matrix theorem of [8] remains a result of independent interest. We only show that the complete-family exclusion does not depend on it.
Setting and the identities used
We use the notation and the probability space of [8 Section 6.1]. Bare equation numbers such as (12) or (27) refer to the numbering of [8]; this note has no numbered equations of its own. Suppose that are mutually unbiased orthonormal bases of . Choose a reference basis uniformly, a uniform permutation of its vectors, and a uniform bijection from the labels to the remaining bases. Let be the transition matrices and the normalised character sums [8 (1)]. Let be the mixed moments [8 (28)] and the ordered-pair injective moments [8 (29)]. All of these are real numbers of modulus at most one [8 (30)]. For a uniformly sampled column of , the test functions are
[8 (31)].
Identities used. We use exactly the following valid relations.
(R1) The contraction relations of [8 Lemma 3.2], eqs. (5) and (6). These come from orthogonality of a single transition matrix. We use the formal seven-row relations (6) generated from exponent matrices with seven formal rows, two to four columns and total order 7 to 11, enumerated as in [8 Section 7.2, (i)–(iii)]. (R2) The vanishing of every moment that has a charge of order one. This is row orthogonality, the first half of [8 (12)], and the first half of [8 Lemma 6.2(ii)]. When such a moment has a single label, its pair expansion is recorded as a homogeneous equation, as in [8 Section 7.3(ii)].
(R3) The singleton reduction [8 Lemma 6.2(iii), (34)]. It follows from the second-moment identity
for charges of order two [8 (12)]. This identity is the projective 2-design property of a complete family [5]. It is the only identity used that requires the family to be complete.
(R4) The symmetries [8 Lemma 6.2(i)] and the single-label expansion [8 Lemma 6.2(iv), (35)]. The latter is the equality-pattern expansion of [8 Lemma 3.3].
(R5) The mixed identities [8 (36), (37)], obtained from column orthogonality and unbiasedness, for all charge triples of total order at most 11, enumerated as in [8 Section 7.3].
(R6) The Gram formulas [8 Lemma 6.1] and the bounds [8 (30)].
Identities omitted. In contrast with [8 (27) and Lemma 6.2(ii)], no moment is set to zero because one of its charges is a permutation of . Neither the identity [8 (10)], nor the dichotomy [8 Corollary 3.1], nor the classification of the cubic case [8 Section 4], nor the cubic-pair obstruction [8 Proposition 5.2] is used.
Lemma 1. For every set of seven mutually unbiased bases of , the relations – hold for the actual moment vector, whether or not the transition matrices satisfy [8 (27)].
Proof. [8 Lemma 6.2] is stated under (27). Its proof, however, invokes (27) only for the second half of part (ii). Parts (i), (iii), (iv) and (v) and the order-one half of (ii) are proved there from uniform sampling, from (12), from the equality-pattern expansion [8 Lemma 3.3], and from (13) and (14). [8 Lemma 3.2] concerns one Hadamard matrix and uses only its orthogonality. [8 Lemma 6.1] uses only the sampling. The bounds (30) hold because each moment averages products of unimodular numbers or of averages of them.
The certificate
Theorem 2. There do not exist seven mutually unbiased bases in . Hence .
The theorem follows from Weiner’s theorem [11] and the following finite statement.
Proposition 3. Let run over the six lists of [8 Section 6.3], unchanged, and let be the Young-symmetrised vectors of test functions [8 (39)]. Let be the symmetric integer matrices in the accompanying data file, given by their lower triangles as in [8 (38)]. Then:
(a) each is positive definite;
(b) using only –, for every complete family,
where each is one of the original moments or , and
Proof (computer-assisted, exact). The computation follows [8 Section 7] step by step, with the three changes listed below.
The pair relations (R1) are generated and reduced by the contraction (5). Exact substitution rules are then certified by [8 Lemma 7.1]: modular elimination proposes the rules, and an exact integer check on the witness rows proves them.
The mixed relations (R5) are generated for every charge triple of total order at most 11, after the reductions (R2)–(R4), and certified in the same way.
Each is expanded by [8 (39)], and each Gram entry is replaced using [8 Lemma 6.1]. The factor 5 is the one in [8 (40)]. The reductions and the certified substitutions are then applied.
The changes relative to [8] are these. The zero rule for permutations of is switched off. The coordinates are ordered by the default sort rather than by the priority lists of [8 Section 7.1]; this choice changes which coordinates are free, but adds no relation. The matrices are new. As noted in [8 Section 7.1], the value of depends on the coordinate order, and only its sign for the order actually used matters. The matrices here were computed for the default order, which the verifier fixes deterministically, so no priority list has to be supplied.
The pair table has 8871 nonzero equations, 909 coordinates and 897 witness rows, as in [8]. The mixed table has 40618 equations, 3583 coordinates and 2842 witness rows; with vanishing, [8] has 46041, 3429 and 2767. After substitution, 481 non-constant symbols remain: 10 pair moments, 320 two-label and 151 three-label mixed moments. All of them are irreducible original moments, and 21 of the mixed ones contain a permuted . Under (27) these 21 would vanish; here they are bounded by 1 like all the others.
Positive definiteness is checked by exact rational Schur complements, as in [8 Section 7.4]. All 98 pivots are positive. The smallest pivot of each matrix lies between 0.97 and 11.6, for entries of size up to about . The check is exact, so this small margin is immaterial. Real Gram forms with a positive definite real matrix are nonnegative on complex vectors. The bounds (30) then give
a contradiction. Every arithmetic step that certifies an equality or an inequality is integer or rational. □
The relative margin is large: , against for the certificate of [8], which uses vanishing. Both ratios are invariant under rescaling of the matrices .
Verification
The verifier verify_d6.py and the data file cert_d6_data.json accompany this note. Their SHA-256 checksums are
8149c17e766767712fa6a1724a9d9f22154ce2f473df8b38a652aad19d86df19 (verify_d6.py)
2f7b602b28799008f78a28744ae6c0e0ee02633038b0f21bf8707a4c7080f3e7 (cert_d6_data.json)
The verifier uses only the Python standard library (Python ) and regenerates every relation from scratch. The command is python3 -I verify_d6.py cert_d6_data.json. The internal checks are assertions, so the program refuses to run when assertions are disabled (-O or PYTHONOPTIMIZE); the final sign test is an explicit check. On one core of a Linux workstation (Intel Core i7-13700KF) the run takes 45 s with CPython 3.10.12 and uses about 0.38 GB of memory. It prints the rank tables and the values of Proposition 3. An earlier revision of the verifier, without these two safeguards, printed the same output on a Linux cluster node with CPython 3.9.18 (82 s) and CPython 3.12.4 (69 s).
The program is an adaptation of the verifier printed in [8 Appendix A]. After normalising formatting, 34 of its functions are unchanged. These include the canonical forms, the contraction, the modular elimination and rational reconstruction, the exact substitution check, the Young-symmetrised test functions and the Gram tagging. The mathematical differences are exactly the following. The test for a permuted always returns false. The single-matrix check of [8 (10)] is absent. No priority lists are used. The pivot test is , the exact criterion for positive definiteness, instead of the margin of [8]. A key that never occurs in a relation is kept as a free symbol instead of raising an error. This is sound, because such a symbol is still an actual moment bounded by 1, and with the present data it never happens.
As a cross-check, an independent C++ re-implementation of the relation generation and exact elimination reproduces exactly the pair table, the mixed table of [8] with vanishing together with its published values of and , and the mixed table without vanishing used here. That program has not been run on the matrices of this note. The proof rests on verify_d6.py alone.
Remarks
Remark 4 (Lower order; computational observation). The six lists of [8] need relations of total order 11. With the vanishing switched off, our pipeline finds no certificate for them when all relations are cut at total order 9 or 10, nor when only the pair relations are cut at order 8. The rationalised bounds are then positive: +74.6, +77.5 and +3.66.
A systematic family does better. Take Young-symmetrised test functions for all partitions of 6, with acting on all six coordinates, , and running over balanced charges of order at most 3. Optionally split each block into its real and imaginary parts using . With all relations of total order at most 9 (pair table 197/55/45, mixed table 5363/465/335), our pipeline produced exact rational upper bounds of −8.54, and of −9.57 with the split. A variant that keeps the relations as equality constraints with rational multipliers, instead of eliminating them, gave −9.97. Test functions symmetrised only over the five non-reference coordinates gave −13.24.
In each case the floating-point dual of a semidefinite program was rationalised, and the resulting bound was then evaluated exactly. These certificates have not yet been packaged as standalone data with a verifier, and they have not been independently checked. We therefore record them as a computational observation, not as part of the proof. With test charges of order at most 2 the corresponding semidefinite program was numerically feasible, so no certificate of this type was found at that order.
Remark 5 (Prime-power dimensions). The relation generator is written for a general dimension . It uses formal rows in (R1), the factor in (34), and the Gram weights and . We ran the same pipeline with order-3 systematic test functions, in several symmetrisation variants, and relations of total order at most 11 for . In these dimensions complete families exist, so no certificate can exist. In every case the exact bound was positive, so the pipeline found no certificate, as it must. The positive values ranged from () to ().
In several of these runs the semidefinite solver reported a slightly negative optimum, for example for . Only the exact evaluation decides. This is a consistency check of the generator and of the exact evaluation, not evidence about the strength of the method.
Further computational observations
The following refines Remark 4. Its status is that of a computational observation, not of a theorem. Each item was obtained with our independent C++ implementation of the relation generator (the program of the cross-check in § 4). The floating-point dual of a semidefinite program was rationalised, and the bound was then evaluated in exact rational arithmetic, in the variant of Remark 4 that keeps the relations as equality constraints with rational multipliers. The rationalised Gram matrices are checked to be positive semidefinite by exact elimination, and the bound is computed exactly; a negative value is a contradiction, provided that the generated relations are correct. The values depend on the normalisation of the matrices, so only their signs matter. None of these certificates has been packaged as standalone data with a verifier or checked independently, and none is used in the proof of Theorem 2.
The test functions are those of Remark 4: for all partitions of 6, , and balanced charges of order at most 3, with each block split into real and imaginary parts. Up to permutation and negation, these charges fall into 11 orbit types. “Relation order” is the total order up to which (R1)–(R5) are generated. The vanishing (27) is never used.
(a) Relation order 7 suffices. With all relations of total order at most 7 (243 relation rows), the exact bound is ; at order 8 it is .
(b) Two orbit types suffice at relation order 9. At relation order 9, if all test charges of type , or all of type , are removed, the pipeline finds no certificate; if any one of the other nine types is removed, it still finds one. The charges of these two types alone, with and all , give the exact bounds at relation order 9 and at order 11. At relation order 7 the pipeline finds no certificate for these two types alone, whereas it does for the full family by (a).
(c) The single-matrix relations are not needed. Omitting from (R1) all formal seven-row relations (6) of [8 Lemma 3.2] still gives certificates, with exact bounds at relation order 9 and at order 11.
(d) The orbit of is not needed. Removing all test charges that are permutations of still gives certificates, with exact bounds at relation order 9 and at order 11. Conversely, for test charges of order at most 2 together with the orbit of , the pipeline finds no certificate at relation order 9 or 11.
Data availability
The verifier verify_d6.py, the certificate data cert_d6_data.json, the checksums and the logs of the reference runs are included as ancillary files of the arXiv version of this note. The verifier is an adaptation of the verifier of [8], which is distributed under the Apache License 2.0; it is released under the Apache License 2.0, with the modifications listed in § 4. The certificate data are released under the Creative Commons Attribution 4.0 International licence (CC BY 4.0).
Use of AI
Claude Opus 5.5 (Anthropic) was used to write and run the code (the adaptation of the verifier, the relation generator and the semidefinite-programming pipeline). The authors take full responsibility for the content.
References
- [1]M. Cárdenes Wuttig and J. Tindall, Using topology to reveal all mutually unbiased bases in dimension six, arXiv:2611.06917 [quant-ph], 4 November 2026.
- [2]I. D. Ivanović, Geometrical description of quantal state determination, J. Phys. A: Math. Gen. 14 (1981), 3241–3245.
- [3]M. Jaswal, An independent certified computation excluding four mutually unbiased bases in dimension six, Zenodo preprint, version 2, 21 October 2026, https://doi.org/10.5281/zenodo.23533384.
- [4]M. Jaswal, Triplets of mutually unbiased bases in dimension six, Zenodo preprint, 6 November 2026, https://doi.org/10.5281/zenodo.23991331.
- [5]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.
- [6]M. Matolcsi, A Fourier analytic approach to the problem of mutually unbiased bases, Studia Sci. Math. Hungar. 49 (2012), 482–491.arxiv.org/abs/1009.2407
- [7]M. Matolcsi, I. Z. Ruzsa and M. Weiner, Systems of mutually unbiased Hadamard matrices containing real and complex matrices, Australas. J. Combin. 55 (2013), 35–47.
- [8]OpenAI, Exact Fourier certificates for complex Hadamard matrices of order six, OpenAI Math Release preprint, dated 24 September 2026 (released 6 October 2026), https://github.com/openai/math/tree/main/preprints/Exact-Fourier-certificates-for-complex-Hadamard-matrices-of-order-six-September-24-2026.
- [9]OpenAI, The maximum number of mutually unbiased bases in dimension six, OpenAI Math Release preprint, dated 24 September 2026 (released 6 October 2026), https://github.com/openai/math/tree/main/preprints/The-maximum-number-of-mutually-unbiased-bases-in-dimension-six-September-24-2026.
- [10]T. Tao, Fuglede’s conjecture is false in 5 and higher dimensions, Math. Res. Lett. 11 (2004), 251–258.DOI
- [11]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
- [12]W. K. Wootters and B. D. Fields, Optimal state-determination by mutually unbiased measurements, Ann. Physics 191 (1989), 363–381.DOI
- [13]G. Zauner, Quantum designs: foundations of a noncommutative design theory, Int. J. Quantum Inf. 9 (2011), 445–507.DOI