The 24-cell conjecture states that every Voronoi cell of a packing of unit balls in R4 has volume at least 8, the volume of the regular 24-cell. It implies that no packing of equal balls in R4 is denser than the lattice packing D4. We prove the conjecture for two classes of cells; in each class, equality holds only for the D4 root configuration. The first consists of the cells whose ball touches every ball centred at distance less than 22 from its centre, and its proof uses a certificate at the second level of the Lasserre hierarchy. The second consists of the cells whose 24 nearest centres lie in an explicit neighbourhood of the D4 configuration, whatever the other centres are. For them the volume is at least 8 plus a fixed multiple of the total outward displacement of the 24 centres, which we prove by a quantitative rigidity estimate and a second-order expansion of the volume about D4. Our proofs for both classes are computer-assisted and partly formalised in Lean.
Author comments: Published on Zenodo: https://doi.org/10.5281/zenodo.23720963
A mix of human-written and AI-generated textHuman understanding: all parts
The kissing number τn is the largest number of non-overlapping unit balls in Rn that can touch a common unit ball. The best bounds so far were 40≤τ5≤44 and 126≤τ7≤134; both upper bounds come from high-precision floating-point computations of the three-point bound by Mittelmann and Vallentin. We prove τ5≤43 and τ7≤132. Each proof is an exact rational solution of the second level of the Lasserre hierarchy for spherical codes at degrees (d1,d2,δ)=(14,20,20). We use the formulation of de Laat, Leijenhorst and de Muinck Keizer and check the solutions with their verification program. Each certificate is obtained by rounding a convex combination of two dual iterates of the interior-point method, one on each side of the target value. Exact certificates for the same hierarchy also give A(4,arccos31)≤14, A(5,arccos31)≤22 and A(6,arccos31)≤34 for spherical codes, improving the previous bounds 15, 23 and 35; with a known code, A(4,arccos31)=14. The largest number of points at angle π/3 from a point of a kissing configuration in Rn equals A(n−1,arccos31), so in R5 this number is 14. We also prove the five bounds in the Lean proof assistant, assuming only the positivity of the zonal kernels given by the zonal polynomial files: once by checking the exact certificates with compiled code, and once, with every computation done by Lean's kernel, from the certificates rounded to 30 decimal places.
Author comments: Published on Zenodo: https://doi.org/10.5281/zenodo.23992400
A mix of human-written and AI-generated textHuman understanding: all parts
Cohn, Kumar and Minton constructed a four-parameter family of tight 39-point codes in the octonionic projective plane. We show that near each member outside a closed nowhere-dense set of resonant parameters, the labeled tight codes modulo F4 form exactly four ten-dimensional families, any two meeting in the four-parameter family, although the first-order deformations of the member form a 28-dimensional space modulo F4. The originally constructed code is resonant, and we do not determine the tight codes near it; but every neighborhood of it in the space of unlabeled tight 39-point codes modulo F4 contains a nonempty open subset homeomorphic to an open subset of R10, so any neighborhood of it that is a manifold has dimension ten.
Author comments: 55 pages, 5 figures. Ancillary files: anc/native-data.json holds the exact coefficients of Appendix C, and anc/finite-identities.py replays the finite identities of Proposition 6.5 from it in exact arithmetic. Lean source: https://github.com/macgregor-sudo/tight-39-code at commit 46f01c1.
A sequence of points in the unit cube is quasi-uniform if its first N points are spread like a grid for every N: no two of them are closer, and no empty ball between them is larger, than a constant times N−1/s in dimension s. Dick, Goda and Suzuki asked for such sequences among the digital sequences of quasi-Monte Carlo integration, and Goda, Hofer and Suzuki asked whether any Halton-type sequence is quasi-uniform. We show that in dimension two and higher no Halton-type sequence, no Sobol' sequence, no classical Niederreiter sequence and no Faure sequence is quasi-uniform, and that linear scrambles of all coordinates but one do not help: a carry in the last digits of one coordinate leaves the leading digits of all others unchanged, giving explicit pairs of points that are far too close. We then show that one fixed linear scramble of a generalized Faure sequence, the same for all numbers of points, repairs this. In it, the digits of all coordinates at one scale are residues of one rational function, each taken at the coordinate's own point of the projective line, so the residue theorem ties the coordinates together and a carry in one coordinate is seen by the others. For every dimension s and every prime base p>2s this gives explicit quasi-uniform digital (0,s)-sequences, hence of low discrepancy, whose first N=psK−1 points are at least (1−2p−1/s)N−1/s apart, for every K; in dimensions two and three, for every prime p≥5, a variant keeps its first N points at least (1−2/p)N−1/s apart whenever N is a power of ps2. For such scrambles, exact linear rules that place every point inside its cell as a function of the last few digits of the cell's index can hold at no more than about one in s of the sizes psK; with allowed ranges instead of exact rules we certify all of these sizes, with points at least (1−O(1/p))N−1/s apart for large p. Quasi-uniform digital sequences follow in every prime base p≥3 in dimension three, and in every prime base in dimension two, base two included. Some steps are computer-assisted.
A mix of human-written and AI-generated textHuman understanding: some partsHalton sequencesLean 4 formalizationcovering radiusdigital sequenceslow discrepancyminimum distancequasi-Monte Carloquasi-uniformityresidue theorem
Contributed by Deep Bhattacharjee, Ushashi Bhattacharya
We prove that no packing of unit balls in R5 whose centres form two translates of a lattice is denser than the D5 lattice packing. Up to congruence exactly three such packings attain this density: 2D5, a layered packing of Leech, and the packing with Szöllősi's kissing configuration.
A mix of human-written and AI-generated textHuman understanding: some parts$D_5$ latticeRyshkov polyhedronSphere packingVoronoi celldeep holeperiodic packing
We develop an alternative route to the solutions of the sphere-packing
problem in dimensions 8 and 24, through the auxiliary functions
in the Cohn--Elkies Fourier bound. In each dimension we construct a
radial Schwartz function with the prescribed lattice zeros by solving
a countable system of linear equations in absolutely summable
coefficients. We prove existence and uniqueness after normalization,
without using the known magic functions or Fourier interpolation.
Through a Cauchy boundary problem, solutions of the homogeneous
equations are identified with global sections of a rank-six
holomorphic vector bundle over the compactified modular curve
of the full modular group, with its orbifold structure retained.
Riemann--Roch shows that this bundle has Euler characteristic
one, forcing a nonzero global section and hence a nontrivial
solution of the equations. A separate argument proves uniqueness
after normalization.
An equivalent positive operator equation has a bounded inverse.
The construction is motivated by the coefficient method for planar
circle packing. A subsequent comparison identifies our solutions
with the functions of Viazovska and of Cohn, Kumar, Miller,
Radchenko, and Viazovska. The route to sphere packing is not yet
independent in its entirety: we do not prove the required global
sign inequalities from the coefficient equations. Those inequalities
are inherited from the known functions after identification.
Primarily AI-generated textHuman understanding: some partsSphere packing
We give an explicit cover of a regular tetrahedron by 4096 cylinders with compact triangular perpendicular bases. The sum of their base areas, divided by the tetrahedron's minimum orthogonal projection area, lies in [0.499613830783714148,0.499613830783714149). Thus the saving below 1/2 exceeds 0.000386169216285851. The construction uses a finite tilt in the angular-sector framework of OpenAI's September 2026 preprint and rationally specified radial cutoffs. Coverage is certified by exact linear-programming dual witnesses, and independently verified by exhaustive rational polyhedron clipping. We describe the geometric reduction, the certificate obligations, and reproducible verification of the full closed tetrahedron and the perpendicular base areas. No optimality is asserted.
Primarily AI-generated textHuman understanding: some parts
Dick, Goda and Suzuki asked for low-discrepancy digital nets that are quasi-uniform, with minimum distance δ and covering radius both of the optimal order N−1/s, and stated the problem as open in dimension s≥3. We construct such nets in every dimension. For every s≥3, every base q whose prime factors are at least s−1, and every k≥1, we give explicit digital (0,sk,s)-nets over Z/qZ with N=qsk points, covering radius at most sN−1/s, and δN1/s≥1−as∗/q once q≥5as∗+4. The constant as∗≤eslogs+O(s) is computed from s by a deterministic polynomial-time lattice algorithm. In dimension three every N=pm, m≥3, is covered for each prime p≥37, and, with computer assistance, every N=2m and N=3m. In prime bases these families are well separated in the sense of Dick, Goda and Suzuki, and they stay so under every digital shift. One mechanism underlies all constructions. For N=qsk it reads as follows. The first k base-q digits of a point's coordinates select a cube of side q−k, and a (0,sk,s)-net has exactly one point in each such cube. Digit k+1 of each coordinate is a linear function of the first k digits of all coordinates, because the net is digital. We choose this linear function so that, between the planes where this digit wraps around modulo q or where a carry occurs among the first k digits, the points lie on a translate of one lattice; for the main theorem this lattice is a slightly distorted copy of Zs. All later digits move a point by less than 1/q of the cube side. Part~II studies the constant K in δN1/s=1−K/q+O(q−2). We prove lower bounds on K when the linear maps have bounded coefficients. At N=q3 we find the optimal K among nets whose points lie on lines parallel to the main diagonal. With the cube-tiling theorems of Haj\'os and Keller we show that idealized lattices with one point per cube, whose points sit independently inside their cubes, cannot exceed δN1/s=1. When the positions inside the cubes are coupled, the value 1 can be exceeded: δN1/3→281/16 at N=p3, δN1/5>1 for (0,5,5)-nets from p=37 on, and δN1/s→min(s/2,2) at N=psk, which is the E8 value at s=8 (limits as p→∞). The results are formalized in Lean~4, apart from some finite computations, which are checked by scripts.
We show that for every convex body K⊂Rn (n≥2), every 0<ε<1, and every prescribed density ρ≥Aεnlogn, there is an arrangement of translates of K along a single lattice with mean multiplicity exactly ρ and multiplicity at least (1−ε)ρ everywhere. Consequently, lattice k-fold coverings exist with density at most max{(1+ε)k,Cεnlogn}, where Cε depends only on ε. In sufficiently large dimension n, a version with prescribed common covolume holds simultaneously for up to exp(nc0) bodies whose uniform distributions have a common covariance matrix, with
c0>0 absolute. The proof builds on OpenAI's horizontal–vertical construction for single-lattice coverings and uses Gaussian layer averaging and Laplace bounds for covering counts.
A mix of human-written and AI-generated textHuman understanding: all partsGeometry of numbersconvex bodiesconvex geometrycovering densitycovering multiplicitylattice coveringmulti-covering
We deduce a new asymptotic upper bound Klat(n)≤(2e/π+o(1))n on the maximal lattice kissing number Klat(n) in dimension n, using the explicit auxiliary functions constructed in OpenAI's "Ten Advances" preprint. The same bound holds for the average contact degree of a finite packing of congruent balls. The bound's base-2 exponential rate rounds to 0.3956, matching the value extrapolated empirically by Afkhami-Jeddi, Cohn, Hartman, de Laat, and Tajdini in 2020.
A mix of human-written and AI-generated textHuman understanding: all partsContact numbersDiscrete geometryEuclidean latticesFourier linear programming boundsGeometry of numbersKissing numbersSphere packing