This closes an order stated as open in the 2025 census of Đoković and Seberry: a symmetric Hadamard matrix of order 8740 = 4·2185. It comes from a symmetric version of Seberry’s 1980 recursion: whenever q ≡ 3 (mod 4) is a prime power and a symmetric Hadamard matrix of order q + 5 exists, there are symmetric Hadamard matrices of orders (q + 1)(q + 4)qr for every r. Seberry had these orders for ordinary matrices, and Kumar the symmetric case r = 0; 8740 is the case q = 19, r = 1.
The 2025 census of Đoković and Seberry lists orders 2116 = 4·529 and 4356 = 4·1089 as undecided for symmetric Hadamard matrices. A 1972 theorem of Wallis and Whiteman, with Mathon’s and Gritsenko’s conference matrices of orders 46 and 66, settles both. No new construction is involved: we found 4356 independently and then learned that SageMath already implements the construction. The note gives the explicit matrix of order 4356 and an exact check.
A new regular Hadamard matrix, of order 4·9332 = 3481956, with every row and column sum 1866. The parameter 933 had been listed as undecided since 2007. It comes from a completion criterion applied to the half-lines Leung, Ma and Schmidt used at 311.
Four new orders of skew Hadamard matrices: 4040, 5192, 5960 and 6216. Each is 23 times an odd part, 505, 649, 745 or 777, for which the smallest known skew order so far needed a higher power of two, as high as 211 for 505. The matrices come from prime fields with p ≡ 17 (mod 32), and one ratio of binomial coefficients decides which primes work for four fixed patterns. A sequel to order 2120.
Skew Hadamard matrices are conjectured to exist at every order divisible by 4. For odd part 265 the least order known so far was 4240 = 24 × 265. This gives one of order 2120 = 23 × 265. The same method gives skew Hadamard matrices of order 4(p2n + 1) for p = 7, 23, 263 and every odd n.
Earlier work built skew Hadamard matrices of orders 4(343d + 1), d odd, from order‑18 cyclotomic patterns in the Wallis–Whiteman array, but left open whether they are new up to equivalence. They are, for every comparison made here: none is equivalent to a Kronecker product, a skew doubling, its own transpose, or a matrix built from sixth-power classes. Up to an overall sign, every symmetry fixes the border and applies one map x ↦ ax343j, a an eighteenth power, to all eight field copies; there are 19 at order 1,376. A second pattern set extends the construction to every prime p ≡ 7 or 31 (mod 36), giving at least eight inequivalent matrices per order and sixteen at order 1,376.
Three near-Hadamard families of Momihara, Suda and Xiang exceed 80% of the Ehlich–Wojtas determinant bound at every order n ≥ 30, answering Problem 15 in a 2026 survey. The paper also proves the limiting ratio √e/2 and determines each family’s exact minimum. Novelty remains unconfirmed: a closely related paper was not openly accessible at the time of writing.
Integrating a function over each interval of a nested subdivision of [0,1] and adding the result back on that interval defines a linear operator with one term per interval, whose only nonzero eigenvalue is that interval’s length. Earlier work on a sparse Mertens kernel showed that the eigenvalues above t track the intervals longer than t when each piece becomes a vanishing fraction of the interval it was cut from, and asked whether that condition can be dropped. Because the terms overlap, it cannot: this paper shows that the ratio of the two counts can oscillate forever even when that fraction is merely kept below a fixed bound, however small, and proves other conditions under which the counts agree, including for the intervals defined by continued-fraction digits.
This page applies data envelopment analysis, a technique built on convex analysis, to find the Pareto frontier of cost and benchmark scores, then the cheapest model or blend on it that matches your model’s scores. Based on a 2022 paper I coauthored.
This paper gives explicit tests κ and μ for whether a curve of cyclic n-roots can leave a boundary point, and finds order-26 points where symmetry forces κ = 0 but μ ≠ 0. The subtle point is that the symmetry makes the first test inconclusive but not the second.
Earlier work used hypergraphs to force large ℓ1 diversity distortion; this paper shows that planar diamond graphs already need distortion of order log n, the largest possible order, so OpenAI’s constant bound for planar distances fails for sets of vertices.
Earlier work showed that the singular values of a sparse Mertens matrix are determined, one at a time, by the eigenvalues of an explicit limiting kernel; this paper shows that the r-th eigenvalue of that kernel is asymptotic to 6/(π² r log r).
Earlier work found how the largest eigenvalue of the totient Gram matrix grows; this paper shows that the second, third and later eigenvalues, divided by N², each settle to a positive limit given by an explicit operator.
OpenAI’s historic October 2026 release of hundreds of new math results left a bound unstated, and I made it explicit for about $15. Expertise that took decades to develop is now available for a modest monthly subscription.
This paper gives a different proof of an averaged form of a classical bound for the error in the prime number theorem, using a differential equation derived from Selberg’s formula.
The paper gives the first explicit length beyond which a polynomial with coefficients ±1 can have modulus within a factor 1 ± ε of √N on the whole unit circle, making quantitative OpenAI’s October 2026 construction, which gave no rate.
This paper proves the prime number theorem by comparing how often each integer remains after deleting either K or K + 1 largest prime factors from every squarefree integer up to a common bound.
The paper describes the singular spectrum of a sparse matrix with determinant equal to the Mertens sum, counts exactly how many singular values equal one, and replaces earlier bounds for the smallest with an unconditional asymptotic formula.
Resolves a 2020 conjecture on the Perron vector of the totient Gram matrix and sharpens that paper’s eigenvalue and Möbius-correlation bounds to asymptotics.
An affine-plane hypergraph with a polynomial-size description requires ℓ1 diversity distortion of order √n, refuting Conjecture 1.7 of Jozefiak and Shepherd.
Weighting the Liouville sum by a power of the square part of n moves its real pole without moving the zeros of ζ, so the sign of the sum reads a zero-free half-plane, and a real-character twist puts 1/L(σ,χ) into the residue.
Matching several histograms on a line at the cost of each unit’s bin span: weighted range costs turn out to be the only symmetric Monge costs with hull invariance and Minkowski additivity, and every vertex of the transport dual is counted.
A short note giving elementary proofs of several exact identities for weighted Möbius and Liouville sums, organized around a single convolution factorization, together with a two-way character-transfer bound.
A numerically certified face-by-face solver fills the low-ratio gap in Mangasarian’s Newton method on synthetic tests, matches production-solver timing on the largest cases, and narrows the gap on the Netlib benchmark through a three-method frontier, though it still trails HiGHS.
We construct an infinite tower of skew Hadamard matrices of orders 4(76t+3 + 1), t ≥ 0, beginning with order 1,376, using one fixed order-18 finite-field construction.
A systematic attack on Turyn’s 1972 converse conjecture closes every remaining non-prime-power case through q = 2000 and proves a semiprime nonexistence theorem, while the full conjecture remains open.
The paper extends earlier work on harmonic divisor matrices to a much sparser version of the Redheffer matrix, bounding its smallest singular value and determining its largest.
An iterated convex relaxation always converges to a fixed point, can stop finitely near regular CAZAC sequences, and has an exact regularity count at Zadoff–Chu points.
The paper settles two conjectures about harmonic divisor matrices and describes the extreme eigenvalues and singular values of a family linking them to the classical Redheffer matrix.