Publications & Papers

Research & Formal Writing

A collection of formal papers, proofs, and research notes spanning mathematics, computation, and the intersections between them.

Filter
Papers 6
Probability 2026 10.5281/zenodo.24180369
New

The Curveball Chain on Regular Graphs: Spectral Gap at Most 2/n and a Comparison with Random Transpositions

Follows up the October 2026 paper on sharp switch and Curveball gaps, which proved a Curveball gap of at least 2/C(n,2) on every graphical degree sequence and left open how much larger the gap is on particular sequences. For the Curveball chain that performs a uniform trade at a uniformly chosen vertex pair of a d-regular graph on n vertices, 1 − 2/n is an eigenvalue for every n and d, so the spectral gap is at most 2/n and the relaxation time is at least n/2 trades. The chain dominates one half of the random-transposition walk on vertex labels, which bounds every eigenvalue outside two isotypic components and shows that the gap equals 2/n unless a smaller eigenvalue has an eigenfunction depending only on the rooted isomorphism class of the graph. The gap equals 2/n for d ∈ {1, n − 2} and, by an exact computer-assisted computation, for every regular degree sequence with n ≤ 11; equality for all n and d is open. The manuscript, its exact verification and the certificates were produced with AI assistance; the release includes the written proofs, exact certificates and ancillary programs.

Probability 2026 10.5281/zenodo.24174781
New

Spectral Gaps for Transpositions with Forbidden Fixed Points

Gives explicit spectral gaps for the random-transposition walk restricted to permutations of m symbols that avoid t specified fixed points. For every m ≥ 4 the unit-rate transposition Laplacian has spectral gap at least max{1, m − min(t, 4)}, so the walk that holds with probability 1/2 and otherwise proposes a uniformly chosen transposition, rejecting a forbidden result, has gap at least max{1, m − min(t, 4)}/[2·C(m,2)]. With exactly two forbidden fixed points the unit-rate gap is exactly m − 2 for every m ≥ 3. With three it is exactly (3(m − 2) − √((m − 2)2 + 8))/2 for every m ≥ 4, where the matching lower bound uses an exact computer-algebra certificate. The bounds carry over to the switch and Curveball chains on loopless directed graphs whose in- and out-degrees are all 0 or 1. The manuscript, its exact verification and the computer-algebra certificate were produced with AI assistance; the release includes the written proofs, exact certificates and ancillary programs.

Probability 2026 10.5281/zenodo.23295359
Preprint

Sharp Spectral Gaps for the Switch and Curveball Chains on Every Graphical Degree Sequence

Gives the best possible universal spectral gaps for two Markov chains that sample simple graphs with a prescribed degree sequence. For every graphical degree vector on n ≥ 4 labeled vertices with more than one realization, the lazy switch chain has spectral gap at least 1/(6·C(n,4)), and the undirected Curveball chain at least 2/C(n,2). Equality holds in both bounds for d = (2, 2, 1, 1, 0, …, 0). Consequently the total-variation mixing times at distance 1/4 are at most n6/4 + 1 for the switch chain and n4/16 + 1 for the Curveball chain. This sharpens the pair-resampling argument of the OpenAI Math Release manuscript, which proved polynomial mixing for every graphical degree sequence with a switch-chain gap of [24n2·C(n,4)]−1 and a mixing bound of 2n8. Both gap bounds are kernel-checked in Lean 4; the repository also holds the finite numerical checks.

Algorithms 2026 10.5281/zenodo.18713449
Preprint

TreeKernelSum: Fast Approximate All-Pairs Kernel Summation on Weighted Trees

Addresses the all-pairs kernel summation problem on weighted trees — computing, for every node, a kernel-weighted sum over all other nodes. Using centroid decomposition and a sum-of-exponentials (SOE) approximation, the algorithm runs in O(r n log n) time, achieving a 972× speedup over the naïve O(n²) baseline at n = 50,000 nodes. Supports Gaussian, Ornstein-Uhlenbeck, Matérn, and inverse multiquadric kernels. Validated on real phylogenetic datasets spanning 5,000+ mammalian species, achieving R² = 0.955 in Nadaraya-Watson body mass prediction. Scales to 1 million nodes in approximately 12 seconds.

Number Theory · Algorithms 2026 10.5281/zenodo.18716922
Preprint

Longest Increasing Subsequences of Collatz Trajectories: A Null Model Decomposition

Uses the longest increasing subsequence (LIS) as a pseudorandomness metric for Collatz trajectories. The mean LIS-to-trajectory ratio is 0.567 for starting values up to 106 — substantially below the 0.812 ratio observed in random sequences of comparable length. Value shuffling restores the ratio to ~0.813, isolating temporal ordering as the driver of the deviation. A hierarchical null model analysis attributes roughly one-third of the disparity to the 31.6% up-step fraction, with the descending suffix responsible for a further reduction. The remaining unexplained variance of ~0.05 persists across generalized 3n+b variants, pointing to deeper structural geometry in Collatz dynamics.

Number Theory 2025 10.5281/zenodo.15738443
Published

Unified p-Adic Bounds and Computation for the Erdős–Moser Equation

A Mahler-Interpolation Framework and Exhaustive Computational Survey

Introduces a closed-form p-adic valuation formula for odd-prime moduli and a Mahler-interpolation lemma for even exponents, together forming a unified framework for the p-adic valuations of power sums Sk(m). Computational verification finds no counterexamples for odd primes 3 ≤ p < 200 with even exponents 2 ≤ k ≤ 20. An exhaustive survey of the Erdős–Moser ratio over 2 ≤ m ≤ 2000 and 2 ≤ k ≤ 12 identifies no solutions beyond the classical case (m, k) = (3, 1). Includes two Python validation scripts and a full PDF manuscript.

No papers in this category yet.

What research is published here?

Six papers across probability, algorithms, and number theory. Three 2026 preprints bound the spectral gaps of Markov chains. The October paper proves best possible universal spectral gaps for the switch and Curveball chains that sample graphs with a given degree sequence, with both gap bounds kernel-checked in Lean 4. Two November papers continue that work: one shows that the Curveball gap on d-regular graphs is at most 2/n and equals 2/n for every regular degree sequence with n ≤ 11, and the other gives explicit spectral gaps for transpositions on permutations that avoid specified fixed points. Two earlier 2026 preprints: TreeKernelSum, which derives an O(r n log n) algorithm for all-pairs kernel summation on weighted trees using centroid decomposition; and a null-model analysis of longest increasing subsequences in Collatz trajectories quantifying structural deviation from random sequences. The 2025 paper establishes unified p-adic valuation bounds for the Erdős–Moser equation via Mahler interpolation. All papers include code and are archived on Zenodo with DOIs.

Who writes these papers?

Mark Short, an independent researcher and developer working under Seveneves. Each paper has a single author. The manuscripts and verification of the three Markov chain papers were produced with AI assistance, as each paper's acknowledgments state; their releases include the Lean project, exact certificates or scripts needed to check the results. Papers are self-published on Zenodo under open licenses (MIT or CC-BY-4.0) with no institutional affiliation.

More from Seveneves

Cite this paper