Spherical codes in five dimensions with 486 to 495 points
There are spherical codes in R^5 (S^4) with N = 483, ..., 495 points whose maximal cosines (0.855185, 0.855392, 0.855766, 0.856019, 0.856286, 0.856682, 0.856807, 0.857094, 0.857130, 0.857291, 0.857550, 0.857738, 0.857973, each rounded up at the 6th decimal) improve the best known values in Cohn's spherical-codes table for these thirteen N (spherical-codes.org: A. Jay, 2026, for 484, 485 and 495; Vahteristo, arXiv:2610.19988, for the others). Each was verified by an exact integer check of all pairs. This record holds the codes for N = 486 to 495 (version 1 had 486 to 489; version 2, 18 November 2026, added 490 to 494; version 3, 18 November 2026, adds 495). The codes for N = 483, 484 and 485 are in separate records: https://doi.org/10.5281/zenodo.24342564 (483), https://doi.org/10.5281/zenodo.24343765 (484) and https://doi.org/10.5281/zenodo.24344672 (485). The table is Henry Cohn, Table of spherical codes, https://spherical-codes.org/ (archived data set: https://hdl.handle.net/1721.1/153543); the comparison uses the rows listed on 18 November 2026. The previous codes for N = 486 to 494 are due to Ilmari Vahteristo (arXiv:2610.19988), with listed maximal cosines 0.856651783818, 0.856831067866, 0.856932073246, 0.857147443193, 0.857315360501, 0.857419798181, 0.857605788666, 0.857744196019 and 0.857798082994; ours are 0.85601842144..., 0.85628553569..., 0.85668102703..., 0.85680658159..., 0.85709309534..., 0.85712928632..., 0.85729066375..., 0.85754920091... and 0.85773798532... . For N = 495 the previous code is due to A. Jay (2026), with listed maximal cosine 0.857974025889; ours is 0.85797224450... (a margin of about 1.8e-6). The A. Jay 2026 values are from the program of A. Jay, M. Poberznik, M. Brut, A. Hemeryck, L.-H. Hoang, M. Gunde, N. Salles, G. Cruchon and M. Roux, "Equidistribution of points on the hypersphere including symmetries", preprint, 2026, https://doi.org/10.2139/ssrn.6427919 . Construction: the 483-point code was found with GPT-6.2 (ChatGPT, OpenAI). Each later code was built from the previous one by adding one point at its deepest hole, relaxing all points jointly on the sphere and polishing with L-BFGS (build_insert_relax.py, build_polish_lbfgs.py and chain.sh, which need numpy and scipy and are given for the record only). This is a simple greedy method. Each single cell is a small improvement: the table's source preprint (arXiv:2610.19988) alone improved 4,577 cells. Verification: the final coordinates are integers; each row v of directions_N.csv stands for the unit vector v/|v|. Run python3 own_check_exact.py directions_N.csv N TABLE_VALUE. It uses only the Python standard library and compares the exact squared cosines of all pairs with positive product as fractions. The certificate does not depend on the numerical optimization. These are constructions at single cardinalities. They say nothing new about the kissing number in dimension five, and the optimal angles for these N remain open. The chain was continued to N = 500 under a fixed compute cap (F91_LOG.txt): 495 was improved, but 496 to 500 were not (for example, our best 496-point code has maximal cosine about 0.858180, above the listed 0.858096902788). The chain is closed. AI use: this note comes from an AI-assisted research lab directed by the author. The 483-point starting code was found by GPT-6.2 (ChatGPT, OpenAI); the later constructions, the exact verification and the text were produced with Claude (Anthropic) under the author's direction, and the check was repeated by a separate session of the same AI system, so this is not independent human review. Not peer reviewed. Contact: [email protected].
Authors
- Учрал Ням-Осор
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.5281/zenodo.24348686
- Primary Topic
- Mathematical Approximation and Integration
- Type
- preprint