Optimal cosine measure, quantitative rigidity and local optimality for direct-sum positive bases
Unrefereed research candidate. Orthogonal regular simplices are optimal among direct sums and locally optimal among all configurations. The question. Direct-search methods minimise a function without derivatives by trying a fixed set of directions from the current point. This works only if, whatever the downhill direction is, one trial direction points roughly the same way. The cosine measure scores a set of unit directions by its worst case: the cosine of the largest angle between any direction in space and the nearest trial direction. Bigger is better. Maximising it is the same as covering the sphere by equal caps of smallest radius. Hare, Jarry-Bolduc and Planiden (2023) asked whether the best positive bases built from minimal positive bases of complementary subspaces (their class Ωn,s) must have mutually orthogonal subspaces. Results. (1) If the subspaces have dimensions r1, …, rk, the cosine measure is at most (∑ rj2)−1/2, with equality exactly for regular simplices in mutually orthogonal subspaces (Theorem 3.1). This answers the question for every dimension and size; the optimum over Ωn,s uses balanced block dimensions (Corollary 3.2). (2) Exact identities split the loss into nonnegative parts and give quantitative rigidity: near-optimal configurations are nearly orthogonal and nearly regular, and the optimal stability exponent changes from 1/2 to 1 at size s = ⌈3n/2⌉ (Theorem 6.1). (3) Every orthogonal union of regular simplices is a strict local maximiser of the cosine measure among all configurations of the same number of unit vectors, up to rotation, with quadratic growth of the loss (Theorem 7.1). So the configurations that Böröczky and Wintsche (2003) conjectured to give optimal sphere coverings by few equal caps are strict local optima in every dimension, and no nearby positive basis, with or without critical vectors, does better. Unbalanced unions are local optima too: in R4 with six directions, a tetrahedron plus an orthogonal pair (cosine measure 1/√10) is a local maximiser although two orthogonal triangles reach 1/√8 (Corollary 7.2). (4) The averaging argument extends to shifted blocks whose shifts lie in the convex hull of the preceding vectors; an explicit positive basis of R4 shows this condition can fail for every decomposition (Section 8). The random-vertex averaging device appears in public figures by Jarry-Bolduc and Krishan Lal for triangles and pairs and is credited to them. Limits. The global optimum over all positive bases or all coverings remains open for sizes n+3 to 2n−2; the local result does not give the size of the neighbourhood; constants are not optimal; the shifted-block extension needs the convex-hull condition. Contents. The paper (LaTeX source and PDF, with one figure); verify_geometry.py (6301 explicit floating-point checks on 312 configurations, with independent vertex enumeration and high-precision finite differences, exiting nonzero on any failure, also under python -O) and its recorded output; plot_stability.py and the figure; a prior-art note; the bibliography from registry records; a machine-readable index; a research-gate record; a rating-free review summary; a replay receipt and SHA-256 sums. Assurance. Complete written proofs; the computations are checks, not proofs. Internal AI-assisted review only (five reviewer roles, an adversarial review and an audit by a second AI model, clean replays, a re-review and a proof audit): not peer review, independent reproduction or formal verification. Creator: Anonymous. Text and data CC0-1.0; code MIT.
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24242262
- Primary Topic
- Mathematical Approximation and Integration
- Type
- preprint