On two questions of Shapiro on Jacobi scaling
For a symmetric positive definite matrix B of order n, let R*(n) be the supremum of κ(J(B))/min_D κ(DBD), where J(B) is the Jacobi scaling of B and D ranges over positive diagonal matrices. Van der Sluis showed R*(n) ≤ n, and Shapiro showed R*(n) ≥ n/2 and asked for the exact value, the limit of R*(n)/n, and whether R*(n)/n increases with n. We prove that R*(n) = n/2 for 2 ≤ n ≤ 5, with no extremal matrix for 3 ≤ n ≤ 5, and that R*(n) > n/2 for every n ≥ 6, with 3.02244 ≤ R*(6) ≤ 3.02753. For every n ≥ 3, −1/(4√n − 2) ≤ R*(n) − (n − 2√n + 2) < 1/(4n), and R*(n) < n − 2√n + 121/60; for n ≥ 6, R*(n) = n + 2 − min_m(m + n/m) + O(1/n), with an explicit constant, and R*(k²) = k² − 2k + 2 + 1/(4k²) + O(k⁻⁴). The upper bounds come from a relaxation of a variational formula for R*(n); for n ≥ 6 we determine the value of this relaxation in closed form. All proofs are by hand. Hence R*(n)/n → 1, and R*(n)/n is constant for 2 ≤ n ≤ 5 and strictly increasing from n = 5 on. We also show that κ(J(B)) ≤ 1 + R*(n)(min_D κ(DBD) − 1). For block-Jacobi scaling we lower the factor in Demmel's bound to max{1, ϑ(Ḡ) − 1}, where ϑ(Ḡ) is the Lovász theta function of the complement of the block graph; for b blocks of size at least 2, the resulting factor b − 1 is sharp. Files: the paper (PDF) and sources_and_supplement.zip (LaTeX source, bibliography, build script and the verification scripts with their outputs). This paper builds on version 3 of the author's preprint doi:10.5281/zenodo.24293244 and is cited by its version 4, doi:10.5281/zenodo.24315153; a companion paper on the exact SGD rate is doi:10.5281/zenodo.24315183.
Authors
- Gabriel Costa de Mello Paiva
Institutions
- Elizabethtown Community Hospital (US)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24315180
- Citations
- 2
- Primary Topic
- Matrix Theory and Algorithms
- Type
- preprint