Work

Worst-case root-exponential decay for low-rank updates of matrix square roots

Year: 2026
Type: preprint
Abstract: For 0 <= A <= B=A+UU* <= I with rank(U) <= k, let D=B^(1/2)-A^(1/2) and E_k(j)=sup s_(jk+1)(D), with the supremum over all finite dimensions and no positive spectral gap assumption. For j,k >= 1, we prove exp(-2*pi*sqrt(j))/(8*(j+1)^2) <= E_k(j) <= min(1,6*exp(-sqrt(2*j))), and E_k(0)=1. Explicit finite rank-one positive semidefinite updates give the lower estimate through a resolvent comparison and a weighted Cauchy matrix. Direct sums yield arbitrary k, and zero padding gives singular inputs. Newman rational approximation and the rational update rank bound give the upper estimate, establishing the optimal root-exponential order.
Author: Xu, Zihuan
Language: English
Open Access status: green
OpenAlex ID: W7227710449
Cites:
Cited by:
Related to: 10