A new analysis separates the complexity of John ellipsoid algorithms into certification, identification, and accuracy. It shows that the \varepsilon^{-1} factor arises only in certification via average iterates, not accuracy. After an \varepsilon-independent setup, accuracy depends only on \log\log(1/\varepsilon), with identification remaining an open problem.
Decoupling Costs in John Ellipsoid Approximation
from English