Jul 2026
Improved Error Bounds for Pure Differentially Private Continual Counting via Matrix Factorization
This work improves the leading constant in the maximum squared error and the mean squared error, and proves an $\Omega(\epsilon^{-2}\log^3 n)$ lower bound for the class of factorizations whose matrices have entries in $\{0,1\}$, matching the upper-bound asymptotics for this class.
P. Arkhipov, Nikita P. Kalinin
· arXiv.org · 2 citations
· ⚡1