New private and non-private estimators achieve per-instance KL error within constant factors of the best possible error over a new additive local neighborhood.
For brevity, denoteˆd = ⌊ dsmall(L′) 2 ⌋ ≥2 and ˆp = P ˆd i=1 pi
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
stat.ML 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Instance-Optimality for Private KL Distribution Estimation
New private and non-private estimators achieve per-instance KL error within constant factors of the best possible error over a new additive local neighborhood.