{"id":"arxiv/2609.05034","url":"https://arxiv.org/abs/2609.05034","tags":["math.ST","stat.ML"],"score":0,"title":"Faster Learning under Relaxed Local Differential Privacy","domain":"math.ST","source":"arxiv","authors":["Cristina Butucea","Huiyun Tang","Marie-Luce Taupin"],"reviews":0,"subject":"math.ST","summary":"We consider density estimation under the relaxed local differential privacy condition that the privatized distributions are $α$-close in total variation distance. We show that adding independent noise with a convenient symmetrized Gamma distribution to each sensitive observation attains the $α$-TV-LDP. We prove that the deconvolution estimator of $r$-Sobolev smooth functions attains the pointwise rate $(nα)^{-\\frac{2r-1}{2r}}$ up to log factors which is faster than $(nα^2)^{-\\frac{2r-1}{2r+1}}$ under the classical $α$-LDP and closer to the nonprivate minimax rate $n^{-\\frac{2r-1}{2r}}$.…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-09-04","reactions":{}}