{"ID":2862031,"CreatedAt":"2026-06-01T04:54:23.091178241Z","UpdatedAt":"2026-06-01T04:54:23.091178241Z","DeletedAt":null,"paper_url":"https://arxiv.org/abs/2510.00790","arxiv_id":"2510.00790","title":"Differentially Private Learning of Exponential Distributions: Simple Algorithms and Tight Bounds","abstract":"We study the problem of learning exponential distributions under differential privacy. Given $n$ i.i.d.\\ samples from $\\mathrm{Exp}(λ)$, the goal is to privately estimate $λ$ so that the learned distribution is close in total variation distance to the truth. We present a simple pure $ε$-differentially private algorithm that avoids the classical dependence on the true value of $λ$. Our method leverages a structural property of the exponential distribution: its $(1-1/e)$-quantile equals $1/λ$, allowing us to estimate the rate parameter directly via private quantile estimation. The resulting learner is both conceptually simple and sample-efficient, achieving near-optimal guarantees. We further extend the method to Pareto distributions via a logarithmic reduction, prove nearly matching lower bounds using group privacy arguments, and show how approximate $(ε,δ)$-DP removes the need for externally supplied parameter bounds. Together, these results give the first tight characterization of exponential distribution learning under differential privacy using a simple $λ$-free approach.","short_abstract":"We study the problem of learning exponential distributions under differential privacy. Given $n$ i.i.d.\\ samples from $\\mathrm{Exp}(λ)$, the goal is to privately estimate $λ$ so that the learned distribution is close in total variation distance to the truth. We present a simple pure $ε$-differentially private algorithm...","url_abs":"https://arxiv.org/abs/2510.00790","url_pdf":"https://arxiv.org/pdf/2510.00790v2","authors":"[\"Bar Mahpud\",\"Or Sheffet\"]","published":"2025-10-01T11:40:33Z","proceeding":"cs.DS","tasks":"[\"cs.DS\"]","methods":"[]","has_code":false}