REVIEW 2 major objections 2 minor 1 cited by
Pointwise adaptive kernel density estimation under local approximate differential privacy
T0 review · 2 major / 2 minor · reviewed 2026-05-24 · grok-4.3
Pith's one-line read Averaging Laplace- or Gaussian-noised kernel density estimators from n data owners attains the minimax rate n^{-(2s-1)/(2s+1)} for pointwise mean squared error over Sobolev classes under local approximate differential privacy.
desk verdict This paper derives the local-DP minimax rate for pointwise KDE and gives a privacy-compatible Lepski adaptation that recovers it up to logs. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
Private kernel density estimator formed by adding scaled Laplace or Gaussian noise to the standard kernel estimator, then averaged across n independent data owners, with bandwidth selected by a privacy-adapted variant of Lepski's method.
What would settle it
A direct calculation or simulation in which the averaged private estimator's pointwise mean squared error fails to match the rate n to the power of minus (2s minus 1) over (2s plus 1) for some fixed s greater than 1/2 when the bandwidth is optimal, or in which the Lepski-adapted version exceeds that rate by more than logarithmic factors.
Extended reading notes
Core claim
We obtain minimax type results over Sobolev classes indexed by a smoothness parameter s>1/2 for the mean squared error at a fixed point. In particular, we show that taking the average of private kernel density estimators from n different data owners attains the optimal rate of convergence if the bandwidth parameter is correctly specified. Notably, the optimal convergence rate in terms of the sample size n is n^{-(2s-1)/(2s+1)} under local differential privacy and thus deteriorated to the rate n^{-(2s-1)/(2s)} which holds without privacy restrictions. A variant of Lepski's method tailored to the privacy setup provides adaptive estimators that attain the optimal rate up to extra logarithmic因素.
Load-bearing premise
The observations are i.i.d. draws from an unknown density belonging to a Sobolev ball of smoothness s greater than one half, and each data owner applies an independent, correctly scaled privacy mechanism before any further processing occurs.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript develops kernel density estimators under local approximate differential privacy by adding scaled Laplace noise or Gaussian processes to per-user KDEs. It establishes minimax-type upper bounds for pointwise MSE over Sobolev balls with smoothness s > 1/2. The average of n independent private estimators attains the rate n^{-(2s-1)/(2s+1)} when the bandwidth is correctly specified, while a Lepski-type adaptive procedure operating on the privatized data attains the same rate up to logarithmic factors.
Significance. If the derivations hold, the work supplies the first explicit minimax rates for pointwise nonparametric density estimation under local DP and quantifies the precise privacy-induced deterioration relative to the non-private rate n^{-(2s-1)/(2s)}. The oracle inequalities for the adaptive estimator and the fact that the procedure remains compatible with downstream statistical tasks are concrete strengths.
major comments (2)
- [Section deriving the non-adaptive upper bound] The upper-bound derivation for the non-adaptive estimator (the rate n^{-(2s-1)/(2s+1)}) must explicitly track how the privacy noise variance enters the pointwise variance term and produces the +1 in the denominator; without that calculation the claimed deterioration from the non-private exponent cannot be verified as load-bearing.
- [Section on the adaptive Lepski procedure] The oracle inequality for the Lepski variant must confirm that the privacy-induced variance is absorbed into the threshold without inflating the logarithmic factor beyond what is stated; the specific inequality should be displayed with the (ε,δ) dependence made explicit.
minor comments (2)
- [Notation and assumptions] The Sobolev-ball definition and the precise scaling of the Laplace/Gaussian mechanisms should be restated in the notation section for self-contained reading.
- [Introduction] The abstract states the rates but does not indicate whether lower bounds are proved or only upper bounds; this distinction should be clarified in the introduction.
Simulated Author's Rebuttal
We thank the referee for the careful reading, positive assessment, and constructive suggestions. We address the two major comments below and will incorporate the requested clarifications into the revised manuscript.
read point-by-point responses
-
Referee: [Section deriving the non-adaptive upper bound] The upper-bound derivation for the non-adaptive estimator (the rate n^{-(2s-1)/(2s+1)}) must explicitly track how the privacy noise variance enters the pointwise variance term and produces the +1 in the denominator; without that calculation the claimed deterioration from the non-private exponent cannot be verified as load-bearing.
Authors: We agree that an explicit step-by-step accounting of the privacy noise contribution is needed to make the rate derivation fully transparent. In the revision we will expand the variance calculation for the averaged private kernel estimator to isolate the term arising from the scaled Laplace (or Gaussian) noise, show how it combines with the usual kernel variance, and derive the resulting pointwise MSE bound that yields the exponent -(2s-1)/(2s+1). This will directly exhibit the privacy-induced shift from the non-private exponent -(2s-1)/(2s). revision: yes
-
Referee: [Section on the adaptive Lepski procedure] The oracle inequality for the Lepski variant must confirm that the privacy-induced variance is absorbed into the threshold without inflating the logarithmic factor beyond what is stated; the specific inequality should be displayed with the (ε,δ) dependence made explicit.
Authors: We will display the full oracle inequality for the private Lepski estimator with the (ε,δ) parameters appearing explicitly in the threshold. Because the threshold is constructed from the variance of the already-privatized estimators, the additional privacy variance is absorbed without enlarging the logarithmic factor beyond the one already present in the non-private case. The revised display will make this absorption and the (ε,δ) dependence immediate. revision: yes
Circularity Check
No significant circularity
full rationale
The derivation chain relies on standard minimax arguments for kernel density estimators, bias-variance trade-offs over Sobolev balls, and an adaptation of Lepski's method to the local privacy setting with per-user Laplace or Gaussian noise calibrated to sensitivity. The claimed rate n^{-(2s-1)/(2s+1)} is obtained directly from the privacy-induced variance term added to the usual KDE variance, without any parameter fitting to the target quantity, self-definitional loops, or load-bearing self-citations. The adaptive oracle inequalities follow from the same concentration bounds applied to the privatized estimators. The paper is self-contained against external benchmarks and contains no reduction of its central claims to its own inputs by construction.
Assumptions & free parameters
free parameters (2)
- bandwidth parameter
- privacy noise scale (epsilon, delta)
assumptions (2)
- domain assumption Data consist of n i.i.d. samples from an unknown density in a Sobolev class with smoothness s > 1/2
- standard math Standard kernel with appropriate moment and smoothness properties is used
Cite this review
Pith. "Pith review of Pointwise adaptive kernel density estimation under local approximate differential privacy." pith.science (2026). https://pith.science/paper/CT7ECGAO
@misc{pith2026190706233,
author = {Pith},
title = {Pith review of: Pointwise adaptive kernel density estimation under local approximate differential privacy},
year = {2026},
howpublished = {\url{https://pith.science/paper/CT7ECGAO}},
note = {Machine review of arXiv:1907.06233}
}
abstract
We consider non-parametric density estimation in the framework of local approximate differential privacy. In contrast to centralized privacy scenarios with a trusted curator, in the local setup anonymization must be guaranteed already on the individual data owners' side and therefore must precede any data mining tasks. Thus, the published anonymized data should be compatible with as many statistical procedures as possible. We suggest adding Laplace noise and Gaussian processes (both appropriately scaled) to kernel density estimators to obtain approximate differential private versions of the latter ones. We obtain minimax type results over Sobolev classes indexed by a smoothness parameter $s>1/2$ for the mean squared error at a fixed point. In particular, we show that taking the average of private kernel density estimators from $n$ different data owners attains the optimal rate of convergence if the bandwidth parameter is correctly specified. Notably, the optimal convergence rate in terms of the sample size $n$ is $n^{-(2s-1)/(2s+1)}$ under local differential privacy and thus deteriorated to the rate $n^{-(2s-1)/(2s)}$ which holds without privacy restrictions. Since the optimal choice of the bandwidth parameter depends on the smoothness $s$ and is thus not accessible in practice, adaptive methods for bandwidth selection are necessary and must, in the local privacy framework, be performed directly on the anonymized data. We address this problem by means of a variant of Lepski's method tailored to the privacy setup and obtain general oracle inequalities for private kernel density estimators. In the Sobolev case, the resulting adaptive estimator attains the optimal rate of convergence at least up to extra logarithmic factors.
Forward citations
Cited by 1 Pith paper
-
Differentially Private Nonparametric Modal Learning with Applications to Regression and Clustering
A private gradient-ascent algorithm estimates all density modes with nearly minimax-optimal error under differential privacy.
Reviewed May 24, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.