Pith. sign in

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 →

arxiv 1907.06233 v1 pith:CT7ECGAO submitted 2019-07-14 math.ST cs.CRstat.TH

classification math.STcs.CRstat.TH
keywords localapproximatedifferentialprivacykerneldensityestimationpointwiseadaptivebandwidthselectionSobolevclassesLepski'smethodminimaxrates
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper studies nonparametric density estimation where each of n data owners must apply local approximate differential privacy to their observations before any shared analysis occurs. It constructs private kernel density estimators by adding scaled Laplace noise or Gaussian process noise to the standard kernel estimator. Averaging these independent private estimators produces an estimator whose pointwise mean squared error attains the optimal rate n^{-(2s-1)/(2s+1)} over Sobolev balls of smoothness s greater than 1/2 when the bandwidth is correctly specified. A variant of Lepski's method, adapted to work directly on the anonymized data, yields an adaptive bandwidth choice whose risk matches the same rate up to logarithmic factors.

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.

Watch

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.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

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)
  1. [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.
  2. [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)
  1. [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.
  2. [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

2 responses · 0 unresolved

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
  1. 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

  2. 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

0 steps flagged · score 0.0 of 10

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 2 free parameters · 2 assumptions · 0 invented entities

Central claims rest on standard domain assumptions from non-parametric statistics and differential privacy; no new entities are postulated and the only free parameters are the usual bandwidth and privacy scales.

free parameters (2)
  • bandwidth parameter
    Must be chosen adaptively because it depends on unknown smoothness s; appears in the rate statements.
  • privacy noise scale (epsilon, delta)
    Determines the magnitude of added Laplace or Gaussian noise to meet the approximate DP guarantee.
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
    Invoked to obtain the minimax results over these classes.
  • standard math Standard kernel with appropriate moment and smoothness properties is used
    Required for the bias-variance analysis of the kernel density estimator.

how reviews work

0 comments
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.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Differentially Private Nonparametric Modal Learning with Applications to Regression and Clustering

    math.ST 2026-07 conditional novelty 7.0 of 10

    A private gradient-ascent algorithm estimates all density modes with nearly minimax-optimal error under differential privacy.

Pith tools

Reviewed May 24, 2026 · model on record in the stance chip above.