acceptodds
Under review as a conference paper at ICLR 2027

Empirical Likelihood for the Differentially Private Mean

Abstract

We develop a nonparametric framework for confidence intervals for a population mean under differential privacy (DP), without requiring variance estimation or a parametric model for the underlying distribution. We partition clipped observations into disjoint blocks, privatize each block mean with Laplace noise, and construct confidence intervals using empirical likelihood (EL) and adjusted empirical likelihood (AEL). We establish Wilks-type results showing that the EL and AEL likelihood-ratio statistics converge to the standard distribution under DP. The resulting intervals are calibrated for the clipped population mean , the estimand targeted by fixed-clipping bounded-sensitivity mechanisms. We explicitly characterize the clipping bias and provide a bias-aware extension for inference on the raw mean. Simulations across various distributions, together with real-data experiments, show that the proposed methods maintain coverage close to the nominal level while typically producing shorter confidence intervals than existing approaches.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Empirical likelihood, differentially privacy

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6L8eOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies confidence intervals for a population mean under differential privacy. It first clips the observations and analyzes the resulting bias, as well as the non-identifiability of the original mean under fixed clipping without additional assumptions. It then partitions the data into disjoint blocks, adds Laplace noise to the block means, and applies empirical likelihood and adjusted empirical likelihood to construct confidence intervals. The theoretical results establish asymptotic chi-squared distributions for the likelihood-ratio statistics at the clipped population mean. The paper also discusses an extension to the original mean using a bound on the clipping bias. The experiments compare several methods across different distributions, including heavy-tailed settings, and include real-data applications.

Critical strengths
  1. The paper explicitly distinguishes the clipped population mean from the original mean and analyzes the consequences of this distinction. The treatment of clipping bias and non-identifiability helps clarify what the proposed procedure can actually infer.
  2. The paper is relatively concise, and the empirical evaluation covers several competing methods, different distributional settings, and real-data applications. Reporting both coverage and interval length provides useful evidence about the finite-sample behavior of EL and AEL.
Critical weaknesses
  1. The additional methodological and theoretical insight is not sufficiently clear. The construction combines clipping, privatized block means, and established EL/AEL procedures. Given the relative simplicity of this construction, I would expect a clearer explanation of what we learn beyond the validity of this particular combination. The paper emphasizes a Wilks-type result for a triangular array with vanishing row variance, but it is not sufficiently clear which aspects require substantially new arguments beyond existing empirical-likelihood theory. This makes the significance of the theoretical contribution difficult to assess.
  2. The cost of privacy and the resulting efficiency trade-offs need a more explicit characterization. The paper discusses privacy-noise variance and varies the privacy budget in Table 2, but these results do not sufficiently clarify the loss relative to a comparable non-private procedure or separate the cost of privacy from the additional cost introduced by the block construction. In particular, the number of blocks affects both calibration and privacy noise. A clearer account of how sample size, privacy budget, block count, and clipping bounds jointly determine interval length and coverage would substantially strengthen the contribution.
Questions for authors
  1. The algorithm is relatively simple, but I do not see sufficient new insight beyond applying existing EL/AEL methods to privatized block means. The additional contribution needs to be clearer.
  2. My main concern is the cost of privacy. Although the paper varies the privacy budget experimentally, it does not sufficiently explain the trade-off between privacy and statistical efficiency or quantify the loss relative to a non-private baseline.
  3. Figure 1 is difficult to interpret because the curves are barely distinguishable. I cannot clearly see the differences between the settings or how much AEL improves over EL.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 6L8e

Response to W1 / Q1

What we learn beyond the validity of this combination.

  1. Inference needs only the privatized block means. The χ12\chi^2_1 limit has no nuisance parameter, so EL and AEL need no parametric model, no explicit variance estimator and no input about the noise variance; the whole budget goes to the block means. By contrast, a studentized interval needs an explicit variance estimate, and with privatized block means that estimate must account for the Laplace noise. To see whether this extra step helps, we compare AEL with a studentized interval. Without any variance step, AEL is shorter in every setting and closer to the nominal level:
DesignnELAELNoise-aware studentized
Normal2000.935/3.510.944/3.580.961/3.67
4000.937/3.290.947/3.350.955/3.41
6000.938/3.060.945/3.110.955/3.15
Exp2000.950/4.160.957/4.250.968/4.34
4000.946/3.610.954/3.680.960/3.73
6000.944/3.190.949/3.250.958/3.27
Beta2000.940/0.850.950/0.870.965/0.89
4000.939/0.770.948/0.780.955/0.79
6000.941/0.700.947/0.710.957/0.72
Lognormal2000.941/5.040.949/5.150.962/5.27
4000.939/4.480.948/4.570.955/4.63
6000.940/4.050.946/4.110.955/4.14
Gamma2000.942/3.370.953/3.450.966/3.53
4000.942/3.000.950/3.050.958/3.10
6000.942/2.700.945/2.750.957/2.76
Weibull2000.942/1.760.950/1.800.964/1.84
4000.942/1.650.950/1.680.960/1.70
6000.943/1.530.948/1.550.958/1.57
  1. With fixed clipping, the clipped mean is the identifiable target. For mean inference from data clipped to a fixed interval, μc\mu_c is identifiable, whereas μ\mu is not without a support or tail assumption (Prop. 2.2). We give the exact signed bias μc−μ\mu_c-\mu (Prop. 2.1) and an assumption-indexed interval for μ\mu (Remark 2.4). We now evaluate this adjustment empirically: | Method | Normal | Exp | Beta | Lognormal | Gamma | Weibull | |---|---|---|---|---|---|---| | Unadjusted | 0.955/3.62 | 0.958/4.29 | 0.946/0.87 | 0.959/5.23 | 0.951/3.45 | 0.953/1.81 | | Adjusted | 0.958/3.78 | 0.962/4.49 | 0.947/0.88 | 0.973/6.32 | 0.953/3.53 | 0.955/1.84 |

The adjusted intervals achieve empirical coverage of 0.947–0.973 for the raw mean, compared with 0.946–0.959 for the unadjusted intervals. The principal benefit of the adjustment is that it explicitly accounts for clipping bias under an additional moment assumption, providing a theoretically justified extension of inference to the raw mean.

Which aspects require new arguments.

The theoretical contribution is the Wilks theorems for EL and AEL on privatized block means (Theorems 2.3 and 2.5). The released values x~j\tilde x_j are block means plus Laplace noise, with variance σB2=σc2/m+2w2/(m2ϵ2)→0\sigma_B^2=\sigma_c^2/m+2w^2/(m^2\epsilon^2)\to0. They form a triangular array whose law changes with n and whose noise has heavier-than-Gaussian tails, so the classical fixed-distribution Wilks theorem does not apply directly, and the existing DP EL theory (Liu & Li, 2025) concerns local estimates released by separate machines. To establish the χ2\chi^2 limits for this array, we prove:

  • uniform bounds on the standardized moments of x~j\tilde x_j along every sequence with B→∞B\to\infty, m→∞m\to\infty (Lemma L.1), which give Lyapunov's condition and S/σB2→p1S/\sigma_B^2\to_p1;
  • the maximal-term condition max⁡j∣x~j−μc∣/σB=op(B1/2)\max_j|\tilde x_j-\mu_c|/\sigma_B=o_p(B^{1/2}) despite the Laplace tails;
  • feasibility (μc\mu_c in the convex hull of the x~j\tilde x_j) with probability tending to one;
  • the case σc2=0\sigma_c^2=0, where the clipped data are constant (e.g., the clipping range misses the data, so μc\mu_c lies at a clipping bound). All variation then comes from the Laplace mechanism, and normalization by the clipped sampling variance is unavailable;
  • the AEL limit (Theorem 2.5). AEL adds the pseudo-point −aBYˉ-a_B\bar Y, which depends on the candidate θ\theta and on all B released values, with aB=max⁡(1,log⁡B/2)a_B=\max(1,\log B/2) growing with B. We derive the exact empirical first and second moments of the augmented sample of B + 1 points, bound its maximal term including the pseudo-point (which requires aB=o(B)a_B=o(B)), and show that for these privatized summaries the pseudo-point changes the statistic only by a factor that tends to one.
AuthorsOpenReview ↗

Response to Reviewer 6L8e (Cont.)

Response to W2 / Q2

We separate the cost of blocking from the cost of privacy, first in closed form and then empirically.

With B equal blocks of size m = n/B that use all n observations, clipping width w and σc2=Var(clip(X))\sigma_c^2=\mathrm{Var}(\mathrm{clip}(X)), the average of the privatized block means x~ˉ\bar{\tilde x} (the point at which the EL statistic is zero) has exact variance

Var(x~ˉ)=σc2n+2w2Bn2ϵ2,and, when σc2>0,Var(x~ˉ)σc2/n=1+2w2Bnϵ2σc2.\mathrm{Var}(\bar{\tilde x})=\frac{\sigma_c^2}{n}+\frac{2w^2B}{n^2\epsilon^2},\qquad\text{and, when }\sigma_c^2>0,\qquad \frac{\mathrm{Var}(\bar{\tilde x})}{\sigma_c^2/n}=1+\frac{2w^2B}{n\epsilon^2\sigma_c^2}.

  • Blocking alone preserves the sample mean and its variance exactly. The sampling term does not depend on B; finite-sample EL/AEL calibration and interval length can still change with B.
  • Privacy adds a variance inflation that grows linearly in B and in 1/ϵ21/\epsilon^2.
  • Under the paper's asymptotics (B≍n/log⁡nB\asymp n/\log n, fixed ϵ\epsilon and w), the privacy-noise variance is of order 1/(nlog⁡n)=o(n−1)1/(n\log n)=o(n^{-1}), i.e. its standard deviation is o(n−1/2)o(n^{-1/2}). So when σc2>0\sigma_c^2>0, EL/AEL are first-order efficient relative to the non-private sampling variance. The decomposition concerns x~ˉ\bar{\tilde x}; finite-sample interval length also reflects χ2\chi^2 calibration, and the privacy term matters most at small ϵ\epsilon.

Empirical decomposition (n = 400, Normal(1,1), 2,000 replications):

Methodϵ\epsilon = 1ϵ\epsilon = 2ϵ\epsilon = 3ϵ\epsilon = 4ϵ\epsilon = 5Coverage
Non-private, 400 observationsEL0.1880.1880.1880.1880.1880.946
AEL0.1900.1900.1900.1900.1900.948
Non-private, 20 block meansEL0.1860.1860.1860.1860.1860.941
AEL0.2040.2040.2040.2040.2040.958
Non-private, 80 block meansEL0.1890.1890.1890.1890.1890.948
AEL0.1950.1950.1950.1950.1950.954
DP, B = 20EL0.310.220.200.200.190.925–0.945
AEL0.340.240.220.210.210.949–0.963
DP, B = 80EL0.540.310.250.230.210.938–0.945
AEL0.550.320.260.230.220.946–0.949
  • Cost of blocking vs. cost of privacy. Comparing rows with the same method separates the two. Blocking alone costs essentially nothing for EL (0.186 and 0.189 against 0.188); for AEL it adds a little at B = 20 (0.204 against 0.190), the price of its small-sample adjustment. Privacy is the main cost, and it falls quickly with ϵ\epsilon: at ϵ\epsilon = 1 the DP intervals are about 1.6–2.9 times the non-private length, and at ϵ\epsilon = 5 only 2–16% longer than the same method on all 400 clipped observations. Against non-private AEL on the same blocks, the ϵ\epsilon = 5 privacy cost is 3% at B = 20 and 13% at B = 80. AEL covers at 0.946–0.963 in every DP row; EL covers at 0.925–0.945 and is below nominal at B = 20, ϵ\epsilon = 1.
  • The formula approximates these lengths. For B = 20 at ϵ\epsilon = 1, the variance factor is 1+2⋅16⋅20/(400⋅0.92)=2.741+2\cdot16\cdot20/(400\cdot0.92)=2.74. To first order this gives an EL length of 0.1882.74=0.310.188\sqrt{2.74}=0.31, against 0.31 observed for EL and 0.34 for AEL.
  • How the quantities interact:
    • larger n and larger ϵ\epsilon shrink the privacy term;
    • larger B inflates it, but small B worsens χ2\chi^2 calibration: at B = 20, plain EL covers at 0.925–0.945, while AEL covers at 0.949–0.963 (table above). AEL therefore allows a smaller B and a shorter interval: at B = 20 its lengths are 0.21–0.34, against 0.22–0.55 at B = 80;
    • a wider clipping range inflates it through w2w^2 (and changes σc2\sigma_c^2), so the interval lengthens while coverage of μc\mu_c stays at 0.947–0.955 (Table 4 of the paper), but it reduces the tail-clipping error bound of Prop. 2.1.

Response to Q3

  1. Figure 1 checks the χ2\chi^2 approximation of Theorems 2.3 and 2.5 in finite samples. The curves overlap because the empirical distributions are close to the χ12\chi^2_1 limit, as the theorems predict.
  2. The figure also shows that AEL is closer to χ12\chi^2_1 than EL. The rejection rates at the nominal levels make this concrete :
Nominal levelEL, n = 200/400/600AEL, n = 200/400/600
10%12.3 / 11.8 / 11.110.4 / 10.8 / 10.5
5%6.7 / 5.6 / 5.45.4 / 4.8 / 4.8
1%1.8 / 1.5 / 1.21.3 / 1.2 / 1.1

EL over-rejects, most at n = 200 (6.7% at the 5% level), while AEL is closer to the nominal level in all nine comparisons. This is the finite-sample improvement of AEL, and it matches the coverage gap between EL and AEL in Table 1.

Thank you for the detailed response. The variance decomposition and the non-private comparisons address my concern about the cost of privacy. The rejection rates clarify the improvement in AEL calibration that was difficult to see in Figure 1. I also appreciate the comparison with the noise-aware studentized interval and the explanation of the additional proof steps for privatized block means.

My remaining concern is the significance of the theoretical contribution. Although the response clarifies why this setting requires additional analysis, I still view the advance over existing EL/AEL theory as modest. I therefore maintain my original recommendation of weak rejection.

Reviewer_R1kDOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper provides a method for constructing differentially private confidence intervals for clipped means for continuous variables. They propose to 1) clip the values based on predetermined bounds, 2) divide the data into equal-sized B non-overlapping blocks, 3) apply Laplace noise to the block mean, and 4) use EL or AEL to construct CI based on the asymptotic chi-square distribution.

Critical strengths

This paper provides a simple, statistically justifiable method for private mean estimation. The proposed idea is based on the well-established asymptotic likelihood results.

The empirical studies cover many distributional scenarios, privacy parameters, and sample sizes. Also, the real data example adds practical value to the proposed method.

Critical weaknesses

The main contribution is only on the 'clipped mean', which itself is somewhat dependent on user input (clipping bounds). This can create user-induced clipping bias, which needs to be investigated.

They discuss an adjustment on how the raw (unclipped) mean CI can be constructed, but it is not empirically evaluated. Appendix only shows the coverage of the original CI (for clipped mean) against the raw mean.

Questions for authors

The average lengths in the tables are presented without any standard errors. I wonder how much Monte Carlo variability there is for each method.

The average lengths of some methods seem unfairly large, like more than 1 for beta distribution. Since the method is based on the overall clipping of the data, I think they should consider applying the same clipping (maybe even post-hoc) to all compared methods.

4 comments
AuthorsOpenReview ↗

Response to Reviewer R1kD

Thank you for a careful and constructive review. We respond to each point below.

Response to W1

Clipping bias and user-chosen bounds. Proposition 2.1 characterizes the exact signed bias μc−μ\mu_c-\mu. Our original sensitivity analysis (Section 3.2 and Table 4) examines coverage of μc\mu_c under different clipping bounds. We now extend this analysis to report the clipping bias and coverage of the original mean μ\mu. Specifically, we vary the bounds as [(1−δ)xmin⁡,(1+δ)xmax⁡][(1-\delta)x_{\min},(1+\delta)x_{\max}], with n=200n=200, B=40B=40, ϵ=0.1\epsilon=0.1, and 2,000 replications (bias / coverage of μ\mu):

δ\deltaNormalExpBetaLognormalGammaWeibull
−0.3−0.065 / 0.949−0.030 / 0.955−0.001 / 0.946−0.231 / 0.950−0.014 / 0.943−0.042 / 0.945
−0.2−0.032 / 0.948−0.018 / 0.9600 / 0.961−0.191 / 0.948−0.007 / 0.939−0.021 / 0.950
−0.1−0.012 / 0.940−0.011 / 0.9560 / 0.958−0.160 / 0.947−0.003 / 0.953−0.010 / 0.950
00 / 0.955−0.007 / 0.9580 / 0.946−0.135 / 0.959−0.002 / 0.951−0.004 / 0.953
+0.10.007 / 0.953−0.004 / 0.9590 / 0.949−0.115 / 0.956−0.001 / 0.952−0.002 / 0.957
+0.20.013 / 0.950−0.002 / 0.9600 / 0.955−0.099 / 0.9570 / 0.961−0.001 / 0.947
+0.30.018 / 0.947−0.002 / 0.9600 / 0.951−0.085 / 0.9510 / 0.9610 / 0.943

The results show that clipping bias depends on the selected bounds and the underlying distribution. At ϵ=0.1\epsilon=0.1, AEL coverage of the original mean remains approximately 0.94–0.96 across the considered bounds.

Data-dependent clipping bounds. We also evaluate privately selected clipping bounds using differentially private estimates of the 1% and 99% quantiles. In the same-sample design, 25% of the privacy budget is allocated to quantile estimation and the remaining 75% to AEL inference (n=400n=400, B=80B=80, ϵ=0.5\epsilon=0.5). Coverage of μc\mu_c at the selected bounds / coverage of μ\mu / average length:

NormalExpBetaLognormalGammaWeibull
0.954/0.953/3.930.943/0.935/2.220.946/0.943/0.480.953/0.942/3.100.941/0.941/1.900.956/0.954/1.11

Across six distributions, empirical coverage of the clipped mean at the selected bounds ranges from 0.941 to 0.956, while coverage of the original mean ranges from 0.935 to 0.954. An independent-pilot design yields broadly similar results. These experiments illustrate the feasibility of privately selecting clipping bounds while highlighting the distinction between inference for the selected clipped mean and inference for μ\mu.

Response to W2

Raw-mean adjustment. By Proposition 2.1(a), the raw population mean satisfies

μ=μc−Emax⁡(xmin⁡−X,0)+Emax⁡(X−xmax⁡,0).\mu = \mu_c - \mathbb E\max(x_{\min}-X,0) + \mathbb E\max(X-x_{\max},0).

Using the moment bounds in Remark 2.4, we construct the adjusted interval

[L−Δlo, U+Δhi],[L-\Delta_{\mathrm{lo}},\ U+\Delta_{\mathrm{hi}}],

where [L,U][L,U] is the AEL interval for μc\mu_c, and

Δlo=Mp,ap(a−xmin⁡)p−1,Δhi=Mp,ap(xmax⁡−a)p−1,\Delta_{\mathrm{lo}} = \frac{M_{p,a}}{p(a-x_{\min})^{p-1}}, \qquad \Delta_{\mathrm{hi}} = \frac{M_{p,a}}{p(x_{\max}-a)^{p-1}},

with Mp,a=E∣X−a∣pM_{p,a}=\mathbb E|X-a|^p and aa denoting the midpoint of the clipping range.

Simulation results. We evaluate the adjustment using the same replications as the unadjusted AEL interval, with n=200n=200, B=40B=40, ϵ=0.1\epsilon=0.1, and 2,000 replications. The results report empirical coverage of the raw mean μ\mu and average interval length.

MethodNormalExpBetaLognormalGammaWeibull
Unadjusted0.955/3.620.958/4.290.946/0.870.959/5.230.951/3.450.953/1.81
Adjusted0.958/3.780.962/4.490.947/0.880.973/6.320.953/3.530.955/1.84

The adjusted intervals achieve empirical coverage of 0.947–0.973 for the raw mean, compared with 0.946–0.959 for the unadjusted intervals. The principal benefit of the adjustment is that it explicitly accounts for clipping bias under an additional moment assumption, providing a theoretically justified extension of inference to the raw mean.

AuthorsOpenReview ↗

Response to Reviewer R1kD (Cont.)

Response to Q1

Monte Carlo standard errors. The following table reports the Monte Carlo standard errors of the average interval lengths in Table 1, based on 2,000 replications per setting.

DesignnELAELKVBHAWPriDELBAPriDEL
Normal2000.0180.0170.0060.0200.0090.0140.012
4000.0160.0150.0110.0220.0130.0140.013
6000.0140.0140.0150.0170.0150.0140.014
Exp2000.0260.0250.0160.0220.0160.0180.016
4000.0260.0260.0220.0260.0220.0200.019
6000.0240.0240.0250.0250.0140.0210.021
Beta2000.0050.005<0.0010.0040.0030.0030.003
4000.0050.0050.0010.0050.0040.0040.004
6000.0040.0040.0010.0050.0030.0040.004
Lognormal2000.0300.0280.0210.0250.0170.0210.018
4000.0300.0290.0270.0310.0270.0230.022
6000.0270.0270.0300.0310.0200.0250.024
Gamma2000.0200.0190.0110.0190.0120.0140.012
4000.0200.0200.0150.0220.0170.0160.015
6000.0180.0180.0170.0200.0120.0170.016
Weibull2000.0090.0080.0010.0090.0050.0070.006
4000.0080.0080.0030.0100.0070.0070.006
6000.0070.0070.0040.0110.0080.0080.007

The Monte Carlo standard errors are small relative to the corresponding average interval lengths, ranging from less than 0.001 to 0.031 across the considered settings.

Significance of interval-length differences. We further compare the average interval lengths of the competing methods against AEL and EL. The following tables summarize the numbers of settings in which each method produces significantly longer or shorter intervals, across the 18 settings.

Comparison against AEL:

MethodSignificantly longer than AELSignificantly shorter than AELNo significant difference
PriDEL1800
BAPriDEL1800
KV1800
BH1521
AW963

Comparison against EL:

MethodSignificantly longer than ELSignificantly shorter than ELNo significant difference
PriDEL1800
BAPriDEL1800
KV1800
BH1620
AW1053

The results indicate that PriDEL, BAPriDEL, and KV produce significantly longer intervals than EL and AEL in all 18 settings. BH also produces significantly longer intervals in most settings, whereas the comparisons with AW are mixed. In particular, AW produces significantly longer intervals than AEL in 9 settings. However, AW consistently exhibits substantial overcoverage (0.984–0.991), while AEL achieves coverage closer to the nominal 0.95 level (0.944–0.957). Overall, these additional analyses support the interval-length advantages of EL/AEL over several existing approaches.

AuthorsOpenReview ↗

Response to Reviewer R1kD (Cont.)

Response to Q2

Matched clipping. In our original experiments, the EL, AEL, KV, and BH confidence intervals were intersected with the public clipping range, whereas those from AW, PriDEL, and BAPriDEL were not. We now apply the same rule to all methods. The results are shown below.

DesignnELAELAWPriDELBAPriDEL
Normal2000.935/3.510.944/3.580.986/3.910.960/3.840.973/3.87
4000.937/3.290.947/3.350.988/3.680.960/3.800.967/3.83
6000.938/3.060.945/3.110.986/3.190.958/3.760.963/3.78
Exp2000.950/4.160.957/4.250.991/4.720.963/4.780.979/4.84
4000.946/3.610.954/3.680.987/3.450.963/4.700.969/4.73
6000.944/3.190.949/3.250.988/2.550.963/4.610.964/4.64
Beta2000.940/0.850.950/0.870.989/0.960.965/0.960.976/0.97
4000.939/0.770.948/0.780.986/0.770.960/0.940.968/0.95
6000.941/0.700.947/0.710.985/0.590.959/0.930.962/0.93
Lognormal2000.941/5.040.949/5.150.989/5.720.965/5.720.978/5.78
4000.939/4.480.948/4.570.989/4.630.962/5.620.967/5.67
6000.940/4.050.946/4.110.987/3.470.960/5.530.963/5.57
Gamma2000.942/3.370.953/3.450.990/3.810.965/3.840.978/3.87
4000.942/3.000.950/3.050.987/2.960.962/3.770.969/3.79
6000.942/2.700.945/2.750.987/2.240.959/3.700.962/3.73
Weibull2000.942/1.760.950/1.800.987/1.950.961/1.920.973/1.93
4000.942/1.650.950/1.680.984/1.810.961/1.900.968/1.92
6000.943/1.530.948/1.550.987/1.530.954/1.870.961/1.89

After matched clipping, AEL maintains empirical coverage close to the nominal 0.95 level (0.944–0.957). PriDEL and BAPriDEL produce longer intervals than AEL in all 18 settings. AW produces longer intervals in 10 of 18 settings, with similar average interval lengths (2.89 for AW versus 2.87 for AEL). However, AW consistently exhibits substantial overcoverage (0.984–0.991), whereas AEL achieves coverage closer to the nominal level.

I appreciate the authors' response to my comments. I'll maintain my evaluation.

Reviewer_3WSfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper develops EL and AEL confidence intervals under DP. It partitions clipped observations into disjoint blocks, adds Laplace noise to each block mean, and constructs intervals through likelihood-ratio inversion. The main results establish asymptotic chi-squared calibration for the clipped population mean. The paper also characterizes clipping bias and provides a conservative extension to the raw mean under an explicit external moment bound. Validation includes simulations and real-data resampling experiments comparing coverage and interval length with existing methods. The authors report that AEL generally improves calibration over EL and that both methods often produce shorter intervals than the selected baselines.

Critical strengths
  1. Careful asymptotic justification for privatized block summaries.
  2. Explicit and consequential treatment of clipping-induced target changes.
Critical weaknesses
  1. The central efficiency advantage is not established by the current comparisons.

Algorithms 1–2 restrict EL/AEL intervals to the public clipping range, whereas several reported baseline lengths exceed that range’s width. For the clipped-mean target, applying the same intersection to every baseline preserves coverage and can reduce interval length, so the comparison does not isolate a methodological advantage. Additionally, a simple Hoeffding–Laplace benchmark, derived separately in this review, gives shorter intervals with a finite-sample coverage guarantee in the main normal-design settings. Together, these observations substantially weaken the paper’s principal practical claim.

  1. The incremental contribution is not supported by sufficiently discriminating comparisons.

The paper acknowledges existing DP EL methods and adopts the established AEL adjustment. The additional benefit of EL/AEL remains unclear without a studentized interval constructed from the same private block summaries and fully specified comparisons with the closest alternatives. Consequently, the current evidence does not demonstrate that the theoretical adaptation delivers a sufficiently substantial practical advance to support acceptance.

3.It is not clear how to convince readers to use the proposed DP mean method, especially given that there are already numerous DP mean estimation and inference methods with various elegant theoretical properties or practical applicability.

Questions for authors
  1. How are block sizes and Laplace noise scales determined when n=400n = 400 and B=70B = 70?

  2. When the clipping bounds in Appendix F are estimated from the same sample used for inference, what parameter does the interval target, and what coverage guarantee applies?

  3. At B=20B = 20 in Table 3, how much of EL’s undercoverage is attributable to convex-hull infeasibility versus error in the chi-squared approximation?

  4. Are the interval endpoints included in the outputs of Algorithms 1–2, particularly when the clipped mean equals a public clipping bound?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 3WSf

Thank you for a careful and constructive review. We respond to each point below.

Response to W1

Matched clipping. In our original experiments, the EL, AEL, KV and BH intervals were intersected with the public clipping range, whereas those from AW, PriDEL and BAPriDEL were not. We now apply the same rule to all methods. The results are shown in the following table.

DesignnELAELAWPriDELBAPriDEL
Normal2000.935/3.510.944/3.580.986/3.910.960/3.840.973/3.87
4000.937/3.290.947/3.350.988/3.680.960/3.800.967/3.83
6000.938/3.060.945/3.110.986/3.190.958/3.760.963/3.78
Exp2000.950/4.160.957/4.250.991/4.720.963/4.780.979/4.84
4000.946/3.610.954/3.680.987/3.450.963/4.700.969/4.73
6000.944/3.190.949/3.250.988/2.550.963/4.610.964/4.64
Beta2000.940/0.850.950/0.870.989/0.960.965/0.960.976/0.97
4000.939/0.770.948/0.780.986/0.770.960/0.940.968/0.95
6000.941/0.700.947/0.710.985/0.590.959/0.930.962/0.93
Lognormal2000.941/5.040.949/5.150.989/5.720.965/5.720.978/5.78
4000.939/4.480.948/4.570.989/4.630.962/5.620.967/5.67
6000.940/4.050.946/4.110.987/3.470.960/5.530.963/5.57
Gamma2000.942/3.370.953/3.450.990/3.810.965/3.840.978/3.87
4000.942/3.000.950/3.050.987/2.960.962/3.770.969/3.79
6000.942/2.700.945/2.750.987/2.240.959/3.700.962/3.73
Weibull2000.942/1.760.950/1.800.987/1.950.961/1.920.973/1.93
4000.942/1.650.950/1.680.984/1.810.961/1.900.968/1.92
6000.943/1.530.948/1.550.987/1.530.954/1.870.961/1.89

The results show that AEL maintains coverage close to the nominal 0.95 level (0.944–0.957). PriDEL and BAPriDEL produce longer intervals than AEL in all 18 settings. AW produces longer intervals than AEL in 10 of 18 settings, with similar average interval lengths across the settings (2.89 for AW versus 2.87 for AEL). However, AW consistently overcovers (0.984–0.991), whereas AEL achieves coverage closer to the nominal level.

Hoeffding–Laplace benchmark. We additionally compare EL and AEL with the Hoeffding–Laplace interval across three privacy budgets (ϵ\epsilon = 0.1, 1, 2) and six distributions at n = 200 (EL and AEL with B = 40). The results are shown in the following table.

ϵ\epsilonMethodNormalExpBetaLognormalGammaWeibull
0.1EL0.935/3.510.950/4.160.940/0.850.941/5.040.942/3.370.942/1.76
AEL0.944/3.580.957/4.250.950/0.870.949/5.150.953/3.450.950/1.80
HoefLap0.994/2.310.996/2.420.996/0.550.997/3.160.995/2.100.996/1.15
1EL0.939/0.760.930/0.920.932/0.190.940/1.140.935/0.730.947/0.38
AEL0.950/0.800.943/0.970.946/0.200.953/1.200.950/0.770.959/0.40
HoefLap1.000/0.981.000/1.231.000/0.251.000/1.471.000/0.981.000/0.49
2EL0.931/0.440.937/0.520.941/0.100.935/0.670.932/0.410.933/0.22
AEL0.945/0.470.950/0.550.955/0.100.949/0.700.950/0.430.945/0.23
HoefLap1.000/0.911.000/1.141.000/0.231.000/1.361.000/0.911.000/0.46

Hoeffding–Laplace exhibits substantial overcoverage across all privacy budgets (0.994–1.000), suggesting that its intervals are overly conservative, whereas EL and AEL achieve coverage much closer to the nominal level (0.930–0.950 and 0.943–0.959, respectively). At ϵ\epsilon = 0.1, Hoeffding–Laplace produces shorter intervals than EL/AEL. However, at ϵ\epsilon = 1 and 2, both EL and AEL produce shorter intervals across all six distributions.

AuthorsOpenReview ↗

Response to Reviewer 3WSf (Cont.)

Response to W2

(i) Differences from existing DP EL methods and AEL.

Liu and Li (2025), PriDEL/BAPriDEL. These methods consider KK physically separate machines that release local estimates under (ϵ,δ)(\epsilon,\delta)-DP using Gaussian noise, with the first machine's estimate left unprivatized. Their theoretical analysis assumes that the sensitivity of each local estimate is o(m−1/2)o(m^{-1/2}) (Condition A4). Their framework does not explicitly characterize the inferential target or bias induced by fixed clipping.

Our setting. We consider pure ϵ\epsilon-DP: every block mean is privatized with Laplace noise, and privacy follows from parallel composition over disjoint blocks. We also characterize the consequences of clipping: the signed bias (Prop. 2.1), the non-identifiability of μ\mu (Prop. 2.2) and an assumption-indexed interval for μ\mu (Remark 2.4).

AEL. We establish the asymptotic χ12\chi_1^2 limit of the AEL statistic for privatized block means (Thm. 2.5). The adjustment introduces a candidate-dependent pseudo-point with coefficient aBa_B growing with B, and we prove that this augmentation preserves the limiting distribution. This differs from the two-point augmentation used in BAPriDEL. Empirically, AEL produces shorter intervals than BAPriDEL in all 18 settings reported in Table 1.

(ii) Benefits relative to studentized intervals and the closest alternatives.

Studentized interval. we construct the studentized interval using the same privatized block means.

DesignnELAELNoise-aware studentized
Normal2000.935/3.510.944/3.580.961/3.67
4000.937/3.290.947/3.350.955/3.41
6000.938/3.060.945/3.110.955/3.15
Exp2000.950/4.160.957/4.250.968/4.34
4000.946/3.610.954/3.680.960/3.73
6000.944/3.190.949/3.250.958/3.27
Beta2000.940/0.850.950/0.870.965/0.89
4000.939/0.770.948/0.780.955/0.79
6000.941/0.700.947/0.710.957/0.72
Lognormal2000.941/5.040.949/5.150.962/5.27
4000.939/4.480.948/4.570.955/4.63
6000.940/4.050.946/4.110.955/4.14
Gamma2000.942/3.370.953/3.450.966/3.53
4000.942/3.000.950/3.050.958/3.10
6000.942/2.700.945/2.750.957/2.76
Weibull2000.942/1.760.950/1.800.964/1.84
4000.942/1.650.950/1.680.960/1.70
6000.943/1.530.948/1.550.958/1.57

The results show that the studentized interval is longer than AEL in all 18 settings. Its empirical coverage ranges from 0.954 to 0.968, compared with 0.944–0.957 for AEL. Thus, AEL gives shorter intervals and coverage closer to 0.95. Unlike studentized inference, EL/AEL construct intervals by inverting a likelihood-ratio statistic, without requiring an explicit variance estimator or a separately privatized variance estimate.

Closest alternatives. As detailed in our response to W1, AEL produces shorter intervals than PriDEL and BAPriDEL in all 18 settings. Compared with AW, AEL has a slightly shorter average interval length but substantially more accurate nominal coverage.

(iii) Theoretical arguments required for the privatized block means.

The classical fixed-distribution Wilks theorem does not apply directly to the released block means. To establish the χ12\chi^2_1 limits for our mechanism, we prove the following.

Vanishing variance and changing distributions. Each released observation is a clipped block mean plus independent Laplace noise, with variance

σB2=σc2m+2w2m2ϵ2⟶0.\sigma_B^2=\frac{\sigma_c^2}{m}+\frac{2w^2}{m^2\epsilon^2}\longrightarrow 0.

Consequently, the released observations form a triangular array whose distribution changes with sample size. Lemma L.1 establishes uniform standardized moment bounds along sequences with B→∞B\to\infty and m→∞m\to\infty. These bounds yield the required central limit theorem and maximal-term condition,

max⁡1≤j≤B∣x~j−μc∣σB=op(B),\frac{\max_{1\le j\le B}|\tilde x_j-\mu_c|}{\sigma_B}=o_p(\sqrt B),

despite the Laplace noise.

Feasibility and AEL. We prove that μc\mu_c lies in the convex hull of the privatized block means with probability tending to one, ensuring asymptotic feasibility of EL. For AEL, we prove that the candidate-dependent pseudo-point preserves the required moment and maximal-term conditions, yielding the χ12\chi_1^2 limit in Thm. 2.5.

Degenerate sampling variance. Our analysis also covers σc2=0\sigma_c^2=0, where the variability of the released observations arises entirely from Laplace noise.

Together, these results give asymptotically calibrated EL/AEL inference from privatized block means under pure ϵ\epsilon-DP.

Reference

Liu, Q. and Li, Z. (2025). Distributed EL inference with privacy guarantees. Statistics and Computing, 35, 88.

AuthorsOpenReview ↗

Response to Reviewer 3WSf (Cont.)

Response to W3

EL/AEL provide nonparametric, asymptotically calibrated confidence intervals for the clipped population mean under pure ϵ\epsilon-DP, without requiring an explicit variance estimator or an additional privacy release.

  1. No parametric model or explicit variance estimation. Unlike KV, which assumes normally distributed data, and AW, which simulates privatized statistics under a specified parametric model, EL/AEL do not require a parametric distributional assumption. Their likelihood-ratio statistics converge to the nuisance-free χ12\chi_1^2 distribution, allowing confidence intervals to be constructed directly from the privatized block means. No separate variance release is needed, and the full privacy budget is allocated to privatizing the block means. By contrast, AW spends part of ϵ\epsilon on a noisy variance, and KV splits ϵ\epsilon into three parts to estimate a scale before the mean.

  2. Accurate coverage and competitive interval lengths. In the 18 settings reported in our response to W1, AEL is the only method among the compared methods whose empirical coverage remains within 0.95±0.010.95\pm0.01 in every setting (0.944–0.957). AEL also produces shorter intervals than PriDEL and BAPriDEL in all settings. AW exhibits substantial overcoverage (0.984–0.991). These results demonstrate the favorable coverage calibration and competitive interval lengths of AEL.

  3. Explicit treatment of clipping. We characterize the exact signed clipping bias (Proposition 2.1), establish the non-identifiability of the raw mean under fixed clipping without additional assumptions (Proposition 2.2), and provide an assumption-indexed interval for the raw mean (Remark 2.4).

Together, these contributions provide a nonparametric approach to differentially private mean inference that provides asymptotic coverage calibration, competitive interval lengths, and explicit treatment of clipping.

Response to Q1

  • Block sizes and noise at n = 400, B = 70. Here n/B = 40/7 is not an integer, so our implementation approximates the design: each of the 70 blocks contains 5 records, and the Laplace scale is wB/(nϵ)=w/(407ϵ)wB/(n\epsilon)=w/(\tfrac{40}{7}\epsilon). Since a 5-record block mean has sensitivity w/5w/5, this approximation provides 87ϵ\tfrac87\epsilon-DP. The B = 70 results therefore show empirically how the method behaves when B does not divide n.
  • B dividing n. When B divides n, the scale w/(mϵ)w/(m\epsilon) equals sensitivity/ϵ\epsilon, so the release is exactly ϵ\epsilon-DP. We therefore reran the n = 400 experiments with B = 80.
  • Results with B = 80 (n = 400, ϵ\epsilon = 0.1):
NormalExpBetaLognormalGammaWeibull
EL, B = 700.937/3.290.946/3.610.939/0.770.939/4.480.942/3.000.942/1.65
EL, B = 800.938/3.360.945/3.720.942/0.790.942/4.610.943/3.080.942/1.68
AEL, B = 700.947/3.350.954/3.680.948/0.780.948/4.570.950/3.050.950/1.68
AEL, B = 800.945/3.410.949/3.790.947/0.800.948/4.690.948/3.140.950/1.71

The results with B = 80 are very close to those with B = 70: coverage changes by at most 0.005 and lengths by 2–3%, and the conclusions are unchanged.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.