acceptodds
Under review as a conference paper at ICLR 2027

Infinite quantiles inference under Local Differential Privacy

Abstract

We study simultaneous quantile inference under local differential privacy for a number of target probabilities that grows with the sample size. Each user participates once, so learning unknown quantile locations and calibrating their joint uncertainty must share the same private information. We fuse complementary cumulative distribution function (CDF) scores from a single tree report using their exact covariance; the variance at the correct anchor is no larger than that of either score, and anchor error incurs only quadratic excess variance. Three consecutive cohorts then localize the quantiles, estimate their densities, and supply fresh scores for a one-step correction and joint multiplier calibration. Under local density regularity, uniform identification on a public bracket, and explicit growth conditions, a uniform representation controls localization and density errors and yields asymptotically valid simultaneous coverage. A fixed-privacy regime permits every polynomially growing target grid. Simulations show consistent same-report precision gains, while comparisons with CDF-band inversion expose the sample cost of learning target locations. Census microdata illustrates how the resulting joint intervals support wage-percentile ratios.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Gaussian approximation, Growing dimension, Local differential privacy, Statistical infernece, Quantile

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_mtj7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates the problem of simultaneously inferring a large number of quantiles under Local Differential Privacy (LDP). The objective is to infer a set of quantiles that are predefined but whose number can increase with the sample size.

Critical strengths
  1. Utility improvement exceeds 20%
  2. For the first time, the minimum variance unbiased fusion is constructed using the exact covariance of tree-OUE prefix/suffix scores, without increasing privacy cost. This is an improvement at the decoding level and holds independent value.
  3. Separating localization, density estimation, and inference calibration makes Stage 3 conditionally independent and identically distributed, facilitating Gaussian approximation and multiplier calibration. This is a clean solution to the inference challenge under adaptive queries.
Critical weaknesses
  1. The scope of application is limited, for example, wages may not meet the theoretical assumptions of this article
  2. hard to follow. Leave me a hard time to review
  3. Observations are independent and identically distributed, which is not applicable to time series or clustering data.
  4. It does not support settings with outliers.
Questions for authors

see w

3 comments
AuthorsOpenReview ↗

Response to Reviewer mtj7

Thank you for your time and valuable comments, and for recognizing the decoding gain, the exact-covariance minimum-variance fusion and the separation of the three stages.

W1. The wage application uses a continuous i.i.d. working model for log wages. The 123531 log wages take 7260 distinct values; the largest point mass is 3.2%, and the masses at the 19 target quantiles are at most 2.7%. Fitted densities at the target locations are 0.123–0.507, with no density clipping or zero-variance fallback. In a stress test with an atom of the same size (n=160000n=160000, k=31k=31, ε=1\varepsilon=1, 500 repetitions, unchanged tuning), simultaneous coverage is 0.928, compared with 0.934 for continuous normal data. Revised Appendix A.6 reports these diagnostics and Table 10 the stress tests.

The same tests show how coverage responds to rounding and point masses:

InputCoverage
Continuous normal0.934
Rounding to mesh 0.020.944
Normal data with a 3.2% point mass at zero0.928
Rounding to mesh 0.100.728
Normal data with a 10% point mass at zero0.332
Rounding to mesh 0.500.000

A second stress test at n=40000n=40000, k=19k=19 shows the same pattern: 0.930 for normal data, 0.922 with the 3.2% atom, and 0.883 and 0.121 for meshes 0.1 and 0.5 (revised Table 14). In these simulated laws, fine rounding and the 3.2% atom retain near-nominal coverage, while coarse rounding and the 10% atom produce substantial undercoverage. The continuous-normal benchmark above is an independently seeded run (coverage/RMSE 0.934/0.08846; submitted Table 1 gives 0.932/0.0897). We present the wage application as a private release supporting joint percentile intervals and ratios; population-level inference for California would additionally require the survey weights and design.

W2. In plain terms, the procedure has three steps, each using a different group of users. The first group locates the target quantiles; the second estimates the data density at those locations; the third, fresh group produces the final estimates and the joint confidence band. Separating the groups makes the final step conditionally i.i.d., so standard Gaussian approximation and multiplier calibration apply. Each user sends a single privatized report, from which we decode two complementary estimates of the same probability and combine them with the variance-minimizing weight; this fusion is the source of the accuracy gain.

To make the construction easier to follow, the revised paper adds a worked four-leaf example in Section 2.1, with the complete step-by-step algorithm in Appendix A.1. Appendix A.3 gives the default tuning and how to choose it, and Appendix E explains the role of each assumption. Alongside the intervals, we report estimated densities, active truncations and the zero-variance fallback frequency.

W3. Each user participates once and sends one report, so repeated measurements of the same user do not enter the data; the requirement is independence across users, which is the standard sampling model for local differential privacy. Given that, the cohort separation supplies the conditional independence used for Gaussian approximation and multiplier calibration. Dependence across users, such as serial or clustered sampling, is a genuine limitation: it would require dependence-aware covariance estimation and calibration, as the revised Concluding Remarks state. For the same reason, we describe the wage intervals under a continuous i.i.d. working model rather than as survey-population inference.

W4. The method accommodates outliers and heavy tails. Observations need not be bounded: the public bracket only has to contain the target quantiles, observations outside it fall into the outer categories, and every privatized score stays bounded. The analysis therefore uses only local regularity near the targets, not tail moments of the raw data. The submitted experiments confirm this at (n,k,ε)=(640000,63,1)(n,k,\varepsilon)=(640000,63,1) (revised Table 7):

DistributionFused coverageFused RMSEBalanced-decoder RMSE
Standardized t3t_30.9480.04830.0597
0.9N(0,1)+0.1N(0,16)0.9N(0,1)+0.1N(0,16)0.9380.08420.1053

Fused inference keeps near-nominal coverage and reduces RMSE by about 19–20% relative to balanced decoding under both heavy tails and contamination; the mixture experiment targets the mixture's quantiles. An atom at a target quantile, of any size, violates the positive continuous local-density assumption; the stress tests in W1 describe the resulting finite-sample behavior for the simulated laws.

Thank you again for your valuable feedback. We hope our responses address your concerns and welcome any further questions or discussion.

Reviewer_mtj7OpenReview ↗

Keep the score

I do not question the innovativeness of this paper, but its writing needs further improvement. Moreover, I believe this paper may be more suitable for a conference related to databases

Thank you for your further reading and for recognizing the paper's originality. The central question is how to quantify uncertainty about a distribution when individuals can release only locally private information. We study simultaneous confidence intervals for an increasing collection of quantiles, linking the accuracy of private estimation to coverage across that collection. This connects the work to statistical learning under privacy constraints. The worked example and the additional coverage experiments make the target and the role of the assumptions more concrete.

Your feedback is helpful as we work to communicate the contribution more clearly. We appreciate your careful consideration of the work and the time you have spent on the discussion.

Reviewer_ug1dOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the joint estimation of multiple quantiles from a sample under local differential privacy. The construction uses a binary tree to define the node-level statistics ψv(Z)\psi_v(Z). These yield two statistics, AA and BB: one estimates the CDF directly through a prefix sum, while the other estimates the same CDF through the complement of a suffix sum. A weighted combination of these statistics produces the fused score GG. Proposition 2.1 characterizes the variance reduction achieved by this combination. The optimal weights are obtained by minimizing the variance of the linear combination. Sections 2.2 and 2.3 describe how these scores are used to estimate quantile locations and local densities, and subsequently construct simultaneous confidence intervals. Theorem 3.1 establishes a uniform representation resembling a Bahadur representation, while Theorem 3.2 establishes the validity of the joint calibration as the number of target quantiles increases. The paper also includes simulations and a real-data application.

Critical strengths

The paper addresses simultaneous inference for multiple quantiles under local differential privacy, allowing the number of target quantiles to grow with the sample size.

The complementary-score construction provides a concrete way to reduce variance using the same private report. The subsequent uniform representation and joint calibration connect this construction to simultaneous confidence intervals.

Critical weaknesses
  1. I found the presentation of the tree-based mechanism difficult to follow. In particular, the connection between the tree structure, the privatized report, the node-level statistics, and the two CDF scores could be explained more intuitively. A small worked example would help.

  2. Although the algorithm and default parameter settings are provided, the procedure involves several tuning choices across three stages. Their practical selection and their effects on coverage are not sufficiently clear to me, particularly when the underlying distribution is unknown.

Questions for authors
  1. Contribution of the fusion step. The variance-minimizing weight follows from a standard quadratic minimization. Could the authors explain more clearly which aspects of the exact covariance calculation and fusion analysis constitute the main technical contribution?

  2. Tuning and implementation. How should practitioners choose the stage allocation, learning-rate parameters, bandwidth, and density stabilization when the distribution is unknown? The existing allocation comparison indicates that these choices can affect coverage. More guidance on this sensitivity would be useful.

  3. Extreme quantiles and theoretical scope. The growing-grid theory restricts the target probabilities to a fixed interior interval. This restriction should be emphasized when describing “infinite quantiles,” since an increasingly dense interior grid does not establish validity for quantile levels approaching zero or one.

  4. Privacy–accuracy trade-off and evaluation. The experiments compare methods at matched privacy budgets and sample sizes. Could the authors additionally clarify how much statistical accuracy is lost relative to a non-private benchmark? As a supplementary evaluation, quantile loss and results separated by quantile level could help explain performance beyond the pooled RMSE.

  5. Discrete observations and the application. The wage data contain repeated values, while the inference relies on a continuous working model and local density regularity. How sensitive is simultaneous coverage to rounding and concentrated probability masses? The distinction between inference for recorded wages and inference under the continuous working model would benefit from further clarification.

6 comments
AuthorsOpenReview ↗

Response to Reviewer ug1d (1/2)

Thank you for your time and valuable comments.

W1. Here is a four-leaf example. Thresholds x1<x2<x3x_1 < x_2 < x_3 form leaves C0,…,C3C_0,\ldots,C_3; level 1 has nodes v01v_{01}, v23v_{23}. A user samples a level uniformly, encodes the containing node as a one-hot vector, and randomizes each bit independently: the true-node bit is one with probability 1/21/2, any other bit with probability (eε+1)−1(e^\varepsilon+1)^{-1}. The node decoder ψv\psi_v debiases its bit on the sampled level, rescales by two, and is zero otherwise; its conditional expectation is the indicator of node vv. At the first two thresholds, A1=ψC0,B1=1−ψC1−ψv23,A2=ψv01,B2=1−ψv23.A_1=\psi_{C_0},\quad B_1=1-\psi_{C_1}-\psi_{v_{23}},\qquad A_2=\psi_{v_{01}},\quad B_2=1-\psi_{v_{23}}. With p=F(x1)p=F(x_1) and ν=4eε/(eε−1)2\nu=4e^\varepsilon/(e^\varepsilon-1)^2, disjoint-node cross-products have expectation zero, so Var⁡(A1)=2ν+4p−p2,Var⁡(B1)=4ν+4(1−p)−(1−p)2,Cov⁡(A1,B1)=p(1−p).\operatorname{Var}(A_1)=2\nu+4p-p^2,\quad \operatorname{Var}(B_1)=4\nu+4(1-p)-(1-p)^2,\quad \operatorname{Cov}(A_1,B_1)=p(1-p). The fusion weights at a public anchor uu are 4ν+3(1−u)/(6ν+3)\\{4\nu+3(1-u)\\}/(6\nu+3) on A1A_1 and 2ν+3(1−u)/(4ν+3)\\{2\nu+3(1-u)\\}/(4\nu+3) on A2A_2. The revised PDF adds this example to Section 2.1, with the complete algorithm in Appendix A.1.

W2. We ran sensitivity analyses for allocation, initialization, bandwidth, gain and density stabilization. At n=40000n=40000, k=19k=19, ε=1\varepsilon=1, the study uses bracket [−3,3][-3,3], 256 repetitions and 799 conditional multiplier draws. Bandwidth multipliers 0.75, 1 and 1.25 give coverage 0.926, 0.930 and 0.926 and RMSE 0.3535, 0.3195 and 0.2856. Gain multipliers 0.5 and 2 give coverage 0.887 and 0.922. Relative to default coverage 0.930 and width 1.1043, split (0.25,0.20,0.55)(0.25,0.20,0.55) gives coverage 0.867 and width 0.9275, while split (0.45,0.30,0.25)(0.45,0.30,0.25) gives coverage 0.945, width 1.2732 and RMSE 0.3742. We use (0.35,0.30,0.35)(0.35,0.30,0.35) as the starting allocation: increasing the inference share reduces its variance, but the smaller learning cohorts can impair coverage.

A separate paired study at n=40000n=40000, k=31k=31, bracket [−3,3][-3,3], uses 500 repetitions and 500 multipliers, with fresh seeds independent of the submitted Table 2 experiment. Density floor/cap multipliers 0.5, 1 and 2 give coverage 0.926, 0.928 and 0.934 with RMSE 0.3671, 0.2663 and 0.2144; halving the gain gives coverage 0.876. The default in submitted Table 2 has coverage 0.942 and RMSE 0.2704.

At n=320000n=320000, k=63k=63, bandwidth multipliers 0.75, 1 and 1.25 give coverage 0.930, 0.944 and 0.970; the multiplier-one result is the default experiment in submitted Table 1. A separate 500-repetition initialization comparison pairs the seeds across starts and gives coverage 0.940 for linear interpolation versus 0.946 for a common midpoint. The paired-data nonprivate comparison in Q4 is another independently seeded experiment. Frequent projection or density clipping is a reason to examine localization and density estimation; coverage still requires the theorem's regularity and rate conditions. These analyses appear in revised Appendices A.3 and B.1 (Tables 11 and 12).

Q1. The contribution is the exact same-report covariance and the control of anchor error. Prefix and complementary suffix scores from one report have covariance pr(1−pr)p_r(1-p_r), and Var⁡Gr(u)−Var⁡Gr(pr)=(2Hd−1)2(u−pr)2Hd(wr−+wr+)νε+2Hd−1≤(2Hd−1)(u−pr)2.\operatorname{Var}\\{G_r(u)\\}-\operatorname{Var}\\{G_r(p_r)\\}=\frac{(2H_d-1)^2(u-p_r)^2}{H_d(w_r^-+w_r^+)\nu_\varepsilon+2H_d-1}\le(2H_d-1)(u-p_r)^2. The feasible anchor u=τu=\tau therefore costs only quadratic excess variance, and conditional unbiasedness holds throughout localization because the weight depends on pre-report history. Optimality is among combinations of the two scores with weights summing to one.

Q2. The defaults use public quantities: cohort fractions (0.35,0.30,0.35)(0.35,0.30,0.35), full Stage-1 averaging, gain 0.1(U−L)t−2/30.1(U-L)t^{-2/3}, bandwidth h=B2k2log⁡2k(n+1)/n21/4h=\\{B_{2k}^2\log\\{2k(n+1)\\}/n_2\\}^{1/4} and floor ϖn=log⁡(n+e)−3/2\varpi_n=\\{\log(n+e)\\}^{-3/2} with reciprocal upper cap. The gain condition is γ0≤(2LF)−1\gamma_0\le(2L_F)^{-1}. When public information supplies an upper bound on LFL_F, choose γ0\gamma_0 no larger than the reciprocal of twice that bound. For the normal designs, LF=ϕ(0)L_F=\phi(0) verifies the default. Without such information, the public default and prespecified nearby choices provide a sensitivity analysis; verifying the gain condition remains model-dependent. The bracket comes from public knowledge or a separately budgeted private pilot, whose users and privacy enter the total budget. We report estimated densities, active density floor/cap events and zero-variance fallback frequency. In the ACS illustration, fitted densities range from 0.123 to 0.507, with no density clipping or zero-variance fallback. Revised Appendix A.3 states these defaults and the gain rule.

AuthorsOpenReview ↗

Response to Reviewer ug1d (2/2)

Q3. The revised PDF uses the title Simultaneous Quantile Inference on Growing Grids under Local Differential Privacy. The theory requires every target to lie in a fixed [τ−,τ+]⊂(0,1)[\tau_-,\tau_+]\subset(0,1), within which the grid may become increasingly dense. Probabilities approaching zero or one require separate control of density lower bounds, localization rates and remainders.

Q4. We compared estimators on independent private/nonprivate simulation streams from the same normal design, each with n=40000n=40000, k=19k=19, ε=1\varepsilon=1, bracket [−3,3][-3,3], 256 repetitions and 799 multipliers. Pooled RMSE is 0.3195 (fused), 0.4920 (balanced), and 0.0074 (nonprivate); mean population excess check loss, multiplied by 10610^6, is 11904, 27506, and 7.23, respectively. At τ=0.05,0.50,0.95\tau=0.05,0.50,0.95, fused RMSE is 0.5877, 0.1241, 0.8322, versus nonprivate 0.0101, 0.0062, 0.0104. The nonprivate estimator uses all nn raw observations. In the two larger normal settings below, private and nonprivate estimators use the same observations, bracket [Φ−1(0.02),Φ−1(0.98)][\Phi^{-1}(0.02),\Phi^{-1}(0.98)], ε=1\varepsilon=1, 500 repetitions and 500 multipliers. These runs are independently seeded from submitted Table 1, whose corresponding coverage/RMSE entries are 0.932/0.0897 and 0.944/0.0698.

nnkkCoverageFused RMSENonprivate RMSEFused lossNonprivate loss
160000310.9340.088460.00364902.481.78
320000630.9280.069880.00258587.320.92

Loss is expected excess check loss ×106\times10^6, averaged over target probabilities and repetitions. For a fitted estimate qq, it is evaluated over a fresh observation from the generating law, equivalently ∫QτqF(t)−τ,dt\int_{Q_\tau}^q\\{F(t)-\tau\\}\\,dt. The private RMSE is about 24–27 times the nonprivate error in the two tabulated settings, and about 43 times at (40000,19)(40000,19): the joint cost of report noise and allocation to learning. Coverage 0.928 in the second tabulated setting indicates finite-sample undercoverage. Revised Section 4.2 summarizes this comparison, and Appendix B.1 (Tables 8, 9 and 13) gives the details.

Q5. The local expansion uses a positive continuous density at each target. The rounding experiment compares mesh rounding and a 3.2% atom at n=40000n=40000, k=19k=19: coverage/RMSE 0.930/0.3195 (normal), 0.883/0.3111 (mesh 0.1), 0.121/0.3813 (mesh 0.5) and 0.922/0.3593 (atom), evaluated against each input law's generalized-inverse quantiles. At n=160000n=160000, k=31k=31, recorded-target coverage is 0.944 (mesh 0.02), 0.728 (mesh 0.10), 0.000 (mesh 0.50), 0.928 (3.2% atom) and 0.332 (10% atom). The ACS illustration uses a continuous working model for log wages; the discrete file distribution and California population coverage require methods for jumps and for the survey design. Revised Appendix B.1 (Tables 10 and 14) and Appendix A.6 report these results and the ACS diagnostics.

Thank you again for your valuable feedback. We hope our responses address your concerns and welcome any further questions or discussion.

Thank you for the worked example, the clarification of the theoretical scope, and the additional experiments. These address several of my concerns. However, the new coverage results leave two important questions about practical reliability.

  1. Tuning when the distribution is unknown. Coverage drops from 0.930 under the default allocation to 0.867 when more observations are allocated to inference. Meanwhile, verifying the gain condition still requires information about the unknown distribution. Can the reported diagnostics—such as projection frequency and density floor/cap events—identify these low-coverage settings? Please explain how these diagnostics behave in the failure cases already evaluated, and what actionable guidance a practitioner can obtain from them.

  2. Interpretation of the ACS application. The rounding experiments show substantial undercoverage: at n = 160,000, coverage falls from 0.944 for mesh 0.02 to 0.728 for mesh 0.10 and 0.000 for mesh 0.50. Given these results, what evidence supports treating rounding and concentrated probability masses as negligible at the target quantiles in the ACS data? Please specify the continuous estimand, the assumptions connecting it to recorded wages, and the intended interpretation of the reported intervals. The absence of density clipping alone does not establish that the continuous approximation is adequate.

Reviewer_4zE9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors consider simultaneous inference for quantiles under LDP. They allow the number of target quantiles to grow with the sample size (hence "infinite" in the title). The target quantile levels are fixed in advance, and a public finite range is assumed to contain the target quantiles. The authors' approach proceeds in three stages, each using a separate cohort of users. First, they combine the tree mechanism of Cormode et al. (2019) with the stochastic-approximation approach of Liu et al. (2023) to produce LDP quantile estimates jointly, rather than running Liu et al. separately for each quantile; they also fuse the prefix- and suffix-based CDF estimates from the tree to reduce variance. Second, they estimate the density at each estimated quantile. Third, using fresh reports at the frozen estimates, they apply a one-step correction and estimate the asymptotic variance and covariance parameters. Finally, they use a Gaussian multiplier (wild) bootstrap, i.e., Monte Carlo from the estimated asymptotic distribution of the maximum, for simultaneous inference. To justify this, they show that localization and density-estimation errors are uniformly negligible, so that the high-dimensional Gaussian approximation of CCK applies with the dimension k growing with n.

Critical strengths

Probably the biggest strength is the simultaneous inference for a growing number of quantiles. Often inference isn't prioritized in the privacy literature, or if it is, the inference often treats the privacy noise as negligible. Neither is the case here and the cost is fairly modest in terms of error growth in the number of quantiles.

Critical weaknesses

The paper comes across as fairly complete, and I appreciate some of the candid empirical results about the their approach vs CDF inversion. However, I don't see where their approach "wins", which seems important when describing their approach as a kind of tradeoff.

On a more minor note, I really dislike using the term "infinite" here to sell the paper. To me, infinite would mean that one could achieve LDP with a growing number of quantiles and an error that doesn't continue to grow with kk. This is actually possible in other settings (e.g. Hall, Rinaldo, and Wasserman, 2013). I get that their error scales with log(k) (e.g. Cor 3.3), but that still isn't "infinite".

Questions for authors

See comments above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 4zE9

Thank you for your time and valuable comments.

W1. The submitted experiments demonstrate gains in decoder accuracy and simultaneous-interval width.

ComparisonSetting (n,k,ε)(n,k,\varepsilon)Fused: coverage / RMSE / widthComparator: coverage / RMSE / width
Decoder comparison(1280000,127,1)(1280000,127,1)0.950 / 0.0386 / —Balanced: 0.952 / 0.0494 / —
Complete inferential procedures(640000,127,1)(640000,127,1)0.948 / 0.0561 / 0.3574Student–Wishart: 0.944 / 0.0570 / 0.5902

In the submitted decoder experiment, fusion reduces RMSE by 21–23% while fixing the reporting channel, privacy budget, sample size and cohort allocation. Table 1 compares decoder rules on independent simulation streams; “same-report” refers to decoding the prefix and complementary suffix scores from one user's report. For those scores A,BA,B, the variance-minimizing combination wA+(1−w)BwA+(1-w)B has w=Var⁡(B)−Cov⁡(A,B)/Var⁡(A)+Var⁡(B)−2Cov⁡(A,B)w=\\{\operatorname{Var}(B)-\operatorname{Cov}(A,B)\\}/\\{\operatorname{Var}(A)+\operatorname{Var}(B)-2\operatorname{Cov}(A,B)\\}. The exact covariance is p(1−p)p(1-p) at CDF value pp, so the gain follows from exploiting the dependence between the two scores.

At k=127k=127, the complete-procedure comparison gives approximately 39% narrower intervals at similar coverage and RMSE. At k=31,63k=31,63, Student–Wishart has smaller RMSE but wider intervals. This comparison captures the combined effects of the decoder, bracket, restart, averaging, allocation and calibration; Student–Wishart uses a public model correlation matrix.

CDF inversion also uses the fused decoder. Under N(0,1)N(0,1) on [−3,3][-3,3], with one ε=1\varepsilon=1 report per user, the complete-procedure comparison is:

(n,k)(n,k)Direct: coverage / RMSE / widthCDF inversion: coverage / RMSE / width
(40000,31)(40000,31)0.942 / 0.2704 / 1.11121.000 / 0.1064 / 0.8216
(160000,31)(160000,31)0.950 / 0.1134 / 0.51460.994 / 0.0512 / 0.5067
(320000,63)(320000,63)0.918 / 0.0809 / 0.43260.992 / 0.0395 / 0.3556

CDF inversion has smaller RMSE and narrower intervals in these settings, with conservative coverage of 0.992–1.000. Direct coverage at (320000,63)(320000,63) is 0.918, indicating finite-sample undercoverage. The accuracy gain comes from complementary decoding, which both routes use. Direct inference learns locations and densities for prespecified probability targets, with a uniform one-step representation on growing grids. CDF inversion instead chooses public spatial points: at (320000,63)(320000,63), refining its inspection grid from d=63d=63 to 127127 leaves the simulated width essentially unchanged (0.3556 vs 0.3564) while increasing mean upload length from 24.0 to 39.3 bits, versus 28.6 bits for the direct procedure. The direct procedure also updates public queries during localization, so this is an upload comparison, not total communication.

W2. The revised PDF uses the title Simultaneous Quantile Inference on Growing Grids under Local Differential Privacy. At every sample size the target grid is finite, and its size may increase while all probabilities remain in a fixed interior interval, as required in Section 3 and Assumption D.1 of the submission.

Hall, Rinaldo and Wasserman (2013), Differential privacy for functions and functional data, study function-valued releases. Their Gaussian-process construction uses a database-level sensitivity bound uniform over finite evaluation sets, with an RKHS formulation in Section 3.4, and provides approximate database-level DP. Our procedure provides pure local DP and simultaneous quantile inference. The privacy models and inferential targets therefore differ.

At fixed privacy, a=2/3a=2/3, β=1\beta=1, fixed positive cohort fractions and the remaining regularity, bandwidth and gain conditions of Corollary 3.3, the sufficient restriction log⁡92kn(n+1)n⟶0\frac{\log^9\\{2k_n(n+1)\\}}{n}\longrightarrow0 permits every polynomial growth rate of knk_n. With fixed density lower bound and cohort fractions, the interval-width upper bound is of order Hknlog⁡(2kn)/n=O(log⁡kn)3/2/nH_{k_n}\sqrt{\log(2k_n)}/\sqrt n=O\\{(\log k_n)^{3/2}/\sqrt n\\}. Here HknH_{k_n} comes from the tree-score scale and log⁡(2kn)\sqrt{\log(2k_n)} from bounding the Gaussian maximum in simultaneous calibration. Whether this dependence on grid size is necessary is a separate lower-bound question; Corollary 3.3 establishes a sufficient upper bound at the prespecified horizon.

Thank you again for your valuable feedback. We hope our responses address your concerns and welcome any further questions or discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.