Pith. sign in

REVIEW 3 major objections 4 minor 103 references

Differentially Private Nonparametric Modal Learning with Applications to Regression and Clustering

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read Differentially private mode estimation is nearly minimax optimal for smooth densities.

desk verdict First DP method for multiple density modes; upper bounds look solid, but the minimax lower bound (Appendix A.4) has a wrong score formula and an invalid score-to-mode reduction, so the near-optimality claim is unproven as written. read the letter →

arxiv 2607.29675 v1 pith:XTDQVZ5M submitted 2026-07-31 math.ST cs.LGstat.MEstat.MLstat.TH

classification math.STcs.LGstat.MEstat.MLstat.TH MSC 62G0762G2062G0862G0562H30
keywords densitymodesdifferentialprivacymeanshiftkernelestimationminimaxratemodalregressionmodeclusteringscore
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 introduces DP-GRAMS, a differentially private algorithm that recovers all modes of a multimodal density while guaranteeing (ε, δ)-differential privacy. It shows that private mode estimation is possible with mean-squared error that separates into the classical non-private rate (log n/n)^(2(β−1)/(d+2β)) plus a privacy term (polylog(n,δ)/(n²ε²))^((β−1)/(d+β)), and it proves matching lower bounds up to logarithmic factors. The method is mean-shift inspired: noisy gradient ascent on a clipped, floored KDE score, with Gaussian noise correlated across parallel starts to avoid a privacy penalty for multiple initializations. A density-aware private initialization (DAP) places at least one start in every modal basin with high probability. The same machinery yields private modal regression (DP-PMS) and mode-based clustering (DP-GRAMS-C).

What carries the argument

The stabilized score estimator ŝ_{A,p_floor}(x) = clip_A(∇p̂(x)) / max{p̂(x), p_floor}, where clip_A bounds the KDE gradient and p_floor keeps the denominator away from zero; gradient ascent on this field is the private analogue of mean shift. Privacy is carried by Gaussian noise with covariance σ²K_t across the k starts at each iteration, where K_t is the exponential kernel matrix of current iterates; the DAP initialization uses the exponential mechanism on local empirical mass with successive suppression radii, guaranteeing basin coverage with k ≍ M log n draws.

What would settle it

Run DP-GRAMS on a fixed multimodal distribution with a deliberately misspecified public box Q that excludes one mode; if the exclusion is known, the missing mode is never recovered, violating the coverage claim. Alternatively, a concrete calculation: check whether the conditional MSE scales with n and ε as predicted on a dense grid of a 2-Gaussian mixture with β=3 and d=2; significant deviation from the predicted exponents would falsify the rate.

Watch

Extended reading notes

Core claim

On the paper's own terms: modes of a density, characterized as zeros of ∇log p with negative Hessian, can be recovered under (ε,δ)-DP by running clipped, floored KDE score ascent from private initializations. Theorem 4.5 states that, on a high-probability global good event, the merged estimator M-hat contains, for each true mode μ_j, a point μ-hat_j with conditional MSE ≤ C_nonDP (log n/n)^(2(β−1)/(d+2β)) + C_DP (T d polylog(n,δ)/(n² ε_modes²))^((β−1)/(d+β)) under β-Hölder smoothness β>2, curvature, and separation. Theorem 4.6 gives minimax lower bounds of order n^(−2(β−1)/(d+2β)) + (nε)^(−2(β−1)/(d+β)), so the upper bound is nearly optimal up to log factors. The construction treats the scor

Load-bearing premise

The algorithm assumes a known public box Q containing every modal basin and a public grid inside it; if that box is wrong or unknown, the initialization coverage proof collapses, and the paper offers no analysis of how to find the domain privately.

Editorial extensions

If this is right

  • All population modes are recovered with high probability, and the per-mode MSE matches the minimax lower bound up to log factors.
  • The privacy cost enters only as an additive term of order (n²ε²)^(−(β−1)/(d+β)); with ε large enough the non-private rate dominates.
  • Higher-order kernels (order ⌊β⌋) are necessary to exploit smoothness β>2; nearest-neighbour estimators do not benefit from such smoothness.
  • The same score-ascent primitives transfer to private modal regression and to private clustering via released centers with deterministic post-processing assignments.
  • The novel differentially private score estimation lower bound is of independent interest.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the public box assumption fails (e.g., domain unknown), DAP coverage breaks and the algorithm must spend privacy budget on domain discovery; the paper does not analyze this, so real deployments would need domain priors or a two-stage private box estimation.
  • The correlated-noise trick suggests a general recipe: releasing a function evaluated at nearby points can be done for the price of one Gaussian mechanism via RKHS kernels; this likely extends beyond mode estimation to other iterative nonparametric procedures.
  • The rate separation hints that private mode estimation is essentially as hard as private score estimation; any better private score estimator (e.g., deep score networks with DP) would directly improve mode recovery.
  • The ε_modes vs ε_init budget split and the threshold ε_thr suggest an adaptive privacy-budget allocation rule: spend just enough on initialization to guarantee coverage, and put the rest into ascent noise.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes DP-GRAMS, a differentially private multi-mode estimation algorithm built on noisy gradient ascent over a stabilized KDE score, with a density-aware private initialization (DAP), correlated Gaussian noise across multiple starts, and a final merging step. The main theoretical claims are: (i) Proposition 4.4, DAP covers every modal basin with high probability under a public-box/public-grid assumption; (ii) Theorem 4.3 and Theorem 4.5, conditional on a high-probability 'good event,' each true mode is recovered with conditional MSE of order (log n/n)^{2(β−1)/(d+2β)} + (T d polylog(n,δ)/(n²ε²_modes))^{(β−1)/(d+β)}; and (iii) Theorem 4.6, a minimax lower bound of the same two terms (up to constants and log factors), leading the abstract to claim near-minimax optimality. The paper also presents DP-PMS for modal regression and DP-GRAMS-C for clustering, with an extensive experimental section on synthetic and real data.

Significance. If the results hold, this is an important contribution to differentially private nonparametric statistics: it provides the first smoothness-adaptive private mode estimator for multiple modes, with a nontrivial privacy mechanism (correlated noise over parallel starts) and a plausible initialization theory. The upper-bound analysis is elaborate and largely coherent, and the experiments are broad and well reported, including sensitivity analyses and a reproducibility repository. The claimed minimax optimality is the headline theoretical achievement, and it rests on the lower-bound proof in Appendix A.4. That proof contains concrete algebraic and inferential errors, so the near-optimality claim is currently unproven. The DAP coverage result also depends on a known public box and public grid, an explicit limitation that is acknowledged in the design but should be prominently stated. On balance, the central contribution is defensible but needs substantial repair before the main claims can be accepted.

major comments (3)
  1. [Appendix A.4, Eq. (18) and surrounding derivation] The score formula for p_θ is misprinted. Direct differentiation of p_θ(x)=p_0(x)(1+θh^β ψ(x/h)) gives s_{p_θ}(x)=∇p_0(x)/p_θ(x)+θh^{β−1}∇ψ(x/h)/(1+θh^β ψ(x/h)). The printed expression in the proof, 'spθ(x) = 1/pθ(x)∇p0(x) + θh^{β−1}ψ(x/h)∇ψ(x/h)', has an extra ψ factor and omits the denominator. This is not a cosmetic typo: the stated separation bound (18) is used to convert the score lower bound into the mode lower bound. The corrected formula still gives separation at x=0 (up to a factor 2/(1−h^{2β})∇ψ(0)), so this part is likely repairable, but the proof as written cannot be followed.
  2. [Appendix A.4, reduction from score estimation to mode estimation] The step '0 = ŝ(x̂) = ŝ(x̂)−s(x̂)+s(x̂) = ŝ(x̂)−s(x̂)+s(0)+(H(ξ))(x̂−0)' is invalid in two ways. First, it assumes an arbitrary (ε,δ)-DP estimator x̂ is a root of the estimated score, which is not guaranteed by the statement of Theorem 4.6. Second, for the alternatives p_θ used immediately above, s_{p_θ}(0)=θh^{β−1}∇ψ(0)/(1+θh^β ψ(0))≠0, so the expansion cannot be centered at 0; the mode of p_θ is itself shifted from 0 by Θ(h^{β−1}). Consequently the displayed inequality E∥x̂−0∥² ≥ [Eλ_max(H(ξ))]⁻² E∥ŝ(x̂)−s(x̂)∥² does not follow. Since Theorem 4.6 is the basis for the abstract's 'nearly minimax optimal' claim, this is load-bearing. A two-point or Le Cam argument comparing the shifted modes directly may reproduce the intended rates, but the current proof is incomplete.
  3. [Section 4.2.2, Proposition 4.4] The DAP coverage guarantee requires a known public box Q containing all modal basins B(μ_j,r_j) and a public h_DAP-grid Z_n. If Q is misspecified or is not public, Proposition 4.4 fails, and the algorithm must spend privacy budget to find the domain; that case is not analyzed. This is an explicit and honest design assumption, not a proof error, but it is essential to the global recovery theorem. The main text should state prominently that the practical applicability of the coverage guarantee is conditional on this public-domain knowledge, since the asymptotic statement of Theorem 4.5 otherwise overstates the algorithm's autonomy.
minor comments (4)
  1. [Definition 5] The notation \nabla^s \hat\ell(x)-\nabla^s \ell(x) for s=0,1,2 is ambiguous: for s=0 one must mean absolute value, and for s=1,2 the operator norm. Please define the norm explicitly.
  2. [Section A.2, Lemma 5] The statement 'Suppose Assumptions 2, 4, and 4 hold' contains a duplicated assumption number; likely one of them should refer to Definition 4 or a different numbered assumption.
  3. [Eq. (11)] The expression for ε_thr mixes powers of n and log n in a way that is hard to parse; please rewrite it with explicit exponents and state which regime assumptions are used in the proof of Theorem 4.3.
  4. [Theorem 4.5 and Theorem 4.6 comparison] The upper bound is conditional on the high-probability good event X∈X_global, while the lower bound is an unconditional minimax risk. The comparison 'nearly minimax' is therefore only informal unless the failure probability is handled by a bounded-loss argument or an explicit risk decomposition. Please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation is self-contained; the lower-bound proof issue is a correctness concern, not a circularity.

full rationale

The paper's central derivation chain is not circular. The upper bound (Theorems 4.3 and 4.5) is obtained from explicit bias-variance and privacy-noise decompositions: deterministic sensitivity of the stabilized score (Lemma 1), Gaussian/advanced composition (Lemmas 2–3), uniform KDE derivative control (Lemmas 6–8), a stopped basin recursion (Proposition A.2), high-probability basin retention (Proposition A.3), and an exponential-mechanism coverage argument for DAP initialization (Proposition 4.4). None of these steps fits a parameter to the target modes or to the final released estimator; the bandwidth and noise scale are set analytically from the stated smoothness, privacy budget, and sample size. Self-citations are not load-bearing: the only author-overlapping citation (Auddy et al. 2025) appears in the discussion of future adaptivity and is unrelated to the main theorems. The DAP public-box and public-grid assumptions are explicit suppositions used to derive coverage, not conclusions assumed from the target. The minimax lower bound in Theorem 4.6 is an independent construction based on perturbed densities, KL/TV contraction under privacy, and a two-point argument; it is not derived from the upper bound or from fitted outputs. The fact that Appendix A.4 contains a concrete mathematical error in the reduction from score estimation to mode estimation (the claimed s(0)=0 step and the score formula) is a proof-correctness problem, not circularity, because it does not make the claimed result equivalent to its inputs by construction. Accordingly, no circular step meets the evidentiary standard of the review rules.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The theoretical analysis introduces no data-fitted parameters; all hyperparameters are deterministic functions of the sample size and model constants. The central claim rests on standard nonparametric assumptions (Hölder smoothness, separation, curvature) plus a known public domain, plus standard DP composition tools. No new physical or conceptual entities are postulated.

free parameters (5)
  • DAP draw count k
    Set as k ≍ M log n; deterministic function of M and n, not fitted to data.
  • Suppression radius ρ_init
    Set as (log n)^{-1/d} in the DAP design.
  • Ascent bandwidth h_mode
    Set via the order-optimal rule (11); depends on β, d, n, ε but not on data.
  • Step size η
    Any constant 0 < η ≤ η̄_j; not fitted.
  • Clipping level A and density floor p_floor
    Chosen in Proposition A.1 as constants depending on population quantities, not fitted to data.
assumptions (7)
  • domain assumption Assumption 1: kernel regularity (order ℓ kernel with moment conditions and tail decay)
    Needed for bias control and RKHS sensitivity integrals in Lemmas 1, 3, 6.
  • domain assumption Assumption 2: density is locally β-Hölder with β>2, positive near modes, modes separated by c0
    Defines the statistical model and is used throughout.
  • domain assumption Assumption 3: bandwidth condition h→0, n h^{d+4}/log n → ∞
    Needed for uniform KDE derivative rates.
  • domain assumption Assumption 4: local strong concavity of log p at modes (curvature)
    Yields inward drift toward the mode.
  • domain assumption Known public box Q containing all modal basins and public h_DAP-grid
    The DAP initialization is built on a public candidate set; without a known Q, the coverage guarantee fails.
  • standard math Standard DP tools: advanced composition, exponential mechanism, Gaussian mechanism, subsampling amplification
    Used for privacy accounting.
  • standard math Bernstein/Bennett inequalities for uniform KDE deviation bounds
    Used in Lemma 6.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Differentially Private Nonparametric Modal Learning with Applications to Regression and Clustering." pith.science (2026). https://pith.science/paper/XTDQVZ5M

@misc{pith2026260729675,
  author       = {Pith},
  title        = {Pith review of: Differentially Private Nonparametric Modal Learning with Applications to Regression and Clustering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XTDQVZ5M}},
  note         = {Machine review of arXiv:2607.29675}
}
abstract

Density modes provide a localized and interpretable summary of multimodal distributions, but their estimation under rigorous differential privacy constraints remains largely unexplored. We study differentially private recovery of density modes for multivariate distributions under local smoothness, curvature, and separation conditions. We propose DP-GRAMS, a mean-shift inspired method that performs noisy ascent on a differentially private score estimator. Assuming the density belongs locally to a H\"older class with smoothness parameter $\beta > 2$, our score estimator uses bias-reducing higher-order kernels, and then enforces privacy in the gradient ascent steps via gradient clipping and calibrated Gaussian noise. A private initialization scheme combines a density-aware utility with a suppression rule and, with $k\asymp M\log n$ draws over a public $h_{\mathrm{DAP}}$-grid and suppression radius $\rho_{\mathrm{init}}\asymp (\log n)^{-1/d}$, achieves high-probability coverage of the modal basins by successively suppressing selected local neighborhoods in competitive regions, while correlated noise across multiple starts enables joint release under a single $(\varepsilon,\delta)$-differential privacy guarantee. We prove that all population modes are recovered with high probability and establish asymptotic error rates of the form $O\!\left((\tfrac{\log n}{n})^{\frac{2(\beta-1)}{d+2\beta}}\right) + O\!\left((\tfrac{\mathrm{polylog}(n,\delta)}{n^2\varepsilon^2})^{\frac{\beta-1}{d+\beta}}\right)$. We also provide minimax lower bounds for private mode estimation, and show that our estimators are nearly optimal, up to a logarithmic factor in the MSE. We present two natural extensions: DP-PMS, a private modal-regression method, and DP-GRAMS-C, a clustering pipeline. Extensive experiments on synthetic and real data demonstrate favorable privacy-utility trade-offs relative to common baselines.

Figures

Figures reproduced from arXiv: 2607.29675 by the authors.

Figure 1
Figure 1. (a) Estimated KDE surface with true modes overlaid. (b) DAP initialization: candidate points colored by the local empirical-mass utility uj in (3), together with the privately selected anchors used as the starting points for DP-GRAMS. (c) Contour plot comparing true (green), mean shift (blue), and DP￾GRAMS (orange) mode estimates on a single dataset. (d) Privacy–utility tradeoff: MSE vs. ε on a log scale for n ∈ {70… view at source ↗
Figure 2
Figure 2. [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗
Figure 3
Figure 3. Private mode estimation on the five-component t-mixture. Panels (a) and (b) use one repre￾sentative dataset with n = 1200 and (ε, δ) = (1, 10−6 ). (a) Estimated KDE surface with true modes overlaid. (b) Contour plot comparing true (green), mean shift (blue), and DP-GRAMS (orange) mode estimates. (c) Privacy–utility tradeoff: MSE in (14) versus ε on a log scale for n ∈ {700, 1000, 2000, 5000} and ε ∈ {0.1, 0.25, 0.5,… view at source ↗
Figures from the paper (26 more)
Figure 4
Figure 4. Figure 4: Private modal regression on sinusoidal two-component mixture data. Panels (a) and (b) use one representative dataset with n = 500; panel (b) uses (ε, δ) = (1, 10−5 ). (a) PMS captures the two conditional modes, whereas LOWESS averages across the mixture components. (b)…
Figure 5
Figure 5. Figure 5: Private clustering on a four-component blobs dataset (n = 1000). Each panel shows cluster assignments, true centroids, and estimated centroids for mean shift, DP-GRAMS-C, k-means, and DP-k￾Means, with private methods run at ε = 1. 0.4 0.5 0.6 0.7 ARI n = 700 DP-GRAMS-C…
Figure 6
Figure 6. Figure 6: Privacy–utility tradeoff for private clustering on blobs across n ∈ {700, 1000, 2000, 5000} and ε ∈ {0.1, 0.2, 0.5, 1, 5}. Each panel shows a 2 × 2 grid over sample size, plotting ARI, NMI, or centroid MSE versus ε on a log scale for DP-GRAMS-C and DP-k-Means. Points s…
Figure 7
Figure 7. Figure 7: MNIST dataset with public auxiliary candidates. Two-dimensional visualization of the five￾dimensional PCA clustering representation, comparing mean shift, DP-GRAMS-C, k-means, and DP-k￾Means, with private methods run at ε = 1. True class centroids and estimated centroi…
Figure 8
Figure 8. Figure 8: Privacy–utility on MNIST with public auxiliary candidates in the whitened five-dimensional PCA representation: ARI, NMI, and centroid MSE versus ε on a log scale for DP-GRAMS-C and DP-k-Means, with ε ∈ {0.05, 0.1, 0.2, 0.5, 1}. Points show averages over 20 runs with st…
Figure 9
Figure 9. Figure 9: Cancer RNA-Seq dataset after gene-wise standardization and projection to six principal compo￾nents. Two-dimensional PCA visualization of the six-dimensional clustering representation, comparing mean shift, DP-GRAMS-C, k-means, and DP-k-Means, with private methods run a…
Figure 10
Figure 10. Figure 10: Privacy–utility on Cancer RNA-Seq after gene-wise standardization and projection to six princi￾pal components: ARI, NMI, and centroid MSE versus ε on a log scale for DP-GRAMS-C and DP-k-Means, with ε ∈ {0.5, 1, 2, 5, 10}. Points show averages over 20 runs with standar…
Figure 11
Figure 11. Figure 11: Bivariate 4-modal Gaussian mixture: grid of 20 [PITH_FULL_IMAGE:figures/full_fig_p079_11.png]
Figure 12
Figure 12. Figure 12: Hyperparameter sensitivity for DP-GRAMS on the 4-modal Gaussian mix￾ture. (a) Effect of clipping multiplier clip multiplier on MSE for n ∈ {700, 1000, 2000, 5000} at fixed ε = 1. (b) Effect of minibatch size m on MSE across the same sample sizes. The MSE is relatively…
Figure 13
Figure 13. Figure 13: Step-size sensitivity for DP-GRAMS on the 4-modal Gaussian mixture. The figure reports MSE versus step size η across n ∈ {700, 1000, 2000, 5000} at fixed ε = 1. The sweep does not show sharp degradation near the selected default. X2,i ∼ Uniform(0.4, 0.7), Y2,i ∼ N (2,…
Figure 14
Figure 14. Figure 14: Bivariate 5-modal t-mixture: grid of 20 DP-GRAMS runs on one fixed dataset. Each subplot shows KDE contours with true modes (green stars), non-private mean-shift estimates (blue crosses), and DP-GRAMS estimates (orange circles). The grid visualizes run-to-run variabil…
Figure 15
Figure 15. Figure 15: Hyperparameter sensitivity for DP-GRAMS on the 5-modal t-mixture. (a) Effect of clipping multiplier clip multiplier on MSE for n ∈ {700, 1000, 2000, 5000} at fixed ε = 1. (b) Effect of minibatch size m on MSE across the same sample sizes. The sweeps do not show sharp …
Figure 16
Figure 16. Figure 16: Step-size sensitivity for DP-GRAMS on the 5-modal t-mixture. The figure reports MSE versus step size η across n ∈ {700, 1000, 2000, 5000} at fixed ε = 1. The sweep does not show sharp degradation near the selected default, although the heavier-tailed setting exhibits …
Figure 17
Figure 17. Figure 17: Private modal regression on three-component piecewise-constant mixture data. Panels (a) and (b) use one representative dataset with n = 500; panel (b) uses (ε, δ) = (1, 10−5 ). (a) PMS tracks the modal branches, whereas LOWESS smooths across them. (b) DP-PMS recovers …
Figure 18
Figure 18. Figure 18: Hyperparameter sensitivity for DP-PMS on the three-component mixture. (a) Effect of the clipping multiplier clip multiplier on oracle MSE for n ∈ {200, 500, 1000, 2000} at fixed ε = 1. (b) Effect of minibatch size m on oracle MSE across the same sample sizes. The swee…
Figure 19
Figure 19. Figure 19: Hyperparameter sensitivity for DP-PMS on sinusoidal two-component mix￾ture data. (a) Effect of the clipping multiplier clip multiplier on oracle MSE for n ∈ {200, 600, 1200, 2400} at fixed ε = 1. (b) Effect of minibatch size m on oracle MSE across the same sample size…
Figure 20
Figure 20. Figure 20: Effect of clipping on DP-GRAMS-C for blobs. Subsampling-effect study for C∗ ∈ {0.01, 0.1, 0.5, 1.0, 2.0} at ε = 1 across n ∈ {700, 1000, 2000, 5000}. ARI, NMI, and centroid MSE vary smoothly with C∗, with a broad range of clipping multipliers (including the default C∗…
Figure 21
Figure 21. Figure 21: Effect of minibatch size on DP-GRAMS-C for blobs. Sample-size grid for m ∈ {0.01n, 0.05n, 0.1n, 0.2n, n} at ε = 1 and C∗ = 1 across n ∈ {700, 1000, 2000, 5000}. The metrics do not show sharp deterioration near the selected minibatch default. 89 [PITH_FULL_IMAGE:figur…
Figure 22
Figure 22. Figure 22: Digits dataset. Two-dimensional visualization of the six-dimensional PCA clustering representa￾tion, comparing non-private mean shift, DP-GRAMS-C, k-means, and DP-k-Means, with private methods run at ε = 1. True class centroids and estimated centroids are overlaid. bu…
Figure 23
Figure 23. Figure 23: Privacy–utility on Digits: ARI, NMI, and centroid MSE versus ε on a log scale for DP-GRAMS￾C and DP-k-Means, with ε ∈ {0.25, 0.5, 1, 2.5, 5}. Points show averages over 20 runs with standard-error bars [PITH_FULL_IMAGE:figures/full_fig_p091_23.png]
Figure 24
Figure 24. Figure 24: Digits, DP-GRAMS-C: centroid MSE, ARI, and NMI versus clipping multiplier clip multiplier at (ε, δ) = (1, 10−5 ). Points show averages over 20 runs with standard-error bars. MNIST public-candidate sensitivity. For MNIST, the main experiment uses a stratified pub￾lic a…
Figure 25
Figure 25. Figure 25: Digits, DP-GRAMS-C: centroid MSE, ARI, and NMI versus minibatch size m on a log scale at (ε, δ) = (1, 10−5 ). Points show averages over 20 runs with standard-error bars. perimental data into a whitened five-dimensional PCA representation, and runs clustering in that s…
Figure 26
Figure 26. Figure 26: MNIST, DP-GRAMS-C: ARI, NMI, and centroid MSE versus clipping multiplier clip multiplier at (ε, δ) = (1, 10−5 ). Points show averages over 20 runs with standard-error bars. 10 3 10 4 Minibatch size m 0.306 0.308 0.310 0.312 0.314 0.316 0.318 ARI DP-GRAMS-C on MNIST: A…
Figure 27
Figure 27. Figure 27: MNIST, DP-GRAMS-C: ARI, NMI, and centroid MSE versus minibatch size m on a log scale at (ε, δ) = (1, 10−5 ). Points show averages over 20 runs with standard-error bars. Together, the Digits and MNIST results provide two complementary real-image checks for the private …
Figure 28
Figure 28. Figure 28: Cancer RNA-Seq, DP-GRAMS-C: centroid MSE, ARI, and NMI versus clipping multiplier clip multiplier at (ε, δ) = (1, 10−5 ). Points show averages over 20 runs with standard-error bars. 10 2 Minibatch size m 1000 2000 3000 4000 5000 6000 7000 Centroid MSE DP-GRAMS-C on Ge…
Figure 29
Figure 29. Figure 29: Cancer RNA-Seq, DP-GRAMS-C: centroid MSE, ARI, and NMI versus minibatch size m on a log scale at (ε, δ) = (1, 10−5 ). Points show averages over 20 runs with standard-error bars. 94 [PITH_FULL_IMAGE:figures/full_fig_p094_29.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

103 extracted references · 10 linked inside Pith

  1. [1]

    9th Innovations in Theoretical Computer Science Conference (ITCS 2018) , volume=

    Finite Sample Differentially Private Confidence Intervals , author=. 9th Innovations in Theoretical Computer Science Conference (ITCS 2018) , volume=. 2018 , organization=

  2. [2]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Robustness Guarantees for Mode Estimation with an Application to Bandits , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  3. [3]

    arXiv preprint arXiv:2406.06755 , year=

    Optimal federated learning for nonparametric regression with heterogeneous distributed differential privacy constraints , author=. arXiv preprint arXiv:2406.06755 , year=

  4. [4]

    The Thirty Seventh Annual Conference on Learning Theory , pages=

    Optimal score estimation via empirical bayes smoothing , author=. The Thirty Seventh Annual Conference on Learning Theory , pages=. 2024 , organization=

  5. [5]

    arXiv preprint arXiv:2502.20123 , year=

    Stein's unbiased risk estimate and Hyv " arinen's score matching , author=. arXiv preprint arXiv:2502.20123 , year=

  6. [6]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  7. [7]

    Festschrift for Lucien Le Cam: research papers in probability and statistics , pages=

    Assouad, fano, and le cam , author=. Festschrift for Lucien Le Cam: research papers in probability and statistics , pages=. 1997 , publisher=

  8. [8]

    The Journal of Machine Learning Research , volume=

    On the estimation of the gradient lines of a density and the consistency of the mean-shift algorithm , author=. The Journal of Machine Learning Research , volume=. 2016 , publisher=

Show all 103 references
  1. [9]

    Journal of Machine Learning Research , volume=

    Locally private k-means clustering , author=. Journal of Machine Learning Research , volume=

  2. [10]

    International Conference on Machine Learning , pages=

    Differentially private clustering in high-dimensional euclidean spaces , author=. International Conference on Machine Learning , pages=. 2017 , organization=

  3. [11]

    Advances in Neural Information Processing Systems , volume=

    Differentially private clustering: Tight approximation ratios , author=. Advances in Neural Information Processing Systems , volume=

  4. [12]

    arXiv preprint arXiv:1405.7085 , year=

    Differentially private empirical risk minimization: Efficient algorithms and tight error bounds , author=. arXiv preprint arXiv:1405.7085 , year=

  5. [13]

    Introduction to Nonparametric Estimation , pages=

    Nonparametric estimators , author=. Introduction to Nonparametric Estimation , pages=. 2008 , publisher=

  6. [14]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Mean shift, mode seeking, and clustering , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 1995 , publisher=

  7. [15]

    IEEE Transactions on information theory , volume=

    The estimation of the gradient of a density function, with applications in pattern recognition , author=. IEEE Transactions on information theory , volume=. 1975 , publisher=

  8. [16]

    Foundations and Trends

    The algorithmic foundations of differential privacy , author=. Foundations and Trends. 2014 , publisher=

  9. [17]

    arXiv preprint arXiv:2210.00597 , year=

    Composition of differential privacy & privacy amplification by subsampling , author=. arXiv preprint arXiv:2210.00597 , year=

  10. [18]

    Journal of Privacy and Confidentiality , volume=

    Calibrating noise to sensitivity in private data analysis , author=. Journal of Privacy and Confidentiality , volume=

  11. [19]

    and Wasserman, Larry , title =

    Chen, Yen-Chi and Genovese, Christopher R. and Wasserman, Larry , title =. Electronic Journal of Statistics , year =

  12. [20]

    , author=

    Practicing differential privacy in health care: A review. , author=. Trans. Data Priv. , volume=

  13. [21]

    Harvard Data Science Review , volume=

    Differential privacy and social science: An urgent puzzle , author=. Harvard Data Science Review , volume=. 2020 , publisher=

  14. [22]

    Biostatistics & Epidemiology , volume=

    A tutorial on kernel density estimation and recent advances , author=. Biostatistics & Epidemiology , volume=. 2017 , publisher=

  15. [23]

    2018 , publisher=

    Density estimation for statistics and data analysis , author=. 2018 , publisher=

  16. [24]

    Statistical Analysis and Data Mining: The ASA Data Science Journal , volume=

    A fast and efficient Modal EM algorithm for Gaussian mixtures , author=. Statistical Analysis and Data Mining: The ASA Data Science Journal , volume=. 2021 , publisher=

  17. [25]

    2007 IEEE 11th international conference on computer vision , pages=

    Mode-seeking by medoidshifts , author=. 2007 IEEE 11th international conference on computer vision , pages=. 2007 , organization=

  18. [26]

    Joint IAPR International Workshops on Statistical Techniques in Pattern Recognition (SPR) and Structural and Syntactic Pattern Recognition (SSPR) , pages=

    Mode seeking clustering by KNN and mean shift evaluated , author=. Joint IAPR International Workshops on Statistical Techniques in Pattern Recognition (SPR) and Structural and Syntactic Pattern Recognition (SSPR) , pages=. 2012 , organization=

  19. [27]

    IEEE Transactions on pattern analysis and machine intelligence , volume=

    Mean shift: A robust approach toward feature space analysis , author=. IEEE Transactions on pattern analysis and machine intelligence , volume=. 2002 , publisher=

  20. [28]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Gaussian mean-shift is an EM algorithm , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2007 , publisher=

  21. [29]

    Journal of the royal statistical society: series B (methodological) , volume=

    Maximum likelihood from incomplete data via the EM algorithm , author=. Journal of the royal statistical society: series B (methodological) , volume=. 1977 , publisher=

  22. [30]

    Journal of Machine Learning Research , volume=

    A nonparametric statistical approach to clustering via mode identification , author=. Journal of Machine Learning Research , volume=

  23. [31]

    The fitting of multifunctions: an approach to nonparametric multimodal regression , author=. A. Rizzi, MV, editor, COMPSTAT 2006, Proceedings in Computational Statistics , pages=

  24. [32]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Ensemble tracking , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2007 , publisher=

  25. [33]

    IEEE Transactions on pattern analysis and machine intelligence , volume=

    Kernel-based object tracking , author=. IEEE Transactions on pattern analysis and machine intelligence , volume=. 2003 , publisher=

  26. [34]

    European conference on computer vision , pages=

    Image and video segmentation by anisotropic kernel mean shift , author=. European conference on computer vision , pages=. 2004 , organization=

  27. [35]

    2009 16th IEEE International Conference on Image Processing (ICIP) , pages=

    Bayesian image segmentation with mean shift , author=. 2009 16th IEEE International Conference on Image Processing (ICIP) , pages=. 2009 , organization=

  28. [36]

    2009 International Conference on Computational Science and Engineering , volume=

    Edge detection based on fast adaptive mean shift algorithm , author=. 2009 International Conference on Computational Science and Engineering , volume=. 2009 , organization=

  29. [37]

    2005 International Conference on Neural Networks and Brain , volume=

    Mean shift-based edge detection for color image , author=. 2005 International Conference on Neural Networks and Brain , volume=. 2005 , organization=

  30. [38]

    2007 IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Object tracking by asymmetric kernel mean shift with automatic scale and orientation selection , author=. 2007 IEEE Conference on Computer Vision and Pattern Recognition , pages=. 2007 , organization=

  31. [39]

    IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) , volume=

    Robust fusion of uncertain information , author=. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) , volume=. 2005 , publisher=

  32. [40]

    2012 26th Biennial Symposium on Communications (QBSC) , pages=

    On noisy source vector quantization via a subspace constrained mean shift algorithm , author=. 2012 26th Biennial Symposium on Communications (QBSC) , pages=. 2012 , organization=

  33. [41]

    Genovese and Ryan J

    Yen-Chi Chen and Christopher R. Genovese and Ryan J. Tibshirani and Larry Wasserman , title =. The Annals of Statistics , number =. 2016 , doi =

  34. [42]

    Cryptology ePrint Archive , year=

    Approximate and probabilistic differential privacy definitions , author=. Cryptology ePrint Archive , year=

  35. [43]

    Annual international conference on the theory and applications of cryptographic techniques , pages=

    Our data, ourselves: Privacy via distributed noise generation , author=. Annual international conference on the theory and applications of cryptographic techniques , pages=. 2006 , organization=

  36. [44]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Convergence analysis of mean shift , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2024 , publisher=

  37. [45]

    Advances in neural information processing systems , volume=

    Privacy amplification by subsampling: Tight analyses via couplings and divergences , author=. Advances in neural information processing systems , volume=

  38. [46]

    Proceedings of the 2016 ACM SIGSAC conference on computer and communications security , pages=

    Deep learning with differential privacy , author=. Proceedings of the 2016 ACM SIGSAC conference on computer and communications security , pages=

  39. [47]

    European conference on computer vision , pages=

    Quick shift and kernel methods for mode seeking , author=. European conference on computer vision , pages=. 2008 , organization=

  40. [48]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Gridshift: A faster mode-seeking algorithm for image segmentation and object tracking , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  41. [49]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    MeanShift++: Extremely fast mode-seeking with applications to segmentation and object tracking , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  42. [50]

    Proceedings of the sixth ACM conference on data and application security and privacy , pages=

    Differentially private k-means clustering , author=. Proceedings of the sixth ACM conference on data and application security and privacy , pages=

  43. [51]

    2010 IEEE 51st annual symposium on foundations of computer science , pages=

    Boosting and differential privacy , author=. 2010 IEEE 51st annual symposium on foundations of computer science , pages=. 2010 , organization=

  44. [52]

    Advances in neural information processing systems , volume=

    Differentially private generalized linear models revisited , author=. Advances in neural information processing systems , volume=

  45. [53]

    arXiv preprint arXiv:2007.05157 , year=

    Differentially private simple linear regression , author=. arXiv preprint arXiv:2007.05157 , year=

  46. [54]

    arXiv preprint arXiv:1803.02596 , year=

    Revisiting differentially private linear regression: optimal and adaptive prediction & estimation in unbounded domain , author=. arXiv preprint arXiv:1803.02596 , year=

  47. [55]

    International Conference on Machine Learning , pages=

    Differentially private ordinary least squares , author=. International Conference on Machine Learning , pages=. 2017 , organization=

  48. [56]

    International Conference on Machine Learning , pages=

    Fast private kernel density estimation via locality sensitive quantization , author=. International Conference on Machine Learning , pages=. 2023 , organization=

  49. [57]

    arXiv preprint arXiv:2409.01688 , year=

    Differentially private kernel density estimation , author=. arXiv preprint arXiv:2409.01688 , year=

  50. [58]

    Journal of the American Statistical Association , volume=

    A statistical framework for differential privacy , author=. Journal of the American Statistical Association , volume=. 2010 , publisher=

  51. [59]

    The Journal of Machine Learning Research , volume=

    Differential privacy for functions and functional data , author=. The Journal of Machine Learning Research , volume=. 2013 , publisher=

  52. [60]

    Genovese and Marco Perone-Pacifico and Isabella Verdinelli and Larry Wasserman , title =

    Christopher R. Genovese and Marco Perone-Pacifico and Isabella Verdinelli and Larry Wasserman , title =. The Annals of Statistics , number =. 2014 , doi =

  53. [61]

    Journal of the American statistical association , volume=

    Randomized response: A survey technique for eliminating evasive answer bias , author=. Journal of the American statistical association , volume=. 1965 , publisher=

  54. [62]

    arXiv preprint arXiv:1907.02444 , year=

    Diffprivlib: the IBM differential privacy library , author=. arXiv preprint arXiv:1907.02444 , year=

  55. [63]

    1998 , publisher=

    Applied regression analysis , author=. 1998 , publisher=

  56. [64]

    2005 , publisher=

    Quantile regression , author=. 2005 , publisher=

  57. [65]

    International colloquium on automata, languages, and programming , pages=

    Differential privacy , author=. International colloquium on automata, languages, and programming , pages=. 2006 , organization=

  58. [66]

    Biometrical Journal , volume=

    Parametric mode regression for bounded responses , author=. Biometrical Journal , volume=. 2020 , publisher=

  59. [67]

    Journal of Biopharmaceutical Statistics , volume=

    A collection of parametric modal regression models for bounded data , author=. Journal of Biopharmaceutical Statistics , volume=. 2021 , publisher=

  60. [68]

    Biometrical Journal , volume=

    Parametric modal regression with varying precision , author=. Biometrical Journal , volume=. 2020 , publisher=

  61. [69]

    Journal of Econometrics , volume=

    Mode regression , author=. Journal of Econometrics , volume=. 1989 , publisher=

  62. [70]

    IEEE computational intelligence magazine , volume=

    A survey on differentially private machine learning , author=. IEEE computational intelligence magazine , volume=. 2020 , publisher=

  63. [71]

    Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security , pages=

    Privacy-preserving density-based clustering , author=. Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security , pages=

  64. [72]

    Proceedings of the ACM on Management of Data , volume=

    Approximate DBSCAN under Differential Privacy , author=. Proceedings of the ACM on Management of Data , volume=. 2025 , publisher=

  65. [73]

    and Varoquaux, G

    Pedregosa, F. and Varoquaux, G. and Gramfort, A. and Michel, V. and Thirion, B. and Grisel, O. and Blondel, M. and Prettenhofer, P. and Weiss, R. and Dubourg, V. and Vanderplas, J. and Passos, A. and Cournapeau, D. and Brucher, M. and Perrot, M. and Duchesnay, E. , journal=. S...

  66. [74]

    Mironov, Ilya , booktitle=. R. 2017 , organization=

  67. [75]

    The Annals of Statistics , volume=

    The cost of privacy: Optimal rates of convergence for parameter estimation with differential privacy , author=. The Annals of Statistics , volume=. 2021 , publisher=

  68. [76]

    Theory of cryptography conference , pages=

    Concentrated differential privacy: Simplifications, extensions, and lower bounds , author=. Theory of cryptography conference , pages=. 2016 , organization=

  69. [77]

    Problemy Peredachi Informatsii , volume=

    Recursive estimation of the mode of a multivariate distribution , author=. Problemy Peredachi Informatsii , volume=. 1990 , publisher=

  70. [78]

    48th Annual IEEE Symposium on Foundations of Computer Science (FOCS'07) , pages=

    Mechanism design via differential privacy , author=. 48th Annual IEEE Symposium on Foundations of Computer Science (FOCS'07) , pages=. 2007 , organization=

  71. [79]

    , author=

    Estimation of non-normalized statistical models by score matching. , author=. Journal of Machine Learning Research , volume=

  72. [80]

    Annales de l'Institut Henri Poincare (B) Probability and Statistics , volume=

    Rates of strong uniform consistency for multivariate kernel density estimators , author=. Annales de l'Institut Henri Poincare (B) Probability and Statistics , volume=. 2002 , organization=

  73. [81]

    Lecture notes , year=

    Lecture notes on nonparametrics , author=. Lecture notes , year=

  74. [82]

    Electronic Journal of Statistics , volume=

    Quantile regression approach to conditional mode estimation , author=. Electronic Journal of Statistics , volume=

  75. [83]

    Naval research logistics quarterly , volume=

    The Hungarian method for the assignment problem , author=. Naval research logistics quarterly , volume=. 1955 , publisher=

  76. [84]

    Journal of the society for industrial and applied mathematics , volume=

    Algorithms for the assignment and transportation problems , author=. Journal of the society for industrial and applied mathematics , volume=. 1957 , publisher=

  77. [85]

    arXiv preprint arXiv:1907.06233 , year=

    Pointwise adaptive kernel density estimation under local approximate differential privacy , author=. arXiv preprint arXiv:1907.06233 , year=

  78. [86]

    Theory of Probability & Its Applications , volume=

    On a problem of adaptive estimation in Gaussian white noise , author=. Theory of Probability & Its Applications , volume=. 1991 , publisher=

  79. [87]

    Journal of Machine Learning Research , volume=

    Adaptive approximation and generalization of deep neural network with intrinsic dimensionality , author=. Journal of Machine Learning Research , volume=

  80. [88]

    Advances in neural information processing systems , volume=

    Generative modeling by estimating gradients of the data distribution , author=. Advances in neural information processing systems , volume=

  81. [89]

    International Conference on Machine Learning , pages=

    Diffusion models are minimax optimal distribution estimators , author=. International Conference on Machine Learning , pages=. 2023 , organization=

  82. [90]

    The Annals of Statistics , pages=

    Optimal spatial adaptation to inhomogeneous smoothness: an approach based on kernel estimates with variable bandwidth selectors , author=. The Annals of Statistics , pages=. 1997 , publisher=

  83. [91]

    arXiv preprint arXiv:2206.07663 , year=

    Adaptive pointwise density estimation under local differential privacy , author=. arXiv preprint arXiv:2206.07663 , year=

  84. [92]

    Bernoulli , volume=

    Local differential privacy: Elbow effect in optimal density estimation and adaptation over Besov ellipsoids , author=. Bernoulli , volume=. 2020 , publisher=

  85. [93]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , pages=

    Minimax and adaptive transfer learning for nonparametric classification under distributed differential privacy constraints , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , pages=. 2025 , publisher=

  86. [94]

    science , volume=

    Clustering by fast search and find of density peaks , author=. science , volume=. 2014 , publisher=

  87. [95]

    arXiv preprint arXiv:1611.06777 , year=

    Effective Deterministic Initialization for k -Means-Like Methods via Local Density Peaks Searching , author=. arXiv preprint arXiv:1611.06777 , year=

  88. [96]

    Advances in Neural Information Processing Systems , volume=

    K-median clustering via metric embedding: towards better initialization with differential privacy , author=. Advances in Neural Information Processing Systems , volume=

  89. [97]

    2016 , howpublished =

    Fiorini, Samuele , title =. 2016 , howpublished =

  90. [98]

    ACM Computing Surveys (CSUR) , volume=

    Membership inference attacks on machine learning: A survey , author=. ACM Computing Surveys (CSUR) , volume=. 2022 , publisher=

  91. [99]

    Statistical Science , volume=

    A Population Background for Nonparametric Density-Based Clustering , author=. Statistical Science , volume=

  92. [100]

    Proceedings of the twenty-second ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems , pages=

    Revealing information while preserving privacy , author=. Proceedings of the twenty-second ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems , pages=

  93. [101]

    Proceedings of the National Academy of Sciences , volume=

    Confidence-ranked reconstruction of census microdata from published statistics , author=. Proceedings of the National Academy of Sciences , volume=. 2023 , publisher=

  94. [102]

    2006 , publisher=

    Gaussian processes for machine learning , author=. 2006 , publisher=

  95. [103]

    Learning theory from first principles Lecture 6: Kernel methods , author=

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.