acceptodds
Under review as a conference paper at ICLR 2027

Private Conditional Quantile Functions through Shared Location Learning

Abstract

Conditional quantile functions support prediction at uncertainty levels chosen after training. A direct approach under differential privacy fits separate models at several levels. Training these models on the same records divides a fixed privacy budget among them, increasing the noise required for each fit. In this study, we propose DP-ShareQ, which combines a shared input-dependent location with a learned shift for each quantile level. We develop two learners for this representation. Joint learns the location and these offsets together, using quantile ordering to calibrate gradient noise. CDF first learns a private location predictor, then privately estimates its residual distribution and uses its quantiles as offsets. For both learners, we derive prediction bounds that quantify the approximation cost of shared quantile spacing and the error of private learning. Across six datasets with public linear and additive features, Joint and CDF reduce average normalized integrated pinball loss relative to output-perturbed separate fits by 25.9–30.3% and 31.0–31.5%, respectively, across privacy budgets \(\varepsilon\in{1,2,4}\).

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Differential privacy, conditional quantile regression

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

This paper studies conditional quantile learning under differential privacy by representing the quantile function as a shared input-dependent location plus a level-specific offset, q(x,τ)=ϕ(x)⊤β+d(τ)q(x,\tau)=\phi(x)^\top\beta+d(\tau), so that multiple quantile levels can be learned without repeatedly splitting the privacy budget. The authors propose two methods: Joint, which learns the shared location and offsets jointly while exploiting quantile ordering to reduce privacy noise, and CDF, which first privately learns the location and then privately estimates the residual distribution to obtain quantile offsets. They also derive prediction-error bounds that separate model-sharing approximation error from private learning error, and show empirical gains over separate private quantile fits.

Critical strengths

S1. The paper addresses a clear and privacy-utility trade-off in conditional quantile learning.

S2. The shared-location formulation is simple and interpretable.

S3. The empirical study is extensive and includes several useful ablations.

Critical weaknesses

W1. The main structural assumption is very restrictive for a paper framed as learning conditional quantile functions.
DP-ShareQ assumes that all conditional quantiles share the same input effect: q(x,τ)=ϕ(x)⊤β+d(τ).q(x,\tau)=\phi(x)^\top\beta+d(\tau). Equivalently, quantile gaps do not depend on xx. This rules out heteroskedasticity and more general changes in conditional shape, which are precisely situations where conditional quantiles are most useful. When this assumption fails, the learned curve is not the true conditional quantile function.

W2. The methodological novelty is somewhat incremental at the component level.
The common-slope idea is closely related to composite quantile regression, while the two proposed learners combine established tools such as private noisy-gradient optimization, Huber regression, private distribution estimation, Gaussian covariance design, and isotonic projection. The integration of these ideas and the ordered-gradient analysis are useful, but the paper should more clearly articulate which methodological ingredient is fundamentally new beyond combining existing components for this particular shared quantile model.

W3. The theoretical guarantees rely on strong distributional assumptions, especially for CDF.
The optimal shared-curve result assumes conditionally symmetric residuals, while the Huber-to-quantile transfer additionally requires conditional residual densities to be unimodal/nonincreasing away from zero and imposes a bound on the pooled density. These assumptions are difficult to verify in applications and exclude many common asymmetric or multimodal conditional distributions.

W4. The real-data uncertainty assessment is based on one fixed train/validation/test split.
For each real dataset, the paper fixes a single partition and repeats only the private-learning noise 100 times. The reported standard deviations therefore quantify privacy randomness but not sampling or split variability. This is particularly concerning for small datasets such as Real Estate. Repeated data splits or cross-validation would provide a much more reliable assessment of the empirical gains.

Questions for authors

See weaknesses.

6 comments

Thank you for your careful review and comments on our model and methodological contribution. We believe that addressing all reviewers' comments has substantially improved the paper. We have uploaded a revised manuscript incorporating our responses to all comments. Equation, section, table, and figure numbers below refer to this revision. Changes are marked in blue, and updated numbers in green.

C1/W1. Input-dependent quantile gaps. Sharing introduces approximation error when quantile spacing varies with the input. However, under privacy constraints, this cost is traded against the gain from reduced DP noise. As Section 4 explains, sharing reduces the number of privately learned coefficients and the resulting DP noise. We assess this trade-off by comparing with Full and an input-dependent scale extension under the same privacy budget.

DP-ShareQ can also capture input-dependent uncertainty by adding scale coefficients. Appendix G.2 evaluates LS-CDF, q(x,τ)=l(x)+s(x)d(τ)q(x,\tau)=l(x)+s(x)d(\tau), under the same total privacy budget. In Figure 8(d), at γ=0.75\gamma=0.75, nIPL ×100 is 3.685 for LS-CDF, 3.882 for CDF, and 3.800 for Full. The extension improves the shared model under input-dependent scale and outperforms Full. On the real datasets, however, CDF has 2.0%, 6.2%, and 1.9% lower mean nIPL than LS-CDF at ε=0.1,0.5,4 (Table 18). This comparison supports retaining CDF as the main model at the evaluated budgets.

C2/W2. Methodological contribution. DP-ShareQ is a new differentially private quantile regression algorithm. We have clarified the contributions of Joint and CDF in the Introduction.

The first challenge is to learn many quantiles under one privacy budget. Separate fits divide the budget across levels. Fitting all levels together still requires many noisy coefficients when the model allows different input effects at each level. Joint and CDF both learn q(x,τ)=ϕ(x)⊤β+d(τ)q(x,\tau)=\phi(x)^\top\beta+d(\tau), sharing one input effect across levels. Joint fits the location and offsets together. CDF fits the location and obtains all offsets from one private residual distribution.

The second challenge is to reduce privacy noise beyond coefficient sharing, which alone does not ensure better prediction (Appendix F.4). We derive tighter sensitivity bounds by ruling out changes that cannot occur together when one record is replaced.

In Joint, a record's residual sign changes at most once across ordered quantile levels. We use the resulting gradient set (Equation (3)) to calibrate noise and allocate its variance across location and offset directions. Table 2 shows the prediction gains from both the ordered bound and noise allocation.

In CDF, all Fourier coefficients are functions of the same residual. Bounding their changes together requires less noise than bounding each coefficient separately. For Fourier orders 16, 32, and 64, noise standard deviations fall by 22–24% (Section 3.2). Table 2 reports the corresponding reduction in prediction loss.

C3/W3. Distributional assumptions. Revised Section 4 removes conditional symmetry, unimodality, and density assumptions while retaining the original big-O learning-error rates for bounded i.i.d. records and fixed learning settings. Proposition 3 is reformulated without conditional symmetry, and Theorem 4 retains its original bound. The substantive change concerns the Huber-to-quantile risk comparison in the original Theorem 5. We now bound this discrepancy without the original distributional shape assumptions.

Proposition 3. For every choice of location coefficients, pooled residual quantiles give the optimal shared offsets. Minimizing this population risk over the location defines the optimal shared curve qβQq_{\beta_Q} used in both learning bounds. The original result identified this curve through the symmetry-defined location β∗\beta_*.

Theorem 4. The bound is unchanged. Conditional symmetry was used only to identify the optimal shared reference, which we now define directly by population risk minimization.

Proposition 5, replacing the original Theorem 5. Equation (14) links Huber loss to quantile risk without the original shape assumptions. Huber loss equals quantile risk with fixed offsets up to a constant, and fitting residual quantiles reduces this risk. Appendix A.6 then bounds the gap from the optimal shared curve.

Theorem 6. The original EC\mathcal E_C contained ridge cost λB02/(2w)\lambda B_0^2/(2w), with B0≥∣β∗∣2B_0\geq\\|\beta_*\\|_2, and finite-CDF approximation. Equation (44) now gives

EC=Γh,B(βQ)+λ2∣βQ∣22+2BCwM,\mathcal E_C=\Gamma_{h,B}(\beta_Q)+\frac{\lambda}{2}\\|\beta_Q\\|_2^2+\frac{2B_C}{wM},

where Γh,B(β)\Gamma_{h,B}(\beta) is the nonnegative gap between the two quantile risks in Equation (14). Equation (42) retains the difference Γh,B(βQ)−EΓh,B(β^)\Gamma_{h,B}(\beta_Q)-\mathbb E\Gamma_{h,B}(\widehat\beta) between the gains from fitting residual quantiles at the optimal shared location and at the learned location. Appendix A.10 gives the full proof.

C4/W4. Split variability. We have rerun the real-data comparisons throughout the main text and appendix on five data splits, with ten noise repetitions per split. The updated results report means and standard deviations across the five splits. At ε=1, both methods have lower mean nIPL than all three reference methods in all 12 dataset–map settings (Table 1). On Real Estate with linear features, nIPL ×100 is 5.159 ± 0.381 for Joint and 5.307 ± 0.603 for CDF, compared with 7.562 ± 0.330 for Full and 7.181 ± 0.846 for Sep-Out.

We welcome any further comments or questions. If our responses resolve your concerns, we would be grateful if you would consider reassessing your score.

Thank you. I find that the authors have substantially addressed most of my previous concerns, and the revisions largely improve the original manuscript.

I have two more questions. I appreciate the additional experiments in Appendix B.6, but after accounting for the privacy cost of tuning parameter selection, the average nIPL have largely increased. Does this evidence indicate that this cost is non-negligible? Besides, Appendix B.6 uses 60% of the data for training and does not retrain the selected model, whereas the main experiments retrain the models on 75% of the data after tuning parameter selection. Please consider to clarify this distinction and the reason.

Reviewer_26UvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a shared-location framework for conditional quantile estimation under differential privacy. It represents conditional quantiles using common regression coefficients in a fixed feature space, together with quantile-specific offsets. It develops two learning approaches: Joint optimizes a smoothed quantile loss using noisy proximal-gradient updates, while CDF learns a location predictor and then estimates its residual distribution privately. The paper provides theoretical prediction-error bounds that explicitly include privacy terms, and evaluates the methods through simulations and real-data experiments.

Critical strengths
  1. The theoretical bounds explicitly separate the approximation error introduced by sharing from the errors of private learning. This provides useful insight into the trade-off between model restrictions and the cost of privacy.
  2. The empirical evaluation is reasonably extensive. It includes multiple datasets, comparisons with alternative methods, component ablations, and simulations examining when the shared structure becomes less effective.
Critical weaknesses
  1. The shared-location structure is restrictive: quantile gaps do not vary with the input. The theoretical analysis also requires conditional symmetry, with additional unimodality and density conditions for the CDF risk-transfer result. I remain uncertain how well these assumptions describe the real datasets. Good average predictive performance alone does not establish their suitability.
  2. The choice of privacy budgets ε∈{1,2,4}\varepsilon\in\{1,2,4\} needs clearer justification. The abstract emphasizes improvements at these budgets, but I find their practical significance difficult to interpret without a clearer explanation of the selected privacy regime and performance under stronger privacy constraints.
Questions for authors
  1. What evidence supports the suitability of the shared-location structure and the residual assumptions on the real datasets? Can the theoretical bounds provide a more explicit criterion for when the reduction in privacy noise outweighs the approximation error from sharing?
  2. Could the main text explain the motivation for Huber location learning and quadratic regularization more directly? I also found “transfer” confusing: the connection between controlling Huber excess risk and controlling quantile prediction risk needs a more intuitive explanation.
  3. The existing ablations and failure-case simulations are helpful. How sensitive are the results to the regularization strength, smoothing bandwidth, learning rate, and iteration count? I would particularly like to see optimization convergence curves and a clearer account of when performance deteriorates.

2:27

6 comments

Thank you for your careful review and helpful questions. We believe that addressing all reviewers' comments has substantially improved the paper. We have uploaded a revised manuscript incorporating our responses to all comments. Equation, section, table, and figure numbers below refer to this revision. Changes are marked in blue, and updated numbers in green.

B1. Shared location structure and residual assumptions. Sharing introduces approximation error when quantile spacing varies with the input. However, under privacy constraints, this cost is traded against the gain from reduced DP noise. As Section 4 explains, sharing reduces the number of privately learned coefficients and the resulting DP noise. We assess this trade-off by comparing with Full and an input-dependent scale extension under the same privacy budget.

DP-ShareQ can also capture input-dependent uncertainty by adding scale coefficients. Appendix G.2 evaluates LS-CDF, q(x,τ)=l(x)+s(x)d(τ)q(x,\tau)=l(x)+s(x)d(\tau), under the same total privacy budget. In Figure 8(d), at γ=0.75\gamma=0.75, nIPL ×100 is 3.685 for LS-CDF, 3.882 for CDF, and 3.800 for Full. The extension improves the shared model under input-dependent scale and outperforms Full. On the real datasets, however, CDF has 2.0%, 6.2%, and 1.9% lower mean nIPL than LS-CDF at ε=0.1,0.5,4 (Table 18). This comparison supports retaining CDF as the main model at the evaluated budgets.

Residual assumptions. Revised Section 4 removes conditional symmetry, unimodality, and density assumptions while retaining the original big-O learning-error rates for bounded i.i.d. records and fixed learning settings. Proposition 3 is reformulated without conditional symmetry, and Theorem 4 retains its original bound. The substantive change concerns the Huber-to-quantile risk comparison in the original Theorem 5. We now bound this discrepancy without the original distributional shape assumptions.

Proposition 3. For every choice of location coefficients, pooled residual quantiles give the optimal shared offsets. Minimizing this population risk over the location defines the optimal shared curve qβQq_{\beta_Q} used in both learning bounds. The original result identified this curve through the symmetry-defined location β∗\beta_*.

Theorem 4. The bound is unchanged. Conditional symmetry was used only to identify the optimal shared reference, which we now define directly by population risk minimization.

Proposition 5, replacing the original Theorem 5. Equation (14) links Huber loss to quantile risk without the original shape assumptions. Huber loss equals quantile risk with fixed offsets up to a constant, and fitting residual quantiles reduces this risk. Appendix A.6 then bounds the gap from the optimal shared curve.

Theorem 6. The original EC\mathcal E_C contained ridge cost λB02/(2w)\lambda B_0^2/(2w), with B0≥∣β∗∣2B_0\geq\\|\beta_*\\|_2, and finite-CDF approximation. Equation (44) now gives

EC=Γh,B(βQ)+λ2∣βQ∣22+2BCwM,\mathcal E_C=\Gamma_{h,B}(\beta_Q)+\frac{\lambda}{2}\\|\beta_Q\\|_2^2+\frac{2B_C}{wM},

where Γh,B(β)\Gamma_{h,B}(\beta) is the nonnegative gap between the two quantile risks in Equation (14). Equation (42) retains the difference Γh,B(βQ)−EΓh,B(β^)\Gamma_{h,B}(\beta_Q)-\mathbb E\Gamma_{h,B}(\widehat\beta) between the gains from fitting residual quantiles at the optimal shared location and at the learned location. Appendix A.10 gives the full proof.

B2. Privacy regime. The original ε=1,2,4 grid was an initial evaluation range. We agree that stronger privacy should also be evaluated and extended the comparison to ε=0.1,0.25,0.5. At ε=0.1 and 0.5, CDF has lower mean nIPL than all three reference methods in all 24 dataset–map settings, and Joint in 23 (Tables 7–8).

B3. Explicit sharing criterion. Appendix A.11 derives the comparison from the detailed prediction bounds. For the averaged Joint (j=Jj=J) and Full (j=Fj=F) predictors at the training quantile levels, write Uj=Aj+bj+PjU_j=A_j+b_j+P_j, where AjA_j is approximation error, bjb_j collects the other learning costs, and

Pj=4ηjTjtr⁡(Ωj)(2−ηjLj)2n2μ(ε,δ)2,j∈J,F.P_j=\frac{4\eta_jT_j\operatorname{tr}(\Omega_j)}{(2-\eta_jL_j)^2n^2\mu(\varepsilon,\delta)^2},\qquad j\in\\{J,F\\}.

Joint's bound is smaller precisely when

(AJ−AF)+(bJ−bF)<PF−PJ.(A_J-A_F)+(b_J-b_F)<P_F-P_J.

Appendix A.11 also derives the corresponding criterion for CDF from Theorem 6.

B4 to B5. Huber loss and quadratic regularization. Huber loss combines a bounded score with a Lipschitz gradient. These properties make it a common choice in DP regression [1,2]. In our scaling, replacing one record changes the average gradient by at most 1/n1/n. We also considered other objectives. The smoothed composite-pinball loss used by Joint introduces auxiliary offset parameters that the residual stage would estimate again. The pairwise rank alternative couples each record to multiple residual pairs, increasing the gradient-sensitivity bound. Huber retains only the pp location coefficients with a bounded-gradient calibration.

Quadratic regularization supplies strong convexity, which limits how much the fitted coefficients change when one record is replaced. This stability helps control the prediction error caused by privacy noise and is a standard tool in private ERM analysis [3]. We added these motivations for Huber loss and quadratic regularization to Section 3.2.

References

[1] Marco Avella-Medina, Casey Bradshaw, and Po-Ling Loh. Differentially private inference via noisy optimization. Annals of Statistics, 2023.

[2] Jinyuan Chang et al. Adapting to Noise Tails in Private Linear Regression. JASA, 2026.

[3] Kamalika Chaudhuri, Claire Monteleoni, and Anand D. Sarwate. Differentially Private Empirical Risk Minimization. JMLR, 2011.

B6. Connecting Huber loss and quantile risk. We used “transfer” to describe the link between these two objectives. CDF minimizes Huber loss for location learning, whereas Theorem 6 bounds quantile prediction risk. Equation (14) expresses Huber loss as quantile risk with fixed offsets, up to a constant. Appendix A.6 uses this identity to bound the quantile-risk gap between the Huber location and the optimal shared curve after fitting residual quantiles. In Appendix A.10, the same identity combines with the private Huber learning bound and residual estimation error to prove Theorem 6, Equation (15). We renamed the subsection “CDF: location learning and residual estimation.”

B7. Sensitivity to the learning settings. Figure 5 (Appendix B.5) varies one parameter at a time at ε=1. Regularization has the largest effect: increasing cλc_\lambda from 1 to 10 raises mean nIPL ×100 from 6.01 to 9.29 for Joint and from 5.69 to 7.52 for CDF. Increasing cTc_T from 1 to 4 lowers Joint's loss from 7.25 to 5.73, while increasing it from 4 to 16 raises CDF's loss from 5.06 to 5.62. Smoothing and learning-rate changes have smaller effects over the tested ranges.

The penalty multiplier cλc_\lambda and horizon parameter cTc_T are defined under “Validation selection” in Appendix B.2. They determine the candidate grids for regularization strength λ\lambda and iteration count TT according to each fit's DP noise level.

B8. Convergence in the scale simulation. Convergence failure does not appear to be the main cause of the prediction gap from Full. Figure 7 in Appendix G.1 shows Joint and CDF for the γ=0.75\gamma=0.75 case of Figure 4(d), using their selected horizons. DP noise leaves larger objective gaps than in the noise-free runs, but both methods reduce their objectives without divergence. At n=2048n=2048 and ε=4, final nIPL ×100 is 3.882 for both methods, compared with 3.800 for Full. Removing privacy noise gives 3.877 and 3.880. The gap persists without privacy noise, consistent with the approximation–privacy trade-off under input-dependent scale in Section 4.

We welcome any further comments or questions. If our responses resolve your concerns, we would be grateful if you would consider reassessing your score.

I thank the authors for the detailed clarification and additional experiments. These address several of my concerns about the methodological choices and empirical evaluation.

However, my concern about the depth and significance of the theoretical contribution remains. The sharing criterion compares error upper bounds, while the revised Huber-to-quantile analysis retains an additional gap term whose magnitude and implications remain unclear to me. Overall, the revisions improve the paper, but do not sufficiently strengthen its central insights to change my assessment.

I therefore maintain my original recommendation of weak rejection.

Reviewer_A5f8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Previous approaches to differentially private conditional quantile function fit separate models at different levels, thus training of same records divides the privacy budget amongst the molds, increasing the noise required for each fit. This paper proposes a framework called DP-ShareQ to avoid this. The crux is to learn a shared input-dependent location function and quantile-dependent offsets, reducing the number of parameters. Furthermore to solve this, the paper introduces two algorithms DP-ShareQ-Joint and DP-ShareQ-CDF. To optimize location and offsets using monotonicity-aware gradient sensitivity and directional Gaussian noise DP-ShareQ-Joint is used. Whereas DP-ShareQ-CDF, learns a private location predictor through Huber regression and estimates a private residual distribution using Fourier coefficients. Furthermore experimental results are provided for six regression models to validate the theoretical guarantees.

Critical strengths

Under a fixed privacy budget, learning quantiles at fixed uncertainty levels is quite an important problem. It’s quite an interesting idea to share model parameters to reduce privacy expenditure. The gradient sensitivity analysis and directional noise contributions look quite novel and meaningful contributions and they are backed by the ablation experiments in section 5.2 which provides evidence that they substantially improve prediction accuracies compared to conventional mechanisms. Finally the experiments are quite extensive as the paper compares against multiple private learning approaches across six datasets and three privacy budgets. The additional experiments also provide useful insights about the methods' limitations and robustness.

Critical weaknesses

The proposed methodology about the methods' limitations and robustness, preventing it from capturing any input-dependent changes. The results from section 5.3 show a more flexible baseline that outperforms when sample size and privacy budget increases and reason might be the method's failure to model heteroscedasticity. I am concerned about the practicality of DP-ShareQ when conditional uncertainty varies substantially across inputs. In appendix E.4, the paper studies sharing under output perturbation, the improvement from sharing alone is relatively modest compared with the headline gains. Also the full baseline uses isotropic gradient noise, making it difficult to isolate the benefits of parameter sharing. Hyperparameter selections are also excluded from the main results. Appendix B.6 addresses this private selection but also reports significant increase in prediction loss. Finally the experimental setup only considers fixed linear and additive feature representations with a single train-test split. I would like to see how the competitiveness holds across variable data splits and non-linear representations.

Questions for authors

See weaknesses.

2 comments

Thank you for your careful review and questions about model flexibility and evaluation. We believe that addressing all reviewers' comments has substantially improved the paper. We have uploaded a revised manuscript incorporating our responses to all comments. Equation, section, table, and figure numbers below refer to this revision. Changes are marked in blue, and updated numbers in green.

A1. Input-dependent uncertainty. We agree that Full can be better with larger samples and privacy budgets. In Figure 4, the reversal occurs at n=2048n=2048, ε=4, and γ=0.75\gamma=0.75. Joint and CDF have lower loss than Full at every tested scale variation for n=512n=512 at both budgets and for n=2048n=2048 at ε=1. As Section 4 explains, sharing reduces the number of privately learned coefficients and the resulting DP noise.

DP-ShareQ can also capture input-dependent uncertainty by adding scale coefficients. Appendix G.2 evaluates LS-CDF, q(x,τ)=l(x)+s(x)d(τ)q(x,\tau)=l(x)+s(x)d(\tau), under the same total privacy budget. In Figure 8(d), at γ=0.75\gamma=0.75, nIPL ×100 is 3.685 for LS-CDF, 3.882 for CDF, and 3.800 for Full. The extension improves the shared model in the region raised by the reviewer and outperforms Full. We retain CDF as the main model because its mean loss on the real datasets is lower at all three budgets (Table 18).

A2. Sharing within output perturbation. Appendix E.4 (Appendix F.4 in the revision) shows that reducing DP noise through sharing does not by itself guarantee better prediction, further highlighting the advantages of our DP-ShareQ algorithms. DP-ShareQ combines sharing with learning and noise calibration tailored to the shared model. Joint learns the location and ordered offsets jointly, using their ordered gradient set to calibrate sensitivity and allocating noise by coefficient direction. CDF learns a shared location by Huber regression and estimates the offsets by inverting a private residual CDF, with noise calibrated to the joint changes of its Fourier coefficients. Table 2 quantifies the contribution of these learning and noise designs.

We further analyzed why sharing gives only modest gains within Sep-Out. At ε=0.1, sharing reduces prediction noise RMS by 22.5%, and loss is lower before rearrangement. The comparison reverses after the common rearrangement. Sep-Out's independently perturbed predictions benefit more from sorting, whereas noise in the shared input effect moves all levels together. DP-ShareQ's predictions are already ordered and remain unchanged by rearrangement.

A3. Noise geometry. Applying non-isotropic noise to Full is nontrivial because Joint's covariance calibration relies on quantile ordering, which Full does not enforce. A valid calibration for Full requires a separate sensitivity analysis. Table 14 instead compares Full with Joint after removing each calibration component separately. The table below summarizes mean nIPL ×100 over datasets and feature maps.

Methodε=0.1ε=0.5ε=4
Full10.9007.7445.170
Joint, isotropic9.4446.6464.425
Joint, without ordered bound9.9847.1454.577
Joint8.8406.2764.332

Joint has lower mean loss than Full with isotropic noise, and also when the ordered-gradient bound is removed instead. Neither calibration component alone accounts for its advantage over Full. Directional noise provides further improvement.

A4. Hyperparameter selection. The main experiments isolate the effect of learning design from private hyperparameter selection. All methods use the same per-fit privacy budget and select settings by validation nIPL before test evaluation.

Appendix B.6 evaluates private selection within the same total privacy budget. Accounting for repeated candidate training reduces the budget available per fit and increases training noise. Limiting the number of candidates controls this cost but may miss better settings, while private validation can select a suboptimal candidate. DP-ShareQ's gains persist under private selection. The performance loss from private selection does not diminish our contribution. Table 6 shows higher loss for every method, while Joint and CDF retain lower mean loss than all three reference methods at each reported budget.

A5. Data splits and nonlinear features. We have rerun the real-data comparisons throughout the main text and appendix on five data splits, with ten noise repetitions per split. The updated results report means and standard deviations across the five splits. At ε=1, both methods have lower mean nIPL than all three reference methods in all 12 dataset–map settings (Table 1). On Real Estate with linear features, nIPL ×100 is 5.159 ± 0.381 for Joint and 5.307 ± 0.603 for CDF, compared with 7.562 ± 0.330 for Full and 7.181 ± 0.846 for Sep-Out.

Appendix D also evaluates 64 and 128 random Fourier features approximating a Gaussian kernel. Joint and CDF retain their advantage over Full, with CDF reducing mean nIPL by 25.9%, 19.3%, and 18.4% at ε=0.1,0.5,4, respectively.

We welcome any further comments or questions. If our responses resolve your concerns, we would be grateful if you would consider reassessing your score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Responses

Thank you again for your time and comments. We have posted our responses and uploaded the revised paper.

We revised the theoretical analysis to remove the conditional symmetry, unimodality, and density assumptions, while explicitly accounting for the Huber-to-quantile comparison cost. The real-data experiments now use five splits, with additional evaluations under stronger privacy, private hyperparameter selection, nonlinear features, and input-dependent scale. We also clarified the methodological contributions and ablations to separate the benefits of sharing and noise calibration.

Before the discussion closes, we would appreciate hearing whether our responses address your concerns and which points still need clarification. We would be happy to discuss further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.