acceptodds
Under review as a conference paper at ICLR 2027

Differentially Private Inference for Geodesic Regression on Riemannian Manifolds

Abstract

In this paper, we study inference for geodesic regression with Euclidean predictors and a manifold-valued response under Gaussian differential privacy (GDP). We develop a one-shot GDP inference procedure via sample splitting. The released outputs include a private parameter estimator, a private Wald region for the footpoint and shooting-vector parameters, and pointwise private confidence regions for the fitted geodesic. We derive a range of theoretical results, including sensitivity bounds and GDP guarantee, consistency, a central limit theorem, and a consistent private covariance estimator. We also illustrate our methodology on several important manifolds and explain how the main theorems specialize in these settings, including Euclidean spaces, spheres, Hadamard manifolds, affine-invariant or log-Euclidean SPD manifolds, and Grassmann manifolds. Extensive numerical experiments provide positive support for the proposed GDP inference framework. We further apply the methodology to the tropical-cyclone track data.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Differential privacy, Geodesic regression, Manifold data, Statistical inference

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_vjxhOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper provides mechanisms for differentially private inference in the setting of manifold-valued regression. In particular, it does a one-shot DP release, followed by post-processing to obtain pointwise and Wald confidence regions.

Critical strengths
  1. The work is substantial in the context of the niche DP-Geometric Statistics community, with potential advanced applications involving manifold-valued data.
  2. The theory is reasonably done, although some questions remain as covered in the review.
Critical weaknesses
  1. How is K chosen in practice though? What kind of apriori information is needed to get a guess for a reasonable K? 2.) Why is the error lower for the private (noised) setting than the no-noise setting? Seems counter-intuitive. What insight are you trying to emphasize as a take away there?
Questions for authors
  1. How is K chosen in practice though? What kind of apriori information is needed to get a guess for a reasonable K?
  2. The description of some terminologies needs substantial care, before using them. -For instance, the phrase shooting-vectors is not known to the wider generic community overall, and is used about 18 times without explaining it. -Similarly, the concept of clipping for sensitivity bounding is well known in the DP community. But you use the concept of active and inactive clipping, without explaining what they refer to. 3.) Why is the error lower for the private (noised) setting than the no-noise setting? Seems counter-intuitive. What insight are you trying to emphasize as a take away there?
3 comments

Thank you for your time and valuable comments. We appreciate the opportunity to clarify our work and improve the manuscript.

W1. Please see Question 1.

W2. Please see Question 3.

Q1. In our experiments, we specified KK by setting ranges for the footpoint and shooting-vector coordinates based on the units of the variables and the scope of the study. The choice of KK should reflect the interpretation of the parameters, estimation uncertainty, and the geometric conditions required by the model. It generally does not require knowledge of the true parameters or the full data-generating distribution. A formal approach is to use a private pilot computed from a sample independent of the subsequent inference sample. Suppose an error bound rr is available such that the distance between the pilot and the true parameter is at most rr with high probability. One can then choose K=\left\\{\eta: \\|\eta-\widetilde{\eta}\\|\le 2r\right\\}. The factor of 2 just leaves a positive margin, and it's not an optimal constant. Related ideas also appear in CoinPress (Biswas et al., 2020).

In practice, one can also specify parameter ranges directly from the units of the variables and the scope of the study. For geodesic regression, the main considerations are the plausible location of the footpoint pp and the magnitude of the shooting vector vv. For example, with a single predictor whose study range is scaled to [−1,1][-1,1], imposing ∣v∣≤s\\|v\\|\le s ensures that the fitted curve travels a distance of at most ss from x=0x=0 to either endpoint. Thus, a useful starting point is to consider how much the fitted response could reasonably change over the predictor range, and then translate this into a bound on vv. Together with a plausible range for pp, this gives a specification of KK in the chosen coordinates. These ranges should allow room for estimation uncertainty while satisfying the geometric conditions.

The relevant prior information therefore concerns the predictor units and range, plausible footpoint locations, and plausible variation of the fitted response. Public reference data can help supply this information but are not necessary. A publicly specified KK can also be a parameter range fixed before the analysis and independently of the protected observations.

Q2. We will explain these terms when they first appear. A shooting vector is a tangent vector specifying the initial velocity of a geodesic. We use norm clipping for per-sample scores, Hessians, and score second moments. Clipping is active when a contribution's norm exceeds its threshold and the vector or matrix is scaled down to that threshold. Clipping is inactive when its norm is at or below the threshold and the contribution remains unchanged.

Q3. In the main simulations, Full LS applies no privatization at all and minimizes the geodesic least-squares loss over the full sample. We use it as a nonprivate error benchmark. No-noise keeps the same private pilot and sample split as Ours, but adds no Gaussian perturbation at the inference stage. This gives a benchmark for the cost of privatizing the inference stage, which is exactly the stage we focus on given a private pilot. The other methods (Ours, Seq-E, and Seq-M) privatize both the pilot and the inference stage. In Table 1, all three fully private methods have higher mean curve error than both No-noise and Full LS in every reported configuration, and No-noise also has higher mean curve error than Full LS. This is consistent with the accuracy cost of privatization. To avoid confusion, we will rename No-noise and write Full LS as Full LS (nonprivate).

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

[1] Sourav Biswas, Yihe Dong, Gautam Kamath, and Jonathan Ullman. CoinPress: Practical Private Mean and Covariance Estimation.

Thank you again for your thoughtful review and helpful suggestions. We hope our responses and revisions have addressed the issues you raised, and we would be glad to clarify any remaining points. If you are satisfied that your main concerns have been resolved, we would be grateful if you would reconsider your assessment and score in light of the revisions.

AuthorsOpenReview ↗

Follow-up on our responses

As the discussion period draws to a close, we would like to thank you again for your time and careful review. If time permits, could you let us know whether any questions remain after reading our responses? We would be happy to clarify any remaining points. We appreciate your consideration and the valuable feedback you have provided.

Reviewer_Kc4EOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary
  • Paper studies differentially private inference (not just estimation) for geodesic regression where predictors (x) are on Euclidean space and response (y) lives on a Riemannian manifold
  • proposed procedure has two phases 1) Pilot phase where rough private estimate of regression parameters is done 2) inference phase, where a noisy gradient, Hessian are released once and used to refine the pilot and get private confidence intervals. It is shown to have a GDP guarantee and asymptotic normality.
  • It is tested on manifolds (sphere S2, 2x2 SPD matrices) and datasets (simulated data, two tropical-cyclone tracks).
Critical strengths

The paper's main strength is filling a real gap: private uncertainty quantification for manifold-valued regression.

  • Releasing once and getting the whole fitted curve and its confidence regions by post-processing is clean, and the privacy cost does not grow with the number of queries.
  • Coverage is close to nominal 95% for moderate n and privacy budget.
Critical weaknesses

The paper's main weaknesses are the cost of sample splitting and weak experiments.

  • The no-split baseline (Seq-E) has lower curve error than the proposed method in almost every setting (e.g., 0.056 vs 0.067 at n=400, µ=2), with similar coverage. So the splitting needed for the theory costs accuracy in practice.
  • Experiments are small (S2 and SPD(2), n ≤ 400), and the real-data example (two cyclone tracks) is neither privacy-sensitive nor iid.
Questions for authors
  • Why use sample splitting if Seq-E performs better? Can the theory be extended to the no-split version?
  • Is there a real application with actually sensitive data (e.g., medical imaging or diffusion tensors)?
  • Some relevant work is missing [1, 2]. Though not directly related, it is worth discussing for the benefit of the reader, since there is very little work on DP on Riemannian manifolds.

Refs:

[1] Differentially private Riemannian optimization

[2] Differentially Private Fréchet Mean on the Manifold of Symmetric Positive Definite (SPD) Matrices with log-Euclidean Metric

4 comments

Thank you for your time and valuable comments. We appreciate the opportunity to clarify our work and improve the manuscript.

W1. We agree that sample splitting can reduce statistical efficiency. We use it so that the second stage can be analyzed as a separate inference step, with a coverage guarantee conditional on the completed first stage. The pilot may be constructed by different methods, as long as it is independent of the inference sample and meets the stated accuracy and regularity conditions. Once the pilot has been obtained, we treat the first-stage information as given. The independent sample D2D_2 then supplies the data for the one-step correction and confidence regions. Under our assumptions, \Pr\\!\left\\{ \theta_0\in C_\theta^{1-\alpha} \mid \mathcal F_1 \right\\} \xrightarrow{p}1-\alpha, \qquad \mathcal F_1=\sigma(D_1,\widetilde\eta_n), with an analogous result for each fixed fitted-response point.

The practical meaning of the conditional guarantee is: the second-stage inference remains asymptotically calibrated given the first-stage information already obtained. Its justification does not rest only on averaging coverage over repeated runs with different pilots. The same second-stage analysis can therefore be used with different independent pilot estimators satisfying our assumptions, without requiring a particular pilot algorithm. With the parameter of interest and the confidence-interval procedure fixed in advance, we can use the first-stage private pilot to decide whether to report the second-stage results. The reported intervals retain asymptotic coverage. The split also permits maximum-composition GDP accounting across the two stages.

W2. We have extended the main simulations to n=800n=800 and 16001600. The table below reports curve error, parameter coverage, and average pointwise coverage for all five methods. Curve error decreases with sample size on both manifolds. We agree that the cyclone example has limited privacy motivation and that its temporal dependence is not covered by our iid theory. We have conducted a new biomedical analysis using 899 participants from AOMIC-ID1000. The response is a regional diffusion tensor in SPD(3)\mathrm{SPD}(3) from the genu of the corpus callosum, and the predictor is biological sex. Each participant contributes one predictor–response pair and constitutes one privacy unit. Please see our response to Q2.

Table 1. Simulation results for n=800n=800 and 16001600 over 500 replications on S2\mathbb S^2, with log-Euclidean SPD(2)\mathrm{SPD}(2) in parentheses. Nominal coverage is 0.950.95.

(a) Mean curve error

nnμ\muOursFull LSNo-noiseSeq-ESeq-M
8000.50.149 (0.176)0.014 (0.018)0.046 (0.025)0.125 (0.152)0.137 (0.175)
1.00.072 (0.089)0.014 (0.017)0.026 (0.025)0.059 (0.075)0.068 (0.086)
2.00.038 (0.047)0.013 (0.017)0.018 (0.023)0.031 (0.039)0.035 (0.045)
16000.50.074 (0.086)0.010 (0.012)0.026 (0.017)0.059 (0.073)0.067 (0.086)
1.00.036 (0.045)0.010 (0.012)0.014 (0.016)0.030 (0.037)0.035 (0.043)
2.00.021 (0.026)0.010 (0.012)0.013 (0.016)0.018 (0.021)0.020 (0.024)

(b) Parameter coverage

nnμ\muOursFull LSNo-noiseSeq-ESeq-M
8000.50.924 (0.948)0.950 (0.936)0.750 (0.988)0.922 (0.934)0.938 (0.934)
1.00.940 (0.942)0.940 (0.938)0.906 (0.980)0.942 (0.950)0.946 (0.942)
2.00.934 (0.932)0.952 (0.940)0.948 (0.972)0.948 (0.940)0.954 (0.956)
16000.50.938 (0.944)0.948 (0.942)0.820 (0.984)0.946 (0.946)0.958 (0.946)
1.00.936 (0.942)0.936 (0.932)0.946 (0.980)0.922 (0.940)0.938 (0.942)
2.00.930 (0.920)0.936 (0.932)0.954 (0.962)0.924 (0.928)0.924 (0.926)

(c) Average pointwise coverage

nnμ\muOursFull LSNo-noiseSeq-ESeq-M
8000.50.933 (0.937)0.951 (0.948)0.804 (0.978)0.930 (0.932)0.951 (0.941)
1.00.946 (0.954)0.943 (0.943)0.920 (0.974)0.952 (0.948)0.949 (0.956)
2.00.950 (0.947)0.948 (0.950)0.957 (0.969)0.951 (0.950)0.957 (0.955)
16000.50.934 (0.956)0.943 (0.945)0.849 (0.980)0.948 (0.940)0.960 (0.943)
1.00.935 (0.948)0.944 (0.937)0.946 (0.976)0.936 (0.946)0.942 (0.944)
2.00.927 (0.942)0.944 (0.943)0.945 (0.958)0.934 (0.948)0.934 (0.950)

Q1. Yes. The same-sample version can have valid unconditional inference. We keep Assumptions 1–5 and 7, replacing n2n_2 by nn, and use fixed positive privacy budgets and fresh release noises independent of the data and pilot. The pilot may use the same observations as the inference stage, provided that ∣η~n−η0∣=op(n−1/4)\\|\widetilde\eta_n-\eta_0\\|=o_p(n^{-1/4}). The Taylor expansion holds for each dataset. Its remainder is bounded by (LH/2)∣η~n−η0∣2(L_H/2)\\|\widetilde\eta_n-\eta_0\\|^2, so pilot independence is not needed for the first-order cancellation in the one-step update. This gives \widehat\eta_n^{\mathrm{dp}}-\eta_0 = -A_0^{-1} \left\\{ \frac1n\sum_{i=1}^n\psi(W_i,\eta_0)+Z_{\psi,n} \right\\} +o_p(n^{-1/2}). Hence the estimator is consistent and asymptotically normal, and n(η^ndp−η0)⇒N(0,Ω),nV^ndp→pΩ,Ω=A0−1J0A0−1.\sqrt n\bigl(\widehat\eta_n^{\mathrm{dp}}-\eta_0\bigr) \Rightarrow N(0,\Omega), \qquad n\widehat V_n^{\mathrm{dp}}\xrightarrow{p}\Omega, \qquad \Omega=A_0^{-1}J_0A_0^{-1}. Under the corresponding nondegeneracy conditions, these results give marginal Wald coverage for the parameter and each fixed fitted-response point.

The conditional guarantees do not follow in general. A same-sample pilot can contain information about the empirical-score fluctuation used in the second stage. Conditioning on that pilot changes the distribution of this fluctuation. Therefore, the conditional covariance formula in Theorem 3, Equation (3), and the conditional-coverage guarantee of the split construction need not hold under the same assumptions, even when marginal coverage is correct. The privacy accounting also changes. For a μ0\mu_0-GDP pilot, sequential composition gives μtot=μ02+μψ2+μH2+μM2\mu_{\mathrm{tot}} = \sqrt{\mu_0^2+\mu_\psi^2+\mu_H^2+\mu_M^2}. We will add the same-sample result, making clear which unconditional conclusions extend and which conditional conclusions require the independence used in the split design.

Q2. We have conducted an additional experiment using the Amsterdam Open MRI Collection (AOMIC-ID1000). AOMIC provides structural, functional, and diffusion MRI data, together with demographic and psychometric measurements from healthy participants. We analyze 899 participants, including 462 females and 437 males. The response is a regional diffusion tensor in SPD(3)\mathrm{SPD}(3), obtained by spatially aligning and reorienting the officially preprocessed tensors and taking their log-Euclidean mean within an anatomically specified region of the genu of the corpus callosum. Recorded biological sex is the predictor, coded as Xi=−1X_i=-1 for females and Xi=1X_i=1 for males. Each participant contributes one predictor–response pair and constitutes one privacy unit. We use the log-Euclidean metric, a 40/60 sample split, the joint KNG pilot, and the original inference-budget allocation and regularization settings, with an overall GDP budget of μ=2\mu=2. For both sex groups, the private fitted diffusion tensors are close to the Full-LS estimates, while the private confidence regions are wider. In the parameterization Y^(x)=exp⁡(A^+xB^)\widehat{Y}(x) = \exp(\widehat{A}+x\widehat{B}), exp⁡(A^)\exp(\widehat{A}) is the log-Euclidean midpoint of the two fitted group tensors, and 2B^2\widehat{B} represents their male-minus-female difference in matrix logarithms. The estimated diagonal contrast is (−0.0098,,0.0204,,0.1087)(-0.0098,\\,0.0204,\\,0.1087). The largest fitted change in the third diagonal log-coordinate, with a nominal 95% marginal interval of (0.0249,0.1925)(0.0249,0.1925). Thus, the resulting male fitted tensor has a slightly less elongated diffusion ellipsoid than the female fitted tensor.

Q3. We will discuss both in the second paragraph of the Introduction, after the discussion of Riemannian Gaussian mechanisms and before the work on private geodesic regression. We will add: “Han et al. (2024) develop a general framework for private empirical risk minimization on Riemannian manifolds using gradient perturbation. They establish privacy guarantees and derive utility bounds for geodesically convex and strongly convex, general nonconvex objectives, and objectives satisfying the Riemannian PL condition. Utpala et al. (2023) develop a tangent Gaussian mechanism for log-Euclidean Fréchet means on SPD manifolds under approximate DP. Their mechanism allows direct sampling and improves the dimension dependence of the expected squared perturbation error.”

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

Thank you again for your thoughtful review and helpful suggestions. We hope our responses and revisions have addressed the issues you raised, and we would be glad to clarify any remaining points. If you are satisfied that your main concerns have been resolved, we would be grateful if you would reconsider your assessment and score in light of the revisions.

AuthorsOpenReview ↗

Follow-up on our responses

As the discussion period draws to a close, we would like to thank you again for your time and careful review. If time permits, could you let us know whether any questions remain after reading our responses? We would be happy to clarify any remaining points. We appreciate your consideration and the valuable feedback you have provided.

Comments on the paper

7 comments

Thank you for your careful reading of our manuscript and the constructive feedback.

W1. Fixed-threshold clipping of per-observation gradients has also been studied in private estimation. For example, Avella-Medina, Bradshaw and Loh (2023) use fixed clipping thresholds in their simulations. However, we agree that this issue needs to be handled explicitly in our inference procedure. We have now rerun the experiments using the Appendix F corrections, and report the score, Hessian, and moment clipping fractions in our response to Q1. For both main simulation settings, we replaced Gaussian observation noise with bounded symmetric noise: uniform tangent-disk noise of radius 0.40.4 on S2\mathbb S^2 and uniform ball noise of radius 0.250.2\sqrt{5} in the three-dimensional SPD log coordinates. Both retain the original noise covariance 0.04I0.04I. All derivatives are now analytic or obtained by automatic differentiation.

W2. The main simulation experiments used the exact rejection sampler in Appendix B. The proposal, acceptance, and timing diagnostics in Q2 establish its computational feasibility in these low-dimensional settings. The higher-dimensional experiments used two sequential random-walk Metropolis–Hastings runs. We will correct that implementation description in Appendix C.4 and revise the associated claims. Our main contribution is a framework for private inference in geodesic regression that allows different independent pilot. We will further investigate how existing private estimation and optimization methods can be adapted to construct computationally practical pilots for our framework.

W3. The revised sphere experiment now uses the intrinsic squared-distance loss and analytic derivatives on S2\mathbb S^2. We have also run an affine-invariant SPD(2)\mathrm{SPD}(2) experiment. Its design and results are reported in our response to Q4.

W4. Our covariance estimator combines a private sandwich estimator with an explicit correction for score-release noise. This construction is closely related to that of Avella-Medina, Bradshaw and Loh (2023). They add Gaussian noise to both the Hessian and the score second-moment matrix, project both matrices onto a positive-definite cone, and use them to construct a private sandwich estimator. Their additional covariance correction retains only the gradient-noise variance. Xia, Zhang and Cai (2025) also use a similar construction for DP-SGD. They add Gaussian noise to the Hessian and score second-moment estimates and establish consistency of the resulting private sandwich estimator by bounding the effects of both matrix-release noises. Their additional variance correction explicitly accounts for gradient noise, without separate additive corrections for Hessian or second-moment release noise. The asymptotically negligible ridge regularization is also used in other private inferences. For example, Raj et al. (2020) add a diagonal regularizer to a privately estimated covariance matrix before inversion and let it tend to zero. The resulting test statistic retains the usual asymptotic null distribution. In our implementation, we choose 6/ninfer6/n_{\mathrm{infer}}, which satisfies Assumption 7 and makes its contribution negligible in the first-order expansion. For projection effect, although the models and projection schemes are not exactly the same as ours, Avella-Medina, Bradshaw and Loh (2023) and Pastore and Gastpar (2021) also show that projection effects are asymptotically negligible under similar regularity conditions. Their results cover the positive-definite projection of noisy Hessians and the projection of private probability estimates onto the simplex, respectively. Also, the following table reports the Hessian PSD and final parameter projection activation rates for n=400,800,1600n=400,800,1600. No Hessian PSD activation was observed in these settings and final parameter projection becomes less frequent as nn increases. These results are consistent with the asymptotic inactivity established in our analysis. Regarding coverage, Avella-Medina, Bradshaw and Loh (2023) also report that asymptotically valid private sandwich intervals can undercover in small samples.

On tuning, we included sensitivity analyses in the original manuscript and have now repeated and extended them under the revised simulation settings. Our response to Q3 gives concrete recommendations based on the theoretical conditions, related literature, and these experiments.

Table 2. Projection activation rates (%) in the main simulations for n=400,800,1600n=400,800,1600. Results are for S2\mathbb S^2, with log-Euclidean SPD(2)\mathrm{SPD}(2) in parentheses.

nnμ\muHessian PSD projectionFinal parameter projection
40010.0 (0.0)20.8 (9.8)
40020.0 (0.0)7.6 (0.2)
80010.0 (0.0)7.8 (0.4)
80020.0 (0.0)0.4 (0.0)
160010.0 (0.0)0.2 (0.0)
160020.0 (0.0)0.0 (0.0)

W5. We have conducted an additional real-data analysis using AOMIC-ID1000, with one regional diffusion tensor and one predictor value per participant. Each participant is the privacy unit and is treated as iid. We describe the analysis in our response to Q5.

Q1. We reran the main simulations, and the tables below report selected results and clipping fractions. The performance trends are essentially the same as in the paper.

Table 3. Simulation results for Ours and Full LS on S2\mathbb S^2 over 500 replications. The three metrics are mean curve error, parameter coverage, and average pointwise coverage. Nominal coverage is 0.950.95.

nnμ\muError (Ours)Error (Full LS)Param. cov. (Ours)Param. cov. (Full LS)Pointwise cov. (Ours)Pointwise cov. (Full LS)
1000.50.6940.0390.9840.9240.9840.936
1.00.4410.0400.9440.9060.9630.929
2.00.2400.0380.9620.9280.9630.936
2000.50.4720.0270.9460.9380.9630.944
1.00.2600.0270.9520.9340.9600.940
2.00.1330.0280.9420.9140.9590.938
4000.50.2710.0190.9300.9400.9520.952
1.00.1390.0200.9340.9560.9470.954
2.00.0720.0190.9520.9520.9600.954
8000.50.1490.0140.9240.9500.9330.951
1.00.0720.0140.9400.9400.9460.943
2.00.0380.0130.9340.9520.9500.948
16000.50.0740.0100.9380.9480.9340.943
1.00.0360.0100.9360.9360.9350.944
2.00.0210.0100.9300.9360.9270.944

Table 4. Clipping percentages for Ours on S2\mathbb S^2 over 500 replications, for n=200,400,800n=200,400,800.

nnμ\muScore (%)Hessian (%)Moment (%)
2000.50.00500.00000.0050
1.00.00170.00170.0017
2.00.00000.00000.0000
4000.50.00080.00080.0008
1.00.00000.00000.0000
2.00.00000.00000.0000
8000.50.00000.00000.0000
1.00.00000.00000.0000
2.00.00000.00000.0000

Q2 For Table 1 in the main text, we used the exact rejection sampler in Appendix B for all private pilots. The table below reports diagnostics for the main experiments. Acceptance is the total number of accepted proposals divided by the total number of proposals. Time is the mean elapsed time per pilot call, excluding subsequent inference. These measurements establish computational feasibility for the low-dimensional settings in Table 1. Our description of the higher-dimensional experiments was inaccurate. Those experiments used two sequential random-walk Metropolis–Hastings runs, and is not covered by the GDP proof in Appendix B. We apologize for the inaccurate description and will correct it.

Table 5. Exact-pilot diagnostics of the main simulation. Each entry is S2\mathbb S^2 / SPD(2)\mathrm{SPD}(2). Proposal counts are means, with maxima in parentheses.

nnμ\muProposal counts: mean (maximum)Acceptance (%)Time (ms)
1001.04.00 (21) / 5.37 (31)25.025 / 18.6150.064 / 0.048
1002.013.83 (91) / 27.34 (149)7.230 / 3.6580.140 / 0.072
2001.08.98 (54) / 13.59 (73)11.131 / 7.3570.244 / 0.076
2002.051.24 (342) / 128.30 (971)1.952 / 0.7790.810 / 0.232
4001.027.81 (183) / 65.76 (370)3.596 / 1.5210.928 / 0.152
4002.0270.23 (1516) / 1084.88 (7340)0.370 / 0.0928.192 / 1.412

Q3. For clipping levels, common approaches include using fixed constants from previous studies (Bu et al., 2024) or choosing thresholds from public data and prior information (Cummings et al., 2024). This does not consume any privacy budget. Alternatively, thresholds can also be selected adaptively by the data using a private mechanism, with its cost included in the total privacy budget (Chang et al., 2026). In our simulations, we used separate reference samples generated independently from the same simulation model. In the real-data analysis, the thresholds were informed by the practical meaning of the variables. Our sensitivity experiments show that the performance is reasonably stable across the tested clipping levels.

For the ridge parameter, we recommend a ridge parameter of the same order as ninfer−1n_{\mathrm{infer}}^{-1} to satisfy Assumption 7, namely a/ninfera/n_{\mathrm{infer}}, where a>0a > 0 is a fixed constant. Our sensitivity experiments show that the results are stable across aa.

The 40/60 pilot–inference split is an empirical choice that balances pilot accuracy against the size of the independent inference sample. It is consistent with Liu, Yu and Li (2022), who recommend allocating 40%–60% of the sample to testing in an estimation-then-testing procedure. Based on the literature and our sensitivity experiments, we recommend allocating 40%–60% of the sample to inference.

We use fixed budget weights proportional to (0.45,0.78,0.45)(0.45,0.78,0.45) for the score, Hessian, and second-moment releases, respectively. This is an empirical choice, and gives more budget to the Hessian to improve the stability of its inverse, which enters both the one-step update and the covariance estimator. We choose this allocation because previous work has discussed how Hessian noise can destabilize Newton updates and compromise positive definiteness (Ji et al., 2014; Avella-Medina, Bradshaw and Loh, 2023). Our sensitivity experiments show that the method is fairly stable across different budget allocations, except when most of the budget is assigned to the score.

Table 6. Sensitivity to clipping thresholds on S2\mathbb S^2 at n=400n=400 and μ=1\mu=1.

Scale aaCurve errorParameter coveragePointwise coverage
0.850.1230.9240.936
1.000.1390.9340.947
1.500.1630.9560.964
2.000.2070.9620.964

Table 7. Sensitivity to ridge regularization on S2\mathbb S^2 at n=400n=400 and μ=1\mu=1.

RidgeCurve errorParameter coveragePointwise coverage
1/ninfer1/n_{\mathrm{infer}}0.1460.9320.947
3/ninfer3/n_{\mathrm{infer}}0.1430.9420.946
6/ninfer6/n_{\mathrm{infer}}0.1390.9340.947
9/ninfer9/n_{\mathrm{infer}}0.1370.9460.949

Table 8. Sensitivity to the pilot sample fraction on S2\mathbb S^2 at n=400n=400 and μ=1\mu=1.

n1/nn_1/nCurve errorParameter coveragePointwise coverage
0.20.1260.8900.903
0.30.1280.9300.938
0.40.1390.9340.947
0.50.1520.9660.964
0.60.1620.9620.968

Table 9. Sensitivity to privacy-budget allocation on S2\mathbb S^2 at n=400n=400 and μ=1\mu=1.

wwCurve errorParameter coveragePointwise coverage
(0.45,0.78,0.45)(0.45,0.78,0.45)0.1390.9340.947
(0.58,0.58,0.58)(0.58,0.58,0.58)0.1240.9320.930
(0.90,0.35,0.35)(0.90,0.35,0.35)0.1310.7820.807
(0.35,0.90,0.35)(0.35,0.90,0.35)0.1680.9560.962
(0.35,0.35,0.90)(0.35,0.35,0.90)0.1930.9240.938

Q4. We added affine-invariant SPD(2)\mathrm{SPD}(2) regression with X∼Unif⁡[−1,1]X\sim\operatorname{Unif}[-1,1], s0=(0.2,−0.1,0.12)⊤s_0=(0.2,-0.1,0.12)^\top, and b0=(0.3,−0.2,0.08)⊤b_0=(0.3,-0.2,0.08)^\top. Define svec⁡(abbc)=(a,c,2b)⊤\operatorname{svec}\left(\begin{smallmatrix}a & b \\\\ b & c\end{smallmatrix}\right)=(a,c,\sqrt{2}b)^\top, S0=svec⁡−1(s0)S_0=\operatorname{svec}^{-1}(s_0), and B0=svec⁡−1(b0)B_0=\operatorname{svec}^{-1}(b_0). With P0=eS0P_0=e^{S_0} and V0=P01/2B0P01/2V_0=P_0^{1/2}B_0P_0^{1/2}, the true curve is Q0(x)=Exp⁡P0(xV0)=eS0/2exB0eS0/2Q_0(x)=\operatorname{Exp}_{P_0}(xV_0)=e^{S_0/2}e^{xB_0}e^{S_0/2}. We generate Y=Q0(X)1/2exp⁡svec⁡−1(e)Q0(X)1/2Y=Q_0(X)^{1/2}\exp\\{\operatorname{svec}^{-1}(e)\\}Q_0(X)^{1/2}, with ee independently uniform in the three-dimensional ball of radius 0.250.2\sqrt{5}. Ours uses the exact rejection pilot, a 40/60 split, inference-budget weights (0.45,0.78,0.45)(0.45,0.78,0.45) normalized to squared sum μ2\mu^2, clipping thresholds (2.5,4,6.25)(2.5,4,6.25), and ridge 6/ninfer6/n_{\mathrm{infer}}.

To our knowledge, existing work does not provide a directly applicable procedure combining private point estimation with established coverage guarantees for geodesic regression under the GDP setting considered here. Developing such inference is the problem addressed by our paper. We adapted Avella-Medina, Bradshaw and Loh (2023), denoted AMBL, to the same intrinsic loss in local coordinates. We report it as an empirical reference. The following table reports the performance of the two methods. Overall, the performance is similar to that observed in the main simulations.

Table 10. Affine-invariant SPD(2)\mathrm{SPD}(2) regression over 500 replications.

nnμ\muCurve error — OursCurve error — AMBLParameter coverage — OursParameter coverage — AMBLPointwise coverage — OursPointwise coverage — AMBL
1001.00.6110.3950.9340.9120.9790.976
2.00.3200.2120.9260.8660.9520.946
2001.00.3270.2290.9340.8480.9650.917
2.00.1660.1180.9520.8160.9610.919
4001.00.1720.1190.9380.7980.9550.890
2.00.0870.0640.9380.8000.9520.891

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.