betaFeedback

WorksComputer Science

Optimal Conversion from Zero-Concentrated to Gaussian Differential Privacy

Abstract

Gaussian differential privacy (GDP) and zero-concentrated differential privacy (zCDP) are two competing privacy accounting frameworks, but they share similar minimax lower bounds in various estimation problems. Moreover, it is known that a mechanism satisfying μμ-GDP satisfies (μ2/2)(μ^2/2)-zCDP, whereas the converse fails to hold at the same calibration. In this paper, we establish the optimal reverse conversion, showing that every ρρ-zCDP mechanism satisfies μzCDP(ρ)μ_{\mathrm{zCDP}}(ρ)-GDP, where μzCDP(ρ)=2Φ−1((1+e−tρ)−1)μ_{\mathrm{zCDP}}(ρ)=2Φ^{-1}((1+e^{-t_ρ})^{-1}) and tρt_ρ uniquely solves the equation ρ=tρtanh⁡(tρ/2)ρ=t_ρ\tanh(t_ρ/2). We also obtain simplified bounds πρ≤μzCDP(ρ)≤8ρ\sqrt{πρ}\leqμ_{\mathrm{zCDP}}(ρ)\leq\sqrt{8ρ}, with sharp constants, which imply equivalence of the two frameworks up to constants. This explains why their minimax risks agree up to a constant rescaling of the privacy budget. We show that the squared conversion is strictly convex, which implies that converting zCDP to GDP before applying composition or group privacy yields a tighter GDP guarantee than converting afterward. We also derive the sharp conversion from bounded-range and logistic DP mechanisms (e.g., exponential mechanism) to GDP, where converting before composition but after group privacy gives the tighter guarantee.

Introduction

Differential privacy (DP) limits what a randomized computation reveals about any individual’s contribution, which creates plausible deniability and thereby provides privacy protection [15, 17]. To satisfy DP, it requires the output distributions on neighboring datasets to be similar, and this similarity can be formalized either through divergences or through hypothesis testing. Concentrated differential privacy and Rényi differential privacy (RDP) bound divergences between the output distributions [9, 24]. The ff-DP framework instead quantifies their similarity through hypothesis tests that distinguish them, requiring that at type I error α\alpha, every such test has type II error at least f(α)f(\alpha), for a given function ff [18]. Within this framework, Gaussian differential privacy (GDP) is the primary single-parameter framework, with ff given by testing between unit-variance Gaussian distributions with mean difference μ\mu.

Zero-concentrated differential privacy (zCDP) and GDP are both one-parameter frameworks motivated by the Gaussian mechanism, and their parameters have similar interpretations. Moreover, μ\mu-GDP implies (μ2/2)(\mu^2/2)-zCDP, and the Gaussian mechanism attains both bounds exactly [18]. While zCDP is more commonly used in the computer science literature [9, 21, 7, 12], where it offers additive composition, covers bounded-range mechanisms such as the exponential mechanism through standard conversions, and underlies the noisy measurements of the 2020 Census [AACM+22], GDP is more common in statistics [18, 8, 2, 5, 4, 32, 13, 27, 26], where its hypothesis-testing guarantee is directly interpretable, and [18] claim it is the tightest guarantee for the Gaussian mechanism, based on an argument on informativeness. Despite the similarities, it has not been established which GDP guarantee, if any, a zCDP budget implies, or whether the two frameworks are equivalent up to constants. Without such a result, a mechanism analyzed under the zCDP framework is hard to report in the more interpretable GDP framework of privacy accounting, and risk bounds proved in one framework cannot be transferred to the other.

The known implication from GDP to zCDP gives one direction, but its converse at the same calibration fails [18]. RDP-to-testing conversions derive trade-off functions from Rényi constraints, most recently the optimal one implied by RDP bounds at all orders [28], but they do not identify the sharp GDP parameter implied by a zCDP budget.

Our results

In this paper, we establish the sharp zCDP-to-GDP conversion and use it to compare the mechanism classes and their minimax risks. We also derive privacy accounting consequences and a sharp bounded-range to GDP calculation. Specifically, our results are as follows:

Optimal conversion. Let Φ(t)\Phi(t) be the cumulative distribution function (CDF) of a standard Gaussian. Then, for every ρ>0\rho> 0, the tightest (smallest) GDP parameter guaranteed by ρ\rho-zCDP is

μzCDP(ρ)=2Φ−1(σ(tρ)),\mu_{\mathrm{zCDP}}(\rho) = 2\Phi^{-1}\bigl(\sigma(t_\rho)\bigr),

where tρt_\rho uniquely solves the calibration equation ρ=tρtanh⁡(tρ/2)\rho= t_\rho\tanh(t_\rho/2) and σ(t)=(1+e−t)−1\sigma(t) = (1 + e^{-t})^{-1} is the logistic function. Our proof is by contraposition, and the tightness of the conversion is achieved by the binary randomized response mechanism.

Conversion loss and statistical equivalence. We simplify the above conversion to obtain πρ≤μzCDP(ρ)≤8ρ\sqrt{\pi\rho} \le\mu_{\mathrm{zCDP}}(\rho) \le\sqrt{8\rho}, where μzCDP(ρ)/ρ\mu_{\mathrm{zCDP}}(\rho)/\sqrt{\rho} increases from π\sqrt{\pi} to 8\sqrt{8} over ρ>0\rho> 0, so both constants are tight. Together with the known conversion from GDP to zCDP, this yields,

μ-GDP⟹(μ2/2)-zCDP⟹2μ-GDP.\mu\text{-GDP} \quad\Longrightarrow\quad(\mu^2/2)\text{-zCDP} \quad\Longrightarrow\quad2\mu\text{-GDP}.

Figure 1. The three trade-off curves, G2ρG_{\sqrt{2\rho}}, the numerical ρ\rho-zCDP testing boundary, and GμzCDP(ρ)G_{\mu_{\mathrm{zCDP}}(\rho)} illustrate the GDP–zCDP round-trip loss.

This shows that zCDP and GDP are equivalent up to constants. In particular, for any estimation problem, this yields a general minimax comparison between GDP and zCDP up to constant-factor changes in the privacy budget.

Composition and group privacy. We prove that the squared conversion map ρ↦μzCDP(ρ)2\rho\mapsto\mu_{\mathrm{zCDP}}(\rho)^2 is strictly convex. This implies that converting individual zCDP budgets to GDP before applying composition or group privacy improves the resulting GDP bound.

GDP result for bounded-range and logistic mechanisms. Using similar techniques, we also prove the sharp conversion from logistic DP (including bounded-range mechanisms such as the exponential mechanism) to GDP. The square of the conversion map is strictly convex, favoring conversion to GDP before composition, while its concavity implies that group privacy should be applied before converting to GDP.

Remark 1 Figure 1 illustrates the conversion from ρ\rho-zCDP to GDP at ρ=1\rho= 1. The trade-off function for ρ\rho-zCDP is computed as follows: For a type I error α\alpha and power qq, let KzCDP(α,q)K_{\mathrm{zCDP}}(\alpha,q) denote the smallest ρ\rho for which Bern⁡(α)\operatorname{Bern}(\alpha) and Bern⁡(q)\operatorname{Bern}(q), viewed as the output distributions of a mechanism on two neighboring datasets, satisfy ρ\rho-zCDP. The figure shows the ρ\rho-zCDP testing boundary TzCDPρ(α)=1−sup⁡{q∈[α,1]:KzCDP(α,q)≤ρ}T_{\mathrm{zCDP}}^{\rho}(\alpha)=1-\sup\{q\in[\alpha,1]:K_{\mathrm{zCDP}}(\alpha,q)\leq\rho\}, which is the smallest type II error at level α\alpha among such Bernoulli pairs, computed numerically by solving KzCDP(α,q)=ρK_{\mathrm{zCDP}}(\alpha,q)=\rho for qq [BBG+20, RGC+26]. This boundary lies above the curve GμzCDP(ρ)G_{\mu_{\mathrm{zCDP}}(\rho)}, where GμG_\mu gives the smallest type II error at each type I error level when testing N(0,1)N(0,1) against N(μ,1)N(\mu,1), and touches it where the type I and type II errors both equal σ(−tρ)\sigma(-t_\rho), so μzCDP(ρ)\mu_{\mathrm{zCDP}}(\rho) cannot be decreased. The shaded region is the loss when a 2ρ\sqrt{2\rho}-GDP guarantee, such as that of the Gaussian mechanism, is converted to ρ\rho-zCDP and back.

Related work

Divergence-based guarantees, such as zCDP and Rényi DP, control the moments of the privacy loss, which is the log-likelihood ratio between the output distributions on neighboring datasets [9, 24]. zCDP is equivalent to bounding the Rényi divergence at every order by a fixed multiple of the order, which is based on the Rényi divergence between two unit-variance Gaussians with differing means. Conversions from divergence-based guarantees to approximate DP, which bounds the likelihood ratio except on an event of small probability, have been established in [9, 24] and were sharpened by [12]. Pure DP, which bounds the privacy loss uniformly, converts sharply to RDP and zCDP, and binary randomized response attains these bounds [29]. Exact zCDP budgets have been derived for several mechanisms, including bounded-range mechanisms, whose privacy loss lies in an interval of fixed width [20, 14].

The ff-DP framework, which includes GDP, was introduced by [18], although previously the testing perspective has been used to study the composition of differential privacy mechanisms in [22]. [18] show that all popularly used divergence-based guarantees can be computed from ff-DP without any loss. The reverse direction, from divergence bounds to testing errors, is studied by [BBG+20, ALC+21], where the latter also characterize the optimal approximate DP guarantee implied by an RDP bound, and by [33] through intersections of the sets of achievable type I and type II errors. The tightest ff-DP guarantee implied by RDP bounds at all orders, including zCDP, has been characterized using Bernoulli experiments [28], while numerical methods have been developed for identifying and reporting the GDP guarantees of specific mechanisms [23, 19].

Appendix § 6 provides additional context on the equations and formulae from the conversions discussed above.

Privacy definitions and randomized response

Let MM be a Markov kernel (randomized mechanism) from an input measurable space to an output measurable space (Ω,F)(\Omega, \mathcal{F}), with a symmetric neighboring relation ∼\sim on its input space. For datasets of a fixed size, the usual choice is x∼x′x \sim x' when dH(x,x′)=1d_H(x,x') = 1, where dHd_H denotes the Hamming distance. For fixed neighboring datasets x∼x′x \sim x', write P=M(x)P = M(x) and Q=M(x′)Q = M(x'). Let σ(t)=(1+e−t)−1\sigma(t) = (1 + e^{-t})^{-1} denote the logistic function, and let Φ\Phi and ϕ\phi denote the standard normal CDF and density functions respectively.

Privacy guarantees

In this section, we define the two guarantees we compare, first zCDP and then GDP. For probability measures P,QP, Q dominated by a common measure ν\nu, with densities p=dP/dνp = dP/d\nu and q=dQ/dνq = dQ/d\nu, the Rényi divergence of order η>1\eta> 1 is Dη(P∥Q)=1η−1log⁡∫pηq1−η dνD_\eta(P\|Q) = \frac{1}{\eta-1} \log\int p^\eta q^{1-\eta}\,d\nu [VEH14]. Moreover, lim⁡η↓1Dη(P∥Q)\lim_{\eta\downarrow1} D_\eta(P\|Q) results in the KL divergence. Rényi divergences are finite only if Q(A)=0Q(A) = 0 implies P(A)=0P(A) = 0 for every measurable set AA, denoted as the absolute continuity requirement P≪QP \ll Q. We say a mechanism MM satisfies (η,ε)(\eta,\varepsilon)-RDP if Dη(M(x)∥M(x′))≤εD_\eta(M(x)\|M(x')) \le\varepsilon for every ordered neighboring pair [24]. Zero-concentrated DP (zCDP) requires such a bound at all η>1\eta> 1 simultaneously, with ε=ρη\varepsilon= \rho\eta growing linearly in η\eta.

Definition 2 (Zero-concentrated DP [9]) A mechanism MM satisfies ρ\rho-zCDP if for every x∼x′x \sim x' and every η>1\eta> 1, Dη(M(x)∥M(x′))≤ρηD_\eta(M(x)\|M(x')) \le\rho\eta.

This definition is motivated by the Gaussian mechanism. If M(x)=f(x)+ZM(x) = f(x) + Z with Z∼N(0,σ02I)Z \sim N(0,\sigma_0^2 I), then Dη(M(x)∥M(x′))=η∥f(x)−f(x′)∥22/(2σ02)D_\eta(M(x)\|M(x')) = \eta\|f(x)-f(x')\|_2^2/(2\sigma_0^2). Thus, if ff has ℓ2\ell_2-sensitivity Δ\Delta, that is sup⁡x∼x′∥f(x)−f(x′)∥2≤Δ\sup_{x\sim x'}\|f(x)-f(x')\|_2 \le\Delta, then the mechanism satisfies ρ\rho-zCDP with ρ=Δ2/(2σ02)\rho= \Delta^2/(2\sigma_0^2). Intuitively, a mechanism satisfies zCDP if it is “more private” than the Gaussian mechanism, in terms of the Rényi divergences. Smaller ρ\rho means that neighboring output distributions are closer, providing stronger privacy, and ρ\rho-zCDP implies ρ′\rho'-zCDP for any ρ′≥ρ\rho' \ge\rho.

For hypothesis testing, a measurable randomized test ϕ:Ω→[0,1]\phi:\Omega\to[0,1] rejects PP with probability ϕ(ω)\phi(\omega), based on the observation ω\omega. Its type I error and power are EPϕE_P\phi and EQϕE_Q\phi. The trade-off function is

T(P,Q)(α)=inf⁡{ϕ:EPϕ≤α}(1−EQϕ)T(P,Q)(\alpha)=\inf_{\{\phi:\mathbb{E}_{P}\phi\leq\alpha\}}(1-\mathbb{E}_{Q}\phi)

References

  1. [2][AD22] Jordan Awan and Jinshuo Dong. Log-concave and multivariate canonical noise distributions for differential privacy. In Advances in Neural Information Processing Systems, volume 35, pages 34229–34240, 2022.arxiv.org/abs/2206.04572
  2. [4][AR24] Jordan Awan and Aishwarya Ramasethu. Optimizing noise for f-differential privacy via anti-concentration and stochastic dominance. Journal of Machine Learning Research, 25(351):1–32, 2024.arxiv.org/abs/2308.08343
  3. [5][AV23] Jordan Awan and Salil Vadhan. Canonical noise distributions and private hypothesis tests. Annals of Statistics, 51(2):547–572, 2023.arxiv.org/abs/2108.04303
  4. [7][BDKU20] Sourav Biswas, Yihe Dong, Gautam Kamath, and Jonathan Ullman. Coinpress: Practical private mean and covariance estimation. Advances in Neural Information Processing Systems, 33:14475–14485, 2020.arxiv.org/abs/2006.06618
  5. [8][BDLS20] Zhiqi Bu, Jinshuo Dong, Qi Long, and Weijie J. Su. Deep learning with Gaussian differential privacy. Harvard Data Science Review, 2(3), 2020.DOI
  6. [9][BS16] Mark Bun and Thomas Steinke. Concentrated differential privacy: Simplifications, extensions, and lower bounds. In Theory of Cryptography Conference, pages 635–658. Springer, 2016.arxiv.org/abs/1605.02065
  7. [10][BW18] Borja Balle and Yu-Xiang Wang. Improving the Gaussian mechanism for differential privacy: Analytical calibration and optimal denoising. In International Conference on Machine Learning, pages 394–403. PMLR, 2018.arxiv.org/abs/1805.06530
  8. [11][Car72] B. C. Carlson. The logarithmic mean. The American Mathematical Monthly, 79(6):615–618, 1972.DOI
  9. [12][CKS20] Clément L. Canonne, Gautam Kamath, and Thomas Steinke. The discrete Gaussian for differential privacy. In Advances in Neural Information Processing Systems, 2020. See also the expanded arXiv version for Proposition 12 and Corollary 13.DOI
  10. [13][CPA26] Yu-Wei Chen, Raghu Pasupathy, and Jordan Awan. Near-optimal private tests for simple and MLR hypotheses. arXiv preprint arXiv:2601.21959, 2026.arxiv.org/abs/2601.21959
  11. [14]Jinshuo Dong, David Durfee, and Ryan Rogers. Optimal differential privacy composition for exponential mechanisms. In Proceedings of the 37Th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 2597–2606, 2020.
  12. [15]Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity in private data analysis. In Theory of Cryptography Conference, pages 265–284. Springer, 2006.
  13. [16]Jinshuo Dong. Gaussian differential privacy and related techniques. PhD thesis, University of Pennsylvania, 2020.
  14. [17]Cynthia Dwork and Aaron Roth. The algorithmic foundations of differential privacy. Foundations and Trends in Theoretical Computer Science, 9(3–4):211–407, 2014.
  15. [18]Jinshuo Dong, Aaron Roth, and Weijie J Su. Gaussian differential privacy. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 84(1):3–37, 2022.arxiv.org/abs/1905.02383
  16. [19]Juan Felipe Gomez, Bogdan Kulynych, Georgios Kaissis, Flavio P. Calmon, Jamie Hayes, Borja Balle, and Antti Honkela. Gaussian DP for reporting differential privacy guarantees in machine learning. In IEEE Conference on Secure and Trustworthy Machine Learning (Satml), 2026. Preprint arXiv:2503.10945, version 3, June 2026; first posted in 2025.arxiv.org/abs/2503.10945
  17. [20]Charlie Harrison and Pasin Manurangsi. Exact zCDP characterizations for fundamental differentially private mechanisms. In 7Th Symposium on Foundations of Responsible Computing (Forc 2026), volume 368 of Leibniz International Proceedings in Informatics (Lipics), 2026. Article 3; preprint first posted in 2025.
  18. [21]Gautam Kamath, Jerry Li, Vikrant Singhal, and Jonathan Ullman. Privately learning high-dimensional distributions. In Proceedings of the 32Nd Conference on Learning Theory, volume 99 of Proceedings of Machine Learning Research, pages 1853–1902, 2019.arxiv.org/abs/1805.00216
  19. [22]Peter Kairouz, Sewoong Oh, and Pramod Viswanath. The composition theorem for differential privacy. In International Conference on Machine Learning, pages 1376–1385. PMLR, 2015.
  20. [23]Yi Liu, Ke Sun, Bei Jiang, and Linglong Kong. Identification, amplification and measurement: A bridge to Gaussian differential privacy. In Advances in Neural Information Processing Systems, volume 35, pages 11410–11422, 2022.arxiv.org/abs/2210.09269
  21. [24]Ilya Mironov. Rényi differential privacy. In 30Th IEEE Computer Security Foundations Symposium (CSF), pages 263–275. IEEE, 2017.DOI
  22. [25]National Institute of Standards and Technology. NIST digital library of mathematical functions. Section 4.36, Equation 4.36.2, 2024. Infinite Products and Partial Fractions.
  23. [26]Aaradhya Pandey. Differential privacy and machine unlearning: A hypothesis testing approach. PhD thesis, Princeton University, 2026.
  24. [27]Aaradhya Pandey, Arian Maleki, and Sanjeev Kulkarni. Infinitely divisible privacy and beyond I: Resolution of the s² = 2k conjecture, 2025.arxiv.org/abs/2512.00734
  25. [28][RGC+26] Anneliese Riess, Juan Felipe Gomez, Flavio du Pin Calmon, Julia Anne Schnabel, and Georgios Kaissis. Optimal conversion from Rényi differential privacy to f-differential privacy. arXiv preprint arXiv:2602.04562, 2026. Version 3, May 29, 2026.arxiv.org/abs/2602.04562
  26. [29][Ste24] Thomas Steinke. Tight RDP & zCDP bounds from pure DP. DifferentialPrivacy.org, May 2024. Published May 27, 2024.
  27. [30][Tor91] Erik Torgersen. Comparison of Statistical Experiments, volume 36 of Encyclopædia of Mathematics and its Applications. Cambridge University Press, March 1991. Cambridge Core digital edition ISBN 9780511666353; paperback ISBN 9780521102827.DOI
  28. [31][vEH14] Tim van Erven and Peter Harremoës. Rényi divergence and Kullback–Leibler divergence. IEEE Transactions on Information Theory, 60(7):3797–3820, 2014.
  29. [32][WCA25] Zhanyu Wang, Guang Cheng, and Jordan Awan. Differentially private bootstrap: New privacy analysis and inference strategies. Journal of Machine Learning Research, 26, 2025.
  30. [33][ZDW22] Yuqing Zhu, Jinshuo Dong, and Yu-Xiang Wang. Optimal accounting of differential privacy via characteristic function. In International Conference on Artificial Intelligence and Statistics, pages 4782–4817. PMLR, 2022.arxiv.org/abs/2106.08567

Paper details

vlnc:W_jymqqmqc
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Wolfram|Alpha, ChatGPT

View statement
Looking for related papers…
Contents