Pith. sign in

REVIEW 3 major objections 7 minor 33 references

Laplace Transform Interpretation of Differential Privacy

T0 review · 3 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper shows that a mechanism's privacy profile and Rényi divergence curve determine each other through Laplace and inverse-Laplace transforms, yielding an exactly tight adaptive composition theorem for (ε,δ)-DP for every ε.

desk verdict A useful Laplace-transform toolkit for DP, but the headline exactly-tight composition theorem is unproven as written for δ>0 because the proof applies a distributional identity that drops the +∞ atom of the randomized-response privacy loss. read the letter →

arxiv 2411.09142 v1 pith:5CCSIGMU submitted 2024-11-14 cs.LG cs.CR

Rishav Chourasia , Uzair Javaid , Biplap Sikdar This is my paper · ORCID
classification cs.LGcs.CR
keywords differentialprivacyLaplacetransformprofileRényidivergenceadaptivecompositionlossdistributionrandomizedresponsefunctional
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

This paper argues that the privacy profile and the Rényi divergence curve are two views of the same object: the privacy loss distribution, read through its bilateral Laplace transform. Recognizing familiar integral expressions as Laplace transforms lets the paper connect the two curves exactly, show that the Rényi order $q$ can be complex, and derive a composition rule for privacy profiles that is exactly tight. The main payoff is an adaptive composition theorem for $(\varepsilon,\delta)$-DP: any sequence of mechanisms each satisfying $(\varepsilon_i,\delta_i)$-DP satisfies a continuous curve $\delta_{\otimes k}(\varepsilon)$ given by a simple recursion, reproducing the known optimal discrete guarantees and filling in the curve between them. The paper also argues that symmetrizing $f$-DP under subsampling breaks equivalence among functional notions, and proposes keeping the natural asymmetry instead.

What carries the argument

The machinery is the bilateral Laplace transform of the privacy loss distribution (PLD), the distribution of $Z = \log(P(\Theta)/Q(\Theta))$ for $\Theta\sim P$. The paper reads the standard expectation formulas for the privacy profile and the Rényi exponential as $\mathcal{B}\{f_Z\}(s)$ evaluated at specific complex frequencies, then uses Laplace uniqueness, convolution, and derivative properties to move between time-domain curves and frequency-domain curves. The load-bearing identity is $e^{(q-1)R_q} = q(q-1)\mathcal{B}\{\delta_{P|Q}\}(1-q)$; its inverse form $\delta_{P|Q}(\varepsilon) = \mathcal{L}^{-1}\{e^{-sR_{1-s}(P\|Q)}/(s(s-1))\}(\varepsilon)$ is what turns a Rényi curve back into a privacy profile.

What would settle it

For two independent randomized-response mechanisms, each $(0.1,10^{-8})$-DP, compute the exact privacy profile of their product by direct convolution of the privacy loss distributions (which is well-defined even though the Laplace transforms diverge) and compare it pointwise with the recursive $\delta_{\otimes 2}(t)$ from Theorem 4.6; if the recursion is not pointwise equal to this exact profile, the claim of constant-tightness for $\delta_i>0$ is false.

Watch

Extended reading notes

Core claim

The paper's central claim is that the privacy profile $\delta_{P|Q}(\varepsilon)$ and the Rényi divergence curve $R_q(P\|Q)$ are the same information viewed in the time and frequency domains. Through Laplace-transform identities, $\delta_{P|Q}(\varepsilon) = \mathcal{L}\{1-F_Z(t+\varepsilon)\}(1)$ and $e^{(q-1)R_q(P\|Q)} = \mathcal{B}\{f_Z\}(1-q)$, the two are shown to invert one another via $e^{(q-1)R_q(P\|Q)} = q(q-1)\mathcal{B}\{\delta_{P|Q}(t)\}(1-q)$, so that $R_q$ exists for complex orders on the region of convergence and $\delta_{P|Q}$ can be recovered by inverse Laplace transform. On composition, the privacy profile of a product distribution is the convolution $\delta_{P|Q} = \delta_{P_1|Q_1} \circledast (\ddot{\delta}_{P_2|Q_2} - \dot{\delta}_{P_2|Q_2})$, which yields the recursive exactly tight composition bound $\delta_{\otimes l}(t) = \delta_l + \frac{1-\delta_l}{e^{\varepsilon_l}+1}[e^{\varepsilon_l}\delta_{\otimes l-1}(t-\varepsilon_l) + \delta_{\otimes l-1}(t+\varepsilon_l)]$ for any sequence of $(\varepsilon_i,\delta_i)$-DP mechanisms. The paper also claims that preserving the natural asymmetry in $f$-DP subsampling, rather than symmetrizing, keeps all functional DP notions equivalent.

Load-bearing premise

The load-bearing premise is that the convolution identity for privacy profiles remains valid for randomized-response profiles with $\delta_i>0$, where the Laplace transforms diverge; Theorem 4.2 is stated under absolute continuity in at least one direction, and the paper notes in Remark 4.3 that formally proving this extension appears challenging.

Editorial extensions

If this is right

  • The recursive bound $\delta_{\otimes k}(\varepsilon)$ is a continuous curve, so for any target $\delta$ budget the tightest $\varepsilon$ can be read off directly, whereas the earlier optimal discrete theorem only gives a finite set of points.
  • The same composition is exactly tight for adaptive and heterogeneous mechanisms under the standard worst-case conditional-profile domination assumption.
  • For homogeneous composition, the closed form $\delta_{\otimes k}(t) = 1-(1-\delta)^k(1 - \mathbb{E}_{Y\sim\mathrm{Binomial}(k,p)}[1 - e^{t-\varepsilon(2Y-k)}]_+)$ gives an analytical accountant.
  • Complex-order Rényi divergence is well-defined on the region of convergence, so the privacy profile can be reconstructed from the Rényi curve along a vertical line in the complex plane rather than only from real orders $q>1$.
  • Keeping asymmetry in functional DP avoids the slack introduced by symmetrization and keeps $f$-DP, privacy profiles, privacy loss distributions, and Rényi curves equivalent under subsampling.

Reading between the lines

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

  • If the distributional step for $\delta_i>0$ is formalized, the Laplace toolkit should extend to exact composition of arbitrary functional curves beyond the randomized-response worst case, potentially replacing numerical PLD accountants with closed-form recursions.
  • The failure of Rényi dominance to imply privacy-profile dominance suggests that accounting should be done in the privacy-profile domain whenever the final guarantee is $(\varepsilon,\delta)$, because converting through real-order Rényi bounds provably loses tightness.
  • A testable extension is to use the inverse-Laplace formula with complex-order Rényi curves as a definition for profiles that are not absolutely continuous, and check whether it reproduces known profiles for mixtures such as subsampled mechanisms.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. This paper proposes interpreting privacy profiles δ_{P|Q}(ε) and Rényi divergence R_q(P||Q) as (two-sided) Laplace transforms of the privacy loss distribution. Theorems 3.2 and 3.3 record identities connecting δ, R_q, and E_q; Eq. (37) gives an inverse-Laplace expression for δ in terms of complex-order Rényi divergence. Section 4 derives Theorem 4.2, an exactly tight convolution-type composition rule for privacy profiles under product distributions, and uses it to prove Theorem 4.6, a recursive adaptive composition theorem for (ε_i,δ_i)-DP point guarantees claimed to match Kairouz et al. in constants and to yield a continuous optimal curve. Section 5 argues that symmetrization in f-DP subsampling breaks equivalence across functional DP notions and proposes a skew convention.

Significance. If the main results were fully proved, the paper would make a useful contribution: it unifies several DP formalisms through standard transform calculus, extends Rényi divergence to complex orders in a natural ROC-based way, and offers a closed-form continuous composition curve that aligns with numerical accountants. The identities in Theorems 3.2 and 3.3 are standard integration-by-parts calculations and check out for absolutely continuous pairs; the recursive formula (62) is explicit, has no fitted parameters, and is directly checkable against numerical PLD accountants, which is a strength. However, the proof of the central composition theorem for the practically relevant δ_i>0 case rests on an unsupported and in fact false extension of Theorem 4.2 to non-absolutely-continuous profiles. The contribution is therefore conditional on a repaired proof; the final recursion is likely correct, but the manuscript as written does not prove it.

major comments (3)
  1. [Appendix A.3, proof of Theorem 4.6, Eqs. (159)-(177)] The induction step applies Theorem 4.2 to the randomized-response profile δ_{ε_l,δ_l}^{RR} with δ_l>0. This violates Theorem 4.2's hypothesis, which requires absolute continuity in at least one direction, and Remark 4.3 concedes that the extension is unproved. The extension is in fact false: the PLD of M_{ε,δ}^{RR} has a point mass δ at +∞, whereas the kernel δ̈_{RR}−δ̇_{RR} computed in Eqs. (163)-(177) contains only the two finite atoms (1−δ)e^ε/(e^ε+1) δ_{t=ε} and (1−δ)/(e^ε+1) δ_{t=−ε}, with the J2 and J3 terms vanishing under convolution. Consequently inequality (160) is false: with ε=1, δ=0.1 and δ⊗0(t)=[1−e^t]_+, the right-hand side at t=0 equals 0.416, while the actual randomized-response profile from Theorem 3.4 gives δ_RR(0)=0.516, the difference being exactly δ. The proof of Theorem 4.6 therefore does not establish recursion (62) for δ_i>0. The recursion itself is consistent with the PLD-convolution recurrence that includes the +∞ atom and with Kairouz et al.'s optimal composition, so a direct time-domain proof is likely to salvage the theorem, but it is not in the manuscript.
  2. [Section 3.1, proof of Theorem 3.6] The second part of Theorem 3.6 is proved by visual inspection of Figure 1: the text says 'From the leftmost plot, we can see...' and 'their privacy profiles cross one another.' A logical implication should not rest on a plot; the authors should provide a closed-form expression for the crossing point or an analytic inequality. This is a local proof gap rather than a challenge to the conclusion, but it should be fixed for the theorem to be considered proved.
  3. [Section 4 and Remark 4.3] The paragraph after Theorem 4.2 states that 'even when the Laplace transform is undefined everywhere, the frequency-domain manipulations performed on it still correspond to valid manipulation steps in the time domain.' This assertion is the entire basis for applying Theorem 4.2 to randomized-response profiles with δ_i>0, and it is contradicted by the concrete counterexample above: the ordinary distributional derivative of δ_RR carries no atom at +∞, so the missing point mass of the PLD is lost. The paper should either remove this assertion and prove Theorem 4.6 by a separate time-domain argument, or restrict Theorem 4.6 to the case where the relevant Laplace transforms have nonempty ROC.
minor comments (7)
  1. [Throughout] Typos and name errors should be corrected: 'Heaveside' should be 'Heaviside' (Appendix A.3), 'Frenchel duality' should be 'Fenchel duality' (Section 4.1), 'Kairozu' should be 'Kairouz' (Figure 2 caption), and 'Oppenhiem' should be 'Oppenheim' (references).
  2. [Section 3, Eq. (37)] The inverse-Laplace formula writes the exponential factor as esε; this should be e^{sε} or exp(sε) to avoid ambiguity with a product of e, s, and ε.
  3. [Section 3, Definition 3.1 and Theorem 3.2] The 'generalized density' f_X is defined informally through Dirac deltas. Since later proofs perform distributional differentiation and convolution of such objects, the authors should state precisely in what space these manipulations are carried out and cite the relevant distributional calculus.
  4. [Section 3, Theorem 3.5 proof] The extension from real q to complex orders is justified only by the sentence 'From dominated convergence theorem the theorem statement holds for complex orders as well.' This is too terse; the authors should invoke analyticity of both sides on the ROC strip or give a dominated-convergence argument with the strip made explicit.
  5. [Section 4.1, comparison claims] The text says the bound 'surpasses the optimal composition result in Kairouz et al.' Because the recursion is claimed to match the existing discrete optimal points and merely adds intermediate ε values, 'surpasses' is misleading; 'extends to a continuous curve while matching the discrete optimal points' would be accurate once a proof is supplied.
  6. [Remark 4.1] The statement that Gopi et al. 'seem to incorrectly assert their Theorem 5.5 to be valid under adaptivity' is a strong claim about a published result and is not substantiated. The authors should either provide a precise counterexample or soften the remark.
  7. [Figures 2 and 4] The comparison with Google's PLDAccountant and Microsoft's PRVAccountant reports numerical gaps, but the manuscript does not state the discretization parameters or provide code. Adding this information would make the numerical comparisons reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Laplace identities are exact reformulations; the admitted gap in Remark 4.3 is a proof/correctness issue, not a circular step.

full rationale

No significant circularity. The Laplace-transform identities are exact mathematical rewritings of the PLD definitions: the privacy profile delta_{P|Q}(eps)=E[(1-e^{eps-Z})_+] is presented as a Laplace transform of the survival function, and Renyi divergence is the log of the bilateral Laplace transform of the PLD density (Eqs. (1)-(2), Theorem 3.2). Theorem 3.3 derives the delta-to-Renyi equivalence by linearity, shifting, reversal, and derivative properties of the Laplace transform; it does not assume the claimed equivalence as an input. The composition theorem (Theorem 4.2) is proved in the absolutely-continuous case from convolution of PLDs and Renyi additivity, then converted back through the same transform identities; the only external result of this type is the dominating randomized-response profile from Kairouz et al. (Theorem 4.5), which is an independent, standard result and is not a self-citation. No parameter is fitted, and no output is a renamed fit. The paper itself flags in Remark 4.3 that the extension of Theorem 4.2 to non-absolutely-continuous profiles, which is needed for delta_i>0 in Theorem 4.6, is not formally proved; this is an admitted proof gap and a possible correctness issue, but it is not circular because the recursive composition formula (62) is not shown to be an input of the derivation. The theorem is benchmarked externally against Kairouz et al. and against numerical PLD accountants, so the central claim has independent content. Hence circularity score 0.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

All structural assumptions are standard probability and Laplace-transform machinery. The main extra weight is the unproven distributional extension for non-absolutely-continuous privacy profiles and the asserted complex-order Rényi branch choice. No free parameters are fitted and no new entities are introduced.

assumptions (6)
  • standard math Privacy loss random variables of product distributions compose as an independent sum, so their densities convolve.
    Used in the proof of Theorem 4.2, equation (49); follows from the product structure of P and Q.
  • domain assumption Generalized densities and distributional derivatives of privacy profiles can be manipulated under one-sided and two-sided Laplace transforms, including Dirac impulses at discontinuities.
    Used in Theorem 4.2 via property (73) and in the induction proof of Theorem 4.6 where δ_RR has kinks at ±ε; requires distributional calculus that is not fully formalized in the paper.
  • domain assumption Zhu et al. [32, Theorem 27] adaptive domination lemma holds.
    Invoked as Lemma 4.4 to extend product composition to adaptive composition; not proven inside this paper.
  • domain assumption Steinke [30, Corollary 24] decomposition of any (ε,δ)-DP pair as post-processing of randomized response holds.
    Used in the proof of Theorem 4.5 through Lemma A.1 to show that randomized response dominates all (ε,δ)-DP pairs in privacy profile.
  • ad hoc to paper Laplace transform uniqueness and inversion remain valid for privacy profiles even when the ROC is empty; formal frequency-domain manipulations are valid in the time domain.
    Needed to apply Theorem 4.2 to randomized response with δ>0 despite no absolute continuity and a divergent Laplace transform; the authors flag this as unresolved in Remark 4.3.
  • ad hoc to paper Rényi divergence for complex orders q=γ+iω is well-defined via a consistent branch of log on the region of convergence.
    After Theorem 3.3 the paper asserts existence for imaginary orders; no analysis is given for zeros or branch cuts of E_q=∫P^qQ^{1-q}dθ.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Laplace Transform Interpretation of Differential Privacy." pith.science (2026). https://pith.science/paper/5CCSIGMU

@misc{pith2026241109142,
  author       = {Pith},
  title        = {Pith review of: Laplace Transform Interpretation of Differential Privacy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5CCSIGMU}},
  note         = {Machine review of arXiv:2411.09142}
}
abstract

We introduce a set of useful expressions of Differential Privacy (DP) notions in terms of the Laplace transform of the privacy loss distribution. Its bare form expression appears in several related works on analyzing DP, either as an integral or an expectation. We show that recognizing the expression as a Laplace transform unlocks a new way to reason about DP properties by exploiting the duality between time and frequency domains. Leveraging our interpretation, we connect the $(q, \rho(q))$-R\'enyi DP curve and the $(\epsilon, \delta(\epsilon))$-DP curve as being the Laplace and inverse-Laplace transforms of one another. This connection shows that the R\'enyi divergence is well-defined for complex orders $q = \gamma + i \omega$. Using our Laplace transform-based analysis, we also prove an adaptive composition theorem for $(\epsilon, \delta)$-DP guarantees that is exactly tight (i.e., matches even in constants) for all values of $\epsilon$. Additionally, we resolve an issue regarding symmetry of $f$-DP on subsampling that prevented equivalence across all functional DP notions.

Figures

Figures reproduced from arXiv: 2411.09142 by the authors.

Figure 1
Figure 1. Comparison between the indistinguishability characteristic of Gaussian mechanism of The [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Comparison of (ε, δ)-DP bounds from various composition theorems for k-fold composition of a (0.1, 10−8 )-DP point guarantee, with the budget constraint δ < 10−6 . The spikes in the right plot, showing exact δ < 10−6 values from Kairouz et al. [19, Theorem 3.3], occur because, out of a set of ⌊k/2⌋ DP point guarantees by their result, we select the smallest ε corresponding to the largest δ < 10−6 in the set, which f… view at source ↗
Figure 3
Figure 3. Visualization of the four functional notions of DP, namely privacy profile [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of (ε, δ)-DP bounds between numerical accountants and our Theorem 4.7 for 100-fold composition of a (0.1, 10−10)-DP point guarantee, with the budget constraint δ < 10−8 . Note that at k = 100, the constraint on δ cannot be satisfied and so the corresponding …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 28 canonical work pages

  1. [1]

    Abadi, A

    M. Abadi, A. Chu, I. Goodfellow, H. B. McMahan, I. Mironov, K. Talwar, and L. Zhang. Deep learning with differential privacy. InProceedings of the 2016 ACM SIGSAC conference on computer and communications security, pages 308–318, 2016

    work page 2016

  2. [2]

    Asoodeh, J

    S. Asoodeh, J. Liao, F. P. Calmon, O. Kosut, and L. Sankar. A better bound gives a hundred rounds: Enhanced privacy guarantees via f-divergences. In2020 IEEE International Symposium on Information Theory (ISIT), pages 920–925. IEEE, 2020

    work page 2020

  3. [3]

    Balle and Y.-X

    B. Balle and Y.-X. Wang. Improving the gaussian mechanism for differential privacy: Analytical calibration and optimal denoising. InInternational Conference on Machine Learning, pages 394–403. PMLR, 2018

    work page 2018

  4. [4]

    Balle, G

    B. Balle, G. Barthe, and M. Gaboardi. Privacy amplification by subsampling: Tight analyses via couplings and divergences.Advances in neural information processing systems, 31, 2018

    work page 2018

  5. [5]

    Balle, G

    B. Balle, G. Barthe, and M. Gaboardi. Privacy profiles and amplification by subsampling. Journal of Privacy and Confidentiality, 10(1), 2020

    work page 2020

  6. [6]

    Balle, G

    B. Balle, G. Barthe, M. Gaboardi, J. Hsu, and T. Sato. Hypothesis testing interpretations and renyi differential privacy. InInternational Conference on Artificial Intelligence and Statistics, pages 2496–2506. PMLR, 2020

    work page 2020

  7. [7]

    Bassily, A

    R. Bassily, A. Smith, and A. Thakurta. Private empirical risk minimization: Efficient algorithms and tight error bounds. In2014 IEEE 55th annual symposium on foundations of computer science, pages 464–473. IEEE, 2014

    work page 2014

  8. [8]

    Bun and T

    M. Bun and T. Steinke. Concentrated differential privacy: Simplifications, extensions, and lower bounds. In Theory of Cryptography Conference, pages 635–658. Springer, 2016

    work page 2016

Show all 33 references
  1. [9]

    C. L. Canonne, G. Kamath, and T. Steinke. The discrete gaussian for differential privacy. Advances in Neural Information Processing Systems, 33:15676–15688, 2020

    2020

  2. [10]

    A. M. Cohen.Numerical methods for Laplace transform inversion, volume 5. Springer Science & Business Media, 2007

    2007

  3. [11]

    J. Dong, A. Roth, and W. J. Su. Gaussian differential privacy.arXiv preprint arXiv:1905.02383, 2019

    1905 arXiv

  4. [12]

    Doroshenko, B

    V. Doroshenko, B. Ghazi, P. Kamath, R. Kumar, and P. Manurangsi. Connect the dots: Tighter discrete approximations of privacy loss distributions.arXiv preprint arXiv:2207.04380, 2022

    2022 arXiv

  5. [13]

    C. Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1–12. Springer, 2006

    2006

  6. [14]

    Dwork and G

    C. Dwork and G. N. Rothblum. Concentrated differential privacy. arXiv preprint arXiv:1603.01887, 2016

    2016 arXiv

  7. [15]

    Dwork, G

    C. Dwork, G. N. Rothblum, and S. Vadhan. Boosting and differential privacy. In2010 IEEE 51st annual symposium on foundations of computer science, pages 51–60. IEEE, 2010

    2010

  8. [16]

    Dwork, A

    C. Dwork, A. Roth, et al. The algorithmic foundations of differential privacy.Foundations and Trends® in Theoretical Computer Science, 9(3–4):211–407, 2014. 20

    2014

  9. [17]

    P. P. Dyke and P. Dyke.An introduction to Laplace transforms and Fourier series. Springer, 2001

    2001

  10. [18]

    S. Gopi, Y. T. Lee, and L. Wutschitz. Numerical composition of differential privacy.Advances in Neural Information Processing Systems, 34:11631–11642, 2021

    2021

  11. [19]

    Kairouz, S

    P. Kairouz, S. Oh, and P. Viswanath. The composition theorem for differential privacy. In International conference on machine learning, pages 1376–1385. PMLR, 2015

    2015

  12. [20]

    Koskela, J

    A. Koskela, J. Jälkö, and A. Honkela. Computing tight differential privacy guarantees using fft. In International Conference on Artificial Intelligence and Statistics, pages 2560–2569. PMLR, 2020

    2020

  13. [21]

    N. Li, W. Qardaji, and D. Su. On sampling, anonymization, and differential privacy or, k-anonymization meets differential privacy. InProceedings of the 7th ACM Symposium on Information, Computer and Communications Security, pages 32–33, 2012

    2012

  14. [22]

    J. F. Miller. Moment-generating functions and laplace transforms.Journal of the Arkansas Academy of Science, 4(1):97–100, 1951

    1951

  15. [23]

    I. Mironov. Rényi differential privacy. In 2017 IEEE 30th computer security foundations symposium (CSF), pages 263–275. IEEE, 2017

    2017

  16. [24]

    Murtagh and S

    J. Murtagh and S. Vadhan. The complexity of computing the optimal composition of differential privacy. InTheory of Cryptography Conference, pages 157–175. Springer, 2015

    2015

  17. [25]

    A. V. Oppenhiem, A. S. Willsky, and S. H. Nawab. Signals and systems, 1996

    1996

  18. [26]

    J. Orloff. Uniqueness of laplace transform, 2015

    2015

  19. [27]

    D. Privacy. Tight RDP & zCDP Bounds from Pure DP — differentialprivacy.org.https: //differentialprivacy.org/pdp-to-zcdp/. [Accessed 12-11-2024]

    2024

  20. [28]

    Sason and S

    I. Sason and S. Verdú.f-divergence inequalities.IEEE Transactions on Information Theory, 62 (11):5973–6006, 2016

    2016

  21. [29]

    Sommer, S

    D. Sommer, S. Meiser, and E. Mohammadi. Privacy loss classes: The central limit theorem in differential privacy.Cryptology ePrint Archive, 2018

    2018

  22. [30]

    T. Steinke. Composition of differential privacy & privacy amplification by subsampling.arXiv preprint arXiv:2210.00597, 2022

    2022 arXiv

  23. [31]

    Y.-X. Wang, S. Fienberg, and A. Smola. Privacy for free: Posterior sampling and stochastic gradient monte carlo. In International Conference on Machine Learning, pages 2493–2502. PMLR, 2015

    2015

  24. [32]

    e(q−1)t q − 1 − eqt q #−ε −∞ + 1 1 + eε

    Y. Zhu, J. Dong, and Y.-X. Wang. Optimal accounting of differential privacy via characteristic function. In International Conference on Artificial Intelligence and Statistics, pages 4782–4817. PMLR, 2022. 21 A Appendix A.1 Table of Properties of Laplace Transform Table 1: Prop...

    2022

  25. [33]

    δ⊗l−2(t − lX i=l−1 ε(2 · Yi − 1)) # (183) = δ lX i=1 (1 − δ)i−1 + (1 − δ)l E Yl←Bernoulli(p) Yl−1←Bernoulli(p)

    tells us that we can simulate these two distributions as post-processing of the randomized re- sponse mechanism Mε,δ RR. To see this, imagine thatP = M(D) and Q = M(D′) are the output distributions of some mechanismM. Define another mechanism G : {0, 1} × {⊥, ⊤} →Ω with output...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.