pith. sign in

arxiv: 2605.23879 · v1 · pith:NCX2YHC5new · submitted 2026-05-22 · 📊 stat.ML · cs.CR· cs.LG· math.ST· stat.TH

On the Stability of Spherical Hellinger-Kantorovich Flows and Their Implications for Differential Privacy

Aratrika Mustafi , Soumya Mukherjee This is my paper

Pith reviewed 2026-05-25 02:52 UTC · model grok-4.3

classification 📊 stat.ML cs.CRcs.LGmath.STstat.TH
keywords spherical Hellinger-Kantorovich flowsperturbation theorydifferential privacyexponential mechanismRényi divergenceKL divergencebirth-death Langevin dynamicsgradient flows
0
0 comments X

The pith

A uniform perturbation bound on spherical Hellinger-Kantorovich flows yields dimension-free control of log-likelihood ratios and Rényi divergences for differential privacy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a perturbation theory for SHK gradient flows that couple transport and reaction and coincide with birth-death Langevin dynamics. For two potentials, it compares the flows started from the same initialization and tracks how discrepancies in the potentials propagate along the trajectories. The resulting uniform bound controls the pointwise log-likelihood ratio and Rényi divergence independently of dimension. These controls are then used to obtain explicit time-dependent privacy guarantees when the flows approximate sampling from the exponential mechanism.

Core claim

For two potentials V and V', the associated SHK flows from a common initialization are compared to quantify propagation of potential discrepancies over time. A uniform perturbation bound yields dimension-free, pointwise control of the log-likelihood ratio and Rényi divergence, while additional structure allows bounds for the KL divergence as well. These results are applied to approximate sampling for the exponential mechanism, providing time-dependent Pure-DP guarantees via the likelihood-ratio control and Approximate-DP certificates via hockey-stick divergence, together with a utility bound that separates intrinsic suboptimality from finite-time sampling error.

What carries the argument

The uniform perturbation bound on SHK gradient flows, which quantifies how discrepancies between two potentials propagate from a shared initialization into controlled differences between the evolving measures.

If this is right

  • Explicit time-dependent Pure-DP guarantees for SHK-based samplers follow directly from the likelihood-ratio control.
  • The KL bound supplies Approximate-DP certificates through the hockey-stick divergence.
  • A utility bound separates the intrinsic suboptimality of the exponential mechanism from the error due to finite-time sampling.
  • The dimension-free pointwise control of the log-likelihood ratio and Rényi divergence holds uniformly along the trajectories.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same perturbation approach could be tested on other reaction-diffusion geometries to obtain comparable privacy certificates.
  • The separation of utility into intrinsic and sampling components allows direct optimization of the stopping time for a target privacy level.
  • Because the bounds are dimension-free, they remain informative even when the underlying data dimension grows large.

Load-bearing premise

The perturbation analysis is assumed to extend from the common initialization to the full trajectory without additional regularity conditions that would break dimension-freeness.

What would settle it

A numerical experiment in which the observed pointwise log-likelihood ratio between two SHK flows for concrete potentials exceeds the dimension-free upper bound predicted by the perturbation analysis at some positive time.

Figures

Figures reproduced from arXiv: 2605.23879 by Aratrika Mustafi, Soumya Mukherjee.

Figure 1
Figure 1. Figure 1: Experiment 1 For the plot in the right panel of [PITH_FULL_IMAGE:figures/full_fig_p038_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Experiment 2 is essentially minimized, but the finite-time t = 2 value is not minimized. The experiment shows that c = 1/2 is the correct asymptotic balancing choice, but a finite-time certificate may benefit from optimizing c numerically for the runtime of interest. This is a useful practical note for using the KL theorem in certificates. Also, the large value of the KL bound at finite time is due to a co… view at source ↗
Figure 3
Figure 3. Figure 3: Experiment 3 Experiment 4: This experiment compares SHK/WFR against the ordinary Langevin/Fokker￾Planck flow on a metastable bimodal target. The goal is to show that SHK’s reaction term can give much faster convergence to the target, which improves the utility side of the privacy-utility story. The base potential is V (x) = 2.5(1 − cos(2x)). The perturbation is V ′ (x) = V (x) + 0.35 sin x. The experiment … view at source ↗
Figure 4
Figure 4. Figure 4: At t = 6, the SHK pointwise log-ratio is 0.362 , below the SHK theorem envelope 0.414, demonstrating that the SHK theorem bound remains valid and reasonably tight. The stability of SHK and Langevin is similar in this particular perturbation diagnostic, but SHK converges to the target much faster and achieves a much lower value of the KL divergence from the Gibbs target as well, since SHK reaches KL(ρt∥π) =… view at source ↗
Figure 4
Figure 4. Figure 4: Experiment 4 42 [PITH_FULL_IMAGE:figures/full_fig_p042_4.png] view at source ↗
read the original abstract

Gradient-flow sampling interprets a Gibbs distribution as the minimizer of an energy functional over probability measures and generates dynamics converging to this target. Under spherical Hellinger-Kantorovich (SHK) geometry, the flow couples transport and reaction and coincides with birth-death Langevin dynamics. In this work, we develop a perturbation theory for SHK gradient flows. For two potentials $V$ and $V^{\prime}$, we compare the associated flows from a common initialization and quantify how potential discrepancies propagate over time. A uniform perturbation bound yields dimension-free, pointwise control of the log-likelihood ratio and R\'enyi divergence, while additional structure allows us to derive bounds for the KL divergence as well. We apply these results to approximate sampling for the exponential mechanism in differential privacy. The likelihood-ratio control provides explicit time-dependent Pure-DP guarantees for SHK-based samplers, while the KL bound yields Approximate-DP certificates via hockey-stick divergence. We also derive a utility bound separating intrinsic exponential-mechanism suboptimality from finite-time sampling error.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper develops a perturbation theory for spherical Hellinger-Kantorovich (SHK) gradient flows. For potentials V and V', it compares flows from a common initialization, quantifies discrepancy propagation, and obtains a uniform perturbation bound that yields dimension-free pointwise control of the log-likelihood ratio and Rényi divergence (with additional structure yielding KL bounds). These stability results are applied to SHK-based samplers for the exponential mechanism, producing explicit time-dependent Pure-DP guarantees via likelihood-ratio control, Approximate-DP certificates via hockey-stick divergence from the KL bound, and a utility bound that separates intrinsic suboptimality from finite-time sampling error. The flows are noted to coincide with birth-death Langevin dynamics.

Significance. If the dimension-free bounds are rigorously established without hidden dimension-dependent regularity, the work would supply a useful stability framework for non-Euclidean gradient flows and concrete, time-dependent DP certificates for a class of sampling algorithms. The explicit separation of mechanism suboptimality from discretization error and the link to Rényi/KL control are potentially valuable for private sampling analyses.

major comments (1)
  1. [Perturbation theory and main stability result (section deriving the uniform bound and its propagation)] The central dimension-free claim requires that the uniform perturbation bound (comparing SHK flows for V and V') propagates over the entire trajectory while remaining dimension-free. The abstract invokes coincidence with birth-death Langevin dynamics and continuity equations under the SHK metric, but provides no indication that Gronwall-type estimates or regularity conditions (e.g., uniform Lipschitz constants or curvature bounds on the potentials) are controlled without dimension dependence; such conditions would typically scale with dimension and invalidate the claimed pointwise log-likelihood and Rényi control.
minor comments (2)
  1. Clarify the precise 'additional structure' needed for the KL divergence bound, as this is referenced only in the abstract.
  2. The utility bound separating intrinsic exponential-mechanism suboptimality from finite-time sampling error is mentioned but not detailed in the abstract; a brief statement of its form would aid readability.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading of our manuscript and for recognizing the potential utility of the stability results for non-Euclidean gradient flows and private sampling. We address the major comment below.

read point-by-point responses

  1. Referee: [Perturbation theory and main stability result (section deriving the uniform bound and its propagation)] The central dimension-free claim requires that the uniform perturbation bound (comparing SHK flows for V and V') propagates over the entire trajectory while remaining dimension-free. The abstract invokes coincidence with birth-death Langevin dynamics and continuity equations under the SHK metric, but provides no indication that Gronwall-type estimates or regularity conditions (e.g., uniform Lipschitz constants or curvature bounds on the potentials) are controlled without dimension dependence; such conditions would typically scale with dimension and invalidate the claimed pointwise log-likelihood and Rényi control.

    Authors: We appreciate the referee's concern about ensuring the dimension-freeness throughout the analysis. The perturbation theory in the relevant section derives the uniform bound by considering the evolution of the discrepancy between the two flows under the SHK metric. Because the SHK geometry is defined via a combination of Hellinger distance and Kantorovich transport on the sphere, the resulting continuity equation has coefficients that depend on the gradient of the potential in a normalized manner. The Lipschitz constants for the vector field are bounded using the sup-norm difference of the potentials, which is independent of the dimension by the problem setup. Curvature bounds, when needed, arise from the spherical structure and do not introduce explicit dimension factors. The Gronwall-type estimate is obtained by integrating a differential inequality whose growth rate is controlled solely by these dimension-independent quantities, yielding a bound that propagates uniformly over time without dimension dependence. This underpins the pointwise log-likelihood ratio and Rényi divergence controls. We can add a clarifying paragraph in the revision if the referee finds the current presentation insufficiently explicit on this point. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The provided abstract and claims describe a perturbation analysis for SHK flows comparing two potentials from common initialization, yielding dimension-free bounds on log-likelihood ratios and divergences via the flow dynamics. No equations, self-citations, fitted parameters renamed as predictions, or ansatzes are shown that reduce the central result to its inputs by construction. The derivation chain is presented as independent analysis of the continuity equation and Gronwall-type estimates under the SHK metric, with applications to DP following from those bounds. This is the normal case of a self-contained theoretical development.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; full text would be required to populate the ledger.

pith-pipeline@v0.9.0 · 5726 in / 1155 out tokens · 33921 ms · 2026-05-25T02:52:03.013137+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · 3 internal anchors

  1. [1]

    Mechanism Design via Differential Privacy , year=

    McSherry, Frank and Talwar, Kunal , booktitle=. Mechanism Design via Differential Privacy , year=

    work page

  2. [2]

    Accelerating Langevin Sampling with Birth-death

    Accelerating Langevin sampling with birth-death , author=. arXiv preprint arXiv:1905.09863 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1905

  3. [3]

    arXiv preprint arXiv:2302.00766 , year=

    Privacy Risk for anisotropic Langevin dynamics using relative entropy bounds , author=. arXiv preprint arXiv:2302.00766 , year=

    work page arXiv

  4. [4]

    SIAM journal on mathematical analysis , volume=

    The variational formulation of the Fokker--Planck equation , author=. SIAM journal on mathematical analysis , volume=. 1998 , publisher=

    work page 1998

  5. [5]

    SIAM Journal on Mathematical Analysis , volume =

    Jordan, Richard and Kinderlehrer, David and Otto, Felix , title =. SIAM Journal on Mathematical Analysis , volume =. 1998 , doi =. https://doi.org/10.1137/S0036141096303359 , abstract =

    work page doi:10.1137/s0036141096303359 1998

  6. [6]

    Schlichting, Andr. Poincar. Entropy , volume=. 2019 , publisher=

    work page 2019

  7. [7]

    Advances in Neural Information Processing Systems , volume=

    Privacy of noisy stochastic gradient descent: More iterations without more privacy loss , author=. Advances in Neural Information Processing Systems , volume=

    work page

  8. [8]

    SIAM/ASA Journal on Uncertainty Quantification , volume=

    Accelerate langevin sampling with birth-death process and exploration component , author=. SIAM/ASA Journal on Uncertainty Quantification , volume=. 2025 , publisher=

    work page 2025

  9. [9]

    SIAM Journal on Mathematical Analysis , volume=

    Spherical Hellinger--Kantorovich Gradient Flows , author=. SIAM Journal on Mathematical Analysis , volume=. 2019 , publisher=

    work page 2019

  10. [10]

    1986 , publisher=

    Logarithmic Sobolev inequalities and stochastic Ising models , author=. 1986 , publisher=

    work page 1986

  11. [11]

    Inventiones mathematicae , volume=

    Optimal entropy-transport problems and a new Hellinger--Kantorovich distance between positive measures , author=. Inventiones mathematicae , volume=. 2018 , publisher=

    work page 2018

  12. [12]

    An Interpolating Distance between Optimal Transport and Fisher-Rao

    An interpolating distance between optimal transport and Fisher-Rao , author=. arXiv preprint arXiv:1506.06430 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    IEEE Transactions on Information Theory , volume=

    f -divergence Inequalities , author=. IEEE Transactions on Information Theory , volume=. 2016 , publisher=

    work page 2016

  14. [14]

    Foundations and trends

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

    work page 2014

  15. [15]

    Advances in neural information processing systems , volume=

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

    work page

  16. [16]

    Hockey-Stick GAN , author=

    work page

  17. [17]

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

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

    work page 2016

  18. [18]

    Advances in Neural Information Processing Systems , volume=

    Differential privacy dynamics of langevin diffusion and noisy gradient descent , author=. Advances in Neural Information Processing Systems , volume=

    work page

  19. [19]

    Faster differentially private samplers via R

    Ganesh, Arun and Talwar, Kunal , journal=. Faster differentially private samplers via R

    work page

  20. [20]

    arXiv preprint arXiv:2204.01585 , year=

    Differentially private sampling from rashomon sets, and the universality of langevin diffusion for convex optimization , author=. arXiv preprint arXiv:2204.01585 , year=

    work page arXiv

  21. [21]

    2017 IEEE International Symposium on Information Theory (ISIT) , pages=

    Information and estimation in Fokker-Planck channels , author=. 2017 IEEE International Symposium on Information Theory (ISIT) , pages=. 2017 , organization=

    work page 2017

  22. [22]

    arXiv preprint arXiv:2204.06106 , year=

    Optimal membership inference bounds for adaptive composition of sampled gaussian mechanisms , author=. arXiv preprint arXiv:2204.06106 , year=

    work page arXiv

  23. [23]

    2019 IEEE symposium on security and privacy (SP) , pages=

    Exploiting unintended feature leakage in collaborative learning , author=. 2019 IEEE symposium on security and privacy (SP) , pages=. 2019 , organization=

    work page 2019

  24. [24]

    Mathematical Notes , volume=

    The Kantorovich and variation distances between invariant measures of diffusions and nonlinear stationary Fokker-Planck-Kolmogorov equations , author=. Mathematical Notes , volume=. 2014 , publisher=

    work page 2014

  25. [25]

    Bogachev and M

    V.I. Bogachev and M. Röckner and S.V. Shaposhnikov , keywords =. Distances between transition probabilities of diffusions and applications to nonlinear Fokker–Planck–Kolmogorov equations , journal =. 2016 , issn =. doi:https://doi.org/10.1016/j.jfa.2016.05.016 , url =

    work page doi:10.1016/j.jfa.2016.05.016 2016

  26. [26]

    Texts in applied mathematics , volume=

    Stochastic processes and applications , author=. Texts in applied mathematics , volume=. 2014 , publisher=

    work page 2014

  27. [27]

    Mattingly and A.M

    J.C. Mattingly and A.M. Stuart and D.J. Higham , keywords =. Ergodicity for SDEs and approximations: locally Lipschitz vector fields and degenerate noise , journal =. 2002 , issn =. doi:https://doi.org/10.1016/S0304-4149(02)00150-3 , url =

    work page doi:10.1016/s0304-4149(02)00150-3 2002

  28. [28]

    Gaussian Approximations of Small Noise Diffusions in Kullback-Leibler Divergence

    Gaussian approximations of small noise diffusions in Kullback-Leibler divergence , author=. arXiv preprint arXiv:1605.05878 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    Malrieu, F. , year=. Logarithmic Sobolev inequalities for some nonlinear PDE’s , volume=. Stochastic Processes and their Applications , publisher=. doi:10.1016/s0304-4149(01)00095-3 , number=

    work page doi:10.1016/s0304-4149(01)00095-3

  30. [30]

    Ledoux, M. , year=. Logarithmic Sobolev Inequalities for Unbounded Spin Systems Revisited , ISBN=. doi:10.1007/978-3-540-44671-2_13 , booktitle=

    work page doi:10.1007/978-3-540-44671-2_13

  31. [31]

    Concentration of measure and logarithmic Sobolev inequalities , ISBN=

    Ledoux, Michel , year=. Concentration of measure and logarithmic Sobolev inequalities , ISBN=. doi:10.1007/bfb0096511 , booktitle=

    work page doi:10.1007/bfb0096511

  32. [32]

    Symmetric Markov Diffusion Operators , ISBN=

    Bakry, Dominique and Gentil, Ivan and Ledoux, Michel , year=. Symmetric Markov Diffusion Operators , ISBN=. doi:10.1007/978-3-319-00227-9_3 , booktitle=

    work page doi:10.1007/978-3-319-00227-9_3