pith. sign in

arxiv: 2508.05462 · v2 · submitted 2025-08-07 · 📊 stat.CO · math.PR

Piecewise Deterministic Sampling for Constrained Distributions

Jo\"el Tatang Demano , Paul Dobson , Konstantinos Zygalakis This is my paper

Pith reviewed 2026-05-19 00:27 UTC · model grok-4.3

classification 📊 stat.CO math.PR
keywords piecewise deterministic Markov processesconstrained samplingmirror mapsconvex setsexact subsamplingunbiased samplingsampling algorithms
0
0 comments X

The pith

A new class of piecewise deterministic Markov processes adapts mirror maps to sample unbiasedly from distributions supported on convex sets.

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

The paper develops piecewise deterministic Markov processes that incorporate mirror maps to handle probability distributions confined to a convex domain. This construction produces samples that remain inside the constraints by design and stay unbiased relative to the target distribution. The same framework supports exact subsampling, avoiding the bias that often appears when data are thinned in other samplers. Readers would care because many statistical models impose hard constraints such as non-negativity or probability simplex membership, and existing methods based on stochastic differential equations frequently struggle with both accuracy and speed in these settings.

Core claim

By adapting the mirror map from convex optimisation to define the deterministic flow and jump mechanism of a PDMP on the convex set M, the resulting process targets the desired distribution pi supported on M, generates samples that exactly respect the boundary of M, remains unbiased, and permits exact subsampling of the underlying data points.

What carries the argument

The mirror map, which transforms the constrained sampling dynamics into an equivalent problem while preserving the target distribution and the unbiasedness and exact-subsampling properties.

If this is right

  • The algorithms produce samples that respect the convex constraints exactly by construction.
  • Exact subsampling is available without introducing bias into the samples.
  • The methods outperform state-of-the-art stochastic differential equation samplers on a range of constrained sampling tasks.
  • Unbiased samples from the target distribution are obtained for any valid mirror map.

Where Pith is reading between the lines

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

  • The exact-subsampling property may reduce computational cost when the data set is large and the constraint set is simple to describe.
  • If mirror maps can be found for additional convex bodies, the same construction could apply to other common constrained domains such as boxes or balls.
  • The approach could be paired with existing PDMP techniques to handle problems that mix constrained and unconstrained variables.

Load-bearing premise

A suitable mirror map exists for the given convex set that correctly transforms the sampling dynamics while preserving the target distribution and enabling unbiasedness together with exact subsampling.

What would settle it

Generate many samples from a simple test distribution such as uniform on the probability simplex using the proposed algorithm and compare the empirical distribution to the known target; systematic deviation or constraint violations would falsify the unbiasedness and constraint-respect claims.

read the original abstract

In this paper, we propose a novel class of Piecewise Deterministic Markov Processes (PDMPs) that are designed to sample from probability distributions $\pi$ supported on a convex set $\mathcal{M}$. This class of PDMPs adapts the concept of a mirror map from convex optimisation to address sampling problems. The corresponding algorithms provide unbiased samples that respect the constraints and, moreover, allow for exact subsampling. We demonstrate the advantages of these algorithms against a range of constrained sampling problems where the proposed algorithms outperform state of the art stochastic differential equation-based methods.

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 / 3 minor

Summary. The manuscript proposes a novel class of Piecewise Deterministic Markov Processes (PDMPs) for sampling from probability distributions π supported on convex sets M. It adapts mirror maps from convex optimization to define the sampling dynamics, claiming that the resulting algorithms produce unbiased samples that respect the constraints and permit exact subsampling. Empirical advantages over state-of-the-art SDE-based methods are demonstrated on a range of constrained sampling problems.

Significance. If the mirror-map construction is shown to preserve the target measure π exactly, the work would constitute a meaningful advance in constrained sampling by extending PDMP methods to handle convex constraints while retaining unbiasedness and exact subsampling. These features could improve efficiency in high-dimensional or large-scale settings where standard SDE approaches struggle with constraints.

major comments (1)
  1. [Section 3 (PDMP construction via mirror map)] The central claim that the PDMP yields unbiased samples respecting the constraints rests on the mirror map ψ (with φ = ∇ψ) preserving the target measure π. However, the change-of-variables formula requires an explicit Jacobian correction by |det D²ψ| (or its inverse) in the transformed density or generator. The manuscript does not appear to include or derive this term in the process definition, which would make the stationary distribution a distorted version of π rather than exactly π.
minor comments (3)
  1. [Abstract] The abstract would benefit from a one-sentence statement of the key technical step (Jacobian adjustment or equivalent) that ensures measure preservation.
  2. [Section 2 (Preliminaries)] Notation for the mirror map and its Hessian should be introduced with a brief reminder of the change-of-variables formula to aid readers unfamiliar with the optimization literature.
  3. [Section 5 (Numerical experiments)] Experimental figures would be clearer if they included error bars or multiple independent runs to support the outperformance claims.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading and constructive feedback. We address the major comment below and will incorporate the necessary clarifications in the revised manuscript.

read point-by-point responses

  1. Referee: [Section 3 (PDMP construction via mirror map)] The central claim that the PDMP yields unbiased samples respecting the constraints rests on the mirror map ψ (with φ = ∇ψ) preserving the target measure π. However, the change-of-variables formula requires an explicit Jacobian correction by |det D²ψ| (or its inverse) in the transformed density or generator. The manuscript does not appear to include or derive this term in the process definition, which would make the stationary distribution a distorted version of π rather than exactly π.

    Authors: We appreciate the referee highlighting this important technical point regarding measure preservation. The mirror map construction in Section 3 transforms the constrained sampling problem into an equivalent unconstrained one in the dual space, with the PDMP dynamics (including event rates and flows) defined to target the pushforward measure. We agree that the manuscript would benefit from an explicit derivation showing how the change-of-variables Jacobian |det D²ψ| is accounted for in the generator or density to ensure the stationary distribution is exactly π. We will add this derivation to Section 3 in the revision, including the correction term where appropriate, to make the unbiasedness argument fully rigorous. revision: yes

Circularity Check

0 steps flagged

No circularity: mirror-map adaptation to PDMPs is externally grounded

full rationale

The paper introduces a class of PDMPs by adapting mirror maps from convex optimization literature to enforce constraints while targeting distribution π. The abstract and description frame this as a direct transfer of an existing optimization tool into a probabilistic sampling setting, with claims of unbiasedness and exact subsampling resting on the transformation properties rather than any self-referential definition or fitted input. No equations or steps reduce by construction to the paper's own outputs, and no load-bearing self-citations or uniqueness theorems imported from the authors' prior work are indicated. The derivation chain remains self-contained against external benchmarks from optimization and PDMP theory.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities are stated in the provided text.

pith-pipeline@v0.9.0 · 5615 in / 1092 out tokens · 42706 ms · 2026-05-19T00:27:26.119350+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages · 1 internal anchor

  1. [1]

    Bayesian computation: a perspective on the current state, and sampling backwards and forwards

    Green, P.J., Łatuszyński, K., Pereyra, M., Robert, C.P. : Bayesian computation: a perspective on the current state, and sampling backwards a nd forwards. arXiv preprint arXiv:1502.01148 (2015)

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  2. [2]

    Leimkuhler, B.a.: Molecular Dynamics, 1st ed. 2015. edn . Molecular Dynamics, vol. 39. Springer, Cham : (2015)

    work page 2015

  3. [3]

    Acta Numerica 19, 451– 559 (2010) https://doi.org/10.1017/S0962492910000061

    Stuart, A.M.: Inverse problems: A bayesian perspective . Acta Numerica 19, 451– 559 (2010) https://doi.org/10.1017/S0962492910000061

    work page doi:10.1017/s0962492910000061 2010

  4. [4]

    The journal of chemical physics 21(6), 1087–1092 (1953)

    Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Te ller, A.H., Teller, E.: Equation of state calculations by fast computing machines. The journal of chemical physics 21(6), 1087–1092 (1953)

    work page 1953

  5. [5]

    Bernoulli 2(4), 341–363 (1996)

    Roberts, G.O., Tweedie, R.L.: Exponential convergence of Langevin distributions and their discrete approximations. Bernoulli 2(4), 341–363 (1996)

    work page 1996

  6. [6]

    Bernoulli 25(4A), 2854–2882 (2019) https://doi.org/10.3150/18-BEJ1073

    Durmus, A., Moulines, É.: High-dimensional Bayesian in ference via the unadjusted Langevin algorithm. Bernoulli 25(4A), 2854–2882 (2019) https://doi.org/10.3150/18-BEJ1073

    work page doi:10.3150/18-bej1073 2019

  7. [7]

    Annals of Applied Probabilit y 27, 1551–1587 (2017)

    Durmus, A., Moulines, E.: Nonasymptotic convergence an alysis for the un- adjusted Langevin algorithm. Annals of Applied Probabilit y 27, 1551–1587 (2017)

    work page 2017

  8. [8]

    Journal of Machine Learning Research 22(242), 1–37 (2021)

    Sanz-Serna, J.M., Zygalakis, K.C.: Wasserstein distan ce estimates for the distri- butions of numerical approximations to ergodic stochastic differential equations. Journal of Machine Learning Research 22(242), 1–37 (2021)

    work page 2021

  9. [9]

    Stochastic Processes and t heir Applications 129(12), 5278–5311 (2019) https://doi.org/10.1016/j.spa.2019.02.016

    Dalalyan, A.S., Karagulyan, A.: User-friendly guarant ees for the Langevin Monte Carlo with inaccurate gradient. Stochastic Processes and t heir Applications 129(12), 5278–5311 (2019) https://doi.org/10.1016/j.spa.2019.02.016

    work page doi:10.1016/j.spa.2019.02.016 2019

  10. [10]

    Journal of the Royal Statistic al Society: Series B (Methodological) 46(3), 353–376 (1984)

    Davis, M.H.: Piecewise-deterministic Markov process es: A general class of non- diffusion stochastic models. Journal of the Royal Statistic al Society: Series B (Methodological) 46(3), 353–376 (1984)

    work page 1984

  11. [11]

    The Annals of Applied Probability 29(4), 2266–2301 (2019)

    Bierkens, J., Roberts, G.O., Zitt, P.-A.: Ergodicity o f the Zig-Zag process. The Annals of Applied Probability 29(4), 2266–2301 (2019)

    work page 2019

  12. [12]

    Journal of the American Statistical Association 113(522), 855–867 (2018)

    Bouchard-Côté, A., Vollmer, S.J., Doucet, A.: The boun cy particle sampler: A nonreversible rejection-free Markov chain Monte Carlo met hod. Journal of the American Statistical Association 113(522), 855–867 (2018)

    work page 2018

  13. [13]

    In: 28 International Conference on Machine Learning, pp

    Bierkens, J., Grazzi, S., Kamatani, K., Roberts, G.: Th e boomerang sampler. In: 28 International Conference on Machine Learning, pp. 908–918 (2020). PMLR

    work page 2020

  14. [14]

    Statistics and Computing 30(3), 721–730 (2020)

    Wu, C., Robert, C.P.: Coordinate sampler: a non-revers ible Gibbs-like MCMC sampler. Statistics and Computing 30(3), 721–730 (2020)

    work page 2020

  15. [15]

    Statist ical Science 33(3), 386–412 (2018)

    Fearnhead, P., Bierkens, J., Pollock, M., Roberts, G.O .: Piecewise deterministic Markov processes for continuous-time Monte Carlo. Statist ical Science 33(3), 386–412 (2018)

    work page 2018

  16. [16]

    Annals of Applied Probability, 726–752 (2000)

    Diaconis, P., Holmes, S., Neal, R.M.: Analysis of a nonr eversible Markov chain sampler. Annals of Applied Probability, 726–752 (2000)

    work page 2000

  17. [17]

    S tatistics and Computing 26(6), 1213–1228 (2016)

    Bierkens, J.: Non-reversible Metropolis-Hastings. S tatistics and Computing 26(6), 1213–1228 (2016)

    work page 2016

  18. [18]

    The Annals of Sta tistics 47(3), 1288– 1320 (2019)

    Bierkens, J., Fearnhead, P., Roberts, G.: The Zig-Zag p rocess and super-efficient sampling for Bayesian analysis of big data. The Annals of Sta tistics 47(3), 1288– 1320 (2019)

    work page 2019

  19. [19]

    SIAM-S ociety for Industrial and Applied Mathematics, Philadelphia, PA, USA (2017)

    Beck, A.: First-Order Methods in Optimization. SIAM-S ociety for Industrial and Applied Mathematics, Philadelphia, PA, USA (2017)

    work page 2017

  20. [20]

    In: Cortes, C., Lawrence, N., Lee, D., Sugiyama , M., Garnett, R

    Bubeck, S., Eldan, R., Lehec, J.: Finite-time analysis of projected Langevin Monte Carlo. In: Cortes, C., Lawrence, N., Lee, D., Sugiyama , M., Garnett, R. (eds.) Advances in Neural Information Processing System s, vol. 28. Curran Associates, Inc., Red Hook, NY, USA (2015)

    work page 2015

  21. [21]

    S IAM Journal on Imaging Sciences 11(1), 473–506 (2018)

    Durmus, A., Moulines, E., Pereyra, M.: Efficient Bayesia n computation by prox- imal Markov chain Monte Carlo: when Langevin meets Moreau. S IAM Journal on Imaging Sciences 11(1), 473–506 (2018)

    work page 2018

  22. [22]

    In: Conference on Learning Theory, pp

    Brosse, N., Durmus, A., Moulines, É., Pereyra, M.: Samp ling from a log-concave distribution with compact support with proximal Langevin M onte Carlo. In: Conference on Learning Theory, pp. 319–342 (2017). PMLR

    work page 2017

  23. [23]

    Statisti cs and Computing 33(4), 85 (2023)

    Eftekhari, A., Vargas, L., Zygalakis, K.C.: The forwar d–backward envelope for sampling with the overdamped Langevin algorithm. Statisti cs and Computing 33(4), 85 (2023)

    work page 2023

  24. [24]

    Advances in Neural Information Processing Systems 31 (2018)

    Hsieh, Y.-P., Kavis, A., Rolland, P., Cevher, V.: Mirro red Langevin dynamics. Advances in Neural Information Processing Systems 31 (2018)

    work page 2018

  25. [25]

    In: Abernethy, J., Agarwal, S

    Zhang, K.S., Peyré, G., Fadili, J., Pereyra, M.: Wasser stein control of mir- ror langevin monte carlo. In: Abernethy, J., Agarwal, S. (ed s.) Proceedings of Thirty Third Conference on Learning Theory. Proceedings of Machine Learning Research, vol. 125, pp. 3814–3841. PMLR, New York, NY, USA (2 020) 29

    work page

  26. [26]

    In: Pr oceedings of the 34th International Conference on Neural Information Processin g Systems

    Chewi, S., Le Gouic, T., Lu, C., Maunu, T., Rigollet, P., Stromme, A.: Ex- ponential ergodicity of mirror-langevin diffusions. In: Pr oceedings of the 34th International Conference on Neural Information Processin g Systems. NIPS ’20. Curran Associates Inc., Red Hook, NY, USA (2020)

    work page 2020

  27. [27]

    Journal of the Royal Statistical Society Ser ies B: Statistical Methodology 73(2), 123–214 (2011)

    Girolami, M., Calderhead, B.: Riemann manifold langev in and hamiltonian monte carlo methods. Journal of the Royal Statistical Society Ser ies B: Statistical Methodology 73(2), 123–214 (2011)

    work page 2011

  28. [28]

    Journal of Machin e Learning Research 26(71), 1–50 (2025)

    Bharath, K., Lewis, A., Sharma, A., Tretyakov, M.V.: Sa mpling and estimation on manifolds using the langevin diffusion. Journal of Machin e Learning Research 26(71), 1–50 (2025)

    work page 2025

  29. [29]

    In: Cortes, C., Lawrence, N., Lee, D., Sugiyama, M., Garnett , R

    Ma, Y.-A., Chen, T., Fox, E.: A complete recipe for stoch astic gradient mcmc. In: Cortes, C., Lawrence, N., Lee, D., Sugiyama, M., Garnett , R. (eds.) Advances in Neural Information Processing Systems, vol. 28, pp. 1–9. Curran Associates, Inc., Red Hook, NY, USA (2015)

    work page 2015

  30. [30]

    Springer, Berlin (1992)

    Kloeden, P.E., Platen, E.: Numerical Solution of Stoch astic Differential Equations. Springer, Berlin (1992)

    work page 1992

  31. [31]

    Princeton m athematical series 28 (1970)

    Tyrrell Rockafellar, R.: Convex analysis. Princeton m athematical series 28 (1970)

    work page 1970

  32. [32]

    Wiley, USA (1983)

    Nemirovsky, A.S., Yudin, D.B.: Problem Complexity and Method Efficiency in Optimization. Wiley, USA (1983)

    work page 1983

  33. [33]

    In: International Conference o n Algorithmic Learning Theory, pp

    Li, R., Tao, M., Vempala, S.S., Wibisono, A.: The mirror langevin algorithm con- verges with vanishing bias. In: International Conference o n Algorithmic Learning Theory, pp. 718–742 (2022). PMLR

    work page 2022

  34. [34]

    Naval research logistics quarterly 26(3), 403–413 (1979)

    Lewis, P.W., Shedler, G.S.: Simulation of nonhomogene ous poisson processes by thinning. Naval research logistics quarterly 26(3), 403–413 (1979)

    work page 1979

  35. [35]

    The Annals of Applied Probability 30(5), 2069–2098 (2020) https://doi.org/10.1214/19-AAP1552

    Durmus, A., Guillin, A., Monmarché, P.: Geometric ergo dicity of the Bouncy Particle Sampler. The Annals of Applied Probability 30(5), 2069–2098 (2020) https://doi.org/10.1214/19-AAP1552

    work page doi:10.1214/19-aap1552 2069

  36. [36]

    Annals of Statistics 3, 1268–1287 (2019)

    Deligiannidis, G., Bouchard-Côté, A., Doucet, A.: Exp onential ergodicity of the bouncy particle sampler. Annals of Statistics 3, 1268–1287 (2019)

    work page 2019

  37. [37]

    In: Proceedings of the 28th International Conferen ce on International Conference on Machine Learning

    Welling, M., Teh, Y.W.: Bayesian learning via stochast ic gradient Langevin dy- namics. In: Proceedings of the 28th International Conferen ce on International Conference on Machine Learning. ICML’11, pp. 681–688. Omni press, Madison, WI, USA (2011) 30

    work page 2011

  38. [38]

    Advances in Neural Information Processing System s 34, 28405–28418 (2021)

    Ahn, K., Chewi, S.: Efficient constrained sampling via th e mirror-langevin al- gorithm. Advances in Neural Information Processing System s 34, 28405–28418 (2021)

    work page 2021

  39. [39]

    SIA M Journal on Imaging Sciences 16(3), 1195–1234 (2023)

    Melidonis, S., Dobson, P., Altmann, Y., Pereyra, M., Zy galakis, K.: Efficient Bayesian computation for low-photon imaging problems. SIA M Journal on Imaging Sciences 16(3), 1195–1234 (2023)

    work page 2023

  40. [40]

    Journal of Computational and Graphic al Statistics 23(2), 518–542 (2014)

    Pakman, A., Paninski, L.: Exact Hamiltonian Monte Carl o for truncated mul- tivariate Gaussians. Journal of Computational and Graphic al Statistics 23(2), 518–542 (2014)

    work page 2014

  41. [41]

    Statistics and Computing 16, 5–14 (2006)

    McAuliffe, J.D., Blei, D.M., Jordan, M.I.: Nonparametr ic empirical bayes for the Dirichlet process mixture model. Statistics and Computing 16, 5–14 (2006)

    work page 2006

  42. [42]

    In: 2005 IEEE Computer Society Conference on Co mputer Vision and Pattern Recognition (CVPR’05), vol

    Fei-Fei, L., Perona, P.: A Bayesian hierarchical model for learning natural scene categories. In: 2005 IEEE Computer Society Conference on Co mputer Vision and Pattern Recognition (CVPR’05), vol. 2, pp. 524–531 (2005). IEEE

    work page 2005

  43. [43]

    Journal of machine Learning research 3(Jan), 993–1022 (2003)

    Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent Dirichlet al location. Journal of machine Learning research 3(Jan), 993–1022 (2003)

    work page 2003

  44. [44]

    The Anna ls of Applied Proba- bility 33(6A), 4693–4746 (2023)

    Vasdekis, G., Roberts, G.O.: Speed up Zig-Zag. The Anna ls of Applied Proba- bility 33(6A), 4693–4746 (2023)

    work page 2023

  45. [45]

    SIAM Journal on Ma thematics of Data Science 5(2), 558–587 (2023) https://doi.org/10.1137/22M1508613

    Tan, H.Y., Mukherjee, S., Tang, J., Schönlieb, C.-B.: D ata-driven mirror de- scent with input-convex neural networks. SIAM Journal on Ma thematics of Data Science 5(2), 558–587 (2023) https://doi.org/10.1137/22M1508613

    work page doi:10.1137/22m1508613 2023

  46. [46]

    Bernoulli (2020)

    Bertazzi, A., Bierkens, J.: Adaptive schemes for piece wise deterministic monte carlo algorithms. Bernoulli (2020)

    work page 2020

  47. [47]

    Statistics and Computing 32(6), 107 (2022) https://doi.org/10.1007/s11222-022-10142-x

    Corbella, A., Spencer, S.E.F., Roberts, G.O.: Automat ic zig-zag sampling in practice. Statistics and Computing 32(6), 107 (2022) https://doi.org/10.1007/s11222-022-10142-x

    work page doi:10.1007/s11222-022-10142-x 2022

  48. [48]

    Stocha stic Processes and their Applications 154, 91–153 (2022) https://doi.org/10.1016/j.spa.2022.09.004

    Bertazzi, A., Bierkens, J., Dobson, P.: Approximation s of piecewise deterministic markov processes and their convergence properties. Stocha stic Processes and their Applications 154, 91–153 (2022) https://doi.org/10.1016/j.spa.2022.09.004

    work page doi:10.1016/j.spa.2022.09.004 2022

  49. [49]

    arXiv preprint arXiv:2301.02537 (2023)

    Bertazzi, A., Dobson, P., Monmarché, P.: Splitting sch emes for second or- der approximations of piecewise-deterministic Markov pro cesses. arXiv preprint arXiv:2301.02537 (2023)

    work page arXiv 2023

  50. [50]

    Statistics and Computing 34(1), 61 (2024) https://doi.org/10.1007/s11222-023-10363-8 31

    Pagani, F., Chevallier, A., Power, S., House, T., Cotte r, S.: Nuzz: Numer- ical zig-zag for general models. Statistics and Computing 34(1), 61 (2024) https://doi.org/10.1007/s11222-023-10363-8 31

    work page doi:10.1007/s11222-023-10363-8 2024

  51. [51]

    : Lectures on Convex Optimization vol

    Nesterov, Y., et al. : Lectures on Convex Optimization vol. 137. Springer, Switzerland (2018) 32

    work page 2018