Pith. sign in

REVIEW 4 major objections 6 minor 28 references

Bouncy particle sampler with infinite exchanging parallel tempering

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read By taking parallel-tempering exchanges to the infinite rate inside the bouncy particle sampler, the paper constructs a sampler whose reordered samples converge to multimodal target distributions.

desk verdict Useful, honest transfer of infinite swapping to BPS, but the fast-swap limit is assumed rather than proved and the compute-adjusted gains are still unclear. read the letter →

arxiv 2509.02003 v1 pith:5LUQCMM6 submitted 2025-09-02 cs.LG stat.ML

Yohei Saito , Shun Kimura , Koujin Takeda This is my paper · ORCID
classification cs.LGstat.ML MSC 65C0560J2562F15
keywords bouncyparticlesamplerparalleltemperinginfiniteswappingpiecewise-deterministicMarkovprocessmultimodalsamplingchainMonteCarloBayesianposterior
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

The paper tries to show that parallel tempering, normally run with finite exchange rates, can be pushed to the infinite-swap limit inside the bouncy particle sampler (BPS), a piecewise-deterministic MCMC method. In this limit, the permutation of temperatures equilibrates instantaneously, and the sampler uses temperature-averaged bounce and jump rates instead of explicit replica exchanges. The authors argue the process still has the target product distribution as its stationary law after reordering particles, and their simulations on a 24-dimensional Gaussian mixture show cluster probabilities and KS statistics far closer to the target than ordinary BPS. This matters because BPS is known for being easy to tune, and the result extends it to posterior distributions with separated modes, where plain BPS tends to get stuck.

What carries the argument

The central object is the generator of the infinite-exchange BPS-PT process, Eq. (22), with rates Eqs. (23)-(25): a weighted average over all temperature permutations, weighted by omega_sigma, the stationary probability of each permutation given the current particle positions. This object carries the argument by replacing discrete swapping events with averaged dynamics; the same weights define the reordering rule Eq. (27), which converts the permutation-symmetric stationary mixture into samples from the intended target.

What would settle it

Run the proposed infinite-rate BPS-PT on a two-temperature Gaussian mixture with known cluster weights and compare the reordered empirical distribution to the target for increasing sample size; if the stationarity claim is wrong, the KS statistic will plateau above Monte Carlo error. A sharper check is to evaluate the generator identity integral L_infinity f p_st = 0 for a simple test function on a one-dimensional two-mode target, since failure of this identity directly falsifies the claimed invariant distribution.

Watch

Extended reading notes

Core claim

The central discovery is a construction: write down the infinitesimal generator of BPS when each of L particles carries a temperature label and labels swap at rate alpha_s, then let alpha_s go to infinity. The raw dynamics has the symmetric stationary distribution of Eq. (19), a uniform mixture over temperature permutations. Because the labels mix instantly, one replaces the indicator of the current permutation by the stationary permutation weights omega_sigma(x,y), giving the effective bounce rate of Eq. (23) and transition rate of Eq. (24). After reordering the particles according to the permutation weights at each time, the empirical measure converges to the original target product distri

Load-bearing premise

The load-bearing premise is that the temperature-swap rate can be sent to infinity inside the bouncing dynamics, so the current temperature ordering can be replaced by its instantaneous equilibrium weights; the paper treats this interchange as intuitive and cites an analogous result for smoother dynamics, not for the piecewise-deterministic bouncing process.

Editorial extensions

If this is right

  • For continuous targets, the infinite-exchange BPS-PT has roughly the same cost as finite-rate parallel tempering, because the bounce rate only requires computing lambda_b and not all discrete transition probabilities.
  • If the claimed invariance holds, the sampler extends the infinite-swap guarantee from smooth Langevin-type dynamics to piecewise-deterministic bouncing dynamics, so the reordered samples converge to the target despite the raw symmetric stationary law.
  • For large permutation groups, the paper's generating-subgroup scheme gives implementable upper bounds for the event rates, making a lightweight version feasible when enumerating all L! permutations is impossible.
  • On multimodal targets, the method avoids the failure mode of plain BPS in which transitions between separated clusters are so rare that the cluster probabilities of the samples are badly biased.
  • For simple unimodal or weakly coupled targets, the added computation is not worth the gain; the paper's simulations show only a marginal improvement in KS statistics while runtime grows by more than an order of magnitude.

Reading between the lines

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

  • The same weighted-permutation averaging could be applied to other piecewise-deterministic samplers with velocity reflections, as long as an analogous generator limit can be justified.
  • A direct stationarity check of the infinite-rate generator on a small two-temperature example would turn the paper's intuitive limiting argument into a theorem for bouncing dynamics; this is a natural next step the paper leaves open.
  • Because cost grows factorially with the number of temperatures, an adaptive scheme that chooses the temperature subsets and the switching time t_beta based on the target geometry would be a practical extension of the generating-subgroup idea.
  • The construction suggests interpreting the infinite-swap BPS as a continuous-time simulated-tempering process in which each particle's temperature fluctuates with coordinate-dependent weights, possibly enabling asynchronous variants where only a subset of temperatures is updated.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a parallel-tempering (PT) variant of the bouncy particle sampler (BPS), including an 'infinite-exchange' limit. Section 2 reviews BPS for continuous and mixed continuous-discrete targets. Section 3.1 constructs a finite-rate BPS-PT by assigning inverse temperatures to particles through a permutation and exchanging those permutations; Eq. (19) is stated as the stationary distribution. Section 3.2 writes down the infinite-exchange-rate generator (22) with position- and permutation-averaged rates (23)-(25), justified by saying that at infinite exchange rate the permutation reaches its conditional stationary distribution and replacing 1_{z=sigma} in Eq. (11) by omega_sigma. To avoid summing over the full permutation group, a lightweight subgroup approximation is introduced via two generating subgroups G, G' and a switching interval t_beta. Section 4 reports numerical experiments on a 24-dimensional Gaussian mixture model and a mixed continuous-discrete model, comparing BPS-PT with BPS and with Metropolis-augmented HMC (MAHMC), and claims BPS-PT improves multimodal sampling.

Significance. If the infinite-swap generator (22)-(25) is correct, the paper would extend infinite-swap parallel tempering beyond Langevin/HMC dynamics to piecewise-deterministic Markov process samplers, potentially improving BPS on multimodal targets. The paper offers pseudocode, an explicit upper bound for thinning, and experiments in a nontrivial multimodal example. However, the central theoretical step is asserted rather than proved, and the numerical experiments evaluate an approximate subgroup version; the method's practical advantage is not established on a cost-adjusted basis. The contribution is therefore conditional: the idea is plausible and potentially useful, but the manuscript as it stands does not supply the rigorous support needed for the claimed infinite-swap limit or for the finite-subgroup approximation.

major comments (4)
  1. [Sec. 3.2, Eqs. (22)-(25)] The infinite-exchange generator is obtained by replacing 1_{z=sigma} in Eq. (11) with omega_sigma(x,y), described as 'intuitively' true and citing Ref. [22]. Ref. [22] proves this averaging for Langevin dynamics, not for a piecewise-deterministic Markov process with velocity reflections, discrete-state transitions, and velocity refreshes. The paper supplies no semigroup or martingale argument that the alpha_s -> infinity limit commutes with the BPS jump dynamics, and no finite-rate convergence experiment. Since Eqs. (23)-(25) are the core proposal, this missing proof is load-bearing.
  2. [Sec. 3.2, Eqs. (28)-(32) and Algs. 5-6] The finite-t_beta subgroup approximation is not shown to preserve the target distribution. Algorithm 6 alternates between subgroups G and G', draws sigma from omega_{l,sigma} at the end of each interval t_beta, and reorders particles accordingly, but no stationarity or detailed-balance argument is given for this discrete reordering. Since t_beta is finite, the permutation does not reach its stationary distribution, and the statement that smaller t_beta is 'closer' to using S_L is heuristic. The numerical results in Section 4 therefore test an approximate algorithm, and cannot be used to validate the exact infinite-swap limit unless the approximation is proven correct.
  3. [Sec. 4.1, Table 2] The claim that BPS-PT is 'more useful' for multimodal distributions is not supported on a cost-adjusted basis. Table 2 reports computation time (5.1 +/- 0.2)x10 s for BPS vs (3.2 +/- 0.1)x10^3 s for BPS-PT, about a 60x increase, while ESS per sample only improves by a factor of about 9.5 (4e-4 to 3.8e-3). Thus ESS per unit computation time is actually worse for BPS-PT. The KS and KLD improvements at a fixed number of samples do not account for the much larger runtime. The comparison should report effective sample size per second or otherwise adjust for computational cost before claiming practical usefulness.
  4. [Sec. 3.1, Eq. (19) and Sec. 3.2, Eq. (27)] The invariance of Eq. (19) under the finite-rate BPS-PT generator (11) and the convergence of the reordered empirical measures (20) and (27) are stated without proof. While these are plausibly standard, the paper's central claims depend on the limiting statements in Eq. (27), and the absence of an ergodicity or convergence argument makes it impossible to assess the conditions under which the proposed sampler is guaranteed to converge to the target.
minor comments (6)
  1. [Sec. 3.2, paragraph after Eq. (26)] Typo: 'BPT-PT' should be 'BPS-PT'.
  2. [Eq. (4)] The Gaussian density uses dimension d, while the paper earlier defines the state space as R^n; unify the notation.
  3. [Sec. 4, ESS definition] The quantity 'ESS per sample' is not formally defined. Since very small absolute values are reported (e.g., 4e-4), clarify whether this is the ratio ESS / number of samples and how autocorrelation is estimated.
  4. [Alg. 7] The pseudocode uses Delta t_ev and i_ev at lines 27-29, but these variables are not clearly defined if the while loop exits without an accepted event. Also the velocity-refresh line 28 appears reachable after a bounce or transition has already been returned; specify the control flow more precisely.
  5. [Eq. (34)] The upper bound for the reflection rate uses theta(v_1), which is not defined in the text.
  6. [Figs. 1-2] The figures lack axis labels for the vertical axis; specify that it is the maximum KS test value.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity detected: the infinite-swap rates are an external-theorem transfer, and the numerical claims are benchmarked against known targets and independent samplers.

full rationale

The paper's derivation chain does not reduce to its own inputs. Section 2 reviews BPS and its mixed continuous-discrete extension, with stationary distributions checked via Eq. (5) and the balance condition Eq. (7). Section 3.1 constructs finite-rate BPS-PT: the swap rate Eq. (10) is the standard Metropolis ratio for the permutation state, so the swap block is reversible with respect to Eq. (19), and each BPS block preserves the corresponding temperature marginal; Eq. (19) is thus an invariant measure of Eq. (11), not a fitted output. Section 3.2 obtains the infinite-exchange generator by replacing the indicator 1_{z=sigma} with the conditional stationary permutation weight omega_sigma(x,y) (Eqs. (23)-(25)), explicitly citing the external infinite-swapping theory of Dupuis et al. [22]. That is an external result, not a self-citation, and the rates are not predictions fitted to any reported KS/KLD/ESS value. The numerical section compares samples against the known Gaussian-mixture and Neal-model targets and against the external MAHMC sampler; the metrics are diagnostics, not calibration targets. The only substantive caveat is the 'Intuitively...' sentence in Section 3.2 (Eq. (26) and following): the alpha_s to infinity averaging is asserted rather than proved for a piecewise-deterministic process. This is an unverified transfer of a theorem and hence a correctness/rigor risk, not a circularity: the argument does not rely on the authors' own prior work, and no equation is defined in terms of the conclusion it is supposed to establish. The finite-subgroup approximation with switching time t_beta is likewise an acknowledged approximation, not a circular step.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

No physical entities are invented. The central claim rests on an unproved infinite-swap limit for piecewise-deterministic processes, hand-chosen temperature ladders and subgroup schedules, and a finite-t_beta subgroup approximation whose bias is unquantified.

free parameters (8)
  • Inverse temperature ladder for Gaussian mixture (L=10) = beta_i = 1.0 - 0.1*(i-1), i=1..10
    Hand-chosen schedule; no derivation or tuning, directly affects PT mixing.
  • Inverse temperature ladder for mixed model (L=5) = beta_i = 1.0 - 0.2*(i-1), i=1..5
    Hand-chosen schedule.
  • Partition pair G,G' for Gaussian mixture = {{1..4},{5..8},{9,10}} and {{1,2},{3..6},{7..10}}
    Choice of generating subgroups; no rule for selecting partitions is given.
  • Partition pair G,G' for mixed model = {{1,2,3},{4,5}} and {{1,2},{3,4,5}}
    Choice of generating subgroups.
  • Subgroup switching time t_beta = 0.1 in both experiments
    Controls approximation of the full symmetric group; shorter is more accurate but more costly, and no bias analysis is provided.
  • Sampling interval tsamp = 10*t_beta = 1
    Set as an integer multiple of t_beta; affects autocorrelation and ESS estimates.
  • BPS rates for Gaussian mixture = alpha_b=1, alpha_j=4, lambda_ref=1
    Set without tuning; these rates control reflection, discrete transition, and refresh.
  • BPS rates for mixed model = alpha_b=1, alpha_j=20, lambda_ref=0.1
    Set without tuning.
assumptions (5)
  • standard math BPS with generator Eq (3) has stationary distribution p(x)psi(v) (Eq 5).
    Background from Refs [12,19], used as the base sampler.
  • domain assumption Discrete variable updates satisfying balance Eq (7) preserve the target distribution for the mixed model.
    Standard extension used in Ref [19].
  • domain assumption The alpha_s to infinity limit of finite-rate BPS-PT (Eq 11) equals the averaged generator Eq (22) with stationary distribution Eq (19).
    The paper says 'intuitively' in Section 3.2; no proof is given, and Ref [22] only proves this type of limit for Langevin dynamics.
  • ad hoc to paper Alternating generating subgroups G and G' at interval t_beta yields samples from the same target distribution.
    No proof of invariant distribution or bias as t_beta grows; all numerical results use this finite-time approximation.
  • standard math A polynomial upper bound for the BPS reflection rate can always be found by Taylor expansion for the tested models.
    Needed for the thinning Algorithm 7; the GMM bound is not stated explicitly.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bouncy particle sampler with infinite exchanging parallel tempering." pith.science (2026). https://pith.science/paper/5LUQCMM6

@misc{pith2026250902003,
  author       = {Pith},
  title        = {Pith review of: Bouncy particle sampler with infinite exchanging parallel tempering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5LUQCMM6}},
  note         = {Machine review of arXiv:2509.02003}
}
read the original abstract

Bayesian inference is useful to obtain a predictive distribution with a small generalization error. However, since posterior distributions are rarely evaluated analytically, we employ the variational Bayesian inference or sampling method to approximate posterior distributions. When we obtain samples from a posterior distribution, Hamiltonian Monte Carlo (HMC) has been widely used for the continuous variable part and Markov chain Monte Carlo (MCMC) for the discrete variable part. Another sampling method, the bouncy particle sampler (BPS), has been proposed, which combines uniform linear motion and stochastic reflection to perform sampling. BPS was reported to have the advantage of being easier to set simulation parameters than HMC. To accelerate the convergence to a posterior distribution, we introduced parallel tempering (PT) to BPS, and then proposed an algorithm when the inverse temperature exchange rate is set to infinity. We performed numerical simulations and demonstrated its effectiveness for multimodal distribution.

Figures

Figures reproduced from arXiv: 2509.02003 by the authors.

Figure 1
Figure 1. We show the maximum KS test values with respect to the com [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. We show the maximum KS test values with respect to the com [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 27 canonical work pages

  1. [22]

    On the infi nite swapping limit for parallel tempering

    Paul Dupuis, Yufei Liu, Nuria Plattner, and Jimmie D Doll. On the infi nite swapping limit for parallel tempering. Multiscale Modeling & Simulation, 10(3):986–1022, 2012

    work page 2012

  2. [1]

    Algebraic geometry and statistical learning theory, volume 25

    Sumio Watanabe. Algebraic geometry and statistical learning theory, volume 25. Cambridge university press, 2009

    work page 2009

  3. [2]

    Development of a bayesian inference method for the analysis of x-ray reflectivit y data

    Atsushi Machida, Takeshi Watanabe, Masaichiro Mizumaki, Kenj i Nagata, and Masato Okada. Development of a bayesian inference method for the analysis of x-ray reflectivit y data. Japanese Journal of Applied Physics, 63(1):017005, 2024

    work page 2024

  4. [3]

    Practical bayesian inference in neuroscience: Or how i learned to stop worrying and embrace the distribution

    Brandon S Coventry and Edward L Bartlett. Practical bayesian inference in neuroscience: Or how i learned to stop worrying and embrace the distribution. Eneuro, 11(7), 2024

    work page 2024

  5. [4]

    Bayesian forecasting in economics and finance: A modern review

    Gael M Martin, David T Frazier, Worapree Maneesoonthorn, Rub ´ en Loaiza-Maya, Florian Huber, Gary Koop, John Maheu, Didier Nibbering, and Anastasios Panagiotelis. Bayesian forecasting in economics and finance: A modern review. International Journal of Forecasting, 40(2):811–839, 2024

    work page 2024

  6. [5]

    Hands- on bayesian neural networks—a tutorial for deep learning users

    Laurent Valentin Jospin, Hamid Laga, Farid Boussaid, Wray Buntin e, and Mohammed Bennamoun. Hands- on bayesian neural networks—a tutorial for deep learning users. IEEE Computational Intelligence Magazine, 17(2):29–48, 2022

    work page 2022

  7. [6]

    A tutorial on adaptive mcmc

    Christophe Andrieu and Johannes Thoms. A tutorial on adaptive mcmc. Statistics and computing, 18:343–373, 2008. 12

    work page 2008

  8. [7]

    Brownian dyna mics as smart monte carlo simulation

    Peter J Rossky, Jimmie D Doll, and Harold L Friedman. Brownian dyna mics as smart monte carlo simulation. The Journal of Chemical Physics, 69(10):4628–4633, 1978

    work page 1978

Show all 28 references
  1. [8]

    Mcmc using hamiltonian dynamics

    Radford M Neal et al. Mcmc using hamiltonian dynamics. Handbook of markov chain monte carlo, 2(11):2, 2011

    2011

  2. [9]

    Mixed hamiltonian monte carlo for mixed discrete a nd continuous variables

    Guangyao Zhou. Mixed hamiltonian monte carlo for mixed discrete a nd continuous variables. Advances in Neural Information Processing Systems, 33:17094–17104, 2020

    2020

  3. [10]

    Metropolis augmented hamiltonian monte carlo

    Guangyao Zhou. Metropolis augmented hamiltonian monte carlo. arXiv preprint arXiv:2201.08044, 2022

    2022 arXiv

  4. [11]

    Rejection-free monte carlo samp ling for general potentials

    Elias AJF Peters and G de With. Rejection-free monte carlo samp ling for general potentials. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 85(2):026703, 2012

    2012

  5. [12]

    The bouncy particle sampler: A non- reversible rejection-free markov chain monte carlo method

    Alexandre Bouchard-Cˆ ot´ e, Sebastian J Vollmer, and Arnaud Doucet. The bouncy particle sampler: A non- reversible rejection-free markov chain monte carlo method. Journal of the American Statistical Association, 113(522):855–867, 2018

    2018

  6. [13]

    Randomized hamiltonian monte carlo as scaling limit of the bouncy particle sampler and dimension -free convergence rates

    George Deligiannidis, Daniel Paulin, Alexandre Bouchard-Cˆ ot´ e, and Arnaud Doucet. Randomized hamiltonian monte carlo as scaling limit of the bouncy particle sampler and dimension -free convergence rates. The Annals of Applied Probability, 31(6):2612–2662, 2021

    2021

  7. [14]

    Mcmc using bouncy hamiltonian d ynamics: A unifying frame- work for hamiltonian monte carlo and piecewise deterministic markov p rocess samplers

    Andrew Chin and Akihiko Nishimura. Mcmc using bouncy hamiltonian d ynamics: A unifying frame- work for hamiltonian monte carlo and piecewise deterministic markov p rocess samplers. arXiv preprint arXiv:2405.08290, 2024

    2024

  8. [15]

    A discrete bouncy partic le sampler

    Chris Sherlock and Alexandre H Thiery. A discrete bouncy partic le sampler. Biometrika, 109(2):335–349, 2022

    2022

  9. [16]

    Stochastic gradient piecewise deterministic monte carlo samplers

    Paul Fearnhead, Sebastiano Grazzi, Chris Nemeth, and Garet h O Roberts. Stochastic gradient piecewise deterministic monte carlo samplers. arXiv preprint arXiv:2406.19051, 2024

    2024 arXiv

  10. [17]

    Stochastic bouncy particle sam pler for bayesian neural networks

    Ethan Goan and Clinton Fookes. Stochastic bouncy particle sam pler for bayesian neural networks. Stochastic Bouncy Particle Sampler for Bayesian Neural Networks, 2021

    2021

  11. [18]

    Piecewise deterministic generative models

    Andrea Bertazzi, Dario Shariatian, Umut Simsekli, Eric Moulines, a nd Alain Durmus. Piecewise deterministic generative models. Advances in Neural Information Processing Systems, 37:93336–93373, 2024

    2024

  12. [19]

    Vollmer, and Arnaud Doucet

    Alexandre Bouchard-Cˆ ot´ e, Sebastian J. Vollmer, and Arnaud Doucet. The bouncy particle sampler: A non- reversible rejection-free markov chain monte carlo method. arXiv preprint arXiv:1510.02451.v1, 2017

    2017 arXiv

  13. [20]

    Replica monte carlo sim ulation of spin-glasses

    Robert H Swendsen and Jian-Sheng Wang. Replica monte carlo sim ulation of spin-glasses. Physical review letters, 57(21):2607, 1986

    1986

  14. [21]

    Sampling from multimodal distributions using temp ered transitions

    Radford M Neal. Sampling from multimodal distributions using temp ered transitions. Statistics and computing, 6:353–366, 1996

    1996

  15. [23]

    Monte carlo implementation

    Radford M Neal. Monte carlo implementation. Bayesian learning for neural networks, pages 55–98, 1996

    1996

  16. [24]

    The no-u-turn sample r: adaptively setting path lengths in hamiltonian monte carlo

    Matthew D Hoffman, Andrew Gelman, et al. The no-u-turn sample r: adaptively setting path lengths in hamiltonian monte carlo. J. Mach. Learn. Res., 15(1):1593–1623, 2014

    2014

  17. [25]

    Simulation of nonhomogeneous po isson processes by thinning

    PA W Lewis and Gerald S Shedler. Simulation of nonhomogeneous po isson processes by thinning. Naval research logistics quarterly, 26(3):403–413, 1979

    1979

  18. [26]

    Introduction to markov chain monte carlo

    Charles J Geyer. Introduction to markov chain monte carlo. Handbook of markov chain monte carlo, 20116022(45):22, 2011

    2011

  19. [27]

    Non-reversibly updating a uniform [0, 1] value fo r metropolis accept/reject decisions

    Radford M Neal. Non-reversibly updating a uniform [0, 1] value fo r metropolis accept/reject decisions. arXiv preprint arXiv:2001.11950, 2020

    2001 arXiv

  20. [28]

    Markov chain monte carlo metho d without detailed balance

    Hidemaro Suwa and Synge Todo. Markov chain monte carlo metho d without detailed balance. Physical review letters, 105(12):120603, 2010. 13

    2010

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.