Pith. sign in

REVIEW 3 major objections 4 minor 96 references

This paper proposes Compound Auxiliary Metropolis, a reversible MCMC method that builds auxiliary variables into the multi-candidate proposal of multiple-try Metropolis, aiming to show that one transition can combine local random-walk moves

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 23:09 UTC pith:DPWMBEGP

load-bearing objection Solid, honest extension of MTM with a clean involutive-MCMC proof, but the NUTS comparison is undercut by per-iteration ESS under a fixed wall-clock budget and per-problem tuning; treat the empirical claim as provisional. the 3 major comments →

arxiv 2607.15499 v1 pith:DPWMBEGP submitted 2026-07-16 stat.CO

Compound Auxiliary Metropolis: Incorporating Auxiliary Variables into Multi-Candidate MCMC

Renny Doig , Liangliang Wang This is my paper
classification stat.CO
keywords Markov chain Monte Carlomultiple-try Metropolisauxiliary variablesannealed sequential Monte Carlodetailed balancemultimodal samplingeffective sample sizeinvolutive MCMC
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that injecting auxiliary information into the multi-candidate step of multiple-try Metropolis yields a valid, reversible sampler that can explore multimodal and strongly non-Gaussian targets that purely local proposals miss. It introduces Compound Auxiliary Metropolis (CAM), where at each iteration some candidates are drawn from local proposals conditioned on the current state and others from proposals conditioned on an auxiliary variable drawn from a user-specified auxiliary generating distribution. The authors prove CAM preserves the target as its stationary distribution by embedding it in an involutive extended-state construction, and they demonstrate empirically that CAM maintains low distributional error on a five-mode Gaussian mixture across large intermodal distances, sustains effective sample size on a severely curved banana target, and stays competitive on the eight-schools hierarchical model. The practical upshot is that practitioners can feed cheap approximations of a hard target—a grid, an annealed-SMC particle cloud, or even external estimates—into a reversible MCMC kernel to escape local traps.

Core claim

The central mathematical claim is that CAM's acceptance probability—the ratio of target density times reverse proposal-selection probability over the forward quantities—exactly equals the acceptance ratio derived from an involutive construction on an extended state space that includes auxiliary variables, candidate pools, and reverse samples. Because that construction is reversible by design, the CAM kernel satisfies detailed balance and leaves the target distribution invariant. The central empirical claim is that the auxiliary-driven candidates are what make global exploration possible: on a two-dimensional five-mode Gaussian mixture with intermodal distance parameter up to 30, CAM with an

What carries the argument

The load-bearing object is the partitioned compound proposal: M candidates are split into Mx local proposals depending on the current state and Mz auxiliary proposals depending only on an auxiliary variable z drawn from an auxiliary generating distribution (AGD). The paper instantiates three AGDs—an uninformed uniform discrete grid, and two informed empirical distributions built from annealed sequential Monte Carlo: the final weighted particle set and the pooled mixture of all intermediate annealed sets. The theoretical engine is the involutive extended-state construction: the involution swaps the current state with the selected candidate and swaps the unselected candidates with the reverse

Load-bearing premise

The auxiliary generating distribution must place enough probability near every mode of interest: if it drops a mode, the chain will not reliably recover it even though the kernel remains correctly targeted.

What would settle it

Construct a two-mode target and set the auxiliary generating distribution to a point mass on one mode; simulate CAM and record the empirical frequency of visits to the other mode. If the chain visits the uncovered mode with positive frequency, the claim that performance is limited by auxiliary quality would be wrong—the paper's own Figure 2 indicates the chain will not visit it.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • CAM gives a reversible, target-preserving transition for any valid auxiliary generating distribution and any strictly positive weight function, so users can add auxiliary information to an MTM-style sampler without re-deriving a kernel each time.
  • With a suitable auxiliary (e.g., the pooled annealed-SMC empirical distribution combined with component-wise updates), a single chain can recover all five modes of the Gaussian mixture across intermodal distances δ=5 to 30, a regime where standard MTM, Metropolis–Hastings, and the tuned Hamiltonian Monte Carlo benchmark all fail.
  • On the banana target, the full-block CAM configuration with an uninformed discrete grid maintains stable effective sample size per iteration across the non-Gaussianity gradient, outperforming the Hamiltonian benchmark at the most extreme curvature.
  • On the eight-schools hierarchical model, CAM is competitive with the Hamiltonian benchmark—comparable per-iteration effective sample size on the location parameters, somewhat lower on the variance parameter—while offering substantially better performance on the pathological targets.
  • The framework formally subsumes several existing samplers as special cases: Metropolis–Hastings, independence Metropolis–Hastings, standard MTM, independent MTM, and two interacting-chain MTM variants, depending on how many local versus auxiliary candidates are used.

Where Pith is reading between the lines

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

  • The paper gestures at but does not develop a natural extension: making the AGD adaptive across the run—starting with an uninformed grid to discover modes, then switching to an informed approximation once modes are located—could combine the global coverage of the grid with the higher acceptance of an informed auxiliary.
  • The mode-dropping failure mode suggests a practical real-time diagnostic: track the empirical selection rate of auxiliary candidates in different regions of the target; if a mode is never proposed by the auxiliary, the auxiliary-candidate acceptance rate for that region will be zero, a signal a practitioner could monitor while the chain runs.
  • Because the AGD is a free input, CAM offers a principled mechanism to inject external information—meta-analytic summaries, variational approximations, or normalizing-flow samples—directly into a reversible transition; this could be tested in Bayesian problems where a cheap surrogate posterior is available.
  • The paper finds that the optimal proposal architecture flips between the multimodal and banana experiments (component-wise on the mixture, full-block on the banana), hinting that the best coupling of local and auxiliary proposals depends on the target's correlation structure; a systematic study across target families could identify the governing factors.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes Compound Auxiliary Metropolis (CAM), an extension of multiple-try Metropolis in which auxiliary variables drawn from user-specified auxiliary generating distributions (AGDs) are incorporated into the compound proposal. The transition kernel draws auxiliary variables, generates M candidates split into local and auxiliary proposals, selects one candidate via weighted selection, forms reverse samples, and accepts with the MTM-style probability in Eq. (2). The authors prove detailed balance by embedding CAM in the involutive MCMC framework: the acceptance probability is shown to equal the extended-state density ratio, with the AGD factors cancelling. Three AGD examples are considered: a discrete uniform grid, an ASMC target-only empirical distribution, and a pooled ASMC path distribution. The paper reports simulation studies on a five-component Gaussian mixture, the banana distribution, and the eight schools model, comparing CAM with MH, MTM, and NUTS/Stan. The central claims are that CAM preserves the target as stationary distribution, that auxiliary-driven candidates improve global multimodal exploration relative to MTM, and that CAM matches or beats NUTS on the hardest tested targets.

Significance. If the empirical claims hold, CAM is a useful contribution to the MCMC toolbox: it is a simple and general mechanism for injecting auxiliary information into multi-candidate proposals, with a clean detailed-balance proof and flexible design choices. The theoretical derivation is self-contained and appears correct; the involutive-MCMC argument does not depend on tuning parameters or fitted constants and is a genuine strength. The paper also ships a public code repository, which supports reproducibility. The potential significance is moderate: CAM is a generalization of existing MTM and independent Metropolis constructions, and the empirical benefits over MTM on multimodal targets are plausible. However, the headline empirical comparison to NUTS is currently not established because the reported metric (ESS per iteration) is incompatible with the stated fixed wall-clock experimental design.

major comments (3)
  1. [Section 3, Section 3.2, Figures 5–8] The experiments are run under a strictly fixed wall-clock time budget, but efficiency is reported as ESS per iteration. Under a time budget, methods complete different numbers of iterations, and ESS per iteration is a valid efficiency measure only if per-iteration costs are equal. They are not: the selected CAM banana configuration uses M' = 1000, i.e., M = 2000 candidates per iteration with target-density evaluations for weights, whereas a NUTS iteration performs an adaptive number of leapfrog steps with gradients. The paper does not report completed iteration counts or total ESS, so the claim that CAM 'shows better performance for the most difficult settings' relative to NUTS is not supported by the data as presented. Please report wall-clock-normalized ESS or total ESS (with iteration counts), and/or run an equal-iteration-count comparison.
  2. [Section 3.1–3.3] The 'best CAM configuration' (AGD type, candidate count, proposal architecture) is selected for each benchmark using the same KSD and ESS metrics that are subsequently used to compare CAM against external baselines. Since the selection is performed on the test problems themselves, the external comparison is optimistic and likely inflates the apparent gains. A fair evaluation should prespecify the CAM configuration (or use a separate selection/tuning set), or at minimum report the sensitivity of the external benchmark to the chosen configuration.
  3. [Section 3.1, Figure 2, Section 4] The paper's own proof-of-concept shows the central limitation: when the ASMC auxiliary misses two of five mixture modes, the CAM chain never visits them. This is acknowledged in Section 4 ('the possible benefits from using CAM are limited by the quality of the auxiliary variables'), but it directly bounds the central empirical claim about global exploration. The claim should be qualified accordingly: CAM can explore multiple modes only if the auxiliary coverage includes them; for uninformed grid auxiliaries, performance depends on user-specified grid bounds and spacing, which in practice encode prior information about the target. Please add a robustness analysis over auxiliary quality/coverage — for example, repeated ASMC runs or multiple grid choices — and report how often the selected informed auxiliary drops modes.
minor comments (4)
  1. [Section 2.1, Eq. (3)] In the extended-state distribution, the product over reverse samples is written as m ≠ J, but in the displayed ratio it becomes a product over all m with the convention x*_J = x. This is correct but worth making explicit at the definition of ˜π(˜x) to avoid confusion.
  2. [Section 4] The statement that 'in all experimental settings, the integration of auxiliary information ... demonstrated consistent performance gains over standard MTM' is stronger than what Figure 8 shows: for the parameter τ in the eight schools model, MTM has higher per-iteration ESS than CAM. Please qualify this summary claim.
  3. [Figures 3, 5, 7] The captions are inconsistent about interval levels: Figure 3 uses 95%, Figures 5 and 7 use 90%. Please harmonize and make the legend clear.
  4. [Section 2.4] The sentence 'CAM an open framework for future algorithm development' is missing a verb. Also, the phrase 'independent multiple-try Metropolis algorithm' may deserve a reference or a one-line definition.

Circularity Check

0 steps flagged

No definitional circularity: the CAM acceptance probability is verified by a self-contained involutive-MCMC ratio, and the empirical claims are benchmarks. Minor self-citations and test-set configuration selection are present but do not reduce the result to its inputs.

full rationale

The central mathematical chain is not circular. Algorithm 1 defines the transition with acceptance probability Eq. 2; Section 2.1 constructs the extended target Eq. 3 and the involution g, computes the ratio ~pi(g)/~pi, shows that the AGD densities and unselected-candidate terms cancel, and obtains exactly Eq. 2. That is a detailed-balance proof, not a renaming: the expression is parameter-free and would not be valid for arbitrary selection weights. No fitted constant enters the acceptance probability or the stationarity claim. The empirical portion likewise does not derive benchmark outputs from fitted inputs: CAM configurations are compared on KSD and ESS against MTM and NUTS, and the paper itself displays a mode-drop failure in Figure 2 and states in Section 4 that the benefits are 'limited by the quality of the auxiliary variables.' The design guidance imported from the authors' own Doig and Wang (2026) benchmark (locally balanced weights, CW vs FB architecture) is a mild self-reliance, but it is not load-bearing: Eq. 2's validity does not depend on it, and the paper's own six-configuration experiments independently drive the final choices. Section 2.4's recovery of MTM and IC-MTM as special cases is explicitly a recovery, not a presentation of existing results as new predictions. The selection of the best CAM configuration on the same test problems and metrics used for the external benchmark is a post-selection concern, and the fixed wall-clock budget combined with ESS-per-iteration reporting is an efficiency-comparability concern; under the hard quotation-reduction standard, neither is a definitional circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The theoretical derivation is parameter-free: the acceptance probability contains no fitted constants. All fitted or hand-chosen numbers live in the experimental design — grid bounds, candidate counts, ASMC settings, auxiliary variance — not in the proof. The most load-bearing scientific contingency is that the auxiliary distribution covers the target's modes; the paper documents this both as the source of its gains (ASMC auxiliaries) and as its failure mode (Figure 2).

free parameters (5)
  • Grid bounds L, U and spacing Δ (uninformed AGD) = chosen by hand per target (e.g., δ=5 demo)
    The uniform grid auxiliary requires the user to pick a bounding box; Section 2.2 acknowledges 'the choice of L and U would contain some information about π', and Section 4 stresses 'well-chosen boundaries'. Errors in bounds bias exploration.
  • Number of candidates M (and split Mx=Mz) = M'∈{2,10,100,1000,10000}; best selected per experiment (10, 1000, 10000)
    Candidate pool size is tuned per target after inspecting benchmark metrics (Figs 3, 5, 7) — a post-hoc selection on test data that loads the headline comparisons.
  • Auxiliary proposal variance = unit (co)variance Gaussians
    Auxiliary candidates use fixed standard Gaussians 'with unit (co)variance' (Sec 3); no sensitivity analysis is provided.
  • ASMC tuning (particle count K, rCESS φ=0.8, resampling ε=0.5) = φ=0.8, ε=0.5, K unspecified in main text
    ASMC preprocessing settings are chosen by hand to give a 'coarse approximation' (Supplementary A); this approximation quality directly controls informed-AGD performance.
  • Balanced selection adaptation hyperparameters = β=100, eps=-15, L=50
    Taken from Yang et al. (2019) for component-wise proposal variance adaptation (Supplementary C); treated as recommended defaults rather than fitted to the present targets.
axioms (4)
  • domain assumption The target π is known pointwise up to a normalizing constant, and proposals can be drawn from it
    Standard MCMC setting; stated in the Introduction as 'density ... can be evaluated point-wise up to an unknown normalizing constant'.
  • domain assumption The involutive MCMC framework (Neklyudov et al. 2020) correctly yields valid reversible kernels for deterministic involutions with unit Jacobian
    CAM's detailed-balance proof (Sec 2.1) reduces to verifying that its transition is an instance of involutive MCMC; no standalone proof of detailed balance is given.
  • domain assumption ASMC produces a weighted particle approximation of the target and its annealed path
    The informed AGDs (Eqs 5-6) sample from ASMC empirical measures; the correctness of ASMC is imported from Moral et al. (2006) and the authors' prior papers (Supplementary A).
  • domain assumption The auxiliary variable z is drawn from a fixed distribution fi independent of the current state x
    The extended-state distribution (Eq 3) and the acceptance ratio cancel the fi(zi) factors only because z does not depend on x; a dynamic AGD (suggested in the Discussion) would break the presented proof.

pith-pipeline@v1.3.0-alltime-deepseek · 17113 in / 14251 out tokens · 141277 ms · 2026-08-01T23:09:02.110859+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Compound Auxiliary Metropolis: Incorporating Auxiliary Variables into Multi-Candidate MCMC." pith.science (2026). https://pith.science/paper/DPWMBEGP

@misc{pith2026260715499,
  author       = {Pith},
  title        = {Pith review of: Compound Auxiliary Metropolis: Incorporating Auxiliary Variables into Multi-Candidate MCMC},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DPWMBEGP}},
  note         = {Machine review of arXiv:2607.15499}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multiple-try Metropolis (MTM) is a Markov chain Monte Carlo (MCMC) algorithm that improves local transition efficiency by evaluating multiple candidate draws at each iteration. However, for complicated target distributions exhibiting severely non-Gaussian topography or multiple well-separated modes, locally optimal transitions may be insufficient for effective global exploration. In this work, we propose compound auxiliary Metropolis (CAM), a general multi-candidate MCMC method that incorporates both the local state of the chain and auxiliary information into the multi-candidate framework of MTM. Using an auxiliary generating distribution, CAM accommodates a flexible definition of auxiliary information. As examples, we consider three different auxiliary variables: one that promotes state-independent exploration and two that use a reference distribution to improve mixing. These auxiliaries are tested against distributions that present challenging targets for modern MCMC methods. In particular, we focus on the challenges presented by multiple well-separated modes and topography that requires long mixing for local MCMC moves. We find that CAM is able to sample effectively from these distributions, using MTM as a baseline to evaluate the benefit introduced by the auxiliary information. CAM also compares favourably with the No-U-Turn Sampler, showing similar performance for milder test distributions and better performance for the most difficult settings.

Figures

Figures reproduced from arXiv: 2607.15499 by Liangliang Wang, Renny Doig.

Figure 1
Figure 1. Figure 1: CAM with an uninformed discrete auxiliary. Left: Traceplots of the final 2,500 samples for both coordinates with red points indicating transitions sampled from the auxiliary variable. Top-right: Target distribution contours (black) with the discrete auxiliary grid (crosses). Bottom-right: The resulting CAM samples overlaid with grey points. This proof-of-concept highlights the core utility of the CAM frame… view at source ↗
Figure 2
Figure 2. Figure 2: CAM with an informed ASMC auxiliary. Left: Traceplots of the final 2,500 samples for both coordinates, with red points indicating auxiliary transitions. Top-right: Target contours (black) with ASMC auxiliary samples (crosses). Bottom-right: The resulting CAM samples overlaid in grey points. 3.2 Banana Distribution The banana distribution is a canonical benchmark for evaluating MCMC algorithms due to its hi… view at source ↗
Figure 3
Figure 3. Figure 3: Two-sample KSD for the Gaussian mixture distribution across internal CAM configurations. Points represent the median over all independent runs, with thick lines and thin lines denoting the 50% and 95% intervals, respectively. 0.0 0.2 0.4 0.6 10 20 30 Kolmogorov−Smirnov Distance Intermodal distance parameter (δ) Sampler CAM MH MTM NUTS [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Two-sample KSD benchmark comparing CAM against external MCMC methods on the Gaussian mixture distribution. Points represent the median over all runs, with 50% (thick lines) and 95% (thin lines) intervals [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: ESS per iteration for the first coordinate of the banana distribution across all CAM configurations. Points represent the median over all independent runs, with thick lines and thin lines denoting the 50% and 90% intervals, respectively. geometries (lower B), its performance degraded sharply as the target curvature became more extreme. CAM however, maintained consistent sampling efficiency across the entir… view at source ↗
Figure 6
Figure 6. Figure 6: ESS per iteration for the first coordinate of the banana distribution, benchmarking CAM against other MCMC methods. Points represent the median over all runs, with 50% (thick lines) and 90% (thin lines) intervals. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: ESS per iteration for each parameter in the eight schools model across internal CAM configurations. Points represent the median over converged runs, with thick and thin vertical lines denoting the 50% and 90% intervals, respectively. 0.00 0.05 0.10 0.15 0.20 θ1 θ2 θ3 θ4 θ5 θ6 θ7 θ8 µ τ Parameter ESS per iteration Sampler CAM MH MTM NUTS [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Benchmark ESS per iteration for each parameter in the eight schools model. Points represent the median across 50 independent runs, with thick and thin vertical lines denoting the 50% and 90% intervals, respectively. cases, using an uninformed discrete grid, with well-chosen boundaries, ensures a more complete exploration of the state-space, albeit at a lower per-iteration efficiency. When the target exhibi… view at source ↗

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

96 extracted references · 2 linked inside Pith

  1. [1]

    Andrieu, C. and J. Thoms (2008). A tutorial on adaptive MCMC . Statistical Computing\/ 18 , 343--373

    2008

  2. [2]

    Edelman, S

    Bezanson, J., A. Edelman, S. Karpinski, and V. B. Shah (2017). Julia: A fresh approach to numerical computing. SIAM review\/ 59\/ (1), 65--98

    2017

  3. [3]

    Surjanovic, S

    Biron-Lattes, M., N. Surjanovic, S. Syed, T. Campbell, and A. Bouchard-C\^ot\'e (2024, 02--04 May). autoMALA : Locally adaptive M etropolis-adjusted L angevin algorithm. In S. Dasgupta, S. Mandt, and Y. Li (Eds.), Proceedings of The 27th International Conference on Artificial Intelligence and Statistics , Volume 238 of Proceedings of Machine Learning Rese...

    2024

  4. [4]

    Craiu, and F

    Casarin, R., R. Craiu, and F. Leisen (2013). Interacting multiple try algorithms with different proposal distribution. Statistical Computing\/ 23 , 185--200

    2013

  5. [5]

    Chopin, N. (2004). Central limit theorem for sequential M onte C arlo methods and its application to B ayesian inference. The Annals of Statistics\/ 32\/ (6), 2385--2411

    2004

  6. [6]

    Doig, R. and L. Wang (2026). A unified framework for multiple-try M etropolis: Construction and empirical benchmarks

    2026

  7. [7]

    Douc, R. and O. Capp \'e (2005). Comparison of resampling schemes for particle filtering. In Image and Signal Processing and Analysis, 2005. ISPA 2005. Proceedings of the 4th International Symposium on , pp.\ 64--69. IEEE

    2005

  8. [8]

    Doucet, A. and A. Johansen (2009). A tutorial on particle filtering and smoothing: Fifteen years later. Handbook of Nonlinear Filtering\/ 12

    2009

  9. [9]

    Edwards, R. G. and A. D. Sokal (1988). Generalization of the F ortuin- K asteleyn- S wendsen- W ang representation and M onte C arlo algorithm. Physical Review Letters\/ 38 , 2009--2012

    1988

  10. [10]

    Fontaine, S. and M. B\' e ard (2022). An adaptive multiple-try M etropolis algorithm. Bernoulli\/ 28\/ (3), 1986--2011

    2022

  11. [11]

    Maire, and G

    Gagnon, P., F. Maire, and G. Zanella (2023). Improving multiple-try M etropolis with local balancing. Journal of Machine Learning Research\/ 24 , 1--59

    2023

  12. [12]

    Carlin, H

    Gelman, A., J. Carlin, H. Stern, D. Dunson, A. Vehtari, and D. Rubin (2013). Bayesian Data Analysis\/ (3 ed.). Boca Raton, USA: Chapman Hall/CRC

    2013

  13. [13]

    Goedman, R. J. and others (2024). Stan.jl

    2024

  14. [14]

    Goodman, J. and J. Weare (2010). Ensemble samplers with affine invariance. Communications in Applied Mathematics and Computational Science\/ 5\/ (1), 65--80

    2010

  15. [15]

    Green, P. J. (1992). A note on the S wendsen- W ang algorithm for ordered colours. Technical report, University of Bristol, Statistics Group

    1992

  16. [16]

    Ghahramani, and R

    Gu, S., Z. Ghahramani, and R. Turner (2015). Neural adaptive sequential M onte C arlo. In Advances in Neural Information Processing Systems

    2015

  17. [17]

    Hastings, W. K. (1970). Monte C arlo sampling methods using M arkov chains and their applications. Biometrika\/ 57\/ (1), 97--109

    1970

  18. [18]

    Higdon, D. M. (1998). Auxiliary variable methods for M arkov chain M onte C arlo with applications. Journal of the American Statistical Association\/ 93\/ (442)

    1998

  19. [19]

    Hoffman, M. D. and A. Gelman (2014). The N o- U - T urn S ampler: Adaptively setting path lengths in H amiltonian M onte C arlo. Journal of Machine Learning Research\/ 15 , 1593--1623

    2014

  20. [20]

    Liang, F. and W. H. Wong (2001). Real-parameter evolutionary M onte C arlo with applications to B ayesian mixture models. Journal of the American Statistical Association\/ 96\/ (454), 653--666

    2001

  21. [21]

    Liu, J. S., F. Liang, and W. H. Wong (2000). The multiple-try method and local optimization in M etropolis sampling. Journal of the American Statistical Association\/ 95\/ (449), 121--134

    2000

  22. [22]

    Marinari, E. and G. Parisi (1992). Simulated tempering: a new M onte C arlo scheme

    1992

  23. [23]

    Martino, L. (2018). A review of multiple try MCMC algorithms for signal processing. Digital Signal Processing\/ 75 , 134--152

    2018

  24. [24]

    Metropolis, N., A. W. Rosenbluth, M. Rosenbluth, A. H. Teller, and E. Teller (1953). Equation of state calculations by fast computing machines. Journal of Chemical Physics\/ 21 , 1087--1092

    1953

  25. [25]

    Moral, P. D., A. Doucet, and A. Jasra (2006). Sequential M onte C arlo samplers. J.R. Statist. Soc. B\/ 68 , 411--436

    2006

  26. [26]

    Neal, R. M. (2001). Annealed importance sampling. Statistics and Computing\/ 11 , 125--139

    2001

  27. [27]

    Neal, R. M. (2003). Slice sampling. The Annals of Statistics\/ 31\/ (3), 705--767

    2003

  28. [28]

    Welling, E

    Neklyudov, K., M. Welling, E. Egorov, and D. Vetrov (2020). Involutive MCMC : a unifying framework. In 37th International Conference on Machine Learning

    2020

  29. [29]

    Bartolucci, and N

    Pandolfi, S., F. Bartolucci, and N. Friel (2014). A generalized multiple-try version of the reversible jump algorithm. Computational Statistics and Data Analysis\/ 72 , 298--314

    2014

  30. [30]

    R: A Language and Environment for Statistical Computing

    R Core Team (2022). R: A Language and Environment for Statistical Computing . Vienna, Austria: R Foundation for Statistical Computing

    2022

  31. [31]

    Rubin, D. (1981). Estimation in parallel randomized experiments. Journal of Educational Statistics\/ 6\/ (4), 377--401

    1981

  32. [32]

    Salimans, T., D. P. Kingma, and M. Welling (2019). Markov chain M onte C arlo and variational inference: bridging the gap. In International Conference on Machine Learning

    2019

  33. [33]

    Smith, B. J. and others (2024). Mcmcchains.jl

    2024

  34. [34]

    Stan Reference Manual

    Stan Development Team (2024). Stan Reference Manual

    2024

  35. [35]

    Surjanovic, N., S. Syed, A. Bouchard-C\^ot\'e, and T. Campbell (2022). Parallel tempering with a variational reference. In Advances in Neural Information Processing Systems

    2022

  36. [36]

    Tanner, M. A. and W. H. Wong (1987). The calculation of posterior distributions by data augmentation. Journal of the American Statistical Association\/ 82\/ (398)

    1987

  37. [37]

    Gelman, D

    Vehtari, A., A. Gelman, D. Simpson, B. Carpenter, and P.-C. B\:urkner (2021). Rank-normalization, folding, and localization: An improved \^ R for assessing convergence of MCMC (with discussion). Bayesian Analysis\/ 16\/ (2), 667--718

    2021

  38. [38]

    Wang, and A

    Wang, L., S. Wang, and A. Bouchard-C\^ o t\' e (2020). An annealed sequential M onte C arlo method for B ayesian phylogenetics. Systematic Biology\/ 69\/ (1), 155--183

    2020

  39. [39]

    Wang, S., S. Ge, R. Doig, and L. Wang (2021). Adaptive semiparametric B ayesian differential equations via sequential M onte C arlo. Journal of Computational and Graphical Statistics\/ 31\/ (2)

    2021

  40. [40]

    Wickham, H. (2016). ggplot2 : Elegant Graphics for Data Analysis . New York City, USA: Springer-Verlag New York

    2016

  41. [41]

    Yang, J., E. Levi, R. Craiu, and J. S. Rosenthal (2019). Adaptive component-wise multiple-try M etropolis sampling. Journal of Computational and Graphical Statistics\/ 28\/ (2), 276--289

    2019

  42. [42]

    Zhou, Y., A. M. Johansen, and J. A. Aston (2016). Toward automatic model comparison: an adaptive sequential M onte C arlo approach. Journal of Computational and Graphical Statistics\/ 25\/ (3), 701--726

    2016

  43. [43]

    Reversible jump

    Green, Peter J , journal=. Reversible jump. 1995 , publisher=

    1995

  44. [44]

    Some adaptive

    Tierney, Luke and Mira, Antonietta , journal=. Some adaptive. 1999 , publisher=

    1999

  45. [45]

    Delayed rejection in reversible jump

    Green, Peter J and Mira, Antonietta , journal=. Delayed rejection in reversible jump. 2001 , publisher=

    2001

  46. [46]

    Delayed rejection

    Modi, Chirag and Barnett, Alex and Carpenter, Bob , journal=. Delayed rejection. 2023 , publisher=

    2023

  47. [47]

    Real-parameter evolutionary

    Liang, Faming and Wong, Wing Hung , journal=. Real-parameter evolutionary. 2001 , publisher=

    2001

  48. [48]

    Communications in Applied Mathematics and Computational Science , volume=

    Ensemble samplers with affine invariance , author=. Communications in Applied Mathematics and Computational Science , volume=. 2010 , publisher=

    2010

  49. [49]

    Liu and Faming Liang and Wing Hung Wong , title =

    Jun S. Liu and Faming Liang and Wing Hung Wong , title =. Journal of the American Statistical Association , volume = 95, number = 449, pages =

  50. [50]

    Statistical Computing , volume = 23, pages =

    Roberto Casarin and Radu Craiu and Fabrizio Leisen , title =. Statistical Computing , volume = 23, pages =

  51. [51]

    Silvia Pandolfi and Francesco Bartolucci and Nial Friel , title =

  52. [52]

    Computational Statistics and Data Analysis , volume = 72, pages =

    Silvia Pandolfi and Francesco Bartolucci and Nial Friel , title =. Computational Statistics and Data Analysis , volume = 72, pages =

  53. [53]

    An adaptive multiple-try

    Simon Fontaine and Myl\`. An adaptive multiple-try. Bernoulli , volume = 28, number = 3, pages =

  54. [54]

    A multiple-try

    Xin Luo and H\/. A multiple-try. Computational Statistics , volume = 34, pages =

  55. [55]

    Rosenthal , title =

    Jinyoung Yang and Evgeny Levi and Radu Craiu and Jeffrey S. Rosenthal , title =. Journal of Computational and Graphical Statistics , volume = 28, number = 2, pages =

  56. [56]

    The Annals of Applied Probability , volume = 8, number = 1, pages =

    Luke Tierney , title =. The Annals of Applied Probability , volume = 8, number = 1, pages =

  57. [57]

    Green , journal =

    Peter J. Green , journal =. Reversible jump

  58. [58]

    A Unified Framework for Multiple-Try

    Renny Doig and Liangliang Wang , year=. A Unified Framework for Multiple-Try. 2503.11583 , archivePrefix=

    Pith/arXiv arXiv

  59. [59]

    Sequential

    Pierre Del Moral and Arnaud Doucet and Ajay Jasra , journal =. Sequential

  60. [60]

    An annealed sequential

    Liangliang Wang and Shijia Wang and Alexandre Bouchard-C\^. An annealed sequential. Systematic Biology , year = 2020, volume = 69, number = 1, pages =

    2020

  61. [61]

    Statistics and Computing , volume = 11, year = 2001, pages =

    Annealed importance sampling , author =. Statistics and Computing , volume = 11, year = 2001, pages =

    2001

  62. [62]

    An adaptive sequential

    Pierre del Moral and Arnaud Doucet and Ajay Jasra , journal =. An adaptive sequential

  63. [63]

    Pigeons.jl:

    Surjanovic, Nikola and Biron-Lattes, Miguel and Tiede, Paul and Syed, Saifuddin and Campbell, Trevor and Bouchard-C. Pigeons.jl:. arXiv:2308.09769 , year=

    Pith/arXiv arXiv

  64. [64]

    Adaptive semiparametric

    Shijia Wang and Shufei Ge and Renny Doig and Liangliang Wang , journal =. Adaptive semiparametric

  65. [65]

    Neal , title =

    Radford M. Neal , title =. The Annals of Statistics , volume = 31, number = 3, year = 2003, pages =

    2003

  66. [66]

    Multivariate output analysis for

    Dootika Vats and James M Flegal and Galin L Jones , journal =. Multivariate output analysis for

  67. [67]

    Journal of Educational Statistics , year = 1981, volume = 6, number = 4, pages =

    Estimation in parallel randomized experiments , author =. Journal of Educational Statistics , year = 1981, volume = 6, number = 4, pages =

    1981

  68. [68]

    Bayesian Data Analysis , author =

  69. [69]

    Improving multiple-try

    Philippe Gagnon and Florian Maire and Giacomo Zanella , journal =. Improving multiple-try

  70. [70]

    Rosenbluth and Marshall Rosenbluth and Augusta H

    Nicholas Metropolis and Arianna W. Rosenbluth and Marshall Rosenbluth and Augusta H. Teller and Edward Teller , title =. Journal of Chemical Physics , volume = 21, pages =

  71. [71]

    W. K. Hastings , title =. Biometrika , volume = 57, number = 1, pages =

  72. [72]

    Nikola Surjanovic and Saifuddin Syed and Alexandre Bouchard-C\^ot\'e and Trevor Campbell , title =

  73. [73]

    Advances in Neural Information Processing Systems , year =

    Shixiang Gu and Zoubin Ghahramani and Richard Turner , title =. Advances in Neural Information Processing Systems , year =

  74. [74]

    Kingma and Max Welling , title =

    Tim Salimans and Diederik P. Kingma and Max Welling , title =

  75. [75]

    Stan Reference Manual , year = 2024, version =

    2024

  76. [76]

    Hoffman and Andrew Gelman , title =

    Matthew D. Hoffman and Andrew Gelman , title =. Journal of Machine Learning Research , volume = 15, pages =

  77. [77]

    Goedman and

    Rob J. Goedman and. Stan.jl , url =

  78. [78]

    2022 , url =

    R: A Language and Environment for Statistical Computing , author =. 2022 , url =

    2022

  79. [79]

    2016 , isbn =

    Hadley Wickham , title =. 2016 , isbn =

    2016

  80. [80]

    Smith and

    Brian J. Smith and. MCMCChains.jl , url =

Showing first 80 references.