REVIEW 3 major objections 6 minor 24 references
Phase space sampling with Markov Chain Monte Carlo methods
T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash
Pith's one-line read MCMC can sample the fully differential cross section of Z+3 jets without bias
desk verdict Useful feasibility study of MCMC phase-space sampling, but the unbiased-sampling claim needs the logit-Jacobian addressed before I'd trust it. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the interface between the BAT.jl MCMC package and the Sherpa event generator. A modified Sherpa executable initializes the process, then halts and waits for random numbers; BAT.jl proposes a point in the (logit-transformed) unit hypercube, Sherpa maps those numbers to four-momenta via Rambo or its multi-channel phase-space mapping, evaluates the matrix element and phase-space factors, and returns the event weight as the MCMC target density. The logit transformation $q(x)=\log(x/(1-x))$ is the main device for making the target tractable, and the effective-sample-size correction with randomized resampling is the device that converts correlated chain output into usable, approximately independent events.
What would settle it
Run the same interface on a low-dimensional toy target with a known density concentrated near the boundary, apply the logit transformation with and without the Jacobian factor, and compare the recovered marginal distribution to the true one; the match identifies the correct implementation. Alternatively, rerun the Z+3-jet comparison with the Jacobian intentionally omitted and check whether the leading-lepton $p_T$ and $m(\ell\ell)$ distributions shift away from the standalone-generator reference.
Extended reading notes
Core claim
The paper claims that MCMC algorithms can be used to sample the phase space of a scattering process and produce distributions of differential cross sections without bias. For the concrete case of $gg\to e^+e^- d\bar d g$ at $\sqrt{s}=13$ TeV, the authors run a Metropolis–Hastings sampler with RAM proposal tuning and the Gelman–Rubin convergence criterion, using the event generator to evaluate the matrix element for every proposed phase-space point. Sampling is performed on logit-transformed variables $q(x)=\log(x/(1-x))$ to move probability mass away from the boundaries, and the chain is run long enough that the effective sample size, estimated from autocorrelation, reaches the target. A final randomized resampling turns the correlated chain into approximately independent events; comparing the $p_T$ of the leading lepton and the invariant mass $m(\ell\ell)$ against a standalone generator run shows agreement within statistical uncertainties, with no bias observed.
Load-bearing premise
The unbiasedness result rests on the MCMC target density being exactly proportional to the differential cross section after the logit coordinate change; the paper never states whether the required Jacobian factor $|\mathrm{d}x/\mathrm{d}q|$ was included, so if it was omitted the chain would converge to a biased distribution.
Editorial extensions
If this is right
- MCMC phase-space sampling can produce physical distributions without constructing an explicit importance-sampling channel for every resonance or singularity.
- The same interface works with any sampler provided by the MCMC toolkit, so Metropolis–Hastings is only one of several algorithms that can drive the exploration.
- Autocorrelation is the main price: raw chain samples are not independent, and the number of useful events equals the effective sample size, not the chain length.
- Because output is written in HEPMC format, MCMC-generated events can enter the same detector-simulation and analysis chain as conventional generator events.
Reading between the lines
- The paper does not state whether the target density used by the sampler includes the Jacobian factor $|\mathrm{d}x/\mathrm{d}q|$ of the logit transformation; if that factor is missing, the chain would converge to a biased density, so reproducing the result requires checking this implementation detail.
- The reported autocorrelation lags of 300–1000 to reach 10% decorrelation imply that higher-multiplicity final states will need substantially longer chains, making wall-clock time per independent event the real test of practical viability.
- The same randomized-resampling recipe could be used as a principled unweighting procedure for MCMC samples, drawing events according to their weight instead of relying on a maximum-weight unweighting step.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper presents an interface between the BAT.jl MCMC library and the Sherpa event generator, and uses it to sample the phase space of a Z+3 jets process (specifically the partonic channel gg -> e+e- d dbar g at 13 TeV) with a Metropolis-Hastings algorithm and RAM tuning. A logit transformation is applied to the phase-space variables to improve sampling efficiency, and the resulting samples are compared with a standalone Sherpa run via two one-dimensional distributions (leading-lepton pT and dilepton invariant mass). The authors conclude that MCMC algorithms can sample the fully differential cross section without bias.
Significance. If substantiated, this is a useful proof-of-principle for MCMC-based phase space sampling as an alternative to traditional importance-sampling event generators, with potential benefits for multi-modal and high-dimensional phase spaces. The work leverages existing, well-tested tools (BAT.jl, Sherpa) and performs a direct comparison against an independent generator, which is a sound self-consistency check. The paper also raises a legitimate practical issue, autocorrelation, and attempts to account for it via effective sample size matching. However, the central unbiasedness claim currently rests on a small amount of validation and on a missing technical detail (the Jacobian of the logit transformation), so the evidence is not yet commensurate with the strength of the conclusion.
major comments (3)
- [Section 3] The logit transformation is introduced without stating the Jacobian of the transformation. For a target density p(x) defined on the unit hypercube, sampling in q = log(x/(1-x)) requires the transformed density p(q) = p(x(q)) |dx/dq| = p(x(q)) x(q)(1-x(q)). If this Jacobian factor is omitted, the chain's stationary distribution is biased by a factor 1/[x(1-x)] relative to p(x). Since the unbiasedness claim in Section 4 rests on sampling in this transformed space, the paper must state explicitly how the Jacobian is handled (e.g., in the target function passed to the sampler) and provide the corresponding derivation or a precise reference to the implementation.
- [Section 3, Fig. 4] The conclusion that the MCMC samples reproduce the differential cross section 'without bias' is supported only by visual agreement of two one-dimensional distributions. The p-values from the claimed chi2 test are not reported in the text or figure, and no other observables or multi-dimensional correlations are shown. For a 13-dimensional phase space, agreement in these two marginals is necessary but not sufficient to establish unbiased sampling. Please report the p-values and chi2/dof values, and add further validation such as jet transverse momenta, angular separations, and at least one two-dimensional distribution or a multivariate two-sample test.
- [Section 2.2, Eq. (7)] The effective sample size formula is incorrect as written: ESS = 1/(1 + 2*N*sum_{k} rho(k)) should read ESS = N/(1 + 2*sum_{k} rho(k)). The current formula would give a dimensionally inconsistent result and would misestimate the number of independent samples. Because the comparison in Section 3 matches the number of Sherpa events to the ESS, this must be corrected and the impact on the uncertainty estimate and on the comparison checked.
minor comments (6)
- [Section 3] The number of MCMC samples, burn-in length, number of chains, and the Gelman-Rubin convergence values are not reported; these details are needed to assess the reliability of the sampling and the convergence criterion.
- [Section 3, Fig. 4] The p-values mentioned in the text and figure caption are not displayed; please provide the numerical p-values (and chi2/dof) either in the figure, its caption, or the text.
- [Section 3] The paper states that the number of Sherpa events is matched to the ESS, but it does not specify the actual numbers or the target ESS used; please provide these values to make the statistical comparison reproducible.
- [Throughout] The text contains numerous LaTeX artifacts from the source, such as 'di fferent', 'di fferential', and 'S HERPA'; these should be cleaned before publication.
- [Section 3] The process definition should clarify whether the Z boson is treated as on-shell or whether the dilepton invariant mass distribution is generated directly with off-shell effects; the cut m(l+l-) > 66 GeV suggests the latter, but this is not stated.
- [Section 3] It is not stated whether the events written to the HEPMC file are unweighted (as the MCMC samples would be after convergence) or whether event weights are included; this should be clarified for the comparison with Sherpa.
Circularity Check
No significant circularity: the MCMC validation is a self-consistency check against the same Sherpa-computed target, with no fitted parameters forcing the result.
full rationale
The paper's central claim is that MCMC algorithms implemented in BAT.jl can sample phase space according to a Sherpa-computed differential cross section. The target density is obtained by evaluating the Sherpa matrix element on proposed phase-space points, and the benchmark is Sherpa's own sampling of the same density. This is a self-consistency check, not a circular reduction: the MCMC sampler is an independent algorithm, its proposal is tuned by the standard RAM scheme, and no parameter of the comparison is fitted to force agreement. The conclusion 'without bias' is properly scoped to agreement with Sherpa, and that agreement is meaningful evidence that the sampler reaches the intended target. Prior work by the same authors (BAT.jl, the earlier MC^3 study, and nested-sampling investigations) is cited as context and software provenance, but the central derivation does not rest on any uniqueness theorem or fitted quantity. The missing Jacobian factor for the logit transform noted in review is a documentation and correctness concern rather than a circularity, because it does not make the output equal to an input by construction. The comparison is between two independent sampling algorithms for the same target, so no circularity is present.
Assumptions & free parameters
assumptions (4)
- standard math Metropolis-Hastings with RAM tuning converges to the target distribution
- domain assumption The Sherpa matrix element evaluation provides the correct differential cross section for Z+3 jets
- ad hoc to paper The logit-transformed target density is correctly implemented including any Jacobian factor
- domain assumption Randomized resampling of autocorrelated samples yields approximately i.i.d. samples
Cite this review
Pith. "Pith review of Phase space sampling with Markov Chain Monte Carlo methods." pith.science (2026). https://pith.science/paper/I5UUFFWV
@misc{pith2026241212963,
author = {Pith},
title = {Pith review of: Phase space sampling with Markov Chain Monte Carlo methods},
year = {2026},
howpublished = {\url{https://pith.science/paper/I5UUFFWV}},
note = {Machine review of arXiv:2412.12963}
}
abstract
We present a study on using Markov Chain Monte Carlo (MCMC) techniques to explore the high-dimensional and multi-modal phase space of scattering events at high-energy particle colliders. To this end, we combine the BAT.jl package that provides implementations of a variety of MCMC algorithms with the Sherpa event generator framework. We discuss technical aspects of the implementation and the resulting algorithm and present first results for the process of $Z+3$ jets production at the LHC.
Figures
Figures from the paper (1 more)
Reference graph
Works this paper leans on
-
[1]
V .D. Elvira et al., The Future of High Energy Physics Software and Computing , in Snowmass 2021 (2022), 2210.05822
arXiv 2022
-
[2]
J.M. Campbell et al., SciPost Phys. 16, 130 (2024), 2203.11110
arXiv 2024
-
[3]
S. Amoroso et al. (HSF Physics Event Generator WG), Comput. Softw. Big Sci. 5, 12 (2021), 2004.13687
arXiv 2021
- [4]
-
[5]
N. Metropolis, A.W. Rosenbluth, M.N. Rosenbluth, A.H. Teller, E. Teller, J. Chem. Phys. 21, 1087 (1953)
work page 1953
-
[6]
T. Gleisberg, S. Höche, F. Krauss, M. Schönherr, S. Schumann, F. Siegert, J. Winter, JHEP 02, 007 (2009), 0811.4622
arXiv 2009
- [7]
- [8]
Show all 24 references
-
[9]
Schulz, F
O. Schulz, F. Beaujean, A. Caldwell, C. Grunwald, V . Hafych, K. Kröninger, S. La Cagnina, L. Röhrig, L. Shtembari, SN Comput. Sci. 2, 1 (2021), 2008.03132
2021 arXiv
- [10]
- [11]
- [12]
-
[13]
Bothmann, T
E. Bothmann, T. Janßen, M. Knobbe, T. Schmale, S. Schumann, SciPost Phys. 8, 069 (2020), 2001.05478
2020 arXiv
-
[14]
C. Gao, S. Höche, J. Isaacson, C. Krause, H. Schulz, Phys. Rev. D 101, 076002 (2020), 2001.10028
2020 arXiv
-
[15]
Hastings, Biometrika 57, 97 (1970)
W.K. Hastings, Biometrika 57, 97 (1970)
1970
-
[16]
Geman, D
S. Geman, D. Geman, IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-6, 721 (1984)
1984
-
[17]
Kharraziha, S
H. Kharraziha, S. Moretti, Comput. Phys. Commun. 127, 242 (2000), [Erratum: Com- put.Phys.Commun. 134, 136–138 (2001)], hep-ph/9909313
2000 arXiv
-
[18]
Kroeninger, S
K. Kroeninger, S. Schumann, B. Willenberg, Comput. Phys. Commun. 186, 1 (2015), 1404.4328
2015 arXiv
-
[19]
Yallup, T
D. Yallup, T. Janßen, S. Schumann, W. Handley, Eur. Phys. J. C 82, 8 (2022), 2205.02030
2022 arXiv
- [20]
-
[21]
Buckley, P
A. Buckley, P. Ilten, D. Konstantinov, L. Lönnblad, J. Monk, W. Pokorski, T. Przedzin- ski, A. Verbytskyi, Comput. Phys. Commun. 260, 107310 (2021), 1912.08005
2021 arXiv
- [22]
-
[23]
Vihola, Statistics and Computing 22, 997 (2012)
M. Vihola, Statistics and Computing 22, 997 (2012)
2012
-
[24]
Gelman, D.B
A. Gelman, D.B. Rubin, Statistical Science 7, 457 (1992) 8
1992
Reviewed August 11, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.