acceptodds
Under review as a conference paper at ICLR 2027

Quenched Ensemble Sampling

Abstract

Some of the sharpest challenges in sampling from the energy functions of physical systems arise at phase transitions, where the density of states changes abruptly and many sampling algorithms stall. Nested sampling is a particle method that traverses the density of states under a hard energy constraint and is known to be robust to such transitions, but its application in high dimension is limited by the difficulty of sampling under that constraint. In this work we introduce Quenched Ensemble Sampling, which generalises the hard constraint to a family of repulsive potentials at the energy boundary. This preserves the quenched path of monotonically decreasing energy while making the constrained target amenable to scalable gradient-based kernels. We demonstrate on synthetic models of phase transitions that our method estimates the marginal likelihood and draws posterior samples across a first-order transition where popular alternatives such as tempering fail. We apply the procedure to marginal likelihood estimation in Bayesian neural networks, enabling model comparison between network architectures. Finally, in a high-dimensional continuous lattice field theory, we show that this method traverses a first-order transition and estimates the partition function.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Sampling, Boltzmann distributions, Sequential Monte Carlo, Probabilistic ML, Bayesian methods

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Vt1TOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Quenched Ensemble Sampling (QES), a sequential Monte Carlo method that aims to extend nested-sampling-style inference to high-dimensional problems. QES replaces the hard energy constraints of traditional nested sampling with softened constraints, allowing gradient-based MCMC kernels while preserving the quenched energy path. It uses SMC importance weights to estimate marginal likelihoods and reconstruct posterior samples.

The authors evaluate QES on synthetic distributions, a high-dimensional lattice model, and Bayesian neural networks on real datasets. QES successfully traverses first-order phase transitions where the tested tempered SMC methods fail, while achieving comparable predictive performance and marginal likelihood estimates on real-world BNN tasks. The paper's contributions are primarily methodological and empirical, with limited new theoretical performance guarantees.

Critical strengths

1. Well-motivated and mathematically grounded methodology: The paper addresses a clear limitation of traditional nested sampling: the difficulty of using gradient-based MCMC kernels under hard energy constraints. QES combines softened energy constraints with SMC-based volume estimation, providing a coherent framework for posterior sampling and marginal likelihood estimation.

2. Promising results on challenging phase-transition problems: QES demonstrates substantial improvements over tempered SMC on the phase-transition tasks. In particular, it accurately estimates the marginal likelihood in high-dimensional spike-and-slab problems where tempered SMC misses an important mode. These results suggest that QES could be valuable for sampling problems where conventional tempering struggles.

Critical weaknesses
  1. The softened energy constraints are closely related to existing microcanonical and demon-based nested sampling methods, while SMC-based volume estimation has also been studied previously. Although QES combines these components into a practical framework, the paper does not sufficiently establish what fundamentally new capabilities this combination provides, neither theoretically nor practically. In particular, the choice of a dimension-independent softness parameter appears to be a design choice rather than a substantial methodological innovation.

  2. A central motivation of this paper is to make nested sampling scalable to high-dimensional problems. However, the scalability comparison is mainly conducted on a simple Gaussian target, while the more challenging high-dimensional phase-transition experiments omit competitive NS baselines. Although the authors attribute these omissions to computational cost, they could report the performance of NS methods under a fixed computational budget rather than omitting the comparisons entirely. Furthermore, comparing gradient evaluations for QES with likelihood evaluations for NS does not establish actual computational efficiency. These limitations weaken the evidence for QES's main claimed advantage over existing NS methods.

Questions for authors
  1. Could the authors clarify the key methodological differences between QES and existing microcanonical or demon-based nested sampling methods, particularly Habeck (2015)? Beyond the dimension-independent softness parameter and SMC-based volume estimation, what specific advantages does QES offer that these approaches do not already provide?

  2. The paper omits NS results on the more challenging high-dimensional benchmarks due to computational cost. Could the authors clarify what computational budgets were attempted and how the NS baselines performed before termination? Even incomplete runs under a fixed budget could help assess whether QES provides a meaningful computational advantage over existing NS methods.

  3. The paper uses substantially deeper, problem-specific stopping thresholds for the phase-transition experiments to ensure that QES reaches the low-energy phase. In practical applications where the location of a phase transition is unknown, how should users determine an appropriate stopping criterion? Could the default criterion terminate before discovering an important low-energy phase, potentially leading to inaccurate evidence estimates?

I would be happy to raise my score if the authors can satisfactorily address these concerns.

2 comments

We thank the reviewer for their thorough reading of our paper. We think the questions raised are very useful, and appreciate the opportunity to provide clarifications.

1. Implications of prior work

The key point is that Habeck does not concretely demonstrate an advantage over classical gradient-free nested sampling. In the 32×3232\times32 Potts comparison, “Demonic nested sampling took approximately one third longer than standard nested sampling”, with similar evidence errors, leading to the conclusion that “there seems to be no advantage for the particular case of the Potts model…”. We think our results help isolate why the exponent matters: we demonstrate that this is a family of paths interpolating between annealing and nested sampling, and empirically identify when the quenched advantage is retained (Appendix B, Fig. 5). We think this is useful pedagogy alongside the empirical and implementation novelty.

Secondly, we focus on continuous targets, rather than the main emphasis of demonic NS, and give a simple identity integrating out the soft boundary term, which admits a wider range of MCMC kernels. Compared with Habeck’s bespoke HMC, we found MALA more effective: HMC trajectories still have to respect the energy boundary, so long trajectories can be rejected and wasted. Langevin sampling was also easier to tune.

Beyond softness, we emphasise improved SMC accounting (important for consistent results at scale), a parallel implementation suited to modern accelerators, and a framing connecting to modern MCMC variants. We think the overall package delivers a substantial practical advance over previous methods, demonstrating efficient sampling and evidence estimation on challenging high-dimensional problems where classical NS is prohibitively expensive and tempered SMC fails. This is supported by experiments that isolate the important components, alongside a synthesis of research threads less familiar at venues like ICLR.

2. Demonstration of scaling of classical nested sampling

We thank the reviewer for this suggestion, which we have explored. The NS baselines stop at D=102D=10^2 primarily for cost reasons. With the standard 2D2D slice steps, versus approximately D\sqrt{D} MALA steps, QES has a more favourable mutation budget, and NSS already shows significantly more bias on the 100D100D Gaussian.

Beyond wall time, linear scaling of slice steps need not maintain mixing at scale. Theoretical arguments give spectral-gap bounds with dimension dependence as unfavourable as cubic for an idealised hybrid slice sampler (Power et al.), suggesting that the usual linear heuristic is not necessarily sufficient. Even gradient-guided NS, such as GGNS (Lemos et al., 2024), using reflective Hamiltonian moves, demonstrates unimodal Gaussian examples up to 256D256D. We are therefore unaware of convincing general-purpose NS demonstrations at the dimensions reached by QES.

We have now tested NSS on the 40×4040\times40 lattice, using our other NS baseline configuration: 1,000 live points, 10% replacement per iteration, and 2D=32002D=3200 slice steps per replacement. We stopped after one hour on an RTX 5090, whereas QES converges in approximately seven minutes on the same GPU. NSS failed to converge: slice moves repeatedly hit a conservative bracket-shrinkage cap, mixing deteriorated, and after an hour the particles had degenerated to a single point. The NSS run consumed a similar order of target evaluations to QES (∼109\sim10^9, albeit likelihood versus gradient evaluations), yet reported an ordered share of zero, failing on the primary metric of interest.

We will include this theoretical context and practical test in the revision.

Weak Poincaré inequality comparisons for ideal and hybrid slice sampling, Power et al.

3. Termination

We agree that the draft gives too little guidance on termination and that a run can stop before reaching a low-energy phase, especially near first-order transitions. There are useful practical mitigations.

Firstly, runs can be resumed and extended until log⁡Z\log Z and relevant observables settle. In materials or molecular applications, independent experimental or simulation information often indicates the stable energy, so a quench ending above it should continue. A known global lower bound on energy can also replace the lowest observed energy in the criterion, making it more conservative.

Because QES is a quench, one can also continue until walkers collapse onto a single energy. This is expensive, but may be a natural endpoint when the ground state itself is of interest, although it is not a guarantee of global convergence. For Bayesian benchmarks we use the nested-sampling criterion for a fairer comparison with tempered SMC, with deeper follow-up runs where warranted.

A possible future extension is curvature-based termination, inspired by the use of L-BFGS to quench molecular configurations. This could provide a useful convergence diagnostic, though not a guarantee against undiscovered phases.

Thank you for the clarifications and additional experiments. The new comparison addresses my main concern about computational performance. Although the comparison remains limited to one NS implementation, it strengthens the evidence for QES’s practical contribution. Overall, while some concerns about novelty and the scope of the empirical claims remain, the additional evidence has improved my assessment. I will raise my score.

Reviewer_74dcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes an MCMC sampling algorithm for estimating marginal likelihoods. The method builds on nested sampling and uses quenching to progressively reach low-energy regions. Experiments are performed on toy distributions, UCI classification tasks (exhibiting phase transitions in model training) and a high-dimensional ϕ6\phi^6 theory lattice. The proposed method is compared against baselines such as the tempered sequential Monte Carlo algorithm [2001].

Critical strengths

The paper presents an algorithm for sampling – using quenching to sample Boltzmann distributions especially for systems at phase transitions. Quenching provides an algorithm to progressively sample low energy regions. It could be useful for sampling near phase transitions or multi-modal distributions without introducing biases.

The proposed method uses MALA and involves no learning, making it computationally efficient.

Critical weaknesses

About evaluation: The original claim of the paper is to sample systems at phase transitions. But there are no rigorous experiments around phase transitions. There are various recent methods attempting the same problem that are not compared with; the current baselines are weak. Moreover, the metrics chosen for evaluation are not clear. Standard metrics for evaluating sampling at phase transitions are missing.

Presentation: The paper is very difficult to follow. Terms are used without being defined. I am unable to understand many sentences even after careful reading and skimming through some of the cited papers. This makes the core method difficult to follow.

Questions for authors

Sec 1:

  • Line 54: What are kernels? E.g., MCMC kernels, MALA kernels.
  • Line 57: Motivations from statistical physics: Please pinpoint exactly what concepts come from statistical physics (as used in this paper) which are not yet known/popular in Bayesian inference community?

Sec 3:

  • Line 233: What are weighs, wiw_i, of the particles? Maybe line 254 talks about it though not very clearly.
  • Line 234: What is mutation?
  • Line 253: What is bisection?
  • Line 261: What is quadrature?

Sec 4:

  • Line 426: What is the prior π0(ϕ)\pi_0(\phi); what is exactly samplable free field? Also, in typical Boltzmann distributions, the term ϕ0\phi_0 does not appear while estimating the partition function ZZ, I think.
  • Line 430: What is coexistence? Some questions about experimentation:
  • What are the baselines? There are many recent works on sampling around critical regions (phase transitions). E.g.
  • “Conditional normalizing flow for Markov chain Monte Carlo sampling in the critical region of lattice field theory,” DOI: 10.1103/PhysRevD.107.014512
  • “Multilevel Generative Samplers for Investigating Critical Phenomena,” ICLR 2025.

It will be interesting to know how the proposed sampling algorithm compares to such state-of-the-art methods. A common problem observed during phase transitions is mode collapse, i.e., the distribution has multiple modes, but the model samples only some of them, leading to bias. There are several recent papers addressing this problem, such as:

  • “Progressive Inference-Time Annealing of Diffusion Models for Sampling from Boltzmann Densities,” NeurIPS 2025.
  • “Flow Annealed Importance Sampling Bootstrap,” ICLR 2023.

Is the proposed method robust to mode-collapse?

  • What are the metrics? The currently reported metrics are not explained. E.g., for ϕ6\phi^6 experiments, what is the ground truth and how much is the estimation error? These are not clear from the paper. There are certain metrics useful to assess sampling at phase transitions; such as statistics (Earth mover’s distance, mean squared error, etc.) of different observables (energy, susceptibility, etc.). Metrics such as negative log-likelihood and reverse negative log-likelihood are useful for assessing how the model samples across multiple modes. See the above references for details.
3 comments

We thank the reviewer for the thorough reading of our work, and we appreciate the time that has been put into this review. Before responding to the specific questions there are a couple of strong assertions in this review that we think aren’t correct that we would push back on.

Firstly the weakness claims there are no rigorous experiments around phase transitions. This isn’t true, the phi6phi^6 lattice and spike-slab analytic example are both exceptionally challenging high dimensional problems and have clear first order phase transitions built in. Secondly there is a claim made about weak baselines and missing citations, the key distinction here is first order, which we are very clear on throughout the draft that we are targeting this class of problems. There is a lot of interesting things to say about second order phase transitions, and indeed we do think that both our baseline SMC and QES would be useful here, however this is a separate future work. There is very little literature at venues like ICLR on first order phase transitions, despite the high scientific impact of these kinds of problems. We will deal with baselines at more length in this comment before getting onto the specific questions.

As this is a subtle point we will include a bit more of a retrospective on first and second order phase transitions. We will add a paragraph to the appendix laying out some thoughts on particle methods and second order phase transitions and what the different challenges are. Similarly we do recognise that some of the technical description in this draft is too dense and this does need to be made more accessible, so we will work through the draft and review this, perhaps relegating some of the denser technical description to the appendix.

Baselines and prior art

One of the questions is how useful QES is on multimodal distributions. The answer is very! The bimodal target tests sharply bimodal target mode recovery up to 200D (Fig 6.). The BNN sampling problems are massively multimodal (labelling degeneracy in MLP networks), whilst mode recovery is not tested, robustness of tuning is certainly sharply tested here. And both the spike and slab and lattice targets are sharply multimodal, fig 3. shows coordinate views on both where this is very apparent. Running on some other targets such as lattice ϕ4\phi^4 or alanine dipeptide that commonly appear in venues like ICLR is on the radar.

Baselines is one of the areas that the review score seems to rest heavily on and we have a lot to say here, we will try and keep this short. We think whilst we do not compare to a broad battery of samplers, we chose a very precise baseline that runs throughout that isolates the particular novelty of this work, the path of distributions. Once appropriately tuned, adaptive tempered SMC is exceptionally strong as a baseline, and often SMC is underreported/represented in venues like ICLR, often using a very undertuned random walk based implementation. We provide well written self tuning, performant jax implementations of both main algorithms and we hope these are a useful byproduct. We also include NUTS and MCLMC baselines on BNN experiments which are generally understood to be state of the art.

As to why we don’t include any learned sampler baselines, these do not tend to appear in advanced MCMC method submissions at these kinds of venue and we do not think they would be particularly instructive. Firstly despite the successes of learned samplers, general purpose learned samplers typically take significantly longer in walltimes and don’t demonstrate sufficiently attractive scaling yet to be relevant to the regimens of 40x40 lattices or training small BNNs.

On the subject of prior art at venues like ICLR related to phase transitions, the two works provided are both concerned with critical slowing down, and hence second order phase transitions, this is a different family of problems to the ones this draft focuses on so these aren’t particularly relevant to the main text, however we will include these in the appendix discussion we propose to add.

Lastly a key point is that a number of the provided refs build their contribution around annealing, a central contribution of this work is that there are problems, and we do think we sufficiently demonstrate this, where annealing itself is the problem. The inner MCMC procedure chosen, whether thats a tuned MALA as in our SMC, or a learned proposal in FAB or the like doesn’t change the fact that for first order phase transitions, the canonical Boltzmann target or tempered variants seriously struggle.

Questions:

General and notation

  • A number of these questions relate to definitions that we think personally are reasonable to assume as prerequisite knowledge for a submission to probabilistic methods at ICLR. We cite multiple pedagogical textbook sources (Doucet et al 2001 and Fearnhead et al 2025, we will additionally include Chopin & Papaspiliopoulos 2020, “An Introduction to Sequential Monte Carlo”, as this is a useful modern standard reference that we didn’t initially include). We think we quite clearly differentiate where we are leaning on standard SMC/MCMC results in the current draft, and don’t think these suggestions would improve the exposition.

  • We used the phrase “We formulate this procedure using motivations from statistical physics” semantically this just means we will use language of statistical physics to help motivate this discussion which we think we do. Table 5 and various parts of the exposition use concepts such as the density of states to motivate the construction. We are not claiming that these concepts are unknown to the Bayesian inference community, nor do we think that is implied by the quoted sentence. Indeed, the connections between the two fields are part of what we are trying to emphasise. We are therefore not entirely sure what the second part of the question is asking.

Lattice Field theory

  • On the question about chosen metrics, we are assuming this refers pretty much exclusively to the lattice result? On the synthetic targets we think ln⁡Z\ln Z is quite clear as the primary target, or mode balance for the bimodal target, estimating what most of the literature calls the intractable normalising constant is an extremely strict metric that is hard to cheat. On BNNs we include NLL which is standard. For the lattices the question seems to be why not use various point cloud distance metrics such as MMD and W2. We were initially not clear what method would be able to form a reasonable reference cloud on these problems, however we can add a reference calculation by forcing initialisation of two separate tempered SMC runs to be in the ordered and disordered phase and pool them appropriately, we will detail this in the appendix
MethodScheduleLevelslog⁡Z^\log \hat{Z}Ordered shareW2(m)W_2(m)W2(q)W_2(q)
QES0.9511759±211759 \pm 2−53.49±0.35-53.49 \pm 0.350.634±0.1380.634 \pm 0.1380.359±0.1670.359 \pm 0.1670.469±0.2190.469 \pm 0.219
Tempered SMC0.9520-54.540.0001.0441.378
Tempered SMC0.9944-54.540.0001.0441.378
Tempered SMC0.999137-54.540.0001.0441.378
Tempered SMCFixed10,000-54.540.0001.0441.378
NUTS———0.0001.0431.376
Reference——−53.48±0.02-53.48 \pm 0.020.655±0.0050.655 \pm 0.005——

We propose rewriting the results table for this as above, which uses the order parameters m and q to reduce to appropriate coordinates, as doing full point cloud W2 in 1600D is quite noisy (and using these order parameters is quite standard for lattice field theory, and match fig3.). We think this is a very useful addition for this venue as it uses a more familiar style of metric but retains what we think is a more appropriate definition for the evaluation coordinates.

This was a very helpful prompt for us, previously we were setting coexistence point as detailed in A.3, but we wrongly stated that coexistence implied a 50:50 split, but this doesnt take into account that there are two ordered and 1 disordered phase, so the split should be 2:1. We now independently measure what the chosen hyperparams give as the actual split between sectors (as we are only roughly determining this) and we now measure the actual phase proportions between ordered and disordered which is closer to 2:1 than 1:1. We will use this as an opportunity to expand on the concept of coexistence in the draft as suggested. The QES results have also changed, the runs were not being properly reseed so the scatter was too tight, having fixed this the inter seed variance does increase but the results are now consistent with the reference calculation that’s included, and the other potential methods all demonstrably fail.

For completeness, coexistence can be sufficiently understood as the point in $ \mu, \lambda, \eta $ space where the different phases carry equal probability. All of this language is just saying we choose a point in physics parameter space that is particularly challenging as it has a manifest first order phase transition.

We do recognise that the description around the lattice experiment is a quite terse, and we detailed in response to tj1n what our proposed solution is to clear up some of the confusion around this

Area_Chair_ytdQOpenReview ↗

rebuttal response

The rebuttal period is quickly approaching the end. Please go over the authors’ rebuttal and revision, and add acknowledgment and further comments if needed.

Reviewer_tj1nOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper tackles the problem of sampling from a posterior distribution which is given by a product between the prior and a target distribution. Their method builds upon nested sampling, which samples states from a distribution under hard energy constraints, that can be realized through sequential monte carlo. The overall target distribution is however constructed by a product distribution between a distribution over energy levels and a the aforementioned conditional distribution under hard energy constraint on this energy level. Thus, the second challenge which this algorithm seems to tackle is to sample from the energy level distribution. The paper is evaluated on several challenging sampling tasks such as Bayesian Neural Networks and the Lattice Φ\Phi 6 model at first order phase transition. Compared to evaluated baseline methods their method is competitive or even outperforms tempering based algorithms on modeling first order phase transitions.

Critical strengths
  1. The paper studies an interesting problem of sampling from intractable target distributions
  2. their algorithm seems have better mode coverage than baseline methods
Critical weaknesses

Due to poor writing I think the paper is hard to understand. While the authors use a lot of math in the Method section to explain their method, I believe the authors do a poor job at explaining why they are deriving these equations and how they are used in the end (How they map to the resulting algoritm). Then in the Algorithm section they mostly describe their algorithm through text and thus it is hard to parse which part of the algorithm corresponds to which equation. While there is a pseudoalgorithm in the appendix it seems to use terms that were not introduced before and thus it is not very helpful. Examples are the terms "donor weight" and "pre-mutation position", which only seem to appear in the pseudoalgorithm.

Additionally some equations are not sufficiently defined in order to understand what they are for. For example I do not understand what Eq. 13 is for and σ\sigma and MM seem not to be defined anywhere.

The authors introduce the background using Bayesian inference but then in the experiments they also experiment on the \phi-6 model. So I am wondering why they do not introduce the problem in a more general way and then in the experiments they could simply explain Bayesian Inference as an example.

Questions for authors
  1. Can the algorithm also be used to sample from any target distribution or does it have to be a product between a target and a prior distribution? In Sec. 4.3 is sampling performed on the product between prior and target or only the target? It seems sampling is performed on a product distribution. Isn't the goal there to sample from the target distribution without the prior?
  2. If I understand correctly the algorithm constructs the posterior by sampling a level p(E) and a particle from the corresponding energy constrained distribution p(x|E). is that correct? While I understand how the algorithm samples from p(x|E) I do not understand why and how the algorithm samples from p(E). can the authors please explain which part of the algorithm realizes this and why?
  3. Where is MM and σ\sigma which is used around Eq. 13 defined and properly explained?
3 comments

Many thanks to the reviewer for taking the time to read our work. We think the overall score of this review is exceptionally harsh, there are no issues with novelty or the empirical results and the issues raised seem to be mostly around the exposition. An alternative to tempering at scale that we don’t think has ever been convincingly demonstrated before is, we think, a really cool result! We appreciate there are places where the writing is quite dense, and we appreciate the opportunity to help improve the readability. However, we have presented this work in local meetings and generally had good feedback on the presentation and content. We spent quite a bit of time on the general higher level exposition in the intro, conclusion and background, so we are quite surprised to see this receive the score it does, although we appreciate there are always areas where particularly the denser technical methods writing can be improved.

To answer the explicit questions:

  1. Aren’t all posterior distributions a product of prior and likelihood? Yes QES can be used to sample from “any target distribution” in the same sense that any SMC/MCMC method can. Like these other sampling methods, QES does require a practical means to initialise particles (which is often a normalised reference distribution), however this does not restrict the target to Bayesian posterior distributions, and the reference measure does not necessarily need a Bayesian interpretation.

We appreciate the text around the lattice field theory in the main text is quite compressed, the Gaussian free field is defined by the Klein-Gordon Lagrangian so the total (integral form) of the action is,

S[ϕ]=∫d2x[12(∇ϕ)2+m022ϕ2+μ2−m022ϕ2+λϕ4+ηϕ6].S[\phi] = \int d^2x [\frac12(\nabla\phi)^2 + \frac{m_0^2}{2}\phi^2 + \frac{\mu^2-m_0^2}{2}\phi^2 + \lambda\phi^4 + \eta\phi^6].

Which is then split up into the pieces (by grouping the reference mass term and the kinetic term into the free field term), we use to define our sampling target as,

p(ϕ)∝π0(ϕ)e−U(ϕ)∝e−S[ϕ].p(\phi) \propto \pi_0(\phi)e^{-U(\phi)} \propto e^{-S[\phi]}.

We can certainly include the above equations in the draft, we think this does improve the clarity of this section. We do note however that splitting the free field and interacting potential out and describing it as the current draft does is quite standard for lattice field theory problems.

  1. The algorithm never samples from P(E)P(E), you effectively get the importance weight of each energy level from the size of the jump between successive levels, and the running estimate of G(E)G(E). The technical description of this essentially follows through section 3.2. The analog in more typical tempered SMC is laid out in Appendix C.3, where if you wanted to pool samples from all levels (which is sometimes done), you would assign an importance weight to each level in a similar fashion without needing to actually sample P(β)P(\beta), neither construction need to sample from the marginal distribution of P(E)P(E) or P(β)P(\beta) to reconstruct the target

We appreciate that this part is hard to grasp so we are reviewing this method section to make it more accessible, we are quite honest in the draft that this is standard SMC construction in essence, and we therefore lean on the existing literature and assume some familiarity with these methods.

  1. σ\sigma is defined by equation 13 itself? We use this notation as often the mass matrix would be derived from the standard deviation of the particle cloud, but as we motivate later in the paragraph this is not ideal for multimodal distributions so the score based metric we design for this paper is more robust. MM is defined in section 2.2 as the mass matrix/pre-conditioner of the MALA kernel we implement, we think quite clearly? In terms of not understanding what (13) is for, we deliberately kept the description of the mutation kernel as brief as possible, as we emphasise really any gradient based kernel is permitted, but we do have to give our concrete implementation choice. We think this should be a lot clearer if the main MALA equation in 2.2 is numbered not inline and then is creffed to later, so we will do that.

We like the suggestion to link each part of the algo psuedocode to the right equations, and we agree it could be a useful addition to the main text. We will clean up the descriptions used there as a result.

Reviewer_tj1nOpenReview ↗

Answer to Rebuttal

After reading the review of Reviewer 74dc and looking again into the paper I still believe the writing is very off and is very hard to understand to human readers due to vague and not well introduced terminology as outlined in my review and in the review of Reviewer 74dc. Also, although the authors could have uploaded an updated version of their paper in order to improve the presentation and writing the authors did not do so and thus I will maintain my score.

We updated the draft very shortly after this comment was posted; we took some time to collect all the feedback and implement it together, though the rebuttal already contained direct responses to every question raised. If we had been given more to go on in terms of what remained difficult to follow, we would have happily addressed that too.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision after rebuttal

We have now had a chance to integrate our proposed changes and have uploaded a new revision. The paper is mostly unchanged in Abstract, Introduction and Background as we think these were already in good shape. We have made changes to the exposition throughout the Method and the Lattice Results to improve the readability, and reduce the assumed technical knowledge to make it more appropriate for this venue. We highlight the main changes as follows (mostly in 3.2 3.3 and 4.3)

Summary of changes to the paper

Method (Section 3)

  • Added a short roadmap at the start of Section 3 setting out the order of the method construction.
  • Included a new ref to Chopin & Papaspiliopoulos (2020) as the standard for defining the SMC terminology, and made efforts to highlight which parts we consider standard, and what we have introduced.
  • Greatly expanded Section 3.2: the level-volume ratio, its estimator and the reweighting step are now set out step by step as numbered equations, and the splitting step (previously called branching but we changed this to be more clear) is more clearly stated as the main deviation from standard SMC, appendix covers this precisely in more detail.
  • Moved the full algorithm from the appendix into the main text (Algorithm 1, Section 3.3), with every step linked to the equation it implements. The MALA proposal is now a numbered equation referenced from the algorithm.
  • Included extra details where the previous text was too terse, to clarify what the equations in 3.2 and 3.3 mean more colloquially.

Lattice ϕ6\phi^6 experiment (Section 4.3)

  • Stated the full lattice action and how it splits into an exactly samplable Gaussian free-field prior and the interaction energy passed to the sampler.
  • Defined coexistence precisely (the disordered phase and each of the two ordered sectors carry equal probability, giving a 2:1 ordered share)
  • Updated Table 2 with re-run QES results and added Wasserstein-2 distances on the order parameters mm and qq, a NUTS baseline and an independent reference calculation. We also added a reference calculation with details on it’s construction (forced mode occupancy) in the appendix
  • Figure 3 now shows histograms of both order parameters, mm and qq, with the phase cut marked.
  • Included a reference here for the textbook reference on statistical field theory (Kardar 2007), to point to which parts are standard, and improved the description of some the technical concepts to be more colloquial. For example coexistence is now much more simply defined rather than the more technical definition before.

We have made no changes to the other experiments in the main text or the conclusion

Appendix

  • Added a description of the lattice reference calculation (two phase-restricted tempered SMC runs, pooled by their estimated partition functions) and a definition of the W2W_2 metric.
  • Renamed "branching" to "splitting" and made its definition and use consistent with the main text.

Presentation

  • In moving the algorithm into the main text we also cleaned up a few places where the notation was missing, for example, we updated dlogZ and target ESS to have common symbols throughout

We hope to add an appendix on phase transitions: first- versus second-order transitions, the different challenges each poses for particle methods, and the critical-slowing-down work raised in review. This however does take some time to draft properly so we will upload the key clarity changes now.

EDIT: We have now also included an appendix A on "Phase transitions”, to improve the pedagogy of why tempering methods fail at first order transitions, what they are, and a bit of an outlook on second order transitions

We think the description is now much clearer and more venue appropriate of the method and the lattice results, we hope the reviewers also share this opinion!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.