acceptodds
Under review as a conference paper at ICLR 2027

Multilevel Plaquette-Space Sampling for Lattice Gauge Theories with Local Constraint Solves

Abstract

Lattice gauge theories are an important class of physical models with high-dimensional structured distributions that underpin first-principles calculations in particle, nuclear, and condensed-matter physics. Recent advances in generative modeling have opened new avenues for sampling Boltzmann distributions in lattice gauge theories, offering the potential to alleviate limitations of traditional Monte Carlo methods, including critical slowing down and topological freezing. However, generative samplers for lattice gauge theories are typically constructed in link space, where correlations become increasingly long-ranged toward weak coupling, posing a challenge for learning. Plaquettes provide a more natural representation, as the action is local in these variables. However, exact Bianchi constraints restrict them to a lower-dimensional manifold, complicating direct generative modeling. We introduce a multilevel normalizing-flow construction that samples directly in plaquette space while satisfying these constraints exactly. The key idea is a coarse-to-fine factorization that transforms a globally coupled constraint problem into a sequence of local solves: at each refinement, every determined plaquette depends on at most four newly generated variables, while the size of the constraint problem remains independent of the lattice size. We validate the construction for in two and four dimensions and for in two dimensions. Our multilevel Plaquette-Space Sampler (PSS) substantially outperforms link-space baselines, with the advantage increasing toward weak coupling, where the gauge coupling becomes small. This regime is particularly challenging for generative sampling and, in asymptotically free gauge theories, is relevant to continuum studies.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
lattice field theory, Boltzmann sampling, normalizing flows, generative modeling, multilevel sampling, constrained sampling, statistical sampling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_WXYNOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a multilevel Plaquette-Space Sampler (PSS) for Boltzmann distributions in lattice gauge theories. Instead of generating link variables directly, it samples plaquette variables through a coarse-to-fine hierarchy of conditional normalizing flows. At each level, the flows generate free variables, while deterministic refinement maps enforce Bianchi constraints and preserve coarse plaquettes as Wilson loops of the refined configuration. This construction keeps constraint solves local while retaining tractable sampling densities, enabling reverse-KL training using the known action and subsequent importance reweighting. Experiments on two-dimensional U(1)U(1), four-dimensional U(1)U(1), and two-dimensional SU(2)SU(2) demonstrate improved sampling quality over the evaluated link-space and single-level plaquette-space baselines, with advantages becoming more pronounced at larger lattice sizes or toward weak coupling.

Critical strengths
  1. The method is well matched to the problem structure. Plaquette variables provide a natural representation of the action, while the multilevel construction enforces Bianchi constraints and coarse-to-fine consistency exactly.

  2. The empirical results support the proposed design. Across two-dimensional U(1)U(1), four-dimensional U(1)U(1), and two-dimensional SU(2)SU(2), the method improves upon the evaluated link-space baselines, demonstrating the potential benefits of combining plaquette-space sampling with a multilevel construction.

Critical weaknesses
  1. The broader applicability remains insufficiently established. The construction exploits geometric and algebraic properties of lattice gauge theories. Although the principle of turning global constraints into local refinement problems could interest the broader Scientific ML community, the paper does not clearly characterize the conditions under which this strategy extends to other constrained sampling problems. A more explicit discussion of these conditions, ideally supported by another application, would strengthen the methodological contribution.

  2. The practical computational benefits remain unclear. The paper reports training schedules but lacks a systematic comparison of preprocessing costs, training time, sampling throughput, and total time to achieve a prescribed statistical accuracy. The authors explicitly acknowledge that they do not claim a wall-clock advantage over MCMC. However, the missing measurements limit assessment of the method’s practical value, particularly when training must be repeated for different target distributions.

Questions for authors

Request for clarification:

  1. What structural properties must a constrained sampling problem possess for the proposed local refinement strategy to apply? Which parts depend specifically on lattice gauge theory, and could the authors identify a concrete application beyond this setting?

  2. Could the authors report preprocessing time, training time, hardware, and sampling throughput, including density evaluation and constraint handling? A comparison with an appropriate traditional Monte Carlo baseline in terms of total time to reach the same accuracy for representative observables would be particularly informative. For the amortized setting, how many samples would be required to cover the initial training cost?

Suggestion:

  1. Could another tractable-density model, such as an autoregressive model, replace the flows while retaining exact constraint satisfaction and valid importance weights?
6 comments
AuthorsOpenReview ↗

Response to Reviewer WXYN

We thank the reviewer for the questions and for the suggestion, which we ran as an experiment. New results are in the revised PDF, in red.

W1 + Q1. Structural properties; what is specific to gauge theory; applications beyond it

The construction rests on three structural facts: the constraint (the Bianchi identity) is local, blocking commutes with it, and each refinement has unimodular coordinates with a bounded stencil. Specific to gauge theory are the group variables with their Haar measure, parallel transport, and the Wilson action. Lattice gauge theory is the application of this paper and we claim nothing beyond it; within it the construction is general in gauge group and dimension (exactness verified in d=2,…,5d=2,\dots,5, Table 5).

W2 + Q2. Computational cost

We measured the cost of each stage on one A100 GPU per model (new Appendix J.1):

4D U(1), 444^4, β=2\beta=22D U(1), 32232^22D SU(2), 16216^2
preprocessingsecondsnonenone
training4545 h1010 h2929–3333 h
5×1055\times10^5 samples99 min4949 s≤18\le18 min
per effective sample7.27.2 ms0.150.15 ms2.42.4–5.15.1 ms

Training is a one-time cost per coupling and volume. As for any learned sampler, it pays off where Markov chains freeze: on the 2D line of constant physics plain HMC makes no topological transition in 8×50,0008\times50{,}000 trajectories at L=8L=8, 1616 or 3232, while the sampler reaches 0.70.7% accuracy on χtop\chi_{\rm top} at 32232^2 from 5×1055\times10^5 samples generated in 4949 s. At the four-dimensional volumes studied here we make no wall-clock claim against Monte Carlo; there the comparison is with other learned samplers (reverse KL 0.660.66 against 192192 nats for the multiscale link-space flow on 444^4).

Q3 (suggestion). Could an autoregressive model replace the flows?

Yes. We replaced every conditional flow by an autoregressive model with the same refinement maps and parameter count (new Appendix N.2). The constraint stays exact and every reweighted observable agrees with its exact value, because exactness comes from the refinement maps alone. With the same training schedule the ESS is higher:

ESS (16216^2, 5×1055\times10^5 samples)autoregressivecoupling flow
2D U(1), β=48\beta=480.8940.894, 0.8980.8980.8170.817–0.8770.877
2D SU(2), β=8\beta=8, 25,00025{,}000 steps0.7460.746, 0.7300.7300.4240.424

The price is sequential generation within a level: a U(1) training step takes 2.52.5 s against 0.20.2 s for the flow, which is why the paper uses flows (its SU(2) flow, trained for 100,000100{,}000 steps, reaches ESS 0.820.82).

Presentation

Fig. 2 now opens with an overview of the pipeline that marks what is learned and what is fixed. Appendix F opens with the construction at a glance (with a notation table) and a worked four-dimensional refinement with a figure.

We are glad to clarify anything further.

Thank you for the responses and additional experiments. I have two more budget questions:

  1. [minor] AR versus coupling flows. AR generation can be batched, while coupling flows also require multiple sequential layers. Matching parameter counts and training steps does not match computational budgets. Could you provide a small architecture sweep comparing accuracy/ESS at matched training time, together with batched sampling throughput?
  2. HMC versus the proposed sampler. Was the 8×50,0008\times50{,}000-trajectory HMC budget comparable to the approximately 10 hours of training plus 49 seconds of generation at L=32L=32? Please report actual HMC runtimes and hardware, and compare total costs at a specified observable accuracy. The break-even sample count should be evaluated against the finite sampling requirement for that accuracy; unlimited future sampling cannot simply be assumed to amortize training. My main concern about the significance of the contribution remains unaddressed. I acknowledge the value of the local, exact constraint construction. However, higher-dimensional and non-Abelian sampling are demonstrated separately, and local constraint handling alone does not establish scalable sampling efficiency.

The current evidence leaves me unconvinced that the construction provides a sufficiently substantial methodological or practical advance for more challenging gauge-sampling problems. Therefore, I maintain my original score for now.

AuthorsOpenReview ↗

Response to the follow-up of Reviewer WXYN

We thank the reviewer for the follow-up.

1. Autoregressive model versus coupling flows at matched budget

We ran the sweep. Every model below was trained on the full 16216^2 U(1) ladder (β=48\beta=48) for the same time, 3.53.5 to 3.93.9 h on one A100, and evaluated on 5×1055\times10^5 configurations:

conditional modelparametersESSχtop\chi_{\rm top} deviationconfigurations per second
coupling flow, 4 layers0.780.78 M0.480.480.6σ0.6\sigma64,00064{,}000
coupling flow, 8 layers (as in the paper)1.271.27 M0.880.880.2σ0.2\sigma38,00038{,}000
coupling flow, 16 layers2.252.25 M0.850.850.1σ0.1\sigma21,00021{,}000
autoregressive, 1 spline1.251.25 M0.040.040.9σ0.9\sigma17,00017{,}000
autoregressive, 2 splines1.251.25 M0.270.270.3σ0.3\sigma9,4009{,}400
autoregressive, 3 splines (as in Appendix N.2)1.251.25 M0.370.370.9σ0.9\sigma7,4007{,}400

At matched training time the coupling flow is ahead in ESS and in throughput, and every model remains unbiased. The autoregressive model reaches its ESS of 0.900.90 (Table R7) only with the full schedule, about 2828 h of training. It shows that the density model can be exchanged while the constraint stays exact; the paper uses flows because of this cost.

2. HMC versus the proposed sampler; significance

We have now run HMC for the same wall-clock time as our total cost, 10 h of training plus 49 s of generation:

32232^2, β=192\beta=192hardwaregenerated in 10 h + 49 ssector changesχtop\chi_{\rm top} / exact
HMC, 8 chains4 CPU cores1.8×1071.8\times10^7 trajectories0240±80240\pm80
HMC, 4096 chainsone A1008.9×1088.9\times10^8 trajectories0257±6257\pm6
multilevel PSSone A1005×1055\times10^5 configurationsindependent samples1.005±0.0071.005\pm0.007

Training is the slow part of our method. At equal cost, however, no HMC chain left its initial topological sector, so the HMC estimator of χtop\chi_{\rm top} is clearly biased. A longer run does not help: the tunnelling rate, measured on the same lattice at β=4\beta=4 to 1010, falls by a factor of about six per unit of β\beta (below 10−14010^{-140} per trajectory at β=192\beta=192). Multilevel PSS reaches 0.70.7% accuracy on χtop\chi_{\rm top} within this budget, training included.

We agree that a local constraint solve alone does not give scalable efficiency. Non-Abelian gauge theory in four dimensions needs a more detailed design and training procedure, which we leave for future work; the refinement itself is verified exact for SU(2) up to d=5d=5 (Table 5). The practical advance we claim is this regime: where HMC freezes, the sampler gives χtop\chi_{\rm top} within 0.7σ0.7\sigma of the exact value up to 64264^2.

Reviewer_JMSmOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes multilevel plaquette space sampling (PSS) for generative models sampling Boltzmann distribution in various lattice gauge theories. While traditional approaches to this kind of sampling are done with MCMC, recent advances in generative modeling have amortized this process to a neural network to allow for efficient sampling and prospects for sampling in much higher dimensions. However, existing generative models usually to emit configurations in link space; this approach has been shown to struggle to capture long range correlations and degrades towards weak coupling as the number of lattice sites increases. The authors argue that because the Wilson action only depends on ordered products of links around elementary squares (called plaquettes), that it is more natural to try to directly emit plaquette configurations. (Singha et al, 2026b) do this for U(1)U(1) lattice gauge theories in 2 dimensions, and this work seeks to expand this scope. However, unlike link space configurations, not any configuration of plaquettes is physically valid. The Bianchi identities impose global constraints that would normally require solving a linear system which grows with the number of plaquettes, limiting the scalability of the method. To get around this, the authors introduce a multilevel approach, whereby an initial coarse grid of plaquettes satisfying the Bianchi identities is generated, and then each of these individual plaquette spaces is iteratively refined into smaller plaquette spaces, only requiring a local consistency solve to maintain the validity of the overall configuration. This avoids ever having to do a global solve, and furthermore allows for long range correlations to be effectively captured. The authors demonstrate their method on U(1)U(1) lattice gauge theories in 2 and 4 dimensions, along with an SU(2)SU(2) lattice gauge theory in 2 dimensions, showing that their method extends to the noncommutative setting. They compare this against (Abbott et al., 2024a), which implements a multiscale normalizing flow construction in link space.

Critical strengths

-The paper provides a novel multilevel sampling approach for plaquette space sampling, which offers a clever solution to be scalable and avoid ever having to solve a global linear system which grows linearly in the number of plaquettes. It seems that there is evidence for both i) multilevel PSS is not only more scalable than single level PSS, but also more expressive (achieves lower KL divergence and higher ESS), and ii) PSS seems to give better results compared with single scale and the multiscale link space sampling done by (Abbot et al, 2024a).

Critical weaknesses

-It is unclear to me what other baselines exist to compare against. There is a lot of cited work in the paper that also attempts to solve the same problem, but the only comparisons provided are to single level PSS and the (Abbot et al, 2024a) work. It would be nice to see what the other methods are capable of to get a better sense of the innovative-ness of this work. My questions below also bring up specific concerns that I don't feel like were adequately discussed in the paper.

-As the authors also mention in the limitations, all lattices considered seem to be relatively small and only a single training run per model. More expansive evaluations could provide the reader with a better understanding of the multilevel PSS's effectiveness in these regimes.

Questions for authors

-Is it possible to compare multilevel PSS against other baselines to get a better sense of where the methodology fits into the larger body of research in this area? For example, I would have expected some type of comparison to the other plaquette space sampling technique done by (Singha et al, 2026b) as this is the other plaquette space sampling approach discussed in the paper.

-Were there any experiments done in the toroidal domains where the Bianchi identities give a more nontrivial relationship between the plaquettes (Equation 5)? Or more generally, were any experiments done in more general lattice topologies?

-Was there any particular reason why there was no experiment done for a U(1)U(1) lattice gauge theory in 3 dimensions?

-It would be interesting to understand how the runtime varies as a function of the number of sampling levels, along with lattice sites/dimension or even lattice gauge theory. I couldn't find any plots describing this, so it was hard to understand how much of a computational win this would be over the other methods (including the MCMC reference).

3 comments
AuthorsOpenReview ↗

Response to Reviewer JMSm

We thank the reviewer for the constructive questions. New results are in the revised PDF, in red.

W1 + Q1. Other baselines; comparison with Singha et al. (2026b)

The comparison is in the paper: single-level PSS, the baseline in Figs. 3–5 and 7, is our implementation of the plaquette-space flow of Singha et al. (2026b), extended to four dimensions and to SU(2); this is now stated in Section 4 of the revision. Their mixture over topological sectors is specific to 2D U(1), so it is not part of the baseline. We also compare with published exact-density samplers on each benchmark (Abbott et al. 2024a; Kanwar et al. 2020; Gerdes et al. 2025). Multilevel samplers for spin systems and ϕ4\phi^4 do not carry over directly to gauge theory; we discuss them as related work (Appendix A, new Table R1).

W2. Small lattices; a single training run

Seeds. We retrained the two-dimensional ladder twice with new seeds (new Appendix N.1):

LLKL paperKL seed 1KL seed 2ESS paperESS seed 1ESS seed 2
80.102±0.0070.102\pm0.0070.108±0.0070.108\pm0.0070.103±0.0060.103\pm0.0060.9230.9230.9220.9220.9080.908
160.215±0.0240.215\pm0.0240.222±0.0260.222\pm0.0260.236±0.0260.236\pm0.0260.8770.8770.8750.8750.8170.817
320.632±0.0920.632\pm0.0920.666±0.1050.666\pm0.1050.538±0.0680.538\pm0.0680.6710.6710.7130.7130.5350.535

The KL agrees within errors, and the lowest ESS at 32232^2 is still nine times that of the multiscale link-space flow.

Size. The four-dimensional experiments now reach 848^4: at β=2\beta=2 the reverse KL is 0.660.66 nats at 444^4 (Fig. 4) and 35.435.4 nats at 848^4 (new Appendix G.3), both below the 192192 nats that the multiscale link-space flow reports on 444^4. The two-dimensional ladder reaches 64264^2.

Q2. Toroidal domains and other topologies

Yes: every experiment is on a periodic torus, so Eq. (5) is active throughout. In two dimensions it is the only constraint and every level satisfies it exactly; all topological sectors are generated, and χtop\chi_{\rm top} agrees with the exact torus value within 0.7σ0.7\sigma at every size (Fig. 10). In four dimensions the torus adds six quantised fluxes, fixed at the coarsest level, and four holonomies that do not enter the action.

Q3. Why no experiment in three dimensions?

There is no obstacle; it was a choice of where to spend training: 2D has exact references for a scaling study, 4D is the physical case with a published benchmark. Exactness is verified numerically in d=2,…,5d=2,\dots,5 (Table 5), and in 3D the refinement solve stays local at every extent.

Q4. Runtime and gain over other methods

Runtime grows close to linearly with the volume, and levels add nothing beyond that because the finest level dominates. Generating 5×1055\times10^5 configurations with their exact density on one A100 GPU takes (new Appendix J.1):

theorylatticerefinementstimeper site and configuration
2D U(1)16216^231313 s0.100.10 μ\mus
2D U(1)32232^244949 s0.100.10 μ\mus
2D U(1)64264^25316316 s0.150.15 μ\mus
2D SU(2)16216^23≤18\le18 min≤8.4\le8.4 μ\mus
4D U(1)444^4499 min4.24.2 μ\mus
4D U(1)848^48158158 min4.64.6 μ\mus

The gain over other methods:

  • Link-space flows. 4D U(1), 444^4, β=2\beta=2: reverse KL 0.660.66 against 192192 nats for the multiscale flow. 2D U(1), 32232^2: 0.150.15 ms per effective sample against 1111 ms. 2D SU(2): at most 2.22.2 ms per configuration against 0.090.09–1.11.1 s for the continuous flow.
  • Single-level PSS needs a preprocessing elimination that grows faster than the volume (824824 s at L=10L=10 in 4D); the multilevel chain needs seconds.
  • Monte Carlo. The gain of a learned sampler is where Markov chains freeze: on the 2D line of constant physics plain HMC makes no topological transition in 8×50,0008\times50{,}000 trajectories at L=8,16,32L=8,16,32, while our samples are independent. At the four-dimensional volumes studied here we make no wall-clock claim against Monte Carlo; there the comparison is with other learned samplers.

We are glad to clarify anything further.

Reviewer_JMSmOpenReview ↗

Response to authors

Thank you for helping me to understand the paper better. I think given the updated revisions, this paper is solid and so I have increased my score.

Dear Reviewer JSMS, thank you for reading our responses and the revision so closely, and for reconsidering your assessment. Your questions helped us improve the paper.

The authors

Reviewer_2e3qOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a multilevel plaquette-space sampler for lattice gauge theories. Conditional flows generate free plaquettes, while deterministic refinements enforce Bianchi constraints and preserve a tractable density. Experiments on 2D and 4D U(1) and 2D SU(2) show improved sampling efficiency over link-space and single-level plaquette baselines.

Critical strengths
  1. The multilevel construction makes constraint solves local while retaining exact constraint satisfaction and density evaluation. This addresses a concrete obstacle to plaquette-space sampling beyond two dimensions.

  2. The sampling gains are substantial and supported by observable checks. The topological susceptibility and SU(2) Wilson loops in Appendix N provide stronger validation than ESS alone.

Critical weaknesses
  1. High-dimensional scalability remains insufficiently tested. The 4D experiment only considers 24→442^4 \rightarrow 4^4. Local constraint solves do not ensure that errors in the learned conditionals or importance-weight variance remain controlled as volume grows.

  2. The conditioning analysis provides a plausible explanation for the gains, but the empirical evidence for this mechanism is limited. Fig. 7 compares single-level and multilevel models, changing both the constraint structure and the conditional factorization. It therefore supports the overall construction without establishing how much of the improvement comes from better conditioning.

Questions for authors
  1. Why were the coordinates in Table 2 preferred over those in Eq. 39? Is there empirical evidence that distinguishes the benefit of better conditioning from that of the multilevel conditional factorization?

  2. At β=1.5\beta=1.5 and 22, how do the reweighted monopole counts and uncertainties compare with the heatbath references in Appendix K?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2e3q

We thank the reviewer for the precise questions. New results are in the revised PDF, in red.

W1. Scalability beyond 444^4

We extended the four-dimensional experiments to 848^4. At β=2\beta=2 the reverse KL is 0.660.66 nats at 444^4 (Fig. 4) and 35.435.4 nats at 848^4 (new Appendix G.3), both below the 192192 nats that the multiscale link-space flow reports on 444^4.

Imperfect proposals lose ESS with volume, which our construction does not remove. It removes the constraint-specific difficulty: every refinement has a solve support and a parameter count independent of the lattice extent.

Q1(a). Why the coordinates of Table 2 and not those of Eq. (39)?

Both are coordinates of the same refinement, related by a unimodular change of variables, so they define the same exact sampler. The models of the submission use the coordinates of Table 2 (support at most nine), which come from a general integer elimination; Eq. (39) is the closed form, used for the new levels beyond 444^4 (Appendix G.3). The revision states which coordinates the models use.

W2 + Q1(b). Conditioning versus conditional factorization

The reviewer is right that Fig. 7 tests the construction as a whole. In our construction the refinement hierarchy and the locality of the constraint solve are coupled by design. Each refinement doubles one direction, which is what allows the new plaquettes to close on the previous level within a bounded stencil. Fig. 7 therefore evaluates the resulting construction and does not isolate numerical conditioning as the sole cause of the improvement; we have clarified this in the revision.

We ran the missing control, a hierarchy without the local solve: an isotropic refinement 24→442^4\to4^4 in one step (support up to 47), with the same root flow and parameter count as the multilevel model of Fig. 7 (new Appendix G.2, Table R3).

model at β=2\beta=2, 444^4solvereverse KL [nats]ESS
single-level PSS (no hierarchy)global2.052.050.0150.015
isotropic refinement (two seeds)non-local1.971.97, 2.062.060.0080.008, 0.0030.003
directional refinements (Fig. 7)local1.001.000.0470.047

The gain appears only with the local solve.

Q2. Reweighted monopole counts at β=1.5\beta=1.5 and β=2\beta=2

At these two couplings the reweighted monopole number is below the heatbath reference, while the plaquette is within 1.4×10−31.4\times10^{-3} and 1.4×10−41.4\times10^{-4} of its reference in relative terms (Fig. 4(b)). The paper is about the action density and the observables built from it. Monopoles at weak coupling are rare discrete defects, which a model trained with the reverse KL proposes too rarely for reweighting to correct; they need a dedicated treatment, which the submitted paper already lists as a limitation. Where monopoles are dense (β=0.5\beta=0.5, 0.80.8) the counts agree with heatbath (Appendix M).

β\betaheatbath referencemultilevel PSS, reweighted
0.50.5352.969±0.043352.969\pm0.043352.990±0.034352.990\pm0.034
0.80.8250.18±0.06250.18\pm0.06250.11±0.20250.11\pm0.20
1.51.50.411±0.0040.411\pm0.0040.233±0.0140.233\pm0.014
2.02.00.0094±0.00070.0094\pm0.00070.0033±0.00040.0033\pm0.0004

(Monopoles per configuration; 500,000500{,}000 samples.)

We are glad to clarify anything further.

I thank the authors for their rebuttal. Most of my concerns have been addressed, and I would like to maintain my original ratings and positive assessment.

Dear Reviewer 2e3q, thank you for your careful review and your positive assessment. We appreciate the time you gave to our paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.