Evaluating generative samplers for lattice field theory by scale

arXiv:2611.11108 · hep-lat, cs.LG · Submitted 2026-11-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: High Energy Physics: Lattice. The day in lattice field theory: what the simulations put on the table.

Kai: We're looking at the paper: "Evaluating generative samplers for lattice field theory by scale".

Mira: Here's the gist of the paper. Generative samplers for lattice field theory are commonly evaluated by their effective sample size (ESS) We show that a small global ESS need not imply inaccurate long-distance observables Errors spread over many short-distance modes can dominate this diagnostic We introduce a scale-resolved diagnostic based on the overlap between coarse-grained model and target distributions At the lattice cutoff, this overlap determines the asymptotic ESS fraction; lower cutoffs isolate long-wavelength errors In the exactly solvable free theory, two samplers with comparable ESS yield a 32% bias and no bias in the susceptibility In two-dimensional ϕ 4 theory near criticality, trained diffusion samplers show the same separation of scales: their path-weight ESS fraction falls rapidly over the volumes studied, while low moments of the magnetization remain close to their target values.

Kai: Scale-Resolved Diagnostics The paper develops a framework for evaluating sampler errors by scale, examining both which part of the mismatch determines the ESS and how to use infrared variables to reduce errors in full-configuration proposals.

Mira: The asymptotic ESS fraction is given by e−D2(p∥q), where D2 is the Rényi divergence of order two.

Kai: Removing short-distance modes gives a monotone overlap flow D2(Λ): its value at the lattice cutoff determines the global ESS, while smaller cutoffs isolate errors at longer wavelengths.

Title and authors: Kai: So, we're looking at this paper today, "Evaluating generative samplers for lattice field theory by scale." It seems to be diving into how we actually measure the quality of these AI models that generate things like lattice field theory configurations.

Mira: Exactly. The title suggests they are moving beyond just looking at the total effective sample size, or ESS, which is a common way we check if a sampler is working well. They argue that relying only on global ESS can be misleading because errors spread across many short-distance modes can hide the real problem affecting long-distance observables.

Lev: From my side, I'm thinking about how this relates to what we actually build for hardware. If we only check the total number of samples taken, we miss where the proposal is failing locally, and that’s where the overhead really lands on a real system.

Kai: Right. So what they introduce is this scale-resolved diagnostic based on comparing a coarse-grained model to the target distribution. They suggest that looking at how much these distributions overlap at different length scales tells us something more specific about the errors.

Mira: It’s about isolating which part of the mismatch causes the ESS problem, and using those infrared variables to actually reduce errors in full configuration proposals instead of just throwing more computational power at them.

Lev: That makes sense for hardware, because if you can pinpoint a scale where things are wrong, you can target your error correction precisely. It means we don't have to treat every mode equally when designing the sampler.

The paper's summary: Kai: So this paper is essentially saying that the total effective sample size isn't the whole story when you are sampling complex systems like lattice field theory. They show that a small global ESS doesn't mean your long-distance observables are accurate, because errors in many short-distance modes can dominate what you measure.

Mira: They propose this new scale-resolved diagnostic, which uses the overlap between a coarse grained model and the target distribution to measure how good the sampler is at different length scales. At the lattice cutoff, this overlap gives you the asymptotic ESS fraction, and using lower cutoffs lets you isolate errors that are at longer wavelengths.

Lev: I see why that matters for us running simulations. If we only check one number, like a global ESS, we might think our system is good when it's actually failing on the modes that matter most for physics at long distances.

Kai: And they test this framework in two different scenarios. In the exactly solvable free theory, they found that two samplers with similar global ESS can still have a thirty-two percent bias in something like susceptibility, while they might not even have a bias in the susceptibility itself.

Mira: Then they look at two dimensional phi four theory near criticality, and their trained diffusion samplers show this same scale separation: the fraction of path weight ESS falls really fast over the volumes studied, but low moments of things like magnetization stay close to what they should be.

Lev: So it means that for critical systems, even if we have many samples but some parts are way off at fine scales, the macroscopic physics we care about might still be okay. It’s a nuance that matters when designing our error correction algorithms.

The paper's improvements: Kai: One big improvement they point out is the marginal test. Instead of dealing with the exponential penalty of global overlap for all observables, this test shows how you can use just a given target marginal—like magnetization—to do full configuration sampling without that huge penalty.

Mira: That’s significant because it means you don't have to worry about the whole distribution matching perfectly for every single observable you care about; you only need the important parts to match well enough for the observables being studied. They say that with exact marginal density ratios, this marginal test accepts at least as often as the full independence test at their stationary distributions.

Lev: From an engineering standpoint, that sounds like a massive reduction in complexity for running proposals on real hardware because you don't have to calculate that massive global overlap measure every time. It simplifies the proposal step considerably.

Kai: And they also discuss how you can incorporate infrared information directly into the proposal mechanism itself, or use a marginal test on selected variables to correct errors in those modes instead of just relying on expensive full configuration proposals.

Mira: They also talk about building composite Markov chains where an exact local kernel update happens after a marginal test, which helps reduce the remaining mismatch at finite relaxation length in lattice field theory simulations.

Lev: That composite chain idea sounds promising for real hardware because it means we can use local updates to fix the error without needing the entire system to be perfectly matched immediately. It suggests a way to handle the mismatch progressively.

Conclusion: Kai: So to wrap up, this paper resolves the first big question by showing that you should resolve the overlap by scale. The asymptotic effective sample size, which they call e minus D two of p parallel q, measures the aggregate mismatch across all scales, but a long-distance observable only probes the marginal distribution of variables on which it depends.

Mira: And for the second question, they show that this marginal test is useful because you can use a given target marginal to do full configuration sampling without paying that exponential penalty from global overlap for the observables you are actually interested in.

Lev: The composite chain generally has a biased stationary distribution at finite relaxation length, but it gets closer and closer to the full target as the local kernel converges, which is good news for convergence speed.

Kai: And finally, they provide a free-field solution that bounds the number of local updates needed to get accuracy on mode variances independently of volume at fixed correlation length and with uniformly bounded relative errors in proposal mode variances.

Mira: The authors do admit some limitations though, specifically regarding infrared relaxation and estimating marginal densities accurately. They flag that interactions can transfer ultraviolet errors into the zero mode during relaxation, so you should probably relax the ultraviolet before correcting the infrared.

Lev: So for real applications, it means we still need to deal with those slow modes near criticality and make sure our estimates of those marginal densities are good, because if they aren't, you get a systematic error.

Kai: The paper "Evaluating generative samplers for lattice field theory by scale" gives us a better way to think about how to evaluate these AI samplers by looking at scale rather than just the total number of samples.

Mira: It shifts the focus toward using targeted tests, like marginal tests, and building composite chains that incorporate local updates based on those tests.

Lev: It points us toward needing better infrared relaxation techniques because that's where the remaining challenges lie for making these methods practical for large-scale simulations.

Satoshi Iso, Lingxiao Wang

RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences · KEK Theory Center · Institute for Physics of Intelligence, Graduate School of Science, The University of Tokyo

hep-lat, cs.LG

Submitted: 2026-11-08

Updated: 2026-11-08

Comments: 27 pages, 11 figures, comments are welcome!

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 98/100

The gist: Generative samplers for lattice field theory are commonly evaluated by their effective sample size (ESS) We show that a small global ESS need not imply inaccurate long-distance observables Errors…

Key concepts

Asymptotic ESS Fraction
This measures the overall mismatch between a proposed model and the target distribution, given by $e^{-D^2(p||q)}$. It quantifies how much information is lost when comparing two distributions. By removing short-distance modes, this fraction tells us about the long-distance errors.
Scale-Resolved Diagnostic
This framework evaluates sampler errors by examining which part of the mismatch causes issues at different scales. It uses a flow based on coarse-graining to isolate relevant (long-wavelength) mismatches from irrelevant (short-distance) lattice artifacts, allowing for targeted error correction.
Marginal Test
This test checks if a specific target marginal distribution—like the magnetization—can be used effectively in sampling. It shows that using exact marginal density ratios can yield acceptance rates comparable to full independence tests, reducing the exponential penalty of global overlap.
Composite Chain
This is a method where an exact local kernel update is performed after a marginal test. It helps reduce the remaining mismatch at finite relaxation lengths in lattice field theory simulations by iteratively refining the proposal towards the target distribution.

Terminology

Summary

Generative samplers for lattice field theory are commonly evaluated by their effective sample size (ESS) We show that a small global ESS need not imply inaccurate long-distance observables Errors spread over many short-distance modes can dominate this diagnostic We introduce a scale-resolved diagnostic based on the overlap between coarse-grained model and target distributions At the lattice cutoff, this overlap determines the asymptotic ESS fraction; lower cutoffs isolate long-wavelength errors In the exactly solvable free theory, two samplers with comparable ESS yield a 32% bias and no bias in the susceptibility In two-dimensional ϕ 4 theory near criticality, trained diffusion samplers show the same separation of scales: their path-weight ESS fraction falls rapidly over the volumes studied, while low moments of the magnetization remain close to their target values. p. 0

Where the paper says it

In the exactly solvable free theory, two samplers with comparable ESS yield a 32% bias and no bias in the susceptibility, respectively. In two-dimensional ϕ 4 theory near criticality, trained diffusion samplers show the same separation of scales: their path-weight ESS fraction falls rapidly over the volumes studied, while low moments of the magnetization remain close to their target values.

Scale-Resolved Diagnostics

The paper develops a framework for evaluating sampler errors by scale, examining both which part of the mismatch determines the ESS and how to use infrared variables to reduce errors in full-configuration proposals. The asymptotic ESS fraction is given by e−D2(p∥q), where D2 is the Rényi divergence of order two. p. 1

Where the paper says it

We develop a framework for evaluating sampler errors by scale, the main subject of this paper, and examine both questions in the free field. The asymptotic ESS fraction is e −D2(p∥q) , where D2 is the Rényi divergence of order two.

Removing short-distance modes gives a monotone overlap flow D2(Λ): its value at the lattice cutoff determines the global ESS, while smaller cutoffs isolate errors at longer wavelengths. p. 1
Where the paper says it

The asymptotic ESS fraction is e −D2(p∥q) , where D2 is the Rényi divergence of order two. Removing short-distance modes gives a monotone overlap flow D2(Λ): its value at the lattice cutoff determines the global ESS, while smaller cutoffs isolate errors at longer wavelengths.

The distinction between relevant and irrelevant mismatches is shown by comparing mass errors (relevant) with lattice artifacts (irrelevant) in the free field. p. 4
Where the paper says it

(8) For small errors each mode contributes ε 2 k /2 regardless of wavelength. We compare a mass error, strongest at long wavelengths, with a lattice artifact, strongest at short wavelengths: relevant (mass): εk = c mˆ 2 ˆk 2 + m2 , irrelevant (lattice artifact): εk = c6 ˆk 4 ˆk 2 + m2 , (9) 10 −1 10 0 cutoff Λ (modes with ̂ k≤ Λ kept) 10 −4 10 −3 10 −2 10 −1 10 0 10 1 D2(Λ) global D2 ≃10 Λ =1/ξ relevant: δm2 (ĉ=0.

How the training signal is built

The paper tests trained OSQ samplers for two-dimensional ϕ 4 theory near criticality against benchmark ensembles. The reference parameters are fixed from the benchmark, and they reproduce ⟨M⟩ and ⟨M2⟩ to a few percent while their path-weight ESS fractions fall below 10−3 at L = 64. p. 1

Where the paper says it

We test this picture with trained OSQ samplers for two-dimensional ϕ 4 theory near criticality. With reference parameters fixed from the benchmark, they reproduce ⟨|M|⟩ and ⟨M2 ⟩ to a few percent while their path-weight ESS fractions fall below 10−3 at L = 64; the test on M still accepts about 0.

The log-weight variance grows approximately with volume, much of it apparently from path fluctuations rather than endpoint mismatch. p. 0
Where the paper says it

The exponential volume dependence has a simple origin. If the model deviates slightly from the target in each correlation volume, log w is a sum of many weakly correlated contributions, and its variance grows approximately in proportion to V .

Where the gains come from

The marginal test shows how a given target marginal can be used in full-configuration sampling without the exponential penalty of global overlap for the observables studied here. With exact marginal density ratios, the marginal test accepts at least as often as the full independence test at their respective stationary distributions. p. 1

Where the paper says it

The distinction also explains why weighting only the modes on which an observable depends avoids noise from other modes, in the spirit of the local factorization of Ref. For the second question, we derive the mean acceptance rate of an independence test and show that a test on a few collective variables (low-dimensional functions of the configuration such as the magnetization), which we call the marginal test, accepts at least as often as the full test at their respective stationary distributions.

p. 1
Where the paper says it

The distinction also explains why weighting only the modes on which an observable depends avoids noise from other modes, in the spirit of the local factorization of Ref. For the second question, we derive the mean acceptance rate of an independence test and show that a test on a few collective variables (low-dimensional functions of the configuration such as the magnetization), which we call the marginal test, accepts at least as often as the full test at their respective stationary distributions.

p. 1
Where the paper says it

The distinction also explains why weighting only the modes on which an observable depends avoids noise from other modes, in the spirit of the local factorization of Ref. For the second question, we derive the mean acceptance rate of an independence test and show that a test on a few collective variables (low-dimensional functions of the configuration such as the magnetization), which we call the marginal test, accepts at least as often as the full test at their respective stationary distributions.

The free-field solution bounds the number of local updates needed for a fixed accuracy of the mode variances independently of volume, at fixed correlation length and uniformly bounded relative errors in the proposal mode variances. p. 0
Where the paper says it

Combined with exact local updates, it reduces the remaining mismatch, although residual bias remains at finite relaxation length. In the free field, uniformly bounded relative errors in the proposal mode variances give a volume-independent bound on the number of updates needed for a prescribed accuracy of the chain’s stationary mode variances at fixed correlation length.

p. 0
Where the paper says it

Combined with exact local updates, it reduces the remaining mismatch, although residual bias remains at finite relaxation length. In the free field, uniformly bounded relative errors in the proposal mode variances give a volume-independent bound on the number of updates needed for a prescribed accuracy of the chain’s stationary mode variances at fixed correlation length.

p. 0
Where the paper says it

Combined with exact local updates, it reduces the remaining mismatch, although residual bias remains at finite relaxation length. In the free field, uniformly bounded relative errors in the proposal mode variances give a volume-independent bound on the number of updates needed for a prescribed accuracy of the chain’s stationary mode variances at fixed correlation length.

p. 0
Where the paper says it

Combined with exact local updates, it reduces the remaining mismatch, although residual bias remains at finite relaxation length. In the free field, uniformly bounded relative errors in the proposal mode variances give a volume-independent bound on the number of updates needed for a prescribed accuracy of the chain’s stationary mode variances at fixed correlation length.

p. 0
Where the paper says it

Combined with exact local updates, it reduces the remaining mismatch, although residual bias remains at finite relaxation length. In the free field, uniformly bounded relative errors in the proposal mode variances give a volume-independent bound on the number of updates needed for a prescribed accuracy of the chain’s stationary mode variances at fixed correlation length.

Limitations and Future Directions

The remaining limitations concern infrared relaxation and marginal-density estimation. p. 1

Where the paper says it

The order of the two corrections matters when the proposal’s ultraviolet is far from the target: its relaxation shifts the tested variables, so the ultraviolet should be relaxed before the infrared is corrected, or with the tested variables held fixed. The remaining limitations concern the untested low modes, their coupling to the tested variables, and errors in the estimated marginal densities.

p. 1
Where the paper says it

The order of the two corrections matters when the proposal’s ultraviolet is far from the target: its relaxation shifts the tested variables, so the ultraviolet should be relaxed before the infrared is corrected, or with the tested variables held fixed. The remaining limitations concern the untested low modes, their coupling to the tested variables, and errors in the estimated marginal densities.

p. 1
Where the paper says it

The order of the two corrections matters when the proposal’s ultraviolet is far from the target: its relaxation shifts the tested variables, so the ultraviolet should be relaxed before the infrared is corrected, or with the tested variables held fixed. The remaining limitations concern the untested low modes, their coupling to the tested variables, and errors in the estimated marginal densities.

p. 1
Where the paper says it

The order of the two corrections matters when the proposal’s ultraviolet is far from the target: its relaxation shifts the tested variables, so the ultraviolet should be relaxed before the infrared is corrected, or with the tested variables held fixed. The remaining limitations concern the untested low modes, their coupling to the tested variables, and errors in the estimated marginal densities.

Interactions can transfer a proposal’s ultraviolet error into the tested zero mode during relaxation, so the ultraviolet is best relaxed before the infrared is corrected, as with the relaxed proposals here, or with the tested variables held fixed. p. 1
Where the paper says it

In the composite chain, about ten HMC trajectories remove most of the ultraviolet mismatch at every volume studied. The order of the two corrections matters when the proposal’s ultraviolet is far from the target: its relaxation shifts the tested variables, so the ultraviolet should be relaxed before the infrared is corrected, or with the tested variables held fixed.

Untested low modes can remain slow near criticality, and inaccurate marginal ratios leave a systematic error. p. 0
Where the paper says it

In the exactly solvable free theory, two samplers with comparable ESS yield a 32% bias and no bias in the susceptibility, respectively. In two-dimensional ϕ 4 theory near criticality, trained diffusion samplers show the same separation of scales: their path-weight ESS fraction falls rapidly over the volumes studied, while low moments of the magnetization remain close to their target values.

Conclusion

The first question is answered by resolving the overlap by scale. The asymptotic ESS fraction, e−D2(p∥q), measures the aggregate mismatch, whereas a long-distance observable probes the marginal distribution of the variables on which it depends. p. 1

Where the paper says it

We construct two samplers with similar global ESS: a mass error biases the susceptibility by 32%, whereas a short-distance lattice artifact leaves it unbiased. The distinction also explains why weighting only the modes on which an observable depends avoids noise from other modes, in the spirit of the local factorization of Ref.

p. 1
Where the paper says it

We construct two samplers with similar global ESS: a mass error biases the susceptibility by 32%, whereas a short-distance lattice artifact leaves it unbiased. The distinction also explains why weighting only the modes on which an observable depends avoids noise from other modes, in the spirit of the local factorization of Ref.

For the second question, the marginal test shows how a given target marginal can be used in full-configuration sampling without the exponential penalty of global overlap for the observables studied here. p. 1
Where the paper says it

For the second question, we derive the mean acceptance rate of an independence test and show that a test on a few collective variables (low-dimensional functions of the configuration such as the magnetization), which we call the marginal test, accepts at least as often as the full test at their respective stationary distributions. Its acceptance stays bounded away from zero if the overlap of the chosen marginal distributions does so as the volume grows.

p. 1
Where the paper says it

For the second question, we derive the mean acceptance rate of an independence test and show that a test on a few collective variables (low-dimensional functions of the configuration such as the magnetization), which we call the marginal test, accepts at least as often as the full test at their respective stationary distributions. Its acceptance stays bounded away from zero if the overlap of the chosen marginal distributions does so as the volume grows.

p. 1
Where the paper says it

For the second question, we derive the mean acceptance rate of an independence test and show that a test on a few collective variables (low-dimensional functions of the configuration such as the magnetization), which we call the marginal test, accepts at least as often as the full test at their respective stationary distributions. Its acceptance stays bounded away from zero if the overlap of the chosen marginal distributions does so as the volume grows.

The composite chain generally has a biased stationary distribution at finite relaxation length and approaches the full target as the local kernel converges. p. 1
Where the paper says it

The marginal test alone has the correct stationary distribution for the tested variables when their density ratio is known; at fixed values of these variables it retains the model’s conditional distribution. The composite chain is generally biased at finite relaxation length.

p. 1
Where the paper says it

The marginal test alone has the correct stationary distribution for the tested variables when their density ratio is known; at fixed values of these variables it retains the model’s conditional distribution. The composite chain is generally biased at finite relaxation length.

p. 1
Where the paper says it

The marginal test alone has the correct stationary distribution for the tested variables when their density ratio is known; at fixed values of these variables it retains the model’s conditional distribution. The composite chain is generally biased at finite relaxation length.

The free-field solution bounds the number of local updates needed for a fixed accuracy of the mode variances independently of volume, at fixed correlation length and uniformly bounded relative errors in the proposal mode variances. The remaining limitations concern infrared relaxation and marginal-density estimation.

Acknowledgments

We thank Gert Aarts, Yuto Ashida, Anna Hasenfratz, Gurtej Kanwar, Jan Pawlowski and Phiala Shanahan for valuable discussions. We also thank the members of the DM-QFT Collaboration and the DEEP-IN WG at RIKEN-iTHEMS for discussions. p. 0

Where the paper says it

RIKEN-iTHEMS-Report-26 Evaluating generative samplers for lattice field theory by scale Satoshi Iso1, 2, ∗ and Lingxiao Wang1, 3, † 1RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences (iTHEMS), Wako, Saitama 351-0198, Japan 2KEK Theory Center, Institute of Particle and Nuclear Studies, KEK, Oho 1-1, Tsukuba, Ibaraki 305-0801, Japan 3 Institute for Physics of Intelligence, Graduate School of Science, The University of Tokyo, Bunkyo-ku, Tokyo 113-0033, Japan (Dated: November 10, 2026) Generative samplers for lattice field theory are commonly evaluated by their effective sample size (ESS). We show that a small global ESS need not imply inaccurate long-distance observables: errors spread over many short-distance modes can dominate this diagnostic.

We acknowledge collaboration with the Google DeepMind team.

Data Availability

The code of the OSQ samplers, on which our samplers are based, is publicly available at https://github.com/DM-QFT/OSQ-Scalar. p. 0

Where the paper says it

Diffusion models have been applied to scalar field theories and sigma models [11–14] and to gauge theories [15–18], also with dynamical fermions [19, 20]; they have been extended to theories with a sign problem [21] and trained without reference ensembles [13, 22]. Other samplers are based on stochastic quantization formulated as optimal control (OSQ) [23], on stochastic path sampling [24] or on non-equilibrium transport with a learned drift [25].

The scripts used for the analysis in this work, including the training with a trainable reference, are available from the authors upon request. p. 6
Where the paper says it

[12], which are publicly available [63]. They consist of 10240 configurations per volume, generated with Fourier-accelerated HMC combined with Wolff cluster updates, and the integrated autocorrelation times of the observables used here are below 0.

The benchmark ensembles are publicly available at https://huggingface. p. 6
Where the paper says it

As benchmark we use the ensembles of Ref. [12], which are publicly available [63].

co/datasets/phi4-generative/phi4-critical-configs. p. 0
Where the paper says it

RIKEN-iTHEMS-Report-26 Evaluating generative samplers for lattice field theory by scale Satoshi Iso1, 2, ∗ and Lingxiao Wang1, 3, † 1RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences (iTHEMS), Wako, Saitama 351-0198, Japan 2KEK Theory Center, Institute of Particle and Nuclear Studies, KEK, Oho 1-1, Tsukuba, Ibaraki 305-0801, Japan 3 Institute for Physics of Intelligence, Graduate School of Science, The University of Tokyo, Bunkyo-ku, Tokyo 113-0033, Japan (Dated: November 10, 2026) Generative samplers for lattice field theory are commonly evaluated by their effective sample size (ESS). We show that a small global ESS need not imply inaccurate long-distance observables: errors spread over many short-distance modes can dominate this diagnostic.

Appendix A

The Rényi divergence and coarse graining section details the overlap identities and the monotonicity of the flow. The independence and marginal tests section covers acceptance rates and stationary distributions for these tests. p. 1

Where the paper says it

The distinction also explains why weighting only the modes on which an observable depends avoids noise from other modes, in the spirit of the local factorization of Ref. For the second question, we derive the mean acceptance rate of an independence test and show that a test on a few collective variables (low-dimensional functions of the configuration such as the magnetization), which we call the marginal test, accepts at least as often as the full test at their respective stationary distributions.

The overlap and reweighting in the free field section discusses single Gaussian modes, correlation-volume law, and reweighting estimators. p. 5
Where the paper says it

Reweighting Q Let O depend only on a set A of modes, and wA = k∈A wk, where wk = pk(xk)/qk(xk) is the normalized single-mode weight. In the free field, provided the required reweighted moments are finite (proof in Appendix A 3): Proposition 2 (Three estimators).

The composite chain in the free field section details the exact solution for a zero mode and its relaxation properties. p. 1
Where the paper says it

In the free field, the resulting composite chain reaches a given accuracy of the mode variances with a number of updates independent of volume at fixed correlation length and uniformly bounded relative errors in the proposal mode variances. Thus, once the infrared input is supplied, poor global overlap need not imply exponentially costly relaxation to a prescribed accuracy.

Appendix C examines an alternative determination of pM from constrained simulations of the action. p. 2
Where the paper says it

VII concludes. p. 2

Where the paper says it

Section VI discusses what evaluating samplers by scale implies for their assessment and design, and Sec. VII concludes.

Derivations are collected in Appendix A, Appendix B studies the composite chain at fixed correlation length, and Appendix C examines how to estimate and use the target marginal.

Appendix B

The composite chain at fixed correlation length discusses how the runs of Sec. V D do not separate the dependence on ξ from that on L. p. 0

Where the paper says it

This poor scaling has been attributed to the multiscale structure of the problem near critical points, which networks of moderate size do not capture efficiently [28]; in diffusion models the same structure makes the score ill-conditioned, so that the number of time steps of the generating process can grow, a difficulty mitigated by factorizing the distribution across scales [29]. The dependence on the volume is even more severe: the quality of a fixed model built from local transformations degrades exponentially under volume transfer [7], and for local theories a single global independence proposal can become impractical at realistic box sizes [30].

Appendix C

Appendix C examines an alternative determination of pM from constrained simulations of the action.

The first question is answered by resolving the overlap by scale. The asymptotic ESS fraction, e−D2(p∥q), measures the aggregate mismatch, whereas a long-distance observable probes the marginal distribution of the variables on which it depends. For the second question, the marginal test shows how a given target marginal can be used in full-configuration sampling without the exponential penalty of global overlap for the observables studied here. The composite chain generally has a biased stationary distribution at finite relaxation length and approaches the full target as the local kernel converges. The free-field solution bounds the number of local updates needed for a fixed accuracy of the mode variances independently of volume, at fixed correlation length and uniformly bounded relative errors in the proposal mode variances. The remaining limitations concern infrared relaxation and marginal-density estimation. The paper is organized as follows. p. 1

Where the paper says it

The two could be combined, but such a hybrid has not been tested (Sec. The paper is organized as follows.

Sections II and III give the general framework: Sec. p. 1
Where the paper says it

The paper is organized as follows. Sections II and III give the general framework: Sec.

II relates the ESS to the Rényi divergence and introduces the overlap flow, and Sec. p. 2
Where the paper says it

3 Rényi divergence and introduces the overlap flow, and Sec. III constructs the marginal test and the composite chain.

III constructs the marginal test and the composite chain. p. 2
Where the paper says it

3 Rényi divergence and introduces the overlap flow, and Sec. III constructs the marginal test and the composite chain.

Sections IV and V test both parts in parallel, exactly in the free field and numerically in two-dimensional ϕ 4 theory near its critical point. p. 2
Where the paper says it

III constructs the marginal test and the composite chain. Sections IV and V test both parts in parallel, exactly in the free field and numerically in two-dimensional ϕ 4 theory near its critical point.

Section VI discusses what evaluating samplers by scale implies for their assessment and design, and Sec. p. 2
Where the paper says it

Sections IV and V test both parts in parallel, exactly in the free field and numerically in two-dimensional ϕ 4 theory near its critical point. Section VI discusses what evaluating samplers by scale implies for their assessment and design, and Sec.

VII concludes. The first question is answered by resolving the overlap by scale The asymptotic ESS fraction, e−D2(p∥q), measures the aggregate mismatch, whereas a long-distance observable probes the marginal distribution of the variables on which it depends The trained OSQ samplers provide interacting evidence over the volumes studied. With reference parameters fixed from the benchmark, they reproduce ⟨M⟩ and ⟨M2⟩ to a few percent while their path log-weight variance grows approximately with volume, much of it apparently from path fluctuations rather than endpoint mismatch. p. 1
Where the paper says it

We test this picture with trained OSQ samplers for two-dimensional ϕ 4 theory near criticality. With reference parameters fixed from the benchmark, they reproduce ⟨|M|⟩ and ⟨M2 ⟩ to a few percent while their path-weight ESS fractions fall below 10−3 at L = 64; the test on M still accepts about 0.

For the second question, the marginal test shows how a given target marginal can be used in full-configuration sampling without the exponential penalty of global overlap for the observables studied here. Observables of M alone already follow from pM; the chain is useful for other infrared modes and for observables coupling them to M, subject to the remaining conditional error. p. 3
Where the paper says it

(7) The collective variables have their target distribution; the remaining variables retain the model’s conditional distribution. If pΦ is known, observables of Φ alone can already be evaluated directly.

The composite chain generally has a biased stationary distribution at finite relaxation length and approaches the full target as the local kernel converges. With exact marginal density ratios, the marginal test accepts at least as often as the full independence test at their respective stationary distributions. The free-field solution bounds the number of local updates needed for a fixed accuracy of the mode variances independently of volume, at fixed correlation length and uniformly bounded relative errors in the proposal mode variances. The remaining limitations concern infrared relaxation and marginal-density estimation.

References

[1] K. Cranmer, G. Kanwar, S. Racanière, D. J. Rezende, and P. E. Shanahan, Advances in machine-learning-based sampling motivated by lattice quantum chromodynamics [2] G. p. 0

Where the paper says it

The target marginal must be supplied or computed separately; estimating it from the samplers’ own weights can inherit the poor global overlap. INTRODUCTION Generative models are increasingly used to sample field configurations in lattice field theory; for reviews of machine-learning methods in lattice QCD, see Refs.

Aarts, K. Fukushima, T. Hatsuda, A. Ipp, S. Shi, L Wang, and K Zhou, Physics-driven learning for inverse problems in quantum chromodynamics [3] A Tomiya, Machine learning for lattice QCD [4] M. p. 0
Where the paper says it

The target marginal must be supplied or computed separately; estimating it from the samplers’ own weights can inherit the poor global overlap. INTRODUCTION Generative models are increasingly used to sample field configurations in lattice field theory; for reviews of machine-learning methods in lattice QCD, see Refs.

S Albergo, G. Kanwar, and P. E Shanahan, Flowbased generative models for Markov chain Monte Carlo in lattice field theory [5] G. p. 0
Where the paper says it

The target marginal must be supplied or computed separately; estimating it from the samplers’ own weights can inherit the poor global overlap. INTRODUCTION Generative models are increasingly used to sample field configurations in lattice field theory; for reviews of machine-learning methods in lattice QCD, see Refs.

Kanwar, M. S Albergo, D Boyda, K Cranmer, D C Hackett, S Racanière, D J Rezende, and P E Shanahan Equivariant flow-based sampling for lattice gauge theory [6] L Del Debbio, J Marsh Rossney, and M Wilson Efficient modeling of trivializing maps for lattice phi 4 theory using normalizing flows [7] R Abbott, M. p. 0
Where the paper says it

The central obstacle to this program is the scaling with the number of degrees of freedom. For flow-based samplers of ϕ 4 theory, the cost of training a model to a given sampling efficiency grows extremely rapidly as the continuum limit is approached [6], and flow-based trivializing maps with a fixed architecture reduce the autocorrelation ∗ satoshi.

S Albergo, A Botev, D Boyda, K Cranmer, D C Hackett, A G. D. G. Matthews Equivariant flow-based sampling of lattice QCD [8] M C N Cheng and N Stratikopoulou Lecture notes on normalizing flows for lattice quantum field theories [9] L Wang, G Aarts, and K Zhou Diffusion models as stochastic quantization in lattice field theory [10] K Fukushima and S Kamata Stochastic quantization and diffusion models [11] O Vega, J Komijani, A El-Khadra, and M Marinkovic Group-equivariant diffusion models for lattice field theory [12] Y. p. 0
Where the paper says it

INTRODUCTION Generative models are increasingly used to sample field configurations in lattice field theory; for reviews of machine-learning methods in lattice QCD, see Refs. Examples include normalizing flows [4–8] and diffusion models, which are closely related to stochastic quantization [9, 10].

-y. Tan, G Aarts, D E Habibi, B Lucini, and L Wang Diffusion models for sampling near criticality in lattice field theories [13] D Grinwald, A Singha, K Kahouli, K Jansen, and S Nakajima Data free training of diffusion models for lattice field theories [14] D Schubring Diffusion models for the O(3) nonlinear sigma model [15] Q Zhu, G Aarts, W Wang, K Zhou, and L Wang Physics-conditioned diffusion models for lattice gauge theory [16] G Aarts, D E Habibi, A Ipp, D I Müller, T R. Ranner L Wang W Wang and Q Zhu Generalizable equivariant diffusion models for non-Abelian lattice gauge theory [17] H Alharazin J Yu Panteleeva and B. p. 0
Where the paper says it

Diffusion models have been applied to scalar field theories and sigma models [11–14] and to gauge theories [15–18], also with dynamical fermions [19, 20]; they have been extended to theories with a sign problem [21] and trained without reference ensembles [13, 22]. Other samplers are based on stochastic quantization formulated as optimal control (OSQ) [23], on stochastic path sampling [24] or on non-equilibrium transport with a learned drift [25].

-D Sun Diffusion models for SU(2) lattice gauge theory in two dimensions [18] J Komijani, M K Marinkovic, and L Turgut Diffusion model for SU(N) gauge theories [19] O Vega and A X El-Khadra Sampling the Schwinger model with gauge-equivariant diffusion in lattice field theory [20] A Tomiya, Continuous Hasenbusch transport towards gauge diffusion with fermions [21] G Aarts, D E Habibi, L Wang, and K Zhou Combining complex Langevin dynamics with score-based and energybased diffusion models [22] A Tomiya Lattice configuration generation with a selflearning diffusion model [23] L Wang Stochastic quantization as optimal control Diffusion models for sampling near criticality in lattice field theories [24] S Chen, M Qian, G Aarts, B Lucini, and K Zhou Stochastic path sampler for lattice field theory [25] M S Albergo and E Vanden-Eijnden NETS: A nonequilibrium transport sampler [26] K A Nicoli, S Nakajima, N Strodthoff, W Samek, K. p. 0
Where the paper says it

Diffusion models have been applied to scalar field theories and sigma models [11–14] and to gauge theories [15–18], also with dynamical fermions [19, 20]; they have been extended to theories with a sign problem [21] and trained without reference ensembles [13, 22]. Other samplers are based on stochastic quantization formulated as optimal control (OSQ) [23], on stochastic path sampling [24] or on non-equilibrium transport with a learned drift [25].

-R Müller, and P Kessel Asymptotically unbiased estimation of physical observables with neural samplers [27] D Albandea L Del Debbio P Hernández R Kenway J Marsh Rossney A Ramos Learning trivializing flows [28] J Komijani M K Marinkovic Generative models for scalar field theories how to deal with poor scaling? p. 0
Where the paper says it

The target marginal must be supplied or computed separately; estimating it from the samplers’ own weights can inherit the poor global overlap. INTRODUCTION Generative models are increasingly used to sample field configurations in lattice field theory; for reviews of machine-learning methods in lattice QCD, see Refs.

[29] F Guth, S Coste V De Bortoli and S Mallat Wavelet score-based generative modeling in lattice field theory [30] J Finkenrath Fine grinding localized updates via gauge equivariant flows in the 2D Schwinger model [31] R Abbott D Boyda Y Fu D C Hackett G Kanwar F Romero-López P E Shanahan Variance reduction in lattice QCD observables via normalizing flows [32] R Abbott Y Fu D C Hackett G Kanwar F RomeroLópez and P E Shanahan Learning the generating functional for variance reduction in lattice QCD [33] M Caselle E Cellini A Nada and M Panero Stochastic normalizing flows as non-equilibrium transformations [34] A Bulgarelli E Cellini and A Nada Scaling of stochastic normalizing flows in SU(3) lattice gauge theory [35] C Bonanno A Bulgarelli E Cellini A Nada D Panfalone D Vadacchino and L Verzichelli Scaling flowbased approaches for topology sampling in SU(3) gauge theory [36] G Kanwar and O Vega Annealed sampling of diffusion models in lattice field theory [37] D Sexty Accurate sampling from diffusion models [38] M Gerdes P de Haan C Rainone R Bondesan and M C N Cheng Learning lattice quantum field theories with equivariant continuous flows [39] S Bacchio P Kessel S Schaefer and L Vaitl Learning trivializing gradient flows for lattice gauge theories [40] J Finkenrath Tackling critical slowing down using global correction steps with equivariant flows the case of the Schwinger model [41] S.-H. Li and L Wang Neural network renormalization group [42] T Marchand M Ozawa G Biroli and S Mallat Multiscale data-driven energy estimation and generation [43] R Abbott M S Albergo D Boyda D C Hackett G Kanwar F Romero-López P E Shanahan Multiscale normalizing flows for gauge theories [44] M Bauer R Kapust J M Pawlowski and F L Temmen Super-resolving normalising flows for lattice field theories [45] A Singha E Cellini K A Nicoli K Jansen S Kühn and S Nakajima Multilevel generative samplers for investigating critical phenomena in International Conference on Learning Representations (ICLR 2025) [46] A Singha J Kauffmann E Cellini K Jansen and S Nakajima Scalable generative sampling and multilevel estimation for lattice field theories near criticality [47] J Cotler S Rezchikov Renormalizing diffusion models [48] K Masuki Y Ashida Generative diffusion model with inverse renormalization group flows [49] K Masuki Y Ashida Renormalization group flow matching for scalable local generative modeling

The first question is answered by resolving the overlap by scale The asymptotic ESS fraction, e−D2(p∥q), measures the aggregate mismatch, whereas a long-distance observable probes the marginal distribution of the variables on which it depends. For the second question, the marginal test shows how a given target marginal can be used in full-configuration sampling without the exponential penalty of global overlap for the observables studied here The composite chain generally has a biased stationary distribution at finite relaxation length and approaches the full target as the local kernel converges. p. 1

Where the paper says it

The marginal test alone has the correct stationary distribution for the tested variables when their density ratio is known; at fixed values of these variables it retains the model’s conditional distribution. The composite chain is generally biased at finite relaxation length.

With exact marginal density ratios, the marginal test accepts at least as often as the full independence test at their respective stationary distributions. The free-field solution bounds the number of local updates needed for a fixed accuracy of the mode variances independently of volume, at fixed correlation length and uniformly bounded relative errors in the proposal mode variances. The remaining limitations concern infrared relaxation and marginal-density estimation.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Scale-resolved diagnostics for generative samplers can replace global effective sample size (ESS) with a measure based on overlap between coarse-grained model and target distributions, allowing researchers to distinguish between errors spread over many short-distance modes and those affecting long-distance observables.

  2. AI systems can utilize a marginal test on a few collective variables, such as magnetization, which maintains an acceptance rate of order one at their stationary distributions even when global overlap is low, effectively reducing the exponential penalty associated with global reweighting for specific observables.

  3. Generative samplers can be designed to incorporate infrared information directly into the proposal mechanism or by using a marginal test on selected variables to correct errors in those modes, rather than relying solely on expensive full-configuration proposals.

  4. AI systems can implement composite Markov chains where an exact local kernel update is used after a marginal test, allowing them to reduce the remaining mismatch at finite relaxation length in lattice field theory simulations.

  5. The system can be equipped with a mechanism to dynamically adjust reference parameters during training, allowing the model to learn the infrared distributions of collective variables and adapt its proposal quality without requiring fixed, pre-determined target distributions.

Abstract

Generative samplers for lattice field theory are commonly evaluated by their effective sample size (ESS). We show that a small global ESS need not imply inaccurate long-distance observables: errors spread over many short-distance modes can dominate this diagnostic. We introduce a scale-resolved diagnostic based on the overlap between coarse-grained model and target distributions. At the lattice cutoff, this overlap determines the asymptotic ESS fraction; lower cutoffs isolate long-wavelength errors. In the exactly solvable free theory, two samplers with comparable ESS yield a 32% bias and no bias in the susceptibility, respectively. In two-dimensional ϕ 4 theory near criticality, trained diffusion samplers show the same separation of scales: their path-weight ESS fraction falls rapidly over the volumes studied, while low moments of the magnetization remain close to their target values. Varying the reference parameters substantially changes infrared observables, with little effect on the log-weight variance. Given the target magnetization distribution, a marginal Metropolis test maintains an acceptance rate of order one in our tests. Combined with exact local updates, it reduces the remaining mismatch, although residual bias remains at finite relaxation length. In the free field, uniformly bounded relative errors in the proposal mode variances give a volume-independent bound on the number of updates needed for a prescribed accuracy of the chain's stationary mode variances at fixed correlation length. The target marginal must be supplied or computed separately; estimating it from the samplers' own weights can inherit the poor global overlap.

Sources

Related papers