Scalable Generative Sampling and Multilevel Estimation for Lattice Field Theories Near Criticality
summary
The gist
Scalable Generative Sampling and Multilevel Estimation for Lattice Field Theories Near Criticality introduces a multiscale generative sampler inspired by renormalization-group ideas that models the…
In short
This work introduces a multiscale generative sampler for lattice field theories near criticality to overcome critical slowing down. It models the Boltzmann distribution using a hierarchy of coarse-to-fine levels, employing conditional Gaussian mixture models and continuous normalizing flows to efficiently sample complex distributions. This method achieves significantly faster sampling and variance reduction than standard Hybrid Monte Carlo methods on large lattices.
Key concepts
- Critical Slowing Down
- This is a problem where simulation methods struggle near phase transitions because the system takes an extremely long time to explore its state space. Standard Markov chain methods become very inefficient, requiring exponentially more computation to get independent samples as the lattice size increases.
- Multiscale Generative Sampler
- This technique builds a sampling strategy by breaking down the problem across different length scales. It starts with coarse levels that capture long-wavelength physics and progressively adds finer levels to model short-distance fluctuations, allowing for efficient generation of complex field configurations.
- Conditional Gaussian Mixture Model (GMM)
- A GMM is a tractable mathematical model used at intermediate and fine scales to approximate the local dependence of new variables on the previously sampled coarse fields. It provides an explicit, easy-to-sample density function that helps guide the sampling process.
- Continuous Normalizing Flow (CNF)
- A CNF is a type of neural network used to refine samples generated by simpler models like the GMM. It uses a continuous mathematical process to map noisy or imperfect samples into a more accurate distribution, improving the quality of the final field configurations.
Terminology used across episodes
This episode discusses
- Scalable Generative Sampling and Multilevel Estimation for Lattice Field Theories Near Criticality · Paper Radio
- FLAG Review 2024
- The anomalous magnetic moment of the muon in the Standard Model
- The anomalous magnetic moment of the muon in the Standard Model: an update
- Critical slowing down and error analysis in lattice QCD simulations
- Neural Ordinary Differential Equations · Paper Radio
- Equivariant flow-based sampling for lattice gauge theory
- Flow-based sampling for fermionic lattice field theories
- Applications of flow models to the generation of correlated lattice QCD ensembles
- Variance reduction in lattice QCD observables via normalizing flows
- Improvement of Heatbath Algorithm in LFT using Generative models
- Toward Scalable Normalizing Flows for the Hubbard Model
- Tackling the Sign Problem in the Doped Hubbard Model with Normalizing Flows · Paper Radio
- Learning Trivializing Gradient Flows for Lattice Gauge Theories
- Sampling the lattice Nambu-Goto string using Continuous Normalizing Flows
- Non-Perturbative Trivializing Flows for Lattice Gauge Theories
- Stochastic normalizing flows as non-equilibrium transformations
- Numerical determination of the width and shape of the effective string using Stochastic Normalizing Flows
- Flow-Based Sampling for Entanglement Entropy and the Machine Learning of Defects
- Scaling of Stochastic Normalizing Flows in SU(3) lattice gauge theory
- Scaling flow-based approaches for topology sampling in SU(3) gauge theory
The paper
Scalable Generative Sampling and Multilevel Estimation for Lattice Field Theories Near Criticality · Read on arXiv
Berlin Institute for the Foundations of Learning and Data (BIFOLD) · Technische Universität Berlin · Higgs Centre for Theoretical Physics, School of Physics and Astronomy, The University of Edinburgh · Deutsches Elektronen-Synchrotron (DESY) · RIKEN Center for Advanced Intelligence Project (AIP)
Sampling lattice field theories near criticality is severely hindered by critical slowing down, which makes standard Markov chain methods increasingly inefficient at large lattice volumes. We introduce a multiscale generative sampler, inspired by renormalization-group ideas, that models the Boltzmann distribution through a coarse-to-fine hierarchy across length scales. At each level, a conditional Gaussian mixture model captures the main local dependence of newly introduced variables on the already-sampled coarse field, while a masked continuous normalizing flow refines the remaining conditional structure. Coarse levels encode the dominant long-wavelength modes, and finer levels progressively add short-distance fluctuations. In addition, because the architecture preserves coarse fields exactly during refinement, it provides exact restriction maps at no additional computational cost and directly enables asymptotically unbiased Multilevel Monte Carlo (MLMC) variance reduction. For the two-dimensional scalar ϕ 4 theory near criticality, the method achieves integrated autocorrelation times orders of magnitude smaller than Hybrid Monte Carlo (HMC) on large volumes, maintains high importance-sampling efficiency relative to other generative baselines, and reproduces asymptotically unbiased physical observables in statistical agreement with long HMC simulations.
DOI: 10.1103/jjss-lwzv
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: "Scalable Generative Sampling and Multilevel Estimation for Lattice Field Theories Near Criticality".
Mira: Here's the gist of the paper. Scalable Generative Sampling and Multilevel Estimation for Lattice Field Theories Near Criticality introduces a multiscale generative sampler inspired by renormalization-group ideas that models the Boltzmann distribution through a coarse-to-fine hierarchy across length scales, achieving orders of magnitude smaller integrated autocorrelation times than Hybrid Monte Carlo on large volumes for the two-dimensional scalar theory at criticality Near Criticality The Core Problem and Proposed Solution Sampling lattice field theories near criticality is severely hindered by critical slowing down, which makes standard Markov chain methods increasingly inefficient at large lattice volumes The authors introduce a multiscale generative sampler that models the Boltzmann distribution through a coarse-to-fine hierarchy across length scales At each level, a conditional Gaussian mixture model captures the main local dependence of newly introduced variables on the already-sampled coarse field, while a masked continuous normalizing flow refines the remaining conditional structure Coarse levels encode the dominant long-wavelength modes, and finer levels progressively add short-distance fluctuations Multiscale Generative Framework The method generalizes a multilevel generative strategy to continuous lattice field theories by representing the Boltzmann measure through a hierarchy of coarse-to-fine conditional densities The exact distribution can then be factorized as p(ϕ) = p(ϕ(0)) / Ymax sum l=1 p(ϕ(l)) ϕ(≤l−1) At intermediate and finest levels, the exact coarse-to-fine conditional distributions are approximated using a conditional Gaussian mixture model (GMM) This GMM provides an explicit, tractable density that can be sampled and evaluated efficiently, making it well-suited as a local conditional prior within the hierarchical framework Refinement Mechanism and Exact Restriction Maps To model the remaining mismatch after the GMM approximation, a conditional continuous normalizing flow (CNF) is used to refine GMM samples The map fl is generated by integrating a conditional neural ordinary differential equation, dϕ(t) dt = vlϕ(t), t; ϕ(≤l−1), t ∈
0, 1: This hierarchical construction enables scalable coarse-to-fine sampling at large lattice volumes, particularly in the critical regime where both conventional sampling methods and single-scale generative models become increasingly inefficient The resulting hierarchical construction enables the use of an MLMC estimator, which provides variance reduction near criticality beyond plain importance sampling Performance and Variance Reduction Gains The method achieves integrated autocorrelation times orders of magnitude smaller than Hybrid Monte Carlo (HMC) on large volumes for the two-dimensional scalar theory at criticality The hierarchical construction induces a built-in coupling between neighboring resolutions, providing exactly the structure needed to construct multilevel Monte Carlo (MLMC) estimators In comparison to plain importance sampling (IS) at equal wall-clock budget, the MLMC estimator achieves lower variance than IS at both lattice sizes The MLMC estimator achieves lower variance than plain IS at the same computational budget at both lattice sizes, with z < 2 indicating statistical consistency with the IS reference Computational Efficiency and Scaling The wall-clock time required to match the effective number of independent samples obtained in the HMC reference run shows that our method achieves speedups of 48× at L = 16, 1245× at L = 64, and 66× at L = 256 relative to HMC The hierarchical construction yields a substantial computational advantage at large lattice sizes when both sampling efficiency and generation cost are taken into account The training ESS/N for our method reaches higher efficiency in less training time at both L = 64 and L = 128 Our method is the only one that avoids all three failure modes simultaneously in the tested regime Physical Consistency and Observables The Dense CNF and our method reproduce the HMC reference values for connected susceptibility χconn and ⟨m⟩ within 1–2σ across all tested lattice sizes By contrast, SR-NF estimates become increasingly unreliable in large volumes: at L = 256, where IS efficiency drops below.
Paper summary: Kai: So we're looking at this paper now titled Scalable Generative Sampling and Multilevel Estimation for Lattice Field Theories Near Criticality. It seems like they are tackling that problem of critical slowing down in lattice field theories, which makes standard Markov chain methods really slow when you try to sample large volumes <ref:2604.10209#pg1>. Mira, what's the main idea here?
Mira: The main thing is they introduce a multiscale generative sampler that uses ideas from renormalization group theory to model the Boltzmann distribution across different length scales <ref:2604.10209#pg1>. They are essentially building a system that goes from coarse to fine, and at each level, they use a conditional Gaussian mixture model for the local part and then a continuous normalizing flow to fix the rest of the structure <ref:2604.10209#pg1>.
Lev: From an engineering standpoint, modeling it that way means they are trying to handle those long-wavelength modes at coarse levels and then progressively adding short-distance fluctuations at finer levels <ref:2604.10209#pg2>. That sounds like a way to manage the complexity of the whole field configuration rather than just sampling it all at once <ref:2604.10209#pg3>.
Kai: Exactly, so they are trying to get better generation procedures without needing an impossibly large number of samples, and they claim this is a way to achieve orders of magnitude smaller integrated autocorrelation times compared to Hybrid Monte Carlo on large volumes for the two-dimensional scalar theory at criticality <ref:2604.10209#pg1>.
Mira: That's a big claim because we know how slow those standard methods get near criticality, where the correlation length diverges <ref:2604.10209#pg3>. They are saying this multiscale approach is fundamentally different in terms of how it handles that divergence <ref:2604.10209#pg3>.
Lev: If their method works, it means we could potentially run simulations on much larger lattices without getting stuck waiting for the autocorrelation time to drop significantly <ref:2604.10209#pg3>. I'd wonder about the overhead of setting up all those levels, though <ref:2604.10209#pg3>.
Kai: That's where I get curious about how computationally expensive that generation procedure actually is compared to what we have on superconducting qubits Kai. The paper mentions they use a masked continuous normalizing flow to refine the Gaussian mixture model at finer levels <ref:2604.10209#pg1>.
Mira: They are using the conditional Gaussian mixture model as an explicit, tractable density for sampling, and then the flow refines that approximation for the rest of the structure <ref:2604.10209#pg1>. It's a two-step process to keep things manageable <ref:2604.10209#pg1>.
Paper summary: Lev: So they are trading off tractability for refinement, which is a trade we deal with constantly when we try to map complex physical systems onto simpler models for error correction Lev. The paper says this hierarchical construction enables the use of a multilevel Monte Carlo estimator, or MLMC estimator <ref:2604.10209#pg3>.
Kai: And that MLMC part is interesting because it provides variance reduction near criticality beyond just plain importance sampling when you look at the numbers <ref:2604.10209#pg3>. They found that this MLMC estimator achieves lower variance than plain importance sampling at both lattice sizes for the two-dimensional scalar theory at criticality <ref:2604.10209#pg3>.
Mira: That variance reduction is a key point, because it means you get the same precision with fewer total samples, which is huge when you're already fighting critical slowing down <ref:2604.10209#pg3>. They also report that for the connected susceptibility and the expectation value of the absolute mass, they reproduce Hybrid Monte Carlo reference values within one to two sigma across all tested lattice sizes <ref:2604.10209#pg3>.
Lev: Reproducing HMC values is a strong consistency check, especially when you're dealing with something as sensitive as critical points <ref:2604.10209#pg3>. But they do flag that the Conditional Neural Ordinary Differential Equation used to generate the flow has a certain cost associated with it <ref:2604.10209#pg1>.
Kai: It seems like a trade-off between generation cost and sampling efficiency, which is something we deal with constantly in hardware design Kai. They also mention that the wall-clock time required to match the effective number of independent samples from the HMC reference run shows speedups ranging from forty-eight times at L equals sixteen up to one thousand two hundred forty-five times at L equals sixty-four <ref:2604.10209#pg3>.
Mira: The scaling is impressive, especially reaching six hundred sixty times speedup at a lattice size of two fifty-six <ref:2604.10209#pg3>. That suggests the multiscale structure really pays off when you push to larger volumes <ref:2604.10209#pg3>. It’s the multiscale structure that enables this massive gain in efficiency for them <ref:2604.10209#pg1>.
Lev: From an error correction view, if we could generate these states efficiently, it would drastically reduce the overhead of generating new configurations in simulations Lev. The paper does mention that their method is one of the few that avoids all three failure modes simultaneously in the regime they tested <ref:2604.10209#pg3>.
Kai: So to summarize this paper on Scalable Generative Sampling and Multilevel Estimation for Lattice Field Theories Near Criticality, it proposes a multiscale generative sampler inspired by renormalization group ideas to tackle critical slowing down <ref:2604.10209#pg1>. The core idea is building a coarse-to-fine hierarchy where conditional Gaussian mixture models capture local dependence and continuous normalizing flows refine the structure <ref:2604.10209#pg1>.
Paper summary: Mira: What this means for us is that they are getting much better sampling efficiency near phase transitions by using this hierarchical approach instead of standard methods like Hybrid Monte Carlo, which suffer from very slow integrated autocorrelation times <ref:2604.10209#pg3>. The results show that this method achieves orders of magnitude smaller autocorrelation times than HMC on large volumes for the two-dimensional scalar theory at criticality <ref:2604.10209#pg1>.
Lev: It implies that we might be able to study these critical systems on much larger scales than currently feasible with conventional MCMC, which is a big deal for testing error correction schemes Lev. The paper also notes that the Dense Continuous Normalizing Flow reproduces Hybrid Monte Carlo reference values for connected susceptibility and the expectation value of the absolute mass within one to two sigma across all tested lattice sizes <ref:2604.10209#pg3>.
Kai: So, when we look at this paper, it’s about finding a way to make generative sampling scalable for critical systems by using a coarse-to-fine structure <ref:2604.10209#pg1>. It tackles the problem of critical slowing down by modeling the distribution across different length scales with conditional Gaussian mixtures and flow refinement <ref:2604.10209#pg1>.
Mira: And it shows that this method provides a variance reduction benefit when using multilevel Monte Carlo estimators compared to plain importance sampling at equal wall-clock budgets <ref:2604.10209#pg3>. This suggests a way to get more reliable results for physical observables near criticality without needing an overwhelming number of samples <ref:2604.10209#pg3>.
Lev: The authors do mention that the method is only tested in a specific regime, and they don't claim it works everywhere <ref:2604.10209#pg3>. They specifically say that SR-NF estimates become increasingly unreliable in large volumes, even at two fifty-six <ref:2604.10209#pg3>. That’s a clear limitation they admit <ref:2604.10209#pg3>.
Kai: So the title Scalable Generative Sampling and Multilevel Estimation for Lattice Field Theories Near Criticality points to this entire approach <ref:2604.10209#pg1>. It’s about using that multiscale structure to build a sampling method that scales well and provides good results near criticality <ref:2604.10209#pg3>.
Mira: Ultimately, the implication is that we can get more reliable statistics for these hard-to-reach critical points by combining generative modeling with multilevel estimation techniques <ref:2604.10209#pg3>. It gives us a new tool to handle the challenges of sampling complex lattice field theories near criticality <ref:2604.10209#pg1>.
Lev: That’s the gist of it, focusing on how they build this scalable sampler and what it means for studying those critical points <ref:2604.10209#pg3>.
Conclusion: Kai: So we’re wrapping up this look at "Scalable Generative Sampling and Multilevel Estimation for Lattice Field Theories Near Criticality." It’s about using a specific way to sample these complex field theories near phase transitions that avoids the usual problems with critical slowing down.
Mira: Right, so they're taking renormalization group ideas and building this multiscale sampler. They use coarse-to-fine levels to model the physics across different length scales, which is supposed to tame how slow standard methods get when you look at large volumes.
Lev: From an error correction standpoint, if this works as advertised, it means we could potentially run simulations on lattices much larger than what’s currently practical with traditional Markov chain methods. The overhead of setting up these levels is the real question for hardware implementation.
Kai: Exactly, and they show some really strong numbers. They claim orders of magnitude smaller integrated autocorrelation times than hybrid monte carlo on large volumes for the two dimensional scalar theory at criticality. That’s a huge deal for getting reliable physics out of simulations.
Mira: It does mean that you can get better statistics near these critical points without needing an impossibly huge number of samples. They also used this structure to build a multilevel monte carlo estimator, which gives them variance reduction when compared to plain importance sampling at the same wall-clock budget.
Lev: That variance reduction is what makes it interesting for real hardware, because it means you get that precision with less computational time. But they do have their limits; they admit that the method becomes less reliable in very large volumes, specifically above two hundred fifty six lattice units.
Kai: So the authors are essentially showing a way to make these complex critical systems more accessible computationally by layering generative models on top of one another. It’s a new tool for studying those hard-to-reach critical points.
More episodes
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2606.20029-Finite-Volume Step Scaling at Fixed Truncation in (2+1)D Hamiltonian U(1) Lattice Gauge Theory
- 2603.16266-Lattice QCD study of the K*(892) resonance at the physical point
- 2607.15742-Path optimization method for the sign problem: Insights from random matrix models
- The day in conversation for 2026-11-18
- Daily Summary for 2026-11-18
- Daily Summary for 2026-11-17
- 2608.18239-Searching for symmetric mass generation with staggered fermions in four dimensions
- The day in conversation for 2026-11-17