Generative sampling with physics-informed kernels

arXiv:2610.41864 · hep-lat, physics.data-an · Submitted 2026-10-30 · 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: "Generative sampling with physics-informed kernels".

Mira: Here's the gist of the paper. Generative sampling with physics-informed kernels constructs a generative network for Monte-Carlo sampling in lattice field theories and beyond, where layerwise propagation is computed and optimized independently on each layer The architecture uses physics-informed renormalisation group flows that provide access to the layerwise propagation step from one layer to the next in terms of a simple first order partial differential equation for the respective renormalisation group kernel through a given layer Thus, it transforms the generative task into that of solving once the set of independent and linear differential equations for the kernels of the transformation As these equations are analytically known, the kernels can be refined iteratively This allows us to structurally tackle out-of-domain problems generally encountered in generative models and opens the path to further optimisation The core mechanism of the architecture The key idea underlying the physics-informed RG flows is to exploit the full generality of the Wegner flow (3) for the distribution or related RG equations for other generating functions Instead of solving the flows for the distribution p t(phi), one solves them for pairs F t(phi), phi t(phi) A key idea underlying the physics-informed RG flows

9: is to exploit the full generality of the Wegner flow (3) for the distribution or related RG equations for other generating functions The key point of the PIRG in this approach is to analytically fix the action S t along this transformation Then, the underlying renormalisation group equation is a simple linear differential equation for the kernel phi i at each time step Addressing out-of-domain challenges The architecture is designed to structurally resolve both of the OOD problems mentioned in the beginning The two OOD problems are: 1.

Kai: related to the high dimensionality of the transformation between the two theories, which grows with the number of lattice sites and typically renders the learned distribution inaccurate This originates in the fact that interpolation problems in high dimensions are equally ill-conditioned as extrapolation problems in low dimensions 2.

Mira: a standard extrapolation problem and may be explained as follows: assume that the generative model has been trained with a given samples size related to a limited information content Systematic error control through kernel refinement The properties of the PIK-architecture include: (i) Independent Kernels: PIKs convert the global task of finding a precise invertible map from the base distribution p0 to the final distribution p1 within a deep network into independent tasks of computing the infinitesimal maps phi t(phi; F t) at each time step (ii) OOD resolution: The Wegner equation (7) provides us with an error control of this map In particular, curing the OOD problem with the PIRG pair (S t, phi t<) boils down to locally improving the solution phi t< to a known linear differential equation (iii) Optimisation: The properties (i) and (ii) are relevant for the setup and solution of PIKs for a given pair (F t, phi t<) Practical implementation in scalar field theories In practice, the independent kernel structure is illustrated using a path for the action S t that connects a simple base distribution to the targeted one For zero-dimensional phi 4-theory, this involves choosing time-dependent mass m t and coupling lambda t to transport a Gaußian into a multimodal interacting theory The flow in RG-time is solved numerically by a standard fourth order Runge-Kutta scheme with a step size of Delta t = 1/25, leaving us with N t = 51 independent kernels phi<t to be determined Figure 2 shows a comparison of the targeted (orange) and modelled (black) zero-dimensional phi 4-theory distribution at the final time t = 1 Systematic improvement workflow To systematically improve the sampling, one computes the residual of (11) with an approximate kernel phi<(0)t A correction Delta phi<(0)t that satisfies L phi<Delta phi<(0)t(phi) = r(0) c(phi), will lead to a corrected kernel phi<(1)t = phi<(0)t + Delta phi<(0)t that solves the Wegner equation (11) accurately This procedure can also be iterated as the Wegner equation is linear in phi<t, hence the superscripts, indicating the first iteration Future directions and applications The paper concludes by outlining future applications of the PIK-architecture The first prominent direction is the application of PIKs to higher-dimensional field theories as well as the extension to field theories with gauge or fermionic degrees of freedom Moreover, as the PIK-architecture enables general transformations of the measure it is tailor-made to tackle sign problems, where the action becomes complex Furthermore, the PIK-architecture enables further optimisations of the sampling process Lastly, going beyond sampling, the PIK-architecture is also envisioned to infer the action or Hamiltonian in a symbolic way from data Acknowledgments We thank Gert Aarts, Kenji Fukushima, Sander Hummerich, Thore Kolja Joswig, Ullrich Köthe, Timoteo Lee, Manfred Salmhofer, Robert Scheichl, Lingxiao Wang and Kai Zhou for discussions and collaborations on related subjects We acknowledge the DM-LFT Collaboration for stimulating discussions and continuous support throughout this work It is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy EXC 2181/1 - 390900948 (the Heidelberg STRUCTURES Excellence Cluster) We also acknowledge support by the state of Baden-Württemberg through bwHPC References

1: U.

Kai: Wolff, CRITICAL SLOWING DOWN, Nucl.

Paper summary: Kai: So to summarize this paper, they are focusing on efficiently computing high-dimensional integrals that are usually done with Markov chain Monte Carlo approaches. They see critical slowing down and topological freezing as problems where traditional sampling algorithms lose their efficiency when you approach the continuum limit of the theory

one–three: <ref:2610.41864#pg2,the continuum limit of the theory 1–3>.

Mira: These problems mean the autocorrelation time of those Markov chains diverges, which requires huge amounts of computational resources to overcome. So, these models aim to find transformations that move configurations from a theory with efficient sampling to configurations of a targeted theory

four–seven: <ref:2610.41864#pg2,these models aim to find transformations that move configurations from a theory>.

Lev: The paper tackles two out-of-domain problems that hamper progress in generative Machine Learning models for Monte Carlo simulations of lattice field theories. First, it’s related to the high dimensionality of the transformation between two theories, which grows with the number of lattice sites and makes the learned distribution inaccurate

two–three: .

Kai: And second is a standard extrapolation problem, which they explain by saying that a generative model trained on a limited sample size has limited information content. This limits how far you can extrapolate to new regions.

Mira: The physics-informed approach addresses this by using the Wegner equation to provide an error control of the map between distributions. Specifically, using the physics-informed RG pair, they can locally improve that kernel solution phi to a known linear differential equation.

Lev: That means curing the out-of-domain problem boils down to locally improving that kernel solution phi to a known linear differential equation, which is really a concrete way of controlling the error eight.

Conclusion: Kai: So, looking at the title "Generative sampling with physics-informed kernels," this paper is essentially proposing a new way to sample from complex lattice theories by using these physics-informed flows. The authors are Ihssen, Kapust, and Pawlowski.

Mira: What they really mean is that instead of trying to map the entire distribution in one go, they break the mapping into independent steps governed by simple linear differential equations for each layer's kernel. This makes the process structurally solvable rather than just a black box learning task.

Lev: For practical hardware, this suggests a way to handle complex sampling where you can check and correct things layer by layer instead of getting stuck trying to solve the whole high-dimensional map simultaneously on real hardware eighteen. The complexity is managed because each step is linear in terms of the kernel update.

Kai: It opens up a path for further optimization because once you have those analytically known differential equations, you can refine those kernels iteratively. This iterative refinement is key to making the generative models more robust when they encounter problems outside of their training data.

Mira: The implications are that this architecture could be applied to higher-dimensional field theories or even theories with gauge or fermionic degrees of freedom twenty <ref:2610.41864#pg1>. Plus, since it allows for general transformations of the measure, it is specifically tailored to tackle sign problems where the action becomes complex.

Lev: And beyond just sampling, they envision using this architecture symbolically to infer the action or Hamiltonian directly from data. That moves it past just generating samples and into a potential tool for understanding the physics itself twenty-seven.

Friederike Ihssen, Renzo Kapusta, §Jan M. Pawlowski

Institut für Theoretische Physik, University Heidelberg · ExtreMe Matter Institute EMMI, GSI

hep-lat, physics.data-an

Submitted: 2026-10-30

Updated: 2026-10-30

Comments: Proceedings of 42nd International Symposium on Lattice Field Theory (LATTICE2025), November 2-8 2025, TIFR, Mumbai, India

Journal ref: PoS LATTICE2025 (2026) 055

DOI: 10.22323/1.518.0055

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

Importance score: 83/100

The gist: Generative sampling with physics-informed kernels constructs a generative network for Monte-Carlo sampling in lattice field theories and beyond…

Key concepts

Physics-Informed RG Flows
This mechanism exploits the Wegner flow to transform the problem. Instead of solving for the probability distribution directly, it solves flows for generating functions, analytically fixing the action along this transformation. This simplifies the complex renormalization group equation into a simple linear differential equation for each kernel.
Independent Kernels
The architecture decomposes the global task of finding a map between two distributions into independent tasks at each time step. Each layer computes its own infinitesimal map, $\phi_t(\phi; F_t)$, based on the flow information, making the overall process more manageable.
Out-of-Domain (OOD) Resolution
The physics-informed RG pair $(S_t, \phi_t^<)$ provides error control for the transformation map. By locally improving the solution $\phi_t^<$ to satisfy a known linear differential equation, the architecture resolves high dimensionality and extrapolation issues common in generative models.

Terminology

Summary

Generative sampling with physics-informed kernels constructs a generative network for Monte-Carlo sampling in lattice field theories and beyond, where layerwise propagation is computed and optimized independently on each layer The architecture uses physics-informed renormalisation group flows that provide access to the layerwise propagation step from one layer to the next in terms of a simple first order partial differential equation for the respective renormalisation group kernel through a given layer Thus, it transforms the generative task into that of solving once the set of independent and linear differential equations for the kernels of the transformation As these equations are analytically known, the kernels can be refined iteratively This allows us to structurally tackle out-of-domain problems generally encountered in generative models and opens the path to further optimisation

The core mechanism of the architecture

The key idea underlying the physics-informed RG flows is to exploit the full generality of the Wegner flow (3) for the distribution or related RG equations for other generating functions Instead of solving the flows for the distribution p t(phi), one solves them for pairs F t(phi), phi t(phi) A key idea underlying the physics-informed RG flows [9] is to exploit the full generality of the Wegner flow (3) for the distribution or related RG equations for other generating functions The key point of the PIRG in this approach is to analytically fix the action S t along this transformation Then, the underlying renormalisation group equation is a simple linear differential equation for the kernel phi i at each time step

Addressing out-of-domain challenges

The architecture is designed to structurally resolve both of the OOD problems mentioned in the beginning The two OOD problems are:

  1. related to the high dimensionality of the transformation between the two theories, which grows with the number of lattice sites and typically renders the learned distribution inaccurate This originates in the fact that interpolation problems in high dimensions are equally ill-conditioned as extrapolation problems in low dimensions

  2. a standard extrapolation problem and may be explained as follows: assume that the generative model has been trained with a given samples size related to a limited information content

Systematic error control through kernel refinement

The properties of the PIK-architecture include:

(i) Independent Kernels: PIKs convert the global task of finding a precise invertible map from the base distribution p0 to the final distribution p1 within a deep network into independent tasks of computing the infinitesimal maps phi t(phi; F t) at each time step

(ii) OOD resolution: The Wegner equation (7) provides us with an error control of this map In particular, curing the OOD problem with the PIRG pair (S t, phi t<) boils down to locally improving the solution phi t< to a known linear differential equation

(iii) Optimisation: The properties (i) and (ii) are relevant for the setup and solution of PIKs for a given pair (F t, phi t<)

Practical implementation in scalar field theories

In practice, the independent kernel structure is illustrated using a path for the action S t that connects a simple base distribution to the targeted one For zero-dimensional phi 4-theory, this involves choosing time-dependent mass m t and coupling lambda t to transport a Gaußian into a multimodal interacting theory The flow in RG-time is solved numerically by a standard fourth order Runge-Kutta scheme with a step size of Delta t = 1/25, leaving us with N t = 51 independent kernels phi<t to be determined Figure 2 shows a comparison of the targeted (orange) and modelled (black) zero-dimensional phi 4-theory distribution at the final time t = 1

Systematic improvement workflow

To systematically improve the sampling, one computes the residual of (11) with an approximate kernel phi<(0)t A correction Delta phi<(0)t that satisfies L phi<Delta phi<(0)t(phi) = r(0) c(phi), will lead to a corrected kernel phi<(1)t = phi<(0)t + Delta phi<(0)t that solves the Wegner equation (11) accurately This procedure can also be iterated as the Wegner equation is linear in phi<t, hence the superscripts, indicating the first iteration

Future directions and applications

The paper concludes by outlining future applications of the PIK-architecture The first prominent direction is the application of PIKs to higher-dimensional field theories as well as the extension to field theories with gauge or fermionic degrees of freedom Moreover, as the PIK-architecture enables general transformations of the measure it is tailor-made to tackle sign problems, where the action becomes complex Furthermore, the PIK-architecture enables further optimisations of the sampling process Lastly, going beyond sampling, the PIK-architecture is also envisioned to infer the action or Hamiltonian in a symbolic way from data

Acknowledgments

We thank Gert Aarts, Kenji Fukushima, Sander Hummerich, Thore Kolja Joswig, Ullrich Köthe, Timoteo Lee, Manfred Salmhofer, Robert Scheichl, Lingxiao Wang and Kai Zhou for discussions and collaborations on related subjects We acknowledge the DM-LFT Collaboration for stimulating discussions and continuous support throughout this work It is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy EXC 2181/1 - 390900948 (the Heidelberg STRUCTURES Excellence Cluster) We also acknowledge support by the state of Baden-Württemberg through bwHPC

References

[1] U. Wolff, CRITICAL SLOWING DOWN, Nucl. p. 8

Where the paper says it

References [1] U. Wolff, CRITICAL SLOWING DOWN, Nucl.

Phys. B Proc. Suppl. 17 (1990) 93

[2] B. Alles, G. Boyd, M. D’Elia, A Di Giacomo and E Vicari, Hybrid Monte Carlo and topological modes of full QCD, Phys Lett. p. 8

Where the paper says it

Di Giacomo and E. Vicari, Hybrid Monte Carlo and topological modes of full QCD, Phys.

B 389 (1996) 107 [hep-lat/9607049]

[3] L Del Debbio, G.M Manca and E Vicari, Critical slowing down of topological modes, Phys Lett. p. 8

Where the paper says it

Manca and E. Vicari, Critical slowing down of topological modes, Phys.

B 594 (2004) 315 [hep-lat/0403001]

[4] M Luscher, Trivializing maps, the Wilson flow and the HMC algorithm, Commun. Math. Phys. 293 (2010) 899 [0907. p. 8

Where the paper says it

Luscher, Trivializing maps, the Wilson flow and the HMC algorithm, Commun. 293 (2010) 899 [0907.

5491]

[5] D J Rezende and S Mohamed, Variational Inference with Normalizing Flows, 5, 2015 [1505.05770]

[6] M S Albergo, G Kanwar and P E Shanahan, Flow-based generative models for Markov chain Monte Carlo in lattice field theory, Phys Rev. D 100 (2019) 034515 [1904. p. 8

Where the paper says it

Shanahan, Flow-based generative models for Markov chain Monte Carlo in lattice field theory, Phys. D 100 (2019) 034515 [1904.

12072]

[7] K A Nicoli, C J Anders, L Funcke, T Hartung, K Jansen, P Kessel et al., Estimation of Thermodynamic Observables in Lattice Field Theories with Deep Generative Models, Phys Rev. p. 8

Where the paper says it

Kessel et al. , Estimation of Thermodynamic Observables in Lattice Field Theories with Deep Generative Models, Phys.

Lett. 126 (2021) 032001 [2007. p. 8
Where the paper says it

, Estimation of Thermodynamic Observables in Lattice Field Theories with Deep Generative Models, Phys. 126 (2021) 032001 [2007.

07115]

[8] R Kapust, F Ihssen and J M Pawlowski, Generative sampling with physics-informed kernels, PoS LATTICE2025 (2026) 055 [2510.26678]

[9] F Ihssen and J M Pawlowski, Physics-informed renormalisation group flows, Annals Phys. 481 (2025) 170177 [2409. p. 8

Where the paper says it

Pawlowski, Physics-informed renormalisation group flows, Annals Phys. 481 (2025) 170177 [2409.

13679]

[10] F J Wegner, Some invariance properties of the renormalization group, J. Phys. C 7 (1974) 2098

[11] J Cotler and S Rezchikov, Renormalizing Diffusion Models, 2308.12355

[12] M S Albergo and E Vanden-Eijnden, NETS: A Non-Equilibrium Transport Sampler, 2410.02711

[13] B Máté and F Fleuret, Learning Interpolations between Boltzmann Densities, 2301.07388

[14] D C Hackett, C.-C Hsieh, S Pontula, M S Albergo, D Boyda, J-W Chen et al. p. 9

Where the paper says it

Pontula, M. Albergo, D.

, Flow-based sampling for multimodal and extended-mode distributions in lattice field theory, 2107. p. 9
Where the paper says it

Chen et al. , Flow-based sampling for multimodal and extended-mode distributions in lattice field theory, 2107.

00734

[15] S Chen, O Savchuk, S Zheng, B Chen, H Stoecker et al., Fourier-flow model generating Feynman paths, Phys Rev. p. 9

Where the paper says it

Wang et al. , Fourier-flow model generating Feynman paths, Phys.

D 107 (2023) 056001 [2211. p. 9
Where the paper says it

, Fourier-flow model generating Feynman paths, Phys. D 107 (2023) 056001 [2211.

03470]

[16] K A Nicoli, C J Anders, T Hartung, K Jansen, P Kessel and S Nakajima, Detecting and mitigating mode-collapse for flow-based sampling of lattice field theories, Phys Rev. D 108 (2023) 114501 [2302. p. 9

Where the paper says it

Nakajima, Detecting and mitigating mode-collapse for flow-based sampling of lattice field theories, Phys. D 108 (2023) 114501 [2302.

14082]

[17] J Köhler, L Klein and F Noé, Equivariant flows: Exact likelihood generative learning for symmetric densities, 2020

[18] M Gerdes, P de Haan, C Rainone, R Bondesan and M C N Cheng, Learning lattice quantum field theories with equivariant continuous flows, SciPost Phys. 15 (2023) 238 [2207. p. 9

Where the paper says it

Cheng, Learning lattice quantum field theories with equivariant continuous flows, SciPost Phys. 15 (2023) 238 [2207.

00283]

[19] R T Q Chen, Y Rubanova, J Bettencourt and D Duvenaud, Neural Ordinary Differential Equations, 1806.07366

[20] A Bulgarelli, E Cellini and A Nada, Scaling of stochastic normalizing flows in SU(3) lattice gauge theory, Phys Rev. D 111 (2025) 074517 [2412. p. 9

Where the paper says it

Nada, Scaling of stochastic normalizing flows in SU(3) lattice gauge theory, Phys. D 111 (2025) 074517 [2412.

00200]

[21] G Aarts, D E Habibi, A Ipp, D I Müller, T R Ranner, L Wang et al., Generalizable Equivariant Diffusion Models for Non-Abelian Lattice Gauge Theory, 2601. p. 9

Where the paper says it

Wang et al. , Generalizable Equivariant Diffusion Models for Non-Abelian Lattice Gauge Theory, 2601.

19552

[22] A Singha, D Chakrabarti and V Arora, Conditional normalizing flow for Markov chain Monte Carlo sampling in the critical region of lattice field theory, Phys Rev. D 107 (2023) 014512 [2207. p. 9

Where the paper says it

Arora, Conditional normalizing flow for Markov chain Monte Carlo sampling in the critical region of lattice field theory, Phys. D 107 (2023) 014512 [2207.

00980]

[23] J M Pawlowski, Aspects of the functional renormalisation group, Annals Phys. 322 (2007) 2831 [hep-th/0512261]

[24] Y Song, J Sohl-Dickstein, D P Kingma, A Kumar, S Ermon and B Poole, Score-Based Generative Modeling through Stochastic Differential Equations, 2011.13456

[25] Y Lipman, R T Q Chen, H Ben-Hamu, M Nickel and M Le, Flow Matching for Generative Modeling, 2210.02747

[26] 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, 3 [2503.08918]

[27] F Ihssen, R Kapust and J M Pawlowski, Solving real sign problems, in preparation (2026).

Conclusion

The PIK-architecture reduces the task of finding transformations between two distributions into a decoupled chain of layerwise transformations that are given by an analytically known renormalisation group equation This setup paves the way for many interesting applications and improvements of the PIKarchitecture The first prominent direction is the application of PIKs to higher-dimensional field theories as well as the extension to field theories with gauge or fermionic degrees of freedom Moreover, as the PIK-architecture enables general transformations of the measure it is tailor-made to tackle sign problems, where the action becomes complex Lastly, going beyond sampling, the PIK-architecture is also envisioned to infer the action or Hamiltonian in a symbolic way from data

--- Page 9 ---

The paper is 2610.41864539.

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. Physics Informed Kernels (PIKs) allow for structural tackle out-of-domain problems generally encountered in generative models by transforming generative tasks into solving once the set of independent and linear differential equations for the kernels of the transformation.

  2. The architecture enables independent kernels where each layerwise propagation step is computed and optimized separately, facilitating a trivial parallelisation as well as the nature of the analytically known differential equation for the kernel at a given action.

  3. The system can achieve OOD resolution by using analytical access to the Wegner equation (7) and (11) to compute a correction Δφt that satisfies Lφt Δφt = r(0), leading to a corrected kernel φt=φ(0) + Δφt(0) that solves the Wegner equation (11) accurately.

  4. The system allows for optimisation through parameter-conditional flow, enabling the enforcement of the same functional form of the action at each time step, allowing observables to be computed continuously without sampling the configurations for each parameter.

  5. The PIK architecture is tailored to tackle sign problems, where the action becomes complex, making it particularly relevant for real-time simulations and finite density field theories.

Abstract

We construct a generative network for Monte-Carlo sampling in lattice field theories and beyond, for which the layerwise propagation is computed and optimised independently on each layer. The architecture uses physics-informed renormalisation group flows that provide access to the layerwise propagation step from one layer to the next in terms of a simple first order partial differential equation for the respective renormalisation group kernel through a given layer. Thus, it transforms the generative task into that of solving once the set of independent and linear differential equations for the kernels of the transformation. As these equations are analytically known, the kernels can be refined iteratively. This allows us to structurally tackle out-of-domain problems generally encountered in generative models and opens the path to further optimisation. We illustrate the practical feasibility of the architecture within simulations in scalar field theories.

Sources

Related papers