Diffusion models for the O(3) nonlinear sigma model

arXiv:2610.39568 · cond-mat.stat-mech, cs.LG, hep-lat · Submitted 2026-10-29 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Diffusion models for the O(3) nonlinear sigma model".

Mira: Here's the gist of the paper. A score-based diffusion model is trained on the two-dimensional classical Heisenberg model to investigate tradeoffs between computational cost and accuracy in generating configurations for lattice field theories Model Application and Test Case The diffusion model is trained on the two-dimensional classical Heisenberg model, which serves as a test case for evaluating diffusion models on lattice field theories The O(3) model shares qualitative features with four-dimensional Yang–Mills theory, including asymptotic freedom, emergence of a mass gap, and topological instantons This model is useful as a benchmark for assessing diffusion-based sampling methods with potential applications to the higher CP N−1 family and the qualitatively related 4D lattice gauge theories Methodological Investigations The authors investigate two complementary methodological questions motivated by the computational cost of score evaluations First, they explore superresolution methods and multiscale sampling, where the diffusion model generates coarse-grained configurations on a smaller lattice which are then super-resolved to the original lattice Second, they introduce Stein diagnostics based on Stein identities to assess the quality of the learned score without relying on reverse diffusion sampling Technical Challenges in Spherical Geometry Additional technical challenges arise because the O(3) model's fields take values on the nonlinear manifold S2 Diffusion models can be naturally extended to Riemannian manifolds by replacing the Gaussian transition kernel with the heat kernel associated with that manifold The short-time asymptotic expansion of the spherical heat kernel becomes singular near the antipodal point, necessitating a separate boundary-layer expansion to obtain an accurate approximation in this region Diagnostic Tools and Training Loss Stein diagnostics are introduced to probe the accuracy of the learned score in reproducing two-point correlations The Stein residual, defined by the identity ⟨D

O; sp: ⟩ p = 0, measures the rate at which Langevin correction changes the expectation of an observable This residual is incorporated into an additional training loss, LStein = Z 1/0 du ⟨D

O; sθ(u): ⟩ p(u) 2, to penalize errors in the learned score that affect particular observables Superresolution and Multiscale Sampling The computational cost can be reduced by sampling on smaller lattices, motivating a multiscale approach where configurations are generated using inexpensive coarse-lattice diffusion models and super-resolved to finer lattices Several super-resolution schemes are investigated, such as Method B2, which gives the best overall agreement with the reference correlations at comparable computational cost Equilibration and Refinement Diffusion-generated configurations can accelerate equilibration under heat-bath Monte Carlo through restricted refinement using the learned score on finer lattices The results demonstrate that multiscale diffusion sampling can improve accuracy without increasing computational cost Discussion and Future Directions The paper demonstrates that diffusion models can reproduce the equilibrium distribution of the O(3) model with reasonable accuracy, and their practical use requires confronting the computational cost of evaluating the learned score Future directions include improving numerical integration and interpolation procedures to reduce fine-lattice refinement requirements A further direction is combining multiscale sampling with diffusion processes that introduce spatial structure progressively across different length scales, such as the SKILD method The code associated with this work is publicly available at https://github.com/daschubring/o3-cp1-diffusion Short-time Heat Kernel Asymptotics The heat kernel on S2 can be approximated using the Minakshisundaram-Pleijel-Schwinger-DeWitt (MPSD) expansion for small diffusion times, which is singular at theta = pi and requires a boundary-layer expansion near the antipodal point The MPSD form of the heat kernel on S2 is given by F(θ, τ) ≈ 1 + 1 - θ cot θ 4θ squared τ + O τ squared The boundary-layer expansion provides an accurate approximation in the region where π − θ ∼ O(τ) Score Parameterization The learned score must be tangent to S2, and this is enforced by having the neural network output coefficients in a basis of tangent vector fields, such as the overcomplete set of three rotational generators eR a = ˆra×n An alternative parameterization involves lifting the data from n ∈ S2 to z ∈ S3, which is consistent with the construction underlying the CP1 model The neural net outputs coefficients of the left-invariant fields, giving score sθ = ω 1 θeL 1 + ω 2 θeL 2 Equilibrium with Heat-Bath Monte Carlo Configurations generated using DM16 can accelerate equilibration under heat-bath Monte Carlo through restricted refinement, showing that approximately 40 restricted sweeps reproduced the original correlations within statistical uncertainty Diffusion-generated configurations begin closer to equilibrium than random initial conditions, although accounting for initialization cost makes their overall computational performance roughly comparable to conventional heat-bath sampling The comparison of B2 with B1, which closely resembles SDEdit, may illustrate the importance of Langevin correction with fixed coarse spins at the intermediate lattice Topological Observables The total topological charge Q(a) is defined by summing local topological charges q i(a) over all plaquettes i The distribution of the topological charge Q and the number of dislocations are shown to exhibit systematic bias in direct samples, which can be corrected after a few heat bath sweeps Stein diagnostics allow for the investigation of these observables independently of reverse diffusion sampling Two-Point Correlation Function Diagnostics The Stein diagnostic for the correlation function CStein(j) ≡ 1/4 ⟨sθ,i · ni+j + ni · sθ,i+j⟩ is used to estimate the correlation function calculated directly from the model score A positive residual (CStein(j) − C(j)) indicates that the learned score would instantaneously increase the value of 13 The Stein residual measures the rate at which Langevin correction changes the expectation of an observable Training with Stein Loss The Stein residual can be used to construct an additional training loss, LStein = Z 1/0 du ⟨D

O; sθ(u): ⟩ p(u) 2, which is added to the usual denoising score matching loss This approach aims to penalize errors in the learned score that affect particular observables, and it has been shown to produce more accurate Stein diagnostics than control models The final loss is L = LDSM + λLStein Computational Cost Analysis The computational cost can be reduced by sampling on smaller lattices, as a score evaluation of DM16 is only about twice as expensive as a heat-bath sweep on the 642 lattice The diffusion model is used to generate coarsegrained configurations on a smaller, cheaper lattice and then super-resolved to the original lattice The results demonstrate that multiscale diffusion sampling can improve accuracy without increasing computational cost Diffusion models can reproduce the equilibrium distribution of the O(3) model with reasonable accuracy, and superresolution schemes can simultaneously improve the accuracy of sampled correlation functions and modestly reduce computational cost compared to direct diffusion sampling on the finest lattice The paper concludes that multiscale diffusion sampling can improve accuracy without increasing computational cost Figure 7 Comparison Figure 7 compares the nearest-neighbor correlation C(1), long-distance correlation C(32), magnetic susceptibility χm, and topological susceptibility χt for hot, cold, and diffusion-generated configurations on the 642 lattice Diffusion-generated configurations are obtained using DM16 with RK2 or mixed SDE sampling, followed by prolongation and restricted heat-bath refinement The nearest-neighbor correlation C(1) and topological susceptibility χt approach equilibrium relatively quickly for all initial conditions The long-distance correlation C(32) and magnetic susceptibility χm require substantially more heat-bath sweeps Figure 4 Comparison Figure 4 compares approximations to the scalar score ∂z log K(θ, τ), showing the spectral expansion with lmax = 10 at large τ and MPSD and boundary-layer approximations at small τ The switching time was chosen by requiring the heat-kernel-weighted relative mean squared score error to fall below a prescribed threshold The boundary-layer expansion is developed to handle difficulties near θ = π Figure 2 Comparison Figure 2 shows the distributions of the topological charge Q and the number of lattice dislocations for the two-dimensional O(3) model at β = 1.5 on a 642 lattice, comparing equilibrium heat-bath reference with diffusion model output Diffusion-generated configurations exhibit a systematic bias in the number of dislocations and a slightly reduced probability near Q = 0 Table 1 Comparison Table 1 compares observables on the coarse-grained 162 lattice with samples from both the diffusion model and O(3) models at effective couplings β′ tuned to reproduce either the correlation length or C(1) of the coarse-grained reference ensemble The diffusion model reproduces both short and long-distance correlations in these comparisons Table 4 Comparison Table 4 compares the resulting correlation functions with those obtained from the reference ensemble, direct diffusion model sampling, and various super-resolution methods For DM32, both super-resolution methods produce correlations substantially closer to equilibrium than direct diffusion sampling at slightly lower computational cost Method B2 gives the best overall agreement with the reference correlations, improving upon direct sampling at comparable computational cost The comparison of B2 with B1, which closely resembles SDEdit, may illustrate the importance of Langevin correction with fixed coarse spins at the intermediate 322 lattice Figure 5 Comparison Figure 5 shows the two-point correlation function C(r) on the 642 lattice, where solid lines show equilibrium correlations at diffusion times u = 0, 0.5, and 0.75 The Stein diagnostic calculated with a well-trained model closely tracks the equilibrium correlation function over a wide range of diffusion times including u = 0 Figure 3 Visualization Figure 3 illustrates both our method for visualizing lattice spin configurations and an application of the multiscale methods discussed later in Sec.

Kai: 5 The resulting image has a value of C(1) and density of lattice dislocations close to their equilibrium values, as reported in the figure caption Figure 7 Equilibration Dynamics Figure 7 shows that diffusion-generated configurations are obtained using DM16 with RK2 or mixed SDE sampling, followed by prolongation and restricted heat-bath refinement The nearest-neighbor correlation C(1) and topological susceptibility χt approach equilibrium relatively quickly for all initial conditions The long-distance correlation C(32) and magnetic susceptibility χm require substantially more heat-bath sweeps Diffusion-generated ensembles begin closer to equilibrium, accounting for the initialization cost makes their overall computational performance roughly comparable to conventional heat-bath sampling Final Conclusion The paper demonstrates that score-based diffusion models can reproduce the equilibrium distribution of the O(3) model with reasonable accuracy The authors conclude that multiscale diffusion sampling can improve accuracy without increasing computational cost There remains considerable room for improving these methods, particularly in reducing the amount of fine-lattice refinement required The code associated with this work is publicly available at https://github.com/daschubring/o3-cp1-diffusion Short-time Heat Kernel Asymptotics The heat kernel on S2 can be approximated using the MPSD expansion for small diffusion times, which is singular at theta = pi and requires a boundary-layer expansion near the antipodal point The heat kernel on SO(3) in Camporesi’s normalization is given by KSO(3)(ϕ, τ) = e τ /4 2(4πτ) 3/2 X n∈Z (−1)n ϕ + 2πn sin(ϕ/2) exp − (ϕ + 2πn) 2 /4τ The heat kernel on S squared is obtained by integrating the SO(3) heat kernel over the isotropy subgroup SO(2) The MPSD expansion for the heat kernel on S squared is given by F(θ, τ) ≈ 1 + 1 - θ cot θ 4θ squared τ + O τ squared The boundary-layer expansion provides an accurate approximation in the region where π − θ ∼ O(τ) The heat kernel on SO(3) is obtained by integrating the SO(3) heat kernel over the isotropy subgroup SO(2) Invariant Vector Fields and Tangent Score Representations The neural net generates the score sθ from the pushforward of a right-invariant vector field, sθ = X a ω a θ eR a = ωθ × n Alternatively, one might generate the score from a left-invariant vector field on SO(3), eL a ≡ d/dtπ(Retomegaa) t=0 = R(ˆra × ˆr3) The neural net outputs coefficients of the left-invariant fields, giving score sθ = ω 1 θeL 1 + ω 2 θeL 2 To implement this parameterization, we use random representatives of the CP1 field z, or equivalently random quaternions q ∈ S3 as inputs to the neural network The score sθ is constructed from the pushforward of a right-invariant vector field, sθ = X a ω a θ eR a = ωθ × n The component ω 3 θ would be redundant, so the neural net need only generate two components The choice of tangent frame introduces a U(1) redundancy in the input Symmetry and Redundancy The double covering S3 ≃ SU(2) → SO(3) is parametrized by unit quaternions q = (q0, q1, q2, q3) The projection π: SO(3) → S2 is defined by π(R) = Rˆr3 ≡ n, with ˆra the standard Cartesian basis This is consistent with the definition of unit complex vector z in the CP1 model, z = (q0 + iq1, q2 + iq3) T, n = z†σz The pushforward of the right-invariant vector field is eR a ≡ d/dtπ(eTomegaaR) t=0 = ˆra × n Code Availability The code associated with this work, including trained diffusion models and Jupyter notebooks illustrating the numerical methods, is publicly available at https://github.com/daschubring/o3-cp1-diffusion The code associated with this work, including trained diffusion models and Jupyter notebooks illustrating the numerical methods, is publicly available at https://github.

Title and authors: Kai: Moving on to the title and authors of this paper, "Diffusion models for the O(three) nonlinear sigma model" <ref:2610.39568#pg1,Diffusion models for the O(3) nonlinear sigma model>. The core idea here is applying diffusion models to a specific physical system—the two dimensional classical Heisenberg model.

Mira: It’s interesting because they aren't just using it as a toy problem; they are using it as a testing ground for methods developed to study nonperturbative gauge theories, specifically things like the four dimensional lattice gauge theories.

Lev: That connection to Yang-Mills theory is important because those 4D theories have much more complex topological features and confinement, which is where the real difficulty lies <ref:2610.39568#pg2>.

Kai: So the paper establishes this O three model as a useful benchmark for assessing diffusion-based sampling methods before trying to apply them to those more difficult four dimensional systems.

Mira: They specifically mention that the O three model shares qualitative features with four dimensional Yang-Mills theory, like asymptotic freedom and the emergence of a mass gap, which makes it a good comparison point.

Lev: That’s helpful because those features tell us what kind of physical behavior we are trying to capture in these lattice field theories.

Kai: The authors also mention that the O three model can be equivalently expressed in terms of an S three valued field with gauge redundancy, which is what they call the CP one model.

Mira: And they point out that this CP one interpretation admits a large N generalization, called the CP N minus one model.

Lev: The authors mention that conventional Monte Carlo simulations struggle with topological freezing in those higher CP N minus one models, so using the O three case is valuable because it has efficient MCMC methods.

Kai: So they are setting up this paper as a bridge: starting simple, testing the diffusion model on O three, and then showing how those results might apply to the more complex CP N minus one families.

The paper's summary: Mira: To summarize what the authors present in "Diffusion models for the O(three) nonlinear sigma model", they are proposing a score-based diffusion model trained on the two dimensional classical Heisenberg model <ref:2610.39568#pg1,Diffusion models for the O(3) nonlinear sigma model>.

Kai: This model is then used to generate configurations, and they investigate how to use that generation process to improve both accuracy and computational cost when sampling these lattice field theories.

Lev: They are specifically looking at two methodological questions driven by the high cost of evaluating those scores, focusing on superresolution and multiscale sampling.

Mira: And they also introduce Stein diagnostics, which they use to check the quality of the learned score using identities related to Langevin corrections, adding that as an additional training loss.

Lev: So, in a nutshell, it’s about improving how we generate configurations and how we validate those generated scores without having to run expensive reverse diffusion sampling for every single observable.

Kai: The paper shows they can produce the equilibrium distribution of the O three model with reasonable accuracy even with these modifications.

Mira: That accuracy is what makes it useful; they show that multiscale diffusion sampling can improve accuracy without increasing computational cost on the finest lattice, which is a key finding for them.

Lev: But we have to remember that this improvement depends heavily on the specific superresolution scheme you choose, like Method B2 which they found gave the best overall agreement with reference correlations at comparable computational cost.

Kai: So it’s not just about diffusion; it’s about the whole pipeline—from generation to refinement to validation.

The paper's improvements: Lev: Let's talk about the specific improvements they suggest in "Diffusion models for the O(three) nonlinear sigma model" <ref:2610.39568#pg1,Diffusion models for the O(3) nonlinear sigma model>. They focus heavily on how to make the diffusion process more efficient.

Mira: They explore superresolution methods and multiscale sampling, as mentioned before, where you generate coarse-grained configurations on a smaller lattice and then superresolve them back up.

Kai: That’s a way to reduce the cost of generation because the coarse-grained part is cheaper, which is exactly what they need for practical application.

Lev: And then they talk about equilibration under heat bath Monte Carlo, where they use restricted refinement using the learned score on finer lattices to speed up how fast configurations settle down.

Mira: They found that this accelerated equilibration helps because diffusion-generated ensembles start closer to equilibrium, though accounting for the initialization cost makes their overall performance comparable to conventional heat bath sampling.

Lev: That means you get a head start in terms of statistical quality, even if the setup overhead is still there.

Kai: And they introduce Stein diagnostics and the related training loss, which allows them to penalize errors in the learned score that affect specific observables directly.

Mira: This targeted loss term is designed to improve things like short-distance correlations, specifically C one, by penalizing errors in the score that mess up those measurements.

Lev: So they’re not just relying on a single loss function anymore; they’s using these tools to get more specific control over the model's behavior.

Kai: It sounds like a complete package: better generation, better validation, and targeted training losses.

Conclusion: Mira: So wrapping up with "Diffusion models for the O(three) nonlinear sigma model", the main conclusion is that score-based diffusion models can reproduce the equilibrium distribution of the O three model with reasonable accuracy <ref:2610.39568#pg1,Diffusion models for the O(3) nonlinear sigma model>.

Kai: And they conclude that multiscale diffusion sampling offers a way to improve accuracy without increasing computational cost compared to direct diffusion sampling on the finest lattice.

Lev: It’s a practical finding, but they also flag that there's still considerable room for improvement in reducing the amount of fine-lattice refinement you need.

Mira: They mention that their code is available publicly, which is helpful for verifying these methods in practice.

Kai: Overall, this work shows that diffusion models are a viable tool for studying these kinds of systems, and they’re moving toward using them to sample more complex lattice field theories.

Lev: For me, the main thing is that the efficiency gains are conditional on the specific superresolution scheme you pick.

Mira: And we'll keep an eye on how these techniques evolve for tackling those higher CP N minus one models next.

City College of New York, CUNY

cond-mat.stat-mech, cs.LG, hep-lat

Submitted: 2026-10-29

Updated: 2026-11-09

Comments: 29 pages, 7 figures

Code: https://github.com/daschubring/o3-cp1-diffusion

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: A score-based diffusion model is trained on the two-dimensional classical Heisenberg model to investigate tradeoffs between computational cost and accuracy in generating configurations for lattice…

Key concepts

Diffusion Model
A generative model trained to reverse a diffusion process, which starts from noise and gradually transforms it into a desired configuration (like a physical state of the O(3) model). It learns how to map random noise back to meaningful physical configurations.
Multiscale Sampling
A technique where configurations are first generated using inexpensive diffusion models on a coarse lattice, and then these coarse states are 'super-resolved' onto the original, finer lattice. This reduces computational cost while aiming to maintain high accuracy in the final configuration.
Stein Diagnostics
Mathematical tools based on Stein identities used to assess how well a learned score function reproduces physical correlations. They provide an independent way to check if the model's learned score is accurate without needing reverse diffusion sampling, helping pinpoint specific errors in the model's learning.
Heat Kernel Expansion
A mathematical tool used to approximate the probability distribution of configurations on curved spaces like S2 (the manifold where O(3) fields live). It involves an asymptotic expansion that needs careful handling near singular points, requiring boundary-layer approximations for accurate results.

Terminology

Summary

A score-based diffusion model is trained on the two-dimensional classical Heisenberg model to investigate tradeoffs between computational cost and accuracy in generating configurations for lattice field theories

Model Application and Test Case

The diffusion model is trained on the two-dimensional classical Heisenberg model, which serves as a test case for evaluating diffusion models on lattice field theories The O(3) model shares qualitative features with four-dimensional Yang–Mills theory, including asymptotic freedom, emergence of a mass gap, and topological instantons This model is useful as a benchmark for assessing diffusion-based sampling methods with potential applications to the higher CP N−1 family and the qualitatively related 4D lattice gauge theories

Methodological Investigations

The authors investigate two complementary methodological questions motivated by the computational cost of score evaluations First, they explore superresolution methods and multiscale sampling, where the diffusion model generates coarse-grained configurations on a smaller lattice which are then super-resolved to the original lattice Second, they introduce Stein diagnostics based on Stein identities to assess the quality of the learned score without relying on reverse diffusion sampling

Technical Challenges in Spherical Geometry

Additional technical challenges arise because the O(3) model's fields take values on the nonlinear manifold S2 Diffusion models can be naturally extended to Riemannian manifolds by replacing the Gaussian transition kernel with the heat kernel associated with that manifold The short-time asymptotic expansion of the spherical heat kernel becomes singular near the antipodal point, necessitating a separate boundary-layer expansion to obtain an accurate approximation in this region

Diagnostic Tools and Training Loss

Stein diagnostics are introduced to probe the accuracy of the learned score in reproducing two-point correlations The Stein residual, defined by the identity ⟨D [O; sp]⟩ p = 0, measures the rate at which Langevin correction changes the expectation of an observable This residual is incorporated into an additional training loss, LStein = Z 1/0 du ⟨D [O; sθ(u)]⟩ p(u) 2, to penalize errors in the learned score that affect particular observables

Superresolution and Multiscale Sampling

The computational cost can be reduced by sampling on smaller lattices, motivating a multiscale approach where configurations are generated using inexpensive coarse-lattice diffusion models and super-resolved to finer lattices Several super-resolution schemes are investigated, such as Method B2, which gives the best overall agreement with the reference correlations at comparable computational cost

Equilibration and Refinement

Diffusion-generated configurations can accelerate equilibration under heat-bath Monte Carlo through restricted refinement using the learned score on finer lattices The results demonstrate that multiscale diffusion sampling can improve accuracy without increasing computational cost

Discussion and Future Directions

The paper demonstrates that diffusion models can reproduce the equilibrium distribution of the O(3) model with reasonable accuracy, and their practical use requires confronting the computational cost of evaluating the learned score Future directions include improving numerical integration and interpolation procedures to reduce fine-lattice refinement requirements A further direction is combining multiscale sampling with diffusion processes that introduce spatial structure progressively across different length scales, such as the SKILD method The code associated with this work is publicly available at https://github.com/daschubring/o3-cp1-diffusion

Short-time Heat Kernel Asymptotics

The heat kernel on S2 can be approximated using the Minakshisundaram-Pleijel-Schwinger-DeWitt (MPSD) expansion for small diffusion times, which is singular at theta = pi and requires a boundary-layer expansion near the antipodal point The MPSD form of the heat kernel on S2 is given by F(θ, τ) ≈ 1 + 1 - θ cot θ 4θ squared τ + O τ squared The boundary-layer expansion provides an accurate approximation in the region where π − θ ∼ O(τ)

Score Parameterization

The learned score must be tangent to S2, and this is enforced by having the neural network output coefficients in a basis of tangent vector fields, such as the overcomplete set of three rotational generators eR a = ˆra×n An alternative parameterization involves lifting the data from n ∈ S2 to z ∈ S3, which is consistent with the construction underlying the CP1 model The neural net outputs coefficients of the left-invariant fields, giving score sθ = ω 1 θeL 1 + ω 2 θeL 2

Equilibrium with Heat-Bath Monte Carlo

Configurations generated using DM16 can accelerate equilibration under heat-bath Monte Carlo through restricted refinement, showing that approximately 40 restricted sweeps reproduced the original correlations within statistical uncertainty Diffusion-generated configurations begin closer to equilibrium than random initial conditions, although accounting for initialization cost makes their overall computational performance roughly comparable to conventional heat-bath sampling The comparison of B2 with B1, which closely resembles SDEdit, may illustrate the importance of Langevin correction with fixed coarse spins at the intermediate lattice

Topological Observables

The total topological charge Q(a) is defined by summing local topological charges q i(a) over all plaquettes i The distribution of the topological charge Q and the number of dislocations are shown to exhibit systematic bias in direct samples, which can be corrected after a few heat bath sweeps Stein diagnostics allow for the investigation of these observables independently of reverse diffusion sampling

Two-Point Correlation Function Diagnostics

The Stein diagnostic for the correlation function CStein(j) ≡ 1/4 ⟨sθ,i · ni+j + ni · sθ,i+j⟩ is used to estimate the correlation function calculated directly from the model score A positive residual (CStein(j) − C(j)) indicates that the learned score would instantaneously increase the value of 13 The Stein residual measures the rate at which Langevin correction changes the expectation of an observable

Training with Stein Loss

The Stein residual can be used to construct an additional training loss, LStein = Z 1/0 du ⟨D [O; sθ(u)]⟩ p(u) 2, which is added to the usual denoising score matching loss This approach aims to penalize errors in the learned score that affect particular observables, and it has been shown to produce more accurate Stein diagnostics than control models The final loss is L = LDSM + λLStein

Computational Cost Analysis

The computational cost can be reduced by sampling on smaller lattices, as a score evaluation of DM16 is only about twice as expensive as a heat-bath sweep on the 642 lattice The diffusion model is used to generate coarsegrained configurations on a smaller, cheaper lattice and then super-resolved to the original lattice The results demonstrate that multiscale diffusion sampling can improve accuracy without increasing computational cost

Summary of Findings

Diffusion models can reproduce the equilibrium distribution of the O(3) model with reasonable accuracy, and superresolution schemes can simultaneously improve the accuracy of sampled correlation functions and modestly reduce computational cost compared to direct diffusion sampling on the finest lattice The paper concludes that multiscale diffusion sampling can improve accuracy without increasing computational cost

Figure 7 Comparison

Figure 7 compares the nearest-neighbor correlation C(1), long-distance correlation C(32), magnetic susceptibility χm, and topological susceptibility χt for hot, cold, and diffusion-generated configurations on the 642 lattice Diffusion-generated configurations are obtained using DM16 with RK2 or mixed SDE sampling, followed by prolongation and restricted heat-bath refinement The nearest-neighbor correlation C(1) and topological susceptibility χt approach equilibrium relatively quickly for all initial conditions The long-distance correlation C(32) and magnetic susceptibility χm require substantially more heat-bath sweeps

Figure 4 Comparison

Figure 4 compares approximations to the scalar score ∂z log K(θ, τ), showing the spectral expansion with lmax = 10 at large τ and MPSD and boundary-layer approximations at small τ The switching time was chosen by requiring the heat-kernel-weighted relative mean squared score error to fall below a prescribed threshold The boundary-layer expansion is developed to handle difficulties near θ = π

Figure 2 Comparison

Figure 2 shows the distributions of the topological charge Q and the number of lattice dislocations for the two-dimensional O(3) model at β = 1.5 on a 642 lattice, comparing equilibrium heat-bath reference with diffusion model output Diffusion-generated configurations exhibit a systematic bias in the number of dislocations and a slightly reduced probability near Q = 0

Table 1 Comparison

Table 1 compares observables on the coarse-grained 162 lattice with samples from both the diffusion model and O(3) models at effective couplings β′ tuned to reproduce either the correlation length or C(1) of the coarse-grained reference ensemble The diffusion model reproduces both short and long-distance correlations in these comparisons

Table 4 Comparison

Table 4 compares the resulting correlation functions with those obtained from the reference ensemble, direct diffusion model sampling, and various super-resolution methods For DM32, both super-resolution methods produce correlations substantially closer to equilibrium than direct diffusion sampling at slightly lower computational cost Method B2 gives the best overall agreement with the reference correlations, improving upon direct sampling at comparable computational cost The comparison of B2 with B1, which closely resembles SDEdit, may illustrate the importance of Langevin correction with fixed coarse spins at the intermediate 322 lattice

Figure 5 Comparison

Figure 5 shows the two-point correlation function C(r) on the 642 lattice, where solid lines show equilibrium correlations at diffusion times u = 0, 0.5, and 0.75 The Stein diagnostic calculated with a well-trained model closely tracks the equilibrium correlation function over a wide range of diffusion times including u = 0

Figure 3 Visualization

Figure 3 illustrates both our method for visualizing lattice spin configurations and an application of the multiscale methods discussed later in Sec. 5 The resulting image has a value of C(1) and density of lattice dislocations close to their equilibrium values, as reported in the figure caption

Figure 7 Equilibration Dynamics

Figure 7 shows that diffusion-generated configurations are obtained using DM16 with RK2 or mixed SDE sampling, followed by prolongation and restricted heat-bath refinement The nearest-neighbor correlation C(1) and topological susceptibility χt approach equilibrium relatively quickly for all initial conditions The long-distance correlation C(32) and magnetic susceptibility χm require substantially more heat-bath sweeps Diffusion-generated ensembles begin closer to equilibrium, accounting for the initialization cost makes their overall computational performance roughly comparable to conventional heat-bath sampling

Final Conclusion

The paper demonstrates that score-based diffusion models can reproduce the equilibrium distribution of the O(3) model with reasonable accuracy The authors conclude that multiscale diffusion sampling can improve accuracy without increasing computational cost There remains considerable room for improving these methods, particularly in reducing the amount of fine-lattice refinement required The code associated with this work is publicly available at https://github.com/daschubring/o3-cp1-diffusion

The heat kernel on S2 can be approximated using the MPSD expansion for small diffusion times, which is singular at theta = pi and requires a boundary-layer expansion near the antipodal point The heat kernel on SO(3) in Camporesi’s normalization is given by KSO(3)(ϕ, τ) = e τ /4 2(4πτ) 3/2 X n∈Z (−1)n ϕ + 2πn sin(ϕ/2) exp − (ϕ + 2πn) 2 /4τ The heat kernel on S squared is obtained by integrating the SO(3) heat kernel over the isotropy subgroup SO(2) The MPSD expansion for the heat kernel on S squared is given by F(θ, τ) ≈ 1 + 1 - θ cot θ 4θ squared τ + O τ squared The boundary-layer expansion provides an accurate approximation in the region where π − θ ∼ O(τ) The heat kernel on SO(3) is obtained by integrating the SO(3) heat kernel over the isotropy subgroup SO(2)

Invariant Vector Fields and Tangent Score Representations

The neural net generates the score sθ from the pushforward of a right-invariant vector field, sθ = X a ω a θ eR a = ωθ × n Alternatively, one might generate the score from a left-invariant vector field on SO(3), eL a ≡ d/dtπ(Retomegaa) t=0 = R(ˆra × ˆr3) The neural net outputs coefficients of the left-invariant fields, giving score sθ = ω 1 θeL 1 + ω 2 θeL 2 To implement this parameterization, we use random representatives of the CP1 field z, or equivalently random quaternions q ∈ S3 as inputs to the neural network The score sθ is constructed from the pushforward of a right-invariant vector field, sθ = X a ω a θ eR a = ωθ × n The component ω 3 θ would be redundant, so the neural net need only generate two components The choice of tangent frame introduces a U(1) redundancy in the input

Symmetry and Redundancy

The double covering S3 ≃ SU(2) → SO(3) is parametrized by unit quaternions q = (q0, q1, q2, q3) The projection π: SO(3) → S2 is defined by π(R) = Rˆr3 ≡ n, with ˆra the standard Cartesian basis This is consistent with the definition of unit complex vector z in the CP1 model, z = (q0 + iq1, q2 + iq3) T, n = z†σz The pushforward of the right-invariant vector field is eR a ≡ d/dtπ(eTomegaaR) t=0 = ˆra × n

Code Availability

The code associated with this work, including trained diffusion models and Jupyter notebooks illustrating the numerical methods, is publicly available at https://github.com/daschubring/o3-cp1-diffusion The code associated with this work, including trained diffusion models and Jupyter notebooks illustrating the numerical methods, is publicly available at https://github. p. 1

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. Multiscale sampling for diffusion models can be used to generate coarse-grained configurations on smaller lattices which are then super-resolved to larger lattices, improving accuracy without increasing computational cost.

  2. Stein diagnostics based on identities like 0 = Z C d sigma mu g mu nu dnu O log p indicates a discrepancy between the learned and target scores that can be detected independently of the sampling method.

  3. A training loss term LStein = Z 1/0 du d < D [O; s theta(u)] p(u) 2 penalizes errors in the learned score that affect particular observables, allowing for targeted improvements in short-distance correlations like C(1).

  4. The model can be extended to Riemannian manifolds by replacing the Gaussian transition kernel with the heat kernel associated with the manifold, using boundary-layer approximations to handle singularities near antipodal points on S2.

  5. The learned score can be parameterized using an overcomplete set of tangent vector fields, such as e R a = hatra x n, allowing for a network output s theta = X a omega a theta e R a = omega theta x n.

Sources

Related papers