Stochastic Path Sampler For Lattice Field Theory
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: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Stochastic Path Sampler For Lattice Field Theory".
Mira: Here's the gist of the paper. In lattice field theory, target distributions are known only up to normalization, while Markov chain Monte Carlo simulations often become inefficient near phase transitions or the continuum limit due to critical slowing down <ref:2606.13790#pg6> The central idea of the SPS is to establish a trajectory-level balance for learnable forward and backward stochastic dynamics between two equilibrium states—the prior and target distributions <ref:2606.13790#pg9> In two-dimensional phi 4 theory, we demonstrate that our neural sampler can achieve the same sampling quality as HMC but with a much shorter autocorrelation time in the pseudocritical region <ref:2606.13790#pg10> This sampler offers a novel stochastic-quantization-inspired route to data-free proposal construction for lattice field theory by leveraging a variational free-energy principle derived from path-space irreversibility <ref:2606.13790#pg5>.
Kai: Nonequilibrium Thermodynamics on Path Space The Clausius inequality provides a mathematical formulation of the Second Law of Thermodynamics, stating that irreversible processes are accompanied by nonnegative entropy production <ref:2606.13790#pg6> In this work, we adopt its stochastic-thermodynamic counterpart, namely entropy production on path space <ref:2606.13790#pg6> For a stochastic process evolving between microscopic configurations, entropy production measures the asymmetry between the forward trajectory ensemble and its corresponding reversed ensemble <ref:2606.13790#pg6> The stochastic entropy production of the system is given by an expression involving the accumulated affinity and a boundary term <ref:2606.13790#pg6> This demonstrates irreversibility of the nonequilibrium system; the larger Σ(τ), the stronger the forward–backward inconsistency, and hence the more irreversible the process <ref:2606.13790#pg6>.
Mira: How the training signal is built The SPS networks are trained by minimizing the path-space KL divergence between the learned forward path measure and the unnormalized backward path measure <ref:2606.13790#pg6> The loss function is defined as LSPS(θ) = EqF
log qθ(sT; τ) − log ˜π∗(sT): <ref:2606.13790#pg6> Since the partition function of the target distribution is not required in Eq.
Kai: (B.18), the SPS can be trained without reference samples from HMC or any other Markov chain sampler <ref:2606.13790#pg6>.
Paper summary: Kai: So we're looking at this paper called "Stochastic Path Sampler For Lattice Field Theory". Basically, they're tackling the problem that in lattice field theory, you usually only know the target distribution up to a normalization constant and the partition function is too hard to calculate directly.
Mira: That’s right. And Markov chain Monte Carlo simulations often get really slow near phase transitions or when you're trying to reach the continuum limit because of this critical slowing down <ref:2606.13790#pg6>. What this paper proposes is a new way to sample that target distribution without needing reference samples from methods like Hybrid Monte Carlo.
Lev: From an engineering standpoint, that's huge because HMC can take forever when you're near those critical points or trying to get very fine details of the physics. We’re talking about making the sampling much faster in those tricky regions <ref:2606.13790#pg6>.
Kai: The core idea they use is based on nonequilibrium thermodynamics, which is a bit outside our usual quantum hardware focus, but it sets up this path-space balance between the prior and the target distribution <ref:2606.13790#pg9>. They claim their Stochastic Path Sampler SPS can achieve sampling quality comparable to HMC in two dimensional phi four theory, but with a much shorter autocorrelation time in the pseudocritical region <ref:2606.13790#pg10>.
Mira: That’s interesting because they are using a variational free-energy principle derived from path-space irreversibility to build their proposal construction, which is a novel route for data free proposal construction in lattice field theory <ref:2606.13790#pg5>. It’s leveraging the fact that entropy production on path space measures the asymmetry between forward and reversed ensembles <ref:2606.13790#pg6>.
Lev: If they can control that path-space KL divergence, which is what they minimize when training the SPS networks, then they are essentially reducing irreversibility in a way that improves sampling efficiency <ref:2606.13790#pg6>. But we need to be careful about how much of that reduction actually translates to usable speed on real hardware <ref:2606.13790#pg8>.
Kai: The results show that for SPS plus an Independence Metropolis–Hastings correction, the integrated autocorrelation time for the absolute magnetization drops from around one hundred sixty in HMC down to about zero point five with this new method <ref:2606.13790#pg6>. That’s a substantial reduction in how much you need to sample to get reliable statistics <ref:2606.13790#pg6>.
Mira: And they also looked at the effective sample size, which is defined as the normalized effective sample size being controlled by the same quantity, sigma squared, which is the variance of the log weights <ref:2606.13790#pg6>. This connects how well you're sampling to a fundamental measure of weight distribution <ref:2606.13790#pg6>.
Lev: But they also found that while the normalized effective sample size shows a deficit that grows linearly with the lattice volume, the acceptance rate deficit in their correction step only grows as the square root of the volume <ref:2606.13790#pg6>. That explains why you can still get decent results even when you increase your lattice size <ref:2606.13790#pg8>.
Kai: So, to recap this paper, they proposed the Stochastic Path Sampler for Lattice Field Theory by using a neural network method based on nonequilibrium thermodynamics to learn the forward and backward paths of the process <ref:2606.13790#pg6>. They train these networks by minimizing the KL divergence in path space, and then they add an Independence Metropolis–Hastings correction step to ensure accuracy <ref:2606.13790#pg8>.
Mira: The main point is that this SPS with the IMH correction reproduces all physical quantities computed by Hybrid Monte Carlo with high accuracy across different phases, including the symmetric phase, the pseudocritical region, and the symmetry broken phase <ref:2606.13790#pg6>.
Lev: It’s a strong claim because it says this method works where traditional methods struggle most—near critical points—and it shows that for real hardware running on larger volumes, the acceptance rate stays relatively high, even with that volume dependence <ref:2606.13790#pg8>.
Kai: The paper also touched on Nelson’s duality relation in the continuous limit, where they show how the learned path-space balance relates to a local relation between forward and backward drifts <ref:2606.13790#pg6>. That’s a nice theoretical connection that isn't immediately obvious <ref:two thousand six hundred six point one three seven nine zero#pg4.
Mira: It suggests the network isn't just learning paths in isolation, but it's learning the integrated path-space balance rather than forcing a pointwise duality between the two drifts <ref:2606.13790#pg6>.
Lev: For someone building error correction hardware, this means we might be able to use these learned dynamics as a guide for constructing more efficient simulation steps, rather than relying purely on the traditional HMC integration scheme <ref:two thousand six hundred six point one three seven nine zero#pg2.
Kai: So the title "Stochastic Path Sampler For Lattice Field Theory" points to a method that uses path integrals and thermodynamics to solve a problem about sampling distributions where you don't know the target distribution well enough to sample directly <ref:two thousand six hundred six point one three seven nine zero#pg1.
Mira: And the implication is that we can get better, faster results from lattice field theory simulations by using these learned dynamics instead of relying solely on existing Markov chain methods <ref:two thousand six hundred six point one three seven nine zero#pg3.
Lev: It’s a practical idea because it's about making the computational cost manageable when you are dealing with complex many-body systems where you can't just take shortcuts <ref:two thousand six hundred six point one three seven nine zero#pg2.
Kai: So, we're seeing a method that uses path space entropy production to guide the training of neural networks for sampling distributions in lattice field theory <ref:two thousand six hundred six point one three seven nine zero#pg5.
Mira: And the main takeaway is that this approach, when combined with the Independence Metropolis–Hastings correction, gives us very accurate results across different physical regimes <ref:two thousand six hundred six point one three seven nine zero#pg6.
Lev: It shows that we can tackle critical slowing down in these simulations by learning a trajectory-level balance between two equilibrium states, which is what the Stochastic Path Sampler is all about <ref:two thousand six hundred six point one three seven nine zero#pg9.
Conclusion: Kai: So we're wrapping up this look at the Stochastic Path Sampler for Lattice Field Theory by Kai and Mira today.
Mira: They used nonequilibrium thermodynamics to build a way to sample those distributions, but they’re focusing on how that path space balance translates into actual sampling efficiency.
Lev: The core idea is learning the forward and backward paths of the process separately, then minimizing a certain divergence between them, which they call the path-space KL divergence.
Kai: It seems like their main achievement is showing that this method can match Hybrid Monte Carlo quality in two dimensional phi four theory, but with a much faster autocorrelation time when you're near phase transitions.
Mira: They’re using an independence metropolis hastings correction step to make sure the results are accurate, and they found that this method works across different phases, including the symmetry broken one.
Lev: The numbers are interesting; their integrated autocorrelation time for absolute magnetization drops from one hundred sixty in HMC down to about zero point five with their SPS plus correction.
Kai: So it’s not just a theoretical trick; they show a tangible speedup in how quickly you get good statistics on the simulation.
Mira: It means that instead of relying solely on traditional Markov chain methods, we might be able to use these learned dynamics to make lattice field theory simulations much more efficient.
Lev: This is important for real hardware because it suggests a path toward constructing more efficient simulation steps based on learning those dynamics rather than just brute force integration.
Kai: And they did look at Nelson’s duality relation in the continuous limit, which connects the learned path-space balance to a local relation between forward and backward drifts.
Mira: That connection is interesting because it shows the network isn't just learning paths; it’s learning a sort of integrated path-space balance.
Lev: It suggests we can use this kind of learned dynamics as a guide for error correction hardware, aiming for better simulation steps.
Centre for Quantum Fields and Gravity, Swansea University · School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen (CUHKShenzhen) · Johannes Gutenberg University Mainz · School of Mathematical Sciences, Queen Mary University of London
hep-lat
Submitted: 2026-06-11
Updated: 2026-11-11
Comments: 13 figures, 39 pages
Code: https://github.com/Chern2/Stochastic-Path-Sampler-For-Lattice-F
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 89/100
The gist: In lattice field theory, target distributions are known only up to normalization, while Markov chain Monte Carlo simulations often become inefficient near phase transitions or the continuum limit due…
Key concepts
- Stochastic Path Sampler (SPS)
- A neural network method that learns the discretized Langevin dynamics paths for both the forward and backward processes in lattice field theory. It trains by minimizing path-space KL divergence to establish a trajectory-level balance between prior and target distributions.
- Entropy Production on Path Space
- This concept uses the Clausius inequality from thermodynamics to measure irreversibility in stochastic processes. It quantifies the asymmetry between the forward trajectory ensemble and its reversed ensemble, showing how much the process moves away from equilibrium.
- Autocorrelation Time
- A measure of how long a Markov chain takes to explore its state space effectively. In this work, SPS+IMH showed a significantly lower integrated autocorrelation time (around 0.5) compared to HMC (around 160), demonstrating improved sampling efficiency, especially near critical regions.
Terminology
Summary
In lattice field theory, target distributions are known only up to normalization, while Markov chain Monte Carlo simulations often become inefficient near phase transitions or the continuum limit due to critical slowing down <ref:2606. p. 013790#pg6> The central idea of the SPS is to establish a trajectory-level balance for learnable forward and backward stochastic dynamics between two equilibrium states—the prior and target distributions <ref:2606. p. 013790#pg9> In two-dimensional phi 4 theory, we demonstrate that our neural sampler can achieve the same sampling quality as HMC but with a much shorter autocorrelation time in the pseudocritical region <ref:2606. p. 013790#pg10> This sampler offers a novel stochastic-quantization-inspired route to data-free proposal construction for lattice field theory by leveraging a variational free-energy principle derived from path-space irreversibility <ref:2606. p. 013790#pg5>.
Nonequilibrium Thermodynamics on Path Space
The Clausius inequality provides a mathematical formulation of the Second Law of Thermodynamics, stating that irreversible processes are accompanied by nonnegative entropy production <ref:2606.13790#pg6> In this work, we adopt its stochastic-thermodynamic counterpart, namely entropy production on path space <ref:2606. p. 513790#pg6> For a stochastic process evolving between microscopic configurations, entropy production measures the asymmetry between the forward trajectory ensemble and its corresponding reversed ensemble <ref:2606. p. 513790#pg6> The stochastic entropy production of the system is given by an expression involving the accumulated affinity and a boundary term <ref:2606. p. 513790#pg6> This demonstrates irreversibility of the nonequilibrium system; the larger Σ(τ), the stronger the forward–backward inconsistency, and hence the more irreversible the process <ref:2606. p. 613790#pg6>.
How the training signal is built
The SPS networks are trained by minimizing the path-space KL divergence between the learned forward path measure and the unnormalized backward path measure <ref:2606.13790#pg6> The loss function is defined as LSPS(θ) = EqF[log qθ(sT; τ) − log ˜π∗(sT)] <ref:2606.13790#pg6> Since the partition function of the target distribution is not required in Eq. p. 0 (B.18), the SPS can be trained without reference samples from HMC or any other Markov chain sampler <ref:2606. p. 013790#pg6>. The expectation value in Eq. p. 7 (B.18) is estimated with mini-batches of stochastic trajectories <ref:2606. p. 313790#pg6>.
Where the gains come from
The gains come from minimizing the forward–backward path-space KL divergence, which is equivalent to reducing irreversibility and improving sampling efficiency <ref:2606.13790#pg6> Via the Jarzynski equality, under perfect training, the KL divergence admits an entropy-free energy decomposition <ref:2606. p. 713790#pg6> This leads to QT −1i=0 qF(si+1 si) π0(s0) ∝ π˜∗(sT) <ref:2606. p. 713790#pg6> The application of the IMH correction significantly reduces these discrepancies, bringing the SPS+IMH results into agreement with HMC across all parameter values <ref:2606.13790#pg6>.
Autocorrelation Time and Efficiency
The autocorrelation function serves as a crucial observable for investigating critical slowing down as κ approaches the pseudocritical region <ref:2606.13790#pg6> For SPS+IMH, the integrated autocorrelation time τM,int ≃ 0. p. 75, which is significantly lower than that for HMC, τM,int ≃ 160 <ref:2606. p. 013790#pg6>. This substantial reduction highlights the effectiveness of SPS in mitigating the critical slowing down problem <ref:2606. p. 013790#pg6>.
Effective Sample Size and Acceptance Rate
The normalized effective sample size is defined as ESS/Ns = PNs i=1 wi 2 Ns PNs i=1 w2 i <ref:2606.13790#pg6> Under the log-normal approximation for the SPS importance weights, both the normalized ESS and the mean acceptance rate are controlled by the same quantity, σ 2 = Var(log w) <ref:2606.13790#pg6> This result is used in the main part of the paper in Fig. 6 <ref:2606. p. 013790#pg6>. The relation derived above also determines the volume dependence of the IMH acceptance rate, showing that although the normalized ESS exhibits a deficit that grows linearly with the lattice volume, the IMH acceptance deficit in Eq. p. 1 (D.18) grows only as the square root of the volume <ref:2606. p. 9 p. 913790#pg6>.
Conclusion
In this paper we proposed a neural network method based on nonequilibrium thermodynamics, called the Stochastic Path Sampler (SPS) <ref:2606.13790#pg6> It employs two independent neural networks to learn the discretized Langevin dynamics paths of the forward process and the backward process, respectively <ref:2606. p. 013790#pg6> The method defines the logarithm of the ratio of the forward and backward path probabilities over the entire trajectory as an entropy, and trains the networks by minimizing the KL divergence in path space <ref:2606. p. 013790#pg6> To guarantee sampling accuracy, an Independence Metropolis–Hastings correction step based on the full trajectory is introduced <ref:2606. p. 813790#pg6>. The results show that SPS with IMH correction reproduces all physical quantities computed by HMC with high accuracy, including their behavior in the symmetric phase, the pseudocritical region, and the symmetry-broken phase <ref:2606. p. 913790#pg6>. Moreover, in the pseudocritical region (κ = 0. p. 027), the integrated autocorrelation time of the absolute magnetization for SPS+IMH is significantly lower than that for HMC <ref:2606.13790#pg6>. The study focuses on key physical observables such as magnetization, susceptibility, and free energy density <ref:2606. p. 113790#pg6>.
A SPS with Nelson’s Duality Relation
In the continuous limit, where st+1 − st ≃ ds, the trajectory sum approximates a line integral <ref:2606.13790#pg6> If we assume that at any time slice t, the marginal distribution is P(s, t), and the forward/backward processes are its time-forward/backward processes, then the integral over any closed loop must be zero, yielding the local relation Kθ,F(s, t) + Kθ,B(s, t) = ∇s log P(s, t) <ref:2606. p. 613790#pg6> This is Nelson’s duality relation <ref:2606. p. 113790#pg6>. In SPS, the network learns the integrated path-space balance rather than explicitly enforcing a pointwise duality relation between the two drifts <ref:2606. p. 713790#pg6>.
B SPS Network Architecture, Training, and Generation
We employ a time-conditional convolution neural network to parameterize the forward and backward drifts of the diffusion process on a two-dimensional lattice <ref:2606.13790#pg6> The network takes as state si a lattice configuration R L×8 together with a diffusion time variable t, and outputs a drift for the forward/backward Langevin process <ref:2606. p. 513790#pg6> The final drift field is produced by applying a cyclic padding convolution with kernel size [L/4+1, 3] to the output of the third block <ref:2606.13790#pg6>. To respect the Z2 symmetry of the lattice phi 4 theory, the learned drift satisfies Kθ,F/B(−ϕ, t) = −Kθ,F/B(ϕ, t) <ref:2606. p. 913790#pg6>.
C Magnetization Histogram
The magnetization histograms shown in this appendix are obtained from the SPS proposals before the IMH correction <ref:2606.13790#pg6> These demonstrate that, near the pseudocritical region and in the broken phase, the SPS proposal already populates both Z2-related magnetization sectors, indicating that the learned Z2-equivariant dynamics does not collapse onto a single mode <ref:2606. p. 313790#pg6>.
D Effective Sample Size and Acceptance Rate
The relation derived above also determines the volume dependence of the IMH acceptance rate, showing that although the normalized ESS exhibits a deficit that grows linearly with the lattice volume, the IMH acceptance deficit in Eq. (D.18) grows only as the square root of the volume <ref:2606.13790#pg6>. This explains why the ESS can exhibit a visible volume dependence while the acceptance rate remains comparatively high over the lattice sizes considered here <ref:2606. p. 113790#pg6>. The derived mean acceptance rate is given by ⟨A⟩ ≃ erfc 1/2 s − ln ESS/Ns
<ref:2606.13790#pg6>. The final result is that under the log-normal approximation for the SPS importance weights, both the normalized ESS and the mean acceptance rate are controlled by the same quantity, σ 2 = Var(log w) <ref:2606.13790#pg6>. The relation derived above also determines the volume dependence of the IMH acceptance rate <ref:2606. p. 113790#pg6>.
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
-
Enhanced sampling efficiency near critical points: The Stochastic Path Sampler (SPS) can achieve
the same sampling quality as HMC but with a much shorter autocorrelation time in the pseudocritical region
for two-dimensional phi 4 theory. This means the improved system can generate configurations with significantly less correlation than traditional methods when studying phase transitions. -
Data-free proposal construction: The SPS offers a
novel stochastic-quantization-inspired route to data-free proposal construction for lattice field theory by leveraging a variational free energy principle derived from path-space irreversibility.
This allows the AI system to sample target distributions without needing reference training data from HMC or other Markov chain samplers. -
Topological sector coverage: The SPS networks are constructed to be Z2 equivariant, which
helps the learned proposal maintain support over both Z2-related modes and reduces the risk of mode collapse onto a single sector.
This ensures that the generated configurations explore all relevant topological sectors of the target distribution, addressing a known issue with normalizing flows. -
Exact sampling guarantee: The inclusion of an
extended-space Independence Metropolis–Hastings step
guarantees thatthe final samples strictly adhere to the target physical distribution,
even when dealing with distributions exhibitingnontrivial topological features.
This corrects residual bias from finite network expressivity and time discretization errors. -
Improved acceptance rate prediction: The ESS-based prediction for the acceptance rate, given by erfc (1/2) s − ln ESS/Ns, is used to estimate the mean acceptance rate, showing that
the normalized ESS and the mean acceptance rate are controlled by the same quantity, sigma 2 = Var(log w).
This provides a theoretically grounded way to predict sampling efficiency based on trajectory statistics. -
Quantifying critical slowing down mitigation: The SPS+IMH correction significantly reduces the integrated autocorrelation time of absolute magnetization from approximately 160 for HMC to about 0.5 for SPS, demonstrating
the efficiency of SPS in generating decorrelated samples near the pseudocritical point.
Sources
- Advances in machine-learning-based sampling motivated by lattice quantum chromodynamics
- Physics-Driven Learning for Inverse Problems in Quantum Chromodynamics
- Continuous-mixture Autoregressive Networks for efficient variational calculation of many-body systems
- Flow-based generative models for Markov chain Monte Carlo in lattice field theory
- Estimation of Thermodynamic Observables in Lattice Field Theories with Deep Generative Models
- Flow-based sampling for fermionic lattice field theories
- Equivariant flow-based sampling for lattice gauge theory
- HMC with Normalizing Flows
- Flow-based sampling in the lattice Schwinger model at criticality
- Regressive and generative neural networks for scalar field theory
- Learning Lattice Quantum Field Theories with Equivariant Continuous Flows
- Exploring Generative Networks for Manifolds with Non-Trivial Topology
- Fourier-Flow model generating Feynman paths
- Exploring QCD matter in extreme conditions with Machine Learning
- Variational Autoregressive Networks Applied to $\phi^4$ Field Theory Systems
- Scaling Up Machine Learning For Quantum Field Theory with Equivariant Continuous Flows
- Flow-based sampling for multimodal and extended-mode distributions in lattice field theory
- Stochastic normalizing flows as non-equilibrium transformations
- Stochastic normalizing flows for lattice field theory
- Numerical determination of the width and shape of the effective string using Stochastic Normalizing Flows
Related papers
- Heavy-Meson Bag Parameters using Gradient Flow
- Impact of Irrep Truncations on Quantum Simulations of QCD
- Evaluating generative samplers for lattice field theory by scale
- Renormalized symmetry probes for emergent symmetries and non-chiral lattice fermions
- Scalable Generative Sampling and Multilevel Estimation for Lattice Field Theories Near Criticality
- Lattice QCD study of the K*(892) resonance at the physical point