Learning Low Modes of the Wilson Dirac Operator with Gauge-Equivariant Networks
summary
The gist
Learning Low Modes of the Wilson Dirac Operator with Gauge-Equivariant Networks is about training gauge-equivariant neural networks to generate test vectors that approximate the low-mode subspace of…
In short
The research trains gauge-equivariant neural networks using a Rayleigh-Ritz loss to learn test vectors that approximate the low-mode subspace of the Wilson Dirac operator. This allows these learned vectors to replace costly adaptive V-cycle refinement in multigrid methods, significantly reducing setup time while maintaining or improving solve time.
Key concepts
- Wilson Dirac Operator (D)
- This is a mathematical operator central to Lattice QCD calculations. Solving linear equations involving this operator is computationally expensive, especially when dealing with small quark masses where low-mode components cause convergence issues in iterative solvers.
- Gauge-Equivariant Neural Networks
- These are specialized neural networks designed to respect the symmetries of the underlying lattice gauge theory. They use specific layers, like parallel-transport layers, to map gauge fields into fermion fields in a way that preserves these physical symmetries during training.
- Rayleigh-Ritz Loss (LRR)
- This is the mathematical objective function used to train the network. It minimizes the Ritz values of D†D within the subspace spanned by the network's output vectors, effectively forcing the learned vectors to span a low-mode subspace of A[U].
- Adaptive Multigrid (AMG) Test Vectors
- In adaptive multigrid, test vectors are crucial for capturing error components on coarse grids. This paper uses a neural network to generate these test vectors, aiming to better approximate the low-mode error components that conventional methods struggle with.
Terminology used across episodes
This episode discusses
- Learning Low Modes of the Wilson Dirac Operator with Gauge-Equivariant Networks · Paper Radio
- Local coherence and deflation of the low quark modes in lattice QCD
- Adaptive Aggregation Based Domain Decomposition Multigrid for the Lattice Wilson Dirac Operator
- Adaptive Multigrid Algorithm for Lattice QCD
- Adaptive multigrid algorithm for the lattice Wilson-Dirac operator
- Gauge-equivariant neural networks as preconditioners in lattice QCD
- A novel gauge-equivariant neural-network architecture for preconditioners in lattice QCD
- Gauge-equivariant pooling layers for preconditioners in lattice QCD
- Spectral Inference Networks: Unifying Deep and Spectral Learning
- JuliaQCD: Portable lattice QCD package in Julia language
- Lattice Gauge Theory via LLVM-Level Automatic Differentiation
The paper
Learning Low Modes of the Wilson Dirac Operator with Gauge-Equivariant Networks · Read on arXiv
Minjae Choi, Hiroshi Ohno
Graduate School of Science and Technology, University of Tsukuba · Center for Computational Sciences, University of Tsukuba
In adaptive multigrid, test vectors that better approximate the low-mode subspace of the Wilson Dirac operator D can reduce the solve time, but their construction can be costly. In some cases, the setup takes longer than a single solve. To reduce the cost of generating test vectors, we train networks built from gauge-equivariant parallel-transport layers once on a quenched ensemble, using a Rayleigh-Ritz loss for D dagger D. We then test whether the trained networks maintain their performance at other hopping parameters and on configurations not used in training. In these tests, the network outputs are used as test vectors, and the costly V-cycle refinement in the adaptive setup is omitted to reduce the setup time. This substantially reduces the setup time while leaving the solve time largely unchanged in most cases. Near the critical hopping parameter κ c, the solve time is also reduced in some cases.
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: "Learning Low Modes of the Wilson Dirac Operator with Gauge-Equivariant Networks".
Mira: Here's the gist of the paper. Learning Low Modes of the Wilson Dirac Operator with Gauge-Equivariant Networks is about training gauge-equivariant neural networks to generate test vectors that approximate the low-mode subspace of the Wilson Dirac operator to reduce setup time in adaptive multigrid methods Core Problem and Solution Lattice QCD requires enormous computational costs, and one of its main bottlenecks is solving the linear equation D u = b, where D is the Wilson Dirac operator, b a given source, and u the fermion field to be computed <ref:2610.30982#pg3> Low-mode components of the error hinder the convergence of conventional iterative solvers, particularly at small quark masses <ref:2610.30982#pg3> Algebraic multigrid (AMG) addresses this difficulty by treating these components on a coarse grid <ref:2610.30982#pg4> The effectiveness of adaptive multigrid depends on how well the coarse space spanned by its test vectors captures the low-mode components of the error <ref:2610.30982#pg5> In this work, we use a gauge-equivariant neural network to construct test vectors, aiming to reduce the computational cost relative to the conventional adaptive setup <ref:2610.30982#pg6> Network Architecture and Training Signal The architecture adopts the parallel-transport (PT) and linear layers introduced by
6: <ref:2610.30982#pg8> The PT layer maps n input fermion fields i,..., n to m output fields a,..., m as psi a (x) = sum over b=1 to n W a b T p (b)
U: phi b (x), where W a b in C 4×4 are trainable spin-mixing matrices and T p (b)
U: parallel transports phi b along the fixed path p (b) <ref:2610.30982#pg9> The network maps the gauge field U to K fermion fields v1,..., vK, which are trained to span an approximated low-mode subspace of D†D The network layers are described in the order in which they are applied Training Objective and Generalization The network is trained by minimizing the Ritz values of D†D, computed using the Rayleigh-Ritz procedure in the subspace spanned by its output vectors The training loss is LRR = sum over k=1 to K lambda k = tr(S−1Aˆ) = tr(V†V−1V†A
U: V) Minimizing LRR pushes the columns of V toward a low-mode subspace of A
U: The loss depends on V only through this K × K problem The network is trained on quenched gauge configurations generated with the heatbath algorithm at gauge coupling beta = 6.0 on a 163 × 32 lattice Performance and Efficiency Gains The learned vectors can replace the adaptive V-cycle test-vector loop in two-level multigrid This lowers the setup time per configuration by a factor that grows with K For K ≥ 24 the solve time stays comparable to or shorter than with the conventional setup, whereas for K = 12 the learned vectors are less effective, in particular near kappa c The setup time is thus reduced by a factor that increases with K, from about 1.5–2 at K = 12 to 8–11 at K = 64 Training the network for 4000 steps takes about 19–25 h per model Case Study Results The learned vectors give shorter solve times than the conventional setup for K = 48 and 64 on both configurations at kappa = 0.157 Near kappa c, the learned vectors shorten the solve for K = 24 on both configurations The network setup, including the network evaluation, takes 5–19 s and grows slowly with K Training the model for 4000 steps takes about 19–25 h per model The one-time network training is excluded from the final wall time measurements We trained gauge-equivariant networks with a Rayleigh-Ritz loss to produce approximate low modes of D†D As a case study, the network outputs can replace the adaptive V-cycle test-vector loop in two-level multigrid Future work includes extending the study to other ensembles and larger volumes, as well as further lattice QCD applications beyond the MG benchmark presented here We also plan to investigate how power series in hopping operators act on the eigenvectors of the Dirac operator and how the choice of parallel-transport paths affects the learned low-mode subspace Acknowledgments We thank Akio Tomiya for fruitful discussions This work is supported by JSPS KAKENHI Grant No.
Kai: 22H05112 Numerical work was performed on GPU resources at the Center for Computational Sciences, University of Tsukuba References
1: M.
Mira: Lüscher, Local coherence and deflation of the low quark modes in lattice QCD, JHEP 07 (2007) 081
arXiv:0706.2298: .
2: A.
Paper summary: Kai: So we've been looking at this paper, "Learning Low Modes of the Wilson Dirac Operator with Gauge-Equivariant Networks," which shows a way to generate test vectors that approximate the low-mode subspace of the Wilson Dirac operator D using gauge-equivariant neural networks.
Mira: The core thesis is that training these networks on quenched ensembles using a Rayleigh-Ritz loss for D†D allows you to create high-quality test vectors that significantly reduce setup time in adaptive multigrid methods <ref:2610.30982#pg1>.
Lev: The implication is that if you can afford the initial nineteen to twenty-five hours of training time, you get a much faster setup phase for solving D u = b when using two-level multigrid <ref:2610.30982#pg2>.
Kai: They demonstrated this works for various parameters, showing that the learned vectors maintain performance even on configurations they weren't trained on, which is a good sign for their reliability <ref:2610.30982#pg1>.
Mira: The main takeaway is that this approach tackles the configuration-dependent nature of building coarse spaces in adaptive multigrid by replacing that costly loop with a pre-trained network output <ref:2610.30982#pg2>.
Lev: They also pointed out that the effectiveness changes based on K, showing it's not a one-size-fits-all solution, especially when you are near the critical hopping parameter kappa c <ref:2610.30982#pg1>.
Kai: So, while it doesn't change the solve time much in most cases, for large K, like sixty-four it makes the setup time much smaller than conventional methods <ref:2610.30982#pg1>.
Mira: Ultimately, this work provides a method for reducing the computational overhead associated with constructing test vectors in adaptive multigrid solvers by leveraging trained gauge-equivariant networks <ref:2610.30982#pg1>.
Conclusion: Kai: So we're wrapping up this piece on "Learning Low Modes of the Wilson Dirac Operator with Gauge-Equivariant Networks." Basically, they trained these networks to make test vectors that are good at representing those low modes of the Dirac operator D.
Mira: It really boils down to using gauge-equivariant neural networks to build a shortcut for adaptive multigrid setups. They aren't just guessing what the coarse grid should look like; they are learning a subspace that matters for convergence.
Lev: From an error correction standpoint, this means we can potentially skip that expensive V-cycle refinement step in two-level multigrid. If the network gives us good test vectors, we save setup time per configuration.
Kai: That’s what they showed with the numbers: for larger K values, like sixty-four, the solve time stays competitive with conventional methods. It's not about getting a perfect solution every single time, but about making the setup phase much quicker to do repeatedly on massive lattices.
Mira: I think the real assumption here is that this learned subspace is robust enough to work even when you change things like the hopping parameter kappa. They tested it outside of their training set, which suggests it's not just memorizing a specific configuration.
Lev: Yeah, but we have to watch how K affects it. If K is too small, say twelve, the learned vectors aren't as good near the critical hopping parameter kappa c. That's where things get tricky and you might need more than just a trained network to handle the physics correctly.
Kai: So while this reduces setup time substantially—from maybe fifteen minutes down to a few seconds for large K —we still have to worry about those low-mode components themselves. Does it actually help solve the underlying QCD problem faster, or is it just making the iteration setup cheaper?
Mira: It's primarily about making the iterative solver work better by giving it a better starting point, which lowers the number of iterations needed for convergence. It doesn't change how hard the physics is; it just changes how fast we can test that physics.
Lev: Exactly. For real hardware, if you have to train that network once for a specific lattice setup, that initial cost is high—about twenty-five hours for training—but then each subsequent solve setup becomes tiny. That's the trade-off we’re looking at here on the engineering side.
Kai: So it’s a massive upfront investment in training, but if you need to run many different problems, especially those near phase transitions, this pre-training pays off by slashing the time you spend just getting started.
Mira: It's a smart way to handle the complexity of adaptive multigrid test vector construction without having to do that heavy lifting every single time we want to check something new.
Lev: And it opens up avenues for testing other things, though, like how different parallel-transport paths affect the learned subspace itself. That’s where the next batch of work needs to go if we’re serious about using this in production codes.
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