acceptodds
Withdrawn from ICLR 2027

Learning Low-Order Approximations of the Quark Propagator in Lattice QCD

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The hopping expansion is an important approximation of the quark propagator in lattice quantum chromodynamics, but it diverges for light quarks, which limits the scope of strong-coupling and finite-density simulations. We present an artificial scientific discovery framework using gauge-equivariant neural networks in the Clifford basis to extract interpretable approximations. Our approach yields improved low-order approximations that outperform the hopping expansion at identical computational cost. In the parameter regime where the hopping expansion fails, our approach identifies convergent behavior, thus suggesting a novel series expansion. The discovered approximations generalize to unseen gauge fields and volumes, enabling immediate implementation as plug-in replacements and a new theoretical understanding of quark-propagator expansions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Machine Learning for Physics, Lattice QCD, Artificial Scientific Discovery, Equivariant Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_85g3OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper trains gauge-equivariant neural networks to approximate the inverse Wilson Dirac operator D−1D^{-1}. The goal is to beat the order-nn hopping expansion D−1≈2κ∑ℓ=0nκℓHℓD^{-1} \approx 2\kappa \sum_{\ell=0}^{n} \kappa^\ell H^\ell at the same number of hop operations.

There are two architectures. HC models alternate hopping layers with Clifford layers, whose spin-mixing weights are expanded in the 16-element Clifford basis {Γk}\{\Gamma_k\}. Restricted models use layer-wise weights (wII4±wγγμ)(w_I I_4 \pm w_\gamma \gamma_\mu). Training minimizes the stochastic error ε(M,D)=∥MD−I∥F2/∥I∥F2\varepsilon(M, D) = \|MD - I\|_F^2 / \|I\|_F^2 at fixed mm.

The authors report three main results: (i) At equal hop count, the models have lower error than both the hopping expansion and nn-step GMRES across −5≤m≤2-5 \le m \le 2. (ii) The models transfer to larger volumes and other gauge ensembles. (iii) For m≲−0.5m \lesssim -0.5 the error decreases geometrically with depth, including where the hopping expansion diverges.

Expanding the trained models into path coefficients cp,kc_{p,k}, the authors find that the large-mm models are retuned hopping expansions. The small-mm models acquire new terms, such as the staple coefficient. The authors interpret this as evidence for a novel convergent series expansion of the quark propagator.

Critical strengths
  1. Well-posed problem with an interpretable design. Whether truncated hopping expansions are optimal at fixed cost is a concrete and practically relevant question. Built-in gauge equivariance and the Clifford-basis parametrization make the learned weights directly comparable to the hopping expansion through the path decomposition M=∑p∑kcp,kΓkTpM = \sum_p \sum_k c_{p,k} \Gamma_k T_p.

  2. Careful empirical study with deployable output. The study includes: (i) a dense scan over 36 masses, (ii) five seeds with reported uncertainties and held-out test configurations, (iii) transfer checks across volume (83×16→163×328^3 \times 16 \to 16^3 \times 32) and across ensembles (quenched, dynamical, Haar).

  3. The large-mm formulas (eq. 16, Tables 1–3) can be dropped into existing codes without any ML infrastructure. An interesting observation at small mass. For m≲−1m \lesssim -1 the models keep a convergence rate b<1b < 1, while the hopping expansion diverges (b≥1b \ge 1) and GMRES stagnates (b≈1b \approx 1). Since the HC model is not restricted to polynomials in DD, this result lies outside the Krylov-in-DD class and deserves an explanation.

Critical weaknesses
  1. The claim that GMRES lower-bounds any fixed polynomial in DD is incorrect. GMRES minimizes the residual ∥Dx−a∥\|Dx - a\|. The metric evaluates ∥MDv−v∥\|MDv - v\| with a=Dva = Dv, which is the error ∥Ma−D−1a∥\|Ma - D^{-1}a\|, not the residual. The order-nn hopping expansion is itself a degree-nn polynomial in DD, since H=(I−2κD)/κH = (I - 2\kappa D)/\kappa. If GMRES were such a bound, it could never lose to the hopping expansion. Yet in Fig. 3 (left, m=1.4m = 1.4) it loses at every order. There is also a likely degree mismatch: nn GMRES steps give a polynomial of degree n−1n-1 applied to aa.

  2. Missing classical baselines leave the “novel expansion” claim unsubstantiated. Fixed-coefficient polynomials fitted by least squares on the training configurations are cheap and natural competitors at matched hop count: (a) polynomials p(D)p(D), such as least-squares or Chebyshev fits; (b) polynomials in the normal operator, D†p(DD†)D^\dagger p(DD^\dagger); (c) polynomials in the Hermitian operator, p(γ5D)p(\gamma_5 D). The latter two converge whenever DD is invertible, even when the spectrum of DD encircles the origin, which is exactly where polynomials in DD fail. The Clifford basis contains γ5\gamma_5 and iγμγ5i\gamma_\mu\gamma_5, so HC models can represent such operators. The restricted model has a simpler possible explanation. Learning wI≠wγw_I \neq w_\gamma is equivalent to a Neumann series for a Wilson operator with a rescaled Wilson parameter r=wI/wγr = w_I / w_\gamma; at m=1m = 1, r≈0.88r \approx 0.88.

  3. “Identical computational cost” is asserted, not measured. Production kernels exploit the projectors (I4±γμ)/2(I_4 \pm \gamma_\mu)/2 to communicate half-spinors. With wI≠wγw_I \neq w_\gamma, or with general Clifford weights, full spinors must be communicated, roughly doubling halo traffic. Each 9→99 \to 9 Clifford layer applies 8181 dense 4×44 \times 4 spin matrices per site and keeps 9 full fields in memory. Wall-clock times, or at least communication and flop counts, are needed.

  4. The error metric is insensitive to light-quark physics. ε\varepsilon averages over N=12VN = 12V modes, so the few near-zero modes that dominate light-quark propagators and determinants barely contribute at low order. The reported geometric convergence for n≤16n \le 16 may therefore be pre-asymptotic. This is consistent with the hopping expansion appearing to converge down to m≈−1.2m \approx -1.2, below the quenched β=6\beta = 6 critical mass mc≈−0.82m_c \approx -0.82. The strongest “novel expansion” evidence comes from m∈[−5,−3]m \in [-5, -3]. That region lies well below mcm_c, where doubler branches become light, and is not the light-quark regime the introduction motivates. The value of mhm_h used to split the reported percentages is never stated.

  5. The “series expansion” framing is overstated, and the ML contribution is limited. Models at each depth are trained independently, so they do not form nested truncations. Low-order coefficients drift with depth; for example, c[],1c_{[],1} at m=−2m = -2 moves from ≈0.27\approx 0.27 to ≈0.40\approx 0.40 (Fig. 13). Coefficient uncertainties grow to ∼15\sim 15–20%20\% at n≥4n \ge 4 (Tables 2–3), which points to non-identifiability. The initialization switches from identity to hopping expansion at exactly m=−0.5m = -0.5, the same place the coefficients begin to deviate. The regime boundary could therefore be an optimization artifact. Methodologically, the architecture largely follows Pfahler et al. (2026). The Clifford parametrization adds no expressivity, and no closed-form mass dependence is actually discovered.

Questions for authors
  1. Baselines. At matched hop count, how do the models compare to fixed-coefficient least-squares polynomials p(D)p(D), D†p(DD†)D^\dagger p(DD^\dagger) and p(γ5D)p(\gamma_5 D) fitted on the training configurations? Do the small-mm HC models carry significant weight on the γ5\gamma_5 or iγμγ5i\gamma_\mu\gamma_5 components? Please also correct the GMRES comparison: use matched polynomial degree, and add either an error-norm-optimal Krylov method or a residual-norm evaluation.

  2. Cost. Can you report wall-clock time, or communication volume and local flops per layer, for the HC and restricted models against a standard half-spinor hopping kernel on a realistic multi-node setup?

  3. Low modes and physical regime. Do the convergence rates bb in Fig. 9 survive when the error is restricted to the lowest eigenmodes of D†DD^\dagger D, or when measured in operator norm? What is mcm_c for each ensemble? How do the models perform in the physically relevant window just above mcm_c, for instance on a pion correlator?

Reviewer_YJdWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper uses gauge-equivariant neural networks to discover improved low-order approximations of the quark propagator in lattice QCD, focusing on the limitations of the conventional hopping expansion. The authors construct networks from hopping operators and Clifford-basis transformations, allowing the learned weights to be extracted as explicit coefficients of gauge-covariant paths.

Two architectures are investigated: a restricted model closely related to the conventional hopping expansion and a more expressive HC model that combines hopping and Clifford layers. Experiments show lower approximation errors than truncated hopping expansions and fixed-step GMRES across the tested quark-mass range. The authors also report apparently convergent behavior where the conventional expansion diverges, along with transfer to unseen gauge configurations and larger lattice volumes.

Critical strengths
  1. Rather than using a neural network as a black-box approximation of the quark propagator, the authors extract explicit formulas that can be implemented independently of the trained model. I think this is a useful approach to scientific discovery, especially when the resulting approximations can be incorporated into existing numerical methods.
  2. The improvements over the conventional hopping expansion are substantial, particularly in the mass regime where the latter diverges. The results in Figures 2 and 3 are interesting, and the inclusion of GMRES provides an additional point of comparison.
  3. The use of gauge equivariance is appropriate for the problem, and constructing the layers directly in the Clifford basis makes the learned parameters interpretable. I also appreciate the experiments on different lattice volumes and gauge-field distributions, since they help establish that the learned formulas are not merely fitting individual gauge configurations.
Critical weaknesses
  1. Although the paper may offer useful progress for lattice QCD, I have difficulty identifying a comparably significant contribution to machine learning. The proposed architectures build heavily on existing gauge-equivariant neural networks and on the known structure of the Wilson Dirac operator, while the main results concern the coefficients and convergence behavior of specific quark-propagator approximations. These findings may be important to lattice QCD practitioners, but their significance depends on rather specialized knowledge of the physical problem. I do not think using neural networks to obtain improved formulas is, by itself, sufficient to establish a strong ML contribution, particularly since ML-based scientific discovery is already widely studied in computational physics. There is also little investigation of what the proposed approach teaches us about equivariant representation learning or interpretable scientific discovery more generally. As a result, I find the contribution to the broader ICLR audience rather limited.
  2. In Section 5.2 and Appendix E, the authors observe decreasing approximation errors as the number of layers increases, even where the conventional hopping expansion diverges. This is encouraging, but the conclusion about discovering a new convergent series seems premature. Models of different depths are trained separately, so the resulting approximations do not necessarily correspond to successive truncations of one well-defined infinite expansion. The apparent geometric convergence is also inferred from a finite range of depths, reaching at most 16 layers. Figure 5 shows that selected coefficients stabilize, but the behavior of a few coefficients is not sufficient to establish convergence of the full operator expansion. I would be more convinced if the authors could characterize the proposed expansion mathematically, or at least demonstrate that coefficients extracted at smaller orders remain consistent when higher-order terms are added. Otherwise, the evidence supports a family of increasingly accurate finite-order approximations, which is a weaker but still useful result.
  3. The comparison at equal computational cost is based mainly on the number of hopping operations, as motivated in Section 4.1. While communication associated with these operations may dominate in typical lattice QCD implementations, the HC model also introduces Clifford transformations and multiple intermediate fields. It is unclear whether these additional operations and their memory requirements remain negligible in an actual implementation, particularly when compared with a conventional hopping expansion or GMRES. Since improved accuracy at the same computational cost is one of the central claims, some direct runtime and memory measurements would be useful.
  4. I am also not sure whether the current experiments adequately separate the benefit of learning from the benefit of enlarging the approximation family. The HC architecture includes the hopping expansion and polynomial approximations in the Dirac operator as special cases, so it is not surprising that optimizing a more expressive family can produce lower approximation errors. The interesting question is whether the neural parameterization provides a particularly effective way of finding these coefficients. A comparison against directly optimized low-order polynomial or rational approximations, where feasible, would help clarify this. This seems especially relevant within the convergence region of the conventional hopping expansion, where the restricted model mainly learns modified coefficients rather than a substantially different structure.
  5. The transfer experiments show that the learned approximations remain effective on larger lattices and unseen gauge configurations, which is valuable, but the training still uses a very specific physical setting. In particular, the models are trained separately for each mass parameter, and Section 5.2 notes that substantially different gauge-field distributions can require retraining. I would like to better understand how much training is needed to obtain useful formulas over a new physical parameter regime, and how this cost compares with the benefit of subsequently reusing the extracted approximation.
  6. The paper evaluates the approximations through the normalized Frobenius residual of the inverse operator, but does not establish their impact on a complete downstream lattice QCD calculation. The connection to practical applications is plausible, since the resulting formulas can replace existing low-order approximations. Still, an example demonstrating reduced cost or improved accuracy in an actual simulation would make the significance of the discovered coefficients much clearer.
Questions for authors
  1. Is there evidence that the learned approximations form successive truncations of a consistent infinite series, rather than a collection of independently optimized finite-order operators? In particular, how stable are the coefficients of lower-order paths when increasing the model depth?
  2. Could the authors clarify the computational-cost comparison with the hopping expansion and GMRES? Have the extracted formulas been benchmarked in terms of wall-clock time and memory, including the additional Clifford operations and intermediate fields?
  3. How much of the improvement comes from the expressive approximation family itself, rather than from the proposed neural parameterization and training procedure? Would a direct optimization of polynomial coefficients provide a useful additional baseline?
  4. Can the authors give a concrete example of how the extracted formulas improve a downstream lattice QCD calculation, beyond reducing the operator-approximation error? This would help establish the practical importance of the result.
Reviewer_a27tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Gauge equivariant architectures are trained to approximate the inverse of the Dirac operator by minimizing the solution error averaged over probe vectors. Performance is then assessed on this same metric for a range of bare masses and compared to the hopping expansion and GMRES at same depth: a smaller error is achieved everywhere, with bigger improvements towards smaller mass parameters. This is taken as evidence for a novel converging series expansion. Existing gauge equivariant architectures are used, with the added trick of expanding model weights in the Clifford basis. This does not change final model performance but is claimed to improve interpretability.

Critical strengths

The paper gives a clear report on applying a simple gauge equivariant architecture to a problem of interest in the lattice QCD community. Improvements over the hopping expansion is convincingly demonstrated. Inspecting architecture weights after training is a useful methodological idea, and may lead to interesting insights.

Critical weaknesses

The paper lacks depth and does not demonstrate actual improvements on any downstream task. The utilized solution error seems unobjectionable as a loss function, but if used as a final metric, justification of it's sufficiency is needed. A deeper numerical analysis of the performance for near singular directions and whether they might be missed by the Frobenius norm of the solution error, for example, is lacking.

The "artificial scientific discovery framework" is primarily realized by inspecting model parameters. An early statement that "our work relies heavily on symbolic regression" does not appear to be realized. Analytic formulas for mass dependence, for example, is left to future work. Further, claiming evidence for a novel series expansion and in particular that "our models discover convergent approximations" seems too strong. The latter is a claim (presumably) about the large depth limit and cannot be established from training at finite depth alone. The former appears to be primarily based on having found the neural network approximations to improve with depth. In that sense it is perhaps unsurprising that improved "series expansion" exists. More rigorously, the network approximations would have to be connected to a nested structure found in series expansions. Figure 5 points in that direction, but with mixed evidence; some coefficients seem to drift meaningfully with depth.

Besides expansion in terms of Clifford basis as an interpretability device, ML methodological novelty is limited.

Questions for authors
  • Can the solution error, used as loss, be directly connected to downstream tasks? How can we bound the potentially large error on near singular directions, and rare gauge configurations?

  • Outside the hopping expansion’s convergence radius, our models discover convergent approximations.

    How is "convergent" justified in this sentence? Does it refer to the empirical evidence that the loss decreases with depth? The main text mentions that in the presumably simpler high mass parameter regime, no improvement is observed beyond 6 layers. Figure 3 is cut off at that number, not showing 6-16 layers; could the authors show 8–16 layers and comment on whether training becomes unstable?

  • Were wider and potentially more expressive architectures considered? Why was width limited to 9; only to match the hopping expansion, or was this the result of a more thorough ablation study? Obviously the number of hopping operators is limited, but each could be applied to multiple "copies" of the "quark field" hidden layers. Could more channels per direction with fewer layers outperform the presented architecture at fixed cost?

  • Should "same computational cost" refer to "same number of hoppings"? Else, did you count the number of operations or numerically study runtimes to support this stronger claim? The hopping expansion uses a projection, not the full Clifford basis. Presumably that makes it cheaper and matched depth?

  • GMRES minimizes the residual, not the solution error. Does minimizing the paper's loss for a polynomial ansatz explicitly give a lower (true) lower bound in that family?

  • What is the critical mass for each of the ensembles? Seeing results presented centered around the physically relevant window would be helpful.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.