- 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 . The goal is to beat the order- hopping expansion 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 . Restricted models use layer-wise weights . Training minimizes the stochastic error at fixed .
The authors report three main results: (i) At equal hop count, the models have lower error than both the hopping expansion and -step GMRES across . (ii) The models transfer to larger volumes and other gauge ensembles. (iii) For the error decreases geometrically with depth, including where the hopping expansion diverges.
Expanding the trained models into path coefficients , the authors find that the large- models are retuned hopping expansions. The small- 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
-
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 .
-
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 () and across ensembles (quenched, dynamical, Haar).
-
The large- formulas (eq. 16, Tables 1–3) can be dropped into existing codes without any ML infrastructure. An interesting observation at small mass. For the models keep a convergence rate , while the hopping expansion diverges () and GMRES stagnates (). Since the HC model is not restricted to polynomials in , this result lies outside the Krylov-in- class and deserves an explanation.
Critical weaknesses
-
The claim that GMRES lower-bounds any fixed polynomial in is incorrect. GMRES minimizes the residual . The metric evaluates with , which is the error , not the residual. The order- hopping expansion is itself a degree- polynomial in , since . If GMRES were such a bound, it could never lose to the hopping expansion. Yet in Fig. 3 (left, ) it loses at every order. There is also a likely degree mismatch: GMRES steps give a polynomial of degree applied to .
-
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 , such as least-squares or Chebyshev fits; (b) polynomials in the normal operator, ; (c) polynomials in the Hermitian operator, . The latter two converge whenever is invertible, even when the spectrum of encircles the origin, which is exactly where polynomials in fail. The Clifford basis contains and , so HC models can represent such operators. The restricted model has a simpler possible explanation. Learning is equivalent to a Neumann series for a Wilson operator with a rescaled Wilson parameter ; at , .
-
“Identical computational cost” is asserted, not measured. Production kernels exploit the projectors to communicate half-spinors. With , or with general Clifford weights, full spinors must be communicated, roughly doubling halo traffic. Each Clifford layer applies dense spin matrices per site and keeps 9 full fields in memory. Wall-clock times, or at least communication and flop counts, are needed.
-
The error metric is insensitive to light-quark physics. averages over modes, so the few near-zero modes that dominate light-quark propagators and determinants barely contribute at low order. The reported geometric convergence for may therefore be pre-asymptotic. This is consistent with the hopping expansion appearing to converge down to , below the quenched critical mass . The strongest “novel expansion” evidence comes from . That region lies well below , where doubler branches become light, and is not the light-quark regime the introduction motivates. The value of used to split the reported percentages is never stated.
-
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, at moves from to (Fig. 13). Coefficient uncertainties grow to – at (Tables 2–3), which points to non-identifiability. The initialization switches from identity to hopping expansion at exactly , 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
-
Baselines. At matched hop count, how do the models compare to fixed-coefficient least-squares polynomials , and fitted on the training configurations? Do the small- HC models carry significant weight on the or 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.
-
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?
-
Low modes and physical regime. Do the convergence rates in Fig. 9 survive when the error is restricted to the lowest eigenmodes of , or when measured in operator norm? What is for each ensemble? How do the models perform in the physically relevant window just above , for instance on a pion correlator?