{"id":"2949679b-63a1-4d86-9e37-90049062f524","arxiv_id":"2509.10378","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A matrix-free neural preconditioner learns to map gauge configurations to modified configurations whose Dirac operators approximate the inverse, halving CG iterations and transferring across lattice sizes.","lead":"This paper trains neural networks to turn a gauge field into a modified gauge field whose Dirac operator acts as a preconditioner, speeding up iterative solves of lattice Dirac equations without ever building large matrices. The method roughly halves solver iterations on a 2D test model and transfers to larger lattices and different couplings without retraining.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Learned mapping collapses to trivial U~ = -0.5U; analytic preconditioner nearly matches NN, undermining the operator-learning claim.","rationale":"The reader's weakest assumption focused on the structural constraint of M^{-1} and the unsuitability of the random-projection loss. My concern is more specific: the structural constraint combined with the learned mapping results in a nearly trivial transformation, which is a concrete manifestation of that assumption. The paper's empirical claims (halved iterations, zero-shot transfer) are supported by the data, but the significance of the neural-network component is weakened by the fact that a simple analytic transformation performs almost as well. This does not invalidate the results but does call for a quantitative comparison and tests in regimes where the simple relation should fail. Since the reader's verdict is already CONDITIONAL (asking for code release, timing, etc.), my concern reinforces the need for conditions but does not shift the verdict to a different category. Therefore, UNCHANGED is appropriate, with the additional condition that the authors must address the trivial-mapping issue by comparing against the analytic baseline systematically.","tokens_in":16505,"tokens_out":4694,"duration_ms":50214,"concrete_test":"Compute the mean squared error between Gθ(U) and -0.5U for all test configurations and compare CG iteration counts and condition numbers of the NN preconditioner against the analytic U~=-0.5U preconditioner for a wider range of κ and β, including κ away from criticality and nontrivial gauge fields. If the MSE is small and the iteration counts are statistically indistinguishable, the neural network provides no measurable advantage over the analytic formula, and the operator-learning claim is unsupported. Additionally, test on SU(3) or configurations with topological features where -0.5U is unlikely to be optimal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's own analysis in Section 4.3 shows the learned transformation Gθ(U) is approximately U~ = -0.5U with a phase shift of π. Table 4 demonstrates that explicitly constructing U~_simple = -0.5U and using D†D[U~_simple] as a preconditioner yields CG iteration counts nearly identical to the neural-network preconditioner: for L=8, 43 vs 40; for L=16, 104 vs 99; for L=32, 278 vs 267. This suggests the neural network is recovering a simple pointwise scaling and rotation, not a complex nonlocal mapping. Consequently, the central claim that the operator-learning framework learns a general mapping enabling effective preconditioning is not supported by the experiments: a parameter-free analytic transformation achieves almost the same performance. The zero-shot transfer may be trivial because the mapping is local and universal. The paper acknowledges the simple relationship but provides no quantitative comparison of Gθ(U) to -0.5U over all test configurations, nor does it test regimes where the simple relation should fail (e.g., different κ far from criticality, different gauge groups). Without such evidence, the added value of the neural network over the analytic preconditioner is unclear, and the framework's broader relevance is questionable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a matrix-free neural preconditioner for Wilson-Dirac normal equations in lattice gauge theory. Given a gauge configuration U, a neural network Gθ (FNO or FCN) outputs a transformed configuration Ũ; the preconditioner M^{-1}=D†D[Ũ] is then applied as a linear operator inside preconditioned CG. Training uses a random-projection loss that minimizes ||M^{-1}D†D v - v||² without ever forming matrices explicitly. Experiments in the 2D U(1) Schwinger model with two Wilson fermions show that, for lattices L=8–64 at κ=0.276, β=2.0, the preconditioner reduces the condition number and roughly halves CG iterations; a model trained at L=16 transfers zero-shot to other volumes and to modified κ and β. Section 4.3 reports that the learned transformation is approximately Ũ = -1/2 e^{iπ} U, and Table 4 shows that the manually constructed analytic preconditioner D†D[-1/2 U] gives nearly identical CG iteration counts to the trained networks.","tokens_in":16830,"tokens_out":5643,"duration_ms":63703,"significance":"If the claimed gains hold, the matrix-free construction is a useful step toward ML preconditioners for large-scale lattice calculations, since it avoids explicit matrix storage and setup/decomposition costs and offers zero-shot volume transfer. The paper is transparent about its limitations: the loss is an acknowledged proxy, the simple learned map is identified, and higher powers of M^{-1} are proposed but not fully analyzed. However, the central added value of the learned operator over the analytic transformation is not established. The paper's own Table 4 shows the neural network outperforms the analytic pointwise mapping by only 3–11 CG iterations out of 78–548, which is a small fraction of the total solve. This weakens the 'operator learning' claim: the evidence suggests the network has essentially rediscovered a simple scaling/phase relation. The framework is plausible and the empirical methodology is sound, but the current manuscript does not yet demonstrate that the neural component is necessary or advantageous relative to a parameter-free analytic preconditioner.","major_comments":[{"comment":"This section reports that the learned maps satisfy |Ũ|≈1/2 and a phase difference of π, i.e., Ũ≈-1/2 U. Table 4 then shows that the manually constructed preconditioner D†D[-0.5U] gives CG iteration counts of 43±2, 104±7, 278±20 for L=8,16,32, while the trained FNO gives 40±2, 99±7, 267±21. The differences are only 3, 5, and 11 iterations, out of unpreconditioned counts of 78, 201, and 548. This is load-bearing: the paper's main contribution is that an operator-learning framework learns a useful preconditioner, but the evidence indicates the network has recovered a simple pointwise scaling/rotation that can be implemented without any training. The authors should (i) quantify how much Gθ deviates from -0.5U over the whole test set rather than a single L=64 instance (Fig. 7), (ii) test regimes where the small-κ expansion argument in §4.3 is expected to fail (e.g., κ further from criticality","section":"Section 4.3, Table 4"},{"comment":"The abstract and Section 4.1 claim computational efficiency and reduced complexity, and the Remark states that the NN preconditioner avoids IChol's setup cost and triangular solve. However, no wall-clock timings are reported anywhere. Figure 4 shows only iteration counts, even though the text in §4.1 says 'we compare the average number of iterations and the time required to reach convergence.' Since CG per-iteration cost differs between one sparse matrix-vector product with M^{-1} and an IChol triangular solve, iteration counts alone do not establish a computational advantage. The paper should report setup time, per-iteration cost, and total solve time on the same hardware (A100) for unpreconditioned, IChol, even-odd, and NN-preconditioned solves, at least for L=32 and L=64. Without these data, the 'computationally efficient' claim is unsubstantiated.","section":"Section 4.1 and Remark after Table 2"}],"minor_comments":[{"comment":"The affiliation 'Lawrance Berkeley National Laboratory' should be 'Lawrence Berkeley National Laboratory'.","section":"Title page / affiliations"},{"comment":"The fermion action term is rendered as '1X f=0' due to a formatting issue; it should be a sum over flavor index f=0,1.","section":"Eq. (2)"},{"comment":"The notation 'U(1)^{X×T×d} ⊂ C^{X×T×d}' is confusing: U(1) is the target space of each component, not a power of U(1). Clarify the spaces. In Fig. 6, the scatter plots would be more informative with a quantitative similarity metric such as a correlation coefficient or normalized distance.","section":"Eq. (5) and Fig. 6"},{"comment":"The text says the code is available at the given URL, but the footnotes state the repository 'will be publicly accessible upon acceptance.' These statements should be made consistent.","section":"Code availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honestly written and the experiments are reproducible in principle, but the analytic baseline in Table 4 undercuts the central 'operator learning' contribution. The authors should be asked to strengthen the comparison (e.g., quantitative deviation from -0.5U, regimes where the simple map fails, wall-clock timing) or to reframe the paper as an analysis of a simple scaling preconditioner. This seems fixable within the scope of a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you want to see a careful attempt at a matrix-free neural preconditioner for Wilson-Dirac normal equations. The authors train FNO/FCN networks to map a gauge configuration U to a transformed configuration U~, then use D^†D[U~] as a preconditioner in CG. Experiments are on the 2D Schwinger model (U(1) gauge theory, two flavors). What is genuinely new: the matrix-free formulation (no explicit matrices anywhere, only operator applications), the random-projection loss that avoids condition-number computation, and zero-shot transfer across lattice sizes and coupling values. The paper is honest: it reports iteration counts with error bars, compares against IChol and even-odd preconditioners, and acknowledges the random-projection loss is an unproven proxy.\n\nThe main soft spot is the one the paper exposes in Section 4.3. The learned transformation is essentially U~ = -0.5U with a π phase shift. The authors test this simple analytic map explicitly and find it gives 43 vs 40 CG iterations for L=8, 104 vs 99 for L=16, 278 vs 267 for L=32. The neural network buys only a few percent over a parameter-free transformation. That doesn't invalidate the framework, but it does mean the 'operator learning' contribution is, for this test case, mostly learning a pointwise scaling. The paper notes the small gap suggests correlations across sites and directions matter, but it doesn't quantify how much or show a regime where the simple map fails. Without such evidence, the advantage of the network over the analytic baseline is not established.\n\nSecond soft spot: no wall-clock timings. The paper calls the preconditioner 'computationally efficient,' but applying M^-1 costs one D^†D application, the same complexity as applying A. If PCG iterations halve, time per iteration roughly doubles, so the wall-clock gain is unclear. A timing table would settle this. Also, the code is not yet public, and the higher-power results (Table 3) show modest gains at increased per-iteration cost.\n\nWho should read this: people working on ML preconditioners for lattice field theory or sparse iterative solvers. It is a legitimate proof of concept with honest reporting, and the matrix-free idea may matter for 4D QCD if it transfers. The central empirical claim (iteration reduction and zero-shot transfer) holds up as an existence proof. But I would not let the paper's framing overstate the neural network's role before comparing to the analytic baseline in the target regime.\n\nRecommendation: send for peer review. The method deserves discussion, and the flaws are addressable: add timings, compare against U~=-0.5U systematically across parameter ranges, and either show a case where the network outperforms the simple map or temper the claims accordingly.","headline":"A well-executed proof-of-concept for a matrix-free learned preconditioner, but the paper's own Section 4.3 shows a trivial analytic map does almost as well, which weakens the operator-learning claim.","tokens_in":17295,"tokens_out":3213,"would_cite":true,"duration_ms":35603,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F08","65F10","68T07","81T25"],"pacs":["11.15.Ha","12.38.Gc"],"model":"deepseek-v4-flash","headline":"A matrix-free neural network that rewrites the gauge field roughly halves Dirac solver iterations in the Schwinger model and transfers to new lattice sizes.","keywords":["preconditioner","neural operator","lattice gauge theory","Wilson-Dirac operator","conjugate gradient","matrix-free","zero-shot transfer","Schwinger model"],"falsifier":"Take the analytic rule Ũ = -½U, build M^{-1}=D†D[Ũ], and measure the condition number of M^{-1/2} A M^{-1/2} (or the CG iteration count) on SU(3) Wilson-Dirac operators at physical parameters in 4D. If the iteration reduction does not approach the roughly twofold gain seen in the Schwinger model—or if the spectrum of M^{-1}A develops clusters far from 1—then the learned mapping and the simple analytic form are artifacts of the Abelian 2D model, and the zero-shot transfer claim is restricted to that setting.","tokens_in":16451,"feed_emoji":"⚛️","tokens_out":13757,"duration_ms":119862,"temperature":0.7,"pith_summary":"The paper proposes a way to precondition the large, ill-conditioned linear systems that dominate lattice gauge theory calculations, without ever forming or storing a matrix. The central idea is to train an operator-learning network to map a gauge field configuration U to a new configuration Ũ such that the Wilson-Dirac normal operator built from Ũ, M^{-1}=D†D[Ũ], acts as an approximate inverse of the original operator A=D†D[U]. Because M^{-1} has the same sparse stencil as A, applying it costs one matrix-vector product, and the whole pipeline is matrix-free from training through solver use. In the 2D Schwinger model the preconditioner reduces the condition number by roughly a factor of four and cuts CG iterations approximately in half, and a network trained on one lattice size works on other sizes and couplings without retraining. The learned mapping turns out to be nearly Ũ ≈ -½U, a simple analytic rule that itself already provides much of the acceleration.","feed_headline":"Neural preconditioner halves Dirac solves, transfers to new lattices","feed_subtitle":"Trained once on a small lattice, it halves CG iterations and transfers to larger lattices with no retraining.","key_machinery":"The load-bearing object is the map Gθ taking a U(1) gauge field U to a transformed field Ũ, together with the fixed rule that the preconditioner is M^{-1}=D†D[Ũ]. This ties the preconditioner to the same sparse stencil as A, so matrix-vector products are cheap and no matrix is ever assembled; two operator-learning architectures (a local convolutional network and a global spectral-mode network) are trained to produce Ũ. Training uses an unsupervised random-projection loss, Eq. (8): the squared norm of M^{-1}A v - v averaged over K random vectors v, which proxies for the Frobenius norm of M^{-1}A - I without computing spectra. A final piece is the identity A=D†D[U]=|1-2κH|², with H linear in U","core_discovery":"The central claim is that an effective preconditioner for Wilson-Dirac normal equations can be learned as a function of the gauge field alone. The construction fixes M^{-1}=D†D[Ũ(x)] with Ũ(x)=Gθ(U(x)), so the preconditioner is the same discretized normal operator evaluated on a transformed gauge field and therefore shares the original matrix's sparse structure. Training is unsupervised and matrix-free: the loss is the average squared norm of (M^{-1}A - I) applied to K random vectors from an isotropic standard normal distribution. With this loss, a network trained on 1600 configurations of the two-flavor Schwinger model reduces the condition number of A by roughly a factor of four across L=8","pith_inferences":["If the learned mapping generalizes to non-Abelian groups, the simple form Ũ ≈ -½U is unlikely to survive because link variables are unitary matrices rather than phases; a non-Abelian analog would more plausibly be a group-specific redefinition, so the preconditioner structure should be tailored to the gauge group.","The random-projection loss, which the paper acknowledges lacks theoretical grounding, could be replaced by a spectral proxy such as stochastic estimates of the condition number; the paper's own sensitivity study suggests K=128 is a pragmatic choice, and a principled objective might close the remaining gap to classical incomplete-factorization preconditioners.","The zero-shot volume transfer indicates the mapping depends mostly on local features of the gauge field, consistent with the local-network's superior transfer at L=8. A testable extension is that the preconditioner will continue to transfer to even larger lattices (e.g., L=128) with bounded loss of efficiency; if instead iteration counts grow as in the unpreconditioned case, the local-feature assu","One could apply the same construction to other fermion discretizations (staggered, domain-wall) by using the corresponding normal operator as the fixed stencil; this would test whether the learned mapping again reduces to a simple field redefinition or requires nontrivial correlations."],"forward_implications":["Preconditioners for lattice fermion solves can be generated on the fly from gauge fields with no matrix assembly or setup step, making the approach scalable to the very large operators (dimension greater than 10^8) encountered in lattice QCD.","A single training run on a small lattice can be amortized across many action parameters and lattice geometries: zero-shot volume transfer means no retraining for unseen configurations within the explored parameter ranges.","Since per CG iteration the neural preconditioner costs one matrix-vector product and no triangular solve, it avoids the per-iteration overhead of incomplete-factorization or even-odd preconditioners, at the cost of somewhat more iterations than the incomplete-factorization approach.","Using (M^{-1})^p with p=2, 3, or 4 as the preconditioner reduces iterations further (e.g., the FCN-based model at L=16 drops from 99 to 79–81 iterations), providing a tunable iteration-vs-per-step-cost knob.","The explicit analytic choice Ũ = -½U already gives most of the acceleration, implying that a useful preconditioner may be constructible without training at all in the Schwinger-model setting."],"fun_headline_variants":["Neural net halves Dirac iterations, works on larger lattices","Matrix-free neural preconditioner cuts CG iterations by 50%","Zero-shot neural preconditioner for Dirac operator halves solves","AI-learned preconditioner speeds lattice QCD, no retraining","Neural preconditioner: 2x faster Dirac solves, transfers across sizes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that forcing the preconditioner to have exactly the same sparse structure as the original Wilson-Dirac normal operator—namely, evaluating D†D on a transformed gauge field—still leaves enough flexibility for M^{-1} to approximate A^{-1} well enough to accelerate conjugate gradient, and that the training loss, which checks the error only on random vectors, is a faithful enough proxy for that approximation.","fun_headline_variants_meta":{"raw":{"variants":["Neural net halves Dirac iterations, works on larger lattices","Matrix-free neural preconditioner cuts CG iterations by 50%","Zero-shot neural preconditioner for Dirac operator halves solves","AI-learned preconditioner speeds lattice QCD, no retraining","Neural preconditioner: 2x faster Dirac solves, transfers across sizes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1316,"prompt_tokens":757,"completion_tokens":559,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":471}},"tokens_in":501,"tokens_out":559,"duration_ms":6152,"temperature":1.0,"reasoning_tokens":471,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T17:51:36.540144+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the analytic rule Ũ = -½U, build M^{-1}=D†D[Ũ], and measure the condition number of M^{-1/2} A M^{-1/2} (or the CG iteration count) on SU(3) Wilson-Dirac operators at physical parameters in 4D. If the iteration reduction does not approach the roughly twofold gain seen in the Schwinger model—or if the spectrum of M^{-1}A develops clusters far from 1—then the learned mapping and the simple analytic form are artifacts of the Abelian 2D model, and the zero-shot transfer claim is restricted to that setting.","supporting_citations":[],"review_version":1}