{"id":"7a008dd4-6175-4f0d-a4f9-1708e5fb4c95","arxiv_id":"2510.24728","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A Minkowski-space PINN trained against a dispersive DSE benchmark reproduces B(p²) in quenched rainbow QED for α ≤ 0.6, but the validation is circular and the abstract's stronger spectral-positivity claims are absent from the body.","lead":"The paper trains a physics-informed neural network to reproduce known Dyson–Schwinger solutions for the fermion mass function in Minkowski-space quenched QED, and claims agreement with a dispersive benchmark. A generalist should read it to see whether neural solvers can be trusted for nonperturbative QFT calculations on the real axis — but the benchmark agreement is partly baked into the training loss.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's central claim about positivity-constrained ansätze and the Fukuda–Kugo zero crossing above α_c=π/3 appears nowhere in the body; the Conclusion explicitly defers such a study to future work.","rationale":"The reader's REJECT verdict is supported: the abstract overclaims a result that the body neither reports nor prepares for. My most load-bearing concern differs from the reader's weakest_assumption (unquantified A(p²)≈1): I focus on the missing positivity-constrained experiment, because the abstract's 'Thus, spectral positivity should be treated as a diagnostic...' cannot be true unless that experiment is actually performed. The reader's weakest_assumption about A≈1 is also real and would matter for the surrogate claim, but it is secondary: even if A=1 were exact, the paper would still not support the positivity-constraint conclusion. Because the central claim as stated fails on evidentiary grounds, the appropriate disposition is rejection; a rewrite that removes the unsupported abstract statements and reframes the work as a neural fit to a dispersive benchmark could support a conditional accept at best. I recommend no change to the reader's verdict.","tokens_in":11504,"tokens_out":7284,"duration_ms":61722,"concrete_test":"Add a positivity-constrained run to the same M-PINN setup by parameterizing Bθ = exp(gθ) (or softplus(gθ)) while keeping losses (17)-(28), hyperparameters, and renormalization fixed, for α ∈ {0.9, 1.0, 1.1} around α_c=π/3. If the constrained network also produces the zero crossing, or the free network fails, the abstract's dichotomy is false. Alternatively, inspect the manuscript and any ancillary code for any positivity-constrained reconstruction; its complete absence settles that the claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract asserts two results: (i) the Fukuda–Kugo benchmark shows a zero crossing above α_c=π/3, and (ii) free-output neural reconstructions reproduce this while positivity-constrained ansätze fail in the supercritical regime, implying spectral positivity should not be imposed blindly. Full-text search shows no implementation of a positivity-constrained ansatz: no network parameterization, loss term, figure, or table compares constrained vs free outputs. The only related statement is in Section IV: 'a systematic study of ... strong-coupling regimes near the loss of Lehmann positivity is still needed.' That sentence explicitly places the supercritical positivity study in the future, contradicting the abstract's past-tense claim. The body's actual benchmark (Figs 1-2, Eqs 17-28) trains Bθ against the dispersive solution via L_data, L_mid_reg, L_high_reg, so agreement is partly by construction; it validates a neural interpolant, not a new result about positivity constraints. Consequently the central claim as stated in the abstract is unsupported by the manuscript's own evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a physics-informed neural network (M-PINN) for the quenched rainbow QED Dyson–Schwinger equation in Minkowski space, targeting the scalar fermion dressing B(p^2). The body develops a dispersive solver based on Lehmann representations and subtracted dispersion relations as a benchmark, then trains the network with a loss that combines the DSE residual with multi-scale regularization and, critically, direct data terms matching the dispersive benchmark. Figures 1–2 show agreement between the PINN and the dispersive solution for α = 0.1, 0.2, 0.4, 0.6 on timelike and spacelike momenta. The abstract additionally claims that free-output neural reconstructions reproduce a Fukuda–Kugo zero crossing above α_c = π/3 while positivity-constrained ansätze fail in the supercritical regime, and draws a conclusion about spectral positivity.","tokens_in":11843,"tokens_out":6480,"duration_ms":54739,"significance":"The methodological idea of combining a dispersive benchmark with a residual-based multi-scale PINN loss for Minkowski DSEs is useful and could be a step toward differentiable surrogates for propagator equations. The explicit comparison of a neural solver with a traditional spectral solver under a common truncation is a strength. However, the abstract's central physical claim about positivity-constrained ansätze is absent from the body, and the body's main agreement is largely trained into the network via data-loss terms, so the independent validation value is limited. If the claims were supported, the paper would be significant; in its current form, the primary advertised results are not established.","major_comments":[{"comment":"The abstract asserts: 'Neural reconstructions with free output reproduce this behavior, while positivity-constrained ansätze fail in the supercritical regime' and concludes that spectral positivity should not be imposed blindly. No positivity-constrained ansatz is defined, implemented, or tested anywhere in the body. The only related statement is Section IV: 'a systematic study of ... strong-coupling regimes near the loss of Lehmann positivity is still needed,' which explicitly defers such a study to future work. The Fukuda–Kugo zero crossing above α_c = π/3 is also not demonstrated in the text; it is only mentioned through reference [31]. The abstract therefore makes a central claim that the manuscript does not support.","section":"Abstract vs. Section IV"},{"comment":"The network is trained with L_data^TL/SL = ⟨(B_θ − B_trad)²⟩, L_mid_reg, and L_high_reg, which directly minimize the squared difference between the network output and the dispersive benchmark over the full momentum range. Consequently, the agreement shown in Figs. 1–2 is largely by construction: the network is fitted to the benchmark, so the 'reproduction' is not an independent check. To support the claim that the M-PINN 'reproduces the dispersive solution,' the authors should either train without the benchmark data terms (or on a held-out subset) and compare on untouched intervals, or report that the physical residual L_phys is small and show that the solution satisfies the DSE independently. Without this, the central validation claim is circular.","section":"§III, Eqs. (19)–(21)"},{"comment":"The paper assumes A(p²) ≈ 1, so that M(p²) = B(p²), and the residual in Eq. (17) contains only B. However, the dispersive solver itself computes A(p²) from ρ_v via Eq. (15), and Eq. (6) defines the dynamical mass as M = B/A. The manuscript never quantifies how close A(p²) is to 1 for the couplings studied (α = 0.1–0.6). If A deviates from unity, the reduced equation solved by the PINN is not the full rainbow DSE, and the benchmark comparison on B alone does not validate the physical mass function. This assumption should be tested by evaluating A from the dispersive spectral solution and reporting it alongside B.","section":"§III, Eq. (17) and Eq. (6)"}],"minor_comments":[{"comment":"The title advertises 'spectral functions,' and the abstract mentions 'Fukuda–Kugo equation, the spectral unitary equations, and the modified unitary equations,' but the body does not present a separation of these equations or any spectral-density results. No figure shows ρ_s, ρ_v, or σ_s, σ_v. Consider either including such quantities or adjusting the title/abstract to match the actual content.","section":"Title/Abstract"},{"comment":"The term 1_TL in Eq. (29) is not defined. If it is an indicator function for the timelike branch, it should be explicitly introduced; otherwise the reader cannot tell how B_UV is used in the spacelike loss L_tail.","section":"§III, Eq. (29)"},{"comment":"The loss weights (w_phys, w_data^TL/SL, w_mid, w_high, etc.) are listed symbolically, but their numerical values are never given. The statement 'the same weights are used for all couplings' is not reproducible without these values.","section":"§III, Eq. (27)–(28)"},{"comment":"The abstract mentions a comparison of 'on-shell and momentum-subtraction schemes,' but the body appears to use a single renormalization condition (µ² = m²) throughout. No explicit comparison of two schemes is presented in the results.","section":"Abstract vs. Body"},{"comment":"The spectral solver equations are quoted without derivation or a pointer to the specific equations of Ref. [28]. Since these equations are the benchmark, the authors should state their provenance explicitly and note any approximations, ranges of validity, and the treatment of thresholds (e.g., s0) and UV cutoffs.","section":"§III, Eqs. (10)–(16)"},{"comment":"There are several presentation issues: 'Facuty' in the affiliation, missing superscripts in Eq. (17), incomplete reference [53] ('8 2025'), and minimal figure captions that do not identify line styles or list the couplings. These should be corrected.","section":"General"}],"recommendation":"reject","confidential_remarks":"The manuscript has a useful methodological core, but the published abstract makes a specific physical claim about positivity-constrained ansätze failing in the supercritical regime that is not backed by any experiment in the body, and the body's central validation is partly circular because the network is trained against the benchmark it is supposed to reproduce. The A(p²)≈1 approximation further weakens the physical interpretation. These are load-bearing issues. The paper could potentially be reworked into a more modest benchmarking study with the overclaims removed and the circularity addressed, but that would require substantial new numerical experiments. I recommend rejection in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The useful part is the M-PINN setup: extending the author's Euclidean PINN DSE program to Minkowski kinematics, with separate networks for spacelike and timelike branches, Fourier features for the oscillatory tail, and a UV template. The dispersive solver description, while quoted from prior work without derivation, gives enough detail to be reproduced. If the paper were just \"here is a differentiable surrogate trained to match the dispersive solution,\" that would be a modest but legitimate contribution for people working on DSEs.\n\nThe problems are real. The abstract claims two things the body does not do: the Fukuda–Kugo zero crossing above α_c = π/3 and the failure of positivity-constrained ansätze. No positivity-constrained network appears anywhere; the conclusion explicitly says that study is still needed. That's not a minor overstatement, it's the headline result.\n\nThe body's validation is also circular. Equations (19)–(21) include L_data, L_mid_reg, and L_high_reg, which directly minimize (Bθ − B_trad)² against the dispersive benchmark. So of course the figures agree. The residual L_phys is in the loss, but with data terms dominating, agreement is by construction, not an independent check. Without an ablation showing L_phys alone gets close, the claim \"reproduces the dispersive solution\" is just curve fitting.\n\nAlso, A(p²) ≈ 1 is assumed without quantification. The network learns B and the paper calls it the mass function. At the couplings tested (α up to 0.6), the error in A could matter, but no estimate is given. And no code or data is provided, so the numbers cannot be checked independently.\n\nIf the abstract were corrected and the paper reframed as \"a neural surrogate trained to the dispersive benchmark, with a comparison of training strategies,\" it could be a useful methods note for the DSE/ML community. As it stands, the central claim is unsupported. I would still send it to peer review rather than desk reject, because the underlying numerical work is worth scrutiny and the authors should have the chance to fix the framing.","headline":"Useful neural-surrogate recipe, but the abstract claims a positivity result the body doesn't contain and the central 'reproduction' is trained in, so the paper needs major reframing before it is trustworthy.","tokens_in":12319,"tokens_out":2588,"would_cite":false,"duration_ms":23543,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A physics-informed neural network reproduces the dispersive QED Dyson–Schwinger solution across momentum scales; the paper also argues spectral positivity should be a diagnostic, not a neural constraint.","keywords":["Dyson–Schwinger equations","Minkowski spacetime","physics-informed neural networks","spectral functions","Lehmann representation","quantum electrodynamics","dynamical mass generation","rainbow truncation"],"falsifier":"Solve the full coupled A(p^2)–B(p^2) rainbow system at alpha = 0.6 in the same scheme; if the vector dressing A deviates from 1 by more than a few percent at the scales plotted, then the reduced equation solved by the network is not the DSE being benchmarked, and agreement on B alone does not validate the dynamical mass M = B/A.","tokens_in":11304,"feed_emoji":"🧠","tokens_out":9786,"duration_ms":64842,"temperature":0.7,"pith_summary":"Working in quenched rainbow QED with Landau gauge, the paper sets out to show that a Minkowski-space physics-informed neural network (M-PINN) can solve the Dyson–Schwinger equation for the fermion scalar dressing B(p^2) directly on the real momentum axis, in both spacelike and timelike regions. The network is trained with a residual loss that embeds the reduced rainbow integral equation, together with multiscale regularisation, Fourier features for the timelike branch, and monotonicity/smoothing penalties for the spacelike branch. Against a dispersive benchmark built from Lehmann spectral representations and subtracted dispersion relations, the M-PINN matches quantitatively from infrared to ultraviolet scales for couplings alpha = 0.1, 0.2, 0.4, 0.6, in both on-shell and momentum-subtraction renormalisation schemes. The abstract further claims that neural reconstructions with free output reproduce the zero crossing of the Fukuda–Kugo criterion above the critical coupling alpha_c = pi/3, while positivity-constrained ansätze fail there, leading to the methodological conclusion that spectral positivity is a diagnostic of the Lehmann representation, not a constraint to be imposed blindly. If these claims hold, residual-based neural solvers offer a compact, differentiable route to Minkowski-space propagators and ultimately to more realistic truncations.","feed_headline":"Matches QED spectral benchmark in Minkowski space","feed_subtitle":"A physics-informed network reproduces the fermion mass function from infrared to ultraviolet at four couplings.","key_machinery":"The load-bearing objects are (i) the Lehmann spectral representation of the fermion propagator together with once-subtracted dispersion relations, which turn the DSE into coupled unitary integral equations for the spectral weights sigma_v and sigma_s, and (ii) the reduced rainbow residual R[B](p^2) that the M-PINN minimises. The dispersive solver discretises the spectral variable on a log grid, handles Cauchy principal values by an extrapolation scheme, and reconstructs B(p^2) by subtracted dispersion relations. The neural solver represents B(p^2) with a multilayer perceptron, using Fourier features on the timelike branch, a multiscale loss with IR/intermediate/UV windows, a perturbative UV","core_discovery":"The central claim is that the M-PINN—a neural network B_theta trained by minimising the residual of the one-dimensional split rainbow equation R[B](p^2)=B(p^2) - (3 alpha/4 pi)[integral terms]—reproduces the B(p^2) computed by a spectral dispersive solver in quenched rainbow QED in Landau gauge. The training fixes the same truncation, gauge, and subtraction point as the benchmark, and the agreement holds across the full momentum domain for alpha in {0.1, 0.2, 0.4, 0.6}, including the timelike branch with its mild UV oscillations, for both on-shell and momentum-subtraction renormalisation. The paper treats M(p^2)=B(p^2)/A(p^2) with A(p^2)≈1, so the learned function is identified with the dyna","pith_inferences":["Because the body's comparison is between the free-output M-PINN and the dispersive solver, the abstract's stronger claim—that positivity-constrained ansätze fail above alpha_c—is not directly demonstrated by the figures; a constrained-versus-free run on the same benchmark would settle it.","The identification M(p^2)=B(p^2) rests on an unquantified A(p^2)≈1; computing the coupled A and B equations at the larger couplings would show whether the reduced residual is sufficient or whether the benchmark is only validating a surrogate of the scalar dressing.","The paper's framing suggests a broader methodological moral: in theories like QCD where loss of spectral positivity is a physical signal, imposing positivity as a hard neural constraint could suppress exactly the nonperturbative dynamics under study; one could test this by running the same M-PINN on a model with known positivity violation.","A differentiable, continuous surrogate for B(p^2) could make renormalization-group improvements and uncertainty quantification straightforward, since the UV template in the loss could be replaced by an RG-motivated tail and retrained without changing the solver, an extension the paper lists as a next step."],"forward_implications":["For the four tested couplings and both renormalisation schemes, the M-PINN reproduces the dispersive B(p^2) from IR to UV, so residual-based neural solvers can handle nonlocal integral equations with realistic renormalisation directly on the real axis.","Because the neural solution is continuous and differentiable, it can serve as a compact surrogate for the spectral solver, easing extensions to dressed vertices, unquenched photons, and uncertainty-aware variants.","The abstract's claim that free-output networks capture the zero crossing above alpha_c = pi/3 while positivity-constrained ansätze fail implies that spectral positivity should be treated as a diagnostic of the Lehmann representation, not a constraint imposed a priori.","The same architecture, trained on lattice correlators instead of DSE residuals, could reconstruct spectral densities and bypass ill-posed analytic continuation, connecting DSE and lattice approaches."],"fun_headline_variants":["Neural reconstruction matches QED spectral benchmark","M-PINN reproduces QED mass function across couplings","Positivity constraint fails in supercritical QED","Neural net solves Minkowski QED spectral function","QED spectral function from neural nets, positivity not"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The argument assumes A(p^2) ≈ 1 so that M(p^2) = B(p^2); if A deviates appreciably at the couplings studied, the network is solving a reduced equation, not the full rainbow DSE, and the benchmark agreement does not validate the dynamical mass.","fun_headline_variants_meta":{"raw":{"variants":["Neural reconstruction matches QED spectral benchmark","M-PINN reproduces QED mass function across couplings","Positivity constraint fails in supercritical QED","Neural net solves Minkowski QED spectral function","QED spectral function from neural nets, positivity not"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000244,"raw_usage":{"total_tokens":1338,"prompt_tokens":682,"completion_tokens":656,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":426,"completion_tokens_details":{"reasoning_tokens":580}},"tokens_in":426,"tokens_out":656,"duration_ms":5221,"temperature":1.0,"reasoning_tokens":580,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T11:22:48.051484+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the full coupled A(p^2)–B(p^2) rainbow system at alpha = 0.6 in the same scheme; if the vector dressing A deviates from 1 by more than a few percent at the scales plotted, then the reduced equation solved by the network is not the DSE being benchmarked, and agreement on B alone does not validate the dynamical mass M = B/A.","supporting_citations":[],"review_version":1}