{"id":"1e64e3f1-eeb6-41aa-9848-692c054ac5cd","arxiv_id":"2512.13812","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"For Z2 GGFPEPS Monte Carlo in 2+1D, updating 1/4–1/2 of links per step is fastest in wall-clock time, gauge fixing generally slows convergence, and explicit spatial averaging helps the magnetic energy error.","lead":"This paper studies the numerical performance of a tensor-network method (GGFPEPS) for simulating lattice gauge theories, testing how Monte Carlo error shrinks when update sizes, gauge fixing, and spatial averaging are varied. It gives practical recipes — update roughly a quarter to half the links per step, avoid heavy gauge fixing in Monte Carlo — that make larger gauge-theory simulations more feasible.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Update-size optimum is established at one coupling and one bond dimension; without a scan over λ and D, the 1/4–1/2 links-per-step recommendation may be an artifact of the tested regime.","rationale":"The gauge-fixing construction (Sec. VII, Apps. A-B) is internally coherent: tree fixing is possible, p0(G) and F_O(G) are gauge invariant, and the EC cost reduction 2^{Nlinks}→2^{Nlinks/2+1} for Z2 follows from fixing ~L^2 links. So the EC part of the abstract is sound. The empirical recommendations, which are the main practical value of the paper, are less secure. The update-size claim rests on a single parameter point and no explicit bond-dimension scan; Section VIII's electric-energy result is explicitly tied to the per-link Pfaffian overhead in the current implementation. Since the paper offers no code/data release, these results cannot be independently audited. The reader's CONDITIONAL verdict is appropriate; my proposed check would upgrade it only if the optimum persists across couplings and bond dimensions. I also note an internal inconsistency: the text and the caption of Fig. 11 appear to state opposite λ-dependencies for the all-plaquette vs single-plaquette total-energy EOM; this should be corrected, though I do not base my verdict on it.","tokens_in":19482,"tokens_out":9844,"duration_ms":91522,"concrete_test":"Rerun the Section VI update-size scan on a 6×6 lattice (Nlinks=72) at λ = g_E/2 = 2/g_B ∈ {0.39, 1.25, 2.0}, for at least two bond dimensions (e.g., one vs. two virtual copies per leg), using 10 independent MCMC seeds per setting. At 10^5 measurement steps, record wall-clock EOM (same rebinning procedure) for each update fraction (1, 1/8, 1/4, 1/2, 3/4 Nlinks). If the fastest setting leaves the [1/4, 1/2] interval in any tested (λ, D) combination or is not distinguishable across seeds, the headline recommendation fails to generalize; if it remains inside and is separated by >1σ across seeds, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section VI (Fig. 3) is the sole basis for the headline update-size recommendation. It scans L=2,4,6 at fixed g_int=g_mass=1, g_E=1/g_B=1.25, with no variation of the gauge coupling λ=g_E/2=2/g_B and no stated bond dimension/ansatz parameters. The wall-clock optimum is set by the balance between decorrelation gain per step and sequential per-link cost; that cost is O(N^2) for the weight update (Woodbury/determinant lemma), but the electric-energy measurement adds a Pfaffian per averaged link (Sec. VIII). Both the autocorrelation time and the relative cost of updates vs. measurements change with λ and with bond dimension, so there is no a priori reason the optimal fraction is universal. Moreover, EOM values in Figs. 3/6/7/12/13 come from individual chains, using an autocorrelation cutoff of 10^-2 for rebinning, and are reported without error bars; the observed orderings may not be statistically significant. The claim is therefore not yet supported as a general GGFPEPS property, only as a property of these runs/implementation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies algorithmic choices in variational Monte Carlo simulations with gauged Gaussian fermionic projected entangled pair states (GGFPEPS), for the Z2 lattice gauge theory in 2+1 dimensions. It develops a gauge-fixing construction over trees, claims a reduction of exact-contraction cost from O(|G|^{N_links}) to O(|G|^{N_links/2+1}), and reports numerical experiments on update-step size, gauge-fixed sampling, and spatial averaging of energy observables. The main reported findings are: an optimal update size of roughly 1/4 to 1/2 of the lattice links per step in wall-clock time; gauge fixing generally slows Monte Carlo convergence; and not exploiting translation invariance for electric-energy measurement can improve time-to-error in some cases.","tokens_in":19811,"tokens_out":7301,"duration_ms":71071,"significance":"If the numerical conclusions hold, the paper provides practical recipes for making GGFPEPS variational searches more efficient and larger-scale, a direction that is relevant to the field. The gauge-fixing construction in Appendices A and B is coherent: the transformation of |ψ(G)⟩ under pure gauge transformations is derived explicitly, and the square-root reduction in exact-contraction complexity is concrete and plausible. The 'chessboard' tree idea is a useful insight, and the paper correctly identifies a trade-off between per-step decorrelation and per-step cost. The main weakness is that the numerical support for the headline recommendations is narrow and lacks statistical characterization; the general claims in the abstract and conclusions are stronger than the evidence presented.","major_comments":[{"comment":"The central recommendation of an 'optimal update size' of 1/4 to 1/2 of links per step is based on a single setting: g_int = g_mass = 1, g_E = 1/g_B = 1.25, and unspecified ansatz parameters/bond dimension. No scan over λ or bond dimension is reported, although the per-step cost, autocorrelation time, and EOM all depend on these. The EOM curves are single-chain point estimates (based on Eq. (30) with an autocorrelation cutoff of 10^-2), and differences between 1/4, 1/2, and 3/4 of links are not shown to be statistically significant. This does not yet establish a general property of GGFPEPS MC; it establishes a property of the tested runs. The abstract should be correspondingly qualified, or additional scans and error bars should be provided.","section":"§VI, Fig. 3"},{"comment":"The entire numerical comparison relies on the EOM estimator defined by rebinning with a decay-time cutoff of 10^-2. The autocorrelation often only decays to the 10^-2 to 10^-3 level before fluctuating, so the estimated τ and hence the rebinned EOM are sensitive to this ad hoc cutoff. No error bars on the EOM values are given, and no multiple independent chains are reported. The ordering of curves in Figs. 3, 6, 7, 8, 12, and 13 may therefore not be statistically robust. This issue is load-bearing because the paper's practical recommendations are all derived from these orderings.","section":"§V, Eq. (30)"},{"comment":"The claim that 'gauge fixing generally slows MC convergence' is stronger than the data support. The tested trees are a maximal tree, fixed rows, and a chessboard tree. Fig. 7 shows that for many λ values the chessboard tree is comparable to or slightly better than no gauge fixing, and the EOM values have no error bars. The qualitative explanation in terms of non-local updates after fixing is plausible and supported by Fig. 9, but the quantitative claim of a 'general' slowdown needs either a more systematic parameter scan or a more cautious wording ('in the settings studied here').","section":"§VII, Figs. 5–8"},{"comment":"The paper does not specify the bond dimension (number of virtual copies μ per leg), the ansatz parameter values at which the simulations are run, or the exact warm-up and measurement protocols beyond '10^5 warmup steps and 10^5 measurement steps'. These details are essential for assessing the transferability of the update-size and averaging recommendations, and for reproducibility. Citations to Refs. [42,44] are not sufficient, since the algorithmic conclusions depend on these implementation choices.","section":"§VI–§VIII, numerical setup"},{"comment":"The paper claims that gauge fixing 'makes it possible' to run 4×4 Z2 exact-contraction ground-state searches, but no exact-contraction results, timings, or demonstration are reported. The complexity reduction from O(|G|^{N_links}) to O(|G|^{N_links/2+1}) is a solid theoretical statement, but the practical feasibility claim would benefit from at least one actual 4×4 exact-contraction computation or timing benchmark.","section":"§VII, exact contraction"}],"minor_comments":[{"comment":"The denominator in F_Mf(G) is written as '⟨ψ(G)|ψ(G)|ψ(G)|ψ(G)⟩' and should be '⟨ψ(G)|ψ(G)⟩'.","section":"Eq. (37)"},{"comment":"Typo: 'Monte Calro' should be 'Monte Carlo'.","section":"§IX"},{"comment":"'The last data point of maximal tree for λ=2 was committed for clarity' should presumably read 'omitted'.","section":"Fig. 8 caption"},{"comment":"Typo: 'couplindsg' and 'plaquttes' should be 'couplings' and 'plaquettes'.","section":"Fig. 10 caption"},{"comment":"The sentence 'The colors indicate which link each operator acts on.' is repeated twice.","section":"Fig. 15 caption"},{"comment":"Typo: 'samples samples' should be 'samples'.","section":"§V"}],"recommendation":"major_revision","confidential_remarks":"The mathematical and algorithmic core—gauge fixing over trees and the square-root EC complexity reduction—is sound and novel enough for this journal. However, the numerical evidence for the headline 'optimal update size' and 'gauge fixing generally slows MC' claims is too thin in its current form: single chains, no error bars on EOM, one coupling, no bond-dimension scan. I would not reject, because the deficiencies are fixable within the manuscript's scope by adding more systematic numerical studies and/or softening the general claims. The paper's heavy reliance on the same group's earlier GGFPEPS papers for ansatz details is a minor concern, but not a correctness issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick take on arXiv:2512.13812. The genuinely useful parts are the gauge-fixing construction and the square-root reduction in exact-contraction cost, which makes 4x4 Z2 with dynamical fermions contractible. The MC convergence findings (update size around 1/4 to 1/2 links, gauge fixing slows convergence, spatial averaging helps magnetic but not electric) are plausible and worth having, but they are empirical recipes from a fairly narrow set of runs.\n\nWhat's new: the tree-gauge-fixing procedure for GGFPEPS, the O(|G|^{Nlinks/2+1}) EC scaling, and the first systematic look at update size and translation-invariance trade-offs in non-action-based MC for this ansatz. The appendices A-B are coherent; the explanation for why gauge fixing hurts MC convergence — local updates become non-local in the fixed configuration, with the chessboard tree mitigating it — is convincing and consistent with the data. Credit also for the EC feasibility at 4x4, which is a real practical milestone for benchmarking.\n\nSoft spots: the EOM curves in Figs 3/6/7/12/13 come from individual chains and are reported without error bars on the EOM itself. The rankings may be real, but the statistical significance is unknown. The update-size recommendation is based on one coupling point (g_E=1/g_B=1.25) and one bond dimension/ansatz setting; the stress-test concern that the optimum could shift with lambda and D is legitimate. The translation-invariance section also shows some scatter across ansatz parameters. No code or data is released, which makes the recipes hard to reproduce. None of this is fatal — the findings are plausible and internally consistent — but the paper overstates slightly by presenting the update-size range as a general recipe rather than a property of the tested regime.\n\nBottom line: this is a solid, honest algorithmic study that deserves a serious referee. The referee should ask for error bars on EOM, a broader scan in lambda and bond dimension, and ideally code/data availability. I'd cite it if working on GGFPEPS; for a general reading group it's niche, but for the tensor-network/LGT crowd it's relevant.","headline":"Useful algorithmic study of GGFPEPS Monte Carlo; the gauge-fixing and exact-contraction results are solid, but the update-size recommendation rests on a narrow parameter scan and EOM curves without error bars.","tokens_in":20263,"tokens_out":2115,"would_cite":true,"duration_ms":19201,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["11.15.Ha","02.70.Uu"],"model":"deepseek-v4-flash","headline":"For GGFPEPS variational Monte Carlo on Z2 lattice gauge theory in 2+1 dimensions, the fastest wall-clock convergence comes from updating roughly a quarter to half of the links per MC step, while gauge fixing that reduces exact-contraction c","keywords":["GGFPEPS","lattice gauge theory","variational Monte Carlo","gauge fixing","translation invariance","Z2 gauge theory","Gaussian fermionic PEPS","tensor networks"],"falsifier":"Run the same update-size scan on a 6x6 Z2 GGFPEPS at g_E=1/g_B=0.39, measuring error-on-the-mean versus wall time for 1/8, 1/4, 1/2, and 3/4-of-links updates: if the fastest fraction leaves the 1/4–1/2 window, the headline recipe fails. For gauge fixing, compare exact-contraction wall time at 4x4 and 6x6 with and without maximal-tree fixing to test the claimed square-root scaling, and for a fixed-row tree verify directly that a one-link flip in the unfixed space requires a row-length number of link changes in the fixed space.","tokens_in":19421,"feed_emoji":"⚛️","tokens_out":5907,"duration_ms":53235,"temperature":0.7,"pith_summary":"The paper asks how to make variational Monte Carlo searches with gauged Gaussian fermionic PEPS (GGFPEPS), a sign-problem-free tensor-network ansatz for lattice gauge theories, numerically efficient. Using Z2 gauge theory with dynamical fermions in 2+1 dimensions, it finds that updating about a quarter to half of the lattice links per Monte Carlo step gives the fastest error reduction per unit wall-clock time. It also shows that fixing the gauge over trees cuts exact-contraction cost from exponential in the number of links to roughly its square root, making exact 4x4 Z2 contractions feasible, but that the same gauge fixing generally slows Monte Carlo convergence because local updates become effectively non-local. Finally, it shows that averaging the magnetic energy over all plaquettes accelerates convergence, while the electric energy is best evaluated on a single link. If these findings carry beyond the tested couplings, they give concrete recipes for scaling GGFPEPS ground-state simulations to larger systems.","feed_headline":"Quarter-to-half link updates fastest for gauge-theory Monte Carlo","feed_subtitle":"Gauge fixing cuts exact-contraction cost to its square root but slows sampling; a disconnected 'chessboard' fix is the nearly harmless excep","key_machinery":"The machinery is the GGFPEPS representation, in which the state is expanded as a sum or integral over gauge configurations, |Psi> = Integral DG |G>|psi(G)>, so the weight p(G) = <psi(G)|psi(G)>/Z is non-negative and can be computed efficiently from fermionic covariance matrices. Monte Carlo updates are made affordable by the Woodbury matrix identity and the matrix determinant lemma, which update the weight and inverse covariance after a local link change in O(N^2) instead of O(N^3). Gauge fixing works through a pure-gauge transformation h(x) that pins tree links to the identity, removing roughly half the links from the configuration sum. The central mechanism behind the gauge-fixing slowdown","core_discovery":"On its own terms, the paper claims that three algorithmic choices control the efficiency of GGFPEPS variational Monte Carlo for a Z2 lattice gauge theory with dynamical fermions in 2+1 dimensions. First, there is an optimal update size: updating 1/4 to 1/2 of the links per step balances reduced autocorrelation against the sequential cost of local updates, giving the fastest decay of the error on the mean in wall-clock time. Second, gauge fixing over trees is a valid reduction because both the weight and the observables are pure-gauge invariant; it reduces exact-contraction cost from O(|G|^N_links) to O(|G|^(N_links/2+1)), enabling exact 4x4 Z2 contraction, but it generally slows MC convergen","pith_inferences":["If the update-size optimum is controlled by the ratio of per-link update cost to per-step autocorrelation decay rather than by the gauge group itself, then for larger bond dimensions or non-Abelian groups the same wall-clock analysis is the right way to pick the fraction, even if the numerical optimum shifts.","The results suggest a two-track workflow: gauge-fixed exact contraction for benchmarking and debugging, and unfixed or chessboard-fixed Monte Carlo for production runs, so that the two techniques complement rather than compete.","The electric-energy result hints that observables whose evaluation is not memory-bound should be spatially averaged only when per-site cost is negligible; otherwise single-site sampling wins, so future implementations should optimize observable evaluation before adding averaging.","Because the connected-tree gauge-fixing penalty grows linearly with the length of the fixed string, its Monte Carlo slowdown likely worsens with system size; testing this scaling explicitly would sharpen the practical recommendation."],"forward_implications":["Future GGFPEPS ground-state searches can use 1/4 to 1/2 link updates per step as a default starting point, the update range that won the wall-clock error-on-the-mean comparison across tested lattice sizes and couplings.","Gauge fixing makes exact contraction of 4x4 Z2 systems feasible, providing an MC-independent benchmark for the ansatz and a debugging tool that does not suffer from sampling errors.","The gauge-fixing slowdown can be largely avoided by fixing disconnected single-link trees (the chessboard pattern), which nearly matches unfixed MC convergence while still reducing the configuration count.","The paper's recipe for total-energy sampling is to average the magnetic energy over all plaquettes and the electric energy over a single link, combining the better convergence behavior of each.","Together these choices shift the practical scale limit of GGFPEPS variational studies, potentially enabling larger lattices and, later, extensions to higher dimensions or non-Abelian gauge groups."],"fun_headline_variants":["Quarter-to-half link updates win for gauge Monte Carlo","Gauge fixing slows MC but slashes contraction cost","Skipping translation symmetry speeds error convergence","Optimal update size found for fermionic gauge simulations","Chessboard gauge fix nearly harmless for Z2 lattice"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper's recommendations assume that the tested settings—Z2 gauge group, 2+1D square lattice with periodic boundaries, couplings g_int=g_mass=1 and g_E=1/g_B in roughly 0.39–1.25, and the specific ansatz parameters—are representative enough that the observed rankings (update-size optimum of 1/4 to 1/2 of links, gauge fixing slowing MC, single-link electric-energy averaging winning) are general GGFPEPS properties rather than artifacts of these runs.","fun_headline_variants_meta":{"raw":{"variants":["Quarter-to-half link updates win for gauge Monte Carlo","Gauge fixing slows MC but slashes contraction cost","Skipping translation symmetry speeds error convergence","Optimal update size found for fermionic gauge simulations","Chessboard gauge fix nearly harmless for Z2 lattice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000633,"raw_usage":{"total_tokens":2777,"prompt_tokens":779,"completion_tokens":1998,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":1937}},"tokens_in":523,"tokens_out":1998,"duration_ms":13054,"temperature":1.0,"reasoning_tokens":1937,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T16:20:39.813958+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same update-size scan on a 6x6 Z2 GGFPEPS at g_E=1/g_B=0.39, measuring error-on-the-mean versus wall time for 1/8, 1/4, 1/2, and 3/4-of-links updates: if the fastest fraction leaves the 1/4–1/2 window, the headline recipe fails. For gauge fixing, compare exact-contraction wall time at 4x4 and 6x6 with and without maximal-tree fixing to test the claimed square-root scaling, and for a fixed-row tree verify directly that a one-link flip in the unfixed space requires a row-length number of link changes in the fixed space.","supporting_citations":[],"review_version":1}