{"id":"911205ef-0103-4ead-8a73-ea64416a26df","arxiv_id":"1908.09359","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A variational Monte Carlo scheme for finite PEPS, with a sequential spin-pair update, accurately simulates 32x32 Heisenberg and 24x24 frustrated J1-J2 lattices.","lead":"This paper presents a faster way to simulate quantum magnets with two-dimensional tensor networks, using Monte Carlo sampling to handle states with up to 32 by 32 spins. The method matches gold-standard quantum Monte Carlo and DMRG results, and could help study frustrated magnetic materials where those methods fail.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Accuracy claims for L=18–32 and for the frustrated 24×24 rest on an unverified near-global optimum; the authors' own report of random L-dependent errors (Sec. II C) signals incomplete optimization, so the agreement with QMC/DMRG could be partly coincidental.","rationale":"The reader's weakest assumption is the same one I would flag: the largest-system results are only as good as the optimization. The paper contains genuine supporting evidence: a detailed comparison of sequential and random sampling on 8×8 (Fig. 1), agreement with ED on small triangular and kagome systems, and DMRG-quality correlations on 8×28. These show the method works when independent reference data are available. The danger is concentrated precisely in the regime the paper advertises as new—24×24 frustrated, beyond QMC/DMRG—where no direct check exists. The authors' own admission of randomly varying errors as a function of L is an in-text warning that the optimization is not demonstrably converged. A multi-seed and continued-optimization test is cheap relative to the original 500-core runs and would either validate the conditional acceptance or expose a systematic bias. Because the reader already marked the paper CONDITIONAL with the same concern, I would not change the verdict.","tokens_in":20003,"tokens_out":10769,"duration_ms":125707,"concrete_test":"Run the same D=8 optimization for L=24 Heisenberg and for 24×24 J1-J2=0.5 from at least three independent simple-update initializations (different random seeds and different SU schedules), with the same MC sweep budget and step-length schedule. Then continue one run of each with δ reduced to 0.0001 (or smaller) until the energy change is below 1e-6 for 20 consecutive gradient steps, recording the final energy and its MC error. If the run-to-run spread or the continued-optimization gain exceeds the MC error (≈1e-5) for either model, the reported energies are not converged variational minima and the claimed agreement with QMC/DMRG should be re-evaluated; if both are within MC error, the incomplete-optimization concern is settled for the largest systems.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the D=8 VMC-PEPS states are accurate enough that their energies and correlations can be compared with QMC/DMRG at up to 32×32 (Heisenberg) and 24×24 (J1-J2). That claim requires the stochastic gradient optimization (Eq. 10), initialized by simple update, to reach a near-global minimum of the variational energy at every system size. The evidence for this is weak: the stopping rule is only 'energy decreases very slowly' (Sec. II B), no gradient-norm or energy-variance diagnostic is reported, and Sec. II C explicitly says the errors versus QMC 'vary randomly with respect to L, possibly because of incomplete optimization.' This random pattern is exactly what one expects from under-optimized states: a converged D=8 variational family should have an energy bias that changes smoothly with L, not one that jumps around. For the frustrated 24×24 case there is no direct benchmark, so the claimed 'excellent agreement' relies on extrapolations against DMRG, vQMC, and iPEPS whose own spread in energy is larger than the claimed 5e-5 error bar. If some of the reported energies are local minima, the finite-size extrapolations and the comparison are not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a variational Monte Carlo (VMC) scheme for finite projected entangled pair states (PEPS) and applies it to the square-lattice spin-1/2 Heisenberg model up to 32x32 and the frustrated J1-J2 model up to 24x24, together with small triangular- and kagome-lattice benchmarks. The central methodological innovations are a sequential spin-pair update that reduces the cost of an MC sweep to O(ND^4Dc^2), a spin-inversion-symmetrized sampling weight, and a stochastic-gradient optimization initialized from simple-update states. The authors report ground-state energies and spin correlations that agree with QMC, DMRG, vQMC, and exact-diagonalization results to within roughly 1e-4 in energy for the unfrustrated case and claim excellent agreement for the frustrated case as well. The paper concludes that the method substantially advances finite-PEPS simulations and can address systems beyond the reach of QMC and DMRG.","tokens_in":20326,"tokens_out":7674,"duration_ms":73535,"significance":"If the central claim holds, this is a significant methodological advance: it substantially enlarges the system sizes accessible to finite-PEPS simulations while retaining benchmark-quality accuracy. The strengths of the paper are its validation against independent QMC (Sandvik), DMRG (Gong et al.), vQMC (Hu et al.), and exact-diagonalization results, with no parameters fitted to reproduce those benchmarks; the documented convergence of the cutoffs Dc1 and Dc2 in the appendices; and the honest admission in Sec. II C that the errors vary randomly with L, possibly because of incomplete optimization. However, the accuracy claims for the largest and frustrated systems rest on optimization and finite-size-scaling assumptions that are not fully supported: the SIS estimator is not specified precisely, no convergence diagnostic (gradient norm, energy variance, or restart tests) is provided, and the spread among central-bulk extrapolations is not folded into the quoted errors.","major_comments":[{"comment":"The authors state that the errors of energy and magnetization 'vary randomly with respect to L, possibly because of incomplete optimization' (Sec. II C), yet no optimization-convergence diagnostic is reported. The stopping criterion is only that the energy decreases very slowly (Sec. II B); no gradient norm, energy variance, or restart-from-different-initialization test is provided for the large systems. This matters because the central claim that D=8 PEPS is accurate up to 32x32 requires that the reported energies are representative of the variational optimum. Concretely, Table IV shows energy errors of 4.3e-5 (L=14), 7.6e-5 (L=20), 8.5e-5 (L=24), and 7.9e-5 (L=32), which fluctuate rather than decreasing smoothly with L. Please add a quantitative convergence diagnostic and at least one independent initialization test for a representative large system, and discuss how the random L-dependent error affects the FSS extrapolations.","section":"Sec. II C, Fig. 3, Table IV"},{"comment":"The spin-inversion-symmetrized state Φ(S)=Ψ(S)+Ψ(S_bar) is introduced in Sec. II B, and the text says that all observables are evaluated with the new weight |Φ(S)|^2. However, Eq. (3) defines the local energy as Σ_{S'} [Ψ(S')/Ψ(S)] <S'|H|S>, not with Φ ratios. If the Monte Carlo average uses weights |Φ(S)|^2 while Eloc(S) is still computed with Ψ(S')/Ψ(S), the resulting estimator is not an unbiased estimate of <Φ|H|Φ>/<Φ|Φ>. The manuscript does not state which ratio is used in the SIS runs. Since all final tables (e.g., Table IV) use SIS, this ambiguity affects every reported number and must be clarified.","section":"Sec. II B, Eq. (3)"},{"comment":"In the thermodynamic-limit extrapolation for the Heisenberg model, Table I gives E(∞) values from -0.66940(2) for L~=L to -0.66926(6) for L~=L-8. The spread of 1.4e-4 is about seven times the quoted fitting error of 2e-5, and the L~=L-8, L-10, and L-12 values lie about 1.6e-4 to 1.8e-4 below the QMC value -0.669437. Quoting -0.66940(2) as the extrapolated energy while reporting the other bulk choices as 'all in excellent agreement' does not account for the systematic dependence on the central-bulk choice. The FSS uncertainty should include this spread.","section":"Table I, Fig. 4"},{"comment":"For the frustrated 24x24 system at J2/J1=0.5, the extrapolated PEPS energy -0.49635(5) (or the window [-0.4964,-0.4962]) is compared with DMRG energy -0.4968 and vQMC energy -0.4961. The difference between the PEPS value and the DMRG value is about 4.5e-4, almost an order of magnitude larger than the reported PEPS fitting error of 5e-5, and the mutual spread of the comparison references is about 7e-4. Treating the two references as lower bounds does not remove this discrepancy. The claim of 'excellent agreement' for the frustrated 24x24 case needs a combined uncertainty estimate that includes the spread of the reference methods, or an explicit statement that the agreement is only at the 5e-4 level.","section":"Sec. III A, Fig. 6"}],"minor_comments":[{"comment":"There are numerous typographical errors, including 'sequetially' (Sec. II), 'furstrated' and 'kgaome' (Sec. III), 'persit' (Sec. III A), 'cofigurations' (Sec. II A), 'centall' (Table I caption), and 'Intutively' and 'geomertric' (Sec. IV). A careful proofread is needed.","section":"General"},{"comment":"For the sequential visiting scheme, the acceptance probability in Eq. (9) contains no K-factor analogous to Eq. (8), and the text does not explicitly prove detailed balance for the deterministic bond-by-bond sweep. A brief explanation of why the proposal probabilities cancel (or why detailed balance still holds) would remove ambiguity.","section":"Sec. II A, Eq. (9)"},{"comment":"The optimization update in Eq. (10) uses the sign of the gradient with a random magnitude r·δ(i), but no information is given about the variance of the stochastic gradient estimator or the number of sweeps M used per gradient step beyond the specific value 45000 for 32x32. A short discussion of estimator noise and its effect on the convergence criterion would strengthen the optimization section.","section":"Sec. II B, Eq. (10)"},{"comment":"In Table II the text says 'the energy persit E and spin order converge gradually'; presumably 'persist' is intended. Also, the reported MC sampling errors for spin orders are all of similar size; it would be helpful to state how these errors were estimated (e.g., binning or blocking).","section":"Sec. III A, Table II"}],"recommendation":"major_revision","confidential_remarks":"The algorithm is a worthwhile methodological contribution and the benchmarks are credible. The main issues are fixable in revision: the SIS estimator ambiguity, the missing optimization-convergence diagnostics, and the treatment of systematic spreads in the finite-size extrapolations. My recommendation of major revision is driven by these load-bearing gaps rather than by any suspicion about the numerical results. I would encourage the editor to request a revision that addresses these points explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before the next tensor-network discussion: it delivers the largest well-benchmarked VMC-PEPS calculations to date, reaching 32x32 Heisenberg and 24x24 frustrated J1-J2 with D=8, and the energies match QMC/DMRG to roughly 1e-4 where a direct comparison exists. The core algorithmic novelty, a sequential bond-sweep update for generating Monte Carlo configurations, cuts the per-sweep cost from O(N^2 D^4 Dc^2) to O(N D^4 Dc^2) and is convincingly validated against the older random-flip scheme on 8x8. That is a real practical advance, not just a longer run.\n\nThe strength of the paper is its empirical honesty. The Heisenberg benchmark over L=8 to 32 is systematic, and the 8x28 frustrated stripe comparison against DMRG is a good sanity check. The spin-inversion symmetrization is simple and appears to give a modest, mostly consistent improvement. The authors also admit that their energy and magnetization errors vary randomly with L, possibly from incomplete optimization of the millions of variational parameters. That admission is refreshing, but it is also the main soft spot: the stopping rule is only \"energy decreases very slowly,\" and no gradient-norm or energy-variance diagnostic is reported. For the 24x24 frustrated case, which has no direct benchmark, this leaves real uncertainty about whether the reported energy is the best the D=8 ansatz can do. The finite-size extrapolations across central-bulk choices give a window around [-0.4964, -0.4962] for J2/J1=0.5, but the spread between choices (about 2e-4) is not folded into the quoted errors. That should be fixed.\n\nI disagree with the reader on one minor point: the SIS estimator is stated, though briefly. The bigger reproducibility issue is that no code or data are released, which matters for a method paper. Still, the central claim holds up: the method is validated on multiple sizes and models against independent methods, and the 24x24 result is consistent with DMRG, vQMC, and iPEPS within a few 1e-4. The stress-test worry about local minima is legitimate but does not sink the paper; it just means the largest-system numbers should be treated as strong evidence, not proof.\n\nThis paper deserves a serious referee. I would recommend acceptance after revision, with requests to report optimization diagnostics (e.g., a restart test or gradient-norm history), propagate the FSS spread into quoted thermodynamic-limit errors, and ideally release code or at least detailed convergence data. Who is this for? Anyone building finite-size tensor-network solvers and anyone studying frustrated quantum magnets where QMC fails.","headline":"A genuinely useful finite-PEPS method with real benchmark validation at 32x32; the 24x24 frustrated numbers are encouraging but the optimization and FSS error bars need tightening.","tokens_in":20868,"tokens_out":2588,"would_cite":true,"duration_ms":29043,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["82B80","81-08"],"pacs":[],"model":"deepseek-v4-flash","headline":"Variational Monte Carlo sampling of finite PEPS reproduces quantum Monte Carlo and DMRG results on 32x32 and 24x24 lattices.","keywords":["projected entangled pair states","variational Monte Carlo","finite PEPS","frustrated Heisenberg model","tensor network simulation","stochastic gradient optimization","spin correlations","quantum lattice models"],"falsifier":"Repeat the D=8 optimization for the 32x32 Heisenberg model from several random seeds and step-length schedules while keeping the same boundary conditions; if the lowest energies found differ by more than about 1e-4 per site or systematically miss the quantum Monte Carlo value -0.65633(1), the claim that the method reliably reaches near-exact ground states for large systems is not supported.","tokens_in":19839,"feed_emoji":"🎲","tokens_out":10767,"duration_ms":92927,"temperature":0.7,"pith_summary":"This paper develops a way to simulate two-dimensional quantum lattice models with projected entangled pair states (PEPS), a tensor-network ansatz whose computational cost has usually limited it to small systems. The key move is to treat the PEPS wave function with variational Monte Carlo sampling, generating spin configurations by sequentially flipping antiparallel nearest-neighbor pairs and symmetrizing the final state under global spin inversion. On the square-lattice Heisenberg model the method reaches 32x32 sites with bond dimension D=8 and reproduces quantum Monte Carlo ground-state energies to within about 8e-5 per site; on the frustrated J1-J2 model it reaches 24x24 sites and matches DMRG energies and correlations closely. The authors argue this substantially advances finite PEPS calculations and opens a route to frustrated and translation-symmetry-broken systems beyond the reach of QMC and DMRG.","feed_headline":"Monte Carlo sampling pushes tensor networks to 32x32","feed_subtitle":"Ground-state energies and spin correlations match quantum Monte Carlo and DMRG references.","key_machinery":"The engine is single-layer Monte Carlo sampling of the PEPS amplitude, with bond dimension D (the size of the virtual index connecting neighboring tensors). For a spin configuration $|S\\rangle$ the amplitude $\\Psi(S)$ is a tensor trace over the bond indices; energy and gradients are written as Monte Carlo averages of the local energy $E_{\\text{loc}}(S)$ and of logarithmic derivatives of $\\Psi(S)$. A Metropolis sweep visits every lattice bond sequentially and tries to flip each antiparallel nearest-neighbor spin pair, accepting with probability $\\min(1, |\\Psi(S_b)|^2/|\\Psi(S_a)|^2)$; because only ratios are needed, proposals can be computed with a smaller cutoff $D_{c1}=2D$ while observables use $D_{c2}=3D$, and a boundary-MPS contraction makes the sweep cost $O(ND^4D_c^2)$. The optimization begins from a simple-update imaginary-time-evolved state and proceeds by stochastic gradient descent with random per-element steps and decreasing step lengths. Afterward, observables are reweighted by the spin-inversion-symmetrized amplitude $\\Phi(S)=\\Psi(S)+\\Psi(\\bar S)$, and long-distance correlations are extracted from many Monte Carlo sweeps.","core_discovery":"The paper's central claim is that a variational Monte Carlo treatment of finite projected entangled pair states removes the practical bottlenecks that have kept PEPS simulations to small or infinite-unit-cell settings, making large finite systems both affordable and accurate. Concretely, with bond dimension D=8, a sequentially visited spin-pair Metropolis sweep at reduced contraction cutoff for proposal generation, and a final spin-inversion symmetrization of the wave function, the algorithm reproduces quantum Monte Carlo ground-state energies for the square-lattice Heisenberg model on lattices from 8x8 to 32x32 with energy errors at or below about 8e-5 per site and staggered-magnetization errors about 1e-3. For the frustrated J1-J2 model, where QMC has a sign problem, the D=8 energies on 8x28 stripes agree with DMRG to about 1e-4, and 24x24 simulations at J2/J1=0.5 extrapolate to -0.49635(5), consistent with DMRG, variational QMC, and iPEPS estimates. The authors take the convergence of spin orders with increasing D, including recovery of SU(2) symmetry and x/y isotropy, as evidence that the optimization reaches the correct ground-state subspace, and they argue the approach extends to fermionic systems and to time evolution.","pith_inferences":["Because the acceptance ratio during sampling is only compared with a random number, the cutoff used to generate configurations can be much smaller than the one used for observables; a natural extension the paper does not explore is to push this asymmetry further and correct any residual bias by reweighting, which would cut the dominant sampling cost.","Spin-inversion symmetrization is applied only after optimization; applying the same idea to lattice symmetries (translations, rotations, reflections) during or after optimization could further reduce variance and improve long-distance correlations, at the price of more sampling.","The near-size-independent Monte Carlo convergence suggests the algorithm's scaling is dominated by per-sweep cost and number of gradient steps, so the practical path to, say, 40x40 or D=10 is parallel hardware rather than algorithmic innovation; this is a prediction, not a claim of the paper.","The consistency among several central-bulk choices in the finite-size-scaling analysis implies that open-boundary finite PEPS can serve as a controlled estimator of bulk properties; testing this on the Heisenberg point at even larger system sizes would isolate how much of the residual error comes from incomplete optimization versus the D=8 truncation."],"forward_implications":["Finite PEPS simulations can now reach 32x32 (unfrustrated) and 24x24 (frustrated) with D=8, and because Monte Carlo convergence is nearly size-independent, larger lattices are mainly a matter of computing resources.","For the J1-J2 model at J2/J1=0.5, bulk energies from open-boundary systems show much smaller finite-size effects than DMRG or vQMC, giving an efficient way to estimate thermodynamic-limit energies.","The systematic study from D=4 to D=8 at J2/J1=0.55 shows energy and spin orders converging while SU(2) symmetry and x/y axis isotropy are gradually recovered, which the paper reads as evidence the optimization lands in the correct ground-state subspace.","The method extends directly to fermionic systems (Hubbard/t-J models) via Grassmann-number tensor networks and to time evolution in the t-VMC scheme, both stated as immediate generalizations.","Finite PEPS without a predefined unit cell can describe translation-symmetry-broken phases, incommensurate orders, and trapped cold atoms, and can be compared directly with DMRG on the same finite geometry."],"supporting_citations":[{"why":"Introduces the VMC sampling of tensor network states and the gradient-based optimization this paper builds on.","marker":"[75, 76]"},{"why":"Provides the simple-update imaginary-time evolution used to initialize the PEPS before gradient optimization.","marker":"[58]"},{"why":"Supplies the QMC ground-state energy for the square Heisenberg model used as the thermodynamic-limit benchmark.","marker":"[81]"},{"why":"Supplies finite-size QMC energies and magnetizations for L=8 to 32 used as the accuracy reference.","marker":"[82]"},{"why":"Supplies DMRG energies and spin correlations for the J1-J2 model used to benchmark the frustrated case.","marker":"[87]"},{"why":"Supplies variational QMC plus Lanczos energies for J1-J2 comparison at J2/J1=0.5.","marker":"[88]"},{"why":"Supplies an iPEPS D=8 energy estimate used as an independent comparison at J2/J1=0.5.","marker":"[90]"},{"why":"Earlier VMC-PEPS studies that reached 16x16, the scale this work extends to 32x32.","marker":"[78, 79]"}],"fun_headline_variants":["Monte Carlo sampling makes PEPS accurate for large 2D lattices","Tensor networks go big: VMC enables 32x32 PEPS simulations","VMC-powered tensor networks reach 32x32 finite systems","Variational Monte Carlo boosts PEPS to 32x32 lattices","Accurate 2D tensor network simulation via Monte Carlo sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy claim depends on the assumption that the gradient-based optimization, beginning from the simple-update state, gets near the best variational answer for every lattice size; if optimization stalls in a local minimum for some sizes, the close agreement with reference methods would be partly a coincidence.","fun_headline_variants_meta":{"raw":{"variants":["Monte Carlo sampling makes PEPS accurate for large 2D lattices","Tensor networks go big: VMC enables 32x32 PEPS simulations","VMC-powered tensor networks reach 32x32 finite systems","Variational Monte Carlo boosts PEPS to 32x32 lattices","Accurate 2D tensor network simulation via Monte Carlo sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000582,"raw_usage":{"total_tokens":2747,"prompt_tokens":963,"completion_tokens":1784,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":1690}},"tokens_in":579,"tokens_out":1784,"duration_ms":13921,"temperature":1.0,"reasoning_tokens":1690,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:14:36.837128+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the D=8 optimization for the 32x32 Heisenberg model from several random seeds and step-length schedules while keeping the same boundary conditions; if the lowest energies found differ by more than about 1e-4 per site or systematically miss the quantum Monte Carlo value -0.65633(1), the claim that the method reliably reaches near-exact ground states for large systems is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies DMRG energies and spin correlations for the J1-J2 model used to benchmark the frustrated case."},{"cited_title":"Haghshenas and D","cited_arxiv_id":null,"evidence_quote":"Supplies an iPEPS D=8 energy estimate used as an independent comparison at J2/J1=0.5."}],"review_version":1}