{"id":"df318ba5-8ffd-42fc-9987-c282e64d24c5","arxiv_id":"2603.25492","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A tensor-network RG scheme that preserves lattice rotation/reflection and PT symmetries is formulated and validated on the hard-square lattice gas, yielding improved critical-point and scaling-dimension estimates.","lead":"This paper develops a way to build lattice rotation, reflection, and PT symmetries into tensor-network renormalization group calculations, and tests it on the hard-square lattice gas model. If it works, tensor-network RG can now be applied reliably to a broader class of phase transitions, including ones where the broken symmetry is not a simple on-site spin flip.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Negative-z accuracy rests on per-χ tuning of ε_inv (Table III); without a selection rule, reported 1e-8 errors may be tuning artifacts.","rationale":"The central claim is that a symmetry-preserving TNRG with loop optimization gives accurate critical data. The proposed symmetric-SVD construction is mathematically sound, and the equivalence to TRG truncation is plausible. The main risk is that the entanglement-filtering step, which is essential for accuracy at criticality, is not a controlled approximation. The paper's own disclosure in Sec. V and the ε_inv values in Table III show the negative-z results are obtained with per-χ tuning of the Moore-Penrose inverse. If the loop-optimization update (A4) is sensitive to ε_inv, the reported 1e-8 error at χ = 20 could be a selected outcome rather than a representative accuracy. This is load-bearing because the paper's validation of the central claim rests on these numerical estimates; however, it does not invalidate the symmetry definitions or the symmetric TRG framework. Therefore the verdict remains CONDITIONAL: the method is promising but not yet established as robust. A fixed-ε_inv rerun would settle whether the tuning is essential.","tokens_in":24665,"tokens_out":7321,"duration_ms":77132,"concrete_test":"Rerun the negative-z analysis (Table II) at χ = 10, 12, 16, 20 with a single fixed ε_inv (e.g., 1e-9) and with the hard inverse (A6). Also scan ε_inv around each Table III value by factors of 2 and 0.5. If any χ fails to produce a stable SSB fixed point, or if z_c− estimates shift by more than the claimed relative errors, the reported accuracy is not robust. This isolates whether ε_inv tuning is responsible for the gains.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The weakest point is the EF loop-optimization step for the negative-z transition. Appendix A defines two implementations of the Moore-Penrose inverse: the 'hard inverse' (A6) used for positive z, and a 'regulated inverse' (A7) with regulator ε_inv for negative z. Table III lists ε_inv = 1e-8, 1e-10, 5e-11, 6e-12 for χ = 10, 12, 16, 20, and Sec. V admits performance is 'very sensitive' to ε_inv and that the updating rule 'does not guarantee an increase of the fidelity.' The headline negative-z number, a 4e-8 relative error at χ = 20, is produced after this per-χ tuning. Since no criterion is given for selecting ε_inv a priori, one cannot distinguish an intrinsic accuracy gain of the symmetry-preserving RG map from a favorable hyperparameter search. The non-monotonic errors in Table II (2e-6, 1e-5, 4e-6, 4e-8) reinforce this concern. For the central claim of accurate estimation to hold, the EF step must be a robust, well-defined map, not a tuned run.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a tensor-network renormalization group (TNRG) scheme that explicitly defines and imposes lattice-reflection, lattice-rotation, and PT symmetries on coarse-grained tensor networks, using a symmetric SVD splitting plus loop optimization. It benchmarks the method on the nearest-neighbor hard-square lattice gas, reporting highly accurate estimates of the two critical activities (relative errors down to 7e-7 for z_c^+ at chi=12 and 4e-8 for z_c^- at chi=20) and scaling dimensions consistent with the 2D Ising and Yang-Lee edge universality classes. The paper also provides numerical experiments showing that the degeneracy-index RG flow at the symmetry-broken fixed point is more stable when the relevant symmetry is preserved.","tokens_in":24925,"tokens_out":4091,"duration_ms":47390,"significance":"If the claims hold, this is the first EF-enhanced TNRG that systematically incorporates lattice rotation and PT symmetries, going beyond the well-studied on-site global symmetries. The symmetric SVD splitting and the weak/strong symmetry definitions are clean and potentially reusable for other lattice models. The numerical gains over plain TRG at equal bond dimension are large and consistent for the positive-activity transition. The availability of reproducible Python code strengthens the paper. However, the negative-activity results rely on a hand-tuned regulator in the loop-optimization step, which is a load-bearing caveat for the central claim of accurate estimation at both transitions.","major_comments":[{"comment":"The reported accuracy for z_c^- is produced by per-bond-dimension tuning of epsilon_inv in Eq. (A7). Table III lists epsilon_inv = 1e-8, 1e-10, 5e-11, 6e-12 for chi = 10,12,16,20, and Sec. V admits performance is 'very sensitive' to this regulator. Since no a priori selection rule is given, the 4e-8 error at chi=20 cannot be cleanly attributed to the symmetry-preserving RG map; it may be a favorable hyperparameter choice. The non-monotonic errors in Table II (2e-6, 1e-5, 4e-6, 4e-8) reinforce this concern. Please provide a principled criterion for epsilon_inv or a sensitivity analysis showing a plateau around the chosen values.","section":"Sec. IV B, Appendix A, Table III"},{"comment":"The loop-optimization update rule does not guarantee an increase of fidelity because Q and Upsilon in Eq. (A3) depend on v-tilde_Lambda; the convex-combination trick in Eq. (A5) only prevents a decrease for a fixed environment, not under the true nonlinear map. This is acknowledged in Sec. V ('does not guarantee an increase of the fidelity'). Since the loop optimization is the essential EF step that tames RG errors at criticality (Figs. 6 and 8), the paper should present stronger convergence diagnostics and, in particular, show that the negative-z results are not artifacts of a favorable initialization and regulator. A robustness test across random initializations or a demonstration that the fidelity actually increases for the reported runs would address this.","section":"Appendix A, Eqs. (A4)-(A7); Sec. V"},{"comment":"The loop approximation introduces an uncontrolled error: the optimized v-tilde_Lambda no longer satisfies the strong lattice symmetry of Eq. (22), and the paper does not bound the deviation. The claim that this is the only approximation in the first half of the RG step is formally true but does not quantify how much the loop optimization distorts the critical tensor. Given that the accuracy gain over TRG is the main evidence for the method, a quantitative comparison of the optimized tensor with the SVD initialization (e.g., fidelity curves versus RG step) would make the claim more convincing.","section":"Sec. III D and Eq. (28)"}],"minor_comments":[{"comment":"The errors for the proposed method are not monotonic in chi (e.g., z_c^+ error worsens from chi=12 to chi=18 before improving at chi=20; z_c^- error worsens from chi=10 to chi=12). The text says 'there is a clear trend of improvement' but the non-monotonicity deserves a comment, especially in relation to the regulator sensitivity.","section":"Tables I-II"},{"comment":"The scaling-dimension estimates are only presented graphically. Since these are central quantitative outputs, a table with numerical values and their convergence in RG step would be more useful to readers and reviewers.","section":"Figs. 7 and 9"},{"comment":"The notation 'TNRG' is used inconsistently; sometimes it refers to the general method and other times to the specific TRG-based scheme. This can confuse readers comparing with the literature.","section":"Throughout"},{"comment":"The statement that 'no EF-enhanced TNRG scheme is suitable for studying a phase transition where some lattice symmetries are spontaneously broken' should be moderated; Ref. [17] treats lattice-reflection symmetry with EF, and the distinction is more about rotation/reflection combinations and PT symmetry. A sentence clarifying the incremental contribution would help.","section":"Sec. V"}],"recommendation":"major_revision","confidential_remarks":"The positive-activity part of the paper is solid and presents a useful methodological advance. My main concern is the negative-activity transition: the headline accuracy depends on a per-chi tuned regulator, and the paper itself admits the optimizer is fragile. I do not think this is grounds for rejection, because the issue may be fixable with a principled regulator selection or a sensitivity analysis, or by reframing the claims for the negative-z case as a benchmark of a tuned EF step. Please ask the authors to address the regulator criterion and the robustness of the loop-optimization convergence before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Xinliang, here's my read of Lyu's TNRG paper. The genuinely new piece is the symmetric SVD splitting that carries the sign of eigenvalues into a diagonal bond matrix, plus the rotation trick that lets the scheme impose rotation and reflection symmetries while keeping tensors real. That is a real step beyond the symmetric loop-TNR in Ref. [28], and the derivation in Sec. III is clean. The demonstration that HOTRG flows away from the SSB fixed point while TRG stays stable is physically instructive.\n\nThe positive-activity results are the strongest part of the paper. The loop optimization for that transition uses a 'hard inverse' with no free regulator, and the reported errors (7e-7 at chi=12) and the tamed RG error growth are credible. The improvements over plain TRG are large and consistent.\n\nNow the soft spot. For the negative-activity transition, the loop optimization switches to a 'regulated inverse' with epsilon_inv, and Table III shows you need a different value at every bond dimension: 1e-8, 1e-10, 5e-11, 6e-12. The paper itself says the method is 'very sensitive' to it and that the updating rule 'does not guarantee an increase of the fidelity.' The headline negative-z number, 4e-8 at chi=20, is produced after that per-chi tuning. With no a priori selection criterion, you cannot tell whether the accuracy is intrinsic to the RG map or a favorable hyperparameter search. The non-monotonic errors in Table II (2e-6, 1e-5, 4e-6, 4e-8) underline that. This is a load-bearing issue for that section, not a cosmetic one.\n\nThere are smaller issues: no error bars, the code is referenced without a commit hash, and the claimed equivalence to standard TRG is only sketched. But those are minor next to the epsilon_inv question.\n\nThe author is honest about the rudimentary state of the optimization, which I credit. The central method is new and the math is sound; the positive-z results stand on their own. The negative-z claims need either a principled way to select epsilon_inv or a robustness study showing the accuracy doesn't depend on fine-tuning. Who should read this: anyone doing TNRG with lattice symmetries, or studying hard-core lattice gases. It deserves a serious referee, but the referee should push for the epsilon_inv selection rule and a rerun with error bars. I'd accept this to peer review with major revision.","headline":"A genuinely new symmetry-preserving TNRG scheme with a clean derivation and strong positive-z results, but the negative-z accuracy rests on per-χ tuning of ε_inv and shouldn't be sold as intrinsic.","tokens_in":25420,"tokens_out":2313,"would_cite":false,"duration_ms":23531,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A tensor-network renormalization group that keeps lattice reflection, rotation, and PT symmetries explicit at every coarse-graining step can locate both phase transitions of the nearest-neighbor hard-square lattice gas with relative errors","keywords":["hard-square lattice gas","tensor-network renormalization group","lattice symmetry","PT symmetry","spontaneous symmetry breaking","loop optimization","Yang-Lee edge singularity","critical exponents"],"falsifier":"Run the published implementation at the negative-activity critical point with bond dimension χ=20 but replace the tuned regulator ϵ_inv=6×10^-12 by 1×10^-8, keeping everything else fixed. If the estimated z_c^- does not degrade from 4×10^-8 toward the plain-TRG error of 2×10^-5, the paper's stated sensitivity to ϵ_inv—and its implication that the reported accuracy depends on hand-tuning—is falsified. If it does degrade, that confirms the load-bearing role of the loop-optimization regulator.","tokens_in":24488,"feed_emoji":"🎯","tokens_out":8051,"duration_ms":85205,"temperature":0.7,"pith_summary":"This paper shows how to build a tensor-network renormalization group (TNRG) that keeps the lattice-reflection, lattice-rotation, and parity-time (PT) symmetries of a statistical model explicit at every coarse-graining step, rather than letting numerical approximations break them. Using the nearest-neighbor hard-square lattice gas as a benchmark, it defines what these symmetries mean for a coarse-grained tensor network and constructs a scheme that both preserves and imposes them through a symmetric singular-value decomposition and a loop-optimization entanglement filter. The payoff is concrete: the critical activities of both the Ising-type ordering transition at positive activity and the Yang-Lee edge singularity at negative activity are estimated with relative errors as small as 7e-7 and 4e-8, roughly two orders of magnitude better than plain TRG at the same bond dimension, and scaling dimensions stop drifting with RG step. The central insight is that the sign information carried by bond matrices—which ordinary tensor renormalization absorbs away or ignores—is exactly what must be tracked to keep PT and lattice symmetries alive through coarse graining.","feed_headline":"Preserving lattice and PT symmetries sharpens tensor RG 100-fold","feed_subtitle":"Hard-square critical activities land at 7e-7 and 4e-8 relative error, with stable scaling dimensions.","key_machinery":"The central object is the symmetric SVD splitting: the 4-leg tensor A is viewed as a symmetric matrix across the diagonal l1 axis and decomposed by truncated eigendecomposition; the signs of the eigenvalues become a new diagonal bond matrix σ′, while their absolute values are split evenly into two 3-leg tensors vΛ. Because the signs are carried by σ′ rather than absorbed into orthogonal matrices, the PT symmetry (reality of all tensors) and the lattice reflection and rotation symmetries remain explicit. The loop optimization then varies vΛ to remove the corner-double-line entanglement while keeping A, σ, and σ′ fixed; the weak form of lattice symmetry survives, and the strong form is re-impo","core_discovery":"The paper's central claim is that lattice-reflection, lattice-rotation, and PT symmetries can be incorporated into a two-dimensional tensor-network renormalization group in a way that is both exact and computationally useful. The key is to work with a coarse-grained tensor network of the form in Eq. (16), in which two tensors A and B sit on alternating sublattices joined by diagonal ±1 bond matrices σ. Lattice symmetries are expressed in a weak form, where B is determined by rotations and reflections of A, and a strong form, where A itself carries a diagonal SWAP-gauge matrix g. The symmetric SVD splitting replaces ordinary SVD by an eigendecomposition of the tensor treated as a matrix acros","pith_inferences":["Editorial inference: the same diagonal sign-and-gauge bookkeeping could encode other discrete symmetries, such as glide or point-group operations, by choosing different diagonal matrices on the bonds, making symmetry-preserving TNRG available for models whose order parameters break those symmetries.","Editorial inference: because the method works at the Yang-Lee edge despite the failure of the standard entanglement-entropy argument, it suggests that entanglement filtering in TNRG can be effective at non-unitary fixed points for reasons unrelated to area-law entanglement; testing it at another non-unitary fixed point would show whether this is generic.","Editorial inference: the hand-tuned regulator for the negative-activity transition means the method is not yet turnkey; a principled way to set or eliminate that regulator, for instance from the spectrum of the environment tensor, is the difference between a demonstration and a robust numerical tool.","Editorial inference: the weak-form lattice symmetry requirement may be enough for other tensor-network algorithms that update tensors sequentially, pointing toward a route for making entanglement filtering symmetry-aware beyond the specific TRG-based scheme proposed here."],"forward_implications":["Any model whose tensor network can be cast in the two-sublattice form with real tensors and diagonal ±1 bond matrices can use this RG without modifying the core algorithm; the paper singles out hard-core lattice gases with longer exclusion ranges as the immediate next step.","The sign-tracking bond matrix makes the method applicable to non-unitary critical points such as the Yang-Lee edge, where the partition function can become negative; this extends TNRG beyond the usual positive Boltzmann weight setting.","Because the strong form of lattice symmetry is imposed at the end of every RG step, RG-relevant perturbations from algorithm artifacts are eliminated, so the spontaneous-symmetry-breaking fixed points become strictly stable and critical tensors can be found reliably by a bisection method.","The truncation error at criticality stops growing and converges below 1e-6 rather than growing to 1e-2, so scaling dimensions extracted from near-fixed-point tensors are stable with respect to RG step rather than drifting."],"fun_headline_variants":["Lattice and PT symmetries refine tensor-network RG","Symmetry-enriched tensor RG improves critical estimates","Tensor RG with lattice symmetries: a hard-square test","PT-symmetric tensor networks sharpen phase transition data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the loop-optimization update, in which an inverse of a certain environment matrix is used to improve the split tensors, can be made to reliably remove redundant short-range entanglement without losing critical information; the paper itself cautions that this update is rudimentary, does not guarantee improved fidelity, and works for the negative-activity transition only with carefully hand-tuned regulator values.","fun_headline_variants_meta":{"raw":{"variants":["Lattice and PT symmetries refine tensor-network RG","Symmetry-enriched tensor RG improves critical estimates","Tensor RG with lattice symmetries: a hard-square test","PT-symmetric tensor networks sharpen phase transition data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1240,"prompt_tokens":768,"completion_tokens":472,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":410}},"tokens_in":512,"tokens_out":472,"duration_ms":5589,"temperature":1.0,"reasoning_tokens":410,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T17:22:05.266913+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the published implementation at the negative-activity critical point with bond dimension χ=20 but replace the tuned regulator ϵ_inv=6×10^-12 by 1×10^-8, keeping everything else fixed. If the estimated z_c^- does not degrade from 4×10^-8 toward the plain-TRG error of 2×10^-5, the paper's stated sensitivity to ϵ_inv—and its implication that the reported accuracy depends on hand-tuning—is falsified. If it does degrade, that confirms the load-bearing role of the loop-optimization regulator.","supporting_citations":[],"review_version":1}