{"id":"e9090e60-af90-4ef7-ae01-13701d336fb4","arxiv_id":"1908.08833","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A DMRG-style PEPS optimizer that builds an approximate canonical form and solves regular eigenvalue problems is tested on the 2D Heisenberg model, converging close to quantum Monte Carlo energies within about ten sweeps.","lead":"This paper implements a density-matrix-renormalization-group style algorithm for optimizing two-dimensional tensor network states (PEPS), combining canonical-form construction, regular eigenvalue solves, and large local updates. It demonstrates the method on the square-lattice Heisenberg model, reaching energies close to quantum Monte Carlo results.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Approximate column canonization (fidelity only >0.99) means the 'regular eigenvalue' solve optimizes a distorted PEPS energy, so the DMRG-equivalence claim needs a controlled convergence test.","rationale":"The most load-bearing assumption is not the environment bond dimension χ=12 or the absence of code, but the fidelity threshold of the column canonization. The entire DMRG analogy rests on the local update being a regular eigenvalue problem in an orthonormal basis, which is only true if M can be replaced by Q R without changing the state. The paper explicitly concedes QR differs from M and that 'occasional increases in energy' result. That concession is important: a DMRG sweep in MPS never increases the energy if the eigensolver is run to the ground state of the effective Hamiltonian, because the canonical form is exact. Here an energy increase means the algorithm is not minimizing a single fixed energy functional; the objective changes whenever R is folded into the next column. The positive comparison to QMC shows the approximation is often small, but it does not demonstrate that the regular eigenvalue solve is acting on the true PEPS energy, which is the paper's claimed mechanism. My proposed test—tightening the fidelity threshold and doubling the Q/R bond dimensions—would directly quantify how much the 0.99 canonization error distorts the optimized energy. If the result is insensitive, the concern is resolved and the paper's claim is substantially supported; if sensitive, the method should be presented as an approximate environment scheme rather than a faithful DMRG generalization. This matches the reader's CONDITIONAL verdict, so I recommend no change.","tokens_in":7542,"tokens_out":6939,"duration_ms":75793,"concrete_test":"Rerun the 10x10, D=6 Heisenberg calculation with the canonization fidelity threshold raised from 0.99 to 0.999 and 0.9999, and with the internal bond dimensions of Q and R doubled, recording per-sweep and final energies. If the final energy shifts by more than the inset gap to QMC, or if the acknowledged 'occasional increases in energy' persist at the higher threshold, then the 0.99-fidelity column replacement is a dominant uncontrolled error and the advertised regular-eigenvalue/variational mechanism is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the three DMRG features—canonical/orthonormal basis, regular eigenvalue solves, and large stable updates—carry over to PEPS. The linchpin is the column canonization in the 'Canonization of PEPS' section: a column M is replaced by Q R with Q exactly unitary, but the text admits 'the product of Q times R generally differs from the original, input column by a small amount' and the stopping criterion is Tr[M†QR]/Tr[M†M] > 0.99. When R is folded into the neighboring column, the wavefunction is changed. The subsequent local solve is a regular eigenvalue problem Hv = λv for that modified state, not for the true PEPS energy, and the energy can increase—the paper explicitly attributes 'occasional increases in energy' to canonization failure. Therefore the method does not inherit DMRG's variational guarantee: updates are not guaranteed to lower a single global energy functional. The reported agreement with QMC is encouraging evidence that the approximation is small in practice, but no data are given on the actual canonization error per sweep or on the sensitivity of the final energy to the 0.99 threshold. Without such control, the central mechanism claimed—regular eigenvalue optimization in an orthonormal basis—is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a finite-PEPS ground-state optimization algorithm that directly adapts the one-site DMRG workflow. After optimizing a column, the column is approximately canonized into a unitary column Q and an MPO remainder R; R is absorbed into the neighboring column, and additional within-column canonization is used so that each local tensor update is formulated as a regular eigenvalue problem solved with Davidson. Sweeps over all columns are repeated until convergence. The method is tested on the square-lattice spin-1/2 Heisenberg model with open boundaries for L = 4, 6, 8, 10, reporting energies versus PEPS bond dimension D and versus sweep number, compared with QMC energies, and nearest-neighbor spin-spin correlators compared with MPS-DMRG. The authors claim that the three features that make MPS-DMRG powerful---enforcing a canonical/orthonormal basis, solving regular eigenvalue problems with iterative solvers, and making large stable updates---are realized for PEPS.","tokens_in":7828,"tokens_out":4342,"duration_ms":45485,"significance":"If the method is validated, it would provide a competitive variational finite-PEPS approach that inherits the favorable convergence and stability properties of DMRG, and it would be a useful tool for finite-size two-dimensional systems. The paper's strengths are the clean benchmark setting (Heisenberg model compared with essentially exact QMC and MPS-DMRG references), the clear outline of the algorithmic pipeline, and the availability of an ITensor-based implementation. However, the approximate column canonization lies at the heart of the method, and the manuscript does not supply enough numerical control over this approximation to establish that the regular eigenvalue solve optimizes the true PEPS energy. With additional convergence data the central claim would be credible; as written it is promising but not fully demonstrated.","major_comments":[{"comment":"The local update is a regular eigenvalue problem H v = lambda v only for the canonized, approximate wavefunction: the text states that 'the product of Q times R generally differs from the original, input column by a small amount' and that the stopping criterion is a fidelity Tr[M^dag Q R]/Tr[M^dag M] of only > 0.99. Because R is folded into the neighboring column, the effective Hamiltonian solved at each step differs from the Hamiltonian in the true PEPS basis, so DMRG's variational guarantee and the claimed property of 'representing the Hamiltonian within an orthonormal basis during the entire calculation' are not inherited. The paper also attributes 'occasional increases in energy' to this same approximation, which directly weakens the claim of large, stable updates. Please report the actual canonization fidelity achieved per sweep rather than only the >0.99 threshold, show convergence of the final energy as the allowed canonization error is reduced, and provide evidence that the energy is controlled by the approximation error. Without this, the central mechanism claimed in the abstract and introduction is not established.","section":"Canonization of PEPS (Fig. 3)"},{"comment":"No environment bond dimension chi is reported for the energy-versus-D curves in Fig. 4, and Fig. 5 uses chi = 12 without a convergence study. The approximate boundary-MPS contraction is an independent source of bias, so the agreement with QMC in Fig. 4 could depend on chi as well as D. Please provide energies versus chi for at least the largest D shown and for each system size, or otherwise demonstrate that the reported numbers are converged in chi.","section":"Application to the Heisenberg Model (Fig. 4 and Fig. 5)"},{"comment":"The reproducibility and verification of the canonization step depend on details placed in the Supplementary Material: the full canonization algorithm, the MPO-MPS contraction procedure, and the intra-column canonization. As written, the main text does not specify the internal bond dimensions of Q and R used in the canonization, the number of canonization passes, or the contraction parameters, so the reader cannot assess whether the reported 'small amount' of canonization error is typical or contingent. Please include these parameters in the main text or in a clearly labeled table, and report the actually achieved fidelity for the calculations shown in Figs. 4 and 5.","section":"Canonization of PEPS and Supplemental Material"}],"minor_comments":[{"comment":"The section heading reads 'Projected Pair Entangled States'; this should be 'Projected Entangled Pair States'.","section":"Projected Entangled Pair States"},{"comment":"The phrase 'A promising ways to break through this scaling issue' is grammatically incorrect and should be 'A promising way'.","section":"Introduction"},{"comment":"The correlator comparison is shown only for L = 4; please state whether similar accuracy holds for L = 6, 8, 10, or clarify that the correlator benchmark was only performed for the smallest system.","section":"Application to the Heisenberg Model (Fig. 5)"},{"comment":"The number of initial simple-update sweeps is not specified, although the text says the optimization starts with 'simple update sweeps of the system for only the first several sweeps.' Please state this number and any associated parameters, since it is part of the optimization protocol.","section":"Application to the Heisenberg Model"},{"comment":"The bottom subfigure appears to contain several curves with labels 3, 4, 5, 6, but the text says the lower plot is for D = 6; please clarify which curves correspond to which bond dimension, including the QMC reference.","section":"Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The central claim is plausible and the benchmark results are encouraging, but the approximate canonization is the load-bearing element and is currently controlled only by a 0.99-fidelity threshold. The requested additional data (actual per-sweep fidelity, sensitivity to the fidelity threshold, and chi convergence) are straightforward to obtain and would substantially increase confidence in the method. I also note that the relationship to the recently proposed canonical PEPS constructions of Refs. [25,26] should be made explicit; the paper claims a different canonization route but does not compare its accuracy or cost."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a genuine step forward for finite PEPS optimization. The paper assembles all three DMRG ingredients—canonical form, regular eigenvalue solves with an iterative solver, and large single-tensor updates—into one working algorithm, and shows it on the square-lattice Heisenberg model up to 10x10. The energies approach QMC values as D grows, and the sweeps converge in about ten passes. That is real evidence the method works, not just a proposal. They also take a different canonization route from Zaletel-Pollmann and Haghshenas et al., which is worth spelling out.\n\nWhere it gets soft is the canonization. The column replacement Q R is only approximately equal to the original column, with fidelity >0.99, and the paper plainly says the energy sometimes rises when the unitary approximation is poor. That matters: the local Davidson solve is a regular eigenvalue problem for the modified state, not for the true PEPS energy, so the method does not inherit DMRG's variational guarantee. The stress-test concern lands. It is not fatal—the QMC comparison suggests the distortion is small in practice—but the paper should give the canonization error per sweep and show the final energy is insensitive to the 0.99 threshold. Without that, the 'DMRG-equivalence' framing is stronger than what is demonstrated. To their credit, the authors explicitly own the energy increases and the approximate nature of the canonization.\n\nOther soft spots are more routine: chi is fixed at 12 and not converged, some essential details live only in the supplementary material, no code is released, and the correlators at D=6 visibly deviate from exact DMRG. Those are the kinds of things a referee should ask for, not reasons to reject. The citation pattern is fine; the benchmarks are external QMC/MPS-DMRG, so there is no fitting-to-target circularity.\n\nWho is this for? People who build or use PEPS algorithms. It is a methods paper that advances the toolkit even if it does not settle the best canonization scheme. I would bring it to a reading group and cite it. It deserves peer review. My recommendation: send it out, ask for the canonization-error characterization, code/data release, and a chi-convergence check, then accept if those come through.","headline":"The paper delivers a working DMRG-style PEPS optimizer, but the approximate column canonization (fidelity >0.99) weakens the variational guarantee; the numerics are good enough to warrant a serious referee.","tokens_in":8359,"tokens_out":3114,"would_cite":true,"duration_ms":27786,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that DMRG, the algorithm that made one-dimensional tensor networks practical, carries over to two-dimensional PEPS by canonizing one column at a time and solving a regular eigenvalue problem for each local tensor.","keywords":["tensor networks","PEPS","DMRG","canonical form","variational optimization","iterative eigensolver","Heisenberg model","projected entangled pair states"],"falsifier":"Run the same Heisenberg calculation on a larger or more frustrated lattice while recording, for every column, the canonization fidelity $\\mathrm{Tr}[M^\\dagger Q R]/\\mathrm{Tr}[M^\\dagger M]$ and the energy after every sweep; if any optimized column has fidelity below $0.99$ or if the energy rises persistently instead of converging within about ten sweeps, the regular-eigenvalue form of the algorithm is not delivering the claimed stable convergence.","tokens_in":7309,"feed_emoji":"🧮","tokens_out":6174,"duration_ms":61180,"temperature":0.7,"pith_summary":"This paper tries to establish that the density matrix renormalization group (DMRG) algorithm, the workhorse for optimizing one-dimensional tensor networks, can be transplanted successfully to projected entangled pair states (PEPS), the natural tensor-network ansatz for two-dimensional quantum systems. The authors argue that all three features that make DMRG powerful, keeping an orthonormal basis via a canonical form, solving a regular eigenvalue problem at each local update with an iterative eigensolver, and allowing large stable tensor updates, can be implemented for PEPS. If correct, this gives a variational method for finite two-dimensional systems that converges in about ten sweeps of the lattice, demonstrated on the square-lattice Heisenberg model. A sympathetic reader would care because it extends a mature, reliable optimization technique from one to two dimensions without the exponential bond-dimension growth that MPS approaches suffer on wide ladders.","feed_headline":"DMRG's core steps now work on 2D tensor networks","feed_subtitle":"Canonizing PEPS one column at a time turns the energy problem into fast, stable eigenvalue solves.","key_machinery":"The central object is an approximate column canonization for PEPS. A column $M$ of the tensor network is decomposed into $Q R$, where $Q$ obeys the exact canonical condition and $R$ is an MPO-like remainder folded into the next column. The decomposition is found by iterating polar decompositions of environment tensors until the fidelity $\\mathrm{Tr}[M^\\dagger Q R]/\\mathrm{Tr}[M^\\dagger M]$ exceeds $0.99$. A second, intra-column canonization using SVD or QR makes the environments above and below the active tensor orthonormal, so the effective normalization matrix becomes the identity. This machinery converts the generalized eigenvalue problem for a PEPS update into a regular eigenvalue problem, which iterative solvers like Davidson can handle quickly and stably, allowing large tensor updates without losing numerical control.","core_discovery":"The central claim is that there is no obstacle in principle to a PEPS optimization with all of DMRG's technical machinery. Working on finite PEPS with open boundary conditions, the authors canonize the network one column at a time: each optimized column $M$ is factored into a unitary column $Q$, which carries the physical indices and stays in the network, and a non-unitary remainder $R$, which is multiplied into the neighboring column. A second, intra-column canonization then imposes orthogonality above and below the tensor being updated. With these two canonization steps, the normalization matrix $\\hat N$ for the environment becomes the identity to good approximation, so the local energy minimization becomes a regular eigenvalue problem $\\hat H v = \\lambda v$ solved by a Davidson iteration, exactly as in MPS-DMRG. On the square-lattice spin-$1/2$ Heisenberg model, the resulting algorithm converges in roughly ten sweeps, reproduces nearest-neighbor correlators at accuracy comparable to state-of-the-art infinite-PEPS methods, and shows residual errors controlled by the bond dimension $D$.","pith_inferences":["If canonization fidelity can be pushed reliably above $0.99$, the occasional energy increases the paper reports should disappear, and the convergence in sweeps might approach the exponential rate that DMRG achieves for gapped one-dimensional systems.","A two-site variant that dynamically grows the bond dimension, which the paper names as future work, would remove the need to guess $D$ in advance and should also be better at escaping local minima; the one-site version presented here freezes the bond dimension throughout.","Because the algorithm works on finite open-boundary systems, it could be paired with finite-size scaling to extract bulk quantities; the paper does not report such extrapolated values, but its boundary-pinned Heisenberg data would support that analysis.","A head-to-head comparison of this finite PEPS-DMRG method with translation-invariant iPEPS on the same bulk observable would settle whether the finite-size, DMRG-style approach can compete with infinite-system optimization; the paper flags this comparison as an open question."],"forward_implications":["Finite PEPS ground states can be optimized with all three DMRG advantages: an orthonormal background from canonization, a regular eigenvalue problem at each site, and large tensor updates without losing stability.","For the square-lattice Heisenberg model up to $10 \\times 10$, the PEPS-DMRG energy converges to within about $10^{-2}$ of the converged value within roughly ten sweeps, with the remaining error controlled by the bond dimension $D$.","Unlike MPS-DMRG on ladders, where the required bond dimension grows exponentially with the ladder width, PEPS-DMRG uses a modest bond dimension independent of system size for fixed accuracy.","Nearest-neighbor spin-spin correlators from PEPS-DMRG match essentially exact MPS-DMRG results to an accuracy comparable to state-of-the-art iPEPS calculations at similar bond dimension.","The algorithm's runtime is dominated by tensor contractions and MPO-MPO multiplications, identifying concrete targets for performance optimization in future implementations."],"supporting_citations":[{"why":"Introduces the DMRG algorithm whose three key technical features are the template being translated to PEPS.","marker":"[1]"},{"why":"Defines projected entangled pair states as the two-dimensional generalization of MPS, the ansatz optimized throughout the paper.","marker":"[21]"},{"why":"Supplies the standard iPEPS imaginary-time evolution method and the boundary-MPS contraction technique used to evaluate observables.","marker":"[22]"},{"why":"Existing PEPS optimization by gradient or conjugate-gradient descent, representing one of the methods lacking DMRG's canonical-form structure.","marker":"[23]"},{"why":"Variational PEPS optimization with a generalized eigenvalue solver, the approach the paper improves on by making the local problem regular.","marker":"[24]"},{"why":"One of the recent PEPS canonical-form constructions the paper contrasts with its own column canonization approach.","marker":"[25]"},{"why":"Another recent PEPS canonical-form construction, providing alternative canonization methods and framing the open question of which approach is best.","marker":"[26]"}],"fun_headline_variants":["DMRG's core algorithm now optimizes 2D PEPS networks","Canonizing PEPS columns enables stable DMRG-style optimization","2D tensor network optimization gets all DMRG steps","PEPS optimization via column canonization and eigenvalue solves"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the approximate column canonization being accurate enough, fidelity above $0.99$ in practice, that replacing a PEPS column by its unitary factor $Q$ times the remainder $R$ only distorts the Hamiltonian slightly; if that approximation fails, the regular eigenvalue solve can minimize a distorted energy and the variational guarantee is lost.","fun_headline_variants_meta":{"raw":{"variants":["DMRG's core algorithm now optimizes 2D PEPS networks","Canonizing PEPS columns enables stable DMRG-style optimization","2D tensor network optimization gets all DMRG steps","PEPS optimization via column canonization and eigenvalue solves"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00053,"raw_usage":{"total_tokens":2506,"prompt_tokens":852,"completion_tokens":1654,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":1582}},"tokens_in":468,"tokens_out":1654,"duration_ms":11094,"temperature":1.0,"reasoning_tokens":1582,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:29:00.840707+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same Heisenberg calculation on a larger or more frustrated lattice while recording, for every column, the canonization fidelity $\\mathrm{Tr}[M^\\dagger Q R]/\\mathrm{Tr}[M^\\dagger M]$ and the energy after every sweep; if any optimized column has fidelity below $0.99$ or if the energy rises persistently instead of converging within about ten sweeps, the regular-eigenvalue form of the algorithm is not delivering the claimed stable convergence.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"One of the recent PEPS canonical-form constructions the paper contrasts with its own column canonization approach."},{"cited_title":"Haghshenas, M","cited_arxiv_id":null,"evidence_quote":"Another recent PEPS canonical-form construction, providing alternative canonization methods and framing the open question of which approach is best."}],"review_version":1}