{"id":"4cbebbb1-14a1-480e-9995-0ffd047aa684","arxiv_id":"2412.04133","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Tensor-network DMRG simulations of SU(2) and small-SU(N) bosonic and supersymmetric matrix models give convergent ground states and entanglement measures, with costs that appear to grow polynomially with the number of matrices.","lead":"This paper applies matrix product states and the density matrix renormalization group to simulate quantum mechanical matrix models, including bosonic and supersymmetric toy models. It reports polynomial computational cost and uses the method to extract ground-state energies and entanglement properties.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SU(N) scalability claim rests on unconverged DMRG: for N≥3 the target truncation error is never reached, so the asserted polynomial scaling in N is unsupported.","rationale":"I read the paper in good faith and find the SU(2) convergence studies internally consistent: the truncation error is controlled at 10^-10 to 10^-11, δE(Λ) decays roughly exponentially, and ⟨G²⟩ decays with Λ, all of which support the feasibility of MPS/DMRG for the SU(2) bosonic and N=2 SUSY models. The entanglement analysis is plausible and the MPO bond-dimension study in App. E gives useful layout guidance. The load-bearing weakness is not the method per se but the extrapolation from the SU(2) regime to the claimed general scalability. For N=3,4,5 the authors explicitly cannot reach their target truncation error and resort to heuristic 1/D fits; for N=6 they give a single rough point. This is a stated limitation in Sec. 4.3 and, together with the Sec. 4.1 disclaimer that only MPO bond dimension was analyzed, it means the central claim that computational cost scales polynomially with matrix size N is not supported by the presented evidence. I therefore agree with the reader's conditional verdict, but I would place slightly more emphasis on the fact that the missing quantity is D_req(N), the bond dimension needed for fixed accuracy, rather than on the extrapolated energy values themselves: even perfect agreement with Ref. [14] would not establish polynomial scaling. The concrete test I propose directly measures the scaling quantity that would settle the concern. I do not see an internal inconsistency in the SU(2) numerics, and I am not relying on disagreement with any external consensus; the issue is an unverified extrapolation inside the paper's own error-control framework.","tokens_in":22810,"tokens_out":4107,"duration_ms":44812,"concrete_test":"For the two-matrix bosonic SU(N) model at fixed Λ and g^2N=1, determine D_req(N) for a fixed truncation error (e.g., TE=10^-8, likely reachable for N=3,4) by running DMRG at successively larger bond dimensions until the estimated truncation error is below the target, and then repeat the current 1/D extrapolation for N=3,4,5 at D=600 and D=800. If D_req(N) grows faster than a low-degree polynomial in N, the Sec. 5 polynomial-scaling claim fails. If the extrapolated E0/N^2 shifts by more than the quoted error bars when the fit window or fit order changes, the 1/D extrapolation is biased and the reported SU(N) numbers are unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Sec. 5 is that the computational cost 'scales polynomially' with the number and size of matrices. The evidence for scaling with the number of matrices is an SU(2) saturation of Dmax (Fig. 2), which is credible. But the evidence for scaling with N is missing. Sec. 4.3 states that 'the available bond dimension was insufficient to reach the fixed truncation error of 10^-10 except for N = 2'; for N = 3,4,5 the paper uses 1/D extrapolations of ground-state observables, and for N = 6 only a rough fixed-D=200 estimate is given. Thus the bond dimension required to reach a fixed accuracy as a function of N is never measured. The D→∞ extrapolation of E0/N^2 cannot certify polynomial cost: a linear/quadratic/cubic fit in 1/D over D=100-400 (or 150-350 for N=5) is consistent with many functional forms, and the quoted error bars are the difference between two low-order fits, not a controlled truncation-error estimate. This is exactly the regime where the scalability claim must be tested: if D_req(N) grows faster than polynomially, the method is not scalable even if the extrapolated energies happen to agree with Ref. [14]. The paper itself flags the missing full cost analysis in Sec. 4.1 ('we only focus on the MPO bond dimension analysis') and the inability to converge N≥3 in Sec. 4.3; both stated limitations undercut the headline conclusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper applies matrix product states and the density matrix renormalization group to bosonic and supersymmetric matrix models. The authors map matrix degrees of freedom onto one-dimensional chains, compare several layout schemes through the MPO bond dimension (App. E), and compute ground states for SU(2) bosonic models with up to nine matrices, for SU(N) bosonic models with N=3,...,6, and for a minimal supersymmetric SU(2) model. They report exponential convergence in the Fock-space cutoff Lambda for SU(2), saturating bond dimensions as the number of matrices grows, and extrapolated energies for larger N that agree qualitatively with Ref. [14]. The paper concludes that tensor networks are for the first time demonstrated to be applicable to matrix models, with computational cost scaling polynomially in the number and size of matrices.","tokens_in":23158,"tokens_out":5592,"duration_ms":49329,"significance":"The SU(2) results are a genuine and useful proof of principle: the energy and gauge-violation converge exponentially with Lambda, the required bond dimension saturates with the number of matrices, and the entanglement observables are computed cleanly. The layout-scheme comparison in App. E is also a practical contribution. The broader claim that the method scales polynomially with N is not established: for N>=3 the target truncation error is never reached, the 1/D extrapolations are uncontrolled, and the external benchmark is not independent. As it stands, the paper supports a feasibility statement for SU(2)-type models and an exploratory statement for SU(N); the conclusion overstates the evidence.","major_comments":[{"comment":"The SU(N) results for N=3,4,5 are obtained from 1/D extrapolations over D=100-400 (N=3,4) or D=150-350 (N=5), with error bars defined as the difference between quadratic and cubic (or linear and quadratic) fits; for N=6 only a fixed-D=200 estimate is provided. The paper itself states that the available bond dimension was insufficient to reach the target truncation error of 10^-10 except for N=2. Low-order polynomial fits over a narrow D window cannot certify the D->infty limit, and the quoted errors are not controlled truncation-error estimates. Consequently, the reported E0/N^2 and gauge-violation values, and the claimed agreement with Ref. [14], are not robust. Please either provide a controlled extrapolation (for example, validating the extrapolation procedure on N=2, where converged D values are available) or explicitly downgrade these numbers to preliminary estimates.","section":"Sec. 4.3, Fig. 5"},{"comment":"The central conclusion that \"the computational cost scales polynomially with the number and size of matrices\" is not supported for the size N. Fig. 2 and the MPO bond-dimension tables show scaling with the number of matrices at fixed N=2, but the required MPS bond dimension D_req(N,Lambda) for fixed accuracy is never measured for N>=3; Sec. 4.3 states that the target truncation error of 10^-10 is not reached for N>=3. Sec. 4.1 explicitly says the paper \"only focus[es] on the MPO bond dimension analysis\" and leaves a full cost analysis to future work. Since the two-site DMRG cost in Eq. (3.22) grows as Dmax^3, a polynomial-cost claim requires bounds on D_req(N), not only on the MPO bond dimension. Please either supply this data or restrict the claim to the evidence actually presented.","section":"Sec. 5 and Sec. 4.1/4.3"},{"comment":"The validation against Ref. [14] is presented as \"excellent agreement\" but is only qualitative: no table or quantitative comparison of E0/N^2 or <G^2>/N^2 is given, and Ref. [14] shares an author with the present paper and is itself an approximate variational method (neural-network VMC). This is a consistency check between two approximate methods, not an independent benchmark. Moreover, for N=5 and N=6 the gauge-invariance violation <G^2>/N^2 increases with N, which the paper attributes to limiting Lambda to 6; the singlet constraint is therefore not restored for those data. A quantitative comparison table and a clear statement of the non-independence of the benchmark are needed to support the SU(N) claims.","section":"Sec. 4.3, Fig. 6"}],"minor_comments":[{"comment":"The sentence \"Figure 5 demonstrates the convergence of the ground state energy and the gauge-invariance violation expectation value for two matrices\" is ambiguous; the figure shows several SU(N) values for models with two bosonic matrices. Please clarify the distinction between the number of matrices and N.","section":"Sec. 4.3"},{"comment":"The phrase \"expected from the 3.5\" should read \"expected from Eq. (3.5)\".","section":"Sec. 4.5.1"},{"comment":"The phrase \"throughout computational cost analysis\" should be \"thorough computational cost analysis\".","section":"Sec. 4.1"},{"comment":"The MPO compression truncation error of 10^-15 is stated only in App. E.1; it should be given in the main text alongside Tables 1 and 2.","section":"Tables 1 and 2"},{"comment":"The DMRG stopping criterion (energy difference 10^-8) is stated in Sec. 4.2 but not in the caption of Fig. 3; the caption should include it because the saturation plateaus in Fig. 3 are attributed to this criterion.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The report is based on the overclaim in Sec. 5 relative to the evidence in Secs. 4.1 and 4.3, and on the uncontrolled 1/D extrapolations. For the editor: the paper would benefit from a revised framing that clearly separates the SU(2) convergence study from the preliminary SU(N) extrapolations, and from a careful statement that the Ref. [14] comparison is a consistency check between two approximate methods rather than an independent validation. This is a calibration issue, not a question of author conduct."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a serious attempt to bring MPS/DMRG to matrix models, and the SU(2) results are the real meat. The layout scheme comparison and MPO bond-dimension analysis are useful, and the exponential convergence in the Fock-space cutoff Lambda, the decay of gauge violation, and the saturation of bond dimension with the number of matrices are all credible and well presented. That part of the paper is a genuine contribution.\n\nThe problems are in the SU(N) section and the conclusions built on it. For N=3,4,5 the target truncation error was never reached; the authors resort to 1/D extrapolations with error bars defined as the difference between two low-order fits. That is not a controlled truncation-error estimate, and for N=6 there is only a fixed-D=200 guess. The agreement with Ref. [14] is qualitative, and that paper shares an author with this one, so it is a consistency check between two approximate methods, not an independent validation.\n\nThe bigger issue is the headline claim in Sec. 5 that computational cost scales polynomially with the number and size of matrices. The evidence for scaling with the number of matrices (SU(2) saturation of Dmax) is solid. But the evidence for scaling with N is missing: the bond dimension required to reach a fixed accuracy as N grows is never measured. The paper itself flags this in Sec. 4.1 (\"we only focus on the MPO bond dimension analysis\") and in Sec. 4.3 (\"the available bond dimension was insufficient\"). These limitations undercut the central scalability conclusion. The \"first time\" wording in Sec. 5 also overreaches; no dedicated prior-art search is reported.\n\nThat said, the SU(2) feasibility result holds up, and the paper is worth taking seriously. My recommendation: send it to peer review, but require the authors to either converge N>=3 with feasible bond dimensions, provide a controlled extrapolation with honest error bars, or substantially soften the polynomial-scaling and \"first time\" claims. The paper would be solid if reframed as a detailed feasibility study with open questions about SU(N) scalability rather than a demonstration of scalable tensor network simulations.","headline":"Solid SU(2) feasibility study; the SU(N) scalability claim goes beyond what the data show.","tokens_in":23676,"tokens_out":1302,"would_cite":true,"duration_ms":15401,"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 demonstrates that tensor networks, specifically matrix product states and the density matrix renormalization group, can simulate bosonic and supersymmetric matrix models with computational cost that scales polynomially in the…","keywords":["matrix models","tensor networks","matrix product states","density matrix renormalization group","BFSS matrix model","BMN matrix model","supersymmetric matrix models","entanglement entropy"],"falsifier":"Run the same SU(3) (or SU(4)) model at a bond dimension well beyond the fitted range, for example $D=800$ with the same $\\Lambda$, and check whether the ground-state energy falls on the $1/D$ extrapolation curve within the quoted error bars; alternatively compare the extrapolated $E_0/N^2$ with a rigorous bootstrap bound at the same cutoff. A high-$D$ point that deviates by more than the reported uncertainty would show the extrapolation is biased.","tokens_in":22604,"feed_emoji":"🕸️","tokens_out":13298,"duration_ms":116753,"temperature":0.7,"pith_summary":"Matrix models are quantum-mechanical theories of matrices that, through gauge/gravity duality, encode information about higher-dimensional gauge theories, string theory, and quantum black holes; simulating their ground states is a route into regimes that analytical methods cannot reach. This paper argues that tensor networks provide a scalable numerical route: it represents ground states as matrix product states, optimizes them with the density matrix renormalization group, and reports that the bond dimension needed for fixed truncation error saturates as the number of matrices grows. The result, if correct, is a polynomial-cost simulation method for bosonic matrix models with up to nine matrices for SU(2) and, with a bond-dimension extrapolation, for SU($N$) matrices up to $N=6$, as well as for the supersymmetric minimal BMN-like model. The computed ground-state energies, gauge-invariance violations, and entanglement quantities are presented as benchmarks that align with earlier simulations.","feed_headline":"First tensor-network simulation of matrix models scales polynomially","feed_subtitle":"Tensor networks reach N=6 matrix models and match earlier Monte Carlo benchmarks.","key_machinery":"The central objects are matrix product states (MPS), tensor networks shaped as a one-dimensional chain whose bond dimension $D$ controls the accuracy, and the density matrix renormalization group (DMRG), a variational algorithm that sweeps through the chain, updates two neighbouring tensors at a time, and truncates the growing bond dimension by singular value decomposition with a controlled truncation error. The Hamiltonian is written as a matrix product operator (MPO), and each bosonic oscillator's infinite Fock space is truncated to $\\Lambda$ states per site. The scalability argument rests on the observed saturation of the required bond dimension with the number of matrices and on the weak dependence on $\\Lambda$, which turns the exponential growth of the Hilbert space into polynomial cost. For the larger SU(N) models the load-bearing device is a $1/D$ extrapolation of observables to $D\\to\\infty$.","core_discovery":"The authors claim to demonstrate for the first time that matrix product states and the density matrix renormalization group are applicable to matrix models. They map the degrees of freedom of a bosonic model with $D$ matrices and of the minimal BMN-like supersymmetric model onto one-dimensional chains, represent the Hamiltonian as a matrix product operator, and find ground states variationally. For SU(2) bosonic models with two to nine matrices, the largest bond dimension $D_{\\max}$ required to hold the truncation error below $10^{-10}$ or $10^{-11}$ saturates as the number of matrices grows and depends only weakly on the Fock-space cutoff $\\Lambda$, which implies polynomial computational cost. For SU(N) models with $N=3,4,5$, where the fixed truncation error could not be reached, the ground-state energy and gauge violation are obtained by extrapolating in $1/D$ to the infinite-bond-dimension limit and agree with earlier variational Monte Carlo results; for $N=6$ a rough estimate at fixed $D=200$ is given. The paper concludes that tensor networks are a promising and scalable tool for simulating matrix models.","pith_inferences":["If the polynomial scaling extends beyond $N=6$, tensor-network simulations could reach matrix models at couplings or temperatures where Monte Carlo sampling suffers from a sign problem, without requiring a quantum computer.","The $1/D$ extrapolation used for $N=3,4,5$ could be validated at modest cost by computing a single high-$D$ point for $N=3$; this is a direct test the paper does not perform.","The observed large gap between the first two entanglement-spectrum levels hints that entanglement data may carry information about the emergent structure of these ground states; the paper reports the gap but defers physical interpretation to future work.","Because the gauge-invariance violation is already smaller than the benchmark at all $N$ studied, enforcing the singlet constraint exactly through a penalty or projected basis might become unnecessary at large $\\Lambda$, simplifying future simulations."],"forward_implications":["For SU(2) bosonic matrix models with two to nine matrices, the ground-state energy and gauge-invariance violation can be computed with truncation errors as small as $10^{-11}$, with finite-cutoff effects decaying exponentially in $\\Lambda$.","The saturation of the required bond dimension means the cost of adding more matrices is polynomial, not exponential, so simulations with many matrices remain feasible at fixed accuracy.","For $N=3,4,5$ the extrapolated energies reproduce earlier variational Monte Carlo results while keeping the gauge-invariance violation $\\langle G^2\\rangle/N^2$ smaller, indicating the method is competitive at larger gauge group rank.","The same framework handles at least one supersymmetric model, the SU(2) minimal BMN-like model, where gauge and angular-momentum violations decay exponentially with $\\Lambda$.","Entanglement entropy and entanglement spectrum are obtained directly from the matrix product state, giving low-energy observables beyond the energy itself."],"supporting_citations":[{"why":"Supplies the benchmark values for the rescaled ground-state energy and gauge-invariance violation that the SU(N) results are compared against.","marker":"[14]"},{"why":"Supplies the earlier quantum-computing, deep-learning and lattice Monte Carlo simulations whose reach the polynomial scaling is contrasted with, and the singlet-constraint formulation.","marker":"[13]"},{"why":"Supplies the density matrix renormalization group algorithm used to optimize the matrix product states.","marker":"[58]"},{"why":"Supplies the tensor-network software used for MPO compression and DMRG sweeps.","marker":"[69]"},{"why":"Supplies the classification of massive super Yang-Mills quantum mechanics from which the supersymmetric minimal BMN-like model is taken.","marker":"[55]"},{"why":"Supplies the Hamiltonian formulation of the matrix models and the argument that no spatial lattice discretisation is needed.","marker":"[15]"}],"fun_headline_variants":["Tensor networks simulate matrix models polynomially","Polynomial-cost tensor network simulation of matrix models","Matrix models yield to tensor networks with polynomial scaling","Scalable tensor network simulation of matrix models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For the SU(N) results with $N=3,4,5$ the reported ground-state energies depend on a $1/D$ extrapolation to infinite bond dimension using low-order polynomial fits, and for $N=6$ on a single fixed-$D=200$ estimate; if those fits are not representative, the quoted $E_0/N^2$ values and their agreement with the benchmark comparison are unsupported.","fun_headline_variants_meta":{"raw":{"variants":["Tensor networks simulate matrix models polynomially","Polynomial-cost tensor network simulation of matrix models","Matrix models yield to tensor networks with polynomial scaling","Scalable tensor network simulation of matrix models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000247,"raw_usage":{"total_tokens":1492,"prompt_tokens":840,"completion_tokens":652,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":595}},"tokens_in":456,"tokens_out":652,"duration_ms":6788,"temperature":1.0,"reasoning_tokens":595,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:43:42.655303+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same SU(3) (or SU(4)) model at a bond dimension well beyond the fitted range, for example $D=800$ with the same $\\Lambda$, and check whether the ground-state energy falls on the $1/D$ extrapolation curve within the quoted error bars; alternatively compare the extrapolated $E_0/N^2$ with a rigorous bootstrap bound at the same cutoff. A high-$D$ point that deviates by more than the reported uncertainty would show the extrapolation is biased.","supporting_citations":[{"cited_title":"White, Density matrix formulation for quantum renormalization groups , Physical review letters 69 (1992) 2863","cited_arxiv_id":null,"evidence_quote":"Supplies the density matrix renormalization group algorithm used to optimize the matrix product states."},{"cited_title":"Fishman, S","cited_arxiv_id":null,"evidence_quote":"Supplies the tensor-network software used for MPO compression and DMRG sweeps."}],"review_version":1}