{"id":"d722b350-50d5-4f5f-85b5-56d998a1d04d","arxiv_id":"1908.06524","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TDVP time evolution with insufficient bond dimension spuriously overestimates delocalization and entanglement in the MBL crossover, and correcting this lowers the estimated critical disorder to Wc = 4.2 ± 0.3.","lead":"This paper compares two matrix product state time-evolution methods on disordered spin chains and shows that the newer TDVP method can mislead when the bond dimension is too small near the many-body localization transition. It matters because it changes the recommended numerical method and lowers the estimated critical disorder from about 5.5 to 4.2.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Wc≈4.2±0.3 rests on an arbitrary β threshold and an assumed power-law decay; the paper admits β=0.01 would give Wc≈5, so the contradiction with [41] is not settled by the convergence corrections.","rationale":"The reader's weakest_assumption—the unsupported power-law form and arbitrary β threshold—is exactly the load-bearing weak point. The algorithm-comparison claim survives contact with exact benchmarks; the Wc estimate does not. The paper's own text contains the decisive admission: the difference from [41] originates mainly from the choice of β cutoff, not from the new convergence corrections. My independent check (model selection on exact long-time data) would determine whether the power-law form is even identifiable in the relevant window and how much the inferred Wc moves with the fit window. This does not change the reader's CONDITIONAL verdict: the paper remains a useful numerical study with an unjustified quantitative transition point. No adversarial or ad hominem angle is intended; the concern is internal to the paper's procedure.","tokens_in":15969,"tokens_out":6028,"duration_ms":63090,"concrete_test":"On the exact Chebyshev L=26 data, for W=3, 3.5, 4, 4.5, 5 compute AIC/BIC for power-law I(t)=a1 t^{-a2}, logarithmic I(t)=a1+a2 ln t, and stretched-exponential fits in windows [100,200], [200,400], and [100,500]. Also record β(W) for each window. If the logarithmic model is within ΔAIC<2 of the power-law model at W≈4, or if β at fixed W changes by more than 0.01 between the [100,200] and [200,400] windows, the β=0.02 crossing is not a robust identifier of Wc and the Wc=4.2±0.3 estimate should be replaced by an explicit sensitivity range (e.g., Wc over β thresholds 0.01–0.03).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The quantitative headline—Wc≈4.2±0.3 versus Wc≈5.5 in [41]—is extracted in Sec. V by fitting I(t)∝t^{-β} in t∈[100,200] and reading off where β crosses the hand-chosen threshold β=0.02. The paper itself states that there is \"no real theoretical foundation\" for the power-law conjecture, and Fig. 13 shows that on exact L=26 data at W=3.5 a logarithmic fit matches the power-law fit to within RMS for t up to 1000. More directly, the text admits that using [41]'s cutoff β≈0.01 shifts the apparent Wc close to 5 even for L=26. Thus the claimed discrepancy with [41] is dominated by the threshold choice, not by the corrected convergence behavior. The entropy-overshoot and imbalance-direction observations for TDVP at insufficient χ are well supported by the exact Chebyshev benchmarks (Figs. 2–5) and are not the vulnerable part of the paper; the vulnerable part is the translation of corrected dynamics into a critical disorder value.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares two matrix-product-state time-evolution algorithms, tDMRG and TDVP, on the disordered Heisenberg chain, using exact Chebyshev propagation for L=26 as a benchmark. It reports that TDVP is more accurate than tDMRG in the delocalized regime, but that in the MBL crossover tDMRG gives more predictable, controllable errors, while TDVP with insufficient bond dimension can spuriously accelerate imbalance decay and overestimate the long-time entanglement entropy. The paper then studies larger systems (L=50, 200) and, from power-law fits of the imbalance decay in t∈[100,200] with a hand-chosen threshold β=0.02, estimates Wc≈4.2±0.3, arguing that the previous Wc≈5.5 estimate of Ref. [41] is affected both by insufficient bond dimension and by the β cutoff. It also reports an open-boundary-condition level-statistics estimate Wc=3.29(9) and argues that time-dynamics and level-statistics estimates need not coincide.","tokens_in":16254,"tokens_out":5857,"duration_ms":54483,"significance":"If the algorithm-comparison result holds, it is a valuable caution for the many TDVP-based studies of MBL dynamics: unconverged TDVP data in the crossover region are biased in a counter-intuitive direction, making the system appear more delocalized, and the entanglement entropy can be overestimated at long times. The benchmark against exact Chebyshev evolution for L=26 with multiple bond dimensions is a clear strength, as is the authors' explicit discussion of the fitting ambiguities in the transition-point estimate. The quantitative claim Wc≈4.2±0.3, however, is not on the same footing as the algorithm comparison, because it depends on an assumed power-law decay and a hand-chosen threshold whose sensitivity the authors themselves demonstrate.","major_comments":[{"comment":"The central quantitative claim that the transition lies at Wc≈4.2±0.3, contradicting the Wc≈5.5 estimate of Ref. [41], is not supported by the analysis as presented. The threshold β=0.02 is chosen by hand, and the text explicitly states that using the β≈0.01 cutoff of Ref. [41] shifts the apparent Wc close to 5 even for the L=26 Chebyshev data. Since Fig. 13 shows that a logarithmic fit matches the power-law fit to within RMS error up to t=1000, and the authors state that the power-law decay has \"no real theoretical foundation,\" the difference between Wc≈4.2 and Wc≈5.5 is dominated by the fitting convention rather than by the corrected convergence behavior of TDVP. The quoted error bar ±0.3 does not include this systematic uncertainty. The manuscript should either remove or substantially weaken the quantitative Wc claim, report a range that explicitly includes the threshold sensitivity (e.g., roughly 4 to 5), or provide an objective, data-driven criterion for β_c.","section":"V, Fig. 15"},{"comment":"The χ→∞ extrapolation used in the flowing-β analysis is uncontrolled. A third-order polynomial in 1/χ is fitted to five data points with no stated uncertainty for the extrapolated value, and the authors note that at W=3 the TDVP and tDMRG extrapolations agree only up to moderate times and deteriorate at longer times. The flowing-β analysis in Fig. 16, which is used to argue that W=4.5 is above Wc, relies on these extrapolated data up to t=450. Please provide a systematic error estimate for the extrapolation (for example, testing higher-order fits or varying the set of χ values included) or restrict the conclusions to statements that do not depend on the extrapolated regime.","section":"IV, Figs. 7–10 and 16"}],"minor_comments":[{"comment":"\"state-of-art estimate\" should read \"state-of-the-art estimate\" in both the abstract and the introduction.","section":"Abstract and I"},{"comment":"The statement \"Fitting errors are less than 10^-10\" is unclear: if this refers to the residual of the polynomial fit in I(t), the value is implausibly small for a five-point fit, and if it refers to the extrapolated value, the basis should be stated. Please clarify or correct.","section":"IV, Fig. 7 and Fig. 10 captions"},{"comment":"The main text says \"Shaded regions are discrepancies appearing in the delocalized regime,\" while the Fig. 14 caption says the exact Chebyshev data lie within a shaded area for L=26; please reconcile these descriptions.","section":"V, Fig. 14"}],"recommendation":"major_revision","confidential_remarks":"The methodological comparison in Sections III–IV is solid and well benchmarked, and I would not question that part of the paper. The main issue is that the title, abstract, and Section V present Wc≈4.2±0.3 as a definite result even though the authors themselves show that the value shifts toward 5 if Ref. [41]'s β criterion is used. In revision, the authors should either substantially soften the Wc claim or re-center the paper on the algorithm comparison and the demonstrated sensitivity of transition-point estimates to fitting conventions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper is that it finds a genuinely counter-intuitive numerical fact: in the MBL crossover, TDVP with insufficient bond dimension makes the system look spuriously delocalized, and the entanglement entropy can overshoot the exact value at long times. That is checked against exact Chebyshev data for L=26, with 200 disorder realizations, and it holds up. tDMRG, by contrast, underestimates entropy in a controllable way and is the safer method in the localized and crossover regimes. This is a useful caution for anyone doing MPS time evolution in disordered systems, and it goes beyond the usual folklore that TDVP is always the better algorithm.\n\nThe authors are also honest about their own uncertainties. They say plainly that the power-law decay of the imbalance has no real theoretical foundation, that a logarithmic fit matches the exact data about as well, and that the choice of threshold β=0.02 is arbitrary to some degree. They give Wc=4.2±0.3 with a large error bar because of that. The flowing-β analysis and the boundary-condition dependence of level statistics (Wc=3.29(9) for OBC) are genuinely interesting and make the paper more than a single numerical claim.\n\nThe soft spot is exactly where the stress-test note lands. The headline correction to the transition point is dominated by the arbitrary β threshold, not by the convergence corrections. The authors themselves admit that using the earlier paper's β≈0.01 cutoff pushes Wc close to 5 even for L=26. The time window is short, t∈[100,200], and the functional-form ambiguity is real. So the paper does not settle whether the true critical disorder is closer to 4.2 or 5.5; it shows that the question is more ambiguous than the earlier estimate suggested. That is a legitimate and useful result, but it should be framed as a convergence-and-ambiguity study, not as a definitive new critical value.\n\nA few minor things: 200 disorder realizations is on the low side for some of the noise-sensitive fits, though they compensate with 400 and 800 in later sections, and the 1/χ extrapolations are not justified by any theory. These are secondary.\n\nMy recommendation: this deserves a serious referee. The algorithmic observation is solid, the benchmarks are reproducible in principle, and the paper engages fairly with the competing results. I would send it out, with the strong suggestion that the quantitative Wc estimate be softened or explicitly presented as one possible reading of ambiguous data.","headline":"The TDVP entanglement overshoot at small bond dimension in the MBL crossover is a real, benchmarked finding worth knowing; the paper's specific Wc≈4.2 estimate is too threshold-dependent to settle the disagreement with the earlier Wc≈5.5.","tokens_in":16718,"tokens_out":1991,"would_cite":true,"duration_ms":21949,"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 argues that TDVP time evolution with insufficient bond dimension makes systems in the many-body localization crossover look spuriously delocalized, and that correcting this shifts the critical disorder of large Heisenberg…","keywords":["many-body localization","matrix product states","TDVP","tDMRG","entanglement entropy","imbalance decay","critical disorder","Heisenberg spin chain"],"falsifier":"Compare the long-time imbalance decay for L=26 at W=3.5 using exact Chebyshev propagation out to t≈$10^{4}$ with a large number of disorder realizations and perform a formal model comparison between power-law and logarithmic fits; if the logarithmic fit is preferred, the β-threshold definition of Wc loses its foundation. Alternatively, run L=50 TDVP at χ=512 and check whether the entanglement entropy still exceeds the tDMRG lower bound for t>350, which would confirm the overestimation mechanism.","tokens_in":15770,"feed_emoji":"🧲","tokens_out":8879,"duration_ms":76241,"temperature":0.7,"pith_summary":"The paper compares two matrix-product-state time-evolution methods for the disordered Heisenberg spin chain and finds that the \"better\" method depends on the regime. Benchmarking against exact Chebyshev propagation for L=26, it shows that TDVP with too-small bond dimension makes systems look spuriously delocalized: the imbalance decays too fast and the entanglement entropy is overestimated at long times. tDMRG makes the opposite error, underestimating the entropy and giving a predictable lower bound. Using converged data (χ=384) for L=50 and L=200, a power-law fit of the imbalance in t∈[100,200] yields a critical disorder $W_c = 4.2 \\pm 0.3$, substantially lower than the previous TDVP-based estimate $W_c \\approx 5.5$. The authors stress that the fit form and the threshold used to define the transition are the main sources of uncertainty.","feed_headline":"MBL critical disorder drops to 4.2 once TDVP converges","feed_subtitle":"Unconverged matrix-product-state dynamics make localized Heisenberg chains look delocalized.","key_machinery":"The central objects are matrix product states of bond dimension $\\chi$ evolved by two algorithms: tDMRG, a TEBD variant using the Sornborger-Stewart decomposition that grows the bond dimension and truncates by singular values, and TDVP (one- and two-site), which projects the Hamiltonian onto the tangent space of the MPS manifold. The load-bearing mechanism is the different error accumulation: tDMRG's truncation discards the most entangled components, systematically underestimating entanglement entropy, while TDVP's tangent-projection error in disordered systems biases the state toward delocalized-looking dynamics and overestimates entanglement entropy at long times when $\\chi$ is insufficient. Exact Chebyshev propagation for L=26 serves as the benchmark that reveals which method is trusted.","core_discovery":"The central claim is that unconverged TDVP results in the MBL crossover are qualitatively wrong in a specific direction: with insufficient bond dimension $\\chi$, the imbalance decays faster than the exact result while the half-chain entanglement entropy overshoots the true value at long times, so the data look \"more delocalized\" rather than simply under-resolved. tDMRG, by contrast, converges monotonically and its entanglement entropy serves as a lower bound. On this basis the paper revisits the earlier L=100 TDVP study, showing that its $W_c \\approx 5.5$ estimate is inflated by the combination of insufficient $\\chi$ and a permissive $\\beta$ threshold, and estimates $W_c = 4.2 \\pm 0.3$ from imbalance decay for L=50 and L=200, with only weak system-size dependence. The paper also reports that open boundary conditions shift the level-statistics critical disorder to $W_c = 3.29(9)$, and argues that level-statistics and finite-time dynamics probes need not yield the same transition point.","pith_inferences":[],"forward_implications":["In the MBL crossover, TDVP results must be checked for bond-dimension convergence before drawing physical conclusions; a single $\\chi$ value can place the system on the wrong side of the transition.","The critical disorder of the disordered Heisenberg chain, as inferred from finite-time imbalance dynamics, is near $W_c \\approx 4.2 \\pm 0.3$ and depends only weakly on system size between L=50 and L=200, contrary to the strong size dependence reported previously.","Because power-law and logarithmic fits of the imbalance are nearly indistinguishable on accessible time scales, critical-disorder estimates from curve fitting carry an irreducible ambiguity that should be reported.","tDMRG's entanglement entropy is a reliable lower bound, so bracketing the true entropy between tDMRG and converged TDVP data gives a practical convergence test.","The flowing-$\\beta$ analysis indicates that W=4.5 is already on the localized side, supporting imbalance saturation in the thermodynamic limit.","Existing TDVP-based studies of entanglement growth in the MBL crossover may overestimate growth exponents if their bond dimension was insufficient; re-analyzing published data with tDMRG lower bounds could revise reported exponents.","The strong boundary-condition dependence of the level-statistics critical disorder (3.29 with open boundaries versus 3.72 with periodic boundaries) suggests that open-boundary finite-size scaling is not a clean thermodynamic probe, which may explain some discrepancies between level-statistics and dynamics-based transition estimates.","A direct test of the paper's convergence picture would be to compute operator spreading or density propagators with both methods at matched $\\chi$; the method-dependent direction of error should carry over if the mechanism is general."],"supporting_citations":[{"why":"The prior TDVP-based L=100 study whose Wc≈5.5 estimate this paper corrects.","marker":"[41]"},{"why":"Supplies the Chebyshev expansion method used as the exact benchmark.","marker":"[42]"},{"why":"Provides the two-site/one-site TDVP algorithm used in the simulations.","marker":"[16]"},{"why":"Review defining the hybrid TDVP protocol and motivating method comparisons.","marker":"[17]"},{"why":"Introduces the tDMRG algorithm with the second-order decomposition used here.","marker":"[9]"},{"why":"Finite-size scaling of level statistics for small systems, giving the Wc=3.72(6) reference.","marker":"[47]"},{"why":"Independent estimate of critical disorder W=4.5(1), supporting the paper's value.","marker":"[55]"},{"why":"Claims questioning MBL in the thermodynamic limit, which the flowing-β analysis addresses.","marker":"[43]"},{"why":"Suggests the power-law decay form for imbalance used to define Wc.","marker":"[51]"},{"why":"Experimental measurement of imbalance decay fitted by a power law, motivating the fit form.","marker":"[52]"}],"fun_headline_variants":["MBL transition drops to 4.2 when TDVP converges","Unconverged TDVP inflates MBL critical disorder estimate","tDMRG beats TDVP for many-body localization dynamics","Large-system MBL transition revisited: Wc=4.2 from imbalance","TDVP convergence flaws skew MBL critical disorder to 5.5"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The estimate $W_c = 4.2 \\pm 0.3$ rests on the assumption that the imbalance decays as a power law $t^{-\\beta}$ on the delocalized side; the paper concedes this form has no real theoretical foundation, and a logarithmic decay would shift the inferred transition toward larger disorder.","fun_headline_variants_meta":{"raw":{"variants":["MBL transition drops to 4.2 when TDVP converges","Unconverged TDVP inflates MBL critical disorder estimate","tDMRG beats TDVP for many-body localization dynamics","Large-system MBL transition revisited: Wc=4.2 from imbalance","TDVP convergence flaws skew MBL critical disorder to 5.5"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000791,"raw_usage":{"total_tokens":3464,"prompt_tokens":905,"completion_tokens":2559,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":2466}},"tokens_in":521,"tokens_out":2559,"duration_ms":18935,"temperature":1.0,"reasoning_tokens":2466,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:41:48.659094+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the long-time imbalance decay for L=26 at W=3.5 using exact Chebyshev propagation out to t≈$10^{4}$ with a large number of disorder realizations and perform a formal model comparison between power-law and logarithmic fits; if the logarithmic fit is preferred, the β-threshold definition of Wc loses its foundation. Alternatively, run L=50 TDVP at χ=512 and check whether the entanglement entropy still exceeds the tDMRG lower bound for t>350, which would confirm the overestimation mechanism.","supporting_citations":[{"cited_title":"We also analyse this issue in this work, but first we com- pare the performance of tDMRG and TDVP in systems close to localized regime","cited_arxiv_id":null,"evidence_quote":"The prior TDVP-based L=100 study whose Wc≈5.5 estimate this paper corrects."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Independent estimate of critical disorder W=4.5(1), supporting the paper's value."},{"cited_title":"Time dynamics with matrix product states: Many-body localization transition of large systems revisited","cited_arxiv_id":"1908.06524","evidence_quote":"Claims questioning MBL in the thermodynamic limit, which the flowing-β analysis addresses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Experimental measurement of imbalance decay fitted by a power law, motivating the fit form."}],"review_version":1}