{"id":"edf35305-6707-46f5-93da-7ecb33ecfdcb","arxiv_id":"2601.12446","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"For the disordered XXZ chain, the Pauli-basis operator length, computed exactly from MPO marginals, grows logarithmically in the interacting MBL regime and saturates in the Anderson-localized case.","lead":"This paper introduces the “operator length”—the typical furthest site reached by a time-evolving operator’s Pauli expansion—and shows it can be computed exactly with tensor networks. In a disordered interacting spin chain it grows logarithmically in time, distinguishing many-body localization from Anderson localization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'arbitrarily weak interactions' claim rests on Δ≥0.2 MPS data with no small-Δ ED check; the log law may be a transient.","rationale":"Reader's weakest assumption concerns MPS truncation; I agree that convergence checks are needed, but I think the more load-bearing gap is the logical extrapolation to Δ→0. The ED data for Δ=1 already establishes the log-growth effect in the strongly interacting MBL regime, so a truncation artifact cannot overturn the paper's core finding. The distinctiveness and significance of the paper hinge on the statement in Sec. 4.2 and the abstract that 'even arbitrarily weak interactions' produce unbounded logarithmic spreading. This statement is used to argue that Anderson localization is singular at Δ=0 and that operator length is a sharp diagnostic of MBL. Yet the smallest Δ simulated is 0.2, which is not arbitrarily weak, and the simulated time t≤500 spans less than three decades. For a slow logarithmic law, identifying the asymptotic regime requires both smaller Δ and longer times. The ℓ-bit model has a phenomenological parameter κ and is not fitted to the XXZ slopes; the text only says the 'behavior' agrees, not a quantitative comparison. A single ED study at Δ=0.01–0.1 up to t=10^5 would settle whether the log growth persists at arbitrarily weak coupling or whether a crossover/saturation appears. If saturation occurs, the paper's strongest physical conclusion would need to be weakened to Δ above some (possibly small) threshold. Therefore the reader's CONDITIONAL verdict is appropriate; my concern sharpens the required check without changing the verdict.","tokens_in":22785,"tokens_out":17364,"duration_ms":188061,"concrete_test":"Use exact diagonalization for L=12 at W=6.5 with Δ=0.01, 0.05, 0.1, and 0.2, averaging over at least 48 disorder realizations, and compute h(t) up to t=10^5 (as in Fig. 1). Fit h(t) vs ln t over the final two decades and check for saturation. Separately, repeat the MPS runs for Δ=0.2 with bond dimension doubled (χ→2χ) at L=20 and compare h(t); if the two disagree, the small-Δ MPS data are not converged. If h(t) saturates for Δ≤0.05 or the log fit degrades, the 'arbitrarily weak interactions' claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central observation of logarithmic operator-length growth is well supported by exact diagonalization at Δ=1 (Fig. 1, L=12, t≤10^5). The load-bearing step is the abstract's assertion that the effect occurs for 'arbitrarily weak interactions.' This is inferred from Fig. 3, which shows Δ=0.2 and 0.4 only, for L=12,20 and t≤500, with no error bars and no documented MPS truncation error. The growth for Δ=0.2 is extremely small (h increases by ~0.2 over three time decades); in such a window a slow transient or a truncation-induced drift could easily masquerade as ln t. The ℓ-bit model (Appendix A) has a free localization parameter κ and is not quantitatively compared to the XXZ slopes; it therefore cannot validate the Δ→0 extrapolation. If for sufficiently small Δ the operator length saturates instead of growing as a ln t, the paper's sharpest claim—that the interacting term is a relevant perturbation for any Δ>0—would be wrong. This is not resolved by the existing data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new Pauli-basis diagnostic, the operator length h(t), complementary to the operator mass m(t), and shows how both can be computed exactly within an MPS/MPO representation of the time-evolved operator, avoiding stochastic sampling. The method is applied to the disordered XXZ chain. For disorder strengths W≳5, the interaction strength Δ=1 gives a clear logarithmic growth of h(t), m(t), and the operator entanglement entropy over several decades in exact diagonalization at L=12 up to t=10^5, while Δ=0 saturates quickly. The authors claim that even arbitrarily weak interactions produce unbounded logarithmic operator spreading and that this is quantitatively captured by a phenomenological ℓ-bit model. An experimental protocol based on Choi-state preparation and Bell-pair classical shadows is also proposed.","tokens_in":23095,"tokens_out":2780,"duration_ms":34023,"significance":"If the conclusions hold, the operator length is a useful, experimentally accessible probe that sharply distinguishes Anderson localization from MBL, complementing the OTOC and operator entanglement. The exact MPO-computation part is an important technical contribution: Eqs. (20)–(21) and the mass generating function in Eqs. (26)–(27) give deterministic, polynomial-cost access to marginals that would otherwise require sampling. The ED data for Δ=1 at strong disorder are clean and support the central logarithmic-growth observation. However, the two most striking claims—arbitrary-weak-interaction relevance and quantitative agreement with the ℓ-bit model—are not established by the presented evidence, and the tensor-network convergence is not documented. These issues are fixable and do not invalidate the core methodology.","major_comments":[{"comment":"The abstract and Sec. 4.2 claim that operator spreading occurs for arbitrarily weak interactions, but the smallest simulated interaction is Δ=0.2. For Δ=0.2, h(t) rises by only about 0.2 over three time decades (Fig. 3b), and the curves have no disorder-averaging error bars and no MPO truncation-error estimate. In that window a slow transient or a truncation-induced drift can masquerade as a logarithmic law. Please add ED data at smaller Δ (e.g., Δ=0.02, 0.05, 0.1) for L=12 to longer times, and include bootstrap or jackknife error bars. Without that, the singular-at-Δ=0 claim is not supported.","section":"Sec. 4.2, Fig. 3"},{"comment":"The text says the ℓ-bit model 'quantitatively reproduces' or 'quantitatively captures' the XXZ behavior, but no quantitative comparison is made. The model contains a free coupling-decay parameter κ and its own disorder amplitude W; Appendix A reports slopes for κ=1,0.5,0.32 but does not fix κ from the XXZ localization length and does not compare the slopes to the a(W) extracted from Fig. 1. Please fit the XXZ h(t) curves to a ln t + b and compare the slopes with the ℓ-bit prediction at the same effective κ, or tone the claim down to qualitative agreement.","section":"Appendix A, Sec. 4.2"},{"comment":"The assertion that in MBL the MPO bond dimension grows only linearly in time, 'enabling faithful simulations', is not accompanied by convergence tests. No comparison is shown between different bond dimensions, different Trotter time steps, or between ED and MPS at L=12 over the MPS time window. Since the Δ=0.2 logarithmic trend covers a very small variation of h(t), truncation could affect the extracted slope. Please add truncation-error plots and an ED/MPS cross-check on h(t) and m(t).","section":"Sec. 4.2, Sec. 3"}],"minor_comments":[{"comment":"Typos and formatting: 'folowing' (Sec. 4.1), 'nonlcal' (end of Sec. 2), 'su bstituting' (Appendix A), 'Choirepresentation' (Sec. 3), and 'emergentℓ-bits' spacing. Please proofread.","section":"General"},{"comment":"The notation h(L,W) and m(L,W) for time averages conflicts with the operator length h(t) and mass m(t); it would be clearer to use something like ⟨h⟩_L(W) or a separate symbol.","section":"Eq. (37)"},{"comment":"The figure shows three disorder strengths but no error bars; reporting the standard error over the 48 realizations would help assess the significance of the slope and is standard practice.","section":"Fig. 1"},{"comment":"The experimental section could state the number of shots needed to estimate P(l) and P(m) to a given accuracy; the discussion of statistical efficiency is qualitative.","section":"Sec. 5"}],"recommendation":"major_revision","confidential_remarks":"The methodological core is solid and the Δ=1 ED result is convincing. The requested additions are numerical controls rather than conceptual corrections, so the paper is likely publishable after a revision that substantiates or softens the 'arbitrarily weak interactions' and 'quantitative ℓ-bit capture' claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the observable: operator length as a Pauli-basis marginal, computed exactly from an MPO via the S̃2(l) recurrence in Eqs. (18)–(21). That is clean and not something I have seen before. The mass marginal via the generating function is also neat. For a reader who works on operator spreading or MBL probes, this is a worthwhile toolbox paper, and the experimental-shadow section is a credible proposal rather than hand-waving.\n\nThe ED result at Δ=1, W≥5 is the strongest part: L=12 out to t=10^5, 48 disorder samples, and a clear logarithmic trend in h(t). I would bet on that being real. The distinction between Anderson saturation and interacting logarithmic growth is also clearly demonstrated. The method itself deserves credit: it avoids stochastic sampling and gives exact marginals for a given MPO, which is a real advantage over SRE sampling.\n\nNow the soft spots, in order of seriousness. First, the abstract's 'arbitrarily weak interactions' claim rests on Fig. 3: Δ=0.2 and 0.4, t≤500, no error bars, no MPS truncation checks. The Δ=0.2 growth is tiny — h increases by about 0.2 over three decades. A slow transient or truncation-induced drift could masquerade as ln t. The stress-test note is right that this is not resolved by the current data. A small-Δ ED run (L=12, t~10^4) would settle it, and I am surprised they did not do it. Second, the claim that the ℓ-bit model 'quantitatively captures' the XXZ result is not supported: κ is a free parameter, and the appendix never compares the XXZ slope to the ℓ-bit slope. It shows the ℓ-bit model also produces logarithmic growth, which is consistency, not quantitative agreement. Third, the MPS convergence claim — 'bond dimension grows only linearly in time, enabling faithful simulations' — is asserted without a single convergence test against larger χ or against ED. That is fixable but should have been included.\n\nNone of this sinks the core contribution. The method is sound, the Δ=1 logarithmic law is well supported, and the paper is honest about the known logarithmic light cone context. I would send it to a serious referee, with instructions to ask for small-Δ ED data, MPS convergence checks, and a direct slope comparison with the ℓ-bit model. After that, it would be a solid publication — not a breakthrough, but a genuinely useful one.","headline":"Smart, useful MPO diagnostics for operator spread; the Δ=1 log law looks solid, but the 'arbitrarily weak interactions' claim is extrapolated beyond the data.","tokens_in":23597,"tokens_out":1221,"would_cite":true,"duration_ms":16743,"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":"The paper introduces an operator length — the averaged rightmost non-identity site in the Pauli expansion — and shows that in the many-body localized regime it grows logarithmically in time for arbitrarily weak interactions, while in Anders","keywords":["operator length","operator mass","many-body localization","Anderson localization","matrix product operators","Pauli basis","operator entanglement entropy","classical shadows"],"falsifier":"Run the same MPS time evolution at L = 20, W = 6.5, Delta = 1 with bond dimension chi and 2 chi, and compare h(t) at t = 500; if the logarithmic slope changes by more than the reported error bars, the logarithmic law is a truncation artifact rather than a property of the exact dynamics.","tokens_in":22693,"feed_emoji":"📏","tokens_out":4834,"duration_ms":50607,"temperature":0.7,"pith_summary":"The paper is trying to establish that operator delocalization in a disordered spin chain can be read off from two simple Pauli-basis marginals — operator mass and the newly introduced operator length — and that these quantities sharply separate Anderson localization from many-body localization. The central numerical claim is that in the interacting disordered XXZ chain at strong disorder, the operator length grows as h(t) ~ a ln t + b over many time decades, while in the non-interacting case it rapidly saturates; the same logarithmic growth is found for operator mass and operator entanglement entropy, and it persists for arbitrarily weak interactions. The paper further claims that this growth is quantitatively captured by an effective l-bit model and that both marginals can be computed exactly and efficiently from a matrix-product-operator representation, avoiding stochastic sampling. A sympathetic reader would care because this offers an exact, experimentally accessible diagnostic that distinguishes the two localized regimes without needing full operator tomography.","feed_headline":"Operator spreading is logarithmic in MBL, flat in Anderson","feed_subtitle":"Exact Pauli-basis marginals give a sharp, experiment-ready fingerprint of the many-body localized phase.","key_machinery":"The central object is the Pauli-string expansion of the time-evolved operator. Each string is assigned a mass (number of non-identity Pauli matrices) and a length (position of the rightmost non-identity site); averaging these against the squared Hilbert-Schmidt amplitudes defines m(t) and h(t). In the MPS representation these become expectation values of simple diagonal operators: h(t) is obtained from the Renyi-2 entropies of the operator state, and m(t) from a factorized mass superoperator. The log growth is explained by the l-bit Hamiltonian with exponentially decaying couplings J_{jl} = W_{jl} e^{-kappa |j-l|}, whose Heisenberg solution gives h(t) ~ (1/kappa) ln t until finite-size satur","core_discovery":"On the paper's own terms: for the disordered XXZ chain, after starting from a single-site sigma-z operator, the average operator length h(t) exhibits a clear logarithmic increase, h(t) ~ a ln t + b, over several time decades for W greater than about 5 and Delta = 1, while for Delta = 0 it rapidly saturates. Even arbitrarily weak interactions produce unbounded logarithmic operator spreading, and the behavior is quantitatively captured by an effective l-bit model. The mechanism is dephasing between exponentially localized integrals of motion: an effective coupling J_{1r} ~ e^{-r/xi} induces dephasing on timescales t_r ~ e^{r/xi}, so inverting gives r(t) ~ xi ln t. The paper also claims the nov","pith_inferences":["Going beyond the paper: if the logarithmic growth is truly unbounded for arbitrarily weak interactions, then in the thermodynamic limit the operator length should eventually saturate only at the system boundary; comparing the observed log slope at different L can test whether the MBL regime survives at asymptotically large sizes or is cut off by rare-region or avalanche effects.","Going beyond the paper: the exact marginals could be repurposed as a benchmark for nonstabilizerness estimators, since both are defined on the same Pauli distribution; discrepancies between exact marginals and sampled magic quantities would identify where sampling overhead enters.","Going beyond the paper: the same MPO-marginal technique transfers directly to monitored circuits, where the operator length distribution could track measurement-induced entanglement transitions, a direction the paper itself gestures toward.","Going beyond the paper: a direct experimental test on a 12-20 qubit device — measuring h(t) via Bell-pair shadows in both Delta = 0 and Delta = 1 disorder — would either confirm or falsify the logarithmic separation at finite times."],"forward_implications":["If the central claim is correct, the operator length is a sharp dynamical probe: it grows logarithmically in the MBL regime and saturates in the Anderson-localized regime, cleanly separating the two phases.","Because the marginals are computed exactly from the MPO, the full probability distributions of operator mass and length are available without stochastic sampling; this removes a sampling bottleneck for these diagnostics.","The logarithmic growth of operator entanglement implies the MPO bond dimension grows only linearly in time in the MBL regime, so long-time simulations remain polynomially efficient and are reliable at the accessible sizes.","The proposed experimental protocol — Choi-state preparation, controlled forward/backward evolution, and Bell-pair classical shadows — gives a shot-efficient route to measuring the operator marginals on current quantum platforms.","The time-averaged operator length and mass, rescaled by ln L, collapse onto a disorder-dependent curve, providing a clean numerical signature of the MBL logarithmic light cone."],"fun_headline_variants":["Operator length logs MBL spread, flat in Anderson","Exact MPS shows log operator growth in MBL phase","Logarithmic operator spreading marks MBL, not Anderson","MBL operator length grows as log(t), Anderson saturates","Operator delocalization: log in MBL, flat in Anderson"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the MPS/MPO time evolution is converged enough that the observed logarithmic growth of h(t) at L = 20 and t < about 500 is physical, not an artifact of bond-dimension truncation; the paper asserts the bond dimension grows only linearly in time but provides no explicit convergence tests against larger bond dimensions or exact data.","fun_headline_variants_meta":{"raw":{"variants":["Operator length logs MBL spread, flat in Anderson","Exact MPS shows log operator growth in MBL phase","Logarithmic operator spreading marks MBL, not Anderson","MBL operator length grows as log(t), Anderson saturates","Operator delocalization: log in MBL, flat in Anderson"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1393,"prompt_tokens":775,"completion_tokens":618,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":533}},"tokens_in":519,"tokens_out":618,"duration_ms":6837,"temperature":1.0,"reasoning_tokens":533,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T09:47:31.244049+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same MPS time evolution at L = 20, W = 6.5, Delta = 1 with bond dimension chi and 2 chi, and compare h(t) at t = 500; if the logarithmic slope changes by more than the reported error bars, the logarithmic law is a truncation artifact rather than a property of the exact dynamics.","supporting_citations":[],"review_version":1}