{"id":"46798ae0-4ae0-4ba3-970a-59b5fa6437cd","arxiv_id":"2412.14261","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Compressed matrix product states from random and monitored circuits spread correlations over xi_eff ~ log chi^alpha, and near-zero transfer-matrix eigenvalues detect the measurement-induced entanglement transition.","lead":"Random quantum circuits compressed into matrix product states with fixed bond dimension are studied through their transfer-matrix spectra. The paper finds that correlations spread over a length set by the full spectrum, and a near-zero-eigenvalue order parameter detects measurement-induced entanglement transitions even for small bond dimensions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MIPT order parameter is unvalidated: the paper uses P(|λ|<ρ) at ρ=10^-6 as a proxy for transfer-matrix nullity, but this proxy is nonzero in the volume-law phase at small χ and the ρ→0 limit is not established; the central transition claim depends on this proxy.","rationale":"The reader's weakest assumption (self-averaging of Tr(ρ^k)) is a real gap, but it mainly endangers the analytic exponent α=2 for RMPS, a secondary result. The most load-bearing claim is the detection of the MIPT via the transfer-matrix null space, stated in the abstract and title; if the order parameter is not the true nullity and does not separate phases at finite χ, the paper's headline contribution is unsupported. The paper ships no code or data, so the numerical checks in Fig. 4 cannot be independently audited. It deserves credit for the nontrivial consistency that pc(χ) approaches the known 0.16, but that is not a substitute for a direct nullity computation or a scaling collapse. A direct nullity test is computationally cheap and would settle whether the order parameter is valid, so the appropriate verdict remains CONDITIONAL, unchanged from the reader.","tokens_in":13086,"tokens_out":16056,"duration_ms":151914,"concrete_test":"Reproduce the monitored-circuit MPS simulation at N=60 (and ideally N=40,80) for p∈[0,0.3] and χ∈{20,30,40,50}. For each trajectory, compute the full SVD of the transfer matrix and define the nullity as the number of singular values below a relative threshold ε·s_max with ε=10^-12. Average the nullity over trajectories and compare it with the paper's P(|λ|<10^-6). Check whether the averaged nullity is exactly zero for p<p_c(χ,N) and positive for p>p_c(χ,N), and whether the crossing sharpens toward p_c=0.16 as N and χ grow. If the nullity is zero for all p at any fixed χ, or if P(|λ|<10^-6) disagrees with the nullity, the proposed order parameter is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MPS ensembles exhibit an MIPT detectable via the transfer-matrix null-space dimension rests on the quantity P(|λ|<ρ) extrapolated to ρ→0+. The paper never computes the actual nullity (rank deficiency) of the transfer matrix; instead it counts eigenvalues with |λ|<10^-6. These are not the same object: for non-normal matrices, small eigenvalues need not correspond to an exact null space, and truncation artifacts can produce spurious near-zero eigenvalues. The data in Fig. 4 show precisely such a problem: in the volume-law phase, P(|λ|<10^-6) is nonzero for small χ (e.g., p=0.1, χ=20; p=0.15, χ=20,30,40) and vanishes only as χ grows. Thus the proxy does not cleanly separate the phases at finite χ, and the extracted pc(χ) may be a truncation crossover rather than a thermodynamic transition. The extrapolation ρ→0+ is not described (no fit form, no stability analysis), and no scaling collapse with system size N is presented. The convergence pc(χ)→0.16 is encouraging but could be coincidental given the arbitrary threshold, N=60, and χ≤50. If the true nullity is zero for all p at fixed χ (because a generic projected MPS is full rank), the 'transition at any χ' claim fails.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies ensembles of matrix product states (MPSs) of fixed bond dimension χ generated by random sequential circuits, random brickwork circuits, and hybrid circuits with projective measurements. The authors analyze the eigenvalues of the MPS transfer matrix and use the Rényi-2 mutual information to define an effective correlation length ξeff. They claim that ξeff grows as log χ^α with α depending on the microscopic dynamics: α = 2 for random MPSs, α ≈ 1.8 for translation-invariant Haar circuits, and a measurement-rate-dependent α(p) that decreases to zero across the measurement-induced entanglement transition (MIPT). For monitored circuits they propose that the averaged dimension of the transfer-matrix null space, probed through the eigenvalue counting function P(|λ| < ρ), acts as an order parameter for the MIPT at finite χ, with a critical rate pc(χ) approaching the known value pc ≃ 0.16 as χ grows.","tokens_in":13413,"tokens_out":3540,"duration_ms":36427,"significance":"If the central claims hold, the paper would establish a useful phenomenology: the full transfer-matrix spectral density, not just the leading eigenvalue, controls spatial correlations in compressed MPS descriptions of chaotic dynamics, and a finite-χ MPS truncation can still detect the MIPT. The analytic Weingarten computation in End Matter A leading to Eq. (6), the comparison of RMPS and Haar-TI spectra, and the explicit prediction ξeff ∼ log χ^2 for RMPSs are concrete and testable strengths. The numerical observation that the inferred pc(χ) approaches the established MIPT value is suggestive. However, the validity of the analytic derivation and of the proposed order parameter is not yet established to the standard required for the paper's main conclusions.","major_comments":[{"comment":"The derivation of Eq. (6) relies on the replacement E log[Tr(ρ_X^k)] ≈ log[E Tr(ρ_X^k)], stated in the text as an assumption that statistical fluctuations are small. This assumption is load-bearing for the analytic value α = 2 and for the claim that Eq. (5), derived for translation-invariant MPSs, can be applied to non-translation-invariant RMPSs after averaging. The manuscript does not provide any numerical check of the variance of Tr(ρ_X^k) over the Haar ensemble, nor a comparison of the averaged mutual information with the mutual information of typical samples. I ask the authors to validate this concentration assumption numerically for the values of χ used in Fig. 3 and, if it fails for some k or χ, to state the resulting limitations on the analytic formula.","section":"Effective correlation length in translational invariant MPSs and in RMPSs; End Matter A"},{"comment":"The central MIPT claim uses P(|λ| < ρ) at ρ = 10^-6 as a proxy for the dimension of the transfer-matrix null space, but the paper never computes the actual nullity. For non-normal transfer matrices, eigenvalues with modulus below a fixed threshold need not correspond to an exact null space, and truncation can produce spurious near-zero eigenvalues. Figure 4(b) itself shows nonzero P(|λ| < 10^-6) in the volume-law phase at small χ (for example p = 0.1, χ = 20 and p = 0.15, χ = 20, 30, 40), so the proxy does not cleanly separate the phases at finite χ. The claimed extrapolation ρ → 0+ is not described: no fit form, threshold-independence analysis, or stability check is given. To support the conclusion that a finite null space appears for any χ at p > pc(χ), the authors should either compute the rank deficiency of the transfer matrix directly or provide a controlled extrapolation in ρ together with a scaling analysis in the system size N.","section":"Monitored dynamics; Figure 4"},{"comment":"The exponents α reported for the Haar-TI circuit (α ≈ 1.8) and for monitored circuits (α(p) in Fig. B.1) are extracted from linear fits of Rényi-2 mutual information versus log χ, but no error bars, fit ranges, or sensitivity to the choice of r = 1 are reported. Since the separation between α = 2 and α ≈ 1.8 is quantitative and underlies the claim that the spectral density encodes model-dependent information, the fits need to be documented sufficiently for the reader to assess whether the difference is significant. Similarly, the convergence of α(p) to 0 for p > pc is inferred from the χmin dependence without an extrapolation; please provide a concrete fitting procedure and associated uncertainties.","section":"Figures 3 and B.1"}],"minor_comments":[{"comment":"The notation P(|λ| < 0+) is not defined; the text should state explicitly that it means lim_{ρ→0+} P(|λ| < ρ) and describe how this limit is estimated from the data.","section":"Monitored dynamics; Figure 4"},{"comment":"The 'Haar-TI' ensemble is not fully specified: it should be stated whether translation invariance is imposed over sites only, over gates within a layer, and whether the same random gate is repeated at each time step or only within a layer.","section":"Spectra of TI-Haar circuits"},{"comment":"The statement that contributions from states with small gaps 'effectively cancel out in the sum (5) as dephasing angles' is asserted without demonstration; Eq. (5) contains complex overlaps involving left and right eigenvectors, so the cancellation is not obvious and should either be proved or presented as a conjecture.","section":"Unitary and Monitored Dynamics"},{"comment":"There are duplicated references: [23] and [61] are both Skinner, Ruhman, and Nahum, Phys. Rev. X 9, 031009 (2019), and [22] and [60] are both Chan, Nandkishore, Pretko, and Smith, Phys. Rev. B 99, 224307 (2018).","section":"References"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the paper gives a genuinely new way to read physical information out of compressed MPS—the full transfer-matrix spectrum, not just the gap, controls correlation spreading, and the RMPS result (Eq. 6) is a real analytic handle. But the headline claim that a near-zero-eigenvalue count detects the MIPT at any finite chi is the weak link: the proxy they use, P(|lambda|<rho) at rho->0+, is not validated as a nullity measure, and their own data show it does not cleanly separate phases at small chi.\n\nWhat is new: Eq. (6) is a parameter-free derivation of the Renyi-2 mutual information for random MPSs by Weingarten averaging, predicting xi_eff ~ log chi^2. The observation that the full spectral sum in Eq. (5), rather than the leading eigenvalue, governs correlations in volume-law MPS ensembles, with model-dependent exponents alpha, is a useful framework. The paper is clearly written and honest: the self-averaging approximation E log ~ log E is stated explicitly, and the End Matter A diagonalization for k=2,3,4 partially supports the RMPS exponent.\n\nSoft spots: the MIPT order parameter is the biggest issue. They count eigenvalues with |lambda|<10^-6 and extrapolate rho->0+, but never compute the actual rank deficiency of the transfer matrix. For non-normal matrices, small eigenvalues are not the same as exact zero modes, and truncation can create spurious near-zero eigenvalues. Their own Fig. 4 caption admits the limit is nonzero in the volume-law phase for small chi (p=0.1, chi=20; p=0.15, chi=20-40), so the proxy does not cleanly separate the phases at finite chi. The extrapolation has no fit form or stability analysis, there is no scaling collapse with system size, and the convergence pc(chi)->0.16 could be fortuitous. A referee should demand a direct nullity computation, finite-size scaling, and a justification of the rho->0 limit. The RMPS derivation's self-averaging assumption is also not numerically validated (though it is a stated assumption, not a hidden one). Exponents alpha come from slope fits without error bars.\n\nWho it is for: people studying tensor-network dynamics, monitored circuits, and MIPT. The RMPS result is citable; the nullity order parameter needs substantial further work. My recommendation: send it to peer review, but ask the authors to validate the order parameter and provide error bars or code. Right now it reads like a promising but unfinished story.","headline":"A useful and partly original RMPS result, but the headline MIPT claim rests on an unvalidated proxy and needs stronger numerical support.","tokens_in":13943,"tokens_out":4428,"would_cite":true,"duration_ms":36049,"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 claims that fixed-bond-dimension matrix product states produced by random circuits retain a spectral memory of the dynamics, with correlations spreading as $\\xi_{\\mathrm{eff}} \\sim \\log \\chi^\\alpha$ and a measurement-induced…","keywords":["matrix product states","measurement-induced entanglement transition","transfer matrix spectrum","random quantum circuits","effective correlation length","Rényi mutual information","monitored quantum circuits","bond dimension truncation"],"falsifier":"Compute the variance of $\\log \\mathrm{Tr}(\\rho_X^2)$ across Haar realizations for random MPSs at $\\chi = 16, 32, 64, 128$; if the variance does not decay relative to the mean, the ensemble-average approximation that yields $\\alpha = 2$ fails. A second check is to extrapolate, in a brickwork monitored circuit, the location where $\\lim_{\\rho \\to 0^+} P(|\\lambda| < \\rho)$ turns on, and see whether $p_c(\\chi)$ converges to $0.16$ as $\\chi$ grows.","tokens_in":12908,"feed_emoji":"🧩","tokens_out":7501,"duration_ms":66373,"temperature":0.7,"pith_summary":"The paper studies ensembles of matrix product states (MPSs) obtained by running random quantum circuits and then truncating the state back to a fixed bond dimension $\\chi$ after each layer. It claims that these compressed states are not featureless: the spectrum of their transfer matrix carries a fingerprint of the original unitary dynamics, and the physically relevant correlation length is not the spectral-gap length but an effective length $\\xi_{\\mathrm{eff}} \\sim \\log \\chi^\\alpha$ whose exponent $\\alpha$ depends on the type of circuit. For random MPSs the paper derives analytically that $\\alpha = 2$, and for translation-invariant Haar random circuits it measures $\\alpha \\simeq 1.8$, a difference that comes from the concentration of small eigenvalues rather than from the largest eigenvalue. The central new result is that in circuits with projective measurements at rate $p$, the transfer matrix develops a null space for $p$ above a critical value, and the size of this null space acts as an order parameter for the measurement-induced entanglement transition, with the critical rate extrapolating to $p_c \\simeq 0.16$ as $\\chi$ grows.","feed_headline":"Measurement-induced transition survives bond-dimension truncation","feed_subtitle":"Transfer-matrix spectra keep the circuit's fingerprint: correlations spread as log χ^α with α encoding the dynamics.","key_machinery":"The load-bearing object is the MPS transfer matrix $T_i = \\sum_{\\sigma} (A^\\sigma_i)^* \\otimes A^\\sigma_i$, whose eigenvalues encode how correlations decay in space. The argument is carried by the $k$-replicated transfer matrices $T^{(k)}_{\\alpha}$, which enter the R\\'enyi-$k$ mutual information formula and whose full eigenvalue sum, not just the leading term, sets the effective correlation length. For random MPSs, Haar averaging with Weingarten calculus reduces these replicated matrices to small $k!\\times k!$ matrices, giving the analytic result $\\xi_{\\mathrm{eff}} \\sim \\log(\\chi^2)$. The null-space order parameter is the probability $P(|\\lambda| < \\rho)$ of transfer-matrix eigenvalues near zero; in the area-law phase this probability remains non-zero as $\\rho \\to 0$ because the MPS description becomes exact and many eigenvalues vanish.","core_discovery":"The central discovery is that truncating quantum dynamics to a finite bond dimension does not erase its physical content. For all ensembles considered, the full spectral density of the MPS transfer matrix, not just the leading eigenvalue, controls spatial correlations. In the volume-law regime, correlations spread over an effective distance $\\xi_{\\mathrm{eff}} \\sim \\log \\chi^\\alpha$, with $\\alpha = 2$ for random MPSs, $\\alpha \\simeq 1.8$ for translation-invariant Haar circuits, and $\\alpha$ decreasing as the measurement rate approaches the transition. In the monitored case, the relevant spectral feature is the density of near-zero eigenvalues: for $p$ above a $\\chi$-dependent critical rate, the probability $P(|\\lambda| < \\rho)$ remains finite as $\\rho \\to 0$, indicating a finite null space, and this quantity serves as the effective order parameter of the measurement-induced entanglement transition. The paper therefore claims that the MIPT can be detected in circuits whose states are constrained to have fixed bond dimension, at any finite $\\chi$, with $p_c(\\chi)$ approaching $0.16$.","pith_inferences":["The paper leaves implicit that if $\\xi_{\\mathrm{eff}} \\sim \\log \\chi^\\alpha$ is generic, then classical tensor-network simulations that truncate bond dimension capture the long-time entanglement structure meaningfully: the mutual information of the simulated state can be compared with exact evolution over a window growing like $\\log \\chi$.","The null-space density $\\lim_{\\rho \\to 0^+} P(|\\lambda| < \\rho)$ may serve as a practical finite-$\\chi$ diagnostic of the measurement-induced transition that is computed locally from the transfer matrix and could be less sensitive to boundary effects than direct entanglement-entropy estimates.","The same transfer-matrix logic could extend to two-dimensional tensor networks, where a finite null-space dimension might count constraints or topological content rather than a phase transition; the paper gestures at PEPS but does not develop this.","The side remark about random quantum channels suggests an analogous effective thermalization time growing like $\\log d \\, N$ in open systems, which would be a testable extension of the correlation-spreading result beyond pure state dynamics."],"forward_implications":["Correlation lengths read from the largest transfer-matrix eigenvalue alone will under-predict how far mutual information reaches in these states; the full spectral sum sets an effective length $\\xi_{\\mathrm{eff}} \\sim \\log \\chi^\\alpha$.","Different random circuit architectures can be distinguished by transfer-matrix spectral density even when their spectral gaps and spectral supports are identical, meaning the exponent $\\alpha$ encodes microscopic dynamical information.","The measurement-induced entanglement transition is visible in fixed-$\\chi$ MPS ensembles for any finite bond dimension, with the extrapolated critical rate $p_c(\\chi) \\to 0.16$ as $\\chi \\to \\infty$, so tensor-network simulations can detect the transition.","In the area-law phase, correlations stop spreading with $\\chi$, giving $\\xi_{\\mathrm{eff}} \\sim \\log \\chi^0$, consistent with an exact MPS description for $\\chi > \\bar{\\chi}(p)$.","The spectral densities of these truncated chaotic states are radially uniform rather than peaked around quasiparticle frequencies, so they require a different interpretive paradigm from MPSs near critical ground states."],"supporting_citations":[{"why":"Supplies the canonical MPS gauge and the transfer-matrix eigen-decomposition used to express correlation functions and the standard spectral-gap correlation length.","marker":"[35]"},{"why":"Establishes that spectra of random quantum channels, which the random-MPS transfer matrix realizes, concentrate in a disk of radius $1/\\sqrt{d}$.","marker":"[48]"},{"why":"Provides the Weingarten calculus used for the Haar averages that produce the analytic mutual-information formula and the exponent $\\alpha = 2$.","marker":"[53]"},{"why":"Introduces the shifted replicated transfer-matrix construction for random MPSs on which the $k=2$ analytic computation is built.","marker":"[30]"},{"why":"Defines the measurement-induced entanglement transition in hybrid circuits, the benchmark whose critical rate $p_c \\simeq 0.16$ the paper's null-space order parameter must reproduce.","marker":"[21]"},{"why":"Establishes the volume-law to area-law entanglement transition in monitored random circuits, providing the standard phase structure the paper compares against.","marker":"[23]"}],"fun_headline_variants":["MIPT survives bond-dimension truncation","Transfer-matrix spectra reveal circuit fingerprints","Entanglement transition detected at fixed bond dimension","Correlation spread exponent alpha encodes circuit type"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The calculation assumes that one typical random state behaves like the average: the log of the average purity moment can stand in for the average of the log, and after averaging the replicated transfer matrices look translation invariant, so the formula derived for translation-invariant MPSs applies to non-translation-invariant random MPSs.","fun_headline_variants_meta":{"raw":{"variants":["MIPT survives bond-dimension truncation","Transfer-matrix spectra reveal circuit fingerprints","Entanglement transition detected at fixed bond dimension","Correlation spread exponent alpha encodes circuit type"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000739,"raw_usage":{"total_tokens":3289,"prompt_tokens":923,"completion_tokens":2366,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":2312}},"tokens_in":539,"tokens_out":2366,"duration_ms":16723,"temperature":1.0,"reasoning_tokens":2312,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:23:07.196715+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the variance of $\\log \\mathrm{Tr}(\\rho_X^2)$ across Haar realizations for random MPSs at $\\chi = 16, 32, 64, 128$; if the variance does not decay relative to the mean, the ensemble-average approximation that yields $\\alpha = 2$ fails. A second check is to extrapolate, in a brickwork monitored circuit, the location where $\\lim_{\\rho \\to 0^+} P(|\\lambda| < \\rho)$ turns on, and see whether $p_c(\\chi)$ converges to $0.16$ as $\\chi$ grows.","supporting_citations":[{"cited_title":"Anticoncentration and state design of random tensor networks","cited_arxiv_id":"2409.13023","evidence_quote":"Establishes that spectra of random quantum channels, which the random-MPS transfer matrix realizes, concentrate in a disk of radius $1/\\sqrt{d}$."},{"cited_title":"Carollo and V","cited_arxiv_id":null,"evidence_quote":"Introduces the shifted replicated transfer-matrix construction for random MPSs on which the $k=2$ analytic computation is built."}],"review_version":1}