{"id":"bb2520f7-bddf-41e8-9dac-38ad5e08127b","arxiv_id":"2512.14583","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Sequential weak measurement records generically saturate in the mutual information they carry about the initial qubit state, and the paper derives scaling functions and perturbative formulas for the saturation plateaus.","lead":"This paper computes how much information a sequence of weak measurements carries about a qubit's starting state, for two practical readout schemes. It finds that the information plateaus below the ideal one bit and identifies the optimal number of measurements before readout stops improving.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central saturation claim rests on unproven finite-ξ conjecture; analytic 'bounds' are uncontrolled low-efficiency extrapolations.","rationale":"The reader's weakest assumption identifies the finite-ξ-to-imperfect-recovery step as the unproven bridge; my analysis agrees that this is the most load-bearing concern. The paper itself labels this step a conjecture, and the chain rule argument shows that the exponentially decaying marginal I(S,A_T) is insufficient to bound the total MI. The additional issue of extrapolating the low-efficiency expansion to η=1 further weakens the claim of obtaining 'bounds.' However, the numerical simulations for Model I and Model II directly exhibit plateau behavior below log 2, so the central claim is plausible for the studied cases. The concern does not warrant rejection; it supports a conditional verdict: the authors should either prove the finite-ξ conjecture or temper the language from 'bounds' to 'estimates.' Since the reader already recommended CONDITIONAL, I see no reason to change the verdict; I mark UNCHANGED.","tokens_in":28146,"tokens_out":9136,"duration_ms":481022,"concrete_test":"Using the authors' own Hoeffding-bounded MI estimator, compute the exact (η=1) total mutual information I(S;A_{1:T}) for a minimal non-commuting scheme without invariant subspaces, e.g., a deterministic alternating sequence of weak X and Z measurements on a qubit prepared in ±Z, for small x and T up to ~10^6. If MI approaches log 2 while ξ remains finite, the finite-ξ conjecture is refuted; if it saturates strictly below log 2 and the deficit does not vanish as x→0, the conjecture survives this test. As a secondary check, compare the η=1 numerical plateau against the η→0 analytical formula to test whether the low-efficiency extrapolation is quantitatively reliable at η=1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that under generic circumstances the information about the initial state saturates and does not allow perfect recovery—depends critically on the conjecture in Section IV B after Eq. (24): 'ξ being finite implies imperfect recovery of information about S.' The authors explicitly state they lack a rigorous way to establish the conditions under which information loss occurs. The preceding analytic result, I(S,A_T)=O(T e^{-2T/ξ}) (or O(e^{-2T/ξ})), concerns only the marginal mutual information of the T-th measurement. It does not control the conditional terms I(S,A_j|A_{j+1:T}) in the chain rule, which sum to the total MI. Therefore no upper bound on I(S,A_{1:T}) follows from finite ξ. The quantitative plateau values are also derived from a perturbative expansion in √η about η=0 (Section IV C) and then evaluated at η=1; this is an uncontrolled extrapolation, so the resulting curves are estimates, not bounds. Thus the abstract's phrase 'bounds on information extraction' overstates what is actually derived. The numerical evidence for the two specific models is credible and supports saturation, but the generalization to 'generic circumstances' and the claimed bounds are not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies how much information about a qubit's initial state can be extracted from a record of sequential weak measurements, focusing on two models: (I) informationally complete measurements of all three Pauli operators with no intrinsic dynamics, and (II) informationally incomplete Z-measurements with transverse-field unitary evolution. Using discrete-time simulations and continuous-time SME descriptions, the authors compute mutual information I(S,A_{1:T}) with Hoeffding-based error bounds, observe scaling collapse as a function of x^2 T (with φ/x^2 fixed), and identify plateaus below the ideal 1 bit. They supplement the numerics with analytic results: an exponential decay bound for the marginal mutual information of the T-th measurement characterized by a correlation length ξ, and a low-efficiency (η≪1) perturbative calculation of the mutual information plateau using a binary-input AWGN channel approximation. They connect these results to the Bayes-optimal readout and to overfitting in physics-agnostic learning. The central claim is that, generically, the measurement record does not allow perfect recovery of the initial state even as T→∞.","tokens_in":28475,"tokens_out":2554,"duration_ms":25283,"significance":"If the central claim holds, the paper establishes an important information-theoretic limitation for weak-measurement-based qubit readout that goes beyond specific estimation schemes. The numerical work is careful: MI estimates carry explicit Hoeffding concentration bounds, the scaling collapse is demonstrated across multiple parameter values, and the model-II analysis addresses a realistic non-QND readout setting. The analytic γ(t) formulas in Sec. IV C are parameter-free predictions (given τ, η, ω, φ) and their agreement with simulations in Fig. 3 is a genuine strength. The connection between information saturation and overfitting in supervised readout is practically relevant. The main weakness is that the analytic route to the saturation claim is incomplete: the finite-ξ conjecture in Sec. IV B is explicitly unproved, and the low-efficiency expansion is extrapolated to η=1 without control. Thus the paper is a valuable contribution with a credible central phenomenon, but the claimed 'bounds' are not yet fully established.","major_comments":[{"comment":"The central saturation claim—that finite ξ implies imperfect recovery—is stated as a conjecture: 'we conjecture that, in our problem, ξ being finite implies imperfect recovery of information about S from arbitrarily long sequence of measurements.' The preceding calculation bounds only the marginal mutual information I(S,A_T)=O(T e^{-2T/ξ}), not the total I(S,A_{1:T}). By the chain rule, Eq. (21), the total MI includes conditional terms I(S,A_j|A_{j+1:T}) that are not controlled by the marginal decay. Therefore no upper bound on the total MI follows from finite ξ. The abstract's phrase 'bounds on information extraction' overstates what is derived. To support the claim, the authors need either a proof of the conjecture (e.g., via a data-processing argument for the conditional terms) or a clearly stated weakening of the claim to 'numerical evidence suggests' for the specific models.","section":"Sec. IV B, after Eq. (24)"},{"comment":"The analytic mutual-information plateaus are obtained from a perturbative expansion in √η about η=0 (Eq. (29) and Appendix E) and then compared to simulations at η=1 in Fig. 3, with model II also at η=0.89, 0.5, 0.1. This is an uncontrolled extrapolation: the expansion parameter √η is not small at η=1. The agreement in Fig. 3(a,e) is empirical evidence, not a derivation. Calling these results 'bounds' (as in the abstract) is not justified; they are approximations. The manuscript should explicitly state this limitation and avoid the word 'bounds' unless a rigorous inequality is proven.","section":"Sec. IV C and Fig. 3"},{"comment":"The general claim that 'under generic circumstances' information saturates is inferred from two specific models, one with no intrinsic dynamics (Model I) and one with a single-axis measurement plus rotation (Model II). The conjecture in Sec. IV A about non-commuting Kraus operators is plausible but not established. The scaling function f(x^2T, φ/x^2) and the limit lim_{b→∞}f(b,a) are partly based on the 'naive scaling' assumption stated in Sec. III B. While the numerical collapse is convincing for the studied parameter range, the extrapolation to infinite T is not proven. The paper should carefully delimit the generality of the claim to the models and parameter ranges studied, or provide additional argument for the generic-case statement.","section":"Sec. III B and Sec. IV A"}],"minor_comments":[{"comment":"The definition of ξ via e^{-1/ξ}=max_{λ≠1}|λ| appears in Eq. (24) and again in Appendix B Eq. (B6). The notation is consistent, but the main text might benefit from a pointer to the derivation of the O(T e^{-2T/ξ}) bound, which is only sketched in Sec. IV B and detailed in Appendix B 1.","section":"Eq. (24) and Appendix B"},{"comment":"The error-kernel noise model p_success=(1+(n−1)√η)/n is a specific choice. The paper should note that other noise models could lead to different efficiency parameterizations; this would help the reader judge the generality of the η-dependence.","section":"Appendix D, Eq. (D1)"},{"comment":"In Eq. (32), the logarithms are presumably base 2 to give bits, but it is not explicitly stated. Please clarify the base of log in Eq. (32) and in the surrounding text.","section":"Sec. IV C, Eq. (32)"},{"comment":"Several typos: 'eignevalues' in Appendix B; 'the curved in (e-h) peal off the plateau' in Sec. III B; 'GSKL' should be 'GKSL' (Gorini–Kossakowski–Sudarshan–Lindblad). Also, in Sec. I B, 'information complete' should be 'informationally complete' for consistency.","section":"Throughout"},{"comment":"The ratio plots in Fig. 3(a,e) would be more informative with quantitative error bars or shaded regions reflecting the Hoeffding bounds of the numerical estimates, since the claim of agreement is based on these ratios.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"This is a solid and interesting paper with credible numerics and a clear practical motivation. The main issue is that the central analytic claim—the saturation bound—rests on an explicit conjecture and on an uncontrolled low-efficiency extrapolation. I think the paper is suitable for publication in a good quantum-information journal after the authors either (a) prove or substantially strengthen the finite-ξ conjecture, or (b) clearly reframe the central claim as a numerical finding with supporting analytic approximations, removing the word 'bounds' from the abstract. The current version overstates the rigor of the results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a real result, not a bounded proof. The key new findings are the scaling collapse I(S,A_{1:T}|x,phi) ≈ f(x^2 T, phi/x^2) and the low-efficiency plateau formula via a binary-AWGN channel. Both are supported by careful numerics — MI estimates are Hoeffding-bounded, the collapse is clean, and the gamma(t) formulas are parameter-free given the model and match simulations. That is worth taking seriously. The paper also gives a general Fano-based upper bound on any predictor's accuracy, a small but useful step beyond filter-specific readout bounds.\n\nNow the soft spots, in proportion. The analytic claim that finite xi implies imperfect recovery is a conjecture, stated honestly in Section IV B after Eq. (24). The decay of I(S,A_T) does not, by itself, control the conditional terms in the chain rule that make up the total MI. So the plateau is proven numerically for the two models studied, but not established \"under generic circumstances.\" Second, the quantitative plateau formula comes from a perturbation in sqrt(eta) about eta=0, then evaluated at eta=1. That is an uncontrolled extrapolation. It works surprisingly well in Fig. 3, but calling the result a \"bound\" (abstract) oversells it. The authors themselves mostly say \"estimate\" in the body; the abstract should match. I would also like to see the code and data — the Hoeffding bounds are nice, but reproducibility would be better with release.\n\nThe stress-test note is mostly right, but one caveat: the authors explicitly flag both gaps. This is not a hidden circularity. The derivation starts from the same SME parameters as the numerics, and no constants are fit to the target MI. So the circularity burden is low.\n\nWho is this for? Anyone working on dispersive readout, monitored qubits, or ML decoders for readout. The overfitting point for physics-agnostic classifiers is a useful practical warning. The paper deserves a serious referee — it is important enough, and the numerics are solid enough, that the right response is heavy revision, not a desk reject. I would ask for a proof or a sharpened conjecture for the finite-xi step, or a more honest abstract, and temper \"bounds\" to \"estimates\" unless the extrapolation can be controlled. Reading group: maybe; good for a discussion of information-theoretic limits, but I would pair it with a follow-up on what remains open.","headline":"Solid numerical evidence for a saturation phenomenon in sequential weak-measurement readout, but the analytic argument for a general bound rests on an explicit conjecture and a small-efficiency extrapolation; the abstract oversells the rigor.","tokens_in":28914,"tokens_out":2133,"would_cite":true,"duration_ms":18749,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P15","81P47","94A17"],"pacs":["03.65.Ta","03.67.-a"],"model":"deepseek-v4-flash","headline":"The paper claims that for generic sequential weak measurements, the mutual information between the initial qubit state and the measurement record saturates below one bit, so perfect readout is impossible however long the record is.","keywords":["weak measurement","qubit readout","mutual information","quantum trajectories","information extraction","measurement record","Bayes optimal estimation","quantum measurement limits"],"falsifier":"Compute exactly (e.g., by direct numerical simulation of the full channel) the mutual information I(S,A1:T) for Model II with ϕ≠0 at T ≫ ξ, with x small but finite; if it exceeds the plateau value and approaches log 2 — or any single non-commuting scheme with finite ξ achieves perfect asymptotic recovery — the saturation claim fails. Alternatively, an explicit construction of simultaneously diagonalizable but non-identity Kraus operators with an invariant subspace of dimension less than 2 that still yields mutual information reaching log 2 would refute the conjecture.","tokens_in":28026,"feed_emoji":"⚛️","tokens_out":3924,"duration_ms":32599,"temperature":0.7,"pith_summary":"The paper tries to establish a fundamental limit on reading out a qubit's initial state from a time-ordered record of weak measurements: for generic measurement schemes, the mutual information between the initial state and the measurement record saturates at a value strictly below one bit, so perfect recovery is impossible no matter how long the record is. It argues that this saturation is a generic feature of non-commuting weak measurement plus intrinsic dynamics, visible as a plateau in the scaling function f(x²T, φ/x²). The plateau value can be estimated analytically in a low-efficiency expansion, and it matches numerics. This matters because it sets an upper bound on any readout procedure, including machine-learning-based ones, and identifies an optimal measurement duration before extra data becomes noise.","feed_headline":"Weak measurement records cap qubit readout below 100 percent","feed_subtitle":"Mutual information saturates below one bit for generic weak measurements; longer records only add noise.","key_machinery":"The averaged single-step measurement channel E — the superoperator obtained by averaging the Kraus operators over measurement outcomes — and its subleading eigenvalue, which defines an exponential correlation length ξ via e^{-1/ξ} = max_{λ≠1}|λ|. Finite ξ means late measurements carry exponentially little information about the initial state. A second piece is the continuum scaling limit (T∼x^{-2}, ϕ∼x²) that collapses the numerical data onto scaling curves, and a third is the low-efficiency (η≪1) approximation that reduces the readout problem to a binary-input additive white Gaussian noise channel whose signal-to-noise ratio γ(t) yields an explicit mutual-information plateau.","core_discovery":"The central claim is that, under generic circumstances, information about the initial qubit state effectively saturates past a certain number of observations and does not allow perfect recovery. Concretely, in the scaling limit the mutual information obeys I(S,A1:T|x,φ) ≈ f(x²T, φ/x²), and for generic non-commuting cases lim_{b→∞} f(b,a) < log 2. As a result, the Bayes-optimal readout fidelity is strictly below 100% even with perfect knowledge of the dynamics and an unlimited record. The paper also shows that late measurements are nearly independent of the initial state, decaying exponentially with a correlation length ξ, and conjectures that finite ξ implies imperfect recovery. An analytic","pith_inferences":["The saturation bound likely extends to multi-qubit readout: any fixed non-commuting weak-measurement record defines a finite correlation length, so per-qubit information is bounded below the Holevo limit; the efficient-sampling mutual-information estimator in the paper is already adapted to that setting.","The correlation length ξ, computed from the Lindblad-averaged channel, could serve as a practical design criterion: hardware tuned so that the subleading eigenvalue of E approaches 1 should push the plateau toward one bit, while strongly non-commuting dynamics lowers it.","The plateau phenomenon is plausibly the measurement-side counterpart of dynamical purification and learnability transitions in monitored quantum circuits; a testable extension would be to look for a sharp learnability transition as the efficiency parameter η is tuned."],"forward_implications":["Beyond a timescale set by ξ, additional weak measurements add negligible information about the initial state; there is an optimal measurement duration for readout.","Any readout scheme, including machine learning classifiers, has accuracy strictly below 100% for generic non-commuting weak-measurement schemes, by Fano's inequality from the mutual-information bound.","Physics-agnostic supervised learning on long records overfits to late measurements that are effectively independent of the initial state; a Bayes-optimal classifier aware of the dynamics avoids this.","In some regimes weaker measurements extract more information than projective ones, producing nonmonotonic behavior in measurement strength."],"fun_headline_variants":["Weak measurements cap qubit readout below perfect","Qubit readout saturates below 100% with weak measurements","Info extraction limit for weak qubit measurements","Weak measurements set ceiling on qubit readout fidelity","Sequential weak measurements cannot fully recover qubit state"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The conjecture, stated after Eq. (24), that a finite correlation length ξ always implies imperfect recovery of initial-state information from an arbitrarily long measurement sequence.","fun_headline_variants_meta":{"raw":{"variants":["Weak measurements cap qubit readout below perfect","Qubit readout saturates below 100% with weak measurements","Info extraction limit for weak qubit measurements","Weak measurements set ceiling on qubit readout fidelity","Sequential weak measurements cannot fully recover qubit state"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000582,"raw_usage":{"total_tokens":2619,"prompt_tokens":830,"completion_tokens":1789,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":1711}},"tokens_in":574,"tokens_out":1789,"duration_ms":10405,"temperature":1.0,"reasoning_tokens":1711,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T16:00:44.579919+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute exactly (e.g., by direct numerical simulation of the full channel) the mutual information I(S,A1:T) for Model II with ϕ≠0 at T ≫ ξ, with x small but finite; if it exceeds the plateau value and approaches log 2 — or any single non-commuting scheme with finite ξ achieves perfect asymptotic recovery — the saturation claim fails. Alternatively, an explicit construction of simultaneously diagonalizable but non-identity Kraus operators with an invariant subspace of dimension less than 2 that still yields mutual information reaching log 2 would refute the conjecture.","supporting_citations":[],"review_version":1}