{"id":"d636ec51-8741-4ce6-883e-f58d7b7e202e","arxiv_id":"2506.06075","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Stepwise, one-parameter-at-a-time quantum estimation can beat the joint-estimation precision limit when the quantum Fisher information matrix is near singular.","lead":"This paper proposes a one-parameter-at-a-time strategy for quantum sensors and shows it can beat the standard all-at-once measurement limit when the parameters are hard to separate. If correct, the approach could restore quantum-enhanced precision in many-body sensors near phase transitions, where joint estimation loses its advantage.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stepwise precision bound in Eq. (3) omits the first-stage error that propagates into the second-stage estimate, so it is an unattainable lower bound; the sufficiency condition (6) and the L^{-1.8} scaling claim therefore do not establish that any stepwise protocol actually beats joint…","rationale":"The paper has three advertised results: a sufficient condition for the SE lower bound to fall below the JE lower bound, a many-body scaling advantage for the SE bound, and a Bayesian demonstration. The algebra behind the sufficient condition in SM Theorem 1 is internally consistent, and the Bayesian simulations are suggestive numerical evidence. However, the load-bearing analytical step is Eq. (3), which treats the two stages as if the second stage were a single-parameter estimation problem with lambda_1 known to the estimator. In a genuine two-step protocol, the second-stage estimate must absorb the uncertainty in the first-stage estimate. That propagation term is of order 1/m_1, i.e., of order 1/M after optimization, and is therefore not a negligible correction in the asymptotic regime the paper considers. As a result, Eq. (3) is a valid but loose lower bound rather than a tight, attainable precision bound; the theorem proves only that an unattainable SE bound can lie below a JE bound. The Gaussian counterexample makes the failure concrete: in a simple two-parameter model satisfying the QFIM condition, the actual stepwise estimator is strictly worse than the JE bound even though Eq. (3) predicts the opposite. I therefore agree with the reader's rejection, but would phrase the flaw as unattainability of the SE bound rather than invalidity of the inequality itself. The Bayesian implementations are interesting, but they compare biased posterior variances against a frequentist bound and do not implement the unbiased estimator whose variance Eq. (3) purports to bound, so they do not close this gap.","tokens_in":16104,"tokens_out":24468,"duration_ms":253750,"concrete_test":"In the qubit example of Fig. 1, implement a frequentist two-step maximum-likelihood estimator for parameters in the red region where Eq. (6) holds (e.g., alpha = pi/4, beta = 3*pi/8, lambda_1 = lambda_2 = 0.5): use m_1 outcomes of the x-basis measurement to estimate lambda_1, then m_2 outcomes of the z-basis measurement to estimate lambda_2 with lambda_1 replaced by its MLE, choosing m_1, m_2 near the authors' optimal gamma. Repeat for about 10^4 independent runs, compute the sum of variances of the two unbiased estimators, and compare M times this variance with the JE bound mu from Eq. (2) and with the SE bound mu_tilde from Eq. (5). If the empirical M*Var exceeds mu in a region where mu_tilde < mu, the analytical claim of beating joint estimation is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The first main result rests entirely on Eq. (3) (and SM Eq. (S14)): the claimed stepwise uncertainty is [Q^{-1}]_{11}/m_1 + 1/(m_2 Q_{22}). The first term is a legitimate scalar Cramér-Rao bound for estimating lambda_1 with lambda_2 as a nuisance. The second term is the single-parameter quantum Cramér-Rao bound for lambda_2 only when lambda_1 is known exactly. In the actual two-step protocol lambda_1 is replaced by the random estimate lambda_est_1, whose error is O(1/sqrt(m_1)). Any second-stage estimator inherits this error: whenever the second-stage outcome distribution depends on lambda_1, the plug-in estimator lambda_hat_2 contains an additional variance term C/m_1. Because m_1 = gamma M, this propagation term is O(1/M), the same order as the terms retained in Eq. (3), and it cannot be neglected precisely in the strongly correlated regime where condition (6) is invoked. Thus Eq. (3) is not a tight attainable precision bound for any unbiased stepwise procedure; it is an optimistic lower bound. The sufficiency condition (6) and the many-body scaling comparison in Fig. 3(e) therefore compare an unattainable SE bound with a (generally unattainable) JE bound, and they do not demonstrate that stepwise estimation actually beats joint estimation. A concrete Gaussian model with QFIM Q = [[1+c^2, c], [c, 1]], measurements x ~ N(lambda_1,1) and y ~ N(c lambda_1 + lambda_2,1), and an equal split m_1 = m_2 = M/2 illustrates the failure: the actual stepwise variance sum is 2(2+c^2)/M, whereas Eq. (3) gives 4/M; for c^2 = 5 the paper's condition predicts SE wins (4/M < 7/M), but the actual stepwise estimator is strictly worse than the JE bound (14/M > 7/M).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a stepwise estimation (SE) protocol for two-parameter quantum metrology: allocate m1 measurement rounds to estimate one parameter, then feed the resulting point estimate into a single-parameter estimation of the second parameter over m2 rounds. It derives a precision bound in Eq. (3) and a sufficient condition Eq. (6) under which this bound lies below the joint-estimation (JE) bound. The authors illustrate the claimed advantage for a qubit probe, a three-level Landau-Zener probe, a coherent-state Gaussian probe, and a many-body mixed Ising model, including Bayesian simulations. The main advertised results are that SE can beat JE near singular QFIMs and can restore quantum-enhanced scaling (claimed L^{-1.8}) near a first-order critical point where joint estimation scales only as L^{-1}.","tokens_in":16424,"tokens_out":4633,"duration_ms":52926,"significance":"If the central bound were valid, the paper would address two recognized problems in multiparameter quantum metrology: the non-attainability of the QFIM bound due to measurement incompatibility and the degradation of precision when the QFIM is near singular. The paper is explicit and offers several worked examples, including an analytic sufficiency condition and a discussion of the Holevo CRB in the supplemental material. However, the entire framework rests on Eq. (3), and that expression is not a valid lower bound for the two-stage protocol as described. The claimed sufficiency condition, the scaling comparison in Fig. 3(e), and the Bayesian demonstrations therefore do not establish that any stepwise procedure actually outperforms joint estimation. The idea is interesting and the examples are clearly presented, but the core technical claim needs a corrected derivation before the conclusions can be accepted.","major_comments":[{"comment":"The right-hand side of Eq. (3) is not a lower bound on the variance sum of the two-stage procedure described in the text. The first term is a legitimate scalar Cramer-Rao bound for estimating lambda1 with lambda2 as a nuisance parameter. The second term, 1/(m2 Q22(lambda_est1, lambda2)), is the single-parameter QCRB for lambda2 only when lambda1 is known exactly. In the actual protocol lambda1 is replaced by the random estimator lambda_est1 whose variance is O(1/m1). Any estimator of lambda2 based on the second-stage data inherits an additional variance term of order 1/m1 whenever the second-stage outcome distribution depends on lambda1. Since m1 = gamma M, this propagation term is O(1/M), the same order as the terms retained in Eq. (3). A concrete Gaussian model with QFIM Q = [[1+c^2, c], [c, 1]], measurements x ~ N(lambda1,1) and y ~ N(c lambda1 + lambda2,1), and m1 = m2 = M/2 gives an actual stepwise variance sum of 2(2+c^2)/M, whereas Eq. (3) predicts 4/M. The omitted c^2 term is of the same order and is largest exactly in the strongly correlated regime where Eq. (6) is invoked.","section":"Eq. (3) and SM Eq. (S14)"},{"comment":"The sufficiency condition Q12^2/(Q11 Q22) > 2*sqrt(2)-2 is derived by optimizing the invalid bound in Eq. (3), so the theorem does not establish that any unbiased stepwise estimator outperforms joint estimation. The algebraic steps from Eq. (S16) to Eq. (S23) are internally consistent, but they manipulate an incorrect starting inequality. The supplemental discussion of the Holevo CRB does not repair this gap because the SE bounds mu' and mu'' used in Eqs. (S24)-(S31) are still based on the same unjustified additivity assumption.","section":"Theorem 1 and Eq. (6)"},{"comment":"The claimed restoration of quantum-enhanced scaling near the first-order critical point of the mixed Ising model compares the joint bound mu ~ L^{-1} with the stepwise bound mu_tilde ~ L^{-1.8}. Since mu_tilde is computed from Eq. (3), the comparison is between a generally unattainable JE bound and an unattainable, overly optimistic SE bound. A corrected two-stage bound that includes the first-stage error propagation could scale differently, particularly near a near-singular QFIM where the propagation term is enhanced. The statement that SE 'restores' the quantum advantage is therefore not supported by the presented calculations.","section":"Fig. 3(e) and the L^{-1.8} scaling claim"},{"comment":"The Bayesian demonstration compares posterior variances against frequentist Cramer-Rao bounds without accounting for the contribution of the prior. Uniform priors with widths pi/5, 0.2, and 1 for the three examples are additional free inputs, and the true parameter values are preselected from the region where Eq. (6) holds. The observed posterior variances are not a test of the claimed SE bound: a Bayesian estimator with informative prior can appear to 'reach' a frequentist bound or even fall below it for reasons unrelated to the two-stage protocol. This undermines the third main result as a validation of the analytical claims.","section":"Bayesian Implementation (Fig. 4 and SM 'Bayesian Scheme')"}],"minor_comments":[{"comment":"The weight matrix in Eq. (4) is written as W1=diag(0,1); it should be W2=diag(0,1) to match the notation introduced for the two ordering strategies.","section":"Eq. (4)"},{"comment":"The phrase 'gap-closing []' contains an empty citation placeholder; a reference to the gap-closing condition or to a relevant review should be inserted.","section":"Introduction, Landau-Zener paragraph"},{"comment":"The Bayesian update formula writes lambda_est1 = sum over lambda1 of lambda1 p1(lambda1|x1), which is only valid for a discrete grid of parameter values; the continuous analogue should be stated for clarity, since the main text treats the parameters as continuous.","section":"SM Eq. (S33)"},{"comment":"The captions for the right-hand panels state that 'Var x M' is plotted, but the text in the main body sometimes refers to 'sum of Bayesian variances multiplied by M'; the notation should be made consistent so the reader can identify which quantity is compared with mu and mu_tilde.","section":"Figure 4 captions"}],"recommendation":"reject","confidential_remarks":"The paper is clearly written and the stepwise idea is potentially interesting, and the authors do provide explicit calculations for several models. My main concern is not the presentation but the validity of the core inequality: Eq. (3) omits the error propagation from the first stage to the second stage, and this is not a minor technicality because the omitted term is the same order in M as the terms retained. Since the sufficiency condition, the scaling claim, and the Bayesian 'validation' all rest on this inequality, the central result is not established. If the authors can derive a correct two-stage Cramer-Rao-type bound and rework the examples with that bound, a resubmission could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The stepwise idea is worth a look: splitting multiparameter estimation into sequential single-parameter steps is a natural way to dodge measurement incompatibility, and the paper works through several concrete models with a Bayesian implementation. That part is solid. But the analytical engine, Eq. (3) (and Eq. S14), is not a valid lower bound for the proposed two-stage procedure. The second term, 1/(m2 Q22), is the single-parameter QCRB for lambda2 only when lambda1 is known exactly. In the actual protocol lambda1 is replaced by a random estimate whose error is O(1/sqrt(m1)). For any second-stage measurement whose distribution depends on lambda1, that error propagates into the lambda2 estimate and contributes a variance term of order 1/m1, the same order as the terms kept in Eq. (3). The stress-test Gaussian model makes this concrete: with Q = [[1+c^2, c],[c,1]] and an equal split, the actual two-step variance sum is 2(2+c^2)/M, while Eq. (3) gives 4/M; for c^2=5 the paper's condition predicts SE wins, but the actual stepwise estimator is worse than the JE bound. So the sufficiency condition (6) and the L^-1.8 scaling claim compare an unattainable SE bound against the JE bound; they do not establish that any stepwise protocol beats joint estimation.\n\nWhat the paper does well: it is clearly written, the sequential protocol is well defined, and the worked examples — qubit, Landau-Zener, mixed Ising, coherent-state probe — give a concrete feel for where near-singular QFIMs arise. The Bayesian demonstrations are useful as existence proofs that some sequential protocol can perform well, but they compare Bayesian posterior variance against frequentist QCRBs, and they are preselected in regions where the (flawed) bound predicts an advantage. The necessary condition for the Holevo bound (Theorem 2) is a separate result and may survive, but it does not fix the sufficiency condition.\n\nBottom line: the core claim is not supported as written. The fix isn't trivial — you need to either include the propagation term in the SE bound (and then the condition will look different), or identify special cases where the second-stage measurement is insensitive to the first parameter, or prove a valid bound for a specific adaptive estimator. If the authors can do that, the idea could have legs. As it stands, this is a well-posed but unproven conjecture.\n\nRecommendation: I'd send it to peer review — the topic matters and the flaw is instructive — but with a clear expectation that the main theorem must be reworked. It should not be published in its current form.","headline":"Sensible idea undermined by a missing error-propagation term in the central bound; the paper is worth reviewing but not publishable as is.","tokens_in":17069,"tokens_out":2921,"would_cite":false,"duration_ms":30974,"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":"Stepwise estimation—splitting the measurement budget into two sequential single-parameter rounds—beats the joint-estimation quantum bound whenever the two-parameter quantum Fisher information matrix is nearly singular, and restores L^-1.8…","keywords":["stepwise estimation","multiparameter quantum metrology","quantum Fisher information matrix","joint estimation","Bayesian quantum sensing","many-body criticality","Landau-Zener model","mixed Ising model"],"falsifier":"Simulate two-step estimation with a first-stage Gaussian error of variance $σ1^{2}$ and a second stage whose Q22 is evaluated at the estimated parameter, including the propagation term (∂Q22/∂λ1)^2 $σ1^{2}$/(m2 $Q22^{2}$) in the total variance; if at parameters satisfying r > 0.828 the corrected stepwise total exceeds the joint bound µ/M, the central claim fails. At the mixed Ising point (1.9, 0.28), the SE variance must scale as $L^{-1}$.8, not $L^{-1}$.","tokens_in":15853,"feed_emoji":"⚛️","tokens_out":6003,"duration_ms":58532,"temperature":0.7,"pith_summary":"The paper proposes stepwise estimation (SE) for two-parameter quantum metrology: spend m1 of M measurement rounds estimating one parameter, then use the remaining m2 rounds to estimate the second parameter at the first estimate. It proves a sufficient condition, $Q12^{2}$/(Q11Q22) > 2√2−2 ≈ 0.828, under which the SE precision bound is strictly below the joint estimation (JE) bound, meaning SE is guaranteed to win when the quantum Fisher information matrix is nearly singular. In a mixed Ising many-body probe near the first-order critical point, the SE bound decreases as $L^{-1}$.8 while the JE bound only as $L^{-1}$, so the quantum-enhanced scaling that parameter correlations destroy for joint estimation is recovered. Bayesian simulations with fixed, nonadaptive measurements on qubit, three-level Landau-Zener, and mixed Ising probes reach or approach the SE bound and beat the JE bound. The paper's central claim is that incompatible measurements and near-singular Fisher matrices need not be a barrier: sequential single-parameter strategies can outperform the ultimate but generally unreachable joint-estimation limit.","feed_headline":"Stepwise sensing beats joint quantum limits near singular points","feed_subtitle":"Splitting measurements into two single-parameter rounds beats joint estimation and restores quantum scaling.","key_machinery":"The central object is the two-parameter quantum Fisher information matrix (QFIM), whose inverse sets the Cramér-Rao lower bound on the covariance of unbiased estimators. The key identities are the joint bound µ = Tr[$Q^{{-1}}$] and the stepwise bound ˜μ = min_γ [ (Q22/(Q11Q22−$Q12^{2}$))/γ + 1/((1−γ)Q22) ] (or the analogous expression with the order reversed), where γ = m1/M is the measurement-budget split. The off-diagonal correlation r = $Q12^{2}$/(Q11Q22) measures how close the QFIM is to singular; the proof of the sufficiency condition optimizes over γ and shows joint estimation can beat stepwise for all γ only when r ≤ 2√2−2. The many-body scaling result uses QFIM elements computed in the mixed Ising ground state, where criticality makes individual elements large but the matrix nearly singular, so joint estimation loses the quantum enhancement while stepwise estimation does not.","core_discovery":"For any two-parameter estimation problem, if the off-diagonal correlation of the quantum Fisher information matrix satisfies r = $Q12^{2}$/(Q11Q22) ≥ 2√2−2 ≈ 0.828, the stepwise estimation bound is strictly lower than the joint estimation bound: the optimal SE bound, minimized over how the M rounds are split, falls below Tr[$Q^{{-1}}$]/M. Because r = 1 for a singular QFIM, the criterion selects models whose QFIM is close to singular, precisely the regime where joint estimation fails. In the mixed Ising ground state near the first-order critical point at (1.9, 0.28), the SE bound scales as $L^{-1}$.8 whereas the JE bound scales only as $L^{-1}$, restoring the quantum-enhanced sensitivity that near-singular parameter correlations remove from joint estimation. The paper also states a complementary result: for D-invariant models with nearly diagonal QFIM, the saturable Holevo Cramér-Rao bound for joint estimation always beats any stepwise strategy, so the SE advantage is tied to strong parameter correlation. Concrete Bayesian implementations using fixed measurement bases confirm that the SE advantage is realizable, not merely formal.","pith_inferences":["The threshold 2√2−2 was derived for equal weighting of the two parameters (W = I2); if one parameter matters more than the other, the threshold would shift, and mapping how the threshold depends on the weight matrix is a natural extension of the paper's result.","The stepwise bound assumes no error propagation from the first-stage estimate to the second stage; a more detailed two-step model including the propagation term (∂Q22/∂λ1)^2 σ1^2/(m2 Q22^2) could shrink the practical advantage region, so testing this is an open problem.","The Bayesian comparison uses posterior variance against a frequentist Cramér-Rao bound without explicitly separating prior information from estimator bias; a fully Bayesian risk bound such as the van Trees inequality might give a fairer benchmark for the numerical demonstrations.","Theorem 2's statement that joint estimation wins near diagonal QFIMs is proven for D-invariant models; whether some non-D-invariant models admit stepwise advantages even near diagonal QFIMs remains an open question."],"forward_implications":["In any two-parameter quantum model with QFIM correlation r > 0.828, the standard joint-estimation precision bound is not the full story: a sequential strategy is guaranteed a lower error bound.","Near-singular QFIMs, which arise naturally in many-body impurity probes, optical spatial-modulation sensing, and private sensor networks, are precisely where stepwise estimation should be used instead of joint estimation.","In many-body critical probes such as the mixed Ising model near the first-order transition, stepwise estimation restores quantum-enhanced scaling: L^-1.8 instead of the L^-1 shot-noise scaling that joint estimation yields.","The Bayesian demonstrations indicate the advantage is practically accessible with fixed, simple measurement bases, so the benefit does not require adaptive feedback or complicated joint measurements.","The optimal budget split approaches γ ≈ 0.5 for large system sizes, so an approximately equal division of measurement rounds between the two parameters is near-optimal in the scalable regime."],"supporting_citations":[{"why":"Supplies the classical Cramér lower bound that underlies the multiparameter Cramér-Rao inequality in Eq. (1).","marker":"[61]"},{"why":"Supplies the Rao form of the Cramér-Rao bound used to define the joint-estimation precision bound µ/M.","marker":"[62]"},{"why":"Provides the Holevo Cramér-Rao bound, the asymptotically saturable version of the joint-estimation bound that the paper compares against in the near-diagonal regime.","marker":"[63]"},{"why":"Gives the semidefinite-programming formulation of the Holevo bound, cited as the standard way to compute saturable bounds when closed forms are unavailable.","marker":"[64]"},{"why":"Provides Suzuki's explicit two-parameter qubit formula for the Holevo bound, used for the analytical comparisons in the supplement.","marker":"[65]"},{"why":"Defines the D-invariant models used in Theorem 2, where the paper proves joint estimation beats stepwise estimation near diagonal QFIMs.","marker":"[68]"},{"why":"Supplies the QFIM for the coherent-state squeezing-and-rotation model used in the supplement to test stepwise versus joint estimation.","marker":"[73]"}],"fun_headline_variants":["Stepwise sensing beats joint limits when correlations dominate","Split measurements improve quantum metrology near critical points","Quantum estimation: stepwise wins over joint for strongly correlated parameters","Stepwise scheme restores quantum scaling lost in joint estimation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The stepwise bound treats the first-stage estimate as exact when computing the second-stage Fisher information, so the claimed advantage assumes no error from the first stage leaks into the second; if it does, the total error can be larger than the bound says.","fun_headline_variants_meta":{"raw":{"variants":["Stepwise sensing beats joint limits when correlations dominate","Split measurements improve quantum metrology near critical points","Quantum estimation: stepwise wins over joint for strongly correlated parameters","Stepwise scheme restores quantum scaling lost in joint estimation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000853,"raw_usage":{"total_tokens":3671,"prompt_tokens":870,"completion_tokens":2801,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":2737}},"tokens_in":486,"tokens_out":2801,"duration_ms":19528,"temperature":1.0,"reasoning_tokens":2737,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T06:03:37.966692+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate two-step estimation with a first-stage Gaussian error of variance $σ1^{2}$ and a second stage whose Q22 is evaluated at the estimated parameter, including the propagation term (∂Q22/∂λ1)^2 $σ1^{2}$/(m2 $Q22^{2}$) in the total variance; if at parameters satisfying r > 0.828 the corrected stepwise total exceeds the joint bound µ/M, the central claim fails. At the mixed Ising point (1.9, 0.28), the SE variance must scale as $L^{-1}$.8, not $L^{-1}$.","supporting_citations":[{"cited_title":"Cramér, Scandinavian Actuarial Journal 1946, 85 (1946)","cited_arxiv_id":null,"evidence_quote":"Supplies the classical Cramér lower bound that underlies the multiparameter Cramér-Rao inequality in Eq. (1)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Rao form of the Cramér-Rao bound used to define the joint-estimation precision bound µ/M."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Holevo Cramér-Rao bound, the asymptotically saturable version of the joint-estimation bound that the paper compares against in the near-diagonal regime."},{"cited_title":"Suzuki, Journal of Mathematical Physics 57 (2016)","cited_arxiv_id":null,"evidence_quote":"Provides Suzuki's explicit two-parameter qubit formula for the Holevo bound, used for the analytical comparisons in the supplement."},{"cited_title":"Bakmou and M","cited_arxiv_id":null,"evidence_quote":"Supplies the QFIM for the coherent-state squeezing-and-rotation model used in the supplement to test stepwise versus joint estimation."}],"review_version":1}