{"id":"063e9c3e-5b2d-48aa-b1c5-8072fd0cb18d","arxiv_id":"2607.06431","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Combining causal, retrodictive, and smoothed trajectories reconstructs the full conditional covariance of a continuously measured mechanical oscillator, removing the forward-backward symmetry bias that afflicts the conventional two-trajectory estimator.","lead":"A three-trajectory estimator — forward, retrodictive, and smoothed — recovers the conditional covariance of a continuously measured mechanical oscillator without the forward-backward symmetry assumption that biases the standard two-trajectory method. On a 7.71-mg suspended mirror it agrees with a fixed-parameter model prediction while the conventional estimator shows a covariance-space bias of about 3.5.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Experimental validation is a consistency test: f_eff, Q_eff, and δ are fixed from the same 10-s record used for verification, so correlated model error can preserve closure; abstract's 'fixed independently' overclaims.","rationale":"The theoretical derivation of the three-trajectory identity is mathematically sound for a correctly specified linear-Gaussian model; the orthogonality argument leading to Eq. (S33) is standard two-filter smoothing. The load-bearing weakness lies in the experimental validation: the key model parameters (f_eff, Q_eff, δ) are fixed from the same 10-s record that is then used to verify the covariance reconstruction. This creates a correlated-error path: any model misspecification that preserves the displacement spectrum will also preserve the closure between the empirical SV(ω) and the model SV(ω), so the reported agreement does not independently confirm the estimator's unbiasedness. The paper is transparent about model dependence in the discussion, but the abstract's wording 'fixed independently' and 'exact' outruns these qualifications. The reader's conditional verdict already captures this concern, and a leave-one-out reanalysis would provide a decisive check. No additional internal inconsistency or fatal flaw was found; the mathematical identity holds and the experimental analysis is honestly caveated. Therefore the verdict remains CONDITIONAL, and no change to the reader's assessment is needed.","tokens_in":31615,"tokens_out":12784,"duration_ms":122420,"concrete_test":"Leave-one-interval-out validation: Re-estimate f_eff, Q_eff, and δ using only intervals 1–5 of the six selected analysis intervals (or from auxiliary calibrations only, without the displacement spectrum), then apply the three-trajectory estimator to the held-out interval 6 and compute the finite-band closure χ²_V against the Riccati reference derived from the held-out parameters. If the held-out closure degrades significantly (e.g., p < 0.05), the reported agreement is an artifact of fixing parameters on the verification record; if the closure remains statistically consistent, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central experimental claim is that the three-trajectory reconstruction agrees with a Riccati prediction 'based on parameters fixed independently.' But the body text shows that key parameters — f_eff, Q_eff, and δ — were fixed from the ordinary displacement spectrum of the same 10-s record used for verification (Table S1; 'A common set of representative values, feff=283.5 Hz and Qeff=250, was fixed for all six intervals before evaluating any trajectory-difference covariance'). The Riccati reference and the three-trajectory estimator both use the same filters built from these parameters. If the reduced state-space model is misspecified in a way that leaves the displacement spectrum unchanged, the empirical SV(ω) and the matched finite-band model SV(ω) will still agree, because the parameters are effectively calibrated to reproduce that same spectrum. The paper itself concedes 'A single-output spectrum cannot, by itself, identify an arbitrary internal realization.' Consequently, the reported χ²_V = 0.66 (p = 0.884) and frequency-resolved χ²_SV = 86.2 (p = 0.198) confirm only that the estimator and the model are internally consistent; they do not independently verify that the reconstructed V is the true conditional covariance. The abstract's 'fixed independently' is therefore stronger than what the methods support, and the central validation claim is not established beyond the self-consistency level.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives and experimentally applies a three-trajectory estimator for the forward conditional covariance V of a continuously measured linear-Gaussian (optomechanical) system, using the causal (filtered), future-likelihood (retrodictive), and two-filter smoothed trajectories. The central identity, Eq. (3)/Eq. (S33), V = [Var(x_fwd−x_bwd) + Var(x_sm−x_fwd) − Var(x_sm−x_bwd)]/2, follows from the orthogonality of the smoothing error to record-derived trajectories and cancels the unknown future-likelihood covariance V^(E) without assuming V^(E)≃V, while recovering off-diagonal elements. On data from a 7.71-mg suspended-mirror cavity, the reconstructed covariance agrees with the matched finite-band Riccati prediction (χ²_V=0.66, p=0.884; frequency-resolved χ²_SV=86.2, p=0.198), whereas the conventional two-trajectory estimator shows the predicted systematic bias (d_M≈3.47±0.48 vs model d_M=3.65). The paper also predicts that the conventional estimator would misclassify the state as separable in a proposed bipartite entanglement configuration (Fig. 1 lower panel).","tokens_in":31896,"tokens_out":27449,"duration_ms":223274,"significance":"The theoretical core is clean and genuinely useful. Eq. (3), derived in Sec. S1 from the orthogonality of the smoothing error to record-derived trajectories, removes the V^(E)≃V assumption that limits the conventional retrodictive estimator, recovers the full covariance including off-diagonal terms, and is parameter-free within the linear-Gaussian model — no fitted bias correction is involved. The discrete-time sample-assignment subtlety is addressed correctly. The experimental execution is careful: 300 paired nuisance draws, coherent displacement-calibration rank-one propagation, paired experiment–model residuals, boundary trimming, and an alternative no-time-domain-filter reanalysis are all reported. The predicted covariance-space bias (d_M≃3.65) and the entanglement misclassification scenario (Fig. 1 lower panel) are falsifiable model-based predictions, and the experimental d_M=3.47±0.48 is consistent with them. The authors are candid about limitations: line-excised covariance optimism, the unmeasured boost-on frequency-noise residual, and the non-identifiability of a single-output spectrum. If the validation claims are reframed as a covariance-closure consistency test, this is","major_comments":[{"comment":"The abstract and Summary say the reconstruction 'agrees with a Riccati prediction based on parameters fixed independently.' The verification is a closed-loop consistency test: Table S1 shows f_eff, Q_eff, and δ were fixed from the ordinary displacement spectrum of the same 10-s record used for verification, and Sec. S3.4 gives S_V(ω)=W(ω)S_y(ω). Since the parameters were calibrated to reproduce S_y(ω), the matched finite-band prediction tracks the estimator largely by construction. The quoted statistics (χ²_V=0.66, p=0.884; χ²_SV=86.2, p=0.198) establish internal consistency and correct filter implementation, not independent certification of V; the paper concedes 'A single-output spectrum cannot, by itself, identify an arbitrary internal realization.' Please reframe the abstract/Summary accordingly.","section":"Abstract; Sec. S2.1/Table S1; Sec. S3.4"},{"comment":"The frequency-resolved test is presented as 'more restrictive' than integrated covariance agreement, but its conclusion depends strongly on the modeled off-diagonal structure of the residual covariance. In the no-time-domain-filter reanalysis (Sec. S2.2), χ²_SV=29.6 (ν=76, p≈1) with the full propagated covariance but χ²=155.5 (p=2.09×10^-7) with diagonal-only covariance; the paper calls the full covariance 'conservative.' The main-text result (χ²_SV=86.2, p=0.198) and the claim of 'no resolved systematic residual' therefore carry limited power and hinge on retained correlations. Please quantify the sensitivity of the main-text statistic to the covariance model, or temper the 'more restrictive' claim.","section":"Sec. S2.2; Fig. 4"}],"minor_comments":[{"comment":"The typeset text contains duplicated blocks: the opening of the Theory section, the experimental data-analysis paragraph, and the captions of Figs. 1–4 each appear twice. This must be cleaned in the revised version.","section":"Main text (general)"},{"comment":"Refs. [14] and [34] are the same Mayne paper; deduplicate. Ref. [28] should read 'Peres–Horodecki' with proper capitalization.","section":"References"},{"comment":"In Table S2 the experimental entry '0.400±1.09×10^3' is hard to parse; use consistent notation, e.g., (0.400±1.09)×10^3, and align the model/experiment columns.","section":"Sec. S2.2, Table S2"},{"comment":"The phrase 'exact linear-Gaussian estimator' should be glossed as 'exact within the calibrated linear-Gaussian model' to match the body's careful qualification in the Discussion.","section":"Abstract"},{"comment":"The conventional two-trajectory estimator is compared against the full-band Riccati target while the three-trajectory estimator is compared against the matched finite-band prediction. Because the estimators are band-limited (130 Hz–2 kHz), please also report the matched finite-band forward target for the two-trajectory comparison to separate band truncation from estimator bias.","section":"Results"},{"comment":"The ×4 'mode-mixing' factor in Eq. (S105) is carried over from Ref. [21]; please state whether any uncertainty is associated with this factor, since it scales the thermal-force noise level entering the filters.","section":"Sec. S2.3"}],"recommendation":"major_revision","confidential_remarks":"The central estimator is sound; the required changes are claim-calibration and robustness quantification rather than new theory. I see no grounds for rejection. Note that the Discussion already contains the appropriate framing ('Its advance is not model independence'); the abstract should be harmonized with it. The v2 text contains duplicated passages and duplicated references, suggesting editorial cleanup is needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: Eq. (3) is the real thing. It is a recombination of classical two-filter smoothing identities, but the combination—canceling the unknown future-likelihood covariance V^(E) while preserving the off-diagonal terms—is new to this literature, and I don't see a flaw in the algebra. The discrete-time sample-assignment detail is handled honestly. The experimental section is careful, with a real noise budget, paired nuisance draws, and unusually candid caveats, including the admission that the method is model-dependent and that the boost-on frequency-noise is not measured out of loop. Credit where due.\n\nThe soft spot is the validation. f_eff, Q_eff, and delta are fixed from the same displacement record that is then used to check the reconstructed covariance. The Riccati reference and the estimators use the same filters built from those parameters, so a model error that leaves the displacement spectrum unchanged shifts both sides in a correlated way. The paper's own statement—'a single-output spectrum cannot identify an arbitrary internal realization'—concedes the limit. The frequency-resolved SV test is more demanding than a single integrated number, and the chi-squares are fine, but it is still the same model against the same record. So the experiment demonstrates internal consistency, not an independent verification of V. The abstract's 'fixed independently' is not supported; the body's 'assigned from the displacement spectrum before covariance reconstruction' is the accurate description.\n\nI would not call this a fatal flaw. The estimator stands on its own as a theoretical contribution, and the experimental consistency is a legitimate demonstration that the protocol can be implemented and that the conventional estimator shows the predicted bias. But the paper gives more weight to the closure test than the evidence can carry. A revision should either rephrase the claims or, better, do a held-out test: fix parameters on one subset and verify on another, and find an out-of-loop measurement for the frequency-noise channel.\n\nThe paper deserves a serious referee. The derivation is crisp and the problem is important for entanglement verification in optomechanics. I'd send it to review with instructions to focus on the validation claims. For my own work I'd cite it for the estimator, with a note that the experiment is a consistency check.","headline":"The three-trajectory covariance identity is a real and clean result; the experimental 'verification' is a guarded consistency test, so the abstract's 'fixed independently' overstates what the paper shows.","tokens_in":32536,"tokens_out":1733,"would_cite":true,"duration_ms":18497,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E11","81V80"],"pacs":[],"model":"deepseek-v4-flash","headline":"A three-trajectory estimator recovers the conditional covariance of a continuously measured quantum oscillator without assuming forward–backward covariance symmetry.","keywords":["conditional covariance","quantum trajectory","Kalman filtering","retrodiction","smoothing","optomechanics","feedback cooling","entanglement verification"],"falsifier":"Generate a synthetic linear-Gaussian measurement record with a known forward covariance V and a deliberately asymmetric future-likelihood covariance V^(E), e.g., strong feedback damping plus colored thermal noise, then apply Eq. (3) and the conventional half-difference estimator: if the three-trajectory estimate does not recover the known V within statistical error, or if it still shows a bias comparable to the conventional estimator's dM ≈ 3.5, the identity or its implementation is wrong. On the experimental side, an out-of-loop measurement of the stabilized laser-frequency noise—or use of a","tokens_in":31372,"feed_emoji":"🪞","tokens_out":5177,"duration_ms":46868,"temperature":0.7,"pith_summary":"Continuous measurement can condition a macroscopic mechanical oscillator into a nearly pure state, but verifying the conditional covariance—the matrix that determines purity and entanglement—has been biased: the standard two-trajectory estimator assumes the future-record likelihood has the same covariance as the forward conditional state, an assumption that fails under feedback damping and colored noise. This paper derives an exact linear-Gaussian estimator that adds a third trajectory, the smoothed estimate, and combines the three pairwise trajectory differences so that the unknown future-likelihood covariance cancels algebraically. The result recovers the full forward covariance, including off-diagonal elements, without separate bias corrections. Applied to a 7.71-mg suspended mirror, the reconstruction agrees with a Riccati prediction based on independently fixed parameters (χ²V = 0.66, p = 0.884), while the conventional estimator exhibits a systematic covariance-space bias dM ≃ 3.5 that record length cannot remove. If correct, the method gives a pathway to unbiased macroscopic entanglement certification in tabletop and gravitational-wave-scale optomechanical systems.","feed_headline":"Three trajectories replace two to verify quantum mirror covariance","feed_subtitle":"A new identity cancels the unknown future-likelihood term that biases the standard two-trajectory estimator.","key_machinery":"The load-bearing object is the three-trajectory covariance identity, Eq. (3)/Eq. (S33): V = ½[Var(fwd−bwd) + Var(sm−fwd) − Var(sm−bwd)], built from the causal Kalman filter, the maximum-likelihood future-record retrodiction (the 'effect' variable), and the two-filter smoother. All three trajectories are constructed from the same measured record through a calibrated linear-Gaussian state-space model that includes auxiliary Ornstein–Uhlenbeck states for colored structural-damping thermal noise and a coherent laser-intensity-noise channel. The orthogonality property of the smoothing error—the residual of the complete-record conditional mean is uncorrelated with any function of the record—is wha","core_discovery":"The central claim is an identity: for a linear-Gaussian system under continuous measurement, the forward conditional covariance V equals one half times [Var(x_fwd − x_bwd) + Var(x_sm − x_fwd) − Var(x_sm − x_bwd)], where x_fwd is the causal Kalman estimate, x_bwd is the future-likelihood retrodiction, and x_sm is the two-filter smoothed estimate. The identity holds because the smoothing error is orthogonal to every record-derived trajectory, giving V = V(s) + Var(x_sm − x_fwd) and V^(E) = V(s) + Var(x_sm − x_bwd); combining with Var(x_fwd − x_bwd) = V + V^(E) cancels the unknown future-likelihood covariance V^(E). The estimator therefore does not require V^(E) ≃ V and retains the off-diagonal","pith_inferences":["Because the identity relies only on linear-Gaussian orthogonality and not on optomechanical specifics, it should transfer to any continuously monitored linear system—classical control plants, circuit-QED, or cold-atom sensors—where the future-likelihood covariance deviates from the forward covariance; a numerical simulation with a known asymmetric V^(E) would test this directly.","The experimental closure is a consistency test of the reduced state-space model as much as of the estimator: f_eff, Q_eff, and δ are fixed from the same 10-s record used for verification, so any model error enters the filters and the Riccati reference in a correlated way; a stronger test would use a held-out record or an out-of-loop measurement of the stabilized laser-frequency noise.","Since the two-trajectory bias is systematic and does not vanish with record length, previously reported conditional-covariance results obtained in feedback-cooled or detuned systems may need re-examination; the paper itself does not survey that body of work.","A natural extension is to apply the three-trajectory estimator to the bipartite two-mirror geometry of the paper's Fig. 1 and ask whether the reconstructed two-mirror covariance violates the positive-partial-transpose criterion, providing a direct macroscopic entanglement witness."],"forward_implications":["The systematic bias of the conventional two-trajectory estimator, quantified as dM ≈ 3.5 at the operating point, is removed without needing numerical correction factors.","All covariance elements, including the off-diagonal q–p correlations that cancel in the half-difference method, are reconstructed from the record.","Feedback cooling can be used to reduce macroscopic unconditional motion without invalidating the covariance estimate through a forward–backward symmetry assumption.","In the two-mirror entanglement geometry analyzed by the paper, the conventional estimator would classify some entangled configurations as separable (EN = 0 while the true value is EN = 0.1), so this estimator is needed for reliable entanglement certification.","The method applies to both milligram-scale tabletop mirrors and kilogram-scale gravitational-wave test masses, since it needs only a calibrated linear-Gaussian model and a single measurement record."],"fun_headline_variants":["Three trajectories beat two for unbiased mirror covariance","Quantum mirror covariance: three trajectories, zero bias","Three-way identity removes covariance bias in optomechanics","Unbiased quantum covariance from three Kalman trajectories","Two trajectories biased; three give exact mirror covariance"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reduced state-space model—the 1/f structural-damping thermal force approximated by three Ornstein–Uhlenbeck poles, feedback cooling absorbed into an effective damping rate, and no independently measured residual laser-frequency-noise state—must accurately describe the apparatus, and f_eff, Q_eff, and δ are fixed from the same 10-s record used for verification, so any model error enters both the estimator filters and the Riccati comparison in a correlated way.","fun_headline_variants_meta":{"raw":{"variants":["Three trajectories beat two for unbiased mirror covariance","Quantum mirror covariance: three trajectories, zero bias","Three-way identity removes covariance bias in optomechanics","Unbiased quantum covariance from three Kalman trajectories","Two trajectories biased; three give exact mirror covariance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000779,"raw_usage":{"total_tokens":3257,"prompt_tokens":700,"completion_tokens":2557,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":444,"completion_tokens_details":{"reasoning_tokens":2500}},"tokens_in":444,"tokens_out":2557,"duration_ms":18360,"temperature":1.0,"reasoning_tokens":2500,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T08:15:13.868269+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a synthetic linear-Gaussian measurement record with a known forward covariance V and a deliberately asymmetric future-likelihood covariance V^(E), e.g., strong feedback damping plus colored thermal noise, then apply Eq. (3) and the conventional half-difference estimator: if the three-trajectory estimate does not recover the known V within statistical error, or if it still shows a bias comparable to the conventional estimator's dM ≈ 3.5, the identity or its implementation is wrong. On the experimental side, an out-of-loop measurement of the stabilized laser-frequency noise—or use of a","supporting_citations":[],"review_version":2}