{"id":"bfd6c78b-3ff8-4c1b-b293-e4030e37b1f3","arxiv_id":"2608.11276","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A full-covariance Bayesian pipeline that combines time-domain TDI simulations with a frequency-domain likelihood recovers injected stochastic gravitational-wave backgrounds for a Taiji-like mission, though its detector-function validation is internal to the simulation code.","lead":"This paper builds and tests a data-analysis pipeline for the planned Taiji space gravitational-wave detector, combining realistic simulated time streams with a Bayesian spectral model that keeps all detector-channel correlations. A generalist might care because the method is designed to separate faint stochastic background signals from instrumental noise and galactic foregrounds, a key step for future space-based detectors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Hann-window normalization is absent from the PSD estimator and likelihood; without a window-gain correction the few-percent validation and matched recovery cannot be assessed.","rationale":"The reader identified the self-referential use of Triangle-Simulator and the first-segment-only direct validation as the weakest assumption. Those are legitimate limitations, acknowledged by the authors, and they affect the strength of the claim but not necessarily its internal correctness. The Hann-window normalization is a different, unacknowledged issue that directly bears on the quantitative few-percent claim and on the unbiased recovery. If the window gain is not compensated, the PSD estimator is biased by a constant factor of about 3/8, which would make the reported agreement and recovery impossible unless the same factor appears in both the simulated data and the calculated model. The paper does not specify such a factor, and the residual-variance derivation in Appendix B assumes unwindowed Fourier coefficients. The central claim is therefore not fully reproducible from the text. A concrete numerical check with white noise would settle the issue. My recommendation remains the reader's CONDITIONAL verdict, but the required condition should explicitly include a demonstration of the window-normalization correction; hence the verdict is unchanged.","tokens_in":25011,"tokens_out":9518,"duration_ms":88325,"concrete_test":"Inspect the released code or reproduce the pipeline: inject a stationary noise with known one-sided PSD P0 into Triangle-Simulator, apply the paper's exact Hann window and the normalization of Eq. (2.6), and average over 1000 realizations. If the averaged spectrum recovers P0 within about 1/sqrt(Nreal), the window gain is accounted for and the concern is resolved; if it returns approximately 0.375 P0, the normalization is missing and the validation and PE results are biased. Additionally, compare the residual scatter to Eq. (2.21) with and without windowing to confirm the reference curve.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing unaddressed issue is the Hann-window normalization. Section 2.2 applies a Hann window to every segment before the DFT, but the PSD estimator in Eq. (2.6), the covariance in Eq. (2.22), and the residual variance in Eq. (2.21) are written for an unwindowed periodogram. For a one-sided PSD P(f), the expected periodogram of N samples windowed by w_n is (sum w_n^2 / N) P(f), i.e., roughly 0.375 P(f) for Hann. The text never introduces the compensating factor 1/(sum w_n^2/N). If the factor is missing, the simulated spectra used in the validation and in the PE likelihood are biased low by a factor of about 8/3, so the claimed few-percent agreement and the unbiased recovery in Tables 2-3 would be spurious. Windowing also correlates neighboring Fourier bins, so the assumed independent-complex-Gaussian likelihood in Eq. (2.23) is not the exact distribution of the windowed coefficients; the reference curve in Eq. (2.21) therefore cannot be the correct residual scale. The paper cites Ref. [42] for the normalization, but does not state the window-gain correction, leaving the method irreproducible and the central quantitative claim unverifiable from the text alone.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a Bayesian spectral-inference framework for stochastic gravitational-wave backgrounds (SGWB) in Taiji-like missions, combining second-generation time-delay interferometry (TDI) time-domain simulations from Triangle-Simulator with a frequency-domain, full 3x3 XYZ covariance likelihood. The detector response and noise transfer functions are evaluated per one-day orbital segment, validated at the few-percent level against 1000-realization averages of component-isolated simulations for the first segment, and then used in nested-sampling parameter estimation for three matched configurations (equal-arm FD, equal-arm TD, unequal-arm TD). The reported benchmarks include a four-parameter recovery test, an eight-parameter Galactic-foreground plus astrophysical-background analysis, and a ten-parameter phase-transition sound-wave search.","tokens_in":25244,"tokens_out":3059,"duration_ms":34304,"significance":"If the central validation claim holds, the framework is a useful contribution to LISA/Taiji stochastic-background analyses: it retains the full XYZ covariance with segment-dependent response functions, provides a controlled comparison of FD and TD pipelines, and includes a careful finite-realization variance prediction in Appendix B. The paper also goes beyond single-component recovery by jointly fitting instrumental noise, a Galactic foreground, an astrophysical background, and a cosmological component, and it includes an explicit model-mismatch test that visibly demonstrates the dangers of an inconsistent detector model. The analytical derivations in Appendices B and C are clear and machine-checkable in principle, and the controlled injection setup with matched likelihoods is a strength.","major_comments":[{"comment":"A Hann window is applied to every segment before the DFT, but the PSD estimator in Eq. (2.6) uses the unwindowed normalization 2/(Tseg fs^2) and no window-gain correction is introduced anywhere in the text. For a Hann window, the expected periodogram of a stationary process is (sum w_n^2 / N) times the true one-sided PSD, approximately 0.375 P(f). Unless a compensating factor 1/(sum w_n^2/N) is applied in the code, the simulated spectra used in the validation and in the PE likelihood are biased low by a factor of about 8/3. Since all quantitative claims—the few-percent validation, the unbiased recovery in Tables 2 and 3, and the evidence values—depend on the absolute scale of the spectral estimates, the paper must state the window-gain factor explicitly and, if it is missing, rerun the affected analyses.","section":"§2.2, Eq. (2.6)"},{"comment":"The likelihood in Eq. (2.23) models the windowed Fourier coefficients as independent circular complex Gaussians with covariance C^kappa(fk)=Tseg fs^2/2 P^kappa(fk). Even after a correct window-gain normalization, Hann windowing introduces correlations between neighboring frequency bins, so the coefficients are not exactly independent and the covariance is not exactly diagonal. The paper should state that Eq. (2.23) is an approximate Whittle likelihood for windowed data, quantify or justify the approximation (e.g., by checking the effective number of independent bins), and confirm that the validation residual reference in Eq. (2.21) remains the correct scale under windowing.","section":"§2.3, Eqs. (2.22)–(2.23)"},{"comment":"The direct simulation-to-calculation validation is performed only for the first one-day segment of the unequal-arm orbit, while all 360 segments enter the parameter-estimation likelihood with their own orbit-dependent response and noise transfer functions. The paper does not validate that the segment-dependent functions are accurate at later times, where the orbital configuration and arm lengths differ. The authors should either extend the validation to a sample of segments across the year or provide a quantitative argument that the first-segment agreement is representative.","section":"§2.2 and §3"},{"comment":"The manuscript correctly states that the validation is not fully independent because Triangle-Simulator is used both to generate the TDI streams and to evaluate the orbit-dependent response and noise transfer functions. This limits the strength of the validation: it can detect implementation inconsistencies but cannot certify that the simulator itself matches a real detector. The authors should explicitly discuss this limitation in the conclusions and indicate whether any independent checks (e.g., analytic equal-arm limits, comparison with another TDI code, or public LISA simulation outputs) were performed.","section":"§2.2 (self-disclosed circularity)"}],"minor_comments":[{"comment":"The abstract and several headings contain non-ASCII ligature artifacts (e.g., 'efficient') that should be cleaned for production.","section":"Abstract and §2"},{"comment":"The sentence introducing Eq. (2.21) says the curves show the predictions 'multiplied by Nreal' and that they give the expected values of the plotted Nreal(r^kappa_IJ)^2. This is clear, but the text later refers to 'the matched unequal-arm TD expectation from Eq. (2.21) multiplied by Nreal'; consider consistently calling it the scaled expectation to avoid confusion.","section":"§2.2, Eq. (2.21)"},{"comment":"The caption states the ratio diagnostic is applied to the full first-segment XYZ covariance with all components included, but the main text does not reference Figure A.7 in the validation discussion; a cross-reference would help the reader locate this important combined check.","section":"Figure A.7 caption"}],"recommendation":"major_revision","confidential_remarks":"The Hann-window normalization issue is the primary technical concern. It is likely fixable if the implementation already contains the compensation factor, but the manuscript as written does not document it, making the central quantitative claims unverifiable from the text alone. The circularity of the validation is disclosed honestly and is a common situation in this field, but the paper should temper its claims accordingly and, if possible, add an independent cross-check. I recommend major revision rather than rejection because the framework and the statistical machinery appear sound and the missing normalization can be corrected and the analyses rerun."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: worth taking seriously, but the written method has a hole in it. The central idea—segment-dependent full 3x3 XYZ covariance matched to time-domain TDI simulations—is the right way to check whether FD spectral likelihoods survive realistic unequal-arm data. The paper does something genuinely useful: it validates ACC, OMS, and response functions component-wise against 1000 Triangle-Simulator realizations, gives a finite-realization variance prediction in Appendix B, and runs matched equal-arm FD, equal-arm TD, and unequal-arm TD pipelines on controlled injections. The deliberate mismatch test is a nice touch. The authors are also candid that the validation is not independent of Triangle-Simulator.\n\nBut there is a load-bearing omission: the Hann window. Section 2.2 says every segment is Hann-windowed before the DFT, yet Eq. (2.6) and (2.17) use the unwindowed normalization 2/(Tseg fs^2), with no (sum w_n^2)/N factor. For a Hann window that factor is about 0.375, so the PSD estimate is biased low by roughly 8/3 unless compensated. If the compensation is in the code but not the text, the few-percent agreement and unbiased recovery are not reproducible as written; if it is absent, those results are wrong. This is the first thing I'd ask the authors. Windowing also correlates neighboring bins, so the independent-complex-Gaussian likelihood in Eq. (2.23) and the residual reference in Eq. (2.21) are approximate—probably acceptable, but should be stated.\n\nSecondary soft spots: the direct simulation-to-calculation check is only the first one-day segment, while the likelihood uses 360 segments; and no code or data is released. Neither kills the paper, but both limit how much confidence an outsider can place on the numbers.\n\nOverall, the framework is useful and the paper deserves a serious referee. I'd send it to review with a request to state the exact window normalization, show it in the equations, and ideally include one validation segment from later in the year. If the normalization is clarified, I'd cite this.","headline":"Useful simulation-to-inference framework for Taiji SGWB analyses, but the written method omits the Hann-window gain correction, so the central few-percent validation claim needs clarification before the numbers can be trusted.","tokens_in":25783,"tokens_out":3454,"would_cite":false,"duration_ms":31839,"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":"Full 3×3 channel covariance recovers Taiji's stochastic gravitational-wave background.","keywords":["stochastic gravitational wave background","time-delay interferometry","Bayesian inference","Taiji","full covariance likelihood","response functions","noise transfer functions","phase transition gravitational waves"],"falsifier":"Generate a one-year Taiji-like dataset with an independently written simulator that uses different orbital-propagation and beam-pattern conventions, inject a known power-law astrophysical background plus noise, and analyze it with the paper's likelihood using transfer functions computed from that independent simulator. If the realization-averaged spectra deviate from the calculated functions by more than a few percent outside the TDI-null notches, or if the injected background amplitude falls outside the posterior credible intervals in a meaningful fraction of realizations, the few-percent-consistency claim is falsified.","tokens_in":24796,"feed_emoji":"🛰️","tokens_out":9565,"duration_ms":87328,"temperature":0.7,"pith_summary":"This paper establishes a simulation-to-inference pipeline for stochastic gravitational-wave background searches in a Taiji-like space mission. The authors divide simulated second-generation time-delay-interferometry X, Y, Z streams into one-day segments and model each segment's Fourier coefficients with a segment-dependent complex 3×3 covariance built from orbital response functions and instrumental-noise transfer functions. Controlled simulations show that the calculated spectra match the ensemble-averaged simulated spectra at the few-percent level away from the TDI null frequencies. Using a matched full-covariance likelihood, Bayesian parameter estimation recovers injected astrophysical-background, galactic-foreground, instrumental-noise, and, in the ten-parameter benchmark, phase-transition sound-wave parameters consistently across equal-arm frequency-domain, equal-arm time-domain, and unequal-arm time-domain configurations. If this holds for real mission data, the simplified equal-arm frequency-domain likelihood is a trustworthy baseline for detectability forecasts as long as the inference model is kept consistent with the data-generation geometry.","feed_headline":"Full XYZ covariance recovers Taiji's gravitational-wave background","feed_subtitle":"Segment-by-segment 3×3 covariance matches simulated TDI spectra at a few percent and recovers injected signals.","key_machinery":"The load-bearing object is the segment-dependent Hermitian spectral matrix $P^\\kappa_{IJ}(f) = N^\\kappa_{IJ}(f) + S_h(f)R^\\kappa_{IJ}(f)$, where $N^\\kappa_{IJ}$ collects the post-TDI acceleration-noise and optical-metrology noise transfer functions and $R^\\kappa_{IJ}$ is the sky- and polarization-averaged response to an isotropic background, both evaluated for the one-day orbital configuration of segment $\\kappa$. The inference layer is the Whittle Gaussian likelihood over retained Fourier coefficients, $\\ln L = -\\sum_{\\kappa,k}[\\ln\\det(\\pi C^\\kappa(f_k)) + (\\tilde d^\\kappa(f_k))^\\dagger (C^\\kappa(f_k))^{-1}\\tilde d^\\kappa(f_k)]$, with $C^\\kappa = (T_{\\rm seg} f_s^2/2)P^\\kappa$. Because the fixed A, E, T rotation does not diagonalize the covariance for unequal, time-varying arms, the full complex 3×3 matrix is retained, and bins within 0.010 Hz of the nominal null frequencies $f_m = mc/(4L)$ are masked where the transfer functions are small and rapidly varying.","core_discovery":"The paper's central claim is that segment-dependent response and noise transfer functions, inserted into the full 3×3 X-Y-Z covariance of a Whittle likelihood, reproduce the realization-averaged spectra of second-generation TDI simulations within the finite-realization scatter (about 3 percent for 1000 realizations) over the retained band away from TDI nulls. In matched parameter-estimation runs, the static equal-arm frequency-domain, equal-arm time-domain, and unequal-arm time-domain configurations all return posterior means close to the injected astrophysical-background amplitude and spectral index after marginalizing over an effective Galactic foreground; the ten-parameter phase-transition benchmark also recovers the sound-wave peak amplitude and peak frequency. A deliberate model-mismatch test, in which unequal-arm time-domain data are fit with an equal-arm spectral model, displaces or pushes several noise and foreground posteriors to the prior boundary; the paper presents this as evidence that the simulation and inference geometries must be internally consistent.","pith_inferences":["Because the same simulation package generates the time-domain streams and supplies the transfer functions used in the likelihood, the few-percent agreement is a test of internal consistency; an independently implemented simulator with different orbital and beam-pattern conventions could expose model error larger than the paper measures.","The notch mask around the TDI null frequencies removes exactly the bands where unequal-arm effects are most visible, so a future analysis using alternative TDI combinations with fewer nulls could recover information the current likelihood discards.","The same segment-dependent full-covariance construction should transfer to other space-based interferometer concepts once their orbits and noise models are supplied, but the matched-pipeline comparison would need to be repeated for each geometry.","The small residual offset in the optical-metrology noise parameter found in the ten-realization validation suggests that aggressive high-frequency notching weakens noise calibration; a real search may need an informative prior on that amplitude or extra high-frequency data."],"forward_implications":["The static equal-arm frequency-domain likelihood can serve as the fast baseline for Taiji stochastic-background forecasts, since it gives posterior precision, SNR, and Bayesian-evidence trends consistent with the full unequal-arm time-domain pipeline.","Unequal-arm time-domain data should not be analyzed with an equal-arm spectral model: the paper's mismatch test shows that such a model can displace instrumental-noise and foreground parameters to the prior boundary.","An astrophysical background with amplitude as low as $\\log_{10}\\Omega_{\\rm ast} = -12.5$ remains strongly favored over the null model after marginalizing over instrumental noise and the effective Galactic foreground, with $\\ln BF$ above 115 in all three configurations.","A phase-transition sound-wave component at peak amplitude $\\log_{10}\\Omega_0 = -11.5$ and peak frequency $10^{-2.25}$ Hz is recoverable in the full ten-parameter model, though the evidence is weaker in the unequal-arm configuration ($\\ln BF \\approx 5.8$) than in the equal-arm ones ($\\ln BF \\approx 10$)."],"supporting_citations":[{"why":"supplies the second-generation TDI simulation package that generates the X, Y, Z time streams used for validation and parameter estimation.","marker":"[36]"},{"why":"derives the post-TDI acceleration-noise and optical-metrology noise transfer functions for the full XYZ spectral matrix.","marker":"[63]"},{"why":"provides the isotropic-SGWB response-function formalism used in Eq. (2.16).","marker":"[30]"},{"why":"supplies the detector sensitivity and noise-model conventions adapted to the Taiji benchmark.","marker":"[31]"},{"why":"defines the three spectral components (Galactic foreground, astrophysical power law, sound-wave template) used in the likelihood.","marker":"[74]"},{"why":"fixes the finite-segment Fourier normalization and Whittle-likelihood convention.","marker":"[42]"},{"why":"demonstrates segmented full non-diagonal covariance inference for stochastic backgrounds, the direct methodological predecessor of this analysis.","marker":"[37]"}],"fun_headline_variants":["Full XYZ covariance nails Taiji GW background","3×3 covariance matches TDI to few percent","Taiji's full covariance recovers injected signals","Segment-wise XYZ covariance recovers GW background"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulation package used to generate the time-domain data streams is a correct and complete model of the Taiji detector, because the same package also supplies the orbital response and noise transfer functions used in the likelihood, and the paper explicitly notes that this comparison is not fully independent.","fun_headline_variants_meta":{"raw":{"variants":["Full XYZ covariance nails Taiji GW background","3×3 covariance matches TDI to few percent","Taiji's full covariance recovers injected signals","Segment-wise XYZ covariance recovers GW background"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000989,"raw_usage":{"total_tokens":4206,"prompt_tokens":971,"completion_tokens":3235,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":3177}},"tokens_in":587,"tokens_out":3235,"duration_ms":21497,"temperature":1.0,"reasoning_tokens":3177,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:12:55.639306+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a one-year Taiji-like dataset with an independently written simulator that uses different orbital-propagation and beam-pattern conventions, inject a known power-law astrophysical background plus noise, and analyze it with the paper's likelihood using transfer functions computed from that independent simulator. If the realization-averaged spectra deviate from the calculated functions by more than a few percent outside the TDI-null notches, or if the injected background amplitude falls outside the posterior credible intervals in a meaningful fraction of realizations, the few-percent-consistency claim is falsified.","supporting_citations":[],"review_version":1}