{"id":"f48581c0-ed6c-4165-badd-75ab436a0682","arxiv_id":"2508.11938","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Wavelet smoothing of periodograms is shown to give higher frequency resolution and moderate improvements in matched-filter SNR and Bayesian evidence over Welch PSD for GW150914, while median wavelet packet PSD (especially WDM) is claimed more robust for non-stationary noise.","lead":"This paper proposes using wavelet smoothing to estimate the noise power spectrum of gravitational wave detectors, aiming to beat the standard Welch method in accuracy and resolution. It also tests a wavelet-packet-based median method for non-stationary noise, claiming better robustness than the traditional median periodogram.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Wavelet smoothing PSD's claimed advantage over Welch is not statistically established because the significance level s is tuned on the same data to force Var[eps]=pi^2/6, making validation circular.","rationale":"The reader's weakest assumption (GGD) is plausible, but on closer reading the thresholds are empirical percentiles rather than a strict parametric GGD fit, so the GGD assumption is less brittle than it first appears. The more load-bearing issue is that the significance level s is selected on the same data using the circular criterion Var[eps]=pi^2/6, and the performance metrics (matched-filter SNR, Bayes factor, Q) are single-point evaluations with no uncertainty. A simulation against a known PSD with a train/test split would settle whether the wavelet smoothing method genuinely beats Welch. I therefore keep the reader's conditional verdict but emphasize the circular tuning and lack of statistical error bars.","tokens_in":12845,"tokens_out":10855,"duration_ms":125324,"concrete_test":"Run a Monte Carlo validation on simulated stationary Gaussian noise with a known PSD (e.g., an aLIGO design curve at 4096 Hz for 32 s). For 100 realizations, compute the wavelet smoothing PSD with the paper's nominal settings (Tukey alpha=0.08, sym7, 6 levels, reflection mode, soft thresholding) and the Welch PSD (4 s segments, 50% overlap, Blackman window). First, fix s=0.12 and measure the mean integrated squared error (MISE) and bias of log-PSD against the known PSD. Second, split the 100 realizations into a tuning set and a test set; choose s on the tuning set to match Var[eps]=pi^2/6, then evaluate on the test set. If the wavelet smoothing PSD does not yield significantly lower MISE than Welch, or if the tuned s performs no better on test data than a fixed s=0.1, the paper's central outperformance claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central stationary-noise claim (Section III) rests on the ability of percentile soft thresholding to remove the Gumbel fluctuation eps(f) in Eq. (9) while preserving the log-PSD. The significance level s is not predetermined; it is chosen so that the variance of the residual eps(f)=ln Pxx + gamma - ln(PSD_wavelet) is close to pi^2/6 (Section III.B, Table III). This is circular: s is tuned on the same 32 s GW150914 segment that is then used to demonstrate performance (Tables II and IV). The 'actual distribution' of eps(f) in Fig. 5 is therefore the residual after tuning, not an independent validation of the Gumbel model. Because the reported matched-filter SNR (19.38 vs 19.27) and Bayes factor (311.71 vs 285.50) are single-event point estimates with no uncertainty, and the quality factor Q is computed from a single PSD curve (whose variance across frequency bins conflates spectral shape with estimator variance), the claimed 'high frequency resolution and low variance' advantage over Welch is not statistically supported. The method may be reasonable, but the evidence provided does not rule out that the observed improvement is noise or overfitting.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes wavelet-based PSD estimation for gravitational-wave detector noise. For stationary noise, it applies discrete wavelet transform smoothing to the log-periodogram with percentile soft thresholding, avoiding the segmentation used by Welch's method, and claims this yields PSD estimates with both high frequency resolution and low variance, outperforming the Welch PSD in matched-filter SNR (Table II) and Bayesian evidence (Table IV). For non-stationary noise, it takes medians of wavelet packet coefficients, especially using the Wilson-Daubechies-Meyer (WDM) basis, and claims greater robustness than the median periodogram method (Fig. 8). The paper validates the stationary method on 32 s of GW150914 H1 data and the non-stationary method on 256 s of GW170817 L1 data. The theoretical starting point is the standard log-periodogram model of Eq. (9), where the fluctuation term follows a Gumbel distribution with variance π²/6.","tokens_in":13187,"tokens_out":4181,"duration_ms":51789,"significance":"If the claims were fully established, the wavelet-smoothing PSD would be a useful addition to GW data analysis, offering a computationally cheap non-parametric estimator with no segmentation loss, and the median WDM PSD could provide a robust option for non-stationary noise. The paper builds on a well-known statistical model (Eqs. 8–10) and demonstrates the method on public LIGO data, which is a strength. However, the current evidence is not yet at the level needed to support the main performance claims: the stationary analysis is based on a single event segment, the significance level is tuned on the same data used for evaluation, and no error bars are provided for the matched-filter SNR, quality factor, or Bayes factors. The non-stationary comparison is also performed on a single segment and the WDM basis is selected after seeing sym20 artifacts on that same dataset. The method is plausible, but the paper needs independent, repeated validation to establish the claimed advantages.","major_comments":[{"comment":"The significance level s is calibrated on the same 32 s GW150914 H1 segment that is later used for performance evaluation. The text states that s=0.12 is chosen so that Var[ε(f)] is close to π²/6 (see Fig. 5 and Table III), and then Tables II and III report matched-filter SNR and quality factor for that same segment. This is a circular evaluation: the method is tuned to the data and then compared on the same data. Please provide out-of-sample validation, e.g., synthetic stationary noise with known PSD, multiple independent noise realizations, or bootstrap/segmentation resampling, and report the distribution of Q, SNR, and Bayes factors rather than point estimates only.","section":"§III.B, Table III"},{"comment":"The quality factor is defined as Q=[E(PSD)]²/Var(PSD), but the text does not explain how E and Var are estimated. If they are computed across frequency bins of a single PSD curve, as appears to be the case, Q conflates spectral shape with estimator variance and cannot support the claim that the wavelet smoothing PSD is 'smoother' than Welch. Please specify the estimator (e.g., ensemble or segment-based) and provide uncertainty. Without this, the numerical difference (0.001528 vs 0.001353) is not interpretable.","section":"§III.C, Table II"},{"comment":"The WDM basis is adopted after observing spurious spectral lines with sym20 on the same GW170817 dataset (p. 8–9), and the robustness comparison in Fig. 8 is performed on the same non-stationary segment. This is a post hoc model selection: the claimed robustness of the median WDM PSD is not demonstrated independently. Please pre-specify the wavelet basis or test on a separate dataset, and add a quantitative robustness metric (e.g., distance to a reference PSD or injection-recovery performance) with error bars.","section":"§IV, Fig. 8"},{"comment":"The main denoising step relies on the assumption that the detail coefficients follow a generalized Gaussian distribution (GGD). The paper only shows qualitative PDF overlays in Fig. 3 and states the distributions 'generally conform' to Eq. (11). Since the theoretical distribution of ε(f) is Gumbel (Eq. 10), it is not obvious that thresholding based on a GGD fit removes the fluctuation term while preserving the true log-PSD. Please provide quantitative goodness-of-fit information or sensitivity tests. This is load-bearing for the reconstruction: if the GGD model is inaccurate for other detector noise, the reconstructed PSD will be biased.","section":"§III.B, Eq. (11)"}],"minor_comments":[{"comment":"Typo: 'empoly' should be 'employ'.","section":"§I"},{"comment":"Typo: 'centeral frequency' should be 'central frequency'.","section":"§II"},{"comment":"Please clarify what p(y) is in Eq. (12) — presumably the fitted GGD — and state explicitly how the threshold values in Table I are obtained from the empirical coefficient distribution.","section":"§III.B, Eqs. (12)–(13)"},{"comment":"The 'normalization factor' for the median wavelet packet PSD is mentioned but never defined. Please specify how the squared wavelet packet coefficients are normalized so that the resulting PSD is on an absolute power scale.","section":"§IV"},{"comment":"The conclusion states that 'only the significance level has a notable effect' on the wavelet smoothing PSD, but Table III shows that the wavelet basis changes both Q (0.001284–0.001595) and SNR (18.83–19.73). Please reconcile these statements.","section":"§V vs Table III"},{"comment":"Figure 2 would benefit from labeled axes with units and a clear indication of the frequency bin indices; Figure 3 would benefit from overlaid fitted GGD curves rather than only empirical densities.","section":"Figures 2–3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a relevant methodological problem and uses public LIGO data, which is positive. However, the main performance claims are based on a single event/segment, with tuning on the evaluation data and no error bars. This is a load-bearing issue that cannot be fixed by copyediting. The authors need to add independent validation, quantify uncertainty, and pre-specify or cross-validate the choice of significance level and wavelet basis. If that is provided, the paper could be publishable. At present, the evidence is insufficient to support the claimed advantages over the Welch method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version: this paper is a competent application of well-known wavelet denoising to PSD estimation for gravitational-wave data. The math is standard and the exposition is clear. But the performance claims are supported by one event and a calibration that is partly circular. The paper deserves a referee, but not at face value.\n\nWhat is new is the specific application: using percentile soft thresholding on the log-periodogram to get a stationary-noise PSD that keeps full frequency resolution, and using the median of wavelet packet coefficients for non-stationary noise. The WDM variant for non-stationary noise is a natural extension of Cornish's work. The authors show that on GW150914 their PSD gives a slightly higher matched-filter SNR (19.38 vs 19.27) and a higher Bayes factor (311.7 vs 285.5) than Welch. That is the kind of result that would interest the community.\n\nThe paper does several things well. It correctly derives the Gumbel fluctuation model for the log-periodogram, and the idea of using the known variance pi^2/6 to set the threshold is clever. The figures, especially Fig. 4, give a real feel for what wavelet smoothing does.\n\nThe soft spots are real. First, the significance level s is chosen on the same 32 s GW150914 segment that is later used for the SNR and Bayes factor comparisons. Tuning the threshold to make the residual variance match theory on the test data is a form of overfitting, and the observed improvement could be inflated. Table III shows that other wavelets give larger SNR gains (19.73 for sym6) but were not chosen; it is not clear why sym7 was selected. Second, there is only one stationary event and one non-stationary segment. SNR and Bayes factors from a single event have run-to-run scatter; the differences here are small, and no error bars or repeated analyses are given. Third, the WDM choice for non-stationary data is made after seeing sym20 produce artifacts on the same dataset. That is post-hoc selection, and the robustness claim is based on visual inspection. Fourth, there is no code or data release, although the data are public.\n\nNone of these issues is disqualifying. The method is plausible, and the statistical foundation is sound. But the evidence as presented is not enough to claim a proven advantage over Welch. The authors need to validate on multiple events, use hold-out data for any tuning, and report uncertainties.\n\nWho should read it? People working on PSD estimation in GW data analysis, and anyone interested in wavelet methods for spectral estimation. I would send it to a referee, but a careful one who will ask for independent validation. I would not cite it yet in my own work.","headline":"A competent application of known wavelet denoising to GW PSD estimation, with a clever variance-matching threshold, but the performance claims rest on a single event and a significance level tuned on the same data.","tokens_in":13680,"tokens_out":3219,"would_cite":false,"duration_ms":37265,"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":"Wavelet smoothing of the log-periodogram yields gravitational-wave noise PSDs with high frequency resolution and low variance at once, beating the Welch PSD; for non-stationary noise, a median WDM wavelet-packet PSD is more robust.","keywords":["power spectral density","Welch method","wavelet smoothing","wavelet packet transform","WDM transform","gravitational wave data analysis","matched filtering","non-stationary noise"],"falsifier":"Take simulated stationary noise with a known PSD containing a narrow line; compute wavelet smoothing PSDs over many realizations. If, with $s$ chosen so $\\mathrm{Var}[\\epsilon]\\approx\\pi^2/6$, the line's width is broadened or the residual noise variance does not approach $\\pi^2/6$ without bias, the central denoising claim fails.","tokens_in":12759,"feed_emoji":"🌊","tokens_out":9215,"duration_ms":97753,"temperature":0.7,"pith_summary":"Gravitational-wave data analysis depends on knowing the detector noise power spectral density (PSD). The paper tries to establish that wavelets can estimate this PSD better than the standard Welch method: take the full-length log-periodogram, remove its known Gumbel fluctuation term by wavelet thresholding, and reconstruct the log-PSD without segmenting the data. The resulting wavelet smoothing PSD keeps the periodogram's full frequency resolution while lowering variance, and on GW150914 H1 data it yields a higher matched-filter SNR and a larger Bayes factor than Welch. For non-stationary noise, the paper claims that taking the median of wavelet-packet coefficients in each frequency band—especially with the Wilson–Daubechies–Meyer basis—gives a more transient-robust PSD than the median periodogram. If these claims stand, PSD estimation becomes faster and more accurate across both detection and parameter estimation.","feed_headline":"Wavelet smoothing beats Welch for gravitational-wave noise spectra","feed_subtitle":"A wavelet trick keeps full frequency resolution while reducing variance, and a median variant suppresses noise transients.","key_machinery":"The load-bearing object is the log-periodogram identity $\\ln P_{xx}(f)+\\gamma=\\ln S_{xx}(f)+\\epsilon(f)$, with $\\gamma$ the Euler–Mascheroni constant and $\\mathrm{Var}[\\epsilon]=\\pi^2/6$. A discrete wavelet transform decomposes the left side into detail and approximation coefficients; percentile soft thresholding under a generalized Gaussian model removes the $\\epsilon$ fluctuation from each detail level, and the inverse transform reconstructs $\\ln S_{xx}(f)$. For non-stationary noise, the machinery is the wavelet packet transform read as an evolutionary power spectrum: taking the median across time of the squared coefficients in each frequency band yields the median wavelet packet PSD, and","core_discovery":"The paper's central discovery is that a periodogram can be denoised instead of averaged: after $\\ln P_{xx}(f)+\\gamma=\\ln S_{xx}(f)+\\epsilon(f)$, the fluctuation term is Gumbel with variance $\\pi^2/6$, so wavelet thresholding of detail coefficients removes $\\epsilon(f)$ while preserving resolution. On GW150914 H1 data, wavelet smoothing achieves quality factor $Q=0.001528$ versus $0.001353$, matched-filter SNR $19.38$ versus $19.27$, and on 128 s of off-source data a Bayes factor $311.71$ versus $285.50$ when compared with Welch. For non-stationary noise, a median wavelet-packet PSD—especially the median WDM PSD—is more robust than the median periodogram at lines near 60, 300, and 500 Hz.","pith_inferences":["Editorial inference: if the generalized-Gaussian model for detail coefficients is stable across detector noise realizations, the same variance-matching rule for $s$ could be used on other observatories and frequency bands without retuning; the paper only demonstrates it on one H1 stretch and one L1 stretch.","Editorial inference: the glitch-robustness demonstration in Fig. 9 suggests a practical transient-removal step—whiten with the median WDM PSD, smooth, and subtract—but the paper does not develop it as a detection pipeline.","Editorial inference: the resolution advantage could help continuous-wave searches that rely on long, quiet stretches, where Welch-style segmentation currently blurs narrow instrumental lines; testing on such a search would be a natural follow-up."],"forward_implications":["On the GW150914 H1 stretch tested, wavelet smoothing raises matched-filter SNR from 19.27 to 19.38 and frequency resolution from 1/4 Hz to 1/32 Hz compared with Welch (Table II).","In parameter estimation on 128 s of off-source data, wavelet smoothing raises the Bayes factor from 285.50 to 311.71 while using a finer 1/128 Hz resolution (Table IV).","The significance level can be chosen by a variance-matching rule: pick $s$ so that $\\mathrm{Var}[\\epsilon(f)]\\approx\\pi^2/6$, eliminating manual tuning of the smoothing strength.","Because no segmentation is required, wavelet smoothing can estimate a PSD from a single short stretch that would leave Welch with either high variance or poor resolution.","For glitchy data, the median WDM PSD suppresses transient-induced spectral bumps more effectively than the median periodogram, giving a PSD closer to the quasi-stationary reference."],"supporting_citations":[{"why":"Defines the Welch averaged-periodogram PSD that serves as the baseline for all comparisons.","marker":"[9]"},{"why":"Derives the log-periodogram representation and the variance $\\pi^2/6$ of the fluctuation term used to set the threshold.","marker":"[31]"},{"why":"Supplies the earlier wavelet-thresholding approach to power spectrum estimation that the paper extends to log-periodograms of GW data.","marker":"[32]"},{"why":"Provides the generalized Gaussian model for wavelet detail coefficients that justifies percentile soft thresholding.","marker":"[34]"},{"why":"Provides the wavelet shrinkage / soft-thresholding technique used to remove the fluctuation coefficients.","marker":"[38]"},{"why":"Introduces the WDM transform and the time-frequency / evolutionary-spectrum framework used for non-stationary noise.","marker":"[23]"},{"why":"Introduces the fast Wilson-Daubechies transform whose median-coefficient estimate is compared with the median periodogram.","marker":"[24]"},{"why":"Supplies the Bilby parameter-estimation pipeline used to obtain the Bayes-factor comparison between Welch and wavelet smoothing PSDs.","marker":"[13]"},{"why":"Supplies the matched-filter SNR definition used to compare detection performance.","marker":"[4]"}],"fun_headline_variants":["Wavelet smoothing of periodograms gives better GW spectra than Welch","For non-stationary noise, median wavelet packet PSD wins","GW150914: wavelet PSD yields higher SNR and Bayes factor than Welch","Drop Welch's averaging; denoise the periodogram with wavelets","Wavelet PSD: high resolution, low variance, and robust to transients"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the noise left in the log-periodogram can be separated from the true spectrum by assuming the wavelet detail coefficients follow a generalized Gaussian shape; if that shape assumption fails, the thresholding will distort the spectral estimate.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet smoothing of periodograms gives better GW spectra than Welch","For non-stationary noise, median wavelet packet PSD wins","GW150914: wavelet PSD yields higher SNR and Bayes factor than Welch","Drop Welch's averaging; denoise the periodogram with wavelets","Wavelet PSD: high resolution, low variance, and robust to transients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000329,"raw_usage":{"total_tokens":1672,"prompt_tokens":744,"completion_tokens":928,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":833}},"tokens_in":488,"tokens_out":928,"duration_ms":9178,"temperature":1.0,"reasoning_tokens":833,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T19:41:19.035437+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take simulated stationary noise with a known PSD containing a narrow line; compute wavelet smoothing PSDs over many realizations. If, with $s$ chosen so $\\mathrm{Var}[\\epsilon]\\approx\\pi^2/6$, the line's width is broadened or the residual noise variance does not approach $\\pi^2/6$ without bias, the central denoising claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Welch averaged-periodogram PSD that serves as the baseline for all comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Derives the log-periodogram representation and the variance $\\pi^2/6$ of the fluctuation term used to set the threshold."},{"cited_title":"Moulin, Wavelet thresholding techniques for power spectrum estimation, IEEE Transactions on Signal Pro- cessing 42, 3126 (1994)","cited_arxiv_id":null,"evidence_quote":"Supplies the earlier wavelet-thresholding approach to power spectrum estimation that the paper extends to log-periodograms of GW data."},{"cited_title":"Vidakovic, Statistical Modeling by Wavelets(John Wi- ley & Sons, Inc., 1999)","cited_arxiv_id":null,"evidence_quote":"Provides the generalized Gaussian model for wavelet detail coefficients that justifies percentile soft thresholding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the wavelet shrinkage / soft-thresholding technique used to remove the fluctuation coefficients."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the WDM transform and the time-frequency / evolutionary-spectrum framework used for non-stationary noise."},{"cited_title":"Necula, S","cited_arxiv_id":null,"evidence_quote":"Introduces the fast Wilson-Daubechies transform whose median-coefficient estimate is compared with the median periodogram."},{"cited_title":"Ashton et al., Bilby: A User-friendly Bayesian In- ference Library for Gravitational-wave Astronomy, The Astrophysical Journal Supplement Series 241, 27 (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies the Bilby parameter-estimation pipeline used to obtain the Bayes-factor comparison between Welch and wavelet smoothing PSDs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the matched-filter SNR definition used to compare detection performance."}],"review_version":1}