Pith. sign in

REVIEW 2 major objections 5 minor 5 cited by

The wavelet-domain noise covariance matrix for gravitational wave data is well approximated as diagonal whenever the noise power spectrum changes slowly across each time-frequency pixel, with off-diagonal terms set by the derivatives of the

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 22:22 UTC pith:J3W6DLXU

load-bearing objection First explicit WDM noise covariance derivations in two limits; the general slowly-varying claim is honest but unproven — still worth a referee. the 2 major comments →

arxiv 2511.10632 v2 pith:J3W6DLXU submitted 2025-11-13 gr-qc

Non-stationary noise in gravitational wave analyses: The wavelet domain noise covariance matrix

Neil J. Cornish This is my paper
classification gr-qc
keywords gravitational wave data analysiswavelet packet transformnoise covariance matrixnon-stationary noiselocally stationary noisedynamic power spectrumtime-frequency analysislikelihood
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper shows that for the WDM wavelet packet basis, the noise covariance matrix—the object whose inverse appears in gravitational wave likelihood calculations—remains nearly diagonal when noise properties vary slowly in both time and frequency. The off-diagonal terms are controlled by the derivatives of the dynamic spectral model S(f,t), and stay at the one-percent level when the fractional change of S(f,t) across a pixel is about ten percent. If correct, long-duration searches could use a diagonal wavelet-domain noise matrix, avoiding dense matrix inversions, and optionally add a few off-diagonal stripes for accuracy. The same qualitative picture is argued to apply to other discrete wavelet transforms with filters compact in time and frequency.

Core claim

The paper derives the structure of the WDM noise covariance matrix in two tractable limits: wide-sense stationary colored noise and uncorrelated non-stationary noise. In both cases the matrix is banded and diagonal dominant, with off-diagonal entries proportional to the logarithmic derivatives of the spectrum in frequency (s_k) and time (µ_k). Taylor expanding the dynamic spectrum about each wavelet pixel shows that correlations between pixels vanish when the spectrum is locally flat, and remain small when the spectrum changes by less than roughly ten percent across a pixel. The paper concludes that for locally stationary noise the WDM noise correlation matrix is well approximated as diagona

What carries the argument

The central object is the WDM wavelet packet basis, a complete orthogonal transform that tiles the time-frequency plane into pixels of area ΔTΔF=1/2 and uses compact time-domain filters with smoothed top-hat frequency windows. The argument is carried by Taylor-expanding the dynamic spectral model S(f,t) about each pixel center and computing the resulting noise correlation matrix. Orthogonality of the wavelets makes the zeroth-order matrix diagonal for locally flat spectra; the leading off-diagonal terms are set by the dimensionless logarithmic slopes s_k and µ_k.

Load-bearing premise

The load-bearing premise is that for noise varying in both time and frequency, cross-derivative terms such as ∂_t∂_f S(f,t) remain small enough that the diagonal approximation derived from the two separate limits still holds; the paper states this extrapolation rather than proving it.

What would settle it

Construct a locally stationary noise process whose dynamic spectrum has comparable time and frequency gradients—for instance S(f,t)=S_0 (f/f_0)^α exp(β (f−f_0)(t−t_0)/T)—compute the WDM noise covariance matrix numerically with the window parameters used in the paper, and check whether any off-diagonal entry exceeds roughly one percent when the fractional change of S per pixel is about ten percent.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Long-duration gravitational wave searches can treat the WDM noise covariance matrix as diagonal when the noise drift is slow, cutting the cost of likelihood evaluations dramatically.
  • The leading off-diagonal correlations can be predicted from gradients of S(f,t); keeping about ten stripes captures the most significant terms at modest extra cost.
  • Steep power spectra in low-frequency layers, sharp spectral lines, and glitches break the diagonal approximation; the paper recommends pre-whitening or feature subtraction before the wavelet transform.
  • The same diagonal-dominance picture is expected to generalize to other discrete wavelet transforms with filter functions compact in time and frequency.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the diagonal approximation survives realistic locally-stationary noise, wavelet-domain likelihoods become nearly separable across pixels, which could simplify stochastic sampling and grid-based parameter estimation for signals lasting days to years.
  • A direct numeric test with a dynamic spectrum S(f,t) having comparable time and frequency gradients would settle whether cross-derivative terms stay subdominant—the main open gap in the paper's argument.
  • A practical self-calibrating scheme could estimate S(f,t) from the data, compute the derivative-driven stripes once, and iterate; the approximate banded inverse needs no expensive matrix operations.
  • If the diagonal approximation holds, it may also simplify transforming waveform models: signals only need to be computed in the wavelet basis once, rather than re-whitened by a dense covariance matrix at every likelihood call.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper studies the noise covariance matrix of the Wilson-Daubechies-Meyer (WDM) wavelet packet transform for non-stationary, colored Gaussian noise. It derives analytic expressions for two limiting cases: wide-sense stationary (WSS) colored noise, where the dynamic spectrum S depends only on frequency, and uncorrelated non-stationary noise, where it depends only on time. In these limits it shows that the wavelet-domain covariance matrix is approximately diagonal when the fractional change in S across a wavelet pixel is small, with off-diagonal terms proportional to dimensionless derivatives of S. For locally flat spectra the matrix is exactly diagonal. The paper then argues that the same diagonal approximation holds for general locally stationary noise, discusses pre-whitening strategies for steep spectra or fast time variation, and proposes a striped, banded approximation with a leading-order inverse. The central practical claim is that so long as S(f,t) varies by less than about 10% across a pixel, off-diagonal elements are at the percent level and can be neglected or included with a few off-diagonal stripes.

Significance. If the central claim is correct, this paper provides a useful theoretical foundation for moving gravitational-wave analyses into the wavelet domain in the presence of slowly varying non-stationary noise. The derivations for the two separable limits are clear and explicit, and the locally-flat limit recovering a perfectly diagonal matrix is an important sanity check. The quantitative smallness parameters (s_k, mu_k) are also well chosen. The main weakness is that the general locally stationary case—where S varies simultaneously in time and frequency—is not derived, and the paper explicitly relies on an unproved extrapolation. Because the paper's abstract and introduction state the general condition as its main result, this gap is load-bearing. The paper also provides no code or numerical validation, though the analytic results are reproducible in principle.

major comments (2)
  1. [§III.A, Eq. (26) and §III.B, Eq. (35)] The central claim—that the WDM covariance matrix is approximately diagonal whenever S(f,t) varies slowly across a wavelet pixel in both time and frequency—is not derived for the general locally stationary case. Only the two separable limits S(f) (Eq. 26) and S(t) (Eq. 35) are treated. The paper states 'Presumably, the general expression will also include cross terms such as \partial_t\partial_f S(f,t)' and 'It may be possible to generalize the calculations presented here.' This is an explicit admission of missing support. Since the practical recommendation—diagonal noise covariance for 10% per-pixel variation—relies on this combined case, please provide either a general derivation (e.g., using the formalism of Ref. [40]) or a numerical demonstration for a model with simultaneous time and frequency dependence, verifying that the off-diagonal terms are the sum of the two limits plus subdom
  2. [§III.A, Eq. (26) and §III.B, Eq. (35)] The numerical coefficients in Eqs. (26) and (35) (e.g., 0.21, 0.0058, 0.25, etc.) are central to the quantitative statements that a 10% variation yields ~1% off-diagonal elements. However, the details of the numerical evaluation are not given beyond the Meyer window parameters (d=6, A=0, B=∆Ω) and q=8 in the time-domain case. No code or pseudocode is provided, and the integrals are not specified sufficiently for independent reproduction. Please include a reproducible script (or at least a complete description of the discretization and summation ranges) so readers can verify the quoted coefficients and their dependence on the wavelet parameters.
minor comments (5)
  1. [Abstract] Typo: 'timeandfrequency' should be 'time and frequency'.
  2. [§III.B, after Eq. (34)] In the definition of µ_k, the text says 'd f^k' but it should be 'dt^k'. This is confusing in a time-domain Taylor expansion.
  3. [§III.A, after Eq. (26)] Typo: 'these old amount to∼20%' should be 'these would amount to ~20%'.
  4. [§IV] The sentence 'The expressions given in equations (26) and (35) are valid when either S(f) or S(t)' is grammatically incomplete. It should read '...are valid when S depends only on f or only on t.'
  5. [§IV, Eq. (37)] The approximation C^{-1} ≈ D^{-1} - D^{-1}OD^{-1} is the first-order Neumann series for a diagonally dominant matrix. The convergence condition (e.g., ||D^{-1}O|| < 1) is not stated. While the paper's examples satisfy this, making the condition explicit would aid readers applying Eq. (37) to more extreme spectra.

Circularity Check

0 steps flagged

No significant circularity: the WDM covariance results are direct analytic derivations; the one self-citation only sets non-critical numerical prefactors.

full rationale

The paper's central claims are derived from stated noise models, not fitted or assumed. For WSS colored noise, the covariance is computed analytically from E[x~j x~*k] = (1/2) δjk S[j], followed by a Taylor expansion of S(f) about the pixel center; the resulting off-diagonal coefficients in Eq. (26) are consequences of that expansion. For uncorrelated non-stationary noise, the same is done from E[x[i]x[j]] = δij σ²[i], yielding Eq. (35). No parameter is fit to data and then renamed a prediction. The extension to general locally stationary noise is explicitly left open: 'Presumably, the general expression will also include cross terms such as ∂t∂f S(f,t).' That is an acknowledged extrapolation, not a circular reduction. The only self-citation used for numerical evaluation is Ref. [13], which supplies the Meyer-window values d=6, A=0, B=ΔΩ; however, the paper states 'The results changed by a just a few tens of percent across a wide range of values,' so this parameter choice is not load-bearing. Thus the derivation is self-contained and any concern about generality is a completeness/correctness issue, not circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

No free parameters are fitted to data; the only hand-chosen numbers are wavelet design parameters and the example q=8. The main circularity-adjacent concern is the unproved extrapolation to the general locally stationary case, which is an axiom rather than a fitted parameter.

free parameters (2)
  • Meyer window parameters (d, A, B) = d=6, A=0, B=ΔΩ (adopted from Pearson & Cornish [13])
    Hand-chosen design parameters; numeric prefactors in Eqs. (26) and (35) depend on them. The paper states the results change by only tens of percent over a wide range, so they are not tuned to data.
  • Time-domain filter half-width q = q=8 in the example
    Sets the temporal support of the wavelet filter (T_filt=2qΔT). The fall-off of temporal correlations in Eq. (35) depends on q; q=8 is an example rather than fitted to a measurement.
axioms (4)
  • standard math WDM wavelets form a complete orthonormal basis (Eq. 3), and the transform can be computed via FFT.
    Used for the expansion (Eq. 1) and coefficient formulas (Eq. 9,10); standard wavelet theory [14,21].
  • domain assumption After glitch removal, the Gaussian noise component is zero-mean and locally stationary.
    This is the standard LIGO/Virgo model (ref [4]) and the premise that motivates time-frequency bases.
  • domain assumption The dynamic spectrum S(f,t) changes slowly enough that Taylor expansions (22) and (33) truncated at low order are valid over a wavelet pixel.
    The diagonal approximation and off-diagonal estimates assume |s_k|,|µ_k| ≪ 1; if this fails, Eq. (26)/(35) don't apply.
  • ad hoc to paper The general locally stationary colored-noise case is a benign combination of the WSS colored-noise limit and the uncorrelated non-stationary limit; cross terms ∂_t∂_f S(f,t) are subdominant.
    No derivation of the mixed case is given; the paper extrapolates from §III.A and §III.B and says 'Presumably, the general expression will also include cross terms...' in §IV. This is the main unproved premise behind the abstract's broad claim.

pith-pipeline@v1.3.0-alltime-deepseek · 10063 in / 14525 out tokens · 118827 ms · 2026-08-03T22:22:10.286877+00:00 · methodology

0 comments
read the original abstract

Gravitational wave detectors produce time series of the gravitational wave strain co-added with instrument noise. For evenly sampled data, such as from laser interferometers, it has been traditional to Fourier transform the data and perform analyses in the frequency domain. The motivation being that the Fourier domain noise covariance matrix will be diagonal if the noise properties are constant in time, which greatly simplifies and accelerates the analysis. However, if the noise is non-stationary this advantage is lost. It has been proposed that the time-frequency or wavelet domain is better suited for studying non-stationary noise, at least when the time variation is suitably slow, since then the wavelet domain noise covariance matrix is, to a good approximation, diagonal. Here we investigate the conditions under which the diagonal approximation is appropriate for the case of the Wilson-Daubechies-Meyer (WDM) wavelet packet basis, which is seeing increased use in gravitational wave data analysis. We show that so long as the noise varies slowly across a wavelet pixel, in both time {\em and} frequency, the WDM noise correlation matrix is well approximated as diagonal. The off-diagonal terms are proportional to the time and frequency derivatives of the dynamic spectral model. The same general picture should apply to other discrete wavelet transforms with wavelet filters that are suitably compact in time and frequency. Strategies for handling data with rapidly varying noise that violate these assumptions are discussed.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Modeling non-stationary noise: applications in gravitational wave astronomy

    gr-qc 2026-07 conditional novelty 6.0

    A positive dynamic spectrum S(f,t) generalizes the stationary power spectrum by defining Gramian closed-form noise covariances in Fourier and Wilson-Daubechies wavelet bases for gravitational wave data.

  2. An explicit and differentiable Wilson-Daubechies-Meyer transform for gravitational-wave data analysis

    gr-qc 2026-06 conditional novelty 4.0

    An explicit, GPU-ready WDM wavelet-packet package reproduces frequency-domain LISA galactic-binary posteriors under stationary noise and documents the full discrete construction.

  3. The WDM Time-Frequency Transform in Gravitational-Wave Data Analysis I: Formalism

    gr-qc 2026-06 unverdicted novelty 3.0

    Derives and documents the WDM basis properties, forward/inverse transforms, and time-frequency likelihood for gravitational-wave data analysis.

  4. The WDM Time-Frequency Transform in Gravitational-Wave Data Analysis I: Formalism

    gr-qc 2026-06 unverdicted novelty 3.0

    A self-contained formalism for the WDM time-frequency basis in gravitational-wave analysis, including transforms, localization, edge effects, and the time-frequency noise likelihood.

  5. An explicit and differentiable Wilson-Daubechies-Meyer transform for gravitational-wave data analysis

    gr-qc 2026-06 unverdicted novelty 3.0

    Open-source WDM transform package with JAX support and numerical validation of equivalence to frequency-domain likelihoods for a LISA binary under stationary noise.

Reference graph

Works this paper leans on

40 extracted references · 4 canonical work pages · cited by 3 Pith papers

  1. [1]

    Buikema et al

    A. Buikema et al. (aLIGO), Phys. Rev. D102, 062003 (2020), 2008.01301

    Pith/arXiv arXiv 2020

  2. [2]

    Acernese et al

    F. Acernese et al. (Virgo), Phys. Rev. Lett.123, 231108 (2019)

    2019

  3. [3]

    A. G. Abac et al. (LIGO Scientific, VIRGO, KAGRA) (2025), 2508.18082

    Pith/arXiv arXiv 2025

  4. [4]

    B. P. Abbott et al. (LIGO Scientific, Virgo), Class. Quant. Grav.37, 055002 (2020), 1908.11170

    Pith/arXiv arXiv 2020

  5. [5]

    Evans et al

    M. Evans et al. (2021), 2109.09882

    Pith/arXiv arXiv 2021

  6. [6]

    Abac et al

    A. Abac et al. (ET) (2025), 2503.12263

    Pith/arXiv arXiv 2025

  7. [7]

    Colpi et al

    M. Colpi et al. (LISA) (2024), 2402.07571

    Pith/arXiv arXiv 2024

  8. [8]

    Klimenko, S

    S. Klimenko, S. Mohanty, M. Rakhmanov, and G. Mitsel- makher, Phys. Rev. D72, 122002 (2005), gr-qc/0508068

    Pith/arXiv arXiv 2005

  9. [9]

    Klimenko et al., Phys

    S. Klimenko et al., Phys. Rev. D93, 042004 (2016), 1511.05999

    Pith/arXiv arXiv 2016

  10. [10]

    N. J. Cornish (2020), 2009.00043

    Pith/arXiv arXiv 2020

  11. [11]

    M. C. Digman and N. J. Cornish, Phys. Rev. D108, 023022 (2023), 2212.04600

    Pith/arXiv arXiv 2023

  12. [12]

    M. C. Digman and N. J. Cornish, Astrophys. J.940, 10 (2022), 2206.14813

    Pith/arXiv arXiv 2022

  13. [13]

    Pearson and N

    N. Pearson and N. J. Cornish (2025), 2509.05479. 7

    Pith/arXiv arXiv 2025

  14. [14]

    Necula, S

    V. Necula, S. Klimenko, and G. Mitselmakher, Jour- nal of Physics: Conference Series363, 012032 (2012), URLhttps://doi.org/10.1088%2F1742-6596% 2F363%2F1%2F012032

    2012

  15. [15]

    Karhunen, Ann

    K. Karhunen, Ann. Acad. Sci. Finnicae, Ser. A1, 34 (1946)

    1946

  16. [16]

    Lo` eve,Probability theory.(Van Nostrand, Prince- ton, N.J., 1960), URLhttps://catalog.hathitrust

    M. Lo` eve,Probability theory.(Van Nostrand, Prince- ton, N.J., 1960), URLhttps://catalog.hathitrust. org/Record/000580390

    arXiv 1960

  17. [17]

    Mallat, G

    S. Mallat, G. Papanicolaou, and Z. Zhang, The Annals of Statistics26, 1 (1998), ISSN 00905364, 21688966, URL http://www.jstor.org/stable/119978

    1998

  18. [18]

    Tenorio and D

    R. Tenorio and D. Gerosa, Phys. Rev. D111, 104044 (2025), 2502.11823

    Pith/arXiv arXiv 2025

  19. [19]

    M. Du, Z. Luo, and P. Xu, Phys. Rev. D112, 083036 (2025), 2506.10599

    Pith/arXiv arXiv 2025

  20. [20]

    Daubechies, S

    I. Daubechies, S. Jaffard, and J. L. Journe, SIAM J. Math. Anal.22, 554 (1991)

    1991

  21. [21]

    Meyer,Ondelettes et op´ erateurs: Ondelettes, Actualit´ es math´ ematiques (Hermann, 1990), ISBN 9782705661250, URLhttps://books.google.com/ books?id=2vtBuwEACAAJ

    Y. Meyer,Ondelettes et op´ erateurs: Ondelettes, Actualit´ es math´ ematiques (Hermann, 1990), ISBN 9782705661250, URLhttps://books.google.com/ books?id=2vtBuwEACAAJ

    1990

  22. [22]

    A. M. Sintes and B. F. Schutz, Phys. Rev. D58, 122003 (1998), gr-qc/9810004

    Pith/arXiv arXiv 1998

  23. [23]

    L. S. Finn and S. Mukherjee, Phys. Rev. D63, 062004 (2001), [Erratum: Phys.Rev.D 67, 109902 (2003)], gr- qc/0009012

    arXiv 2001

  24. [24]

    Kimpson, S

    T. Kimpson, S. Suvorova, H. Middleton, C. Liu, A. Melatos, R. J. Evans, and W. Moran, Phys. Rev. D 110, 122004 (2024), 2412.01058

    Pith/arXiv arXiv 2024

  25. [25]

    T. B. Littenberg and N. J. Cornish, Phys. Rev. D 91, 084034 (2015), URLhttps://link.aps.org/doi/ 10.1103/PhysRevD.91.084034

    doi:10.1103/physrevd.91.084034 2015

  26. [26]

    Siegel, M

    H. Siegel, M. Isi, and W. M. Farr, Phys. Rev. D111, 044070 (2025), 2410.02704

    Pith/arXiv arXiv 2025

  27. [27]

    N. J. Cornish and T. B. Littenberg, Class. Quant. Grav. 32, 135012 (2015), 1410.3835

    Pith/arXiv arXiv 2015

  28. [28]

    N. J. Cornish, T. B. Littenberg, B. B´ ecsy, K. Chatziioan- nou, J. A. Clark, S. Ghonge, and M. Millhouse (2020), 2011.09494

    Pith/arXiv arXiv 2020

  29. [29]

    Gupta and N

    T. Gupta and N. J. Cornish, Phys. Rev. D109, 064040 (2024), 2312.11808

    Pith/arXiv arXiv 2024

  30. [30]

    N. J. Cornish, In Preparation (2025)

    2025

  31. [31]

    Lentati, P

    L. Lentati, P. Alexander, M. P. Hobson, S. Taylor, J. Gair, S. T. Balan, and R. van Haasteren, Phys. Rev. D87, 104021 (2013), 1210.3578

    Pith/arXiv arXiv 2013

  32. [32]

    N. Laal, S. R. Taylor, R. van Haasteren, W. G. Lamb, and X. Siemens (2024), 2410.11944

    Pith/arXiv arXiv 2024

  33. [33]

    Gundersen and N

    A. Gundersen and N. J. Cornish, Phys. Rev. D112, 083035 (2025), 2412.13379

    arXiv 2025

  34. [34]

    Bickel and M

    P. Bickel and M. Lindner, Theory of Prob- ability & Its Applications56, 1 (2012), https://doi.org/10.1137/S0040585X97985224, URL https://doi.org/10.1137/S0040585X97985224

    work page doi:10.1137/s0040585x97985224 2012

  35. [35]

    Boito and Y

    P. Boito and Y. Eidelman, Numerical Linear Algebra with Applications32(2025), ISSN 1099-1506, URL http://dx.doi.org/10.1002/nla.70044

    work page doi:10.1002/nla.70044 2025

  36. [36]

    G. A. Gravvanis, Engineering Compu- tations16, 337 (1999), ISSN 0264- 4401, https://www.emerald.com/ec/article- pdf/16/3/337/739754/02644409910266485.pdf, URL https://doi.org/10.1108/02644409910266485

    work page doi:10.1108/02644409910266485 1999

  37. [37]

    Pan and R

    V. Pan and R. Schreiber, SIAM Journal on Sci- entific and Statistical Computing12, 1109 (1991), https://doi.org/10.1137/0912058, URLhttps://doi. org/10.1137/0912058

    work page doi:10.1137/0912058 1991

  38. [38]

    G. Schulz, ZAMM - Journal of Applied Mathe- matics and Mechanics / Zeitschrift f¨ ur Ange- wandte Mathematik und Mechanik13, 57 (1933), https://onlinelibrary.wiley.com/doi/pdf/10.1002/zamm.19330130111, URLhttps://onlinelibrary.wiley.com/doi/abs/10. 1002/zamm.19330130111

    doi:10.1002/zamm.19330130111 1933

  39. [39]

    Toutounian and F

    F. Toutounian and F. Soleymani, Applied Mathemat- ics and Computation224, 671 (2013), ISSN 0096- 3003, URLhttps://www.sciencedirect.com/science/ article/pii/S0096300313009545

    2013

  40. [40]

    Dechant and E

    A. Dechant and E. Lutz, Physical Review Letters115 (2015), ISSN 1079-7114, URLhttp://dx.doi.org/10. 1103/PhysRevLett.115.080603

    2015