Pith. sign in

REVIEW 2 major objections 4 minor 30 references

Non-stationary noise can be modeled exactly by a positive dynamic spectrum S(f,t)=s(f,t)^2, whose square root defines a Gramian noise covariance in both Fourier and wavelet bases.

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

2026-08-02 06:01 UTC pith:RSMPFNBD

load-bearing objection Solid, carefully derived framework for non-stationary noise; the math is sound, but the applicability claims outrun the evidence on one point—the truncation-correlation bound. the 2 major comments →

arxiv 2607.13168 v1 pith:RSMPFNBD submitted 2026-07-14 gr-qc

Modeling non-stationary noise: applications in gravitational wave astronomy

Neil J. Cornish This is my paper
classification gr-qc
keywords dynamic spectrumnon-stationary noisenoise covarianceGramian matrixgravitational wave data analysisWilson-Daubechies waveletWhittle likelihoodLISA
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 claims that non-stationary detector noise is fully described by a positive time-frequency function S(f,t)=s(f,t)^2, the dynamic spectrum, without needing negative-valued spectra or ad hoc smoothing. From this function the full noise covariance matrix follows in closed form in both the Fourier domain and the Wilson-Daubechies wavelet domain, and it is positive semidefinite by construction. Because the covariance is a Gram matrix, non-stationary noise can be simulated directly and likelihoods can be computed with the exact inverse rather than a diagonal approximation. The framework is aimed at gravitational wave analyses, where noise from windows, orbitally modulated confusion foregrounds, and general time-varying spectra otherwise biases parameter inference. A sympathetic reader should care because it replaces the stationarity assumption with a tractable, interpretable model that preserves the 'when' of noise variations.

Core claim

The central claim is Eq. (3): for discretely sampled finite data whose dynamic spectrum is S(f_k,t_n)=s_k[n]^2, the covariance between Fourier coefficients is C_ml = (1/N) sum_{k=0}^{N-1} \tilde{s}_k[{m-k}_N] \tilde{s}_k^*[{l-k}_N]. Each basis frequency k contributes an independent stationary unit-variance noise component modulated by the time-domain amplitude s_k[n]; the modulation spreads power across Fourier bins and the off-diagonal entries carry the time information. The same Gram construction yields an exact closed-form Wilson-Daubechies wavelet covariance, Eq. (37), built from filtered versions of the same amplitude transforms. The dynamic spectrum thus generalizes the stationary powe

What carries the argument

The Gram-factor model: set S(f,t)=s(f,t)^2 and treat s(f_k,t) as the time-domain envelope multiplying an independent stationary Gaussian noise process for each frequency bin. The covariance between discrete Fourier coefficients is then the Gramian of the matrix A_{ak} = N^{-1/2} \tilde{s}_k[{a-k}_N], making the covariance positive semidefinite by construction and giving a ready square-root factor for simulation. The same factor, after projection through the Wilson-Daubechies window functions, produces the wavelet-domain covariance. This Gram structure turns an O(N^3) covariance computation into structured O(N^2 log N) operations.

Load-bearing premise

The model assumes the measured noise can be split into independent unit-variance stationary noise components per frequency bin, each multiplied by an amplitude s(f_k,t), and that correlations induced by cutting a longer continuous noise stream are negligible for the detector band of interest; if real noise violates this decomposition, Eq. (3) will not be the true covariance.

What would settle it

Take a stretch of real or simulated detector noise with a known time-varying gain, estimate the dynamic spectrum s(f_k,t), build C from Eq. (3), and compare the off-diagonal entries of the sample covariance (or the full inverse-covariance likelihood of a known signal) with that predicted C; a systematic mismatch—especially from correlated broadband transients or steep low-frequency spectral slopes—would falsify the model.

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

If this is right

  • Standard diagonal-approximation likelihoods miscompute SNRs when noise is non-stationary; with windowed data, the full covariance keeps SNR flat while diagonal approximations show losses and overshoots in the roll-off region.
  • For LISA galactic confusion noise modulated at 1/yr harmonics, SNR of a 1-day signal varies by a factor of two over a year, an effect the diagonal approximation misses entirely.
  • Even when off-diagonal Fourier covariance entries look small, dropping them removes all knowledge of non-stationarity; pre-whitening by the time-averaged spectrum plus a diagonal wavelet approximation recovers SNR within about 0.002%.
  • If the model is right, any positive dynamic spectrum defines a valid, simulatable noise covariance, so parameter estimation can be done with the exact likelihood in Fourier or wavelet domain.
  • The same covariance-square-root factor that computes likelihoods also generates noise realizations without Cholesky decomposition.

Where Pith is reading between the lines

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

  • This construction may translate to other time-series fields where time-varying noise power matters, such as radio astronomy or environmental monitoring, wherever one can estimate a time-frequency spectrum and needs a positive-definite covariance rather than a diagonal approximation.
  • Because the covariance is Gramian and built from s(f,t), one could imagine a nonparametric estimator that fits s(f,t) to data via likelihood maximization with a smoothness prior, directly connecting the model to existing spectrogram or modulation estimation methods.
  • The paper's claim that inter-frequency correlations from truncation are negligible rests on smooth, not too steep spectra; for low-frequency steep noise (e.g., pulsar timing arrays) the same Gram form would need a correction, and extending Eq. (3) to include continuum spectral windows is a natural next step.
  • A testable prediction: for real detector noise with an injected time-varying gain, the sample covariance of Fourier coefficients should match Eq. (3) with the estimated dynamic spectrum, rather than a diagonal or Wigner-Ville covariance.

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 / 4 minor

Summary. The paper introduces a Gram-factor model for non-stationary Gaussian noise in gravitational wave data. Given a positive dynamic spectrum S(f,t)=s^2(f,t), the noise is defined as a sum of independent stationary components, each modulated by a time-dependent amplitude. For a finite discrete Fourier representation, the noise covariance is shown to be Eq. (3), a Gramian sum over frequency bins. A closely analogous closed-form expression is derived for Wilson-Daubechies wavelet coefficients, Eq. (37). The paper argues that this construction generalizes the stationary power spectrum, enables efficient simulation via a square-root factor, and provides computationally focused covariance evaluations. Applications include window-function effects, modulated LISA galactic confusion noise, and a non-separable time-frequency dynamic spectrum. The mathematical derivation of the model covariance is self-consistent; the main caveat is the asserted but unquantified neglect of finite-window inter-frequency correlations in Sec. III.B.

Significance. If the applicability claims are supported, this would be a useful, interpretable framework for non-stationary noise in LISA/LIGO analyses. The construction has clear strengths: the Gramian form guarantees positive semidefiniteness by construction; the Fourier and wavelet covariance expressions are explicit; the simulation algorithm directly uses the same square-root factor; and the LISA confusion-noise example shows a large SNR variation that is missed by the standard diagonal approximation. The derivation is parameter-free, with no fitted constants. The main weakness is that the model's fidelity to real detector noise rests on an unquantified approximation in Sec. III.B; the paper's own examples use the model covariance rather than the true continuum covariance, so they do not yet validate that approximation.

major comments (2)
  1. [Sec. III.B, Eq. (19)] The paper asserts that finite-window inter-frequency correlations can be ignored for ground- and space-based detectors, but no quantitative bound is provided. For a stationary continuum process, the true covariance is Eq. (19); the model covariance in Eq. (3) with s_k = sqrt(S(f_k)) is exactly diagonal, so the discrepancy is entirely the truncation correlation. This error scales with the fractional change of S(ν) over the window width 1/T inside the sensitive band. The examples in Sec. V compute the model covariance, not the true continuum covariance, so they do not validate the assertion. Please provide an explicit numerical estimate or bound using representative LIGO and LISA sensitivity curves, including the lowest in-band frequency bins.
  2. [Sec. III.C, Eqs. (25)-(27)] The Schmidt-expansion argument is invoked to claim that the construction is 'quite general,' but the model in Eq. (24) is a diagonal-in-frequency sum of modulated stationary components, not a separable expansion of s(f,t). The connection between the Schmidt decomposition and the representable class of covariance matrices is not established. Please clarify the exact class of non-stationary processes covered by the model and whether any representational limitations are intrinsic to the model or only to the finite-window truncation approximation.
minor comments (4)
  1. [Sec. II, Eq. (4)] The notation F^{-1} = F^†/N is correct for the unnormalized DFT, but the subsequent complexity statement 'reduced from N^3 to N^2 log N' could be clarified by distinguishing the cost of forming C versus applying C to a vector.
  2. [Sec. III.C, Eq. (24)] The symbol k is used both as the frequency-bin index and as the summation index in \tilde{s}_k. Consider using a different letter for the summation index to avoid confusion.
  3. [Sec. V, Figs. 4 and 7] The SNR curves for the full covariance and the diagonal approximation would benefit from different line styles or markers; the captions do not indicate which curve is which beyond the text.
  4. [References] Refs. [13], [22], and [23] are 'In Preparation' or current-year preprints; if any of the main results depend on them, please state that explicitly.

Circularity Check

0 steps flagged

No significant circularity: the covariance formulas are derived exactly from an explicit construction, with no fitted parameters and no load-bearing self-citations.

full rationale

The central result Eq. (3) is not a fitted prediction but an algebraic consequence of the explicit generative model x[n] = (1/√N) Σ_k s_k[n] η_k e^{2πikn/N} (Eq. 46), and the paper says so: 'we are ready to define a model of non-stationary noise that proceeds by construction' (Sec. III.C). Eq. (24) follows by taking E[η_k η_l^*]=δ_kl, and the wavelet covariance Eq. (37) is obtained by the same Gram construction under a linear WD transform; no constant is fitted to data and no target result is assumed. The agreement with Wigner-Ville in the slowly varying limit is explicitly presented as an approximation (Eqs. 44-45), not as a derivation of the model. The checks against the author's prior special cases [27] are consistency checks on a limiting case, not load-bearing; the same is true of the WD mapping from Ref. [21]. The admitted limitation that finite-window inter-frequency correlations are neglected (Sec. III.B) is an assumption about applicability, not a circular step: Eq. (19) is the exact continuum-truncation covariance and Eq. (3) is a different, approximate model. The absence of a quantitative bound for LISA/LIGO noise curves is a completeness/correctness concern, not evidence that the derivation reduces to its inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The framework's central axioms are the decomposability of noise into independent frequency-bin modulations and the negligibility of finite-duration truncation correlations for ground/space-based detectors. Both are explicitly stated and scoped. No fitted parameters enter the central construction; the example parameters are illustrative only.

axioms (5)
  • domain assumption Noise can be decomposed as x(t)=sum_a n_a(t) g_a(t) with stationary, mutually uncorrelated processes n_a (Eqs. 20-23).
    This is the core modeling assumption: it is what makes Eq. (24) follow. Real detector noise may include correlated broad-band transients that do not fit this form; the paper acknowledges the model is approximate.
  • domain assumption Inter-frequency correlations from truncating a continuous stationary process are negligible for ground- and space-based detectors (Sec. III.B).
    Eq. (19) shows these correlations exist and matter for steep spectra (PTA); the framework ignores them for the target detectors, with a qualitative argument rather than a quantitative bound.
  • domain assumption The dynamic spectrum S(f,t) is non-negative and square-integrable; s=sqrt(S) is well-defined (Sec. II).
    Positivity is required for the Gram-factor construction; any physical power spectrum should be non-negative, and the paper enforces this by construction.
  • standard math The Wilson-Daubechies wavelet transform uses window functions obeying the partition-of-unity and orthogonality relations (Sec. III.D).
    The WD covariance formulas rely on the standard properties of the WD/Meyer window construction; these are established in Refs. [19,20,22,23].
  • standard math Schmidt expansion: any square-integrable g(f,t) can be written as a (truncated) sum of separable functions (Sec. III.C).
    Used to argue the model is general; this is the continuous SVD theorem, but its use here is heuristic since the covariance model is not shown to be universal.

pith-pipeline@v1.3.0-alltime-deepseek · 12147 in / 20754 out tokens · 193526 ms · 2026-08-02T06:01:18.290779+00:00 · methodology

0 comments
read the original abstract

In an ideal world, the measurement noise in gravitational wave data would be stationary and Gaussian. In reality, neither of these conditions holds. Here a general framework is introduced that can be used to model non-stationary noise in an easily interpretable way, using a dynamic power spectrum $S(f,t)$. The construction is a Gram-factor model for the noise covariance matrix that is positive semi-definite by construction. This construction generalizes the familiar stationary power spectrum $S(f)$. The dynamic spectrum encodes the properties of the noise covariance matrix in any basis, including the frequency domain, time domain, and time-frequency wavelet domain. Closed form expressions are given for discrete Fourier representations of the data, and for discrete Wilson-Daubechies wavelet representations of the data. Both take the form of Gramian matrices. Examples are provided, including the non-stationarity caused by window functions, the modulated response to galactic binary signals for space-based detectors, and other, more general types of non-stationarity.

Figures

Figures reproduced from arXiv: 2607.13168 by Neil J. Cornish.

Figure 1
Figure 1. Figure 1: FIG. 1. The noise correlation matrix and its inverse in a [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. The SNR as a function of central time for a well [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. The SNR as a function of central time for a well [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. The galactic confusion noise correlation matrix and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. The diagonal entries of the WDM noise correlation [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. The SNR as a function of central time for a Gaus [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. The inverse of the noise correlation matrix for a dy [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

30 extracted references · 19 linked inside Pith

  1. [1]

    [27] the WDM noise covariance matrix was cal- culated analytically for the special casesS(f, t) =S f (f) andS(f, t) =S t(t)

    T aylor expansion of the WDM noise covariance matrix In Ref. [27] the WDM noise covariance matrix was cal- culated analytically for the special casesS(f, t) =S f (f) andS(f, t) =S t(t). Now with the full expression for the WD noise covariance in hand we are in a position to com- plete the calculation to include cross terms in the Taylor expansion ofS(f, t...

  2. [2]

    Generating noise realizations The same factorization used to compute the WD noise correlation matrix efficiently gives a direct way to sim- ulate frequency-domain data with dynamic spectrum S(f, t). The usual way to generate Gaussian data with a specified covariance matrix is to find a matrix square root, for example by a Cholesky decomposition or by di- ...

  3. [3]

    L. S. Finn, Phys. Rev. D46, 5236 (1992), gr-qc/9209010

    Pith/arXiv arXiv 1992

  4. [4]

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

    Pith/arXiv arXiv 2020

  5. [5]

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

    Pith/arXiv arXiv 2015

  6. [6]

    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

  7. [7]

    Chatziioannou, N

    K. Chatziioannou, N. Cornish, M. Wijngaarden, and T. B. Littenberg, Phys. Rev. D103, 044013 (2021), 2101.01200

    Pith/arXiv arXiv 2021

  8. [8]

    Hourihane, K

    S. Hourihane, K. Chatziioannou, M. Wijngaarden, D. Davis, T. Littenberg, and N. Cornish, Phys. Rev. D 106, 042006 (2022), 2205.13580

    Pith/arXiv arXiv 2022

  9. [9]

    Zackay, T

    B. Zackay, T. Venumadhav, J. Roulet, L. Dai, and M. Zaldarriaga (2019), 1908.05644

    Pith/arXiv arXiv 2019

  10. [10]

    Mozzon, L

    S. Mozzon, L. K. Nuttall, A. Lundgren, T. Dent, S. Ku- mar, and A. H. Nitz, Classical and Quantum Gravity 37, 215014 (2020), ISSN 1361-6382, URLhttp://dx. doi.org/10.1088/1361-6382/abac6c

    doi:10.1088/1361-6382/abac6c 2020

  11. [11]

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

    Pith/arXiv arXiv 2022

  12. [12]

    Mozzon, G

    S. Mozzon, G. Ashton, L. K. Nuttall, and A. R. Williamson, Phys. Rev. D106, 043504 (2022), 2110.11731

    Pith/arXiv arXiv 2022

  13. [13]

    Allen and J

    B. Allen and J. D. Romano, Phys. Rev. Lett.134, 031401 (2025), 2407.10968

    Pith/arXiv arXiv 2025

  14. [14]

    Crisostomi, R

    M. Crisostomi, R. van Haasteren, P. M. Meyers, and M. Vallisneri (2025), 2506.13866

    Pith/arXiv arXiv 2025

  15. [15]

    Ghosh and N

    S. Ghosh and N. J. Cornish, In Preparation (2026). 11

    2026

  16. [16]

    E. P. Wigner, Physical Review40, 749 (1932)

    1932

  17. [17]

    Ville, Cˆ ables et Transmission2A, 61 (1948)

    J. Ville, Cˆ ables et Transmission2A, 61 (1948)

    1948

  18. [18]

    Martin and P

    W. Martin and P. Flandrin, IEEE Transactions on Acoustics, Speech, and Signal Processing33, 1461 (1985)

    1985

  19. [19]

    Lu and N

    W. Lu and N. Vaswani,The wiener-khinchin theorem for non-wide sense stationary random processes(2009), 0904.0602, URLhttps://arxiv.org/abs/0904.0602

    Pith/arXiv arXiv 2009

  20. [20]

    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

  21. [21]

    Daubechies, S

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

    1991

  22. [22]

    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

  23. [23]

    N. J. Cornish (2020), 2009.00043

    Pith/arXiv arXiv 2020

  24. [24]

    Johnson, K

    A. Johnson, K. Chatziioannou, and J. Summers (2026), 2606.21473

    Pith/arXiv arXiv 2026

  25. [25]

    Vajpeyi, G

    A. Vajpeyi, G. Mentasti, Q. Baghi, O. Burke, and L. Speri (2026), 2606.20269

    Pith/arXiv arXiv 2026

  26. [26]

    Talbot, E

    C. Talbot, E. Thrane, S. Biscoveanu, and R. Smith, Phys. Rev. Res.3, 043049 (2021), 2106.13785

    Pith/arXiv arXiv 2021

  27. [27]

    Talbot et al., Class

    C. Talbot et al., Class. Quant. Grav.42, 235023 (2025), 2508.11091

    Pith/arXiv arXiv 2025

  28. [28]

    Schmidt, Mathematische Annalen63, 433 (1907), URLhttp://eudml.org/doc/158296

    E. Schmidt, Mathematische Annalen63, 433 (1907), URLhttp://eudml.org/doc/158296

    1907

  29. [29]

    N. J. Cornish (2025), 2511.10632

    Pith/arXiv arXiv 2025

  30. [30]

    Pearson and N

    N. Pearson and N. J. Cornish, Phys. Rev. D113, 064033 (2026), 2509.05479

    Pith/arXiv arXiv 2026