Pith. sign in

REVIEW 3 major objections 5 minor 22 references

Optimal design of interpolation methods for time-delay interferometry

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A cosine-sum interpolation kernel with 22 coefficients and a continuous first derivative satisfies LISA's 1 pm requirement while suppressing the time-varying-delay glitch that order-41 Lagrange interpolation cannot.

desk verdict A real, previously missed interpolation defect in TDI plus a clean fix, with the main caveat that the numerical validation ran at half the design delay-rate. read the letter →

arxiv 2412.14884 v1 pith:FUPA3AHL submitted 2024-12-19 astro-ph.IM gr-qc

Martin Staab , Jean-Baptiste Bayle , Olaf Hartwig , Aurélien Hees , Marc Lilley , Graham Woan , Peter Wolf This is my paper · ORCID
classification astro-ph.IMgr-qc
keywords LISAtime-delayinterferometryinterpolationkernelLagrangecosine-sumspectralleakagelaserfrequencynoisegravitational-wavedetection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that Lagrange interpolation, long the standard way to time-shift LISA data in time-delay interferometry (TDI), no longer meets the mission's needs once delays vary in time. Because the Lagrange kernel's first derivative is discontinuous, a time-varying delay that sweeps across an integer sample count produces a glitch-like transient whose spectral leakage breaches LISA's 1 pm noise budget at low frequencies. As a remedy, the authors design a cosine-sum kernel whose smoothness is imposed at design time (continuous first derivative) and whose coefficients are optimized to keep in-band interpolation error below the same budget. With 22 coefficients, the new kernel matches the in-band error performance of a 42-coefficient Lagrange interpolator while suppressing the glitch, halving the computational cost and shrinking edge effects.

What carries the argument

The load-bearing object is the interpolation kernel k(τ) that defines the time-shift operation as a discrete convolution with samples of the kernel. The paper derives the constant-delay interpolation error from the kernel's Fourier transform, and for slowly varying delays it linearizes the delay near an integer-sample crossing and decomposes the output into a constant-delay part, a continuous correction, and a discontinuous part proportional to the sign function. The magnitude of the discontinuous part is set by the size of the jump in the lowest-order discontinuous derivative, transferred through the factor |Δh_q(f)|², and its spectral leakage falls as $f^{{-(q+1)}}$. The cosine-sum kernel, defined as a finite cosine series times a rectangular window, is optimized by a weighted minimax algorithm that minimizes in-band deviation while enforcing vanishing even derivatives at the kernel boundaries; with N=22 and one boundary-smoothness constraint the first derivative is continuous.

What would settle it

Compute the periodogram of the X2 Michelson variable using the cosine-sum kernel in LISA simulations with a round-trip delay rate near 1e-6 s/s (an order of magnitude above the assumed 1e-7) or with a visibly accelerating delay; if excess power appears above the TDI 1 pm reference curve in the 0.1 mHz–1 Hz band, the model's claim that the lowest-order discontinuity dominates would be contradicted.

Watch

Extended reading notes

Core claim

The central claim is that for TDI with time-varying delays, an interpolation kernel's boundary smoothness is as important as its frequency-domain flatness. The paper builds a model of the interpolation glitch: when the delay passes through an integer multiple of the sampling time, the lowest-order discontinuous derivative of the kernel drives spectral leakage that decays only algebraically in frequency. Order-41 Lagrange interpolation, whose kernel has a discontinuous first derivative, passes the constant-delay error test but fails this leakage test; a cosine-sum kernel with N=22 coefficients and a continuous first derivative passes both, keeping the worst-case constant-delay in-band error and the additional leakage below LISA's single-link and TDI 1 pm reference curves. Numerical simulations of the second-generation Michelson combination X2 confirm the glitch for Lagrange and its absence for the cosine-sum kernel.

Load-bearing premise

The glitch model assumes that near an integer-sample crossing the delay evolves linearly in time and that the lowest-order discontinuous derivative of the kernel dominates the spectral leakage, with LISA's round-trip delay rate near 1e-7 s/s and a 10,000 s observing window; if higher-order terms, a faster delay rate, or other orbit geometries produce larger leakage, the 22-coefficient kernel may no longer stay below the 1 pm curve.

Editorial extensions

If this is right

  • TDI processing for LISA can switch from order-41 Lagrange interpolation to the 22-coefficient cosine-sum kernel, roughly halving the floating-point cost of the time-shift operation.
  • The narrower kernel shortens the stretches of invalid samples at the boundaries of the data and around gaps, increasing robustness against data loss.
  • The glitch-induced low-frequency leakage disappears without needing an extra sharp low-pass filter between 1 Hz and 2 Hz to suppress out-of-band laser noise.
  • The same design procedure applies with higher smoothness requirements: requesting continuity of the second derivative raises N or lowers in-band accuracy, and the paper gives the trade-off through the L parameter.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The design principle transfers to other future space-based gravitational-wave observatories (TianQin, DECIGO) whose arm-length flexing forces time-varying delays; the passband, sampling rate, and delay-rate parameters would have to be re-optimized, but the kernel family and leakage model carry over.
  • The leakage model gives a diagnostic fingerprint: a pair of glitches separated by the round-trip time produces nulls at multiples of the inverse separation, so monitoring the periodogram of TDI variables could flag when the assumed delay rate is exceeded.
  • One could test a hybrid kernel that keeps Lagrange's exact zero-crossings at integer samples but tapers the kernel tails smoothly, potentially preserving exact sample-crossing behavior while still achieving a continuous first derivative; the paper does not explore such hybrids.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper develops a framework for designing interpolation kernels for TDI in LISA, where laser frequency noise must be suppressed by applying accurate time-varying delays to discretely sampled measurements. The authors show that Lagrange interpolation, although near-optimal for constant delays, produces glitch-like spectral leakage when a time-varying delay crosses an integer sample shift, because its first derivative is discontinuous. They propose a family of 'cosine-sum' kernels with tunable smoothness and optimize the kernel to satisfy LISA's 1 pm in-band requirement with N=22 coefficients and a continuous first derivative. The central claim is that this kernel suffices both for constant-delay interpolation error and for suppressing transient glitch leakage. The claim is supported by an analytic transfer-function/leakage model and by LISA Instrument/PyTDI simulations at a single orbit epoch.

Significance. If the sufficiency claim holds, the paper offers a practically useful replacement for the standard Lagrange interpolation in TDI pipelines, cutting the number of interpolation coefficients roughly in half while avoiding a newly identified time-domain glitch mechanism. The analytic model for glitch-induced spectral leakage (Eq. 34) is a useful contribution, and the design procedure based on weighted Chebyshev optimization is transparent and reproducible (coefficients are tabulated). The simulations use publicly available LISA simulation tools, which strengthens the work. The main weakness is that the numerical validation of glitch suppression is performed at a delay derivative about a factor of two smaller than the design value, so the central sufficiency claim is not yet fully backed by simulation at the assumed worst case.

major comments (3)
  1. [Sec. 6, Fig. 9] The numerical validation of the cosine-sum kernel's glitch suppression is carried out at a delay derivative of approximately |dot_d| = 5e-8, while the design analysis in Sec. 3 and the sufficiency statement in Sec. 7 assume |dot_d| = 1e-7. Since the leakage model in Eq. (34) scales as dot_d^(2*q_hat) with q_hat=2, the modeled leakage power increases by a factor of 16 (a factor of 4 in ASD) at the design value. The authors should either run the simulation at |dot_d| = 1e-7 (or report the actual |dot_d| from the ESA orbit file in Sec. 6) and show that the periodogram still lies below the 1 pm reference curve, or provide a quantitative bound showing that the margin in Fig. 7 is sufficient to absorb this increase. As it stands, the only direct numerical evidence for the 'adequate suppression' claim is at a less demanding operating point.
  2. [Sec. 3.2, Eq. (34)] The leakage model rests on two unverified assumptions: (i) the delay can be linearized near the integer-sample crossing as in Eq. (23), and (ii) the leading-order discontinuous derivative q_hat dominates, with higher-q terms negligible. The paper does not provide a quantitative check of these approximations against numerical results, aside from a single visual comparison in Fig. 9 for Lagrange interpolation. I ask the authors to validate the model for a range of delay rates and orbit epochs (at least two or three realizations), and to estimate the contribution of the next-order terms (q>q_hat) or of the time variation of dot_d over the 10,000 s window. This is load-bearing because the conclusion that 'a continuous first derivative suffices' is derived directly from this model.
  3. [Sec. 5, Fig. 7] The in-band performance of the cosine-sum kernel is demonstrated using the same weighted error function (A.5)-(A.7) that is optimized in the design. This is reasonable design practice, but it means the statement 'the kernel respects the 1 pm requirement' is partly ensured by construction rather than independently tested. The time-domain simulation in Fig. 9 provides an independent check, but only for one epoch and one realization. I recommend either adding a small scan of orbit epochs in the simulation (e.g., three to five realizations) or stating explicitly that the in-band requirement is guaranteed by the optimization, with the simulation serving only as a cross-check.
minor comments (5)
  1. [Sec. 3.2] The text uses 'Through out' where 'Throughout' is intended; please correct the typo.
  2. [Sec. 5] In the paragraph describing Fig. 6, 'opposed to' should be 'unlike' when contrasting the cosine-sum kernel with the sinc and Lagrange kernels.
  3. [Sec. 6] Please state explicitly the value of |dot_d| used in the simulation, rather than saying 'approximately 5e-8'. This is important because the sufficiency claim depends on the delay rate.
  4. [Fig. 9] The legend entry 'cosine-sum with glitch' is slightly misleading; the time-series does not contain a visible glitch, so consider renaming it, e.g., 'cosine-sum at integer-delay crossing'.
  5. [Sec. 4] The derivation of the Lagrange kernel in Eq. (37) could be clearer: the reader must infer that the indexing m runs only over the half-width. A short sentence adding that the kernel is symmetric for negative offsets would improve readability.

Circularity Check

1 steps flagged · score 4.0 of 10

In-band 'sufficiency' is the minimax design objective restated; glitch leakage has independent simulation support but only at |dot_d|≈5e-8.

  1. fitted input called prediction [Appendix A, Eqs. (A.5)-(A.6), and Sec. 5, Fig. 7; cf. Eq. (22)]
    "The weighting function W (f ) is defined as W(f) = (f+fmin)^{-1} if 0 ≤ f ≤ fpass, 10 · f^{-1}_pass · (f/fstop)^3 else ... Consequently, the maximum error δ/W(f) ... is proportional to the Fourier frequency in band which is desired to respect the LISA requirement curve, which exhibits the same frequency dependency."

    The cosine-sum coefficients {a_n} are obtained by minimizing max |W(ξ)(k~(ξ)-D(ξ))| (Eq. A.6), with W(f) chosen to mirror the 1 pm curve's f-slope. The 'worst-case in-band interpolation error' in Fig. 7 is evaluated from Eq. (22), which is exactly the in-band |k~-1| and out-of-band aliasing terms that the optimizer minimizes. Thus the claimed in-band adequacy of the N=22 kernel is a restatement of the design constraint, not an independent prediction. The independent content is the time-domain TDI glitch simulation (Sec. 6), which was not part of the optimization.

full rationale

The paper's central glitch-suppression claim is supported by an analytic model (Eq. 34) and by a numerical TDI simulation with realistic ESA orbit files in which the optimized cosine-sum kernel is actually applied. That simulation is not fitted to the model: the dot_d value used for the Lagrange leakage comparison is taken from the simulated orbit (≈5e-8), and the cosine-sum result is shown to be free of excess power. The 1 pm reference is cited from the authors' prior paper [9], but the underlying requirement is external to this paper and the TDI propagation is a standard calculation, so I do not count that self-citation as load-bearing circularity. The main circular element is the in-band part of the sufficiency claim: the kernel is designed by minimizing the same weighted error metric that is later used to demonstrate compliance with the 1 pm curve (Appendix A vs. Sec. 5/Fig. 7). This is a constructive design result rather than a falsifiable prediction, but it is partial circularity under the fitted-input-called-prediction pattern. A further non-circular weakness is that the numerical validation uses |dot_d|≈5e-8 while the design analysis assumes 1e-7; because the q=2 leakage scales as dot_d^4, a factor two in dot_d changes the modeled leakage by 16, so the simulation does not fully close the gap at the assumed worst case. That is a validation/risk issue, not a circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central design rests on three types of unpaid inputs: the 22 optimized kernel coefficients and the hand-chosen weighting parameters (free parameters), the signal-processing and LISA modeling assumptions (axioms), and no invented physical entities. The optimization target itself is the LISA 1 pm requirement from the authors' earlier paper, so the 'sufficiency' margin is in part a design outcome rather than an independent measurement.

free parameters (3)
  • cosine-sum kernel coefficients a_0 ... a_21 = Listed in Table A1, e.g., a0=4.545e-2, a1=9.091e-2, ..., a21=5.280e-11
    These 22 coefficients are the output of the weighted Chebyshev optimization (Eqs. A.5-A.7) and define the central artifact; they are chosen to meet the design constraints, not derived from first principles.
  • design band edges and weights = fpass/fs=1/4, fstop/fs=3/4, fmin=0.25e-4, stop-band weight proportional to 10*(f/fstop)^3
    These choices in Appendix A shape the minimax error to follow the 1 pm requirement curve; they are hand-picked and influence whether N=22 satisfies the requirement.
  • kernel width N and smoothness order L = N=22, L=2
    N is the stated minimum width found to meet the requirements and L=2 imposes continuity of the first derivative; both are design choices tested for one LISA-operations scenario.
assumptions (5)
  • standard math Whittaker-Shannon sampling theorem: band-limited continuous signals are exactly recoverable from samples
    Used in Eq. (14) to derive the continuous-time representation of the delayed signal and the transfer function (Eqs. 15-19).
  • domain assumption LISA laser frequency noise is band-limited, stationary white noise with ASD 30 Hz/sqrt(Hz) up to Nyquist
    Used for the interpolation-error PSD calculations and in numerical simulations (Secs. 3.1, 6).
  • domain assumption Time-varying delay can be linearized as d(t)=d0+dot_d*(t-t0) near an integer-sample crossing
    Equation (23), the basis of the glitch spectral-leakage model Eq. (34); if the delay curvature is not negligible the leakage estimate changes.
  • domain assumption Second-order couplings of interpolation errors to interpolation residuals are negligible
    Stated in Sec. 2 after Eq. (9) when defining delta X2^D.
  • domain assumption The 1 pm requirement curve from reference [9] and parameters T=10,000 s, dot_d=1e-7, fs=4 Hz, Kaiser beta=30 represent LISA conditions
    Determines whether the designed kernel is 'sufficient'; alternative requirements would change N and L.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimal design of interpolation methods for time-delay interferometry." pith.science (2026). https://pith.science/paper/FUPA3AHL

@misc{pith2026241214884,
  author       = {Pith},
  title        = {Pith review of: Optimal design of interpolation methods for time-delay interferometry},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FUPA3AHL}},
  note         = {Machine review of arXiv:2412.14884}
}
read the original abstract

Time-delay interferometry (TDI) suppresses laser frequency noise by forming linear combinations of time-shifted interferometric measurements. The time-shift operation is implemented by interpolating discretely sampled data. To enable in-band laser noise reduction by eight to nine orders of magnitude, interpolation has to be performed with high accuracy. Optimizing the design of those interpolation methods is the focus of this work. Previous research that studied constant time-shifts suggested Lagrange interpolation as the interpolation method for TDI. Its transfer function performs well at low frequency but requires a high number of coefficients. Furthermore, when applied in TDI we observed prominent time-domain features when a time-varying shift scanned over a pure integer sample shift. To limit this effect we identify an additional requirement for the interpolation kernel: when considering time-varying shifts the interpolation kernel must be sufficiently smooth to avoid unwanted time-domain transitions that produce glitch-like features in power spectral density estimates. The Lagrange interpolation kernel exhibits a discontinuous first derivative by construction, which is insufficient for the application to LISA or other space-based GW observatories. As a solution we propose a novel design method for interpolation kernels that respect a predefined requirement on in-band interpolation residuals and that possess continuous derivatives up to a prescribed order. Using this method we show that an interpolation kernel with 22 coefficients is sufficient to respect LISA's picometre-requirement and to allow for a continuous first derivative which suppresses the magnitude of the time-domain transition adequately. The reduction from 42 (Lagrange interpolation) to 22 coefficients enables us to save computational cost and increases robustness against artefacts in the data.

Figures

Figures reproduced from arXiv: 2412.14884 by the authors.

Figure 1
Figure 1. Illustration of the working principle of TDI. Increasingly longer virtual photon paths are constructed by hierarchically building up the intermediary variables π, ρ and σ as defined in equations 4, 7 and 8. In the final step the two round-trip variables σ12 and σ13 are subtracted to form the second-generation Michelson variable X2. The two synthesized counter-propagating photon path have almost identical length, hen… view at source ↗
Figure 2
Figure 2. Illustration of the post-processing delay operation. In the upper panel, the solid blue line depicts a simple linear interpolation between discrete samples (blue dots) of the continuous band-limited signal x(t) (dashed blue line). The black line shows the corresponding triangular interpolation kernel. To interpolate at a given time (black cross) the neighbouring samples are weighted by the corresponding readings of … view at source ↗
Figure 3
Figure 3. One-sided representation of the modified spectral window magnitude |v˜q(f)| assuming a Kaiser window with β of 30 and an observation time T of 10 000 s. Results for q = 0 (blue), q = 1 (red) and q = 2 (yellow) are shown. The dashed lines show the limit f ≫ 1 T . In (34) we limit ourselves to only the leading order contribution in ˙d, i.e., the degree q = ˆq of the lowest-order derivative that is discontinuous. For q… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Lagrange kernel functions (upper panel) and their corresponding Fourier transform (lower panel) for orders 1 (blue) and 5 (red). For comparison, we also plot the result for the sinc-kernel which is the limit for Lagrange kernels of increasing order. In the lower panel …
Figure 5
Figure 5. Figure 5: Amplitude spectral densities (ASDs) of interpolation errors for Lagrange interpolation order 41 operating on laser noise (in yellow) with an ASD of 30 Hz/ √ Hz. The light red line shows the worst case interpolation error (see (22)) which stays below the single link (SL…
Figure 6
Figure 6. Figure 6: Comparison of the ideal sinc-kernel in yellow, the order 41 Lagrange kernel (N = 42) in blue and the cosine-sum kernel (N = 22) in red. The upper panel presents the time-domain kernel functions and their different behaviours at integer times (highlighted in the zoomed-…
Figure 7
Figure 7. Figure 7: ASDs of interpolation errors for the cosine-sum kernel with otherwise identical parameters as figure 5. For this choice of kernel both the worst-case in￾band interpolation error (red) and the ASD estimate of the leading order discontinuity (blue), i.e. in the second de…
Figure 8
Figure 8. Figure 8: Second-generation Michelson combination X2 computed using Lagrange interpolation (order 41). In dark red the low-pass filtered time series is shown exhibiting a prominent glitch at the centre. The dashed lines represent the round-trip delays d12131 (light blue) and d13…
Figure 9
Figure 9. Figure 9: ASD of second-generation Michelson combination X2 for different interpolation schemes. Periodograms of 10 000 s-datasets including the critical time (c.f. figure 8) are shown for numerical TDI calculations using the cosine-sum kernel (dark blue) and Lagrange interpolat…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 10 canonical work pages

  1. [1]

    2024 LISA Definition Study Report (Preprint 2402.07571)

    Colpi M et al. 2024 LISA Definition Study Report (Preprint 2402.07571)

    arXiv 2024

  2. [2]

    (LISA) 2017 Laser Interferometer Space Antenna (Preprint 1702.00786)

    Amaro-Seoane P et al. (LISA) 2017 Laser Interferometer Space Antenna (Preprint 1702.00786)

    arXiv 2017

  3. [3]

    (LISA) 2023 Living Rev

    Seoane P Aet al. (LISA) 2023 Living Rev. Rel. 26 2 (Preprint 2203.06016)

    arXiv 2023

  4. [4]

    Astronaut

    Martens W and Joffre E 2021J. Astronaut. Sci. 68 402–443

    work page

  5. [5]

    Vallisneri M 2005 Phys. Rev. D 72 042003 [Erratum: Phys.Rev.D 76, 109903 (2007)] (Preprint gr-qc/0504145)

    arXiv 2007

  6. [6]

    Tinto M and Dhurandhar S V 2021Living Rev. Rel. 24 1

    work page

  7. [7]

    Shaddock D A, Tinto M, Estabrook F B and Armstrong J W 2003 Phys. Rev. D 68 061303 (Preprint gr-qc/0307080)

    arXiv 2003

  8. [8]

    Cornish N J and Hellings R W 2003Class. Quant. Grav. 20 4851–4860 (Preprint gr-qc/0306096)

    arXiv

Show all 22 references
  1. [9]

    StaabM,Lilley M,BayleJBandHartwigO2024 Phys. Rev. D 109043040( Preprint 2306.11774)

    arXiv

  2. [10]

    BayleJB,LilleyM,PetiteauAandHalloinH2019 Phys. Rev. D 99084023( Preprint 1811.01575)

    arXiv

  3. [11]

    thesis Leibniz U., Hannover

    Staab M 2023Time-delay interferometric ranging for LISA Ph.D. thesis Leibniz U., Hannover

  4. [12]

    Reinhardt J N, Staab M, Yamamoto K, Bayle J B, Hees A, Hartwig O, Wiesner K, Shah S and Heinzel G 2024Phys. Rev. D 109 022004 (Preprint 2307.05204)

    arXiv

  5. [13]

    Shaddock D A, Ware B, Spero R E and Vallisneri M 2004Phys. Rev. D 70 081101 (Preprint gr-qc/0406106) Optimal design of interpolation methods for TDI 27

    arXiv

  6. [14]

    Halloin H 2018 Note on decimation and interpolation filters for LISA simulation Tech. Rep. LISA- APC-INST-TN-003 APC

    2018

  7. [15]

    Jenkins G M and Watts D G 1968Spectral analysis and its applications Holden-Day series in time series analysis (San Francisco: Holden-Day) ISBN 978-0-8162-4464-5

  8. [16]

    Tinto M, Estabrook F B and Armstrong J W 2004Phys. Rev. D 69 082001 (Preprint gr-qc/ 0310017)

  9. [17]

    Nuttall A 1981IEEE Transactions on Acoustics, Speech, and Signal Processing 29 84–91

  10. [18]

    Parks T and McClellan J 1972IEEE Transactions on Circuit Theory 19 189–194

  11. [19]

    Bayle J B, Hartwig O and Staab M 2024 Lisa instrument URL https://doi.org/10.5281/ zenodo.13809621

    2024

  12. [20]

    Bayle J B, Hartwig O and Staab M 2021Phys. Rev. D 104 023006 (Preprint 2103.06976)

    arXiv

  13. [21]

    Staab M, Bayle J B and Hartwig O 2023 Pytdi URLhttps://doi.org/10.5281/zenodo.8429119

    2023 doi

  14. [22]

    Albrecht H H 2010 The Open Signal Processing Journal 3 20–29 ISSN 18768253 URL http: //benthamopen.com/ABSTRACT/TOSIGPJ-3-20

    2010

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.