REVIEW 4 major objections 6 minor 5 references
PAD\'E FILTERING, Principles and Use: an Introductory Report
T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash
Pith's one-line read This report argues that the reliable way to separate a signal from noise in a data stream is to study the complex poles and zeros of the data's z-transform via Padé approximants, rather than the real peaks of its Fourier energy spectrum.
desk verdict Honest, useful introduction to Padé Filtering with solid toy-model mathematics and a candid admission of the open Froissart-separation problem, but the Virgo 'proof of concept' outruns what the data demonstration actually supports. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the Padé Approximant $[L/M](z)=P_L(z)/Q_M(z)$ of the z-transform $T_N(\mathcal{D})(z)$, whose numerator zeros and denominator poles are tracked in the complex plane. Two probabilistic phenomena give the method its power: the Kac phenomenon (random polynomials have a positive density of real roots, so the real axis attracts PA poles in the noisy case) and the Froissart phenomenon (noise in the Taylor coefficients creates short zero-pole doublets, the Froissart doublets, located near the unit circle). The practical step that carries all the applications is separating these doublets from faithful poles and erasing them, which yields the compressed rational transfer functions used on Virgo data.
What would settle it
To test the central claim, synthesize a damped cosine with known resonance $z_0=e^{-\lambda\tau}e^{i\omega\tau}$, add white Gaussian noise, and run the paper's [10/10] Padé filtering with $N=20$ and many realizations: if the pole-density maximum does not converge to $z_0$ and remain separated from the Froissart doublets, the method fails. A second decisive check is the Virgo transfer property: the pole-histogram maxima of the seismic channel must reappear in the F0 suspension channel on independent non-overlapping data chunks; if they do not, the [5/5] skeleton is an artifact of the chosen chunk.
Extended reading notes
Core claim
The central claim is that the z-transform $T_N(\mathcal{D})(z)=\sum_{n=0}^{N-1} d_n z^n$ encodes the physics of a data set in its complex analytic structure, and that (possibly noisy) Padé Approximants $[L/M](z)=P_L(z)/Q_M(z)$ reveal that structure even for finite $N$. For the paper's model signals the faithful poles are explicit: a constant signal has a pole at $z=1$; a damped cosine has the complex-conjugate pair $1/a$ and $1/\bar a$; AR(1) pink noise has a pole at $1/a$ outside the unit circle; AR(2) red noise has two poles at $r e^{\pm i\varphi}$. White Gaussian noise, by contrast, makes the unit circle a natural boundary of the infinite-$N$ z-transform, and the PA's respond by placing a dense set of poles there, broken near $\pm 1$, together with Froissart doublets—zero-pole pairs with $O(\varepsilon)$ separation that are the signature of randomness. The paper's motto is 'Better study the complex poles of the z-Transform of a process rather than the real peaks of its energy spectrum,' and it supports this by showing that Virgo pole-density maxima of one channel reappear in another and that erasing doublets leaves a [5/5] rational skeleton for the seismic-to-suspension transfer function.
Load-bearing premise
The whole method rests on being able to tell the faithful poles from the spurious Froissart doublets in real noisy data; the paper explicitly notes that this separation is still under discussion and gives no precise rule for the threshold used when erasing doublets.
Editorial extensions
If this is right
- Gravitational-wave searches could stop hunting for peaks in the energy spectrum and instead ask whether a particular pole of the z-transform appears consistently across realizations of the noisy data.
- Noise types become identifiable by their complex-plane signatures: white noise fills the unit circle with poles except near $\pm 1$, AR(1) noise concentrates poles near $1/a$, AR(2) noise creates a pair at $r e^{\pm i\varphi}$, and these signatures persist when a signal is added.
- Erasing Froissart doublets gives a data-compression tool: the Virgo seismic-to-suspension coupling is represented as a [5/5] rational transfer function rather than 432,000-sample channels, and the same idea generalizes to other linear apparatus.
- For modeled gravitational-wave signals the PA reproduces the expected analytic structures—the chirp's accumulating singularity, the ringdown's [1/2] fraction with one zero and complex-conjugate poles, and the pulsar's slowly drifting frequency as secondary peaks near the unit circle.
Reading between the lines
- If the Froissart doublet separation can be made quantitative, the density and angular spread of doublets could serve as a noise-level estimator, giving the method an error bar on the cleaned signal.
- Comparing complex singularity patterns rather than time series might provide a new coincidence test between separate detectors (LIGO, Virgo, KAGRA), especially in the low signal-to-noise regime where spectral peaks are ambiguous.
- The contrast between the real-axis pole attraction and the unit-circle condensation suggests a cheap diagnostic for non-Gaussian noise: real-axis PA pole density should be sensitive to non-Gaussian contaminants, a prediction that can be checked with simulated non-Gaussian noise.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This manuscript is an introductory report on 'Padé Filtering' (PF), the practice of using Padé Approximants of the z-Transform of a data set to locate complex poles and zeros and thereby separate signal from noise. Part A motivates the approach: a two-Lorentzian example (Section 3) and AR(1)/AR(2) computations (Section 8) show that real spectral peaks are ambiguous while complex-pole structures are stable, and the Kac and Froissart phenomena (Section 4) are presented as the probabilistic backdrop. Part B develops toy-model mathematics: the random z-Transform of white Gaussian data (Section 5), random polynomials (Section 7), and explicit noisy rational-interpolation and Padé models in which a 'Froissart polynomial' controls the statistics of pole-zero doublets (Sections 10-11). Part C applies PF to prototypical signals and noises (Section 14), to gravitational-wave templates (chirp, ringdown, pulsar; Section 15), and to Virgo data, where a [5/5] transfer-function 'skeleton' between a seismicity channel and a suspension channel is claimed (Sections 13 and 16), and the open state of the 'Froissart debate' is acknowledged (Section 17). The paper is explicitly a mixture of review and research report.
Significance. If it can be made operational, PF would give gravitational-wave analysts a genuinely different diagnostic: a data-compression and signal/noise-separation tool whose resolution is governed by complex-pole statistics rather than by spectral-peak separation. The paper's strengths are its explicit, checkable toy-model calculations; I verified the Lorentz two-line threshold S >= 2/sqrt(3) (Section 3), the AR(1)/AR(2) energy spectra and the AR(2) peak-transition curve r_2(phi) (Section 8), and the mean energy E[E_N] = 1 for white noise (Section 5). The Section 10 derivation of an explicit Cauchy doublet density with condensation probability 2/3 inside the interpolation interval is a concrete and useful result, and Section 17's candid statement of the open Froissart question is a model of scientific honesty. However, the two applied demonstrations -- the channel-to-channel pole reappearance (Section 13) and the Virgo [5/5] skeleton (Section 16) -- rest on exactly the faithful-versus-Froissart separation that Section 17 concedes is not settled, and neither demonstration is currently quantitative or reproducible.
major comments (4)
- [Section 17] Section 17, 'The Froissart debate': the manuscript itself states that the clear separation between faithful poles and spurious Froissart doublets is 'the corner stone of the Padé Filtering method' and that 'the situation is not clear cut and is still under discussion.' Both the channel-to-channel proof of concept (Section 13) and the Virgo [5/5] extraction (Section 16) interpret PA pole/zero statistics using precisely that separation, so the central applied claim of the report is conditional on a premise the paper leaves open. The report should either supply an operational, validated separation criterion for the noisy-data regime of Sections 13 and 16, or re-frame the Virgo and channel-cascade results as illustrations of an intended workflow whose validity awaits resolution of the Froissart question.
- [Section 13] Section 13, displayed z-transform relation: the sentence 'the singularities of T_{O_j}(z), for each random realization, have to reappear among the singularities of T_{O_{j+1}}(z)' does not follow from the displayed identity T_{O_{j+1}} = (T_{E_j} T_{L_{j+1}} / T_{L_j}) T_{O_j} + T_{E_j} T_{L_{j+1}} T_{M_j} T_{I_j}. A pole of T_{O_j} at z_0 is cancelled if T_{E_j} or T_{L_{j+1}} has a zero at z_0, and the second term contributes singularities of its own; moreover, for the finite data sets used here the z-transforms are polynomials, so the statement can only be meant at the level of PA poles, while PA's are nonlinear functions of the data. The histogram version ('the local maxima of the polar histogram ... have to reappear') is stronger still and is supported only by the two Virgo figures, whose content is qualitative. Please either prove the claim under stated genericity conditions (no pole-zero cancellations, known transfer functions) or test it on a synthetic cascade with known stages, and state what the histograms are expected to show in the presence of shared versus channel-specific poles.
- [Section 16] Section 16, [5/5] skeleton: the procedure is not reproducible as written. The doublet-erasure threshold is not specified ('we erase those doublets'), the number of erased doublets is not reported, the [20/20] PA is applied to a single realization, and the promised 'check on other chunks of data' is not presented. As a result, the claimed compression of the seismicity-to-suspension transfer function to a [5/5] rational fraction cannot be verified. Please specify the erasure criterion (distance in |z|, residue size, or histogram-based rule), report the counts and the variability across realizations, and show the held-out-chunk validation, or explicitly label the [5/5] as a single-realization illustration pending that validation.
- [Section 16, Table 16] Section 16, Table 16: the correlation coefficients and Kullback-Leibler divergences are point estimates from single realizations with no uncertainty quantification and no stated null model. The within-channel KL entries are printed with an unclear exponent, and the histogram binning and number of draws are not given, so it is not possible to decide whether the contrasts (for example 6.14 versus 5.4) are statistically meaningful. For a statistical method whose output is a two-dimensional histogram, the proof of concept should include bootstrap or jackknife error bars (over draws or over disjoint time chunks) and a statement of the histogram resolution.
minor comments (6)
- [Section 11] Section 11: in the displayed formula for the [n-1/n] PA of the geometric series, the Froissart polynomial appears once as K_{n-1} and once as K_k; please use a single symbol and define its degree, and add an explicit inline citation to Gilewicz and Pindor (JCAM 105, 1999, already in the bibliography) at the point where the general rational case is asserted.
- [Section 15] Section 15, pulsar paragraph: the authors are explicit that the test uses b=4 as 'a model of the model' because the physical b=9 regime is not resolvable by the PA; the section heading 'testing on signals that are expected from the Gravitational Universe' should state up front that the pulsar test is a surrogate, so that readers do not mistake it for a test of the physical signal.
- [Section 2] Section 2: 'We prove the validity of the concept on actual Virgo data' overstates what Sections 13 and 16 establish given the caveats of Section 17; suggest 'provide a proof of concept subject to the open Froissart question' or an equivalent formulation.
- [Table 16] Table 16: print the diagonal KL entries in decimal form, state the histogram binning, the number of Monte Carlo draws, and the exact definition of the correlation factor between two pole distributions, so that the numbers can be reproduced.
- [Bibliography and Section 4] Bibliography and Sections 4/IX: the surname 'Kahn' should read 'Kahane' (J.-P. Kahane), and the year of the Porquerolles school is given as 2003 in Section 4 and in the bibliography but as 2005 in Section IX for E. Saff's lecture; please unify these references.
- [Section 5] Section 5a: the asymptotic statement that the z-Transform 'explodes like 2^{(N-1)/2}(1-rho)^{-1/2}' does not follow from the preceding variance expression E[T_N^2] = (1 - z^{2N})/(1 - z^2), whose square root in the regime rho^{2N} << 1 and rho roughly 1 scales as (1-rho)^{-1/2} with no exponential-in-N factor; please correct the prefactor or its typesetting.
Circularity Check
No significant circularity: the analytic derivations and known-signal benchmarks are self-contained; self-citations are not load-bearing, and the main gaps are unproven or unvalidated assertions, not circular reductions.
full rationale
The report's load-bearing analytic core is rederived in-text rather than assumed. Section 10 derives the Froissart formula for noisy rational interpolants from Cramer determinants ('Using formal manipulations on the corresponding Cramer determinants, one can prove the formula...'), and Section 11 does the same for noisy PA's ('A combination of elementary manipulation of those determinants allows one to prove the formula...'). These formulas are then checked against external benchmarks: the ringdown template has an exact [1/2] rational z-transform and the paper notes 'The PA [L/M] to this fraction is the fraction itself' (Section 15); the AR(1)/AR(2) pole locations are computed independently in Section 8 and compared with PF pole densities in Section 14. This is self-contained evidence, not a fit renamed as a prediction. The self-citations to Gilewicz-Pindor and Fournier-Pindor support extension of the Froissart formalism, but Section 2 expressly says those earlier works 'are not logical prerequisites for this report's main topic', and the basic formulas are present in the text. What remains are validation gaps, not circularity: Section 13's assertion that singularities of T_Oj 'have to reappear' among those of T_O(j+1) ignores possible cancellation by the added second term; Section 16's [5/5] skeleton is produced by an unspecified erasure of doublets and the promised 'check on other chunks of data' is not reported; and Section 17 concedes that the faithful/spurious pole separation, 'the corner stone' of PF, 'is still under discussion.' These are important correctness and reproducibility limitations, but none of them makes a stated result equal to its input by construction.
Assumptions & free parameters
free parameters (5)
- PA degrees [L/M] (choice of [10/10], [20/20], [50/50])
- Number of Monte Carlo draws M (e.g., 5000, 1300)
- Data window length N (e.g., 20 for toy patterns; 432,000 for Virgo)
- Froissart doublet erasure threshold (unspecified)
- Pulsar model parameter b set to 4 instead of physical b=9 =
b=4
assumptions (6)
- standard math Kac formula for the average density of real roots of Gaussian random polynomials is admitted.
- standard math For an infinite iid Gaussian process, the z-transform has almost surely a natural boundary on the unit circle.
- domain assumption A GW interferometer can be modeled as a linear multi-stage apparatus with convolutions and linear filters.
- ad hoc to paper Poles and zeros of noisy PA separate into faithful and Froissart doublet classes, with doublets located near the unit circle and reflecting noise.
- ad hoc to paper Pole histogram peaks of one output channel reappear in another channel for the same realization, enabling transfer-function extraction.
- domain assumption GW signal templates (chirp, ringdown, pulsar) are represented by the lowest-order models cited.
Cite this review
Pith. "Pith review of PAD\'E FILTERING, Principles and Use: an Introductory Report." pith.science (2026). https://pith.science/paper/4I7G4CYX
@misc{pith2026241208254,
author = {Pith},
title = {Pith review of: PAD\'E FILTERING, Principles and Use: an Introductory Report},
year = {2026},
howpublished = {\url{https://pith.science/paper/4I7G4CYX}},
note = {Machine review of arXiv:2412.08254}
}
read the original abstract
This report aims to provide gravitational waves data analysts with an introduction to the ideas and practice of the Pad\'e Filtering method for disentangling a signal from the noise. Technically it comes to the tracking of the zeros and singularities of random z-Transforms by noisy Pad\'e Approximants.
Figures
Figures from the paper (14 more)
Reference graph
Works this paper leans on
-
[4]
Two odd probabilistic phenomena in the complex In this paragraph we assume that the reader is somehow familiar with probability theory and deterministic Padé approximation. a. The Kac phenomenon. We consider the real roots of a real random polynomial 𝑃;(𝑧)= _𝑎< 𝑧< ,; <*" where the coefficients 𝑎< are real random gaussian variables. One observes that, for ...
work page 1943
-
[9]
Szegö polynomials, deterministic and random. a. Spectral identification via deterministic Szegö polynomials. We assume that the reader is familiar with the classical theory of orthogonal polynomials on the line. We recall here only that orthogonality is usually defined using the inner product of two functions 𝑓 and 𝑔 <𝑓,𝑔> = ¶𝑓(𝑥)𝑔(𝑥)𝑑𝜇(𝑥) . where 𝜇(𝑥) is...
work page 2005
-
[14]
∈ℝ∗ 𝑛=0 ,...𝑁−1 32 and their z-Transform satisfies Ti(z)= d
Prototypical analyticity patterns: T(z) and PA’s to T(z). Before proceeding with this description, we remind the reader that while our goal is to characterize the z-Transform in the complex, we rely in most concrete cases only on approximate hints pointed by PA’s. Moreover while 𝑇𝒟 (𝑧) is a linear transform of the data, the PA is not linear and is admitte...
work page 2016
-
[16]
Testing on Virgo output channels subject to environmental noises. We now turn to the study of actual data collected as the Virgo interferometer***. We concentrate on the coupling between the state of the detector and the environmental perturbation it is subjected to. Among them the most critical one is the micro-seismicity which governs the bad low freque...
-
[17]
Dealing with the singularities of analytic functions
Actual implementation. Generalizations. The Froissart debate. Actual implementation. As for any processing of actual data collected at an interferometric GW detector, we advise to perform the routine preprocessing, namely: - undersampling from the usual 20 kHz to 1 kHz - bandpass filtering [20 Hz, 500 Hz] 42 - remove the 50 Hz component (60 Hz in the US),...
work page 2014
Reviewed August 11, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.