Statistical inference based on band-limited kernels: Rational-infinitely divisible distributions and beyond

Anton Ryabchenko; Vladimir Panov

REVIEW 3 major objections 5 minor 17 references

Statistical inference based on band-limited kernels: Rational-infinitely divisible distributions and beyond

T0 review · 3 major / 5 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read Band-limited kernels recover discrete atoms, continuous density and quasi-Lévy measure from mixtures, with near-parametric rates.

desk verdict Solid first nonparametric rates for discrete/continuous mixtures in class Q via band-limited kernels; the math holds under the usual separation assumption. read the letter →

arxiv 2607.05048 v1 pith:EA5XFTL5 submitted 2026-07-06 stat.ME math.PR

Vladimir Panov , Anton Ryabchenko This is my paper

classification stat.MEmath.PR MSC 62G0762G2060E07

keywords rational-infinitelydivisibledistributionsquasi-Lévymeasureband-limitedkernelsmixtureestimationalmostperiodicfunctionsempiricalcharacteristicfunctionnonparametricrates

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 studies mixture laws that combine a discrete component and an absolutely continuous component, with special attention to the larger class of rational-infinitely divisible distributions (those whose characteristic functions are ratios of infinitely divisible ones). Its goal is non-parametric recovery of the mixing weight, the atoms and their masses, the continuous density, and—when the law is rational-infinitely divisible—the signed quasi-Lévy measure. The method convolves the empirical characteristic function with band-limited kernels (functions whose Fourier transforms have compact support). Because the continuous part of the convolution vanishes at high frequencies while the discrete part remains almost periodic, the atoms can be isolated by a simple weighted average of the modulus and argument of the convolution; the continuous density and the quasi-Lévy measure are then recovered by Fourier inversion. Under a minimal separation condition on the atoms and mild smoothness or moment assumptions, the estimators attain polynomial rates, and for the discrete parameters the rates become parametric up to a logarithmic factor. Numerical experiments on Poisson–exponential mixtures illustrate the practical behaviour of the procedure.

What carries the argument

Band-limited kernels K_{c,δ} whose Fourier transforms are indicators of intervals of width 2δ. Their convolution with the characteristic function isolates individual atoms (when atoms are separated by at least 2δ) and produces an almost-periodic function whose modulus and argument yield the atom masses and locations; the same kernels also regularise the inverse-Fourier estimator of the quasi-Lévy measure.

What would settle it

Generate samples from a discrete-plus-continuous mixture whose atoms are closer than the design separation 2δ, run the band-limited estimator, and check whether the recovered atom masses and locations systematically merge or bias; any consistent failure of isolation falsifies the claimed rates.

Watch

Extended reading notes

Core claim

For mixture distributions of discrete-plus-absolutely-continuous type, and for the subclass of rational-infinitely divisible laws, convolution with band-limited kernels separates the discrete atoms from the continuous background in the Fourier domain; the resulting plug-in estimators of atom locations and masses, mixing weight, continuous density and quasi-Lévy measure converge at the polynomial (sometimes near-parametric) rates stated in Theorems 1–4.

Load-bearing premise

The atoms of the discrete component must be separated by a known minimal distance 2δ; if two atoms fall inside the same band of width 2δ the isolation argument fails.

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.

Desk Editor's Note

read the letter

This is the first clean rate result for estimating both pieces of a discrete-plus-absolutely-continuous mixture when the law sits in the rational-infinitely-divisible class Q, plus the quasi-Lévy measure. Prior Fourier work covered ordinary ID laws (Belomestny–Reiss) or the contamination model with weight >1/2; here they handle the pure mixture form (1.1) by convolving the empirical characteristic function with band-limited kernels (sinc-type) that isolate atoms under a known minimal separation 2δ.

What works: the isolation identity (3.7)–(3.8) is elementary once the kernel support is compact, the plug-in estimators for atom masses, locations and weight are explicit, and Theorems 1–4 give matching polynomial rates (parametric up to log for the discrete part when Un=n; ordinary-smooth and supersmooth rates for the continuous density; H^{-1} rates for the signed measure). Proofs in Section 6 are standard empirical-process arguments (Hoeffding + truncation for derivatives, Plancherel) and look correct. Simulations on a Poisson–exponential mixture that satisfies the Q condition illustrate the procedure and improve with n.

Soft spots are real but proportional. Assumption (A1) that atoms are separated by a known 2δ is load-bearing; if two atoms fall in the same band the isolation collapses. The fourth-moment condition for the quasi-Lévy rates is restrictive (they note it). Free tuning parameters (Un, Vn, Wn, p°) are chosen theoretically but not data-driven. No code, no real data. None of these break the theorems under the stated classes S and M.

For anyone working on nonparametric inference for Lévy-type or quasi-ID models this is useful and citable. It deserves a serious referee; I would accept it after the usual polishing on reproducibility and a short discussion of how sensitive the method is when δ is misspecified.

Referee Report

3 major / 5 minor

Summary. The paper develops nonparametric estimators for the discrete component µ_d, continuous density g_ac and mixing weight ω of a mixture µ=ω µ_d+(1-ω)µ_ac of type (1.1), and, when µ lies in the class Q of rational-infinitely divisible distributions, for the associated quasi-Lévy measure ν and its atomic part. The estimators are constructed by convolving the empirical characteristic function with band-limited kernels (compactly supported Fourier transforms) and by an inverse-Fourier procedure for the second derivative of the characteristic exponent. Under a known minimal atom separation (A1), bounded support of µ_d and mild smoothness/moment conditions, Theorems 1–2 give nearly parametric rates (up to logs) for the discrete parameters and polynomial rates for g_ac; Theorems 3–4 give polynomial rates in the H^{-1} norm for the signed measures ¯ν=x^{2}ν and ¯ν_d. Finite-sample behaviour is illustrated on a Poisson–exponential mixture.

Significance. The work supplies the first systematic frequentist nonparametric procedure for mixtures of type (1.1) inside the recently introduced class Q, complementing earlier Bayesian analysis and Fourier methods for contamination models. The nearly parametric rates for atom locations and masses under separation, the systematic use of almost-periodic functions and band-limited kernels, and the uniform H^{-1} bounds over the moment class M are genuine technical contributions. Full proofs (Section 6) via standard empirical-process and Fourier arguments, together with explicit simulation examples, make the results immediately usable for the financial, insurance and physical models that motivate the class Q. The band-limited-kernel technique itself may transfer to other almost-periodic estimation problems.

major comments (3)

[Section 3.1, Assumption (A1) and Eqs. (3.7)–(3.11)] The isolation identity F[K_{c_j,δ}](-x_k)=I{j=k} (Eq. 3.7) and the subsequent optimisation problems (3.9)–(3.11) rest entirely on a known minimal separation 2δ>0. While the class S incorporates (A1), the paper treats δ as a free modelling parameter without any discussion of how it would be chosen or estimated from data, nor of the consequences of misspecification (two atoms falling inside the same band of width 2δ). This assumption is load-bearing for every discrete-component estimator and for the subsequent quasi-Lévy estimator of ν_d; a data-driven grid construction or a robustness analysis is needed.
[Sections 3.2–4.2 and Theorems 1–4] The cut-offs U_n, V_n, W_n and the detection threshold p° enter the rates Q_n and the density/H^{-1} bounds, yet only theoretical recommendations (U_n=n, V_n=n^{1/(2(α-1))}, p°∼U_n^{-1}, W_n=n) are supplied. No practical selection rule, cross-validation scheme or sensitivity study is given. In Section 5 the authors simply fix numerical values; without guidance the claimed rates remain difficult to realise on real data.
[Section 5] The numerical study is confined to a single Poisson–exponential mixture (with and without the Q condition) and fixed tuning parameters. While the box-plots and Fourier plots illustrate consistency, they do not probe the boundary of the assumptions (smaller separation, ordinary-smooth densities with α near 3, ω near 0 or 1, unbounded discrete support). Additional experiments would better support the practical relevance of the polynomial rates.

minor comments (5)

[Global] Throughout the manuscript the Lévy symbol is rendered with extraneous spaces (“L ´evy”, “quasi-L ´evy”). Consistent typesetting would improve readability.
[Introduction and Section 2] The class notation QQQ / III is visually heavy and occasionally inconsistent with the surrounding text; a simpler bold or calligraphic font would be preferable.
[Section 5] Figures 1–6 lack quantitative error measures (e.g., integrated absolute error or H^{-1} distances) and reproducibility details (random seeds, exact numerical quadrature). Adding these would strengthen the empirical claims.
[Section 4.2, Remark 1] Remark 1 correctly notes that the fourth-moment assumption is restrictive; a short discussion of possible weakenings (or of the price paid for weaker moments) would be useful.
[Section 5.3, Lemma 2] In Section 5.3 the authors switch to the contamination representation of Proposition 1(ii) even though the model is of type (1.1). A one-sentence clarification that the two representations coincide under the stated condition on λ_1 would avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: estimators and rates are derived from the empirical characteristic function, band-limited kernels, and standard concentration/Fourier arguments under explicit modelling assumptions.

full rationale

The paper defines the estimators (bp_j, bx_j, bω, bg_ac, bν̄) directly from the empirical characteristic function convolved with known band-limited kernels K_{c_j,δ} whose Fourier transforms are indicators of fixed-width intervals (eqs. 3.5–3.15, 4.2). The isolation identity F[K_{c_j,δ}](-x_k)=I{j=k} (3.7) follows immediately from the compact support of the kernel and the modelling separation assumption (A1); it is not obtained by fitting a free parameter that is later re-used as a prediction. Convergence rates in Theorems 1–4 are proved from external empirical-process bounds (Lemmas 3–4, Hoeffding + Markov) and standard H^{-1}/Plancherel arguments already employed for Lévy measures (Belomestny–Reiss). Self-citations (Panov–Ryabchenko 2026, Belomestny–Reiss) supply background tools and the quasi-Lévy representation, not the target rates themselves. No uniqueness theorem is imported to forbid alternatives, no ansatz is smuggled via citation, and no fitted quantity is renamed a prediction. The derivation is therefore self-contained once the stated assumptions (A1, class S, class M) are granted.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claims rest on a short list of modelling assumptions (atom separation, density smoothness, moment bounds, membership in class Q) that are standard in the nonparametric literature, plus a handful of tuning sequences chosen by the statistician. No new physical entities are postulated; the free parameters are the usual bandwidth-type sequences of nonparametric Fourier estimation.

free parameters (5)

Un (frequency cut-off for discrete estimation)
Increasing sequence chosen by the user; theoretical rates require Un o∞ and Un=n for near-parametric rates.
Vn (frequency cut-off for continuous density)
Increasing sequence; optimal choice depends on smoothness class (n^{1/(2(α−1))} or √log n).
Wn (frequency cut-off for quasi-Lévy measure)
Increasing sequence; balanced against sample size to obtain the n^{−1/4} or n^{−1/6} rates.
p° (threshold for atom detection)
Positive sequence (suggested p°=c/Un) that decides which estimated masses are retained; chosen by hand.
δ (minimal atom separation)
Fixed positive constant assumed known; determines the width of the band-limited kernels.

assumptions (4)

domain assumption Atoms of the discrete component are separated by at least 2δ>0 (Assumption A1).
Used from Section 3.1 onward to guarantee that each band-limited kernel isolates at most one atom.
domain assumption The continuous density g_ac and its derivative are uniformly bounded by C (class S).
Controls the bias term I2 in the proof of Theorem 1 and the remainder in Lemma 5.
standard math Characteristic function of a distribution in class Q is bounded away from zero on the real line.
Cited from Berger–Kutlu / Khartov; used to justify the logarithm and the inverse-Fourier step for the quasi-Lévy measure.
domain assumption Fourth-moment bound E|X|^4≤C1 for the quasi-Lévy estimator (class M).
Required for the empirical-process bounds on the second derivative of the characteristic function (Lemma 4, Theorem 3).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Statistical inference based on band-limited kernels: Rational-infinitely divisible distributions and beyond." pith.science (2026). https://pith.science/paper/EA5XFTL5

@misc{pith2026260705048,
  author       = {Pith},
  title        = {Pith review of: Statistical inference based on band-limited kernels: Rational-infinitely divisible distributions and beyond},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EA5XFTL5}},
  note         = {Machine review of arXiv:2607.05048}
}

read the original abstract

This paper investigates the problem of statistical inference for a mixture distribution consisting of a discrete and a continuous component, with a particular focus on the class of rational-infinitely divisible distributions. We consider non-parametric estimation of both components of the mixture as well as the quasi-L{\'e}vy measure, assuming that the mixture belongs to the class of rational-infinitely divisible distributions. We propose an estimation framework based on band-limited kernels, which are the functions characterized by compactly supported Fourier transform. Under mild assumptions, the proposed estimators are theoretically shown to achieve polynomial (and in some cases even almost parametric) convergence rates. Finally, we demonstrate the numerical performance of the algorithm on simulated examples.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 1 linked inside Pith

[1]

Alexeev and A

I. Alexeev and A. Khartov. Spectral representations of characteristic functions of discrete probability laws. Bernoulli, 29(2):1392–1409, 2023

2023
[2]

Belomestny and M

D. Belomestny and M. Reiss. Estimation and calibration of L ´evy models via fourier methods.L ´evy Matters IV: Estimation for Discretely Observed L´evy Processes, pages 1–76, 2015

2015
[3]

Berger and M

D. Berger and M. Kutlu. Quasi-infinite divisibility of a class of distributions with discrete part.Proceedings of the American Mathematical Society, 151(5):2211–2224, 2023

2023
[4]

Berger and A

D. Berger and A. Lindner. A Cram ´er–Wold device for infinite divisibility ofZd-valued distributions.Bernoulli, 28(2):1276–1283, 2022

2022
[5]

Corduneanu.Almost Periodic Oscillations and Waves

C. Corduneanu.Almost Periodic Oscillations and Waves. Springer Science & Business Media, 2009

2009
[6]

R. Cuppens. Quelques nouveaux r ´esultats en arithm ´etique des lois de probabilit ´e.Les probabilit ´es sur les structures alg ´ebriques (Actes Colloq. Internat. CNRS, No. 186, Clermont-Ferrand, 1969), pages 97–112, 1970

1969
[7]

Demni and Z

N. Demni and Z. Mouayn. Analysis of generalized Poisson distributions associated with higher Landau levels. Infinite Dimensional Analysis, Quantum Probability and Related Topics, 18(4):1550028, 2015

2015
[8]

A. Khartov. The class Q and mixture distributions with dominated continuous singular parts.Pacific Journal of Mathematics, 343(1):139–177, 2026

2026

Show all 17 references

[9]

Lindner, L

A. Lindner, L. Pan, and K. I. Sato. On quasi-infinitely divisible distributions.Transactions of the American Mathematical Society, 370(12):8483–8520, 2018

2018
[10]

Madan, W

D. Madan, W. Schoutens, and K. Wang. Modeling the bid and ask prices of options.Journal of Computational Finance, 26(4):1–36, 2023

2023
[11]

Meister.Deconvolution Problems in Nonparametric Statistics

A. Meister.Deconvolution Problems in Nonparametric Statistics. Berlin Heidelberg: Springer Science & Business Media, 2019

2019
[12]

Nakamura

T. Nakamura. Zeta distributions generated by Dirichlet series and their (quasi) infinite divisibility.arXiv preprint, 2209.13257, 2022

2022 arXiv
[13]

Neumann and M

M. Neumann and M. Reiss. Nonparametric estimation for L ´evy processes from low-frequency observations. Bernoulli, 15(1):223–248, 2009

2009
[14]

Panov and A

V . Panov and A. Ryabchenko. Statistical inference for quasi-infinitely divisible distributions via Fourier methods.Statistics, pages 1–28, 2026

2026
[15]

Passeggeri

R. Passeggeri. On quasi-infinitely divisible random measures.Bayesian Analysis, 18(1):253–286, 2023

2023
[16]

E. T. Whittaker. On the functions which are represented by the expansions of the interpolation theory. Proceedings of the Royal Society of Edinburgh, 35:181–194, 1915

1915
[17]

Zhang, Y

H. Zhang, Y . Liu, and B. Li. Notes on discrete compound Poisson model with applications to risk theory. Insurance: Mathematics and Economics, 59:325–336, 2014

2014

Pith tools

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

[1] [1]

Alexeev and A

I. Alexeev and A. Khartov. Spectral representations of characteristic functions of discrete probability laws. Bernoulli, 29(2):1392–1409, 2023

2023

[2] [2]

Belomestny and M

D. Belomestny and M. Reiss. Estimation and calibration of L ´evy models via fourier methods.L ´evy Matters IV: Estimation for Discretely Observed L´evy Processes, pages 1–76, 2015

2015

[3] [3]

Berger and M

D. Berger and M. Kutlu. Quasi-infinite divisibility of a class of distributions with discrete part.Proceedings of the American Mathematical Society, 151(5):2211–2224, 2023

2023

[4] [4]

Berger and A

D. Berger and A. Lindner. A Cram ´er–Wold device for infinite divisibility ofZd-valued distributions.Bernoulli, 28(2):1276–1283, 2022

2022

[5] [5]

Corduneanu.Almost Periodic Oscillations and Waves

C. Corduneanu.Almost Periodic Oscillations and Waves. Springer Science & Business Media, 2009

2009

[6] [6]

R. Cuppens. Quelques nouveaux r ´esultats en arithm ´etique des lois de probabilit ´e.Les probabilit ´es sur les structures alg ´ebriques (Actes Colloq. Internat. CNRS, No. 186, Clermont-Ferrand, 1969), pages 97–112, 1970

1969

[7] [7]

Demni and Z

N. Demni and Z. Mouayn. Analysis of generalized Poisson distributions associated with higher Landau levels. Infinite Dimensional Analysis, Quantum Probability and Related Topics, 18(4):1550028, 2015

2015

[8] [8]

A. Khartov. The class Q and mixture distributions with dominated continuous singular parts.Pacific Journal of Mathematics, 343(1):139–177, 2026

2026