Pith. sign in

REVIEW 3 major objections 5 minor 10 references

Stability estimates for adaptive focused time-frequency transforms

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Explicit Lipschitz stability estimates for the adaptive focused time-frequency transform, and continuity of the entropy-based focus functions, aimed at inversion.

desk verdict Useful stability estimate for the focused STFT, but the entropy-based focus functions are not yet admissible and the paper is a research announcement with proofs still to come. read the letter →

arxiv 2506.08637 v1 pith:JDRGIOD3 submitted 2025-06-10 math.CA eess.SP

Pierre Warion (I2M) , Bruno Torrésani (I2M) This is my paper
classification math.CAeess.SP MSC 42A3894A12
keywords adaptivetime-frequencytransformsfocusedshort-timeFouriertransformfocusfunctionsRényientropystabilityestimatesLipschitzcontinuitynonlinearanalysisinverseproblem
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 claims that the adaptive, signal-dependent short-time Fourier transform is stable in the focus function: if two focus functions $\sigma,\kappa$ are close in the uniform norm, then for every square-integrable signal $f$ the two transforms differ by at most an explicit constant times $\|\sigma-\kappa\|_\infty \|f\|_2$. The same Lipschitz-type control is proved for the associated left-inverse operators and for mixed $L^p_t L^q_\omega$ norms. The paper also introduces regularized R\'enyi-entropy focus functions and proves they are continuous in both the reference focus and the analyzed signal when the regularization parameter is set to its minimal value. A sympathetic reader would care because these estimates are exactly the quantitative control needed to approximate the inverse of the nonlinear transform, an open problem. The paper states that only sketches of the proofs are given, with full proofs deferred to a forthcoming companion paper.

What carries the argument

The central object is the time-focused atom $h_{t,\omega,\sigma}(x)=\sqrt{\sigma(t)}e^{2\pi i\omega x}h(\sigma(t)(x-t))$, with the transform $L_\sigma f(t,\omega)=\langle f,h_{t,\omega,\sigma}\rangle$; the focus function $\sigma$ dilates the window $h$ differently at each time $t$. The argument is carried by explicit decay and differentiability assumptions on $h$, which let the authors write the difference of two focused transforms against the kernel $\Sigma_{\sigma,\kappa}(x,t)=\sqrt{\sigma(t)}h_\sigma(x,t)-\sqrt{\kappa(t)}h_\kappa(x,t)$ and bound it in $L^2(dt\,dx)$ using Lagrange's mean value theorem. For the entropy-based focus functions, the machinery is the regularized probability density $\rho^r_{f,\kappa}(t,\omega)=(|L_\kappa f(t,\omega)|^2+r\|f\|_2^2 u(\omega))/(\|L_\kappa f(t,\cdot)\|_2^2+r\|f\|_2^2)$, its R\'enyi entropy $g_{f,\kappa,p}(t)$, and the conversion $\sigma_{f,\kappa,p}(t)=1+A\bigl[g_{f,\kappa,p}(t)-\frac{p}{1-p}\ln\|u\|_p\bigr]$, with amplitude $A>0$ and regularization $r>0$ chosen so the focus function stays in the allowed class.

What would settle it

Take $h(t)=e^{-t^2}$, $u(\omega)=e^{-\omega^2}/\sqrt{\pi}$, a Gaussian $f$, $\kappa=1$, and $p=3$; compute $\sigma_{f,\kappa,p}$ from (15)--(17) with $r=r_{\min}$ as written in Proposition 4. If the result is not bounded below by a positive constant and does not tend to 1 at $\pm\infty$, the admissibility claim is false as stated; the displayed factor $(1-\sigma_{\min})$ in $r_{\min}$ already indicates the calculation will fail for any allowed $\sigma_{\min}<1$.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the nonlinearity of the adaptive transform is mild in the focus variable: the map $\sigma \mapsto L_\sigma f$ is Lipschitz from $(\Phi_{\sigma_{\min}},\|\cdot\|_\infty)$ to $L^2(\mathbb{R}^2)$, with the explicit constant $C_1(h,\sigma,\kappa)=\frac{\sqrt{2}}{2}\frac{\|\psi_h\|_\infty}{\sigma_{\min}}+\sqrt{\frac{\pi}{2}}\frac{\|\phi_h\|_\infty}{\sigma_{\min}^{3/2}}\min\{\|\kappa\|_\infty^{1/2},\|\sigma\|_\infty^{1/2}\}$. It further claims that the entropy-based focus functions $\sigma_{f,\kappa,p}$ satisfy $\|\sigma_{f,\kappa_2,p}-\sigma_{f,\kappa_1,p}\|_\infty \le C_3 \frac{A}{p-1}\|\kappa_2-\kappa_1\|_\infty$ and $\|\sigma_{f,\kappa,p}-\sigma_{g,\kappa,p}\|_\infty \le C_4 \frac{A}{p-1}\|f-g\|_2$, so the full pipeline from signal to focus to transform is continuous. These results are framed as a step toward solving the open inversion problem for the nonlinear transform, by combining an estimated focus function with the stable left inverses of fixed-focus transforms. The paper explicitly notes that the proofs are only sketched and full versions are planned for a later publication.

Load-bearing premise

The load-bearing premise is that the regularized entropy-based focus functions really belong to the class of allowed focus functions, bounded below by a positive constant and tending to 1 at infinity; the paper's own formula for the minimal regularization contains a factor that would force $\sigma_{\min}>1$, which cannot hold for functions that tend to 1, so this admissibility step is not yet coherent.

Editorial extensions

If this is right

  • Any focus estimate that is $\varepsilon$-accurate in the uniform norm produces a transform within $\varepsilon C_1'(h,\sigma_f)\|f\|_2$ of the true adaptive transform, by Corollary 1.
  • The left inverse is stable: applying left inverses built from nearby focus functions changes the reconstruction by at most $C_2\|\sigma-\kappa\|_\infty \|F\|_2$, so an approximate focus function gives a controlled approximate inversion.
  • The entropy-based focus function is insensitive in a controlled way to the choice of reference focus: replacing $\kappa_1$ by $\kappa_2$ moves $\sigma_{f,\kappa,p}$ by $C_3 A/(p-1)\|\kappa_2-\kappa_1\|_\infty$.
  • The focus map is continuous in the signal: $f\to g$ in $L^2$ implies $\sigma_{f,\kappa,p}\to\sigma_{g,\kappa,p}$ in $L^\infty$, so small perturbations of the signal do not cause large jumps in the adapted resolution.
  • The estimates stay explicit and depend on the entropy order $p$ only through constants, so varying $p$ does not change the structure of the bounds.

Reading between the lines

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

  • Editorial inference: the Lipschitz continuity of the focus map has the shape needed for a fixed-point inversion scheme, where one estimates $\sigma$ from the current transform and recomputes; the paper does not show the constants contract, which is the numerical question that would decide the scheme's usefulness.
  • Editorial inference: the inconsistency in Proposition 4, where the formula for $r_{\min}$ contains $1-\sigma_{\min}$ while the remark forces $\sigma_{\min}>1$, suggests a misstated threshold rather than a dead end; a corrected regularization bound or a reformulated focus-function class would restore the applicability of the stability theorems to the entropy-based construction.
  • Editorial inference: because $C_3$ depends on $\|\kappa_1\|_\infty$ and $\|\kappa_2\|_\infty$, choosing a reference focus with small uniform norm should improve the stability constant, a testable prediction for numerical experiments.
  • Editorial inference: the restriction to $L^2$ signals and the explicit dependence of constants on $\sigma_{\min}$ indicate that the stability estimates deteriorate as the allowed focus functions approach zero; extending the class to handle vanishing $\sigma_{\min}$ would require stronger window decay.
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 studies stability properties of the adaptive time-frequency transform introduced by the authors in [10], in which a signal-dependent focus function sigma_f adapts the resolution of a short-time Fourier transform. The main results are an L2 stability estimate for the transform under changes of focus function (Proposition 1), an analogous estimate for the left inverse (Proposition 2), an L_t^infty L_omega^p error bound (Proposition 3), and continuity results for a regularized Rényi-entropy-based family of focus functions with respect to the reference focus (Theorem 2) and with respect to the analyzed signal (Proposition 5). The paper also states an admissibility condition for these entropy-based focus functions (Proposition 4). The authors explicitly say that proofs are only sketched and that complete proofs will appear in a forthcoming paper.

Significance. If the results are correct, the paper establishes that the focused STFT is Lipschitz stable in the focus function, which is an important step toward the inversion scheme described in Section IV. The proposed regularized entropy-based focus functions are a natural construction and, if admissible, give concrete examples to which the stability estimates apply. The paper is honest about relying on the authors' prior work [10] for the foundational norm estimates, and the main new inequalities are plausible. However, the manuscript as submitted does not contain enough proof detail to verify Proposition 2, Proposition 4, and Theorem 2, and the admissibility statement in Proposition 4 contains a sign inconsistency that currently blocks the advertised application to entropy-based focus functions.

major comments (3)
  1. [Section III-A, Proposition 2] Proposition 2, which asserts stability of the left-inverse operator L^+_sigma, is dismissed with the single sentence that the proof essentially follows the same lines as Proposition 1. This is a central advertised stability result and the inversion scheme in Section IV explicitly depends on it. The manuscript should provide a complete proof, or at least a detailed derivation of the constant C2 and the steps that parallel Proposition 1.
  2. [Section III-B, Theorem 2] The proof sketch of Theorem 2 contains an unverifiable norm inequality. The displayed expression for V_p(t) mixes a function Q_i(t,·) with its Lp norm in a way that is not written correctly, and the bound on the log-ratio by | ||Q_1||_p - ||Q_2||_p | / (r_2 ||f||_2^2 ||u||_p) relies on the lower bound ||Q_2||_p >= r_2 ||f||_2^2 ||u||_p, which is not stated. The subsequent estimate of ||Q_1 - Q_2||_p requires applying Proposition 3 with index 2p and combining it with Lemma 1; this step is not explained. Because Theorem 2 is the main continuity result for entropy-based focus functions, the proof needs to be made complete and consistent.
  3. [Section I] The manuscript states that proofs of the main results are only briefly sketched and that complete proofs will be published in a forthcoming paper. For a journal submission, this is not sufficient for the central claims. At minimum, the proofs of Propositions 2, 4, and Theorem 2 must be included in full, or the paper should be explicitly framed as a research announcement with the complete version cited. As it stands, the referees cannot verify the load-bearing results.
minor comments (5)
  1. [Section III-B, Eq. (15) and Proposition 5] The density rho_{f,kappa,p} is undefined when f = 0 because both the numerator and denominator vanish; Proposition 5 should explicitly assume f and g are nonzero, or handle the zero case separately.
  2. [Section III-B, Proposition 4 and Theorem 2] The constant r_min depends on p and A in addition to kappa, h, and ||u||_p; the sentence in the proof sketch referring to a constant C(sigma_min, h, u) should list all parameters on which the constant depends.
  3. [Throughout] There are numerous typos and notation inconsistencies, including 'resularization' for 'regularization' in Section III-B, 'F or' in the Index Terms, and the use of ||sigma||_8 where ||sigma||_infinity is meant in Proposition 1 and Lemma 1.
  4. [Section II-B, Eq. (8)] The definition of the left inverse L^+_sigma = (1/k_sigma) L^*_sigma should clarify that k_sigma is real-valued and bounded below by a positive constant; this is asserted by reference to Theorem 1 but should be stated explicitly in the notation.
  5. [Section III-A, Remark 2] The claim that C1 depends on sigma and kappa only through their L8 norms is imprecise, because C1 also depends on sigma_min, which is a class parameter rather than a norm of the individual focus functions.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the stability estimates are proved from stated hypotheses; the only self-citation is the prior foundation in [10], and the Proposition 4 sign inconsistency is a correctness gap, not circularity.

full rationale

Walking the derivation chain, Proposition 1 bounds ||L_sigma f - L_kappa f||_L2(R2) by expanding the difference as an integral of |f|^2 against Sigma_{sigma,kappa}(x,t), then using decay condition (1) and the mean value theorem; the bound is not fed back into the definition of the difference. Corollary 1, Proposition 2, and Proposition 3 follow from the same direct expansion, with Lemma 1 proved in the paper. The entropy focus functions are explicitly defined in equations (15)-(17), and Lemma 2 and Corollary 2 bound the Renyi entropy using Lemma 1 and known norms of L_kappa f. Theorem 2 derives the Lipschitz continuity of sigma_{f,kappa,p} in kappa by expanding the entropy quotient and invoking Proposition 3, with r_min introduced as a sufficient admissibility threshold, not fitted to the target estimate. The only reliance on the authors' prior work [10] is Theorem 1, which supplies frame-type norm bounds and the original definition of focused atoms; that is an external published result, not a re-derivation of the new stability claims. No fitted parameter is renamed as a prediction, no ansatz is smuggled via citation, and no uniqueness claim is imported from the authors. One genuine defect is not circularity: the displayed r_min in Proposition 4 has denominator (1 - sigma_min), while the following remark says this imposes sigma_min > 1, contradicting the class definition 0 < sigma_min < 1 and the requirement that sigma(t) tends to 1 at infinity. Taken literally, Proposition 4 has no domain of application, so Theorem 2 is not established as written; however, this is a consistency/correctness gap, not an equivalence of output to input by construction. Score 1 reflects a single minor self-citation in the foundations; the central derivation is independent.

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

The paper's central claims rest on the authors' prior transform framework [10], standard harmonic analysis tools, and the specific choice of the regularized entropy construction. The main new objects are the regularized Renyi focus functions, which are mathematical constructions rather than newly postulated physical entities. The free parameters A, r, a, and p shape these functions but are not fitted to data.

free parameters (4)
  • Amplitude A
    Introduced in Eq. (17) to scale the deviation of sigma_{f,kappa,p} from 1. No principled choice is given, and it appears in the lower bound r_min and in the continuity constants.
  • Regularization parameter r = r_min(kappa) = A p / ((p-1)(1-sigma_min)) * ||kappa||_infty^{1-1/p} ||h||_infty^{2/p} ||h||_2^{2-2/p} ||u||_p
    Added to the numerator and denominator in (15) to keep entropy finite. Later fixed to the minimal value in Remark 3, making r a function of A, p, sigma_min, kappa, h, and u.
  • Gaussian scale a
    The regularizing density u(omega)=exp(-omega^2/a^2)/(a*sqrt(pi)) has an unspecified scale a > 0 that influences entropy bounds and the construction.
  • Renyi order p
    Order of the Renyi entropy in (16), required to be p > 2 in the construction. It affects the constants in all bounds and the expression for r_min.
assumptions (4)
  • domain assumption Window h is C^1 and satisfies decay condition (1) with bounded non-negative functions psi_h and phi_h.
    Used throughout, e.g., in Propositions 1 and 3 and Lemma 1, to control differences via the mean value theorem and to ensure Lp bounds.
  • domain assumption Focus functions belong to Phi_{sigma_min}: continuous, sigma_f >= sigma_min > 0, and sigma_f(t) -> 1 as |t| -> infinity.
    Defines the admissible class for the transform, used in Theorem 1 of [10] and in the stability estimates.
  • domain assumption Theorem 1 of [10]: the focused transform satisfies c_f ||f||_2^2 <= ||M_f||^2_L2(R2) <= C_f ||f||_2^2 with c_f and C_f depending on sigma_f.
    The foundational norm estimate from the authors' prior paper is taken as a black box; the present stability estimates build on it.
  • standard math Standard Lp interpolation, Fubini, change of variables, and Parseval-Plancherel theorems.
    Used in the proofs of Proposition 1 and Lemma 1, e.g., interpolation yields the p-bounds for Lemma 1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stability estimates for adaptive focused time-frequency transforms." pith.science (2026). https://pith.science/paper/JDRGIOD3

@misc{pith2026250608637,
  author       = {Pith},
  title        = {Pith review of: Stability estimates for adaptive focused time-frequency transforms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JDRGIOD3}},
  note         = {Machine review of arXiv:2506.08637}
}
read the original abstract

This contribution is a follow-up of a recent paper by the authors on adaptive, non-linear time-frequency transforms, focusing on the STFT based transforms. The adaptivity is provided by a focus function, that depends on the analyzed function or signal, and that adapts dynamically the time-frequency resolution of the analysis. Sticking to the continuous case setting, this work provides new stability results for the transform (stability with respect with the focus function). It also investigates in some details focus functions based upon regularized R{\'e}nyi entropies and show corresponding continuity results.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

10 extracted references · 10 canonical work pages

  1. [10]

    A class of non-line ar adaptive time- frequency transform

    Pierre Warion and Bruno Torr´ esani. A class of non-line ar adaptive time- frequency transform. Sampling Theory, Signal Processing, and Data Analysis, 22(2):18, 2024

    work page 2024

  2. [1]

    Con- tinuous frames in Hilbert space

    Syed Twareke Ali, Jean-Pierre Antoine, and Jean-Pierre Gazeau. Con- tinuous frames in Hilbert space. Annals of Physics, 222(1):1–37, 1993

    work page 1993

  3. [2]

    Inequalities in Fourier analysis

    William Beckner. Inequalities in Fourier analysis. Annals of Mathematics, pages 159–182, 1975

    work page 1975

  4. [3]

    Formulation of the uncertainty relations in terms of the R´ enyi entropies

    Iwo Bialynicki-Birula. Formulation of the uncertainty relations in terms of the R´ enyi entropies. Phys. Rev. A, 74:052101, Nov 2006

    work page 2006

  5. [4]

    Entropy- based algorithms for best basis selection

    Ronald Raphael Coifman and Mladen Victor Wickerhauser. Entropy- based algorithms for best basis selection. IEEE Transactions on Information Theory, 38(2):713–718, March 1992

    work page 1992

  6. [5]

    Continuous warped time-frequency representations—coorbit spaces an d discretiza- tion

    Nicki Holighaus, Christoph Wiesmeyr, and Peter Balazs. Continuous warped time-frequency representations—coorbit spaces an d discretiza- tion. Applied and Computational Harmonic Analysis, 47(3):975–1013, 2019

    work page 2019

  7. [6]

    Nons tationary Gabor frames

    Florent Jaillet, Peter Balazs, and Monika D¨ orfler. Nons tationary Gabor frames. In SAMPTA ’09, pages General–session, 2009

    work page 2009

  8. [7]

    A friendly guide to wavelets

    Gerald Kaiser. A friendly guide to wavelets. Birkhauser Boston Inc., USA, 1994

    work page 1994

Show all 10 references
  1. [8]

    Automatic adaptation of the time-frequency resolut ion for sound analysis and re-synthesis

    Marco Liuni, Axel Robel, Ewa Matusiak, Marco Romito, and Xavier Rodet. Automatic adaptation of the time-frequency resolut ion for sound analysis and re-synthesis. IEEE Transactions on Audio, Speech, and Language Processing, 21(5):959–970, May 2013

    2013

  2. [9]

    Refined support an d entropic uncertainty inequalities

    Benjamin Ricaud and Bruno Torr´ esani. Refined support an d entropic uncertainty inequalities. IEEE Transactions on Information Theory, 59(7):4272–4279, July 2013

    2013

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.