Pith. sign in

REVIEW 3 major objections 7 minor 27 references

Selecting Optimal Sampling Rate for Stable Super-Resolution

T0 review · 3 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper proposes that the (M+1)-th singular value of the sample Toeplitz matrix tracks the minimum node separation in decimated super-resolution, and selecting the rate that maximizes this value yields the best conditioning.

desk verdict A promising heuristic for oracle-free decimation is undercut by a false supporting theorem; the selection rule may be salvageable, but the paper as written does not establish its central claim. read the letter →

arxiv 2502.06673 v1 pith:OWFCNYIQ submitted 2025-02-10 math.NA cs.ITcs.NAmath.IT

Nuha Diab This is my paper · ORCID
classification math.NAcs.ITcs.NAmath.IT MSC 65F1565T4094A12
keywords super-resolutiondecimationToeplitzmatrixsingularvaluesVandermondespiketrainPronymethodpencil
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

This paper proposes a preprocessing rule for super-resolution of spike-train signals: among all admissible decimation rates, choose the one that maximizes the (M+1)-th singular value of the Toeplitz matrix built from the Fourier samples. The paper argues that this singular value scales with the square of the minimal separation between the decimated nodes, so the maximizing rate is the one that best separates the clustered nodes. If the scaling law holds, the rule gives a parameter-free way to improve conditioning and can be plugged into any SR algorithm, yielding faster variants of Prony and matrix-pencil methods that retain accuracy. The paper also presents numerical evidence that the resulting enhanced Prony method attains the known min-max error bounds.

What carries the argument

The central object is the Toeplitz matrix Tρ of the Fourier samples at rate ρ, which admits the factorization Tρ = Vn D Vn*, where Vn is the Vandermonde matrix of the nodes and D encodes the amplitudes and a phase shift. The argument's load is carried by a claimed scaling law: combining a general theorem on singular values of V*DV with a known scaling result for Vandermonde matrices with clustered nodes yields Proposition 1, which transfers the cluster structure of Δ_ρ into the spectrum of Tρ.

What would settle it

Compute the singular values of Q = V*DV for V = [[1,0],[1,1]] and D = diag(1,i). One gets σ_1(Q)=2.189 and σ_2(Q)=0.457, while λ_1(VV*)=2.618 and λ_2(VV*)=0.382, so the equality in Theorem 1 fails; since Proposition 1 is derived directly from Theorem 1, this counterexample is enough to invalidate the scaling law as stated.

Watch

Extended reading notes

Core claim

The paper claims Proposition 1: for any positive decimation rate ρ and a node set forming M clusters, the singular values of the Toeplitz matrix Tρ = T(ρX; n) separate into groups scaling as $Δ_ρ^{0}$, $Δ_ρ^{2}$, ..., $Δ_ρ^{{2(s-1)}}$, and in particular σ_{M+1}(Tρ) ≍ $Δ_ρ^{2}$, where Δ_ρ is the minimal circular distance between decimated nodes. On this basis the paper selects ρ = argmax σ_{M+1}(Tρ) as the optimal rate, and shows numerically that the resulting enhanced Prony method attains the known min-max error bounds.

Load-bearing premise

The proof of the scaling law rests on Theorem 1, the assertion that for any diagonal D and any matrix V, the singular values of V*DV equal θ_i times the eigenvalues of VV*, with θ_i between the extreme singular values of |D|.

Editorial extensions

If this is right

  • Any super-resolution method can adopt the selection rule as a preprocessing step, replacing the oracle-based selection of the decimation parameter used in earlier work.
  • The proposed Enhanced Decimated Prony (EDP) and Decimated Matrix Pencil (DMP) methods run in O(Ω log Ω) time, versus O(Ω^2) for DP and O(Ω^3) for MP.
  • Numerical experiments show EDP attains the min-max error bounds for clustered configurations, indicating the rule preserves optimal recovery accuracy.
  • The rule requires only the (M+1)-th singular value of an n×n Toeplitz matrix for each candidate rate, making the search over the interval I cheap.

Reading between the lines

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

  • The proportionality constant in σ_{M+1}(Tρ) ≍ Δ_ρ^2 is not shown to be uniform in ρ; if it varies with ρ, the argmax rule could select a rate that is not truly optimal.
  • The paper leaves the noisy-sample version of Proposition 1 as future work; a perturbation analysis would be needed to guarantee the singular value ranking survives noise.
  • The method takes the number of clusters M as an input; a wrong M would direct the selection to the wrong singular value, so the rule inherits the difficulty of cluster-count estimation.
  • A direct numerical check of the claimed Δ_ρ^2 scaling over many random configurations and rates would independently test the rule's core assumption.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes a preprocessing method for the discrete super-resolution problem: from noisy Fourier samples, it selects a decimation parameter rho by maximizing sigma_{M+1}(T_rho), the (M+1)-th singular value of a Toeplitz matrix built from the samples, where M is the number of clusters. The theoretical basis is Proposition 1, which asserts that the singular values of T_rho scale as Delta_rho^{0}, Delta_rho^2, ..., Delta_rho^{2(s-1)} according to cluster sizes, with the proof resting on a general linear-algebra statement (Theorem 1) about matrices of the form V*DV. The paper also introduces two algorithms, EDP and DMP, and reports numerical experiments showing speed improvements over DP and MP and optimality of EDP in the min-max sense.

Significance. If the central scaling statement were established, the proposed selection rule would be a simple, parameter-free heuristic, and the reported speed-ups of EDP and DMP over existing methods would be practically valuable. The paper also draws on and credits a body of prior work on Vandermonde matrices and decimation, and it makes concrete algorithmic and empirical contributions. However, the central mathematical result is not established: Theorem 1 is false, and Proposition 1 inherits the gap. The numerical experiments do not compensate for this, because they only demonstrate the scaling in a few examples and do not validate the optimality of the argmax selection rule. The current version therefore lacks a sound basis for its main claim.

major comments (3)
  1. [Theorem 1 (Section III-A)] Theorem 1 is false as stated. Taking V = [[1,0],[1,1]] and D = diag(1,i) gives Q = V*DV = [[1+i, i],[i, i]] with singular values sigma_1(Q) = 2.189 and sigma_2(Q) = 0.457, while VV* has eigenvalues 2.618 and 0.382. Since |D| = I, the theorem would force sigma_i(Q) = lambda_i(VV*) exactly, which is false. The proof is also invalid: the matrix (D*)^{1/2} VV* D^{1/2} is not Hermitian in general for complex diagonal D, and the claimed identity |lambda_i(V*DV)| = sigma_i(V*DV) fails for non-normal matrices. Because this theorem is the only bridge between the Vandermonde singular values and the Toeplitz singular values, the proof of Proposition 1 collapses.
  2. [Proposition 1 (Section III-A)] The proof of Proposition 1 is the single sentence 'Combining Theorems 1 and 4', and it does not explain how singular values of T_rho = V_n D V_n* are controlled by the singular values of V_n* D V_n. Since Theorem 1 is false and the diagonal matrix D in (4) is complex and non-Hermitian, Proposition 1 is unsupported. In addition, Proposition 1 is stated for any rho > 0 without the restrictions of Theorem 2 (eta >= C1 and Delta <= C2/(n^2 nu)); as rho varies, collisions can make Delta_rho = 0 and the scaling cannot hold uniformly. The statement and proof need to be replaced by a correct, self-contained derivation with explicit hypotheses on rho.
  3. [Section IV-A] Even if Proposition 1 were repaired, the selection rule rho = argmax_j sigma_{M+1}(T_{rho_j}) in Section IV-A is not justified. The notation sigma_{M+1}(T_rho) asymptotically proportional to Delta_rho^2 carries an unspecified multiplicative constant, and the paper does not show that this constant is bounded uniformly over the admissible interval. Without uniform constants, maximizing sigma_{M+1} need not maximize Delta_rho or, more importantly, the actual conditioning of the recovery problem. The numerical validation in Section III-B plots the scaling but does not compare the selected rho against the true optimal rate for recovery accuracy, so the optimality claim is not empirically supported either.
minor comments (7)
  1. [Definition 5] There is a typo: 'V andermonde' should be 'Vandermonde'.
  2. [Definition 4] 'there exist an M-partition' should be 'there exists an M-partition'.
  3. [General notation] The notation ≍ is used throughout but never formally defined; please state that it means bounded above and below by constants independent of Delta (and specify which other parameters are held fixed).
  4. [Algorithm 1] In step 5(b), the expression '{\hat{\Phi}_j^\rho, \hat{a}_j \hat{\Phi}_j^t}_{j=1}^n = SRmethod(Mds)' is confusing; clarify what the output of the SR method is for the shifted samples.
  5. [Figure 1] The caption says 'The plotted p sigma_{k+1}(T_rho) values are scaled by C := n/Omega'; this should read sqrt(sigma_{k+1}(T_rho)), and the reason for this scaling should be explained.
  6. [Section IV-A] The set 'Lambda = {rho_j, }^{m}_{j=1}' has an extra comma; also, the role of the coprime condition for rho and t is stated but the noisy matching procedure is only sketched and deserves more detail.
  7. [Remark 3] Saying 'enabling randomness in algorithm 1' without elaboration is unclear; clarify whether the lower bound 1/n^2 is on the measure of admissible rho or on the search complexity.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the selection rule uses σ_{M+1}(Tρ) as a proxy for Δ_ρ^2; the paper's main defect is an invalid auxiliary theorem, not equivalence of outputs to inputs.

full rationale

The central tool is Proposition 1, which asserts σ_{M+1}(Tρ) ≍ Δ_ρ^2 and then yields the algorithm's choice ρ = argmax σ_{M+1}(Tρ_j). Walking the derivation chain: Theorem 2 gives lower bounds on Vandermonde singular values using Theorem 3 of [4] and Lemma 1 of [12]; Theorem 4 converts determinant matching into scaling for Vn; Proposition 1 combines Theorem 1 and Theorem 4 for Tρ = Vn D Vn*. None of these steps defines the target quantity in terms of the selection rule, and no fitted parameter is renamed as a prediction. The self-citation of Lemma 1 from [12] is load-bearing in the proof of Theorem 2, but that lemma is a separate statement about Gram matrices and is not the target claim; it is not invoked as a uniqueness theorem or as an ansatz that smuggles the conclusion. The numerical validation in Figure 1 is an empirical check, not an input. The serious problem is that Theorem 1 is false as stated: with V=[[1,0],[1,1]] and D=diag(1,i), σ(V*DV)=(2.189,0.457) while λ(VV*)=(2.618,0.382), so the claimed θ_i interlacing cannot hold. Thus Proposition 1 is unsupported, and the argmax rule additionally assumes ρ-uniform constants that are not established. These are issues of invalidity and missing justification, not circularity. The paper also flags in Remark 2 that the noisy-sample extension is left for future work, which is a limitation rather than a circular step. Score 2 reflects the presence of author self-citation without any construction-level circular equivalence.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. It relies on several specialized results from the author's own prior work, and the key proof step (Theorem 1) is false as stated. The algorithm also assumes knowledge of M.

free parameters (1)
  • M (number of clusters)
    Algorithm 1 requires M as input; the method does not estimate it from data, so the user must know or guess the number of clusters.
assumptions (6)
  • domain assumption The nodes form a multi-cluster configuration with h(1)=...=h(M)=Δ(X), for some η and ν.
    Used in Theorem 2 and Proposition 1; real spike configurations may not satisfy the cluster separation assumptions.
  • domain assumption Theorem 3 (Prop 7.1 of Batenkov-Goldman 2021) bounding the singular values of clustered Vandermonde matrices.
    Lifted from prior work; its conditions on η and Δ are inherited.
  • domain assumption Lemma 1 (Lemma 3 of Diab-Batenkov 2024) on the eigenvalue splitting of Gram matrices of clustered nodes.
    Imported from an arXiv preprint; the scaling of eigenvalues is a key input to Theorem 4.
  • domain assumption All amplitudes a_j are nonzero and do not cause algebraic cancellation in the Toeplitz matrix; Proposition 1's scaling can fail with opposite-sign amplitudes.
    Not stated in the paper; a concrete configuration with V=[[1,1],[1,1+δ]] and D=diag(1,-1) yields singular values that scale as δ, not δ^2.
  • standard math Ostrowski's theorem for Hermitian matrices (Horn-Johnson Thm 4.5.9).
    Used in the proof of Theorem 1.
  • standard math Lemma 5.6 of [5] bounding |e^{ix}-e^{ix'}| by |x-x'| for |x-x'|≤π/2.
    Used in the determinant computation in Theorem 4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Selecting Optimal Sampling Rate for Stable Super-Resolution." pith.science (2026). https://pith.science/paper/OWFCNYIQ

@misc{pith2026250206673,
  author       = {Pith},
  title        = {Pith review of: Selecting Optimal Sampling Rate for Stable Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OWFCNYIQ}},
  note         = {Machine review of arXiv:2502.06673}
}
read the original abstract

We investigate the recovery of nodes and amplitudes from noisy frequency samples in spike train signals, also known as the super-resolution (SR) problem. When the node separation falls below the Rayleigh limit, the problem becomes ill-conditioned. Admissible sampling rates, or decimation parameters, improve the conditioning of the SR problem, enabling more accurate recovery. We propose an efficient preprocessing method to identify the optimal sampling rate, significantly enhancing the performance of SR techniques.

Figures

Figures reproduced from arXiv: 2502.06673 by the authors.

Figure 1
Figure 1. (right) X has one cluster of size ℓ1 with SRF = 6. (left) p X has two clusters of sizes ℓ1 and ℓ2 with SRF = 3. The plotted σk+1(Tρ) values are scaled by C := n Ω for better visualization. Remark 2. We expect our estimates in Proposition 1 to hold for noisy samples, using standard perturbation analysis for singular values. We leave it for future work. IV. ALGORITHM A decimation parameter λ ∈ I is said to be admissib… view at source ↗
Figure 2
Figure 2. (right) single cluster configuration. (left) multi-cluster con￾figuration. For both experiments the noise level is 10−6 , Nρ = 900 and Nb = 3∆−1 . Remark 4. To obtain higher accuracy for DMP, we can use N ≫ 2n decimated samples if possible. B. Optimality of EDP In [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. EDP - asymptotic optimality. For cluster node x1, Kx1 (left) scales like SRF2ℓ1−2 , while the Ka1 (right) scales like SRF2ℓ1−1 . For the non-cluster node x4, both Kx4 and Ka4 are lower bounded by a constant. These scaling rates are optimal. Acknowledgment. We thank Dr. Dmitry Batenkov and Prof. Ronen Talmon for their insightful comments [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 24 canonical work pages

  1. [1]

    A music algorithm for locating small inclusions buried in a half-space from the scattering amplitude at a fixed frequency

    Habib Ammari, Ekaterina Iakovleva, and Dominique Lesselier. A music algorithm for locating small inclusions buried in a half-space from the scattering amplitude at a fixed frequency. Multiscale Modeling & Simulation, 3(3):597–628, 2005

    work page 2005

  2. [2]

    Super-resolution of generalized spikes and spectra of confluent vandermonde matrices

    Dmitry Batenkov and Nuha Diab. Super-resolution of generalized spikes and spectra of confluent vandermonde matrices. Applied and Computational Harmonic Analysis , 65:181–208, 2023

    work page 2023

  3. [3]

    The spectral properties of vandermonde matrices with clustered nodes

    Dmitry Batenkov, Benedikt Diederichs, Gil Goldman, and Yosef Yomdin. The spectral properties of vandermonde matrices with clustered nodes. Linear Algebra and its Applications , 609:37–72, 2021

    work page 2021

  4. [4]

    Single-exponential bounds for the smallest singular value of vandermonde matrices in the sub-rayleigh regime

    Dmitry Batenkov and Gil Goldman. Single-exponential bounds for the smallest singular value of vandermonde matrices in the sub-rayleigh regime. Applied and Computational Harmonic Analysis , 55:426–439, 2021

    work page 2021

  5. [5]

    Super-resolution of near-colliding point sources

    Dmitry Batenkov, Gil Goldman, and Yosef Yomdin. Super-resolution of near-colliding point sources. Information and inference, 10(2):515–572, 2021

    work page 2021

  6. [6]

    Iii super-resolution by data inversion

    Mario Bertero and Christine De Mol. Iii super-resolution by data inversion. In Progress in optics , volume 36, pages 129–178. Elsevier, 1996

    work page 1996

  7. [7]

    Signal processing for time-of-flight imaging sensors: An introduction to inverse problems in computational 3-d imaging

    Ayush Bhandari and Ramesh Raskar. Signal processing for time-of-flight imaging sensors: An introduction to inverse problems in computational 3-d imaging. IEEE Signal Processing Magazine , 33(5):45–58, 2016

    work page 2016

  8. [8]

    Vexpa: Validated exponential analysis through regular sub-sampling

    Matteo Briani, Annie Cuyt, Ferre Knaepkens, and Wen-shin Lee. Vexpa: Validated exponential analysis through regular sub-sampling. Signal processing, 177:107722–, 2020

    work page 2020

Show all 27 references
  1. [9]

    How to get high resolution results from sparse and coarsely sampled data

    Annie Cuyt and Wen-shin Lee. How to get high resolution results from sparse and coarsely sampled data. Applied and Computational Harmonic Analysis, 48(3):1066–1087, 2020

    2020

  2. [10]

    The recoverability limit for superresolution via sparsity

    Laurent Demanet and Nam Nguyen. The recoverability limit for superresolution via sparsity. arXiv preprint arXiv:1502.01385 , 2015

    2015 arXiv

  3. [11]

    From esprit to es- pira: estimation of signal parameters by iterative rational approximation

    Nadiia Derevianko, Gerlind Plonka, and Markus Petz. From esprit to es- pira: estimation of signal parameters by iterative rational approximation. IMA Journal of Numerical Analysis , 43(2):789–827, 2023

    2023

  4. [12]

    Spectral properties of infinitely smooth kernel matrices in the single cluster limit, with applications to multivariate super-resolution

    Nuha Diab and Dmitry Batenkov. Spectral properties of infinitely smooth kernel matrices in the single cluster limit, with applications to multivariate super-resolution. arXiv preprint arXiv:2407.10600 , 2024

    2024

  5. [13]

    Sparse frequency estimation: Stability and algo- rithms

    Benedikt Diederichs. Sparse frequency estimation: Stability and algo- rithms. PhD thesis, Staats-und Universitätsbibliothek Hamburg Carl von Ossietzky, 2018

    2018

  6. [14]

    D.L. Donoho. Superresolution via sparsity constraints. SIAM Journal on Mathematical Analysis , 23(5):1309–1331, 1992

    1992

  7. [15]

    Efficient algorithms for sparse moment problems without separation

    Zhiyuan Fan and Jian Li. Efficient algorithms for sparse moment problems without separation. In The Thirty Sixth Annual Conference on Learning Theory , pages 3510–3565. PMLR, 2023

    2023

  8. [16]

    Matrix analysis

    Roger A Horn and Charles R Johnson. Matrix analysis . Cambridge university press, 2012

    2012

  9. [17]

    Matrix pencil method for estimating parameters of exponentially damped/undamped sinusoids in noise

    Yingbo Hua and Tapan K Sarkar. Matrix pencil method for estimating parameters of exponentially damped/undamped sinusoids in noise. IEEE Transactions on Acoustics, Speech, and Signal Processing , 38(5):814– 824, 1990

    1990

  10. [18]

    Decimated prony’s method for stable super-resolution

    Rami Katz, Nuha Diab, and Dmitry Batenkov. Decimated prony’s method for stable super-resolution. IEEE Signal Processing Letters , 2023

    2023

  11. [19]

    On the accuracy of prony’s method for recovery of exponential sums with closely spaced exponents

    Rami Katz, Nuha Diab, and Dmitry Batenkov. On the accuracy of prony’s method for recovery of exponential sums with closely spaced exponents. Applied and Computational Harmonic Analysis , 73:101687, 2024

    2024

  12. [20]

    On the condition number of vandermonde matrices with pairs of nearly-colliding nodes

    S Kunis and D Nagel. On the condition number of vandermonde matrices with pairs of nearly-colliding nodes. arxiv e-prints. arXiv preprint arXiv:1812.08645, 2018

    2018 arXiv

  13. [21]

    Super-resolution limit of the esprit algorithm

    Weilin Li, Wenjing Liao, and Albert Fannjiang. Super-resolution limit of the esprit algorithm. IEEE transactions on information theory , 66(7):4593–4608, 2020

    2020

  14. [22]

    Mathematical concepts of optical superresolution

    Jari Lindberg. Mathematical concepts of optical superresolution. Journal of Optics, 14(8):083001, 2012

    2012

  15. [23]

    A theory of computational resolution limit for line spectral estimation

    Ping Liu and Hai Zhang. A theory of computational resolution limit for line spectral estimation. IEEE Transactions on Information Theory , 67(7):4812–4827, 2021

    2021

  16. [24]

    Super-resolution of positive sources: The discrete setup

    Veniamin I Morgenshtern and Emmanuel J Candes. Super-resolution of positive sources: The discrete setup. SIAM Journal on Imaging Sciences, 9(1):412–444, 2016

    2016

  17. [25]

    Super-resolved nuclear magnetic resonance spectroscopy

    Satish Mulleti, Amrinder Singh, Varsha P Brahmkhatri, Kousik Chandra, Tahseen Raza, Sulakshana P Mukherjee, Chandra Sekhar Seelamantula, and Hanudatta S Atreya. Super-resolved nuclear magnetic resonance spectroscopy. Scientific reports, 7(1):9651, 2017

    2017

  18. [26]

    Deterministic sparse sublinear fft with improved numerical stability

    Gerlind Plonka and Therese von Wulffen. Deterministic sparse sublinear fft with improved numerical stability. Results in Mathematics, 76(2):53, 2021

    2021

  19. [27]

    Spectral analysis of signals , volume 452

    Petre Stoica, Randolph L Moses, et al. Spectral analysis of signals , volume 452. Pearson Prentice Hall Upper Saddle River, NJ, 2005

    2005

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.