{"id":"ed282fbe-bf32-4a2e-8dce-e8c9a2086aea","arxiv_id":"2502.06673","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A singular-value criterion for choosing the decimation rate in spike-train super-resolution is proposed, but the central proof is invalid.","lead":"The paper proposes a new way to choose the sampling rate in super-resolution problems: instead of trying many rates, it uses the singular values of a small matrix built from the measurements to pick the rate that best separates the spikes. The idea could make existing super-resolution algorithms faster, but the mathematical proof behind the rule has serious gaps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1's proof rests on Theorem 1, which is false: for 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 equality cannot hold.","rationale":"The reader's verdict of REJECT is strongly supported. The central contribution is the automatic selection of a decimation parameter via σ_{M+1}(Tρ)≍Δ_ρ^2. This scaling is Proposition 1, and its proof is entirely an invocation of Theorems 1 and 4. Theorem 1 is not merely unproven; it is false, and the counterexample is simple and directly in the square complex-diagonal setting required by the application. Because the amplitudes in the spike train are arbitrary complex numbers and the diagonal D in Tρ=V_n D V_n* contains phases e^{i(n-1)ρx_j} multiplying the amplitudes, D is generally non-Hermitian, so the paper cannot retreat to a Hermitian special case. The proof of Theorem 1 confuses singular values of V*DV with eigenvalues of a Hermitian matrix derived from it; the Ostrowski theorem invoked there applies to a different matrix. Thus Proposition 1 is unsupported, and the argmax rule built on it is not justified. Even if a corrected theorem were found, the additional uniformity-in-ρ of the proportionality constant would still need to be established; no argument in the paper addresses it. I find no reason to soften the rejection: the main mathematical claim is not established by the submitted proof, and the numerical validations do not compensate for the false theorem at the proof's core. I keep the reader's verdict unchanged.","tokens_in":7884,"tokens_out":18935,"duration_ms":159071,"concrete_test":"Run the 2×2 check: with V=[[1,0],[1,1]] and D=diag(1,i), compute Q=V*DV and the eigenvalues of VV*. If the singular values of Q are (2.189,0.457) while λ(VV*)=(2.618,0.382), Theorem 1 is false and Proposition 1 has no valid proof. As an additional check on the selection rule, scan ρ over I for a two-cluster configuration with non-real amplitudes and plot σ_{M+1}(Tρ) versus Δ_ρ^2; if the ratio σ_{M+1}/Δ_ρ^2 varies with ρ beyond the stated O(1) constant, the argmax rule is unjustified even if the asymptotic scaling happened to hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Proposition 1, whose proof is the one-line 'Combining Theorems 1 and 4'. Theorem 1 asserts that for any diagonal complex D and square V, σ_i(V*DV)=θ_i λ_i(VV*) with λ_n(|D|)≤θ_i≤λ_1(|D|), where |D|=(DD*)^{1/2}. This is false. Take V=[[1,0],[1,1]] and D=diag(1,i). Then |D|=I, so the theorem forces σ_i(V*DV)=λ_i(VV*) exactly. Direct computation gives Q=V*DV=[[1+i,i],[i,i]] with singular values σ_1=2.189 and σ_2=0.457, whereas VV* has eigenvalues 2.618 and 0.382. The ratios are ordered in the opposite direction, so no θ_i within the claimed interval can realize both. The proof breaks because the Ostrowski step is applied to (D*)^{1/2}VV*D^{1/2}, which is Hermitian, but that matrix is not V*DV when D is not Hermitian; moreover |λ_i(V*DV)| is not equal to σ_i(V*DV) for non-normal V*DV. In the application Tρ=V_n D V_n* with D=diag(a_j e^{i(n-1)ρx_j}) complex non-Hermitian, so Proposition 1 inherits the gap. The subsequent rule ρ=argmax σ_{M+1}(Tρ) additionally assumes the proportionality constant in σ_{M+1}≍Δ_ρ^2 does not depend on ρ; even a corrected version of Theorem 1 would not establish this unless the θ_i are shown uniform in ρ, and the paper does not do so. The central claim selecting an optimal sampling rate is therefore unsupported as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":1633,"tokens_out":1695,"duration_ms":114933,"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":[{"comment":"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.","section":"Theorem 1 (Section III-A)"},{"comment":"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.","section":"Proposition 1 (Section III-A)"},{"comment":"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.","section":"Section IV-A"}],"minor_comments":[{"comment":"There is a typo: 'V andermonde' should be 'Vandermonde'.","section":"Definition 5"},{"comment":"'there exist an M-partition' should be 'there exists an M-partition'.","section":"Definition 4"},{"comment":"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).","section":"General notation"},{"comment":"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.","section":"Algorithm 1"},{"comment":"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.","section":"Figure 1"},{"comment":"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.","section":"Section IV-A"},{"comment":"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.","section":"Remark 3"}],"recommendation":"reject","confidential_remarks":"The paper's central mathematical result depends on a theorem that is demonstrably false, and the main selection rule is not justified even if the scaling statement were true. The numerical experiments are encouraging but do not repair the proof gap. I see no way to accept the manuscript in its current form; a resubmission would need a correct proof of the singular-value scaling (or a different selection mechanism) and a direct validation of the argmax rule."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague —\n\nThe novel idea here is to pick the decimation rate ρ by maximizing σ_{M+1} of the sample Toeplitz matrix, replacing the oracle used in Batenkov–Goldman–Yomdin. That is a real gap in the literature, and the complexity analysis for EDP and DMP (O(Ω log Ω) versus O(Ω²) or O(Ω³)) is a genuine practical selling point. The numerical figures show the heuristic tracking Δ_ρ² in the tested configurations, so the empirical case is not empty.\n\nThe problem is the theoretical foundation. Theorem 1 is false as stated: with V = [[1,0],[1,1]] and D = diag(1,i), Q = V*DV has singular values 2.189 and 0.457, while VV* has eigenvalues 2.618 and 0.382. The claimed θ_i bounds cannot hold. The proof’s step comparing λ_i(V*DV) with λ_i(D^{1/2}VV*D^{1/2}) requires D Hermitian or at least normal; D = diag(a_j e^{i(...)}) is complex diagonal, so the Ostrowski step does not apply. Proposition 1 is then proved by two lines citing Theorems 1 and 4, but the bridge from V*DV to VDV* is missing, and the failure of Theorem 1 means the scaling law is not established.\n\nEven if a corrected theorem appears, the selection rule ρ = argmax σ_{M+1}(T_ρ) requires the proportionality constant in σ_{M+1} ≍ Δ_ρ² to be independent of ρ. The paper does not show that; the constant could vary with ρ and break the argmax. The algorithm also needs M as an input, which is not discussed beyond a remark.\n\nSo: the idea is worth exploring, but the current manuscript does not prove its central claim. I would not cite it in this form. That said, I would not desk-reject it; a referee could demand the proof be repaired or the theorem restricted to the actual Vandermonde setting, and the numerical evidence is enough to justify the effort. A serious editor should send it out, with the expectation of major revision.\n\n— your name","headline":"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.","tokens_in":8812,"tokens_out":3083,"would_cite":false,"duration_ms":26407,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F15","65T40","94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["super-resolution","decimation","Toeplitz matrix","singular values","Vandermonde matrix","spike train","Prony method","matrix pencil"],"falsifier":"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.","tokens_in":7681,"feed_emoji":"🎯","tokens_out":7456,"duration_ms":53918,"temperature":0.7,"pith_summary":"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.","feed_headline":"The best sampling rate maximizes one singular value","feed_subtitle":"For clustered spike trains, the (M+1)-th singular value of the sample Toeplitz matrix tracks node separation, giving a parameter-free…","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the clustered-Vandermonde spectral scaling and the definitions of cluster configurations used in Theorem 4.","marker":"[3]"},{"why":"Provides the lower-bound theorem (Proposition 7.1) used to control the singular values of Vn in Theorem 2.","marker":"[4]"},{"why":"Introduces decimation for super-resolution, defines admissible rates, and provides the min-max error bounds and Lemma 5.6 used in the proofs.","marker":"[5]"},{"why":"Supplies Ostrowski's theorem, the step used inside Theorem 1 to relate singular values of V*DV to eigenvalues of VV*.","marker":"[16]"},{"why":"Gives the de-aliasing technique that resolves the ρ-fold ambiguity in the decimated samples.","marker":"[9]"},{"why":"Provides Lemma 1 on eigenvalue splitting of Gram matrices, used in the proof of Theorem 2.","marker":"[12]"},{"why":"Defines the Decimated Prony baseline that EDP improves upon.","marker":"[18]"}],"fun_headline_variants":["Maximize one singular value to pick optimal sampling rate","For spike trains, optimal sampling rate maximizes a certain singular value","Improve super-resolution by maximizing this singular value","One singular value tells you the optimal sampling rate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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|.","fun_headline_variants_meta":{"raw":{"variants":["Maximize one singular value to pick optimal sampling rate","For spike trains, optimal sampling rate maximizes a certain singular value","Improve super-resolution by maximizing this singular value","One singular value tells you the optimal sampling rate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001206,"raw_usage":{"total_tokens":4869,"prompt_tokens":748,"completion_tokens":4121,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":364,"completion_tokens_details":{"reasoning_tokens":4058}},"tokens_in":364,"tokens_out":4121,"duration_ms":25613,"temperature":1.0,"reasoning_tokens":4058,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T14:43:34.625469+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"The spectral properties of vandermonde matrices with clustered nodes","cited_arxiv_id":null,"evidence_quote":"Supplies the clustered-Vandermonde spectral scaling and the definitions of cluster configurations used in Theorem 4."},{"cited_title":"Single-exponential bounds for the smallest singular value of vandermonde matrices in the sub-rayleigh regime","cited_arxiv_id":null,"evidence_quote":"Provides the lower-bound theorem (Proposition 7.1) used to control the singular values of Vn in Theorem 2."},{"cited_title":"Super-resolution of near-colliding point sources","cited_arxiv_id":null,"evidence_quote":"Introduces decimation for super-resolution, defines admissible rates, and provides the min-max error bounds and Lemma 5.6 used in the proofs."},{"cited_title":"How to get high resolution results from sparse and coarsely sampled data","cited_arxiv_id":null,"evidence_quote":"Gives the de-aliasing technique that resolves the ρ-fold ambiguity in the decimated samples."},{"cited_title":"Spectral properties of infinitely smooth kernel matrices in the single cluster limit, with applications to multivariate super-resolution","cited_arxiv_id":null,"evidence_quote":"Provides Lemma 1 on eigenvalue splitting of Gram matrices, used in the proof of Theorem 2."},{"cited_title":"Decimated prony’s method for stable super-resolution","cited_arxiv_id":null,"evidence_quote":"Defines the Decimated Prony baseline that EDP improves upon."}],"review_version":1}