{"id":"2392a9c6-5698-42b4-9f94-15891bf44d0c","arxiv_id":"2507.10412","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For d-dimensional Cartesian discrete prolate matrices, the number of eigenvalues above a threshold ε is approximately (2MW)^d with an explicit error bound B_d(MW, ε).","lead":"This paper proves that a high-dimensional version of the classic 'prolate matrix' has eigenvalues that cluster tightly near 1 and 0, with a narrow transition region whose size is controlled by the time-bandwidth product. It gives explicit error bounds for images and other multidimensional signals, extending known one-dimensional results.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1.1's error term is not a valid upper bound on the stated range: for MW<1 the RHS is negative, and for ε>1/2 the log(1/ε) factor underestimates the transition width.","rationale":"The reader's weakest assumption identified the reliance on 1D bounds holding uniformly over M≤N and W∈(0,1/2). My concern is in the same region but is more concrete: the paper's own manipulation of those 1D bounds produces a negative or falsely small RHS on allowed parameters. The tensor-product reduction in Lemma 3.2 and the d-dimensional counting structure in Lemmas 4.4–4.6 are otherwise plausible; the decisive defect is in the conversion of the 1D error term, specifically the replacement of log(100MW+25)+constant by log(MW), and the use of log(1/ε) instead of χ(ε) for ε>1/2. These are not mere typos: they make Theorem 1.1 false as stated, as shown by the K=0 example. The underlying result is likely salvageable by stating the theorem only for 0<ε<1/2, using χ(ε)=log(1/(ε(1−ε))) in B_d, and replacing log(MW) with a positive logarithmic factor such as log(100MW+25) or adding an explicit MW≥1 hypothesis. Other issues noted by the reader, including the internal inconsistency in Proposition 3.4 and the overclaimed O((log(MW)log(1/ε))^d) statement in the introduction, are secondary and do not by themselves invalidate the main counting argument, but the failure of (1.3) on allowed parameters is load-bearing and requires rejection of the current version.","tokens_in":20021,"tokens_out":14561,"duration_ms":166538,"concrete_test":"Explicit numerical check: set d=1, N=100, M=5, K=0, so W=1/200 and MW=0.025. Form the 5×5 matrix A with entries 1/100 (the truncation of the K=0 bandlimiting operator); its only positive eigenvalue is 0.05. Take ε=0.01: m_ε=1, (2MW)=0.05, so the LHS of (1.3) is 0.95 while B_1(MW,ε)=log(0.025)·log(100)≈−16.99, so the inequality fails for every C_1≥0. Also test the same matrix with ε=0.8: m_ε=0 and log(1/ε)≈0.223, again contradicting (1.3). After the theorem is corrected by assuming MW≥1 and ε∈(0,1/2), rerun with M=800, N=1000, K=100 (MW=80) and ε=0.4 to confirm the amended bound.","verdict_should_be":"REJECT","load_bearing_attack":"The central bound (1.3) fails on the parameter range stated. In Lemma 4.6, the 1D estimate R_ε(MW) from Theorem 4.1, which is essentially log(100MW+25)·χ(ε)+7, is replaced by the assertion R_ε(MW) ≲ log(MW)·χ(ε). This replacement is false when MW<1 because log(MW)<0 while R_ε(MW)>0, so the resulting quantity B_d(MW,ε) can be negative or zero while the left side is positive. Explicit failure: take d=1, N=100, M=5, K=0, so W=1/200 and MW=0.025. Then A=T_MB_KT_M is the 5×5 matrix with all entries 1/100; its only positive eigenvalue is M/N=0.05. For ε=0.01, m_ε=1, so |m_ε−(2MW)|=0.95, whereas B_1(MW,ε)=log(0.025)·log(100)<0. Inequality (1.3) cannot hold for any C_1. A second independent failure occurs for ε>1/2: Lemma 4.5(i) bounds χ(ε)=log(1/(ε(1−ε))) by 2log(1/ε) only for ε≤1/2, and the proof of Lemma 4.6 tries to use symmetry to extend this to ε>1/2. But as ε→1, χ(ε)→∞ while log(1/ε)→0, and m_ε eventually becomes 0 once ε exceeds the largest eigenvalue. Thus the RHS of (1.3) tends to 0 while the LHS is (2MW)^d>0. The theorem should restrict ε to (0,1/2), use χ(ε) on the RHS, and replace log(MW) by a positive quantity such as log(100MW+25), or explicitly assume MW≥1.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper defines d-dimensional discrete prolate matrices as tensor products of one-dimensional time- and band-limiting projections, proves that their eigenvalues are products of one-dimensional prolate eigenvalues, and derives non-asymptotic eigenvalue-distribution bounds. The main claim, Theorem 1.1, states that for every epsilon>0 the number m_epsilon of eigenvalues above epsilon is within C_d B_d(MW,epsilon) of (2MW)^d, and that the number n_epsilon of eigenvalues in the transition band (epsilon,1-epsilon) is at most C_d B_d(MW,epsilon), where B_d involves log(MW)log(1/epsilon). The proof reduces the d-dimensional counting to known one-dimensional bounds of Karnik et al. and Zhu-Wakin via product-counting lemmas.","tokens_in":20356,"tokens_out":10086,"duration_ms":107969,"significance":"If the theorem is corrected, the paper would provide a useful multidimensional discrete analogue of the classical Slepian eigenvalue-concentration phenomenon, with explicit dimension-dependent transition bounds. The tensor-product reduction in Lemma 3.2 and the product-counting argument in Lemmas 4.4-4.6 are clean and give a credible strategy. The paper also makes a welcome effort to track constants and non-asymptotic dependence on the time-bandwidth product and on epsilon. However, the main theorem as stated is false on part of its declared parameter range, and the proof contains invalid estimates for small MW and for epsilon>1/2. These issues are localized and appear fixable, but they currently block acceptance.","major_comments":[{"comment":"The bound B_d(MW,epsilon) is not a valid upper bound when MW<1, because log(MW) is negative while the left-hand side is nonnegative. This is not a vacuous restriction: for d=1, N=100, M=5, K=0, we have W=1/200 and MW=0.025, and A=T_M B_K T_M is the 5x5 matrix with all entries 1/100. Its eigenvalues are 0.05,0,0,0,0, so for epsilon=0.01 we have m_epsilon=1 and |m_epsilon-(2MW)^d|=0.95, whereas B_1(0.025,0.01)=log(0.025)log(100)<0. Thus (1.3) cannot hold for any C_1. Since MW=M(2K+1)/(2N) can be arbitrarily small, the theorem must either assume MW>=1 or replace log(MW) by a positive majorant such as log(100MW+25).","section":"Theorem 1.1, Eq. (1.3); Lemma 4.6, Eq. (4.2)"},{"comment":"The first inequality of Theorem 1.1 is stated for epsilon in (0,1), but the proof only establishes the estimate for epsilon in (0,1/2). Lemma 4.5(i), which gives chi(epsilon)<=2log(1/epsilon), is valid only for epsilon<=1/2, and the final paragraph of Lemma 4.6 attempts to extend the result to epsilon>1/2 using m_epsilon <= m_{1-epsilon}. That argument is invalid because it imports the nonsymmetric factor log(1/epsilon) into the right-hand side: as epsilon approaches 1, the right side of (1.3) tends to 0 while the left side equals (2MW)^d once epsilon exceeds the largest eigenvalue. The theorem should either restrict the first estimate to epsilon in (0,1/2) or use the symmetric quantity chi(epsilon)=log(1/(epsilon(1-epsilon))) in B_d.","section":"Theorem 1.1, Eq. (1.3); Lemma 4.6, final paragraph"},{"comment":"The proof replaces the one-dimensional bound R_epsilon(MW) from Theorem 4.1 by the assertion R_epsilon(MW) \\lesssim log(MW)chi(epsilon). This is false for MW<1 because Theorem 4.1 gives R_epsilon(MW)=(2/pi^2)log(100MW+25)log(5/(epsilon(1-epsilon)))+7, which is positive and at least 7, while log(MW) tends to -infinity as MW tends to 0. This is precisely the step that produces the negative or zero right-hand side in the counterexample above. A correct proof needs a positive majorant such as log(100MW+25), with an additive constant handled explicitly.","section":"Lemma 4.6, proof around Eq. (4.4)"}],"minor_comments":[{"comment":"Proposition 3.4 appears to be incorrect as stated: the estimate #Z_W=2(M-r) with r approximately MW is inconsistent with the concluding bound |#Z_W-2floor(MW)|<=2, and the proof counts zeros by treating every j with j/(2W)<=M-1 as a zero, although one must additionally require 2WDelta to be an integer. For example, with W=1/4 and M=100, the proof would give #Z_W=98 while 2floor(MW)=50. Since this proposition is not used in the proof of Theorem 1.1, it does not affect the main result, but it should be corrected or removed.","section":"Section 3.0.1, Proposition 3.4"},{"comment":"The last line of the proof of Proposition 4.3 contains the typo '2NW' where '2MW' is intended; the printed statement otherwise uses M and W consistently.","section":"Proposition 4.3, last line"},{"comment":"References [22] and [23] list the same paper by Zhu and Wakin with different page ranges; Theorem 4.2 cites [23] while the related-work discussion cites [22], which will confuse readers tracking the one-dimensional threshold result.","section":"References [22] and [23]"},{"comment":"The displayed formula in Lemma 2.7, '(B_K x)[n] = prod_i (B_{K_i})(n_i)', is not a meaningful operator identity as written; it should either be stated for separable inputs x=x_1 tensor ... tensor x_d or be replaced by the corresponding entrywise product formula for the kernel.","section":"Lemma 2.7"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me get straight to it: the d-dimensional counting result is a legitimate extension, but Theorem 1.1 as stated is false on its declared parameter range. The tensor-product reduction in Lemma 3.2 is clean, and the product-counting inequalities in Lemmas 4.4–4.6 are correct assuming the cited 1D bounds. That is real work, and the paper deserves credit for it.\n\nThe soft spot is not in the reduction; it is in Lemma 4.6's replacement of R_ε(MW) ≈ log(100MW+25) χ(ε) by log(MW) χ(ε). For MW < 1, log(MW) is negative while the left side is nonnegative, so (1.3) fails. The stress-test example is on point: d=1, N=100, M=5, K=0 gives MW=0.025, a single eigenvalue 0.05, and the bound is negative. The same misstep breaks the ε∈(0,1) claim: for ε>1/2, log(1/ε)→0 while m_ε eventually drops to 0, so the RHS cannot dominate (2MW)^d. The theorem is only plausible for MW≥c>0 and ε∈(0,1/2), with χ(ε) kept on the RHS. The fix is simple, but the statement needs to be corrected.\n\nOther issues: Proposition 3.4's statement says 2⌊MW⌋ while the proof gives 2⌊2MW⌋; that inconsistency should be fixed. The introduction's claim that B_d scales as O((log(MW) log(1/ε))^d) is also misleading, since the (2MW)^{d-1} term in the max dominates for large MW.\n\nWho is this for? A specialist in time-frequency analysis or computational harmonic analysis who wants non-asymptotic eigenvalue estimates in higher dimensions. The central idea is right and the proof strategy is sound once the parameter ranges are fixed. It deserves a serious referee—not a desk reject—but the referee should require the parameter correction and a clean restatement. If the authors fix the ranges, I would cite it.","headline":"A legitimate d-dimensional extension of the 1D prolate eigenvalue bounds, but Theorem 1.1 as stated is false on its parameter range; the fix is straightforward.","tokens_in":20994,"tokens_out":3557,"would_cite":false,"duration_ms":37273,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A18","42A38","47B35"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that the spectrum of a d-dimensional prolate matrix concentrates: up to a controlled error, (2MW)^d eigenvalues are ε-close to 1, at most C_d log(MW) log(1/ε) max{(log(MW) log(1/ε))^{d−1}, (2MW)^{d−1}} eigenvalues lie in…","keywords":["prolate matrices","eigenvalue concentration","time-frequency localization","transition band","tensor product","multidimensional DFT","discrete prolate spheroidal sequences","non-asymptotic bounds"],"falsifier":"Compute the eigenvalue counts of the 2D prolate matrix numerically for, say, N = 100, M = 50, K = 20 (so W = 41/200 ≈ 0.205), and count eigenvalues in (0.1, 0.9). If n_{0.1} exceeds C_2 B_2(10.25, 0.1) for any apparent constant, or if m_ε deviates from (2MW)^2 = 420.25 more than the bound allows, the theorem fails. Equivalently, verify the 1D ordering λ_{⌊2MW⌋−1} ≥ 1/2 ≥ λ_{⌊2MW⌋+1} for small M and W near 1/2; any counterexample there would collapse Proposition 4.3.","tokens_in":19711,"feed_emoji":"📊","tokens_out":6563,"duration_ms":62252,"temperature":0.7,"pith_summary":"The paper extends the classical one-dimensional time–frequency concentration analysis of prolate matrices to signals defined on d-dimensional Cartesian grids. It proves a non-asymptotic eigenvalue-distribution theorem: for the multidimensional prolate matrix A = T_M B_K T_M, the number of eigenvalues above ε is pinned to (2MW)^d within an explicit error, and the width of the transition band between eigenvalues near 1 and near 0 is bounded by a quantified quantity involving log(MW), log(1/ε), and dimension d. If correct, this gives a rigorous quantitative version of the 'effective dimension' intuition: the number of jointly time-and-frequency-concentrated degrees of freedom in a d-dimensional discrete signal is approximately the time–bandwidth volume (2MW)^d, independent of the ambient grid size.","feed_headline":"Exactly (2MW)^d eigenvalues sit near 1 in every dimension","feed_subtitle":"New bounds make the d-dimensional transition band grow like (log(MW) log(1/ε))^d.","key_machinery":"The argument rests on three pieces: the multidimensional prolate matrix factors as a d-fold tensor product of one-dimensional prolate matrices, A = A_1 ⊗ ⋯ ⊗ A_d (Lemma 3.2), so its eigenvalues are products of 1D eigenvalues; a sandwiching lemma (Lemma 4.4) that relates the d-dimensional count of eigenvalues above ε to the 1D counts above ε and $ε^{{1/d}}$; and two cited one-dimensional results—the transition-region bound R_ε(MW) (Theorem 4.1, from reference [6]) and the 1/2-threshold ordering (Theorem 4.2, from reference [23])—which together yield Proposition 4.3, a sharp 1D estimate |#{λ > γ} − 2MW| ≲ R_γ(MW). Raising the 1D estimate to the d-th power and controlling the binomial error gives Lemma 4.6, which is the core of Theorem 1.1.","core_discovery":"The central claim, Theorem 1.1, is that for every d ≥ 1, M < N, K ≤ ⌊(N−1)/2⌋, W = (2K+1)/(2N) ∈ (0,1/2), and ε > 0, the eigenvalue counts of the multidimensional prolate matrix satisfy |m_ε(M,K) − (2MW)^d| ≤ C_d B_d(MW,ε) and n_ε(M,K) ≤ C_d B_d(MW,ε), where B_d(MW,ε) = log(MW) log(1/ε) max{(log(MW) log(1/ε))^{d−1}, (2MW)^{d−1}} and C_d depends only on d. In words: the bulk of the eigenvalues cluster near 1 or 0, the number of eigenvalues that are ε-close to 1 is approximately (2MW)^d, and the transition band contains at most O((log(MW) log(1/ε))^d) eigenvalues.","pith_inferences":["The same tensor-product sandwiching method likely extends to rectangular grids with different M_i and W_i per dimension, replacing (2MW)^d by ∏(2M_i W_i) with a sum of per-dimension error terms; the paper does not state this, but it follows from Lemma 3.2.","Tracking the dimension-dependent constant C_d carefully could turn Theorem 1.1 into a practical truncation rule for multidimensional DPSS-based compression, since the bulk eigenvalue count and the transition width are both explicit.","Proposition 3.4, which counts zeros of the multidimensional Dirichlet kernel, hints at a combinatorial counterpart to eigenvalue counting in higher dimensions; a testable extension would compare the transition-band width to zero-crossing counts of the product kernel."],"forward_implications":["For images and volumetric data, the number of significant degrees of freedom in a time–frequency window is quantitatively (2MW)^d, giving a principled dimension reduction for multidimensional signals.","The non-asymptotic error bound means the concentration holds for finite grids, not just in a limiting regime, so algorithms that threshold eigenvalues of prolate matrices in d dimensions have a proven accuracy guarantee.","The transition-band bound scales as O((log(MW) log(1/ε))^d), so the 'plunge region' stays narrow relative to the number of significant eigenvalues even as the time–bandwidth product grows.","The result extends the known 1D estimates of [6] to d dimensions, closing the gap identified in the literature for non-asymptotic higher-dimensional bounds."],"supporting_citations":[{"why":"Supplies the one-dimensional transition-region bound R_ε(MW) used in Theorem 4.1 to control eigenvalues in (ε, 1−ε).","marker":"[6]"},{"why":"Supplies the ordering λ_{⌊2MW⌋−1} ≥ 1/2 ≥ λ_{⌊2MW⌋+1} used in Proposition 4.3 to count eigenvalues above the 1/2 threshold.","marker":"[23]"},{"why":"Provides the continuous higher-dimensional eigenvalue-distribution result whose discrete analogue the paper establishes.","marker":"[5]"},{"why":"Provides the basic 1D facts that eigenvalues are strictly between 0 and 1, strictly decreasing, and that the number near 1 is about 2MW, used in Theorem 3.3 and Proposition 4.3.","marker":"[18]"}],"fun_headline_variants":["Eigenvalue clustering in d dimensions: transition band is tiny","In any dimension, eigenvalues of prolate matrices concentrate near 1 and 0","Prolate matrix eigenvalues cluster near 1 and 0 in all dimensions","Eigenvalue concentration proven for multidimensional prolate matrices","d-dimensional prolate matrices: eigenvalue spectrum sharply two-valued"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that the two cited one-dimensional estimates—the transition-region bound and the 1/2-threshold ordering—hold for every M ≤ N and W ∈ (0,1/2) with the stated error R_ε(MW); if either bound has hidden regime restrictions, the multidimensional result inherits them.","fun_headline_variants_meta":{"raw":{"variants":["Eigenvalue clustering in d dimensions: transition band is tiny","In any dimension, eigenvalues of prolate matrices concentrate near 1 and 0","Prolate matrix eigenvalues cluster near 1 and 0 in all dimensions","Eigenvalue concentration proven for multidimensional prolate matrices","d-dimensional prolate matrices: eigenvalue spectrum sharply two-valued"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001492,"raw_usage":{"total_tokens":6000,"prompt_tokens":968,"completion_tokens":5032,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":4942}},"tokens_in":584,"tokens_out":5032,"duration_ms":40275,"temperature":1.0,"reasoning_tokens":4942,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:32:19.399017+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the eigenvalue counts of the 2D prolate matrix numerically for, say, N = 100, M = 50, K = 20 (so W = 41/200 ≈ 0.205), and count eigenvalues in (0.1, 0.9). If n_{0.1} exceeds C_2 B_2(10.25, 0.1) for any apparent constant, or if m_ε deviates from (2MW)^2 = 420.25 more than the bound allows, the theorem fails. Equivalently, verify the 1D ordering λ_{⌊2MW⌋−1} ≥ 1/2 ≥ λ_{⌊2MW⌋+1} for small M and W near 1/2; any counterexample there would collapse Proposition 4.3.","supporting_citations":[{"cited_title":"Improved bounds for the eigenvalues of prolate spheroidal wave functions and discrete prolate spheroidal sequences","cited_arxiv_id":null,"evidence_quote":"Supplies the one-dimensional transition-region bound R_ε(MW) used in Theorem 4.1 to control eigenvalues in (ε, 1−ε)."},{"cited_title":"Approximating sampled sinusoids and multiband signals using multiband modulated DPSS dictionaries","cited_arxiv_id":null,"evidence_quote":"Supplies the ordering λ_{⌊2MW⌋−1} ≥ 1/2 ≥ λ_{⌊2MW⌋+1} used in Proposition 4.3 to count eigenvalues above the 1/2 threshold."},{"cited_title":"On the eigenvalue distribution of spatio-spectral limiting operators in higher dimensions","cited_arxiv_id":null,"evidence_quote":"Provides the continuous higher-dimensional eigenvalue-distribution result whose discrete analogue the paper establishes."},{"cited_title":"Prolate spheroidal wave functions, Fourier analysis, and uncertainty. V- The discrete case","cited_arxiv_id":null,"evidence_quote":"Provides the basic 1D facts that eigenvalues are strictly between 0 and 1, strictly decreasing, and that the number near 1 is about 2MW, used in Theorem 3.3 and Proposition 4.3."}],"review_version":1}