{"id":"5da7d697-f494-4786-9079-45cf3808deff","arxiv_id":"1908.05788","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Uniform sampling of the spectral symbol does not generally approximate eigenvalues relative to their index, and a simple formula gives the unavoidable relative error of a discretization method.","lead":"This paper shows that a widely used shortcut for estimating the eigenvalues of a differential operator from a discretization, sampling the 'spectral symbol' on a uniform grid, fails in general to give accurate relative errors. It provides a computable condition that tells when a discretization cannot approximate the full spectrum uniformly, and numerical evidence that a specially chosen non-uniform grid can fix the problem.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.1.1's proof asserts that the eigenvalue closest to omega_tilde(x) has index ratio k(n)/n -> x; Theorem 3.4.1 only proves the forward quantile direction, so the bridge from spectral distribution to relative error is not justified as written.","rationale":"The reader's weakest assumption identifies the same load-bearing gap in the proof of Theorem 5.1.1: the quantile-matching step is asserted, not derived. I agree with that diagnosis. This is genuinely load-bearing because it is the only step that converts spectral distribution information into an index-by-index relative error statement. However, the gap appears repairable: the forward statement (3.12) provides a direct way to select k(n) with k(n)/n ~ x, and weak clustering gives enough room to avoid outliers, so the lower-bound inequality is likely still true. The equality direction relies on Corollary 3.4.1, whose uniform conclusion is also only sketched. Since the concern is a substantial but fixable gap in the proof rather than a demonstrated falsehood, the reader's CONDITIONAL verdict remains appropriate and no change is needed.","tokens_in":38500,"tokens_out":9939,"duration_ms":100507,"concrete_test":"Re-derive the lower-bound direction of Theorem 5.1.1 without the nearest-eigenvalue permutation: for each x in (0,1], take k(n)=floor(n*x), discard the o(n) outlier indices allowed by weak clustering, and apply Eq. (3.12) to conclude lambda_{k(n)}/(n+1)^2 -> omega_tilde(x) while lambda_{k(n)}/k(n)^2 -> pi^2/B^2. If this derivation succeeds, the concern is a fixable proof gap rather than a false theorem; if it needs extra hypotheses, such as a uniform version of (3.12) or control of flat portions of omega_tilde, then the theorem statement requires amendment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is in the proof of Theorem 5.1.1 (Section 5.1). After defining sigma_n(k)=1 to be the index whose weighted eigenvalue is closest to omega_tilde(x), the proof says Theorem 3.1.1 gives lambda_{k(n)}/(n+1)^2 -> omega_tilde(x) and Theorem 3.4.1 gives k(n)/n -> x. The first implication is strong attraction and is acceptable; the second is not a consequence of Theorem 3.4.1 as stated. Equation (3.12) is the forward quantile statement: if k(n)/n ~ x and lambda_{k(n)} lies in the essential range, then the eigenvalue tends to omega_tilde(x). It does not assert that the eigenvalue closest to omega_tilde(x) has that quantile index. If omega_tilde has a flat segment on [x1,x2] at level t, eigenvalues near t occupy indices roughly between n*x1 and n*x2, and the closest one may have index ratio x2 rather than an interior x in the flat segment. The proof needs lambda_{k(n)} ~ k(n)^2*pi^2/B^2, so a wrong index ratio changes the predicted error. The lower bound may be repairable by selecting k(n)=floor(n*x) and using (3.12) directly, and the equality side needs a uniform version of Corollary 3.4.1 that is only sketched; but as written the theorem is incomplete at exactly the point connecting the spectral distribution to relative eigenvalue error.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies, for a regular one-dimensional Sturm-Liouville problem discretized by a matrix method, what information the GLT spectral symbol carries about relative spectral approximation. It proves a necessary condition (Theorem 5.1.1): the limit of the maximum relative eigenvalue error is bounded below by max_{x in [0,1]} |tilde_omega(x)/(x^2 pi^2 / B^2) - 1|, with equality if there are no outliers. It also argues, with numerical experiments on Euler-Cauchy problems, that a uniform sampling of the monotone rearrangement of the spectral symbol does not in general give an accurate relative approximation of the continuous spectrum, and that the condition becomes numerically sufficient when a suitable non-uniform grid and increasing method order are used. The L1 coefficient case is presented as an example where even absolute error diverges.","tokens_in":38785,"tokens_out":3444,"duration_ms":34480,"significance":"If the main theorem is fully established, the paper gives a simple, machine-checkable necessary condition for uniform relative spectral approximation, and it sharpens the known informal reading of the spectral symbol as an eigenvalue approximation tool. The Euler-Cauchy example is analytically explicit and the numerical experiments in Tables 4, 6, and 8 directly test the claimed equality. The paper also correctly distinguishes the necessary condition from the heuristic sufficiency statements, and it honestly labels the sufficiency part as numerical evidence rather than a theorem. The limitations are that the central proof has a real gap at the quantile-matching step, and two auxiliary results are either skipped or asserted without proof.","major_comments":[{"comment":"The proof defines k(n) by sigma_n(k(n))=1, i.e. as the index whose weighted eigenvalue is closest to tilde_omega(x), and then asserts that Theorem 3.4.1 gives k(n)/n -> x. However, Theorem 3.4.1, especially equation (3.12), is only a forward quantile statement: if k(n)/n ~ x and lambda_{k(n)} lies in R_omega, then lambda_{k(n)} ~ tilde_omega(x). It does not say that the eigenvalue closest to tilde_omega(x) has that quantile index. If tilde_omega has a flat segment or jump, the closest eigenvalue can correspond to a different index ratio, and the subsequent identity lambda_{k(n)} ~ (k(n)^2 pi^2 / B^2) may fail. This is exactly the step connecting the spectral distribution to the relative error, so the proof of the lower bound is incomplete as written. The equality statement under the no-outlier assumption also needs a uniform version of Corollary 3.4.1, which is only sketched.","section":"Section 5.1, proof of Theorem 5.1.1"},{"comment":"The L1 example asserts that the spectral symbol theorem for the 3-point finite difference discretization 'works fine' for p(x) = x^{-1/2} in L1, citing [24, Theorem 10.5]. The divergence result (4.21) and the estimate (4.18) depend on this extension, but no proof or specific reference for the L1 extension is provided. Since this is the basis for the stronger claim that even absolute error diverges, the extension should be proved or replaced by a precise citation with the stated hypotheses.","section":"Section 4.4, L1 example"},{"comment":"The proof of Theorem 3.2.1(i) is omitted, and Corollary 5.1.4's proof is skipped with the sentence 'It can be proved by direct computation ... we skip the details.' Both statements are load-bearing for the applicability of Theorem 5.1.1 to the FD and IgA methods. The fixed-k convergence in Theorem 3.2.1(i) is hypothesis (a) of Proposition 5.1.1, and Corollary 5.1.4 is what guarantees that the condition (5.3) is actually verified for the methods of Sections 3.2 and 3.3. Please supply complete proofs or precise references that cover exactly these hypotheses.","section":"Section 3.2 and Appendix A; Corollary 5.1.4"}],"minor_comments":[{"comment":"The section heading contains the typo 'B-slpine'; it should read 'B-spline'.","section":"Section 3.3"},{"comment":"The numerical relative error alpha_err_k^{(n)} is defined with a reference discretization of size n' >> n, while Theorem 5.1.1 compares against the continuous eigenvalues lambda_k(L). Please clarify the relation between these two definitions, since Tables 2 and 4 use different comparisons.","section":"Definition 4.0.1 and Section 4.1"},{"comment":"Figure 12's caption says 'discrete differential operator' where it appears to mean the eigenvalues of the discrete matrix; the wording is confusing.","section":"Section 4.4 and Figure 12"},{"comment":"The phrase 'We loose convergence even for the absolute error' should read 'We lose convergence'.","section":"Section 4.1"},{"comment":"The phrase 'we disprove that in general a uniform sampling of the spectral symbol can provide an accurate relative approximation' is stronger than what the text shows; the paper gives counterexamples for specific methods and symbols, not a proof that no method can have this property. Consider softening to 'we show by examples that...'.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The central result is plausible and the numerical validation is convincing, but the proof gap in Theorem 5.1.1 is real and load-bearing. I do not see signs of circularity or of fitted parameters; the experiments compare against exact eigenvalues. The paper should be revisable within its scope if the quantile-matching step is repaired or explicitly assumed, and if the L1 extension and the skipped corollary are supplied. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this paper is right about the main target — uniform sampling of the GLT spectral symbol does not generally give relative eigenvalue accuracy — and it offers a computable necessary condition plus a clean counterexample. The Euler-Cauchy example makes the point explicit: the analytic relative error tends to a positive constant c_{alpha,k} even as the mesh refines. That is a genuine contribution, and the numerical experiments back it up.\n\nWhat is new: the negative result against uniform symbol sampling, the necessary-condition formula in Theorem 5.1.1, and the explicit lower bound c_{alpha,k}. The paper directly challenges the optimistic reading of the review [26] and shows that absolute-error convergence, which can look convincing in plots, does not imply relative accuracy. The numerical work is careful, with tables that check the claimed limits at several values of n, alpha, and order.\n\nWhere the soft spots are: the proof of Theorem 5.1.1 has a real gap. After choosing the eigenvalue closest to omega_tilde(x), the proof invokes the discrete Weyl law to claim that its index ratio k(n)/n tends to x. But Theorem 3.4.1 is a forward quantile statement: if the index ratio goes to x, the eigenvalue goes to the symbol value. It does not say the closest eigenvalue has that index ratio. If omega_tilde has a flat segment, the closest eigenvalue could sit at a different quantile, and the equality (and possibly even the lower bound) is not justified as written. I think the lower bound can likely be repaired by choosing k(n)=floor(n x) and using (3.12) directly, but the proof as it stands is incomplete at the load-bearing step.\n\nTwo auxiliary results are skipped (Theorem 3.2.1(i) and Corollary 5.1.4), and the L1 example in Section 4.4 depends on an unproved extension of the GLT symbol to L1 coefficients. The sufficiency direction — non-uniform grid plus increasing order — is supported only by numerics, which the author explicitly acknowledges. These are real issues, but they are proportionate: the central message is probably correct, and the gaps are fixable.\n\nWho this is for: people working in GLT theory, isogeometric analysis, or eigenvalue solvers that need uniform relative accuracy. It deserves a serious referee; the right outcome is a revision that closes the quantile-matching gap and fills or properly relegates the skipped proofs.\n\nBest,","headline":"Uniform sampling of the GLT spectral symbol fails for relative eigenvalue accuracy; the paper has the right counterexample and a likely-true necessary condition, but the main theorem's proof has a gap.","tokens_in":39332,"tokens_out":2699,"would_cite":false,"duration_ms":25306,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N25","65N06","65N30","65N35","65L15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A spectral symbol's shape sets a hard floor on eigenvalue error","keywords":["spectral symbol","Generalized Locally Toeplitz sequences","Sturm-Liouville eigenvalue problem","relative eigenvalue error","monotone rearrangement","finite difference discretization","isogeometric analysis","Weyl law"],"falsifier":"Take the Euler-Cauchy operator with a fixed $\\alpha$ and the 3-point central finite-difference scheme, compute $\\max_{k\\le n}|\\lambda_k^{(n)}/\\lambda_k-1|$ for increasing $n$, and compare it with the predicted $\\max_{x\\in[0,1]}|\\tilde{\\omega}(x)/(x^2\\pi^2)-1|$; if the observed maximum converges to a strictly smaller limit, the lower bound in Theorem 5.1.1 is false.","tokens_in":38226,"feed_emoji":"📉","tokens_out":8342,"duration_ms":72521,"temperature":0.7,"pith_summary":"This paper asks when a matrix discretization of a regular Sturm-Liouville eigenvalue problem can approximate the whole spectrum uniformly in relative error, not just its low eigenvalues. The answer is controlled by the monotone rearrangement $\\tilde{\\omega}$ of the spectral symbol $\\omega$: the asymptotic maximum relative eigenvalue error is bounded below by $\\max_{x\\in[0,1]} |\\tilde{\\omega}(x)/(x^{2}\\pi^{2}/B^{2})-1|$, and equals this value when no outlier eigenvalues exist. In the Euler-Cauchy example this maximum is positive, so a uniform sampling of the symbol gives a relative error that does not vanish; for each fixed $k$ the analytic relative error tends to $\\alpha/(4k^{2}\\pi^{2}+\\alpha)>0$. The paper argues this necessary condition is also numerically sufficient when the discretization is paired with a symbol-adapted non-uniform grid and an increasing approximation order, and it disproves the common practice of using uniform symbol sampling for accurate relative spectra.","feed_headline":"A spectral symbol's shape sets a hard floor on eigenvalue error","feed_subtitle":"If the symbol's rearrangement misses the square profile, uniform relative accuracy is impossible.","key_machinery":"The load-bearing object is the spectral symbol $\\omega$ of the matrix sequence, a function on $[a,b]\\times[0,\\pi]$ whose level sets describe how the discrete eigenvalues distribute, together with its monotone rearrangement $\\tilde{\\omega}$, defined as the generalized inverse of the cumulative distribution $\\varphi(t)=\\mu\\{(x,\\theta):\\omega(x,\\theta)\\le t\\}/\\mu(D)$. The discrete Weyl law (Theorem 3.4.1) is the mechanism that links the rearranged symbol to the ordered eigenvalues: an eigenvalue whose index is $k(n)$ with $k(n)/n\\to x$ must approach $\\tilde{\\omega}(x)$. Comparing this limit with the exact Sturm-Liouville eigenvalue asymptotics $\\lambda_k\\sim k^2\\pi^2/B^2$ yields the ratio $\\tilde{\\omega}(x)/(x^2\\pi^2/B^2)$, which is exactly the quantity whose maximum controls the asymptotic relative error. The Liouville transformation is used to reduce the problem to a form where $B$ and the square profile $x^2\\pi^2/B^2$ appear naturally.","core_discovery":"For a regular Sturm-Liouville problem of the form $-\\partial_x(p(x)\\partial_x u)+q(x)u=\\lambda w(x)u$ with separated boundary conditions, any matrix method whose weighted discretization has spectral symbol $\\omega$ must obey the following: the limit of the maximum relative eigenvalue error cannot be smaller than the maximum deviation of the symbol's monotone rearrangement $\\tilde{\\omega}$ from the square profile $x^{2}\\pi^{2}/B^{2}$, where $B=\\int_a^b \\sqrt{w/p}\\,dx$. If the method has no outlier eigenvalues for large $n$, this lower bound is an equality. Consequently a method satisfying the usual pointwise convergence for each fixed eigenvalue can still fail to approximate the spectrum uniformly: the relative error is forced to stay positive whenever the symbol rearrangement departs from the Laplacian-like profile. The paper exhibits this failure explicitly for the Euler-Cauchy operator with 3-point finite differences, where $\\tilde{\\omega}(x)\\neq x^2\\pi^2$ and the analytic relative error for eigenvalue $k$ tends to $\\alpha/(4k^2\\pi^2+\\alpha)>0$.","pith_inferences":["A practical design rule follows from the paper's mismatch formula: choose the grid map $\\tau$ so that the symbol's rearrangement becomes $x^2\\pi^2/B^2$, then increase order to shrink the residual error; the Euler-Cauchy computation is one template, and the same rule could be tested on other Sturm-Liouville coefficients.","For higher-dimensional operators the analogous condition would compare the symbol's rearrangement with the exact spectrum quantile function (the paper sketches the 2D Laplacian); checking it on discretizations of the Laplace-Beltrami operator or wave-control problems would be a natural next test.","The equality in Theorem 5.1.1 should persist after removing the fixed number of outlier eigenvalues in IgA methods; a careful outlier-excluded numerical study would turn the current validation into a sharper test.","If the observed sufficiency holds generally, symbol-based analysis could be used to design spectrally accurate schemes for applications such as structural vibration and uniform observability, where uniform eigenvalue fidelity matters more than pointwise convergence."],"forward_implications":["Any discretization whose symbol rearrangement deviates from $x^2\\pi^2/B^2$ on a set of positive measure cannot approximate all eigenvalues uniformly in relative error, however fine the mesh.","On a uniform grid, raising the order $\\eta$ of a finite-difference or IgA scheme does not automatically help: for each fixed $\\eta$ the maximum relative error has a positive lower bound in the Euler-Cauchy case.","A uniform sampling of the spectral symbol can be visually perfect while carrying a fixed positive relative error for the first eigenvalues, so symbol-based eigenvalue estimates need a relative-error check, not just an absolute-error check.","In the absence of outliers, the computable quantity $\\max_x|\\tilde{\\omega}(x)/(x^2\\pi^2/B^2)-1|$ gives the exact asymptotic maximum relative error, making the necessary condition a practical diagnostic.","When the grid is chosen from the symbol (via the Liouville diffeomorphism) and the approximation order increases, the observed maximum relative error tends to zero, so the condition appears sufficient in that setting."],"supporting_citations":[{"why":"Supplies the GLT framework, the spectral symbol definition, and the clustering/attraction results used to identify the symbol of the FD and IgA discretizations.","marker":"[24]"},{"why":"Provides the uniform-distribution result behind the discrete Weyl law, which gives the quantile relation $\\lambda_{k(n)}\\sim\\tilde{\\omega}(k(n)/n)$.","marker":"[38]"},{"why":"Is the symbol-sampling review whose recommendation of uniform sampling the paper targets and disproves with the positive relative-error example.","marker":"[26]"},{"why":"Gives the convergence of fixed eigenvalues for finite-difference discretizations used to identify the continuous spectrum being approximated.","marker":"[27]"},{"why":"Supplies the exact eigenvalues of the discrete Dirichlet Laplacian that anchor the comparison with the symbol in the limit case.","marker":"[48]"},{"why":"Gives the $f_\\eta\\sim\\theta^2$ and uniform-convergence properties for IgA B-spline symbols needed for the monotone rearrangement analysis.","marker":"[20]"},{"why":"Proves the positivity of the trigonometric sum that makes the FD symbol factor $f_\\eta$ monotone, a property required by Corollary 3.2.1.","marker":"[5]"},{"why":"Documents the fixed number of outlier eigenvalues in IgA discretizations, which the numerical tests must exclude when checking equality in Theorem 5.1.1.","marker":"[14]"}],"fun_headline_variants":["Spectral symbol must match square profile for uniform eigenvalue accuracy","Eigenvalue error floor set by symbol's monotone rearrangement","Uniform spectrum approximation requires Laplacian-like symbol","Finite differences on Euler–Cauchy fail uniform relative error","Symbol shape forces minimum relative eigenvalue error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof's quantile-matching step assumes that the discrete eigenvalue closest to a symbol value $\\tilde{\\omega}(x)$ has index $k(n)$ with $k(n)/n\\to x$, a step asserted rather than fully justified; the sufficiency half is additionally supported only by numerical experiments.","fun_headline_variants_meta":{"raw":{"variants":["Spectral symbol must match square profile for uniform eigenvalue accuracy","Eigenvalue error floor set by symbol's monotone rearrangement","Uniform spectrum approximation requires Laplacian-like symbol","Finite differences on Euler–Cauchy fail uniform relative error","Symbol shape forces minimum relative eigenvalue error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1546,"prompt_tokens":1011,"completion_tokens":535,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":458}},"tokens_in":627,"tokens_out":535,"duration_ms":5275,"temperature":1.0,"reasoning_tokens":458,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:05:16.714785+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the Euler-Cauchy operator with a fixed $\\alpha$ and the 3-point central finite-difference scheme, compute $\\max_{k\\le n}|\\lambda_k^{(n)}/\\lambda_k-1|$ for increasing $n$, and compare it with the predicted $\\max_{x\\in[0,1]}|\\tilde{\\omega}(x)/(x^2\\pi^2)-1|$; if the observed maximum converges to a strictly smaller limit, the lower bound in Theorem 5.1.1 is false.","supporting_citations":[{"cited_title":"Garoni, S","cited_arxiv_id":null,"evidence_quote":"Supplies the GLT framework, the spectral symbol definition, and the clustering/attraction results used to identify the symbol of the FD and IgA discretizations."},{"cited_title":"Kuipers, H","cited_arxiv_id":null,"evidence_quote":"Provides the uniform-distribution result behind the discrete Weyl law, which gives the quantile relation $\\lambda_{k(n)}\\sim\\tilde{\\omega}(k(n)/n)$."},{"cited_title":"Garoni, H","cited_arxiv_id":null,"evidence_quote":"Is the symbol-sampling review whose recommendation of uniform sampling the paper targets and disproves with the positive relative-error example."},{"cited_title":"Gary, Computing eigenvalues of ordinary differential equationsby finite differences","cited_arxiv_id":null,"evidence_quote":"Gives the convergence of fixed eigenvalues for finite-difference discretizations used to identify the continuous spectrum being approximated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the exact eigenvalues of the discrete Dirichlet Laplacian that anchor the comparison with the symbol in the limit case."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the $f_\\eta\\sim\\theta^2$ and uniform-convergence properties for IgA B-spline symbols needed for the monotone rearrangement analysis."},{"cited_title":"Askey, J","cited_arxiv_id":null,"evidence_quote":"Proves the positivity of the trigonometric sum that makes the FD symbol factor $f_\\eta$ monotone, a property required by Corollary 3.2.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the fixed number of outlier eigenvalues in IgA discretizations, which the numerical tests must exclude when checking equality in Theorem 5.1.1."}],"review_version":1}