{"id":"6f5e88b0-489e-4310-abc3-deeea436bbf9","arxiv_id":"2505.16345","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"GMRES convergence plateaus near Helmholtz resonances are explained by harmonic Ritz values slowly approximating small eigenvalues, and deflation of the associated modes removes the plateaus.","lead":"This paper analyzes why GMRES, a standard iterative solver, stalls when solving Helmholtz wave problems at frequencies close to resonances. It derives a convergence bound based on harmonic Ritz values and shows that deflating the resonant modes, especially when combined with a shifted Laplacian preconditioner, removes the stagnation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 5 applies Theorem 2.6 to a non-normal scattering matrix without checking Assumption 2.4 or bounding the eigenvalue condition numbers, so the HR-value plateau diagnosis is not yet supported in the non-normal case.","rationale":"The reader's weakest assumption was the diagonalizability requirement of Assumption 2.4, with Theorem 2.9 cited as mitigation. I agree that this is the main soft spot, but I sharpen it: even when A is diagonalizable, the bound in Theorem 2.6 contains the sum of individual eigenvalue condition numbers, which can be arbitrarily large for the non-normal scattering matrix. The paper does not verify that these constants are moderate, and it does not actually use Theorem 2.9 in the Section 5 diagnosis, so the pseudospectral backup does not currently support the numerical interpretation. The normal cavity benchmark is unaffected because kappa_i = 1 there, and the deflation experiments provide independent causal evidence, so the core practical conclusions likely survive. However, the claimed validation of the theoretical framework in the non-normal setting is incomplete. The requested condition is concrete and easy to check: either report the relevant spectral conditioning constants for the Section 5 example, or reframe the diagnosis using the pseudospectral bound. No internal mathematical error was found in the theorems themselves, and the dangling 'Theorem 2.4' reference in Lemma 2.5 is a minor typo, not a load-bearing flaw.","tokens_in":20409,"tokens_out":20198,"duration_ms":159626,"concrete_test":"For the matrix from Section 5 at k=23.591, estimate the departure from normality: use an Arnoldi eigensolver to approximate the eigenvalues and right/left eigenvectors of A (or of a smaller representative subproblem), form the eigenvector matrix V for the relevant spectral window, and compute kappa_2(V) or sum kappa(lambda_i). If the resulting constant is moderate (e.g., below 10^3), the Theorem 2.6 interpretation is applicable and the concern is void. If it is large, instead evaluate the pseudospectral bound of Theorem 2.9 by numerically estimating L(Gamma_c^J)/(2 pi epsilon_c^J) on a contour enclosing the complement eigenvalues; the HR-value explanation carries force only if that constant is also moderate. This settles whether the Section 5 plateau diagnosis is supported by the stated theory.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central diagnostic story for the non-normal scattering benchmark rests on Theorem 2.6. That theorem assumes A is diagonalizable (Assumption 2.4) and its right-hand side contains the factor sum_{lambda_i in Lambda_c^J} kappa(lambda_i), the individual eigenvalue condition numbers. In Section 5, the authors explicitly call A complex and non-normal, then invoke Theorem 2.6 to explain the plateau ('this stagnation can be explained using HR values and Theorem 2.6'). They do not verify diagonalizability, nor do they report kappa(lambda_i) or kappa_2(V). If the eigenvector matrix is ill-conditioned, this sum can be arbitrarily large, making the bound vacuous; the observed correlation between an HR value approaching lambda_1 and the end of the plateau would then be consistent with the bound but not explained by it. Theorem 2.9 provides a pseudospectral alternative, but it is not used in the Section 5 diagnosis, and its constant L/(2 pi epsilon_c^J) can likewise be huge for non-normal A. This is a gap between the theoretical claim and its numerical validation in the non-normal setting, rather than an internal inconsistency of the theorems.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies convergence of GMRES for finite-element Helmholtz systems near resonant wavenumbers. It proves a bound (Theorem 2.6) in which, once harmonic Ritz values at iteration l approximate a subset of eigenvalues Lambda_J, the factor max_{lambda_i in Lambda_c^J} |s_l^J(lambda_i)| approaches 1, so that subsequent convergence is controlled by the remaining eigenvalues; a pseudospectral version (Theorem 2.9) is also given. The paper then discusses deflation of eigenvectors associated with small eigenvalues and its combination with a complex-shifted Laplacian preconditioner with ILU approximation. Numerical experiments on a closed cavity (normal symmetric matrix) and an open-cavity scattering problem with PMLs (complex non-normal matrix) show plateaus in the GMRES residual histories that end when an HR value approaches a small eigenvalue, and demonstrate that deflation removes the associated plateaus.","tokens_in":20640,"tokens_out":10497,"duration_ms":90880,"significance":"The algebraic core of the paper is sound and useful: Lemma 2.5, Theorem 2.6, and Theorem 2.9 are proved by elementary arguments, the bounds are parameter-free in the sense that no quantity is fitted, and the cavity experiments directly connect the end of a plateau to an HR value approaching an eigenvalue. This is a genuinely useful diagnostic for Helmholtz solvers near resonances. The main weakness is that the non-normal scattering benchmark invokes Theorem 2.6 without verifying its diagonalizability assumption or quantifying the eigenvalue condition numbers in the bound; the pseudospectral theorem is not used for the diagnosis. With that gap repaired, the paper would be a solid contribution to the numerical analysis of GMRES for resonant Helmholtz problems.","major_comments":[{"comment":"The plateau in the scattering residual history is explained by 'HR values and Theorem 2.6', but the scattering matrix is explicitly complex and non-normal, and the paper does not verify that Assumption 2.4 (diagonalizability) holds or report the quantities kappa(lambda_i) = ||v_i|| ||w_i|| that appear in the bound. If the eigenvector matrix is ill-conditioned, the sum sum_{lambda_i in Lambda_c^J} kappa(lambda_i) can be arbitrarily large, making the bound vacuous; the observed correlation would then be consistent with the theorem but not explained by it. The authors should either (i) verify diagonalizability and show that the condition-number sum is moderate for the matrices used, (ii) use Theorem 2.9 with a concrete estimate of L(Gamma_c^J)/(2 pi epsilon_c^J), or (iii) explicitly label the Section 5 explanation as heuristic.","section":"Section 5, paragraph 'Influence of the wavenumber on the residual history'; Assumption 2.4; Theorem 2.6"},{"comment":"The diagnosis that an HR value 'approaches the smallest eigenvalue' is only qualitative. For a complex spectrum, 'smallest' is ambiguous, and the figure alone does not show the distance between the HR value and the chosen eigenvalue. Reporting the numerical values of the eigenvalue and the HR value at iterations 850 and 1100, together with the corresponding factor max|s_l^J(lambda_i)|, would make the claimed mechanism quantitative and checkable.","section":"Section 5, Figure 5.3"}],"minor_comments":[{"comment":"The sentence 'Suppose that Theorem 2.4 holds' refers to a non-existent Theorem 2.4; it should say 'Assumption 2.4'.","section":"Section 2.4, Lemma 2.5"},{"comment":"The expression q_m(A) s_l^J(A) p_l(A) is formally undefined if an HR value nu_j^(l) coincides exactly with an eigenvalue of A, since I - A/nu_j is then singular. Please state that the rational expression is understood before cancellation, or handle the equality case by continuity.","section":"Section 2.4, proof of Lemma 2.5"},{"comment":"There are several typos: 'interessed' (Section 1), 'mentionned' (Section 3.3), 'slighlty' (Section 6), and 'the bound in (2.3) less predictive' (Section 2.2) should read 'is less predictive'.","section":"Throughout"},{"comment":"The left panel of Figure 5.3 is hard to read at the scale shown; marking the specific eigenvalue and HR value used in the diagnosis would improve the presentation.","section":"Section 5, Figure 5.3"},{"comment":"It would help to clarify that the system (P_def A A_epsilon^{-1}) tilde{x} = P_def b is solved for tilde{x} and that the reconstruction u = Q b + Q_def A_epsilon^{-1} tilde{x} produces the solution of the preconditioned deflated problem; currently this is only implicit.","section":"Section 3.3, equation (3.7)"}],"recommendation":"major_revision","confidential_remarks":"The paper is fundamentally sound in its algebraic core and in the cavity experiments; the main issue is the overclaim in Section 5 regarding the non-normal case. I would be willing to accept after the authors either verify the spectral assumptions and condition numbers or reposition the non-normal discussion as heuristic. The references to prior work [24] are appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper deserves a serious referee. The core result is a genuine extension of Cao's single-eigenvalue harmonic Ritz bound to subsets of eigenvalues, and the authors use it to give a plausible mechanism for the plateau-then-drop residual histories seen when FEM Helmholtz matrices have a small eigenvalue near a resonance. Once an HR value captures the difficult eigenvalue, the factor in the bound goes to unity and GMRES behaves as if that eigenvalue were deflated. That is a clean, useful explanation, and the deflation guidance follows naturally.\n\nThe proofs are algebraic and appear sound. Lemma 2.5 uses the cancellation between the roots of the GMRES polynomial and the denominator of s_J^l, which is the right observation. Theorem 2.6 correctly separates the residual bound into the contribution of unapproximated eigenvalues and a sum of individual eigenvalue condition numbers. Theorem 2.9 gives a pseudospectral fallback, which is a reasonable response to non-normality. The cavity experiments are convincing: the table of negative eigenvalues and the residual histories line up, and deflating the 11 negative modes removes the plateaus. The restarted GMRES results are practically useful — deflation plus CSL/ILU rescues GMRES(m), and that is a concrete, reproducible claim.\n\nThe soft spots are not fatal. The biggest one is in Section 5. The scattering matrix is explicitly complex and non-normal, and the plateau diagnosis is framed as \"explained using HR values and Theorem 2.6.\" But Theorem 2.6 requires diagonalizability and contains the sum of eigenvalue condition numbers kappa(lambda_i), which can be arbitrarily large. The authors do not report kappa(lambda_i) or kappa_2(V), so in the non-normal case the bound is not shown to be tight, and the correlation between an HR value approaching one eigenvalue and the end of the plateau is plausible but not established by the theorem. They have the pseudospectral tool in hand but do not use it in the diagnosis. This is a gap between the theoretical statement and the numerical validation, not an error in the math. A careful revision should either check the conditioning or explicitly treat the Section 5 story as a heuristic and lean on Theorem 2.9.\n\nMinor things: Lemma 2.5 says \"Suppose Theorem 2.4 holds\" — it should be Assumption 2.4, a dangling reference. No code or data is shipped, but the experiments are small enough to reproduce from the description. The scattering iteration counts are high and the improvement from deflation is modest there, as the authors acknowledge.\n\nWho gets value: researchers working on Helmholtz preconditioners or GMRES convergence theory. It deserves peer review as is, with the non-normal diagnostic gap flagged for revision.\n\nRecommendation: send it to review.","headline":"Subset-based harmonic Ritz bound with a clean plateau diagnosis; the non-normal scattering test needs a conditioning check, but the mathematics holds and the paper deserves serious review.","tokens_in":21163,"tokens_out":2299,"would_cite":true,"duration_ms":17621,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F10","65N22","65N30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Each plateau in GMRES residual histories near a Helmholtz resonance is the Krylov subspace slowly resolving a small eigenvalue, and deflating that eigenvalue's eigenvector removes the plateau.","keywords":["GMRES","harmonic Ritz values","Helmholtz equation","resonance","quasi-resonance","deflation","complex shifted Laplacian preconditioner","convergence bounds"],"falsifier":"Run GMRES on a nearly defective matrix, for example a Helmholtz-like matrix with a Jordan block at the smallest eigenvalue, track harmonic Ritz values, and check whether the residual drops as soon as a harmonic Ritz value matches that eigenvalue; if the plateau persists after the match, or if deflating the corresponding eigenvector fails to remove it, the diagonalizable-based bound is not the operative mechanism.","tokens_in":20220,"feed_emoji":"📉","tokens_out":8990,"duration_ms":71707,"temperature":0.7,"pith_summary":"Near a resonance or quasi-resonance, the finite-element Helmholtz system matrix has a small eigenvalue, and GMRES convergence slows dramatically. The paper argues that the slowdown is not a uniform spectral phenomenon: each plateau in the residual history is the Krylov subspace slowly resolving one small eigenvalue through a harmonic Ritz value, and the plateau ends when that resolution succeeds. Building on an earlier single-eigenvalue bound, it proves a subset version: once harmonic Ritz values approximate a group of eigenvalues, the factor in the convergence bound involving those eigenvalues tends to one, so the remaining eigenvalues alone govern subsequent convergence. Numerical experiments on a resonant cavity and on a quasi-resonant scattering problem confirm that deflating the corresponding eigenvectors removes the plateaus, and that combining deflation with a complex shifted Laplacian preconditioner is the most robust acceleration strategy.","feed_headline":"GMRES stalls until harmonic Ritz values pin down the small eigenvalues","feed_subtitle":"New bounds trace each residual plateau to a resonance eigenvalue being captured; deflating that mode removes the stall.","key_machinery":"The machinery is the harmonic Ritz value, the root of the GMRES minimizing polynomial at iteration $l$; because the polynomial is normalized to one at zero, the residual can be factored as $p_l(z)=\\prod_j(1-z/\\nu_j^{(l)})$. The paper's central object is the rational factor $s_l^J(z)$ obtained by replacing some of those roots by the corresponding eigenvalues; the numerator cancels the denominator so the factor is well defined, and its size on the remaining spectrum controls progress after iteration $l$. A spectral-projector identity shows $s_l^J(A)=P_J^c s_l^J(A)$, which powers the pseudospectral bound, while deflation operators $P_{\\mathrm{def}}=I-AQ$ and $Q_{\\mathrm{def}}=I-QA$ built from approximate eigenvectors remove the captured modes explicitly.","core_discovery":"The central discovery is a convergence bound that tracks GMRES between any two iterations using harmonic Ritz values. For diagonalizable $A$, if at iteration $l$ the harmonic Ritz values in a set $\\mathcal{N}_J^{(l)}$ approximate the eigenvalues in $\\Lambda_J$, then $$\\frac{\\|r_{l+m}\\|_2}{\\|r_l\\|_2} \\le \\Big(\\sum_{\\lambda_i\\in\\Lambda_J^c} \\kappa(\\lambda_i)\\Big) \\max_{\\lambda_i\\in\\Lambda_J^c} |s_l^J(\\lambda_i)| \\min_{q_m\\in\\mathcal{P}$_m^{1}$}\\max_{\\lambda_i\\in\\Lambda_J^c}|q_m(\\lambda_i)|,$$ where $s_l^J(z)=\\prod_{\\lambda_j\\in\\Lambda_J}(1-z/\\lambda_j)/\\prod_{\\nu_j\\in\\mathcal{N}_J^{(l)}}(1-z/\\nu_j)$. When the harmonic Ritz values closely match the eigenvalues in $\\Lambda_J$, the factor $\\max_{\\lambda_i\\in\\Lambda_J^c}|s_l^J(\\lambda_i)|$ approaches one, meaning the approximated eigenvalues cease to hinder convergence and GMRES behaves as if they had been deflated. A pseudospectral analogue replaces the eigenvalue sum by a contour integral of the resolvent, covering cases where non-normality or near-defectiveness makes the eigenbasis story unreliable.","pith_inferences":["Beyond the paper, harmonic Ritz values could be monitored during a solve to trigger deflation dynamically, adding a deflation vector only when a value has converged to a problematic eigenvalue; this would remove the need for a priori spectral information.","The plateau-capture mechanism suggests a practical estimator: the ratio of consecutive residual norms should spike precisely while a harmonic Ritz value approaches a small eigenvalue, which could be used to decide when to enlarge the deflation space.","The bound's mechanism is not tied to the complex shifted Laplacian specifically; the same plateau structure should appear with other preconditioners whenever the preconditioned matrix is diagonalizable, which is a testable prediction for sweeping or domain-decomposition preconditioners.","In 3D problems where exact eigenvectors are unavailable, the scattering benchmark suggests that physically motivated approximate modes, such as closed-cavity eigenmodes extended by zero, may be enough to remove the plateaus."],"forward_implications":["Each plateau in a GMRES residual history near a resonance corresponds to one small eigenvalue being resolved by a harmonic Ritz value; the sharp drop begins once the approximation is good enough for the factor $|s_l^J(\\lambda_i)|$ to approach one on the remaining spectrum.","Deflating the eigenvector of the resonant mode, even an approximate mode from a simplified closed-cavity problem, removes the corresponding plateau and reduces iteration counts across an interval of wavenumbers, not only at the resonant value.","Restarted GMRES near a resonance can fail because each restart discards the Krylov basis before the small eigenvalues are captured; deflation restores convergence, and in the hardest quasi-resonant case only deflation combined with ILU-preconditioned complex shifted Laplacian succeeds.","Because the bound is stage-by-stage, it predicts nonlinear residual histories rather than a single asymptotic rate, explaining why linear-rate bounds are poor descriptive tools for these problems."],"supporting_citations":[{"why":"Provides the single-eigenvalue harmonic Ritz convergence bound that this paper generalizes to subsets of eigenvalues.","marker":"[10]"},{"why":"Introduces the quasimode trapping setting and the frequency-dependent GMRES iteration analysis that motivates the benchmarks.","marker":"[24]"},{"why":"Supplies the deflation-matrix construction and convergence conditions used for the acceleration experiments.","marker":"[33]"},{"why":"Defines the complex shifted Laplacian preconditioner and the analysis this paper combines with deflation.","marker":"[22]"},{"why":"Provides the shifted-Laplacian preconditioning framework and the ILU-based approximation used in the experiments.","marker":"[29]"},{"why":"Defines GMRES and the classical eigenvector-conditioned convergence bound used as the starting point.","marker":"[32]"},{"why":"Provides the harmonic Ritz value definition, pseudospectra, spectral projectors, and eigenvector condition number used in the bounds.","marker":"[34]"},{"why":"Shows that deflated GMRES still minimizes the original residual norm, supporting the deflation arguments.","marker":"[19]"}],"fun_headline_variants":["Harmonic Ritz values break GMRES stagnation","Deflate resonance modes to unstall GMRES","GMRES convergence bound ties stalls to harmonic Ritz values","Harmonic Ritz values unlock GMRES speed near resonances","Harmonic Ritz deflation prevents GMRES stalls"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything spectral in the main bound assumes the system matrix can be diagonalized with a full set of eigenvectors; if the finite-element Helmholtz matrix were defective or nearly defective, the simple eigenvalue-tracking diagnosis of plateaus would not follow in the stated form.","fun_headline_variants_meta":{"raw":{"variants":["Harmonic Ritz values break GMRES stagnation","Deflate resonance modes to unstall GMRES","GMRES convergence bound ties stalls to harmonic Ritz values","Harmonic Ritz values unlock GMRES speed near resonances","Harmonic Ritz deflation prevents GMRES stalls"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001329,"raw_usage":{"total_tokens":5416,"prompt_tokens":965,"completion_tokens":4451,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":4369}},"tokens_in":581,"tokens_out":4451,"duration_ms":26090,"temperature":1.0,"reasoning_tokens":4369,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:03:15.432626+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GMRES on a nearly defective matrix, for example a Helmholtz-like matrix with a Jordan block at the smallest eigenvalue, track harmonic Ritz values, and check whether the residual drops as soon as a harmonic Ritz value matches that eigenvalue; if the plateau persists after the match, or if deflating the corresponding eigenvector fails to remove it, the diagonalizable-based bound is not the operative mechanism.","supporting_citations":[{"cited_title":"Cao , A note on the convergence behavior of GMRES , Applied Numerical Mathematics, 25 (1997), pp","cited_arxiv_id":null,"evidence_quote":"Provides the single-eigenvalue harmonic Ritz convergence bound that this paper generalizes to subsets of eigenvalues."},{"cited_title":"Marchand, J","cited_arxiv_id":null,"evidence_quote":"Introduces the quasimode trapping setting and the frequency-dependent GMRES iteration analysis that motivates the benchmarks."},{"cited_title":"New Convergence Analysis of GMRES with Weighted Norms, Preconditioning and Deflation, Leading to a New Deflation Space","cited_arxiv_id":"2312.13625","evidence_quote":"Supplies the deflation-matrix construction and convergence conditions used for the acceleration experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the shifted-Laplacian preconditioning framework and the ILU-based approximation used in the experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the harmonic Ritz value definition, pseudospectra, spectral projectors, and eigenvector condition number used in the bounds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that deflated GMRES still minimizes the original residual norm, supporting the deflation arguments."}],"review_version":1}