{"id":"1c7ddc6f-e932-452a-b917-761d978be795","arxiv_id":"2608.05625","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"New a priori error estimators predict eigenvalue errors of Craig-Bampton and higher-order Craig-Bampton reduced models using only reduced-order eigensolutions, with numerical validation on three structural models.","lead":"This paper proposes formulas that estimate how wrong a reduced-order dynamic model's vibration frequencies are, without running the full, expensive simulation. The estimators use the gap between two successive levels of the higher-order Craig-Bampton method, and are tested on a plate, a pipe, and a reactor pressure vessel.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CB estimator's central approximation u0^T M_p u0 ≈ 1 is known to break down for high modes (ratios down to 0.76), and the exact-denominator fix is proposed but not validated, so the claim that the estimator is validated across all reported modes is not yet secured.","rationale":"The paper is a solid contribution: the hierarchical HCB construction is coherent, the derivations of Eqs. (42) and (50) are traceable, the numerical experiments are extensive across three geometries, and the limitations are disclosed in Remark 2 and Section 5. The most load-bearing weakness is not a hidden algebraic error but an acknowledged approximation whose failure regime is precisely the high-error regime that an error estimator is meant to certify. The exact-denominator variant would resolve this if validated, but the paper explicitly declines to validate it, leaving the headline claim ('validate both estimators') stronger than the evidence. I agree with much of the reader's analysis, especially the need to validate the exact-denominator variant and to delineate the reliable regime. I differ in emphasis: I see the CB estimator's unvalidated denominator fix as the single most actionable concern, since it is self-contained and testable within the paper's own framework, whereas the HCB-1 estimator's dependence on Neumann-series convergence is a limitation that the paper already documents with quantitative ratios. The reproducibility concern (no code or mesh files) is real but secondary; the proposed concrete test addresses the mathematical claim directly. A conditional verdict is appropriate: acceptance after the exact-denominator variant is validated, or after the paper explicitly restricts the CB estimator's validity to the regime where u0^T M_p u0 ≈ 1 holds and reports ratios for high-error modes accordingly.","tokens_in":22585,"tokens_out":9705,"duration_ms":108198,"concrete_test":"Recompute the CB estimator validation for the cantilever plate, including modes 11–20, using the exact-denominator formula from Remark 2: η_ex_CB,i = (φbar_i^T K00 φbar_i − λbar_i φbar_i^T M00 φbar_i)/(λbar_i φbar_i^T M00 φbar_i), with M00 = T0^T M_p T0 and K00 = T0^T K_p T0. Compare the ratios η_ex/ξ against the reported Eq. (42) ratios. If the exact-denominator ratios recover near 1.0 for all modes, then the degradation is fully explained by the denominator approximation and the central claim can be secured by switching to the exact variant. If ratios remain substantially below 1 (e.g., below 0.9), then an additional source of error exists in the perturbation derivation, and the paper's claim that Eq. (42) is a validated a priori estimator must be revised or restricted.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that Eq. (42) reliably estimates CB relative eigenvalue errors a priori. However, the derivation in Appendix D proceeds through Eq. (D5) and then invokes u0,i^T M_p u0,i ≈ 1 to reach Eq. (37)/(42). This leading-order approximation is exactly what fails when the residual component u_r,i = T_r φbar_i is not small, i.e., precisely for modes where the CB error is large. The paper acknowledges this in Remark 2 and in Section 4: for the cantilever plate, modes 11–20 produce ratios as low as 0.76, meaning the estimator underestimates the true error by about 24% in the regime where an error certificate is most needed. Remark 2 proposes an exact-denominator variant η_ex_CB,i = (φbar_i^T K00 φbar_i − λbar_i φbar_i^T M00 φbar_i)/(λbar_i φbar_i^T M00 φbar_i) that removes the approximation, but explicitly states that validation is left for future work. Thus the numerical validation in Tables 2, 4, and 6 does not actually validate the estimator in the regime where its own derivation is most questionable; it validates a cheaper approximate formula under favorable conditions. The HCB-1 estimator (Eq. (50)) has a related but separate fragility: its accuracy depends on ξ_HCB-2,i being small compared with the HCB-1-to-HCB-2 gap, i.e., on fast Neumann-series convergence. This is honestly reported (ratios 0.85–0.99), and the paper states the limitation. The more load-bearing, unresolved issue for the headline claim is the CB estimator's unvalidated exact-denominator variant, because the abstract and conclusions present Eq. (42) as validated without the regime caveat being quantified or fixed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a priori error estimators for the Craig-Bampton (CB) and higher-order Craig-Bampton (HCB) reduced-order models. The CB estimator, Eq. (42), is derived from a Rayleigh quotient perturbation analysis that decomposes an HCB eigenvector into a base (CB-like) part and a residual part, and the HCB-1 estimator, Eq. (50), is the relative eigenvalue gap between HCB-1 and HCB-2, justified by the nested Ritz subspace structure and Courant-Fischer monotonicity. The estimators require only reduced-order eigenpairs and already-assembled full-order matrices, not the full-order eigenpairs. Numerical experiments on a cantilever plate, an elbow pipe, and a reactor pressure vessel report estimation ratios close to 1 for many modes, with degradation for some high modes and for modes where the Neumann-series convergence is slow.","tokens_in":22942,"tokens_out":4427,"duration_ms":48056,"significance":"The hierarchical estimation idea is genuinely useful: it turns the nested HCB hierarchy into an error reference, avoids solving the full-order eigenproblem, and the HCB-1 estimator has a clean non-negativity property from min-max ordering. The appendices provide a consistent derivation of the residual flexibility, the coupled inertia force, and the decomposition-based estimator, which are reproducible and give confidence in the algebra. If the CB estimator's high-mode limitation were resolved, the framework would offer practitioners a practical, low-cost accuracy indicator for CMS-based ROMs. The paper also honestly reports the per-mode estimation ratios and the Neumann-series convergence dependence, which is commendable.","major_comments":[{"comment":"The central validation claim for the CB estimator is not secured for the modes where the estimator is most needed. The derivation of Eq. (D5) invokes the leading-order approximation u0,i^T M_p u0,i ≈ 1, and the numerical results in Section 4 show exactly where this fails: for the cantilever plate the CB estimator ratio degrades to 0.76 for modes 11–20, and for the pipe to 0.93. The paper's own Remark 2 acknowledges that the exact-denominator variant η_ex_CB,i removes this approximation but states that validation of this variant is left for future work. Consequently, the abstract's statement that both estimators are validated is an overstatement; the current evidence validates Eq. (42) only under favorable conditions (small residual component), and the regime of large CB errors, where an error certificate is most valuable, is left without validated support. The authors should either validate the exact-denominator variant numerically for the high-mode cases or explicitly restrict the claim to the regime u0,i^T M_p u0,i ≈ 1 and provide the condition under which the estimator can be trusted.","section":"Appendix D, Remark 2, and Section 4"},{"comment":"The statement that evaluating Eq. (50) requires only the HCB-1 and HCB-2 reduced-order eigenvalues, 'both of which are obtained by solving reduced eigenvalue problems of the same size as the CB model', is incorrect for the HCB-2 reference. As Eq. (31) shows, constructing the HCB-2 SEREP projection requires solving the augmented HCB-2 eigenvalue problem of size n_CB + 2N_b (sum N_d + 3N_b), not the CB size n_CB; only after SEREP is the final reduced eigenproblem of size n_CB. This misstates the computational overhead of the HCB-1 estimator and undermines the claim that the additional cost is negligible. The overhead discussion in Section 4 and Table 7 should be revised to reflect the true size of the HCB-2 augmented solve, which grows with N_b.","section":"Section 3.3, page 10"}],"minor_comments":[{"comment":"The definition of the coupled inertia force contains a typographical symbol 'B' where an equality sign is intended; it should read M_hat_c = M_ss Ψ_b + M_sb.","section":"Eq. (14)"},{"comment":"The caption of Figure 4 reports CB estimator ratios in the range 0.76–1.00 for all 20 modes, while Table 2, which lists only the first 10 modes, shows ratios 0.94–1.00. The text should reconcile the full-mode range with the table and state that the reported degradation occurs for modes 11–20.","section":"Figure 4 and Table 2"},{"comment":"The HCB-2 row is listed as a shared cost 'required for ROM', but HCB-2 is not needed to compute the CB estimator in Eq. (42) with n=1; it is only needed for the HCB-1 estimator of Eq. (50). The cost attribution in Table 7 would be clearer if the HCB-2 solve were separated according to which estimator it supports.","section":"Section 4, Table 7"},{"comment":"The sentence 'providing a slight underestimate (lower bound) of the true HCB-1 error' is only proved for the case ξ_HCB-2,i ≥ 0, which follows from the min-max ordering; it would be helpful to state explicitly that the lower-bound property relies on the ordering in Eq. (44) and on the same mode correspondence being maintained between HCB-1 and HCB-2.","section":"Section 3.3, Eq. (51)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a genuine contribution, and the paper is more honest than its abstract. The genuinely new piece is the HCB-1 estimator (Eq. 50), which just says the relative eigenvalue error of HCB-1 is approximated by the relative gap between HCB-1 and HCB-2 eigenvalues. Because the HCB subspaces are nested and SEREP preserves the eigenvalues of the augmented problem, the gap is a guaranteed non-negative lower bound. Cheap, simple, and useful for automated mode selection. The derivations in Appendices B–D are consistent, and the reported ratios (0.85–0.99 for HCB-1 across three quite different geometries) match what the theory predicts. That part deserves a serious referee.\n\nThe CB estimator (Eq. 42) is an extension of Ref. [23] to arbitrary HCB order rather than a wholly new object. Its derivation in Appendix D passes through the leading-order approximation u0^T M_p u0 ≈ 1, and the paper itself reports the breakdown: plate modes 11–20 give ratios as low as 0.76. Remark 2 proposes an exact-denominator variant that removes that approximation, but says validation is left for future work. The problem is that the abstract and conclusions claim the estimators are validated without this caveat being quantified. In the regime where an error certificate matters most—large CB errors—the leading-order estimator is least reliable. That is a real soft spot, though not a fatal one: the limitation is acknowledged, and the fix is explicitly identified.\n\nThe other soft spot is reproducibility. The tables look consistent, but the Data Availability statement says 'on request.' For a numerical methods paper, that is not enough. A referee should ask for the mesh/geometry files and, better, the code that produces Tables 2, 4, and 6.\n\nOne minor point: the mode correspondence coefficient c_i in Eq. (45) is defined assuming a one-to-one match between HCB-1 and HCB-2 modes. The authors mention caution for repeated modes, which is fair, but it would be good to see how the estimator behaves when the MAC diagonal drops below, say, 0.9.\n\nWho is this for? Practitioners using HCB in substructuring, especially digital-twin workflows where cheap error estimation drives adaptive mode selection. The math is not deep—this is applied linear algebra—but it is clean and the examples are convincing.\n\nMy recommendation: send it to peer review. It is not a desk reject. The referee should push for validation of the exact-denominator variant and for the data/code, but the central idea and the HCB-1 estimator are solid.","headline":"New HCB-1 eigenvalue error estimator is cheap and works; CB estimator extension is solid but its validation is narrower than the abstract claims, so worth refereeing with requests for exact-denominator validation and data.","tokens_in":23538,"tokens_out":2347,"would_cite":true,"duration_ms":24247,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F15","65N55","74H45"],"pacs":[],"model":"deepseek-v4-flash","headline":"For Craig–Bampton and higher-order Craig–Bampton reduced models, eigenvalue errors can be estimated a priori from reduced-order eigenpairs and already-assembled full matrices, with no full-order solve.","keywords":["component mode synthesis","Craig-Bampton method","higher-order Craig-Bampton method","a priori error estimator","eigenvalue error","residual flexibility","SEREP","reduced-order modeling"],"falsifier":"Compute the true relative errors for a substructuring configuration where the highest tracked eigenvalue is deliberately placed near the smallest discarded substructure frequency (for example, retaining only five normal modes per component in the reactor pressure vessel model); the estimator-to-truth ratios should drop markedly below 0.85 for the HCB-1 estimator and below 0.76 for the CB estimator on those modes, directly falsifying the claimed validity range.","tokens_in":22351,"feed_emoji":"🧮","tokens_out":8496,"duration_ms":81742,"temperature":0.7,"pith_summary":"Model reduction for structural dynamics has a blind spot: to know how far the reduced model is from the full-order truth, you are usually forced to solve the full-order problem, which defeats the purpose of reduction. This paper takes on that problem for the Craig–Bampton (CB) method and its higher-order Craig–Bampton (HCB) extensions, and it derives estimators that return the relative eigenvalue error for each mode without ever solving the full-order eigenproblem. The first estimator generalizes an earlier CB error formula by using an HCB eigenpair as the reference; the second estimates the HCB-1 error directly as the relative gap between the HCB-1 and HCB-2 eigenvalues. Because all HCB models are compressed to the same size as CB by SEREP, the extra cost is almost nothing. If the paper is right, engineers can certify reduced models, choose basis sizes, and keep digital twins up to date without full-order solves.","feed_headline":"Two formulas predict reduced-model errors without a full solve","feed_subtitle":"Uses the nested higher-order Craig-Bampton hierarchy to certify reduced-order accuracy at near-zero extra cost.","key_machinery":"The engine is the nested Ritz subspace hierarchy $\\mathcal{R}(T_{\\mathrm{CB}}) \\subset \\mathcal{R}(T_{\\mathrm{HCB-1}}) \\subset \\mathcal{R}(T_{\\mathrm{HCB-2}})$, built by augmenting the CB basis with Neumann-series residual modes and then compressing back to CB dimension with SEREP. The auxiliary matrix $G(\\mu)=M_p-\\mu^{-1}K_p$ converts the Rayleigh-quotient perturbation terms into products of already-assembled full-order matrices, so the estimator needs only reduced eigenpairs plus matrices that exist from the CMS setup. The HCB-1 estimator's derivation relies on a rank-1 mass-orthogonal projection of each HCB-2 eigenvector onto the corresponding HCB-1 eigenvector, with mode-correspondence coefficient $c_i\\approx 1$; this makes all stiffness cross-terms cancel, leaving the estimator as a pure eigenvalue difference. Courant–Fischer min-max ordering then guarantees that difference is non-negative, so the HCB-1 estimate is a true lower bound on the actual relative error.","core_discovery":"The paper's central claim is that the relative eigenvalue error of a Craig–Bampton reduced model can be predicted by the formula $\\hat{\\eta}^{(n)}_{\\mathrm{CB},i} = 2\\,\\bar{\\phi}_i^T T_0^T G(\\bar\\lambda_i) T_r \\bar\\phi_i + \\bar\\phi_i^T T_r^T G(\\bar\\lambda_i) T_r \\bar\\phi_i$, where $G(\\mu)=M_p-\\mu^{-1}K_p$ is built from the already-assembled full-order matrices, $T_0$ and $T_r$ split the HCB transformation into its CB-like base and its residual correction, and $(\\bar\\lambda_i,\\bar\\phi_i)$ is the $i$-th HCB-$n$ eigenpair after SEREP compression. For the HCB-1 model itself, it claims that the relative eigenvalue error is estimated by $\\hat{\\eta}^{(1)}_{\\mathrm{HCB},i} = (\\lambda^{(1)}_{\\mathrm{HCB},i}-\\lambda^{(2)}_{\\mathrm{HCB},i})/\\lambda^{(2)}_{\\mathrm{HCB},i}$, a non-negative lower bound that follows from the nested Ritz subspaces and a rank-1 mass-orthogonal projection that cancels all stiffness cross-terms. In validation runs on a cantilever plate, a 90-degree elbow pipe, and a reactor pressure vessel, the CB estimator tracks true errors with ratios mostly between 0.76 and 1.06, while the HCB-1 estimator stays between 0.85 and 0.99, and the full-order eigenproblem is never solved.","pith_inferences":["Inference: the HCB-1 estimator's shortfall is controlled by the ratio $\\xi_{\\mathrm{HCB-2}}/\\hat{\\eta}$ through Eq. (53), so iterating the hierarchy one more step could yield a corrected, tighter a priori bound without ever solving the full-order problem.","Inference: the same hierarchical eigenvalue-difference logic should transfer to other nested projection-based reduced models—dual Craig–Bampton, automated multilevel substructuring, or enriched POD bases—wherever Courant–Fischer min-max ordering holds.","Inference: the mode-correspondence coefficient $c_i$ could serve as a cheap reliability indicator, warning users when closely spaced or repeated eigenvalues break the one-to-one mode mapping before the estimator is trusted.","Inference: the CB estimator's $O(N_b^2)$ dense cross-product cost points to interface reduction as a natural companion; combining the two would make the estimator affordable for industrial models with very large boundary DOF counts."],"forward_implications":["Users of CB and HCB reduced models can estimate per-mode relative eigenvalue errors from data they already have—the reduced eigenpairs and assembled full matrices—so no full-order solve is needed for certification.","Because HCB-1 and HCB-2 are compressed to the same dimension as CB, the extra cost of the HCB-1 estimator is just the mode-correspondence coefficients plus a few dense products, negligible compared with the reduced solves.","The HCB-1 estimator is a consistent slight underestimate (lower bound) guaranteed by the monotone eigenvalue ordering, so reported errors will not silently flip sign.","For high modes where CB errors are large, using a higher-order HCB reference (HCB-2 or beyond) should restore estimation accuracy, since the leading-order denominator approximation $u_0^T M_p u_0 \\approx 1$ is what degrades.","The framework turns the HCB hierarchy into a practical certification and basis-sizing tool, which the paper identifies as a promising route to automated adaptive mode selection for digital-twin applications."],"supporting_citations":[{"why":"Defines the Craig–Bampton reduced basis whose eigenvalue errors are the object of estimation.","marker":"[11]"},{"why":"Introduces the HCB residual-mode hierarchy and Neumann-series expansion that supply the higher-order reference solutions the estimators rely on.","marker":"[15]"},{"why":"SEREP reduction compresses every HCB model back to CB dimension, making the estimates cheap and directly comparable across orders.","marker":"[16]"},{"why":"Prior a priori CB error estimator that this paper generalizes through its base/residual transformation split.","marker":"[23]"},{"why":"Residual-based a posteriori alternative requiring full-order quantities, which highlights the a priori nature of the new estimators.","marker":"[21]"},{"why":"Courant–Fischer min-max principle used to guarantee the monotone eigenvalue ordering and the non-negative lower-bound property of the HCB-1 estimator.","marker":"[27–29]"},{"why":"Modal Assurance Criterion used in the validation examples to establish one-to-one mode correspondence between reduced and full-order modes.","marker":"[30]"}],"fun_headline_variants":["A priori error formulas for Craig–Bampton, no full solve","Hierarchical HCB error estimators predict eigenvalue errors","Nested HCB hierarchy yields error estimators with no full solve","Two a priori formulas bound eigenvalue errors in HCB models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the target mode sitting comfortably below the lowest frequency of the modes discarded in each substructure—only then does the Neumann-series expansion converge fast enough that HCB-2 is a much better model than HCB-1, and the estimates' accuracy collapses as this gap narrows.","fun_headline_variants_meta":{"raw":{"variants":["A priori error formulas for Craig–Bampton, no full solve","Hierarchical HCB error estimators predict eigenvalue errors","Nested HCB hierarchy yields error estimators with no full solve","Two a priori formulas bound eigenvalue errors in HCB models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001147,"raw_usage":{"total_tokens":4853,"prompt_tokens":1135,"completion_tokens":3718,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":751,"completion_tokens_details":{"reasoning_tokens":3650}},"tokens_in":751,"tokens_out":3718,"duration_ms":31104,"temperature":1.0,"reasoning_tokens":3650,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T05:12:59.306659+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the true relative errors for a substructuring configuration where the highest tracked eigenvalue is deliberately placed near the smallest discarded substructure frequency (for example, retaining only five normal modes per component in the reactor pressure vessel model); the estimator-to-truth ratios should drop markedly below 0.85 for the HCB-1 estimator and below 0.76 for the CB estimator on those modes, directly falsifying the claimed validity range.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Craig–Bampton reduced basis whose eigenvalue errors are the object of estimation."},{"cited_title":"Kim, S.-H","cited_arxiv_id":null,"evidence_quote":"Introduces the HCB residual-mode hierarchy and Neumann-series expansion that supply the higher-order reference solutions the estimators rely on."},{"cited_title":"O’Callahan, P","cited_arxiv_id":null,"evidence_quote":"SEREP reduction compresses every HCB model back to CB dimension, making the estimates cheap and directly comparable across orders."},{"cited_title":"Kim, K.-H","cited_arxiv_id":null,"evidence_quote":"Prior a priori CB error estimator that this paper generalizes through its base/residual transformation split."},{"cited_title":"Jakobsson, M","cited_arxiv_id":null,"evidence_quote":"Residual-based a posteriori alternative requiring full-order quantities, which highlights the a priori nature of the new estimators."},{"cited_title":"Pastor, M","cited_arxiv_id":null,"evidence_quote":"Modal Assurance Criterion used in the validation examples to establish one-to-one mode correspondence between reduced and full-order modes."}],"review_version":1}