{"id":"5532ad7e-8eb2-49b8-bd18-8a47fd2908f2","arxiv_id":"1908.00652","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Petrov-Galerkin localized orthogonal decomposition method with error-indicator-driven recomputation solves perturbed diffusion problems by reusing reference-coefficient computations.","lead":"An adaptive numerical scheme reuses precomputed multiscale basis functions from a reference diffusion problem to solve many perturbed variants cheaply, using error indicators to decide where to recompute. It targets composite materials with defects or small shifts and comes with an a priori error bound.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 (§5.2) does not implement the analyzed method: its marking rule differs from (3.10) and it computes b_T for marked instead of unmarked elements, so the theorem's assumptions need not hold for the algorithm as printed.","rationale":"The central mathematical result, Theorem 4.1, has a plausible proof for the algorithm defined in Definition 3.5, and the numerical experiments are qualitatively consistent with that theorem. However, the paper's own implementable specification in Algorithm 1 contradicts the theorem's hypotheses and cannot even produce a well-defined assembly as written: the marking condition differs from (3.10), and the perturbed element contributions are computed for the wrong set of elements. A theorem about an algorithm whose pseudocode is not the algorithm analyzed does not establish the central claim as written. This is not an attack on the proof; it is an internal-consistency and reproducibility failure that is more concrete than the scope limitation emphasized by the reader. The reader's weakest_assumption concerned efficiency only for localized perturbations, which I do not dispute, but I do not regard that as the weakest point: even in the favorable local-defect case, a reader following Algorithm 1 would implement a different method. The verdict should remain conditional, with the additional required revision that Algorithm 1 be corrected to match Definition 3.5 and that the numerical experiments be tied to the actual marking rule used in the code.","tokens_in":17138,"tokens_out":37299,"duration_ms":367640,"concrete_test":"Run Experiment 6.1 twice from the published code: once with the exact Definition 3.5 / Theorem 4.1 rule (mark T iff E_QVH,T ≤ TOL, E_f,T + E_Rf,T ≤ TOL(||f|| + |||g|||), and E_Qg,T ≤ TOL(||f|| + |||g|||)), and once with Algorithm 1's line 11 condition and the line 13-14 placement. Record T_ref_H and the assembled stiffness-matrix entries in both cases. If the two marking sets differ for any T at any TOL, or if the Algorithm 1 version leaves any b_T entry undefined for T ∉ T_ref_H, then the published algorithm is not the analyzed one. The paper should be revised so that Algorithm 1 matches Definition 3.5, and the numerical evidence should be tied to the rule actually implemented in the code.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 4.1 analyzes the marking rule (3.10) in Definition 3.5, whose conditions are E_QVH,T ≤ TOL, E_f,T + E_Rf,T ≤ TOL(||f|| + |||g|||), and E_Qg,T ≤ TOL(||f|| + |||g|||). The proof bounds the reuse error only over the marked set T_ref_H. Algorithm 1, however, uses in line 11 the condition max(E_QVH,T (||f||_{L2(T)} + |||g|||), E_f,T, E_Rf,T, E_Qg,T) ≤ TOL. This is not equivalent to (3.10): it never sums E_f,T and E_Rf,T, it drops the global norm on the right-hand side, and it inserts a local f-norm into the first term. An element can therefore be marked as 'reference' even when E_f,T + E_Rf,T > TOL(||f|| + |||g|||), so the premise of Theorem 4.1 can fail for the algorithm as printed. Additionally, lines 13-14 are placed inside the same if-block as line 12: for a marked (reference) element the algorithm computes perturbed correctors and the perturbed element matrix b_T, while line 16 assembles b^ref_T for T ∈ T_ref_H and b_T for T ∉ T_ref_H. Thus b_T is computed for exactly the elements where it is not used, and is never computed for the elements where assembly needs it. As written, Algorithm 1 cannot produce the numerical results and does not implement Definition 3.5. The theorem itself may be correct, and the GitHub code may implement the correct rule, but the paper's central claimed method is internally inconsistent.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an adaptive Petrov--Galerkin localized orthogonal decomposition (PG-LOD) method for elliptic diffusion problems whose coefficient is a perturbation of a reference coefficient. The perturbations considered are local defects and global or local domain mappings, possibly combined. The method precomputes reference correctors and element stiffness contributions, computes four local error indicators for each coarse element, and then decides adaptively which elements need recomputation of correctors using the perturbed data. The central theoretical result, Theorem 4.1, states an error bound of the form |||u - \\tilde u_k||| ≲ k^{d/2}(θ^k + TOL)(||f||_{L^2(Ω)} + |||g|||) under the marking conditions in Definition 3.5. Numerical experiments for random defects, local domain mappings, and a global domain mapping are reported, together with reproducibility information for the implementation. The main technical framework is standard LOD machinery combined with perturbation indicators, and the paper is generally well organized; however, the printed Algorithm 1 does not implement the analyzed method, and the proof of Theorem 4.1 contains a flawed norm estimate that needs correction.","tokens_in":17521,"tokens_out":14806,"duration_ms":146481,"significance":"If the implementation and proof are corrected, the paper would be a useful contribution to multiscale simulation for parametric or perturbed coefficients. The idea of reusing reference-coefficient correctors and element matrices, with computable indicators that trigger local recomputation, is natural and practically relevant. The paper gives explicit indicator definitions, a parameter-free (up to standard LOD constants) theoretical bound, and numerical evidence for three perturbation regimes. The claim that local defects and mild domain mappings can be handled without recomputing the complete multiscale basis is credible and of interest to the LOD community. Credit is due for shipping reproducible code and for reporting the expected degradation in the global-mapping example rather than overclaiming efficiency. The main reservation is not the novelty but the consistency between the analysis and the algorithm as printed.","major_comments":[{"comment":"As printed, Algorithm 1 does not implement the analyzed method. The marking condition in line 11 is max(E_QVH,T (||f||_{L2(T)} + |||g|||), E_f,T, E_Rf,T, E_Qg,T) ≤ TOL, which differs from the marking criterion (3.10) in several load-bearing ways: it does not sum E_f,T and E_Rf,T, it uses the local norm ||f||_{L2(T)} instead of the global norm ||f||_{L2(Ω)}, and it does not scale the right-hand side by ||f||_{L2(Ω)} + |||g|||. Therefore an element can be marked for reference reuse even when E_f,T + E_Rf,T > TOL(||f||_{L2(Ω)} + |||g|||), so the hypothesis of Theorem 4.1 need not hold for the algorithm. Moreover, lines 13–14 compute the perturbed correctors and the perturbed element matrix b_T for the elements T that have just been added to T_ref^H, while line 16 assembles b^ref_T for T ∈ T_ref^H and b_T for T ∉ T_ref^H. Thus b_T is computed exactly for the elements where it is not used and is never computed for the elements where the assembly requires it. As written, the algorithm cannot assemble the matrix in line 16. This inconsistency must be repaired, for example by moving the computation of b_T into an else-branch for T ∉ T_ref^H and by making the marking rule identical to (3.10).","section":"Section 5.2, Algorithm 1 (lines 11–16)"},{"comment":"The displayed estimate for |||u_k - \\tilde u_k||| in the proof of Theorem 4.1 contains sign errors. A norm satisfies the triangle inequality, so the terms -|||Q_k g - \\tilde Q_k g||| (and the analogous negative signs in the following line) should be plus signs. In addition, bounding |||Q_k u^H_k - \\tilde Q_k \\tilde u^H_k||| requires inserting ±\\tilde Q_k u^H_k and then using the boundedness of \\tilde Q_k that is established just before (4.3); as written, the second line of (4.3) omits this step. Since (4.3) is the estimate that connects the reuse error to the coarse-scale error u^H_k - \\tilde u^H_k, the proof of Theorem 4.1 needs this corrected calculation before the main bound can be considered established.","section":"Section 4, Eq. (4.3)"}],"minor_comments":[{"comment":"The hypothesis of Theorem 4.1 is written in a different normalization from the marking criterion (3.10): the first term is multiplied by ||f||_{L2(Ω)} + |||g||| while the other two are not. The two statements are equivalent for nonzero data, but unifying the notation would prevent exactly the kind of mismatch that occurs in Algorithm 1.","section":"Section 4, Theorem 4.1 vs. Definition 3.5"},{"comment":"The text says that the global domain mapping experiment would surely result in 100% recomputation without domain mappings; this is a heuristic observation. Consider stating it as a comparison of indicator values rather than as a categorical claim, since the computational cost of the alternative is not measured in the experiment.","section":"Section 2.1 and Section 6.3"},{"comment":"References [13] and [14] are the same paper and should be merged or one removed.","section":"References"},{"comment":"There are several typos, including 'communcation', 'mapppings', and 'performed'; the abstract and introduction should be proofread.","section":"Abstract and Introduction"},{"comment":"The indentation and block structure of Algorithm 1 is ambiguous: it should be made explicit whether the assembly in line 16 is inside or outside the loop over elements, and which 'end' closes the for loop.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The inconsistency between the pseudocode and the analyzed method is serious but appears fixable within the scope of the paper. The proof issue in Eq. (4.3) is also local. I would recommend a major revision rather than rejection, provided the authors align Algorithm 1 with Definition 3.5 and correct the proof estimate. No concerns about attribution or scope beyond the points stated in the report."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this paper's core idea is good and the analysis is mostly sound, but Algorithm 1 in Section 5.2 does not implement the method that Theorem 4.1 analyzes. As printed, it just doesn't work. That is the thing to know before anything else.\n\nThe actual contribution is extending the 'similar coefficients' PG-LOD setup from Hellman–Målqvist to two perturbation classes—local defects and domain mappings—with computable error indicators for the coefficient, RHS, and boundary correctors, plus an a priori bound for the adaptive reuse strategy. The indicators are nontrivial, and the paper explains how to compute them cheaply (precomputed eigenvalue data, ν_T estimates). The experiments are honest, include the global-mapping case where the method degrades, and the code is on GitHub. That is real work and worth building on.\n\nSoft spots:\n\n1. Algorithm 1 is wrong as written. The marking rule on line 11 (max of four quantities, local f-norm, no sum of E_f and E_Rf) is not equivalent to the marking criteria (3.10) in Definition 3.5, so the theorem's premise need not hold for the algorithm. Worse, lines 13–14 compute the perturbed correctors and b_T inside the same if-block that puts T into T_ref_H—the set for which the method should be using reference correctors. Then line 16 assembles b_T for T ∉ T_ref_H, but b_T was never computed for those elements. This isn't stylistic; the algorithm as printed cannot produce any of the numerical results. The conceptual method in Definition 3.5 is coherent, and the GitHub code presumably implements it, but the paper must be revised so the pseudocode and experiments agree.\n\n2. Smaller items. Equation (4.3) has a minus sign in front of the Qg term that should be a plus (triangle inequality). The statement of Theorem 4.1 mixes scaled and unscaled quantities in the max; it is mathematically equivalent to (3.10), but confusing and should be cleaned up. The paper claims low memory and parallel efficiency but reports no runtime or memory measurements. The ν_T constant used in the RHS indicator is calibrated numerically (ν_T ≈ 0.25H), not derived—acceptable if stated, but it is a heuristic, not a guaranteed bound.\n\nOverall: the central error bound is plausible and the experiments support it, modulo the algorithm bug. This deserves a serious referee but with a clear expectation of major revision. I would want the pseudocode fixed and the efficiency claim backed by measurements before accepting.","headline":"Useful extension of PG-LOD for perturbed coefficients with a sound core analysis, but the printed algorithm does not match the analyzed method and as written cannot reproduce the experiments—fixable, but essential.","tokens_in":18043,"tokens_out":4160,"would_cite":true,"duration_ms":38367,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65N15","65N12","35B27"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes that perturbed diffusion problems can be solved by reusing reference-coefficient computations wherever local error indicators are small, with a global error bound that preserves the multiscale convergence order.","keywords":["multiscale method","localized orthogonal decomposition","Petrov-Galerkin","perturbed diffusion coefficient","local defects","domain mapping","error indicators","numerical upscaling"],"falsifier":"Refine the coarse mesh and compute, on an element next to a fixed defect, the true local corrector error $|||Q_{k,T}v-Q^{\\mathrm{ref}}_{k,T}v|||$ alongside the indicator $E_{QV_H,T}$; if their ratio is unbounded as $H\\to0$, Lemma 4.2 fails and Theorem 4.1 collapses. A reader could run this local test because the paper reports only the global relative error, not the indicator-to-error ratio.","tokens_in":16940,"feed_emoji":"🔄","tokens_out":14069,"duration_ms":130065,"temperature":0.7,"pith_summary":"The paper addresses a practical bottleneck in multiscale simulation: many similar diffusion problems, each with a rapidly varying coefficient that is a perturbation of one reference coefficient. It proposes an adaptive Petrov--Galerkin localized orthogonal decomposition (PG-LOD) method in which the fine-scale basis corrections (correctors) are computed once for the reference coefficient, and for each perturbed problem only the correctors on coarse elements where local error indicators exceed a tolerance are recomputed. The central result, Theorem 4.1, states that if the marking criterion holds, the energy error satisfies $|||u-\\tilde u_k||| \\lesssim k^{d/2}(\\theta^k+\\mathrm{TOL})(\\|f\\|_{L^2(\\Omega)}+|||g|||)$, so the reuse does not destroy the multiscale convergence order. Two perturbation classes are covered: local material defects and coordinate mappings of the reference coefficient. If true, this makes it feasible to simulate many perturbed configurations, such as manufacturing variations, at a fraction of the cost of solving each problem from scratch.","feed_headline":"Reuse cuts cost of perturbed diffusion solves","feed_subtitle":"Local error indicators decide which fine-scale basis corrections to recompute, preserving accuracy without full re-solves.","key_machinery":"The load-bearing object is the localized corrector pair $(Q_{k,T},R_{k,T})$ of the PG-LOD method: for each coarse element $T$, these solve fine-scale patch problems with the perturbed coefficient, and the corresponding reference objects $(Q^{\\mathrm{ref}}_{k,T},R^{\\mathrm{ref}}_{k,T})$ are precomputed once. The mechanism that makes reuse safe is the set of error indicators (3.9), written in terms of $\\delta=A^{-1/2}(A-A_{\\mathrm{ref}})A_{\\mathrm{ref}}^{-1/2}$ and precomputable coarse quantities such as the eigenvalue ratios $\\tilde\\mu_{T,T'}$ of (5.1). Lemma 4.2 shows each indicator controls the true difference between perturbed and reference correctors, and Theorem 4.1 converts the summed indicator errors, together with the exponential localization error $k^{d/2}\\theta^k$, into the global bound. The domain-mapping representation $A=\\det(J)J^{-1}(A_{\\mathrm{ref}}-D)J^{-T}$ is what turns a geometric shift of the material into a value perturbation of the reference coefficient, which is the form the indicators can see.","core_discovery":"The authors claim that reference-coefficient computations can be recycled safely in an adaptive multiscale solver. In precise terms, Theorem 4.1 states that when the error indicators of (3.10) are below a tolerance with $\\mathrm{TOL}=\\tau k^{-d/2}$, the solution $\\tilde u_k$ produced by mixing reference and perturbed correctors obeys $|||u-\\tilde u_k||| \\lesssim k^{d/2}(\\theta^k+\\mathrm{TOL})(\\|f\\|_{L^2(\\Omega)}+|||g|||)$, with constants independent of $H$, $k$, $\\tau$, and $\\mathrm{TOL}$. The proof bounds each local corrector error by a computable indicator (Lemma 4.2) and then shows that the mixed bilinear form remains coercive, so the adaptive update contributes only a term proportional to the tolerance. The numerical experiments show that local defects are detected sharply by the indicators and that the error decreases quickly with the percentage of updated correctors; a global domain mapping is the hard case, where many elements are marked. The intended conclusion is that for the right perturbation classes, the method achieves near full-recomputation accuracy while updating only a small fraction of the local problems.","pith_inferences":["Editorial extension: the same indicator-driven reuse could be applied to sequences of coefficients that drift slowly in time, with the reference stage refreshed only when the marked fraction grows, yielding per-time-step cost proportional to the amount of coefficient change instead of the full problem.","Editorial extension: the eigenvalue precomputation of $\\tilde\\mu_{T,T'}$ suggests a practical screening tool, where after a manufactured sample is measured only the per-element coefficient deviation $\\|\\delta\\|_{L^\\infty}$ is needed to decide which regions require new simulations.","Editorial extension: perturbations that change the topology of the high-conductivity phase are not naturally captured by the defect-plus-mapping form, and the indicators would likely mark a large fraction of elements; representing such perturbations would require a different parametrization or periodic reference updates."],"forward_implications":["If the marking criterion (3.10) is met, the adaptive PG-LOD solution has the same order of accuracy as a full recomputation up to the TOL term, so the reuse is theoretically safe rather than heuristic.","For local defects, the indicators concentrate around the defect and decay exponentially away from it, so most coarse elements keep their reference correctors and the cost per perturbed solve is dominated by a coarse-scale system plus a few local patches.","The reference stage stores only coarse-scale data of size $k^d H^{-d}$ and can be parallelized over elements; each perturbed solve is then independent, so many perturbations can be processed in parallel with low communication.","Global domain mappings are the limiting case: the marked fraction grows, and the reuse advantage shrinks, although the method still converges; for a completely changed coefficient without a mapping, the approach reduces to recomputation.","Periodic reference structures allow corrector reuse across periodic patches, further reducing memory and complexity."],"supporting_citations":[{"why":"Establishes the exponential decay of fine-scale correctors and the localized patch construction on which the reuse strategy relies.","marker":"[20]"},{"why":"Introduces the Petrov-Galerkin formulation of LOD that the paper adapts to mixed reference and perturbed correctors.","marker":"[5]"},{"why":"Develops the similar-coefficient sequence viewpoint and supplies the localization error bound (3.8) used in Theorem 4.1.","marker":"[12]"},{"why":"Introduces the right-hand-side and boundary-data correction operators that the paper reuses.","marker":"[13]"},{"why":"Supplies the stability estimate for the interpolation operator that defines the fine-scale space.","marker":"[23]"}],"fun_headline_variants":["Reuse reference solves for perturbed diffusion","Adaptive corrector reuse for diffusion problems","Diffusion perturbations: reuse beats recompute","Selective updates for perturbed diffusion upscaling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The savings rest on the premise that a large fraction of coarse elements can be marked safe by the local error indicators; if the perturbation is large everywhere or does not fit the defect-plus-coordinate-change representation, nearly every element must be recomputed and the method degenerates to full recomputation.","fun_headline_variants_meta":{"raw":{"variants":["Reuse reference solves for perturbed diffusion","Adaptive corrector reuse for diffusion problems","Diffusion perturbations: reuse beats recompute","Selective updates for perturbed diffusion upscaling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1338,"prompt_tokens":903,"completion_tokens":435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":378}},"tokens_in":519,"tokens_out":435,"duration_ms":5085,"temperature":1.0,"reasoning_tokens":378,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:41:47.651806+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Refine the coarse mesh and compute, on an element next to a fixed defect, the true local corrector error $|||Q_{k,T}v-Q^{\\mathrm{ref}}_{k,T}v|||$ alongside the indicator $E_{QV_H,T}$; if their ratio is unbounded as $H\\to0$, Lemma 4.2 fails and Theorem 4.1 collapses. A reader could run this local test because the paper reports only the global relative error, not the indicator-to-error ratio.","supporting_citations":[{"cited_title":"Localization of elliptic multiscale problems","cited_arxiv_id":null,"evidence_quote":"Establishes the exponential decay of fine-scale correctors and the localized patch construction on which the reuse strategy relies."},{"cited_title":"On multiscale methods in petrov– galerkin formulation","cited_arxiv_id":null,"evidence_quote":"Introduces the Petrov-Galerkin formulation of LOD that the paper adapts to mixed reference and perturbed correctors."},{"cited_title":"Numerical homogenization of elliptic pdes with similar coefficients","cited_arxiv_id":null,"evidence_quote":"Develops the similar-coefficient sequence viewpoint and supplies the localization error bound (3.8) used in Theorem 4.1."},{"cited_title":"Variational multiscale stabilization and the exponential decay of fine- scale correctors","cited_arxiv_id":null,"evidence_quote":"Supplies the stability estimate for the interpolation operator that defines the fine-scale space."}],"review_version":1}