{"id":"c94c238e-5687-474a-b334-b6b32cba1113","arxiv_id":"2607.13888","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For every deterministic fixed-visibility multiscale finite-element construction with bounded support, bounded coefficient-information radius, and bounded local multiplicity, a smooth periodic coefficient and smooth right-hand side exist for which the normalized Galerkin error is bounded below by a p","lead":"Multiscale finite-element constructions that keep their local support and coefficient 'visibility' bounded cannot converge uniformly on rough-coefficient elliptic problems: the worst-case error stays bounded below by a positive constant as the mesh shrinks. The proof shows that at least one local parameter—support radius, coefficient-information radius, or local basis multiplicity—must grow to get uniform optimal accuracy.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified.","rationale":"The reader's verdict is ACCEPT with the weakest assumption identified as fixed visibility. My stress-test pass did not surface a load-bearing technical concern. The main theorem is an impossibility result inside an explicitly delimited information model. The fixed-visibility condition is genuinely load-bearing—without it, Lemma 3.1 fails and the lower-bound construction collapses—but the paper flags this boundary in §8.2 and §9 and does not overclaim. The internal proof steps are coherent: Lemma 3.1's patch nesting is correct; Prop. 4.4's compactness argument is protected by the inradius lower bound; Lemma 5.1's positive-density count does not require mesh alignment; and Prop. 7.1's transfer from cell correctors to global Galerkin error is carried out with the right Lipschitz and orthogonality inequalities. I therefore see no reason to alter the verdict. The concrete test is a lightweight independent verification of the perturbation lemma, since that is the most algebraically delicate step; it is intended as a confirmatory check, not as a response to a specific suspected flaw.","tokens_in":18103,"tokens_out":24293,"duration_ms":237388,"concrete_test":"Independently re-derive Lemma 4.3 for a_t = 1 + t b with b merely nonnegative and not mean-zero, testing Eq. (4.8) with χ_t − t v and using ||∇χ_t||_2 ≤ t||b||_2. Verify that the O(t^2) bound holds with a constant independent of t; this stability estimate is the pivot on which Prop. 4.4's perturbation argument rests.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I read the proof as internally consistent. The only candidate load-bearing point is the fixed-visibility condition in Definition 2.2, Eq. (2.3), which Lemma 3.1 needs to obtain a common local restriction space. This is a deliberate scope restriction, not a hidden flaw: §8.2 and §9 explicitly state that if (2.3) is dropped, the common-local-space reduction is unavailable and the support-only, globally informed problem remains open. Within the model, the step is sound: for T ∈ N_m(K), ω_{m+k}(T) ⊆ ω_{2m+k}(K), so coefficient agreement on the R-layer patch forces equality of the anchored subspaces. I also checked the finite corrector family (Prop. 4.4), the positive-density mesh count (Lemma 5.1), and the corrector-transfer step (Prop. 7.1); each appears valid under the stated hypotheses. The perturbative construction in Lemma 4.3 is stable for nonnegative, non-mean-zero b, and the compactness argument in Prop. 4.4 is protected from degeneration by the inradius lower bound. I found no internal inconsistency or unacknowledged hidden assumption that would threaten Theorem 2.3 within its delimited model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves a minimax lower bound for generalized multiscale finite element methods under a 'fixed-visibility' scalability model. The model requires a deterministic coefficient-to-space rule whose chosen local basis has fixed support radius (in coarse-element layers), fixed anchor multiplicity, and depends on the coefficient only through fixed-size local patches. Theorem 2.3 states that, for any such rule and any quasi-uniform mesh family, there are coefficients in the bounded-contrast class κ∈K_ρ such that the normalized L2-to-energy Galerkin error is at least a positive constant independent of H. The proof constructs a finite family of smooth periodic coefficients and smooth right-hand sides, fixed before H and the rule, and shows via a local dimension bound (Lemma 3.1), a finite corrector family with common local data (Proposition 4.4), a positive-density patch-containment lemma (Lemma 5.1), and strong periodic corrector convergence (Proposition 6.1) that at least one member of this family gives the claimed error for every H<H* and every admissible rule. The authors carefully delimit the scope: the result applies only to fixed-visibility rules; globally informed support-local constructions are left open (§8.2, §9).","tokens_in":18423,"tokens_out":10844,"duration_ms":95696,"significance":"If sustained, this is a significant contribution to the understanding of structural scalability in multiscale finite element methods. It rigorously formalizes a natural information-locality constraint and proves a sharp Θ(1) worst-case error barrier, complementing positive results for LOD, CEM-GMsFEM, and MS-GFEM that rely on growing localization radii or local spectral dimensions. The proof is detailed, largely self-contained, and chains standard ingredients (harmonic-function analyticity, compactness of the simplex family, Allaire's strong corrector theorem) in a coherent way. The paper is careful with quantifier order and explicitly identifies the boundary of its model, including the unresolved support-only globally informed variant. The finite-family construction is an elegant feature: the coefficient profiles and right-hand sides are fixed before H and the rule, and only the index is chosen after seeing the rule. This makes the lower bound robust and non-circular.","major_comments":[],"minor_comments":[{"comment":"The text systematically labels lemmas and propositions as 'Theorem' in cross-references: 'Theorem 3.1' should be Lemma 3.1; 'Theorem 4.2' and 'Theorem 4.3' should be Lemmas 4.2 and 4.3; 'Theorem 5.1' should be Lemma 5.1; 'Theorem 6.1' should be Proposition 6.1; 'Theorem 6.2' should be Lemma 6.2; 'Theorem 7.1' should be Proposition 7.1. These inconsistencies should be corrected before publication.","section":"Throughout"},{"comment":"The opening line 'From Theorem 4.1, we have (4.8)' appears to refer to the cell problem in Definition 4.1 (or Eq. (4.2)), not to a separate theorem. Please rewrite as 'From Definition 4.1' or 'From (4.2)'.","section":"Lemma 4.3"},{"comment":"The phrase 'with parameters (H, m, k, Cdim, Cloc)' includes H as a parameter, but the rule is defined per mesh size H. Consider clarifying in the text that H plays the role of the mesh scale rather than a method parameter; this avoids possible confusion with the fixed structural parameters (m,k,Cdim,Cloc).","section":"Definition 2.2"},{"comment":"In the statement of Proposition 4.4, the dependence list 'd,q,ρ,D,D0,h−,h+,γ' might be read as implying that q is independent of m,k,Cloc; of course q is defined in (3.1) from m and Cloc. A short reminder in the text would be helpful.","section":"Proposition 4.4"}],"recommendation":"minor_revision","confidential_remarks":"This is a solid negative result within a well-defined model. The mathematical argument is internally consistent and the scope restrictions are explicitly acknowledged. The only issues I found are presentation-level (systematic mislabeled cross-references and a few phrasings). After fixing those, the paper is suitable for publication. I see no need for additional technical revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuine negative result, and the proof holds up. For any fixed-visibility GMsFEM-style construction — bounded support radius, bounded coefficient-information radius, bounded local multiplicity — the worst-case normalized Galerkin error stays bounded below by a positive constant as H → 0 over the bounded-contrast measurable coefficient class. That is the first order-one lower bound I know for this class of scalable multiscale constructions, and it does what an impossibility result should: it isolates the feature doing the work (fixed visibility) and does not overclaim.\n\nThe proof is well engineered. Exterior dipoles generate q+1 independent corrector directions inside a common core where all coefficients agree; the perturbation lemma converts those fields into actual corrector gradients with uniform control; the positive-density lemma works for arbitrary quasi-uniform simplicial meshes, not just structured ones; and the transfer to exact solutions via Allaire's strong corrector theorem is legitimate. The quantifier order is clean: the finite family, the right-hand sides, and the constant c* are fixed before H and before the rule; only the index is chosen afterward. Constants are tracked honestly, including the fact that c* deteriorates as ρ → 1.\n\nThe soft spots are mostly scope notes rather than cracks. The fixed-visibility premise is load-bearing — Lemma 3.1 needs it — and the paper says plainly in §8.2 and §9 that the support-only, globally informed problem (fixed support, global coefficient knowledge) remains open. The theorem is also stated for d ≥ 2 only; the 1D case is untouched, and the paper says so in §8.1. A reader who wants an impossibility statement for all locally supported basis constructions has to respect that boundary. The risk is overselling the title by a reader, not the authors overclaiming. One genuinely minor technical gap: Lemma 4.2 chooses the dipole points z_j in Y \\ D, and the mollification step needs positive torus distance from z_j to D. With D open, a point on ∂D would break the step; choosing the points in Y \\ D̄ gives a one-line fix. It does not threaten the argument.\n\nThis paper settles the fixed-visibility case and sharpens the picture: LOD, CEM-GMsFEM, and spectral methods really do need their growing localization radii or dimensions. It deserves a serious referee — I would send it out — and I would engage with it in my own work.","headline":"A genuine order-one lower bound for fixed-visibility GMsFEM-style constructions; the proof is coherent, the scope is honest, one minor fix needed — send it to referees.","tokens_in":18783,"tokens_out":10496,"would_cite":true,"duration_ms":90215,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35B27","65N15","65N30"],"pacs":[],"model":"deepseek-v4-flash","headline":"No deterministic fixed-visibility multiscale finite element construction can converge uniformly over the bounded-contrast coefficient class; the worst-case normalized Galerkin error stays bounded below by a positive constant as the mesh is","keywords":["generalized finite element method","multiscale finite element method","numerical homogenization","rough coefficients","localization","error lower bounds","elliptic PDE","Galerkin method"],"falsifier":"A deterministic fixed-visibility rule with bounded m, k, and Cloc for which the supremum over the bounded-contrast coefficient class of the normalized L2-to-energy Galerkin error tends to zero as H→0 would disprove the theorem. Concretely, in two dimensions one could implement a rule and compute the worst-case error on the finite periodic family constructed in the proof (smooth coefficients a_j(x/(LH)) with right-hand sides f_j); the theorem predicts a positive lower bound as H→0.","tokens_in":18038,"feed_emoji":"📉","tokens_out":3907,"duration_ms":35022,"temperature":0.7,"pith_summary":"The paper asks whether coefficient-adapted multiscale finite element methods can keep their local construction parameters fixed—constant support radius, constant coefficient-information radius, and constant local basis multiplicity—while refining the coarse mesh. It proves an impossibility result: for any deterministic rule satisfying this fixed-visibility condition, there always exists a coefficient in the bounded-contrast measurable class for which the normalized Galerkin error does not tend to zero; it remains bounded below by a positive constant independent of mesh size. This means uniform optimal-order accuracy over the whole class requires at least one local parameter to grow or requires using coefficient information from beyond the fixed local patches. The proof constructs a finite family of smooth periodic coefficients that agree on all relevant local patches yet produce more independent local corrector fields than any fixed-dimensional local space can approximate.","feed_headline":"Fixed-local multiscale FEM fails to converge on rough coefficients","feed_subtitle":"Any rule with bounded local patches hits coefficients it cannot approximate, at any resolution.","key_machinery":"The fixed-visibility consistency condition (Definition 2.2) requires each anchored local subspace to depend on the coefficient only through its restriction to a fixed (m+k)-layer patch. This yields a common local restriction space of dimension at most q for any two coefficients agreeing on the patch (Lemma 3.1). The paper then builds a finite family of periodic coefficients, all equal to 1 on a core region, whose first cell-corrector gradients span q+1 independent directions inside the core; exterior dipole perturbations realize these fields as corrector gradients. A positive-density mesh lemma shows that a fixed fraction of elements have their coefficient-information patches contained in th","core_discovery":"The central claim is Theorem 2.3: for fixed bounds on support radius, coefficient-information radius, and local multiplicity, and for any deterministic fixed-visibility selection rule, there exists a coefficient in the bounded-contrast class such that the normalized L2-to-energy Galerkin error is at least a positive constant c*, independent of the mesh size H. Equivalently, the minimax quantity liminf_{H→0} inf_M sup_κ E_H(κ; V_H^M(κ)) stays above a positive constant. The failure is not merely the loss of an optimal rate; the error does not vanish. The construction uses a finite family of smooth periodic coefficient profiles and smooth compactly supported right-hand sides, chosen before H an","pith_inferences":["The proof suggests that matrix sparsity alone is not the right notion of scalability: sparsity does not encode coefficient-information locality, and the theorem identifies information locality as the restrictive condition.","A plausible next step is to test the support-only conjecture: if global coefficient information is allowed, the common-local-space lemma fails, so new arguments would be needed; the result hints that the answer may depend on how much geometric information can be encoded in basis shapes.","The finite-family construction could serve as a numerical benchmark: for any proposed scalable rule, computing the worst-case error on the periodic family would, by the theorem, yield an order-one lower bound at sufficiently small H.","The theorem leaves randomized rules open, requiring a separate expected-value or high-probability formulation."],"forward_implications":["Any deterministic multiscale finite element construction that keeps support radius, coefficient-information radius, and local multiplicity bounded cannot be uniformly convergent over the full bounded-contrast coefficient class.","Achieving uniform optimal O(H) accuracy necessarily requires at least one local construction parameter to grow as H→0, or requires using coefficient information beyond the prescribed local patches.","The lower bound is worst-case over coefficients but not over right-hand sides: the normalized error remains Θ(1) in the worst case, matching the trivial upper bound from the zero space.","The finite family of coefficients and right-hand sides is chosen before the scale and before the rule, so the lower bound is not an adversarial choice made after seeing the rule.","The result does not settle whether globally informed, locally supported bases can achieve uniform optimal accuracy; that question remains open."],"fun_headline_variants":["Fixed-local multiscale FEM can't converge on rough coefficients","Rough coefficients defeat any fixed-visibility multiscale FEM","Bounded-local multiscale FEM has a constant error floor","Uniform convergence fails for fixed-local multiscale methods","Order-one error floor for scalable multiscale FEM with fixed locality"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof depends on the fixed-visibility assumption that each anchored local basis space is determined by the coefficient only on a fixed (m+k)-layer patch; if a rule can use coefficient values from far away to shape a locally supported basis, the common-local-space argument collapses.","fun_headline_variants_meta":{"raw":{"variants":["Fixed-local multiscale FEM can't converge on rough coefficients","Rough coefficients defeat any fixed-visibility multiscale FEM","Bounded-local multiscale FEM has a constant error floor","Uniform convergence fails for fixed-local multiscale methods","Order-one error floor for scalable multiscale FEM with fixed locality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000569,"raw_usage":{"total_tokens":2569,"prompt_tokens":821,"completion_tokens":1748,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":1663}},"tokens_in":565,"tokens_out":1748,"duration_ms":20908,"temperature":1.0,"reasoning_tokens":1663,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T03:24:53.709657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A deterministic fixed-visibility rule with bounded m, k, and Cloc for which the supremum over the bounded-contrast coefficient class of the normalized L2-to-energy Galerkin error tends to zero as H→0 would disprove the theorem. Concretely, in two dimensions one could implement a rule and compute the worst-case error on the finite periodic family constructed in the proof (smooth coefficients a_j(x/(LH)) with right-hand sides f_j); the theorem predicts a positive lower bound as H→0.","supporting_citations":[],"review_version":1}