{"id":"297ee950-712a-4b62-8ebc-850f4b57d03c","arxiv_id":"2509.09410","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"For real analytic, multiply periodic coefficients, multiscale elliptic homogenization converges at rate epsilon1 + max_i e^{-c epsilon_i/epsilon_{i+1}}, and the exponential part is shown optimal up to the constant c.","lead":"This paper proves that, for elliptic equations whose coefficients are real analytic and oscillate at several microscopic scales, the homogenization error can be made exponentially small in the ratios between successive scales, rather than only polynomially small as in prior work. The result improves the accuracy of effective equations for hierarchical composite materials and introduces a genuinely multiscale method.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Induction in §5.2 assumes block-homogenized coefficients inherit real analyticity (1.7); existing Theorem 3.9 only supplies Lipschitz dependence in the large-scale parameter, so removal of scale separation for n≥3 is not justified.","rationale":"The paper is substantial and the core simultaneous-homogenization argument appears internally coherent: the multiscale ansatz, energy estimates, and flux correctors are developed in detail, and under the explicit scale-separation condition (3.95) the convergence-rate proof is largely complete. The central claim of Theorem 1.1, however, is the removal of that condition, and the proof of the removal rests on an induction whose inductive hypothesis is not verified. The block-homogenized coefficient after simultaneous homogenization of several scales is not shown to satisfy the real-analyticity condition (1.7); the paper asserts this in one sentence. The available parameter-dependent estimates, Theorem 3.9, are only Lipschitz in the parameter and do not control second and higher derivatives. This is a genuine gap for n≥3, not just a matter of exposition, because without analyticity the inductive theorem for n−m scales cannot be applied. The reader's weakest_assumption identified exactly this issue. A secondary concern is that the optimality counterexample in §7.2 uses coefficients whose effective lower-bound exponent is not related quantitatively to the exponent c in Theorem 1.1, so the optimality claim is less rigorous than the upper bound; however, the main theorem's validity does not depend on that example. Given the substantial original work and the plausibly fillable nature of the analyticity-inheritance gap, the appropriate status remains CONDITIONAL rather than full acceptance or rejection.","tokens_in":60561,"tokens_out":11084,"duration_ms":131768,"concrete_test":"Set n=3, m=2. Write \\mathcal A(y1)=⟨A+A\\hat\\nabla_2 X⟩_{y2,y3}, where X solves the two-scale corrector equation (3.101) with parameter x=y1. Prove, using the recursive equations (3.14)–(3.24) at a fixed ratio ε3/ε2, uniform bounds for ∂_{y1}^ℓ X for ℓ=0,1,2 with constants depending only on original characters. If ∥∂_{y1}^2 \\mathcal A∥_{L∞} cannot be bounded by C Λ^2 2!, then the induction hypothesis in §5.2 is missing higher x-derivative assumptions. Alternatively, formulate a generalized induction hypothesis allowing only Lipschitz dependence in x and check whether the error term ε_{n-m+1}/ε_{n-m} absorbs the resulting constants.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing point is the induction in §5.2. To remove scale separation, the proof homogenizes the smallest m scales simultaneously, obtaining a coefficient A_{ε'}(x)=\\mathcal A(x/ε1,...,x/ε_{n-m}) for the remaining large scales, and then invokes the inductive theorem for n−m scales. For that induction to apply, \\mathcal A(y_1,...,y_{n-m}) must itself satisfy the real-analyticity assumption (1.7). The paper only writes 'since A(y_1,...,y_{n-m}) satisfies the assumptions (1.5)-(1.7)' (Proof of Theorem 1.1, §5.2) and gives no proof. The estimates available for the multiscale correctors with a parameter are not sufficient: Theorem 3.9 establishes Lipschitz dependence in the large-scale parameter x (with constant L0), not the full hierarchy of derivative bounds required by (1.7), and the analyticity proved for bA in Theorem 3.6/3.4 covers only the coefficient obtained by homogenizing the single last scale, not the simultaneous m-scale effective matrix. Thus, for n≥3 and m<n, the induction step is not justified as written. This is not an internal inconsistency in the two-scale case (where separation can be assumed with loss of generality), but it is exactly the step that supports the no-scale-separation version of Theorem 1.1.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies multiscale elliptic homogenization for coefficients A(x/ε_1,...,x/ε_n) that are periodic and real-analytic in all scales. It introduces regularized multiscale correctors and flux correctors, proves uniform estimates under a quantitative scale-separation condition, and then attempts to remove that condition by an induction that simultaneously homogenizes blocks of well-separated scales. The main theorem (Theorem 1.1) claims an L^2 convergence rate of order ε_1 + max_i e^{-c ε_i/ε_{i+1}}, with a constant effective matrix depending on the ratios ε_i/ε_1. A uniform Lipschitz estimate under a double-log separation condition and several counterexamples are also given.","tokens_in":60816,"tokens_out":9401,"duration_ms":100249,"significance":"If the proof is completed, the claimed exponential improvement over the classical reiterated-homogenization rate would be a substantial advance, and the detailed corrector/flux-corrector machinery is an interesting contribution in itself. The energy estimates are explicit, and the counterexamples in Section 7 are concrete and testable. However, the step that removes scale separation for n≥3 rests on an unproved inheritance of real analyticity by the block-homogenized coefficient matrix. Since the central new exponent e^{-c ε_i/ε_{i+1}} is precisely a consequence of analyticity, this gap is load-bearing and must be repaired before the main theorem is established in full generality.","major_comments":[{"comment":"The induction step requires the block-homogenized matrix A(y_1,...,y_{n-m}) to satisfy the real-analyticity assumption (1.7). The paper asserts this in the sentence 'since A(y_1,...,y_{n-m}) satisfies the assumptions (1.5)-(1.7)' but gives no proof. The only regularity established for such a matrix is ∥A∥∞≤C and |∇_x A|≤CL_0, from (5.2) and Theorem 3.9; Theorem 3.9 controls X and ∇_x X in L∞/L^2, but not the higher analytic derivatives in the large-scale variables. Thus the induction for n≥3 is not closed as written. One must either prove that A inherits (1.7) with controlled constants, or supply a different argument that avoids full analyticity of the block-homogenized coefficient.","section":"§5.2, Proof of Theorem 1.1"},{"comment":"The estimate (5.18) for the first block-homogenization step contains the factor ε_{n-m+1}/ε_{n-m}. In the first alternative of Lemma 5.6 this factor is of order one, and the paper dismisses the resulting bound as 'trivial'. This is only acceptable because the final right-hand side of (1.8) contains e^{-c ε_{n-m}/ε_{n-m+1}}, which is also of order one in that case. The argument should state this explicitly; as written it is easy to misread the intermediate estimate as being small when it is not.","section":"§5.2, Lemma 5.6 and (5.18)"}],"minor_comments":[{"comment":"The proofs of Theorems 3.4 and 3.6 choose many constants in sequence (C, C_*, bC_*, C_**, eC_j, Λ_j, etc.). A consolidated list or a table of the constraints would substantially improve verifiability.","section":"§3.3–§3.4"},{"comment":"The asserted derivative bound |d^k b_1/dy_1^k| ≤ (2π)^k k! for b_1(y_1) = (β_0!/β_0^{β_0}) sin(2πβ_0 y_1) is not immediate: the frequency β_0 enters, so a short verification using β_0!/β_0^{β_0} ≤ e^{-β_0} (or Stirling's formula) should be supplied.","section":"§7.2"},{"comment":"The operator b∇_n is used before the parameters δ_i are introduced in the surrounding text; define δ_i explicitly when first defining b∇_n.","section":"§2"},{"comment":"In the displayed computation of the explicit solution, '1/α_ε(t)' appears where the coefficient should be denoted a_ε(t). Please correct this typo.","section":"§7.2"}],"recommendation":"major_revision","confidential_remarks":"The main technical gap is confined to the induction in §5.2. The rest of the paper is detailed and the counterexamples are valuable. If the authors can establish the missing analyticity inheritance or replace the induction, the result would be significant; I do not see reasons to doubt the two-scale case or the conditional results under scale separation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: this is a substantial paper, and the reader's conditional verdict is about right. The genuinely new thing is the exponential ratio rate e^{-c ε_i/ε_{i+1}} for analytic coefficients, which is a real improvement over Niu–Shen–Xu's linear ratio bound. The multiscale corrector framework—lifting to a degenerate periodic problem, adding τ as a regularizer, and proving uniform-in-τ estimates from the analyticity—is the right tool, and the main argument under scale separation (Theorem 5.4) is detailed and largely convincing. I also think the uniform Lipschitz estimate under double-log separation is a nice byproduct.\n\nThe soft spots are exactly where the reader puts them, with one exception.\n\nThe §5.2 induction is the load-bearing gap. The proof removes scale separation by homogenizing the smallest m scales simultaneously and then invoking the theorem for the remaining n−m scales. That requires the block effective matrix to inherit the real-analyticity assumption (1.7) with controlled constants. The paper simply asserts this ('On the other hand, since A(y1,...,y_{n−m}) satisfies the assumptions (1.5)–(1.7)'). Theorem 3.9 gives Lipschitz dependence of the correctors on the external parameter, not analyticity of the homogenized matrix with respect to the remaining variables. For two scales the issue can be dodged, but for n≥3 the no-separation version of Theorem 1.1 is not justified as written. I don't see an internal contradiction; the statement is probably true, but the proof needs another theorem.\n\nThe counterexample worry in the reader's report, however, does not land. In §7.2 the coefficient b1 has amplitude β0!/β0^{β0} ~ e^{-β0}. Since e^{-β0} β0^k ≤ k!, the derivative bounds are uniform in β0, and composition with (1+z)^{-1} keeps analyticity constants under control. So the example does speak to optimality within a fixed analytic class.\n\nMinor: the effective matrix depends on τ and the ratios ε_i/ε1; Proposition 4.1 clarifies the relation to the reiterated limit, but the τ-dependence in Theorem 1.1 is implicit.\n\nBottom line: this deserves a serious referee. The scale-separated theorem is well-supported; the full Theorem 1.1 needs the §5.2 gap closed. I'd send it to peer review and ask the authors to either prove the analyticity inheritance or state the theorem under scale separation.","headline":"Real progress on multiscale homogenization rates, but the no-separation proof rests on an unproved analyticity inheritance claim in §5.2.","tokens_in":61393,"tokens_out":4246,"would_cite":true,"duration_ms":47694,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35B27"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that for real-analytic, periodic, uniformly elliptic coefficients oscillating at n scales, the L2 homogenization error is bounded by the largest scale ε1 plus exponentially small terms e^{-c ε_i/ε_{i+1}} between consecutiv","keywords":["multiscale homogenization","convergence rate","correctors","real analytic coefficients","elliptic equations","uniform Lipschitz estimate","reiterated homogenization","effective matrix"],"falsifier":"Take a concrete real-analytic two-scale coefficient such as A(y_1,y_2) = (2 + sin 2π y_1 sin 2π y_2)^{-1}, set ε_1 = ε and ε_2 = ε/β, and measure ||u_ε - u_bar||_{L2} as β grows at fixed small ε. The theorem predicts decay like C(ε + e^{-cβ}); if instead the error decays no faster than C/β for some analytic coefficient of this type, the central claim would be refuted, whereas the Section 7.2 counterexample only rules out enlarging c.","tokens_in":60333,"feed_emoji":"📉","tokens_out":9618,"duration_ms":92972,"temperature":0.7,"pith_summary":"Multiscale elliptic equations with coefficients oscillating at several separated scales traditionally have homogenization errors limited by the slow ratios ε_{i+1}/ε_i, the hallmark of reiterated homogenization. This paper shows that if the coefficient matrix is real analytic in all its scale variables, the ratio part of the error can be improved to max_i e^{-c ε_i/ε_{i+1}}, with ε1 alone remaining as the linear error. The improvement comes from constructing multiscale correctors and flux correctors that homogenize all scales simultaneously, yielding an effective matrix that depends on the scale ratios. If correct, this removes the old ratio bottleneck, makes the exponential rate optimal, and gives uniform Lipschitz regularity under a mild double-log scale-separation condition.","feed_headline":"Exponential accuracy beats slow ratios in multiscale homogenization","feed_subtitle":"For real-analytic coefficients, only the largest scale and exponentially small scale gaps limit the L2 error.","key_machinery":"The load-bearing device is the multiscale corrector X, defined on the n-fold torus as the solution of the tau-regularized degenerate lifted equation -grad_delta · A grad_delta X + tau^2 X = grad_delta·(A v), where grad_delta = Σ δ_i^{-1} ∇_{y_i} with δ_i = ε_i/ε_1. The effective matrix is the cell average A_bar = ⟨A + A grad_delta X⟩. Because the equation is degenerate, the paper solves it by a formal expansion in powers of the smallest ratio δ_n, reducing each step to an equation with n-1 scales; real analyticity supplies the factorial derivative bounds that make the truncated expansion accurate with truncation order k ≃ ε_{n-1}/ε_n, producing the exponential factor e^{-c ε_{n-1}/ε_n}. Flux","core_discovery":"The central claim is Theorem 1.1: under ellipticity, periodicity, and the real-analyticity bound (1.7), the solution u_ε of -∇·A(x/ε_1,...,x/ε_n)∇u_ε = f with Dirichlet data g is within C(ε_1 + max_i e^{-c ε_i/ε_{i+1}})(||f||_{L2} + ||g||_{H^{3/2}}) of the solution u_0 of -∇·A_bar ∇u_0 = f with the same data, where A_bar is a constant matrix determined by A and the ratios ε_i/ε_1. The proof identifies the slow linear ratios in earlier results as an artifact of reiterated homogenization, which homogenizes one scale at a time and misses interactions between close scales. The new effective matrix is built by simultaneous homogenization through multiscale correctors, and a one-dimensional analyt","pith_inferences":["The mechanism suggests a general principle: analyticity makes high-frequency Fourier coefficients of the coefficient matrix decay exponentially, so almost-resonant interactions between close scales have exponentially small amplitude; the same principle should transfer to other analytic quasi-periodic homogenization problems, including parabolic or nonlinear settings, once the partial-homogenizatio","The ratio-dependent effective matrix has a practical consequence for computation: a numerical homogenization scheme that precomputes one classical homogenized matrix will miss the interaction correction; targeting A_bar(ε) instead would capture accuracy gains precisely in the regime where scales are close but separated.","A direct testable extension would quantify the trade-off for finite smoothness: if A is only C^m, the same expansion should give polynomial rates in ε_i/ε_{i+1} with exponent tied to m; the paper notes this qualitatively but does not state the sharp polynomial rate."],"forward_implications":["Quantitative homogenization of analytic multiscale coefficients no longer has to pay the slow ratio ε_{i+1}/ε_i: the only linear error is the largest scale ε_1.","The effective matrix depends on the ratios ε_i/ε_1, not just on A, so the homogenized equation encodes interactions between scales; the difference from the classical reiterated-homogenization matrix is controlled by C(τ^2 + max_j ε_j/ε_{j-1}).","The exponential rate is optimal: the paper's one-dimensional analytic example shows that replacing c by a larger constant fails in general.","Uniform Lipschitz estimates hold under the double-log separation condition ε_i/ε_{i+1} ≥ M log log ε_i^{-1}, far weaker than previously required power separation.","The same approach is stated to cover elliptic systems, x-dependent analytic coefficients with Lipschitz x-dependence, and different boundary conditions."],"fun_headline_variants":["Exponential gains replace ratio-limited errors in homogenization","Multiscale correctors yield optimal exponential convergence rates","Optimal exponential error bounds for multiscale homogenization","From linear to exponential: sharp rates in homogenization","Exponentially better: optimal rates for multiscale homogenization"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof's induction for removing scale separation assumes that the matrix obtained by simultaneously homogenizing the smallest m scales still satisfies the same real-analyticity and derivative-growth assumptions (1.5)-(1.7); Section 5.2 asserts this rather than proving it, and the uniform corrector estimates in Theorem 3.9 alone give only Lipschitz, not analytic, dependence on the large-scale parameters.","fun_headline_variants_meta":{"raw":{"variants":["Exponential gains replace ratio-limited errors in homogenization","Multiscale correctors yield optimal exponential convergence rates","Optimal exponential error bounds for multiscale homogenization","From linear to exponential: sharp rates in homogenization","Exponentially better: optimal rates for multiscale homogenization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000926,"raw_usage":{"total_tokens":3862,"prompt_tokens":856,"completion_tokens":3006,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":2936}},"tokens_in":600,"tokens_out":3006,"duration_ms":26031,"temperature":1.0,"reasoning_tokens":2936,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T19:09:08.885797+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a concrete real-analytic two-scale coefficient such as A(y_1,y_2) = (2 + sin 2π y_1 sin 2π y_2)^{-1}, set ε_1 = ε and ε_2 = ε/β, and measure ||u_ε - u_bar||_{L2} as β grows at fixed small ε. The theorem predicts decay like C(ε + e^{-cβ}); if instead the error decays no faster than C/β for some analytic coefficient of this type, the central claim would be refuted, whereas the Section 7.2 counterexample only rules out enlarging c.","supporting_citations":[],"review_version":1}