{"id":"da381b0d-7b29-49ef-9745-106592ae8677","arxiv_id":"2607.29466","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"On symmetric triangular refinements, even polynomial surface approximations gain one extra order of accuracy in weighted integrals of functions, derivatives, normals, and curvature, while pointwise errors remain at standard orders.","lead":"Even-numbered surface mesh shapes often converge faster in averaged quantities than standard error bounds predict, but only when the mesh has mirror-symmetric triangle pairs. This paper proves the extra order comes from cancellation of leading errors in integrals, not from better pointwise accuracy.","discovery_kind":"first_principles","skeptic_critique":null,"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops an interpolation theory for weighted integral errors on symmetric structured triangulations of surface macro-elements and uses it to explain why even-order piecewise polynomial surface approximations display superconvergent geometric consistency errors. The central object is a macro-element triangulation that can be partitioned into mirror-image pairs of triangles plus an O(h) boundary strip, as is produced by uniform red refinement. The main flat estimate, Theorem 3.2, shows that for degree-k Lagrange interpolation the integral of D^β(f_h - f) over a symmetric macro element is O(h^{k+2-|β|}) whenever k+2-|β| is even, one order better than the standard bound. Weighted versions, first-derivative refinements, and transfers to curved surfaces via the parametrization lifting L and the closest-point projection π are then proved (Theorems 3.9-3.12, Corollary 3.15). These estimates are applied to geometric quantities: improved L1-type integral bounds for the signed distance, normal-vector inner products, the Weingarten map, and Gaussian curvature (Lemma 4.1, Lemma 4.3, Corollary 4.4), while pointwise errors retain standard orders. Numerical experiments on a deformed sphere under red refinement confirm the predicted odd-even pattern, and additional experiments in the appendices test higher derivatives, projected refinement, and newest-vertex bisection, with the latter explicitly outside the proved statement.","tokens_in":34989,"tokens_out":15371,"duration_ms":139485,"significance":"If the estimates hold, the paper gives a genuine first-principles explanation of a repeatedly observed numerical phenomenon, namely that even-order surface parametrizations often integrate geometric quantities one order better than standard approximation theory predicts. The mechanism is transparent and falsifiable: the parity condition k+2-|β| even and the symmetric-pair structure are concrete and are verified by the experiments. Strengths of the paper are that the improved orders are derived rather than fitted, that the hypotheses are stated precisely, and that the numerical code and data are deposited. The authors are also appropriately careful about scope, explicitly stating that newest-vertex bisection with closure is not covered by the proofs and that the projected-refinement experiments in Section B.1 rest on an interpretation whose supporting smooth parametrization is not constructed. The paper should be of immediate interest to researchers working on higher-order surface finite element methods, surface Stokes discretizations, and geometric consistency error analysis.","major_comments":[],"minor_comments":[{"comment":"The scalar case is well explained, but for the derivative case the sentence 'we apply the same argument to the differentiated leading term' skips the treatment of the nonconstant factor D^α f(x) inside the integral; the remainder R_{k+2-|β|} is asserted to absorb it. Two extra lines showing that the linear correction around x0 is of order h times the differentiated basis product would make the proof fully self-contained.","section":"Section 3.1, proof of Theorem 3.2"},{"comment":"The identity ρΓ(Φh)=ρΓ(Φh)(n_h·n + 1/2||n_h-n||^2)=n·e+O(||e||^2) is compressed; a short derivation obtained by dotting the closest-point projection expansion with n would help the reader verify the sign and the O(||e||^2) form.","section":"Section 3.2, Lemma 3.6"},{"comment":"A few flat-test values deviate from the predicted rates (for example b=3, k=5 gives 3.44 versus the predicted 4, and b=3, k=6 gives 3.82 versus 4). Since the text warns about roundoff only for the m=6 surface tests, please add a sentence explaining that these deviations are pre-asymptotic or quadrature-limited.","section":"Section 5.2, Table 1"},{"comment":"The construction of the smooth macro parametrization on the polytopal reference grid should be clarified: the map Φ uses spherical coordinates of the point x on the flat macro triangle, and the reader should be told explicitly why this map is C^{m+2} on each closed macro element, especially if any macro element has a vertex on the polar axis.","section":"Section 5.1"},{"comment":"The caveat that projected refinement is not directly covered by the assumptions because a smooth macro parametrization preserving the projected nodes is not constructed is welcome and important; consider stating this limitation at the beginning of Section B.1 rather than only at its end, so that the appendix tables are not mistaken for proved cases.","section":"Appendix B.1"}],"recommendation":"minor_revision","confidential_remarks":"I have no concerns about novelty, attribution, or scope. The manuscript builds properly on the earlier work of Chien, Zavalani et al., Nedelec, and Demlow, and it is honest about the limits of the proofs. The numerical code is available, and the predicted rates are specific enough to be independently checked. The remaining issues are local clarifications, so I recommend minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear —,\n\nThe short version: this is a real result, and it holds up. The paper explains a genuine observed phenomenon — even-order surface geometry approximations converge at h^{m+2} in integral geometric consistency terms instead of the standard h^{m+1} — and the explanation is the right one: cancellation of the leading odd Taylor term on mirror-symmetric triangle pairs. The core estimate, Theorem 3.2, is elementary and correct. The authors properly extend Chien's parity argument from integration errors to interpolation errors and to normals, Weingarten map, and Gaussian curvature. The transfer to curved surfaces via parametrization lifting is careful; the distinction between the two liftings is handled explicitly. The numerical experiments reproduce the predicted rates and, importantly, show the pointwise errors stay at standard order, which is what separates cancellation from actual approximation.\n\nThe paper is also honest about its limits. The proof needs the symmetric-pair structure of red refinement; newest-vertex bisection is explicitly not covered, though the numerics suggest the effect persists. The red-refinement decomposition lemma is a bit sketched but convincing. The boundary pairing assumption (Remark 2.9) is an extra condition that is true for red refinement and needed for the first-derivative estimates. The higher-order derivative appendix is heavy going, but it's an appendix and the main text doesn't depend on it.\n\nSoft spots, in proportion. The novelty is real but modest — Chien had the core cancellation mechanism for surface integration, and Zavalani et al. extended it. What's new here is the systematic treatment of interpolation errors and geometric quantities, which is the piece the surface-FEM community has been needing. The applicability to full discretizations is limited, as the authors admit: the improved normal estimates don't immediately improve vector-Laplace schemes because the regularity demands don't match. That is stated, not hidden.\n\nWho is this for? Anyone doing a priori error analysis of higher-order surface FEM, especially for surface Stokes, tensor-valued PDEs, or curvature-dependent problems. It will save people from re-deriving these cancellations and from trying to prove them using NVB.\n\nRecommendation: send it to review. I would accept it with minor revisions — mainly tightening the red-refinement lemma and perhaps adding a remark on why the boundary hypothesis is not vacuous. The code and data are available, the claims are proportional to the proofs, and the limitations are honestly stated.","headline":"Solid proof-level explanation of even-order superconvergence on symmetric meshes; correct core argument, honest limitations, deserves review.","tokens_in":35566,"tokens_out":8661,"would_cite":true,"duration_ms":71629,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65N15","65D05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Even-order surface approximations gain an extra order of accuracy in integral quantities because mirror-image mesh pairs cancel the leading interpolation error.","keywords":["surface finite element methods","superconvergence","geometric consistency errors","symmetric triangulations","Lagrange interpolation","curvature approximation","red refinement","parity-dependent convergence"],"falsifier":"On one flat macro triangle with uniform red refinement, take degree-2 Lagrange interpolation of a smooth function and compute $|\\sum_T\\int_T(f_h-f)|$; the symmetric-pair proof predicts $h^4$. If every interior vertex is then randomly perturbed off the symmetric lattice while keeping the mesh size fixed, the leading $h^3$ term should reappear and the measured order should drop to 3.","tokens_in":34865,"feed_emoji":"📐","tokens_out":7101,"duration_ms":59514,"temperature":0.7,"pith_summary":"Standard surface finite element analysis predicts that a degree-$m$ polynomial fit to a smooth surface converges like $h^{m+1}$ in position and one order less per derivative. This paper establishes a systematic exception: on meshes produced by structured refinement that create mirror-image pairs of triangles, the leading interpolation errors cancel in integrals, so even orders $m$ gain one extra order ($h^{m+2}$ for distance, normal, and function integrals; $h^m$ for Weingarten and Gaussian curvature integrals), while odd orders keep the standard rates. The mechanism is a parity-dependent cancellation of Taylor terms over symmetric pairs, proved for red refinement and verified numerically on a deformed sphere. The result matters because even-order geometry is common in surface finite element computations, and observed superconvergent geometric consistency errors have lacked a proof.","feed_headline":"Mirror-image triangles cancel leading errors on even-order meshes","feed_subtitle":"Surface finite element integral errors gain one order for even polynomial degrees on structured symmetric grids.","key_machinery":"The load-bearing object is the symmetric pair decomposition of a flat macro triangulation (Definition 2.10), which uniform red refinement produces by grouping the six triangles around each interior vertex into three mirror-image pairs (Lemma 2.8); the remaining unpaired triangles lie in an $O(h)$ strip along the macro-element boundary, and for first-derivative estimates a uniform induced boundary triangulation pairs boundary intervals as well. On a symmetric pair, the Lagrange basis functions are even under the reflection $R\\bar x = 2\\bar x_0 - \\bar x$ while the Taylor monomial is odd, so the leading error terms cancel in the integral over the pair; the same computation for derivatives cancels exactly when $k+2-|\\beta|$ is even. Counting the boundary strip and applying weighted Poincar\\'e-type estimates turns this local cancellation into global $h^{k+2-|\\beta|}$ bounds, which are then transferred to curved patches through the parametrization lifting and closest-point projection.","core_discovery":"The central claim is Theorem 3.2 and its surface lifts: for degree-$k$ Lagrange interpolation on a symmetric triangulation of a macro element, whenever $0\\le |\\beta|\\le k+2$ and $k+2-|\\beta|$ is even, $\\bigl|\\sum_T \\int_T (D^\\beta f_h - D^\\beta f)\\,d\\sigma\\bigr| \\le C h^{k+2-|\\beta|}$, one power better than the standard bound. Transferred to an even geometry order $m$, this gives $h^{m+2}$ for weighted integrals of the signed distance, for inner products with normal-vector errors and projection terms, and for function interpolation with both parametrization and closest-point liftings, and $h^m$ for Weingarten and Gaussian curvature integrals. Pointwise errors keep their standard orders $h^{m+1}$, $h^m$, and $h^{m-1}$, so the improvement is a genuine integral-level cancellation, not better local approximation. The proof identifies the cancellation on mirror-image triangle pairs produced by uniform red refinement: the leading Taylor term flips sign between paired elements, leaving only an $O(h)$ boundary strip of unpaired triangles to contribute at the standard rate.","pith_inferences":["If the same parity pattern persists for newest-vertex bisection, as the experiments suggest, then proving a pair-decomposition bound for bisection closures would extend the theory to adaptive surface meshes; the paper leaves this open.","The per-pair cancellation could be tested directly by isolating the contributions of symmetric pairs versus boundary triangles, which would separate the cancellation mechanism from other sources of superconvergence.","The improved normal-vector and curvature integral estimates may feed into sharper geometric consistency terms in surface vector-Laplace and Stokes analyses, although the paper notes the extra regularity requirements can block an immediate gain in full discretization rates.","Projected refinements built from structured coarse grids show the same even-order gains numerically; a proof would need smooth patch parametrizations compatible with projected Lagrange nodes, which the paper does not construct."],"forward_implications":["Even geometry order $m$ yields $h^{m+2}$ for weighted integrals of the signed distance and for inner products against normal-vector errors, one order beyond the standard $h^{m+1}$.","Weingarten and Gaussian curvature integrals on even geometries converge as $h^m$, improving the standard $h^{m-1}$ by one order.","Odd geometry orders show no improvement: distance and normal integrals stay $h^{m+1}$ and curvature integrals $h^{m-1}$, so parity is the observable signature of the cancellation.","Pointwise errors remain standard, so any reported superconvergence in $L^\\infty$-type surface quantities must come from a different mechanism.","The improved rates hold for both the parametrization lifting and the closest-point projection, for function-value and first-derivative interpolation integrals on global surfaces under red refinement."],"supporting_citations":[{"why":"Supplies the original symmetric-mesh cancellation argument for surface integrals that this paper extends to interpolation errors.","marker":"Chien (1993)"},{"why":"Uses the same symmetric-pair structure for collocation errors on surfaces.","marker":"Atkinson and Chien (1995)"},{"why":"Gives the numerical surface-integration rates on symmetric meshes that the improved integral estimates generalize.","marker":"Chien (1995)"},{"why":"Recent treatment of convergence rates for integration schemes on closed surfaces that this paper connects to the interpolation theory.","marker":"Zavalani et al. (2024)"},{"why":"Provides the curved finite element interpolation and projected area-element estimates used to lift flat estimates to surfaces.","marker":"Nedelec (1976)"},{"why":"Establishes the standard higher-order surface geometric estimates that the parity-improved bounds beat.","marker":"Demlow (2009)"},{"why":"Gives the Lemma 4.2 baseline for normal-vector projection integrals that Lemma 4.1 improves for even $m$.","marker":"Hansbo et al. (2020)"},{"why":"Documents the observed Gaussian curvature inner-product superconvergence in surface Stokes analysis that this paper proves.","marker":"Hardering and Praetorius (2025)"},{"why":"Supplies the standard Lagrange interpolation estimates that serve as the baseline and fallback bound.","marker":"Ern and Guermond (2021)"},{"why":"Provides the boundary-strip measure estimate used to control unpaired boundary triangles.","marker":"Mittelmann (1977)"}],"fun_headline_variants":["Even-order surface meshes cancel leading integral errors","Mirror-image triangles explain superconvergent surface errors","Even polynomial degree boosts accuracy in surface integrals","Parity-dependent superconvergence from mesh symmetry","Even order on symmetric grids: one extra integral accuracy power"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof collapses if the refined triangulation cannot be decomposed into mirror-image triangle pairs covering the interior with unpaired elements confined to an $O(h)$ boundary strip; general newest-vertex bisection with closure is explicitly not covered, and without the uniform boundary triangulation the first-derivative improvement is not established.","fun_headline_variants_meta":{"raw":{"variants":["Even-order surface meshes cancel leading integral errors","Mirror-image triangles explain superconvergent surface errors","Even polynomial degree boosts accuracy in surface integrals","Parity-dependent superconvergence from mesh symmetry","Even order on symmetric grids: one extra integral accuracy power"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1377,"prompt_tokens":893,"completion_tokens":484,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":409}},"tokens_in":509,"tokens_out":484,"duration_ms":4817,"temperature":1.0,"reasoning_tokens":409,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:23:14.865012+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On one flat macro triangle with uniform red refinement, take degree-2 Lagrange interpolation of a smooth function and compute $|\\sum_T\\int_T(f_h-f)|$; the symmetric-pair proof predicts $h^4$. If every interior vertex is then randomly perturbed off the symmetric lattice while keeping the mesh size fixed, the leading $h^3$ term should reappear and the measured order should drop to 3.","supporting_citations":[],"review_version":2}