{"id":"a8bf02ac-6467-457e-a4ec-b206fea97f80","arxiv_id":"2607.22206","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For 1D Poisson, an Armijo-type mesh-moving algorithm driven by the residual estimator converges, and numerical tests show optimized meshes outperform uniform meshes.","lead":"This paper recasts adaptive mesh movement (r-refinement) as a shape-optimization problem and proves a standard descent algorithm converges when it minimizes either the true finite-element error or a residual-based error estimator in one dimension. Readers may care because it gives a convergence guarantee for a practical mesh-moving strategy and links residual-based adaptivity to shape calculus.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (23) gives a wrong derivative for E: for f=1, P1 on two elements, the formula yields 3.78 times the true derivative; Theorem 2's descent proof collapses.","rationale":"The reader's weakest assumption was loss of injectivity in the limit. That concern is less severe than it appears because Algorithm 1 restricts α_n to 2^{-k} with k≥1 and |∂_xU_n|≤1, so each finite iterate remains injective; degeneracy can occur only in the infinite limit, while the algorithm terminates at finite n for any fixed ε>0. The load-bearing issue is instead Lemma 4's derivative formula. A two-element P1 calculation with f=1 gives an explicit counterexample: the true derivative of E is 3δ(2a-1), while formula (23) with adjoint (24) produces δ(2a-1)[4-(a³+(1-a)³)/2]. These differ for every non-symmetric mesh, so the computed gradient is not the actual gradient of the estimator. Since Lemma 5 and Theorem 2 rely entirely on this derivative for the Armijo sufficient-decrease condition and the convergence argument, the central claim is not merely unproved—it is false as stated. The paper could potentially be repaired by re-deriving the adjoint and derivative, but as written the rigorous foundation for the residual-based algorithm is invalid. I therefore recommend REJECT rather than the reader's CONDITIONAL.","tokens_in":13277,"tokens_out":21446,"duration_ms":188020,"concrete_test":"Run the one-line check: with Ω=(0,1), f=1, P1, mesh nodes {0,a,1}, set E=a³+(1-a)³. Compute true derivative along U(a)=δ as 3δ(2a-1). Solve the adjoint (24) for p_a=-a(1-a)(a³+(1-a)³) and substitute into (23); the result is δ(2a-1)[4-(a³+(1-a)³)/2]. For a=1/4, the coefficients differ: 3 vs 3.78125, a 26% discrepancy. This settles whether the printed derivative is correct.","verdict_should_be":"REJECT","load_bearing_attack":"The central technical result is Lemma 4's derivative formula, on which the Armijo step and Theorem 2 rest. A direct one-dimensional check shows it is incorrect. Take Ω=(0,1), f≡1, P1 elements with nodes 0,a,1. Since y_Φ is the nodal FEM solution and ∂_xx y_Φ=0, the estimator (18) is E(a)=a³+(1-a)³. For U∈V_Φ with U(0)=U(1)=0, U(a)=δ, the true derivative is d/dε E(a+εδ)|_0 = 3δ(2a-1). The proposed adjoint (24) gives p_a=-a(1-a)(a³+(1-a)³), and substituting into (23) yields E'(Φ)[U] = δ(2a-1)[4 - (a³+(1-a)³)/2], which is not 3δ(2a-1) for a≠1/2. Hence Lemma 4 is false as stated; Lemma 5 and Theorem 2, which use this derivative for the sufficient-decrease/descent argument, do not follow. The numerical experiments may still work for the chosen data, but the paper's rigorous convergence claim for the residual-based strategy is unsupported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an r-refinement strategy for finite element Poisson problems, formulated as a shape-optimisation problem over admissible mesh parametrisations. The main theoretical results are: (i) for the exact H^1 or L^2 error, an Armijo-type algorithm generates a sequence whose derivative goes to zero and which has a subsequential limit in a non-injective enlargement of the admissible set (Theorem 1, following [15]); and (ii) in one dimension, a fully computable residual-based estimator E in (18) is claimed to be differentiable, with an explicit derivative (23), an Armijo sufficient-decrease condition (Lemma 5), and a corresponding convergence theorem (Theorem 2). Numerical experiments in one and two dimensions are presented.","tokens_in":13606,"tokens_out":16568,"duration_ms":153477,"significance":"If the main theorem were correct, the paper would give a useful rigorous foundation for residual-based r-refinement, a topic where rigorous convergence results are scarce. The paper is clearly written, and the numerical experiments show that the algorithm improves the approximation in the tested cases. However, the central derivative formula (23) is not correct; an elementary one-dimensional computation gives a different value for the derivative of the estimator. Since the Armijo proof and Theorem 2 rely directly on this formula, the paper's main theoretical contribution is not established. The convergence statements are also weakened by the possibility that the limit parametrisation is not injective. The significance of the paper is therefore conditional on a successful correction of the derivative and a reworking of the subsequent proof.","major_comments":[{"comment":"Lemma 4 is false as stated. Take Ω=(0,1), f≡1, P1 elements with nodes 0<a<1, and U the hat function with U(a)=δ. Then y_Φ(a)=a(1-a)/2 and the estimator (18) is E(a)=a^3+(1-a)^3, so its true derivative is 3δ(2a-1). Solving the adjoint equation (24) gives p_Φ(a)=-a(1-a)E(a). Substituting into (23) and simplifying yields E'(Φ)[U]=δ(2a-1)(4-E(a)/2). For a=1/4 this is -1.890625δ, whereas the true derivative is -1.5δ. Thus (23) is not the derivative of E. Since Lemma 5 and Theorem 2 use this expression in the Armijo sufficient-decrease condition, the convergence proof does not follow.","section":"§3.2, Lemma 4 and Eq. (23)"},{"comment":"The convergence statements only guarantee a subsequential limit in the enlarged space (17)/(41), not in the admissible set U. After (17) the author acknowledges that Φ* need not be injective, and the condition sup_n |DΦ_n^{-1}|<C is stated but never proved. If the limit is non-injective, it is not a valid mesh, so the algorithm has not been shown to produce a usable converged mesh. This limitation should be stated as an explicit assumption or proved as part of the theorem.","section":"§3.1, Theorem 1, and §3.2, Theorem 2"}],"minor_comments":[{"comment":"The statement of Lemma 5 writes J'(Φ)[U] where the objective is E; the derivative should be E'(Φ)[U]. In the expansion at (33) the term I2,3 appears twice; one of these should presumably be I2,2.","section":"§3.2, Lemma 5"},{"comment":"The one-dimensional estimator (18) omits the jump term that appears in the general residual estimator (5). If this is intentional, it should be justified; if not, the estimator is not the standard residual estimator.","section":"§3.2, Eq. (18)"},{"comment":"The definitions of f in §4.1 and §4.3 are inconsistent: §4.1 gives f(x)=-C^2x^2(6-20x-15x^2), while §4.3 gives f(x)=-2Cx^2(6-20x-15x^2). Also, the last sentence of §4.3 ('Choosing less regular data') is incomplete.","section":"§4"}],"recommendation":"reject","confidential_remarks":"The paper is well organised and the numerical experiments are plausible, but the central theoretical claim is invalidated by the elementary counterexample to Lemma 4. I do not see a minor repair: the derivative formula and the Armijo convergence proof would both need substantial reworking. If the author can provide a corrected derivative and a full verification of the descent argument, I would be willing to reconsider the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The novel claim here is that the residual estimator E in 1D is differentiable with respect to the mesh and that an Armijo descent on E converges. That claim does not hold as written. I checked Lemma 4 directly. Take f=1, P1 elements on (0,1) with interior node at a. Then y_Φ is piecewise linear, ∂_xx y_Φ=0, and the estimator reduces to E(a)=a^3+(1-a)^3, whose derivative is 3δ(2a-1). Plugging the same data into the proposed formula (23) gives a different value—for a=1/4, the formula yields about -1.89δ instead of -1.5δ. So Lemma 4 is false, and Lemma 5 and Theorem 2, which rely on that derivative for sufficient decrease, do not follow. This is a load-bearing flaw, not a minor typo. (The stress-test note's specific ratio of 3.78 doesn't match my calculation, but the conclusion is correct: the formula is wrong.)\n\nWhat the paper does well: the idea of viewing r-refinement as shape optimization and applying W^{1,∞} methods is sensible and timely. The exact-error convergence (Theorem 1) is essentially a specialization of the author's prior work [15], but it is cleanly stated. The numerical experiments are encouraging—the optimized meshes do improve on uniform ones in the test cases—but they cannot rescue the theory.\n\nOther soft spots: the limit Φ* is not shown to be injective, so the terminal mesh may be degenerate (the author notes this but handwaves it). The proof leans heavily on [15] and [23], with several typos (e.g., duplicated I_{2,3}, J' vs E' in Lemma 5) that make it harder to audit. No code or data is shipped, and the Data Availability statement says 'no data.'\n\nWho should read this: researchers working on r-refinement or computational shape optimization might find the framing interesting, but they should not rely on Theorem 2 until the derivative is corrected and the injectivity issue is resolved. I would not cite it in its current form.\n\nRecommendation: send it to peer review—a competent referee will catch the derivative error, and the underlying idea may be salvageable. But it needs major revision before publication.","headline":"The paper's central derivative formula (Lemma 4) is incorrect, so the fully practical convergence theorem (Theorem 2) is unsupported; the exact-error part is a clean but mostly prior-work corollary.","tokens_in":14055,"tokens_out":6697,"would_cite":false,"duration_ms":57286,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65N50","49Q10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Residual estimator gives a provably convergent r-refinement in 1D","keywords":["r-refinement","shape optimisation","finite element method","residual error estimator","Poisson equation","Armijo descent","mesh adaptation","one-dimensional"],"falsifier":"Run Algorithm 1 on a one-dimensional Poisson problem with a boundary-layer profile in f and monitor the minimum element length; if an element length converges to zero (so the parametrisation leaves U) before ∥E'(Φ_n)∥ reaches the stopping tolerance, the practical convergence claim without the bounded-inverse assumption fails.","tokens_in":13180,"feed_emoji":"📐","tokens_out":4417,"duration_ms":40821,"temperature":0.7,"pith_summary":"This paper shows that moving the nodes of a finite element mesh to reduce the residual error can be treated as a shape optimisation problem. It proves that a standard Armijo descent algorithm converges when the cost is the true finite element error (in any dimension), and, more importantly, when the cost is a fully computable residual-based error estimator in one dimension. This establishes a rigorous foundation for a practical r-refinement strategy that improves the approximation without changing the mesh topology. Numerical experiments confirm that the optimised meshes beat uniform meshes and are competitive with h-refinement.","feed_headline":"Residual estimator gives a provably convergent r-refinement in 1D","feed_subtitle":"A shape-optimisation view of node movement reduces the residual error without knowing the exact solution.","key_machinery":"The central object is the space of admissible mesh parametrisations U, consisting of piecewise-linear bi-Lipschitz maps that fix the boundary, together with the residual estimator E(Φ) defined on the image mesh. The argument relies on an explicit formula for the derivative E'(Φ)[U] derived via a Lagrangian adjoint approach, and an Armijo step-size rule that guarantees a sufficient decrease at each iteration. This yields the convergence of the descent algorithm and the subsequential limit result of Theorem 2.","core_discovery":"The central claim is that the residual-based error estimator E(Φ) (Equation 18) is differentiable with respect to the mesh parametrisation Φ, and that the Armijo-type Algorithm 1 applied to E(Φ) produces a sequence of parametrisations whose derivative norm tends to zero and which has a convergent subsequence. This makes estimator-based r-refinement a fully computable and provably convergent adaptive strategy in one dimension. The same algorithm applied to the exact H1 or L2 error converges in arbitrary dimensions, as stated in Theorem 1.","pith_inferences":["The differentiability of the residual estimator is established only in one dimension, where the element diameter equals its integral; extending the strategy to higher dimensions likely requires a smoothed element-size measure or a different estimator, as the author notes.","The convergence proof assumes a uniform bound on the inverse derivative of the parametrisation; without it, the limiting map may lose injectivity, so a practical implementation may need to enforce validity of the mesh through projection or constraints.","The numerical comparison to h-refinement suggests that r-refinement could serve as an alternative when mesh topology must remain fixed, for instance in parallel solvers or time-dependent problems where remeshing is costly.","If the residual estimator is accepted as a reliable error proxy, this optimisation viewpoint opens the door to more general estimators, such as dual-weighted residuals, within the same descent framework."],"forward_implications":["For any one-dimensional Poisson problem with f ∈ H², repeated application of Algorithm 1 with the residual estimator terminates or drives the derivative norm to zero, giving a fully computable convergence guarantee.","When the exact H¹ or L² error is used as the objective, the same algorithm converges in arbitrary dimensions, not just one.","Optimised meshes reduce the finite element error compared to uniform meshes at the same number of degrees of freedom, as shown in the numerical experiments.","The method is competitive with a greedy h-refinement strategy in the one-dimensional example, suggesting that fixed-topology node movement can match the error reduction of local refinement.","The results turn r-refinement from a heuristic mesh-moving technique into an optimisation problem with a rigorous descent property."],"fun_headline_variants":["Shape optimization turns residual estimator into a convergent r-refinement","Differentiable residual estimator yields provable convergence in 1D r-refinement","Shape-optimized node movement reduces residual error with proof","r-refinement with residual estimator: shape-optimized convergence","Provably convergent mesh adaptivity via residual estimator in 1D"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The sequence of mesh parametrisations must stay under a uniform Lipschitz bound on the inverse derivative (sup_n |DΦ_n^{-1}| < C), otherwise the limit Φ* may be non-injective and no longer define a valid mesh, so the practical convergence to a usable mesh is not fully established.","fun_headline_variants_meta":{"raw":{"variants":["Shape optimization turns residual estimator into a convergent r-refinement","Differentiable residual estimator yields provable convergence in 1D r-refinement","Shape-optimized node movement reduces residual error with proof","r-refinement with residual estimator: shape-optimized convergence","Provably convergent mesh adaptivity via residual estimator in 1D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000884,"raw_usage":{"total_tokens":3604,"prompt_tokens":642,"completion_tokens":2962,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":386,"completion_tokens_details":{"reasoning_tokens":2874}},"tokens_in":386,"tokens_out":2962,"duration_ms":20070,"temperature":1.0,"reasoning_tokens":2874,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T05:27:13.237808+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on a one-dimensional Poisson problem with a boundary-layer profile in f and monitor the minimum element length; if an element length converges to zero (so the parametrisation leaves U) before ∥E'(Φ_n)∥ reaches the stopping tolerance, the practical convergence claim without the bounded-inverse assumption fails.","supporting_citations":[],"review_version":1}