{"id":"5120d0c4-2b5c-42c3-95c3-326c25470a2a","arxiv_id":"1908.07916","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A discrete total variation of the surface normal, equal to the discrete total mean curvature, is proposed as a shape prior that lets reconstruction recover polyhedral shapes instead of rounding them.","lead":"This paper introduces a way to measure how much a 3D mesh surface bends, by adding up the angles between neighboring flat faces, weighted by edge length. The measure is tested as a regularizer in mesh denoising and in an electrical impedance tomography problem, where it helps recover flat-sided, sharp-edged inclusions such as a cube.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 3.1 is not shown to be a well-defined Riemannian split Bregman method: its d and b subproblem variables live in shape-dependent tangent spaces but are frozen during the shape update, and no convergence analysis is supplied; Section 5 therefore does not yet demonstrate the prior's utility.","rationale":"The reader's weakest assumption targeted the lack of convergence analysis for Algorithm 3.1, which is indeed a load-bearing gap for the numerical utility claim. I sharpen this to a specific, testable issue: the shape-dependent tangent spaces for d and b are not handled consistently inside Step 4, so the algorithm as written is not a mathematically well-defined Riemannian split Bregman iteration. This is more concrete than a generic 'no convergence proof' objection. However, I do not think the concern is severe enough to reject the paper outright: the discrete total variation functional itself is a reasonable and previously used curvature measure, the identification with discrete total mean curvature is standard, and the piecewise-flat motivation is sound. The verdict should remain CONDITIONAL: the numerical demonstration needs to be supported by a correct, reproducible optimization procedure or a convergence/stationarity analysis. I therefore recommend no change to the reader's conditional verdict.","tokens_in":17645,"tokens_out":15497,"duration_ms":175627,"concrete_test":"Reproduce the fixed-connectivity mesh denoising problem (2.8) for a triangulated box with known clean vertices, and implement Algorithm 3.1 in two variants: (i) the paper's convention of freezing d and b during Step 4, and (ii) a variant that parallel transports both d and b to the current tangent space at every gradient evaluation in Step 4. Run each variant from at least three initial guesses, e.g., the sphere, the noisy box, and a random perturbation, and record the final objective value of (2.8) and the vertex positions. If the two variants produce substantially different final shapes or objectives, or if either variant terminates at widely different stationary points across initializations, then the Section 5 experiments do not demonstrate the value of the prior.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The numerical support for the central utility claim (Abstract; Section 5) rests on Algorithm 3.1 solving the nonconvex, nonsmooth problem (3.1). Two concrete problems remain. First, the augmented Lagrangian L in (3.3) is written with d_E and b_E in T_{n+_E} S^2, but Step 4 minimizes L over vertex positions while d^(k) and b^(k) are held fixed. As the vertices move inside the several gradient steps of Step 4, n+_E changes, so d^(k) and b^(k) are no longer elements of the current tangent space; b is parallel transported only after Step 4 completes, and d is never transported during the shape update. Thus the object minimized in Step 4 is not the L of (3.3) evaluated at the current iterate, and the algorithm is not a fully specified Riemannian split Bregman method. Its iterates may depend on an arbitrary ambient identification of frozen tangent vectors. Second, even if this transport issue is repaired, the paper supplies no convergence or stationarity result for the nonconvex, nonsmooth problem; Section 6 explicitly defers an analysis of the ADMM scheme to elsewhere. With hand-selected parameters, a single mesh size, and no code provided, the reconstructions in Figure 5.2 are not yet robust evidence that |n|_DTV 'can help to identify polyhedral shapes' in the claimed sense.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a discrete total variation functional for the unit normal vector field on piecewise flat surfaces, defined as the sum over edges of the geodesic distance between adjacent facet normals weighted by edge length. It claims this functional is the discrete analogue of the smooth total variation of the normal, agrees with discrete total mean curvature, admits piecewise flat minimizers, and can serve as a shape prior in geometric inverse problems. The paper analyzes simple stationary shapes, proposes a Riemannian split Bregman algorithm, and presents mesh denoising and EIT inclusion-detection experiments.","tokens_in":18027,"tokens_out":3836,"duration_ms":38792,"significance":"The functional (2.1) is natural and the connection to discrete total mean curvature via Sullivan (2005) is a real and useful observation. If the utility claim is properly supported, the paper offers an alternative to surface-area regularization that avoids rounding and shrinkage of polyhedral shapes. The paper is honest about open points, including the incomplete characterization of minimizers and the deferred convergence analysis of the ADMM scheme. However, the proof of the smooth-to-discrete limit is currently a sketch with unverified geometric assumptions, and Algorithm 3.1 has a tangent-space consistency gap that affects the interpretation of the numerical results. The central idea is promising, but the evidence presented for the main utility claim is not yet decisive.","major_comments":[{"comment":"The proof of the convergence (2.3) rests on two assumptions that are not established: that in each stripe of the transition region IE,ε the normal vector changes monotonically along the geodesic between n+E and n−E, and that the integrand on the vertex caps BV,ε has order ε−1 while the cap area has order ε2. The sentence \"can be easily evaluated as an iterated integral\" does not supply the required geometric estimates for the mollified surface. Since Theorem 2.1 underpins the claim that (2.1) is the discrete analogue of (1.1), this step needs either a rigorous proof or an explicit statement that it is to be read as a formal calculation.","section":"Section 2, Theorem 2.1"},{"comment":"Step 4 minimizes L(Ωh, d(k), b(k)) over vertex positions while d(k) and b(k) are held fixed, but these vectors are elements of T_{n+E^{(k)}} S2. As the vertices move during the gradient steps, n+E changes, so d(k) and b(k) are no longer tangent vectors at the current normal. Step 5 transports b(k) only after the shape update and d(k) is never transported during the subproblem. Consequently the object minimized in Step 4 is not the augmented Lagrangian (3.3) evaluated at the current iterate, and the algorithm is not fully specified as a Riemannian split Bregman method. This gap affects the validity of the numerical results in Section 5.","section":"Section 3, Algorithm 3.1"},{"comment":"The reconstructions are presented visually with hand-selected regularization parameters, a single mesh size, and no quantitative error metric or parameter sensitivity study. In combination with the absent convergence analysis for Algorithm 3.1 (Section 6 explicitly defers it), the experiments do not yet robustly demonstrate that |n|DTV \"can help to identify polyhedral shapes\" as claimed in the Abstract.","section":"Section 5, Figure 5.2"}],"minor_comments":[{"comment":"In the proof of Theorem 2.2, \"all facets are unilaterial triangles\" should read \"equilateral triangles\".","section":"Section 2.2.1"},{"comment":"The word \"sucessively\" should be \"successively\" in the discussion of refined meshes.","section":"Section 2.2.2"},{"comment":"The sentence \"For the surface area regularization, β|n|TV(Γ1) is replaced by γ∫Γ1 ds\" is confusingly worded, since the left-hand side refers to the prior term in (4.1) rather than to the TV functional itself.","section":"Section 5"},{"comment":"In Step 5, \"Parallely transport\" should be \"transport in parallel\" or \"Parallel transport\".","section":"Section 3, Algorithm 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is at the boundary of the journal's scope: the main novelty is an algorithmic proposal with a partial analysis and numerical demonstrations. The most serious technical concern is the tangent-space consistency gap in Algorithm 3.1, which currently prevents the reconstructions from being interpreted as outputs of the stated augmented Lagrangian. If the authors can fix or clarify the algorithm and strengthen the proof of Theorem 2.1 or appropriately weaken its claim, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take: the paper does not claim a brand new functional—it is openly identified as discrete total mean curvature (Sullivan 2005; Wu et al. 2015; Pellis et al. 2019). The actual new material is the mollification consistency theorem, the stationarity analysis, and the first application of this functional as a prior in a PDE-constrained inverse problem. That is a legitimate and useful contribution.\n\nThe paper does several things well. It is honest about prior work. Theorem 2.1, while not fully proved, gives a plausible argument that the edge-based sum is the limit of the smooth TV of the normal for mollified surfaces; the decomposition into edge strips and vertex caps is reasonable, though the monotone-geodesic and O(ε) vertex-cap estimates are asserted rather than derived. Theorem 2.2 is a solid, explicit stationarity check for the icosahedron and cube, including the non-smooth case via Clarke derivatives. The EIT experiment is a meaningful demonstration: the functional really does preserve flat faces and sharp edges, unlike surface area regularization.\n\nThe soft spots are real. The proof of Theorem 2.1 rests on two unverified assumptions about the mollification geometry; a referee should ask for details. More serious is Algorithm 3.1. During the vertex update in Step 4, d^(k) and b^(k) are frozen vectors in the tangent spaces of the previous shape. As vertices move, those tangent spaces change, so the augmented Lagrangian (3.3) is not well-defined at the current iterate without a parallel transport that is only performed for b, after the update, and never for d. The algorithm as written is therefore not a fully specified Riemannian split Bregman method; the Step 4 subproblem is ambiguous. No convergence analysis is supplied, and the paper explicitly defers it. With no code and hand-picked regularization parameters, the reconstructions in Figure 5.2 are suggestive, not robust evidence.\n\nNone of this kills the paper. The central mathematical content—that this functional is a natural discrete analogue and that it promotes piecewise flat shapes—holds up. But the numerical utility claim needs a well-defined algorithm, and the experiments need more systematic parameter studies or code. This paper is for researchers in shape optimization, geometric inverse problems, and discrete differential geometry. It deserves a serious peer review, but with major revision expected.\n\nMy recommendation: send it to review. Ask the authors to fix the algorithm specification, add details to Theorem 2.1's proof, and provide code or more thorough numerical validation.","headline":"The paper's genuine novelty is the mollification limit and the first PDE-inverse-problem application of discrete total mean curvature; the numerical algorithm has a specification gap that needs addressing.","tokens_in":18536,"tokens_out":4064,"would_cite":true,"duration_ms":37711,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q10","65N21","65K10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces a discrete total variation of the normal-vector field on triangulated surfaces, proves it equals discrete total mean curvature, and shows it recovers polyhedral shapes in PDE-constrained inverse problems where…","keywords":["total variation of the normal","discrete total mean curvature","shape prior","split Bregman iteration","Riemannian optimization","geometric inverse problem","electrical impedance tomography","piecewise flat shapes"],"falsifier":"Construct the noise-free cube-reconstruction experiment from an initial sphere and track both the objective (3.1) and the Bregman residual $\\|d-\\log n-b\\|$ across iterations; if the iteration stops at a rounded shape with positive dihedral angles, or if the residual fails to decrease, the paper's claim that the prior helps identify polyhedral shapes is not supported by its own algorithm.","tokens_in":17471,"feed_emoji":"🧊","tokens_out":9276,"duration_ms":83466,"temperature":0.7,"pith_summary":"The paper proposes the edge-based sum $|n|_{\\mathrm{DTV}}(\\Gamma_h)=\\sum_E d(n^+_E,n^-_E)\\,|E|$, the geodesic distance between adjacent facet normals weighted by edge length, as the discrete substitute for the smooth total variation of the normal. It proves this functional coincides with the discrete total mean curvature and shows analytically that it becomes the smooth TV in the limit of mollified surfaces; numerically it stays within a factor of $\\sqrt{2}$ of the smooth value on spheres, suggesting an anisotropic limit. Used as a regularizer in shape optimization of type (1.3), the functional penalizes curvature concentrated at edges rather than surface area, so its minimizers can be piecewise flat. A Riemannian split-Bregman algorithm is developed, and experiments on mesh denoising and on an electrical-impedance-tomography inclusion problem show that polyhedral objects—flat sides and sharp edges—are reconstructed accurately, in contrast to surface-area regularization which rounds and shrinks the result.","feed_headline":"Facet-angle prior reconstructs polyhedral inclusions","feed_subtitle":"Unlike surface-area priors, this regularizer favors flat facets and sharp edges in PDE-based inverse problems.","key_machinery":"The load-bearing object is the edge functional $|n|_{\\mathrm{DTV}}(\\Gamma_h)=\\sum_E \\arccos((n^+_E)^\\top n^-_E)\\,|E|$, an $\\ell^1$-type sum of exterior dihedral angles weighted by edge length; it doubles as the discrete total mean curvature. Theorem 2.1 justifies it as the limit of smooth normal-TV under mollification, and numerical evidence on spheres indicates it tracks the anisotropic integral $\\int_\\Gamma(|k_1|+|k_2|)\\,ds$. On the optimization side, the paper adapts the split Bregman/ADMM scheme to the sphere: the jump across each edge is represented by the logarithmic map $d_E=\\log_{n^+_E} n^-_E$ in the tangent space $T_{n^+_E}S^2$, multipliers are carried between iterates by parallel transport, and the $d$-update reduces to an explicit vectorial shrinkage step (3.4). This machinery is what makes the nonconvex, nonsmooth shape problem (3.1) tractable numerically.","core_discovery":"The central claim is that the functional $|n|_{\\mathrm{DTV}}(\\Gamma_h) = \\sum_E d(n^+_E, n^-_E)\\,|E|$, formed by summing, over every edge, the geodesic distance between the two adjacent facet normals times the edge length, is the discrete counterpart of the smooth total variation of the normal. The paper proves that this sum coincides exactly with the discrete total mean curvature and that, for a family of mollified smooth surfaces converging to $\\Gamma_h$, the smooth total variation of the normal converges to $|n|_{\\mathrm{DTV}}(\\Gamma_h)$ as the smoothing width goes to zero. As a shape prior it penalizes curvature concentrated at edges rather than surface area, so flat facets and sharp creases become admissible optimizers; numerical experiments on mesh denoising and on an electrical-impedance-tomography problem show that a cube-shaped inclusion is reconstructed with flat lateral faces and sharp edges, whereas surface-area regularization smooths and shrinks the result. For spheres the discrete value exceeds the smooth one by a factor close to $\\sqrt{2}$, which the paper conjectures reflects convergence to $\\int_\\Gamma (|k_1|+|k_2|)\\,ds$ rather than $\\int_\\Gamma\\sqrt{k_1^2+k_2^2}\\,ds$.","pith_inferences":["A natural testable extension is an isotropic variant that measures the joint dihedral angles around each vertex; if it recovers the smooth value on spheres for all meshes, it would eliminate the $\\sqrt{2}$ artifact while keeping piecewise-flat minimizers.","If Algorithm 3.1's convergence were established, the same prior could be applied to other severely ill-posed geometric inverse problems—crack detection, cavity imaging, interface reconstruction—where the ground truth is known to be piecewise flat.","The connectivity dependence of the minimizers (cube with crossed diagonals is stationary) suggests that using this prior with a fixed triangulation biases the reconstruction toward that triangulation's flat patterns; treating connectivity as a design variable could change the recovered shapes.","The prior's effectiveness likely depends on the ratio of the Bregman penalty to the regularization strength and on the number of inner gradient steps; a systematic parameter study on synthetic polyhedral targets would provide a practical calibration rule the paper does not give."],"forward_implications":["Shape optimization problems (1.3) gain a prior that explicitly permits flat facets and sharp edges, so polyhedral inclusions in PDE-constrained inverse problems can be recovered without the rounding produced by surface-area terms.","Because $|n|_{\\mathrm{DTV}}$ equals discrete total mean curvature, results from discrete differential geometry—such as the stationarity of the icosahedron and crossed-diagonal cube under an area constraint—transfer directly to the prior.","The Riemannian split-Bregman algorithm extends ADMM to manifold-valued data and, as the paper notes, can be applied to other nonsmooth total-variation problems with values in $S^2$ or other manifolds.","Uniform scaling of $\\Gamma_h$ scales $|n|_{\\mathrm{DTV}}$ linearly, so any unconstrained use of the prior must be paired with an area constraint or a tracking term to prevent collapse to a point.","The observed $\\sqrt{2}$ gap on spheres points to the conjecture that refined-mesh discrete values converge to the anisotropic measure $\\int_\\Gamma (|k_1|+|k_2|)\\,ds$, implying the discrete prior is an $\\ell^1$ analogue of the smooth $\\ell^2$ functional."],"supporting_citations":[{"why":"Defines the smooth total variation of the normal (1.1) that the discrete functional is designed to approximate, and introduces the Riemannian split-Bregman approach this paper adapts.","marker":"Bergmann, Herrmann, et al., 2019"},{"why":"Supplies the split Bregman/ADMM template whose manifold-valued generalization is Algorithm 3.1.","marker":"Goldstein, Osher, 2009"},{"why":"Establishes $H_E = |E|\\Theta_E$ as discrete total mean curvature, letting the paper identify (1.2) with it.","marker":"Sullivan, 2005"},{"why":"Used the same edge sum as a mesh-denoising regularizer with approximated distances, showing prior use in a different application and motivating the geodesic-distance version here.","marker":"Wu et al., 2015"},{"why":"Provides the observation that optimal meshes have zero dihedral angles and that minimizers depend on connectivity, supporting the piecewise-flat minimizer claim.","marker":"Pellis et al., 2019"},{"why":"Supplies the cotangent-formula gradient (2.7) used in the proof that the icosahedron and crossed-diagonal cube are stationary under an area constraint.","marker":"Crane et al., 2013"}],"fun_headline_variants":["Edge-based curvature prior recovers polyhedra","Normal-vector TV prior preserves sharp creases","Discrete mean curvature as shape prior for 3D inversion","Flat-facet prior: total variation of normals","Curvature-at-edges prior for geometric inverse problems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The numerical demonstrations assume the proposed Riemannian split-Bregman iteration, for which the paper supplies no convergence analysis, actually reaches a useful minimizer of the nonconvex, nonsmooth problem (3.1).","fun_headline_variants_meta":{"raw":{"variants":["Edge-based curvature prior recovers polyhedra","Normal-vector TV prior preserves sharp creases","Discrete mean curvature as shape prior for 3D inversion","Flat-facet prior: total variation of normals","Curvature-at-edges prior for geometric inverse problems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000139,"raw_usage":{"total_tokens":1173,"prompt_tokens":974,"completion_tokens":199,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":122}},"tokens_in":590,"tokens_out":199,"duration_ms":3396,"temperature":1.0,"reasoning_tokens":122,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:22:00.192769+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct the noise-free cube-reconstruction experiment from an initial sphere and track both the objective (3.1) and the Bregman residual $\\|d-\\log n-b\\|$ across iterations; if the iteration stops at a rounded shape with positive dihedral angles, or if the residual fails to decrease, the paper's claim that the prior helps identify polyhedral shapes is not supported by its own algorithm.","supporting_citations":[],"review_version":1}