{"id":"ce263891-9cd8-424c-a216-b443ea84acc8","arxiv_id":"1908.06154","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A 3D circle average is defined for point-normal pairs and plugged into four classical subdivision schemes, enabling shape variations by editing initial normals.","lead":"This paper defines a 3D average of point-normal pairs and uses it to build new subdivision schemes that refine surfaces. Because the initial normals can be edited, the modified schemes create a wider family of shapes from one mesh.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Modified interpolatory schemes are undefined as written: §5.1 binary averaging of negative-mask schemes yields 3D-circle-average weights outside [0,1], where §3.2 does not define the average.","rationale":"The reader's weakest assumption, absence of a convergence proof, is valid and explicitly conceded in Section 6.2. My stress-test identifies a more elementary, testable gap that the reader did not flag: the repeated-binary-averaging rewrite in Section 5.1 produces binary weights outside [0,1] whenever the original mask has negative coefficients. Since the 3D circle average is defined only for ω∈[0,1], the modified Butterfly and K4 schemes are not actually specified by the paper. This does not necessarily destroy the central editing-capability claim for positive-mask schemes such as MLP/MCC, but it does undermine the paper's broader claim to modify any convergent linear scheme and the credibility of the MK4/MBY experiments. Both concerns are fixable in principle—by extending the circle average to the required weight range and by supplying a convergence/proximity proof—so the reader's CONDITIONAL verdict remains appropriate rather than a rejection. The concrete test on the released code would settle whether the negative-weight issue is real or already handled implicitly in the implementation.","tokens_in":9712,"tokens_out":13593,"duration_ms":148572,"concrete_test":"Run the released implementation on the K4 stencil and log every weight argument passed to the circle-average routine during one MK4 refinement step. If any logged weight lies outside [0,1], the construction is undefined as specified and the MK4/MBY results in Figures 8–9 must be reproduced with an explicit extension of the average. If all weights are in [0,1], this concern is refuted and the conditional verdict should rest solely on convergence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The general method claims to modify any convergent linear scheme by rewriting its refinement rule as repeated weighted linear binary averages and replacing each with the 3D circle average. For K4/BY, whose masks contain negative coefficients, the rewrite in §5.1 cannot keep the binary weights in [0,1]. Example: for the K4 stencil q=(9/16)p0+(9/16)p1-(1/16)p2-(1/16)p3, the first binary average B=(p0+p1)/2 is fine, but the next required binary average combines B and p2 with weights 18/17 and -1/17; the 3D circle average is defined only for ω∈[0,1] (Section 3.1). The reordering in §5.1 only avoids zero denominators; it does not keep weights in range. Consequently MK4/MBY are not well-defined by the paper, and the empirical results in §6.2.2 and the claim 'any converging linear scheme' are unsupported. Independently, convergence of the point iterates is not proven—Section 6.2 explicitly disclaims it—so even the positive-mask schemes are conditional.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 3D extension of the authors' earlier 2D 'circle average' for point-normal pairs, defines the 3D average by reducing it to a 2D average in a plane and then translating along the direction n0×n1, and proves several properties: consistency (Eq. (2)), a helix trace, limit behavior as θ→0 or φ→0, and preservation of spheres and of a helical curve on a cylinder. It then modifies four linear subdivision schemes — Catmull-Clark, Kobbelt 4-point, Butterfly, and Loop — by rewriting their refinement rules as repeated binary averages and replacing each binary average with the 3D circle average. A method for assigning 'naive' normals to vertices of a mesh without normals is given. The paper reports numerical comparisons of dihedral angles and discrete curvature, demonstrates editing of initial normals on a fox mesh, and provides code and videos. The authors explicitly state in Section 6.2 that they do not have a convergence proof for the modified point positions.","tokens_in":9941,"tokens_out":5953,"duration_ms":62418,"significance":"If the construction is well-defined and convergent, the editing capability is a real and useful contribution to subdivision surface design: fixing the control mesh and rotating initial normals yields a family of limit surfaces, as demonstrated in Section 6.3. The paper is also honest and concrete: it gives explicit formulas, includes a GitHub repository with code and mesh files, and provides videos. However, the central claim is currently conditional in two ways. First, the modification of schemes with negative mask coefficients is undefined as written, because the required binary weights leave the interval [0,1] on which the circle average is defined. Second, convergence of the point iterates is not proven and is explicitly disclaimed in Section 6.2. These issues do not destroy the value of the positive-mask schemes, but they do mean that the broad claims in Section 1.1 and in the empirical sections are not supported by the manuscript as it stands.","major_comments":[{"comment":"The rewriting method in Section 5.1 does not keep binary weights in [0,1] when the original mask coefficients have mixed signs, but the 3D circle average is defined only for weights ω∈[0,1] in Section 3.1. For the K4 rule q=(9/16)(p0+p1)-(1/16)(p2+p3), the first binary average B=(p0+p1)/2 is valid, but the next required binary average combines B and p2 with weights 18/17 and -1/17, which are outside the allowed range. Reordering coefficients so that partial sums are positive avoids zero denominators but does not fix the out-of-range weights. Consequently MK4 and MBY are not defined by the manuscript, and the empirical results in Section 6.2.2, Table 1 for MK4, and the claim in Section 1.1 that the method can modify 'any convergent linear scheme' are unsupported.","section":"§5.1, §3.1"},{"comment":"The paper states: 'Although we do not have a convergence proof for the modified schemes investigated in this paper, our tests indicate that the generated meshes converge to a surface.' Since the central contribution is that the modified schemes 'generate more geometries' and are 'surface-generating', this missing proof is load-bearing. The cited convergence of normals via manifold-valued geodesic averaging covers only the normal component and does not imply convergence of the point component. To place the main claim on a sound footing, the authors need either a convergence proof (for example by a proximity argument relative to the underlying linear scheme) or an explicit reframing of the convergence statement as a conjecture supported by experiments.","section":"§6.2"},{"comment":"The monotonic decrease of the angles ξi is measured against the 'naive normal in the final mesh' at the last refinement level, not against the exact normal of the limit surface. This is an empirical proxy, and the conclusion that 'the closer are the initial normals to the naive normals, the closer are the limit normals to the normals of the limit surface' is therefore only as strong as the proxy. This should be stated explicitly, especially because no convergence proof is available for the point positions.","section":"§6.3, Table 2"}],"minor_comments":[{"comment":"The word 'Fugure' in 'as is demonstrated in Fugure 8' is a typo and should be corrected.","section":"§6.2.2"},{"comment":"The example 'x=cost, y=sint, z=t, 0≤t≤π' is a helix on a cylinder, not a cylinder surface; calling this 'preservation of cylinders' is misleading and the terminology should be clarified, e.g., 'preservation of circular helices on cylinders'.","section":"§4.4"},{"comment":"The manuscript notes that the performance of a modified scheme is 'almost independent of the order of the summands', but it does not state whether different admissible orders define the same modified scheme in general. Since the circle average is nonlinear, a canonical ordering should be specified in the definition of each modified scheme.","section":"§5.1"},{"comment":"Reference [7] is listed as 'Mathematics of Computation, 1:1–2, 2016', which appears to be an incomplete or incorrect volume/pagination citation; please verify it.","section":"References"},{"comment":"The definition of naive normals uses a_i from cross products of edge vectors; the case where two consecutive edge vectors are collinear is not discussed, and a fallback should be mentioned for robustness.","section":"§5.3"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about negative weights is confirmed by the manuscript text: Section 5.1's algorithm genuinely produces binary weights outside [0,1] for the K4 and Butterfly masks, and Sections 3.1 and 3.2 define the circle average only for ω∈[0,1]. I do not recommend rejection because the 3D circle average itself, the positive-mask modified schemes, and the editing demonstrations are a viable core contribution; the paper needs to either extend the average to the required weight range or restrict all claims to schemes whose decompositions stay inside [0,1], and it needs to clearly separate proven convergence from empirical evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The 3D circle average is a real extension of your earlier 2D construction, and the thought of editing limit surfaces by rotating initial normals is appealing. The consistency argument and the sphere/cylinder preservation results are neat, and you are honest that convergence is not proved. The code, meshes, and videos are a plus.\n\nThe load-bearing problem is in Section 5.1. You claim your rewriting method applies to any convergent linear scheme, but for K4 and BY the mask has negative coefficients. After the first binary average, the next weight is outside [0,1] — for the K4 stencil, the second step needs weights 18/17 and -1/17. The 3D circle average is defined only for weights in [0,1]. Reordering positive coefficients before negative ones avoids division by zero, but it does not keep weights in range. So MK4 and MBY are not defined by the paper as written, and the empirical results in Section 6.2.2 rest on an operation that hasn't been specified. That's not a minor gap; it cuts the claim \"any converging linear scheme\" down to schemes with nonnegative masks.\n\nEven for the positive-mask schemes, convergence is an assumption, not a result. You say the normals converge via known manifold-valued theory, but that doesn't help if the point iterates don't converge. The consistency proof in Section 4.1 is a plausibility argument rather than a full proof, and the numerical tables lack error bars — minor, but they make the comparisons less convincing.\n\nWho should read this? People in CAGD who work with point-normal subdivision and want to see what editing-by-normals can do. The 3D average itself is a legitimate contribution. But the paper needs a serious revision: either restrict the modification claim to schemes with nonnegative masks, or extend the circle average to handle the negative-weight case explicitly. A convergence proof, or at least a clearly stated conjecture with supporting evidence, would help. As it stands, I'd send it to peer review, but the referee should push hard on the undefined interpolatory schemes and the convergence gap.","headline":"A genuinely new 3D circle average and a nice editing idea, but the paper's own methodology makes the modified interpolatory schemes (MK4, MBY) undefined, because negative mask coefficients push binary weights outside [0,1].","tokens_in":660,"tokens_out":892,"would_cite":false,"duration_ms":36190,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D17"],"pacs":[],"model":"deepseek-v4-flash","headline":"Refining point-normal pairs with a 3D circle average turns a fixed control mesh into a family of limit surfaces that can be edited by rotating initial normals.","keywords":["surface-generating subdivision","3D point-normal pairs","3D circle average","surface design","subdivision schemes","normal editing"],"falsifier":"Run one of the modified schemes on a mesh with naive normals for many refinement levels and check whether the maximal dihedral angle and the displacement between successive refined meshes decay to zero; if they oscillate or fail to decay, the modified scheme does not generate a limit surface and the editing claim loses its foundation.","tokens_in":9539,"feed_emoji":"🌀","tokens_out":10943,"duration_ms":93048,"temperature":0.7,"pith_summary":"This paper claims that any surface-generating linear subdivision scheme can be made to refine point-normal pairs instead of bare points, by rewriting its refinement rules as repeated weighted binary averages and replacing each linear average with a newly constructed 3D circle average. In the resulting modified schemes, the control mesh can stay fixed while the initial normals are edited, and the limit surface changes accordingly, giving designers a shape handle that point-only subdivision lacks. The paper proves the 3D circle average is a genuine average with a consistency property, shows that it preserves spheres and cylinders, and demonstrates modified versions of four classical schemes on test meshes. The authors report numerical evidence that the modified meshes converge to surfaces, while explicitly noting that a convergence proof is not provided.","feed_headline":"One mesh, many surfaces: editing normals changes the limit surface","feed_subtitle":"Subdivision refined on point-normal pairs morphs the limit surface without moving a single mesh vertex.","key_machinery":"The load-bearing object is the 3D circle average $\\circledast_\\omega$ of two point-normal pairs, a weighted binary operation built by reducing the construction to a plane: with weight $\\omega$, the point $p_1$ is projected onto the plane through $p_0$ with normal $n_0\\times n_1$, the 2D circle average is evaluated there, and the resulting point is projected onto the parallel plane at distance $\\omega\\hbar$ from $\\Pi_0$. This operation is the substitute for every weighted linear binary average in a rewritten subdivision rule, so the normal of each pair influences where the refined point lands. Its consistency property, $(P_0\\circledast_t P_1)\\circledast_k (P_0\\circledast_s P_1) = P_0\\circledast_{\\omega^*}P_1$ with $\\omega^*=ks+(1-k)t$, is what makes the nonlinear iteration behave like an average, and the limit-case analysis shows that when normals are parallel the average degenerates to the linear average, which is why the modified schemes are expected to match the smoothness of their linear counterparts.","core_discovery":"The paper's central claim is that the limit surface of a modified subdivision scheme depends on the initial normals as well as on the control mesh, so a single scheme with one mesh can generate a continuum of geometries by rotating those normals. The carrier of this claim is the 3D circle average $P_0 \\circledast_\\omega P_1$: project $p_1$ into the plane through $p_0$ normal to $n_0\\times n_1$, apply the 2D circle average there, and lift the result to the parallel plane at fraction $\\omega$ of the spacing. The paper establishes the consistency identity that makes the operation an average, shows that it reduces to the linear average as the normal angle or the chord angle tends to zero, and demonstrates sphere and cylinder preservation. Modified Catmull-Clark, Kobbelt 4-point, Butterfly, and Loop schemes are tested; with naive normals computed from the mesh they show smoother curvature estimates and faster decay of dihedral angles, and the modified interpolatory schemes avoid the self-intersections seen in their linear counterparts. The paper also states that the limit normals generated by the modified schemes are not the normals of the limit surface, and that convergence of the schemes is supported by tests but not yet proved.","pith_inferences":["Because the 3D construction is described for any 2D average of point-normal pairs, the lifting procedure could be applied to other binary averages to create additional editable subdivision schemes.","The monotone decrease of the angle between generated normals and mesh normals as initial normals approach the naive normals suggests using the initial normals as optimization variables, fitting them to steer the limit surface toward a target curvature distribution.","Since the limit normals are not the surface normals, applications that consume normals, such as rendering, shading, and offsetting, would need to recompute them from the refined mesh; designing a binary average whose limit normals coincide with the surface normals is the natural next step.","The helix trace of the average indicates that editing normals twists the interpolated paths around the chords, which could be exploited as a fairing or shape-morphing tool in interactive design."],"forward_implications":["A fixed control mesh yields a family of limit surfaces indexed by the initial normals, so shape editing gains a new degree of freedom alongside vertex positions.","Every convergent linear scheme whose rules can be rewritten as repeated weighted binary averages admits a modified normal-editing version, so the construction extends beyond the four schemes tested.","With naive normals estimated from the mesh, the modified approximating schemes produce smoother discrete-curvature estimates and a faster decrease of maximal dihedral angles than the linear schemes on homogeneous meshes.","The modified interpolatory schemes avoid the self-intersections that the linear Kobbelt 4-point and Butterfly schemes exhibit on meshes with uneven edge lengths.","If a modified scheme converges, its limit surface is conjectured to have the same smoothness as the corresponding linear scheme, because the circle average tends to the linear average as the angle between normals goes to zero."],"supporting_citations":[{"why":"Introduces the 2D circle average of point-normal pairs and its consistency and circle-preservation properties, which the 3D construction extends.","marker":"[11]"},{"why":"Supplies the general convergence result for manifold-valued subdivision schemes based on geodesic averages, invoked for the normal component.","marker":"[7]"},{"why":"Provides the neighborhood-averaging formula used for naive normals and the discrete curvature and area estimates used in comparisons.","marker":"[17]"},{"why":"Documents self-intersections of four-point subdivision on polygons with uneven edge lengths, used to explain the K4 and Butterfly artifacts.","marker":"[5]"},{"why":"Defines the Catmull-Clark linear subdivision scheme that the modified MCC scheme is built from.","marker":"[2]"},{"why":"Defines the Kobbelt 4-point interpolatory scheme that the modified MK4 scheme is built from.","marker":"[10]"},{"why":"Defines the Butterfly interpolatory scheme that the modified MBY scheme is built from.","marker":"[6]"},{"why":"Defines the Loop approximating scheme that the modified MLP scheme is built from.","marker":"[15]"}],"fun_headline_variants":["Point-normal subdivision: rotate normals to redesign the surface","New 3D average for point-normal pairs widens subdivision design","Edit normals, keep vertices: subdivision schemes morph surfaces","Subdivision schemes now refine point-normal pairs for more geometry","Rotate normals to morph the limit surface without moving vertices"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that repeated application of a modified scheme settles down to a surface; the paper provides test evidence for this but no proof, and the normal component depends on a general geodesic-average convergence result.","fun_headline_variants_meta":{"raw":{"variants":["Point-normal subdivision: rotate normals to redesign the surface","New 3D average for point-normal pairs widens subdivision design","Edit normals, keep vertices: subdivision schemes morph surfaces","Subdivision schemes now refine point-normal pairs for more geometry","Rotate normals to morph the limit surface without moving vertices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001153,"raw_usage":{"total_tokens":4782,"prompt_tokens":955,"completion_tokens":3827,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":3743}},"tokens_in":571,"tokens_out":3827,"duration_ms":27754,"temperature":1.0,"reasoning_tokens":3743,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:53:37.356485+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run one of the modified schemes on a mesh with naive normals for many refinement levels and check whether the maximal dihedral angle and the displacement between successive refined meshes decay to zero; if they oscillate or fail to decay, the modified scheme does not generate a limit surface and the editing claim loses its foundation.","supporting_citations":[{"cited_title":"A weighted binary average of point- normal pairs with application to subdivision schemes","cited_arxiv_id":null,"evidence_quote":"Introduces the 2D circle average of point-normal pairs and its consistency and circle-preservation properties, which the 3D construction extends."},{"cited_title":"A global approach to the refinement of manifold data","cited_arxiv_id":null,"evidence_quote":"Supplies the general convergence result for manifold-valued subdivision schemes based on geodesic averages, invoked for the normal component."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the neighborhood-averaging formula used for naive normals and the discrete curvature and area estimates used in comparisons."},{"cited_title":"Floater, and Kai Hormann","cited_arxiv_id":null,"evidence_quote":"Documents self-intersections of four-point subdivision on polygons with uneven edge lengths, used to explain the K4 and Butterfly artifacts."},{"cited_title":"Recursively generated b-spline surfaces on arbitrary topological meshes","cited_arxiv_id":null,"evidence_quote":"Defines the Catmull-Clark linear subdivision scheme that the modified MCC scheme is built from."},{"cited_title":"Interpolatory subdivision on open quadrilateral nets with arbitrary topology","cited_arxiv_id":null,"evidence_quote":"Defines the Kobbelt 4-point interpolatory scheme that the modified MK4 scheme is built from."},{"cited_title":"Gregory, and David Levin","cited_arxiv_id":null,"evidence_quote":"Defines the Butterfly interpolatory scheme that the modified MBY scheme is built from."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Loop approximating scheme that the modified MLP scheme is built from."}],"review_version":1}