{"id":"f7f9f32e-be4f-416c-8362-2ba0331466ac","arxiv_id":"1909.00551","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An iterative implicit B-spline fitting method, I-PIA, converges to the minimum-norm least-squares fit and empirically reconstructs curves and surfaces without spurious sheets.","lead":"This paper applies a progressive-iterative approximation (PIA) iteration to implicit B-spline curve and surface reconstruction, aiming to remove spurious zero-level sets without solving a linear system. It reports up to three orders of magnitude speedup over a total-variation regularized baseline and supports the claims with convergence analysis and numerical examples.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.4 only establishes convergence to a minimum-norm least-squares fit; it says nothing about the topology of the zero level set. The paper's headline claim that I-PIA produces no spurious sheets is therefore unsupported by the proof and rests solely on selected examples.","rationale":"The reader's conditional verdict is appropriate. My stress-test did not find a counterexample, but it did confirm that the main advertised advantage—absence of spurious sheets—is not established by the paper's mathematics. Theorem 3.4 is a standard convergence result for Richardson iteration applied to a singular least-squares problem; it is essentially correct (up to the n vs. 2n slip in Property 2, which can be repaired by replacing n with 2n). However, the theorem's conclusion is about coefficients, not about the zero set of the implicit function. The min-norm solution of B^T B X = B^T b minimizes ||X||_E, a regularization on control coefficients, not a geometric smoothness or topology penalty. Nothing in the proof prevents the resulting tensor-product B-spline from having additional zero-level components far from the data; the claim is only supported by selected examples. The rank error in Property 2 and the unproven practical μ bound are secondary but reinforce that the proof does not tightly govern the implementation. The proposed concrete test—counting zero-level components on a simple benchmark with marching squares/cubes—would settle the concern directly. If the test shows no extra components, the paper's practical claim is supported and the conditional verdict can be lifted; if it shows extra components, the central claim would need to be revised to a claim about typical behavior or a heuristic. Therefore the reader's CONDITIONAL verdict remains the right call and no change in verdict is needed.","tokens_in":13639,"tokens_out":9783,"duration_ms":94640,"concrete_test":"Take a simple benchmark curve, e.g., 200 points sampled from a 'C'-shaped or 'S'-shaped open curve, with one-sided and two-sided offset points as in §4.2, and run the published I-PIA update (18) on a uniform 30×30 bi-cubic B-spline grid until ||C^{(α+1)}−C^{(α)}|| < 1e-10. Then extract the zero level set by marching squares and count connected components; repeat at iterations 1, 5, 10. If any component lies outside a narrow tubular neighborhood of the input curve, the advertised 'no extra zero-level sets' claim is falsified. The same check can be run on a 3D point cloud (e.g., a bumpy sphere) with marching cubes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central practical claim—'No extra zero-level sets exist in the iteration of I-PIA'—is not a consequence of Theorem 3.4. That theorem proves only that the iterates C^{(α)} converge to (B^T B)^+B^T b, the minimum-Euclidean-norm solution of the normal equations B^T B X = B^T b. It gives no control over the number or location of connected components of the zero set {f=0}, either in the limit or at any finite iteration. Minimizing the coefficient norm does not imply any geometric regularity of the tensor-product B-spline function; the fit can still oscillate and create extra branches away from the data. The paper's only support is a set of favorable examples (Figs. 1, 2, 4–8) with no failure cases and no topological count. A secondary correctness issue compounds this: Property 2 states B is n×m, but the collocation matrix is 2n×m because offset points are included. Thus the claimed singularity of B^T B is not guaranteed when n < m ≤ 2n, and the practical weight μ = 2/||B^T B||_∞ is not proven to satisfy the strict bound μ < 2/λ_max(B^T B) required by Property 3. These do not invalidate Theorem 3.4 in the singular overdetermined case, but they show the proof is loose exactly where the geometric claim needs support.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an implicit progressive-iterative approximation (I-PIA) method for reconstructing implicit curves and surfaces from unorganized point clouds with associated normals. The method appends offset points to the data, sets the initial tensor-product B-spline coefficients to zero, and iterates C^{(α+1)} = (I − μB^T B)C^{(α)} + μB^T b, where B is the collocation matrix over the data-plus-offset points and b encodes zero function values at the data points and ε at the offset points. The authors prove that for zero initial coefficients the iteration converges to (B^T B)^+ B^T b, the minimum-Euclidean-norm solution of the normal equations B^T B X = B^T b, and they interpret this as solving a constrained minimization problem with a natural regularization. The paper further claims that I-PIA eliminates extra zero-level sets ('spurious sheets') without explicit minimization and reports speedups of one to three orders of magnitude over a total-variation method on a wide range of curve and surface examples.","tokens_in":13975,"tokens_out":9439,"duration_ms":83432,"significance":"The idea of adapting progressive-iterative approximation to implicit fitting is novel, and the matrix formulation is clean. The convergence theorem (Theorem 3.4) is essentially a Richardson-iteration argument and is correct provided the step-size condition is satisfied; the paper also gives a self-contained derivation rather than citing the result. The experimental coverage is broad, including holes, noise, non-uniform sampling, porous surfaces, and fine details, and the reported speedups are substantial. However, the headline geometric claim that I-PIA produces no spurious sheets is not a consequence of the convergence theorem, which says nothing about the topology of the zero level set. That claim currently rests only on selected favorable examples, and the convergence analysis contains several gaps that need to be closed. If the geometric claim can be supported by a rigorous argument or a much more systematic empirical study, and the technical issues identified below are fixed, this would be a valuable contribution to implicit reconstruction.","major_comments":[{"comment":"Theorem 3.4 establishes only that the coefficient vector C^{(α)} converges to (B^T B)^+ B^T b, the minimum-Euclidean-norm solution of the normal equations. It contains no statement about the number or location of the connected components of the zero set {f^{(α)} = 0}, either in the limit or at any finite iteration. The paper's headline claim that 'No extra zero-level sets exist in the iteration of I-PIA' is therefore not a consequence of the proved result; minimizing the coefficient norm does not by itself imply any geometric regularity of the tensor-product B-spline function. The only support offered is a favorable selection of examples (Figures 1, 2, 4-8) with no negative cases and no counts of zero-level components. Because spurious-sheet elimination is the paper's main advertised advantage, this gap is load-bearing and needs either a theorem or a far more systematic empirical study with explicit topological counts.","section":"Section 3.3, Theorem 3.4; Section 1, contribution bullet 2"},{"comment":"Property 2 states that B is n × m and concludes that B^T B is singular from m > n. But B is defined as a collocation matrix on the 2n points (n data points plus n offset points), so its size is 2n × m. The claimed singularity does not follow from m > n. For the planar examples in Table 1, 2n exceeds m (e.g., Flower has 1018 rows versus 900 columns), so B^T B may well be nonsingular. The convergence theorem itself does not require singularity, but the motivation for the constrained minimum-norm problem (20) and the interpretation of the limit as the solution of a 'regularized' problem rest on this property. The property and the surrounding discussion must be corrected.","section":"Section 3.3, Property 2; Eq. (12), Eq. (17)"},{"comment":"The practical step size μ = 2/||B^T B||_∞ is asserted to satisfy the convergence condition 0 < μ < 2/λ_max(B^T B), but no proof or reference is given for the strict inequality. For the symmetric positive semidefinite matrix B^T B, λ_max ≤ ||B^T B||_∞, so the chosen μ is only known to satisfy μ ≥ 2/λ_max when the bound is attained; the strict inequality used in Property 3 and in the proof of Theorem 3.4 may fail. The manuscript needs either a proof that ||B^T B||_∞ strictly exceeds λ_max for the specific B-spline collocation matrices used here, or a different step-size selection that is provably convergent.","section":"Remarks 3.1 and 3.2"},{"comment":"The algorithm is specified with exactly one offset point per data point (p_l = p_i + σ n_i, l = n+i, i = 1,...,n) and a right-hand side b of length 2n with entries 0 at the data points and +ε at the offset points. However, the experiments in Section 4.2 describe both inside and outside offset sets with signed values (−0.1, −0.2, 0.2, 0.1), and Figure 3 shows four offset layers. This inconsistency between the described method and the implementation used for the reported results affects reproducibility and makes it unclear which variant is being evaluated. The authors should specify the exact offset configuration used in each experiment and align the algorithm description with it.","section":"Sections 3.1 and 4.2"}],"minor_comments":[{"comment":"'we proposed' should be 'we propose', and the sentence 'extra zero-level sets or spurious sheets arise in the reconstruction process makes' is ungrammatical; please revise for clarity.","section":"Abstract and Section 1"},{"comment":"The displayed collocation matrix is internally inconsistent: the second column is B1(x1)B1(y1)B_{N_w}(z1) in the first row but B1(x2)B1(y2)B2(z2) in the second row, and the lexicographic ordering of the basis functions should be stated carefully.","section":"Eq. (17)"},{"comment":"In the surface case, the difference δ_l for l = n+i is defined using f^{(α)}(x_r, y_r, z_r), but the index r is not tied to i; this should be corrected to f^{(α)}(x_i, y_i, z_i).","section":"Section 3.2, definition of δ_l"},{"comment":"The 'Maximum Error' column reports errors for I-PIA only; to support the time comparison, the errors of the comparison method [1] should be reported as well, or a statement should clarify that the accuracy of the two methods is comparable.","section":"Table 2"},{"comment":"The formula C = ||B^T B||_∞ = max_{ij} Σ_k B_i(x_k)B_j(y_k) is not the standard definition of the induced ∞-norm and is ambiguous; the maximum should be over one index only, with summation over the other.","section":"Remark 3.1"},{"comment":"'porous surfac' is a typo for 'porous surface'.","section":"Figure 7 caption"}],"recommendation":"major_revision","confidential_remarks":"The core iterative scheme and the Richardson-iteration convergence result are sound in principle, but the paper's principal selling point — the absence of spurious sheets — is not supported by Theorem 3.4. The mismatch between the one-sided offset description in Section 3.1 and the two-sided offsets used in the experiments of Section 4.2 is a serious reproducibility issue that should be resolved before the paper can be considered further. The paper is within the scope of a geometric modelling journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a modest, mostly sound application of an existing iteration to a new problem class. The paper transfers LSPIA to implicit B-spline reconstruction, proves convergence to the minimum-norm least-squares fit, and demonstrates big speedups over one TV-regularized baseline. The proof is correct as far as it goes, but it does not do what the authors claim it does: it says nothing about the number or location of zero-level components, so the 'no spurious sheets' headline is not supported by the theorem. It rests entirely on the examples.\n\nWhat is genuinely new: PIA has not been used for implicit reconstruction before, and framing the minimum-norm solution as an implicit regularization is a reasonable idea. The experiments are fairly extensive—holes, noise, non-uniform sampling, open surfaces—and the reported speedups (one to three orders of magnitude over Liu et al.) are plausible because the iteration avoids solving linear systems. If the speedups reproduce, that is useful for practitioners.\n\nSoft spots, in order of seriousness. First, the central practical claim is over-stated. Theorem 3.4 only establishes convergence to (B^T B)^+ B^T b. Minimizing coefficient norm does not control the geometry of the zero set; a tensor-product B-spline can still oscillate and produce extra branches. The paper shows only selected models with no failure cases and no topological counts. Second, Property 2 is factually wrong about the size of B: with offset points the collocation matrix is 2n × m, not n × m. The singularity argument then only holds when m > 2n; the case n < m ≤ 2n is not covered, which is exactly where the paper needs the argument to be cleanest. Third, the practical choice μ = 2/||B^T B||_∞ is taken from prior work and not shown to satisfy the strict inequality μ < 2/λ_max(B^T B). This is likely fine numerically but the paper leaves a gap in its own convergence condition. These are addressable weaknesses, not fatal ones. The iteration is coherent, the theorem is right under its stated hypothesis, and the examples are consistent with the claim, even if not proof.\n\nWho gets value: researchers in geometric modeling and CAD who want a fast implicit fitting baseline and a clear statement of the minimum-norm convergence property. The paper deserves a serious referee, but the authors should be asked to fix Property 2, temper the no-spurious-sheets claim, add a failure analysis or at least a topological count on a test suite, and release code/data. I'd send it to review.","headline":"A correct but standard LSPIA convergence proof applied to implicit B-spline fitting; the 'no spurious sheets' claim is empirically supported but not proven, and Property 2 has a size error.","tokens_in":14493,"tokens_out":2410,"would_cite":false,"duration_ms":22206,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D17","65D10","65F20","65F10"],"pacs":[],"model":"deepseek-v4-flash","headline":"I-PIA reconstructs implicit curves and surfaces by iterating B-spline coefficients from zero, converging to the minimum-norm least-squares fit and eliminating spurious sheets without explicit regularization.","keywords":["Implicit curve and surface","Curve and surface fitting","Progressive-iterative approximation","B-spline","Zero-level sets","Spurious sheets","Moore-Penrose pseudoinverse"],"falsifier":"Take a point cloud with two well-separated closed curve components, run I-PIA to convergence, and count the connected components of the zero level set of the reconstructed function; if any component appears away from the input data, the claim that I-PIA produces no extra zero-level sets is false.","tokens_in":13422,"feed_emoji":"📐","tokens_out":6358,"duration_ms":54641,"temperature":0.7,"pith_summary":"The paper's aim is to remove the spurious sheets that appear in implicit curve and surface reconstruction without paying the usual cost of adding explicit regularization. It does this by adapting the progressive-iterative approximation idea to implicit fitting: starting from zero control coefficients, the iteration converges to the minimum-Euclidean-norm solution of the normal equations. That minimum-norm solution acts as a built-in regularizer, so extra zero-level sets do not appear during iteration. The authors demonstrate on 2D and 3D models that the resulting reconstructions are robust to inaccurate distance fields, holes, non-uniform sampling, noise, and open surfaces, and that timings improve by one to three orders of magnitude over a total-variation-regularized baseline.","feed_headline":"Iterative fitting removes spurious sheets at no extra cost","feed_subtitle":"Zero-start B-spline iterations converge to the minimum-norm fit, cutting reconstruction time by one to three orders of magnitude.","key_machinery":"The machinery is the stationary linear iteration $C^{{(α+1)}}$ = (I - μ B^T B) $C^{{(α)}}$ + μ B^T b, where B is the collocation matrix of tensor-product B-spline basis functions evaluated at the data points plus offset points generated by displacing each data point along its normal. The weight μ is chosen so that 0 < μ < 2/λ_max(B^T B), with a practical choice μ = 2 / ||B^T B||_∞. Because B^T B is positive semi-definite and singular (more coefficients than constraints), the iteration is a Richardson iteration on the normal equations; the eigen-decomposition of B^T B makes the error contract in the range of B^T B while the null-space component stays fixed. Starting from C^(0) = 0 therefore selects the projection onto the range, i.e., the minimum-norm least-squares coefficient vector.","core_discovery":"The central claim is Theorem 3.4: with zero initial coefficients C^(0)=0, the I-PIA iteration converges to (B^T B)^+ B^T b, the Moore-Penrose pseudoinverse solution of the singular least-squares system B^T B X = B^T b. Since this is the solution of minimum Euclidean norm among all solutions of the constrained minimization min ||X||_E subject to B^T B X = B^T b, the iteration solves the normally underdetermined implicit-fitting problem with an implicit regularization: it selects the smallest coefficient vector that fits the data and offset points. On the paper's account this is why no extra zero-level sets (spurious sheets) arise in the reconstruction, and why no explicit minimization or regularization step is needed; the regularizer is a byproduct of starting at zero and iterating.","pith_inferences":["Beyond the paper, a direct test of the central practical claim would be to count the connected components of the zero level set on adversarial point clouds; if any spurious component appears away from the data, the advertised advantage would fail.","The min-norm limit depends on the offset distance σ and the sign pattern assigned to offset points; varying σ likely trades fidelity against sheet suppression, a trade the paper does not quantify.","The convergence proof only uses the eigen-decomposition of B^T B, so the same iteration could be accelerated with standard stationary-iteration accelerations (e.g., Chebyshev or conjugate-gradient-like schemes) that the paper does not explore.","Because the argument does not rely on B-spline-specific properties beyond a positive semi-definite Gram matrix, the same zero-start iteration may extend to other implicit bases, such as T-splines or radial basis functions, though the paper only treats B-splines."],"forward_implications":["Each I-PIA iteration is a matrix-vector product, so no linear system is solved during reconstruction; this is the source of the reported one-to-three-orders-of-magnitude speedup over regularized solvers.","Starting from zero control coefficients selects the minimum-norm B-spline coefficient vector among all fits, which is the regularization that suppresses spurious sheets.","Since no explicit regularization term is added, the same iteration applies to both implicit curves (bivariate B-splines) and implicit surfaces (trivariate B-splines).","The method inherits robustness to inaccurate distance fields, holes, non-uniform sampling, and noise because the offset-point construction and the min-norm selection do not depend on precise distance values.","Convergence is guaranteed for any weight in (0, 2/λ_max(B^T B)), so the iteration can be tuned for speed without changing the limit."],"supporting_citations":[{"why":"Supplies the progressive-iterative least-squares framework and the practical weight choice μ used by I-PIA.","marker":"[8]"},{"why":"Gives the solvability condition for the singular normal equations used in the convergence proof.","marker":"[35]"},{"why":"Introduces the offset-point construction along normals that I-PIA uses to avoid the trivial zero solution and fix orientation.","marker":"[20]"},{"why":"Provides the total-variation-regularized baseline whose runtime the paper compares against.","marker":"[1]"},{"why":"Establishes convergence of least-squares PIA when the iterative matrix is singular, the setting I-PIA relies on.","marker":"[11]"},{"why":"Contributes the implicit B-spline surface reconstruction formulation and offset-point fitting strategy I-PIA extends.","marker":"[4]"}],"fun_headline_variants":["Zero-start iteration gives minimum-norm implicit fits","No spurious sheets: start iterative fitting at zero","Implicit surface fitting sped up by zero-init iteration","Minimum-norm regularization from starting at zero"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes, rather than proves, that the smallest-coefficient least-squares B-spline fit to the data and offset points does not contain extra loops or sheets; this is shown only by examples, not proved.","fun_headline_variants_meta":{"raw":{"variants":["Zero-start iteration gives minimum-norm implicit fits","No spurious sheets: start iterative fitting at zero","Implicit surface fitting sped up by zero-init iteration","Minimum-norm regularization from starting at zero"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000582,"raw_usage":{"total_tokens":2686,"prompt_tokens":835,"completion_tokens":1851,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":1791}},"tokens_in":451,"tokens_out":1851,"duration_ms":13157,"temperature":1.0,"reasoning_tokens":1791,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:50:29.385125+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a point cloud with two well-separated closed curve components, run I-PIA to convergence, and count the connected components of the zero level set of the reconstructed function; if any component appears away from the input data, the claim that I-PIA produces no extra zero-level sets is false.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the progressive-iterative least-squares framework and the practical weight choice μ used by I-PIA."},{"cited_title":"James, The generalised inverse, The Mathematical Gazette 62 (420) (1978) 109–114","cited_arxiv_id":null,"evidence_quote":"Gives the solvability condition for the singular normal equations used in the convergence proof."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the offset-point construction along normals that I-PIA uses to avoid the trivial zero solution and fix orientation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the total-variation-regularized baseline whose runtime the paper compares against."},{"cited_title":"The Convergence of Least-Squares Progressive Iterative Approximation with Singular Iterative Matrix","cited_arxiv_id":"1707.09109","evidence_quote":"Establishes convergence of least-squares PIA when the iterative matrix is singular, the setting I-PIA relies on."},{"cited_title":"Rouhani, A","cited_arxiv_id":null,"evidence_quote":"Contributes the implicit B-spline surface reconstruction formulation and offset-point fitting strategy I-PIA extends."}],"review_version":1}