{"id":"7b79cb6b-9408-4f11-9b87-b91531050d7a","arxiv_id":"1908.06417","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"A new 'with memory' variant of the LSPIA method accelerates least-squares B-spline curve and surface fitting by reusing previous iterates, with a proven faster convergence rate for totally positive bases.","lead":"This paper introduces MLSPIA, an iterative method that speeds up least-squares fitting of curves and surfaces by remembering previous iteration steps. It proves the method converges and is faster than the standard LSPIA method, and demonstrates the speedup on five example data sets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 7 is the load-bearing gap: the surface convergence and acceleration promised in the abstract are asserted without proof, and the text says only that the proof is 'similar' to Theorems 2 and 6.","rationale":"After checking the curve-case algebra (Lemma 1, the H matrix in (10), the SVD reduction (14), Lemma 5, and the eigenvalue computation in Theorem 6), I find no error in the main curve result. The numerical comparison protocol is fair, and the reported large speedups are plausible given the proven rate. The genuine weakness is the omitted proof of Theorem 7. The reader's weakest_assumption identifies exactly this gap. The Kronecker facts needed are standard and I expect the theorem can be repaired without changing the rate formula, but the manuscript as submitted does not contain that argument. The placeholder '[ ? ]' in Section 2 and the abstract's 'the' least-squares curve in the rank-deficient case are minor wording and reference issues; the latter is harmless because the fitted curve B P∞ is unique even when P∞ is not. Thus the reader's CONDITIONAL verdict stands.","tokens_in":19091,"tokens_out":17230,"duration_ms":168728,"concrete_test":"Derive the vectorized surface iteration explicitly: write (33)-(35) as (Λ^{k+1}; P^{k+1}) = H_{ω,γ,υ}(B1⊗B2)(Λ^k;P^k) + C with H as in (10), set W = diag(U1⊗U2, V1⊗V2) from the SVDs B1=U1Σ1V1^T, B2=U2Σ2V2^T, and compute W^T H W. Verify that it has the claimed block-diagonal form with coupled 2×2 blocks [(1-ω)-γυ(σ_iμ_j)^2, -ωσ_iμ_j; υσ_iμ_j, 1] for i=1..r, j=1..s, plus scalar blocks 1-ω and 1. Then substitute ω*, γ*, υ* from (37) and confirm the spectral radius is (σ1μ1-σrμs)/(σ1μ1+σrμs). If the SVD does not decouple this way, Theorem 7 needs correction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5 states Theorem 7, extending convergence and the rate formula to tensor-product surfaces, with the sentence: 'The proof to the Theorem 7 is omitted since it is similar to the case for Theorems 2 and 6.' This is an explicit omission of the surface half of the abstract's central claim. The surface iteration (33)-(35) is not literally the curve iteration: it acts on vectorized data and control points through the Kronecker product B1⊗B2, and the proof must establish that after the SVD of B1⊗B2 the iteration matrix decouples into a 2rs-dimensional coupled block, an (m1m2-rs)-dimensional diagonal block, and an (n1n2-rs)-dimensional identity block. It must also show the extremal singular values of B1⊗B2 are σ1μ1 and σrμs. These are standard facts but they are not written down; if the decoupling failed for deficient-rank factors, the stated rate could be wrong. Because the abstract, the conclusion, and Example 5 all rely on the surface result, this gap is load-bearing for the paper's full claim, even though the curve-case analysis itself appears sound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MLSPIA, a progressive and iterative approximation method with memory for least-squares fitting. The iteration (1)-(3) uses three weights to update control points and accumulated difference terms. For curves, the authors prove (Theorem 2) that for any normalized totally positive basis, even when the collocation matrix B has deficient column rank, the iteration converges to a least-squares fit provided the weights satisfy condition (11). They then derive a spectral-radius estimate at a particular weight choice (Theorem 6) and show that this rate is strictly smaller than the optimal LSPIA rate from [7] whenever the relevant extremal singular values differ. A tensor-product surface extension is stated as Theorem 7, and five numerical examples, including one surface example, illustrate the acceleration. The proof of Theorem 7 is explicitly omitted, although the abstract, the conclusions, and Example 5 rely on the surface result.","tokens_in":19197,"tokens_out":9765,"duration_ms":97918,"significance":"If the curve-case theorems are correct, the paper makes a useful contribution: it extends LSPIA convergence to rank-deficient collocation matrices, provides a constructive weight choice with a rigorously faster theoretical rate, and the numerical examples confirm the acceleration with negligible final deviations. The eigenvalue analysis in Lemma 3 and the discriminant computation leading to (29) are sound, and the treatment of the rank-deficient null-space block in Theorem 2 is careful. The omitted surface proof is the main weakness: because the surface theorem is a central advertised claim, the manuscript is not yet complete. A full proof of Theorem 7 is required before the surface claims can be accepted; the curve part alone would be a modest but valid contribution.","major_comments":[{"comment":"The proof of the surface extension is explicitly omitted with the sentence 'The proof to the Theorem 7 is omitted since it is similar to the case for Theorems 2 and 6.' This is a load-bearing gap. The surface iteration (33)-(35) acts through the Kronecker product B1⊗B2, and the proof must establish the singular-value factorization of B1⊗B2, the corresponding block structure of the iteration matrix (a 2rs-dimensional coupled block, an (m1m2-rs)-dimensional diagonal block, and an (n1n2-rs)-dimensional identity block), and the rate formula ρ = (σ1μ1 - σrμs)/(σ1μ1 + σrμs). These are standard facts, but they are not written down, and the abstract, the conclusions, and Example 5 all rely on Theorem 7. Please supply a complete proof, or explicitly restrict the paper's central claims to curves.","section":"Section 5, Theorem 7"},{"comment":"The statement 'When the NTP basis is linearly independent, B is of full column rank' is not true in general. A normalized totally positive basis can be linearly independent as a set of functions while its collocation matrix at the chosen parameters is rank-deficient; for example, a B-spline basis with no data parameter in the support of one basis function yields a zero column. The comparison with the LSPIA rate from [7] is valid under the assumption that B has full column rank. Please restate item 2 in those terms and remove the false implication.","section":"Theorem 6(2) and its proof"}],"minor_comments":[{"comment":"The inequalities in (11) and (36) appear to be missing division signs: the lower bound should read ω - ω/(σ1^2 υ) and the upper bound ω/2 - (ω-2)/(σ1^2 υ); as typeset, the expressions are dimensionally inconsistent. The same correction applies to (36) with σ1μ1 in place of σ1.","section":"Equations (11) and (36)"},{"comment":"The caption of Figure 5 says '(b) Step 1000', while the text says the second and third panels are at steps 200 and 362, and Table 2 reports 362 iterations. Please reconcile these numbers.","section":"Figure 5 and Section 6.3"},{"comment":"The symbols ν* and υ* are used interchangeably for the third weight in Table 2 and the surrounding text; please use a single notation throughout.","section":"Notation, Table 2"},{"comment":"There are several typographical errors, including 'woks', 'dada set', and 'expect' for 'except', and the running header 'Journal of LaTeX Templates' should be removed.","section":"Section 2 and general text"},{"comment":"Reference [10] is cited as an arXiv preprint; if a published version exists, please cite the published source.","section":"Reference [10]"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the omitted proof of Theorem 7. If the authors can provide that proof, the curve-case analysis appears sound and the paper could be acceptable. The false implication about full column rank in Theorem 6 should also be corrected in revision. I recommend major revision rather than rejection because the curve-case mathematics is solid and the surface proof is likely a routine Kronecker-product extension."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core claim is real: the MLSPIA recurrence with two prior-step memory terms and three weights is a legitimate extension of LSPIA, and the paper proves that for any NTP basis, with a rank-deficient collocation matrix, the iteration converges to a least-squares solution. I checked the eigenvalue analysis in Lemma 5 and the discriminant calculation in Theorem 6; they are correct. The rate comparison is also fair: the optimal MLSPIA rate is strictly better than the LSPIA rate when sigma1 > sigma_r, and the numerical experiments confirm the improvement on independent data sets. The paper is honest about what it does not do: Remark 2 admits that no general weight optimization is attempted, and Remark 3 notes that shape issues are out of scope.\n\nThe main problem is Theorem 7. The surface extension, which is part of the abstract's promise, is stated and then dismissed with \"The proof to the Theorem 7 is omitted since it is similar to the case for Theorems 2 and 6.\" That is not adequate. The surface iteration is not literally the same as the curve iteration: it acts through the Kronecker product B1⊗B2, and one needs to show that the SVD decouples the iteration matrix the way it does in the curve case, and that the extremal singular values are products sigma1*mu1 and sigma_r*mu_s. These are standard facts, but they are not written down, and if the null-space block structure did not carry over, the stated surface rate would not follow. The authors could fix this by including the full proof, or by restating the surface claim as conjectural and moving the surface rate out of the abstract.\n\nTwo smaller issues. First, there is a placeholder citation \"[ ? ]\" in Section 2 where a T-spline development is mentioned. That needs to be resolved before any final version. Second, the abstract says the method converges to \"the\" least-squares curve, but in the rank-deficient case the proof establishes convergence to one of possibly several least-squares solutions. The wording should be adjusted.\n\nNone of these problems undermine the curve-case result, which is solid and practically useful. The citation pattern looks normal, and the numerical comparisons are reproducible in spirit, including the use of the same data sets as the original LSPIA paper. The paper does not oversell its incremental contribution; it is a targeted speedup, not a new paradigm.\n\nFor a CAGD audience, this is worth engaging with. I would send it to peer review, expecting a conditional acceptance: the surface proof must be supplied or the claims appropriately narrowed. The placeholder and the abstract wording are minor and easily fixed.","headline":"The curve-case acceleration result is genuine and the convergence proof checks out; the surface extension in Theorem 7 is asserted without proof and should be supplied or softened before publication.","tokens_in":19871,"tokens_out":1114,"would_cite":true,"duration_ms":13260,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D17","65F10","41A15","15B48"],"pacs":[],"model":"deepseek-v4-flash","headline":"An iteration with memory converges to least-squares curve fits at a rate provably faster than LSPIA, even for rank-deficient collocation matrices.","keywords":["progressive and iterative approximation","least squares fitting","normalized totally positive basis","convergence rate","spectral radius","rank-deficient collocation matrix","B-spline curve fitting","tensor product surface fitting"],"falsifier":"Construct a small tensor-product example, say with $B_1$ rank-deficient and $B_2$ a single-column positive basis, compute the spectral radius of the iteration matrix $H_{\\omega^*,\\gamma^*,\\upsilon^*}$ formed from $B_1\\otimes B_2$, and compare it with $(\\sigma_1\\mu_1-\\sigma_r\\mu_s)/(\\sigma_1\\mu_1+\\sigma_r\\mu_s)$; a mismatch in the asymptotic error contraction would show the omitted surface proof does not follow from the stated assumptions.","tokens_in":18761,"feed_emoji":"📐","tokens_out":11147,"duration_ms":95291,"temperature":0.7,"pith_summary":"This paper introduces MLSPIA, a progressive and iterative approximation method for least-squares fitting that stores the previous correction and uses three weights when updating control points. Its central claim is that, for any normalized totally positive basis, the iterated curves converge to the least-squares fit even when the collocation matrix has deficient column rank, provided the weights satisfy an explicit interval condition. At the paper's chosen optimal weights, the asymptotic contraction rate is $(\\sigma_1-\\sigma_r)/(\\sigma_1+\\sigma_r)$, strictly smaller than the optimal rate $(\\sigma_1^2-\\sigma_r^2)/(\\sigma_1^2+\\sigma_r^2)$ of the earlier LSPIA method whenever $\\sigma_1>\\sigma_r$. Because the method never assembles or solves a linear system directly, a better contraction rate translates directly into fewer iterations on very large point sets. A tensor-product surface version is stated with the same kind of rate, with the proof deferred as similar.","feed_headline":"A fitting iteration with memory reaches least-squares fits faster","feed_subtitle":"A memory-efficient variant of least-squares fitting beats the standard iteration's rate even when the data matrix is singular.","key_machinery":"The load-bearing object is the iteration matrix $H_{\\omega,\\gamma,\\upsilon}$ that governs the coupled update of the auxiliary weighted sums $\\Lambda^k$ and the control points $P^k$; the paper works in any normalized totally positive (NTP) basis, meaning basis functions that are nonnegative, sum to one, and have totally positive collocation matrices. An orthogonal change of variables built from the SVD of the collocation matrix $B$ block-diagonalizes $H_{\\omega,\\gamma,\\upsilon}$, leaving eigenvalues $1-\\omega$ together with the roots of a quadratic equation $\\lambda^2+[\\gamma\\upsilon\\sigma_i^2-(2-\\omega)]\\lambda+\\sigma_i^2\\upsilon(\\omega-\\gamma)+1-\\omega=0$ for each positive singular value $\\sigma_i$. A root-bound lemma converts those quadratics into the explicit weight condition (11), and at the paper's optimal weights every active eigenvalue has modulus $(\\sigma_1-\\sigma_r)/(\\sigma_1+\\sigma_r)$, the contraction factor that determines the convergence rate.","core_discovery":"The discovery, on the paper's own terms, is that the least-squares progressive iteration can be accelerated by adding memory: the increment to each control point at step $k$ blends the previous increment with two successive residual-weighted sums, controlled by weights $\\omega$, $\\gamma$, and $\\upsilon$. The paper proves the resulting curves converge to a solution of the least-squares normal equations $B^T B X = B^T Q$ whenever (11) holds, regardless of whether $B$ has full column rank. The proof rewrites the iteration in coupled variables $(\\Lambda^k, P^k)$, block-diagonalizes the iteration matrix by an orthogonal transform, and shows each singular value $\\sigma_i$ contributes a quadratic eigenvalue equation; the root bound of Lemma 4 converts those equations into the explicit condition (11). At the weights $\\omega^*=\\gamma^*=4\\sigma_1\\sigma_r/(\\sigma_1+\\sigma_r)^2$, $\\upsilon^*=1/(\\sigma_1\\sigma_r)$, the spectral radius is $(\\sigma_1-\\sigma_r)/(\\sigma_1+\\sigma_r)$, which is less than the fastest LSPIA rate $(\\sigma_1^2-\\sigma_r^2)/(\\sigma_1^2+\\sigma_r^2)$ whenever $\\sigma_1>\\sigma_r$. Numerical examples with cubic B-spline curves and a tensor-product cubic B-spline surface confirm that fewer iterations are needed.","pith_inferences":["The paper does not analyze how estimation error in knowing $\\sigma_1$ and $\\sigma_r$ affects the promised rate; a practical version would need cheap approximations of these extremes, and the sensitivity of the spectral radius to those estimates is an open question.","The two-step memory is a linear acceleration; storing more past corrections could in principle push the contraction factor further toward a value governed by the square root of the condition number, although the present paper gives no such analysis.","Because the argument relies only on the singular values of the collocation matrix, the same three-weight acceleration should transfer to other fitting settings with totally positive bases, such as generalized B-spline or subdivision-surface fitting, whenever the same spectral structure can be verified."],"forward_implications":["For very large point sets, MLSPIA reaches a fixed fitting tolerance in fewer iterations than LSPIA, because its spectral radius is provably smaller at the recommended weights.","The convergence guarantee covers rank-deficient collocation matrices, so redundant or nearly dependent basis functions no longer make the iterative fit fail or require a separate least-squares solver.","The added memory costs only two extra additions and three extra scalar multiplications per control point per step, so the faster convergence is obtained at a small per-iteration premium.","If Theorem 7's tensor-product claim is accepted, surface fitting inherits the same speed-up, with the extreme singular products $\\sigma_1\\mu_1$ and $\\sigma_r\\mu_s$ replacing $\\sigma_1$ and $\\sigma_r$ in the rate."],"supporting_citations":[{"why":"Defines the LSPIA method and its optimal convergence rate, which this paper uses as the baseline to beat.","marker":"[7]"},{"why":"Establishes convergence of LSPIA-type iterations when the collocation matrix is singular, the deficient-rank setting this paper extends.","marker":"[10]"},{"why":"Supplies the root-bound criterion quoted as Lemma 4, used to convert the quadratic eigenvalue equations into condition (11).","marker":"[35]"},{"why":"Introduces the PIA framework for totally positive bases that LSPIA and MLSPIA are built on.","marker":"[15]"},{"why":"Supplies the knot-placement condition used to choose parameter values in the numerical examples.","marker":"[36]"}],"fun_headline_variants":["Memory-accelerated fitting iteration outpaces least-squares baseline","With-memory progressive iteration converges faster for least squares","Faster least-squares fitting via iterative memory in control points","Progressive iteration with memory beats standard least-squares rate","Memory-driven fitting iteration speeds up even on singular data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The surface result (Theorem 7) is stated without proof and rests on the unstated assumption that the tensor-product iteration behaves exactly like the curve iteration, with singular values of the two factors multiplying pairwise; if that spectral factorization fails for a rank-deficient factor, the stated surface rate does not follow from the paper's arguments.","fun_headline_variants_meta":{"raw":{"variants":["Memory-accelerated fitting iteration outpaces least-squares baseline","With-memory progressive iteration converges faster for least squares","Faster least-squares fitting via iterative memory in control points","Progressive iteration with memory beats standard least-squares rate","Memory-driven fitting iteration speeds up even on singular data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00071,"raw_usage":{"total_tokens":3219,"prompt_tokens":987,"completion_tokens":2232,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":2151}},"tokens_in":603,"tokens_out":2232,"duration_ms":15047,"temperature":1.0,"reasoning_tokens":2151,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:48:21.772308+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a small tensor-product example, say with $B_1$ rank-deficient and $B_2$ a single-column positive basis, compute the spectral radius of the iteration matrix $H_{\\omega^*,\\gamma^*,\\upsilon^*}$ formed from $B_1\\otimes B_2$, and compare it with $(\\sigma_1\\mu_1-\\sigma_r\\mu_s)/(\\sigma_1\\mu_1+\\sigma_r\\mu_s)$; a mismatch in the asymptotic error contraction would show the omitted surface proof does not follow from the stated assumptions.","supporting_citations":[{"cited_title":"Progressive and iterative approximation for least squares B-spline curve and surface fitting","cited_arxiv_id":null,"evidence_quote":"Defines the LSPIA method and its optimal convergence rate, which this paper uses as the baseline to beat."},{"cited_title":"Iterative solution of large linear systems","cited_arxiv_id":null,"evidence_quote":"Supplies the root-bound criterion quoted as Lemma 4, used to convert the quadratic eigenvalue equations into condition (11)."},{"cited_title":"Totally positive bases and progressive iteration approximation","cited_arxiv_id":null,"evidence_quote":"Introduces the PIA framework for totally positive bases that LSPIA and MLSPIA are built on."},{"cited_title":"On polya frequency functions","cited_arxiv_id":null,"evidence_quote":"Supplies the knot-placement condition used to choose parameter values in the numerical examples."}],"review_version":1}