{"id":"720c64e5-8496-4007-a339-8821321a55a6","arxiv_id":"2608.05463","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A slice-sampled tensor can be reconstructed by jointly fitting a known basis along the trajectory and kernel interpolation between slices inside a low-rank Tucker model.","lead":"BMTA reconstructs a large data cube from a few observed slices by combining a known smooth basis with kernel-based interpolation between nearby slices. The paper adds an error bound and shows gains on quantum-chemistry and radio-map tensors, though the proof has gaps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The contraction coefficient in Lemma 1 is a sum of individually contracting bounds, so it can exceed 1 under the stated step-size range; Theorem 1's geometric convergence is unproven even with valid initialization.","rationale":"The reader's verdict is REJECT with moderate confidence, based on the initialization-basin gap and a sampling lemma issue. My analysis identifies a different, more load-bearing defect: even granting the initialization hypothesis, the proof of Lemma 1 sums three separate contraction inequalities, and the resulting coefficient (equation (47)) is not shown to be less than 1. In fact, for the stated step-size interval, the sum generically exceeds 1, and the interval is empty for tensors with condition number above about 1.85. This directly invalidates the contraction argument behind Theorem 1. The paper does provide a concrete algorithm and experiments, but the central theoretical claim is not established. Because this concern strengthens rather than changes the reader's rejection, the verdict remains UNCHANGED. The concrete test is a symbolic recomputation of the coefficient, which can be done with paper and pencil or a computer algebra system in minutes.","tokens_in":48,"tokens_out":8007,"duration_ms":255719,"concrete_test":"Independently recompute the coefficient in equation (47) using the definitions of ρ0 and ρ1 from equation (42), with ρ0=2(1−2ησ_min^2(Σ1))^2 and ρ1 analogous, and with τ=ϵ1σ_min. Choose a tensor with σ_max/σ_min=5 and any admissible η from the stated lower bound up to 1/(2σ_max^2): verify whether the coefficient (1−η)^2+ρ0+ρ1+η^2γ^2(1+τ)^10 is strictly less than 1 for some η in the interval. If no such η exists, Lemma 1's contraction claim is false and Theorem 1's proof collapses; this single symbolic check settles the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Appendix C, the proof bounds each of the three terms of dist^2 separately: inequality (42) gives a bound for the mode-1 factor with coefficient ρ0<1, an analogous bound gives ρ1<1 for the mode-2 factor, and inequality (43) bounds the core term. These are then added, and equation (47) yields the dist^2 coefficient (1−η)^2 + ρ0 + ρ1 + η^2 γ^2(1+τ)^10, with τ=ϵ1σ_min. This coefficient must be less than 1 for the claimed contraction ρ<1. But no argument shows the sum is less than 1: each ρi is positive and the (1−η)^2 term is strictly positive, so the sum is generically larger than any individual term. In fact, for the smallest allowed η=(1−1/√2)/(2 min_k σ_min^2(Σ_k)), the factor 2(1−2ησ_min^2)^2 is close to 1, and adding ρ1 and (1−η)^2 makes the total exceed 1. Moreover, the stated upper bound η≤1/(2 max_k σ_max^2(Σ_k)) requires σ_max^2/σ_min^2 ≤ 1/(1−1/√2) ≈ 3.414, i.e., a condition number below about 1.85; for typical Hessian tensors with larger condition numbers the step-size interval is empty. Thus Lemma 1's contraction factor is not established, and Theorem 1, which relies on Lemma 1, is unsupported even for initializations that satisfy dist(F^0,F)≤ϵ1σ_min. This is a more fundamental gap than the initialization issue alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Basis and Manifold prior Tensor Approximation (BMTA), a three-stage algorithm for reconstructing a third-order tensor from a small set of uniformly sampled lateral slices. The method combines a known global basis representation along the trajectory mode with a kernel-based manifold interpolation prior, and fits the resulting hybrid model through a low-rank Tucker optimization. The main theoretical contribution is Theorem 1, which claims a non-asymptotic reconstruction error bound of the form O(ρ^{t/2}) plus model-mismatch terms, under a sampling complexity condition on the number of observed slices. The paper also reports numerical experiments on synthetic tensors, radio-map reconstruction, and quantum-chemistry Hessian datasets.","tokens_in":24929,"tokens_out":10061,"duration_ms":91702,"significance":"If Theorem 1 were correct, the paper would provide a useful convergence guarantee for tensor recovery under structured lateral-slice sampling, a regime that is common in scientific applications but less studied than random entrywise sampling. The algorithmic idea of combining a global basis prior with a local manifold prior is reasonable, and the experiments show consistent gains over the considered baselines in low-sampling regimes. However, the theoretical analysis is the central advertised contribution, and it contains several load-bearing gaps: the contraction lemma's coefficient is not shown to be less than one, the sampling lemma misapplies a rescaled row-sampling bound to an unnormalized selection matrix, and a key constant in the tensor-space lifting lemma is obtained from a false inequality. These issues prevent the paper from substantiating its main claim as written.","major_comments":[{"comment":"The contraction coefficient in Lemma 1 is not proven to be less than one. Equation (47) adds the mode-1 contraction coefficient ρ0, the mode-2 coefficient ρ1, the term (1-η)^2, and the term η^2 γ^2(1+τ)^10. Even if ρ0<1 and ρ1<1 individually, their sum together with the other positive terms has no reason to be below 1. Moreover, the text states that 2(1-2ησ_min^2)^2<1 follows from the lower bound on η, but the preceding bound actually contains the coefficient 4(1-2ησ_min^2)^2, which can exceed 1 for step sizes at the stated lower end. Consequently, the geometric contraction claimed in Lemma 1 and used in Theorem 1 is unsupported.","section":"Appendix C, Lemma 1 and Eq. (47)-(48)"},{"comment":"The step-size interval in Lemma 1, (1-1/√2)/(2 min_k σ_min^2(Σ_k)) < η < 1/(2 max_k σ_max^2(Σ_k)), is empty unless σ_max^2/σ_min^2 < 1/(1-1/√2) ≈ 3.414, i.e., unless the condition number of the relevant unfoldings is below about 1.85. This is not stated as an assumption in Theorem 1 and is violated in typical numerical regimes. Additionally, Lemma 1 assumes γ(1+ϵ1 σ_min)≤1, where γ is the largest singular value of H_(1) or H_(2); this excludes all tensors with a singular value larger than one and is absent from Theorem 1. The theorem therefore applies only under conditions that are neither stated nor verified.","section":"Lemma 1 and Theorem 1, step-size and spectral assumptions"},{"comment":"Lemma 5's proof applies Lemma 4 to the unnormalized selection matrix Ψ∈{0,1}^{d×n}. Lemma 4, however, is a concentration bound for a row-sampling and rescaling operator; without the rescaling factor, E[(ΨU_S)^T(ΨH^⊥)] = (d/n) U_S^T H^⊥, and the quoted bound does not hold for the raw selection matrix. The Markov step following (22) also does not yield the stated sample complexity: matching the high-probability conclusion requires d≥O(n^2l/(nl+δ^2 ϵ^2(n-1))), not the d≥O(n^2l/(nl+δ ϵ^2(n-1))) stated in Lemma 5 and Theorem 1. Lemma 3 additionally assumes σ_min^2(ΨU_S)≥1/√2 without proving that this subspace embedding condition follows from the assumed d; as a result, the sampling-induced error bound in Lemma 3 is not established.","section":"Lemma 5 and Lemma 3, sampling argument"},{"comment":"The proof of Lemma 2 uses the inequality (τ^2+2τ+2) ≤ (4/3)(1+τ)^3 when passing from (c) to (d). At τ=0 this reads 2 ≤ 4/3, which is false, and the inequality only becomes valid for τ larger than roughly 0.22. Since τ=ϵ1 σ_min can be arbitrarily small, the factor 4/3 in Lemma 2 is not justified. This directly affects the constant in Theorem 1's error bound.","section":"Appendix D, Lemma 2"},{"comment":"The stated closed-form solution for the basis coefficient tensor is written as \\hat{Q} = fold_(2)(Ψ† C_(2)) ×_2 S†. The least-squares problem in (7) is min ||C_(2) - Ψ S \\bar{Q}_{(2)}||_F^2, whose solution is fold_(2)((ΨS)^† C_(2)). Using the pseudoinverse of Ψ alone is dimensionally inconsistent and does not solve the displayed regression; this is a concrete error in the algorithm's definition, not merely a typographical issue in notation.","section":"Algorithm 1, line 4 and Eq. (7)"},{"comment":"Theorem 1 assumes dist(\\hat{F}^0,F) ≤ ϵ1 σ_min, but Algorithm 1 initializes every entry of \\hat{G}, \\hat{X}_1, and \\hat{X}_2 independently from N(0,1). The paper provides no result showing that such a random initialization satisfies the basin-of-attraction condition. Without this, the linear convergence guarantee does not apply to the algorithm as implemented and tested.","section":"Theorem 1 and Algorithm 1, initialization"}],"minor_comments":[{"comment":"The notation σ_min^2(Σ_k) is used where σ_min(Σ_k) or σ_min^2(H_(k)) would be clearer; likewise the upper bound in Theorem 1 and Lemma 1 alternates between min_k σ_max^2(Σ_k) and max_k σ_max^2(Σ_k), which should be reconciled.","section":"Notation throughout"},{"comment":"The lemma states that the bound holds 'with probability δ', but the intended meaning appears to be 'with probability at least 1−δ'; this should be corrected everywhere.","section":"Lemma 5 statement"},{"comment":"The use of an O(·) term inside a displayed high-probability inequality is not a rigorous non-asymptotic bound; the constants hidden in the O should be made explicit if the bound is to be used for comparison with experiments.","section":"Theorem 1 statement"},{"comment":"The text says results 'closely follow the trends predicted by Theorem 1', but the experiments do not estimate the contraction rate ρ or the model-mismatch terms; this claim should be softened or accompanied by a quantitative comparison.","section":"Section VI, Figure 4 caption"},{"comment":"Several supporting lemmas are stated as imported from the authors' prior work ([4], [5], [8], [11], [14]); the paper should state explicitly which results are reused and verify that the cited versions match the assumptions used here.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript's core theoretical contribution is not established: the contraction coefficient in Lemma 1 is not shown to be less than one, the sampling lemma is based on an inapplicable concentration result, and the constants in Lemma 2 are wrong. These are not local presentation issues but affect the main theorem and the algorithm's stated guarantees. A resubmission that fixes the contraction analysis, uses a properly rescaled sampling operator, and either proves the initialization condition or removes the basin-of-attraction assumption would be needed before the theoretical claims can be considered sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful algorithm paper with an unproven theorem. The BMTA idea—combining a global basis prior with a local RBF interpolation prior inside a symmetric Tucker completion—is a reasonable extension of the authors' earlier ICASSP work, and the experiments on quantum chemistry and radio maps are well-motivated. But the contraction lemma that drives Theorem 1 is not valid as written, and the sampling lemma has an operator mismatch. The theorem should not be cited as established until those are fixed.\n\nWhat is actually new: the symmetric Tucker formulation (same factor for modes 1 and 3) with the alternating gradient updates, and the specific non-asymptotic error bound for it. The bound's structure—contraction plus model-mismatch—is standard, but the statement is new. The experiments are thoughtfully designed: ablations against structure-only and interpolation-only, plus TensorCUR and Tucker, and they honestly report that TensorCUR wins at high sampling fractions. That is more honest than the abstract suggests.\n\nSoft spots, in order of severity.\n\nFirst, Lemma 1. The proof bounds the three factor errors separately and adds them. The resulting contraction coefficient is C = (1−η)^2 + ρ0 + ρ1 + η^2 γ^2(1+τ)^10. Each ρi is individually less than 1, but the sum can easily exceed 1. For instance, with σ_min = σ_max = 1 and η at the stated lower bound, ρ0 and ρ1 are close to 1, (1−η)^2 is about 0.72, so C is above 2. The paper simply asserts that the η range makes ρ < 1, but no argument shows it. The step-size interval also appears empty for condition numbers above about 1.85, which is restrictive for Hessian applications. Second, Lemma 5. The paper defines Ψ as a d×n selection matrix without rescaling, but Lemma 4 (from Drineas et al.) applies to the row-sampling-and-rescaling operator. The proof drops the scaling, so the bound does not follow. The d≥... condition also looks off by a factor of δ from the Markov argument. Third, the initialization assumption: Theorem 1 requires dist(F^0,F) ≤ ϵ1 σ_min, but Algorithm 1 starts from Gaussian draws, and there is no result showing that random initialization satisfies this. That is a genuine gap, though less fundamental than the contraction problem.\n\nThe core idea is plausible and the experiments suggest it works. I would send this to peer review, but I would expect major revision. Lemma 1 has to be fixed before the main theorem can be trusted.","headline":"A useful algorithm and honest experiments, but Lemma 1's contraction proof doesn't hold; Theorem 1's guarantee is unsupported as stated.","tokens_in":25412,"tokens_out":7813,"would_cite":false,"duration_ms":65959,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69","65F22"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes BMTA, a tensor-approximation algorithm that reconstructs a full third-order tensor from a small set of lateral slices by combining a known basis model with manifold-guided interpolation, and proves a non-asymptotic…","keywords":["structured tensor approximation","lateral slice sampling","Tucker decomposition","basis prior","manifold interpolation","non-asymptotic error bound","quantum chemistry Hessians","radio map reconstruction"],"falsifier":"Take a ground-truth tensor that exactly satisfies the Tucker, basis, and interpolation models, run Algorithm 1's $N(0,1)$ initialization on it, and compute $\\mathrm{dist}(\\hat F^0,F)$ before the first update: if this distance exceeds $\\epsilon_1\\sigma_{\\min}(H)$ for a non-negligible fraction of random draws, the contraction premise of Theorem 1 fails on those starts. A weaker check is to plot the tensor-space error against iterations for many random seeds; absence of the predicted geometric contraction from the first step would indicate the basin-of-attraction assumption is not being met.","tokens_in":24315,"feed_emoji":"🧮","tokens_out":7664,"duration_ms":68810,"temperature":0.7,"pith_summary":"The paper proposes Basis and Manifold prior Tensor Approximation (BMTA), an algorithm for reconstructing a third-order tensor from only a few randomly chosen lateral slices. BMTA combines two complementary assumptions: the slices evolve smoothly along a known low-dimensional basis such as Legendre polynomials or a DCT, and neighboring slices lie on a smooth manifold that supports kernel interpolation from the sampled slices. The central claim is a high-probability, non-asymptotic bound in which the reconstruction error of BMTA decomposes into a term that contracts geometrically with gradient-descent iterations and an irreducible bias term governed by how well the two priors match the true tensor. Numerical experiments on synthetic data, radio-map reconstruction, and quantum-chemistry Hessian tensors indicate that BMTA attains lower normalized error than the baselines in limited-sampling regimes.","feed_headline":"Sparse lateral slices suffice for tensor recovery with two priors","feed_subtitle":"BMTA combines a known basis with manifold interpolation and proves linear convergence to a bounded error.","key_machinery":"The load-bearing mechanism is an invariant distance on Tucker factor triples, $$\\mathrm{dist}^2((\\hat X_1,\\hat X_2,\\hat G),(X_1,X_2,G)) = \\inf_{Q_k\\in GL(r_k)} \\|\\hat X_1 Q_1 - X_1\\|$_F^{2}$ + \\|\\hat X_2 Q_2 - X_2\\|$_F^{2}$ + \\|($Q_1^{{-1}}$,$Q_2^{{-1}}$,$Q_1^{{-1}}$)\\cdot \\hat G - G\\|$_F^{2}$,$$ together with three supporting lemmas: Lemma 1 shows this distance contracts linearly under gradient descent as long as the current factors stay inside the basin; Lemma 2 lifts the factor-space contraction to a Frobenius-norm bound on the reconstructed tensor; and Lemma 3 bounds the error introduced by random lateral-slice sampling using a subspace-embedding concentration inequality. The known basis matrix $S$ enters in two roles: it defines the regression in (7) that produces the basis coefficient tensor $\\hat Q$, and its spectral and coherence properties set the sampling complexity $d$ in Theorem 1. The interpolation matrix $\\Lambda_\\Omega$ is constructed from Gaussian RBF weights (8) between sampled and unsampled reaction-coordinate points, and the parameter $\\alpha$ balances the two priors in the final low-rank Tucker objective (9).","core_discovery":"On its own terms, the paper's main discovery is an error decomposition (Theorem 1) for BMTA: if the initial Tucker factors lie within a basin of attraction of the true factors and the number of sampled lateral slices is large enough, then with probability $1-\\delta$ the iterates satisfy a bound of the form $$\\|(\\hat X_1^t,\\hat X_2^t,\\hat X_1^t)\\cdot \\hat G^t - H\\|_F \\le \\tfrac43 (1+\\epsilon_1\\sigma_{\\min})^3 \\zeta \\$rho^{{t/2}}$\\,\\mathrm{dist}(\\hat $F^{0}$,F) + O\\bigl((1+\\epsilon_1\\sigma_{\\min})^3 \\zeta\\,(\\epsilon_1+\\|\\$\\alpha$ E_{\\hat QS,(1)}+\\$\\beta$ E_{C\\hat\\Lambda_\\$\\Omega$,(1)}\\|_F)/\\sqrt{1-\\rho}\\bigr).$$ The first term shrinks geometrically, while the second, which does not vanish as iterations grow, is controlled by the quality of the basis representation and the manifold-interpolation model. The theorem thus states that BMTA converges linearly to a neighborhood of the true tensor, and that the size of that neighborhood is governed by model mismatch rather than by optimization. This is the claim the paper is trying to establish.","pith_inferences":["The theorem suggests a practical warm-start rule the paper does not test: initialize the Tucker factors from the Structure-only or Interpolation-only reconstructions rather than from $N(0,1)$, which would make the basin-of-attraction assumption checkable and more likely to hold.","The same two-prior architecture should extend to fourth- and higher-order tensors by replacing the shared mode-1/mode-3 factor with independent factors; the invariant-distance proof only uses the Tucker algebra, which does not change.","The bias term predicts a testable plateau: on synthetic tensors with deliberately mismatched basis or bandwidth, the NMSE should stop improving once the optimization term falls below the mismatch term, even as $t$ grows.","One could estimate the bias term online from the residuals $\\hat Q\\times_2 S - \\hat H$ and $C\\times_2 \\hat\\Lambda_\\Omega - \\hat H$, giving a data-driven criterion for choosing between BMTA and a pure CUR-style method when sampling is abundant."],"forward_implications":["BMTA turns side information about a trajectory (a known basis and coordinate distances) into a provable reconstruction guarantee from structured lateral slices, not just random entries.","When the basis and interpolation priors are well aligned with the true tensor, the asymptotic reconstruction error is small and additional gradient iterations buy exponential improvement down to that floor.","The sampling-complexity condition $d = O(n^2 l/(nl+\\delta\\epsilon_2^2(n-1)))$ quantifies how many lateral slices are needed so that random slice sampling behaves like a near-isometry on the basis column space.","The bound also identifies which failure modes matter: a poor basis or a badly tuned interpolation bandwidth creates an irreducible error floor that no amount of optimization can remove.","On the numerical side, BMTA is argued to outperform Tucker decomposition and TensorCUR when slices are scarce, which is exactly the regime the theorem's contraction-plus-bias structure predicts."],"supporting_citations":[{"why":"Supplies the Tucker decomposition model and the tensor-norm identities that define the reconstruction target.","marker":"[7]"},{"why":"Provides the motivating Hessian-along-reaction-coordinate application that requires lateral slice sampling.","marker":"[13]"},{"why":"Introduces the side-information matrix approximation idea that BMTA's basis prior extends to tensors.","marker":"[5]"},{"why":"Supplies the existence of optimal alignment matrices (Lemma 7) that underpins the invariant distance metric.","marker":"[22]"},{"why":"Supplies the random row-sampling concentration inequality (Lemma 4) used to bound the sampling error in Lemma 3.","marker":"[29]"},{"why":"Grounds the Gaussian RBF kernel construction of the interpolation weights in (8).","marker":"[21]"},{"why":"Gives the TensorCUR baseline that BMTA is compared against in the numerical section.","marker":"[24]"},{"why":"Documents the computational cost of Hessian evaluations in chemical reactions, motivating the limited-slice setting.","marker":"[11]"}],"fun_headline_variants":["BMTA: Tensor recovery from few slices via dual priors","Two priors, few slices: BMTA proves bounded error","Linear convergence to model-limited error in tensor recovery","Sparse slices, dual priors: BMTA's bounded-error path"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire linear-convergence guarantee depends on the initial random factor estimates already lying within a small basin of attraction around the true factors, of radius $\\epsilon_1 \\sigma_{\\min}(H)$, and the paper does not prove that the $N(0,1)$ initialization in Algorithm 1 satisfies this condition.","fun_headline_variants_meta":{"raw":{"variants":["BMTA: Tensor recovery from few slices via dual priors","Two priors, few slices: BMTA proves bounded error","Linear convergence to model-limited error in tensor recovery","Sparse slices, dual priors: BMTA's bounded-error path"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00068,"raw_usage":{"total_tokens":3101,"prompt_tokens":970,"completion_tokens":2131,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":2061}},"tokens_in":586,"tokens_out":2131,"duration_ms":14990,"temperature":1.0,"reasoning_tokens":2061,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T12:37:24.817500+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a ground-truth tensor that exactly satisfies the Tucker, basis, and interpolation models, run Algorithm 1's $N(0,1)$ initialization on it, and compute $\\mathrm{dist}(\\hat F^0,F)$ before the first update: if this distance exceeds $\\epsilon_1\\sigma_{\\min}(H)$ for a non-negligible fraction of random draws, the contraction premise of Theorem 1 fails on those starts. A weaker check is to plot the tensor-space error against iterations for many random seeds; absence of the predicted geometric contraction from the first step would indicate the basin-of-attraction assumption is not being met.","supporting_citations":[{"cited_title":"Interpolated variational transition-state theory by mapping,","cited_arxiv_id":null,"evidence_quote":"Provides the motivating Hessian-along-reaction-coordinate application that requires lateral slice sampling."},{"cited_title":"Matrix approximation with side information: When column sampling is enough,","cited_arxiv_id":null,"evidence_quote":"Introduces the side-information matrix approximation idea that BMTA's basis prior extends to tensors."},{"cited_title":"Fast monte carlo al- gorithms for matrices i: Approximating matrix multiplication,","cited_arxiv_id":null,"evidence_quote":"Supplies the random row-sampling concentration inequality (Lemma 4) used to bound the sampling error in Lemma 3."},{"cited_title":"Kernel methods on riemannian manifolds with gaussian rbf kernels,","cited_arxiv_id":null,"evidence_quote":"Grounds the Gaussian RBF kernel construction of the interpolation weights in (8)."},{"cited_title":"Tensor-cur decomposi- tions for tensor-based data,","cited_arxiv_id":null,"evidence_quote":"Gives the TensorCUR baseline that BMTA is compared against in the numerical section."},{"cited_title":"A matrix completion algorithm for efficient calculation of quantum and variational effects in chemical reactions,","cited_arxiv_id":null,"evidence_quote":"Documents the computational cost of Hessian evaluations in chemical reactions, motivating the limited-slice setting."}],"review_version":1}