{"id":"5b2ad05f-92d6-4515-831e-02abf537b7e9","arxiv_id":"2608.03928","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"R-ItCUR is an iterative t-CUR algorithm that completes low-tubal-rank tensors from cross-concentrated samples corrupted by sparse outliers, using Welsch robust correction and blockwise projected gradient descent.","lead":"This paper proposes a method to fill in a partially observed 3D data array, a tensor, when some observed entries are also badly corrupted, using a sampling pattern that only observes entries inside selected horizontal and vertical slabs. The authors demonstrate on MRI and seismic data that their algorithm, R-ItCUR, is more robust to outliers than the previous cross-sampling completion method.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"t-CCS model mismatch: entrywise corruption experiments do not satisfy the paper's own tube-wise sparsity Assumption 3.1, so the reported robustness is not established for the model claimed.","rationale":"The reader's weakest assumption identified the blockwise projection heuristic (Eq. 12) as the main concern. That is a legitimate algorithmic robustness question: if U's estimated subspaces are inaccurate, the whole reconstruction suffers. But the paper does not claim a guarantee for that projection and explicitly describes it as following t-CUR philosophy; its accuracy is ultimately the empirical success or failure of the method. A more fundamental, load-bearing concern is that the spike corruption used in every single experiment violates the paper's own formal corruption model. Section 3.2 introduces tube-wise α-sparsity as 'central to this work', states that conventional entrywise sparsity does not enjoy the Fourier-support property, and then Section 5 corrupts individual entries, not tubes. Lemma 3.2 is the only theoretical justification tying the sparse-outlier model to the t-SVD/t-CCS framework, and it explicitly does not apply to the experiments run. Since R-ItCUR has no convergence or recovery theorem, the entire support for the abstract's central assertion--'strong robustness to sparse gross corruptions' under t-CCS--rests on experiments that instantiate a different corruption model than the one the paper argues is appropriate. This does not necessarily invalidate the practical findings (entrywise corruption might be the harder or easier model), but it means the evidence does not test the stated claim. There is also a second, smaller mismatch worth noting: the corruption magnitudes are randomly signed in the seismic/MRI descriptions while the synthetic description says Unif[−cμX, cμX] (which is also random-sign), so the 'arbitrarily large' adversarial magnitude claim is untested--though this is less damaging than the tube-wise mismatch. My proposed test is standard and cheap: regenerate the same experiments with full-tube corruption at matched counts and compare. The verdict stays CONDITIONAL because the empirical comparisons can still be reproduced and could be re-run under the correct model; but the condition should include 'experiments must match the stated corruption model.' I partially agree with the reader because their blockwise-projection concern and my model-mismatch concern are both correctness risks, though I view the model mismatch as more central to the paper's contribution, since Section 3.2 explicitly foregrounds it as a design driver. No ad hominem is intended; the critique is about the alignment between the formal model and the experiments, not about author intent.","tokens_in":17706,"tokens_out":2157,"duration_ms":20503,"concrete_test":"Re-run the synthetic, MRI, and seismic experiments under the tube-wise corruption model required by Assumption 3.1: select αn1n2 tubes uniformly at random and corrupt every entry in each selected tube with an independently drawn nonzero value (uniform sign-magnitude or random sign with magnitude cμX), then apply R-ItCUR under identical t-CCS masks. If the median/mean PSNR or RelErr degrades substantially relative to the current entrywise-corruption results (e.g., the MRI α=0.15 gains over ITCURTC shrink or disappear), then the current experiments were not testing the paper's claimed outlier model. Conversely, if tube-wise results match the entrywise results, the tube-wise motivation in Section 3.2 is vacuous and should be re-written.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central theoretical framing (Section 3.2 and Assumption 3.1) requires tube-wise sparsity, not entrywise sparsity, precisely because entrywise spatial outliers spread across all Fourier slices. Yet every numerical experiment corrupts entries: Section 5.1 draws ⌈α|Ω|⌉ observed entries uniformly at random (19); MRI experiments corrupt round(α|Ω|) entries (21); seismic experiments corrupt ⌈α|Ω|⌉ entries. Under Assumption 3.1, tube-wise α-sparsity means at most αn2 tubes per row and αn1 per column are corrupted in full (all n3 entries along the third mode), whereas the experiments corrupt isolated entries. Uniformly random entrywise outliers are overwhelmingly single-entry tubes, so each corrupted tube contains one corrupted entry rather than being fully corrupted. This is not merely a cosmetic mismatch: Lemma 3.2, the justification for applying robust CUR frequency-by-frequency and the motivation for tube-wise sparsity, does not hold for entrywise corruption. Therefore the experiments do not test the model under which the method's robustness strategy is argued to be principled. The claim that R-ItCUR is robust to sparse gross corruptions under t-CCS is an empirical claim about a different outlier model (entrywise) than the one introduced and analyzed in the paper. Since the paper provides no convergence or recovery guarantee, the only evidence for the central claim is these experiments; the mismatch means the central claim as stated is not supported by the provided evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies robust completion of third-order low-tubal-rank tensors from tensor cross-concentrated sampling (t-CCS) when the observations are contaminated by sparse, arbitrarily large outliers. The proposed algorithm, R-ItCUR, stores only the three sampled cross blocks, applies an adaptive blockwise Welsch correction to suppress outliers, updates the intersection block by a truncated t-SVD, and projects the exterior blocks onto the resulting tubal column and row subspaces. The output is an implicit t-CUR representation, so the full tensor is never formed during the iterations. The paper reports synthetic experiments and real-data experiments on a cardiac MRI volume and a 3D seismic data set, comparing R-ItCUR with the nonrobust t-CCS method ITCURTC and with a robust full-tensor iterative hard-thresholding method under both t-CCS and matched uniform sampling.","tokens_in":1618,"tokens_out":1584,"duration_ms":150243,"significance":"If the reported empirical behavior is representative, R-ItCUR is a practically useful and computationally efficient algorithm for robust tensor completion under structured sampling: it extends the t-CUR framework to corrupted observations while preserving the memory savings of working only on the sampled cross. The experimental design is carefully matched across methods, using the same masks, corruption supports, and corruption amplitudes in paired comparisons, and the authors are transparent about the fact that synthetic relative errors do not vanish. The main value of the paper is algorithmic and empirical rather than theoretical, since no convergence or recovery guarantees are provided. The significance is tempered by a mismatch between the formal tube-wise outlier model and the entrywise corruptions used in all experiments, and by the heuristic subspace-propagation step that is not stress-tested. These issues are fixable, but they affect how strongly the central claim can be stated.","major_comments":[{"comment":"Assumption 3.1 defines tube-wise alpha-sparsity: at most alpha n2 full tubes per horizontal direction and alpha n1 per lateral direction may be corrupted. Lemma 3.2 then argues that each Fourier slice has the same sparse support, which is the stated justification for the robust strategy. However, all numerical experiments corrupt individual entries rather than full tubes: Eq. (19) selects ceil(alpha |Omega|) observed entries uniformly at random, Eq. (21) corrupts round(alpha |Omega|) entries, and the seismic experiments corrupt ceil(alpha |Omega|) entries. A randomly chosen corrupted entry is a tube with only one nonzero entry, so the experiments are not instances of Assumption 3.1 and Lemma 3.2 does not apply to them. Because the paper provides no recovery guarantee, the central claim that R-ItCUR is robust under the formal t-CCS sparse-corruption model is supported only for a different, entrywise outlier model. Please add tube-wise corruption experiments that satisfy Assumption 3.1 or reformulate the problem in terms of entrywise sparsity and remove the Fourier-support argument.","section":"Section 3.2 and Sections 5.1-5.3"},{"comment":"The recovery of the exterior blocks depends entirely on the tubal column and row subspaces W^l and V^l estimated from the truncated t-SVD of the intersection block X_U. If the intersection is small, its sampling within the cross is sparse, or its observed entries are heavily corrupted, W^l and V^l can be poor approximations of the true subspaces, and the projection step in (12) propagates this error to X_R' and X_C' with no mechanism for the exterior blocks to correct the subspace estimate. This is a heuristic step, and no perturbation analysis or numerical ablation is provided that varies |I|, |J|, the within-intersection sampling rate, or the concentration of corruptions inside U. Since the paper makes no convergence guarantee, this assumption is load-bearing for the claim that the implicit t-CUR update recovers the tensor.","section":"Section 4.2, Eqs. (11)-(12)"},{"comment":"The function W_sigma(E) = (1 - exp(-E odot E / (2 sigma^2))) odot E is not the standard Welsch influence function. For the Welsch loss rho(x) = 1 - exp(-x^2 / (2 sigma^2)), the influence function is x exp(-x^2 / (2 sigma^2)), which tends to zero for large |x|, whereas the displayed formula tends to E for large |E|. The text states that this correction is 'smooth, bounded, and nonconvex' and that it suppresses the influence of large residuals, which is inaccurate for the displayed function. The algorithm may still be sensible as a smooth surrogate for hard thresholding of residuals, but the naming, the description, and the connection to references [48,49] should be corrected.","section":"Section 4.1, Eq. (8)"}],"minor_comments":[{"comment":"The definition of Omega_U contains a typo: 'Omega_U = (Omega_R cup Omega_U) cap (I times J times [n3])' should presumably read '(Omega_R cup Omega_C) cap (I times J times [n3])'.","section":"Section 4"},{"comment":"Several rows in Table 5 have missing separators between the R-ItCUR and ITCURTC columns, e.g., '27.388.69' should read '27.38  8.69'. Please fix the table formatting.","section":"Table 5"},{"comment":"The caption begins with '[20]', which appears to be a leftover citation reference rather than part of the caption text.","section":"Figure 1 caption"},{"comment":"The synthetic relative errors in Table 1 remain between about 2 and 5 percent, and the text itself notes that 'the reconstruction error does not vanish' and that most trials reach the iteration limit. The abstract's phrase 'accurate recovery' should be qualified in light of these values.","section":"Section 5.1"},{"comment":"The contamination fraction alpha is an input to the algorithm, and all experiments use the true alpha value. No sensitivity analysis is reported for misspecified alpha, which would be relevant for practical use.","section":"Section 4.1 and Algorithm 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable algorithmic contribution, but the mismatch between the formal tube-wise sparsity model and the entrywise corruption experiments is a substantive gap that the authors should address directly. The subspace-propagation step is the weakest point of the method and deserves either a numerical stress test or a theoretical perturbation bound. The Welsch formula issue should be corrected because it affects how the method is described and understood."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: R-ItCUR is a genuinely new algorithmic combination—Welsch robust correction layered onto iterative t-CUR on the sampled cross—and it does demonstrate consistent gains over the nonrobust ITCURTC baseline under t-CCS on MRI and seismic data. Avoiding full-tensor reconstruction throughout the iterations is a real practical advantage, and the clean-case parity with ITCURTC is a good sanity check. The related work is handled transparently, including the authors' own prior t-CUR machinery; self-citation there is not a problem.\n\nNow the soft spots. The largest is the mismatch between the model and the experiments. Assumption 3.1 defines tube-wise sparsity, and Lemma 3.2 uses that to argue Fourier support preservation. But every numerical experiment—synthetic Eq. (19), MRI Eq. (21), and seismic—corrupts isolated entries chosen uniformly at random. So the experiments validate robustness to entrywise outliers, not the tube-wise adversarial model that motivates the Fourier argument. This is not purely cosmetic. The central claim as stated in the abstract, robust recovery under t-CCS with \"sparse gross corruptions,\" is ambiguous, and the evidence does not currently cover the model the paper emphasizes. The fix is straightforward: either add experiments with full-tube corruptions or revise the model to entrywise sparsity and drop the Fourier-support justification.\n\nSecond, there are no convergence or recovery guarantees. The paper admits the reconstruction error does not vanish, and synthetic relative errors sit around 2–5%. That is acceptable for an algorithmic paper, but it makes the experiments load-bearing, and the experimental reporting is not quite airtight: seismic results are reported as means without standard deviations, and the synthetic table gives medians only. Third, the conclusion overclaims the universal advantage of exploiting t-CCS structure: uniform-sampling Robust-IHT beats R-ItCUR on the seismic dataset at the two higher observation fractions, so the claim that explicitly exploiting the sampling structure significantly improves performance needs qualification.\n\nOverall, this is a reasonable, readable paper aimed at researchers working on structured sampling and robust tensor completion. It deserves a serious referee, but it needs a revision that aligns the experiments with the stated corruption model, provides code and data, and tempers the claims. I would not desk-reject it; I would send it out with a request for those changes.","headline":"A useful memory-efficient algorithm for robust t-CCS completion, but the experiments corrupt entrywise while the model assumes tube-wise sparsity, and that mismatch leaves the central claim under-supported.","tokens_in":18553,"tokens_out":2434,"would_cite":true,"duration_ms":29746,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69","15A83","65K10","68T09"],"pacs":[],"model":"deepseek-v4-flash","headline":"R-ItCUR, a tensor-native algorithm, recovers low-tubal-rank tensors from cross-concentrated samples under sparse gross corruption, working only on the observed slice blocks.","keywords":["tensor completion","tubal rank","t-CUR decomposition","cross-concentrated sampling","sparse outliers","Welsch loss","t-product","robust tensor recovery"],"falsifier":"Take a synthetic low-tubal-rank tensor of size roughly $120 \\times 120 \\times 8$, fix the t-CCS cross fraction small enough that the intersection block holds only a few slices, and place all gross outliers inside that intersection rather than spreading them across the cross. If R-ItCUR's relative error then rises well above the clean-case level while a full-tensor robust method on the same observations stays accurate, the blockwise subspace-projection heuristic is the failing link: the corrupted intersection's estimated subspaces would no longer support the exterior blocks.","tokens_in":17472,"feed_emoji":"🧩","tokens_out":17464,"duration_ms":145420,"temperature":0.7,"pith_summary":"The paper asks whether a third-order tensor of low tubal rank — meaning its frontal slices are low-rank matrices in the Fourier domain along the third mode — can still be recovered when observations are confined to a cross-concentrated sampling pattern (t-CCS), in which entries are seen only inside selected horizontal and lateral slices, and some of those observations are corrupted by sparse, arbitrarily large outliers. Its answer is yes, in the form of the R-ItCUR algorithm, which splits the sampled cross into three disjoint blocks, applies an adaptive blockwise Welsch correction to suppress outliers, and updates the low-rank component by projecting the exterior blocks onto subspaces taken from a rank-$r$ t-SVD of the intersection block. Because the algorithm keeps an implicit t-CUR representation, it never reconstructs the full tensor during the iterations, yielding substantial memory and computational savings. Experiments on synthetic tensors, cardiac MRI volumes, and three-dimensional seismic data show accurate recovery under corruption, with the robust version matching the nonrobust solver on clean data and outperforming full-tensor robust baselines under the same t-CCS observations.","feed_headline":"Corrupted low-rank tensors recoverable from cross-concentrated samples","feed_subtitle":"New method works on observed slices only, so recovery skips building the full tensor and withstands gross outliers.","key_machinery":"The load-bearing object is the t-CUR block decomposition of the sampled cross into three disjoint blocks — the horizontal exterior $X_{R'}$, the lateral exterior $X_{C'}$, and the intersection $X_U$ — together with the rank-$r$ truncated t-SVD of the intersection block, $X_U = W \\ast \\Theta \\ast V^\\top$, whose factors define tubal column and row subspaces onto which the exterior blocks are projected via $X_{R'} \\leftarrow W \\ast W^\\top \\ast \\widetilde{X}_{R'}$ and $X_{C'} \\leftarrow \\widetilde{X}_{C'} \\ast V \\ast V^\\top$. The second mechanism is the tube-wise $\\alpha$-sparsity outlier model, whose support survives the Fourier transform unchanged, paired with the blockwise Welsch correction $W_\\sigma(E) = (1 - \\exp(-E \\odot E / (2\\sigma^2))) \\odot E$, a smooth bounded robust loss whose scale $\\sigma$ is read from the empirical $(1-\\alpha)$-quantile of observed residual magnitudes. Together these pieces keep every update on the three stored blocks, so the full tensor is never assembled and the exact t-CUR identity $X = C \\ast U^\\dagger \\ast R$ provides the final reconstruction.","core_discovery":"The central claim is that robust low-tubal-rank tensor completion is achievable under t-CCS sampling provided the outliers are tube-wise sparse: each corrupted tube is entirely corrupted, and at most an $\\alpha$ fraction of tubes in every row and column is affected. The paper argues that this model is the right match for the t-SVD framework because tube-wise sparsity is preserved by the Fourier transform along the third mode, so every frontal slice of the outlier tensor shares one common support. On this basis, R-ItCUR alternates a blockwise Welsch correction, whose scale is set adaptively by an empirical quantile rule that needs only the contamination fraction $\\alpha$, with projected blockwise gradient descent on the three cross blocks. The projection is the load-bearing step: a rank-$r$ truncated t-SVD of the intersection block, $X_U^\\ell = W^\\ell \\ast \\Theta^\\ell \\ast (V^\\ell)^\\top$, yields tubal column and row subspaces, and the exterior blocks are projected onto these subspaces, producing the next t-CUR reconstruction $X^\\ell = X_C^\\ell \\ast (X_U^\\ell)^\\dagger \\ast X_R^\\ell$. The experiments report that this cross-only procedure recovers tensors as accurately as the nonrobust solver on clean data, that accuracy improves as outlier magnitude grows, and that a generic robust full-tensor method degrades under t-CCS, indicating that the cross structure itself must be exploited.","pith_inferences":["Editorial note on the paper's own caveat: in Section 5.1 the authors state that most synthetic trials reach the 100-iteration cap, so the reported curves establish stable residual reduction rather than strict convergence under the stopping tolerance; steady-state behavior, not certified convergence, is what the numerics support.","The subspace-projection step is the natural target for a follow-up theory: if the intersection block satisfied an incoherence or restricted-isometry condition for the t-SVD, the exterior projections might admit a linear-convergence guarantee, an argument this paper does not attempt.","The tube-wise sparsity model sets a real boundary: corruptions scattered as individual entries rather than whole tubes would lose the common-support property across Fourier slices that the analysis relies on, so an entrywise outlier model would likely need a different correction.","Because every update touches only the three stored blocks, the same skeleton — Welsch correction plus subspace projection — could transfer to other sampling designs with an overlap block, such as multiple crosses or fiber-concentrated patterns; the paper does not explore this."],"forward_implications":["Robust t-CCS completion is possible without ever forming the full tensor, so memory and per-iteration cost scale with the sampled cross rather than with the ambient size $n_1 n_2 n_3$.","Setting the contamination fraction $\\alpha = 0$ reduces R-ItCUR to the nonrobust t-CUR solver, and the reported PSNR values coincide in the clean case on both MRI and seismic data, so robustness is added without sacrificing clean-case accuracy.","Under identical t-CCS observations R-ItCUR consistently outperforms a robust full-tensor method, while that same full-tensor method scores higher under uniform sampling, indicating that exploiting the cross-concentrated geometry is itself a source of robustness.","Recovery accuracy improves as outlier magnitude grows — median relative error on synthetic tensors falls from $3.44 \\times 10^{-2}$ at magnitude factor $c=10$ to $2.20 \\times 10^{-2}$ at $c=200$ — because larger outliers separate more cleanly from the low-rank signal under the Welsch correction.","The adaptive quantile rule sets the Welsch scale from the contamination fraction alone, so the user does not need to know the outlier magnitudes in advance."],"supporting_citations":[{"why":"Supplies the t-CCS sampling model, the exact t-CUR decomposition (Theorem 2.4), and the nonrobust ITCURTC solver that R-ItCUR extends.","marker":"[19]"},{"why":"Establishes the matrix analogue that CCS sampling remains robust to sparse outliers and supplies the synthetic tensor generation protocol used in Section 5.1.","marker":"[20]"},{"why":"Defines the t-product, t-SVD, and tubal rank that the entire low-rank model and all t-SVD projections rest on.","marker":"[2]"},{"why":"Provides the t-CUR decomposition perspective underlying the implicit representation and the reconstruction identity X = C * U^† * R.","marker":"[16]"},{"why":"Introduces matrix cross-concentrated sampling, the paradigm that t-CCS generalizes to third-order tensors.","marker":"[18]"},{"why":"The convex robust low-tubal-rank completion method (Robust-IHT) used as the full-tensor comparison baseline that R-ItCUR must beat under identical observations.","marker":"[43]"},{"why":"Introduces the Welsch robust loss that the blockwise correction in Section 4.1 is built on.","marker":"[48]"},{"why":"Unifies outlier rejection with robust statistics, motivating the smooth Welsch correction as a stable alternative to hard support estimates.","marker":"[49]"}],"fun_headline_variants":["Cross-sampled tensor completion withstands gross outliers","Robust tensor recovery from cross-section samples only","Outlier-robust tensor completion on cross-concentrated data","Low-tubal-rank tensor recovery robust to sparse corruptions","R-ItCUR: robust tensor completion without full-matrix construction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the rank-$r$ subspaces estimated from the partially observed, possibly corrupted intersection block are accurate enough to serve as projection bases for the two exterior blocks; if the intersection is too small, too sparsely observed, or too heavily corrupted, errors made there propagate through the subspace projection to the whole reconstruction, and this step is a heuristic without an accompanying guarantee.","fun_headline_variants_meta":{"raw":{"variants":["Cross-sampled tensor completion withstands gross outliers","Robust tensor recovery from cross-section samples only","Outlier-robust tensor completion on cross-concentrated data","Low-tubal-rank tensor recovery robust to sparse corruptions","R-ItCUR: robust tensor completion without full-matrix construction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002005,"raw_usage":{"total_tokens":7876,"prompt_tokens":1051,"completion_tokens":6825,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":6741}},"tokens_in":667,"tokens_out":6825,"duration_ms":48135,"temperature":1.0,"reasoning_tokens":6741,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T00:34:08.841790+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic low-tubal-rank tensor of size roughly $120 \\times 120 \\times 8$, fix the t-CCS cross fraction small enough that the intersection block holds only a few slices, and place all gross outliers inside that intersection rather than spreading them across the cross. If R-ItCUR's relative error then rises well above the clean-case level while a full-tensor robust method on the same observations stays accurate, the blockwise subspace-projection heuristic is the failing link: the corrupted intersection's estimated subspaces would no longer support the exterior blocks.","supporting_citations":[{"cited_title":"Guaranteed Sampling Flexibility for Low-tubal-rank Tensor Completion","cited_arxiv_id":"2406.11092","evidence_quote":"Supplies the t-CCS sampling model, the exact t-CUR decomposition (Theorem 2.4), and the nonrobust ITCURTC solver that R-ItCUR extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces matrix cross-concentrated sampling, the paradigm that t-CCS generalizes to third-order tensors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The convex robust low-tubal-rank completion method (Robust-IHT) used as the full-tensor comparison baseline that R-ItCUR must beat under identical observations."}],"review_version":2}