{"id":"f4263e00-8d09-4a23-82bb-5f7382053df2","arxiv_id":"2501.04565","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"RTPCA-SGD applies ScaledGD to t-SVD robust tensor PCA, claiming linear convergence independent of condition number, with a self-supervised deep-unfolded variant for learned parameters.","lead":"This paper introduces a scaled gradient descent method for separating low-rank and sparse parts of 3D tensor data in the t-SVD framework, with a claimed recovery guarantee. It also unrolls the iterations into a self-supervised network to learn step sizes, and tests it on video denoising and background initialization.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sparse-error bounds in Lemma 12 silently drop I1 I2 factors that the stated α-sparsity condition cannot absorb, so Theorem 1's induction is not established for generic dimensions.","rationale":"The reader's weakest_assumption identifies exactly the same algebraic gap: the I1 I2 factors introduced by combining the α-sparse spectral bound (Lemma 5) with the threshold ζ_{k+1} disappear from the displayed proof, and keeping them forces a much stronger sparsity condition. My independent reading of the supplementary material confirms the gap in Lemma 12's bounds for R2, R3, and R4. This is not a disagreement with an existing consensus; it is an internal correctness issue in the proof of the paper's headline guarantee. The numerical experiments use corruption levels around α = 0.1, which already lie orders of magnitude above even the paper's quoted theoretical bound; although experiments can exceed sufficient conditions, they cannot rescue a missing induction step. The deep-unfolding experiments also lack reproducibility details, but the decisive issue is the unproved convergence theorem. A revised proof with a corrected α dependence might eventually support a conditional acceptance, but as written the central claim is not established, so the reader's REJECT verdict is appropriate.","tokens_in":41236,"tokens_out":2958,"duration_ms":29246,"concrete_test":"Re-derive Eq. (41) without dropping the I1 I2 factors in the R2, R3, and R4 bounds of Lemma 12. Then evaluate the contraction prefactor for a concrete allowed instance, e.g. I1 = I2 = 100, I3 = 50, R = 5, μ = 1, ε = 0.02, η = 2/3, and α = 1/(10^4 μ R^{1.5} I3^{1.5}); if the resulting prefactor on dist^2 exceeds (1 − 0.8η)^2, the induction step in Eq. (41) fails and Theorem 1's stated α bound is insufficient.","verdict_should_be":"REJECT","load_bearing_attack":"The load-bearing gap is in the proof of Lemma 12, which supplies the contraction step for Theorem 1. In bounding the sparse-error terms R2, R3, and R4, the proof uses Lemma 5, giving ∥∆S∥2 ≤ α I3 sqrt(I1 I2) ∥∆S∥∞, together with ζ_{k+1} = 3 sqrt(I1 I2) μ R σmin(X⋆) τ^k. The product of these two quantities contains I1 I2 (and I1^2 I2^2 for R4), yet the displayed bounds on the supplementary pages 23–24 read 12 I3 η(1−η) α μ R^{1.5} τ^k σmin(X⋆) ∥∆LΣ^{1/2}∥F and 36 I3^2 η^2 α^2 μ^2 R^3 τ^{2k} σmin^2(X⋆), with the dimension factors I1 and I2 silently removed. Retaining them, the second-to-last line of Eq. (41) acquires extra terms proportional to I1 I2 sqrt(I3) α μ R^{1.5}/ε and I1^2 I2^2 I3^2 α^2 μ^2 R^3/ε^2, up to constants. With the stated assumption α ≤ 1/(10^4 μ R^{1.5} I3^{1.5}), these become on the order of c I1 I2 sqrt(I3)/ε and c I1^2 I2^2, which are enormous for even moderate dimensions such as I1 = I2 = 100, I3 = 50. The induction in Eq. (41) therefore requires a much smaller corruption fraction, roughly α = O(1/(μ R I1^2 I2^2 I3)), not the bound claimed in Theorem 1. Because this step controls both dist(L_{k+1}, R_{k+1}; L⋆, R⋆) and, through Lemma 13, the entrywise error and support recovery used in Theorem 1, the central 'guaranteed recovery' claim is not proven as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RTPCA-SGD, a scaled gradient descent method for t-SVD based robust tensor PCA, together with a deep-unfolding variant RTPCA-LSGD. The main theoretical claim (Theorem 1) is that under tensor µ-incoherence and α-t-sparsity with α ≤ 1/(10^4 μ R^{1.5} I3^{1.5} κ), the iterates converge linearly to the true low-rank tensor with a rate independent of the condition number, and that the sparse component is exactly support-recovered. The proof is organized as a spectral initialization guarantee (Theorem 3) plus a local linear convergence result (Theorem 2), with auxiliary lemmas in the supplementary material. The paper also reports synthetic and real-data experiments on video denoising and background initialization.","tokens_in":41662,"tokens_out":9661,"duration_ms":89982,"significance":"If the main theorem were established, this would be a useful contribution: it would provide the first ScaledGD-type recovery guarantee in the t-SVD framework, with a condition-number-independent linear rate and per-iteration cost O(I1 I2 I3 log I3 + ceil((I3+1)/2) I1 I2 R). The self-supervised deep-unfolding extension is a practical and timely addition, and the experimental section is extensive and mostly well presented. However, the central theoretical guarantee is not established as written: the proof of the key contraction step silently drops dimension factors that the stated sparsity condition cannot absorb. Since the advertised contribution is precisely the guaranteed recovery theorem, this gap is load-bearing for the paper's main claim.","major_comments":[{"comment":"The bounds for the sparse-error terms R2, R3, and R4 drop the I1 I2 factors that arise from the threshold ζ_{k+1} = 3 sqrt(I1 I2) μ R σ_min(X⋆) τ^k together with Lemma 5's bound ∥ΔS∥_2 ≤ α I3 sqrt(I1 I2) ∥ΔS∥_∞. For example, the R2 bound as displayed is proportional to I3 α μ R^{1.5} τ^k σ_min ∥ΔL Σ^{1/2}∥_F, but the preceding line contains I3 sqrt(I1 I2) sqrt(R) ∥ΔS∥_∞ and ∥ΔS∥_∞ ≤ 2 ζ_{k+1}, producing an extra factor I1 I2. The same omission occurs in R3, and in R4 the omitted factor is I1^2 I2^2. If these factors are retained, the second-to-last display of Eq. (41) gains terms proportional to I1 I2 sqrt(I3) α μ R^{1.5}/ε and I1^2 I2^2 I3^3 α^2 μ^2 R^3/ε^2 (up to constants), which are enormously larger than the displayed coefficients under the stated assumption α ≤ 1/(10^4 μ R^{1.5} I3^{1.5}) for generic dimensions such as I1 = I2 = 100, I3 = 50. Consequently the induction contraction in Eq. (41) is not established, and Theorem 1's linear convergence guarantee does not follow from the proof as written.","section":"Supplementary, Lemma 12 (pages 23–24), Eq. (41)"},{"comment":"The same dimension-factor omission appears in the entrywise-error propagation. The displayed bound T3 ≤ 6 η sqrt(α I3)/(1−ε) · (1/sqrt(I1)) μ R τ^k σ_min(X⋆) is obtained from ∥ΔS∥_{2,∞} ≤ sqrt(α I2 I3) ∥ΔS∥_∞ and ∥ΔS∥_∞ ≤ 2 ζ_{k+1}, but substituting ζ_{k+1} = 3 sqrt(I1 I2) μ R τ^k σ_min gives an additional factor I1 I2. After multiplying by sqrt(I1), the retained bound is proportional to sqrt(α) I1 I2 sqrt(I3) μ R τ^k σ_min rather than sqrt(α I3) μ R τ^k σ_min. This affects the proof of the ∥X_k − X⋆∥_∞ and ∥S_k − S⋆∥_∞ claims in Theorem 2 and hence the corresponding statements in Theorem 1.","section":"Supplementary, Lemma 13 (page 25), bound on T3"}],"minor_comments":[{"comment":"The wording of Assumption 2 ('at most αt fraction of non-zero element in each slices') is imprecise; it should say that each slice contains at most an α fraction of nonzero entries, and the relation between the notation α_t in Assumption 2 and α in Theorem 1 should be stated explicitly.","section":"Assumption 2 and Theorem 1"},{"comment":"The proof cites a matrix bound as '[?, Lemma 1]'; the placeholder reference should be replaced with a proper citation, since Lemma 5 is used in the main proof.","section":"Supplementary, Lemma 5"},{"comment":"The caption of Figure 3 refers to 'ScaleGD iterative steps', but the method is called ScaledGD; this typo should be fixed.","section":"Algorithm 2 and Figure 3"},{"comment":"The relation ζ_{k+1} = τ^k ζ_1 for k ≥ 1 is stated, while the text elsewhere writes ζ_{k+1} = τ ζ_k; the indexing for k = 0 should be clarified.","section":"Algorithm 2 and Theorem 1, threshold indexing"},{"comment":"Lemma 12 concludes with τ = 1 − 0.8η, whereas Theorem 2 and Theorem 1 state τ = 1 − 0.6η; the proof says 'substituting τ = 1 − 0.8η ≤ 1 − 0.6η', but the same symbol τ is used for the contraction rate and the threshold decay, which is confusing and should be disambiguated.","section":"Lemma 12 and Theorem 2, rate notation"}],"recommendation":"reject","confidential_remarks":"The paper is not circular and does not misuse its own results to define the target; the problem is different. The main theorem is not established because the supplementary proof of Lemma 12 omits dimension factors in the sparse-error bounds. This is a load-bearing error in the paper's central claim, and the required correction—making α depend polynomially on I1 and I2, e.g. roughly α = O(1/(μ R I1^2 I2^2 I3))—substantially changes the advertised 'mild assumptions' guarantee. I would encourage the authors to rework the proof and restate the theorem accordingly in a future submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a serious attempt to extend ScaledGD to t-SVD robust tensor PCA, with a genuinely weaker slice-wise sparsity assumption and a full convergence proof. But the proof of the main theorem has a load-bearing gap: in Lemma 12 the sparse-error terms R2, R3, R4 pick up I1 I2 factors from the threshold zeta_{k+1} that are silently dropped in the displayed bounds. I checked the supplementary pages 23-24; the step from the line with alpha I3 sqrt(I1I2) zeta_{k+1} to the next line simply loses the I1 I2 product. Keeping those factors, the induction in Eq. (41) would need alpha <= O(1/(mu R I1^2 I2^2 I3)), not the stated alpha <= 1/(10^4 mu R^1.5 I3^1.5). So Theorem 1 as stated is not proven.\n\nWhat the paper does well: the algorithm is novel for t-SVD RTPCA, the slice-wise sparsity condition is a real improvement over the matrix and Tucker conditions, the proof structure is largely coherent at the level of lemmas, and the experimental section is thorough. The deep-unfolded variant is a nice practical touch, though the reproducibility details are thin.\n\nTwo other soft spots: the synthetic experiments run at corruption levels orders of magnitude above the theorem's bound, so the phase transitions don't actually validate the theoretical claim; and the deep-unfolding results lack code and training details. Both are minor compared to the proof gap.\n\nWho gets value from this: researchers working on ScaledGD variants or t-SVD recovery will find the algorithm and the proof framework useful, even if the guarantee needs repair. It deserves a serious referee: the question is worth an answer, and the gap is fixable, but the current version should not be accepted with this theorem. My recommendation: send it to review, but with the expectation that the authors either fix the dimension dependence or scale back the claims.","headline":"A genuinely useful algorithm and a real sparsity condition, but the main recovery guarantee drops I1 I2 factors and is not proven as stated.","tokens_in":42256,"tokens_out":4331,"would_cite":false,"duration_ms":37445,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69","90C26"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that robust tensor PCA can be solved by scaled gradient descent on a t-SVD factorization, with a linear convergence rate independent of the condition number.","keywords":["robust tensor PCA","scaled gradient descent","t-SVD","tubal rank","tensor incoherence","deep unfolding","self-supervised learning","linear convergence"],"falsifier":"Inspect the contraction estimate in the proof of Lemma 12: combining the slice-sparsity spectral bound of Lemma 5 with $\\zeta_{k+1}=3\\sqrt{I_1I_2}\\,\\mu R\\sigma_{\\min}\\tau^k$ introduces $I_1$ and $I_2$ factors that the proof drops, and keeping them would force $\\alpha$ down to about $O(1/(\\mu R I_1^2 I_2^2 I_3))$. Numerically, run RTPCA-SGD with growing $I_1=I_2$ at $\\alpha$ just below the theorem's stated value and check whether the measured error still contracts at rate $\\tau$; a stall or slowdown with increasing spatial dimensions would falsify the guarantee as stated.","tokens_in":41003,"feed_emoji":"🧮","tokens_out":14777,"duration_ms":127522,"temperature":0.7,"pith_summary":"This paper aims to establish that robust tensor principal component analysis (RTPCA) under the t-SVD framework can be solved by factored scaled gradient descent rather than by convex tensor-nuclear-norm (TNN) minimization. The proposed RTPCA-SGD writes the low-rank tensor as $X = L * R^\\top$ and alternates soft-thresholding of the sparse component with scaled gradient updates of the two factors. The paper's central theoretical claim is that, under tensor $\\mu$-incoherence and a slice-based sparsity condition, the iterates converge linearly to the true low-rank and sparse components at a rate independent of the condition number. A self-supervised deep-unfolding version, RTPCA-LSGD, learns the step size and threshold parameters from the observed data alone. If the guarantee is correct, the practical payoff is a scalable RTPCA that avoids full t-SVD computations in every iteration while retaining recovery guarantees.","feed_headline":"Scaled gradient descent provably separates low-rank and sparse tensors","feed_subtitle":"A factored t-SVD update avoids costly nuclear-norm steps and converges linearly regardless of conditioning.","key_machinery":"The central object is the tensor distance metric $$\\operatorname{dist}(L,R;L_\\star,R_\\star) = \\inf_{Q\\in GL(R)}\\left(\\|(L*Q-L_\\star)*\\Sigma_\\$star^{{1/2}}$\\|$_F^{2}$ + \\|(R*$Q^{{-\\top}}$-R_\\star)*\\Sigma_\\$star^{{1/2}}$\\|$_F^{2}$\\right)^{1/2},$$ which removes the gauge ambiguity between the two factors by optimizing over invertible alignment tensors $Q$; Lemma 1 guarantees the optimal $Q$ exists and is invertible when the factors are close. The scaled gradient updates $$L_{k+1}=L_k-\\eta\\,(L_k*R_k^\\top+S_{k+1}-Y)*R_k*(R_k^\\top R_k)^{-1}, \\quad R_{k+1}=R_k-\\eta\\,(L_k*R_k^\\top+S_{k+1}-Y)^\\top*L_k*(L_k^\\top L_k)^{-1}$$ use the scaling factors $(R_k^\\top R_k)^{-1}$ and $(L_k^\\top L_k)^{-1}$ to remove the dependence on the condition number. The threshold schedule $\\zeta_{k+1}=3\\sqrt{I_1I_2}\\,\\mu R\\sigma_{\\min}(X_\\star)\\tau^k$ keeps the soft-thresholding step above $\\|X_\\star-X_k\\|_\\infty$, so by Lemma 3 it projects onto the support of $S_\\star$; the t-product algebra (circular convolution along the third mode, computed as matrix products in the Fourier domain) is what makes the whole factorization and distance calculus possible.","core_discovery":"Theorem 1 states that when the true low-rank tensor $X_\\star$ has tubal rank $R$ and satisfies tensor $\\mu$-incoherence, and the sparse tensor $S_\\star$ is $\\alpha$-sparse with $\\alpha \\le 1/(10^4\\mu R^{1.5} I_3^{1.5}\\kappa)$, then with threshold values $\\zeta_0 \\in [\\|X_\\star\\|_\\infty, 2\\|X_\\star\\|_\\infty]$, $\\zeta_1 = 3\\sqrt{I_1I_2}\\mu R\\sigma_{\\min}(X_\\star)$, and step size $\\eta \\in [1/4, 2/3]$, the RTPCA-SGD iterates satisfy $$\\|X_k - X_\\star\\|_F \\le 0.03\\sqrt{I_3}\\,\\tau^k \\sigma_{\\min}(X_\\star), \\quad \\|X_k-X_\\star\\|_\\infty \\le 3\\sqrt{I_1I_2}\\,\\mu R\\tau^k\\sigma_{\\min}(X_\\star), \\quad \\|S_k-S_\\star\\|_\\infty \\le 6\\sqrt{I_1I_2}\\,\\mu R\\$tau^{{k-1}}$\\sigma_{\\min}(X_\\star),$$ with rate $\\tau = 1-0.6\\eta$ independent of the condition number $\\kappa$. The theorem also claims that the support of every sparse iterate lies inside the support of $S_\\star$, so the thresholding step never introduces false outliers. The proof rests on a spectral initialization through the top-$R$ t-SVD, a threshold schedule that stays above the current entrywise error, and a tensor distance metric that measures the two factors under their best alignment.","pith_inferences":["The same factorized scaled-gradient machinery should transfer to other invertible transforms and to higher-order t-SVD generalizations, since the distance metric and thresholding arguments depend mainly on the norm identities of the t-product, not on the specific Fourier implementation.","Because the pure algorithm needs $\\mu R\\sigma_{\\min}(X_\\star)$ to set $\\zeta_1$, the deep-unfolded model is implicitly learning a spectral quantity of the unknown tensor; this likely explains why it gains most in heavily corrupted settings where hand-set thresholds are unreliable.","A concrete extension would apply RTPCA-LSGD to tensor completion or tensor regression, where the same low-rank factorization and scaled updates can be used without a sparse component."],"forward_implications":["The per-iteration cost drops from $O(I_1I_2I_3\\log I_3 + \\lceil (I_3+1)/2\\rceil I_1I_2\\min(I_1,I_2))$ for TNN-based RTPCA to $O(I_1I_2I_3\\log I_3 + \\lceil (I_3+1)/2\\rceil I_1I_2R)$ when $R\\ll\\min(I_1,I_2)$.","If Theorem 1 is correct, the recovery error contracts linearly at rate $\\tau=1-0.6\\eta$ regardless of the condition number $\\kappa$, so ill-conditioned tensors are recovered as quickly as well-conditioned ones.","The support containment $\\mathrm{supp}(S_k)\\subseteq\\mathrm{supp}(S_\\star)$ means the sparse estimate never invents nonzero entries outside the true corruption support once the threshold schedule is active.","The $\\alpha$-sparsity assumption applies slice-wise rather than fiber-wise, so it covers corruption patterns that would be fully dense under matrix row/column sparsity.","The self-supervised deep-unfolded variant RTPCA-LSGD can learn $\\zeta_0,\\zeta_1,\\tau,\\eta$ from observed data alone, removing the need for labeled ground truth in practical deployment."],"supporting_citations":[{"why":"Supplies the matrix ScaledGD update and the condition-number-free convergence proof template that the tensor method adapts.","marker":"[43]"},{"why":"Provides the prior Tucker-decomposition extension of ScaledGD for tensors, the comparison for sparsity assumptions, and the proof-strategy baseline.","marker":"[44]"},{"why":"Defines the t-SVD/TNN RTPCA model, the convex baseline, and the definitions and complexity figures the paper contrasts.","marker":"[10]"},{"why":"Introduces the t-product, identity/inverse/orthogonal tensors, and the algebra underlying the factorization.","marker":"[48]"},{"why":"Supplies the tensor incoherence conditions and tensor-basis framework used in Assumption 1.","marker":"[50]"},{"why":"Provides the deep-unfolded RPCA-SGD architecture that RTPCA-LSGD adapts to the tensor setting.","marker":"[45]"},{"why":"Supplies the self-supervised loss and unfolding recipe used by the learnable model.","marker":"[46]"}],"fun_headline_variants":["Robust tensor PCA: scaled gradient descent without nuclear norm","Scaled gradient descent for tensor PCA: guaranteed linear convergence","Learnable tensor PCA via scaled gradient, with recovery guarantees","Skip tensor nuclear norm: scaled gradient descent with guarantees","Tensor PCA: scaled gradient descent, linear rate independent of conditioning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the sparse-error terms in the scaled gradient update can be bounded by constants independent of $I_1$ and $I_2$; if the dimension factors introduced by the threshold schedule are kept, the stated guarantee would require a much smaller corruption fraction, roughly $O(1/(\\mu R I_1^2 I_2^2 I_3))$, than the $\\alpha \\le 1/(10^4\\mu R^{1.5}I_3^{1.5}\\kappa)$ stated in Theorem 1.","fun_headline_variants_meta":{"raw":{"variants":["Robust tensor PCA: scaled gradient descent without nuclear norm","Scaled gradient descent for tensor PCA: guaranteed linear convergence","Learnable tensor PCA via scaled gradient, with recovery guarantees","Skip tensor nuclear norm: scaled gradient descent with guarantees","Tensor PCA: scaled gradient descent, linear rate independent of conditioning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000957,"raw_usage":{"total_tokens":4164,"prompt_tokens":1119,"completion_tokens":3045,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":735,"completion_tokens_details":{"reasoning_tokens":2964}},"tokens_in":735,"tokens_out":3045,"duration_ms":24119,"temperature":1.0,"reasoning_tokens":2964,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:30:46.043657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the contraction estimate in the proof of Lemma 12: combining the slice-sparsity spectral bound of Lemma 5 with $\\zeta_{k+1}=3\\sqrt{I_1I_2}\\,\\mu R\\sigma_{\\min}\\tau^k$ introduces $I_1$ and $I_2$ factors that the proof drops, and keeping them would force $\\alpha$ down to about $O(1/(\\mu R I_1^2 I_2^2 I_3))$. Numerically, run RTPCA-SGD with growing $I_1=I_2$ at $\\alpha$ just below the theorem's stated value and check whether the measured error still contracts at rate $\\tau$; a stall or slowdown with increasing spatial dimensions would falsify the guarantee as stated.","supporting_citations":[{"cited_title":"Fast and provable tensor robust principal component analysis via scaled gradient descent,","cited_arxiv_id":null,"evidence_quote":"Provides the prior Tucker-decomposition extension of ScaledGD for tensors, the comparison for sparsity assumptions, and the proof-strategy baseline."},{"cited_title":"Tensor robust principal component analysis with a new tensor nuclear norm,","cited_arxiv_id":null,"evidence_quote":"Defines the t-SVD/TNN RTPCA model, the convex baseline, and the definitions and complexity figures the paper contrasts."},{"cited_title":"Exact tensor completion using t-SVD,","cited_arxiv_id":null,"evidence_quote":"Supplies the tensor incoherence conditions and tensor-basis framework used in Assumption 1."},{"cited_title":"Deep unfolded tensor robust PCA with self-supervised learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the self-supervised loss and unfolding recipe used by the learnable model."}],"review_version":1}