{"id":"48b08074-715c-4cc9-84bb-e342d0d2915e","arxiv_id":"1908.04924","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TTPUDR tensorizes the LPP mapping with tensor-trains and a Frobenius-norm objective, claiming robust and storage-efficient dimensionality reduction for high-dimensional tensor data.","lead":"The paper introduces TTPUDR, a dimensionality reduction method that combines tensor-train decomposition with locality preserving projections to handle extremely high-dimensional tensor data. It reports higher classification accuracy than PCA and LPP on face and hyperspectral datasets, especially when training samples are scarce and noise levels are high.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unproven convergence of the alternating reweighted scheme (Eqs. 6-7) leaves the reported robustness gains unverified; the paper itself admits no convergence proof.","rationale":"I agree with the reader's weakest assumption. The reweighted scheme in Eqs. (6)-(7) is the computational core of TTPUDR, and its convergence is explicitly left unproven in the manuscript ('there has not been any perfect theoretical proof of the convergence of TTPUDR'). This admission is not merely a stylistic caveat: the algorithm is an IRLS-type fixed-point iteration with solution-dependent weights, and no descent argument is given for the original F-norm objective (5). In particular, division by the projected distance in Eq. (6) can become infinite if two distinct data points are mapped to the same point, and there is no damping or safeguard mentioned. Without a convergence guarantee, the experimental results in Tables I-II and Figure 1 may reflect properties of the initialization and iteration count rather than the solution of the stated robust objective. This makes the central claim unverifiable. Other concerns, such as the false claim that the F-norm approximates the ℓ1-norm and the omission of TLPP/TTNPE baselines, are real but secondary: even a correct motivation and full baselines would not rescue the method if the optimization does not reliably solve the objective.","tokens_in":11523,"tokens_out":15244,"duration_ms":145823,"concrete_test":"Re-run Algorithm 1 on the Extended Yale B dataset with the reported settings (cores 1×4×4, 4×8×7, 7×4×4, 4×8×R_n, R_n=28, k=4, t=0.5, 60% training). Record the value of the original F-norm objective (5) at each outer iteration and the final classification accuracy for at least 10 random initializations. If the objective is not monotonically non-increasing or the final objective/accuracy varies by more than a few percent across initializations, the algorithm's convergence is not established and the claimed robustness gains may be artifacts. Also monitor the denominator in Eq. (6) for zero or near-zero values.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that Algorithm 1 converges to a minimizer of the F-norm objective in Eq. (5). The algorithm computes weights ~s_ij = s_ij / ||L^T(~U)V(X_i) - L^T(~U)V(X_j)||_F (Eq. 6), fixes them, and minimizes the weighted squared objective (Eq. 7) by block coordinate descent over the cores on Stiefel manifolds (Problems 10-12). Because the weights depend on the current solution, the sequence of weighted problems is not a global majorization of (5); there is no guarantee that the original objective decreases, and a zero or near-zero denominator in Eq. (6) can produce unbounded weights. The paper explicitly concedes in Section III that 'there has not been any perfect theoretical proof of the convergence of TTPUDR.' If the iteration does not reliably converge to a stationary point of (5), the reported classification improvements over PCA and LPP cannot be attributed to the robust objective and may instead reflect initialization or early stopping. This is load-bearing because the central claim of robust ultra-dimensionality reduction rests on the algorithm actually solving (5).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TTPUDR, a dimensionality reduction method that parameterizes an LPP-style projection as a tensor-train and replaces the squared-Frobenius LPP objective with a non-squared Frobenius-norm objective (Problem (5)), claimed to be a robust approximation to the ℓ1-norm. The optimization alternates between reweighting the affinity matrix via Equation (6) and solving weighted squared-Frobenius subproblems (Problems (10)–(12)) on Stiefel manifolds. Experiments on Extended Yale B and Indian Pines compare TTPUDR with PCA and LPP, reporting better accuracy on face data and better robustness in a small-sample noisy setting.","tokens_in":11790,"tokens_out":8356,"duration_ms":88472,"significance":"The idea of using a tensor-train parameterization to compress the projection operator is potentially useful, and the reported storage complexity O(nIR^2) is a concrete advantage over the Tucker-based TLPP. The empirical gain on Extended Yale B is large and, if reproducible, would be of interest. However, the paper's central theoretical motivation is mathematically incorrect (the Frobenius norm is not an approximation of the ℓ1-norm), the convergence of the proposed alternating scheme is explicitly unproved, and the experiments do not cover the claimed ultra-high-dimensional regime or the stated state-of-the-art baselines. No code or data are released, and the evaluation protocol contains internal inconsistencies. As it stands, the support for the central claims is insufficient.","major_comments":[{"comment":"The statement that \"[a]n approximation to the ℓ1-norm is the Frobenius norm\" is false: by the paper's own definition, ‖X‖_F = ‖V(X)‖_2, which is an ℓ2-norm, not an ℓ1-norm. This false premise is load-bearing because Contribution 2 and the robustness claims attribute the method's outlier resistance to this alleged ℓ1 approximation. The non-squared Frobenius objective in Problem (5) is a legitimate robust alternative to the squared Frobenius norm, but it is not an ℓ1-type objective, and the paper should either establish the robustness property by a correct argument or substantially weaken the claim.","section":"Section II.A"},{"comment":"The convergence of the alternating reweighted scheme is unproved and, as written, not guaranteed. The weights ~s_ij in Equation (6) depend on the current tensor cores, and fixing them in Problem (7) does not produce a global majorization of Problem (5), so the original objective need not decrease monotonically. A zero or near-zero denominator in Equation (6) can also produce unbounded weights. The paper itself concedes that \"there has not been any perfect theoretical proof of the convergence of TTPUDR.\" Without a proof of convergence to a stationary point of Problem (5), or at least a rigorous empirical convergence analysis with multiple initializations, the reported classification gains cannot be attributed to solving the stated objective rather than to initialization or early stopping.","section":"Section III, Algorithm 1"},{"comment":"The experimental evaluation does not support the abstract's claim that TTPUDR \"significantly outperforms the past methods and the several state-of-the-art methods.\" Only PCA and LPP are compared; TTNPE and TLPP are dismissed without running. On the Indian Pines dataset (Table I), TTPUDR is substantially worse than both PCA and LPP (OA 0.7101 versus 0.7907 and 0.7810). Moreover, the first two experimental settings have data dimension smaller than the number of samples (200 < 10,366 and 1024 < 2414), so they do not test the ultra-high-dimensional regime that is the paper's stated motivation; only the 20%-training noisy case (482 samples vs. 1024 dimensions) violates LPP's assumption, and in that case the LPP baseline is absent. The evidence for the central performance claims is therefore incomplete and partly contradictory.","section":"Section IV"},{"comment":"The output dimension R_n is defined inconsistently. In the text preceding Problem (5), R_n is set to I_1 I_2 ... I_n, which would mean the projection has the same dimensionality as the vectorized input and no reduction occurs. Algorithm 1 line 1 and the experiments instead use R_n = 2, ..., 30 as the number of features. Since the subproblems (10)–(12) depend critically on the meaning of R_n (e.g., the slices Y_k(:, :, r_n, :, :) and the constraint L^⊤(U_n)L(U_n) = I_{R_n}), this inconsistency affects the mathematical formulation of the method and must be resolved.","section":"Section III, Eqs. (5)–(12)"}],"minor_comments":[{"comment":"The data-split description for the noised dataset is contradictory: \"60% and 20% of the data are treated to be the training set and 40% and 80% of the data are set as the test set\" does not define coherent experimental settings and should be rewritten.","section":"Section IV.A"},{"comment":"The evaluation metrics are nonstandard: the average accuracy formula uses TP/(TP+FP) rather than per-class recall, and the Kappa coefficient formula is not the usual Cohen's kappa. Since all methods are evaluated with the same formulas, this may not change relative rankings, but it should be corrected or justified.","section":"Section IV.B"},{"comment":"The results in Tables I and II are reported for a single randomly selected output dimension (R_n = 24 and R_n = 28), while the text says R_n ranges from 2 to 30. Reporting only one dimension makes the comparison fragile and does not show the behavior across the claimed range; the authors should present results across all R_n or justify the selection.","section":"Section IV.C.1"},{"comment":"The claim that the Frobenius norm \"preserves the general spatial relations between elements\" while the ℓ1-norm \"treats each element separately and can probably cause the spatial information loss\" is not substantiated. Both norms are elementwise functions of the tensor entries; neither explicitly exploits spatial layout. If the intention is that the projection operator itself preserves tensor structure, that point should be argued separately.","section":"Section II.A"},{"comment":"The statement that \"the tensor-train decomposition is proved to be able to efficiently and effectively capture the spatial relations\" is too strong without a specific reference or argument. The cited work [7] is about tensor train subspace and neighborhood preserving embedding, not a general proof of spatial-relation capture.","section":"Section I"}],"recommendation":"reject","confidential_remarks":"The manuscript has a useful high-level idea, but the current version contains a false mathematical claim at the core of the robustness motivation, an explicitly unproved convergence property for the central algorithm, and experiments that do not match the paper's stated scope. In my view these are load-bearing issues rather than presentation gaps. If the authors can replace the ℓ1-approximation claim with a correct robustness argument, prove or substantially validate convergence, and add the missing baselines and a proper ultra-high-dimensional evaluation, a fundamentally revised version could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the specific algorithm: projecting tensor data through a TT-parameterized mapping while optimizing an LPP-type objective with an F-norm (not squared) loss on a Stiefel manifold. That combination is distinguishable from TTNPE and TLPP, and the storage-efficiency argument is real—you get a subspace with far fewer parameters than full PCA or LPP when the tensor is high-order. The paper is also refreshingly explicit that a convergence proof is missing. Those are real pluses.\n\nThe soft spots, though, are serious. First, the mathematical motivation is wrong: the paper says the Frobenius norm approximates the ℓ1 norm. It does not. The F-norm of an error is the ℓ2 norm of its vectorized entries, not an approximation to ℓ1, and it is not more robust than ℓ1 in any standard sense. That is not a cosmetic issue; it is the stated reason for the objective. Second, the claimed robustness upgrades from the reweighted scheme in Eqs. (6)–(7) are unverified. The weights depend on the current solution, and the iteration is not a majorization of the original objective, so there is no guarantee of monotone decrease. You can construct cases where a zero or near-zero denominator blows the weights up. The authors concede this in Section III, but then the reported classification gains over PCA and LPP cannot be confidently attributed to solving (5).\n\nThe experiments are the weakest part. Only PCA and LPP are compared—TLPP and TTNPE are dismissed without running them, which may be fair given their computational cost, but it means the paper does not compare against the most relevant tensor-based baselines. On Indiana, TTPUDR is noticeably worse than PCA and LPP; on Yale B it looks much better, but that single bright spot is loaded with hand-chosen ranks, no error bars, and a contradictory description of how Rn was selected (\"randomly selected\" vs. \"best result\"). For the robustness tests, LPP is absent in the high-dimensional regime where it cannot run, so the \"siginificant outperformance\" claim rests on comparisons only with PCA.\n\nWhere does this leave us? There is a plausible algorithmic kernel here, and the paper is honest about its main theoretical gap. But as written, the central claim—that TTPUDR is robust because it solves an F-norm objective—is not verified. The false l1/F-norm statement needs correction, the convergence analysis needs at least a partial result or a proper IRLS-style justification, and the experiments need more baselines, error bars, and a transparent parameter-selection rule.\n\nIf it were my call, I would not desk-reject it. There is enough novelty and enough of a real problem being attacked that a knowledgeable referee could help the authors turn this into a decent paper. But I would send it out with the expectation that heavy revision, or a substantially stronger experimental section, is required before it can be accepted.","headline":"Genuinely new TT-plus-LPP combination with real storage-efficiency potential, but the robustness claim is undercut by a false l1/F-norm motivation, unproven convergence, and weak empirical support.","tokens_in":12279,"tokens_out":2287,"would_cite":false,"duration_ms":26663,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes TTPUDR, which tensorizes the locality-preserving projection as a tensor-train and swaps the squared-error objective for a Frobenius-norm one, yielding a low-parameter projection that preserves spatial structure and…","keywords":["tensor-train decomposition","locality preserving projections","dimensionality reduction","Frobenius norm","robustness to outliers","Stiefel manifold","high-dimensional tensor data","graph embedding"],"falsifier":"Run TTPUDR from many random tensor-core initializations on the noisy Yale B setting and compare the final objective values and classification accuracies; if different initializations settle at materially different values or the objective does not stabilize, the alternating reweighted scheme is not solving the claimed optimization problem.","tokens_in":11348,"feed_emoji":"🧮","tokens_out":8426,"duration_ms":73630,"temperature":0.7,"pith_summary":"The paper proposes a dimensionality reduction method, TTPUDR, that turns the classical locality preserving projections (LPP) mapping into a chain of small tensor cores—a tensor-train—instead of a single large matrix, and measures projection error with the Frobenius norm rather than the squared error. The goal is to make LPP work on ultra-high-dimensional tensor data without first flattening it to a vector, while using far fewer parameters and resisting extreme outliers. The paper reports that on face and hyperspectral image data TTPUDR beats PCA and LPP in classification accuracy when the data dimension exceeds the sample count, and that on data with added block noise it keeps accuracy much higher than the baselines. The load-bearing consequence is that high-order tensor structure can be exploited for dimensionality reduction without paying exponential storage or computation costs.","feed_headline":"Tensor-train LPP cuts parameters and resists extreme outliers","feed_subtitle":"Replacing squared error with a Frobenius-norm tensor-train map keeps spatial structure in ultra-high-dimensional data.","key_machinery":"The central object is the tensor-train (TT) parameterization: an $n$-mode tensor is written as a contraction of $n$ three-way cores $U_1,\\dots,U_n$, which costs $O(nIR^2)$ storage instead of the full tensor size. The argument is carried by replacing the LPP objective's squared Frobenius distance with the plain Frobenius norm and then rewriting it as a squared norm with data-dependent weights $\\tilde{s}_{ij}$; this turns each core update into a small orthogonality-constrained (Stiefel manifold) eigenproblem built from matrices $H_k$, and the affinity weights are refreshed after every sweep. The orthogonality constraints $L^{\\top}(U_k)L(U_k)=I_{R_k}$ guarantee the overall projection has orthogonal columns.","core_discovery":"On its own terms, the paper establishes that the LPP projection, normally a matrix solving a generalized eigenvalue problem, can be replaced by a tensor-train parameterized mapping $\\tilde{U}=U_1 \\times U_2 \\times \\cdots \\times U_n$ whose left unfolding $L(\\tilde{U})$ projects a tensor input by contraction. The squared $F$-norm in the standard LPP objective is replaced by an unsquared Frobenius norm, and the optimization is recast as a sequence of smaller subproblems over orthogonality-constrained tensor cores, with the similarity graph reweighted by $\\tilde{s}_{ij} = s_{ij}/\\|L^{\\top}(\\tilde{U})V(X_i)-L^{\\top}(\\tilde{U})V(X_j)\\|_F$ after each sweep. The paper shows this yields classification models with $O(nIR^2)$ effective parameters, operating where the data dimension is larger than the number of samples, and reports accuracy gains over PCA and LPP on the Extended Yale B face data and Indian Pines hyperspectral data, including under 10% and 20% block noise.","pith_inferences":["Editorial inference: the $\\tilde{s}_{ij}$ reweighting is structurally an iteratively reweighted least-squares scheme, so a proof of monotone decrease or fixed-point convergence would turn TTPUDR into a template for robustifying other graph-embedding objectives.","Editorial inference: because the mapping is a chain of mode-wise cores, the same construction should extend naturally to streaming or time-series tensors, although the paper only lists that as future work.","Editorial inference: the experiments report classification accuracy only; a direct measure of local geometric preservation, such as nearest-neighbor reconstruction error, would test more sharply whether the tensor-train subspace really preserves spatial relations as claimed."],"forward_implications":["Storage cost of the learned projection scales as $O(nIR^2)$, with $n$ modes, maximum mode size $I$, and tensor rank $R$, so the parameter count stays far below the $O(p)$ matrices of PCA and LPP.","The method can run when the data dimension exceeds the training sample size, as demonstrated on 1024-dimensional face patches with 482 training samples, a regime where LPP cannot execute.","On the noisy Extended Yale B data, TTPUDR keeps overall accuracy well above PCA and LPP at both 10% and 20% block noise, and reaches the higher accuracy at a lower reduced dimensionality.","Each tensor-core subproblem is a small Stiefel-manifold optimization, so the algorithm avoids forming the full vectorized data matrix that would be infeasible for ultra-high-dimensional data."],"supporting_citations":[{"why":"Supplies the tensor-train decomposition whose chain of three-way cores is the paper's parameterization.","marker":"[10]"},{"why":"Supplies the locality preserving projections formulation and the graph affinity matrix that TTPUDR tensorizes.","marker":"[12]"},{"why":"Supplies the tensor-train subspace embedding that TTPUDR extends and the orthogonality lemma used for the constraints.","marker":"[7]"},{"why":"Supplies the earlier TT subspace optimization strategy that TTPUDR replaces with a faster per-core scheme.","marker":"[17]"},{"why":"Supplies the Tucker-based tensor LPP baseline whose exponential storage complexity TTPUDR aims to avoid.","marker":"[6]"},{"why":"Supplies the Extended Yale B face dataset used for classification and noise-robustness experiments.","marker":"[18]"},{"why":"Supplies the Indian Pines hyperspectral dataset used to test spatial-structure preservation.","marker":"[19]"},{"why":"Supplies the PCA baseline that TTPUDR must beat in the comparisons.","marker":"[11]"}],"fun_headline_variants":["Tensor-train LPP resists outliers in ultra-high dimensions","Frobenius-norm tensor-train LPP cuts parameters, beats PCA","Ultra-dim reduction via tensor-train LPP robust to noise","Tensor-train parameterization tames high-dim LPP outliers","Robust LPP with tensor-trains for ultra-high-dimensional data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the alternating sweep—recompute the local weights from the current projection, then optimize each tensor core—converges to a minimizer of the stated Frobenius-norm objective, a convergence the paper explicitly says has not been proven.","fun_headline_variants_meta":{"raw":{"variants":["Tensor-train LPP resists outliers in ultra-high dimensions","Frobenius-norm tensor-train LPP cuts parameters, beats PCA","Ultra-dim reduction via tensor-train LPP robust to noise","Tensor-train parameterization tames high-dim LPP outliers","Robust LPP with tensor-trains for ultra-high-dimensional data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1399,"prompt_tokens":947,"completion_tokens":452,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":360}},"tokens_in":563,"tokens_out":452,"duration_ms":4810,"temperature":1.0,"reasoning_tokens":360,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:28:27.232239+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TTPUDR from many random tensor-core initializations on the noisy Yale B setting and compare the final objective values and classification accuracies; if different initializations settle at materially different values or the objective does not stabilize, the alternating reweighted scheme is not solving the claimed optimization problem.","supporting_citations":[{"cited_title":"Tensor-train decomposition,","cited_arxiv_id":null,"evidence_quote":"Supplies the tensor-train decomposition whose chain of three-way cores is the paper's parameterization."},{"cited_title":"Locality preserving projections,","cited_arxiv_id":null,"evidence_quote":"Supplies the locality preserving projections formulation and the graph affinity matrix that TTPUDR tensorizes."},{"cited_title":"Principal Component Analysis with Tensor Train Subspace","cited_arxiv_id":"1803.05026","evidence_quote":"Supplies the tensor-train subspace embedding that TTPUDR extends and the orthogonality lemma used for the constraints."},{"cited_title":"Tensor train neighb orhood pre- serving embedding,","cited_arxiv_id":null,"evidence_quote":"Supplies the earlier TT subspace optimization strategy that TTPUDR replaces with a faster per-core scheme."},{"cited_title":"Tensor embedding methods,","cited_arxiv_id":null,"evidence_quote":"Supplies the Tucker-based tensor LPP baseline whose exponential storage complexity TTPUDR aims to avoid."},{"cited_title":"From fe w to many: Illumination cone models for face recognition under variab le lighting and pose,","cited_arxiv_id":null,"evidence_quote":"Supplies the Extended Yale B face dataset used for classification and noise-robustness experiments."},{"cited_title":"220 band aviris hyperspectral image data set: June 12, 1992 indian pine test site 3,","cited_arxiv_id":null,"evidence_quote":"Supplies the Indian Pines hyperspectral dataset used to test spatial-structure preservation."},{"cited_title":"Liii. on lines and planes of closest fit to sy stems of points in space,","cited_arxiv_id":null,"evidence_quote":"Supplies the PCA baseline that TTPUDR must beat in the comparisons."}],"review_version":1}