{"id":"447ff166-e27d-44c8-8a9d-cc9d428a0d4e","arxiv_id":"2502.09926","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A tensor-pseudoskeleton robust PCA method detects anomalies in space-time data, reporting higher event counts than four baselines on NYC taxi data, with claimed convergence guarantees.","lead":"This paper applies a tensor compression tool called pseudoskeleton decomposition to split New York City taxi trip data into regular patterns and unusual events. The authors report that their method detects more of 20 known urban events than four comparison methods and claim convergence guarantees for the algorithm.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The most load-bearing flaw is internal: Theorem 3's proof replaces the hard-thresholding error ||E(k)||_1 with C sqrt(log dmax) ||S*||_inf despite no support-size bound, so the convergence and error bounds do not follow even under the exact model.","rationale":"The reader's rejection is well supported. My stress test isolates one internal gap rather than a data-model mismatch: even under the exact decomposition and incoherence assumptions, the proof of Theorem 3 does not establish the claimed bound because it never controls the support size of S*. This is more fundamental than asking whether taxi data obey the model, since the theorem's own assumptions are insufficient. I therefore focus the concrete check on a synthetic experiment that varies support size while holding ||S*||_inf fixed. The empirical issue with the Tucker rank (26,6,4,10) on a 24-dimensional first mode and the missing code/parameter values support rejection, but the proof gap alone is decisive. No ad hominem is intended; the manuscript simply does not supply the missing cardinality bound or the derivation connecting ||E(k)||_1 to ||S*||_inf. The reader's weakest_assumption centered on real-data incoherence, hence partial agreement. Since this concern only strengthens the existing REJECT verdict, no verdict change is needed.","tokens_in":8442,"tokens_out":13097,"duration_ms":133466,"concrete_test":"Run Algorithm 1 on synthetic tensors T = L* + S* that satisfy the assumptions of Theorem 3, with L* an incoherent rank-(1,...,1) Tucker tensor and S* supported on s entries of magnitude epsilon, keeping ||S*||_inf fixed and varying s = 10, 100, 1000, 10000. Choose zeta(0) = 2 epsilon sqrt(log dmax) so the threshold condition holds, and set the estimated rank to the true rank. Measure ||L(K)-L*||_F after K iteration counts from Theorem 4. Theorem 3's bound is independent of s, so if the empirical error grows with s (as the unquantified ||E(0)||_1 term predicts), the key inequality in Theorem 3 is false and the theoretical claim collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is that Algorithm 1 converges to L* with the estimation error in Theorem 4. The load-bearing step is in Section III, Theorem 3. Its proof first derives ||E(k)||_1 <= gamma^k ||E(0)||_1 + ... and then, under the condition ||S*||_inf <= zeta(0)/(2 sqrt(log dmax)), asserts ||B(k)E(k)||_F <= C2 sqrt(log dmax) ||S*||_inf. This does not follow. The paper never quantifies the sparsity of S*, so ||E(0)||_1 = ||S*||_1 can be as large as s ||S*||_inf for support size s. Any support with s > C sqrt(log dmax) makes the asserted inequality false; for the NYC tensor D is about 721k, so s can be only a small fraction of entries and still violate it. Lemma 1 repeats the same circular step: it bounds ||B(k)E(k)||_F by ||B(k)||_F (||E(k)||_1 + ||Delta(k)||_F) and then drops the ||E(k)||_1 term. Consequently the contraction recursion and the iteration count in Theorem 4 are unsupported even for synthetic tensors that exactly satisfy T = L* + S* and the incoherence assumptions. This is not a data-model mismatch objection; it is an internal gap in the proof of the paper's own theorem.","agreement_with_reader":"partial"},"referee_report":null,"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take on arXiv:2502.09926.\n\nFor a short paper, this one is easy to summarize: the algorithm is robust tensor CUR from [14] with hard thresholding, applied to anomaly detection on NYC taxi data. The experiments show a modest improvement over baselines, but the theoretical section does not support the convergence claims.\n\nWhat's actually new? Not much. The tensor pseudoskeleton decomposition and its recovery guarantees are quoted from prior work [9,10,14]. The only new content is the empirical application to NYC taxi anomalies and a running-time comparison. That's a routine extension. To the paper's credit, the writing is clear about these debts, and the experiment design is straightforward: a small set of events, a threshold sweep, comparison against four baselines.\n\nThe soft spots are serious. Theorem 2's proof has a dimension mismatch: eU_i = Phi_i U_i is |I_i| x r_i, so eU_i eU_i^T is |I_i| x |I_i|, but the bound is on the spectrum of U_i(I_i,:), which is controlled by eU_i^T eU_i. That can be fixed, but it's sloppy. Theorem 3 is load-bearing and wrong: it asserts ||B(k)E(k)||_F <= C sqrt(log dmax) ||S*||_inf from a bound on ||E(k)||_1, with no support-size bound. ||E(0)||_1 can be s||S*||_inf for support s, so the asserted inequality fails for moderate sparsity. The contraction factor and Theorem 4's iteration count both depend on this step. Theorem 4 also states K = O(log(1/eps)/log(1/rho)) but its proof sets rho^K <= sqrt(log dmax/(rmax dmax)), which doesn't involve eps. These are not minor typos; the main theoretical claims do not follow as written.\n\nReproducibility is also weak. No code is shipped, sampling cardinalities are not given, and the Tucker rank (26,6,4,10) is impossible because the first mode has dimension 24. Table I has no error bars, and the reported gains are one or two detected events out of 20.\n\nIf the authors dropped the theory and presented this as an empirical study, it might be a marginal workshop paper. As is, the central proof gap and the impossible rank make it hard to take seriously. I would not send it to peer review; a desk reject is appropriate.\n\nHope this helps.","headline":"Routine application of robust tensor CUR to NYC taxi anomalies; the convergence proofs have a load-bearing gap and the rank parameter is impossible.","tokens_in":9289,"tokens_out":4538,"would_cite":false,"duration_ms":42090,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69","62H25"],"pacs":[],"model":"deepseek-v4-flash","headline":"Tensor pseudoskeleton decomposition can replace SVD inside tensor robust PCA, yielding a provably convergent anomaly detector that outperforms existing baselines on New York City taxi data.","keywords":["anomaly detection","tensor robust PCA","pseudoskeleton decomposition","tensor CUR","Tucker rank","low-rank plus sparse","hard thresholding","spatiotemporal data"],"falsifier":"A concrete test: build a synthetic tensor $T = L^* + S^*$ with $L^*$ of Tucker rank $(r_1,\\dots,r_n)$ but with one factor row much larger than the rest (so the $\\mu$-incoherence constant is large), and run Algorithm 1 with the theorem's sampling sizes. If the error $\\|L^{(k)} - L^*\\|_F$ does not decrease at the rate $\\rho^k$ claimed in Theorem 3, or if the singular-value bounds in Theorem 2 fail, the central convergence claim is falsified on that class. Conversely, if the method still converges on such tensors, the incoherence assumption is unnecessary and the theorem is not tight.","tokens_in":8215,"feed_emoji":"🚕","tokens_out":8995,"duration_ms":81559,"temperature":0.7,"pith_summary":"The paper's aim is to show that tensor pseudoskeleton decomposition—a sampling-based cousin of SVD—can serve as the low-rank engine of tensor robust PCA, splitting an observed tensor $T$ into a low-Tucker-rank component $L^*$ that carries the regular pattern and a sparse component $S^*$ that carries anomalies. The payoff would be an anomaly detector that is both scalable, because per-mode cost drops from $O(d_i^3)$ to $O(d_i r_i^2 + r_i^3)$, and provably stable, because the paper proves linear convergence of the iterates plus a finite estimation error bound. On New York City taxi arrival data organized as a $24 \\times 7 \\times 53 \\times 81$ tensor, the method is reported to detect more known urban events than LR-STSS, LR-TS, LR-SS, and HoRPCA at every threshold while running faster. The broader claim is that exact low-rank-plus-sparse structure can be recovered without ever forming a full SVD, using only random fiber samples and QR snapshots.","feed_headline":"Tensor pseudoskeleton method beats PCA baselines on taxi anomalies","feed_subtitle":"Splitting the 24×7×53×81 arrival tensor into low-rank pattern plus sparse events finds 20 of 20 known events at top K=3%.","key_machinery":"The central object is the tensor pseudoskeleton (tensor CUR) decomposition: for a tensor $A$ with Tucker ranks $(r_1,\\dots,r_n)$, one selects a subtensor $R=[A]_{I_1,\\dots,I_n}$ and mode-$i$ fibers $C_i=[A_{(i)}]_{:,J_i}$, forms $U_i=[C_i]_{I_i,:}$, and reconstructs $A=R\\times_{i=1}^n (C_i U_i^\\dagger)$ when the ranks agree. Algorithm 1 uses this in each iteration: it hard-thresholds the residual $T-L^{(k)}$ to update the sparse part, then samples fibers of $T-S^{(k+1)}$ to estimate mode-$i$ subspaces by a QR decomposition on the selected rows, and updates $L$ by the mode products $L \\leftarrow L\\times_i (C_i [Q]_{:,:r_i}[R]^\\dagger_{:r_i,:})$. The proof machinery is a matrix concentration inequality (Theorem 2) showing that uniformly sampled rows of a $\\mu$-incoherent factor matrix preserve its singular values, plus a singular-subspace perturbation bound that turns preserved singular values into the contraction factor $\\rho = \\max_i (1 - \\sigma_{\\min}^2(U_i(I_i,:))/2) < 1$. This replaces the cubic $O(d_i^3)$ SVD cost with $O(d_i r_i^2 + r_i^3)$ per mode.","core_discovery":"In the paper's own terms, the claim is that Algorithm 1 computes a sequence $(L^{(k)}, S^{(k)})$ that converges to the true decomposition $T = L^* + S^*$: under a $\\mu$-incoherence condition on the factor matrices of $L^*$ and the sparse bound $\\|S^*\\|_\\infty \\le \\zeta^{(0)}/(2\\sqrt{\\log d_{\\max}})$, Theorem 3 gives $\\|L^{(k+1)} - L^*\\|_F \\le \\rho \\|L^{(k)} - L^*\\|_F + C\\sqrt{\\log d_{\\max}/|I|}\\,\\|S^*\\|_\\infty$ with contraction $\\rho < 1$, and Theorem 4 turns this into an error decomposition with an approximation term $C_1\\sqrt{r_{\\max} d_{\\max}\\log d_{\\max}/|I|}$ and an optimization term $C_2\\|S^*\\|_\\infty\\sqrt{\\log d_{\\max}}$. In experiments, the sparse component $S$ is ranked entrywise and the top-$K\\%$ entries are matched against a compiled list of 20 events; the method achieves 20/20 at the 3% threshold, exceeds the four benchmark methods at all listed thresholds, and has lower running time.","pith_inferences":["Editorial extension: if the incoherence assumption holds only on a subset of rows, adaptive or leverage-score sampling rather than uniform sampling could relax Theorem 2, since the contraction factor is controlled by the smallest sampled singular value.","Editorial extension: adding dense Gaussian noise to the model would introduce an extra term in the recursion similar to $C\\|N\\|_F/\\sqrt{|I|}$; one testable extension is to run Algorithm 1 on synthetic $L^*+S^*+N$ and check whether the empirical error tracks that term as $|I|$ grows.","Editorial extension: the linear contraction suggests a streaming variant in which, once the QR basis is computed, new time slices are projected onto it and thresholded without re-running the full loop; the paper does not test this."],"forward_implications":["On data satisfying the theorem's conditions, the user can choose per-mode sample sizes $|I_i|$ from Corollary 1 to hit a target accuracy $\\epsilon$, and the iteration count $O(\\log(1/\\epsilon)/\\log(1/\\rho))$ gives a concrete stopping rule.","The per-mode cost reduction from $O(d_i^3)$ to $O(d_i r_i^2 + r_i^3)$ makes the method a candidate for tensors with large mode dimensions and small Tucker rank, where full SVD-based tensor robust PCA would be too slow.","Because the sparse tensor $S$ is the anomaly score map, the estimation error bound transfers to detection: entries of $S$ that exceed the threshold are interpretable as localized events, and the recovered count in the NYC experiment is 20 of 20 compiled events at the top 3% threshold.","The method outperforms LR-STSS, LR-TS, LR-SS, and HoRPCA at every listed threshold in the reported experiment, which supports the claim that pseudoskeleton sampling preserves enough structure for downstream ranking."],"supporting_citations":[{"why":"supplies the exact tensor pseudoskeleton reconstruction and the rank-equivalence conditions used in Definition 4.","marker":"[10]"},{"why":"provides the generalized pseudoskeleton decomposition that motivates the sampled-subtensor update.","marker":"[9]"},{"why":"introduces the robust tensor CUR framing and thresholding schedule that Algorithm 1's sparse update builds on.","marker":"[14]"},{"why":"gives the Tucker decomposition, mode-$k$ product, and unfolding definitions used throughout Section I.","marker":"[11]"},{"why":"provides the matrix concentration inequality that yields the sampled singular-value bounds in Theorem 2.","marker":"[22]"},{"why":"supplies the subspace perturbation bound used to turn sampled singular values into the contraction factor.","marker":"[23]"},{"why":"supplies the LR-STSS baseline and the NYC taxi data aggregation and event list used in the experiments.","marker":"[24]"},{"why":"supplies the GLOSS baseline and the compiled event list against which detections are scored.","marker":"[25]"},{"why":"contributes the nonconvex robust PCA thresholding idea behind the hard-thresholding update of $S$.","marker":"[19]"},{"why":"contributes accelerated alternating-projection machinery for robust PCA that informs the alternating update structure.","marker":"[20]"}],"fun_headline_variants":["Tensor pseudoskeleton exposes all 20 NYC taxi anomalies","20/20 taxi anomalies caught by tensor pseudoskeleton method","Tensor pseudoskeleton beats PCA baselines on taxi data","Convergent tensor decomposition detects 20/20 urban events","Pseudoskeleton TRPCA: theory, speed, and 20/20 anomalies"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the observed tensor being exactly a low-Tucker-rank component whose factors spread mass evenly across coordinates (the $\\mu$-incoherence condition) plus a sparse component whose largest entry is below the initial threshold; real data with dense noise, a concentrated low-rank part, or a wrong rank choice fall outside the proof's scope.","fun_headline_variants_meta":{"raw":{"variants":["Tensor pseudoskeleton exposes all 20 NYC taxi anomalies","20/20 taxi anomalies caught by tensor pseudoskeleton method","Tensor pseudoskeleton beats PCA baselines on taxi data","Convergent tensor decomposition detects 20/20 urban events","Pseudoskeleton TRPCA: theory, speed, and 20/20 anomalies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"total_tokens":1594,"prompt_tokens":983,"completion_tokens":611,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":523}},"tokens_in":599,"tokens_out":611,"duration_ms":5838,"temperature":1.0,"reasoning_tokens":523,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T20:02:07.720522+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: build a synthetic tensor $T = L^* + S^*$ with $L^*$ of Tucker rank $(r_1,\\dots,r_n)$ but with one factor row much larger than the rest (so the $\\mu$-incoherence constant is large), and run Algorithm 1 with the theorem's sampling sizes. If the error $\\|L^{(k)} - L^*\\|_F$ does not decrease at the rate $\\rho^k$ claimed in Theorem 3, or if the singular-value bounds in Theorem 2 fail, the central convergence claim is falsified on that class. Conversely, if the method still converges on such tensors, the incoherence assumption is unnecessary and the theorem is not tight.","supporting_citations":[{"cited_title":"Mode-wise tensor decom- positions: Multi-dimensional generalizations of CUR decompositions,","cited_arxiv_id":null,"evidence_quote":"supplies the exact tensor pseudoskeleton reconstruction and the rank-equivalence conditions used in Definition 4."},{"cited_title":"Generalized pseudoskeleton decompositions,","cited_arxiv_id":null,"evidence_quote":"provides the generalized pseudoskeleton decomposition that motivates the sampled-subtensor update."},{"cited_title":"Robust Tensor CUR Decompositions: Rapid Low-Tucker-Rank Tensor Recovery with Sparse Corruption","cited_arxiv_id":"2305.04080","evidence_quote":"introduces the robust tensor CUR framing and thresholding schedule that Algorithm 1's sparse update builds on."},{"cited_title":"An introduction to matrix concentration inequalities,","cited_arxiv_id":null,"evidence_quote":"provides the matrix concentration inequality that yields the sampled singular-value bounds in Theorem 2."},{"cited_title":"Perturbation bounds in connection with singular value decomposition,","cited_arxiv_id":null,"evidence_quote":"supplies the subspace perturbation bound used to turn sampled singular values into the contraction factor."},{"cited_title":"Spatiotemporal group anomaly detection via graph total variation on tensors,","cited_arxiv_id":null,"evidence_quote":"supplies the LR-STSS baseline and the NYC taxi data aggregation and event list used in the experiments."},{"cited_title":"Gloss: Tensor-based anomaly detection in spatiotemporal urban traffic data,","cited_arxiv_id":null,"evidence_quote":"supplies the GLOSS baseline and the compiled event list against which detections are scored."},{"cited_title":"Non- convex robust pca,","cited_arxiv_id":null,"evidence_quote":"contributes the nonconvex robust PCA thresholding idea behind the hard-thresholding update of $S$."},{"cited_title":"Accelerated alternating projections for robust principal component analysis,","cited_arxiv_id":null,"evidence_quote":"contributes accelerated alternating-projection machinery for robust PCA that informs the alternating update structure."}],"review_version":1}